1 //===- ARMISelLowering.cpp - ARM DAG Lowering Implementation --------------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file defines the interfaces that ARM uses to lower LLVM code into a
12 //===----------------------------------------------------------------------===//
14 #include "ARMISelLowering.h"
15 #include "ARMBaseInstrInfo.h"
16 #include "ARMBaseRegisterInfo.h"
17 #include "ARMCallingConv.h"
18 #include "ARMConstantPoolValue.h"
19 #include "ARMMachineFunctionInfo.h"
20 #include "ARMPerfectShuffle.h"
21 #include "ARMRegisterInfo.h"
22 #include "ARMSelectionDAGInfo.h"
23 #include "ARMSubtarget.h"
24 #include "MCTargetDesc/ARMAddressingModes.h"
25 #include "MCTargetDesc/ARMBaseInfo.h"
26 #include "Utils/ARMBaseInfo.h"
27 #include "llvm/ADT/APFloat.h"
28 #include "llvm/ADT/APInt.h"
29 #include "llvm/ADT/ArrayRef.h"
30 #include "llvm/ADT/BitVector.h"
31 #include "llvm/ADT/DenseMap.h"
32 #include "llvm/ADT/STLExtras.h"
33 #include "llvm/ADT/SmallPtrSet.h"
34 #include "llvm/ADT/SmallVector.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/ADT/StringExtras.h"
37 #include "llvm/ADT/StringRef.h"
38 #include "llvm/ADT/StringSwitch.h"
39 #include "llvm/ADT/Triple.h"
40 #include "llvm/ADT/Twine.h"
41 #include "llvm/Analysis/VectorUtils.h"
42 #include "llvm/CodeGen/CallingConvLower.h"
43 #include "llvm/CodeGen/ISDOpcodes.h"
44 #include "llvm/CodeGen/IntrinsicLowering.h"
45 #include "llvm/CodeGen/MachineBasicBlock.h"
46 #include "llvm/CodeGen/MachineConstantPool.h"
47 #include "llvm/CodeGen/MachineFrameInfo.h"
48 #include "llvm/CodeGen/MachineFunction.h"
49 #include "llvm/CodeGen/MachineInstr.h"
50 #include "llvm/CodeGen/MachineInstrBuilder.h"
51 #include "llvm/CodeGen/MachineJumpTableInfo.h"
52 #include "llvm/CodeGen/MachineMemOperand.h"
53 #include "llvm/CodeGen/MachineOperand.h"
54 #include "llvm/CodeGen/MachineRegisterInfo.h"
55 #include "llvm/CodeGen/RuntimeLibcalls.h"
56 #include "llvm/CodeGen/SelectionDAG.h"
57 #include "llvm/CodeGen/SelectionDAGNodes.h"
58 #include "llvm/CodeGen/TargetInstrInfo.h"
59 #include "llvm/CodeGen/TargetLowering.h"
60 #include "llvm/CodeGen/TargetOpcodes.h"
61 #include "llvm/CodeGen/TargetRegisterInfo.h"
62 #include "llvm/CodeGen/TargetSubtargetInfo.h"
63 #include "llvm/CodeGen/ValueTypes.h"
64 #include "llvm/IR/Attributes.h"
65 #include "llvm/IR/CallingConv.h"
66 #include "llvm/IR/Constant.h"
67 #include "llvm/IR/Constants.h"
68 #include "llvm/IR/DataLayout.h"
69 #include "llvm/IR/DebugLoc.h"
70 #include "llvm/IR/DerivedTypes.h"
71 #include "llvm/IR/Function.h"
72 #include "llvm/IR/GlobalAlias.h"
73 #include "llvm/IR/GlobalValue.h"
74 #include "llvm/IR/GlobalVariable.h"
75 #include "llvm/IR/IRBuilder.h"
76 #include "llvm/IR/InlineAsm.h"
77 #include "llvm/IR/Instruction.h"
78 #include "llvm/IR/Instructions.h"
79 #include "llvm/IR/IntrinsicInst.h"
80 #include "llvm/IR/Intrinsics.h"
81 #include "llvm/IR/IntrinsicsARM.h"
82 #include "llvm/IR/Module.h"
83 #include "llvm/IR/PatternMatch.h"
84 #include "llvm/IR/Type.h"
85 #include "llvm/IR/User.h"
86 #include "llvm/IR/Value.h"
87 #include "llvm/MC/MCInstrDesc.h"
88 #include "llvm/MC/MCInstrItineraries.h"
89 #include "llvm/MC/MCRegisterInfo.h"
90 #include "llvm/MC/MCSchedule.h"
91 #include "llvm/Support/AtomicOrdering.h"
92 #include "llvm/Support/BranchProbability.h"
93 #include "llvm/Support/Casting.h"
94 #include "llvm/Support/CodeGen.h"
95 #include "llvm/Support/CommandLine.h"
96 #include "llvm/Support/Compiler.h"
97 #include "llvm/Support/Debug.h"
98 #include "llvm/Support/ErrorHandling.h"
99 #include "llvm/Support/KnownBits.h"
100 #include "llvm/Support/MachineValueType.h"
101 #include "llvm/Support/MathExtras.h"
102 #include "llvm/Support/raw_ostream.h"
103 #include "llvm/Target/TargetMachine.h"
104 #include "llvm/Target/TargetOptions.h"
116 using namespace llvm;
117 using namespace llvm::PatternMatch;
119 #define DEBUG_TYPE "arm-isel"
121 STATISTIC(NumTailCalls, "Number of tail calls");
122 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
123 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
124 STATISTIC(NumConstpoolPromoted,
125 "Number of constants with their storage promoted into constant pools");
128 ARMInterworking("arm-interworking", cl::Hidden,
129 cl::desc("Enable / disable ARM interworking (for debugging only)"),
132 static cl::opt<bool> EnableConstpoolPromotion(
133 "arm-promote-constant", cl::Hidden,
134 cl::desc("Enable / disable promotion of unnamed_addr constants into "
136 cl::init(false)); // FIXME: set to true by default once PR32780 is fixed
137 static cl::opt<unsigned> ConstpoolPromotionMaxSize(
138 "arm-promote-constant-max-size", cl::Hidden,
139 cl::desc("Maximum size of constant to promote into a constant pool"),
141 static cl::opt<unsigned> ConstpoolPromotionMaxTotal(
142 "arm-promote-constant-max-total", cl::Hidden,
143 cl::desc("Maximum size of ALL constants to promote into a constant pool"),
146 static cl::opt<unsigned>
147 MVEMaxSupportedInterleaveFactor("mve-max-interleave-factor", cl::Hidden,
148 cl::desc("Maximum interleave factor for MVE VLDn to generate."),
151 // The APCS parameter registers.
152 static const MCPhysReg GPRArgRegs[] = {
153 ARM::R0, ARM::R1, ARM::R2, ARM::R3
156 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
157 MVT PromotedBitwiseVT) {
158 if (VT != PromotedLdStVT) {
159 setOperationAction(ISD::LOAD, VT, Promote);
160 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
162 setOperationAction(ISD::STORE, VT, Promote);
163 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
166 MVT ElemTy = VT.getVectorElementType();
167 if (ElemTy != MVT::f64)
168 setOperationAction(ISD::SETCC, VT, Custom);
169 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
170 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
171 if (ElemTy == MVT::i32) {
172 setOperationAction(ISD::SINT_TO_FP, VT, Custom);
173 setOperationAction(ISD::UINT_TO_FP, VT, Custom);
174 setOperationAction(ISD::FP_TO_SINT, VT, Custom);
175 setOperationAction(ISD::FP_TO_UINT, VT, Custom);
177 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
178 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
179 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
180 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
182 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
183 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
184 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);
185 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
186 setOperationAction(ISD::SELECT, VT, Expand);
187 setOperationAction(ISD::SELECT_CC, VT, Expand);
188 setOperationAction(ISD::VSELECT, VT, Expand);
189 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
190 if (VT.isInteger()) {
191 setOperationAction(ISD::SHL, VT, Custom);
192 setOperationAction(ISD::SRA, VT, Custom);
193 setOperationAction(ISD::SRL, VT, Custom);
196 // Promote all bit-wise operations.
197 if (VT.isInteger() && VT != PromotedBitwiseVT) {
198 setOperationAction(ISD::AND, VT, Promote);
199 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
200 setOperationAction(ISD::OR, VT, Promote);
201 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
202 setOperationAction(ISD::XOR, VT, Promote);
203 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
206 // Neon does not support vector divide/remainder operations.
207 setOperationAction(ISD::SDIV, VT, Expand);
208 setOperationAction(ISD::UDIV, VT, Expand);
209 setOperationAction(ISD::FDIV, VT, Expand);
210 setOperationAction(ISD::SREM, VT, Expand);
211 setOperationAction(ISD::UREM, VT, Expand);
212 setOperationAction(ISD::FREM, VT, Expand);
214 if (!VT.isFloatingPoint() &&
215 VT != MVT::v2i64 && VT != MVT::v1i64)
216 for (auto Opcode : {ISD::ABS, ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX})
217 setOperationAction(Opcode, VT, Legal);
218 if (!VT.isFloatingPoint())
219 for (auto Opcode : {ISD::SADDSAT, ISD::UADDSAT, ISD::SSUBSAT, ISD::USUBSAT})
220 setOperationAction(Opcode, VT, Legal);
223 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
224 addRegisterClass(VT, &ARM::DPRRegClass);
225 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
228 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
229 addRegisterClass(VT, &ARM::DPairRegClass);
230 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
233 void ARMTargetLowering::setAllExpand(MVT VT) {
234 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
235 setOperationAction(Opc, VT, Expand);
237 // We support these really simple operations even on types where all
238 // the actual arithmetic has to be broken down into simpler
239 // operations or turned into library calls.
240 setOperationAction(ISD::BITCAST, VT, Legal);
241 setOperationAction(ISD::LOAD, VT, Legal);
242 setOperationAction(ISD::STORE, VT, Legal);
243 setOperationAction(ISD::UNDEF, VT, Legal);
246 void ARMTargetLowering::addAllExtLoads(const MVT From, const MVT To,
247 LegalizeAction Action) {
248 setLoadExtAction(ISD::EXTLOAD, From, To, Action);
249 setLoadExtAction(ISD::ZEXTLOAD, From, To, Action);
250 setLoadExtAction(ISD::SEXTLOAD, From, To, Action);
253 void ARMTargetLowering::addMVEVectorTypes(bool HasMVEFP) {
254 const MVT IntTypes[] = { MVT::v16i8, MVT::v8i16, MVT::v4i32 };
256 for (auto VT : IntTypes) {
257 addRegisterClass(VT, &ARM::MQPRRegClass);
258 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
259 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
260 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
261 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
262 setOperationAction(ISD::SHL, VT, Custom);
263 setOperationAction(ISD::SRA, VT, Custom);
264 setOperationAction(ISD::SRL, VT, Custom);
265 setOperationAction(ISD::SMIN, VT, Legal);
266 setOperationAction(ISD::SMAX, VT, Legal);
267 setOperationAction(ISD::UMIN, VT, Legal);
268 setOperationAction(ISD::UMAX, VT, Legal);
269 setOperationAction(ISD::ABS, VT, Legal);
270 setOperationAction(ISD::SETCC, VT, Custom);
271 setOperationAction(ISD::MLOAD, VT, Custom);
272 setOperationAction(ISD::MSTORE, VT, Legal);
273 setOperationAction(ISD::CTLZ, VT, Legal);
274 setOperationAction(ISD::CTTZ, VT, Custom);
275 setOperationAction(ISD::BITREVERSE, VT, Legal);
276 setOperationAction(ISD::BSWAP, VT, Legal);
277 setOperationAction(ISD::SADDSAT, VT, Legal);
278 setOperationAction(ISD::UADDSAT, VT, Legal);
279 setOperationAction(ISD::SSUBSAT, VT, Legal);
280 setOperationAction(ISD::USUBSAT, VT, Legal);
282 // No native support for these.
283 setOperationAction(ISD::UDIV, VT, Expand);
284 setOperationAction(ISD::SDIV, VT, Expand);
285 setOperationAction(ISD::UREM, VT, Expand);
286 setOperationAction(ISD::SREM, VT, Expand);
287 setOperationAction(ISD::CTPOP, VT, Expand);
290 setOperationAction(ISD::VECREDUCE_ADD, VT, Legal);
291 setOperationAction(ISD::VECREDUCE_SMAX, VT, Legal);
292 setOperationAction(ISD::VECREDUCE_UMAX, VT, Legal);
293 setOperationAction(ISD::VECREDUCE_SMIN, VT, Legal);
294 setOperationAction(ISD::VECREDUCE_UMIN, VT, Legal);
297 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
298 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
299 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
300 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
303 // Pre and Post inc are supported on loads and stores
304 for (unsigned im = (unsigned)ISD::PRE_INC;
305 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
306 setIndexedLoadAction(im, VT, Legal);
307 setIndexedStoreAction(im, VT, Legal);
308 setIndexedMaskedLoadAction(im, VT, Legal);
309 setIndexedMaskedStoreAction(im, VT, Legal);
313 const MVT FloatTypes[] = { MVT::v8f16, MVT::v4f32 };
314 for (auto VT : FloatTypes) {
315 addRegisterClass(VT, &ARM::MQPRRegClass);
319 // These are legal or custom whether we have MVE.fp or not
320 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
321 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
322 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getVectorElementType(), Custom);
323 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
324 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
325 setOperationAction(ISD::BUILD_VECTOR, VT.getVectorElementType(), Custom);
326 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Legal);
327 setOperationAction(ISD::SETCC, VT, Custom);
328 setOperationAction(ISD::MLOAD, VT, Custom);
329 setOperationAction(ISD::MSTORE, VT, Legal);
331 // Pre and Post inc are supported on loads and stores
332 for (unsigned im = (unsigned)ISD::PRE_INC;
333 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
334 setIndexedLoadAction(im, VT, Legal);
335 setIndexedStoreAction(im, VT, Legal);
336 setIndexedMaskedLoadAction(im, VT, Legal);
337 setIndexedMaskedStoreAction(im, VT, Legal);
341 setOperationAction(ISD::FMINNUM, VT, Legal);
342 setOperationAction(ISD::FMAXNUM, VT, Legal);
343 setOperationAction(ISD::FROUND, VT, Legal);
345 // No native support for these.
346 setOperationAction(ISD::FDIV, VT, Expand);
347 setOperationAction(ISD::FREM, VT, Expand);
348 setOperationAction(ISD::FSQRT, VT, Expand);
349 setOperationAction(ISD::FSIN, VT, Expand);
350 setOperationAction(ISD::FCOS, VT, Expand);
351 setOperationAction(ISD::FPOW, VT, Expand);
352 setOperationAction(ISD::FLOG, VT, Expand);
353 setOperationAction(ISD::FLOG2, VT, Expand);
354 setOperationAction(ISD::FLOG10, VT, Expand);
355 setOperationAction(ISD::FEXP, VT, Expand);
356 setOperationAction(ISD::FEXP2, VT, Expand);
357 setOperationAction(ISD::FNEARBYINT, VT, Expand);
361 // We 'support' these types up to bitcast/load/store level, regardless of
362 // MVE integer-only / float support. Only doing FP data processing on the FP
363 // vector types is inhibited at integer-only level.
364 const MVT LongTypes[] = { MVT::v2i64, MVT::v2f64 };
365 for (auto VT : LongTypes) {
366 addRegisterClass(VT, &ARM::MQPRRegClass);
368 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
369 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
370 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
372 // We can do bitwise operations on v2i64 vectors
373 setOperationAction(ISD::AND, MVT::v2i64, Legal);
374 setOperationAction(ISD::OR, MVT::v2i64, Legal);
375 setOperationAction(ISD::XOR, MVT::v2i64, Legal);
377 // It is legal to extload from v4i8 to v4i16 or v4i32.
378 addAllExtLoads(MVT::v8i16, MVT::v8i8, Legal);
379 addAllExtLoads(MVT::v4i32, MVT::v4i16, Legal);
380 addAllExtLoads(MVT::v4i32, MVT::v4i8, Legal);
382 // It is legal to sign extend from v4i8/v4i16 to v4i32 or v8i8 to v8i16.
383 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i8, Legal);
384 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Legal);
385 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i32, Legal);
386 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v8i8, Legal);
387 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v8i16, Legal);
389 // Some truncating stores are legal too.
390 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Legal);
391 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Legal);
392 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Legal);
394 // Pre and Post inc on these are legal, given the correct extends
395 for (unsigned im = (unsigned)ISD::PRE_INC;
396 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
397 for (auto VT : {MVT::v8i8, MVT::v4i8, MVT::v4i16}) {
398 setIndexedLoadAction(im, VT, Legal);
399 setIndexedStoreAction(im, VT, Legal);
400 setIndexedMaskedLoadAction(im, VT, Legal);
401 setIndexedMaskedStoreAction(im, VT, Legal);
406 const MVT pTypes[] = {MVT::v16i1, MVT::v8i1, MVT::v4i1};
407 for (auto VT : pTypes) {
408 addRegisterClass(VT, &ARM::VCCRRegClass);
409 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
410 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
411 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
412 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
413 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
414 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
415 setOperationAction(ISD::SETCC, VT, Custom);
416 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
417 setOperationAction(ISD::LOAD, VT, Custom);
418 setOperationAction(ISD::STORE, VT, Custom);
422 ARMTargetLowering::ARMTargetLowering(const TargetMachine &TM,
423 const ARMSubtarget &STI)
424 : TargetLowering(TM), Subtarget(&STI) {
425 RegInfo = Subtarget->getRegisterInfo();
426 Itins = Subtarget->getInstrItineraryData();
428 setBooleanContents(ZeroOrOneBooleanContent);
429 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
431 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetIOS() &&
432 !Subtarget->isTargetWatchOS()) {
433 bool IsHFTarget = TM.Options.FloatABIType == FloatABI::Hard;
434 for (int LCID = 0; LCID < RTLIB::UNKNOWN_LIBCALL; ++LCID)
435 setLibcallCallingConv(static_cast<RTLIB::Libcall>(LCID),
436 IsHFTarget ? CallingConv::ARM_AAPCS_VFP
437 : CallingConv::ARM_AAPCS);
440 if (Subtarget->isTargetMachO()) {
441 // Uses VFP for Thumb libfuncs if available.
442 if (Subtarget->isThumb() && Subtarget->hasVFP2Base() &&
443 Subtarget->hasARMOps() && !Subtarget->useSoftFloat()) {
444 static const struct {
445 const RTLIB::Libcall Op;
446 const char * const Name;
447 const ISD::CondCode Cond;
449 // Single-precision floating-point arithmetic.
450 { RTLIB::ADD_F32, "__addsf3vfp", ISD::SETCC_INVALID },
451 { RTLIB::SUB_F32, "__subsf3vfp", ISD::SETCC_INVALID },
452 { RTLIB::MUL_F32, "__mulsf3vfp", ISD::SETCC_INVALID },
453 { RTLIB::DIV_F32, "__divsf3vfp", ISD::SETCC_INVALID },
455 // Double-precision floating-point arithmetic.
456 { RTLIB::ADD_F64, "__adddf3vfp", ISD::SETCC_INVALID },
457 { RTLIB::SUB_F64, "__subdf3vfp", ISD::SETCC_INVALID },
458 { RTLIB::MUL_F64, "__muldf3vfp", ISD::SETCC_INVALID },
459 { RTLIB::DIV_F64, "__divdf3vfp", ISD::SETCC_INVALID },
461 // Single-precision comparisons.
462 { RTLIB::OEQ_F32, "__eqsf2vfp", ISD::SETNE },
463 { RTLIB::UNE_F32, "__nesf2vfp", ISD::SETNE },
464 { RTLIB::OLT_F32, "__ltsf2vfp", ISD::SETNE },
465 { RTLIB::OLE_F32, "__lesf2vfp", ISD::SETNE },
466 { RTLIB::OGE_F32, "__gesf2vfp", ISD::SETNE },
467 { RTLIB::OGT_F32, "__gtsf2vfp", ISD::SETNE },
468 { RTLIB::UO_F32, "__unordsf2vfp", ISD::SETNE },
470 // Double-precision comparisons.
471 { RTLIB::OEQ_F64, "__eqdf2vfp", ISD::SETNE },
472 { RTLIB::UNE_F64, "__nedf2vfp", ISD::SETNE },
473 { RTLIB::OLT_F64, "__ltdf2vfp", ISD::SETNE },
474 { RTLIB::OLE_F64, "__ledf2vfp", ISD::SETNE },
475 { RTLIB::OGE_F64, "__gedf2vfp", ISD::SETNE },
476 { RTLIB::OGT_F64, "__gtdf2vfp", ISD::SETNE },
477 { RTLIB::UO_F64, "__unorddf2vfp", ISD::SETNE },
479 // Floating-point to integer conversions.
480 // i64 conversions are done via library routines even when generating VFP
481 // instructions, so use the same ones.
482 { RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp", ISD::SETCC_INVALID },
483 { RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp", ISD::SETCC_INVALID },
484 { RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp", ISD::SETCC_INVALID },
485 { RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp", ISD::SETCC_INVALID },
487 // Conversions between floating types.
488 { RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp", ISD::SETCC_INVALID },
489 { RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp", ISD::SETCC_INVALID },
491 // Integer to floating-point conversions.
492 // i64 conversions are done via library routines even when generating VFP
493 // instructions, so use the same ones.
494 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
495 // e.g., __floatunsidf vs. __floatunssidfvfp.
496 { RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp", ISD::SETCC_INVALID },
497 { RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp", ISD::SETCC_INVALID },
498 { RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp", ISD::SETCC_INVALID },
499 { RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp", ISD::SETCC_INVALID },
502 for (const auto &LC : LibraryCalls) {
503 setLibcallName(LC.Op, LC.Name);
504 if (LC.Cond != ISD::SETCC_INVALID)
505 setCmpLibcallCC(LC.Op, LC.Cond);
510 // These libcalls are not available in 32-bit.
511 setLibcallName(RTLIB::SHL_I128, nullptr);
512 setLibcallName(RTLIB::SRL_I128, nullptr);
513 setLibcallName(RTLIB::SRA_I128, nullptr);
516 if (Subtarget->isAAPCS_ABI() &&
517 (Subtarget->isTargetAEABI() || Subtarget->isTargetGNUAEABI() ||
518 Subtarget->isTargetMuslAEABI() || Subtarget->isTargetAndroid())) {
519 static const struct {
520 const RTLIB::Libcall Op;
521 const char * const Name;
522 const CallingConv::ID CC;
523 const ISD::CondCode Cond;
525 // Double-precision floating-point arithmetic helper functions
526 // RTABI chapter 4.1.2, Table 2
527 { RTLIB::ADD_F64, "__aeabi_dadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
528 { RTLIB::DIV_F64, "__aeabi_ddiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
529 { RTLIB::MUL_F64, "__aeabi_dmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
530 { RTLIB::SUB_F64, "__aeabi_dsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
532 // Double-precision floating-point comparison helper functions
533 // RTABI chapter 4.1.2, Table 3
534 { RTLIB::OEQ_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
535 { RTLIB::UNE_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
536 { RTLIB::OLT_F64, "__aeabi_dcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
537 { RTLIB::OLE_F64, "__aeabi_dcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
538 { RTLIB::OGE_F64, "__aeabi_dcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
539 { RTLIB::OGT_F64, "__aeabi_dcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
540 { RTLIB::UO_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
542 // Single-precision floating-point arithmetic helper functions
543 // RTABI chapter 4.1.2, Table 4
544 { RTLIB::ADD_F32, "__aeabi_fadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
545 { RTLIB::DIV_F32, "__aeabi_fdiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
546 { RTLIB::MUL_F32, "__aeabi_fmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
547 { RTLIB::SUB_F32, "__aeabi_fsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
549 // Single-precision floating-point comparison helper functions
550 // RTABI chapter 4.1.2, Table 5
551 { RTLIB::OEQ_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
552 { RTLIB::UNE_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
553 { RTLIB::OLT_F32, "__aeabi_fcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
554 { RTLIB::OLE_F32, "__aeabi_fcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
555 { RTLIB::OGE_F32, "__aeabi_fcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
556 { RTLIB::OGT_F32, "__aeabi_fcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
557 { RTLIB::UO_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
559 // Floating-point to integer conversions.
560 // RTABI chapter 4.1.2, Table 6
561 { RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
562 { RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
563 { RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
564 { RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
565 { RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
566 { RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
567 { RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
568 { RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
570 // Conversions between floating types.
571 // RTABI chapter 4.1.2, Table 7
572 { RTLIB::FPROUND_F64_F32, "__aeabi_d2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
573 { RTLIB::FPROUND_F64_F16, "__aeabi_d2h", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
574 { RTLIB::FPEXT_F32_F64, "__aeabi_f2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
576 // Integer to floating-point conversions.
577 // RTABI chapter 4.1.2, Table 8
578 { RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
579 { RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
580 { RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
581 { RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
582 { RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
583 { RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
584 { RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
585 { RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
587 // Long long helper functions
588 // RTABI chapter 4.2, Table 9
589 { RTLIB::MUL_I64, "__aeabi_lmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
590 { RTLIB::SHL_I64, "__aeabi_llsl", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
591 { RTLIB::SRL_I64, "__aeabi_llsr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
592 { RTLIB::SRA_I64, "__aeabi_lasr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
594 // Integer division functions
595 // RTABI chapter 4.3.1
596 { RTLIB::SDIV_I8, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
597 { RTLIB::SDIV_I16, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
598 { RTLIB::SDIV_I32, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
599 { RTLIB::SDIV_I64, "__aeabi_ldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
600 { RTLIB::UDIV_I8, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
601 { RTLIB::UDIV_I16, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
602 { RTLIB::UDIV_I32, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
603 { RTLIB::UDIV_I64, "__aeabi_uldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
606 for (const auto &LC : LibraryCalls) {
607 setLibcallName(LC.Op, LC.Name);
608 setLibcallCallingConv(LC.Op, LC.CC);
609 if (LC.Cond != ISD::SETCC_INVALID)
610 setCmpLibcallCC(LC.Op, LC.Cond);
613 // EABI dependent RTLIB
614 if (TM.Options.EABIVersion == EABI::EABI4 ||
615 TM.Options.EABIVersion == EABI::EABI5) {
616 static const struct {
617 const RTLIB::Libcall Op;
618 const char *const Name;
619 const CallingConv::ID CC;
620 const ISD::CondCode Cond;
621 } MemOpsLibraryCalls[] = {
623 // RTABI chapter 4.3.4
624 { RTLIB::MEMCPY, "__aeabi_memcpy", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
625 { RTLIB::MEMMOVE, "__aeabi_memmove", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
626 { RTLIB::MEMSET, "__aeabi_memset", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
629 for (const auto &LC : MemOpsLibraryCalls) {
630 setLibcallName(LC.Op, LC.Name);
631 setLibcallCallingConv(LC.Op, LC.CC);
632 if (LC.Cond != ISD::SETCC_INVALID)
633 setCmpLibcallCC(LC.Op, LC.Cond);
638 if (Subtarget->isTargetWindows()) {
639 static const struct {
640 const RTLIB::Libcall Op;
641 const char * const Name;
642 const CallingConv::ID CC;
644 { RTLIB::FPTOSINT_F32_I64, "__stoi64", CallingConv::ARM_AAPCS_VFP },
645 { RTLIB::FPTOSINT_F64_I64, "__dtoi64", CallingConv::ARM_AAPCS_VFP },
646 { RTLIB::FPTOUINT_F32_I64, "__stou64", CallingConv::ARM_AAPCS_VFP },
647 { RTLIB::FPTOUINT_F64_I64, "__dtou64", CallingConv::ARM_AAPCS_VFP },
648 { RTLIB::SINTTOFP_I64_F32, "__i64tos", CallingConv::ARM_AAPCS_VFP },
649 { RTLIB::SINTTOFP_I64_F64, "__i64tod", CallingConv::ARM_AAPCS_VFP },
650 { RTLIB::UINTTOFP_I64_F32, "__u64tos", CallingConv::ARM_AAPCS_VFP },
651 { RTLIB::UINTTOFP_I64_F64, "__u64tod", CallingConv::ARM_AAPCS_VFP },
654 for (const auto &LC : LibraryCalls) {
655 setLibcallName(LC.Op, LC.Name);
656 setLibcallCallingConv(LC.Op, LC.CC);
660 // Use divmod compiler-rt calls for iOS 5.0 and later.
661 if (Subtarget->isTargetMachO() &&
662 !(Subtarget->isTargetIOS() &&
663 Subtarget->getTargetTriple().isOSVersionLT(5, 0))) {
664 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
665 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
668 // The half <-> float conversion functions are always soft-float on
669 // non-watchos platforms, but are needed for some targets which use a
670 // hard-float calling convention by default.
671 if (!Subtarget->isTargetWatchABI()) {
672 if (Subtarget->isAAPCS_ABI()) {
673 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_AAPCS);
674 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_AAPCS);
675 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_AAPCS);
677 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_APCS);
678 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_APCS);
679 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_APCS);
683 // In EABI, these functions have an __aeabi_ prefix, but in GNUEABI they have
684 // a __gnu_ prefix (which is the default).
685 if (Subtarget->isTargetAEABI()) {
686 static const struct {
687 const RTLIB::Libcall Op;
688 const char * const Name;
689 const CallingConv::ID CC;
691 { RTLIB::FPROUND_F32_F16, "__aeabi_f2h", CallingConv::ARM_AAPCS },
692 { RTLIB::FPROUND_F64_F16, "__aeabi_d2h", CallingConv::ARM_AAPCS },
693 { RTLIB::FPEXT_F16_F32, "__aeabi_h2f", CallingConv::ARM_AAPCS },
696 for (const auto &LC : LibraryCalls) {
697 setLibcallName(LC.Op, LC.Name);
698 setLibcallCallingConv(LC.Op, LC.CC);
702 if (Subtarget->isThumb1Only())
703 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
705 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
707 if (!Subtarget->useSoftFloat() && !Subtarget->isThumb1Only() &&
708 Subtarget->hasFPRegs()) {
709 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
710 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
711 if (!Subtarget->hasVFP2Base())
712 setAllExpand(MVT::f32);
713 if (!Subtarget->hasFP64())
714 setAllExpand(MVT::f64);
717 if (Subtarget->hasFullFP16()) {
718 addRegisterClass(MVT::f16, &ARM::HPRRegClass);
719 setOperationAction(ISD::BITCAST, MVT::i16, Custom);
720 setOperationAction(ISD::BITCAST, MVT::i32, Custom);
721 setOperationAction(ISD::BITCAST, MVT::f16, Custom);
723 setOperationAction(ISD::FMINNUM, MVT::f16, Legal);
724 setOperationAction(ISD::FMAXNUM, MVT::f16, Legal);
727 for (MVT VT : MVT::fixedlen_vector_valuetypes()) {
728 for (MVT InnerVT : MVT::fixedlen_vector_valuetypes()) {
729 setTruncStoreAction(VT, InnerVT, Expand);
730 addAllExtLoads(VT, InnerVT, Expand);
733 setOperationAction(ISD::MULHS, VT, Expand);
734 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
735 setOperationAction(ISD::MULHU, VT, Expand);
736 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
738 setOperationAction(ISD::BSWAP, VT, Expand);
741 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
742 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
744 setOperationAction(ISD::READ_REGISTER, MVT::i64, Custom);
745 setOperationAction(ISD::WRITE_REGISTER, MVT::i64, Custom);
747 if (Subtarget->hasMVEIntegerOps())
748 addMVEVectorTypes(Subtarget->hasMVEFloatOps());
750 // Combine low-overhead loop intrinsics so that we can lower i1 types.
751 if (Subtarget->hasLOB()) {
752 setTargetDAGCombine(ISD::BRCOND);
753 setTargetDAGCombine(ISD::BR_CC);
756 if (Subtarget->hasNEON()) {
757 addDRTypeForNEON(MVT::v2f32);
758 addDRTypeForNEON(MVT::v8i8);
759 addDRTypeForNEON(MVT::v4i16);
760 addDRTypeForNEON(MVT::v2i32);
761 addDRTypeForNEON(MVT::v1i64);
763 addQRTypeForNEON(MVT::v4f32);
764 addQRTypeForNEON(MVT::v2f64);
765 addQRTypeForNEON(MVT::v16i8);
766 addQRTypeForNEON(MVT::v8i16);
767 addQRTypeForNEON(MVT::v4i32);
768 addQRTypeForNEON(MVT::v2i64);
770 if (Subtarget->hasFullFP16()) {
771 addQRTypeForNEON(MVT::v8f16);
772 addDRTypeForNEON(MVT::v4f16);
776 if (Subtarget->hasMVEIntegerOps() || Subtarget->hasNEON()) {
777 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
778 // none of Neon, MVE or VFP supports any arithmetic operations on it.
779 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
780 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
781 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
782 // FIXME: Code duplication: FDIV and FREM are expanded always, see
783 // ARMTargetLowering::addTypeForNEON method for details.
784 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
785 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
786 // FIXME: Create unittest.
787 // In another words, find a way when "copysign" appears in DAG with vector
789 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
790 // FIXME: Code duplication: SETCC has custom operation action, see
791 // ARMTargetLowering::addTypeForNEON method for details.
792 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
793 // FIXME: Create unittest for FNEG and for FABS.
794 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
795 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
796 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
797 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
798 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
799 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
800 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
801 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
802 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
803 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
804 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
805 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
806 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
807 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
808 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
809 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
810 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
811 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
814 if (Subtarget->hasNEON()) {
815 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
816 // supported for v4f32.
817 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
818 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
819 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
820 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
821 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
822 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
823 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
824 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
825 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
826 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
827 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
828 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
829 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
830 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
832 // Mark v2f32 intrinsics.
833 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
834 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
835 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
836 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
837 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
838 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
839 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
840 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
841 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
842 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
843 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
844 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
845 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
846 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
848 // Neon does not support some operations on v1i64 and v2i64 types.
849 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
850 // Custom handling for some quad-vector types to detect VMULL.
851 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
852 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
853 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
854 // Custom handling for some vector types to avoid expensive expansions
855 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
856 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
857 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
858 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
859 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
860 // a destination type that is wider than the source, and nor does
861 // it have a FP_TO_[SU]INT instruction with a narrower destination than
863 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
864 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Custom);
865 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
866 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
867 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
868 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Custom);
869 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
870 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Custom);
872 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
873 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
875 // NEON does not have single instruction CTPOP for vectors with element
876 // types wider than 8-bits. However, custom lowering can leverage the
877 // v8i8/v16i8 vcnt instruction.
878 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
879 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
880 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
881 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
882 setOperationAction(ISD::CTPOP, MVT::v1i64, Custom);
883 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
885 setOperationAction(ISD::CTLZ, MVT::v1i64, Expand);
886 setOperationAction(ISD::CTLZ, MVT::v2i64, Expand);
888 // NEON does not have single instruction CTTZ for vectors.
889 setOperationAction(ISD::CTTZ, MVT::v8i8, Custom);
890 setOperationAction(ISD::CTTZ, MVT::v4i16, Custom);
891 setOperationAction(ISD::CTTZ, MVT::v2i32, Custom);
892 setOperationAction(ISD::CTTZ, MVT::v1i64, Custom);
894 setOperationAction(ISD::CTTZ, MVT::v16i8, Custom);
895 setOperationAction(ISD::CTTZ, MVT::v8i16, Custom);
896 setOperationAction(ISD::CTTZ, MVT::v4i32, Custom);
897 setOperationAction(ISD::CTTZ, MVT::v2i64, Custom);
899 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i8, Custom);
900 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i16, Custom);
901 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i32, Custom);
902 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v1i64, Custom);
904 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i8, Custom);
905 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i16, Custom);
906 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
907 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i64, Custom);
909 // NEON only has FMA instructions as of VFP4.
910 if (!Subtarget->hasVFP4Base()) {
911 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
912 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
915 setTargetDAGCombine(ISD::INTRINSIC_VOID);
916 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
917 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
918 setTargetDAGCombine(ISD::SHL);
919 setTargetDAGCombine(ISD::SRL);
920 setTargetDAGCombine(ISD::SRA);
921 setTargetDAGCombine(ISD::FP_TO_SINT);
922 setTargetDAGCombine(ISD::FP_TO_UINT);
923 setTargetDAGCombine(ISD::FDIV);
924 setTargetDAGCombine(ISD::LOAD);
926 // It is legal to extload from v4i8 to v4i16 or v4i32.
927 for (MVT Ty : {MVT::v8i8, MVT::v4i8, MVT::v2i8, MVT::v4i16, MVT::v2i16,
929 for (MVT VT : MVT::integer_fixedlen_vector_valuetypes()) {
930 setLoadExtAction(ISD::EXTLOAD, VT, Ty, Legal);
931 setLoadExtAction(ISD::ZEXTLOAD, VT, Ty, Legal);
932 setLoadExtAction(ISD::SEXTLOAD, VT, Ty, Legal);
937 if (Subtarget->hasNEON() || Subtarget->hasMVEIntegerOps()) {
938 setTargetDAGCombine(ISD::BUILD_VECTOR);
939 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
940 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
941 setTargetDAGCombine(ISD::STORE);
942 setTargetDAGCombine(ISD::SIGN_EXTEND);
943 setTargetDAGCombine(ISD::ZERO_EXTEND);
944 setTargetDAGCombine(ISD::ANY_EXTEND);
947 if (!Subtarget->hasFP64()) {
948 // When targeting a floating-point unit with only single-precision
949 // operations, f64 is legal for the few double-precision instructions which
950 // are present However, no double-precision operations other than moves,
951 // loads and stores are provided by the hardware.
952 setOperationAction(ISD::FADD, MVT::f64, Expand);
953 setOperationAction(ISD::FSUB, MVT::f64, Expand);
954 setOperationAction(ISD::FMUL, MVT::f64, Expand);
955 setOperationAction(ISD::FMA, MVT::f64, Expand);
956 setOperationAction(ISD::FDIV, MVT::f64, Expand);
957 setOperationAction(ISD::FREM, MVT::f64, Expand);
958 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
959 setOperationAction(ISD::FGETSIGN, MVT::f64, Expand);
960 setOperationAction(ISD::FNEG, MVT::f64, Expand);
961 setOperationAction(ISD::FABS, MVT::f64, Expand);
962 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
963 setOperationAction(ISD::FSIN, MVT::f64, Expand);
964 setOperationAction(ISD::FCOS, MVT::f64, Expand);
965 setOperationAction(ISD::FPOW, MVT::f64, Expand);
966 setOperationAction(ISD::FLOG, MVT::f64, Expand);
967 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
968 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
969 setOperationAction(ISD::FEXP, MVT::f64, Expand);
970 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
971 setOperationAction(ISD::FCEIL, MVT::f64, Expand);
972 setOperationAction(ISD::FTRUNC, MVT::f64, Expand);
973 setOperationAction(ISD::FRINT, MVT::f64, Expand);
974 setOperationAction(ISD::FNEARBYINT, MVT::f64, Expand);
975 setOperationAction(ISD::FFLOOR, MVT::f64, Expand);
976 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
977 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
978 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
979 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
980 setOperationAction(ISD::FP_TO_SINT, MVT::f64, Custom);
981 setOperationAction(ISD::FP_TO_UINT, MVT::f64, Custom);
982 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
983 setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i32, Custom);
984 setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i32, Custom);
985 setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::f64, Custom);
986 setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::f64, Custom);
987 setOperationAction(ISD::STRICT_FP_ROUND, MVT::f32, Custom);
990 if (!Subtarget->hasFP64() || !Subtarget->hasFPARMv8Base()) {
991 setOperationAction(ISD::FP_EXTEND, MVT::f64, Custom);
992 setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f64, Custom);
993 if (Subtarget->hasFullFP16()) {
994 setOperationAction(ISD::FP_ROUND, MVT::f16, Custom);
995 setOperationAction(ISD::STRICT_FP_ROUND, MVT::f16, Custom);
999 if (!Subtarget->hasFP16()) {
1000 setOperationAction(ISD::FP_EXTEND, MVT::f32, Custom);
1001 setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f32, Custom);
1004 computeRegisterProperties(Subtarget->getRegisterInfo());
1006 // ARM does not have floating-point extending loads.
1007 for (MVT VT : MVT::fp_valuetypes()) {
1008 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
1009 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
1012 // ... or truncating stores
1013 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
1014 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
1015 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
1017 // ARM does not have i1 sign extending load.
1018 for (MVT VT : MVT::integer_valuetypes())
1019 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
1021 // ARM supports all 4 flavors of integer indexed load / store.
1022 if (!Subtarget->isThumb1Only()) {
1023 for (unsigned im = (unsigned)ISD::PRE_INC;
1024 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
1025 setIndexedLoadAction(im, MVT::i1, Legal);
1026 setIndexedLoadAction(im, MVT::i8, Legal);
1027 setIndexedLoadAction(im, MVT::i16, Legal);
1028 setIndexedLoadAction(im, MVT::i32, Legal);
1029 setIndexedStoreAction(im, MVT::i1, Legal);
1030 setIndexedStoreAction(im, MVT::i8, Legal);
1031 setIndexedStoreAction(im, MVT::i16, Legal);
1032 setIndexedStoreAction(im, MVT::i32, Legal);
1035 // Thumb-1 has limited post-inc load/store support - LDM r0!, {r1}.
1036 setIndexedLoadAction(ISD::POST_INC, MVT::i32, Legal);
1037 setIndexedStoreAction(ISD::POST_INC, MVT::i32, Legal);
1040 setOperationAction(ISD::SADDO, MVT::i32, Custom);
1041 setOperationAction(ISD::UADDO, MVT::i32, Custom);
1042 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
1043 setOperationAction(ISD::USUBO, MVT::i32, Custom);
1045 setOperationAction(ISD::ADDCARRY, MVT::i32, Custom);
1046 setOperationAction(ISD::SUBCARRY, MVT::i32, Custom);
1047 if (Subtarget->hasDSP()) {
1048 setOperationAction(ISD::SADDSAT, MVT::i8, Custom);
1049 setOperationAction(ISD::SSUBSAT, MVT::i8, Custom);
1050 setOperationAction(ISD::SADDSAT, MVT::i16, Custom);
1051 setOperationAction(ISD::SSUBSAT, MVT::i16, Custom);
1053 if (Subtarget->hasBaseDSP()) {
1054 setOperationAction(ISD::SADDSAT, MVT::i32, Legal);
1055 setOperationAction(ISD::SSUBSAT, MVT::i32, Legal);
1058 // i64 operation support.
1059 setOperationAction(ISD::MUL, MVT::i64, Expand);
1060 setOperationAction(ISD::MULHU, MVT::i32, Expand);
1061 if (Subtarget->isThumb1Only()) {
1062 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
1063 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
1065 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
1066 || (Subtarget->isThumb2() && !Subtarget->hasDSP()))
1067 setOperationAction(ISD::MULHS, MVT::i32, Expand);
1069 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
1070 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
1071 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
1072 setOperationAction(ISD::SRL, MVT::i64, Custom);
1073 setOperationAction(ISD::SRA, MVT::i64, Custom);
1074 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1075 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom);
1077 // MVE lowers 64 bit shifts to lsll and lsrl
1078 // assuming that ISD::SRL and SRA of i64 are already marked custom
1079 if (Subtarget->hasMVEIntegerOps())
1080 setOperationAction(ISD::SHL, MVT::i64, Custom);
1082 // Expand to __aeabi_l{lsl,lsr,asr} calls for Thumb1.
1083 if (Subtarget->isThumb1Only()) {
1084 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
1085 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
1086 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
1089 if (!Subtarget->isThumb1Only() && Subtarget->hasV6T2Ops())
1090 setOperationAction(ISD::BITREVERSE, MVT::i32, Legal);
1092 // ARM does not have ROTL.
1093 setOperationAction(ISD::ROTL, MVT::i32, Expand);
1094 for (MVT VT : MVT::fixedlen_vector_valuetypes()) {
1095 setOperationAction(ISD::ROTL, VT, Expand);
1096 setOperationAction(ISD::ROTR, VT, Expand);
1098 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
1099 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
1100 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only()) {
1101 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
1102 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, LibCall);
1105 // @llvm.readcyclecounter requires the Performance Monitors extension.
1106 // Default to the 0 expansion on unsupported platforms.
1107 // FIXME: Technically there are older ARM CPUs that have
1108 // implementation-specific ways of obtaining this information.
1109 if (Subtarget->hasPerfMon())
1110 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
1112 // Only ARMv6 has BSWAP.
1113 if (!Subtarget->hasV6Ops())
1114 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
1116 bool hasDivide = Subtarget->isThumb() ? Subtarget->hasDivideInThumbMode()
1117 : Subtarget->hasDivideInARMMode();
1119 // These are expanded into libcalls if the cpu doesn't have HW divider.
1120 setOperationAction(ISD::SDIV, MVT::i32, LibCall);
1121 setOperationAction(ISD::UDIV, MVT::i32, LibCall);
1124 if (Subtarget->isTargetWindows() && !Subtarget->hasDivideInThumbMode()) {
1125 setOperationAction(ISD::SDIV, MVT::i32, Custom);
1126 setOperationAction(ISD::UDIV, MVT::i32, Custom);
1128 setOperationAction(ISD::SDIV, MVT::i64, Custom);
1129 setOperationAction(ISD::UDIV, MVT::i64, Custom);
1132 setOperationAction(ISD::SREM, MVT::i32, Expand);
1133 setOperationAction(ISD::UREM, MVT::i32, Expand);
1135 // Register based DivRem for AEABI (RTABI 4.2)
1136 if (Subtarget->isTargetAEABI() || Subtarget->isTargetAndroid() ||
1137 Subtarget->isTargetGNUAEABI() || Subtarget->isTargetMuslAEABI() ||
1138 Subtarget->isTargetWindows()) {
1139 setOperationAction(ISD::SREM, MVT::i64, Custom);
1140 setOperationAction(ISD::UREM, MVT::i64, Custom);
1141 HasStandaloneRem = false;
1143 if (Subtarget->isTargetWindows()) {
1145 const RTLIB::Libcall Op;
1146 const char * const Name;
1147 const CallingConv::ID CC;
1148 } LibraryCalls[] = {
1149 { RTLIB::SDIVREM_I8, "__rt_sdiv", CallingConv::ARM_AAPCS },
1150 { RTLIB::SDIVREM_I16, "__rt_sdiv", CallingConv::ARM_AAPCS },
1151 { RTLIB::SDIVREM_I32, "__rt_sdiv", CallingConv::ARM_AAPCS },
1152 { RTLIB::SDIVREM_I64, "__rt_sdiv64", CallingConv::ARM_AAPCS },
1154 { RTLIB::UDIVREM_I8, "__rt_udiv", CallingConv::ARM_AAPCS },
1155 { RTLIB::UDIVREM_I16, "__rt_udiv", CallingConv::ARM_AAPCS },
1156 { RTLIB::UDIVREM_I32, "__rt_udiv", CallingConv::ARM_AAPCS },
1157 { RTLIB::UDIVREM_I64, "__rt_udiv64", CallingConv::ARM_AAPCS },
1160 for (const auto &LC : LibraryCalls) {
1161 setLibcallName(LC.Op, LC.Name);
1162 setLibcallCallingConv(LC.Op, LC.CC);
1166 const RTLIB::Libcall Op;
1167 const char * const Name;
1168 const CallingConv::ID CC;
1169 } LibraryCalls[] = {
1170 { RTLIB::SDIVREM_I8, "__aeabi_idivmod", CallingConv::ARM_AAPCS },
1171 { RTLIB::SDIVREM_I16, "__aeabi_idivmod", CallingConv::ARM_AAPCS },
1172 { RTLIB::SDIVREM_I32, "__aeabi_idivmod", CallingConv::ARM_AAPCS },
1173 { RTLIB::SDIVREM_I64, "__aeabi_ldivmod", CallingConv::ARM_AAPCS },
1175 { RTLIB::UDIVREM_I8, "__aeabi_uidivmod", CallingConv::ARM_AAPCS },
1176 { RTLIB::UDIVREM_I16, "__aeabi_uidivmod", CallingConv::ARM_AAPCS },
1177 { RTLIB::UDIVREM_I32, "__aeabi_uidivmod", CallingConv::ARM_AAPCS },
1178 { RTLIB::UDIVREM_I64, "__aeabi_uldivmod", CallingConv::ARM_AAPCS },
1181 for (const auto &LC : LibraryCalls) {
1182 setLibcallName(LC.Op, LC.Name);
1183 setLibcallCallingConv(LC.Op, LC.CC);
1187 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
1188 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
1189 setOperationAction(ISD::SDIVREM, MVT::i64, Custom);
1190 setOperationAction(ISD::UDIVREM, MVT::i64, Custom);
1192 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
1193 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
1196 if (Subtarget->getTargetTriple().isOSMSVCRT()) {
1197 // MSVCRT doesn't have powi; fall back to pow
1198 setLibcallName(RTLIB::POWI_F32, nullptr);
1199 setLibcallName(RTLIB::POWI_F64, nullptr);
1202 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
1203 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
1204 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
1205 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
1207 setOperationAction(ISD::TRAP, MVT::Other, Legal);
1208 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
1210 // Use the default implementation.
1211 setOperationAction(ISD::VASTART, MVT::Other, Custom);
1212 setOperationAction(ISD::VAARG, MVT::Other, Expand);
1213 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
1214 setOperationAction(ISD::VAEND, MVT::Other, Expand);
1215 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
1216 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
1218 if (Subtarget->isTargetWindows())
1219 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
1221 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
1223 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
1224 // the default expansion.
1225 InsertFencesForAtomic = false;
1226 if (Subtarget->hasAnyDataBarrier() &&
1227 (!Subtarget->isThumb() || Subtarget->hasV8MBaselineOps())) {
1228 // ATOMIC_FENCE needs custom lowering; the others should have been expanded
1229 // to ldrex/strex loops already.
1230 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
1231 if (!Subtarget->isThumb() || !Subtarget->isMClass())
1232 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
1234 // On v8, we have particularly efficient implementations of atomic fences
1235 // if they can be combined with nearby atomic loads and stores.
1236 if (!Subtarget->hasAcquireRelease() ||
1237 getTargetMachine().getOptLevel() == 0) {
1238 // Automatically insert fences (dmb ish) around ATOMIC_SWAP etc.
1239 InsertFencesForAtomic = true;
1242 // If there's anything we can use as a barrier, go through custom lowering
1243 // for ATOMIC_FENCE.
1244 // If target has DMB in thumb, Fences can be inserted.
1245 if (Subtarget->hasDataBarrier())
1246 InsertFencesForAtomic = true;
1248 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
1249 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
1251 // Set them all for expansion, which will force libcalls.
1252 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
1253 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
1254 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
1255 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
1256 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
1257 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
1258 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
1259 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
1260 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
1261 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
1262 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
1263 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
1264 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
1265 // Unordered/Monotonic case.
1266 if (!InsertFencesForAtomic) {
1267 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
1268 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
1272 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
1274 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
1275 if (!Subtarget->hasV6Ops()) {
1276 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
1277 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
1279 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
1281 if (!Subtarget->useSoftFloat() && Subtarget->hasFPRegs() &&
1282 !Subtarget->isThumb1Only()) {
1283 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
1284 // iff target supports vfp2.
1285 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
1286 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
1289 // We want to custom lower some of our intrinsics.
1290 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1291 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
1292 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
1293 setOperationAction(ISD::EH_SJLJ_SETUP_DISPATCH, MVT::Other, Custom);
1294 if (Subtarget->useSjLjEH())
1295 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
1297 setOperationAction(ISD::SETCC, MVT::i32, Expand);
1298 setOperationAction(ISD::SETCC, MVT::f32, Expand);
1299 setOperationAction(ISD::SETCC, MVT::f64, Expand);
1300 setOperationAction(ISD::SELECT, MVT::i32, Custom);
1301 setOperationAction(ISD::SELECT, MVT::f32, Custom);
1302 setOperationAction(ISD::SELECT, MVT::f64, Custom);
1303 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
1304 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
1305 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
1306 if (Subtarget->hasFullFP16()) {
1307 setOperationAction(ISD::SETCC, MVT::f16, Expand);
1308 setOperationAction(ISD::SELECT, MVT::f16, Custom);
1309 setOperationAction(ISD::SELECT_CC, MVT::f16, Custom);
1312 setOperationAction(ISD::SETCCCARRY, MVT::i32, Custom);
1314 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
1315 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
1316 if (Subtarget->hasFullFP16())
1317 setOperationAction(ISD::BR_CC, MVT::f16, Custom);
1318 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
1319 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
1320 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
1322 // We don't support sin/cos/fmod/copysign/pow
1323 setOperationAction(ISD::FSIN, MVT::f64, Expand);
1324 setOperationAction(ISD::FSIN, MVT::f32, Expand);
1325 setOperationAction(ISD::FCOS, MVT::f32, Expand);
1326 setOperationAction(ISD::FCOS, MVT::f64, Expand);
1327 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
1328 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
1329 setOperationAction(ISD::FREM, MVT::f64, Expand);
1330 setOperationAction(ISD::FREM, MVT::f32, Expand);
1331 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2Base() &&
1332 !Subtarget->isThumb1Only()) {
1333 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
1334 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
1336 setOperationAction(ISD::FPOW, MVT::f64, Expand);
1337 setOperationAction(ISD::FPOW, MVT::f32, Expand);
1339 if (!Subtarget->hasVFP4Base()) {
1340 setOperationAction(ISD::FMA, MVT::f64, Expand);
1341 setOperationAction(ISD::FMA, MVT::f32, Expand);
1344 // Various VFP goodness
1345 if (!Subtarget->useSoftFloat() && !Subtarget->isThumb1Only()) {
1346 // FP-ARMv8 adds f64 <-> f16 conversion. Before that it should be expanded.
1347 if (!Subtarget->hasFPARMv8Base() || !Subtarget->hasFP64()) {
1348 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
1349 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
1352 // fp16 is a special v7 extension that adds f16 <-> f32 conversions.
1353 if (!Subtarget->hasFP16()) {
1354 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
1355 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
1359 // Use __sincos_stret if available.
1360 if (getLibcallName(RTLIB::SINCOS_STRET_F32) != nullptr &&
1361 getLibcallName(RTLIB::SINCOS_STRET_F64) != nullptr) {
1362 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1363 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1366 // FP-ARMv8 implements a lot of rounding-like FP operations.
1367 if (Subtarget->hasFPARMv8Base()) {
1368 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
1369 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
1370 setOperationAction(ISD::FROUND, MVT::f32, Legal);
1371 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
1372 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
1373 setOperationAction(ISD::FRINT, MVT::f32, Legal);
1374 setOperationAction(ISD::FMINNUM, MVT::f32, Legal);
1375 setOperationAction(ISD::FMAXNUM, MVT::f32, Legal);
1376 if (Subtarget->hasNEON()) {
1377 setOperationAction(ISD::FMINNUM, MVT::v2f32, Legal);
1378 setOperationAction(ISD::FMAXNUM, MVT::v2f32, Legal);
1379 setOperationAction(ISD::FMINNUM, MVT::v4f32, Legal);
1380 setOperationAction(ISD::FMAXNUM, MVT::v4f32, Legal);
1383 if (Subtarget->hasFP64()) {
1384 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
1385 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
1386 setOperationAction(ISD::FROUND, MVT::f64, Legal);
1387 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
1388 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
1389 setOperationAction(ISD::FRINT, MVT::f64, Legal);
1390 setOperationAction(ISD::FMINNUM, MVT::f64, Legal);
1391 setOperationAction(ISD::FMAXNUM, MVT::f64, Legal);
1395 // FP16 often need to be promoted to call lib functions
1396 if (Subtarget->hasFullFP16()) {
1397 setOperationAction(ISD::FREM, MVT::f16, Promote);
1398 setOperationAction(ISD::FCOPYSIGN, MVT::f16, Expand);
1399 setOperationAction(ISD::FSIN, MVT::f16, Promote);
1400 setOperationAction(ISD::FCOS, MVT::f16, Promote);
1401 setOperationAction(ISD::FSINCOS, MVT::f16, Promote);
1402 setOperationAction(ISD::FPOWI, MVT::f16, Promote);
1403 setOperationAction(ISD::FPOW, MVT::f16, Promote);
1404 setOperationAction(ISD::FEXP, MVT::f16, Promote);
1405 setOperationAction(ISD::FEXP2, MVT::f16, Promote);
1406 setOperationAction(ISD::FLOG, MVT::f16, Promote);
1407 setOperationAction(ISD::FLOG10, MVT::f16, Promote);
1408 setOperationAction(ISD::FLOG2, MVT::f16, Promote);
1410 setOperationAction(ISD::FROUND, MVT::f16, Legal);
1413 if (Subtarget->hasNEON()) {
1414 // vmin and vmax aren't available in a scalar form, so we use
1415 // a NEON instruction with an undef lane instead.
1416 setOperationAction(ISD::FMINIMUM, MVT::f16, Legal);
1417 setOperationAction(ISD::FMAXIMUM, MVT::f16, Legal);
1418 setOperationAction(ISD::FMINIMUM, MVT::f32, Legal);
1419 setOperationAction(ISD::FMAXIMUM, MVT::f32, Legal);
1420 setOperationAction(ISD::FMINIMUM, MVT::v2f32, Legal);
1421 setOperationAction(ISD::FMAXIMUM, MVT::v2f32, Legal);
1422 setOperationAction(ISD::FMINIMUM, MVT::v4f32, Legal);
1423 setOperationAction(ISD::FMAXIMUM, MVT::v4f32, Legal);
1425 if (Subtarget->hasFullFP16()) {
1426 setOperationAction(ISD::FMINNUM, MVT::v4f16, Legal);
1427 setOperationAction(ISD::FMAXNUM, MVT::v4f16, Legal);
1428 setOperationAction(ISD::FMINNUM, MVT::v8f16, Legal);
1429 setOperationAction(ISD::FMAXNUM, MVT::v8f16, Legal);
1431 setOperationAction(ISD::FMINIMUM, MVT::v4f16, Legal);
1432 setOperationAction(ISD::FMAXIMUM, MVT::v4f16, Legal);
1433 setOperationAction(ISD::FMINIMUM, MVT::v8f16, Legal);
1434 setOperationAction(ISD::FMAXIMUM, MVT::v8f16, Legal);
1438 // We have target-specific dag combine patterns for the following nodes:
1439 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
1440 setTargetDAGCombine(ISD::ADD);
1441 setTargetDAGCombine(ISD::SUB);
1442 setTargetDAGCombine(ISD::MUL);
1443 setTargetDAGCombine(ISD::AND);
1444 setTargetDAGCombine(ISD::OR);
1445 setTargetDAGCombine(ISD::XOR);
1447 if (Subtarget->hasV6Ops())
1448 setTargetDAGCombine(ISD::SRL);
1449 if (Subtarget->isThumb1Only())
1450 setTargetDAGCombine(ISD::SHL);
1452 setStackPointerRegisterToSaveRestore(ARM::SP);
1454 if (Subtarget->useSoftFloat() || Subtarget->isThumb1Only() ||
1455 !Subtarget->hasVFP2Base() || Subtarget->hasMinSize())
1456 setSchedulingPreference(Sched::RegPressure);
1458 setSchedulingPreference(Sched::Hybrid);
1460 //// temporary - rewrite interface to use type
1461 MaxStoresPerMemset = 8;
1462 MaxStoresPerMemsetOptSize = 4;
1463 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
1464 MaxStoresPerMemcpyOptSize = 2;
1465 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
1466 MaxStoresPerMemmoveOptSize = 2;
1468 // On ARM arguments smaller than 4 bytes are extended, so all arguments
1469 // are at least 4 bytes aligned.
1470 setMinStackArgumentAlignment(Align(4));
1472 // Prefer likely predicted branches to selects on out-of-order cores.
1473 PredictableSelectIsExpensive = Subtarget->getSchedModel().isOutOfOrder();
1475 setPrefLoopAlignment(Align(1ULL << Subtarget->getPrefLoopLogAlignment()));
1477 setMinFunctionAlignment(Subtarget->isThumb() ? Align(2) : Align(4));
1479 if (Subtarget->isThumb() || Subtarget->isThumb2())
1480 setTargetDAGCombine(ISD::ABS);
1483 bool ARMTargetLowering::useSoftFloat() const {
1484 return Subtarget->useSoftFloat();
1487 // FIXME: It might make sense to define the representative register class as the
1488 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
1489 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
1490 // SPR's representative would be DPR_VFP2. This should work well if register
1491 // pressure tracking were modified such that a register use would increment the
1492 // pressure of the register class's representative and all of it's super
1493 // classes' representatives transitively. We have not implemented this because
1494 // of the difficulty prior to coalescing of modeling operand register classes
1495 // due to the common occurrence of cross class copies and subregister insertions
1497 std::pair<const TargetRegisterClass *, uint8_t>
1498 ARMTargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
1500 const TargetRegisterClass *RRC = nullptr;
1502 switch (VT.SimpleTy) {
1504 return TargetLowering::findRepresentativeClass(TRI, VT);
1505 // Use DPR as representative register class for all floating point
1506 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
1507 // the cost is 1 for both f32 and f64.
1508 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
1509 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
1510 RRC = &ARM::DPRRegClass;
1511 // When NEON is used for SP, only half of the register file is available
1512 // because operations that define both SP and DP results will be constrained
1513 // to the VFP2 class (D0-D15). We currently model this constraint prior to
1514 // coalescing by double-counting the SP regs. See the FIXME above.
1515 if (Subtarget->useNEONForSinglePrecisionFP())
1518 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1519 case MVT::v4f32: case MVT::v2f64:
1520 RRC = &ARM::DPRRegClass;
1524 RRC = &ARM::DPRRegClass;
1528 RRC = &ARM::DPRRegClass;
1532 return std::make_pair(RRC, Cost);
1535 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
1536 switch ((ARMISD::NodeType)Opcode) {
1537 case ARMISD::FIRST_NUMBER: break;
1538 case ARMISD::Wrapper: return "ARMISD::Wrapper";
1539 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
1540 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
1541 case ARMISD::COPY_STRUCT_BYVAL: return "ARMISD::COPY_STRUCT_BYVAL";
1542 case ARMISD::CALL: return "ARMISD::CALL";
1543 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
1544 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
1545 case ARMISD::BRCOND: return "ARMISD::BRCOND";
1546 case ARMISD::BR_JT: return "ARMISD::BR_JT";
1547 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
1548 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
1549 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
1550 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
1551 case ARMISD::CMP: return "ARMISD::CMP";
1552 case ARMISD::CMN: return "ARMISD::CMN";
1553 case ARMISD::CMPZ: return "ARMISD::CMPZ";
1554 case ARMISD::CMPFP: return "ARMISD::CMPFP";
1555 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
1556 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
1557 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
1559 case ARMISD::CMOV: return "ARMISD::CMOV";
1560 case ARMISD::SUBS: return "ARMISD::SUBS";
1562 case ARMISD::SSAT: return "ARMISD::SSAT";
1563 case ARMISD::USAT: return "ARMISD::USAT";
1565 case ARMISD::ASRL: return "ARMISD::ASRL";
1566 case ARMISD::LSRL: return "ARMISD::LSRL";
1567 case ARMISD::LSLL: return "ARMISD::LSLL";
1569 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
1570 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
1571 case ARMISD::RRX: return "ARMISD::RRX";
1573 case ARMISD::ADDC: return "ARMISD::ADDC";
1574 case ARMISD::ADDE: return "ARMISD::ADDE";
1575 case ARMISD::SUBC: return "ARMISD::SUBC";
1576 case ARMISD::SUBE: return "ARMISD::SUBE";
1577 case ARMISD::LSLS: return "ARMISD::LSLS";
1579 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
1580 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
1581 case ARMISD::VMOVhr: return "ARMISD::VMOVhr";
1582 case ARMISD::VMOVrh: return "ARMISD::VMOVrh";
1583 case ARMISD::VMOVSR: return "ARMISD::VMOVSR";
1585 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
1586 case ARMISD::EH_SJLJ_LONGJMP: return "ARMISD::EH_SJLJ_LONGJMP";
1587 case ARMISD::EH_SJLJ_SETUP_DISPATCH: return "ARMISD::EH_SJLJ_SETUP_DISPATCH";
1589 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1591 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1593 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1595 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1597 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1599 case ARMISD::WIN__CHKSTK: return "ARMISD::WIN__CHKSTK";
1600 case ARMISD::WIN__DBZCHK: return "ARMISD::WIN__DBZCHK";
1602 case ARMISD::PREDICATE_CAST: return "ARMISD::PREDICATE_CAST";
1603 case ARMISD::VCMP: return "ARMISD::VCMP";
1604 case ARMISD::VCMPZ: return "ARMISD::VCMPZ";
1605 case ARMISD::VTST: return "ARMISD::VTST";
1607 case ARMISD::VSHLs: return "ARMISD::VSHLs";
1608 case ARMISD::VSHLu: return "ARMISD::VSHLu";
1609 case ARMISD::VSHLIMM: return "ARMISD::VSHLIMM";
1610 case ARMISD::VSHRsIMM: return "ARMISD::VSHRsIMM";
1611 case ARMISD::VSHRuIMM: return "ARMISD::VSHRuIMM";
1612 case ARMISD::VRSHRsIMM: return "ARMISD::VRSHRsIMM";
1613 case ARMISD::VRSHRuIMM: return "ARMISD::VRSHRuIMM";
1614 case ARMISD::VRSHRNIMM: return "ARMISD::VRSHRNIMM";
1615 case ARMISD::VQSHLsIMM: return "ARMISD::VQSHLsIMM";
1616 case ARMISD::VQSHLuIMM: return "ARMISD::VQSHLuIMM";
1617 case ARMISD::VQSHLsuIMM: return "ARMISD::VQSHLsuIMM";
1618 case ARMISD::VQSHRNsIMM: return "ARMISD::VQSHRNsIMM";
1619 case ARMISD::VQSHRNuIMM: return "ARMISD::VQSHRNuIMM";
1620 case ARMISD::VQSHRNsuIMM: return "ARMISD::VQSHRNsuIMM";
1621 case ARMISD::VQRSHRNsIMM: return "ARMISD::VQRSHRNsIMM";
1622 case ARMISD::VQRSHRNuIMM: return "ARMISD::VQRSHRNuIMM";
1623 case ARMISD::VQRSHRNsuIMM: return "ARMISD::VQRSHRNsuIMM";
1624 case ARMISD::VSLIIMM: return "ARMISD::VSLIIMM";
1625 case ARMISD::VSRIIMM: return "ARMISD::VSRIIMM";
1626 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1627 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1628 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1629 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1630 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1631 case ARMISD::VDUP: return "ARMISD::VDUP";
1632 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1633 case ARMISD::VEXT: return "ARMISD::VEXT";
1634 case ARMISD::VREV64: return "ARMISD::VREV64";
1635 case ARMISD::VREV32: return "ARMISD::VREV32";
1636 case ARMISD::VREV16: return "ARMISD::VREV16";
1637 case ARMISD::VZIP: return "ARMISD::VZIP";
1638 case ARMISD::VUZP: return "ARMISD::VUZP";
1639 case ARMISD::VTRN: return "ARMISD::VTRN";
1640 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1641 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1642 case ARMISD::VMOVN: return "ARMISD::VMOVN";
1643 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1644 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1645 case ARMISD::UMAAL: return "ARMISD::UMAAL";
1646 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1647 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1648 case ARMISD::SMLALBB: return "ARMISD::SMLALBB";
1649 case ARMISD::SMLALBT: return "ARMISD::SMLALBT";
1650 case ARMISD::SMLALTB: return "ARMISD::SMLALTB";
1651 case ARMISD::SMLALTT: return "ARMISD::SMLALTT";
1652 case ARMISD::SMULWB: return "ARMISD::SMULWB";
1653 case ARMISD::SMULWT: return "ARMISD::SMULWT";
1654 case ARMISD::SMLALD: return "ARMISD::SMLALD";
1655 case ARMISD::SMLALDX: return "ARMISD::SMLALDX";
1656 case ARMISD::SMLSLD: return "ARMISD::SMLSLD";
1657 case ARMISD::SMLSLDX: return "ARMISD::SMLSLDX";
1658 case ARMISD::SMMLAR: return "ARMISD::SMMLAR";
1659 case ARMISD::SMMLSR: return "ARMISD::SMMLSR";
1660 case ARMISD::QADD16b: return "ARMISD::QADD16b";
1661 case ARMISD::QSUB16b: return "ARMISD::QSUB16b";
1662 case ARMISD::QADD8b: return "ARMISD::QADD8b";
1663 case ARMISD::QSUB8b: return "ARMISD::QSUB8b";
1664 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1665 case ARMISD::BFI: return "ARMISD::BFI";
1666 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1667 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1668 case ARMISD::VBSL: return "ARMISD::VBSL";
1669 case ARMISD::MEMCPY: return "ARMISD::MEMCPY";
1670 case ARMISD::VLD1DUP: return "ARMISD::VLD1DUP";
1671 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1672 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1673 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1674 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1675 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1676 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1677 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1678 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1679 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1680 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1681 case ARMISD::VLD1DUP_UPD: return "ARMISD::VLD1DUP_UPD";
1682 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1683 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1684 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1685 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1686 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1687 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1688 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1689 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1690 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1691 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1692 case ARMISD::WLS: return "ARMISD::WLS";
1693 case ARMISD::LE: return "ARMISD::LE";
1694 case ARMISD::LOOP_DEC: return "ARMISD::LOOP_DEC";
1695 case ARMISD::CSINV: return "ARMISD::CSINV";
1696 case ARMISD::CSNEG: return "ARMISD::CSNEG";
1697 case ARMISD::CSINC: return "ARMISD::CSINC";
1702 EVT ARMTargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1705 return getPointerTy(DL);
1707 // MVE has a predicate register.
1708 if (Subtarget->hasMVEIntegerOps() &&
1709 (VT == MVT::v4i32 || VT == MVT::v8i16 || VT == MVT::v16i8))
1710 return MVT::getVectorVT(MVT::i1, VT.getVectorElementCount());
1711 return VT.changeVectorElementTypeToInteger();
1714 /// getRegClassFor - Return the register class that should be used for the
1715 /// specified value type.
1716 const TargetRegisterClass *
1717 ARMTargetLowering::getRegClassFor(MVT VT, bool isDivergent) const {
1719 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1720 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1721 // load / store 4 to 8 consecutive NEON D registers, or 2 to 4 consecutive
1723 if (Subtarget->hasNEON() || Subtarget->hasMVEIntegerOps()) {
1724 if (VT == MVT::v4i64)
1725 return &ARM::QQPRRegClass;
1726 if (VT == MVT::v8i64)
1727 return &ARM::QQQQPRRegClass;
1729 return TargetLowering::getRegClassFor(VT);
1732 // memcpy, and other memory intrinsics, typically tries to use LDM/STM if the
1733 // source/dest is aligned and the copy size is large enough. We therefore want
1734 // to align such objects passed to memory intrinsics.
1735 bool ARMTargetLowering::shouldAlignPointerArgs(CallInst *CI, unsigned &MinSize,
1736 unsigned &PrefAlign) const {
1737 if (!isa<MemIntrinsic>(CI))
1740 // On ARM11 onwards (excluding M class) 8-byte aligned LDM is typically 1
1741 // cycle faster than 4-byte aligned LDM.
1742 PrefAlign = (Subtarget->hasV6Ops() && !Subtarget->isMClass() ? 8 : 4);
1746 // Create a fast isel object.
1748 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1749 const TargetLibraryInfo *libInfo) const {
1750 return ARM::createFastISel(funcInfo, libInfo);
1753 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1754 unsigned NumVals = N->getNumValues();
1756 return Sched::RegPressure;
1758 for (unsigned i = 0; i != NumVals; ++i) {
1759 EVT VT = N->getValueType(i);
1760 if (VT == MVT::Glue || VT == MVT::Other)
1762 if (VT.isFloatingPoint() || VT.isVector())
1766 if (!N->isMachineOpcode())
1767 return Sched::RegPressure;
1769 // Load are scheduled for latency even if there instruction itinerary
1770 // is not available.
1771 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
1772 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1774 if (MCID.getNumDefs() == 0)
1775 return Sched::RegPressure;
1776 if (!Itins->isEmpty() &&
1777 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1780 return Sched::RegPressure;
1783 //===----------------------------------------------------------------------===//
1785 //===----------------------------------------------------------------------===//
1787 static bool isSRL16(const SDValue &Op) {
1788 if (Op.getOpcode() != ISD::SRL)
1790 if (auto Const = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
1791 return Const->getZExtValue() == 16;
1795 static bool isSRA16(const SDValue &Op) {
1796 if (Op.getOpcode() != ISD::SRA)
1798 if (auto Const = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
1799 return Const->getZExtValue() == 16;
1803 static bool isSHL16(const SDValue &Op) {
1804 if (Op.getOpcode() != ISD::SHL)
1806 if (auto Const = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
1807 return Const->getZExtValue() == 16;
1811 // Check for a signed 16-bit value. We special case SRA because it makes it
1812 // more simple when also looking for SRAs that aren't sign extending a
1813 // smaller value. Without the check, we'd need to take extra care with
1814 // checking order for some operations.
1815 static bool isS16(const SDValue &Op, SelectionDAG &DAG) {
1817 return isSHL16(Op.getOperand(0));
1818 return DAG.ComputeNumSignBits(Op) == 17;
1821 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1822 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1824 default: llvm_unreachable("Unknown condition code!");
1825 case ISD::SETNE: return ARMCC::NE;
1826 case ISD::SETEQ: return ARMCC::EQ;
1827 case ISD::SETGT: return ARMCC::GT;
1828 case ISD::SETGE: return ARMCC::GE;
1829 case ISD::SETLT: return ARMCC::LT;
1830 case ISD::SETLE: return ARMCC::LE;
1831 case ISD::SETUGT: return ARMCC::HI;
1832 case ISD::SETUGE: return ARMCC::HS;
1833 case ISD::SETULT: return ARMCC::LO;
1834 case ISD::SETULE: return ARMCC::LS;
1838 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1839 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1840 ARMCC::CondCodes &CondCode2) {
1841 CondCode2 = ARMCC::AL;
1843 default: llvm_unreachable("Unknown FP condition!");
1845 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1847 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1849 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1850 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1851 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1852 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1853 case ISD::SETO: CondCode = ARMCC::VC; break;
1854 case ISD::SETUO: CondCode = ARMCC::VS; break;
1855 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1856 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1857 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1859 case ISD::SETULT: CondCode = ARMCC::LT; break;
1861 case ISD::SETULE: CondCode = ARMCC::LE; break;
1863 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1867 //===----------------------------------------------------------------------===//
1868 // Calling Convention Implementation
1869 //===----------------------------------------------------------------------===//
1871 /// getEffectiveCallingConv - Get the effective calling convention, taking into
1872 /// account presence of floating point hardware and calling convention
1873 /// limitations, such as support for variadic functions.
1875 ARMTargetLowering::getEffectiveCallingConv(CallingConv::ID CC,
1876 bool isVarArg) const {
1879 report_fatal_error("Unsupported calling convention");
1880 case CallingConv::ARM_AAPCS:
1881 case CallingConv::ARM_APCS:
1882 case CallingConv::GHC:
1883 case CallingConv::CFGuard_Check:
1885 case CallingConv::PreserveMost:
1886 return CallingConv::PreserveMost;
1887 case CallingConv::ARM_AAPCS_VFP:
1888 case CallingConv::Swift:
1889 return isVarArg ? CallingConv::ARM_AAPCS : CallingConv::ARM_AAPCS_VFP;
1890 case CallingConv::C:
1891 if (!Subtarget->isAAPCS_ABI())
1892 return CallingConv::ARM_APCS;
1893 else if (Subtarget->hasVFP2Base() && !Subtarget->isThumb1Only() &&
1894 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1896 return CallingConv::ARM_AAPCS_VFP;
1898 return CallingConv::ARM_AAPCS;
1899 case CallingConv::Fast:
1900 case CallingConv::CXX_FAST_TLS:
1901 if (!Subtarget->isAAPCS_ABI()) {
1902 if (Subtarget->hasVFP2Base() && !Subtarget->isThumb1Only() && !isVarArg)
1903 return CallingConv::Fast;
1904 return CallingConv::ARM_APCS;
1905 } else if (Subtarget->hasVFP2Base() &&
1906 !Subtarget->isThumb1Only() && !isVarArg)
1907 return CallingConv::ARM_AAPCS_VFP;
1909 return CallingConv::ARM_AAPCS;
1913 CCAssignFn *ARMTargetLowering::CCAssignFnForCall(CallingConv::ID CC,
1914 bool isVarArg) const {
1915 return CCAssignFnForNode(CC, false, isVarArg);
1918 CCAssignFn *ARMTargetLowering::CCAssignFnForReturn(CallingConv::ID CC,
1919 bool isVarArg) const {
1920 return CCAssignFnForNode(CC, true, isVarArg);
1923 /// CCAssignFnForNode - Selects the correct CCAssignFn for the given
1924 /// CallingConvention.
1925 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1927 bool isVarArg) const {
1928 switch (getEffectiveCallingConv(CC, isVarArg)) {
1930 report_fatal_error("Unsupported calling convention");
1931 case CallingConv::ARM_APCS:
1932 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1933 case CallingConv::ARM_AAPCS:
1934 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1935 case CallingConv::ARM_AAPCS_VFP:
1936 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1937 case CallingConv::Fast:
1938 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1939 case CallingConv::GHC:
1940 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1941 case CallingConv::PreserveMost:
1942 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1943 case CallingConv::CFGuard_Check:
1944 return (Return ? RetCC_ARM_AAPCS : CC_ARM_Win32_CFGuard_Check);
1948 /// LowerCallResult - Lower the result values of a call into the
1949 /// appropriate copies out of appropriate physical registers.
1950 SDValue ARMTargetLowering::LowerCallResult(
1951 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
1952 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
1953 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
1954 SDValue ThisVal) const {
1955 // Assign locations to each value returned by this call.
1956 SmallVector<CCValAssign, 16> RVLocs;
1957 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1959 CCInfo.AnalyzeCallResult(Ins, CCAssignFnForReturn(CallConv, isVarArg));
1961 // Copy all of the result registers out of their specified physreg.
1962 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1963 CCValAssign VA = RVLocs[i];
1965 // Pass 'this' value directly from the argument to return value, to avoid
1966 // reg unit interference
1967 if (i == 0 && isThisReturn) {
1968 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1969 "unexpected return calling convention register assignment");
1970 InVals.push_back(ThisVal);
1975 if (VA.needsCustom()) {
1976 // Handle f64 or half of a v2f64.
1977 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1979 Chain = Lo.getValue(1);
1980 InFlag = Lo.getValue(2);
1981 VA = RVLocs[++i]; // skip ahead to next loc
1982 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1984 Chain = Hi.getValue(1);
1985 InFlag = Hi.getValue(2);
1986 if (!Subtarget->isLittle())
1988 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1990 if (VA.getLocVT() == MVT::v2f64) {
1991 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1992 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1993 DAG.getConstant(0, dl, MVT::i32));
1995 VA = RVLocs[++i]; // skip ahead to next loc
1996 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1997 Chain = Lo.getValue(1);
1998 InFlag = Lo.getValue(2);
1999 VA = RVLocs[++i]; // skip ahead to next loc
2000 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
2001 Chain = Hi.getValue(1);
2002 InFlag = Hi.getValue(2);
2003 if (!Subtarget->isLittle())
2005 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
2006 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
2007 DAG.getConstant(1, dl, MVT::i32));
2010 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
2012 Chain = Val.getValue(1);
2013 InFlag = Val.getValue(2);
2016 switch (VA.getLocInfo()) {
2017 default: llvm_unreachable("Unknown loc info!");
2018 case CCValAssign::Full: break;
2019 case CCValAssign::BCvt:
2020 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
2024 InVals.push_back(Val);
2030 /// LowerMemOpCallTo - Store the argument to the stack.
2031 SDValue ARMTargetLowering::LowerMemOpCallTo(SDValue Chain, SDValue StackPtr,
2032 SDValue Arg, const SDLoc &dl,
2034 const CCValAssign &VA,
2035 ISD::ArgFlagsTy Flags) const {
2036 unsigned LocMemOffset = VA.getLocMemOffset();
2037 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2038 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2040 return DAG.getStore(
2041 Chain, dl, Arg, PtrOff,
2042 MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset));
2045 void ARMTargetLowering::PassF64ArgInRegs(const SDLoc &dl, SelectionDAG &DAG,
2046 SDValue Chain, SDValue &Arg,
2047 RegsToPassVector &RegsToPass,
2048 CCValAssign &VA, CCValAssign &NextVA,
2050 SmallVectorImpl<SDValue> &MemOpChains,
2051 ISD::ArgFlagsTy Flags) const {
2052 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2053 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2054 unsigned id = Subtarget->isLittle() ? 0 : 1;
2055 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd.getValue(id)));
2057 if (NextVA.isRegLoc())
2058 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1-id)));
2060 assert(NextVA.isMemLoc());
2061 if (!StackPtr.getNode())
2062 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP,
2063 getPointerTy(DAG.getDataLayout()));
2065 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1-id),
2071 /// LowerCall - Lowering a call into a callseq_start <-
2072 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
2075 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2076 SmallVectorImpl<SDValue> &InVals) const {
2077 SelectionDAG &DAG = CLI.DAG;
2079 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2080 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2081 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2082 SDValue Chain = CLI.Chain;
2083 SDValue Callee = CLI.Callee;
2084 bool &isTailCall = CLI.IsTailCall;
2085 CallingConv::ID CallConv = CLI.CallConv;
2086 bool doesNotRet = CLI.DoesNotReturn;
2087 bool isVarArg = CLI.IsVarArg;
2089 MachineFunction &MF = DAG.getMachineFunction();
2090 MachineFunction::CallSiteInfo CSInfo;
2091 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
2092 bool isThisReturn = false;
2093 bool PreferIndirect = false;
2095 // Disable tail calls if they're not supported.
2096 if (!Subtarget->supportsTailCall())
2099 if (isa<GlobalAddressSDNode>(Callee)) {
2100 // If we're optimizing for minimum size and the function is called three or
2101 // more times in this block, we can improve codesize by calling indirectly
2102 // as BLXr has a 16-bit encoding.
2103 auto *GV = cast<GlobalAddressSDNode>(Callee)->getGlobal();
2105 auto *BB = CLI.CS.getParent();
2106 PreferIndirect = Subtarget->isThumb() && Subtarget->hasMinSize() &&
2107 count_if(GV->users(), [&BB](const User *U) {
2108 return isa<Instruction>(U) &&
2109 cast<Instruction>(U)->getParent() == BB;
2114 // Check if it's really possible to do a tail call.
2115 isTailCall = IsEligibleForTailCallOptimization(
2116 Callee, CallConv, isVarArg, isStructRet,
2117 MF.getFunction().hasStructRetAttr(), Outs, OutVals, Ins, DAG,
2119 if (!isTailCall && CLI.CS && CLI.CS.isMustTailCall())
2120 report_fatal_error("failed to perform tail call elimination on a call "
2121 "site marked musttail");
2122 // We don't support GuaranteedTailCallOpt for ARM, only automatically
2123 // detected sibcalls.
2128 // Analyze operands of the call, assigning locations to each operand.
2129 SmallVector<CCValAssign, 16> ArgLocs;
2130 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
2132 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CallConv, isVarArg));
2134 // Get a count of how many bytes are to be pushed on the stack.
2135 unsigned NumBytes = CCInfo.getNextStackOffset();
2138 // For tail calls, memory operands are available in our caller's stack.
2141 // Adjust the stack pointer for the new arguments...
2142 // These operations are automatically eliminated by the prolog/epilog pass
2143 Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, dl);
2147 DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy(DAG.getDataLayout()));
2149 RegsToPassVector RegsToPass;
2150 SmallVector<SDValue, 8> MemOpChains;
2152 // Walk the register/memloc assignments, inserting copies/loads. In the case
2153 // of tail call optimization, arguments are handled later.
2154 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2156 ++i, ++realArgIdx) {
2157 CCValAssign &VA = ArgLocs[i];
2158 SDValue Arg = OutVals[realArgIdx];
2159 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2160 bool isByVal = Flags.isByVal();
2162 // Promote the value if needed.
2163 switch (VA.getLocInfo()) {
2164 default: llvm_unreachable("Unknown loc info!");
2165 case CCValAssign::Full: break;
2166 case CCValAssign::SExt:
2167 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
2169 case CCValAssign::ZExt:
2170 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
2172 case CCValAssign::AExt:
2173 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
2175 case CCValAssign::BCvt:
2176 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2180 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
2181 if (VA.needsCustom()) {
2182 if (VA.getLocVT() == MVT::v2f64) {
2183 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2184 DAG.getConstant(0, dl, MVT::i32));
2185 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2186 DAG.getConstant(1, dl, MVT::i32));
2188 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
2189 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
2191 VA = ArgLocs[++i]; // skip ahead to next loc
2192 if (VA.isRegLoc()) {
2193 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
2194 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
2196 assert(VA.isMemLoc());
2198 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
2199 dl, DAG, VA, Flags));
2202 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
2203 StackPtr, MemOpChains, Flags);
2205 } else if (VA.isRegLoc()) {
2206 if (realArgIdx == 0 && Flags.isReturned() && !Flags.isSwiftSelf() &&
2207 Outs[0].VT == MVT::i32) {
2208 assert(VA.getLocVT() == MVT::i32 &&
2209 "unexpected calling convention register assignment");
2210 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
2211 "unexpected use of 'returned'");
2212 isThisReturn = true;
2214 const TargetOptions &Options = DAG.getTarget().Options;
2215 if (Options.EnableDebugEntryValues)
2216 CSInfo.emplace_back(VA.getLocReg(), i);
2217 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2218 } else if (isByVal) {
2219 assert(VA.isMemLoc());
2220 unsigned offset = 0;
2222 // True if this byval aggregate will be split between registers
2224 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
2225 unsigned CurByValIdx = CCInfo.getInRegsParamsProcessed();
2227 if (CurByValIdx < ByValArgsCount) {
2229 unsigned RegBegin, RegEnd;
2230 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
2233 DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
2235 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
2236 SDValue Const = DAG.getConstant(4*i, dl, MVT::i32);
2237 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
2238 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
2239 MachinePointerInfo(),
2240 DAG.InferPtrAlignment(AddArg));
2241 MemOpChains.push_back(Load.getValue(1));
2242 RegsToPass.push_back(std::make_pair(j, Load));
2245 // If parameter size outsides register area, "offset" value
2246 // helps us to calculate stack slot for remained part properly.
2247 offset = RegEnd - RegBegin;
2249 CCInfo.nextInRegsParam();
2252 if (Flags.getByValSize() > 4*offset) {
2253 auto PtrVT = getPointerTy(DAG.getDataLayout());
2254 unsigned LocMemOffset = VA.getLocMemOffset();
2255 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2256 SDValue Dst = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, StkPtrOff);
2257 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset, dl);
2258 SDValue Src = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, SrcOffset);
2259 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset, dl,
2261 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), dl,
2264 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2265 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
2266 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
2269 } else if (!isTailCall) {
2270 assert(VA.isMemLoc());
2272 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2273 dl, DAG, VA, Flags));
2277 if (!MemOpChains.empty())
2278 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
2280 // Build a sequence of copy-to-reg nodes chained together with token chain
2281 // and flag operands which copy the outgoing args into the appropriate regs.
2283 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2284 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2285 RegsToPass[i].second, InFlag);
2286 InFlag = Chain.getValue(1);
2289 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2290 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2291 // node so that legalize doesn't hack it.
2292 bool isDirect = false;
2294 const TargetMachine &TM = getTargetMachine();
2295 const Module *Mod = MF.getFunction().getParent();
2296 const GlobalValue *GV = nullptr;
2297 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2298 GV = G->getGlobal();
2300 !TM.shouldAssumeDSOLocal(*Mod, GV) && Subtarget->isTargetMachO();
2302 bool isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
2303 bool isLocalARMFunc = false;
2304 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2305 auto PtrVt = getPointerTy(DAG.getDataLayout());
2307 if (Subtarget->genLongCalls()) {
2308 assert((!isPositionIndependent() || Subtarget->isTargetWindows()) &&
2309 "long-calls codegen is not position independent!");
2310 // Handle a global address or an external symbol. If it's not one of
2311 // those, the target's already in a register, so we don't need to do
2313 if (isa<GlobalAddressSDNode>(Callee)) {
2314 // Create a constant pool entry for the callee address
2315 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2316 ARMConstantPoolValue *CPV =
2317 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
2319 // Get the address of the callee into a register
2320 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
2321 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2322 Callee = DAG.getLoad(
2323 PtrVt, dl, DAG.getEntryNode(), CPAddr,
2324 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()));
2325 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
2326 const char *Sym = S->getSymbol();
2328 // Create a constant pool entry for the callee address
2329 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2330 ARMConstantPoolValue *CPV =
2331 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
2332 ARMPCLabelIndex, 0);
2333 // Get the address of the callee into a register
2334 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
2335 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2336 Callee = DAG.getLoad(
2337 PtrVt, dl, DAG.getEntryNode(), CPAddr,
2338 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()));
2340 } else if (isa<GlobalAddressSDNode>(Callee)) {
2341 if (!PreferIndirect) {
2343 bool isDef = GV->isStrongDefinitionForLinker();
2345 // ARM call to a local ARM function is predicable.
2346 isLocalARMFunc = !Subtarget->isThumb() && (isDef || !ARMInterworking);
2347 // tBX takes a register source operand.
2348 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
2349 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
2350 Callee = DAG.getNode(
2351 ARMISD::WrapperPIC, dl, PtrVt,
2352 DAG.getTargetGlobalAddress(GV, dl, PtrVt, 0, ARMII::MO_NONLAZY));
2353 Callee = DAG.getLoad(
2354 PtrVt, dl, DAG.getEntryNode(), Callee,
2355 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
2356 /* Alignment = */ 0, MachineMemOperand::MODereferenceable |
2357 MachineMemOperand::MOInvariant);
2358 } else if (Subtarget->isTargetCOFF()) {
2359 assert(Subtarget->isTargetWindows() &&
2360 "Windows is the only supported COFF target");
2361 unsigned TargetFlags = ARMII::MO_NO_FLAG;
2362 if (GV->hasDLLImportStorageClass())
2363 TargetFlags = ARMII::MO_DLLIMPORT;
2364 else if (!TM.shouldAssumeDSOLocal(*GV->getParent(), GV))
2365 TargetFlags = ARMII::MO_COFFSTUB;
2366 Callee = DAG.getTargetGlobalAddress(GV, dl, PtrVt, /*offset=*/0,
2368 if (TargetFlags & (ARMII::MO_DLLIMPORT | ARMII::MO_COFFSTUB))
2370 DAG.getLoad(PtrVt, dl, DAG.getEntryNode(),
2371 DAG.getNode(ARMISD::Wrapper, dl, PtrVt, Callee),
2372 MachinePointerInfo::getGOT(DAG.getMachineFunction()));
2374 Callee = DAG.getTargetGlobalAddress(GV, dl, PtrVt, 0, 0);
2377 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2379 // tBX takes a register source operand.
2380 const char *Sym = S->getSymbol();
2381 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
2382 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2383 ARMConstantPoolValue *CPV =
2384 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
2385 ARMPCLabelIndex, 4);
2386 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
2387 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2388 Callee = DAG.getLoad(
2389 PtrVt, dl, DAG.getEntryNode(), CPAddr,
2390 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()));
2391 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2392 Callee = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVt, Callee, PICLabel);
2394 Callee = DAG.getTargetExternalSymbol(Sym, PtrVt, 0);
2398 // FIXME: handle tail calls differently.
2400 if (Subtarget->isThumb()) {
2401 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
2402 CallOpc = ARMISD::CALL_NOLINK;
2404 CallOpc = ARMISD::CALL;
2406 if (!isDirect && !Subtarget->hasV5TOps())
2407 CallOpc = ARMISD::CALL_NOLINK;
2408 else if (doesNotRet && isDirect && Subtarget->hasRetAddrStack() &&
2409 // Emit regular call when code size is the priority
2410 !Subtarget->hasMinSize())
2411 // "mov lr, pc; b _foo" to avoid confusing the RSP
2412 CallOpc = ARMISD::CALL_NOLINK;
2414 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
2417 std::vector<SDValue> Ops;
2418 Ops.push_back(Chain);
2419 Ops.push_back(Callee);
2421 // Add argument registers to the end of the list so that they are known live
2423 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2424 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2425 RegsToPass[i].second.getValueType()));
2427 // Add a register mask operand representing the call-preserved registers.
2429 const uint32_t *Mask;
2430 const ARMBaseRegisterInfo *ARI = Subtarget->getRegisterInfo();
2432 // For 'this' returns, use the R0-preserving mask if applicable
2433 Mask = ARI->getThisReturnPreservedMask(MF, CallConv);
2435 // Set isThisReturn to false if the calling convention is not one that
2436 // allows 'returned' to be modeled in this way, so LowerCallResult does
2437 // not try to pass 'this' straight through
2438 isThisReturn = false;
2439 Mask = ARI->getCallPreservedMask(MF, CallConv);
2442 Mask = ARI->getCallPreservedMask(MF, CallConv);
2444 assert(Mask && "Missing call preserved mask for calling convention");
2445 Ops.push_back(DAG.getRegisterMask(Mask));
2448 if (InFlag.getNode())
2449 Ops.push_back(InFlag);
2451 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2453 MF.getFrameInfo().setHasTailCall();
2454 SDValue Ret = DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, Ops);
2455 DAG.addCallSiteInfo(Ret.getNode(), std::move(CSInfo));
2459 // Returns a chain and a flag for retval copy to use.
2460 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
2461 InFlag = Chain.getValue(1);
2462 DAG.addCallSiteInfo(Chain.getNode(), std::move(CSInfo));
2464 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, dl, true),
2465 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
2467 InFlag = Chain.getValue(1);
2469 // Handle result values, copying them out of physregs into vregs that we
2471 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
2472 InVals, isThisReturn,
2473 isThisReturn ? OutVals[0] : SDValue());
2476 /// HandleByVal - Every parameter *after* a byval parameter is passed
2477 /// on the stack. Remember the next parameter register to allocate,
2478 /// and then confiscate the rest of the parameter registers to insure
2480 void ARMTargetLowering::HandleByVal(CCState *State, unsigned &Size,
2481 unsigned Align) const {
2482 // Byval (as with any stack) slots are always at least 4 byte aligned.
2483 Align = std::max(Align, 4U);
2485 unsigned Reg = State->AllocateReg(GPRArgRegs);
2489 unsigned AlignInRegs = Align / 4;
2490 unsigned Waste = (ARM::R4 - Reg) % AlignInRegs;
2491 for (unsigned i = 0; i < Waste; ++i)
2492 Reg = State->AllocateReg(GPRArgRegs);
2497 unsigned Excess = 4 * (ARM::R4 - Reg);
2499 // Special case when NSAA != SP and parameter size greater than size of
2500 // all remained GPR regs. In that case we can't split parameter, we must
2501 // send it to stack. We also must set NCRN to R4, so waste all
2502 // remained registers.
2503 const unsigned NSAAOffset = State->getNextStackOffset();
2504 if (NSAAOffset != 0 && Size > Excess) {
2505 while (State->AllocateReg(GPRArgRegs))
2510 // First register for byval parameter is the first register that wasn't
2511 // allocated before this method call, so it would be "reg".
2512 // If parameter is small enough to be saved in range [reg, r4), then
2513 // the end (first after last) register would be reg + param-size-in-regs,
2514 // else parameter would be splitted between registers and stack,
2515 // end register would be r4 in this case.
2516 unsigned ByValRegBegin = Reg;
2517 unsigned ByValRegEnd = std::min<unsigned>(Reg + Size / 4, ARM::R4);
2518 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
2519 // Note, first register is allocated in the beginning of function already,
2520 // allocate remained amount of registers we need.
2521 for (unsigned i = Reg + 1; i != ByValRegEnd; ++i)
2522 State->AllocateReg(GPRArgRegs);
2523 // A byval parameter that is split between registers and memory needs its
2524 // size truncated here.
2525 // In the case where the entire structure fits in registers, we set the
2526 // size in memory to zero.
2527 Size = std::max<int>(Size - Excess, 0);
2530 /// MatchingStackOffset - Return true if the given stack call argument is
2531 /// already available in the same position (relatively) of the caller's
2532 /// incoming argument stack.
2534 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2535 MachineFrameInfo &MFI, const MachineRegisterInfo *MRI,
2536 const TargetInstrInfo *TII) {
2537 unsigned Bytes = Arg.getValueSizeInBits() / 8;
2538 int FI = std::numeric_limits<int>::max();
2539 if (Arg.getOpcode() == ISD::CopyFromReg) {
2540 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2541 if (!Register::isVirtualRegister(VR))
2543 MachineInstr *Def = MRI->getVRegDef(VR);
2546 if (!Flags.isByVal()) {
2547 if (!TII->isLoadFromStackSlot(*Def, FI))
2552 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2553 if (Flags.isByVal())
2554 // ByVal argument is passed in as a pointer but it's now being
2555 // dereferenced. e.g.
2556 // define @foo(%struct.X* %A) {
2557 // tail call @bar(%struct.X* byval %A)
2560 SDValue Ptr = Ld->getBasePtr();
2561 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2564 FI = FINode->getIndex();
2568 assert(FI != std::numeric_limits<int>::max());
2569 if (!MFI.isFixedObjectIndex(FI))
2571 return Offset == MFI.getObjectOffset(FI) && Bytes == MFI.getObjectSize(FI);
2574 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2575 /// for tail call optimization. Targets which want to do tail call
2576 /// optimization should implement this function.
2577 bool ARMTargetLowering::IsEligibleForTailCallOptimization(
2578 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
2579 bool isCalleeStructRet, bool isCallerStructRet,
2580 const SmallVectorImpl<ISD::OutputArg> &Outs,
2581 const SmallVectorImpl<SDValue> &OutVals,
2582 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG,
2583 const bool isIndirect) const {
2584 MachineFunction &MF = DAG.getMachineFunction();
2585 const Function &CallerF = MF.getFunction();
2586 CallingConv::ID CallerCC = CallerF.getCallingConv();
2588 assert(Subtarget->supportsTailCall());
2590 // Indirect tail calls cannot be optimized for Thumb1 if the args
2591 // to the call take up r0-r3. The reason is that there are no legal registers
2592 // left to hold the pointer to the function to be called.
2593 if (Subtarget->isThumb1Only() && Outs.size() >= 4 &&
2594 (!isa<GlobalAddressSDNode>(Callee.getNode()) || isIndirect))
2597 // Look for obvious safe cases to perform tail call optimization that do not
2598 // require ABI changes. This is what gcc calls sibcall.
2600 // Exception-handling functions need a special set of instructions to indicate
2601 // a return to the hardware. Tail-calling another function would probably
2603 if (CallerF.hasFnAttribute("interrupt"))
2606 // Also avoid sibcall optimization if either caller or callee uses struct
2607 // return semantics.
2608 if (isCalleeStructRet || isCallerStructRet)
2611 // Externally-defined functions with weak linkage should not be
2612 // tail-called on ARM when the OS does not support dynamic
2613 // pre-emption of symbols, as the AAELF spec requires normal calls
2614 // to undefined weak functions to be replaced with a NOP or jump to the
2615 // next instruction. The behaviour of branch instructions in this
2616 // situation (as used for tail calls) is implementation-defined, so we
2617 // cannot rely on the linker replacing the tail call with a return.
2618 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2619 const GlobalValue *GV = G->getGlobal();
2620 const Triple &TT = getTargetMachine().getTargetTriple();
2621 if (GV->hasExternalWeakLinkage() &&
2622 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2626 // Check that the call results are passed in the same way.
2627 LLVMContext &C = *DAG.getContext();
2628 if (!CCState::resultsCompatible(CalleeCC, CallerCC, MF, C, Ins,
2629 CCAssignFnForReturn(CalleeCC, isVarArg),
2630 CCAssignFnForReturn(CallerCC, isVarArg)))
2632 // The callee has to preserve all registers the caller needs to preserve.
2633 const ARMBaseRegisterInfo *TRI = Subtarget->getRegisterInfo();
2634 const uint32_t *CallerPreserved = TRI->getCallPreservedMask(MF, CallerCC);
2635 if (CalleeCC != CallerCC) {
2636 const uint32_t *CalleePreserved = TRI->getCallPreservedMask(MF, CalleeCC);
2637 if (!TRI->regmaskSubsetEqual(CallerPreserved, CalleePreserved))
2641 // If Caller's vararg or byval argument has been split between registers and
2642 // stack, do not perform tail call, since part of the argument is in caller's
2644 const ARMFunctionInfo *AFI_Caller = MF.getInfo<ARMFunctionInfo>();
2645 if (AFI_Caller->getArgRegsSaveSize())
2648 // If the callee takes no arguments then go on to check the results of the
2650 if (!Outs.empty()) {
2651 // Check if stack adjustment is needed. For now, do not do this if any
2652 // argument is passed on the stack.
2653 SmallVector<CCValAssign, 16> ArgLocs;
2654 CCState CCInfo(CalleeCC, isVarArg, MF, ArgLocs, C);
2655 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, isVarArg));
2656 if (CCInfo.getNextStackOffset()) {
2657 // Check if the arguments are already laid out in the right way as
2658 // the caller's fixed stack objects.
2659 MachineFrameInfo &MFI = MF.getFrameInfo();
2660 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2661 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
2662 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2664 ++i, ++realArgIdx) {
2665 CCValAssign &VA = ArgLocs[i];
2666 EVT RegVT = VA.getLocVT();
2667 SDValue Arg = OutVals[realArgIdx];
2668 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2669 if (VA.getLocInfo() == CCValAssign::Indirect)
2671 if (VA.needsCustom()) {
2672 // f64 and vector types are split into multiple registers or
2673 // register/stack-slot combinations. The types will not match
2674 // the registers; give up on memory f64 refs until we figure
2675 // out what to do about this.
2678 if (!ArgLocs[++i].isRegLoc())
2680 if (RegVT == MVT::v2f64) {
2681 if (!ArgLocs[++i].isRegLoc())
2683 if (!ArgLocs[++i].isRegLoc())
2686 } else if (!VA.isRegLoc()) {
2687 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2694 const MachineRegisterInfo &MRI = MF.getRegInfo();
2695 if (!parametersInCSRMatch(MRI, CallerPreserved, ArgLocs, OutVals))
2703 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2704 MachineFunction &MF, bool isVarArg,
2705 const SmallVectorImpl<ISD::OutputArg> &Outs,
2706 LLVMContext &Context) const {
2707 SmallVector<CCValAssign, 16> RVLocs;
2708 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2709 return CCInfo.CheckReturn(Outs, CCAssignFnForReturn(CallConv, isVarArg));
2712 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2713 const SDLoc &DL, SelectionDAG &DAG) {
2714 const MachineFunction &MF = DAG.getMachineFunction();
2715 const Function &F = MF.getFunction();
2717 StringRef IntKind = F.getFnAttribute("interrupt").getValueAsString();
2719 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2720 // version of the "preferred return address". These offsets affect the return
2721 // instruction if this is a return from PL1 without hypervisor extensions.
2722 // IRQ/FIQ: +4 "subs pc, lr, #4"
2723 // SWI: 0 "subs pc, lr, #0"
2724 // ABORT: +4 "subs pc, lr, #4"
2725 // UNDEF: +4/+2 "subs pc, lr, #0"
2726 // UNDEF varies depending on where the exception came from ARM or Thumb
2727 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2730 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2733 else if (IntKind == "SWI" || IntKind == "UNDEF")
2736 report_fatal_error("Unsupported interrupt attribute. If present, value "
2737 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2739 RetOps.insert(RetOps.begin() + 1,
2740 DAG.getConstant(LROffset, DL, MVT::i32, false));
2742 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other, RetOps);
2746 ARMTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2748 const SmallVectorImpl<ISD::OutputArg> &Outs,
2749 const SmallVectorImpl<SDValue> &OutVals,
2750 const SDLoc &dl, SelectionDAG &DAG) const {
2751 // CCValAssign - represent the assignment of the return value to a location.
2752 SmallVector<CCValAssign, 16> RVLocs;
2754 // CCState - Info about the registers and stack slots.
2755 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2758 // Analyze outgoing return values.
2759 CCInfo.AnalyzeReturn(Outs, CCAssignFnForReturn(CallConv, isVarArg));
2762 SmallVector<SDValue, 4> RetOps;
2763 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2764 bool isLittleEndian = Subtarget->isLittle();
2766 MachineFunction &MF = DAG.getMachineFunction();
2767 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2768 AFI->setReturnRegsCount(RVLocs.size());
2770 // Copy the result values into the output registers.
2771 for (unsigned i = 0, realRVLocIdx = 0;
2773 ++i, ++realRVLocIdx) {
2774 CCValAssign &VA = RVLocs[i];
2775 assert(VA.isRegLoc() && "Can only return in registers!");
2777 SDValue Arg = OutVals[realRVLocIdx];
2778 bool ReturnF16 = false;
2780 if (Subtarget->hasFullFP16() && Subtarget->isTargetHardFloat()) {
2781 // Half-precision return values can be returned like this:
2783 // t11 f16 = fadd ...
2784 // t12: i16 = bitcast t11
2785 // t13: i32 = zero_extend t12
2786 // t14: f32 = bitcast t13 <~~~~~~~ Arg
2788 // to avoid code generation for bitcasts, we simply set Arg to the node
2789 // that produces the f16 value, t11 in this case.
2791 if (Arg.getValueType() == MVT::f32 && Arg.getOpcode() == ISD::BITCAST) {
2792 SDValue ZE = Arg.getOperand(0);
2793 if (ZE.getOpcode() == ISD::ZERO_EXTEND && ZE.getValueType() == MVT::i32) {
2794 SDValue BC = ZE.getOperand(0);
2795 if (BC.getOpcode() == ISD::BITCAST && BC.getValueType() == MVT::i16) {
2796 Arg = BC.getOperand(0);
2803 switch (VA.getLocInfo()) {
2804 default: llvm_unreachable("Unknown loc info!");
2805 case CCValAssign::Full: break;
2806 case CCValAssign::BCvt:
2808 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2812 if (VA.needsCustom()) {
2813 if (VA.getLocVT() == MVT::v2f64) {
2814 // Extract the first half and return it in two registers.
2815 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2816 DAG.getConstant(0, dl, MVT::i32));
2817 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2818 DAG.getVTList(MVT::i32, MVT::i32), Half);
2820 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2821 HalfGPRs.getValue(isLittleEndian ? 0 : 1),
2823 Flag = Chain.getValue(1);
2824 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2825 VA = RVLocs[++i]; // skip ahead to next loc
2826 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2827 HalfGPRs.getValue(isLittleEndian ? 1 : 0),
2829 Flag = Chain.getValue(1);
2830 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2831 VA = RVLocs[++i]; // skip ahead to next loc
2833 // Extract the 2nd half and fall through to handle it as an f64 value.
2834 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2835 DAG.getConstant(1, dl, MVT::i32));
2837 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2839 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2840 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2841 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2842 fmrrd.getValue(isLittleEndian ? 0 : 1),
2844 Flag = Chain.getValue(1);
2845 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2846 VA = RVLocs[++i]; // skip ahead to next loc
2847 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2848 fmrrd.getValue(isLittleEndian ? 1 : 0),
2851 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2853 // Guarantee that all emitted copies are
2854 // stuck together, avoiding something bad.
2855 Flag = Chain.getValue(1);
2856 RetOps.push_back(DAG.getRegister(VA.getLocReg(),
2857 ReturnF16 ? MVT::f16 : VA.getLocVT()));
2859 const ARMBaseRegisterInfo *TRI = Subtarget->getRegisterInfo();
2860 const MCPhysReg *I =
2861 TRI->getCalleeSavedRegsViaCopy(&DAG.getMachineFunction());
2864 if (ARM::GPRRegClass.contains(*I))
2865 RetOps.push_back(DAG.getRegister(*I, MVT::i32));
2866 else if (ARM::DPRRegClass.contains(*I))
2867 RetOps.push_back(DAG.getRegister(*I, MVT::getFloatingPointVT(64)));
2869 llvm_unreachable("Unexpected register class in CSRsViaCopy!");
2873 // Update chain and glue.
2876 RetOps.push_back(Flag);
2878 // CPUs which aren't M-class use a special sequence to return from
2879 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2880 // though we use "subs pc, lr, #N").
2882 // M-class CPUs actually use a normal return sequence with a special
2883 // (hardware-provided) value in LR, so the normal code path works.
2884 if (DAG.getMachineFunction().getFunction().hasFnAttribute("interrupt") &&
2885 !Subtarget->isMClass()) {
2886 if (Subtarget->isThumb1Only())
2887 report_fatal_error("interrupt attribute is not supported in Thumb1");
2888 return LowerInterruptReturn(RetOps, dl, DAG);
2891 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, RetOps);
2894 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2895 if (N->getNumValues() != 1)
2897 if (!N->hasNUsesOfValue(1, 0))
2900 SDValue TCChain = Chain;
2901 SDNode *Copy = *N->use_begin();
2902 if (Copy->getOpcode() == ISD::CopyToReg) {
2903 // If the copy has a glue operand, we conservatively assume it isn't safe to
2904 // perform a tail call.
2905 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2907 TCChain = Copy->getOperand(0);
2908 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2909 SDNode *VMov = Copy;
2910 // f64 returned in a pair of GPRs.
2911 SmallPtrSet<SDNode*, 2> Copies;
2912 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2914 if (UI->getOpcode() != ISD::CopyToReg)
2918 if (Copies.size() > 2)
2921 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2923 SDValue UseChain = UI->getOperand(0);
2924 if (Copies.count(UseChain.getNode()))
2928 // We are at the top of this chain.
2929 // If the copy has a glue operand, we conservatively assume it
2930 // isn't safe to perform a tail call.
2931 if (UI->getOperand(UI->getNumOperands()-1).getValueType() == MVT::Glue)
2937 } else if (Copy->getOpcode() == ISD::BITCAST) {
2938 // f32 returned in a single GPR.
2939 if (!Copy->hasOneUse())
2941 Copy = *Copy->use_begin();
2942 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2944 // If the copy has a glue operand, we conservatively assume it isn't safe to
2945 // perform a tail call.
2946 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2948 TCChain = Copy->getOperand(0);
2953 bool HasRet = false;
2954 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2956 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2957 UI->getOpcode() != ARMISD::INTRET_FLAG)
2969 bool ARMTargetLowering::mayBeEmittedAsTailCall(const CallInst *CI) const {
2970 if (!Subtarget->supportsTailCall())
2973 if (!CI->isTailCall())
2979 // Trying to write a 64 bit value so need to split into two 32 bit values first,
2980 // and pass the lower and high parts through.
2981 static SDValue LowerWRITE_REGISTER(SDValue Op, SelectionDAG &DAG) {
2983 SDValue WriteValue = Op->getOperand(2);
2985 // This function is only supposed to be called for i64 type argument.
2986 assert(WriteValue.getValueType() == MVT::i64
2987 && "LowerWRITE_REGISTER called for non-i64 type argument.");
2989 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, WriteValue,
2990 DAG.getConstant(0, DL, MVT::i32));
2991 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, WriteValue,
2992 DAG.getConstant(1, DL, MVT::i32));
2993 SDValue Ops[] = { Op->getOperand(0), Op->getOperand(1), Lo, Hi };
2994 return DAG.getNode(ISD::WRITE_REGISTER, DL, MVT::Other, Ops);
2997 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2998 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2999 // one of the above mentioned nodes. It has to be wrapped because otherwise
3000 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
3001 // be used to form addressing mode. These wrapped nodes will be selected
3003 SDValue ARMTargetLowering::LowerConstantPool(SDValue Op,
3004 SelectionDAG &DAG) const {
3005 EVT PtrVT = Op.getValueType();
3006 // FIXME there is no actual debug info here
3008 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3011 // When generating execute-only code Constant Pools must be promoted to the
3012 // global data section. It's a bit ugly that we can't share them across basic
3013 // blocks, but this way we guarantee that execute-only behaves correct with
3014 // position-independent addressing modes.
3015 if (Subtarget->genExecuteOnly()) {
3016 auto AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3017 auto T = const_cast<Type*>(CP->getType());
3018 auto C = const_cast<Constant*>(CP->getConstVal());
3019 auto M = const_cast<Module*>(DAG.getMachineFunction().
3020 getFunction().getParent());
3021 auto GV = new GlobalVariable(
3022 *M, T, /*isConstant=*/true, GlobalVariable::InternalLinkage, C,
3023 Twine(DAG.getDataLayout().getPrivateGlobalPrefix()) + "CP" +
3024 Twine(DAG.getMachineFunction().getFunctionNumber()) + "_" +
3025 Twine(AFI->createPICLabelUId())
3027 SDValue GA = DAG.getTargetGlobalAddress(dyn_cast<GlobalValue>(GV),
3029 return LowerGlobalAddress(GA, DAG);
3032 if (CP->isMachineConstantPoolEntry())
3033 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
3034 CP->getAlignment());
3036 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
3037 CP->getAlignment());
3038 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
3041 unsigned ARMTargetLowering::getJumpTableEncoding() const {
3042 return MachineJumpTableInfo::EK_Inline;
3045 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
3046 SelectionDAG &DAG) const {
3047 MachineFunction &MF = DAG.getMachineFunction();
3048 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3049 unsigned ARMPCLabelIndex = 0;
3051 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3052 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
3054 bool IsPositionIndependent = isPositionIndependent() || Subtarget->isROPI();
3055 if (!IsPositionIndependent) {
3056 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
3058 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
3059 ARMPCLabelIndex = AFI->createPICLabelUId();
3060 ARMConstantPoolValue *CPV =
3061 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
3062 ARMCP::CPBlockAddress, PCAdj);
3063 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
3065 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
3066 SDValue Result = DAG.getLoad(
3067 PtrVT, DL, DAG.getEntryNode(), CPAddr,
3068 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()));
3069 if (!IsPositionIndependent)
3071 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, DL, MVT::i32);
3072 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
3075 /// Convert a TLS address reference into the correct sequence of loads
3076 /// and calls to compute the variable's address for Darwin, and return an
3077 /// SDValue containing the final node.
3079 /// Darwin only has one TLS scheme which must be capable of dealing with the
3080 /// fully general situation, in the worst case. This means:
3081 /// + "extern __thread" declaration.
3082 /// + Defined in a possibly unknown dynamic library.
3084 /// The general system is that each __thread variable has a [3 x i32] descriptor
3085 /// which contains information used by the runtime to calculate the address. The
3086 /// only part of this the compiler needs to know about is the first word, which
3087 /// contains a function pointer that must be called with the address of the
3088 /// entire descriptor in "r0".
3090 /// Since this descriptor may be in a different unit, in general access must
3091 /// proceed along the usual ARM rules. A common sequence to produce is:
3093 /// movw rT1, :lower16:_var$non_lazy_ptr
3094 /// movt rT1, :upper16:_var$non_lazy_ptr
3098 /// [...address now in r0...]
3100 ARMTargetLowering::LowerGlobalTLSAddressDarwin(SDValue Op,
3101 SelectionDAG &DAG) const {
3102 assert(Subtarget->isTargetDarwin() &&
3103 "This function expects a Darwin target");
3106 // First step is to get the address of the actua global symbol. This is where
3107 // the TLS descriptor lives.
3108 SDValue DescAddr = LowerGlobalAddressDarwin(Op, DAG);
3110 // The first entry in the descriptor is a function pointer that we must call
3111 // to obtain the address of the variable.
3112 SDValue Chain = DAG.getEntryNode();
3113 SDValue FuncTLVGet = DAG.getLoad(
3114 MVT::i32, DL, Chain, DescAddr,
3115 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
3116 /* Alignment = */ 4,
3117 MachineMemOperand::MONonTemporal | MachineMemOperand::MODereferenceable |
3118 MachineMemOperand::MOInvariant);
3119 Chain = FuncTLVGet.getValue(1);
3121 MachineFunction &F = DAG.getMachineFunction();
3122 MachineFrameInfo &MFI = F.getFrameInfo();
3123 MFI.setAdjustsStack(true);
3125 // TLS calls preserve all registers except those that absolutely must be
3126 // trashed: R0 (it takes an argument), LR (it's a call) and CPSR (let's not be
3129 getTargetMachine().getSubtargetImpl(F.getFunction())->getRegisterInfo();
3130 auto ARI = static_cast<const ARMRegisterInfo *>(TRI);
3131 const uint32_t *Mask = ARI->getTLSCallPreservedMask(DAG.getMachineFunction());
3133 // Finally, we can make the call. This is just a degenerate version of a
3134 // normal AArch64 call node: r0 takes the address of the descriptor, and
3135 // returns the address of the variable in this thread.
3136 Chain = DAG.getCopyToReg(Chain, DL, ARM::R0, DescAddr, SDValue());
3138 DAG.getNode(ARMISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
3139 Chain, FuncTLVGet, DAG.getRegister(ARM::R0, MVT::i32),
3140 DAG.getRegisterMask(Mask), Chain.getValue(1));
3141 return DAG.getCopyFromReg(Chain, DL, ARM::R0, MVT::i32, Chain.getValue(1));
3145 ARMTargetLowering::LowerGlobalTLSAddressWindows(SDValue Op,
3146 SelectionDAG &DAG) const {
3147 assert(Subtarget->isTargetWindows() && "Windows specific TLS lowering");
3149 SDValue Chain = DAG.getEntryNode();
3150 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3153 // Load the current TEB (thread environment block)
3154 SDValue Ops[] = {Chain,
3155 DAG.getTargetConstant(Intrinsic::arm_mrc, DL, MVT::i32),
3156 DAG.getTargetConstant(15, DL, MVT::i32),
3157 DAG.getTargetConstant(0, DL, MVT::i32),
3158 DAG.getTargetConstant(13, DL, MVT::i32),
3159 DAG.getTargetConstant(0, DL, MVT::i32),
3160 DAG.getTargetConstant(2, DL, MVT::i32)};
3161 SDValue CurrentTEB = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
3162 DAG.getVTList(MVT::i32, MVT::Other), Ops);
3164 SDValue TEB = CurrentTEB.getValue(0);
3165 Chain = CurrentTEB.getValue(1);
3167 // Load the ThreadLocalStoragePointer from the TEB
3168 // A pointer to the TLS array is located at offset 0x2c from the TEB.
3170 DAG.getNode(ISD::ADD, DL, PtrVT, TEB, DAG.getIntPtrConstant(0x2c, DL));
3171 TLSArray = DAG.getLoad(PtrVT, DL, Chain, TLSArray, MachinePointerInfo());
3173 // The pointer to the thread's TLS data area is at the TLS Index scaled by 4
3174 // offset into the TLSArray.
3176 // Load the TLS index from the C runtime
3178 DAG.getTargetExternalSymbol("_tls_index", PtrVT, ARMII::MO_NO_FLAG);
3179 TLSIndex = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, TLSIndex);
3180 TLSIndex = DAG.getLoad(PtrVT, DL, Chain, TLSIndex, MachinePointerInfo());
3182 SDValue Slot = DAG.getNode(ISD::SHL, DL, PtrVT, TLSIndex,
3183 DAG.getConstant(2, DL, MVT::i32));
3184 SDValue TLS = DAG.getLoad(PtrVT, DL, Chain,
3185 DAG.getNode(ISD::ADD, DL, PtrVT, TLSArray, Slot),
3186 MachinePointerInfo());
3188 // Get the offset of the start of the .tls section (section base)
3189 const auto *GA = cast<GlobalAddressSDNode>(Op);
3190 auto *CPV = ARMConstantPoolConstant::Create(GA->getGlobal(), ARMCP::SECREL);
3191 SDValue Offset = DAG.getLoad(
3192 PtrVT, DL, Chain, DAG.getNode(ARMISD::Wrapper, DL, MVT::i32,
3193 DAG.getTargetConstantPool(CPV, PtrVT, 4)),
3194 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()));
3196 return DAG.getNode(ISD::ADD, DL, PtrVT, TLS, Offset);
3199 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
3201 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
3202 SelectionDAG &DAG) const {
3204 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3205 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
3206 MachineFunction &MF = DAG.getMachineFunction();
3207 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3208 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
3209 ARMConstantPoolValue *CPV =
3210 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
3211 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
3212 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
3213 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
3214 Argument = DAG.getLoad(
3215 PtrVT, dl, DAG.getEntryNode(), Argument,
3216 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()));
3217 SDValue Chain = Argument.getValue(1);
3219 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
3220 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
3222 // call __tls_get_addr.
3225 Entry.Node = Argument;
3226 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
3227 Args.push_back(Entry);
3229 // FIXME: is there useful debug info available here?
3230 TargetLowering::CallLoweringInfo CLI(DAG);
3231 CLI.setDebugLoc(dl).setChain(Chain).setLibCallee(
3232 CallingConv::C, Type::getInt32Ty(*DAG.getContext()),
3233 DAG.getExternalSymbol("__tls_get_addr", PtrVT), std::move(Args));
3235 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
3236 return CallResult.first;
3239 // Lower ISD::GlobalTLSAddress using the "initial exec" or
3240 // "local exec" model.
3242 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
3244 TLSModel::Model model) const {
3245 const GlobalValue *GV = GA->getGlobal();
3248 SDValue Chain = DAG.getEntryNode();
3249 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3250 // Get the Thread Pointer
3251 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
3253 if (model == TLSModel::InitialExec) {
3254 MachineFunction &MF = DAG.getMachineFunction();
3255 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3256 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
3257 // Initial exec model.
3258 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
3259 ARMConstantPoolValue *CPV =
3260 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
3261 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
3263 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
3264 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
3265 Offset = DAG.getLoad(
3266 PtrVT, dl, Chain, Offset,
3267 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()));
3268 Chain = Offset.getValue(1);
3270 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
3271 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
3273 Offset = DAG.getLoad(
3274 PtrVT, dl, Chain, Offset,
3275 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()));
3278 assert(model == TLSModel::LocalExec);
3279 ARMConstantPoolValue *CPV =
3280 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
3281 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
3282 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
3283 Offset = DAG.getLoad(
3284 PtrVT, dl, Chain, Offset,
3285 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()));
3288 // The address of the thread local variable is the add of the thread
3289 // pointer with the offset of the variable.
3290 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
3294 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
3295 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3296 if (DAG.getTarget().useEmulatedTLS())
3297 return LowerToTLSEmulatedModel(GA, DAG);
3299 if (Subtarget->isTargetDarwin())
3300 return LowerGlobalTLSAddressDarwin(Op, DAG);
3302 if (Subtarget->isTargetWindows())
3303 return LowerGlobalTLSAddressWindows(Op, DAG);
3305 // TODO: implement the "local dynamic" model
3306 assert(Subtarget->isTargetELF() && "Only ELF implemented here");
3307 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
3310 case TLSModel::GeneralDynamic:
3311 case TLSModel::LocalDynamic:
3312 return LowerToTLSGeneralDynamicModel(GA, DAG);
3313 case TLSModel::InitialExec:
3314 case TLSModel::LocalExec:
3315 return LowerToTLSExecModels(GA, DAG, model);
3317 llvm_unreachable("bogus TLS model");
3320 /// Return true if all users of V are within function F, looking through
3322 static bool allUsersAreInFunction(const Value *V, const Function *F) {
3323 SmallVector<const User*,4> Worklist;
3324 for (auto *U : V->users())
3325 Worklist.push_back(U);
3326 while (!Worklist.empty()) {
3327 auto *U = Worklist.pop_back_val();
3328 if (isa<ConstantExpr>(U)) {
3329 for (auto *UU : U->users())
3330 Worklist.push_back(UU);
3334 auto *I = dyn_cast<Instruction>(U);
3335 if (!I || I->getParent()->getParent() != F)
3341 static SDValue promoteToConstantPool(const ARMTargetLowering *TLI,
3342 const GlobalValue *GV, SelectionDAG &DAG,
3343 EVT PtrVT, const SDLoc &dl) {
3344 // If we're creating a pool entry for a constant global with unnamed address,
3345 // and the global is small enough, we can emit it inline into the constant pool
3346 // to save ourselves an indirection.
3348 // This is a win if the constant is only used in one function (so it doesn't
3349 // need to be duplicated) or duplicating the constant wouldn't increase code
3350 // size (implying the constant is no larger than 4 bytes).
3351 const Function &F = DAG.getMachineFunction().getFunction();
3353 // We rely on this decision to inline being idemopotent and unrelated to the
3354 // use-site. We know that if we inline a variable at one use site, we'll
3355 // inline it elsewhere too (and reuse the constant pool entry). Fast-isel
3356 // doesn't know about this optimization, so bail out if it's enabled else
3357 // we could decide to inline here (and thus never emit the GV) but require
3358 // the GV from fast-isel generated code.
3359 if (!EnableConstpoolPromotion ||
3360 DAG.getMachineFunction().getTarget().Options.EnableFastISel)
3363 auto *GVar = dyn_cast<GlobalVariable>(GV);
3364 if (!GVar || !GVar->hasInitializer() ||
3365 !GVar->isConstant() || !GVar->hasGlobalUnnamedAddr() ||
3366 !GVar->hasLocalLinkage())
3369 // If we inline a value that contains relocations, we move the relocations
3370 // from .data to .text. This is not allowed in position-independent code.
3371 auto *Init = GVar->getInitializer();
3372 if ((TLI->isPositionIndependent() || TLI->getSubtarget()->isROPI()) &&
3373 Init->needsRelocation())
3376 // The constant islands pass can only really deal with alignment requests
3377 // <= 4 bytes and cannot pad constants itself. Therefore we cannot promote
3378 // any type wanting greater alignment requirements than 4 bytes. We also
3379 // can only promote constants that are multiples of 4 bytes in size or
3380 // are paddable to a multiple of 4. Currently we only try and pad constants
3381 // that are strings for simplicity.
3382 auto *CDAInit = dyn_cast<ConstantDataArray>(Init);
3383 unsigned Size = DAG.getDataLayout().getTypeAllocSize(Init->getType());
3384 unsigned Align = DAG.getDataLayout().getPreferredAlignment(GVar);
3385 unsigned RequiredPadding = 4 - (Size % 4);
3386 bool PaddingPossible =
3387 RequiredPadding == 4 || (CDAInit && CDAInit->isString());
3388 if (!PaddingPossible || Align > 4 || Size > ConstpoolPromotionMaxSize ||
3392 unsigned PaddedSize = Size + ((RequiredPadding == 4) ? 0 : RequiredPadding);
3393 MachineFunction &MF = DAG.getMachineFunction();
3394 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3396 // We can't bloat the constant pool too much, else the ConstantIslands pass
3397 // may fail to converge. If we haven't promoted this global yet (it may have
3398 // multiple uses), and promoting it would increase the constant pool size (Sz
3399 // > 4), ensure we have space to do so up to MaxTotal.
3400 if (!AFI->getGlobalsPromotedToConstantPool().count(GVar) && Size > 4)
3401 if (AFI->getPromotedConstpoolIncrease() + PaddedSize - 4 >=
3402 ConstpoolPromotionMaxTotal)
3405 // This is only valid if all users are in a single function; we can't clone
3406 // the constant in general. The LLVM IR unnamed_addr allows merging
3407 // constants, but not cloning them.
3409 // We could potentially allow cloning if we could prove all uses of the
3410 // constant in the current function don't care about the address, like
3411 // printf format strings. But that isn't implemented for now.
3412 if (!allUsersAreInFunction(GVar, &F))
3415 // We're going to inline this global. Pad it out if needed.
3416 if (RequiredPadding != 4) {
3417 StringRef S = CDAInit->getAsString();
3419 SmallVector<uint8_t,16> V(S.size());
3420 std::copy(S.bytes_begin(), S.bytes_end(), V.begin());
3421 while (RequiredPadding--)
3423 Init = ConstantDataArray::get(*DAG.getContext(), V);
3426 auto CPVal = ARMConstantPoolConstant::Create(GVar, Init);
3428 DAG.getTargetConstantPool(CPVal, PtrVT, /*Align=*/4);
3429 if (!AFI->getGlobalsPromotedToConstantPool().count(GVar)) {
3430 AFI->markGlobalAsPromotedToConstantPool(GVar);
3431 AFI->setPromotedConstpoolIncrease(AFI->getPromotedConstpoolIncrease() +
3434 ++NumConstpoolPromoted;
3435 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
3438 bool ARMTargetLowering::isReadOnly(const GlobalValue *GV) const {
3439 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
3440 if (!(GV = GA->getBaseObject()))
3442 if (const auto *V = dyn_cast<GlobalVariable>(GV))
3443 return V->isConstant();
3444 return isa<Function>(GV);
3447 SDValue ARMTargetLowering::LowerGlobalAddress(SDValue Op,
3448 SelectionDAG &DAG) const {
3449 switch (Subtarget->getTargetTriple().getObjectFormat()) {
3450 default: llvm_unreachable("unknown object format");
3452 return LowerGlobalAddressWindows(Op, DAG);
3454 return LowerGlobalAddressELF(Op, DAG);
3456 return LowerGlobalAddressDarwin(Op, DAG);
3460 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
3461 SelectionDAG &DAG) const {
3462 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3464 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3465 const TargetMachine &TM = getTargetMachine();
3466 bool IsRO = isReadOnly(GV);
3468 // promoteToConstantPool only if not generating XO text section
3469 if (TM.shouldAssumeDSOLocal(*GV->getParent(), GV) && !Subtarget->genExecuteOnly())
3470 if (SDValue V = promoteToConstantPool(this, GV, DAG, PtrVT, dl))
3473 if (isPositionIndependent()) {
3474 bool UseGOT_PREL = !TM.shouldAssumeDSOLocal(*GV->getParent(), GV);
3475 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
3476 UseGOT_PREL ? ARMII::MO_GOT : 0);
3477 SDValue Result = DAG.getNode(ARMISD::WrapperPIC, dl, PtrVT, G);
3480 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
3481 MachinePointerInfo::getGOT(DAG.getMachineFunction()));
3483 } else if (Subtarget->isROPI() && IsRO) {
3485 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT);
3486 SDValue Result = DAG.getNode(ARMISD::WrapperPIC, dl, PtrVT, G);
3488 } else if (Subtarget->isRWPI() && !IsRO) {
3491 if (Subtarget->useMovt()) {
3493 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_SBREL);
3494 RelAddr = DAG.getNode(ARMISD::Wrapper, dl, PtrVT, G);
3495 } else { // use literal pool for address constant
3496 ARMConstantPoolValue *CPV =
3497 ARMConstantPoolConstant::Create(GV, ARMCP::SBREL);
3498 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
3499 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
3500 RelAddr = DAG.getLoad(
3501 PtrVT, dl, DAG.getEntryNode(), CPAddr,
3502 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()));
3504 SDValue SB = DAG.getCopyFromReg(DAG.getEntryNode(), dl, ARM::R9, PtrVT);
3505 SDValue Result = DAG.getNode(ISD::ADD, dl, PtrVT, SB, RelAddr);
3509 // If we have T2 ops, we can materialize the address directly via movt/movw
3510 // pair. This is always cheaper.
3511 if (Subtarget->useMovt()) {
3513 // FIXME: Once remat is capable of dealing with instructions with register
3514 // operands, expand this into two nodes.
3515 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
3516 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
3518 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
3519 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
3521 PtrVT, dl, DAG.getEntryNode(), CPAddr,
3522 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()));
3526 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
3527 SelectionDAG &DAG) const {
3528 assert(!Subtarget->isROPI() && !Subtarget->isRWPI() &&
3529 "ROPI/RWPI not currently supported for Darwin");
3530 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3532 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3534 if (Subtarget->useMovt())
3537 // FIXME: Once remat is capable of dealing with instructions with register
3538 // operands, expand this into multiple nodes
3540 isPositionIndependent() ? ARMISD::WrapperPIC : ARMISD::Wrapper;
3542 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
3543 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
3545 if (Subtarget->isGVIndirectSymbol(GV))
3546 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
3547 MachinePointerInfo::getGOT(DAG.getMachineFunction()));
3551 SDValue ARMTargetLowering::LowerGlobalAddressWindows(SDValue Op,
3552 SelectionDAG &DAG) const {
3553 assert(Subtarget->isTargetWindows() && "non-Windows COFF is not supported");
3554 assert(Subtarget->useMovt() &&
3555 "Windows on ARM expects to use movw/movt");
3556 assert(!Subtarget->isROPI() && !Subtarget->isRWPI() &&
3557 "ROPI/RWPI not currently supported for Windows");
3559 const TargetMachine &TM = getTargetMachine();
3560 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3561 ARMII::TOF TargetFlags = ARMII::MO_NO_FLAG;
3562 if (GV->hasDLLImportStorageClass())
3563 TargetFlags = ARMII::MO_DLLIMPORT;
3564 else if (!TM.shouldAssumeDSOLocal(*GV->getParent(), GV))
3565 TargetFlags = ARMII::MO_COFFSTUB;
3566 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3572 // FIXME: Once remat is capable of dealing with instructions with register
3573 // operands, expand this into two nodes.
3574 Result = DAG.getNode(ARMISD::Wrapper, DL, PtrVT,
3575 DAG.getTargetGlobalAddress(GV, DL, PtrVT, /*offset=*/0,
3577 if (TargetFlags & (ARMII::MO_DLLIMPORT | ARMII::MO_COFFSTUB))
3578 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
3579 MachinePointerInfo::getGOT(DAG.getMachineFunction()));
3584 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
3586 SDValue Val = DAG.getConstant(0, dl, MVT::i32);
3587 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
3588 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
3589 Op.getOperand(1), Val);
3593 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
3595 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
3596 Op.getOperand(1), DAG.getConstant(0, dl, MVT::i32));
3599 SDValue ARMTargetLowering::LowerEH_SJLJ_SETUP_DISPATCH(SDValue Op,
3600 SelectionDAG &DAG) const {
3602 return DAG.getNode(ARMISD::EH_SJLJ_SETUP_DISPATCH, dl, MVT::Other,
3606 SDValue ARMTargetLowering::LowerINTRINSIC_VOID(
3607 SDValue Op, SelectionDAG &DAG, const ARMSubtarget *Subtarget) const {
3609 cast<ConstantSDNode>(
3610 Op.getOperand(Op.getOperand(0).getValueType() == MVT::Other))
3614 return SDValue(); // Don't custom lower most intrinsics.
3615 case Intrinsic::arm_gnu_eabi_mcount: {
3616 MachineFunction &MF = DAG.getMachineFunction();
3617 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3619 SDValue Chain = Op.getOperand(0);
3620 // call "\01__gnu_mcount_nc"
3621 const ARMBaseRegisterInfo *ARI = Subtarget->getRegisterInfo();
3622 const uint32_t *Mask =
3623 ARI->getCallPreservedMask(DAG.getMachineFunction(), CallingConv::C);
3624 assert(Mask && "Missing call preserved mask for calling convention");
3625 // Mark LR an implicit live-in.
3626 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3627 SDValue ReturnAddress =
3628 DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, PtrVT);
3629 std::vector<EVT> ResultTys = {MVT::Other, MVT::Glue};
3631 DAG.getTargetExternalSymbol("\01__gnu_mcount_nc", PtrVT, 0);
3632 SDValue RegisterMask = DAG.getRegisterMask(Mask);
3633 if (Subtarget->isThumb())
3636 ARM::tBL_PUSHLR, dl, ResultTys,
3637 {ReturnAddress, DAG.getTargetConstant(ARMCC::AL, dl, PtrVT),
3638 DAG.getRegister(0, PtrVT), Callee, RegisterMask, Chain}),
3641 DAG.getMachineNode(ARM::BL_PUSHLR, dl, ResultTys,
3642 {ReturnAddress, Callee, RegisterMask, Chain}),
3649 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
3650 const ARMSubtarget *Subtarget) const {
3651 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3654 default: return SDValue(); // Don't custom lower most intrinsics.
3655 case Intrinsic::thread_pointer: {
3656 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3657 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
3659 case Intrinsic::arm_cls: {
3660 const SDValue &Operand = Op.getOperand(1);
3661 const EVT VTy = Op.getValueType();
3663 DAG.getNode(ISD::SRA, dl, VTy, Operand, DAG.getConstant(31, dl, VTy));
3664 SDValue XOR = DAG.getNode(ISD::XOR, dl, VTy, SRA, Operand);
3666 DAG.getNode(ISD::SHL, dl, VTy, XOR, DAG.getConstant(1, dl, VTy));
3668 DAG.getNode(ISD::OR, dl, VTy, SHL, DAG.getConstant(1, dl, VTy));
3669 SDValue Result = DAG.getNode(ISD::CTLZ, dl, VTy, OR);
3672 case Intrinsic::arm_cls64: {
3673 // cls(x) = if cls(hi(x)) != 31 then cls(hi(x))
3674 // else 31 + clz(if hi(x) == 0 then lo(x) else not(lo(x)))
3675 const SDValue &Operand = Op.getOperand(1);
3676 const EVT VTy = Op.getValueType();
3678 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VTy, Operand,
3679 DAG.getConstant(1, dl, VTy));
3680 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VTy, Operand,
3681 DAG.getConstant(0, dl, VTy));
3682 SDValue Constant0 = DAG.getConstant(0, dl, VTy);
3683 SDValue Constant1 = DAG.getConstant(1, dl, VTy);
3684 SDValue Constant31 = DAG.getConstant(31, dl, VTy);
3685 SDValue SRAHi = DAG.getNode(ISD::SRA, dl, VTy, Hi, Constant31);
3686 SDValue XORHi = DAG.getNode(ISD::XOR, dl, VTy, SRAHi, Hi);
3687 SDValue SHLHi = DAG.getNode(ISD::SHL, dl, VTy, XORHi, Constant1);
3688 SDValue ORHi = DAG.getNode(ISD::OR, dl, VTy, SHLHi, Constant1);
3689 SDValue CLSHi = DAG.getNode(ISD::CTLZ, dl, VTy, ORHi);
3691 DAG.getSetCC(dl, MVT::i1, CLSHi, Constant31, ISD::CondCode::SETEQ);
3693 DAG.getSetCC(dl, MVT::i1, Hi, Constant0, ISD::CondCode::SETEQ);
3694 SDValue AdjustedLo =
3695 DAG.getSelect(dl, VTy, HiIsZero, Lo, DAG.getNOT(dl, Lo, VTy));
3696 SDValue CLZAdjustedLo = DAG.getNode(ISD::CTLZ, dl, VTy, AdjustedLo);
3698 DAG.getSelect(dl, VTy, CheckLo,
3699 DAG.getNode(ISD::ADD, dl, VTy, CLZAdjustedLo, Constant31), CLSHi);
3702 case Intrinsic::eh_sjlj_lsda: {
3703 MachineFunction &MF = DAG.getMachineFunction();
3704 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3705 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
3706 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3708 bool IsPositionIndependent = isPositionIndependent();
3709 unsigned PCAdj = IsPositionIndependent ? (Subtarget->isThumb() ? 4 : 8) : 0;
3710 ARMConstantPoolValue *CPV =
3711 ARMConstantPoolConstant::Create(&MF.getFunction(), ARMPCLabelIndex,
3712 ARMCP::CPLSDA, PCAdj);
3713 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
3714 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
3715 SDValue Result = DAG.getLoad(
3716 PtrVT, dl, DAG.getEntryNode(), CPAddr,
3717 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()));
3719 if (IsPositionIndependent) {
3720 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
3721 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
3725 case Intrinsic::arm_neon_vabs:
3726 return DAG.getNode(ISD::ABS, SDLoc(Op), Op.getValueType(),
3728 case Intrinsic::arm_neon_vmulls:
3729 case Intrinsic::arm_neon_vmullu: {
3730 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
3731 ? ARMISD::VMULLs : ARMISD::VMULLu;
3732 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
3733 Op.getOperand(1), Op.getOperand(2));
3735 case Intrinsic::arm_neon_vminnm:
3736 case Intrinsic::arm_neon_vmaxnm: {
3737 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vminnm)
3738 ? ISD::FMINNUM : ISD::FMAXNUM;
3739 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
3740 Op.getOperand(1), Op.getOperand(2));
3742 case Intrinsic::arm_neon_vminu:
3743 case Intrinsic::arm_neon_vmaxu: {
3744 if (Op.getValueType().isFloatingPoint())
3746 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vminu)
3747 ? ISD::UMIN : ISD::UMAX;
3748 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
3749 Op.getOperand(1), Op.getOperand(2));
3751 case Intrinsic::arm_neon_vmins:
3752 case Intrinsic::arm_neon_vmaxs: {
3753 // v{min,max}s is overloaded between signed integers and floats.
3754 if (!Op.getValueType().isFloatingPoint()) {
3755 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmins)
3756 ? ISD::SMIN : ISD::SMAX;
3757 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
3758 Op.getOperand(1), Op.getOperand(2));
3760 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmins)
3761 ? ISD::FMINIMUM : ISD::FMAXIMUM;
3762 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
3763 Op.getOperand(1), Op.getOperand(2));
3765 case Intrinsic::arm_neon_vtbl1:
3766 return DAG.getNode(ARMISD::VTBL1, SDLoc(Op), Op.getValueType(),
3767 Op.getOperand(1), Op.getOperand(2));
3768 case Intrinsic::arm_neon_vtbl2:
3769 return DAG.getNode(ARMISD::VTBL2, SDLoc(Op), Op.getValueType(),
3770 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
3771 case Intrinsic::arm_mve_pred_i2v:
3772 case Intrinsic::arm_mve_pred_v2i:
3773 return DAG.getNode(ARMISD::PREDICATE_CAST, SDLoc(Op), Op.getValueType(),
3778 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
3779 const ARMSubtarget *Subtarget) {
3781 ConstantSDNode *SSIDNode = cast<ConstantSDNode>(Op.getOperand(2));
3782 auto SSID = static_cast<SyncScope::ID>(SSIDNode->getZExtValue());
3783 if (SSID == SyncScope::SingleThread)
3786 if (!Subtarget->hasDataBarrier()) {
3787 // Some ARMv6 cpus can support data barriers with an mcr instruction.
3788 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
3790 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
3791 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
3792 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
3793 DAG.getConstant(0, dl, MVT::i32));
3796 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
3797 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
3798 ARM_MB::MemBOpt Domain = ARM_MB::ISH;
3799 if (Subtarget->isMClass()) {
3800 // Only a full system barrier exists in the M-class architectures.
3801 Domain = ARM_MB::SY;
3802 } else if (Subtarget->preferISHSTBarriers() &&
3803 Ord == AtomicOrdering::Release) {
3804 // Swift happens to implement ISHST barriers in a way that's compatible with
3805 // Release semantics but weaker than ISH so we'd be fools not to use
3806 // it. Beware: other processors probably don't!
3807 Domain = ARM_MB::ISHST;
3810 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
3811 DAG.getConstant(Intrinsic::arm_dmb, dl, MVT::i32),
3812 DAG.getConstant(Domain, dl, MVT::i32));
3815 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
3816 const ARMSubtarget *Subtarget) {
3817 // ARM pre v5TE and Thumb1 does not have preload instructions.
3818 if (!(Subtarget->isThumb2() ||
3819 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
3820 // Just preserve the chain.
3821 return Op.getOperand(0);
3824 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
3826 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
3827 // ARMv7 with MP extension has PLDW.
3828 return Op.getOperand(0);
3830 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
3831 if (Subtarget->isThumb()) {
3833 isRead = ~isRead & 1;
3834 isData = ~isData & 1;
3837 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
3838 Op.getOperand(1), DAG.getConstant(isRead, dl, MVT::i32),
3839 DAG.getConstant(isData, dl, MVT::i32));
3842 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
3843 MachineFunction &MF = DAG.getMachineFunction();
3844 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
3846 // vastart just stores the address of the VarArgsFrameIndex slot into the
3847 // memory location argument.
3849 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
3850 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
3851 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3852 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
3853 MachinePointerInfo(SV));
3856 SDValue ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA,
3857 CCValAssign &NextVA,
3860 const SDLoc &dl) const {
3861 MachineFunction &MF = DAG.getMachineFunction();
3862 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3864 const TargetRegisterClass *RC;
3865 if (AFI->isThumb1OnlyFunction())
3866 RC = &ARM::tGPRRegClass;
3868 RC = &ARM::GPRRegClass;
3870 // Transform the arguments stored in physical registers into virtual ones.
3871 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3872 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
3875 if (NextVA.isMemLoc()) {
3876 MachineFrameInfo &MFI = MF.getFrameInfo();
3877 int FI = MFI.CreateFixedObject(4, NextVA.getLocMemOffset(), true);
3879 // Create load node to retrieve arguments from the stack.
3880 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3881 ArgValue2 = DAG.getLoad(
3882 MVT::i32, dl, Root, FIN,
3883 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI));
3885 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
3886 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
3888 if (!Subtarget->isLittle())
3889 std::swap (ArgValue, ArgValue2);
3890 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
3893 // The remaining GPRs hold either the beginning of variable-argument
3894 // data, or the beginning of an aggregate passed by value (usually
3895 // byval). Either way, we allocate stack slots adjacent to the data
3896 // provided by our caller, and store the unallocated registers there.
3897 // If this is a variadic function, the va_list pointer will begin with
3898 // these values; otherwise, this reassembles a (byval) structure that
3899 // was split between registers and memory.
3900 // Return: The frame index registers were stored into.
3901 int ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
3902 const SDLoc &dl, SDValue &Chain,
3903 const Value *OrigArg,
3904 unsigned InRegsParamRecordIdx,
3905 int ArgOffset, unsigned ArgSize) const {
3906 // Currently, two use-cases possible:
3907 // Case #1. Non-var-args function, and we meet first byval parameter.
3908 // Setup first unallocated register as first byval register;
3909 // eat all remained registers
3910 // (these two actions are performed by HandleByVal method).
3911 // Then, here, we initialize stack frame with
3912 // "store-reg" instructions.
3913 // Case #2. Var-args function, that doesn't contain byval parameters.
3914 // The same: eat all remained unallocated registers,
3915 // initialize stack frame.
3917 MachineFunction &MF = DAG.getMachineFunction();
3918 MachineFrameInfo &MFI = MF.getFrameInfo();
3919 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3920 unsigned RBegin, REnd;
3921 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
3922 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
3924 unsigned RBeginIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
3925 RBegin = RBeginIdx == 4 ? (unsigned)ARM::R4 : GPRArgRegs[RBeginIdx];
3930 ArgOffset = -4 * (ARM::R4 - RBegin);
3932 auto PtrVT = getPointerTy(DAG.getDataLayout());
3933 int FrameIndex = MFI.CreateFixedObject(ArgSize, ArgOffset, false);
3934 SDValue FIN = DAG.getFrameIndex(FrameIndex, PtrVT);
3936 SmallVector<SDValue, 4> MemOps;
3937 const TargetRegisterClass *RC =
3938 AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
3940 for (unsigned Reg = RBegin, i = 0; Reg < REnd; ++Reg, ++i) {
3941 unsigned VReg = MF.addLiveIn(Reg, RC);
3942 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
3943 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
3944 MachinePointerInfo(OrigArg, 4 * i));
3945 MemOps.push_back(Store);
3946 FIN = DAG.getNode(ISD::ADD, dl, PtrVT, FIN, DAG.getConstant(4, dl, PtrVT));
3949 if (!MemOps.empty())
3950 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
3954 // Setup stack frame, the va_list pointer will start from.
3955 void ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
3956 const SDLoc &dl, SDValue &Chain,
3958 unsigned TotalArgRegsSaveSize,
3959 bool ForceMutable) const {
3960 MachineFunction &MF = DAG.getMachineFunction();
3961 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3963 // Try to store any remaining integer argument regs
3964 // to their spots on the stack so that they may be loaded by dereferencing
3965 // the result of va_next.
3966 // If there is no regs to be stored, just point address after last
3967 // argument passed via stack.
3968 int FrameIndex = StoreByValRegs(CCInfo, DAG, dl, Chain, nullptr,
3969 CCInfo.getInRegsParamsCount(),
3970 CCInfo.getNextStackOffset(),
3971 std::max(4U, TotalArgRegsSaveSize));
3972 AFI->setVarArgsFrameIndex(FrameIndex);
3975 SDValue ARMTargetLowering::LowerFormalArguments(
3976 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
3977 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
3978 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
3979 MachineFunction &MF = DAG.getMachineFunction();
3980 MachineFrameInfo &MFI = MF.getFrameInfo();
3982 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3984 // Assign locations to all of the incoming arguments.
3985 SmallVector<CCValAssign, 16> ArgLocs;
3986 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
3988 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForCall(CallConv, isVarArg));
3990 SmallVector<SDValue, 16> ArgValues;
3992 Function::const_arg_iterator CurOrigArg = MF.getFunction().arg_begin();
3993 unsigned CurArgIdx = 0;
3995 // Initially ArgRegsSaveSize is zero.
3996 // Then we increase this value each time we meet byval parameter.
3997 // We also increase this value in case of varargs function.
3998 AFI->setArgRegsSaveSize(0);
4000 // Calculate the amount of stack space that we need to allocate to store
4001 // byval and variadic arguments that are passed in registers.
4002 // We need to know this before we allocate the first byval or variadic
4003 // argument, as they will be allocated a stack slot below the CFA (Canonical
4004 // Frame Address, the stack pointer at entry to the function).
4005 unsigned ArgRegBegin = ARM::R4;
4006 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
4007 if (CCInfo.getInRegsParamsProcessed() >= CCInfo.getInRegsParamsCount())
4010 CCValAssign &VA = ArgLocs[i];
4011 unsigned Index = VA.getValNo();
4012 ISD::ArgFlagsTy Flags = Ins[Index].Flags;
4013 if (!Flags.isByVal())
4016 assert(VA.isMemLoc() && "unexpected byval pointer in reg");
4017 unsigned RBegin, REnd;
4018 CCInfo.getInRegsParamInfo(CCInfo.getInRegsParamsProcessed(), RBegin, REnd);
4019 ArgRegBegin = std::min(ArgRegBegin, RBegin);
4021 CCInfo.nextInRegsParam();
4023 CCInfo.rewindByValRegsInfo();
4025 int lastInsIndex = -1;
4026 if (isVarArg && MFI.hasVAStart()) {
4027 unsigned RegIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
4028 if (RegIdx != array_lengthof(GPRArgRegs))
4029 ArgRegBegin = std::min(ArgRegBegin, (unsigned)GPRArgRegs[RegIdx]);
4032 unsigned TotalArgRegsSaveSize = 4 * (ARM::R4 - ArgRegBegin);
4033 AFI->setArgRegsSaveSize(TotalArgRegsSaveSize);
4034 auto PtrVT = getPointerTy(DAG.getDataLayout());
4036 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
4037 CCValAssign &VA = ArgLocs[i];
4038 if (Ins[VA.getValNo()].isOrigArg()) {
4039 std::advance(CurOrigArg,
4040 Ins[VA.getValNo()].getOrigArgIndex() - CurArgIdx);
4041 CurArgIdx = Ins[VA.getValNo()].getOrigArgIndex();
4043 // Arguments stored in registers.
4044 if (VA.isRegLoc()) {
4045 EVT RegVT = VA.getLocVT();
4047 if (VA.needsCustom()) {
4048 // f64 and vector types are split up into multiple registers or
4049 // combinations of registers and stack slots.
4050 if (VA.getLocVT() == MVT::v2f64) {
4051 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
4053 VA = ArgLocs[++i]; // skip ahead to next loc
4055 if (VA.isMemLoc()) {
4056 int FI = MFI.CreateFixedObject(8, VA.getLocMemOffset(), true);
4057 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
4058 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
4059 MachinePointerInfo::getFixedStack(
4060 DAG.getMachineFunction(), FI));
4062 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
4065 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
4066 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
4067 ArgValue, ArgValue1,
4068 DAG.getIntPtrConstant(0, dl));
4069 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
4070 ArgValue, ArgValue2,
4071 DAG.getIntPtrConstant(1, dl));
4073 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
4075 const TargetRegisterClass *RC;
4078 if (RegVT == MVT::f16)
4079 RC = &ARM::HPRRegClass;
4080 else if (RegVT == MVT::f32)
4081 RC = &ARM::SPRRegClass;
4082 else if (RegVT == MVT::f64 || RegVT == MVT::v4f16)
4083 RC = &ARM::DPRRegClass;
4084 else if (RegVT == MVT::v2f64 || RegVT == MVT::v8f16)
4085 RC = &ARM::QPRRegClass;
4086 else if (RegVT == MVT::i32)
4087 RC = AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass
4088 : &ARM::GPRRegClass;
4090 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
4092 // Transform the arguments in physical registers into virtual ones.
4093 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
4094 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
4096 // If this value is passed in r0 and has the returned attribute (e.g.
4097 // C++ 'structors), record this fact for later use.
4098 if (VA.getLocReg() == ARM::R0 && Ins[VA.getValNo()].Flags.isReturned()) {
4099 AFI->setPreservesR0();
4103 // If this is an 8 or 16-bit value, it is really passed promoted
4104 // to 32 bits. Insert an assert[sz]ext to capture this, then
4105 // truncate to the right size.
4106 switch (VA.getLocInfo()) {
4107 default: llvm_unreachable("Unknown loc info!");
4108 case CCValAssign::Full: break;
4109 case CCValAssign::BCvt:
4110 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
4112 case CCValAssign::SExt:
4113 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
4114 DAG.getValueType(VA.getValVT()));
4115 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
4117 case CCValAssign::ZExt:
4118 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
4119 DAG.getValueType(VA.getValVT()));
4120 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
4124 InVals.push_back(ArgValue);
4125 } else { // VA.isRegLoc()
4127 assert(VA.isMemLoc());
4128 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
4130 int index = VA.getValNo();
4132 // Some Ins[] entries become multiple ArgLoc[] entries.
4133 // Process them only once.
4134 if (index != lastInsIndex)
4136 ISD::ArgFlagsTy Flags = Ins[index].Flags;
4137 // FIXME: For now, all byval parameter objects are marked mutable.
4138 // This can be changed with more analysis.
4139 // In case of tail call optimization mark all arguments mutable.
4140 // Since they could be overwritten by lowering of arguments in case of
4142 if (Flags.isByVal()) {
4143 assert(Ins[index].isOrigArg() &&
4144 "Byval arguments cannot be implicit");
4145 unsigned CurByValIndex = CCInfo.getInRegsParamsProcessed();
4147 int FrameIndex = StoreByValRegs(
4148 CCInfo, DAG, dl, Chain, &*CurOrigArg, CurByValIndex,
4149 VA.getLocMemOffset(), Flags.getByValSize());
4150 InVals.push_back(DAG.getFrameIndex(FrameIndex, PtrVT));
4151 CCInfo.nextInRegsParam();
4153 unsigned FIOffset = VA.getLocMemOffset();
4154 int FI = MFI.CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
4157 // Create load nodes to retrieve arguments from the stack.
4158 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
4159 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
4160 MachinePointerInfo::getFixedStack(
4161 DAG.getMachineFunction(), FI)));
4163 lastInsIndex = index;
4169 if (isVarArg && MFI.hasVAStart())
4170 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
4171 CCInfo.getNextStackOffset(),
4172 TotalArgRegsSaveSize);
4174 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
4179 /// isFloatingPointZero - Return true if this is +0.0.
4180 static bool isFloatingPointZero(SDValue Op) {
4181 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
4182 return CFP->getValueAPF().isPosZero();
4183 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
4184 // Maybe this has already been legalized into the constant pool?
4185 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
4186 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
4187 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
4188 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
4189 return CFP->getValueAPF().isPosZero();
4191 } else if (Op->getOpcode() == ISD::BITCAST &&
4192 Op->getValueType(0) == MVT::f64) {
4193 // Handle (ISD::BITCAST (ARMISD::VMOVIMM (ISD::TargetConstant 0)) MVT::f64)
4194 // created by LowerConstantFP().
4195 SDValue BitcastOp = Op->getOperand(0);
4196 if (BitcastOp->getOpcode() == ARMISD::VMOVIMM &&
4197 isNullConstant(BitcastOp->getOperand(0)))
4203 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
4204 /// the given operands.
4205 SDValue ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
4206 SDValue &ARMcc, SelectionDAG &DAG,
4207 const SDLoc &dl) const {
4208 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
4209 unsigned C = RHSC->getZExtValue();
4210 if (!isLegalICmpImmediate((int32_t)C)) {
4211 // Constant does not fit, try adjusting it by one.
4216 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
4217 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
4218 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
4223 if (C != 0 && isLegalICmpImmediate(C-1)) {
4224 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
4225 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
4230 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
4231 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
4232 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
4237 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
4238 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
4239 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
4244 } else if ((ARM_AM::getShiftOpcForNode(LHS.getOpcode()) != ARM_AM::no_shift) &&
4245 (ARM_AM::getShiftOpcForNode(RHS.getOpcode()) == ARM_AM::no_shift)) {
4246 // In ARM and Thumb-2, the compare instructions can shift their second
4248 CC = ISD::getSetCCSwappedOperands(CC);
4249 std::swap(LHS, RHS);
4252 // Thumb1 has very limited immediate modes, so turning an "and" into a
4253 // shift can save multiple instructions.
4255 // If we have (x & C1), and C1 is an appropriate mask, we can transform it
4256 // into "((x << n) >> n)". But that isn't necessarily profitable on its
4257 // own. If it's the operand to an unsigned comparison with an immediate,
4258 // we can eliminate one of the shifts: we transform
4259 // "((x << n) >> n) == C2" to "(x << n) == (C2 << n)".
4261 // We avoid transforming cases which aren't profitable due to encoding
4264 // 1. C2 fits into the immediate field of a cmp, and the transformed version
4265 // would not; in that case, we're essentially trading one immediate load for
4267 // 2. C1 is 255 or 65535, so we can use uxtb or uxth.
4268 // 3. C2 is zero; we have other code for this special case.
4270 // FIXME: Figure out profitability for Thumb2; we usually can't save an
4271 // instruction, since the AND is always one instruction anyway, but we could
4272 // use narrow instructions in some cases.
4273 if (Subtarget->isThumb1Only() && LHS->getOpcode() == ISD::AND &&
4274 LHS->hasOneUse() && isa<ConstantSDNode>(LHS.getOperand(1)) &&
4275 LHS.getValueType() == MVT::i32 && isa<ConstantSDNode>(RHS) &&
4276 !isSignedIntSetCC(CC)) {
4277 unsigned Mask = cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue();
4278 auto *RHSC = cast<ConstantSDNode>(RHS.getNode());
4279 uint64_t RHSV = RHSC->getZExtValue();
4280 if (isMask_32(Mask) && (RHSV & ~Mask) == 0 && Mask != 255 && Mask != 65535) {
4281 unsigned ShiftBits = countLeadingZeros(Mask);
4282 if (RHSV && (RHSV > 255 || (RHSV << ShiftBits) <= 255)) {
4283 SDValue ShiftAmt = DAG.getConstant(ShiftBits, dl, MVT::i32);
4284 LHS = DAG.getNode(ISD::SHL, dl, MVT::i32, LHS.getOperand(0), ShiftAmt);
4285 RHS = DAG.getConstant(RHSV << ShiftBits, dl, MVT::i32);
4290 // The specific comparison "(x<<c) > 0x80000000U" can be optimized to a
4291 // single "lsls x, c+1". The shift sets the "C" and "Z" flags the same
4293 // FIXME: Add support for ARM/Thumb2; this would need isel patterns, and
4294 // some tweaks to the heuristics for the previous and->shift transform.
4295 // FIXME: Optimize cases where the LHS isn't a shift.
4296 if (Subtarget->isThumb1Only() && LHS->getOpcode() == ISD::SHL &&
4297 isa<ConstantSDNode>(RHS) &&
4298 cast<ConstantSDNode>(RHS)->getZExtValue() == 0x80000000U &&
4299 CC == ISD::SETUGT && isa<ConstantSDNode>(LHS.getOperand(1)) &&
4300 cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue() < 31) {
4302 cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue() + 1;
4303 SDValue Shift = DAG.getNode(ARMISD::LSLS, dl,
4304 DAG.getVTList(MVT::i32, MVT::i32),
4306 DAG.getConstant(ShiftAmt, dl, MVT::i32));
4307 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, ARM::CPSR,
4308 Shift.getValue(1), SDValue());
4309 ARMcc = DAG.getConstant(ARMCC::HI, dl, MVT::i32);
4310 return Chain.getValue(1);
4313 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
4315 // If the RHS is a constant zero then the V (overflow) flag will never be
4316 // set. This can allow us to simplify GE to PL or LT to MI, which can be
4317 // simpler for other passes (like the peephole optimiser) to deal with.
4318 if (isNullConstant(RHS)) {
4322 CondCode = ARMCC::PL;
4325 CondCode = ARMCC::MI;
4330 ARMISD::NodeType CompareType;
4333 CompareType = ARMISD::CMP;
4338 CompareType = ARMISD::CMPZ;
4341 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
4342 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
4345 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
4346 SDValue ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS,
4347 SelectionDAG &DAG, const SDLoc &dl) const {
4348 assert(Subtarget->hasFP64() || RHS.getValueType() != MVT::f64);
4350 if (!isFloatingPointZero(RHS))
4351 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
4353 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
4354 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
4357 /// duplicateCmp - Glue values can have only one use, so this function
4358 /// duplicates a comparison node.
4360 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
4361 unsigned Opc = Cmp.getOpcode();
4363 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
4364 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
4366 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
4367 Cmp = Cmp.getOperand(0);
4368 Opc = Cmp.getOpcode();
4369 if (Opc == ARMISD::CMPFP)
4370 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
4372 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
4373 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
4375 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
4378 // This function returns three things: the arithmetic computation itself
4379 // (Value), a comparison (OverflowCmp), and a condition code (ARMcc). The
4380 // comparison and the condition code define the case in which the arithmetic
4381 // computation *does not* overflow.
4382 std::pair<SDValue, SDValue>
4383 ARMTargetLowering::getARMXALUOOp(SDValue Op, SelectionDAG &DAG,
4384 SDValue &ARMcc) const {
4385 assert(Op.getValueType() == MVT::i32 && "Unsupported value type");
4387 SDValue Value, OverflowCmp;
4388 SDValue LHS = Op.getOperand(0);
4389 SDValue RHS = Op.getOperand(1);
4392 // FIXME: We are currently always generating CMPs because we don't support
4393 // generating CMN through the backend. This is not as good as the natural
4394 // CMP case because it causes a register dependency and cannot be folded
4397 switch (Op.getOpcode()) {
4399 llvm_unreachable("Unknown overflow instruction!");
4401 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
4402 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
4403 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
4406 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
4407 // We use ADDC here to correspond to its use in LowerUnsignedALUO.
4408 // We do not use it in the USUBO case as Value may not be used.
4409 Value = DAG.getNode(ARMISD::ADDC, dl,
4410 DAG.getVTList(Op.getValueType(), MVT::i32), LHS, RHS)
4412 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
4415 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
4416 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
4417 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
4420 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
4421 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
4422 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
4425 // We generate a UMUL_LOHI and then check if the high word is 0.
4426 ARMcc = DAG.getConstant(ARMCC::EQ, dl, MVT::i32);
4427 Value = DAG.getNode(ISD::UMUL_LOHI, dl,
4428 DAG.getVTList(Op.getValueType(), Op.getValueType()),
4430 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value.getValue(1),
4431 DAG.getConstant(0, dl, MVT::i32));
4432 Value = Value.getValue(0); // We only want the low 32 bits for the result.
4435 // We generate a SMUL_LOHI and then check if all the bits of the high word
4436 // are the same as the sign bit of the low word.
4437 ARMcc = DAG.getConstant(ARMCC::EQ, dl, MVT::i32);
4438 Value = DAG.getNode(ISD::SMUL_LOHI, dl,
4439 DAG.getVTList(Op.getValueType(), Op.getValueType()),
4441 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value.getValue(1),
4442 DAG.getNode(ISD::SRA, dl, Op.getValueType(),
4444 DAG.getConstant(31, dl, MVT::i32)));
4445 Value = Value.getValue(0); // We only want the low 32 bits for the result.
4449 return std::make_pair(Value, OverflowCmp);
4453 ARMTargetLowering::LowerSignedALUO(SDValue Op, SelectionDAG &DAG) const {
4454 // Let legalize expand this if it isn't a legal type yet.
4455 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
4458 SDValue Value, OverflowCmp;
4460 std::tie(Value, OverflowCmp) = getARMXALUOOp(Op, DAG, ARMcc);
4461 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4463 // We use 0 and 1 as false and true values.
4464 SDValue TVal = DAG.getConstant(1, dl, MVT::i32);
4465 SDValue FVal = DAG.getConstant(0, dl, MVT::i32);
4466 EVT VT = Op.getValueType();
4468 SDValue Overflow = DAG.getNode(ARMISD::CMOV, dl, VT, TVal, FVal,
4469 ARMcc, CCR, OverflowCmp);
4471 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
4472 return DAG.getNode(ISD::MERGE_VALUES, dl, VTs, Value, Overflow);
4475 static SDValue ConvertBooleanCarryToCarryFlag(SDValue BoolCarry,
4476 SelectionDAG &DAG) {
4477 SDLoc DL(BoolCarry);
4478 EVT CarryVT = BoolCarry.getValueType();
4480 // This converts the boolean value carry into the carry flag by doing
4481 // ARMISD::SUBC Carry, 1
4482 SDValue Carry = DAG.getNode(ARMISD::SUBC, DL,
4483 DAG.getVTList(CarryVT, MVT::i32),
4484 BoolCarry, DAG.getConstant(1, DL, CarryVT));
4485 return Carry.getValue(1);
4488 static SDValue ConvertCarryFlagToBooleanCarry(SDValue Flags, EVT VT,
4489 SelectionDAG &DAG) {
4492 // Now convert the carry flag into a boolean carry. We do this
4493 // using ARMISD:ADDE 0, 0, Carry
4494 return DAG.getNode(ARMISD::ADDE, DL, DAG.getVTList(VT, MVT::i32),
4495 DAG.getConstant(0, DL, MVT::i32),
4496 DAG.getConstant(0, DL, MVT::i32), Flags);
4499 SDValue ARMTargetLowering::LowerUnsignedALUO(SDValue Op,
4500 SelectionDAG &DAG) const {
4501 // Let legalize expand this if it isn't a legal type yet.
4502 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
4505 SDValue LHS = Op.getOperand(0);
4506 SDValue RHS = Op.getOperand(1);
4509 EVT VT = Op.getValueType();
4510 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
4513 switch (Op.getOpcode()) {
4515 llvm_unreachable("Unknown overflow instruction!");
4517 Value = DAG.getNode(ARMISD::ADDC, dl, VTs, LHS, RHS);
4518 // Convert the carry flag into a boolean value.
4519 Overflow = ConvertCarryFlagToBooleanCarry(Value.getValue(1), VT, DAG);
4522 Value = DAG.getNode(ARMISD::SUBC, dl, VTs, LHS, RHS);
4523 // Convert the carry flag into a boolean value.
4524 Overflow = ConvertCarryFlagToBooleanCarry(Value.getValue(1), VT, DAG);
4525 // ARMISD::SUBC returns 0 when we have to borrow, so make it an overflow
4526 // value. So compute 1 - C.
4527 Overflow = DAG.getNode(ISD::SUB, dl, MVT::i32,
4528 DAG.getConstant(1, dl, MVT::i32), Overflow);
4533 return DAG.getNode(ISD::MERGE_VALUES, dl, VTs, Value, Overflow);
4536 static SDValue LowerSADDSUBSAT(SDValue Op, SelectionDAG &DAG,
4537 const ARMSubtarget *Subtarget) {
4538 EVT VT = Op.getValueType();
4539 if (!Subtarget->hasDSP())
4545 bool IsAdd = Op->getOpcode() == ISD::SADDSAT;
4546 switch (VT.getSimpleVT().SimpleTy) {
4550 NewOpcode = IsAdd ? ARMISD::QADD8b : ARMISD::QSUB8b;
4553 NewOpcode = IsAdd ? ARMISD::QADD16b : ARMISD::QSUB16b;
4559 DAG.getNode(NewOpcode, dl, MVT::i32,
4560 DAG.getSExtOrTrunc(Op->getOperand(0), dl, MVT::i32),
4561 DAG.getSExtOrTrunc(Op->getOperand(1), dl, MVT::i32));
4562 return DAG.getNode(ISD::TRUNCATE, dl, VT, Add);
4565 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
4566 SDValue Cond = Op.getOperand(0);
4567 SDValue SelectTrue = Op.getOperand(1);
4568 SDValue SelectFalse = Op.getOperand(2);
4570 unsigned Opc = Cond.getOpcode();
4572 if (Cond.getResNo() == 1 &&
4573 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
4574 Opc == ISD::USUBO)) {
4575 if (!DAG.getTargetLoweringInfo().isTypeLegal(Cond->getValueType(0)))
4578 SDValue Value, OverflowCmp;
4580 std::tie(Value, OverflowCmp) = getARMXALUOOp(Cond, DAG, ARMcc);
4581 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4582 EVT VT = Op.getValueType();
4584 return getCMOV(dl, VT, SelectTrue, SelectFalse, ARMcc, CCR,
4590 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
4591 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
4593 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
4594 const ConstantSDNode *CMOVTrue =
4595 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
4596 const ConstantSDNode *CMOVFalse =
4597 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
4599 if (CMOVTrue && CMOVFalse) {
4600 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
4601 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
4605 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
4607 False = SelectFalse;
4608 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
4613 if (True.getNode() && False.getNode()) {
4614 EVT VT = Op.getValueType();
4615 SDValue ARMcc = Cond.getOperand(2);
4616 SDValue CCR = Cond.getOperand(3);
4617 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
4618 assert(True.getValueType() == VT);
4619 return getCMOV(dl, VT, True, False, ARMcc, CCR, Cmp, DAG);
4624 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
4625 // undefined bits before doing a full-word comparison with zero.
4626 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
4627 DAG.getConstant(1, dl, Cond.getValueType()));
4629 return DAG.getSelectCC(dl, Cond,
4630 DAG.getConstant(0, dl, Cond.getValueType()),
4631 SelectTrue, SelectFalse, ISD::SETNE);
4634 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
4635 bool &swpCmpOps, bool &swpVselOps) {
4636 // Start by selecting the GE condition code for opcodes that return true for
4638 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
4639 CC == ISD::SETULE || CC == ISD::SETGE || CC == ISD::SETLE)
4640 CondCode = ARMCC::GE;
4642 // and GT for opcodes that return false for 'equality'.
4643 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
4644 CC == ISD::SETULT || CC == ISD::SETGT || CC == ISD::SETLT)
4645 CondCode = ARMCC::GT;
4647 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
4648 // to swap the compare operands.
4649 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
4650 CC == ISD::SETULT || CC == ISD::SETLE || CC == ISD::SETLT)
4653 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
4654 // If we have an unordered opcode, we need to swap the operands to the VSEL
4655 // instruction (effectively negating the condition).
4657 // This also has the effect of swapping which one of 'less' or 'greater'
4658 // returns true, so we also swap the compare operands. It also switches
4659 // whether we return true for 'equality', so we compensate by picking the
4660 // opposite condition code to our original choice.
4661 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
4662 CC == ISD::SETUGT) {
4663 swpCmpOps = !swpCmpOps;
4664 swpVselOps = !swpVselOps;
4665 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
4668 // 'ordered' is 'anything but unordered', so use the VS condition code and
4669 // swap the VSEL operands.
4670 if (CC == ISD::SETO) {
4671 CondCode = ARMCC::VS;
4675 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
4676 // code and swap the VSEL operands. Also do this if we don't care about the
4678 if (CC == ISD::SETUNE || CC == ISD::SETNE) {
4679 CondCode = ARMCC::EQ;
4684 SDValue ARMTargetLowering::getCMOV(const SDLoc &dl, EVT VT, SDValue FalseVal,
4685 SDValue TrueVal, SDValue ARMcc, SDValue CCR,
4686 SDValue Cmp, SelectionDAG &DAG) const {
4687 if (!Subtarget->hasFP64() && VT == MVT::f64) {
4688 FalseVal = DAG.getNode(ARMISD::VMOVRRD, dl,
4689 DAG.getVTList(MVT::i32, MVT::i32), FalseVal);
4690 TrueVal = DAG.getNode(ARMISD::VMOVRRD, dl,
4691 DAG.getVTList(MVT::i32, MVT::i32), TrueVal);
4693 SDValue TrueLow = TrueVal.getValue(0);
4694 SDValue TrueHigh = TrueVal.getValue(1);
4695 SDValue FalseLow = FalseVal.getValue(0);
4696 SDValue FalseHigh = FalseVal.getValue(1);
4698 SDValue Low = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseLow, TrueLow,
4700 SDValue High = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseHigh, TrueHigh,
4701 ARMcc, CCR, duplicateCmp(Cmp, DAG));
4703 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Low, High);
4705 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
4710 static bool isGTorGE(ISD::CondCode CC) {
4711 return CC == ISD::SETGT || CC == ISD::SETGE;
4714 static bool isLTorLE(ISD::CondCode CC) {
4715 return CC == ISD::SETLT || CC == ISD::SETLE;
4718 // See if a conditional (LHS CC RHS ? TrueVal : FalseVal) is lower-saturating.
4719 // All of these conditions (and their <= and >= counterparts) will do:
4724 static bool isLowerSaturate(const SDValue LHS, const SDValue RHS,
4725 const SDValue TrueVal, const SDValue FalseVal,
4726 const ISD::CondCode CC, const SDValue K) {
4727 return (isGTorGE(CC) &&
4728 ((K == LHS && K == TrueVal) || (K == RHS && K == FalseVal))) ||
4730 ((K == RHS && K == TrueVal) || (K == LHS && K == FalseVal)));
4733 // Similar to isLowerSaturate(), but checks for upper-saturating conditions.
4734 static bool isUpperSaturate(const SDValue LHS, const SDValue RHS,
4735 const SDValue TrueVal, const SDValue FalseVal,
4736 const ISD::CondCode CC, const SDValue K) {
4737 return (isGTorGE(CC) &&
4738 ((K == RHS && K == TrueVal) || (K == LHS && K == FalseVal))) ||
4740 ((K == LHS && K == TrueVal) || (K == RHS && K == FalseVal)));
4743 // Check if two chained conditionals could be converted into SSAT or USAT.
4745 // SSAT can replace a set of two conditional selectors that bound a number to an
4746 // interval of type [k, ~k] when k + 1 is a power of 2. Here are some examples:
4748 // x < -k ? -k : (x > k ? k : x)
4749 // x < -k ? -k : (x < k ? x : k)
4750 // x > -k ? (x > k ? k : x) : -k
4751 // x < k ? (x < -k ? -k : x) : k
4754 // USAT works similarily to SSAT but bounds on the interval [0, k] where k + 1 is
4757 // It returns true if the conversion can be done, false otherwise.
4758 // Additionally, the variable is returned in parameter V, the constant in K and
4759 // usat is set to true if the conditional represents an unsigned saturation
4760 static bool isSaturatingConditional(const SDValue &Op, SDValue &V,
4761 uint64_t &K, bool &usat) {
4762 SDValue LHS1 = Op.getOperand(0);
4763 SDValue RHS1 = Op.getOperand(1);
4764 SDValue TrueVal1 = Op.getOperand(2);
4765 SDValue FalseVal1 = Op.getOperand(3);
4766 ISD::CondCode CC1 = cast<CondCodeSDNode>(Op.getOperand(4))->get();
4768 const SDValue Op2 = isa<ConstantSDNode>(TrueVal1) ? FalseVal1 : TrueVal1;
4769 if (Op2.getOpcode() != ISD::SELECT_CC)
4772 SDValue LHS2 = Op2.getOperand(0);
4773 SDValue RHS2 = Op2.getOperand(1);
4774 SDValue TrueVal2 = Op2.getOperand(2);
4775 SDValue FalseVal2 = Op2.getOperand(3);
4776 ISD::CondCode CC2 = cast<CondCodeSDNode>(Op2.getOperand(4))->get();
4778 // Find out which are the constants and which are the variables
4779 // in each conditional
4780 SDValue *K1 = isa<ConstantSDNode>(LHS1) ? &LHS1 : isa<ConstantSDNode>(RHS1)
4783 SDValue *K2 = isa<ConstantSDNode>(LHS2) ? &LHS2 : isa<ConstantSDNode>(RHS2)
4786 SDValue K2Tmp = isa<ConstantSDNode>(TrueVal2) ? TrueVal2 : FalseVal2;
4787 SDValue V1Tmp = (K1 && *K1 == LHS1) ? RHS1 : LHS1;
4788 SDValue V2Tmp = (K2 && *K2 == LHS2) ? RHS2 : LHS2;
4789 SDValue V2 = (K2Tmp == TrueVal2) ? FalseVal2 : TrueVal2;
4791 // We must detect cases where the original operations worked with 16- or
4792 // 8-bit values. In such case, V2Tmp != V2 because the comparison operations
4793 // must work with sign-extended values but the select operations return
4794 // the original non-extended value.
4795 SDValue V2TmpReg = V2Tmp;
4796 if (V2Tmp->getOpcode() == ISD::SIGN_EXTEND_INREG)
4797 V2TmpReg = V2Tmp->getOperand(0);
4799 // Check that the registers and the constants have the correct values
4800 // in both conditionals
4801 if (!K1 || !K2 || *K1 == Op2 || *K2 != K2Tmp || V1Tmp != V2Tmp ||
4805 // Figure out which conditional is saturating the lower/upper bound.
4806 const SDValue *LowerCheckOp =
4807 isLowerSaturate(LHS1, RHS1, TrueVal1, FalseVal1, CC1, *K1)
4809 : isLowerSaturate(LHS2, RHS2, TrueVal2, FalseVal2, CC2, *K2)
4812 const SDValue *UpperCheckOp =
4813 isUpperSaturate(LHS1, RHS1, TrueVal1, FalseVal1, CC1, *K1)
4815 : isUpperSaturate(LHS2, RHS2, TrueVal2, FalseVal2, CC2, *K2)
4819 if (!UpperCheckOp || !LowerCheckOp || LowerCheckOp == UpperCheckOp)
4822 // Check that the constant in the lower-bound check is
4823 // the opposite of the constant in the upper-bound check
4824 // in 1's complement.
4825 int64_t Val1 = cast<ConstantSDNode>(*K1)->getSExtValue();
4826 int64_t Val2 = cast<ConstantSDNode>(*K2)->getSExtValue();
4827 int64_t PosVal = std::max(Val1, Val2);
4828 int64_t NegVal = std::min(Val1, Val2);
4830 if (((Val1 > Val2 && UpperCheckOp == &Op) ||
4831 (Val1 < Val2 && UpperCheckOp == &Op2)) &&
4832 isPowerOf2_64(PosVal + 1)) {
4834 // Handle the difference between USAT (unsigned) and SSAT (signed) saturation
4837 else if (NegVal == 0)
4843 K = (uint64_t)PosVal; // At this point, PosVal is guaranteed to be positive
4851 // Check if a condition of the type x < k ? k : x can be converted into a
4852 // bit operation instead of conditional moves.
4853 // Currently this is allowed given:
4854 // - The conditions and values match up
4855 // - k is 0 or -1 (all ones)
4856 // This function will not check the last condition, thats up to the caller
4857 // It returns true if the transformation can be made, and in such case
4858 // returns x in V, and k in SatK.
4859 static bool isLowerSaturatingConditional(const SDValue &Op, SDValue &V,
4862 SDValue LHS = Op.getOperand(0);
4863 SDValue RHS = Op.getOperand(1);
4864 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
4865 SDValue TrueVal = Op.getOperand(2);
4866 SDValue FalseVal = Op.getOperand(3);
4868 SDValue *K = isa<ConstantSDNode>(LHS) ? &LHS : isa<ConstantSDNode>(RHS)
4872 // No constant operation in comparison, early out
4876 SDValue KTmp = isa<ConstantSDNode>(TrueVal) ? TrueVal : FalseVal;
4877 V = (KTmp == TrueVal) ? FalseVal : TrueVal;
4878 SDValue VTmp = (K && *K == LHS) ? RHS : LHS;
4880 // If the constant on left and right side, or variable on left and right,
4881 // does not match, early out
4882 if (*K != KTmp || V != VTmp)
4885 if (isLowerSaturate(LHS, RHS, TrueVal, FalseVal, CC, *K)) {
4893 bool ARMTargetLowering::isUnsupportedFloatingType(EVT VT) const {
4895 return !Subtarget->hasVFP2Base();
4897 return !Subtarget->hasFP64();
4899 return !Subtarget->hasFullFP16();
4903 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
4904 EVT VT = Op.getValueType();
4907 // Try to convert two saturating conditional selects into a single SSAT
4909 uint64_t SatConstant;
4911 if (((!Subtarget->isThumb() && Subtarget->hasV6Ops()) || Subtarget->isThumb2()) &&
4912 isSaturatingConditional(Op, SatValue, SatConstant, SatUSat)) {
4914 return DAG.getNode(ARMISD::USAT, dl, VT, SatValue,
4915 DAG.getConstant(countTrailingOnes(SatConstant), dl, VT));
4917 return DAG.getNode(ARMISD::SSAT, dl, VT, SatValue,
4918 DAG.getConstant(countTrailingOnes(SatConstant), dl, VT));
4921 // Try to convert expressions of the form x < k ? k : x (and similar forms)
4922 // into more efficient bit operations, which is possible when k is 0 or -1
4923 // On ARM and Thumb-2 which have flexible operand 2 this will result in
4924 // single instructions. On Thumb the shift and the bit operation will be two
4926 // Only allow this transformation on full-width (32-bit) operations
4927 SDValue LowerSatConstant;
4928 if (VT == MVT::i32 &&
4929 isLowerSaturatingConditional(Op, SatValue, LowerSatConstant)) {
4930 SDValue ShiftV = DAG.getNode(ISD::SRA, dl, VT, SatValue,
4931 DAG.getConstant(31, dl, VT));
4932 if (isNullConstant(LowerSatConstant)) {
4933 SDValue NotShiftV = DAG.getNode(ISD::XOR, dl, VT, ShiftV,
4934 DAG.getAllOnesConstant(dl, VT));
4935 return DAG.getNode(ISD::AND, dl, VT, SatValue, NotShiftV);
4936 } else if (isAllOnesConstant(LowerSatConstant))
4937 return DAG.getNode(ISD::OR, dl, VT, SatValue, ShiftV);
4940 SDValue LHS = Op.getOperand(0);
4941 SDValue RHS = Op.getOperand(1);
4942 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
4943 SDValue TrueVal = Op.getOperand(2);
4944 SDValue FalseVal = Op.getOperand(3);
4945 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FalseVal);
4946 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TrueVal);
4948 if (Subtarget->hasV8_1MMainlineOps() && CFVal && CTVal &&
4949 LHS.getValueType() == MVT::i32 && RHS.getValueType() == MVT::i32) {
4950 unsigned TVal = CTVal->getZExtValue();
4951 unsigned FVal = CFVal->getZExtValue();
4952 unsigned Opcode = 0;
4954 if (TVal == ~FVal) {
4955 Opcode = ARMISD::CSINV;
4956 } else if (TVal == ~FVal + 1) {
4957 Opcode = ARMISD::CSNEG;
4958 } else if (TVal + 1 == FVal) {
4959 Opcode = ARMISD::CSINC;
4960 } else if (TVal == FVal + 1) {
4961 Opcode = ARMISD::CSINC;
4962 std::swap(TrueVal, FalseVal);
4963 std::swap(TVal, FVal);
4964 CC = ISD::getSetCCInverse(CC, LHS.getValueType());
4968 // If one of the constants is cheaper than another, materialise the
4969 // cheaper one and let the csel generate the other.
4970 if (Opcode != ARMISD::CSINC &&
4971 HasLowerConstantMaterializationCost(FVal, TVal, Subtarget)) {
4972 std::swap(TrueVal, FalseVal);
4973 std::swap(TVal, FVal);
4974 CC = ISD::getSetCCInverse(CC, LHS.getValueType());
4977 // Attempt to use ZR checking TVal is 0, possibly inverting the condition
4978 // to get there. CSINC not is invertable like the other two (~(~a) == a,
4979 // -(-a) == a, but (a+1)+1 != a).
4980 if (FVal == 0 && Opcode != ARMISD::CSINC) {
4981 std::swap(TrueVal, FalseVal);
4982 std::swap(TVal, FVal);
4983 CC = ISD::getSetCCInverse(CC, LHS.getValueType());
4986 TrueVal = DAG.getRegister(ARM::ZR, MVT::i32);
4988 // Drops F's value because we can get it by inverting/negating TVal.
4992 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
4993 EVT VT = TrueVal.getValueType();
4994 return DAG.getNode(Opcode, dl, VT, TrueVal, FalseVal, ARMcc, Cmp);
4998 if (isUnsupportedFloatingType(LHS.getValueType())) {
4999 DAG.getTargetLoweringInfo().softenSetCCOperands(
5000 DAG, LHS.getValueType(), LHS, RHS, CC, dl, LHS, RHS);
5002 // If softenSetCCOperands only returned one value, we should compare it to
5004 if (!RHS.getNode()) {
5005 RHS = DAG.getConstant(0, dl, LHS.getValueType());
5010 if (LHS.getValueType() == MVT::i32) {
5011 // Try to generate VSEL on ARMv8.
5012 // The VSEL instruction can't use all the usual ARM condition
5013 // codes: it only has two bits to select the condition code, so it's
5014 // constrained to use only GE, GT, VS and EQ.
5016 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
5017 // swap the operands of the previous compare instruction (effectively
5018 // inverting the compare condition, swapping 'less' and 'greater') and
5019 // sometimes need to swap the operands to the VSEL (which inverts the
5020 // condition in the sense of firing whenever the previous condition didn't)
5021 if (Subtarget->hasFPARMv8Base() && (TrueVal.getValueType() == MVT::f16 ||
5022 TrueVal.getValueType() == MVT::f32 ||
5023 TrueVal.getValueType() == MVT::f64)) {
5024 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
5025 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
5026 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
5027 CC = ISD::getSetCCInverse(CC, LHS.getValueType());
5028 std::swap(TrueVal, FalseVal);
5033 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
5034 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
5035 // Choose GE over PL, which vsel does now support
5036 if (cast<ConstantSDNode>(ARMcc)->getZExtValue() == ARMCC::PL)
5037 ARMcc = DAG.getConstant(ARMCC::GE, dl, MVT::i32);
5038 return getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
5041 ARMCC::CondCodes CondCode, CondCode2;
5042 FPCCToARMCC(CC, CondCode, CondCode2);
5044 // Normalize the fp compare. If RHS is zero we prefer to keep it there so we
5045 // match CMPFPw0 instead of CMPFP, though we don't do this for f16 because we
5046 // must use VSEL (limited condition codes), due to not having conditional f16
5048 if (Subtarget->hasFPARMv8Base() &&
5049 !(isFloatingPointZero(RHS) && TrueVal.getValueType() != MVT::f16) &&
5050 (TrueVal.getValueType() == MVT::f16 ||
5051 TrueVal.getValueType() == MVT::f32 ||
5052 TrueVal.getValueType() == MVT::f64)) {
5053 bool swpCmpOps = false;
5054 bool swpVselOps = false;
5055 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
5057 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
5058 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
5060 std::swap(LHS, RHS);
5062 std::swap(TrueVal, FalseVal);
5066 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
5067 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
5068 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
5069 SDValue Result = getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
5070 if (CondCode2 != ARMCC::AL) {
5071 SDValue ARMcc2 = DAG.getConstant(CondCode2, dl, MVT::i32);
5072 // FIXME: Needs another CMP because flag can have but one use.
5073 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
5074 Result = getCMOV(dl, VT, Result, TrueVal, ARMcc2, CCR, Cmp2, DAG);
5079 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
5080 /// to morph to an integer compare sequence.
5081 static bool canChangeToInt(SDValue Op, bool &SeenZero,
5082 const ARMSubtarget *Subtarget) {
5083 SDNode *N = Op.getNode();
5084 if (!N->hasOneUse())
5085 // Otherwise it requires moving the value from fp to integer registers.
5087 if (!N->getNumValues())
5089 EVT VT = Op.getValueType();
5090 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
5091 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
5092 // vmrs are very slow, e.g. cortex-a8.
5095 if (isFloatingPointZero(Op)) {
5099 return ISD::isNormalLoad(N);
5102 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
5103 if (isFloatingPointZero(Op))
5104 return DAG.getConstant(0, SDLoc(Op), MVT::i32);
5106 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
5107 return DAG.getLoad(MVT::i32, SDLoc(Op), Ld->getChain(), Ld->getBasePtr(),
5108 Ld->getPointerInfo(), Ld->getAlignment(),
5109 Ld->getMemOperand()->getFlags());
5111 llvm_unreachable("Unknown VFP cmp argument!");
5114 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
5115 SDValue &RetVal1, SDValue &RetVal2) {
5118 if (isFloatingPointZero(Op)) {
5119 RetVal1 = DAG.getConstant(0, dl, MVT::i32);
5120 RetVal2 = DAG.getConstant(0, dl, MVT::i32);
5124 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
5125 SDValue Ptr = Ld->getBasePtr();
5127 DAG.getLoad(MVT::i32, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
5128 Ld->getAlignment(), Ld->getMemOperand()->getFlags());
5130 EVT PtrType = Ptr.getValueType();
5131 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
5132 SDValue NewPtr = DAG.getNode(ISD::ADD, dl,
5133 PtrType, Ptr, DAG.getConstant(4, dl, PtrType));
5134 RetVal2 = DAG.getLoad(MVT::i32, dl, Ld->getChain(), NewPtr,
5135 Ld->getPointerInfo().getWithOffset(4), NewAlign,
5136 Ld->getMemOperand()->getFlags());
5140 llvm_unreachable("Unknown VFP cmp argument!");
5143 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
5144 /// f32 and even f64 comparisons to integer ones.
5146 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
5147 SDValue Chain = Op.getOperand(0);
5148 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
5149 SDValue LHS = Op.getOperand(2);
5150 SDValue RHS = Op.getOperand(3);
5151 SDValue Dest = Op.getOperand(4);
5154 bool LHSSeenZero = false;
5155 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
5156 bool RHSSeenZero = false;
5157 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
5158 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
5159 // If unsafe fp math optimization is enabled and there are no other uses of
5160 // the CMP operands, and the condition code is EQ or NE, we can optimize it
5161 // to an integer comparison.
5162 if (CC == ISD::SETOEQ)
5164 else if (CC == ISD::SETUNE)
5167 SDValue Mask = DAG.getConstant(0x7fffffff, dl, MVT::i32);
5169 if (LHS.getValueType() == MVT::f32) {
5170 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
5171 bitcastf32Toi32(LHS, DAG), Mask);
5172 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
5173 bitcastf32Toi32(RHS, DAG), Mask);
5174 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
5175 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
5176 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
5177 Chain, Dest, ARMcc, CCR, Cmp);
5182 expandf64Toi32(LHS, DAG, LHS1, LHS2);
5183 expandf64Toi32(RHS, DAG, RHS1, RHS2);
5184 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
5185 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
5186 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
5187 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
5188 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
5189 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
5190 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops);
5196 SDValue ARMTargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
5197 SDValue Chain = Op.getOperand(0);
5198 SDValue Cond = Op.getOperand(1);
5199 SDValue Dest = Op.getOperand(2);
5202 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
5204 unsigned Opc = Cond.getOpcode();
5205 bool OptimizeMul = (Opc == ISD::SMULO || Opc == ISD::UMULO) &&
5206 !Subtarget->isThumb1Only();
5207 if (Cond.getResNo() == 1 &&
5208 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
5209 Opc == ISD::USUBO || OptimizeMul)) {
5210 // Only lower legal XALUO ops.
5211 if (!DAG.getTargetLoweringInfo().isTypeLegal(Cond->getValueType(0)))
5214 // The actual operation with overflow check.
5215 SDValue Value, OverflowCmp;
5217 std::tie(Value, OverflowCmp) = getARMXALUOOp(Cond, DAG, ARMcc);
5219 // Reverse the condition code.
5220 ARMCC::CondCodes CondCode =
5221 (ARMCC::CondCodes)cast<const ConstantSDNode>(ARMcc)->getZExtValue();
5222 CondCode = ARMCC::getOppositeCondition(CondCode);
5223 ARMcc = DAG.getConstant(CondCode, SDLoc(ARMcc), MVT::i32);
5224 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
5226 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other, Chain, Dest, ARMcc, CCR,
5233 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
5234 SDValue Chain = Op.getOperand(0);
5235 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
5236 SDValue LHS = Op.getOperand(2);
5237 SDValue RHS = Op.getOperand(3);
5238 SDValue Dest = Op.getOperand(4);
5241 if (isUnsupportedFloatingType(LHS.getValueType())) {
5242 DAG.getTargetLoweringInfo().softenSetCCOperands(
5243 DAG, LHS.getValueType(), LHS, RHS, CC, dl, LHS, RHS);
5245 // If softenSetCCOperands only returned one value, we should compare it to
5247 if (!RHS.getNode()) {
5248 RHS = DAG.getConstant(0, dl, LHS.getValueType());
5253 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
5255 unsigned Opc = LHS.getOpcode();
5256 bool OptimizeMul = (Opc == ISD::SMULO || Opc == ISD::UMULO) &&
5257 !Subtarget->isThumb1Only();
5258 if (LHS.getResNo() == 1 && (isOneConstant(RHS) || isNullConstant(RHS)) &&
5259 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
5260 Opc == ISD::USUBO || OptimizeMul) &&
5261 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5262 // Only lower legal XALUO ops.
5263 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
5266 // The actual operation with overflow check.
5267 SDValue Value, OverflowCmp;
5269 std::tie(Value, OverflowCmp) = getARMXALUOOp(LHS.getValue(0), DAG, ARMcc);
5271 if ((CC == ISD::SETNE) != isOneConstant(RHS)) {
5272 // Reverse the condition code.
5273 ARMCC::CondCodes CondCode =
5274 (ARMCC::CondCodes)cast<const ConstantSDNode>(ARMcc)->getZExtValue();
5275 CondCode = ARMCC::getOppositeCondition(CondCode);
5276 ARMcc = DAG.getConstant(CondCode, SDLoc(ARMcc), MVT::i32);
5278 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
5280 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other, Chain, Dest, ARMcc, CCR,
5284 if (LHS.getValueType() == MVT::i32) {
5286 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
5287 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
5288 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
5289 Chain, Dest, ARMcc, CCR, Cmp);
5292 if (getTargetMachine().Options.UnsafeFPMath &&
5293 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
5294 CC == ISD::SETNE || CC == ISD::SETUNE)) {
5295 if (SDValue Result = OptimizeVFPBrcond(Op, DAG))
5299 ARMCC::CondCodes CondCode, CondCode2;
5300 FPCCToARMCC(CC, CondCode, CondCode2);
5302 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
5303 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
5304 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
5305 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
5306 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
5307 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
5308 if (CondCode2 != ARMCC::AL) {
5309 ARMcc = DAG.getConstant(CondCode2, dl, MVT::i32);
5310 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
5311 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
5316 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
5317 SDValue Chain = Op.getOperand(0);
5318 SDValue Table = Op.getOperand(1);
5319 SDValue Index = Op.getOperand(2);
5322 EVT PTy = getPointerTy(DAG.getDataLayout());
5323 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
5324 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
5325 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI);
5326 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, dl, PTy));
5327 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Table, Index);
5328 if (Subtarget->isThumb2() || (Subtarget->hasV8MBaselineOps() && Subtarget->isThumb())) {
5329 // Thumb2 and ARMv8-M use a two-level jump. That is, it jumps into the jump table
5330 // which does another jump to the destination. This also makes it easier
5331 // to translate it to TBB / TBH later (Thumb2 only).
5332 // FIXME: This might not work if the function is extremely large.
5333 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
5334 Addr, Op.getOperand(2), JTI);
5336 if (isPositionIndependent() || Subtarget->isROPI()) {
5338 DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
5339 MachinePointerInfo::getJumpTable(DAG.getMachineFunction()));
5340 Chain = Addr.getValue(1);
5341 Addr = DAG.getNode(ISD::ADD, dl, PTy, Table, Addr);
5342 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
5345 DAG.getLoad(PTy, dl, Chain, Addr,
5346 MachinePointerInfo::getJumpTable(DAG.getMachineFunction()));
5347 Chain = Addr.getValue(1);
5348 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
5352 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
5353 EVT VT = Op.getValueType();
5356 if (Op.getValueType().getVectorElementType() == MVT::i32) {
5357 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
5359 return DAG.UnrollVectorOp(Op.getNode());
5362 const bool HasFullFP16 =
5363 static_cast<const ARMSubtarget&>(DAG.getSubtarget()).hasFullFP16();
5366 const EVT OpTy = Op.getOperand(0).getValueType();
5367 if (OpTy == MVT::v4f32)
5369 else if (OpTy == MVT::v4f16 && HasFullFP16)
5371 else if (OpTy == MVT::v8f16 && HasFullFP16)
5374 llvm_unreachable("Invalid type for custom lowering!");
5376 if (VT != MVT::v4i16 && VT != MVT::v8i16)
5377 return DAG.UnrollVectorOp(Op.getNode());
5379 Op = DAG.getNode(Op.getOpcode(), dl, NewTy, Op.getOperand(0));
5380 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
5383 SDValue ARMTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const {
5384 EVT VT = Op.getValueType();
5386 return LowerVectorFP_TO_INT(Op, DAG);
5388 bool IsStrict = Op->isStrictFPOpcode();
5389 SDValue SrcVal = Op.getOperand(IsStrict ? 1 : 0);
5391 if (isUnsupportedFloatingType(SrcVal.getValueType())) {
5393 if (Op.getOpcode() == ISD::FP_TO_SINT ||
5394 Op.getOpcode() == ISD::STRICT_FP_TO_SINT)
5395 LC = RTLIB::getFPTOSINT(SrcVal.getValueType(),
5398 LC = RTLIB::getFPTOUINT(SrcVal.getValueType(),
5401 MakeLibCallOptions CallOptions;
5402 SDValue Chain = IsStrict ? Op.getOperand(0) : SDValue();
5404 std::tie(Result, Chain) = makeLibCall(DAG, LC, Op.getValueType(), SrcVal,
5405 CallOptions, Loc, Chain);
5406 return IsStrict ? DAG.getMergeValues({Result, Chain}, Loc) : Result;
5409 // FIXME: Remove this when we have strict fp instruction selection patterns
5411 DAG.mutateStrictFPToFP(Op.getNode());
5417 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5418 EVT VT = Op.getValueType();
5421 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
5422 if (VT.getVectorElementType() == MVT::f32)
5424 return DAG.UnrollVectorOp(Op.getNode());
5427 assert((Op.getOperand(0).getValueType() == MVT::v4i16 ||
5428 Op.getOperand(0).getValueType() == MVT::v8i16) &&
5429 "Invalid type for custom lowering!");
5431 const bool HasFullFP16 =
5432 static_cast<const ARMSubtarget&>(DAG.getSubtarget()).hasFullFP16();
5435 if (VT == MVT::v4f32)
5436 DestVecType = MVT::v4i32;
5437 else if (VT == MVT::v4f16 && HasFullFP16)
5438 DestVecType = MVT::v4i16;
5439 else if (VT == MVT::v8f16 && HasFullFP16)
5440 DestVecType = MVT::v8i16;
5442 return DAG.UnrollVectorOp(Op.getNode());
5446 switch (Op.getOpcode()) {
5447 default: llvm_unreachable("Invalid opcode!");
5448 case ISD::SINT_TO_FP:
5449 CastOpc = ISD::SIGN_EXTEND;
5450 Opc = ISD::SINT_TO_FP;
5452 case ISD::UINT_TO_FP:
5453 CastOpc = ISD::ZERO_EXTEND;
5454 Opc = ISD::UINT_TO_FP;
5458 Op = DAG.getNode(CastOpc, dl, DestVecType, Op.getOperand(0));
5459 return DAG.getNode(Opc, dl, VT, Op);
5462 SDValue ARMTargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const {
5463 EVT VT = Op.getValueType();
5465 return LowerVectorINT_TO_FP(Op, DAG);
5466 if (isUnsupportedFloatingType(VT)) {
5468 if (Op.getOpcode() == ISD::SINT_TO_FP)
5469 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(),
5472 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(),
5474 MakeLibCallOptions CallOptions;
5475 return makeLibCall(DAG, LC, Op.getValueType(), Op.getOperand(0),
5476 CallOptions, SDLoc(Op)).first;
5482 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
5483 // Implement fcopysign with a fabs and a conditional fneg.
5484 SDValue Tmp0 = Op.getOperand(0);
5485 SDValue Tmp1 = Op.getOperand(1);
5487 EVT VT = Op.getValueType();
5488 EVT SrcVT = Tmp1.getValueType();
5489 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
5490 Tmp0.getOpcode() == ARMISD::VMOVDRR;
5491 bool UseNEON = !InGPR && Subtarget->hasNEON();
5494 // Use VBSL to copy the sign bit.
5495 unsigned EncodedVal = ARM_AM::createVMOVModImm(0x6, 0x80);
5496 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
5497 DAG.getTargetConstant(EncodedVal, dl, MVT::i32));
5498 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
5500 Mask = DAG.getNode(ARMISD::VSHLIMM, dl, OpVT,
5501 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
5502 DAG.getConstant(32, dl, MVT::i32));
5503 else /*if (VT == MVT::f32)*/
5504 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
5505 if (SrcVT == MVT::f32) {
5506 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
5508 Tmp1 = DAG.getNode(ARMISD::VSHLIMM, dl, OpVT,
5509 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
5510 DAG.getConstant(32, dl, MVT::i32));
5511 } else if (VT == MVT::f32)
5512 Tmp1 = DAG.getNode(ARMISD::VSHRuIMM, dl, MVT::v1i64,
5513 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
5514 DAG.getConstant(32, dl, MVT::i32));
5515 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
5516 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
5518 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createVMOVModImm(0xe, 0xff),
5520 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
5521 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
5522 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
5524 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
5525 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
5526 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
5527 if (VT == MVT::f32) {
5528 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
5529 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
5530 DAG.getConstant(0, dl, MVT::i32));
5532 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
5538 // Bitcast operand 1 to i32.
5539 if (SrcVT == MVT::f64)
5540 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
5542 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
5544 // Or in the signbit with integer operations.
5545 SDValue Mask1 = DAG.getConstant(0x80000000, dl, MVT::i32);
5546 SDValue Mask2 = DAG.getConstant(0x7fffffff, dl, MVT::i32);
5547 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
5548 if (VT == MVT::f32) {
5549 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
5550 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
5551 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
5552 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
5555 // f64: Or the high part with signbit and then combine two parts.
5556 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
5558 SDValue Lo = Tmp0.getValue(0);
5559 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
5560 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
5561 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
5564 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
5565 MachineFunction &MF = DAG.getMachineFunction();
5566 MachineFrameInfo &MFI = MF.getFrameInfo();
5567 MFI.setReturnAddressIsTaken(true);
5569 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
5572 EVT VT = Op.getValueType();
5574 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5576 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
5577 SDValue Offset = DAG.getConstant(4, dl, MVT::i32);
5578 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
5579 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
5580 MachinePointerInfo());
5583 // Return LR, which contains the return address. Mark it an implicit live-in.
5584 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
5585 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
5588 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
5589 const ARMBaseRegisterInfo &ARI =
5590 *static_cast<const ARMBaseRegisterInfo*>(RegInfo);
5591 MachineFunction &MF = DAG.getMachineFunction();
5592 MachineFrameInfo &MFI = MF.getFrameInfo();
5593 MFI.setFrameAddressIsTaken(true);
5595 EVT VT = Op.getValueType();
5596 SDLoc dl(Op); // FIXME probably not meaningful
5597 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5598 Register FrameReg = ARI.getFrameRegister(MF);
5599 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
5601 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
5602 MachinePointerInfo());
5606 // FIXME? Maybe this could be a TableGen attribute on some registers and
5607 // this table could be generated automatically from RegInfo.
5608 Register ARMTargetLowering::getRegisterByName(const char* RegName, LLT VT,
5609 const MachineFunction &MF) const {
5610 Register Reg = StringSwitch<unsigned>(RegName)
5611 .Case("sp", ARM::SP)
5615 report_fatal_error(Twine("Invalid register name \""
5616 + StringRef(RegName) + "\"."));
5619 // Result is 64 bit value so split into two 32 bit values and return as a
5621 static void ExpandREAD_REGISTER(SDNode *N, SmallVectorImpl<SDValue> &Results,
5622 SelectionDAG &DAG) {
5625 // This function is only supposed to be called for i64 type destination.
5626 assert(N->getValueType(0) == MVT::i64
5627 && "ExpandREAD_REGISTER called for non-i64 type result.");
5629 SDValue Read = DAG.getNode(ISD::READ_REGISTER, DL,
5630 DAG.getVTList(MVT::i32, MVT::i32, MVT::Other),
5634 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Read.getValue(0),
5636 Results.push_back(Read.getOperand(0));
5639 /// \p BC is a bitcast that is about to be turned into a VMOVDRR.
5640 /// When \p DstVT, the destination type of \p BC, is on the vector
5641 /// register bank and the source of bitcast, \p Op, operates on the same bank,
5642 /// it might be possible to combine them, such that everything stays on the
5643 /// vector register bank.
5644 /// \p return The node that would replace \p BT, if the combine
5646 static SDValue CombineVMOVDRRCandidateWithVecOp(const SDNode *BC,
5647 SelectionDAG &DAG) {
5648 SDValue Op = BC->getOperand(0);
5649 EVT DstVT = BC->getValueType(0);
5651 // The only vector instruction that can produce a scalar (remember,
5652 // since the bitcast was about to be turned into VMOVDRR, the source
5653 // type is i64) from a vector is EXTRACT_VECTOR_ELT.
5654 // Moreover, we can do this combine only if there is one use.
5655 // Finally, if the destination type is not a vector, there is not
5656 // much point on forcing everything on the vector bank.
5657 if (!DstVT.isVector() || Op.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5661 // If the index is not constant, we will introduce an additional
5662 // multiply that will stick.
5663 // Give up in that case.
5664 ConstantSDNode *Index = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5667 unsigned DstNumElt = DstVT.getVectorNumElements();
5669 // Compute the new index.
5670 const APInt &APIntIndex = Index->getAPIntValue();
5671 APInt NewIndex(APIntIndex.getBitWidth(), DstNumElt);
5672 NewIndex *= APIntIndex;
5673 // Check if the new constant index fits into i32.
5674 if (NewIndex.getBitWidth() > 32)
5677 // vMTy bitcast(i64 extractelt vNi64 src, i32 index) ->
5678 // vMTy extractsubvector vNxMTy (bitcast vNi64 src), i32 index*M)
5680 SDValue ExtractSrc = Op.getOperand(0);
5681 EVT VecVT = EVT::getVectorVT(
5682 *DAG.getContext(), DstVT.getScalarType(),
5683 ExtractSrc.getValueType().getVectorNumElements() * DstNumElt);
5684 SDValue BitCast = DAG.getNode(ISD::BITCAST, dl, VecVT, ExtractSrc);
5685 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DstVT, BitCast,
5686 DAG.getConstant(NewIndex.getZExtValue(), dl, MVT::i32));
5689 /// ExpandBITCAST - If the target supports VFP, this function is called to
5690 /// expand a bit convert where either the source or destination type is i64 to
5691 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
5692 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
5693 /// vectors), since the legalizer won't know what to do with that.
5694 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG,
5695 const ARMSubtarget *Subtarget) {
5696 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5698 SDValue Op = N->getOperand(0);
5700 // This function is only supposed to be called for i64 types, either as the
5701 // source or destination of the bit convert.
5702 EVT SrcVT = Op.getValueType();
5703 EVT DstVT = N->getValueType(0);
5704 const bool HasFullFP16 = Subtarget->hasFullFP16();
5706 if (SrcVT == MVT::f32 && DstVT == MVT::i32) {
5707 // FullFP16: half values are passed in S-registers, and we don't
5708 // need any of the bitcast and moves:
5710 // t2: f32,ch = CopyFromReg t0, Register:f32 %0
5711 // t5: i32 = bitcast t2
5712 // t18: f16 = ARMISD::VMOVhr t5
5713 if (Op.getOpcode() != ISD::CopyFromReg ||
5714 Op.getValueType() != MVT::f32)
5717 auto Move = N->use_begin();
5718 if (Move->getOpcode() != ARMISD::VMOVhr)
5721 SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1) };
5722 SDValue Copy = DAG.getNode(ISD::CopyFromReg, SDLoc(Op), MVT::f16, Ops);
5723 DAG.ReplaceAllUsesWith(*Move, &Copy);
5727 if (SrcVT == MVT::i16 && DstVT == MVT::f16) {
5730 // SoftFP: read half-precision arguments:
5733 // t7: i16 = truncate t2 <~~~~ Op
5734 // t8: f16 = bitcast t7 <~~~~ N
5736 if (Op.getOperand(0).getValueType() == MVT::i32)
5737 return DAG.getNode(ARMISD::VMOVhr, SDLoc(Op),
5738 MVT::f16, Op.getOperand(0));
5743 // Half-precision return values
5744 if (SrcVT == MVT::f16 && DstVT == MVT::i16) {
5748 // t11: f16 = fadd t8, t10
5749 // t12: i16 = bitcast t11 <~~~ SDNode N
5750 // t13: i32 = zero_extend t12
5751 // t16: ch,glue = CopyToReg t0, Register:i32 %r0, t13
5752 // t17: ch = ARMISD::RET_FLAG t16, Register:i32 %r0, t16:1
5754 // transform this into:
5756 // t20: i32 = ARMISD::VMOVrh t11
5757 // t16: ch,glue = CopyToReg t0, Register:i32 %r0, t20
5759 auto ZeroExtend = N->use_begin();
5760 if (N->use_size() != 1 || ZeroExtend->getOpcode() != ISD::ZERO_EXTEND ||
5761 ZeroExtend->getValueType(0) != MVT::i32)
5764 auto Copy = ZeroExtend->use_begin();
5765 if (Copy->getOpcode() == ISD::CopyToReg &&
5766 Copy->use_begin()->getOpcode() == ARMISD::RET_FLAG) {
5767 SDValue Cvt = DAG.getNode(ARMISD::VMOVrh, SDLoc(Op), MVT::i32, Op);
5768 DAG.ReplaceAllUsesWith(*ZeroExtend, &Cvt);
5774 if (!(SrcVT == MVT::i64 || DstVT == MVT::i64))
5777 // Turn i64->f64 into VMOVDRR.
5778 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
5779 // Do not force values to GPRs (this is what VMOVDRR does for the inputs)
5780 // if we can combine the bitcast with its source.
5781 if (SDValue Val = CombineVMOVDRRCandidateWithVecOp(N, DAG))
5784 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
5785 DAG.getConstant(0, dl, MVT::i32));
5786 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
5787 DAG.getConstant(1, dl, MVT::i32));
5788 return DAG.getNode(ISD::BITCAST, dl, DstVT,
5789 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
5792 // Turn f64->i64 into VMOVRRD.
5793 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
5795 if (DAG.getDataLayout().isBigEndian() && SrcVT.isVector() &&
5796 SrcVT.getVectorNumElements() > 1)
5797 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
5798 DAG.getVTList(MVT::i32, MVT::i32),
5799 DAG.getNode(ARMISD::VREV64, dl, SrcVT, Op));
5801 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
5802 DAG.getVTList(MVT::i32, MVT::i32), Op);
5803 // Merge the pieces into a single i64 value.
5804 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
5810 /// getZeroVector - Returns a vector of specified type with all zero elements.
5811 /// Zero vectors are used to represent vector negation and in those cases
5812 /// will be implemented with the NEON VNEG instruction. However, VNEG does
5813 /// not support i64 elements, so sometimes the zero vectors will need to be
5814 /// explicitly constructed. Regardless, use a canonical VMOV to create the
5816 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, const SDLoc &dl) {
5817 assert(VT.isVector() && "Expected a vector type");
5818 // The canonical modified immediate encoding of a zero vector is....0!
5819 SDValue EncodedVal = DAG.getTargetConstant(0, dl, MVT::i32);
5820 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
5821 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
5822 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5825 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
5826 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
5827 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
5828 SelectionDAG &DAG) const {
5829 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5830 EVT VT = Op.getValueType();
5831 unsigned VTBits = VT.getSizeInBits();
5833 SDValue ShOpLo = Op.getOperand(0);
5834 SDValue ShOpHi = Op.getOperand(1);
5835 SDValue ShAmt = Op.getOperand(2);
5837 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
5838 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
5840 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
5842 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
5843 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
5844 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
5845 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
5846 DAG.getConstant(VTBits, dl, MVT::i32));
5847 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
5848 SDValue LoSmallShift = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
5849 SDValue LoBigShift = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
5850 SDValue CmpLo = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
5851 ISD::SETGE, ARMcc, DAG, dl);
5852 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, LoSmallShift, LoBigShift,
5855 SDValue HiSmallShift = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
5856 SDValue HiBigShift = Opc == ISD::SRA
5857 ? DAG.getNode(Opc, dl, VT, ShOpHi,
5858 DAG.getConstant(VTBits - 1, dl, VT))
5859 : DAG.getConstant(0, dl, VT);
5860 SDValue CmpHi = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
5861 ISD::SETGE, ARMcc, DAG, dl);
5862 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, HiSmallShift, HiBigShift,
5865 SDValue Ops[2] = { Lo, Hi };
5866 return DAG.getMergeValues(Ops, dl);
5869 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
5870 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
5871 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
5872 SelectionDAG &DAG) const {
5873 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5874 EVT VT = Op.getValueType();
5875 unsigned VTBits = VT.getSizeInBits();
5877 SDValue ShOpLo = Op.getOperand(0);
5878 SDValue ShOpHi = Op.getOperand(1);
5879 SDValue ShAmt = Op.getOperand(2);
5881 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
5883 assert(Op.getOpcode() == ISD::SHL_PARTS);
5884 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
5885 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
5886 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
5887 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
5888 SDValue HiSmallShift = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
5890 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
5891 DAG.getConstant(VTBits, dl, MVT::i32));
5892 SDValue HiBigShift = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
5893 SDValue CmpHi = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
5894 ISD::SETGE, ARMcc, DAG, dl);
5895 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, HiSmallShift, HiBigShift,
5898 SDValue CmpLo = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
5899 ISD::SETGE, ARMcc, DAG, dl);
5900 SDValue LoSmallShift = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5901 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, LoSmallShift,
5902 DAG.getConstant(0, dl, VT), ARMcc, CCR, CmpLo);
5904 SDValue Ops[2] = { Lo, Hi };
5905 return DAG.getMergeValues(Ops, dl);
5908 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
5909 SelectionDAG &DAG) const {
5910 // The rounding mode is in bits 23:22 of the FPSCR.
5911 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
5912 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
5913 // so that the shift + and get folded into a bitfield extract.
5915 SDValue Ops[] = { DAG.getEntryNode(),
5916 DAG.getConstant(Intrinsic::arm_get_fpscr, dl, MVT::i32) };
5918 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_W_CHAIN, dl, MVT::i32, Ops);
5919 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
5920 DAG.getConstant(1U << 22, dl, MVT::i32));
5921 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
5922 DAG.getConstant(22, dl, MVT::i32));
5923 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
5924 DAG.getConstant(3, dl, MVT::i32));
5927 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
5928 const ARMSubtarget *ST) {
5930 EVT VT = N->getValueType(0);
5931 if (VT.isVector() && ST->hasNEON()) {
5933 // Compute the least significant set bit: LSB = X & -X
5934 SDValue X = N->getOperand(0);
5935 SDValue NX = DAG.getNode(ISD::SUB, dl, VT, getZeroVector(VT, DAG, dl), X);
5936 SDValue LSB = DAG.getNode(ISD::AND, dl, VT, X, NX);
5938 EVT ElemTy = VT.getVectorElementType();
5940 if (ElemTy == MVT::i8) {
5941 // Compute with: cttz(x) = ctpop(lsb - 1)
5942 SDValue One = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
5943 DAG.getTargetConstant(1, dl, ElemTy));
5944 SDValue Bits = DAG.getNode(ISD::SUB, dl, VT, LSB, One);
5945 return DAG.getNode(ISD::CTPOP, dl, VT, Bits);
5948 if ((ElemTy == MVT::i16 || ElemTy == MVT::i32) &&
5949 (N->getOpcode() == ISD::CTTZ_ZERO_UNDEF)) {
5950 // Compute with: cttz(x) = (width - 1) - ctlz(lsb), if x != 0
5951 unsigned NumBits = ElemTy.getSizeInBits();
5952 SDValue WidthMinus1 =
5953 DAG.getNode(ARMISD::VMOVIMM, dl, VT,
5954 DAG.getTargetConstant(NumBits - 1, dl, ElemTy));
5955 SDValue CTLZ = DAG.getNode(ISD::CTLZ, dl, VT, LSB);
5956 return DAG.getNode(ISD::SUB, dl, VT, WidthMinus1, CTLZ);
5959 // Compute with: cttz(x) = ctpop(lsb - 1)
5963 if (ElemTy == MVT::i64) {
5964 // Load constant 0xffff'ffff'ffff'ffff to register.
5965 SDValue FF = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
5966 DAG.getTargetConstant(0x1eff, dl, MVT::i32));
5967 Bits = DAG.getNode(ISD::ADD, dl, VT, LSB, FF);
5969 SDValue One = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
5970 DAG.getTargetConstant(1, dl, ElemTy));
5971 Bits = DAG.getNode(ISD::SUB, dl, VT, LSB, One);
5973 return DAG.getNode(ISD::CTPOP, dl, VT, Bits);
5976 if (!ST->hasV6T2Ops())
5979 SDValue rbit = DAG.getNode(ISD::BITREVERSE, dl, VT, N->getOperand(0));
5980 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
5983 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
5984 const ARMSubtarget *ST) {
5985 EVT VT = N->getValueType(0);
5988 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
5989 assert((VT == MVT::v1i64 || VT == MVT::v2i64 || VT == MVT::v2i32 ||
5990 VT == MVT::v4i32 || VT == MVT::v4i16 || VT == MVT::v8i16) &&
5991 "Unexpected type for custom ctpop lowering");
5993 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5994 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
5995 SDValue Res = DAG.getBitcast(VT8Bit, N->getOperand(0));
5996 Res = DAG.getNode(ISD::CTPOP, DL, VT8Bit, Res);
5998 // Widen v8i8/v16i8 CTPOP result to VT by repeatedly widening pairwise adds.
5999 unsigned EltSize = 8;
6000 unsigned NumElts = VT.is64BitVector() ? 8 : 16;
6001 while (EltSize != VT.getScalarSizeInBits()) {
6002 SmallVector<SDValue, 8> Ops;
6003 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddlu, DL,
6004 TLI.getPointerTy(DAG.getDataLayout())));
6009 MVT WidenVT = MVT::getVectorVT(MVT::getIntegerVT(EltSize), NumElts);
6010 Res = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, WidenVT, Ops);
6016 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
6017 /// operand of a vector shift operation, where all the elements of the
6018 /// build_vector must have the same constant integer value.
6019 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
6020 // Ignore bit_converts.
6021 while (Op.getOpcode() == ISD::BITCAST)
6022 Op = Op.getOperand(0);
6023 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
6024 APInt SplatBits, SplatUndef;
6025 unsigned SplatBitSize;
6028 !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs,
6030 SplatBitSize > ElementBits)
6032 Cnt = SplatBits.getSExtValue();
6036 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
6037 /// operand of a vector shift left operation. That value must be in the range:
6038 /// 0 <= Value < ElementBits for a left shift; or
6039 /// 0 <= Value <= ElementBits for a long left shift.
6040 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
6041 assert(VT.isVector() && "vector shift count is not a vector type");
6042 int64_t ElementBits = VT.getScalarSizeInBits();
6043 if (!getVShiftImm(Op, ElementBits, Cnt))
6045 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
6048 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
6049 /// operand of a vector shift right operation. For a shift opcode, the value
6050 /// is positive, but for an intrinsic the value count must be negative. The
6051 /// absolute value must be in the range:
6052 /// 1 <= |Value| <= ElementBits for a right shift; or
6053 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
6054 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
6056 assert(VT.isVector() && "vector shift count is not a vector type");
6057 int64_t ElementBits = VT.getScalarSizeInBits();
6058 if (!getVShiftImm(Op, ElementBits, Cnt))
6061 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
6062 if (Cnt >= -(isNarrow ? ElementBits / 2 : ElementBits) && Cnt <= -1) {
6069 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
6070 const ARMSubtarget *ST) {
6071 EVT VT = N->getValueType(0);
6078 // We essentially have two forms here. Shift by an immediate and shift by a
6079 // vector register (there are also shift by a gpr, but that is just handled
6080 // with a tablegen pattern). We cannot easily match shift by an immediate in
6081 // tablegen so we do that here and generate a VSHLIMM/VSHRsIMM/VSHRuIMM.
6082 // For shifting by a vector, we don't have VSHR, only VSHL (which can be
6083 // signed or unsigned, and a negative shift indicates a shift right).
6084 if (N->getOpcode() == ISD::SHL) {
6085 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
6086 return DAG.getNode(ARMISD::VSHLIMM, dl, VT, N->getOperand(0),
6087 DAG.getConstant(Cnt, dl, MVT::i32));
6088 return DAG.getNode(ARMISD::VSHLu, dl, VT, N->getOperand(0),
6092 assert((N->getOpcode() == ISD::SRA || N->getOpcode() == ISD::SRL) &&
6093 "unexpected vector shift opcode");
6095 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
6096 unsigned VShiftOpc =
6097 (N->getOpcode() == ISD::SRA ? ARMISD::VSHRsIMM : ARMISD::VSHRuIMM);
6098 return DAG.getNode(VShiftOpc, dl, VT, N->getOperand(0),
6099 DAG.getConstant(Cnt, dl, MVT::i32));
6102 // Other right shifts we don't have operations for (we use a shift left by a
6103 // negative number).
6104 EVT ShiftVT = N->getOperand(1).getValueType();
6105 SDValue NegatedCount = DAG.getNode(
6106 ISD::SUB, dl, ShiftVT, getZeroVector(ShiftVT, DAG, dl), N->getOperand(1));
6107 unsigned VShiftOpc =
6108 (N->getOpcode() == ISD::SRA ? ARMISD::VSHLs : ARMISD::VSHLu);
6109 return DAG.getNode(VShiftOpc, dl, VT, N->getOperand(0), NegatedCount);
6112 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
6113 const ARMSubtarget *ST) {
6114 EVT VT = N->getValueType(0);
6117 // We can get here for a node like i32 = ISD::SHL i32, i64
6121 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA ||
6122 N->getOpcode() == ISD::SHL) &&
6123 "Unknown shift to lower!");
6125 unsigned ShOpc = N->getOpcode();
6126 if (ST->hasMVEIntegerOps()) {
6127 SDValue ShAmt = N->getOperand(1);
6128 unsigned ShPartsOpc = ARMISD::LSLL;
6129 ConstantSDNode *Con = dyn_cast<ConstantSDNode>(ShAmt);
6131 // If the shift amount is greater than 32 or has a greater bitwidth than 64
6132 // then do the default optimisation
6133 if (ShAmt->getValueType(0).getSizeInBits() > 64 ||
6134 (Con && (Con->getZExtValue() == 0 || Con->getZExtValue() >= 32)))
6137 // Extract the lower 32 bits of the shift amount if it's not an i32
6138 if (ShAmt->getValueType(0) != MVT::i32)
6139 ShAmt = DAG.getZExtOrTrunc(ShAmt, dl, MVT::i32);
6141 if (ShOpc == ISD::SRL) {
6143 // There is no t2LSRLr instruction so negate and perform an lsll if the
6144 // shift amount is in a register, emulating a right shift.
6145 ShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
6146 DAG.getConstant(0, dl, MVT::i32), ShAmt);
6148 // Else generate an lsrl on the immediate shift amount
6149 ShPartsOpc = ARMISD::LSRL;
6150 } else if (ShOpc == ISD::SRA)
6151 ShPartsOpc = ARMISD::ASRL;
6153 // Lower 32 bits of the destination/source
6154 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
6155 DAG.getConstant(0, dl, MVT::i32));
6156 // Upper 32 bits of the destination/source
6157 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
6158 DAG.getConstant(1, dl, MVT::i32));
6160 // Generate the shift operation as computed above
6161 Lo = DAG.getNode(ShPartsOpc, dl, DAG.getVTList(MVT::i32, MVT::i32), Lo, Hi,
6163 // The upper 32 bits come from the second return value of lsll
6164 Hi = SDValue(Lo.getNode(), 1);
6165 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
6168 // We only lower SRA, SRL of 1 here, all others use generic lowering.
6169 if (!isOneConstant(N->getOperand(1)) || N->getOpcode() == ISD::SHL)
6172 // If we are in thumb mode, we don't have RRX.
6173 if (ST->isThumb1Only())
6176 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
6177 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
6178 DAG.getConstant(0, dl, MVT::i32));
6179 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
6180 DAG.getConstant(1, dl, MVT::i32));
6182 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
6183 // captures the result into a carry flag.
6184 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
6185 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), Hi);
6187 // The low part is an ARMISD::RRX operand, which shifts the carry in.
6188 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
6190 // Merge the pieces into a single i64 value.
6191 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
6194 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG,
6195 const ARMSubtarget *ST) {
6196 bool Invert = false;
6198 unsigned Opc = ARMCC::AL;
6200 SDValue Op0 = Op.getOperand(0);
6201 SDValue Op1 = Op.getOperand(1);
6202 SDValue CC = Op.getOperand(2);
6203 EVT VT = Op.getValueType();
6204 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6209 CmpVT = Op0.getValueType().changeVectorElementTypeToInteger();
6211 assert(ST->hasMVEIntegerOps() &&
6212 "No hardware support for integer vector comparison!");
6214 if (Op.getValueType().getVectorElementType() != MVT::i1)
6217 // Make sure we expand floating point setcc to scalar if we do not have
6218 // mve.fp, so that we can handle them from there.
6219 if (Op0.getValueType().isFloatingPoint() && !ST->hasMVEFloatOps())
6225 if (Op0.getValueType().getVectorElementType() == MVT::i64 &&
6226 (SetCCOpcode == ISD::SETEQ || SetCCOpcode == ISD::SETNE)) {
6227 // Special-case integer 64-bit equality comparisons. They aren't legal,
6228 // but they can be lowered with a few vector instructions.
6229 unsigned CmpElements = CmpVT.getVectorNumElements() * 2;
6230 EVT SplitVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, CmpElements);
6231 SDValue CastOp0 = DAG.getNode(ISD::BITCAST, dl, SplitVT, Op0);
6232 SDValue CastOp1 = DAG.getNode(ISD::BITCAST, dl, SplitVT, Op1);
6233 SDValue Cmp = DAG.getNode(ISD::SETCC, dl, SplitVT, CastOp0, CastOp1,
6234 DAG.getCondCode(ISD::SETEQ));
6235 SDValue Reversed = DAG.getNode(ARMISD::VREV64, dl, SplitVT, Cmp);
6236 SDValue Merged = DAG.getNode(ISD::AND, dl, SplitVT, Cmp, Reversed);
6237 Merged = DAG.getNode(ISD::BITCAST, dl, CmpVT, Merged);
6238 if (SetCCOpcode == ISD::SETNE)
6239 Merged = DAG.getNOT(dl, Merged, CmpVT);
6240 Merged = DAG.getSExtOrTrunc(Merged, dl, VT);
6244 if (CmpVT.getVectorElementType() == MVT::i64)
6245 // 64-bit comparisons are not legal in general.
6248 if (Op1.getValueType().isFloatingPoint()) {
6249 switch (SetCCOpcode) {
6250 default: llvm_unreachable("Illegal FP comparison");
6253 if (ST->hasMVEFloatOps()) {
6254 Opc = ARMCC::NE; break;
6256 Invert = true; LLVM_FALLTHROUGH;
6259 case ISD::SETEQ: Opc = ARMCC::EQ; break;
6261 case ISD::SETLT: Swap = true; LLVM_FALLTHROUGH;
6263 case ISD::SETGT: Opc = ARMCC::GT; break;
6265 case ISD::SETLE: Swap = true; LLVM_FALLTHROUGH;
6267 case ISD::SETGE: Opc = ARMCC::GE; break;
6268 case ISD::SETUGE: Swap = true; LLVM_FALLTHROUGH;
6269 case ISD::SETULE: Invert = true; Opc = ARMCC::GT; break;
6270 case ISD::SETUGT: Swap = true; LLVM_FALLTHROUGH;
6271 case ISD::SETULT: Invert = true; Opc = ARMCC::GE; break;
6272 case ISD::SETUEQ: Invert = true; LLVM_FALLTHROUGH;
6274 // Expand this to (OLT | OGT).
6275 SDValue TmpOp0 = DAG.getNode(ARMISD::VCMP, dl, CmpVT, Op1, Op0,
6276 DAG.getConstant(ARMCC::GT, dl, MVT::i32));
6277 SDValue TmpOp1 = DAG.getNode(ARMISD::VCMP, dl, CmpVT, Op0, Op1,
6278 DAG.getConstant(ARMCC::GT, dl, MVT::i32));
6279 SDValue Result = DAG.getNode(ISD::OR, dl, CmpVT, TmpOp0, TmpOp1);
6281 Result = DAG.getNOT(dl, Result, VT);
6284 case ISD::SETUO: Invert = true; LLVM_FALLTHROUGH;
6286 // Expand this to (OLT | OGE).
6287 SDValue TmpOp0 = DAG.getNode(ARMISD::VCMP, dl, CmpVT, Op1, Op0,
6288 DAG.getConstant(ARMCC::GT, dl, MVT::i32));
6289 SDValue TmpOp1 = DAG.getNode(ARMISD::VCMP, dl, CmpVT, Op0, Op1,
6290 DAG.getConstant(ARMCC::GE, dl, MVT::i32));
6291 SDValue Result = DAG.getNode(ISD::OR, dl, CmpVT, TmpOp0, TmpOp1);
6293 Result = DAG.getNOT(dl, Result, VT);
6298 // Integer comparisons.
6299 switch (SetCCOpcode) {
6300 default: llvm_unreachable("Illegal integer comparison");
6302 if (ST->hasMVEIntegerOps()) {
6303 Opc = ARMCC::NE; break;
6305 Invert = true; LLVM_FALLTHROUGH;
6307 case ISD::SETEQ: Opc = ARMCC::EQ; break;
6308 case ISD::SETLT: Swap = true; LLVM_FALLTHROUGH;
6309 case ISD::SETGT: Opc = ARMCC::GT; break;
6310 case ISD::SETLE: Swap = true; LLVM_FALLTHROUGH;
6311 case ISD::SETGE: Opc = ARMCC::GE; break;
6312 case ISD::SETULT: Swap = true; LLVM_FALLTHROUGH;
6313 case ISD::SETUGT: Opc = ARMCC::HI; break;
6314 case ISD::SETULE: Swap = true; LLVM_FALLTHROUGH;
6315 case ISD::SETUGE: Opc = ARMCC::HS; break;
6318 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
6319 if (ST->hasNEON() && Opc == ARMCC::EQ) {
6321 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
6323 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
6326 // Ignore bitconvert.
6327 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
6328 AndOp = AndOp.getOperand(0);
6330 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
6331 Op0 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(0));
6332 Op1 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(1));
6333 SDValue Result = DAG.getNode(ARMISD::VTST, dl, CmpVT, Op0, Op1);
6335 Result = DAG.getNOT(dl, Result, VT);
6342 std::swap(Op0, Op1);
6344 // If one of the operands is a constant vector zero, attempt to fold the
6345 // comparison to a specialized compare-against-zero form.
6347 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
6349 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
6350 if (Opc == ARMCC::GE)
6352 else if (Opc == ARMCC::GT)
6358 if (SingleOp.getNode()) {
6359 Result = DAG.getNode(ARMISD::VCMPZ, dl, CmpVT, SingleOp,
6360 DAG.getConstant(Opc, dl, MVT::i32));
6362 Result = DAG.getNode(ARMISD::VCMP, dl, CmpVT, Op0, Op1,
6363 DAG.getConstant(Opc, dl, MVT::i32));
6366 Result = DAG.getSExtOrTrunc(Result, dl, VT);
6369 Result = DAG.getNOT(dl, Result, VT);
6374 static SDValue LowerSETCCCARRY(SDValue Op, SelectionDAG &DAG) {
6375 SDValue LHS = Op.getOperand(0);
6376 SDValue RHS = Op.getOperand(1);
6377 SDValue Carry = Op.getOperand(2);
6378 SDValue Cond = Op.getOperand(3);
6381 assert(LHS.getSimpleValueType().isInteger() && "SETCCCARRY is integer only.");
6383 // ARMISD::SUBE expects a carry not a borrow like ISD::SUBCARRY so we
6384 // have to invert the carry first.
6385 Carry = DAG.getNode(ISD::SUB, DL, MVT::i32,
6386 DAG.getConstant(1, DL, MVT::i32), Carry);
6387 // This converts the boolean value carry into the carry flag.
6388 Carry = ConvertBooleanCarryToCarryFlag(Carry, DAG);
6390 SDVTList VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
6391 SDValue Cmp = DAG.getNode(ARMISD::SUBE, DL, VTs, LHS, RHS, Carry);
6393 SDValue FVal = DAG.getConstant(0, DL, MVT::i32);
6394 SDValue TVal = DAG.getConstant(1, DL, MVT::i32);
6395 SDValue ARMcc = DAG.getConstant(
6396 IntCCToARMCC(cast<CondCodeSDNode>(Cond)->get()), DL, MVT::i32);
6397 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
6398 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), DL, ARM::CPSR,
6399 Cmp.getValue(1), SDValue());
6400 return DAG.getNode(ARMISD::CMOV, DL, Op.getValueType(), FVal, TVal, ARMcc,
6401 CCR, Chain.getValue(1));
6404 /// isVMOVModifiedImm - Check if the specified splat value corresponds to a
6405 /// valid vector constant for a NEON or MVE instruction with a "modified
6406 /// immediate" operand (e.g., VMOV). If so, return the encoded value.
6407 static SDValue isVMOVModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
6408 unsigned SplatBitSize, SelectionDAG &DAG,
6409 const SDLoc &dl, EVT &VT, bool is128Bits,
6410 VMOVModImmType type) {
6411 unsigned OpCmode, Imm;
6413 // SplatBitSize is set to the smallest size that splats the vector, so a
6414 // zero vector will always have SplatBitSize == 8. However, NEON modified
6415 // immediate instructions others than VMOV do not support the 8-bit encoding
6416 // of a zero vector, and the default encoding of zero is supposed to be the
6421 switch (SplatBitSize) {
6423 if (type != VMOVModImm)
6425 // Any 1-byte value is OK. Op=0, Cmode=1110.
6426 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
6429 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
6433 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
6434 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
6435 if ((SplatBits & ~0xff) == 0) {
6436 // Value = 0x00nn: Op=x, Cmode=100x.
6441 if ((SplatBits & ~0xff00) == 0) {
6442 // Value = 0xnn00: Op=x, Cmode=101x.
6444 Imm = SplatBits >> 8;
6450 // NEON's 32-bit VMOV supports splat values where:
6451 // * only one byte is nonzero, or
6452 // * the least significant byte is 0xff and the second byte is nonzero, or
6453 // * the least significant 2 bytes are 0xff and the third is nonzero.
6454 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
6455 if ((SplatBits & ~0xff) == 0) {
6456 // Value = 0x000000nn: Op=x, Cmode=000x.
6461 if ((SplatBits & ~0xff00) == 0) {
6462 // Value = 0x0000nn00: Op=x, Cmode=001x.
6464 Imm = SplatBits >> 8;
6467 if ((SplatBits & ~0xff0000) == 0) {
6468 // Value = 0x00nn0000: Op=x, Cmode=010x.
6470 Imm = SplatBits >> 16;
6473 if ((SplatBits & ~0xff000000) == 0) {
6474 // Value = 0xnn000000: Op=x, Cmode=011x.
6476 Imm = SplatBits >> 24;
6480 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
6481 if (type == OtherModImm) return SDValue();
6483 if ((SplatBits & ~0xffff) == 0 &&
6484 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
6485 // Value = 0x0000nnff: Op=x, Cmode=1100.
6487 Imm = SplatBits >> 8;
6491 // cmode == 0b1101 is not supported for MVE VMVN
6492 if (type == MVEVMVNModImm)
6495 if ((SplatBits & ~0xffffff) == 0 &&
6496 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
6497 // Value = 0x00nnffff: Op=x, Cmode=1101.
6499 Imm = SplatBits >> 16;
6503 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
6504 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
6505 // VMOV.I32. A (very) minor optimization would be to replicate the value
6506 // and fall through here to test for a valid 64-bit splat. But, then the
6507 // caller would also need to check and handle the change in size.
6511 if (type != VMOVModImm)
6513 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
6514 uint64_t BitMask = 0xff;
6516 unsigned ImmMask = 1;
6518 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
6519 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
6522 } else if ((SplatBits & BitMask) != 0) {
6529 if (DAG.getDataLayout().isBigEndian())
6530 // swap higher and lower 32 bit word
6531 Imm = ((Imm & 0xf) << 4) | ((Imm & 0xf0) >> 4);
6533 // Op=1, Cmode=1110.
6535 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
6540 llvm_unreachable("unexpected size for isVMOVModifiedImm");
6543 unsigned EncodedVal = ARM_AM::createVMOVModImm(OpCmode, Imm);
6544 return DAG.getTargetConstant(EncodedVal, dl, MVT::i32);
6547 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
6548 const ARMSubtarget *ST) const {
6549 EVT VT = Op.getValueType();
6550 bool IsDouble = (VT == MVT::f64);
6551 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
6552 const APFloat &FPVal = CFP->getValueAPF();
6554 // Prevent floating-point constants from using literal loads
6555 // when execute-only is enabled.
6556 if (ST->genExecuteOnly()) {
6557 // If we can represent the constant as an immediate, don't lower it
6558 if (isFPImmLegal(FPVal, VT))
6560 // Otherwise, construct as integer, and move to float register
6561 APInt INTVal = FPVal.bitcastToAPInt();
6563 switch (VT.getSimpleVT().SimpleTy) {
6565 llvm_unreachable("Unknown floating point type!");
6568 SDValue Lo = DAG.getConstant(INTVal.trunc(32), DL, MVT::i32);
6569 SDValue Hi = DAG.getConstant(INTVal.lshr(32).trunc(32), DL, MVT::i32);
6570 if (!ST->isLittle())
6572 return DAG.getNode(ARMISD::VMOVDRR, DL, MVT::f64, Lo, Hi);
6575 return DAG.getNode(ARMISD::VMOVSR, DL, VT,
6576 DAG.getConstant(INTVal, DL, MVT::i32));
6580 if (!ST->hasVFP3Base())
6583 // Use the default (constant pool) lowering for double constants when we have
6585 if (IsDouble && !Subtarget->hasFP64())
6588 // Try splatting with a VMOV.f32...
6589 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
6592 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
6593 // We have code in place to select a valid ConstantFP already, no need to
6598 // It's a float and we are trying to use NEON operations where
6599 // possible. Lower it to a splat followed by an extract.
6601 SDValue NewVal = DAG.getTargetConstant(ImmVal, DL, MVT::i32);
6602 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
6604 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
6605 DAG.getConstant(0, DL, MVT::i32));
6608 // The rest of our options are NEON only, make sure that's allowed before
6610 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
6614 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
6616 // It wouldn't really be worth bothering for doubles except for one very
6617 // important value, which does happen to match: 0.0. So make sure we don't do
6619 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
6622 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
6623 SDValue NewVal = isVMOVModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op),
6624 VMovVT, false, VMOVModImm);
6625 if (NewVal != SDValue()) {
6627 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
6630 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
6632 // It's a float: cast and extract a vector element.
6633 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
6635 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
6636 DAG.getConstant(0, DL, MVT::i32));
6639 // Finally, try a VMVN.i32
6640 NewVal = isVMOVModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op), VMovVT,
6642 if (NewVal != SDValue()) {
6644 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
6647 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
6649 // It's a float: cast and extract a vector element.
6650 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
6652 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
6653 DAG.getConstant(0, DL, MVT::i32));
6659 // check if an VEXT instruction can handle the shuffle mask when the
6660 // vector sources of the shuffle are the same.
6661 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
6662 unsigned NumElts = VT.getVectorNumElements();
6664 // Assume that the first shuffle index is not UNDEF. Fail if it is.
6670 // If this is a VEXT shuffle, the immediate value is the index of the first
6671 // element. The other shuffle indices must be the successive elements after
6673 unsigned ExpectedElt = Imm;
6674 for (unsigned i = 1; i < NumElts; ++i) {
6675 // Increment the expected index. If it wraps around, just follow it
6676 // back to index zero and keep going.
6678 if (ExpectedElt == NumElts)
6681 if (M[i] < 0) continue; // ignore UNDEF indices
6682 if (ExpectedElt != static_cast<unsigned>(M[i]))
6689 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
6690 bool &ReverseVEXT, unsigned &Imm) {
6691 unsigned NumElts = VT.getVectorNumElements();
6692 ReverseVEXT = false;
6694 // Assume that the first shuffle index is not UNDEF. Fail if it is.
6700 // If this is a VEXT shuffle, the immediate value is the index of the first
6701 // element. The other shuffle indices must be the successive elements after
6703 unsigned ExpectedElt = Imm;
6704 for (unsigned i = 1; i < NumElts; ++i) {
6705 // Increment the expected index. If it wraps around, it may still be
6706 // a VEXT but the source vectors must be swapped.
6708 if (ExpectedElt == NumElts * 2) {
6713 if (M[i] < 0) continue; // ignore UNDEF indices
6714 if (ExpectedElt != static_cast<unsigned>(M[i]))
6718 // Adjust the index value if the source operands will be swapped.
6725 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
6726 /// instruction with the specified blocksize. (The order of the elements
6727 /// within each block of the vector is reversed.)
6728 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
6729 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
6730 "Only possible block sizes for VREV are: 16, 32, 64");
6732 unsigned EltSz = VT.getScalarSizeInBits();
6736 unsigned NumElts = VT.getVectorNumElements();
6737 unsigned BlockElts = M[0] + 1;
6738 // If the first shuffle index is UNDEF, be optimistic.
6740 BlockElts = BlockSize / EltSz;
6742 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
6745 for (unsigned i = 0; i < NumElts; ++i) {
6746 if (M[i] < 0) continue; // ignore UNDEF indices
6747 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
6754 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
6755 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
6756 // range, then 0 is placed into the resulting vector. So pretty much any mask
6757 // of 8 elements can work here.
6758 return VT == MVT::v8i8 && M.size() == 8;
6761 static unsigned SelectPairHalf(unsigned Elements, ArrayRef<int> Mask,
6763 if (Mask.size() == Elements * 2)
6764 return Index / Elements;
6765 return Mask[Index] == 0 ? 0 : 1;
6768 // Checks whether the shuffle mask represents a vector transpose (VTRN) by
6769 // checking that pairs of elements in the shuffle mask represent the same index
6770 // in each vector, incrementing the expected index by 2 at each step.
6771 // e.g. For v1,v2 of type v4i32 a valid shuffle mask is: [0, 4, 2, 6]
6772 // v1={a,b,c,d} => x=shufflevector v1, v2 shufflemask => x={a,e,c,g}
6774 // WhichResult gives the offset for each element in the mask based on which
6775 // of the two results it belongs to.
6777 // The transpose can be represented either as:
6778 // result1 = shufflevector v1, v2, result1_shuffle_mask
6779 // result2 = shufflevector v1, v2, result2_shuffle_mask
6780 // where v1/v2 and the shuffle masks have the same number of elements
6781 // (here WhichResult (see below) indicates which result is being checked)
6784 // results = shufflevector v1, v2, shuffle_mask
6785 // where both results are returned in one vector and the shuffle mask has twice
6786 // as many elements as v1/v2 (here WhichResult will always be 0 if true) here we
6787 // want to check the low half and high half of the shuffle mask as if it were
6789 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
6790 unsigned EltSz = VT.getScalarSizeInBits();
6794 unsigned NumElts = VT.getVectorNumElements();
6795 if (M.size() != NumElts && M.size() != NumElts*2)
6798 // If the mask is twice as long as the input vector then we need to check the
6799 // upper and lower parts of the mask with a matching value for WhichResult
6800 // FIXME: A mask with only even values will be rejected in case the first
6801 // element is undefined, e.g. [-1, 4, 2, 6] will be rejected, because only
6802 // M[0] is used to determine WhichResult
6803 for (unsigned i = 0; i < M.size(); i += NumElts) {
6804 WhichResult = SelectPairHalf(NumElts, M, i);
6805 for (unsigned j = 0; j < NumElts; j += 2) {
6806 if ((M[i+j] >= 0 && (unsigned) M[i+j] != j + WhichResult) ||
6807 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != j + NumElts + WhichResult))
6812 if (M.size() == NumElts*2)
6818 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
6819 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
6820 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
6821 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
6822 unsigned EltSz = VT.getScalarSizeInBits();
6826 unsigned NumElts = VT.getVectorNumElements();
6827 if (M.size() != NumElts && M.size() != NumElts*2)
6830 for (unsigned i = 0; i < M.size(); i += NumElts) {
6831 WhichResult = SelectPairHalf(NumElts, M, i);
6832 for (unsigned j = 0; j < NumElts; j += 2) {
6833 if ((M[i+j] >= 0 && (unsigned) M[i+j] != j + WhichResult) ||
6834 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != j + WhichResult))
6839 if (M.size() == NumElts*2)
6845 // Checks whether the shuffle mask represents a vector unzip (VUZP) by checking
6846 // that the mask elements are either all even and in steps of size 2 or all odd
6847 // and in steps of size 2.
6848 // e.g. For v1,v2 of type v4i32 a valid shuffle mask is: [0, 2, 4, 6]
6849 // v1={a,b,c,d} => x=shufflevector v1, v2 shufflemask => x={a,c,e,g}
6851 // Requires similar checks to that of isVTRNMask with
6852 // respect the how results are returned.
6853 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
6854 unsigned EltSz = VT.getScalarSizeInBits();
6858 unsigned NumElts = VT.getVectorNumElements();
6859 if (M.size() != NumElts && M.size() != NumElts*2)
6862 for (unsigned i = 0; i < M.size(); i += NumElts) {
6863 WhichResult = SelectPairHalf(NumElts, M, i);
6864 for (unsigned j = 0; j < NumElts; ++j) {
6865 if (M[i+j] >= 0 && (unsigned) M[i+j] != 2 * j + WhichResult)
6870 if (M.size() == NumElts*2)
6873 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
6874 if (VT.is64BitVector() && EltSz == 32)
6880 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
6881 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
6882 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
6883 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
6884 unsigned EltSz = VT.getScalarSizeInBits();
6888 unsigned NumElts = VT.getVectorNumElements();
6889 if (M.size() != NumElts && M.size() != NumElts*2)
6892 unsigned Half = NumElts / 2;
6893 for (unsigned i = 0; i < M.size(); i += NumElts) {
6894 WhichResult = SelectPairHalf(NumElts, M, i);
6895 for (unsigned j = 0; j < NumElts; j += Half) {
6896 unsigned Idx = WhichResult;
6897 for (unsigned k = 0; k < Half; ++k) {
6898 int MIdx = M[i + j + k];
6899 if (MIdx >= 0 && (unsigned) MIdx != Idx)
6906 if (M.size() == NumElts*2)
6909 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
6910 if (VT.is64BitVector() && EltSz == 32)
6916 // Checks whether the shuffle mask represents a vector zip (VZIP) by checking
6917 // that pairs of elements of the shufflemask represent the same index in each
6918 // vector incrementing sequentially through the vectors.
6919 // e.g. For v1,v2 of type v4i32 a valid shuffle mask is: [0, 4, 1, 5]
6920 // v1={a,b,c,d} => x=shufflevector v1, v2 shufflemask => x={a,e,b,f}
6922 // Requires similar checks to that of isVTRNMask with respect the how results
6924 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
6925 unsigned EltSz = VT.getScalarSizeInBits();
6929 unsigned NumElts = VT.getVectorNumElements();
6930 if (M.size() != NumElts && M.size() != NumElts*2)
6933 for (unsigned i = 0; i < M.size(); i += NumElts) {
6934 WhichResult = SelectPairHalf(NumElts, M, i);
6935 unsigned Idx = WhichResult * NumElts / 2;
6936 for (unsigned j = 0; j < NumElts; j += 2) {
6937 if ((M[i+j] >= 0 && (unsigned) M[i+j] != Idx) ||
6938 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != Idx + NumElts))
6944 if (M.size() == NumElts*2)
6947 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
6948 if (VT.is64BitVector() && EltSz == 32)
6954 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
6955 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
6956 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
6957 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
6958 unsigned EltSz = VT.getScalarSizeInBits();
6962 unsigned NumElts = VT.getVectorNumElements();
6963 if (M.size() != NumElts && M.size() != NumElts*2)
6966 for (unsigned i = 0; i < M.size(); i += NumElts) {
6967 WhichResult = SelectPairHalf(NumElts, M, i);
6968 unsigned Idx = WhichResult * NumElts / 2;
6969 for (unsigned j = 0; j < NumElts; j += 2) {
6970 if ((M[i+j] >= 0 && (unsigned) M[i+j] != Idx) ||
6971 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != Idx))
6977 if (M.size() == NumElts*2)
6980 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
6981 if (VT.is64BitVector() && EltSz == 32)
6987 /// Check if \p ShuffleMask is a NEON two-result shuffle (VZIP, VUZP, VTRN),
6988 /// and return the corresponding ARMISD opcode if it is, or 0 if it isn't.
6989 static unsigned isNEONTwoResultShuffleMask(ArrayRef<int> ShuffleMask, EVT VT,
6990 unsigned &WhichResult,
6993 if (isVTRNMask(ShuffleMask, VT, WhichResult))
6994 return ARMISD::VTRN;
6995 if (isVUZPMask(ShuffleMask, VT, WhichResult))
6996 return ARMISD::VUZP;
6997 if (isVZIPMask(ShuffleMask, VT, WhichResult))
6998 return ARMISD::VZIP;
7001 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
7002 return ARMISD::VTRN;
7003 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
7004 return ARMISD::VUZP;
7005 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
7006 return ARMISD::VZIP;
7011 /// \return true if this is a reverse operation on an vector.
7012 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
7013 unsigned NumElts = VT.getVectorNumElements();
7014 // Make sure the mask has the right size.
7015 if (NumElts != M.size())
7018 // Look for <15, ..., 3, -1, 1, 0>.
7019 for (unsigned i = 0; i != NumElts; ++i)
7020 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
7026 static bool isVMOVNMask(ArrayRef<int> M, EVT VT, bool Top) {
7027 unsigned NumElts = VT.getVectorNumElements();
7028 // Make sure the mask has the right size.
7029 if (NumElts != M.size() || (VT != MVT::v8i16 && VT != MVT::v16i8))
7033 // Look for <0, N, 2, N+2, 4, N+4, ..>.
7034 // This inserts Input2 into Input1
7036 // Look for <0, N+1, 2, N+3, 4, N+5, ..>
7037 // This inserts Input1 into Input2
7038 unsigned Offset = Top ? 0 : 1;
7039 for (unsigned i = 0; i < NumElts; i+=2) {
7040 if (M[i] >= 0 && M[i] != (int)i)
7042 if (M[i+1] >= 0 && M[i+1] != (int)(NumElts + i + Offset))
7049 // If N is an integer constant that can be moved into a register in one
7050 // instruction, return an SDValue of such a constant (will become a MOV
7051 // instruction). Otherwise return null.
7052 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
7053 const ARMSubtarget *ST, const SDLoc &dl) {
7055 if (!isa<ConstantSDNode>(N))
7057 Val = cast<ConstantSDNode>(N)->getZExtValue();
7059 if (ST->isThumb1Only()) {
7060 if (Val <= 255 || ~Val <= 255)
7061 return DAG.getConstant(Val, dl, MVT::i32);
7063 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
7064 return DAG.getConstant(Val, dl, MVT::i32);
7069 static SDValue LowerBUILD_VECTOR_i1(SDValue Op, SelectionDAG &DAG,
7070 const ARMSubtarget *ST) {
7072 EVT VT = Op.getValueType();
7074 assert(ST->hasMVEIntegerOps() && "LowerBUILD_VECTOR_i1 called without MVE!");
7076 unsigned NumElts = VT.getVectorNumElements();
7078 unsigned BitsPerBool;
7082 } else if (NumElts == 8) {
7085 } else if (NumElts == 16) {
7091 // If this is a single value copied into all lanes (a splat), we can just sign
7092 // extend that single value
7093 SDValue FirstOp = Op.getOperand(0);
7094 if (!isa<ConstantSDNode>(FirstOp) &&
7095 std::all_of(std::next(Op->op_begin()), Op->op_end(),
7096 [&FirstOp](SDUse &U) {
7097 return U.get().isUndef() || U.get() == FirstOp;
7099 SDValue Ext = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::i32, FirstOp,
7100 DAG.getValueType(MVT::i1));
7101 return DAG.getNode(ARMISD::PREDICATE_CAST, dl, Op.getValueType(), Ext);
7104 // First create base with bits set where known
7105 unsigned Bits32 = 0;
7106 for (unsigned i = 0; i < NumElts; ++i) {
7107 SDValue V = Op.getOperand(i);
7108 if (!isa<ConstantSDNode>(V) && !V.isUndef())
7110 bool BitSet = V.isUndef() ? false : cast<ConstantSDNode>(V)->getZExtValue();
7112 Bits32 |= BoolMask << (i * BitsPerBool);
7115 // Add in unknown nodes
7116 SDValue Base = DAG.getNode(ARMISD::PREDICATE_CAST, dl, VT,
7117 DAG.getConstant(Bits32, dl, MVT::i32));
7118 for (unsigned i = 0; i < NumElts; ++i) {
7119 SDValue V = Op.getOperand(i);
7120 if (isa<ConstantSDNode>(V) || V.isUndef())
7122 Base = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Base, V,
7123 DAG.getConstant(i, dl, MVT::i32));
7129 // If this is a case we can't handle, return null and let the default
7130 // expansion code take care of it.
7131 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
7132 const ARMSubtarget *ST) const {
7133 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
7135 EVT VT = Op.getValueType();
7137 if (ST->hasMVEIntegerOps() && VT.getScalarSizeInBits() == 1)
7138 return LowerBUILD_VECTOR_i1(Op, DAG, ST);
7140 APInt SplatBits, SplatUndef;
7141 unsigned SplatBitSize;
7143 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7144 if (SplatUndef.isAllOnesValue())
7145 return DAG.getUNDEF(VT);
7147 if ((ST->hasNEON() && SplatBitSize <= 64) ||
7148 (ST->hasMVEIntegerOps() && SplatBitSize <= 32)) {
7149 // Check if an immediate VMOV works.
7151 SDValue Val = isVMOVModifiedImm(SplatBits.getZExtValue(),
7152 SplatUndef.getZExtValue(), SplatBitSize,
7153 DAG, dl, VmovVT, VT.is128BitVector(),
7156 if (Val.getNode()) {
7157 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
7158 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
7161 // Try an immediate VMVN.
7162 uint64_t NegatedImm = (~SplatBits).getZExtValue();
7163 Val = isVMOVModifiedImm(
7164 NegatedImm, SplatUndef.getZExtValue(), SplatBitSize,
7165 DAG, dl, VmovVT, VT.is128BitVector(),
7166 ST->hasMVEIntegerOps() ? MVEVMVNModImm : VMVNModImm);
7167 if (Val.getNode()) {
7168 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
7169 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
7172 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
7173 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
7174 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
7176 SDValue Val = DAG.getTargetConstant(ImmVal, dl, MVT::i32);
7177 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
7183 // Scan through the operands to see if only one value is used.
7185 // As an optimisation, even if more than one value is used it may be more
7186 // profitable to splat with one value then change some lanes.
7188 // Heuristically we decide to do this if the vector has a "dominant" value,
7189 // defined as splatted to more than half of the lanes.
7190 unsigned NumElts = VT.getVectorNumElements();
7191 bool isOnlyLowElement = true;
7192 bool usesOnlyOneValue = true;
7193 bool hasDominantValue = false;
7194 bool isConstant = true;
7196 // Map of the number of times a particular SDValue appears in the
7198 DenseMap<SDValue, unsigned> ValueCounts;
7200 for (unsigned i = 0; i < NumElts; ++i) {
7201 SDValue V = Op.getOperand(i);
7205 isOnlyLowElement = false;
7206 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
7209 ValueCounts.insert(std::make_pair(V, 0));
7210 unsigned &Count = ValueCounts[V];
7212 // Is this value dominant? (takes up more than half of the lanes)
7213 if (++Count > (NumElts / 2)) {
7214 hasDominantValue = true;
7218 if (ValueCounts.size() != 1)
7219 usesOnlyOneValue = false;
7220 if (!Value.getNode() && !ValueCounts.empty())
7221 Value = ValueCounts.begin()->first;
7223 if (ValueCounts.empty())
7224 return DAG.getUNDEF(VT);
7226 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
7227 // Keep going if we are hitting this case.
7228 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
7229 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
7231 unsigned EltSize = VT.getScalarSizeInBits();
7233 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
7234 // i32 and try again.
7235 if (hasDominantValue && EltSize <= 32) {
7239 // If we are VDUPing a value that comes directly from a vector, that will
7240 // cause an unnecessary move to and from a GPR, where instead we could
7241 // just use VDUPLANE. We can only do this if the lane being extracted
7242 // is at a constant index, as the VDUP from lane instructions only have
7243 // constant-index forms.
7244 ConstantSDNode *constIndex;
7245 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
7246 (constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1)))) {
7247 // We need to create a new undef vector to use for the VDUPLANE if the
7248 // size of the vector from which we get the value is different than the
7249 // size of the vector that we need to create. We will insert the element
7250 // such that the register coalescer will remove unnecessary copies.
7251 if (VT != Value->getOperand(0).getValueType()) {
7252 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
7253 VT.getVectorNumElements();
7254 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
7255 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
7256 Value, DAG.getConstant(index, dl, MVT::i32)),
7257 DAG.getConstant(index, dl, MVT::i32));
7259 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
7260 Value->getOperand(0), Value->getOperand(1));
7262 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
7264 if (!usesOnlyOneValue) {
7265 // The dominant value was splatted as 'N', but we now have to insert
7266 // all differing elements.
7267 for (unsigned I = 0; I < NumElts; ++I) {
7268 if (Op.getOperand(I) == Value)
7270 SmallVector<SDValue, 3> Ops;
7272 Ops.push_back(Op.getOperand(I));
7273 Ops.push_back(DAG.getConstant(I, dl, MVT::i32));
7274 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ops);
7279 if (VT.getVectorElementType().isFloatingPoint()) {
7280 SmallVector<SDValue, 8> Ops;
7281 MVT FVT = VT.getVectorElementType().getSimpleVT();
7282 assert(FVT == MVT::f32 || FVT == MVT::f16);
7283 MVT IVT = (FVT == MVT::f32) ? MVT::i32 : MVT::i16;
7284 for (unsigned i = 0; i < NumElts; ++i)
7285 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, IVT,
7287 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), IVT, NumElts);
7288 SDValue Val = DAG.getBuildVector(VecVT, dl, Ops);
7289 Val = LowerBUILD_VECTOR(Val, DAG, ST);
7291 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
7293 if (usesOnlyOneValue) {
7294 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
7295 if (isConstant && Val.getNode())
7296 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
7300 // If all elements are constants and the case above didn't get hit, fall back
7301 // to the default expansion, which will generate a load from the constant
7306 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
7308 SDValue shuffle = ReconstructShuffle(Op, DAG);
7309 if (shuffle != SDValue())
7313 if (ST->hasNEON() && VT.is128BitVector() && VT != MVT::v2f64 && VT != MVT::v4f32) {
7314 // If we haven't found an efficient lowering, try splitting a 128-bit vector
7315 // into two 64-bit vectors; we might discover a better way to lower it.
7316 SmallVector<SDValue, 64> Ops(Op->op_begin(), Op->op_begin() + NumElts);
7317 EVT ExtVT = VT.getVectorElementType();
7318 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElts / 2);
7320 DAG.getBuildVector(HVT, dl, makeArrayRef(&Ops[0], NumElts / 2));
7321 if (Lower.getOpcode() == ISD::BUILD_VECTOR)
7322 Lower = LowerBUILD_VECTOR(Lower, DAG, ST);
7323 SDValue Upper = DAG.getBuildVector(
7324 HVT, dl, makeArrayRef(&Ops[NumElts / 2], NumElts / 2));
7325 if (Upper.getOpcode() == ISD::BUILD_VECTOR)
7326 Upper = LowerBUILD_VECTOR(Upper, DAG, ST);
7328 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Lower, Upper);
7331 // Vectors with 32- or 64-bit elements can be built by directly assigning
7332 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
7333 // will be legalized.
7334 if (EltSize >= 32) {
7335 // Do the expansion with floating-point types, since that is what the VFP
7336 // registers are defined to use, and since i64 is not legal.
7337 EVT EltVT = EVT::getFloatingPointVT(EltSize);
7338 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
7339 SmallVector<SDValue, 8> Ops;
7340 for (unsigned i = 0; i < NumElts; ++i)
7341 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
7342 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
7343 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
7346 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
7347 // know the default expansion would otherwise fall back on something even
7348 // worse. For a vector with one or two non-undef values, that's
7349 // scalar_to_vector for the elements followed by a shuffle (provided the
7350 // shuffle is valid for the target) and materialization element by element
7351 // on the stack followed by a load for everything else.
7352 if (!isConstant && !usesOnlyOneValue) {
7353 SDValue Vec = DAG.getUNDEF(VT);
7354 for (unsigned i = 0 ; i < NumElts; ++i) {
7355 SDValue V = Op.getOperand(i);
7358 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i32);
7359 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
7367 // Gather data to see if the operation can be modelled as a
7368 // shuffle in combination with VEXTs.
7369 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
7370 SelectionDAG &DAG) const {
7371 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
7373 EVT VT = Op.getValueType();
7374 unsigned NumElts = VT.getVectorNumElements();
7376 struct ShuffleSourceInfo {
7378 unsigned MinElt = std::numeric_limits<unsigned>::max();
7379 unsigned MaxElt = 0;
7381 // We may insert some combination of BITCASTs and VEXT nodes to force Vec to
7382 // be compatible with the shuffle we intend to construct. As a result
7383 // ShuffleVec will be some sliding window into the original Vec.
7386 // Code should guarantee that element i in Vec starts at element "WindowBase
7387 // + i * WindowScale in ShuffleVec".
7389 int WindowScale = 1;
7391 ShuffleSourceInfo(SDValue Vec) : Vec(Vec), ShuffleVec(Vec) {}
7393 bool operator ==(SDValue OtherVec) { return Vec == OtherVec; }
7396 // First gather all vectors used as an immediate source for this BUILD_VECTOR
7398 SmallVector<ShuffleSourceInfo, 2> Sources;
7399 for (unsigned i = 0; i < NumElts; ++i) {
7400 SDValue V = Op.getOperand(i);
7403 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
7404 // A shuffle can only come from building a vector from various
7405 // elements of other vectors.
7407 } else if (!isa<ConstantSDNode>(V.getOperand(1))) {
7408 // Furthermore, shuffles require a constant mask, whereas extractelts
7409 // accept variable indices.
7413 // Add this element source to the list if it's not already there.
7414 SDValue SourceVec = V.getOperand(0);
7415 auto Source = llvm::find(Sources, SourceVec);
7416 if (Source == Sources.end())
7417 Source = Sources.insert(Sources.end(), ShuffleSourceInfo(SourceVec));
7419 // Update the minimum and maximum lane number seen.
7420 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
7421 Source->MinElt = std::min(Source->MinElt, EltNo);
7422 Source->MaxElt = std::max(Source->MaxElt, EltNo);
7425 // Currently only do something sane when at most two source vectors
7427 if (Sources.size() > 2)
7430 // Find out the smallest element size among result and two sources, and use
7431 // it as element size to build the shuffle_vector.
7432 EVT SmallestEltTy = VT.getVectorElementType();
7433 for (auto &Source : Sources) {
7434 EVT SrcEltTy = Source.Vec.getValueType().getVectorElementType();
7435 if (SrcEltTy.bitsLT(SmallestEltTy))
7436 SmallestEltTy = SrcEltTy;
7438 unsigned ResMultiplier =
7439 VT.getScalarSizeInBits() / SmallestEltTy.getSizeInBits();
7440 NumElts = VT.getSizeInBits() / SmallestEltTy.getSizeInBits();
7441 EVT ShuffleVT = EVT::getVectorVT(*DAG.getContext(), SmallestEltTy, NumElts);
7443 // If the source vector is too wide or too narrow, we may nevertheless be able
7444 // to construct a compatible shuffle either by concatenating it with UNDEF or
7445 // extracting a suitable range of elements.
7446 for (auto &Src : Sources) {
7447 EVT SrcVT = Src.ShuffleVec.getValueType();
7449 if (SrcVT.getSizeInBits() == VT.getSizeInBits())
7452 // This stage of the search produces a source with the same element type as
7453 // the original, but with a total width matching the BUILD_VECTOR output.
7454 EVT EltVT = SrcVT.getVectorElementType();
7455 unsigned NumSrcElts = VT.getSizeInBits() / EltVT.getSizeInBits();
7456 EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumSrcElts);
7458 if (SrcVT.getSizeInBits() < VT.getSizeInBits()) {
7459 if (2 * SrcVT.getSizeInBits() != VT.getSizeInBits())
7461 // We can pad out the smaller vector for free, so if it's part of a
7464 DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, Src.ShuffleVec,
7465 DAG.getUNDEF(Src.ShuffleVec.getValueType()));
7469 if (SrcVT.getSizeInBits() != 2 * VT.getSizeInBits())
7472 if (Src.MaxElt - Src.MinElt >= NumSrcElts) {
7473 // Span too large for a VEXT to cope
7477 if (Src.MinElt >= NumSrcElts) {
7478 // The extraction can just take the second half
7480 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
7481 DAG.getConstant(NumSrcElts, dl, MVT::i32));
7482 Src.WindowBase = -NumSrcElts;
7483 } else if (Src.MaxElt < NumSrcElts) {
7484 // The extraction can just take the first half
7486 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
7487 DAG.getConstant(0, dl, MVT::i32));
7489 // An actual VEXT is needed
7491 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
7492 DAG.getConstant(0, dl, MVT::i32));
7494 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
7495 DAG.getConstant(NumSrcElts, dl, MVT::i32));
7497 Src.ShuffleVec = DAG.getNode(ARMISD::VEXT, dl, DestVT, VEXTSrc1,
7499 DAG.getConstant(Src.MinElt, dl, MVT::i32));
7500 Src.WindowBase = -Src.MinElt;
7504 // Another possible incompatibility occurs from the vector element types. We
7505 // can fix this by bitcasting the source vectors to the same type we intend
7507 for (auto &Src : Sources) {
7508 EVT SrcEltTy = Src.ShuffleVec.getValueType().getVectorElementType();
7509 if (SrcEltTy == SmallestEltTy)
7511 assert(ShuffleVT.getVectorElementType() == SmallestEltTy);
7512 Src.ShuffleVec = DAG.getNode(ISD::BITCAST, dl, ShuffleVT, Src.ShuffleVec);
7513 Src.WindowScale = SrcEltTy.getSizeInBits() / SmallestEltTy.getSizeInBits();
7514 Src.WindowBase *= Src.WindowScale;
7517 // Final sanity check before we try to actually produce a shuffle.
7518 LLVM_DEBUG(for (auto Src
7520 assert(Src.ShuffleVec.getValueType() == ShuffleVT););
7522 // The stars all align, our next step is to produce the mask for the shuffle.
7523 SmallVector<int, 8> Mask(ShuffleVT.getVectorNumElements(), -1);
7524 int BitsPerShuffleLane = ShuffleVT.getScalarSizeInBits();
7525 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
7526 SDValue Entry = Op.getOperand(i);
7527 if (Entry.isUndef())
7530 auto Src = llvm::find(Sources, Entry.getOperand(0));
7531 int EltNo = cast<ConstantSDNode>(Entry.getOperand(1))->getSExtValue();
7533 // EXTRACT_VECTOR_ELT performs an implicit any_ext; BUILD_VECTOR an implicit
7534 // trunc. So only std::min(SrcBits, DestBits) actually get defined in this
7536 EVT OrigEltTy = Entry.getOperand(0).getValueType().getVectorElementType();
7537 int BitsDefined = std::min(OrigEltTy.getSizeInBits(),
7538 VT.getScalarSizeInBits());
7539 int LanesDefined = BitsDefined / BitsPerShuffleLane;
7541 // This source is expected to fill ResMultiplier lanes of the final shuffle,
7542 // starting at the appropriate offset.
7543 int *LaneMask = &Mask[i * ResMultiplier];
7545 int ExtractBase = EltNo * Src->WindowScale + Src->WindowBase;
7546 ExtractBase += NumElts * (Src - Sources.begin());
7547 for (int j = 0; j < LanesDefined; ++j)
7548 LaneMask[j] = ExtractBase + j;
7552 // We can't handle more than two sources. This should have already
7553 // been checked before this point.
7554 assert(Sources.size() <= 2 && "Too many sources!");
7556 SDValue ShuffleOps[] = { DAG.getUNDEF(ShuffleVT), DAG.getUNDEF(ShuffleVT) };
7557 for (unsigned i = 0; i < Sources.size(); ++i)
7558 ShuffleOps[i] = Sources[i].ShuffleVec;
7560 SDValue Shuffle = buildLegalVectorShuffle(ShuffleVT, dl, ShuffleOps[0],
7561 ShuffleOps[1], Mask, DAG);
7564 return DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
7567 enum ShuffleOpCodes {
7568 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
7577 OP_VUZPL, // VUZP, left result
7578 OP_VUZPR, // VUZP, right result
7579 OP_VZIPL, // VZIP, left result
7580 OP_VZIPR, // VZIP, right result
7581 OP_VTRNL, // VTRN, left result
7582 OP_VTRNR // VTRN, right result
7585 static bool isLegalMVEShuffleOp(unsigned PFEntry) {
7586 unsigned OpNum = (PFEntry >> 26) & 0x0F;
7599 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7600 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7601 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7602 /// are assumed to be legal.
7603 bool ARMTargetLowering::isShuffleMaskLegal(ArrayRef<int> M, EVT VT) const {
7604 if (VT.getVectorNumElements() == 4 &&
7605 (VT.is128BitVector() || VT.is64BitVector())) {
7606 unsigned PFIndexes[4];
7607 for (unsigned i = 0; i != 4; ++i) {
7611 PFIndexes[i] = M[i];
7614 // Compute the index in the perfect shuffle table.
7615 unsigned PFTableIndex =
7616 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
7617 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
7618 unsigned Cost = (PFEntry >> 30);
7620 if (Cost <= 4 && (Subtarget->hasNEON() || isLegalMVEShuffleOp(PFEntry)))
7624 bool ReverseVEXT, isV_UNDEF;
7625 unsigned Imm, WhichResult;
7627 unsigned EltSize = VT.getScalarSizeInBits();
7628 if (EltSize >= 32 ||
7629 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7630 ShuffleVectorInst::isIdentityMask(M) ||
7631 isVREVMask(M, VT, 64) ||
7632 isVREVMask(M, VT, 32) ||
7633 isVREVMask(M, VT, 16))
7635 else if (Subtarget->hasNEON() &&
7636 (isVEXTMask(M, VT, ReverseVEXT, Imm) ||
7637 isVTBLMask(M, VT) ||
7638 isNEONTwoResultShuffleMask(M, VT, WhichResult, isV_UNDEF)))
7640 else if (Subtarget->hasNEON() && (VT == MVT::v8i16 || VT == MVT::v16i8) &&
7641 isReverseMask(M, VT))
7643 else if (Subtarget->hasMVEIntegerOps() &&
7644 (isVMOVNMask(M, VT, 0) || isVMOVNMask(M, VT, 1)))
7650 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
7651 /// the specified operations to build the shuffle.
7652 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
7653 SDValue RHS, SelectionDAG &DAG,
7655 unsigned OpNum = (PFEntry >> 26) & 0x0F;
7656 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
7657 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
7659 if (OpNum == OP_COPY) {
7660 if (LHSID == (1*9+2)*9+3) return LHS;
7661 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
7665 SDValue OpLHS, OpRHS;
7666 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
7667 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
7668 EVT VT = OpLHS.getValueType();
7671 default: llvm_unreachable("Unknown shuffle opcode!");
7673 // VREV divides the vector in half and swaps within the half.
7674 if (VT.getVectorElementType() == MVT::i32 ||
7675 VT.getVectorElementType() == MVT::f32)
7676 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
7677 // vrev <4 x i16> -> VREV32
7678 if (VT.getVectorElementType() == MVT::i16)
7679 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
7680 // vrev <4 x i8> -> VREV16
7681 assert(VT.getVectorElementType() == MVT::i8);
7682 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
7687 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
7688 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, dl, MVT::i32));
7692 return DAG.getNode(ARMISD::VEXT, dl, VT,
7694 DAG.getConstant(OpNum - OP_VEXT1 + 1, dl, MVT::i32));
7697 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
7698 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
7701 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
7702 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
7705 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
7706 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
7710 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
7711 ArrayRef<int> ShuffleMask,
7712 SelectionDAG &DAG) {
7713 // Check to see if we can use the VTBL instruction.
7714 SDValue V1 = Op.getOperand(0);
7715 SDValue V2 = Op.getOperand(1);
7718 SmallVector<SDValue, 8> VTBLMask;
7719 for (ArrayRef<int>::iterator
7720 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
7721 VTBLMask.push_back(DAG.getConstant(*I, DL, MVT::i32));
7723 if (V2.getNode()->isUndef())
7724 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
7725 DAG.getBuildVector(MVT::v8i8, DL, VTBLMask));
7727 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
7728 DAG.getBuildVector(MVT::v8i8, DL, VTBLMask));
7731 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
7732 SelectionDAG &DAG) {
7734 SDValue OpLHS = Op.getOperand(0);
7735 EVT VT = OpLHS.getValueType();
7737 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
7738 "Expect an v8i16/v16i8 type");
7739 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
7740 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
7741 // extract the first 8 bytes into the top double word and the last 8 bytes
7742 // into the bottom double word. The v8i16 case is similar.
7743 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
7744 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
7745 DAG.getConstant(ExtractNum, DL, MVT::i32));
7748 static EVT getVectorTyFromPredicateVector(EVT VT) {
7749 switch (VT.getSimpleVT().SimpleTy) {
7757 llvm_unreachable("Unexpected vector predicate type");
7761 static SDValue PromoteMVEPredVector(SDLoc dl, SDValue Pred, EVT VT,
7762 SelectionDAG &DAG) {
7763 // Converting from boolean predicates to integers involves creating a vector
7764 // of all ones or all zeroes and selecting the lanes based upon the real
7767 DAG.getTargetConstant(ARM_AM::createVMOVModImm(0xe, 0xff), dl, MVT::i32);
7768 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v16i8, AllOnes);
7771 DAG.getTargetConstant(ARM_AM::createVMOVModImm(0xe, 0x0), dl, MVT::i32);
7772 AllZeroes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v16i8, AllZeroes);
7774 // Get full vector type from predicate type
7775 EVT NewVT = getVectorTyFromPredicateVector(VT);
7778 // If the real predicate is an v8i1 or v4i1 (not v16i1) then we need to recast
7779 // this to a v16i1. This cannot be done with an ordinary bitcast because the
7780 // sizes are not the same. We have to use a MVE specific PREDICATE_CAST node,
7781 // since we know in hardware the sizes are really the same.
7782 if (VT != MVT::v16i1)
7783 RecastV1 = DAG.getNode(ARMISD::PREDICATE_CAST, dl, MVT::v16i1, Pred);
7787 // Select either all ones or zeroes depending upon the real predicate bits.
7788 SDValue PredAsVector =
7789 DAG.getNode(ISD::VSELECT, dl, MVT::v16i8, RecastV1, AllOnes, AllZeroes);
7791 // Recast our new predicate-as-integer v16i8 vector into something
7792 // appropriate for the shuffle, i.e. v4i32 for a real v4i1 predicate.
7793 return DAG.getNode(ISD::BITCAST, dl, NewVT, PredAsVector);
7796 static SDValue LowerVECTOR_SHUFFLE_i1(SDValue Op, SelectionDAG &DAG,
7797 const ARMSubtarget *ST) {
7798 EVT VT = Op.getValueType();
7799 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
7800 ArrayRef<int> ShuffleMask = SVN->getMask();
7802 assert(ST->hasMVEIntegerOps() &&
7803 "No support for vector shuffle of boolean predicates");
7805 SDValue V1 = Op.getOperand(0);
7807 if (isReverseMask(ShuffleMask, VT)) {
7808 SDValue cast = DAG.getNode(ARMISD::PREDICATE_CAST, dl, MVT::i32, V1);
7809 SDValue rbit = DAG.getNode(ISD::BITREVERSE, dl, MVT::i32, cast);
7810 SDValue srl = DAG.getNode(ISD::SRL, dl, MVT::i32, rbit,
7811 DAG.getConstant(16, dl, MVT::i32));
7812 return DAG.getNode(ARMISD::PREDICATE_CAST, dl, VT, srl);
7815 // Until we can come up with optimised cases for every single vector
7816 // shuffle in existence we have chosen the least painful strategy. This is
7817 // to essentially promote the boolean predicate to a 8-bit integer, where
7818 // each predicate represents a byte. Then we fall back on a normal integer
7819 // vector shuffle and convert the result back into a predicate vector. In
7820 // many cases the generated code might be even better than scalar code
7821 // operating on bits. Just imagine trying to shuffle 8 arbitrary 2-bit
7822 // fields in a register into 8 other arbitrary 2-bit fields!
7823 SDValue PredAsVector = PromoteMVEPredVector(dl, V1, VT, DAG);
7824 EVT NewVT = PredAsVector.getValueType();
7827 SDValue Shuffled = DAG.getVectorShuffle(NewVT, dl, PredAsVector,
7828 DAG.getUNDEF(NewVT), ShuffleMask);
7830 // Now return the result of comparing the shuffled vector with zero,
7831 // which will generate a real predicate, i.e. v4i1, v8i1 or v16i1.
7832 return DAG.getNode(ARMISD::VCMPZ, dl, VT, Shuffled,
7833 DAG.getConstant(ARMCC::NE, dl, MVT::i32));
7836 static SDValue LowerVECTOR_SHUFFLEUsingMovs(SDValue Op,
7837 ArrayRef<int> ShuffleMask,
7838 SelectionDAG &DAG) {
7839 // Attempt to lower the vector shuffle using as many whole register movs as
7840 // possible. This is useful for types smaller than 32bits, which would
7841 // often otherwise become a series for grp movs.
7843 EVT VT = Op.getValueType();
7844 if (VT.getScalarSizeInBits() >= 32)
7847 assert((VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v16i8) &&
7848 "Unexpected vector type");
7849 int NumElts = VT.getVectorNumElements();
7850 int QuarterSize = NumElts / 4;
7851 // The four final parts of the vector, as i32's
7854 // Look for full lane vmovs like <0,1,2,3> or <u,5,6,7> etc, (but not
7855 // <u,u,u,u>), returning the vmov lane index
7856 auto getMovIdx = [](ArrayRef<int> ShuffleMask, int Start, int Length) {
7857 // Detect which mov lane this would be from the first non-undef element.
7859 for (int i = 0; i < Length; i++) {
7860 if (ShuffleMask[Start + i] >= 0) {
7861 if (ShuffleMask[Start + i] % Length != i)
7863 MovIdx = ShuffleMask[Start + i] / Length;
7867 // If all items are undef, leave this for other combines
7870 // Check the remaining values are the correct part of the same mov
7871 for (int i = 1; i < Length; i++) {
7872 if (ShuffleMask[Start + i] >= 0 &&
7873 (ShuffleMask[Start + i] / Length != MovIdx ||
7874 ShuffleMask[Start + i] % Length != i))
7880 for (int Part = 0; Part < 4; ++Part) {
7881 // Does this part look like a mov
7882 int Elt = getMovIdx(ShuffleMask, Part * QuarterSize, QuarterSize);
7884 SDValue Input = Op->getOperand(0);
7886 Input = Op->getOperand(1);
7889 SDValue BitCast = DAG.getBitcast(MVT::v4i32, Input);
7890 Parts[Part] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32, BitCast,
7891 DAG.getConstant(Elt, dl, MVT::i32));
7895 // Nothing interesting found, just return
7896 if (!Parts[0] && !Parts[1] && !Parts[2] && !Parts[3])
7899 // The other parts need to be built with the old shuffle vector, cast to a
7900 // v4i32 and extract_vector_elts
7901 if (!Parts[0] || !Parts[1] || !Parts[2] || !Parts[3]) {
7902 SmallVector<int, 16> NewShuffleMask;
7903 for (int Part = 0; Part < 4; ++Part)
7904 for (int i = 0; i < QuarterSize; i++)
7905 NewShuffleMask.push_back(
7906 Parts[Part] ? -1 : ShuffleMask[Part * QuarterSize + i]);
7907 SDValue NewShuffle = DAG.getVectorShuffle(
7908 VT, dl, Op->getOperand(0), Op->getOperand(1), NewShuffleMask);
7909 SDValue BitCast = DAG.getBitcast(MVT::v4i32, NewShuffle);
7911 for (int Part = 0; Part < 4; ++Part)
7913 Parts[Part] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
7914 BitCast, DAG.getConstant(Part, dl, MVT::i32));
7916 // Build a vector out of the various parts and bitcast it back to the original
7918 SDValue NewVec = DAG.getBuildVector(MVT::v4i32, dl, Parts);
7919 return DAG.getBitcast(VT, NewVec);
7922 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG,
7923 const ARMSubtarget *ST) {
7924 SDValue V1 = Op.getOperand(0);
7925 SDValue V2 = Op.getOperand(1);
7927 EVT VT = Op.getValueType();
7928 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
7929 unsigned EltSize = VT.getScalarSizeInBits();
7931 if (ST->hasMVEIntegerOps() && EltSize == 1)
7932 return LowerVECTOR_SHUFFLE_i1(Op, DAG, ST);
7934 // Convert shuffles that are directly supported on NEON to target-specific
7935 // DAG nodes, instead of keeping them as shuffles and matching them again
7936 // during code selection. This is more efficient and avoids the possibility
7937 // of inconsistencies between legalization and selection.
7938 // FIXME: floating-point vectors should be canonicalized to integer vectors
7939 // of the same time so that they get CSEd properly.
7940 ArrayRef<int> ShuffleMask = SVN->getMask();
7942 if (EltSize <= 32) {
7943 if (SVN->isSplat()) {
7944 int Lane = SVN->getSplatIndex();
7945 // If this is undef splat, generate it via "just" vdup, if possible.
7946 if (Lane == -1) Lane = 0;
7948 // Test if V1 is a SCALAR_TO_VECTOR.
7949 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
7950 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
7952 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
7953 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
7955 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
7956 !isa<ConstantSDNode>(V1.getOperand(0))) {
7957 bool IsScalarToVector = true;
7958 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
7959 if (!V1.getOperand(i).isUndef()) {
7960 IsScalarToVector = false;
7963 if (IsScalarToVector)
7964 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
7966 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
7967 DAG.getConstant(Lane, dl, MVT::i32));
7970 bool ReverseVEXT = false;
7972 if (ST->hasNEON() && isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
7975 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
7976 DAG.getConstant(Imm, dl, MVT::i32));
7979 if (isVREVMask(ShuffleMask, VT, 64))
7980 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
7981 if (isVREVMask(ShuffleMask, VT, 32))
7982 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
7983 if (isVREVMask(ShuffleMask, VT, 16))
7984 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
7986 if (ST->hasNEON() && V2->isUndef() && isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
7987 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
7988 DAG.getConstant(Imm, dl, MVT::i32));
7991 // Check for Neon shuffles that modify both input vectors in place.
7992 // If both results are used, i.e., if there are two shuffles with the same
7993 // source operands and with masks corresponding to both results of one of
7994 // these operations, DAG memoization will ensure that a single node is
7995 // used for both shuffles.
7996 unsigned WhichResult = 0;
7997 bool isV_UNDEF = false;
7998 if (ST->hasNEON()) {
7999 if (unsigned ShuffleOpc = isNEONTwoResultShuffleMask(
8000 ShuffleMask, VT, WhichResult, isV_UNDEF)) {
8003 return DAG.getNode(ShuffleOpc, dl, DAG.getVTList(VT, VT), V1, V2)
8004 .getValue(WhichResult);
8007 if (ST->hasMVEIntegerOps()) {
8008 if (isVMOVNMask(ShuffleMask, VT, 0))
8009 return DAG.getNode(ARMISD::VMOVN, dl, VT, V2, V1,
8010 DAG.getConstant(0, dl, MVT::i32));
8011 if (isVMOVNMask(ShuffleMask, VT, 1))
8012 return DAG.getNode(ARMISD::VMOVN, dl, VT, V1, V2,
8013 DAG.getConstant(1, dl, MVT::i32));
8016 // Also check for these shuffles through CONCAT_VECTORS: we canonicalize
8017 // shuffles that produce a result larger than their operands with:
8018 // shuffle(concat(v1, undef), concat(v2, undef))
8020 // shuffle(concat(v1, v2), undef)
8021 // because we can access quad vectors (see PerformVECTOR_SHUFFLECombine).
8023 // This is useful in the general case, but there are special cases where
8024 // native shuffles produce larger results: the two-result ops.
8026 // Look through the concat when lowering them:
8027 // shuffle(concat(v1, v2), undef)
8029 // concat(VZIP(v1, v2):0, :1)
8031 if (ST->hasNEON() && V1->getOpcode() == ISD::CONCAT_VECTORS && V2->isUndef()) {
8032 SDValue SubV1 = V1->getOperand(0);
8033 SDValue SubV2 = V1->getOperand(1);
8034 EVT SubVT = SubV1.getValueType();
8036 // We expect these to have been canonicalized to -1.
8037 assert(llvm::all_of(ShuffleMask, [&](int i) {
8038 return i < (int)VT.getVectorNumElements();
8039 }) && "Unexpected shuffle index into UNDEF operand!");
8041 if (unsigned ShuffleOpc = isNEONTwoResultShuffleMask(
8042 ShuffleMask, SubVT, WhichResult, isV_UNDEF)) {
8045 assert((WhichResult == 0) &&
8046 "In-place shuffle of concat can only have one result!");
8047 SDValue Res = DAG.getNode(ShuffleOpc, dl, DAG.getVTList(SubVT, SubVT),
8049 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Res.getValue(0),
8055 // If the shuffle is not directly supported and it has 4 elements, use
8056 // the PerfectShuffle-generated table to synthesize it from other shuffles.
8057 unsigned NumElts = VT.getVectorNumElements();
8059 unsigned PFIndexes[4];
8060 for (unsigned i = 0; i != 4; ++i) {
8061 if (ShuffleMask[i] < 0)
8064 PFIndexes[i] = ShuffleMask[i];
8067 // Compute the index in the perfect shuffle table.
8068 unsigned PFTableIndex =
8069 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
8070 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
8071 unsigned Cost = (PFEntry >> 30);
8075 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
8076 else if (isLegalMVEShuffleOp(PFEntry)) {
8077 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
8078 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
8079 unsigned PFEntryLHS = PerfectShuffleTable[LHSID];
8080 unsigned PFEntryRHS = PerfectShuffleTable[RHSID];
8081 if (isLegalMVEShuffleOp(PFEntryLHS) && isLegalMVEShuffleOp(PFEntryRHS))
8082 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
8087 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
8088 if (EltSize >= 32) {
8089 // Do the expansion with floating-point types, since that is what the VFP
8090 // registers are defined to use, and since i64 is not legal.
8091 EVT EltVT = EVT::getFloatingPointVT(EltSize);
8092 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
8093 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
8094 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
8095 SmallVector<SDValue, 8> Ops;
8096 for (unsigned i = 0; i < NumElts; ++i) {
8097 if (ShuffleMask[i] < 0)
8098 Ops.push_back(DAG.getUNDEF(EltVT));
8100 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
8101 ShuffleMask[i] < (int)NumElts ? V1 : V2,
8102 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
8105 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
8106 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
8109 if (ST->hasNEON() && (VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
8110 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
8112 if (ST->hasNEON() && VT == MVT::v8i8)
8113 if (SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG))
8116 if (ST->hasMVEIntegerOps())
8117 if (SDValue NewOp = LowerVECTOR_SHUFFLEUsingMovs(Op, ShuffleMask, DAG))
8123 static SDValue LowerINSERT_VECTOR_ELT_i1(SDValue Op, SelectionDAG &DAG,
8124 const ARMSubtarget *ST) {
8125 EVT VecVT = Op.getOperand(0).getValueType();
8128 assert(ST->hasMVEIntegerOps() &&
8129 "LowerINSERT_VECTOR_ELT_i1 called without MVE!");
8132 DAG.getNode(ARMISD::PREDICATE_CAST, dl, MVT::i32, Op->getOperand(0));
8133 unsigned Lane = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
8134 unsigned LaneWidth =
8135 getVectorTyFromPredicateVector(VecVT).getScalarSizeInBits() / 8;
8136 unsigned Mask = ((1 << LaneWidth) - 1) << Lane * LaneWidth;
8137 SDValue Ext = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::i32,
8138 Op.getOperand(1), DAG.getValueType(MVT::i1));
8139 SDValue BFI = DAG.getNode(ARMISD::BFI, dl, MVT::i32, Conv, Ext,
8140 DAG.getConstant(~Mask, dl, MVT::i32));
8141 return DAG.getNode(ARMISD::PREDICATE_CAST, dl, Op.getValueType(), BFI);
8144 SDValue ARMTargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
8145 SelectionDAG &DAG) const {
8146 // INSERT_VECTOR_ELT is legal only for immediate indexes.
8147 SDValue Lane = Op.getOperand(2);
8148 if (!isa<ConstantSDNode>(Lane))
8151 SDValue Elt = Op.getOperand(1);
8152 EVT EltVT = Elt.getValueType();
8154 if (Subtarget->hasMVEIntegerOps() &&
8155 Op.getValueType().getScalarSizeInBits() == 1)
8156 return LowerINSERT_VECTOR_ELT_i1(Op, DAG, Subtarget);
8158 if (getTypeAction(*DAG.getContext(), EltVT) ==
8159 TargetLowering::TypePromoteFloat) {
8160 // INSERT_VECTOR_ELT doesn't want f16 operands promoting to f32,
8161 // but the type system will try to do that if we don't intervene.
8162 // Reinterpret any such vector-element insertion as one with the
8163 // corresponding integer types.
8167 EVT IEltVT = MVT::getIntegerVT(EltVT.getScalarSizeInBits());
8168 assert(getTypeAction(*DAG.getContext(), IEltVT) !=
8169 TargetLowering::TypePromoteFloat);
8171 SDValue VecIn = Op.getOperand(0);
8172 EVT VecVT = VecIn.getValueType();
8173 EVT IVecVT = EVT::getVectorVT(*DAG.getContext(), IEltVT,
8174 VecVT.getVectorNumElements());
8176 SDValue IElt = DAG.getNode(ISD::BITCAST, dl, IEltVT, Elt);
8177 SDValue IVecIn = DAG.getNode(ISD::BITCAST, dl, IVecVT, VecIn);
8178 SDValue IVecOut = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, IVecVT,
8179 IVecIn, IElt, Lane);
8180 return DAG.getNode(ISD::BITCAST, dl, VecVT, IVecOut);
8186 static SDValue LowerEXTRACT_VECTOR_ELT_i1(SDValue Op, SelectionDAG &DAG,
8187 const ARMSubtarget *ST) {
8188 EVT VecVT = Op.getOperand(0).getValueType();
8191 assert(ST->hasMVEIntegerOps() &&
8192 "LowerINSERT_VECTOR_ELT_i1 called without MVE!");
8195 DAG.getNode(ARMISD::PREDICATE_CAST, dl, MVT::i32, Op->getOperand(0));
8196 unsigned Lane = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
8197 unsigned LaneWidth =
8198 getVectorTyFromPredicateVector(VecVT).getScalarSizeInBits() / 8;
8199 SDValue Shift = DAG.getNode(ISD::SRL, dl, MVT::i32, Conv,
8200 DAG.getConstant(Lane * LaneWidth, dl, MVT::i32));
8204 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG,
8205 const ARMSubtarget *ST) {
8206 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
8207 SDValue Lane = Op.getOperand(1);
8208 if (!isa<ConstantSDNode>(Lane))
8211 SDValue Vec = Op.getOperand(0);
8212 EVT VT = Vec.getValueType();
8214 if (ST->hasMVEIntegerOps() && VT.getScalarSizeInBits() == 1)
8215 return LowerEXTRACT_VECTOR_ELT_i1(Op, DAG, ST);
8217 if (Op.getValueType() == MVT::i32 && Vec.getScalarValueSizeInBits() < 32) {
8219 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
8225 static SDValue LowerCONCAT_VECTORS_i1(SDValue Op, SelectionDAG &DAG,
8226 const ARMSubtarget *ST) {
8227 SDValue V1 = Op.getOperand(0);
8228 SDValue V2 = Op.getOperand(1);
8230 EVT VT = Op.getValueType();
8231 EVT Op1VT = V1.getValueType();
8232 EVT Op2VT = V2.getValueType();
8233 unsigned NumElts = VT.getVectorNumElements();
8235 assert(Op1VT == Op2VT && "Operand types don't match!");
8236 assert(VT.getScalarSizeInBits() == 1 &&
8237 "Unexpected custom CONCAT_VECTORS lowering");
8238 assert(ST->hasMVEIntegerOps() &&
8239 "CONCAT_VECTORS lowering only supported for MVE");
8241 SDValue NewV1 = PromoteMVEPredVector(dl, V1, Op1VT, DAG);
8242 SDValue NewV2 = PromoteMVEPredVector(dl, V2, Op2VT, DAG);
8244 // We now have Op1 + Op2 promoted to vectors of integers, where v8i1 gets
8245 // promoted to v8i16, etc.
8247 MVT ElType = getVectorTyFromPredicateVector(VT).getScalarType().getSimpleVT();
8249 // Extract the vector elements from Op1 and Op2 one by one and truncate them
8250 // to be the right size for the destination. For example, if Op1 is v4i1 then
8251 // the promoted vector is v4i32. The result of concatentation gives a v8i1,
8252 // which when promoted is v8i16. That means each i32 element from Op1 needs
8253 // truncating to i16 and inserting in the result.
8254 EVT ConcatVT = MVT::getVectorVT(ElType, NumElts);
8255 SDValue ConVec = DAG.getNode(ISD::UNDEF, dl, ConcatVT);
8256 auto ExractInto = [&DAG, &dl](SDValue NewV, SDValue ConVec, unsigned &j) {
8257 EVT NewVT = NewV.getValueType();
8258 EVT ConcatVT = ConVec.getValueType();
8259 for (unsigned i = 0, e = NewVT.getVectorNumElements(); i < e; i++, j++) {
8260 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32, NewV,
8261 DAG.getIntPtrConstant(i, dl));
8262 ConVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ConcatVT, ConVec, Elt,
8263 DAG.getConstant(j, dl, MVT::i32));
8268 ConVec = ExractInto(NewV1, ConVec, j);
8269 ConVec = ExractInto(NewV2, ConVec, j);
8271 // Now return the result of comparing the subvector with zero,
8272 // which will generate a real predicate, i.e. v4i1, v8i1 or v16i1.
8273 return DAG.getNode(ARMISD::VCMPZ, dl, VT, ConVec,
8274 DAG.getConstant(ARMCC::NE, dl, MVT::i32));
8277 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG,
8278 const ARMSubtarget *ST) {
8279 EVT VT = Op->getValueType(0);
8280 if (ST->hasMVEIntegerOps() && VT.getScalarSizeInBits() == 1)
8281 return LowerCONCAT_VECTORS_i1(Op, DAG, ST);
8283 // The only time a CONCAT_VECTORS operation can have legal types is when
8284 // two 64-bit vectors are concatenated to a 128-bit vector.
8285 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
8286 "unexpected CONCAT_VECTORS");
8288 SDValue Val = DAG.getUNDEF(MVT::v2f64);
8289 SDValue Op0 = Op.getOperand(0);
8290 SDValue Op1 = Op.getOperand(1);
8292 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
8293 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
8294 DAG.getIntPtrConstant(0, dl));
8296 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
8297 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
8298 DAG.getIntPtrConstant(1, dl));
8299 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
8302 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG,
8303 const ARMSubtarget *ST) {
8304 SDValue V1 = Op.getOperand(0);
8305 SDValue V2 = Op.getOperand(1);
8307 EVT VT = Op.getValueType();
8308 EVT Op1VT = V1.getValueType();
8309 unsigned NumElts = VT.getVectorNumElements();
8310 unsigned Index = cast<ConstantSDNode>(V2)->getZExtValue();
8312 assert(VT.getScalarSizeInBits() == 1 &&
8313 "Unexpected custom EXTRACT_SUBVECTOR lowering");
8314 assert(ST->hasMVEIntegerOps() &&
8315 "EXTRACT_SUBVECTOR lowering only supported for MVE");
8317 SDValue NewV1 = PromoteMVEPredVector(dl, V1, Op1VT, DAG);
8319 // We now have Op1 promoted to a vector of integers, where v8i1 gets
8320 // promoted to v8i16, etc.
8322 MVT ElType = getVectorTyFromPredicateVector(VT).getScalarType().getSimpleVT();
8324 EVT SubVT = MVT::getVectorVT(ElType, NumElts);
8325 SDValue SubVec = DAG.getNode(ISD::UNDEF, dl, SubVT);
8326 for (unsigned i = Index, j = 0; i < (Index + NumElts); i++, j++) {
8327 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32, NewV1,
8328 DAG.getIntPtrConstant(i, dl));
8329 SubVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, SubVT, SubVec, Elt,
8330 DAG.getConstant(j, dl, MVT::i32));
8333 // Now return the result of comparing the subvector with zero,
8334 // which will generate a real predicate, i.e. v4i1, v8i1 or v16i1.
8335 return DAG.getNode(ARMISD::VCMPZ, dl, VT, SubVec,
8336 DAG.getConstant(ARMCC::NE, dl, MVT::i32));
8339 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
8340 /// element has been zero/sign-extended, depending on the isSigned parameter,
8341 /// from an integer type half its size.
8342 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
8344 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
8345 EVT VT = N->getValueType(0);
8346 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
8347 SDNode *BVN = N->getOperand(0).getNode();
8348 if (BVN->getValueType(0) != MVT::v4i32 ||
8349 BVN->getOpcode() != ISD::BUILD_VECTOR)
8351 unsigned LoElt = DAG.getDataLayout().isBigEndian() ? 1 : 0;
8352 unsigned HiElt = 1 - LoElt;
8353 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
8354 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
8355 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
8356 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
8357 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
8360 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
8361 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
8364 if (Hi0->isNullValue() && Hi1->isNullValue())
8370 if (N->getOpcode() != ISD::BUILD_VECTOR)
8373 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
8374 SDNode *Elt = N->getOperand(i).getNode();
8375 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
8376 unsigned EltSize = VT.getScalarSizeInBits();
8377 unsigned HalfSize = EltSize / 2;
8379 if (!isIntN(HalfSize, C->getSExtValue()))
8382 if (!isUIntN(HalfSize, C->getZExtValue()))
8393 /// isSignExtended - Check if a node is a vector value that is sign-extended
8394 /// or a constant BUILD_VECTOR with sign-extended elements.
8395 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
8396 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
8398 if (isExtendedBUILD_VECTOR(N, DAG, true))
8403 /// isZeroExtended - Check if a node is a vector value that is zero-extended
8404 /// or a constant BUILD_VECTOR with zero-extended elements.
8405 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
8406 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
8408 if (isExtendedBUILD_VECTOR(N, DAG, false))
8413 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
8414 if (OrigVT.getSizeInBits() >= 64)
8417 assert(OrigVT.isSimple() && "Expecting a simple value type");
8419 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
8420 switch (OrigSimpleTy) {
8421 default: llvm_unreachable("Unexpected Vector Type");
8430 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
8431 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
8432 /// We insert the required extension here to get the vector to fill a D register.
8433 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
8436 unsigned ExtOpcode) {
8437 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
8438 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
8439 // 64-bits we need to insert a new extension so that it will be 64-bits.
8440 assert(ExtTy.is128BitVector() && "Unexpected extension size");
8441 if (OrigTy.getSizeInBits() >= 64)
8444 // Must extend size to at least 64 bits to be used as an operand for VMULL.
8445 EVT NewVT = getExtensionTo64Bits(OrigTy);
8447 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
8450 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
8451 /// does not do any sign/zero extension. If the original vector is less
8452 /// than 64 bits, an appropriate extension will be added after the load to
8453 /// reach a total size of 64 bits. We have to add the extension separately
8454 /// because ARM does not have a sign/zero extending load for vectors.
8455 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
8456 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
8458 // The load already has the right type.
8459 if (ExtendedTy == LD->getMemoryVT())
8460 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
8461 LD->getBasePtr(), LD->getPointerInfo(),
8462 LD->getAlignment(), LD->getMemOperand()->getFlags());
8464 // We need to create a zextload/sextload. We cannot just create a load
8465 // followed by a zext/zext node because LowerMUL is also run during normal
8466 // operation legalization where we can't create illegal types.
8467 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
8468 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
8469 LD->getMemoryVT(), LD->getAlignment(),
8470 LD->getMemOperand()->getFlags());
8473 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
8474 /// extending load, or BUILD_VECTOR with extended elements, return the
8475 /// unextended value. The unextended vector should be 64 bits so that it can
8476 /// be used as an operand to a VMULL instruction. If the original vector size
8477 /// before extension is less than 64 bits we add a an extension to resize
8478 /// the vector to 64 bits.
8479 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
8480 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
8481 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
8482 N->getOperand(0)->getValueType(0),
8486 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8487 assert((ISD::isSEXTLoad(LD) || ISD::isZEXTLoad(LD)) &&
8488 "Expected extending load");
8490 SDValue newLoad = SkipLoadExtensionForVMULL(LD, DAG);
8491 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), newLoad.getValue(1));
8492 unsigned Opcode = ISD::isSEXTLoad(LD) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
8494 DAG.getNode(Opcode, SDLoc(newLoad), LD->getValueType(0), newLoad);
8495 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 0), extLoad);
8500 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
8501 // have been legalized as a BITCAST from v4i32.
8502 if (N->getOpcode() == ISD::BITCAST) {
8503 SDNode *BVN = N->getOperand(0).getNode();
8504 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
8505 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
8506 unsigned LowElt = DAG.getDataLayout().isBigEndian() ? 1 : 0;
8507 return DAG.getBuildVector(
8508 MVT::v2i32, SDLoc(N),
8509 {BVN->getOperand(LowElt), BVN->getOperand(LowElt + 2)});
8511 // Construct a new BUILD_VECTOR with elements truncated to half the size.
8512 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
8513 EVT VT = N->getValueType(0);
8514 unsigned EltSize = VT.getScalarSizeInBits() / 2;
8515 unsigned NumElts = VT.getVectorNumElements();
8516 MVT TruncVT = MVT::getIntegerVT(EltSize);
8517 SmallVector<SDValue, 8> Ops;
8519 for (unsigned i = 0; i != NumElts; ++i) {
8520 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
8521 const APInt &CInt = C->getAPIntValue();
8522 // Element types smaller than 32 bits are not legal, so use i32 elements.
8523 // The values are implicitly truncated so sext vs. zext doesn't matter.
8524 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), dl, MVT::i32));
8526 return DAG.getBuildVector(MVT::getVectorVT(TruncVT, NumElts), dl, Ops);
8529 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
8530 unsigned Opcode = N->getOpcode();
8531 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
8532 SDNode *N0 = N->getOperand(0).getNode();
8533 SDNode *N1 = N->getOperand(1).getNode();
8534 return N0->hasOneUse() && N1->hasOneUse() &&
8535 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
8540 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
8541 unsigned Opcode = N->getOpcode();
8542 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
8543 SDNode *N0 = N->getOperand(0).getNode();
8544 SDNode *N1 = N->getOperand(1).getNode();
8545 return N0->hasOneUse() && N1->hasOneUse() &&
8546 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
8551 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
8552 // Multiplications are only custom-lowered for 128-bit vectors so that
8553 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
8554 EVT VT = Op.getValueType();
8555 assert(VT.is128BitVector() && VT.isInteger() &&
8556 "unexpected type for custom-lowering ISD::MUL");
8557 SDNode *N0 = Op.getOperand(0).getNode();
8558 SDNode *N1 = Op.getOperand(1).getNode();
8559 unsigned NewOpc = 0;
8561 bool isN0SExt = isSignExtended(N0, DAG);
8562 bool isN1SExt = isSignExtended(N1, DAG);
8563 if (isN0SExt && isN1SExt)
8564 NewOpc = ARMISD::VMULLs;
8566 bool isN0ZExt = isZeroExtended(N0, DAG);
8567 bool isN1ZExt = isZeroExtended(N1, DAG);
8568 if (isN0ZExt && isN1ZExt)
8569 NewOpc = ARMISD::VMULLu;
8570 else if (isN1SExt || isN1ZExt) {
8571 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
8572 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
8573 if (isN1SExt && isAddSubSExt(N0, DAG)) {
8574 NewOpc = ARMISD::VMULLs;
8576 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
8577 NewOpc = ARMISD::VMULLu;
8579 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
8581 NewOpc = ARMISD::VMULLu;
8587 if (VT == MVT::v2i64)
8588 // Fall through to expand this. It is not legal.
8591 // Other vector multiplications are legal.
8596 // Legalize to a VMULL instruction.
8599 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
8601 Op0 = SkipExtensionForVMULL(N0, DAG);
8602 assert(Op0.getValueType().is64BitVector() &&
8603 Op1.getValueType().is64BitVector() &&
8604 "unexpected types for extended operands to VMULL");
8605 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
8608 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
8609 // isel lowering to take advantage of no-stall back to back vmul + vmla.
8616 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
8617 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
8618 EVT Op1VT = Op1.getValueType();
8619 return DAG.getNode(N0->getOpcode(), DL, VT,
8620 DAG.getNode(NewOpc, DL, VT,
8621 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
8622 DAG.getNode(NewOpc, DL, VT,
8623 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
8626 static SDValue LowerSDIV_v4i8(SDValue X, SDValue Y, const SDLoc &dl,
8627 SelectionDAG &DAG) {
8628 // TODO: Should this propagate fast-math-flags?
8631 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
8632 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
8633 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
8634 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
8635 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
8636 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
8637 // Get reciprocal estimate.
8638 // float4 recip = vrecpeq_f32(yf);
8639 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
8640 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
8642 // Because char has a smaller range than uchar, we can actually get away
8643 // without any newton steps. This requires that we use a weird bias
8644 // of 0xb000, however (again, this has been exhaustively tested).
8645 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
8646 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
8647 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
8648 Y = DAG.getConstant(0xb000, dl, MVT::v4i32);
8649 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
8650 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
8651 // Convert back to short.
8652 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
8653 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
8657 static SDValue LowerSDIV_v4i16(SDValue N0, SDValue N1, const SDLoc &dl,
8658 SelectionDAG &DAG) {
8659 // TODO: Should this propagate fast-math-flags?
8662 // Convert to float.
8663 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
8664 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
8665 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
8666 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
8667 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
8668 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
8670 // Use reciprocal estimate and one refinement step.
8671 // float4 recip = vrecpeq_f32(yf);
8672 // recip *= vrecpsq_f32(yf, recip);
8673 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
8674 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
8676 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
8677 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
8679 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
8680 // Because short has a smaller range than ushort, we can actually get away
8681 // with only a single newton step. This requires that we use a weird bias
8682 // of 89, however (again, this has been exhaustively tested).
8683 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
8684 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
8685 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
8686 N1 = DAG.getConstant(0x89, dl, MVT::v4i32);
8687 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
8688 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
8689 // Convert back to integer and return.
8690 // return vmovn_s32(vcvt_s32_f32(result));
8691 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
8692 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
8696 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG,
8697 const ARMSubtarget *ST) {
8698 EVT VT = Op.getValueType();
8699 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
8700 "unexpected type for custom-lowering ISD::SDIV");
8703 SDValue N0 = Op.getOperand(0);
8704 SDValue N1 = Op.getOperand(1);
8707 if (VT == MVT::v8i8) {
8708 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
8709 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
8711 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
8712 DAG.getIntPtrConstant(4, dl));
8713 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
8714 DAG.getIntPtrConstant(4, dl));
8715 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
8716 DAG.getIntPtrConstant(0, dl));
8717 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
8718 DAG.getIntPtrConstant(0, dl));
8720 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
8721 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
8723 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
8724 N0 = LowerCONCAT_VECTORS(N0, DAG, ST);
8726 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
8729 return LowerSDIV_v4i16(N0, N1, dl, DAG);
8732 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG,
8733 const ARMSubtarget *ST) {
8734 // TODO: Should this propagate fast-math-flags?
8735 EVT VT = Op.getValueType();
8736 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
8737 "unexpected type for custom-lowering ISD::UDIV");
8740 SDValue N0 = Op.getOperand(0);
8741 SDValue N1 = Op.getOperand(1);
8744 if (VT == MVT::v8i8) {
8745 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
8746 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
8748 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
8749 DAG.getIntPtrConstant(4, dl));
8750 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
8751 DAG.getIntPtrConstant(4, dl));
8752 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
8753 DAG.getIntPtrConstant(0, dl));
8754 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
8755 DAG.getIntPtrConstant(0, dl));
8757 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
8758 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
8760 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
8761 N0 = LowerCONCAT_VECTORS(N0, DAG, ST);
8763 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
8764 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, dl,
8770 // v4i16 sdiv ... Convert to float.
8771 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
8772 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
8773 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
8774 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
8775 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
8776 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
8778 // Use reciprocal estimate and two refinement steps.
8779 // float4 recip = vrecpeq_f32(yf);
8780 // recip *= vrecpsq_f32(yf, recip);
8781 // recip *= vrecpsq_f32(yf, recip);
8782 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
8783 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
8785 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
8786 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
8788 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
8789 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
8790 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
8792 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
8793 // Simply multiplying by the reciprocal estimate can leave us a few ulps
8794 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
8795 // and that it will never cause us to return an answer too large).
8796 // float4 result = as_float4(as_int4(xf*recip) + 2);
8797 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
8798 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
8799 N1 = DAG.getConstant(2, dl, MVT::v4i32);
8800 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
8801 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
8802 // Convert back to integer and return.
8803 // return vmovn_u32(vcvt_s32_f32(result));
8804 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
8805 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
8809 static SDValue LowerADDSUBCARRY(SDValue Op, SelectionDAG &DAG) {
8810 SDNode *N = Op.getNode();
8811 EVT VT = N->getValueType(0);
8812 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
8814 SDValue Carry = Op.getOperand(2);
8819 if (Op.getOpcode() == ISD::ADDCARRY) {
8820 // This converts the boolean value carry into the carry flag.
8821 Carry = ConvertBooleanCarryToCarryFlag(Carry, DAG);
8823 // Do the addition proper using the carry flag we wanted.
8824 Result = DAG.getNode(ARMISD::ADDE, DL, VTs, Op.getOperand(0),
8825 Op.getOperand(1), Carry);
8827 // Now convert the carry flag into a boolean value.
8828 Carry = ConvertCarryFlagToBooleanCarry(Result.getValue(1), VT, DAG);
8830 // ARMISD::SUBE expects a carry not a borrow like ISD::SUBCARRY so we
8831 // have to invert the carry first.
8832 Carry = DAG.getNode(ISD::SUB, DL, MVT::i32,
8833 DAG.getConstant(1, DL, MVT::i32), Carry);
8834 // This converts the boolean value carry into the carry flag.
8835 Carry = ConvertBooleanCarryToCarryFlag(Carry, DAG);
8837 // Do the subtraction proper using the carry flag we wanted.
8838 Result = DAG.getNode(ARMISD::SUBE, DL, VTs, Op.getOperand(0),
8839 Op.getOperand(1), Carry);
8841 // Now convert the carry flag into a boolean value.
8842 Carry = ConvertCarryFlagToBooleanCarry(Result.getValue(1), VT, DAG);
8843 // But the carry returned by ARMISD::SUBE is not a borrow as expected
8844 // by ISD::SUBCARRY, so compute 1 - C.
8845 Carry = DAG.getNode(ISD::SUB, DL, MVT::i32,
8846 DAG.getConstant(1, DL, MVT::i32), Carry);
8849 // Return both values.
8850 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Result, Carry);
8853 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
8854 assert(Subtarget->isTargetDarwin());
8856 // For iOS, we want to call an alternative entry point: __sincos_stret,
8857 // return values are passed via sret.
8859 SDValue Arg = Op.getOperand(0);
8860 EVT ArgVT = Arg.getValueType();
8861 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
8862 auto PtrVT = getPointerTy(DAG.getDataLayout());
8864 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
8865 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8867 // Pair of floats / doubles used to pass the result.
8868 Type *RetTy = StructType::get(ArgTy, ArgTy);
8869 auto &DL = DAG.getDataLayout();
8872 bool ShouldUseSRet = Subtarget->isAPCS_ABI();
8874 if (ShouldUseSRet) {
8875 // Create stack object for sret.
8876 const uint64_t ByteSize = DL.getTypeAllocSize(RetTy);
8877 const unsigned StackAlign = DL.getPrefTypeAlignment(RetTy);
8878 int FrameIdx = MFI.CreateStackObject(ByteSize, StackAlign, false);
8879 SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy(DL));
8883 Entry.Ty = RetTy->getPointerTo();
8884 Entry.IsSExt = false;
8885 Entry.IsZExt = false;
8886 Entry.IsSRet = true;
8887 Args.push_back(Entry);
8888 RetTy = Type::getVoidTy(*DAG.getContext());
8894 Entry.IsSExt = false;
8895 Entry.IsZExt = false;
8896 Args.push_back(Entry);
8899 (ArgVT == MVT::f64) ? RTLIB::SINCOS_STRET_F64 : RTLIB::SINCOS_STRET_F32;
8900 const char *LibcallName = getLibcallName(LC);
8901 CallingConv::ID CC = getLibcallCallingConv(LC);
8902 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy(DL));
8904 TargetLowering::CallLoweringInfo CLI(DAG);
8906 .setChain(DAG.getEntryNode())
8907 .setCallee(CC, RetTy, Callee, std::move(Args))
8908 .setDiscardResult(ShouldUseSRet);
8909 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
8912 return CallResult.first;
8915 DAG.getLoad(ArgVT, dl, CallResult.second, SRet, MachinePointerInfo());
8917 // Address of cos field.
8918 SDValue Add = DAG.getNode(ISD::ADD, dl, PtrVT, SRet,
8919 DAG.getIntPtrConstant(ArgVT.getStoreSize(), dl));
8921 DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add, MachinePointerInfo());
8923 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
8924 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
8925 LoadSin.getValue(0), LoadCos.getValue(0));
8928 SDValue ARMTargetLowering::LowerWindowsDIVLibCall(SDValue Op, SelectionDAG &DAG,
8930 SDValue &Chain) const {
8931 EVT VT = Op.getValueType();
8932 assert((VT == MVT::i32 || VT == MVT::i64) &&
8933 "unexpected type for custom lowering DIV");
8936 const auto &DL = DAG.getDataLayout();
8937 const auto &TLI = DAG.getTargetLoweringInfo();
8939 const char *Name = nullptr;
8941 Name = (VT == MVT::i32) ? "__rt_sdiv" : "__rt_sdiv64";
8943 Name = (VT == MVT::i32) ? "__rt_udiv" : "__rt_udiv64";
8945 SDValue ES = DAG.getExternalSymbol(Name, TLI.getPointerTy(DL));
8947 ARMTargetLowering::ArgListTy Args;
8949 for (auto AI : {1, 0}) {
8951 Arg.Node = Op.getOperand(AI);
8952 Arg.Ty = Arg.Node.getValueType().getTypeForEVT(*DAG.getContext());
8953 Args.push_back(Arg);
8956 CallLoweringInfo CLI(DAG);
8959 .setCallee(CallingConv::ARM_AAPCS_VFP, VT.getTypeForEVT(*DAG.getContext()),
8960 ES, std::move(Args));
8962 return LowerCallTo(CLI).first;
8965 // This is a code size optimisation: return the original SDIV node to
8966 // DAGCombiner when we don't want to expand SDIV into a sequence of
8967 // instructions, and an empty node otherwise which will cause the
8968 // SDIV to be expanded in DAGCombine.
8970 ARMTargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor,
8972 SmallVectorImpl<SDNode *> &Created) const {
8973 // TODO: Support SREM
8974 if (N->getOpcode() != ISD::SDIV)
8977 const auto &ST = static_cast<const ARMSubtarget&>(DAG.getSubtarget());
8978 const bool MinSize = ST.hasMinSize();
8979 const bool HasDivide = ST.isThumb() ? ST.hasDivideInThumbMode()
8980 : ST.hasDivideInARMMode();
8982 // Don't touch vector types; rewriting this may lead to scalarizing
8984 if (N->getOperand(0).getValueType().isVector())
8987 // Bail if MinSize is not set, and also for both ARM and Thumb mode we need
8988 // hwdiv support for this to be really profitable.
8989 if (!(MinSize && HasDivide))
8992 // ARM mode is a bit simpler than Thumb: we can handle large power
8993 // of 2 immediates with 1 mov instruction; no further checks required,
8994 // just return the sdiv node.
8996 return SDValue(N, 0);
8998 // In Thumb mode, immediates larger than 128 need a wide 4-byte MOV,
8999 // and thus lose the code size benefits of a MOVS that requires only 2.
9000 // TargetTransformInfo and 'getIntImmCodeSizeCost' could be helpful here,
9001 // but as it's doing exactly this, it's not worth the trouble to get TTI.
9002 if (Divisor.sgt(128))
9005 return SDValue(N, 0);
9008 SDValue ARMTargetLowering::LowerDIV_Windows(SDValue Op, SelectionDAG &DAG,
9009 bool Signed) const {
9010 assert(Op.getValueType() == MVT::i32 &&
9011 "unexpected type for custom lowering DIV");
9014 SDValue DBZCHK = DAG.getNode(ARMISD::WIN__DBZCHK, dl, MVT::Other,
9015 DAG.getEntryNode(), Op.getOperand(1));
9017 return LowerWindowsDIVLibCall(Op, DAG, Signed, DBZCHK);
9020 static SDValue WinDBZCheckDenominator(SelectionDAG &DAG, SDNode *N, SDValue InChain) {
9022 SDValue Op = N->getOperand(1);
9023 if (N->getValueType(0) == MVT::i32)
9024 return DAG.getNode(ARMISD::WIN__DBZCHK, DL, MVT::Other, InChain, Op);
9025 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, Op,
9026 DAG.getConstant(0, DL, MVT::i32));
9027 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, Op,
9028 DAG.getConstant(1, DL, MVT::i32));
9029 return DAG.getNode(ARMISD::WIN__DBZCHK, DL, MVT::Other, InChain,
9030 DAG.getNode(ISD::OR, DL, MVT::i32, Lo, Hi));
9033 void ARMTargetLowering::ExpandDIV_Windows(
9034 SDValue Op, SelectionDAG &DAG, bool Signed,
9035 SmallVectorImpl<SDValue> &Results) const {
9036 const auto &DL = DAG.getDataLayout();
9037 const auto &TLI = DAG.getTargetLoweringInfo();
9039 assert(Op.getValueType() == MVT::i64 &&
9040 "unexpected type for custom lowering DIV");
9043 SDValue DBZCHK = WinDBZCheckDenominator(DAG, Op.getNode(), DAG.getEntryNode());
9045 SDValue Result = LowerWindowsDIVLibCall(Op, DAG, Signed, DBZCHK);
9047 SDValue Lower = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Result);
9048 SDValue Upper = DAG.getNode(ISD::SRL, dl, MVT::i64, Result,
9049 DAG.getConstant(32, dl, TLI.getPointerTy(DL)));
9050 Upper = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Upper);
9052 Results.push_back(Lower);
9053 Results.push_back(Upper);
9056 static SDValue LowerPredicateLoad(SDValue Op, SelectionDAG &DAG) {
9057 LoadSDNode *LD = cast<LoadSDNode>(Op.getNode());
9058 EVT MemVT = LD->getMemoryVT();
9059 assert((MemVT == MVT::v4i1 || MemVT == MVT::v8i1 || MemVT == MVT::v16i1) &&
9060 "Expected a predicate type!");
9061 assert(MemVT == Op.getValueType());
9062 assert(LD->getExtensionType() == ISD::NON_EXTLOAD &&
9063 "Expected a non-extending load");
9064 assert(LD->isUnindexed() && "Expected a unindexed load");
9066 // The basic MVE VLDR on a v4i1/v8i1 actually loads the entire 16bit
9067 // predicate, with the "v4i1" bits spread out over the 16 bits loaded. We
9068 // need to make sure that 8/4 bits are actually loaded into the correct
9069 // place, which means loading the value and then shuffling the values into
9070 // the bottom bits of the predicate.
9071 // Equally, VLDR for an v16i1 will actually load 32bits (so will be incorrect
9075 SDValue Load = DAG.getExtLoad(
9076 ISD::EXTLOAD, dl, MVT::i32, LD->getChain(), LD->getBasePtr(),
9077 EVT::getIntegerVT(*DAG.getContext(), MemVT.getSizeInBits()),
9078 LD->getMemOperand());
9079 SDValue Pred = DAG.getNode(ARMISD::PREDICATE_CAST, dl, MVT::v16i1, Load);
9080 if (MemVT != MVT::v16i1)
9081 Pred = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MemVT, Pred,
9082 DAG.getConstant(0, dl, MVT::i32));
9083 return DAG.getMergeValues({Pred, Load.getValue(1)}, dl);
9086 static SDValue LowerPredicateStore(SDValue Op, SelectionDAG &DAG) {
9087 StoreSDNode *ST = cast<StoreSDNode>(Op.getNode());
9088 EVT MemVT = ST->getMemoryVT();
9089 assert((MemVT == MVT::v4i1 || MemVT == MVT::v8i1 || MemVT == MVT::v16i1) &&
9090 "Expected a predicate type!");
9091 assert(MemVT == ST->getValue().getValueType());
9092 assert(!ST->isTruncatingStore() && "Expected a non-extending store");
9093 assert(ST->isUnindexed() && "Expected a unindexed store");
9095 // Only store the v4i1 or v8i1 worth of bits, via a buildvector with top bits
9096 // unset and a scalar store.
9098 SDValue Build = ST->getValue();
9099 if (MemVT != MVT::v16i1) {
9100 SmallVector<SDValue, 16> Ops;
9101 for (unsigned I = 0; I < MemVT.getVectorNumElements(); I++)
9102 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32, Build,
9103 DAG.getConstant(I, dl, MVT::i32)));
9104 for (unsigned I = MemVT.getVectorNumElements(); I < 16; I++)
9105 Ops.push_back(DAG.getUNDEF(MVT::i32));
9106 Build = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i1, Ops);
9108 SDValue GRP = DAG.getNode(ARMISD::PREDICATE_CAST, dl, MVT::i32, Build);
9109 return DAG.getTruncStore(
9110 ST->getChain(), dl, GRP, ST->getBasePtr(),
9111 EVT::getIntegerVT(*DAG.getContext(), MemVT.getSizeInBits()),
9112 ST->getMemOperand());
9115 static bool isZeroVector(SDValue N) {
9116 return (ISD::isBuildVectorAllZeros(N.getNode()) ||
9117 (N->getOpcode() == ARMISD::VMOVIMM &&
9118 isNullConstant(N->getOperand(0))));
9121 static SDValue LowerMLOAD(SDValue Op, SelectionDAG &DAG) {
9122 MaskedLoadSDNode *N = cast<MaskedLoadSDNode>(Op.getNode());
9123 MVT VT = Op.getSimpleValueType();
9124 SDValue Mask = N->getMask();
9125 SDValue PassThru = N->getPassThru();
9128 if (isZeroVector(PassThru))
9131 // MVE Masked loads use zero as the passthru value. Here we convert undef to
9132 // zero too, and other values are lowered to a select.
9133 SDValue ZeroVec = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
9134 DAG.getTargetConstant(0, dl, MVT::i32));
9135 SDValue NewLoad = DAG.getMaskedLoad(
9136 VT, dl, N->getChain(), N->getBasePtr(), N->getOffset(), Mask, ZeroVec,
9137 N->getMemoryVT(), N->getMemOperand(), N->getAddressingMode(),
9138 N->getExtensionType(), N->isExpandingLoad());
9139 SDValue Combo = NewLoad;
9140 if (!PassThru.isUndef() &&
9141 (PassThru.getOpcode() != ISD::BITCAST ||
9142 !isZeroVector(PassThru->getOperand(0))))
9143 Combo = DAG.getNode(ISD::VSELECT, dl, VT, Mask, NewLoad, PassThru);
9144 return DAG.getMergeValues({Combo, NewLoad.getValue(1)}, dl);
9147 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
9148 if (isStrongerThanMonotonic(cast<AtomicSDNode>(Op)->getOrdering()))
9149 // Acquire/Release load/store is not legal for targets without a dmb or
9150 // equivalent available.
9153 // Monotonic load/store is legal for all targets.
9157 static void ReplaceREADCYCLECOUNTER(SDNode *N,
9158 SmallVectorImpl<SDValue> &Results,
9160 const ARMSubtarget *Subtarget) {
9162 // Under Power Management extensions, the cycle-count is:
9163 // mrc p15, #0, <Rt>, c9, c13, #0
9164 SDValue Ops[] = { N->getOperand(0), // Chain
9165 DAG.getTargetConstant(Intrinsic::arm_mrc, DL, MVT::i32),
9166 DAG.getTargetConstant(15, DL, MVT::i32),
9167 DAG.getTargetConstant(0, DL, MVT::i32),
9168 DAG.getTargetConstant(9, DL, MVT::i32),
9169 DAG.getTargetConstant(13, DL, MVT::i32),
9170 DAG.getTargetConstant(0, DL, MVT::i32)
9173 SDValue Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
9174 DAG.getVTList(MVT::i32, MVT::Other), Ops);
9175 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Cycles32,
9176 DAG.getConstant(0, DL, MVT::i32)));
9177 Results.push_back(Cycles32.getValue(1));
9180 static SDValue createGPRPairNode(SelectionDAG &DAG, SDValue V) {
9181 SDLoc dl(V.getNode());
9182 SDValue VLo = DAG.getAnyExtOrTrunc(V, dl, MVT::i32);
9183 SDValue VHi = DAG.getAnyExtOrTrunc(
9184 DAG.getNode(ISD::SRL, dl, MVT::i64, V, DAG.getConstant(32, dl, MVT::i32)),
9186 bool isBigEndian = DAG.getDataLayout().isBigEndian();
9188 std::swap (VLo, VHi);
9190 DAG.getTargetConstant(ARM::GPRPairRegClassID, dl, MVT::i32);
9191 SDValue SubReg0 = DAG.getTargetConstant(ARM::gsub_0, dl, MVT::i32);
9192 SDValue SubReg1 = DAG.getTargetConstant(ARM::gsub_1, dl, MVT::i32);
9193 const SDValue Ops[] = { RegClass, VLo, SubReg0, VHi, SubReg1 };
9195 DAG.getMachineNode(TargetOpcode::REG_SEQUENCE, dl, MVT::Untyped, Ops), 0);
9198 static void ReplaceCMP_SWAP_64Results(SDNode *N,
9199 SmallVectorImpl<SDValue> & Results,
9200 SelectionDAG &DAG) {
9201 assert(N->getValueType(0) == MVT::i64 &&
9202 "AtomicCmpSwap on types less than 64 should be legal");
9203 SDValue Ops[] = {N->getOperand(1),
9204 createGPRPairNode(DAG, N->getOperand(2)),
9205 createGPRPairNode(DAG, N->getOperand(3)),
9207 SDNode *CmpSwap = DAG.getMachineNode(
9208 ARM::CMP_SWAP_64, SDLoc(N),
9209 DAG.getVTList(MVT::Untyped, MVT::i32, MVT::Other), Ops);
9211 MachineMemOperand *MemOp = cast<MemSDNode>(N)->getMemOperand();
9212 DAG.setNodeMemRefs(cast<MachineSDNode>(CmpSwap), {MemOp});
9214 bool isBigEndian = DAG.getDataLayout().isBigEndian();
9217 DAG.getTargetExtractSubreg(isBigEndian ? ARM::gsub_1 : ARM::gsub_0,
9218 SDLoc(N), MVT::i32, SDValue(CmpSwap, 0)));
9220 DAG.getTargetExtractSubreg(isBigEndian ? ARM::gsub_0 : ARM::gsub_1,
9221 SDLoc(N), MVT::i32, SDValue(CmpSwap, 0)));
9222 Results.push_back(SDValue(CmpSwap, 2));
9225 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
9226 LLVM_DEBUG(dbgs() << "Lowering node: "; Op.dump());
9227 switch (Op.getOpcode()) {
9228 default: llvm_unreachable("Don't know how to custom lower this!");
9229 case ISD::WRITE_REGISTER: return LowerWRITE_REGISTER(Op, DAG);
9230 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
9231 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
9232 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
9233 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
9234 case ISD::SELECT: return LowerSELECT(Op, DAG);
9235 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
9236 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
9237 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
9238 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
9239 case ISD::VASTART: return LowerVASTART(Op, DAG);
9240 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
9241 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
9242 case ISD::SINT_TO_FP:
9243 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
9244 case ISD::STRICT_FP_TO_SINT:
9245 case ISD::STRICT_FP_TO_UINT:
9246 case ISD::FP_TO_SINT:
9247 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
9248 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
9249 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
9250 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
9251 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
9252 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
9253 case ISD::EH_SJLJ_SETUP_DISPATCH: return LowerEH_SJLJ_SETUP_DISPATCH(Op, DAG);
9254 case ISD::INTRINSIC_VOID: return LowerINTRINSIC_VOID(Op, DAG, Subtarget);
9255 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
9257 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG, Subtarget);
9260 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
9261 case ISD::SREM: return LowerREM(Op.getNode(), DAG);
9262 case ISD::UREM: return LowerREM(Op.getNode(), DAG);
9263 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
9264 case ISD::SRL_PARTS:
9265 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
9267 case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
9268 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
9269 case ISD::SETCC: return LowerVSETCC(Op, DAG, Subtarget);
9270 case ISD::SETCCCARRY: return LowerSETCCCARRY(Op, DAG);
9271 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
9272 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
9273 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG, Subtarget);
9274 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG, Subtarget);
9275 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
9276 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG, Subtarget);
9277 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG, Subtarget);
9278 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
9279 case ISD::MUL: return LowerMUL(Op, DAG);
9281 if (Subtarget->isTargetWindows() && !Op.getValueType().isVector())
9282 return LowerDIV_Windows(Op, DAG, /* Signed */ true);
9283 return LowerSDIV(Op, DAG, Subtarget);
9285 if (Subtarget->isTargetWindows() && !Op.getValueType().isVector())
9286 return LowerDIV_Windows(Op, DAG, /* Signed */ false);
9287 return LowerUDIV(Op, DAG, Subtarget);
9289 case ISD::SUBCARRY: return LowerADDSUBCARRY(Op, DAG);
9292 return LowerSignedALUO(Op, DAG);
9295 return LowerUnsignedALUO(Op, DAG);
9298 return LowerSADDSUBSAT(Op, DAG, Subtarget);
9300 return LowerPredicateLoad(Op, DAG);
9302 return LowerPredicateStore(Op, DAG);
9304 return LowerMLOAD(Op, DAG);
9305 case ISD::ATOMIC_LOAD:
9306 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
9307 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
9309 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
9310 case ISD::DYNAMIC_STACKALLOC:
9311 if (Subtarget->isTargetWindows())
9312 return LowerDYNAMIC_STACKALLOC(Op, DAG);
9313 llvm_unreachable("Don't know how to custom lower this!");
9314 case ISD::STRICT_FP_ROUND:
9315 case ISD::FP_ROUND: return LowerFP_ROUND(Op, DAG);
9316 case ISD::STRICT_FP_EXTEND:
9317 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
9318 case ARMISD::WIN__DBZCHK: return SDValue();
9322 static void ReplaceLongIntrinsic(SDNode *N, SmallVectorImpl<SDValue> &Results,
9323 SelectionDAG &DAG) {
9324 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9326 if (IntNo == Intrinsic::arm_smlald)
9327 Opc = ARMISD::SMLALD;
9328 else if (IntNo == Intrinsic::arm_smlaldx)
9329 Opc = ARMISD::SMLALDX;
9330 else if (IntNo == Intrinsic::arm_smlsld)
9331 Opc = ARMISD::SMLSLD;
9332 else if (IntNo == Intrinsic::arm_smlsldx)
9333 Opc = ARMISD::SMLSLDX;
9338 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
9340 DAG.getConstant(0, dl, MVT::i32));
9341 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
9343 DAG.getConstant(1, dl, MVT::i32));
9345 SDValue LongMul = DAG.getNode(Opc, dl,
9346 DAG.getVTList(MVT::i32, MVT::i32),
9347 N->getOperand(1), N->getOperand(2),
9349 Results.push_back(LongMul.getValue(0));
9350 Results.push_back(LongMul.getValue(1));
9353 /// ReplaceNodeResults - Replace the results of node with an illegal result
9354 /// type with new values built out of custom code.
9355 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
9356 SmallVectorImpl<SDValue> &Results,
9357 SelectionDAG &DAG) const {
9359 switch (N->getOpcode()) {
9361 llvm_unreachable("Don't know how to custom expand this!");
9362 case ISD::READ_REGISTER:
9363 ExpandREAD_REGISTER(N, Results, DAG);
9366 Res = ExpandBITCAST(N, DAG, Subtarget);
9371 Res = Expand64BitShift(N, DAG, Subtarget);
9375 Res = LowerREM(N, DAG);
9379 Res = LowerDivRem(SDValue(N, 0), DAG);
9380 assert(Res.getNumOperands() == 2 && "DivRem needs two values");
9381 Results.push_back(Res.getValue(0));
9382 Results.push_back(Res.getValue(1));
9386 Res = LowerSADDSUBSAT(SDValue(N, 0), DAG, Subtarget);
9388 case ISD::READCYCLECOUNTER:
9389 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
9393 assert(Subtarget->isTargetWindows() && "can only expand DIV on Windows");
9394 return ExpandDIV_Windows(SDValue(N, 0), DAG, N->getOpcode() == ISD::SDIV,
9396 case ISD::ATOMIC_CMP_SWAP:
9397 ReplaceCMP_SWAP_64Results(N, Results, DAG);
9399 case ISD::INTRINSIC_WO_CHAIN:
9400 return ReplaceLongIntrinsic(N, Results, DAG);
9402 lowerABS(N, Results, DAG);
9407 Results.push_back(Res);
9410 //===----------------------------------------------------------------------===//
9411 // ARM Scheduler Hooks
9412 //===----------------------------------------------------------------------===//
9414 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
9415 /// registers the function context.
9416 void ARMTargetLowering::SetupEntryBlockForSjLj(MachineInstr &MI,
9417 MachineBasicBlock *MBB,
9418 MachineBasicBlock *DispatchBB,
9420 assert(!Subtarget->isROPI() && !Subtarget->isRWPI() &&
9421 "ROPI/RWPI not currently supported with SjLj");
9422 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
9423 DebugLoc dl = MI.getDebugLoc();
9424 MachineFunction *MF = MBB->getParent();
9425 MachineRegisterInfo *MRI = &MF->getRegInfo();
9426 MachineConstantPool *MCP = MF->getConstantPool();
9427 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
9428 const Function &F = MF->getFunction();
9430 bool isThumb = Subtarget->isThumb();
9431 bool isThumb2 = Subtarget->isThumb2();
9433 unsigned PCLabelId = AFI->createPICLabelUId();
9434 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
9435 ARMConstantPoolValue *CPV =
9436 ARMConstantPoolMBB::Create(F.getContext(), DispatchBB, PCLabelId, PCAdj);
9437 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
9439 const TargetRegisterClass *TRC = isThumb ? &ARM::tGPRRegClass
9440 : &ARM::GPRRegClass;
9442 // Grab constant pool and fixed stack memory operands.
9443 MachineMemOperand *CPMMO =
9444 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(*MF),
9445 MachineMemOperand::MOLoad, 4, 4);
9447 MachineMemOperand *FIMMOSt =
9448 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(*MF, FI),
9449 MachineMemOperand::MOStore, 4, 4);
9451 // Load the address of the dispatch MBB into the jump buffer.
9453 // Incoming value: jbuf
9454 // ldr.n r5, LCPI1_1
9457 // str r5, [$jbuf, #+4] ; &jbuf[1]
9458 Register NewVReg1 = MRI->createVirtualRegister(TRC);
9459 BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
9460 .addConstantPoolIndex(CPI)
9461 .addMemOperand(CPMMO)
9462 .add(predOps(ARMCC::AL));
9463 // Set the low bit because of thumb mode.
9464 Register NewVReg2 = MRI->createVirtualRegister(TRC);
9465 BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
9466 .addReg(NewVReg1, RegState::Kill)
9468 .add(predOps(ARMCC::AL))
9470 Register NewVReg3 = MRI->createVirtualRegister(TRC);
9471 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
9472 .addReg(NewVReg2, RegState::Kill)
9474 BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
9475 .addReg(NewVReg3, RegState::Kill)
9477 .addImm(36) // &jbuf[1] :: pc
9478 .addMemOperand(FIMMOSt)
9479 .add(predOps(ARMCC::AL));
9480 } else if (isThumb) {
9481 // Incoming value: jbuf
9482 // ldr.n r1, LCPI1_4
9486 // add r2, $jbuf, #+4 ; &jbuf[1]
9488 Register NewVReg1 = MRI->createVirtualRegister(TRC);
9489 BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
9490 .addConstantPoolIndex(CPI)
9491 .addMemOperand(CPMMO)
9492 .add(predOps(ARMCC::AL));
9493 Register NewVReg2 = MRI->createVirtualRegister(TRC);
9494 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
9495 .addReg(NewVReg1, RegState::Kill)
9497 // Set the low bit because of thumb mode.
9498 Register NewVReg3 = MRI->createVirtualRegister(TRC);
9499 BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
9500 .addReg(ARM::CPSR, RegState::Define)
9502 .add(predOps(ARMCC::AL));
9503 Register NewVReg4 = MRI->createVirtualRegister(TRC);
9504 BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
9505 .addReg(ARM::CPSR, RegState::Define)
9506 .addReg(NewVReg2, RegState::Kill)
9507 .addReg(NewVReg3, RegState::Kill)
9508 .add(predOps(ARMCC::AL));
9509 Register NewVReg5 = MRI->createVirtualRegister(TRC);
9510 BuildMI(*MBB, MI, dl, TII->get(ARM::tADDframe), NewVReg5)
9512 .addImm(36); // &jbuf[1] :: pc
9513 BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
9514 .addReg(NewVReg4, RegState::Kill)
9515 .addReg(NewVReg5, RegState::Kill)
9517 .addMemOperand(FIMMOSt)
9518 .add(predOps(ARMCC::AL));
9520 // Incoming value: jbuf
9523 // str r1, [$jbuf, #+4] ; &jbuf[1]
9524 Register NewVReg1 = MRI->createVirtualRegister(TRC);
9525 BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
9526 .addConstantPoolIndex(CPI)
9528 .addMemOperand(CPMMO)
9529 .add(predOps(ARMCC::AL));
9530 Register NewVReg2 = MRI->createVirtualRegister(TRC);
9531 BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
9532 .addReg(NewVReg1, RegState::Kill)
9534 .add(predOps(ARMCC::AL));
9535 BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
9536 .addReg(NewVReg2, RegState::Kill)
9538 .addImm(36) // &jbuf[1] :: pc
9539 .addMemOperand(FIMMOSt)
9540 .add(predOps(ARMCC::AL));
9544 void ARMTargetLowering::EmitSjLjDispatchBlock(MachineInstr &MI,
9545 MachineBasicBlock *MBB) const {
9546 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
9547 DebugLoc dl = MI.getDebugLoc();
9548 MachineFunction *MF = MBB->getParent();
9549 MachineRegisterInfo *MRI = &MF->getRegInfo();
9550 MachineFrameInfo &MFI = MF->getFrameInfo();
9551 int FI = MFI.getFunctionContextIndex();
9553 const TargetRegisterClass *TRC = Subtarget->isThumb() ? &ARM::tGPRRegClass
9554 : &ARM::GPRnopcRegClass;
9556 // Get a mapping of the call site numbers to all of the landing pads they're
9558 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2>> CallSiteNumToLPad;
9559 unsigned MaxCSNum = 0;
9560 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
9562 if (!BB->isEHPad()) continue;
9564 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
9566 for (MachineBasicBlock::iterator
9567 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
9568 if (!II->isEHLabel()) continue;
9570 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
9571 if (!MF->hasCallSiteLandingPad(Sym)) continue;
9573 SmallVectorImpl<unsigned> &CallSiteIdxs = MF->getCallSiteLandingPad(Sym);
9574 for (SmallVectorImpl<unsigned>::iterator
9575 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
9576 CSI != CSE; ++CSI) {
9577 CallSiteNumToLPad[*CSI].push_back(&*BB);
9578 MaxCSNum = std::max(MaxCSNum, *CSI);
9584 // Get an ordered list of the machine basic blocks for the jump table.
9585 std::vector<MachineBasicBlock*> LPadList;
9586 SmallPtrSet<MachineBasicBlock*, 32> InvokeBBs;
9587 LPadList.reserve(CallSiteNumToLPad.size());
9588 for (unsigned I = 1; I <= MaxCSNum; ++I) {
9589 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
9590 for (SmallVectorImpl<MachineBasicBlock*>::iterator
9591 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
9592 LPadList.push_back(*II);
9593 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
9597 assert(!LPadList.empty() &&
9598 "No landing pad destinations for the dispatch jump table!");
9600 // Create the jump table and associated information.
9601 MachineJumpTableInfo *JTI =
9602 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
9603 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
9605 // Create the MBBs for the dispatch code.
9607 // Shove the dispatch's address into the return slot in the function context.
9608 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
9609 DispatchBB->setIsEHPad();
9611 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
9612 unsigned trap_opcode;
9613 if (Subtarget->isThumb())
9614 trap_opcode = ARM::tTRAP;
9616 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
9618 BuildMI(TrapBB, dl, TII->get(trap_opcode));
9619 DispatchBB->addSuccessor(TrapBB);
9621 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
9622 DispatchBB->addSuccessor(DispContBB);
9625 MF->insert(MF->end(), DispatchBB);
9626 MF->insert(MF->end(), DispContBB);
9627 MF->insert(MF->end(), TrapBB);
9629 // Insert code into the entry block that creates and registers the function
9631 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
9633 MachineMemOperand *FIMMOLd = MF->getMachineMemOperand(
9634 MachinePointerInfo::getFixedStack(*MF, FI),
9635 MachineMemOperand::MOLoad | MachineMemOperand::MOVolatile, 4, 4);
9637 MachineInstrBuilder MIB;
9638 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
9640 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
9641 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
9643 // Add a register mask with no preserved registers. This results in all
9644 // registers being marked as clobbered. This can't work if the dispatch block
9645 // is in a Thumb1 function and is linked with ARM code which uses the FP
9646 // registers, as there is no way to preserve the FP registers in Thumb1 mode.
9647 MIB.addRegMask(RI.getSjLjDispatchPreservedMask(*MF));
9649 bool IsPositionIndependent = isPositionIndependent();
9650 unsigned NumLPads = LPadList.size();
9651 if (Subtarget->isThumb2()) {
9652 Register NewVReg1 = MRI->createVirtualRegister(TRC);
9653 BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
9656 .addMemOperand(FIMMOLd)
9657 .add(predOps(ARMCC::AL));
9659 if (NumLPads < 256) {
9660 BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
9662 .addImm(LPadList.size())
9663 .add(predOps(ARMCC::AL));
9665 Register VReg1 = MRI->createVirtualRegister(TRC);
9666 BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
9667 .addImm(NumLPads & 0xFFFF)
9668 .add(predOps(ARMCC::AL));
9670 unsigned VReg2 = VReg1;
9671 if ((NumLPads & 0xFFFF0000) != 0) {
9672 VReg2 = MRI->createVirtualRegister(TRC);
9673 BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
9675 .addImm(NumLPads >> 16)
9676 .add(predOps(ARMCC::AL));
9679 BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
9682 .add(predOps(ARMCC::AL));
9685 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
9690 Register NewVReg3 = MRI->createVirtualRegister(TRC);
9691 BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT), NewVReg3)
9692 .addJumpTableIndex(MJTI)
9693 .add(predOps(ARMCC::AL));
9695 Register NewVReg4 = MRI->createVirtualRegister(TRC);
9696 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
9697 .addReg(NewVReg3, RegState::Kill)
9699 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))
9700 .add(predOps(ARMCC::AL))
9703 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
9704 .addReg(NewVReg4, RegState::Kill)
9706 .addJumpTableIndex(MJTI);
9707 } else if (Subtarget->isThumb()) {
9708 Register NewVReg1 = MRI->createVirtualRegister(TRC);
9709 BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
9712 .addMemOperand(FIMMOLd)
9713 .add(predOps(ARMCC::AL));
9715 if (NumLPads < 256) {
9716 BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
9719 .add(predOps(ARMCC::AL));
9721 MachineConstantPool *ConstantPool = MF->getConstantPool();
9722 Type *Int32Ty = Type::getInt32Ty(MF->getFunction().getContext());
9723 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
9725 // MachineConstantPool wants an explicit alignment.
9726 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
9728 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
9729 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
9731 Register VReg1 = MRI->createVirtualRegister(TRC);
9732 BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
9733 .addReg(VReg1, RegState::Define)
9734 .addConstantPoolIndex(Idx)
9735 .add(predOps(ARMCC::AL));
9736 BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
9739 .add(predOps(ARMCC::AL));
9742 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
9747 Register NewVReg2 = MRI->createVirtualRegister(TRC);
9748 BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
9749 .addReg(ARM::CPSR, RegState::Define)
9752 .add(predOps(ARMCC::AL));
9754 Register NewVReg3 = MRI->createVirtualRegister(TRC);
9755 BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
9756 .addJumpTableIndex(MJTI)
9757 .add(predOps(ARMCC::AL));
9759 Register NewVReg4 = MRI->createVirtualRegister(TRC);
9760 BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
9761 .addReg(ARM::CPSR, RegState::Define)
9762 .addReg(NewVReg2, RegState::Kill)
9764 .add(predOps(ARMCC::AL));
9766 MachineMemOperand *JTMMOLd = MF->getMachineMemOperand(
9767 MachinePointerInfo::getJumpTable(*MF), MachineMemOperand::MOLoad, 4, 4);
9769 Register NewVReg5 = MRI->createVirtualRegister(TRC);
9770 BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
9771 .addReg(NewVReg4, RegState::Kill)
9773 .addMemOperand(JTMMOLd)
9774 .add(predOps(ARMCC::AL));
9776 unsigned NewVReg6 = NewVReg5;
9777 if (IsPositionIndependent) {
9778 NewVReg6 = MRI->createVirtualRegister(TRC);
9779 BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
9780 .addReg(ARM::CPSR, RegState::Define)
9781 .addReg(NewVReg5, RegState::Kill)
9783 .add(predOps(ARMCC::AL));
9786 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
9787 .addReg(NewVReg6, RegState::Kill)
9788 .addJumpTableIndex(MJTI);
9790 Register NewVReg1 = MRI->createVirtualRegister(TRC);
9791 BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
9794 .addMemOperand(FIMMOLd)
9795 .add(predOps(ARMCC::AL));
9797 if (NumLPads < 256) {
9798 BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
9801 .add(predOps(ARMCC::AL));
9802 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
9803 Register VReg1 = MRI->createVirtualRegister(TRC);
9804 BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
9805 .addImm(NumLPads & 0xFFFF)
9806 .add(predOps(ARMCC::AL));
9808 unsigned VReg2 = VReg1;
9809 if ((NumLPads & 0xFFFF0000) != 0) {
9810 VReg2 = MRI->createVirtualRegister(TRC);
9811 BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
9813 .addImm(NumLPads >> 16)
9814 .add(predOps(ARMCC::AL));
9817 BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
9820 .add(predOps(ARMCC::AL));
9822 MachineConstantPool *ConstantPool = MF->getConstantPool();
9823 Type *Int32Ty = Type::getInt32Ty(MF->getFunction().getContext());
9824 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
9826 // MachineConstantPool wants an explicit alignment.
9827 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
9829 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
9830 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
9832 Register VReg1 = MRI->createVirtualRegister(TRC);
9833 BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
9834 .addReg(VReg1, RegState::Define)
9835 .addConstantPoolIndex(Idx)
9837 .add(predOps(ARMCC::AL));
9838 BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
9840 .addReg(VReg1, RegState::Kill)
9841 .add(predOps(ARMCC::AL));
9844 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
9849 Register NewVReg3 = MRI->createVirtualRegister(TRC);
9850 BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
9852 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))
9853 .add(predOps(ARMCC::AL))
9855 Register NewVReg4 = MRI->createVirtualRegister(TRC);
9856 BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
9857 .addJumpTableIndex(MJTI)
9858 .add(predOps(ARMCC::AL));
9860 MachineMemOperand *JTMMOLd = MF->getMachineMemOperand(
9861 MachinePointerInfo::getJumpTable(*MF), MachineMemOperand::MOLoad, 4, 4);
9862 Register NewVReg5 = MRI->createVirtualRegister(TRC);
9863 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
9864 .addReg(NewVReg3, RegState::Kill)
9867 .addMemOperand(JTMMOLd)
9868 .add(predOps(ARMCC::AL));
9870 if (IsPositionIndependent) {
9871 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
9872 .addReg(NewVReg5, RegState::Kill)
9874 .addJumpTableIndex(MJTI);
9876 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
9877 .addReg(NewVReg5, RegState::Kill)
9878 .addJumpTableIndex(MJTI);
9882 // Add the jump table entries as successors to the MBB.
9883 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
9884 for (std::vector<MachineBasicBlock*>::iterator
9885 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
9886 MachineBasicBlock *CurMBB = *I;
9887 if (SeenMBBs.insert(CurMBB).second)
9888 DispContBB->addSuccessor(CurMBB);
9891 // N.B. the order the invoke BBs are processed in doesn't matter here.
9892 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
9893 SmallVector<MachineBasicBlock*, 64> MBBLPads;
9894 for (MachineBasicBlock *BB : InvokeBBs) {
9896 // Remove the landing pad successor from the invoke block and replace it
9897 // with the new dispatch block.
9898 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
9900 while (!Successors.empty()) {
9901 MachineBasicBlock *SMBB = Successors.pop_back_val();
9902 if (SMBB->isEHPad()) {
9903 BB->removeSuccessor(SMBB);
9904 MBBLPads.push_back(SMBB);
9908 BB->addSuccessor(DispatchBB, BranchProbability::getZero());
9909 BB->normalizeSuccProbs();
9911 // Find the invoke call and mark all of the callee-saved registers as
9912 // 'implicit defined' so that they're spilled. This prevents code from
9913 // moving instructions to before the EH block, where they will never be
9915 for (MachineBasicBlock::reverse_iterator
9916 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
9917 if (!II->isCall()) continue;
9919 DenseMap<unsigned, bool> DefRegs;
9920 for (MachineInstr::mop_iterator
9921 OI = II->operands_begin(), OE = II->operands_end();
9923 if (!OI->isReg()) continue;
9924 DefRegs[OI->getReg()] = true;
9927 MachineInstrBuilder MIB(*MF, &*II);
9929 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
9930 unsigned Reg = SavedRegs[i];
9931 if (Subtarget->isThumb2() &&
9932 !ARM::tGPRRegClass.contains(Reg) &&
9933 !ARM::hGPRRegClass.contains(Reg))
9935 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
9937 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
9940 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
9947 // Mark all former landing pads as non-landing pads. The dispatch is the only
9949 for (SmallVectorImpl<MachineBasicBlock*>::iterator
9950 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
9951 (*I)->setIsEHPad(false);
9953 // The instruction is gone now.
9954 MI.eraseFromParent();
9958 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
9959 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
9960 E = MBB->succ_end(); I != E; ++I)
9963 llvm_unreachable("Expecting a BB with two successors!");
9966 /// Return the load opcode for a given load size. If load size >= 8,
9967 /// neon opcode will be returned.
9968 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
9970 return LdSize == 16 ? ARM::VLD1q32wb_fixed
9971 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
9973 return LdSize == 4 ? ARM::tLDRi
9974 : LdSize == 2 ? ARM::tLDRHi
9975 : LdSize == 1 ? ARM::tLDRBi : 0;
9977 return LdSize == 4 ? ARM::t2LDR_POST
9978 : LdSize == 2 ? ARM::t2LDRH_POST
9979 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
9980 return LdSize == 4 ? ARM::LDR_POST_IMM
9981 : LdSize == 2 ? ARM::LDRH_POST
9982 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
9985 /// Return the store opcode for a given store size. If store size >= 8,
9986 /// neon opcode will be returned.
9987 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
9989 return StSize == 16 ? ARM::VST1q32wb_fixed
9990 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
9992 return StSize == 4 ? ARM::tSTRi
9993 : StSize == 2 ? ARM::tSTRHi
9994 : StSize == 1 ? ARM::tSTRBi : 0;
9996 return StSize == 4 ? ARM::t2STR_POST
9997 : StSize == 2 ? ARM::t2STRH_POST
9998 : StSize == 1 ? ARM::t2STRB_POST : 0;
9999 return StSize == 4 ? ARM::STR_POST_IMM
10000 : StSize == 2 ? ARM::STRH_POST
10001 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
10004 /// Emit a post-increment load operation with given size. The instructions
10005 /// will be added to BB at Pos.
10006 static void emitPostLd(MachineBasicBlock *BB, MachineBasicBlock::iterator Pos,
10007 const TargetInstrInfo *TII, const DebugLoc &dl,
10008 unsigned LdSize, unsigned Data, unsigned AddrIn,
10009 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
10010 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
10011 assert(LdOpc != 0 && "Should have a load opcode");
10013 BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
10014 .addReg(AddrOut, RegState::Define)
10017 .add(predOps(ARMCC::AL));
10018 } else if (IsThumb1) {
10019 // load + update AddrIn
10020 BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
10023 .add(predOps(ARMCC::AL));
10024 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut)
10025 .add(t1CondCodeOp())
10028 .add(predOps(ARMCC::AL));
10029 } else if (IsThumb2) {
10030 BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
10031 .addReg(AddrOut, RegState::Define)
10034 .add(predOps(ARMCC::AL));
10036 BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
10037 .addReg(AddrOut, RegState::Define)
10041 .add(predOps(ARMCC::AL));
10045 /// Emit a post-increment store operation with given size. The instructions
10046 /// will be added to BB at Pos.
10047 static void emitPostSt(MachineBasicBlock *BB, MachineBasicBlock::iterator Pos,
10048 const TargetInstrInfo *TII, const DebugLoc &dl,
10049 unsigned StSize, unsigned Data, unsigned AddrIn,
10050 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
10051 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
10052 assert(StOpc != 0 && "Should have a store opcode");
10054 BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
10058 .add(predOps(ARMCC::AL));
10059 } else if (IsThumb1) {
10060 // store + update AddrIn
10061 BuildMI(*BB, Pos, dl, TII->get(StOpc))
10065 .add(predOps(ARMCC::AL));
10066 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut)
10067 .add(t1CondCodeOp())
10070 .add(predOps(ARMCC::AL));
10071 } else if (IsThumb2) {
10072 BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
10076 .add(predOps(ARMCC::AL));
10078 BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
10083 .add(predOps(ARMCC::AL));
10087 MachineBasicBlock *
10088 ARMTargetLowering::EmitStructByval(MachineInstr &MI,
10089 MachineBasicBlock *BB) const {
10090 // This pseudo instruction has 3 operands: dst, src, size
10091 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
10092 // Otherwise, we will generate unrolled scalar copies.
10093 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
10094 const BasicBlock *LLVM_BB = BB->getBasicBlock();
10095 MachineFunction::iterator It = ++BB->getIterator();
10097 Register dest = MI.getOperand(0).getReg();
10098 Register src = MI.getOperand(1).getReg();
10099 unsigned SizeVal = MI.getOperand(2).getImm();
10100 unsigned Align = MI.getOperand(3).getImm();
10101 DebugLoc dl = MI.getDebugLoc();
10103 MachineFunction *MF = BB->getParent();
10104 MachineRegisterInfo &MRI = MF->getRegInfo();
10105 unsigned UnitSize = 0;
10106 const TargetRegisterClass *TRC = nullptr;
10107 const TargetRegisterClass *VecTRC = nullptr;
10109 bool IsThumb1 = Subtarget->isThumb1Only();
10110 bool IsThumb2 = Subtarget->isThumb2();
10111 bool IsThumb = Subtarget->isThumb();
10115 } else if (Align & 2) {
10118 // Check whether we can use NEON instructions.
10119 if (!MF->getFunction().hasFnAttribute(Attribute::NoImplicitFloat) &&
10120 Subtarget->hasNEON()) {
10121 if ((Align % 16 == 0) && SizeVal >= 16)
10123 else if ((Align % 8 == 0) && SizeVal >= 8)
10126 // Can't use NEON instructions.
10131 // Select the correct opcode and register class for unit size load/store
10132 bool IsNeon = UnitSize >= 8;
10133 TRC = IsThumb ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
10135 VecTRC = UnitSize == 16 ? &ARM::DPairRegClass
10136 : UnitSize == 8 ? &ARM::DPRRegClass
10139 unsigned BytesLeft = SizeVal % UnitSize;
10140 unsigned LoopSize = SizeVal - BytesLeft;
10142 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
10143 // Use LDR and STR to copy.
10144 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
10145 // [destOut] = STR_POST(scratch, destIn, UnitSize)
10146 unsigned srcIn = src;
10147 unsigned destIn = dest;
10148 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
10149 Register srcOut = MRI.createVirtualRegister(TRC);
10150 Register destOut = MRI.createVirtualRegister(TRC);
10151 Register scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
10152 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
10153 IsThumb1, IsThumb2);
10154 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
10155 IsThumb1, IsThumb2);
10160 // Handle the leftover bytes with LDRB and STRB.
10161 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
10162 // [destOut] = STRB_POST(scratch, destIn, 1)
10163 for (unsigned i = 0; i < BytesLeft; i++) {
10164 Register srcOut = MRI.createVirtualRegister(TRC);
10165 Register destOut = MRI.createVirtualRegister(TRC);
10166 Register scratch = MRI.createVirtualRegister(TRC);
10167 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
10168 IsThumb1, IsThumb2);
10169 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
10170 IsThumb1, IsThumb2);
10174 MI.eraseFromParent(); // The instruction is gone now.
10178 // Expand the pseudo op to a loop.
10181 // movw varEnd, # --> with thumb2
10183 // ldrcp varEnd, idx --> without thumb2
10184 // fallthrough --> loopMBB
10186 // PHI varPhi, varEnd, varLoop
10187 // PHI srcPhi, src, srcLoop
10188 // PHI destPhi, dst, destLoop
10189 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
10190 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
10191 // subs varLoop, varPhi, #UnitSize
10193 // fallthrough --> exitMBB
10195 // epilogue to handle left-over bytes
10196 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
10197 // [destOut] = STRB_POST(scratch, destLoop, 1)
10198 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10199 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10200 MF->insert(It, loopMBB);
10201 MF->insert(It, exitMBB);
10203 // Transfer the remainder of BB and its successor edges to exitMBB.
10204 exitMBB->splice(exitMBB->begin(), BB,
10205 std::next(MachineBasicBlock::iterator(MI)), BB->end());
10206 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
10208 // Load an immediate to varEnd.
10209 Register varEnd = MRI.createVirtualRegister(TRC);
10210 if (Subtarget->useMovt()) {
10211 unsigned Vtmp = varEnd;
10212 if ((LoopSize & 0xFFFF0000) != 0)
10213 Vtmp = MRI.createVirtualRegister(TRC);
10214 BuildMI(BB, dl, TII->get(IsThumb ? ARM::t2MOVi16 : ARM::MOVi16), Vtmp)
10215 .addImm(LoopSize & 0xFFFF)
10216 .add(predOps(ARMCC::AL));
10218 if ((LoopSize & 0xFFFF0000) != 0)
10219 BuildMI(BB, dl, TII->get(IsThumb ? ARM::t2MOVTi16 : ARM::MOVTi16), varEnd)
10221 .addImm(LoopSize >> 16)
10222 .add(predOps(ARMCC::AL));
10224 MachineConstantPool *ConstantPool = MF->getConstantPool();
10225 Type *Int32Ty = Type::getInt32Ty(MF->getFunction().getContext());
10226 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
10228 // MachineConstantPool wants an explicit alignment.
10229 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
10231 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
10232 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
10233 MachineMemOperand *CPMMO =
10234 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(*MF),
10235 MachineMemOperand::MOLoad, 4, 4);
10238 BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci))
10239 .addReg(varEnd, RegState::Define)
10240 .addConstantPoolIndex(Idx)
10241 .add(predOps(ARMCC::AL))
10242 .addMemOperand(CPMMO);
10244 BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp))
10245 .addReg(varEnd, RegState::Define)
10246 .addConstantPoolIndex(Idx)
10248 .add(predOps(ARMCC::AL))
10249 .addMemOperand(CPMMO);
10251 BB->addSuccessor(loopMBB);
10253 // Generate the loop body:
10254 // varPhi = PHI(varLoop, varEnd)
10255 // srcPhi = PHI(srcLoop, src)
10256 // destPhi = PHI(destLoop, dst)
10257 MachineBasicBlock *entryBB = BB;
10259 Register varLoop = MRI.createVirtualRegister(TRC);
10260 Register varPhi = MRI.createVirtualRegister(TRC);
10261 Register srcLoop = MRI.createVirtualRegister(TRC);
10262 Register srcPhi = MRI.createVirtualRegister(TRC);
10263 Register destLoop = MRI.createVirtualRegister(TRC);
10264 Register destPhi = MRI.createVirtualRegister(TRC);
10266 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
10267 .addReg(varLoop).addMBB(loopMBB)
10268 .addReg(varEnd).addMBB(entryBB);
10269 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
10270 .addReg(srcLoop).addMBB(loopMBB)
10271 .addReg(src).addMBB(entryBB);
10272 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
10273 .addReg(destLoop).addMBB(loopMBB)
10274 .addReg(dest).addMBB(entryBB);
10276 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
10277 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
10278 Register scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
10279 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
10280 IsThumb1, IsThumb2);
10281 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
10282 IsThumb1, IsThumb2);
10284 // Decrement loop variable by UnitSize.
10286 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop)
10287 .add(t1CondCodeOp())
10290 .add(predOps(ARMCC::AL));
10292 MachineInstrBuilder MIB =
10293 BuildMI(*BB, BB->end(), dl,
10294 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
10297 .add(predOps(ARMCC::AL))
10298 .add(condCodeOp());
10299 MIB->getOperand(5).setReg(ARM::CPSR);
10300 MIB->getOperand(5).setIsDef(true);
10302 BuildMI(*BB, BB->end(), dl,
10303 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
10304 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
10306 // loopMBB can loop back to loopMBB or fall through to exitMBB.
10307 BB->addSuccessor(loopMBB);
10308 BB->addSuccessor(exitMBB);
10310 // Add epilogue to handle BytesLeft.
10312 auto StartOfExit = exitMBB->begin();
10314 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
10315 // [destOut] = STRB_POST(scratch, destLoop, 1)
10316 unsigned srcIn = srcLoop;
10317 unsigned destIn = destLoop;
10318 for (unsigned i = 0; i < BytesLeft; i++) {
10319 Register srcOut = MRI.createVirtualRegister(TRC);
10320 Register destOut = MRI.createVirtualRegister(TRC);
10321 Register scratch = MRI.createVirtualRegister(TRC);
10322 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
10323 IsThumb1, IsThumb2);
10324 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
10325 IsThumb1, IsThumb2);
10330 MI.eraseFromParent(); // The instruction is gone now.
10334 MachineBasicBlock *
10335 ARMTargetLowering::EmitLowered__chkstk(MachineInstr &MI,
10336 MachineBasicBlock *MBB) const {
10337 const TargetMachine &TM = getTargetMachine();
10338 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
10339 DebugLoc DL = MI.getDebugLoc();
10341 assert(Subtarget->isTargetWindows() &&
10342 "__chkstk is only supported on Windows");
10343 assert(Subtarget->isThumb2() && "Windows on ARM requires Thumb-2 mode");
10345 // __chkstk takes the number of words to allocate on the stack in R4, and
10346 // returns the stack adjustment in number of bytes in R4. This will not
10347 // clober any other registers (other than the obvious lr).
10349 // Although, technically, IP should be considered a register which may be
10350 // clobbered, the call itself will not touch it. Windows on ARM is a pure
10351 // thumb-2 environment, so there is no interworking required. As a result, we
10352 // do not expect a veneer to be emitted by the linker, clobbering IP.
10354 // Each module receives its own copy of __chkstk, so no import thunk is
10355 // required, again, ensuring that IP is not clobbered.
10357 // Finally, although some linkers may theoretically provide a trampoline for
10358 // out of range calls (which is quite common due to a 32M range limitation of
10359 // branches for Thumb), we can generate the long-call version via
10360 // -mcmodel=large, alleviating the need for the trampoline which may clobber
10363 switch (TM.getCodeModel()) {
10364 case CodeModel::Tiny:
10365 llvm_unreachable("Tiny code model not available on ARM.");
10366 case CodeModel::Small:
10367 case CodeModel::Medium:
10368 case CodeModel::Kernel:
10369 BuildMI(*MBB, MI, DL, TII.get(ARM::tBL))
10370 .add(predOps(ARMCC::AL))
10371 .addExternalSymbol("__chkstk")
10372 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
10373 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
10375 RegState::Implicit | RegState::Define | RegState::Dead)
10377 RegState::Implicit | RegState::Define | RegState::Dead);
10379 case CodeModel::Large: {
10380 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
10381 Register Reg = MRI.createVirtualRegister(&ARM::rGPRRegClass);
10383 BuildMI(*MBB, MI, DL, TII.get(ARM::t2MOVi32imm), Reg)
10384 .addExternalSymbol("__chkstk");
10385 BuildMI(*MBB, MI, DL, TII.get(ARM::tBLXr))
10386 .add(predOps(ARMCC::AL))
10387 .addReg(Reg, RegState::Kill)
10388 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
10389 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
10391 RegState::Implicit | RegState::Define | RegState::Dead)
10393 RegState::Implicit | RegState::Define | RegState::Dead);
10398 BuildMI(*MBB, MI, DL, TII.get(ARM::t2SUBrr), ARM::SP)
10399 .addReg(ARM::SP, RegState::Kill)
10400 .addReg(ARM::R4, RegState::Kill)
10401 .setMIFlags(MachineInstr::FrameSetup)
10402 .add(predOps(ARMCC::AL))
10403 .add(condCodeOp());
10405 MI.eraseFromParent();
10409 MachineBasicBlock *
10410 ARMTargetLowering::EmitLowered__dbzchk(MachineInstr &MI,
10411 MachineBasicBlock *MBB) const {
10412 DebugLoc DL = MI.getDebugLoc();
10413 MachineFunction *MF = MBB->getParent();
10414 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
10416 MachineBasicBlock *ContBB = MF->CreateMachineBasicBlock();
10417 MF->insert(++MBB->getIterator(), ContBB);
10418 ContBB->splice(ContBB->begin(), MBB,
10419 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
10420 ContBB->transferSuccessorsAndUpdatePHIs(MBB);
10421 MBB->addSuccessor(ContBB);
10423 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
10424 BuildMI(TrapBB, DL, TII->get(ARM::t__brkdiv0));
10425 MF->push_back(TrapBB);
10426 MBB->addSuccessor(TrapBB);
10428 BuildMI(*MBB, MI, DL, TII->get(ARM::tCMPi8))
10429 .addReg(MI.getOperand(0).getReg())
10431 .add(predOps(ARMCC::AL));
10432 BuildMI(*MBB, MI, DL, TII->get(ARM::t2Bcc))
10435 .addReg(ARM::CPSR);
10437 MI.eraseFromParent();
10441 // The CPSR operand of SelectItr might be missing a kill marker
10442 // because there were multiple uses of CPSR, and ISel didn't know
10443 // which to mark. Figure out whether SelectItr should have had a
10444 // kill marker, and set it if it should. Returns the correct kill
10446 static bool checkAndUpdateCPSRKill(MachineBasicBlock::iterator SelectItr,
10447 MachineBasicBlock* BB,
10448 const TargetRegisterInfo* TRI) {
10449 // Scan forward through BB for a use/def of CPSR.
10450 MachineBasicBlock::iterator miI(std::next(SelectItr));
10451 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
10452 const MachineInstr& mi = *miI;
10453 if (mi.readsRegister(ARM::CPSR))
10455 if (mi.definesRegister(ARM::CPSR))
10456 break; // Should have kill-flag - update below.
10459 // If we hit the end of the block, check whether CPSR is live into a
10461 if (miI == BB->end()) {
10462 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
10463 sEnd = BB->succ_end();
10464 sItr != sEnd; ++sItr) {
10465 MachineBasicBlock* succ = *sItr;
10466 if (succ->isLiveIn(ARM::CPSR))
10471 // We found a def, or hit the end of the basic block and CPSR wasn't live
10472 // out. SelectMI should have a kill flag on CPSR.
10473 SelectItr->addRegisterKilled(ARM::CPSR, TRI);
10477 MachineBasicBlock *
10478 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
10479 MachineBasicBlock *BB) const {
10480 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
10481 DebugLoc dl = MI.getDebugLoc();
10482 bool isThumb2 = Subtarget->isThumb2();
10483 switch (MI.getOpcode()) {
10486 llvm_unreachable("Unexpected instr type to insert");
10489 // Thumb1 post-indexed loads are really just single-register LDMs.
10490 case ARM::tLDR_postidx: {
10491 MachineOperand Def(MI.getOperand(1));
10492 BuildMI(*BB, MI, dl, TII->get(ARM::tLDMIA_UPD))
10494 .add(MI.getOperand(2)) // Rn
10495 .add(MI.getOperand(3)) // PredImm
10496 .add(MI.getOperand(4)) // PredReg
10497 .add(MI.getOperand(0)) // Rt
10499 MI.eraseFromParent();
10503 // The Thumb2 pre-indexed stores have the same MI operands, they just
10504 // define them differently in the .td files from the isel patterns, so
10505 // they need pseudos.
10506 case ARM::t2STR_preidx:
10507 MI.setDesc(TII->get(ARM::t2STR_PRE));
10509 case ARM::t2STRB_preidx:
10510 MI.setDesc(TII->get(ARM::t2STRB_PRE));
10512 case ARM::t2STRH_preidx:
10513 MI.setDesc(TII->get(ARM::t2STRH_PRE));
10516 case ARM::STRi_preidx:
10517 case ARM::STRBi_preidx: {
10518 unsigned NewOpc = MI.getOpcode() == ARM::STRi_preidx ? ARM::STR_PRE_IMM
10519 : ARM::STRB_PRE_IMM;
10520 // Decode the offset.
10521 unsigned Offset = MI.getOperand(4).getImm();
10522 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
10523 Offset = ARM_AM::getAM2Offset(Offset);
10527 MachineMemOperand *MMO = *MI.memoperands_begin();
10528 BuildMI(*BB, MI, dl, TII->get(NewOpc))
10529 .add(MI.getOperand(0)) // Rn_wb
10530 .add(MI.getOperand(1)) // Rt
10531 .add(MI.getOperand(2)) // Rn
10532 .addImm(Offset) // offset (skip GPR==zero_reg)
10533 .add(MI.getOperand(5)) // pred
10534 .add(MI.getOperand(6))
10535 .addMemOperand(MMO);
10536 MI.eraseFromParent();
10539 case ARM::STRr_preidx:
10540 case ARM::STRBr_preidx:
10541 case ARM::STRH_preidx: {
10543 switch (MI.getOpcode()) {
10544 default: llvm_unreachable("unexpected opcode!");
10545 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
10546 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
10547 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
10549 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
10550 for (unsigned i = 0; i < MI.getNumOperands(); ++i)
10551 MIB.add(MI.getOperand(i));
10552 MI.eraseFromParent();
10556 case ARM::tMOVCCr_pseudo: {
10557 // To "insert" a SELECT_CC instruction, we actually have to insert the
10558 // diamond control-flow pattern. The incoming instruction knows the
10559 // destination vreg to set, the condition code register to branch on, the
10560 // true/false values to select between, and a branch opcode to use.
10561 const BasicBlock *LLVM_BB = BB->getBasicBlock();
10562 MachineFunction::iterator It = ++BB->getIterator();
10567 // cmpTY ccX, r1, r2
10569 // fallthrough --> copy0MBB
10570 MachineBasicBlock *thisMBB = BB;
10571 MachineFunction *F = BB->getParent();
10572 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
10573 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
10574 F->insert(It, copy0MBB);
10575 F->insert(It, sinkMBB);
10577 // Check whether CPSR is live past the tMOVCCr_pseudo.
10578 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
10579 if (!MI.killsRegister(ARM::CPSR) &&
10580 !checkAndUpdateCPSRKill(MI, thisMBB, TRI)) {
10581 copy0MBB->addLiveIn(ARM::CPSR);
10582 sinkMBB->addLiveIn(ARM::CPSR);
10585 // Transfer the remainder of BB and its successor edges to sinkMBB.
10586 sinkMBB->splice(sinkMBB->begin(), BB,
10587 std::next(MachineBasicBlock::iterator(MI)), BB->end());
10588 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
10590 BB->addSuccessor(copy0MBB);
10591 BB->addSuccessor(sinkMBB);
10593 BuildMI(BB, dl, TII->get(ARM::tBcc))
10595 .addImm(MI.getOperand(3).getImm())
10596 .addReg(MI.getOperand(4).getReg());
10599 // %FalseValue = ...
10600 // # fallthrough to sinkMBB
10603 // Update machine-CFG edges
10604 BB->addSuccessor(sinkMBB);
10607 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
10610 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), MI.getOperand(0).getReg())
10611 .addReg(MI.getOperand(1).getReg())
10613 .addReg(MI.getOperand(2).getReg())
10616 MI.eraseFromParent(); // The pseudo instruction is gone now.
10621 case ARM::BCCZi64: {
10622 // If there is an unconditional branch to the other successor, remove it.
10623 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
10625 // Compare both parts that make up the double comparison separately for
10627 bool RHSisZero = MI.getOpcode() == ARM::BCCZi64;
10629 Register LHS1 = MI.getOperand(1).getReg();
10630 Register LHS2 = MI.getOperand(2).getReg();
10632 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
10635 .add(predOps(ARMCC::AL));
10636 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
10637 .addReg(LHS2).addImm(0)
10638 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
10640 Register RHS1 = MI.getOperand(3).getReg();
10641 Register RHS2 = MI.getOperand(4).getReg();
10642 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
10645 .add(predOps(ARMCC::AL));
10646 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
10647 .addReg(LHS2).addReg(RHS2)
10648 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
10651 MachineBasicBlock *destMBB = MI.getOperand(RHSisZero ? 3 : 5).getMBB();
10652 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
10653 if (MI.getOperand(0).getImm() == ARMCC::NE)
10654 std::swap(destMBB, exitMBB);
10656 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
10657 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
10659 BuildMI(BB, dl, TII->get(ARM::t2B))
10661 .add(predOps(ARMCC::AL));
10663 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
10665 MI.eraseFromParent(); // The pseudo instruction is gone now.
10669 case ARM::Int_eh_sjlj_setjmp:
10670 case ARM::Int_eh_sjlj_setjmp_nofp:
10671 case ARM::tInt_eh_sjlj_setjmp:
10672 case ARM::t2Int_eh_sjlj_setjmp:
10673 case ARM::t2Int_eh_sjlj_setjmp_nofp:
10676 case ARM::Int_eh_sjlj_setup_dispatch:
10677 EmitSjLjDispatchBlock(MI, BB);
10682 // To insert an ABS instruction, we have to insert the
10683 // diamond control-flow pattern. The incoming instruction knows the
10684 // source vreg to test against 0, the destination vreg to set,
10685 // the condition code register to branch on, the
10686 // true/false values to select between, and a branch opcode to use.
10691 // BCC (branch to SinkBB if V0 >= 0)
10692 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
10693 // SinkBB: V1 = PHI(V2, V3)
10694 const BasicBlock *LLVM_BB = BB->getBasicBlock();
10695 MachineFunction::iterator BBI = ++BB->getIterator();
10696 MachineFunction *Fn = BB->getParent();
10697 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
10698 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
10699 Fn->insert(BBI, RSBBB);
10700 Fn->insert(BBI, SinkBB);
10702 Register ABSSrcReg = MI.getOperand(1).getReg();
10703 Register ABSDstReg = MI.getOperand(0).getReg();
10704 bool ABSSrcKIll = MI.getOperand(1).isKill();
10705 bool isThumb2 = Subtarget->isThumb2();
10706 MachineRegisterInfo &MRI = Fn->getRegInfo();
10707 // In Thumb mode S must not be specified if source register is the SP or
10708 // PC and if destination register is the SP, so restrict register class
10709 Register NewRsbDstReg = MRI.createVirtualRegister(
10710 isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRRegClass);
10712 // Transfer the remainder of BB and its successor edges to sinkMBB.
10713 SinkBB->splice(SinkBB->begin(), BB,
10714 std::next(MachineBasicBlock::iterator(MI)), BB->end());
10715 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
10717 BB->addSuccessor(RSBBB);
10718 BB->addSuccessor(SinkBB);
10720 // fall through to SinkMBB
10721 RSBBB->addSuccessor(SinkBB);
10723 // insert a cmp at the end of BB
10724 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
10727 .add(predOps(ARMCC::AL));
10729 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
10731 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
10732 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
10734 // insert rsbri in RSBBB
10735 // Note: BCC and rsbri will be converted into predicated rsbmi
10736 // by if-conversion pass
10737 BuildMI(*RSBBB, RSBBB->begin(), dl,
10738 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
10739 .addReg(ABSSrcReg, ABSSrcKIll ? RegState::Kill : 0)
10741 .add(predOps(ARMCC::AL))
10742 .add(condCodeOp());
10744 // insert PHI in SinkBB,
10745 // reuse ABSDstReg to not change uses of ABS instruction
10746 BuildMI(*SinkBB, SinkBB->begin(), dl,
10747 TII->get(ARM::PHI), ABSDstReg)
10748 .addReg(NewRsbDstReg).addMBB(RSBBB)
10749 .addReg(ABSSrcReg).addMBB(BB);
10751 // remove ABS instruction
10752 MI.eraseFromParent();
10754 // return last added BB
10757 case ARM::COPY_STRUCT_BYVAL_I32:
10759 return EmitStructByval(MI, BB);
10760 case ARM::WIN__CHKSTK:
10761 return EmitLowered__chkstk(MI, BB);
10762 case ARM::WIN__DBZCHK:
10763 return EmitLowered__dbzchk(MI, BB);
10767 /// Attaches vregs to MEMCPY that it will use as scratch registers
10768 /// when it is expanded into LDM/STM. This is done as a post-isel lowering
10769 /// instead of as a custom inserter because we need the use list from the SDNode.
10770 static void attachMEMCPYScratchRegs(const ARMSubtarget *Subtarget,
10771 MachineInstr &MI, const SDNode *Node) {
10772 bool isThumb1 = Subtarget->isThumb1Only();
10774 DebugLoc DL = MI.getDebugLoc();
10775 MachineFunction *MF = MI.getParent()->getParent();
10776 MachineRegisterInfo &MRI = MF->getRegInfo();
10777 MachineInstrBuilder MIB(*MF, MI);
10779 // If the new dst/src is unused mark it as dead.
10780 if (!Node->hasAnyUseOfValue(0)) {
10781 MI.getOperand(0).setIsDead(true);
10783 if (!Node->hasAnyUseOfValue(1)) {
10784 MI.getOperand(1).setIsDead(true);
10787 // The MEMCPY both defines and kills the scratch registers.
10788 for (unsigned I = 0; I != MI.getOperand(4).getImm(); ++I) {
10789 Register TmpReg = MRI.createVirtualRegister(isThumb1 ? &ARM::tGPRRegClass
10790 : &ARM::GPRRegClass);
10791 MIB.addReg(TmpReg, RegState::Define|RegState::Dead);
10795 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr &MI,
10796 SDNode *Node) const {
10797 if (MI.getOpcode() == ARM::MEMCPY) {
10798 attachMEMCPYScratchRegs(Subtarget, MI, Node);
10802 const MCInstrDesc *MCID = &MI.getDesc();
10803 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
10804 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
10805 // operand is still set to noreg. If needed, set the optional operand's
10806 // register to CPSR, and remove the redundant implicit def.
10808 // e.g. ADCS (..., implicit-def CPSR) -> ADC (... opt:def CPSR).
10810 // Rename pseudo opcodes.
10811 unsigned NewOpc = convertAddSubFlagsOpcode(MI.getOpcode());
10814 const ARMBaseInstrInfo *TII = Subtarget->getInstrInfo();
10815 MCID = &TII->get(NewOpc);
10817 assert(MCID->getNumOperands() ==
10818 MI.getDesc().getNumOperands() + 5 - MI.getDesc().getSize()
10819 && "converted opcode should be the same except for cc_out"
10820 " (and, on Thumb1, pred)");
10824 // Add the optional cc_out operand
10825 MI.addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
10827 // On Thumb1, move all input operands to the end, then add the predicate
10828 if (Subtarget->isThumb1Only()) {
10829 for (unsigned c = MCID->getNumOperands() - 4; c--;) {
10830 MI.addOperand(MI.getOperand(1));
10831 MI.RemoveOperand(1);
10834 // Restore the ties
10835 for (unsigned i = MI.getNumOperands(); i--;) {
10836 const MachineOperand& op = MI.getOperand(i);
10837 if (op.isReg() && op.isUse()) {
10838 int DefIdx = MCID->getOperandConstraint(i, MCOI::TIED_TO);
10840 MI.tieOperands(DefIdx, i);
10844 MI.addOperand(MachineOperand::CreateImm(ARMCC::AL));
10845 MI.addOperand(MachineOperand::CreateReg(0, /*isDef=*/false));
10848 ccOutIdx = MCID->getNumOperands() - 1;
10850 ccOutIdx = MCID->getNumOperands() - 1;
10852 // Any ARM instruction that sets the 's' bit should specify an optional
10853 // "cc_out" operand in the last operand position.
10854 if (!MI.hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
10855 assert(!NewOpc && "Optional cc_out operand required");
10858 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
10859 // since we already have an optional CPSR def.
10860 bool definesCPSR = false;
10861 bool deadCPSR = false;
10862 for (unsigned i = MCID->getNumOperands(), e = MI.getNumOperands(); i != e;
10864 const MachineOperand &MO = MI.getOperand(i);
10865 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
10866 definesCPSR = true;
10869 MI.RemoveOperand(i);
10873 if (!definesCPSR) {
10874 assert(!NewOpc && "Optional cc_out operand required");
10877 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
10879 assert(!MI.getOperand(ccOutIdx).getReg() &&
10880 "expect uninitialized optional cc_out operand");
10881 // Thumb1 instructions must have the S bit even if the CPSR is dead.
10882 if (!Subtarget->isThumb1Only())
10886 // If this instruction was defined with an optional CPSR def and its dag node
10887 // had a live implicit CPSR def, then activate the optional CPSR def.
10888 MachineOperand &MO = MI.getOperand(ccOutIdx);
10889 MO.setReg(ARM::CPSR);
10893 //===----------------------------------------------------------------------===//
10894 // ARM Optimization Hooks
10895 //===----------------------------------------------------------------------===//
10897 // Helper function that checks if N is a null or all ones constant.
10898 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
10899 return AllOnes ? isAllOnesConstant(N) : isNullConstant(N);
10902 // Return true if N is conditionally 0 or all ones.
10903 // Detects these expressions where cc is an i1 value:
10905 // (select cc 0, y) [AllOnes=0]
10906 // (select cc y, 0) [AllOnes=0]
10907 // (zext cc) [AllOnes=0]
10908 // (sext cc) [AllOnes=0/1]
10909 // (select cc -1, y) [AllOnes=1]
10910 // (select cc y, -1) [AllOnes=1]
10912 // Invert is set when N is the null/all ones constant when CC is false.
10913 // OtherOp is set to the alternative value of N.
10914 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
10915 SDValue &CC, bool &Invert,
10917 SelectionDAG &DAG) {
10918 switch (N->getOpcode()) {
10919 default: return false;
10920 case ISD::SELECT: {
10921 CC = N->getOperand(0);
10922 SDValue N1 = N->getOperand(1);
10923 SDValue N2 = N->getOperand(2);
10924 if (isZeroOrAllOnes(N1, AllOnes)) {
10929 if (isZeroOrAllOnes(N2, AllOnes)) {
10936 case ISD::ZERO_EXTEND:
10937 // (zext cc) can never be the all ones value.
10941 case ISD::SIGN_EXTEND: {
10943 EVT VT = N->getValueType(0);
10944 CC = N->getOperand(0);
10945 if (CC.getValueType() != MVT::i1 || CC.getOpcode() != ISD::SETCC)
10949 // When looking for an AllOnes constant, N is an sext, and the 'other'
10951 OtherOp = DAG.getConstant(0, dl, VT);
10952 else if (N->getOpcode() == ISD::ZERO_EXTEND)
10953 // When looking for a 0 constant, N can be zext or sext.
10954 OtherOp = DAG.getConstant(1, dl, VT);
10956 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), dl,
10963 // Combine a constant select operand into its use:
10965 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
10966 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
10967 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
10968 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
10969 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
10971 // The transform is rejected if the select doesn't have a constant operand that
10972 // is null, or all ones when AllOnes is set.
10974 // Also recognize sext/zext from i1:
10976 // (add (zext cc), x) -> (select cc (add x, 1), x)
10977 // (add (sext cc), x) -> (select cc (add x, -1), x)
10979 // These transformations eventually create predicated instructions.
10981 // @param N The node to transform.
10982 // @param Slct The N operand that is a select.
10983 // @param OtherOp The other N operand (x above).
10984 // @param DCI Context.
10985 // @param AllOnes Require the select constant to be all ones instead of null.
10986 // @returns The new node, or SDValue() on failure.
10988 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
10989 TargetLowering::DAGCombinerInfo &DCI,
10990 bool AllOnes = false) {
10991 SelectionDAG &DAG = DCI.DAG;
10992 EVT VT = N->getValueType(0);
10993 SDValue NonConstantVal;
10995 bool SwapSelectOps;
10996 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
10997 NonConstantVal, DAG))
11000 // Slct is now know to be the desired identity constant when CC is true.
11001 SDValue TrueVal = OtherOp;
11002 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
11003 OtherOp, NonConstantVal);
11004 // Unless SwapSelectOps says CC should be false.
11006 std::swap(TrueVal, FalseVal);
11008 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
11009 CCOp, TrueVal, FalseVal);
11012 // Attempt combineSelectAndUse on each operand of a commutative operator N.
11014 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
11015 TargetLowering::DAGCombinerInfo &DCI) {
11016 SDValue N0 = N->getOperand(0);
11017 SDValue N1 = N->getOperand(1);
11018 if (N0.getNode()->hasOneUse())
11019 if (SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes))
11021 if (N1.getNode()->hasOneUse())
11022 if (SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes))
11027 static bool IsVUZPShuffleNode(SDNode *N) {
11028 // VUZP shuffle node.
11029 if (N->getOpcode() == ARMISD::VUZP)
11032 // "VUZP" on i32 is an alias for VTRN.
11033 if (N->getOpcode() == ARMISD::VTRN && N->getValueType(0) == MVT::v2i32)
11039 static SDValue AddCombineToVPADD(SDNode *N, SDValue N0, SDValue N1,
11040 TargetLowering::DAGCombinerInfo &DCI,
11041 const ARMSubtarget *Subtarget) {
11042 // Look for ADD(VUZP.0, VUZP.1).
11043 if (!IsVUZPShuffleNode(N0.getNode()) || N0.getNode() != N1.getNode() ||
11047 // Make sure the ADD is a 64-bit add; there is no 128-bit VPADD.
11048 if (!N->getValueType(0).is64BitVector())
11052 SelectionDAG &DAG = DCI.DAG;
11053 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11055 SDNode *Unzip = N0.getNode();
11056 EVT VT = N->getValueType(0);
11058 SmallVector<SDValue, 8> Ops;
11059 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpadd, dl,
11060 TLI.getPointerTy(DAG.getDataLayout())));
11061 Ops.push_back(Unzip->getOperand(0));
11062 Ops.push_back(Unzip->getOperand(1));
11064 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT, Ops);
11067 static SDValue AddCombineVUZPToVPADDL(SDNode *N, SDValue N0, SDValue N1,
11068 TargetLowering::DAGCombinerInfo &DCI,
11069 const ARMSubtarget *Subtarget) {
11070 // Check for two extended operands.
11071 if (!(N0.getOpcode() == ISD::SIGN_EXTEND &&
11072 N1.getOpcode() == ISD::SIGN_EXTEND) &&
11073 !(N0.getOpcode() == ISD::ZERO_EXTEND &&
11074 N1.getOpcode() == ISD::ZERO_EXTEND))
11077 SDValue N00 = N0.getOperand(0);
11078 SDValue N10 = N1.getOperand(0);
11080 // Look for ADD(SEXT(VUZP.0), SEXT(VUZP.1))
11081 if (!IsVUZPShuffleNode(N00.getNode()) || N00.getNode() != N10.getNode() ||
11085 // We only recognize Q register paddl here; this can't be reached until
11086 // after type legalization.
11087 if (!N00.getValueType().is64BitVector() ||
11088 !N0.getValueType().is128BitVector())
11091 // Generate vpaddl.
11092 SelectionDAG &DAG = DCI.DAG;
11093 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11095 EVT VT = N->getValueType(0);
11097 SmallVector<SDValue, 8> Ops;
11098 // Form vpaddl.sN or vpaddl.uN depending on the kind of extension.
11100 if (N0.getOpcode() == ISD::SIGN_EXTEND)
11101 Opcode = Intrinsic::arm_neon_vpaddls;
11103 Opcode = Intrinsic::arm_neon_vpaddlu;
11104 Ops.push_back(DAG.getConstant(Opcode, dl,
11105 TLI.getPointerTy(DAG.getDataLayout())));
11106 EVT ElemTy = N00.getValueType().getVectorElementType();
11107 unsigned NumElts = VT.getVectorNumElements();
11108 EVT ConcatVT = EVT::getVectorVT(*DAG.getContext(), ElemTy, NumElts * 2);
11109 SDValue Concat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), ConcatVT,
11110 N00.getOperand(0), N00.getOperand(1));
11111 Ops.push_back(Concat);
11113 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT, Ops);
11116 // FIXME: This function shouldn't be necessary; if we lower BUILD_VECTOR in
11117 // an appropriate manner, we end up with ADD(VUZP(ZEXT(N))), which is
11118 // much easier to match.
11120 AddCombineBUILD_VECTORToVPADDL(SDNode *N, SDValue N0, SDValue N1,
11121 TargetLowering::DAGCombinerInfo &DCI,
11122 const ARMSubtarget *Subtarget) {
11123 // Only perform optimization if after legalize, and if NEON is available. We
11124 // also expected both operands to be BUILD_VECTORs.
11125 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
11126 || N0.getOpcode() != ISD::BUILD_VECTOR
11127 || N1.getOpcode() != ISD::BUILD_VECTOR)
11130 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
11131 EVT VT = N->getValueType(0);
11132 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
11135 // Check that the vector operands are of the right form.
11136 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
11137 // operands, where N is the size of the formed vector.
11138 // Each EXTRACT_VECTOR should have the same input vector and odd or even
11139 // index such that we have a pair wise add pattern.
11141 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
11142 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
11144 SDValue Vec = N0->getOperand(0)->getOperand(0);
11145 SDNode *V = Vec.getNode();
11146 unsigned nextIndex = 0;
11148 // For each operands to the ADD which are BUILD_VECTORs,
11149 // check to see if each of their operands are an EXTRACT_VECTOR with
11150 // the same vector and appropriate index.
11151 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
11152 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
11153 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
11155 SDValue ExtVec0 = N0->getOperand(i);
11156 SDValue ExtVec1 = N1->getOperand(i);
11158 // First operand is the vector, verify its the same.
11159 if (V != ExtVec0->getOperand(0).getNode() ||
11160 V != ExtVec1->getOperand(0).getNode())
11163 // Second is the constant, verify its correct.
11164 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
11165 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
11167 // For the constant, we want to see all the even or all the odd.
11168 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
11169 || C1->getZExtValue() != nextIndex+1)
11172 // Increment index.
11178 // Don't generate vpaddl+vmovn; we'll match it to vpadd later. Also make sure
11179 // we're using the entire input vector, otherwise there's a size/legality
11180 // mismatch somewhere.
11181 if (nextIndex != Vec.getValueType().getVectorNumElements() ||
11182 Vec.getValueType().getVectorElementType() == VT.getVectorElementType())
11185 // Create VPADDL node.
11186 SelectionDAG &DAG = DCI.DAG;
11187 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11191 // Build operand list.
11192 SmallVector<SDValue, 8> Ops;
11193 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls, dl,
11194 TLI.getPointerTy(DAG.getDataLayout())));
11196 // Input is the vector.
11197 Ops.push_back(Vec);
11199 // Get widened type and narrowed type.
11201 unsigned numElem = VT.getVectorNumElements();
11203 EVT inputLaneType = Vec.getValueType().getVectorElementType();
11204 switch (inputLaneType.getSimpleVT().SimpleTy) {
11205 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
11206 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
11207 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
11209 llvm_unreachable("Invalid vector element type for padd optimization.");
11212 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, widenType, Ops);
11213 unsigned ExtOp = VT.bitsGT(tmp.getValueType()) ? ISD::ANY_EXTEND : ISD::TRUNCATE;
11214 return DAG.getNode(ExtOp, dl, VT, tmp);
11217 static SDValue findMUL_LOHI(SDValue V) {
11218 if (V->getOpcode() == ISD::UMUL_LOHI ||
11219 V->getOpcode() == ISD::SMUL_LOHI)
11224 static SDValue AddCombineTo64BitSMLAL16(SDNode *AddcNode, SDNode *AddeNode,
11225 TargetLowering::DAGCombinerInfo &DCI,
11226 const ARMSubtarget *Subtarget) {
11227 if (!Subtarget->hasBaseDSP())
11230 // SMLALBB, SMLALBT, SMLALTB, SMLALTT multiply two 16-bit values and
11231 // accumulates the product into a 64-bit value. The 16-bit values will
11232 // be sign extended somehow or SRA'd into 32-bit values
11233 // (addc (adde (mul 16bit, 16bit), lo), hi)
11234 SDValue Mul = AddcNode->getOperand(0);
11235 SDValue Lo = AddcNode->getOperand(1);
11236 if (Mul.getOpcode() != ISD::MUL) {
11237 Lo = AddcNode->getOperand(0);
11238 Mul = AddcNode->getOperand(1);
11239 if (Mul.getOpcode() != ISD::MUL)
11243 SDValue SRA = AddeNode->getOperand(0);
11244 SDValue Hi = AddeNode->getOperand(1);
11245 if (SRA.getOpcode() != ISD::SRA) {
11246 SRA = AddeNode->getOperand(1);
11247 Hi = AddeNode->getOperand(0);
11248 if (SRA.getOpcode() != ISD::SRA)
11251 if (auto Const = dyn_cast<ConstantSDNode>(SRA.getOperand(1))) {
11252 if (Const->getZExtValue() != 31)
11257 if (SRA.getOperand(0) != Mul)
11260 SelectionDAG &DAG = DCI.DAG;
11261 SDLoc dl(AddcNode);
11262 unsigned Opcode = 0;
11266 if (isS16(Mul.getOperand(0), DAG) && isS16(Mul.getOperand(1), DAG)) {
11267 Opcode = ARMISD::SMLALBB;
11268 Op0 = Mul.getOperand(0);
11269 Op1 = Mul.getOperand(1);
11270 } else if (isS16(Mul.getOperand(0), DAG) && isSRA16(Mul.getOperand(1))) {
11271 Opcode = ARMISD::SMLALBT;
11272 Op0 = Mul.getOperand(0);
11273 Op1 = Mul.getOperand(1).getOperand(0);
11274 } else if (isSRA16(Mul.getOperand(0)) && isS16(Mul.getOperand(1), DAG)) {
11275 Opcode = ARMISD::SMLALTB;
11276 Op0 = Mul.getOperand(0).getOperand(0);
11277 Op1 = Mul.getOperand(1);
11278 } else if (isSRA16(Mul.getOperand(0)) && isSRA16(Mul.getOperand(1))) {
11279 Opcode = ARMISD::SMLALTT;
11280 Op0 = Mul->getOperand(0).getOperand(0);
11281 Op1 = Mul->getOperand(1).getOperand(0);
11287 SDValue SMLAL = DAG.getNode(Opcode, dl, DAG.getVTList(MVT::i32, MVT::i32),
11289 // Replace the ADDs' nodes uses by the MLA node's values.
11290 SDValue HiMLALResult(SMLAL.getNode(), 1);
11291 SDValue LoMLALResult(SMLAL.getNode(), 0);
11293 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
11294 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
11296 // Return original node to notify the driver to stop replacing.
11297 SDValue resNode(AddcNode, 0);
11301 static SDValue AddCombineTo64bitMLAL(SDNode *AddeSubeNode,
11302 TargetLowering::DAGCombinerInfo &DCI,
11303 const ARMSubtarget *Subtarget) {
11304 // Look for multiply add opportunities.
11305 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
11306 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
11307 // a glue link from the first add to the second add.
11308 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
11309 // a S/UMLAL instruction.
11312 // V \ [no multiline comment]
11318 // In the special case where only the higher part of a signed result is used
11319 // and the add to the low part of the result of ISD::UMUL_LOHI adds or subtracts
11320 // a constant with the exact value of 0x80000000, we recognize we are dealing
11321 // with a "rounded multiply and add" (or subtract) and transform it into
11322 // either a ARMISD::SMMLAR or ARMISD::SMMLSR respectively.
11324 assert((AddeSubeNode->getOpcode() == ARMISD::ADDE ||
11325 AddeSubeNode->getOpcode() == ARMISD::SUBE) &&
11326 "Expect an ADDE or SUBE");
11328 assert(AddeSubeNode->getNumOperands() == 3 &&
11329 AddeSubeNode->getOperand(2).getValueType() == MVT::i32 &&
11330 "ADDE node has the wrong inputs");
11332 // Check that we are chained to the right ADDC or SUBC node.
11333 SDNode *AddcSubcNode = AddeSubeNode->getOperand(2).getNode();
11334 if ((AddeSubeNode->getOpcode() == ARMISD::ADDE &&
11335 AddcSubcNode->getOpcode() != ARMISD::ADDC) ||
11336 (AddeSubeNode->getOpcode() == ARMISD::SUBE &&
11337 AddcSubcNode->getOpcode() != ARMISD::SUBC))
11340 SDValue AddcSubcOp0 = AddcSubcNode->getOperand(0);
11341 SDValue AddcSubcOp1 = AddcSubcNode->getOperand(1);
11343 // Check if the two operands are from the same mul_lohi node.
11344 if (AddcSubcOp0.getNode() == AddcSubcOp1.getNode())
11347 assert(AddcSubcNode->getNumValues() == 2 &&
11348 AddcSubcNode->getValueType(0) == MVT::i32 &&
11349 "Expect ADDC with two result values. First: i32");
11351 // Check that the ADDC adds the low result of the S/UMUL_LOHI. If not, it
11352 // maybe a SMLAL which multiplies two 16-bit values.
11353 if (AddeSubeNode->getOpcode() == ARMISD::ADDE &&
11354 AddcSubcOp0->getOpcode() != ISD::UMUL_LOHI &&
11355 AddcSubcOp0->getOpcode() != ISD::SMUL_LOHI &&
11356 AddcSubcOp1->getOpcode() != ISD::UMUL_LOHI &&
11357 AddcSubcOp1->getOpcode() != ISD::SMUL_LOHI)
11358 return AddCombineTo64BitSMLAL16(AddcSubcNode, AddeSubeNode, DCI, Subtarget);
11360 // Check for the triangle shape.
11361 SDValue AddeSubeOp0 = AddeSubeNode->getOperand(0);
11362 SDValue AddeSubeOp1 = AddeSubeNode->getOperand(1);
11364 // Make sure that the ADDE/SUBE operands are not coming from the same node.
11365 if (AddeSubeOp0.getNode() == AddeSubeOp1.getNode())
11368 // Find the MUL_LOHI node walking up ADDE/SUBE's operands.
11369 bool IsLeftOperandMUL = false;
11370 SDValue MULOp = findMUL_LOHI(AddeSubeOp0);
11371 if (MULOp == SDValue())
11372 MULOp = findMUL_LOHI(AddeSubeOp1);
11374 IsLeftOperandMUL = true;
11375 if (MULOp == SDValue())
11378 // Figure out the right opcode.
11379 unsigned Opc = MULOp->getOpcode();
11380 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
11382 // Figure out the high and low input values to the MLAL node.
11383 SDValue *HiAddSub = nullptr;
11384 SDValue *LoMul = nullptr;
11385 SDValue *LowAddSub = nullptr;
11387 // Ensure that ADDE/SUBE is from high result of ISD::xMUL_LOHI.
11388 if ((AddeSubeOp0 != MULOp.getValue(1)) && (AddeSubeOp1 != MULOp.getValue(1)))
11391 if (IsLeftOperandMUL)
11392 HiAddSub = &AddeSubeOp1;
11394 HiAddSub = &AddeSubeOp0;
11396 // Ensure that LoMul and LowAddSub are taken from correct ISD::SMUL_LOHI node
11397 // whose low result is fed to the ADDC/SUBC we are checking.
11399 if (AddcSubcOp0 == MULOp.getValue(0)) {
11400 LoMul = &AddcSubcOp0;
11401 LowAddSub = &AddcSubcOp1;
11403 if (AddcSubcOp1 == MULOp.getValue(0)) {
11404 LoMul = &AddcSubcOp1;
11405 LowAddSub = &AddcSubcOp0;
11411 // If HiAddSub is the same node as ADDC/SUBC or is a predecessor of ADDC/SUBC
11412 // the replacement below will create a cycle.
11413 if (AddcSubcNode == HiAddSub->getNode() ||
11414 AddcSubcNode->isPredecessorOf(HiAddSub->getNode()))
11417 // Create the merged node.
11418 SelectionDAG &DAG = DCI.DAG;
11420 // Start building operand list.
11421 SmallVector<SDValue, 8> Ops;
11422 Ops.push_back(LoMul->getOperand(0));
11423 Ops.push_back(LoMul->getOperand(1));
11425 // Check whether we can use SMMLAR, SMMLSR or SMMULR instead. For this to be
11426 // the case, we must be doing signed multiplication and only use the higher
11427 // part of the result of the MLAL, furthermore the LowAddSub must be a constant
11428 // addition or subtraction with the value of 0x800000.
11429 if (Subtarget->hasV6Ops() && Subtarget->hasDSP() && Subtarget->useMulOps() &&
11430 FinalOpc == ARMISD::SMLAL && !AddeSubeNode->hasAnyUseOfValue(1) &&
11431 LowAddSub->getNode()->getOpcode() == ISD::Constant &&
11432 static_cast<ConstantSDNode *>(LowAddSub->getNode())->getZExtValue() ==
11434 Ops.push_back(*HiAddSub);
11435 if (AddcSubcNode->getOpcode() == ARMISD::SUBC) {
11436 FinalOpc = ARMISD::SMMLSR;
11438 FinalOpc = ARMISD::SMMLAR;
11440 SDValue NewNode = DAG.getNode(FinalOpc, SDLoc(AddcSubcNode), MVT::i32, Ops);
11441 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeSubeNode, 0), NewNode);
11443 return SDValue(AddeSubeNode, 0);
11444 } else if (AddcSubcNode->getOpcode() == ARMISD::SUBC)
11445 // SMMLS is generated during instruction selection and the rest of this
11446 // function can not handle the case where AddcSubcNode is a SUBC.
11449 // Finish building the operand list for {U/S}MLAL
11450 Ops.push_back(*LowAddSub);
11451 Ops.push_back(*HiAddSub);
11453 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcSubcNode),
11454 DAG.getVTList(MVT::i32, MVT::i32), Ops);
11456 // Replace the ADDs' nodes uses by the MLA node's values.
11457 SDValue HiMLALResult(MLALNode.getNode(), 1);
11458 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeSubeNode, 0), HiMLALResult);
11460 SDValue LoMLALResult(MLALNode.getNode(), 0);
11461 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcSubcNode, 0), LoMLALResult);
11463 // Return original node to notify the driver to stop replacing.
11464 return SDValue(AddeSubeNode, 0);
11467 static SDValue AddCombineTo64bitUMAAL(SDNode *AddeNode,
11468 TargetLowering::DAGCombinerInfo &DCI,
11469 const ARMSubtarget *Subtarget) {
11470 // UMAAL is similar to UMLAL except that it adds two unsigned values.
11471 // While trying to combine for the other MLAL nodes, first search for the
11472 // chance to use UMAAL. Check if Addc uses a node which has already
11473 // been combined into a UMLAL. The other pattern is UMLAL using Addc/Adde
11474 // as the addend, and it's handled in PerformUMLALCombine.
11476 if (!Subtarget->hasV6Ops() || !Subtarget->hasDSP())
11477 return AddCombineTo64bitMLAL(AddeNode, DCI, Subtarget);
11479 // Check that we have a glued ADDC node.
11480 SDNode* AddcNode = AddeNode->getOperand(2).getNode();
11481 if (AddcNode->getOpcode() != ARMISD::ADDC)
11484 // Find the converted UMAAL or quit if it doesn't exist.
11485 SDNode *UmlalNode = nullptr;
11487 if (AddcNode->getOperand(0).getOpcode() == ARMISD::UMLAL) {
11488 UmlalNode = AddcNode->getOperand(0).getNode();
11489 AddHi = AddcNode->getOperand(1);
11490 } else if (AddcNode->getOperand(1).getOpcode() == ARMISD::UMLAL) {
11491 UmlalNode = AddcNode->getOperand(1).getNode();
11492 AddHi = AddcNode->getOperand(0);
11494 return AddCombineTo64bitMLAL(AddeNode, DCI, Subtarget);
11497 // The ADDC should be glued to an ADDE node, which uses the same UMLAL as
11498 // the ADDC as well as Zero.
11499 if (!isNullConstant(UmlalNode->getOperand(3)))
11502 if ((isNullConstant(AddeNode->getOperand(0)) &&
11503 AddeNode->getOperand(1).getNode() == UmlalNode) ||
11504 (AddeNode->getOperand(0).getNode() == UmlalNode &&
11505 isNullConstant(AddeNode->getOperand(1)))) {
11506 SelectionDAG &DAG = DCI.DAG;
11507 SDValue Ops[] = { UmlalNode->getOperand(0), UmlalNode->getOperand(1),
11508 UmlalNode->getOperand(2), AddHi };
11509 SDValue UMAAL = DAG.getNode(ARMISD::UMAAL, SDLoc(AddcNode),
11510 DAG.getVTList(MVT::i32, MVT::i32), Ops);
11512 // Replace the ADDs' nodes uses by the UMAAL node's values.
11513 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), SDValue(UMAAL.getNode(), 1));
11514 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), SDValue(UMAAL.getNode(), 0));
11516 // Return original node to notify the driver to stop replacing.
11517 return SDValue(AddeNode, 0);
11522 static SDValue PerformUMLALCombine(SDNode *N, SelectionDAG &DAG,
11523 const ARMSubtarget *Subtarget) {
11524 if (!Subtarget->hasV6Ops() || !Subtarget->hasDSP())
11527 // Check that we have a pair of ADDC and ADDE as operands.
11528 // Both addends of the ADDE must be zero.
11529 SDNode* AddcNode = N->getOperand(2).getNode();
11530 SDNode* AddeNode = N->getOperand(3).getNode();
11531 if ((AddcNode->getOpcode() == ARMISD::ADDC) &&
11532 (AddeNode->getOpcode() == ARMISD::ADDE) &&
11533 isNullConstant(AddeNode->getOperand(0)) &&
11534 isNullConstant(AddeNode->getOperand(1)) &&
11535 (AddeNode->getOperand(2).getNode() == AddcNode))
11536 return DAG.getNode(ARMISD::UMAAL, SDLoc(N),
11537 DAG.getVTList(MVT::i32, MVT::i32),
11538 {N->getOperand(0), N->getOperand(1),
11539 AddcNode->getOperand(0), AddcNode->getOperand(1)});
11544 static SDValue PerformAddcSubcCombine(SDNode *N,
11545 TargetLowering::DAGCombinerInfo &DCI,
11546 const ARMSubtarget *Subtarget) {
11547 SelectionDAG &DAG(DCI.DAG);
11549 if (N->getOpcode() == ARMISD::SUBC) {
11550 // (SUBC (ADDE 0, 0, C), 1) -> C
11551 SDValue LHS = N->getOperand(0);
11552 SDValue RHS = N->getOperand(1);
11553 if (LHS->getOpcode() == ARMISD::ADDE &&
11554 isNullConstant(LHS->getOperand(0)) &&
11555 isNullConstant(LHS->getOperand(1)) && isOneConstant(RHS)) {
11556 return DCI.CombineTo(N, SDValue(N, 0), LHS->getOperand(2));
11560 if (Subtarget->isThumb1Only()) {
11561 SDValue RHS = N->getOperand(1);
11562 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS)) {
11563 int32_t imm = C->getSExtValue();
11564 if (imm < 0 && imm > std::numeric_limits<int>::min()) {
11566 RHS = DAG.getConstant(-imm, DL, MVT::i32);
11567 unsigned Opcode = (N->getOpcode() == ARMISD::ADDC) ? ARMISD::SUBC
11569 return DAG.getNode(Opcode, DL, N->getVTList(), N->getOperand(0), RHS);
11577 static SDValue PerformAddeSubeCombine(SDNode *N,
11578 TargetLowering::DAGCombinerInfo &DCI,
11579 const ARMSubtarget *Subtarget) {
11580 if (Subtarget->isThumb1Only()) {
11581 SelectionDAG &DAG = DCI.DAG;
11582 SDValue RHS = N->getOperand(1);
11583 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS)) {
11584 int64_t imm = C->getSExtValue();
11588 // The with-carry-in form matches bitwise not instead of the negation.
11589 // Effectively, the inverse interpretation of the carry flag already
11590 // accounts for part of the negation.
11591 RHS = DAG.getConstant(~imm, DL, MVT::i32);
11593 unsigned Opcode = (N->getOpcode() == ARMISD::ADDE) ? ARMISD::SUBE
11595 return DAG.getNode(Opcode, DL, N->getVTList(),
11596 N->getOperand(0), RHS, N->getOperand(2));
11599 } else if (N->getOperand(1)->getOpcode() == ISD::SMUL_LOHI) {
11600 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
11605 static SDValue PerformABSCombine(SDNode *N,
11606 TargetLowering::DAGCombinerInfo &DCI,
11607 const ARMSubtarget *Subtarget) {
11609 SelectionDAG &DAG = DCI.DAG;
11610 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11612 if (TLI.isOperationLegal(N->getOpcode(), N->getValueType(0)))
11615 if (!TLI.expandABS(N, res, DAG))
11621 /// PerformADDECombine - Target-specific dag combine transform from
11622 /// ARMISD::ADDC, ARMISD::ADDE, and ISD::MUL_LOHI to MLAL or
11623 /// ARMISD::ADDC, ARMISD::ADDE and ARMISD::UMLAL to ARMISD::UMAAL
11624 static SDValue PerformADDECombine(SDNode *N,
11625 TargetLowering::DAGCombinerInfo &DCI,
11626 const ARMSubtarget *Subtarget) {
11627 // Only ARM and Thumb2 support UMLAL/SMLAL.
11628 if (Subtarget->isThumb1Only())
11629 return PerformAddeSubeCombine(N, DCI, Subtarget);
11631 // Only perform the checks after legalize when the pattern is available.
11632 if (DCI.isBeforeLegalize()) return SDValue();
11634 return AddCombineTo64bitUMAAL(N, DCI, Subtarget);
11637 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
11638 /// operands N0 and N1. This is a helper for PerformADDCombine that is
11639 /// called with the default operands, and if that fails, with commuted
11641 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
11642 TargetLowering::DAGCombinerInfo &DCI,
11643 const ARMSubtarget *Subtarget){
11644 // Attempt to create vpadd for this add.
11645 if (SDValue Result = AddCombineToVPADD(N, N0, N1, DCI, Subtarget))
11648 // Attempt to create vpaddl for this add.
11649 if (SDValue Result = AddCombineVUZPToVPADDL(N, N0, N1, DCI, Subtarget))
11651 if (SDValue Result = AddCombineBUILD_VECTORToVPADDL(N, N0, N1, DCI,
11655 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
11656 if (N0.getNode()->hasOneUse())
11657 if (SDValue Result = combineSelectAndUse(N, N0, N1, DCI))
11663 ARMTargetLowering::isDesirableToCommuteWithShift(const SDNode *N,
11664 CombineLevel Level) const {
11665 if (Level == BeforeLegalizeTypes)
11668 if (N->getOpcode() != ISD::SHL)
11671 if (Subtarget->isThumb1Only()) {
11672 // Avoid making expensive immediates by commuting shifts. (This logic
11673 // only applies to Thumb1 because ARM and Thumb2 immediates can be shifted
11675 if (N->getOpcode() != ISD::SHL)
11677 SDValue N1 = N->getOperand(0);
11678 if (N1->getOpcode() != ISD::ADD && N1->getOpcode() != ISD::AND &&
11679 N1->getOpcode() != ISD::OR && N1->getOpcode() != ISD::XOR)
11681 if (auto *Const = dyn_cast<ConstantSDNode>(N1->getOperand(1))) {
11682 if (Const->getAPIntValue().ult(256))
11684 if (N1->getOpcode() == ISD::ADD && Const->getAPIntValue().slt(0) &&
11685 Const->getAPIntValue().sgt(-256))
11691 // Turn off commute-with-shift transform after legalization, so it doesn't
11692 // conflict with PerformSHLSimplify. (We could try to detect when
11693 // PerformSHLSimplify would trigger more precisely, but it isn't
11694 // really necessary.)
11698 bool ARMTargetLowering::shouldFoldConstantShiftPairToMask(
11699 const SDNode *N, CombineLevel Level) const {
11700 if (!Subtarget->isThumb1Only())
11703 if (Level == BeforeLegalizeTypes)
11709 bool ARMTargetLowering::preferIncOfAddToSubOfNot(EVT VT) const {
11710 if (!Subtarget->hasNEON()) {
11711 if (Subtarget->isThumb1Only())
11712 return VT.getScalarSizeInBits() <= 32;
11715 return VT.isScalarInteger();
11718 static SDValue PerformSHLSimplify(SDNode *N,
11719 TargetLowering::DAGCombinerInfo &DCI,
11720 const ARMSubtarget *ST) {
11721 // Allow the generic combiner to identify potential bswaps.
11722 if (DCI.isBeforeLegalize())
11725 // DAG combiner will fold:
11726 // (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
11727 // (shl (or x, c1), c2) -> (or (shl x, c2), c1 << c2
11728 // Other code patterns that can be also be modified have the following form:
11729 // b + ((a << 1) | 510)
11730 // b + ((a << 1) & 510)
11731 // b + ((a << 1) ^ 510)
11732 // b + ((a << 1) + 510)
11734 // Many instructions can perform the shift for free, but it requires both
11735 // the operands to be registers. If c1 << c2 is too large, a mov immediate
11736 // instruction will needed. So, unfold back to the original pattern if:
11737 // - if c1 and c2 are small enough that they don't require mov imms.
11738 // - the user(s) of the node can perform an shl
11740 // No shifted operands for 16-bit instructions.
11741 if (ST->isThumb() && ST->isThumb1Only())
11744 // Check that all the users could perform the shl themselves.
11745 for (auto U : N->uses()) {
11746 switch(U->getOpcode()) {
11756 // Check that the user isn't already using a constant because there
11757 // aren't any instructions that support an immediate operand and a
11758 // shifted operand.
11759 if (isa<ConstantSDNode>(U->getOperand(0)) ||
11760 isa<ConstantSDNode>(U->getOperand(1)))
11763 // Check that it's not already using a shift.
11764 if (U->getOperand(0).getOpcode() == ISD::SHL ||
11765 U->getOperand(1).getOpcode() == ISD::SHL)
11771 if (N->getOpcode() != ISD::ADD && N->getOpcode() != ISD::OR &&
11772 N->getOpcode() != ISD::XOR && N->getOpcode() != ISD::AND)
11775 if (N->getOperand(0).getOpcode() != ISD::SHL)
11778 SDValue SHL = N->getOperand(0);
11780 auto *C1ShlC2 = dyn_cast<ConstantSDNode>(N->getOperand(1));
11781 auto *C2 = dyn_cast<ConstantSDNode>(SHL.getOperand(1));
11782 if (!C1ShlC2 || !C2)
11785 APInt C2Int = C2->getAPIntValue();
11786 APInt C1Int = C1ShlC2->getAPIntValue();
11788 // Check that performing a lshr will not lose any information.
11789 APInt Mask = APInt::getHighBitsSet(C2Int.getBitWidth(),
11790 C2Int.getBitWidth() - C2->getZExtValue());
11791 if ((C1Int & Mask) != C1Int)
11794 // Shift the first constant.
11795 C1Int.lshrInPlace(C2Int);
11797 // The immediates are encoded as an 8-bit value that can be rotated.
11798 auto LargeImm = [](const APInt &Imm) {
11799 unsigned Zeros = Imm.countLeadingZeros() + Imm.countTrailingZeros();
11800 return Imm.getBitWidth() - Zeros > 8;
11803 if (LargeImm(C1Int) || LargeImm(C2Int))
11806 SelectionDAG &DAG = DCI.DAG;
11808 SDValue X = SHL.getOperand(0);
11809 SDValue BinOp = DAG.getNode(N->getOpcode(), dl, MVT::i32, X,
11810 DAG.getConstant(C1Int, dl, MVT::i32));
11811 // Shift left to compensate for the lshr of C1Int.
11812 SDValue Res = DAG.getNode(ISD::SHL, dl, MVT::i32, BinOp, SHL.getOperand(1));
11814 LLVM_DEBUG(dbgs() << "Simplify shl use:\n"; SHL.getOperand(0).dump();
11815 SHL.dump(); N->dump());
11816 LLVM_DEBUG(dbgs() << "Into:\n"; X.dump(); BinOp.dump(); Res.dump());
11821 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
11823 static SDValue PerformADDCombine(SDNode *N,
11824 TargetLowering::DAGCombinerInfo &DCI,
11825 const ARMSubtarget *Subtarget) {
11826 SDValue N0 = N->getOperand(0);
11827 SDValue N1 = N->getOperand(1);
11829 // Only works one way, because it needs an immediate operand.
11830 if (SDValue Result = PerformSHLSimplify(N, DCI, Subtarget))
11833 // First try with the default operand order.
11834 if (SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget))
11837 // If that didn't work, try again with the operands commuted.
11838 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
11841 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
11843 static SDValue PerformSUBCombine(SDNode *N,
11844 TargetLowering::DAGCombinerInfo &DCI,
11845 const ARMSubtarget *Subtarget) {
11846 SDValue N0 = N->getOperand(0);
11847 SDValue N1 = N->getOperand(1);
11849 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
11850 if (N1.getNode()->hasOneUse())
11851 if (SDValue Result = combineSelectAndUse(N, N1, N0, DCI))
11854 if (!Subtarget->hasMVEIntegerOps() || !N->getValueType(0).isVector())
11857 // Fold (sub (ARMvmovImm 0), (ARMvdup x)) -> (ARMvdup (sub 0, x))
11858 // so that we can readily pattern match more mve instructions which can use
11859 // a scalar operand.
11860 SDValue VDup = N->getOperand(1);
11861 if (VDup->getOpcode() != ARMISD::VDUP)
11864 SDValue VMov = N->getOperand(0);
11865 if (VMov->getOpcode() == ISD::BITCAST)
11866 VMov = VMov->getOperand(0);
11868 if (VMov->getOpcode() != ARMISD::VMOVIMM || !isZeroVector(VMov))
11872 SDValue Negate = DCI.DAG.getNode(ISD::SUB, dl, MVT::i32,
11873 DCI.DAG.getConstant(0, dl, MVT::i32),
11874 VDup->getOperand(0));
11875 return DCI.DAG.getNode(ARMISD::VDUP, dl, N->getValueType(0), Negate);
11878 /// PerformVMULCombine
11879 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
11880 /// special multiplier accumulator forwarding.
11881 /// vmul d3, d0, d2
11882 /// vmla d3, d1, d2
11884 /// vadd d3, d0, d1
11885 /// vmul d3, d3, d2
11886 // However, for (A + B) * (A + B),
11893 static SDValue PerformVMULCombine(SDNode *N,
11894 TargetLowering::DAGCombinerInfo &DCI,
11895 const ARMSubtarget *Subtarget) {
11896 if (!Subtarget->hasVMLxForwarding())
11899 SelectionDAG &DAG = DCI.DAG;
11900 SDValue N0 = N->getOperand(0);
11901 SDValue N1 = N->getOperand(1);
11902 unsigned Opcode = N0.getOpcode();
11903 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
11904 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
11905 Opcode = N1.getOpcode();
11906 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
11907 Opcode != ISD::FADD && Opcode != ISD::FSUB)
11915 EVT VT = N->getValueType(0);
11917 SDValue N00 = N0->getOperand(0);
11918 SDValue N01 = N0->getOperand(1);
11919 return DAG.getNode(Opcode, DL, VT,
11920 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
11921 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
11924 static SDValue PerformMULCombine(SDNode *N,
11925 TargetLowering::DAGCombinerInfo &DCI,
11926 const ARMSubtarget *Subtarget) {
11927 SelectionDAG &DAG = DCI.DAG;
11929 if (Subtarget->isThumb1Only())
11932 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
11935 EVT VT = N->getValueType(0);
11936 if (VT.is64BitVector() || VT.is128BitVector())
11937 return PerformVMULCombine(N, DCI, Subtarget);
11938 if (VT != MVT::i32)
11941 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
11945 int64_t MulAmt = C->getSExtValue();
11946 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
11948 ShiftAmt = ShiftAmt & (32 - 1);
11949 SDValue V = N->getOperand(0);
11953 MulAmt >>= ShiftAmt;
11956 if (isPowerOf2_32(MulAmt - 1)) {
11957 // (mul x, 2^N + 1) => (add (shl x, N), x)
11958 Res = DAG.getNode(ISD::ADD, DL, VT,
11960 DAG.getNode(ISD::SHL, DL, VT,
11962 DAG.getConstant(Log2_32(MulAmt - 1), DL,
11964 } else if (isPowerOf2_32(MulAmt + 1)) {
11965 // (mul x, 2^N - 1) => (sub (shl x, N), x)
11966 Res = DAG.getNode(ISD::SUB, DL, VT,
11967 DAG.getNode(ISD::SHL, DL, VT,
11969 DAG.getConstant(Log2_32(MulAmt + 1), DL,
11975 uint64_t MulAmtAbs = -MulAmt;
11976 if (isPowerOf2_32(MulAmtAbs + 1)) {
11977 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
11978 Res = DAG.getNode(ISD::SUB, DL, VT,
11980 DAG.getNode(ISD::SHL, DL, VT,
11982 DAG.getConstant(Log2_32(MulAmtAbs + 1), DL,
11984 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
11985 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
11986 Res = DAG.getNode(ISD::ADD, DL, VT,
11988 DAG.getNode(ISD::SHL, DL, VT,
11990 DAG.getConstant(Log2_32(MulAmtAbs - 1), DL,
11992 Res = DAG.getNode(ISD::SUB, DL, VT,
11993 DAG.getConstant(0, DL, MVT::i32), Res);
11999 Res = DAG.getNode(ISD::SHL, DL, VT,
12000 Res, DAG.getConstant(ShiftAmt, DL, MVT::i32));
12002 // Do not add new nodes to DAG combiner worklist.
12003 DCI.CombineTo(N, Res, false);
12007 static SDValue CombineANDShift(SDNode *N,
12008 TargetLowering::DAGCombinerInfo &DCI,
12009 const ARMSubtarget *Subtarget) {
12010 // Allow DAGCombine to pattern-match before we touch the canonical form.
12011 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
12014 if (N->getValueType(0) != MVT::i32)
12017 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1));
12021 uint32_t C1 = (uint32_t)N1C->getZExtValue();
12022 // Don't transform uxtb/uxth.
12023 if (C1 == 255 || C1 == 65535)
12026 SDNode *N0 = N->getOperand(0).getNode();
12027 if (!N0->hasOneUse())
12030 if (N0->getOpcode() != ISD::SHL && N0->getOpcode() != ISD::SRL)
12033 bool LeftShift = N0->getOpcode() == ISD::SHL;
12035 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0->getOperand(1));
12039 uint32_t C2 = (uint32_t)N01C->getZExtValue();
12040 if (!C2 || C2 >= 32)
12043 // Clear irrelevant bits in the mask.
12049 SelectionDAG &DAG = DCI.DAG;
12052 // We have a pattern of the form "(and (shl x, c2) c1)" or
12053 // "(and (srl x, c2) c1)", where c1 is a shifted mask. Try to
12054 // transform to a pair of shifts, to save materializing c1.
12056 // First pattern: right shift, then mask off leading bits.
12057 // FIXME: Use demanded bits?
12058 if (!LeftShift && isMask_32(C1)) {
12059 uint32_t C3 = countLeadingZeros(C1);
12061 SDValue SHL = DAG.getNode(ISD::SHL, DL, MVT::i32, N0->getOperand(0),
12062 DAG.getConstant(C3 - C2, DL, MVT::i32));
12063 return DAG.getNode(ISD::SRL, DL, MVT::i32, SHL,
12064 DAG.getConstant(C3, DL, MVT::i32));
12068 // First pattern, reversed: left shift, then mask off trailing bits.
12069 if (LeftShift && isMask_32(~C1)) {
12070 uint32_t C3 = countTrailingZeros(C1);
12072 SDValue SHL = DAG.getNode(ISD::SRL, DL, MVT::i32, N0->getOperand(0),
12073 DAG.getConstant(C3 - C2, DL, MVT::i32));
12074 return DAG.getNode(ISD::SHL, DL, MVT::i32, SHL,
12075 DAG.getConstant(C3, DL, MVT::i32));
12079 // Second pattern: left shift, then mask off leading bits.
12080 // FIXME: Use demanded bits?
12081 if (LeftShift && isShiftedMask_32(C1)) {
12082 uint32_t Trailing = countTrailingZeros(C1);
12083 uint32_t C3 = countLeadingZeros(C1);
12084 if (Trailing == C2 && C2 + C3 < 32) {
12085 SDValue SHL = DAG.getNode(ISD::SHL, DL, MVT::i32, N0->getOperand(0),
12086 DAG.getConstant(C2 + C3, DL, MVT::i32));
12087 return DAG.getNode(ISD::SRL, DL, MVT::i32, SHL,
12088 DAG.getConstant(C3, DL, MVT::i32));
12092 // Second pattern, reversed: right shift, then mask off trailing bits.
12093 // FIXME: Handle other patterns of known/demanded bits.
12094 if (!LeftShift && isShiftedMask_32(C1)) {
12095 uint32_t Leading = countLeadingZeros(C1);
12096 uint32_t C3 = countTrailingZeros(C1);
12097 if (Leading == C2 && C2 + C3 < 32) {
12098 SDValue SHL = DAG.getNode(ISD::SRL, DL, MVT::i32, N0->getOperand(0),
12099 DAG.getConstant(C2 + C3, DL, MVT::i32));
12100 return DAG.getNode(ISD::SHL, DL, MVT::i32, SHL,
12101 DAG.getConstant(C3, DL, MVT::i32));
12105 // FIXME: Transform "(and (shl x, c2) c1)" ->
12106 // "(shl (and x, c1>>c2), c2)" if "c1 >> c2" is a cheaper immediate than
12111 static SDValue PerformANDCombine(SDNode *N,
12112 TargetLowering::DAGCombinerInfo &DCI,
12113 const ARMSubtarget *Subtarget) {
12114 // Attempt to use immediate-form VBIC
12115 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
12117 EVT VT = N->getValueType(0);
12118 SelectionDAG &DAG = DCI.DAG;
12120 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
12123 APInt SplatBits, SplatUndef;
12124 unsigned SplatBitSize;
12126 if (BVN && Subtarget->hasNEON() &&
12127 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
12128 if (SplatBitSize <= 64) {
12130 SDValue Val = isVMOVModifiedImm((~SplatBits).getZExtValue(),
12131 SplatUndef.getZExtValue(), SplatBitSize,
12132 DAG, dl, VbicVT, VT.is128BitVector(),
12134 if (Val.getNode()) {
12136 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
12137 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
12138 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
12143 if (!Subtarget->isThumb1Only()) {
12144 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
12145 if (SDValue Result = combineSelectAndUseCommutative(N, true, DCI))
12148 if (SDValue Result = PerformSHLSimplify(N, DCI, Subtarget))
12152 if (Subtarget->isThumb1Only())
12153 if (SDValue Result = CombineANDShift(N, DCI, Subtarget))
12159 // Try combining OR nodes to SMULWB, SMULWT.
12160 static SDValue PerformORCombineToSMULWBT(SDNode *OR,
12161 TargetLowering::DAGCombinerInfo &DCI,
12162 const ARMSubtarget *Subtarget) {
12163 if (!Subtarget->hasV6Ops() ||
12164 (Subtarget->isThumb() &&
12165 (!Subtarget->hasThumb2() || !Subtarget->hasDSP())))
12168 SDValue SRL = OR->getOperand(0);
12169 SDValue SHL = OR->getOperand(1);
12171 if (SRL.getOpcode() != ISD::SRL || SHL.getOpcode() != ISD::SHL) {
12172 SRL = OR->getOperand(1);
12173 SHL = OR->getOperand(0);
12175 if (!isSRL16(SRL) || !isSHL16(SHL))
12178 // The first operands to the shifts need to be the two results from the
12179 // same smul_lohi node.
12180 if ((SRL.getOperand(0).getNode() != SHL.getOperand(0).getNode()) ||
12181 SRL.getOperand(0).getOpcode() != ISD::SMUL_LOHI)
12184 SDNode *SMULLOHI = SRL.getOperand(0).getNode();
12185 if (SRL.getOperand(0) != SDValue(SMULLOHI, 0) ||
12186 SHL.getOperand(0) != SDValue(SMULLOHI, 1))
12190 // (or (srl (smul_lohi ?, ?), 16), (shl (smul_lohi ?, ?), 16)))
12191 // For SMUL[B|T] smul_lohi will take a 32-bit and a 16-bit arguments.
12192 // For SMUWB the 16-bit value will signed extended somehow.
12193 // For SMULWT only the SRA is required.
12194 // Check both sides of SMUL_LOHI
12195 SDValue OpS16 = SMULLOHI->getOperand(0);
12196 SDValue OpS32 = SMULLOHI->getOperand(1);
12198 SelectionDAG &DAG = DCI.DAG;
12199 if (!isS16(OpS16, DAG) && !isSRA16(OpS16)) {
12201 OpS32 = SMULLOHI->getOperand(0);
12205 unsigned Opcode = 0;
12206 if (isS16(OpS16, DAG))
12207 Opcode = ARMISD::SMULWB;
12208 else if (isSRA16(OpS16)) {
12209 Opcode = ARMISD::SMULWT;
12210 OpS16 = OpS16->getOperand(0);
12215 SDValue Res = DAG.getNode(Opcode, dl, MVT::i32, OpS32, OpS16);
12216 DAG.ReplaceAllUsesOfValueWith(SDValue(OR, 0), Res);
12217 return SDValue(OR, 0);
12220 static SDValue PerformORCombineToBFI(SDNode *N,
12221 TargetLowering::DAGCombinerInfo &DCI,
12222 const ARMSubtarget *Subtarget) {
12223 // BFI is only available on V6T2+
12224 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
12227 EVT VT = N->getValueType(0);
12228 SDValue N0 = N->getOperand(0);
12229 SDValue N1 = N->getOperand(1);
12230 SelectionDAG &DAG = DCI.DAG;
12232 // 1) or (and A, mask), val => ARMbfi A, val, mask
12233 // iff (val & mask) == val
12235 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
12236 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
12237 // && mask == ~mask2
12238 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
12239 // && ~mask == mask2
12240 // (i.e., copy a bitfield value into another bitfield of the same width)
12242 if (VT != MVT::i32)
12245 SDValue N00 = N0.getOperand(0);
12247 // The value and the mask need to be constants so we can verify this is
12248 // actually a bitfield set. If the mask is 0xffff, we can do better
12249 // via a movt instruction, so don't use BFI in that case.
12250 SDValue MaskOp = N0.getOperand(1);
12251 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
12254 unsigned Mask = MaskC->getZExtValue();
12255 if (Mask == 0xffff)
12258 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
12259 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
12261 unsigned Val = N1C->getZExtValue();
12262 if ((Val & ~Mask) != Val)
12265 if (ARM::isBitFieldInvertedMask(Mask)) {
12266 Val >>= countTrailingZeros(~Mask);
12268 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
12269 DAG.getConstant(Val, DL, MVT::i32),
12270 DAG.getConstant(Mask, DL, MVT::i32));
12272 DCI.CombineTo(N, Res, false);
12273 // Return value from the original node to inform the combiner than N is
12275 return SDValue(N, 0);
12277 } else if (N1.getOpcode() == ISD::AND) {
12278 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
12279 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
12282 unsigned Mask2 = N11C->getZExtValue();
12284 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
12286 if (ARM::isBitFieldInvertedMask(Mask) &&
12287 (Mask == ~Mask2)) {
12288 // The pack halfword instruction works better for masks that fit it,
12289 // so use that when it's available.
12290 if (Subtarget->hasDSP() &&
12291 (Mask == 0xffff || Mask == 0xffff0000))
12294 unsigned amt = countTrailingZeros(Mask2);
12295 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
12296 DAG.getConstant(amt, DL, MVT::i32));
12297 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
12298 DAG.getConstant(Mask, DL, MVT::i32));
12299 DCI.CombineTo(N, Res, false);
12300 // Return value from the original node to inform the combiner than N is
12302 return SDValue(N, 0);
12303 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
12304 (~Mask == Mask2)) {
12305 // The pack halfword instruction works better for masks that fit it,
12306 // so use that when it's available.
12307 if (Subtarget->hasDSP() &&
12308 (Mask2 == 0xffff || Mask2 == 0xffff0000))
12311 unsigned lsb = countTrailingZeros(Mask);
12312 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
12313 DAG.getConstant(lsb, DL, MVT::i32));
12314 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
12315 DAG.getConstant(Mask2, DL, MVT::i32));
12316 DCI.CombineTo(N, Res, false);
12317 // Return value from the original node to inform the combiner than N is
12319 return SDValue(N, 0);
12323 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
12324 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
12325 ARM::isBitFieldInvertedMask(~Mask)) {
12326 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
12327 // where lsb(mask) == #shamt and masked bits of B are known zero.
12328 SDValue ShAmt = N00.getOperand(1);
12329 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
12330 unsigned LSB = countTrailingZeros(Mask);
12334 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
12335 DAG.getConstant(~Mask, DL, MVT::i32));
12337 DCI.CombineTo(N, Res, false);
12338 // Return value from the original node to inform the combiner than N is
12340 return SDValue(N, 0);
12346 static bool isValidMVECond(unsigned CC, bool IsFloat) {
12363 static SDValue PerformORCombine_i1(SDNode *N,
12364 TargetLowering::DAGCombinerInfo &DCI,
12365 const ARMSubtarget *Subtarget) {
12366 // Try to invert "or A, B" -> "and ~A, ~B", as the "and" is easier to chain
12367 // together with predicates
12368 EVT VT = N->getValueType(0);
12369 SDValue N0 = N->getOperand(0);
12370 SDValue N1 = N->getOperand(1);
12372 ARMCC::CondCodes CondCode0 = ARMCC::AL;
12373 ARMCC::CondCodes CondCode1 = ARMCC::AL;
12374 if (N0->getOpcode() == ARMISD::VCMP)
12375 CondCode0 = (ARMCC::CondCodes)cast<const ConstantSDNode>(N0->getOperand(2))
12377 else if (N0->getOpcode() == ARMISD::VCMPZ)
12378 CondCode0 = (ARMCC::CondCodes)cast<const ConstantSDNode>(N0->getOperand(1))
12380 if (N1->getOpcode() == ARMISD::VCMP)
12381 CondCode1 = (ARMCC::CondCodes)cast<const ConstantSDNode>(N1->getOperand(2))
12383 else if (N1->getOpcode() == ARMISD::VCMPZ)
12384 CondCode1 = (ARMCC::CondCodes)cast<const ConstantSDNode>(N1->getOperand(1))
12387 if (CondCode0 == ARMCC::AL || CondCode1 == ARMCC::AL)
12390 unsigned Opposite0 = ARMCC::getOppositeCondition(CondCode0);
12391 unsigned Opposite1 = ARMCC::getOppositeCondition(CondCode1);
12393 if (!isValidMVECond(Opposite0,
12394 N0->getOperand(0)->getValueType(0).isFloatingPoint()) ||
12395 !isValidMVECond(Opposite1,
12396 N1->getOperand(0)->getValueType(0).isFloatingPoint()))
12399 SmallVector<SDValue, 4> Ops0;
12400 Ops0.push_back(N0->getOperand(0));
12401 if (N0->getOpcode() == ARMISD::VCMP)
12402 Ops0.push_back(N0->getOperand(1));
12403 Ops0.push_back(DCI.DAG.getConstant(Opposite0, SDLoc(N0), MVT::i32));
12404 SmallVector<SDValue, 4> Ops1;
12405 Ops1.push_back(N1->getOperand(0));
12406 if (N1->getOpcode() == ARMISD::VCMP)
12407 Ops1.push_back(N1->getOperand(1));
12408 Ops1.push_back(DCI.DAG.getConstant(Opposite1, SDLoc(N1), MVT::i32));
12410 SDValue NewN0 = DCI.DAG.getNode(N0->getOpcode(), SDLoc(N0), VT, Ops0);
12411 SDValue NewN1 = DCI.DAG.getNode(N1->getOpcode(), SDLoc(N1), VT, Ops1);
12412 SDValue And = DCI.DAG.getNode(ISD::AND, SDLoc(N), VT, NewN0, NewN1);
12413 return DCI.DAG.getNode(ISD::XOR, SDLoc(N), VT, And,
12414 DCI.DAG.getAllOnesConstant(SDLoc(N), VT));
12417 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
12418 static SDValue PerformORCombine(SDNode *N,
12419 TargetLowering::DAGCombinerInfo &DCI,
12420 const ARMSubtarget *Subtarget) {
12421 // Attempt to use immediate-form VORR
12422 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
12424 EVT VT = N->getValueType(0);
12425 SelectionDAG &DAG = DCI.DAG;
12427 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
12430 APInt SplatBits, SplatUndef;
12431 unsigned SplatBitSize;
12433 if (BVN && Subtarget->hasNEON() &&
12434 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
12435 if (SplatBitSize <= 64) {
12437 SDValue Val = isVMOVModifiedImm(SplatBits.getZExtValue(),
12438 SplatUndef.getZExtValue(), SplatBitSize,
12439 DAG, dl, VorrVT, VT.is128BitVector(),
12441 if (Val.getNode()) {
12443 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
12444 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
12445 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
12450 if (!Subtarget->isThumb1Only()) {
12451 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
12452 if (SDValue Result = combineSelectAndUseCommutative(N, false, DCI))
12454 if (SDValue Result = PerformORCombineToSMULWBT(N, DCI, Subtarget))
12458 SDValue N0 = N->getOperand(0);
12459 SDValue N1 = N->getOperand(1);
12461 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
12462 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
12463 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
12465 // The code below optimizes (or (and X, Y), Z).
12466 // The AND operand needs to have a single user to make these optimizations
12468 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
12472 unsigned SplatBitSize;
12475 APInt SplatBits0, SplatBits1;
12476 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
12477 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
12478 // Ensure that the second operand of both ands are constants
12479 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
12480 HasAnyUndefs) && !HasAnyUndefs) {
12481 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
12482 HasAnyUndefs) && !HasAnyUndefs) {
12483 // Ensure that the bit width of the constants are the same and that
12484 // the splat arguments are logical inverses as per the pattern we
12485 // are trying to simplify.
12486 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
12487 SplatBits0 == ~SplatBits1) {
12488 // Canonicalize the vector type to make instruction selection
12490 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
12491 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
12494 N1->getOperand(0));
12495 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
12501 if (Subtarget->hasMVEIntegerOps() &&
12502 (VT == MVT::v4i1 || VT == MVT::v8i1 || VT == MVT::v16i1))
12503 return PerformORCombine_i1(N, DCI, Subtarget);
12505 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
12507 if (N0.getOpcode() == ISD::AND && N0.hasOneUse()) {
12508 if (SDValue Res = PerformORCombineToBFI(N, DCI, Subtarget))
12512 if (SDValue Result = PerformSHLSimplify(N, DCI, Subtarget))
12518 static SDValue PerformXORCombine(SDNode *N,
12519 TargetLowering::DAGCombinerInfo &DCI,
12520 const ARMSubtarget *Subtarget) {
12521 EVT VT = N->getValueType(0);
12522 SelectionDAG &DAG = DCI.DAG;
12524 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
12527 if (!Subtarget->isThumb1Only()) {
12528 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
12529 if (SDValue Result = combineSelectAndUseCommutative(N, false, DCI))
12532 if (SDValue Result = PerformSHLSimplify(N, DCI, Subtarget))
12539 // ParseBFI - given a BFI instruction in N, extract the "from" value (Rn) and return it,
12540 // and fill in FromMask and ToMask with (consecutive) bits in "from" to be extracted and
12541 // their position in "to" (Rd).
12542 static SDValue ParseBFI(SDNode *N, APInt &ToMask, APInt &FromMask) {
12543 assert(N->getOpcode() == ARMISD::BFI);
12545 SDValue From = N->getOperand(1);
12546 ToMask = ~cast<ConstantSDNode>(N->getOperand(2))->getAPIntValue();
12547 FromMask = APInt::getLowBitsSet(ToMask.getBitWidth(), ToMask.countPopulation());
12549 // If the Base came from a SHR #C, we can deduce that it is really testing bit
12550 // #C in the base of the SHR.
12551 if (From->getOpcode() == ISD::SRL &&
12552 isa<ConstantSDNode>(From->getOperand(1))) {
12553 APInt Shift = cast<ConstantSDNode>(From->getOperand(1))->getAPIntValue();
12554 assert(Shift.getLimitedValue() < 32 && "Shift too large!");
12555 FromMask <<= Shift.getLimitedValue(31);
12556 From = From->getOperand(0);
12562 // If A and B contain one contiguous set of bits, does A | B == A . B?
12564 // Neither A nor B must be zero.
12565 static bool BitsProperlyConcatenate(const APInt &A, const APInt &B) {
12566 unsigned LastActiveBitInA = A.countTrailingZeros();
12567 unsigned FirstActiveBitInB = B.getBitWidth() - B.countLeadingZeros() - 1;
12568 return LastActiveBitInA - 1 == FirstActiveBitInB;
12571 static SDValue FindBFIToCombineWith(SDNode *N) {
12572 // We have a BFI in N. Follow a possible chain of BFIs and find a BFI it can combine with,
12574 APInt ToMask, FromMask;
12575 SDValue From = ParseBFI(N, ToMask, FromMask);
12576 SDValue To = N->getOperand(0);
12578 // Now check for a compatible BFI to merge with. We can pass through BFIs that
12579 // aren't compatible, but not if they set the same bit in their destination as
12580 // we do (or that of any BFI we're going to combine with).
12582 APInt CombinedToMask = ToMask;
12583 while (V.getOpcode() == ARMISD::BFI) {
12584 APInt NewToMask, NewFromMask;
12585 SDValue NewFrom = ParseBFI(V.getNode(), NewToMask, NewFromMask);
12586 if (NewFrom != From) {
12587 // This BFI has a different base. Keep going.
12588 CombinedToMask |= NewToMask;
12589 V = V.getOperand(0);
12593 // Do the written bits conflict with any we've seen so far?
12594 if ((NewToMask & CombinedToMask).getBoolValue())
12595 // Conflicting bits - bail out because going further is unsafe.
12598 // Are the new bits contiguous when combined with the old bits?
12599 if (BitsProperlyConcatenate(ToMask, NewToMask) &&
12600 BitsProperlyConcatenate(FromMask, NewFromMask))
12602 if (BitsProperlyConcatenate(NewToMask, ToMask) &&
12603 BitsProperlyConcatenate(NewFromMask, FromMask))
12606 // We've seen a write to some bits, so track it.
12607 CombinedToMask |= NewToMask;
12609 V = V.getOperand(0);
12615 static SDValue PerformBFICombine(SDNode *N,
12616 TargetLowering::DAGCombinerInfo &DCI) {
12617 SDValue N1 = N->getOperand(1);
12618 if (N1.getOpcode() == ISD::AND) {
12619 // (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
12620 // the bits being cleared by the AND are not demanded by the BFI.
12621 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
12624 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
12625 unsigned LSB = countTrailingZeros(~InvMask);
12626 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
12628 static_cast<unsigned>(std::numeric_limits<unsigned>::digits) &&
12629 "undefined behavior");
12630 unsigned Mask = (1u << Width) - 1;
12631 unsigned Mask2 = N11C->getZExtValue();
12632 if ((Mask & (~Mask2)) == 0)
12633 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
12634 N->getOperand(0), N1.getOperand(0),
12636 } else if (N->getOperand(0).getOpcode() == ARMISD::BFI) {
12637 // We have a BFI of a BFI. Walk up the BFI chain to see how long it goes.
12638 // Keep track of any consecutive bits set that all come from the same base
12639 // value. We can combine these together into a single BFI.
12640 SDValue CombineBFI = FindBFIToCombineWith(N);
12641 if (CombineBFI == SDValue())
12644 // We've found a BFI.
12645 APInt ToMask1, FromMask1;
12646 SDValue From1 = ParseBFI(N, ToMask1, FromMask1);
12648 APInt ToMask2, FromMask2;
12649 SDValue From2 = ParseBFI(CombineBFI.getNode(), ToMask2, FromMask2);
12650 assert(From1 == From2);
12653 // First, unlink CombineBFI.
12654 DCI.DAG.ReplaceAllUsesWith(CombineBFI, CombineBFI.getOperand(0));
12655 // Then create a new BFI, combining the two together.
12656 APInt NewFromMask = FromMask1 | FromMask2;
12657 APInt NewToMask = ToMask1 | ToMask2;
12659 EVT VT = N->getValueType(0);
12662 if (NewFromMask[0] == 0)
12663 From1 = DCI.DAG.getNode(
12664 ISD::SRL, dl, VT, From1,
12665 DCI.DAG.getConstant(NewFromMask.countTrailingZeros(), dl, VT));
12666 return DCI.DAG.getNode(ARMISD::BFI, dl, VT, N->getOperand(0), From1,
12667 DCI.DAG.getConstant(~NewToMask, dl, VT));
12672 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
12673 /// ARMISD::VMOVRRD.
12674 static SDValue PerformVMOVRRDCombine(SDNode *N,
12675 TargetLowering::DAGCombinerInfo &DCI,
12676 const ARMSubtarget *Subtarget) {
12677 // vmovrrd(vmovdrr x, y) -> x,y
12678 SDValue InDouble = N->getOperand(0);
12679 if (InDouble.getOpcode() == ARMISD::VMOVDRR && Subtarget->hasFP64())
12680 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
12682 // vmovrrd(load f64) -> (load i32), (load i32)
12683 SDNode *InNode = InDouble.getNode();
12684 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
12685 InNode->getValueType(0) == MVT::f64 &&
12686 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
12687 !cast<LoadSDNode>(InNode)->isVolatile()) {
12688 // TODO: Should this be done for non-FrameIndex operands?
12689 LoadSDNode *LD = cast<LoadSDNode>(InNode);
12691 SelectionDAG &DAG = DCI.DAG;
12693 SDValue BasePtr = LD->getBasePtr();
12695 DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr, LD->getPointerInfo(),
12696 LD->getAlignment(), LD->getMemOperand()->getFlags());
12698 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
12699 DAG.getConstant(4, DL, MVT::i32));
12701 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, LD->getChain(), OffsetPtr,
12702 LD->getPointerInfo().getWithOffset(4),
12703 std::min(4U, LD->getAlignment()),
12704 LD->getMemOperand()->getFlags());
12706 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
12707 if (DCI.DAG.getDataLayout().isBigEndian())
12708 std::swap (NewLD1, NewLD2);
12709 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
12716 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
12717 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
12718 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
12719 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
12720 SDValue Op0 = N->getOperand(0);
12721 SDValue Op1 = N->getOperand(1);
12722 if (Op0.getOpcode() == ISD::BITCAST)
12723 Op0 = Op0.getOperand(0);
12724 if (Op1.getOpcode() == ISD::BITCAST)
12725 Op1 = Op1.getOperand(0);
12726 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
12727 Op0.getNode() == Op1.getNode() &&
12728 Op0.getResNo() == 0 && Op1.getResNo() == 1)
12729 return DAG.getNode(ISD::BITCAST, SDLoc(N),
12730 N->getValueType(0), Op0.getOperand(0));
12734 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
12735 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
12736 /// i64 vector to have f64 elements, since the value can then be loaded
12737 /// directly into a VFP register.
12738 static bool hasNormalLoadOperand(SDNode *N) {
12739 unsigned NumElts = N->getValueType(0).getVectorNumElements();
12740 for (unsigned i = 0; i < NumElts; ++i) {
12741 SDNode *Elt = N->getOperand(i).getNode();
12742 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
12748 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
12749 /// ISD::BUILD_VECTOR.
12750 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
12751 TargetLowering::DAGCombinerInfo &DCI,
12752 const ARMSubtarget *Subtarget) {
12753 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
12754 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
12755 // into a pair of GPRs, which is fine when the value is used as a scalar,
12756 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
12757 SelectionDAG &DAG = DCI.DAG;
12758 if (N->getNumOperands() == 2)
12759 if (SDValue RV = PerformVMOVDRRCombine(N, DAG))
12762 // Load i64 elements as f64 values so that type legalization does not split
12763 // them up into i32 values.
12764 EVT VT = N->getValueType(0);
12765 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
12768 SmallVector<SDValue, 8> Ops;
12769 unsigned NumElts = VT.getVectorNumElements();
12770 for (unsigned i = 0; i < NumElts; ++i) {
12771 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
12773 // Make the DAGCombiner fold the bitcast.
12774 DCI.AddToWorklist(V.getNode());
12776 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
12777 SDValue BV = DAG.getBuildVector(FloatVT, dl, Ops);
12778 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
12781 /// Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
12783 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
12784 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
12785 // At that time, we may have inserted bitcasts from integer to float.
12786 // If these bitcasts have survived DAGCombine, change the lowering of this
12787 // BUILD_VECTOR in something more vector friendly, i.e., that does not
12788 // force to use floating point types.
12790 // Make sure we can change the type of the vector.
12791 // This is possible iff:
12792 // 1. The vector is only used in a bitcast to a integer type. I.e.,
12793 // 1.1. Vector is used only once.
12794 // 1.2. Use is a bit convert to an integer type.
12795 // 2. The size of its operands are 32-bits (64-bits are not legal).
12796 EVT VT = N->getValueType(0);
12797 EVT EltVT = VT.getVectorElementType();
12799 // Check 1.1. and 2.
12800 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
12803 // By construction, the input type must be float.
12804 assert(EltVT == MVT::f32 && "Unexpected type!");
12807 SDNode *Use = *N->use_begin();
12808 if (Use->getOpcode() != ISD::BITCAST ||
12809 Use->getValueType(0).isFloatingPoint())
12812 // Check profitability.
12813 // Model is, if more than half of the relevant operands are bitcast from
12814 // i32, turn the build_vector into a sequence of insert_vector_elt.
12815 // Relevant operands are everything that is not statically
12816 // (i.e., at compile time) bitcasted.
12817 unsigned NumOfBitCastedElts = 0;
12818 unsigned NumElts = VT.getVectorNumElements();
12819 unsigned NumOfRelevantElts = NumElts;
12820 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
12821 SDValue Elt = N->getOperand(Idx);
12822 if (Elt->getOpcode() == ISD::BITCAST) {
12823 // Assume only bit cast to i32 will go away.
12824 if (Elt->getOperand(0).getValueType() == MVT::i32)
12825 ++NumOfBitCastedElts;
12826 } else if (Elt.isUndef() || isa<ConstantSDNode>(Elt))
12827 // Constants are statically casted, thus do not count them as
12828 // relevant operands.
12829 --NumOfRelevantElts;
12832 // Check if more than half of the elements require a non-free bitcast.
12833 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
12836 SelectionDAG &DAG = DCI.DAG;
12837 // Create the new vector type.
12838 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
12839 // Check if the type is legal.
12840 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12841 if (!TLI.isTypeLegal(VecVT))
12845 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
12846 // => BITCAST INSERT_VECTOR_ELT
12847 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
12848 // (BITCAST EN), N.
12849 SDValue Vec = DAG.getUNDEF(VecVT);
12851 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
12852 SDValue V = N->getOperand(Idx);
12855 if (V.getOpcode() == ISD::BITCAST &&
12856 V->getOperand(0).getValueType() == MVT::i32)
12857 // Fold obvious case.
12858 V = V.getOperand(0);
12860 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
12861 // Make the DAGCombiner fold the bitcasts.
12862 DCI.AddToWorklist(V.getNode());
12864 SDValue LaneIdx = DAG.getConstant(Idx, dl, MVT::i32);
12865 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
12867 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
12868 // Make the DAGCombiner fold the bitcasts.
12869 DCI.AddToWorklist(Vec.getNode());
12874 PerformPREDICATE_CASTCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
12875 EVT VT = N->getValueType(0);
12876 SDValue Op = N->getOperand(0);
12879 // PREDICATE_CAST(PREDICATE_CAST(x)) == PREDICATE_CAST(x)
12880 if (Op->getOpcode() == ARMISD::PREDICATE_CAST) {
12881 // If the valuetypes are the same, we can remove the cast entirely.
12882 if (Op->getOperand(0).getValueType() == VT)
12883 return Op->getOperand(0);
12884 return DCI.DAG.getNode(ARMISD::PREDICATE_CAST, dl,
12885 Op->getOperand(0).getValueType(), Op->getOperand(0));
12891 static SDValue PerformVCMPCombine(SDNode *N,
12892 TargetLowering::DAGCombinerInfo &DCI,
12893 const ARMSubtarget *Subtarget) {
12894 if (!Subtarget->hasMVEIntegerOps())
12897 EVT VT = N->getValueType(0);
12898 SDValue Op0 = N->getOperand(0);
12899 SDValue Op1 = N->getOperand(1);
12900 ARMCC::CondCodes Cond =
12901 (ARMCC::CondCodes)cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
12904 // vcmp X, 0, cc -> vcmpz X, cc
12905 if (isZeroVector(Op1))
12906 return DCI.DAG.getNode(ARMISD::VCMPZ, dl, VT, Op0,
12909 unsigned SwappedCond = getSwappedCondition(Cond);
12910 if (isValidMVECond(SwappedCond, VT.isFloatingPoint())) {
12911 // vcmp 0, X, cc -> vcmpz X, reversed(cc)
12912 if (isZeroVector(Op0))
12913 return DCI.DAG.getNode(ARMISD::VCMPZ, dl, VT, Op1,
12914 DCI.DAG.getConstant(SwappedCond, dl, MVT::i32));
12915 // vcmp vdup(Y), X, cc -> vcmp X, vdup(Y), reversed(cc)
12916 if (Op0->getOpcode() == ARMISD::VDUP && Op1->getOpcode() != ARMISD::VDUP)
12917 return DCI.DAG.getNode(ARMISD::VCMP, dl, VT, Op1, Op0,
12918 DCI.DAG.getConstant(SwappedCond, dl, MVT::i32));
12924 /// PerformInsertEltCombine - Target-specific dag combine xforms for
12925 /// ISD::INSERT_VECTOR_ELT.
12926 static SDValue PerformInsertEltCombine(SDNode *N,
12927 TargetLowering::DAGCombinerInfo &DCI) {
12928 // Bitcast an i64 load inserted into a vector to f64.
12929 // Otherwise, the i64 value will be legalized to a pair of i32 values.
12930 EVT VT = N->getValueType(0);
12931 SDNode *Elt = N->getOperand(1).getNode();
12932 if (VT.getVectorElementType() != MVT::i64 ||
12933 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
12936 SelectionDAG &DAG = DCI.DAG;
12938 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
12939 VT.getVectorNumElements());
12940 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
12941 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
12942 // Make the DAGCombiner fold the bitcasts.
12943 DCI.AddToWorklist(Vec.getNode());
12944 DCI.AddToWorklist(V.getNode());
12945 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
12946 Vec, V, N->getOperand(2));
12947 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
12950 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
12951 /// ISD::VECTOR_SHUFFLE.
12952 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
12953 // The LLVM shufflevector instruction does not require the shuffle mask
12954 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
12955 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
12956 // operands do not match the mask length, they are extended by concatenating
12957 // them with undef vectors. That is probably the right thing for other
12958 // targets, but for NEON it is better to concatenate two double-register
12959 // size vector operands into a single quad-register size vector. Do that
12960 // transformation here:
12961 // shuffle(concat(v1, undef), concat(v2, undef)) ->
12962 // shuffle(concat(v1, v2), undef)
12963 SDValue Op0 = N->getOperand(0);
12964 SDValue Op1 = N->getOperand(1);
12965 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
12966 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
12967 Op0.getNumOperands() != 2 ||
12968 Op1.getNumOperands() != 2)
12970 SDValue Concat0Op1 = Op0.getOperand(1);
12971 SDValue Concat1Op1 = Op1.getOperand(1);
12972 if (!Concat0Op1.isUndef() || !Concat1Op1.isUndef())
12974 // Skip the transformation if any of the types are illegal.
12975 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12976 EVT VT = N->getValueType(0);
12977 if (!TLI.isTypeLegal(VT) ||
12978 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
12979 !TLI.isTypeLegal(Concat1Op1.getValueType()))
12982 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
12983 Op0.getOperand(0), Op1.getOperand(0));
12984 // Translate the shuffle mask.
12985 SmallVector<int, 16> NewMask;
12986 unsigned NumElts = VT.getVectorNumElements();
12987 unsigned HalfElts = NumElts/2;
12988 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
12989 for (unsigned n = 0; n < NumElts; ++n) {
12990 int MaskElt = SVN->getMaskElt(n);
12992 if (MaskElt < (int)HalfElts)
12994 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
12995 NewElt = HalfElts + MaskElt - NumElts;
12996 NewMask.push_back(NewElt);
12998 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
12999 DAG.getUNDEF(VT), NewMask);
13002 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP,
13003 /// NEON load/store intrinsics, and generic vector load/stores, to merge
13004 /// base address updates.
13005 /// For generic load/stores, the memory type is assumed to be a vector.
13006 /// The caller is assumed to have checked legality.
13007 static SDValue CombineBaseUpdate(SDNode *N,
13008 TargetLowering::DAGCombinerInfo &DCI) {
13009 SelectionDAG &DAG = DCI.DAG;
13010 const bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
13011 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
13012 const bool isStore = N->getOpcode() == ISD::STORE;
13013 const unsigned AddrOpIdx = ((isIntrinsic || isStore) ? 2 : 1);
13014 SDValue Addr = N->getOperand(AddrOpIdx);
13015 MemSDNode *MemN = cast<MemSDNode>(N);
13018 // Search for a use of the address operand that is an increment.
13019 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
13020 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
13021 SDNode *User = *UI;
13022 if (User->getOpcode() != ISD::ADD ||
13023 UI.getUse().getResNo() != Addr.getResNo())
13026 // Check that the add is independent of the load/store. Otherwise, folding
13027 // it would create a cycle. We can avoid searching through Addr as it's a
13028 // predecessor to both.
13029 SmallPtrSet<const SDNode *, 32> Visited;
13030 SmallVector<const SDNode *, 16> Worklist;
13031 Visited.insert(Addr.getNode());
13032 Worklist.push_back(N);
13033 Worklist.push_back(User);
13034 if (SDNode::hasPredecessorHelper(N, Visited, Worklist) ||
13035 SDNode::hasPredecessorHelper(User, Visited, Worklist))
13038 // Find the new opcode for the updating load/store.
13039 bool isLoadOp = true;
13040 bool isLaneOp = false;
13041 unsigned NewOpc = 0;
13042 unsigned NumVecs = 0;
13044 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
13046 default: llvm_unreachable("unexpected intrinsic for Neon base update");
13047 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
13048 NumVecs = 1; break;
13049 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
13050 NumVecs = 2; break;
13051 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
13052 NumVecs = 3; break;
13053 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
13054 NumVecs = 4; break;
13055 case Intrinsic::arm_neon_vld2dup:
13056 case Intrinsic::arm_neon_vld3dup:
13057 case Intrinsic::arm_neon_vld4dup:
13058 // TODO: Support updating VLDxDUP nodes. For now, we just skip
13059 // combining base updates for such intrinsics.
13061 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
13062 NumVecs = 2; isLaneOp = true; break;
13063 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
13064 NumVecs = 3; isLaneOp = true; break;
13065 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
13066 NumVecs = 4; isLaneOp = true; break;
13067 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
13068 NumVecs = 1; isLoadOp = false; break;
13069 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
13070 NumVecs = 2; isLoadOp = false; break;
13071 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
13072 NumVecs = 3; isLoadOp = false; break;
13073 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
13074 NumVecs = 4; isLoadOp = false; break;
13075 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
13076 NumVecs = 2; isLoadOp = false; isLaneOp = true; break;
13077 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
13078 NumVecs = 3; isLoadOp = false; isLaneOp = true; break;
13079 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
13080 NumVecs = 4; isLoadOp = false; isLaneOp = true; break;
13084 switch (N->getOpcode()) {
13085 default: llvm_unreachable("unexpected opcode for Neon base update");
13086 case ARMISD::VLD1DUP: NewOpc = ARMISD::VLD1DUP_UPD; NumVecs = 1; break;
13087 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
13088 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
13089 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
13090 case ISD::LOAD: NewOpc = ARMISD::VLD1_UPD;
13091 NumVecs = 1; isLaneOp = false; break;
13092 case ISD::STORE: NewOpc = ARMISD::VST1_UPD;
13093 NumVecs = 1; isLaneOp = false; isLoadOp = false; break;
13097 // Find the size of memory referenced by the load/store.
13100 VecTy = N->getValueType(0);
13101 } else if (isIntrinsic) {
13102 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
13104 assert(isStore && "Node has to be a load, a store, or an intrinsic!");
13105 VecTy = N->getOperand(1).getValueType();
13108 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
13110 NumBytes /= VecTy.getVectorNumElements();
13112 // If the increment is a constant, it must match the memory ref size.
13113 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
13114 ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode());
13115 if (NumBytes >= 3 * 16 && (!CInc || CInc->getZExtValue() != NumBytes)) {
13116 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
13117 // separate instructions that make it harder to use a non-constant update.
13121 // OK, we found an ADD we can fold into the base update.
13122 // Now, create a _UPD node, taking care of not breaking alignment.
13124 EVT AlignedVecTy = VecTy;
13125 unsigned Alignment = MemN->getAlignment();
13127 // If this is a less-than-standard-aligned load/store, change the type to
13128 // match the standard alignment.
13129 // The alignment is overlooked when selecting _UPD variants; and it's
13130 // easier to introduce bitcasts here than fix that.
13131 // There are 3 ways to get to this base-update combine:
13132 // - intrinsics: they are assumed to be properly aligned (to the standard
13133 // alignment of the memory type), so we don't need to do anything.
13134 // - ARMISD::VLDx nodes: they are only generated from the aforementioned
13135 // intrinsics, so, likewise, there's nothing to do.
13136 // - generic load/store instructions: the alignment is specified as an
13137 // explicit operand, rather than implicitly as the standard alignment
13138 // of the memory type (like the intrisics). We need to change the
13139 // memory type to match the explicit alignment. That way, we don't
13140 // generate non-standard-aligned ARMISD::VLDx nodes.
13141 if (isa<LSBaseSDNode>(N)) {
13142 if (Alignment == 0)
13144 if (Alignment < VecTy.getScalarSizeInBits() / 8) {
13145 MVT EltTy = MVT::getIntegerVT(Alignment * 8);
13146 assert(NumVecs == 1 && "Unexpected multi-element generic load/store.");
13147 assert(!isLaneOp && "Unexpected generic load/store lane.");
13148 unsigned NumElts = NumBytes / (EltTy.getSizeInBits() / 8);
13149 AlignedVecTy = MVT::getVectorVT(EltTy, NumElts);
13151 // Don't set an explicit alignment on regular load/stores that we want
13152 // to transform to VLD/VST 1_UPD nodes.
13153 // This matches the behavior of regular load/stores, which only get an
13154 // explicit alignment if the MMO alignment is larger than the standard
13155 // alignment of the memory type.
13156 // Intrinsics, however, always get an explicit alignment, set to the
13157 // alignment of the MMO.
13161 // Create the new updating load/store node.
13162 // First, create an SDVTList for the new updating node's results.
13164 unsigned NumResultVecs = (isLoadOp ? NumVecs : 0);
13166 for (n = 0; n < NumResultVecs; ++n)
13167 Tys[n] = AlignedVecTy;
13168 Tys[n++] = MVT::i32;
13169 Tys[n] = MVT::Other;
13170 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs+2));
13172 // Then, gather the new node's operands.
13173 SmallVector<SDValue, 8> Ops;
13174 Ops.push_back(N->getOperand(0)); // incoming chain
13175 Ops.push_back(N->getOperand(AddrOpIdx));
13176 Ops.push_back(Inc);
13178 if (StoreSDNode *StN = dyn_cast<StoreSDNode>(N)) {
13179 // Try to match the intrinsic's signature
13180 Ops.push_back(StN->getValue());
13182 // Loads (and of course intrinsics) match the intrinsics' signature,
13183 // so just add all but the alignment operand.
13184 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands() - 1; ++i)
13185 Ops.push_back(N->getOperand(i));
13188 // For all node types, the alignment operand is always the last one.
13189 Ops.push_back(DAG.getConstant(Alignment, dl, MVT::i32));
13191 // If this is a non-standard-aligned STORE, the penultimate operand is the
13192 // stored value. Bitcast it to the aligned type.
13193 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::STORE) {
13194 SDValue &StVal = Ops[Ops.size()-2];
13195 StVal = DAG.getNode(ISD::BITCAST, dl, AlignedVecTy, StVal);
13198 EVT LoadVT = isLaneOp ? VecTy.getVectorElementType() : AlignedVecTy;
13199 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, dl, SDTys, Ops, LoadVT,
13200 MemN->getMemOperand());
13202 // Update the uses.
13203 SmallVector<SDValue, 5> NewResults;
13204 for (unsigned i = 0; i < NumResultVecs; ++i)
13205 NewResults.push_back(SDValue(UpdN.getNode(), i));
13207 // If this is an non-standard-aligned LOAD, the first result is the loaded
13208 // value. Bitcast it to the expected result type.
13209 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::LOAD) {
13210 SDValue &LdVal = NewResults[0];
13211 LdVal = DAG.getNode(ISD::BITCAST, dl, VecTy, LdVal);
13214 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
13215 DCI.CombineTo(N, NewResults);
13216 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
13223 static SDValue PerformVLDCombine(SDNode *N,
13224 TargetLowering::DAGCombinerInfo &DCI) {
13225 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
13228 return CombineBaseUpdate(N, DCI);
13231 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
13232 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
13233 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
13235 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
13236 SelectionDAG &DAG = DCI.DAG;
13237 EVT VT = N->getValueType(0);
13238 // vldN-dup instructions only support 64-bit vectors for N > 1.
13239 if (!VT.is64BitVector())
13242 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
13243 SDNode *VLD = N->getOperand(0).getNode();
13244 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
13246 unsigned NumVecs = 0;
13247 unsigned NewOpc = 0;
13248 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
13249 if (IntNo == Intrinsic::arm_neon_vld2lane) {
13251 NewOpc = ARMISD::VLD2DUP;
13252 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
13254 NewOpc = ARMISD::VLD3DUP;
13255 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
13257 NewOpc = ARMISD::VLD4DUP;
13262 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
13263 // numbers match the load.
13264 unsigned VLDLaneNo =
13265 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
13266 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
13268 // Ignore uses of the chain result.
13269 if (UI.getUse().getResNo() == NumVecs)
13271 SDNode *User = *UI;
13272 if (User->getOpcode() != ARMISD::VDUPLANE ||
13273 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
13277 // Create the vldN-dup node.
13280 for (n = 0; n < NumVecs; ++n)
13282 Tys[n] = MVT::Other;
13283 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumVecs+1));
13284 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
13285 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
13286 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
13287 Ops, VLDMemInt->getMemoryVT(),
13288 VLDMemInt->getMemOperand());
13290 // Update the uses.
13291 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
13293 unsigned ResNo = UI.getUse().getResNo();
13294 // Ignore uses of the chain result.
13295 if (ResNo == NumVecs)
13297 SDNode *User = *UI;
13298 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
13301 // Now the vldN-lane intrinsic is dead except for its chain result.
13302 // Update uses of the chain.
13303 std::vector<SDValue> VLDDupResults;
13304 for (unsigned n = 0; n < NumVecs; ++n)
13305 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
13306 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
13307 DCI.CombineTo(VLD, VLDDupResults);
13312 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
13313 /// ARMISD::VDUPLANE.
13314 static SDValue PerformVDUPLANECombine(SDNode *N,
13315 TargetLowering::DAGCombinerInfo &DCI) {
13316 SDValue Op = N->getOperand(0);
13318 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
13319 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
13320 if (CombineVLDDUP(N, DCI))
13321 return SDValue(N, 0);
13323 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
13324 // redundant. Ignore bit_converts for now; element sizes are checked below.
13325 while (Op.getOpcode() == ISD::BITCAST)
13326 Op = Op.getOperand(0);
13327 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
13330 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
13331 unsigned EltSize = Op.getScalarValueSizeInBits();
13332 // The canonical VMOV for a zero vector uses a 32-bit element size.
13333 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
13335 if (ARM_AM::decodeVMOVModImm(Imm, EltBits) == 0)
13337 EVT VT = N->getValueType(0);
13338 if (EltSize > VT.getScalarSizeInBits())
13341 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
13344 /// PerformVDUPCombine - Target-specific dag combine xforms for ARMISD::VDUP.
13345 static SDValue PerformVDUPCombine(SDNode *N,
13346 TargetLowering::DAGCombinerInfo &DCI,
13347 const ARMSubtarget *Subtarget) {
13348 SelectionDAG &DAG = DCI.DAG;
13349 SDValue Op = N->getOperand(0);
13351 if (!Subtarget->hasNEON())
13354 // Match VDUP(LOAD) -> VLD1DUP.
13355 // We match this pattern here rather than waiting for isel because the
13356 // transform is only legal for unindexed loads.
13357 LoadSDNode *LD = dyn_cast<LoadSDNode>(Op.getNode());
13358 if (LD && Op.hasOneUse() && LD->isUnindexed() &&
13359 LD->getMemoryVT() == N->getValueType(0).getVectorElementType()) {
13360 SDValue Ops[] = { LD->getOperand(0), LD->getOperand(1),
13361 DAG.getConstant(LD->getAlignment(), SDLoc(N), MVT::i32) };
13362 SDVTList SDTys = DAG.getVTList(N->getValueType(0), MVT::Other);
13363 SDValue VLDDup = DAG.getMemIntrinsicNode(ARMISD::VLD1DUP, SDLoc(N), SDTys,
13364 Ops, LD->getMemoryVT(),
13365 LD->getMemOperand());
13366 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), VLDDup.getValue(1));
13373 static SDValue PerformLOADCombine(SDNode *N,
13374 TargetLowering::DAGCombinerInfo &DCI) {
13375 EVT VT = N->getValueType(0);
13377 // If this is a legal vector load, try to combine it into a VLD1_UPD.
13378 if (ISD::isNormalLoad(N) && VT.isVector() &&
13379 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
13380 return CombineBaseUpdate(N, DCI);
13385 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
13386 // pack all of the elements in one place. Next, store to memory in fewer
13388 static SDValue PerformTruncatingStoreCombine(StoreSDNode *St,
13389 SelectionDAG &DAG) {
13390 SDValue StVal = St->getValue();
13391 EVT VT = StVal.getValueType();
13392 if (!St->isTruncatingStore() || !VT.isVector())
13394 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13395 EVT StVT = St->getMemoryVT();
13396 unsigned NumElems = VT.getVectorNumElements();
13397 assert(StVT != VT && "Cannot truncate to the same type");
13398 unsigned FromEltSz = VT.getScalarSizeInBits();
13399 unsigned ToEltSz = StVT.getScalarSizeInBits();
13401 // From, To sizes and ElemCount must be pow of two
13402 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz))
13405 // We are going to use the original vector elt for storing.
13406 // Accumulated smaller vector elements must be a multiple of the store size.
13407 if (0 != (NumElems * FromEltSz) % ToEltSz)
13410 unsigned SizeRatio = FromEltSz / ToEltSz;
13411 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
13413 // Create a type on which we perform the shuffle.
13414 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
13415 NumElems * SizeRatio);
13416 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
13419 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
13420 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
13421 for (unsigned i = 0; i < NumElems; ++i)
13422 ShuffleVec[i] = DAG.getDataLayout().isBigEndian() ? (i + 1) * SizeRatio - 1
13425 // Can't shuffle using an illegal type.
13426 if (!TLI.isTypeLegal(WideVecVT))
13429 SDValue Shuff = DAG.getVectorShuffle(
13430 WideVecVT, DL, WideVec, DAG.getUNDEF(WideVec.getValueType()), ShuffleVec);
13431 // At this point all of the data is stored at the bottom of the
13432 // register. We now need to save it to mem.
13434 // Find the largest store unit
13435 MVT StoreType = MVT::i8;
13436 for (MVT Tp : MVT::integer_valuetypes()) {
13437 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
13440 // Didn't find a legal store type.
13441 if (!TLI.isTypeLegal(StoreType))
13444 // Bitcast the original vector into a vector of store-size units
13446 EVT::getVectorVT(*DAG.getContext(), StoreType,
13447 VT.getSizeInBits() / EVT(StoreType).getSizeInBits());
13448 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
13449 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
13450 SmallVector<SDValue, 8> Chains;
13451 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, DL,
13452 TLI.getPointerTy(DAG.getDataLayout()));
13453 SDValue BasePtr = St->getBasePtr();
13455 // Perform one or more big stores into memory.
13456 unsigned E = (ToEltSz * NumElems) / StoreType.getSizeInBits();
13457 for (unsigned I = 0; I < E; I++) {
13458 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, StoreType,
13459 ShuffWide, DAG.getIntPtrConstant(I, DL));
13461 DAG.getStore(St->getChain(), DL, SubVec, BasePtr, St->getPointerInfo(),
13462 St->getAlignment(), St->getMemOperand()->getFlags());
13464 DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr, Increment);
13465 Chains.push_back(Ch);
13467 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
13470 // Try taking a single vector store from an truncate (which would otherwise turn
13471 // into an expensive buildvector) and splitting it into a series of narrowing
13473 static SDValue PerformSplittingToNarrowingStores(StoreSDNode *St,
13474 SelectionDAG &DAG) {
13475 if (!St->isSimple() || St->isTruncatingStore() || !St->isUnindexed())
13477 SDValue Trunc = St->getValue();
13478 if (Trunc->getOpcode() != ISD::TRUNCATE)
13480 EVT FromVT = Trunc->getOperand(0).getValueType();
13481 EVT ToVT = Trunc.getValueType();
13482 if (!ToVT.isVector())
13484 assert(FromVT.getVectorNumElements() == ToVT.getVectorNumElements());
13485 EVT ToEltVT = ToVT.getVectorElementType();
13486 EVT FromEltVT = FromVT.getVectorElementType();
13488 unsigned NumElements = 0;
13489 if (FromEltVT == MVT::i32 && (ToEltVT == MVT::i16 || ToEltVT == MVT::i8))
13491 if (FromEltVT == MVT::i16 && ToEltVT == MVT::i8)
13493 if (NumElements == 0 || FromVT.getVectorNumElements() == NumElements ||
13494 FromVT.getVectorNumElements() % NumElements != 0)
13498 // Details about the old store
13499 SDValue Ch = St->getChain();
13500 SDValue BasePtr = St->getBasePtr();
13501 unsigned Alignment = St->getOriginalAlignment();
13502 MachineMemOperand::Flags MMOFlags = St->getMemOperand()->getFlags();
13503 AAMDNodes AAInfo = St->getAAInfo();
13505 EVT NewFromVT = EVT::getVectorVT(*DAG.getContext(), FromEltVT, NumElements);
13506 EVT NewToVT = EVT::getVectorVT(*DAG.getContext(), ToEltVT, NumElements);
13508 SmallVector<SDValue, 4> Stores;
13509 for (unsigned i = 0; i < FromVT.getVectorNumElements() / NumElements; i++) {
13510 unsigned NewOffset = i * NumElements * ToEltVT.getSizeInBits() / 8;
13511 SDValue NewPtr = DAG.getObjectPtrOffset(DL, BasePtr, NewOffset);
13514 DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, NewFromVT, Trunc.getOperand(0),
13515 DAG.getConstant(i * NumElements, DL, MVT::i32));
13516 SDValue Store = DAG.getTruncStore(
13517 Ch, DL, Extract, NewPtr, St->getPointerInfo().getWithOffset(NewOffset),
13518 NewToVT, Alignment, MMOFlags, AAInfo);
13519 Stores.push_back(Store);
13521 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Stores);
13524 /// PerformSTORECombine - Target-specific dag combine xforms for
13526 static SDValue PerformSTORECombine(SDNode *N,
13527 TargetLowering::DAGCombinerInfo &DCI,
13528 const ARMSubtarget *Subtarget) {
13529 StoreSDNode *St = cast<StoreSDNode>(N);
13530 if (St->isVolatile())
13532 SDValue StVal = St->getValue();
13533 EVT VT = StVal.getValueType();
13535 if (Subtarget->hasNEON())
13536 if (SDValue Store = PerformTruncatingStoreCombine(St, DCI.DAG))
13539 if (Subtarget->hasMVEIntegerOps())
13540 if (SDValue NewToken = PerformSplittingToNarrowingStores(St, DCI.DAG))
13543 if (!ISD::isNormalStore(St))
13546 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
13547 // ARM stores of arguments in the same cache line.
13548 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
13549 StVal.getNode()->hasOneUse()) {
13550 SelectionDAG &DAG = DCI.DAG;
13551 bool isBigEndian = DAG.getDataLayout().isBigEndian();
13553 SDValue BasePtr = St->getBasePtr();
13554 SDValue NewST1 = DAG.getStore(
13555 St->getChain(), DL, StVal.getNode()->getOperand(isBigEndian ? 1 : 0),
13556 BasePtr, St->getPointerInfo(), St->getAlignment(),
13557 St->getMemOperand()->getFlags());
13559 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
13560 DAG.getConstant(4, DL, MVT::i32));
13561 return DAG.getStore(NewST1.getValue(0), DL,
13562 StVal.getNode()->getOperand(isBigEndian ? 0 : 1),
13563 OffsetPtr, St->getPointerInfo(),
13564 std::min(4U, St->getAlignment() / 2),
13565 St->getMemOperand()->getFlags());
13568 if (StVal.getValueType() == MVT::i64 &&
13569 StVal.getNode()->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
13571 // Bitcast an i64 store extracted from a vector to f64.
13572 // Otherwise, the i64 value will be legalized to a pair of i32 values.
13573 SelectionDAG &DAG = DCI.DAG;
13575 SDValue IntVec = StVal.getOperand(0);
13576 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
13577 IntVec.getValueType().getVectorNumElements());
13578 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
13579 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
13580 Vec, StVal.getOperand(1));
13582 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
13583 // Make the DAGCombiner fold the bitcasts.
13584 DCI.AddToWorklist(Vec.getNode());
13585 DCI.AddToWorklist(ExtElt.getNode());
13586 DCI.AddToWorklist(V.getNode());
13587 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
13588 St->getPointerInfo(), St->getAlignment(),
13589 St->getMemOperand()->getFlags(), St->getAAInfo());
13592 // If this is a legal vector store, try to combine it into a VST1_UPD.
13593 if (Subtarget->hasNEON() && ISD::isNormalStore(N) && VT.isVector() &&
13594 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
13595 return CombineBaseUpdate(N, DCI);
13600 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
13601 /// can replace combinations of VMUL and VCVT (floating-point to integer)
13602 /// when the VMUL has a constant operand that is a power of 2.
13604 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
13605 /// vmul.f32 d16, d17, d16
13606 /// vcvt.s32.f32 d16, d16
13608 /// vcvt.s32.f32 d16, d16, #3
13609 static SDValue PerformVCVTCombine(SDNode *N, SelectionDAG &DAG,
13610 const ARMSubtarget *Subtarget) {
13611 if (!Subtarget->hasNEON())
13614 SDValue Op = N->getOperand(0);
13615 if (!Op.getValueType().isVector() || !Op.getValueType().isSimple() ||
13616 Op.getOpcode() != ISD::FMUL)
13619 SDValue ConstVec = Op->getOperand(1);
13620 if (!isa<BuildVectorSDNode>(ConstVec))
13623 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
13624 uint32_t FloatBits = FloatTy.getSizeInBits();
13625 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
13626 uint32_t IntBits = IntTy.getSizeInBits();
13627 unsigned NumLanes = Op.getValueType().getVectorNumElements();
13628 if (FloatBits != 32 || IntBits > 32 || (NumLanes != 4 && NumLanes != 2)) {
13629 // These instructions only exist converting from f32 to i32. We can handle
13630 // smaller integers by generating an extra truncate, but larger ones would
13631 // be lossy. We also can't handle anything other than 2 or 4 lanes, since
13632 // these intructions only support v2i32/v4i32 types.
13636 BitVector UndefElements;
13637 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(ConstVec);
13638 int32_t C = BV->getConstantFPSplatPow2ToLog2Int(&UndefElements, 33);
13639 if (C == -1 || C == 0 || C > 32)
13643 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
13644 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
13645 Intrinsic::arm_neon_vcvtfp2fxu;
13646 SDValue FixConv = DAG.getNode(
13647 ISD::INTRINSIC_WO_CHAIN, dl, NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
13648 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32), Op->getOperand(0),
13649 DAG.getConstant(C, dl, MVT::i32));
13651 if (IntBits < FloatBits)
13652 FixConv = DAG.getNode(ISD::TRUNCATE, dl, N->getValueType(0), FixConv);
13657 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
13658 /// can replace combinations of VCVT (integer to floating-point) and VDIV
13659 /// when the VDIV has a constant operand that is a power of 2.
13661 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
13662 /// vcvt.f32.s32 d16, d16
13663 /// vdiv.f32 d16, d17, d16
13665 /// vcvt.f32.s32 d16, d16, #3
13666 static SDValue PerformVDIVCombine(SDNode *N, SelectionDAG &DAG,
13667 const ARMSubtarget *Subtarget) {
13668 if (!Subtarget->hasNEON())
13671 SDValue Op = N->getOperand(0);
13672 unsigned OpOpcode = Op.getNode()->getOpcode();
13673 if (!N->getValueType(0).isVector() || !N->getValueType(0).isSimple() ||
13674 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
13677 SDValue ConstVec = N->getOperand(1);
13678 if (!isa<BuildVectorSDNode>(ConstVec))
13681 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
13682 uint32_t FloatBits = FloatTy.getSizeInBits();
13683 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
13684 uint32_t IntBits = IntTy.getSizeInBits();
13685 unsigned NumLanes = Op.getValueType().getVectorNumElements();
13686 if (FloatBits != 32 || IntBits > 32 || (NumLanes != 4 && NumLanes != 2)) {
13687 // These instructions only exist converting from i32 to f32. We can handle
13688 // smaller integers by generating an extra extend, but larger ones would
13689 // be lossy. We also can't handle anything other than 2 or 4 lanes, since
13690 // these intructions only support v2i32/v4i32 types.
13694 BitVector UndefElements;
13695 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(ConstVec);
13696 int32_t C = BV->getConstantFPSplatPow2ToLog2Int(&UndefElements, 33);
13697 if (C == -1 || C == 0 || C > 32)
13701 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
13702 SDValue ConvInput = Op.getOperand(0);
13703 if (IntBits < FloatBits)
13704 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
13705 dl, NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
13708 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
13709 Intrinsic::arm_neon_vcvtfxu2fp;
13710 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl,
13712 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32),
13713 ConvInput, DAG.getConstant(C, dl, MVT::i32));
13716 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
13717 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
13718 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
13721 // Don't do anything for most intrinsics.
13724 // Vector shifts: check for immediate versions and lower them.
13725 // Note: This is done during DAG combining instead of DAG legalizing because
13726 // the build_vectors for 64-bit vector element shift counts are generally
13727 // not legal, and it is hard to see their values after they get legalized to
13728 // loads from a constant pool.
13729 case Intrinsic::arm_neon_vshifts:
13730 case Intrinsic::arm_neon_vshiftu:
13731 case Intrinsic::arm_neon_vrshifts:
13732 case Intrinsic::arm_neon_vrshiftu:
13733 case Intrinsic::arm_neon_vrshiftn:
13734 case Intrinsic::arm_neon_vqshifts:
13735 case Intrinsic::arm_neon_vqshiftu:
13736 case Intrinsic::arm_neon_vqshiftsu:
13737 case Intrinsic::arm_neon_vqshiftns:
13738 case Intrinsic::arm_neon_vqshiftnu:
13739 case Intrinsic::arm_neon_vqshiftnsu:
13740 case Intrinsic::arm_neon_vqrshiftns:
13741 case Intrinsic::arm_neon_vqrshiftnu:
13742 case Intrinsic::arm_neon_vqrshiftnsu: {
13743 EVT VT = N->getOperand(1).getValueType();
13745 unsigned VShiftOpc = 0;
13748 case Intrinsic::arm_neon_vshifts:
13749 case Intrinsic::arm_neon_vshiftu:
13750 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
13751 VShiftOpc = ARMISD::VSHLIMM;
13754 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
13755 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ? ARMISD::VSHRsIMM
13756 : ARMISD::VSHRuIMM);
13761 case Intrinsic::arm_neon_vrshifts:
13762 case Intrinsic::arm_neon_vrshiftu:
13763 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
13767 case Intrinsic::arm_neon_vqshifts:
13768 case Intrinsic::arm_neon_vqshiftu:
13769 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
13773 case Intrinsic::arm_neon_vqshiftsu:
13774 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
13776 llvm_unreachable("invalid shift count for vqshlu intrinsic");
13778 case Intrinsic::arm_neon_vrshiftn:
13779 case Intrinsic::arm_neon_vqshiftns:
13780 case Intrinsic::arm_neon_vqshiftnu:
13781 case Intrinsic::arm_neon_vqshiftnsu:
13782 case Intrinsic::arm_neon_vqrshiftns:
13783 case Intrinsic::arm_neon_vqrshiftnu:
13784 case Intrinsic::arm_neon_vqrshiftnsu:
13785 // Narrowing shifts require an immediate right shift.
13786 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
13788 llvm_unreachable("invalid shift count for narrowing vector shift "
13792 llvm_unreachable("unhandled vector shift");
13796 case Intrinsic::arm_neon_vshifts:
13797 case Intrinsic::arm_neon_vshiftu:
13798 // Opcode already set above.
13800 case Intrinsic::arm_neon_vrshifts:
13801 VShiftOpc = ARMISD::VRSHRsIMM;
13803 case Intrinsic::arm_neon_vrshiftu:
13804 VShiftOpc = ARMISD::VRSHRuIMM;
13806 case Intrinsic::arm_neon_vrshiftn:
13807 VShiftOpc = ARMISD::VRSHRNIMM;
13809 case Intrinsic::arm_neon_vqshifts:
13810 VShiftOpc = ARMISD::VQSHLsIMM;
13812 case Intrinsic::arm_neon_vqshiftu:
13813 VShiftOpc = ARMISD::VQSHLuIMM;
13815 case Intrinsic::arm_neon_vqshiftsu:
13816 VShiftOpc = ARMISD::VQSHLsuIMM;
13818 case Intrinsic::arm_neon_vqshiftns:
13819 VShiftOpc = ARMISD::VQSHRNsIMM;
13821 case Intrinsic::arm_neon_vqshiftnu:
13822 VShiftOpc = ARMISD::VQSHRNuIMM;
13824 case Intrinsic::arm_neon_vqshiftnsu:
13825 VShiftOpc = ARMISD::VQSHRNsuIMM;
13827 case Intrinsic::arm_neon_vqrshiftns:
13828 VShiftOpc = ARMISD::VQRSHRNsIMM;
13830 case Intrinsic::arm_neon_vqrshiftnu:
13831 VShiftOpc = ARMISD::VQRSHRNuIMM;
13833 case Intrinsic::arm_neon_vqrshiftnsu:
13834 VShiftOpc = ARMISD::VQRSHRNsuIMM;
13839 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
13840 N->getOperand(1), DAG.getConstant(Cnt, dl, MVT::i32));
13843 case Intrinsic::arm_neon_vshiftins: {
13844 EVT VT = N->getOperand(1).getValueType();
13846 unsigned VShiftOpc = 0;
13848 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
13849 VShiftOpc = ARMISD::VSLIIMM;
13850 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
13851 VShiftOpc = ARMISD::VSRIIMM;
13853 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
13857 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
13858 N->getOperand(1), N->getOperand(2),
13859 DAG.getConstant(Cnt, dl, MVT::i32));
13862 case Intrinsic::arm_neon_vqrshifts:
13863 case Intrinsic::arm_neon_vqrshiftu:
13864 // No immediate versions of these to check for.
13871 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
13872 /// lowers them. As with the vector shift intrinsics, this is done during DAG
13873 /// combining instead of DAG legalizing because the build_vectors for 64-bit
13874 /// vector element shift counts are generally not legal, and it is hard to see
13875 /// their values after they get legalized to loads from a constant pool.
13876 static SDValue PerformShiftCombine(SDNode *N,
13877 TargetLowering::DAGCombinerInfo &DCI,
13878 const ARMSubtarget *ST) {
13879 SelectionDAG &DAG = DCI.DAG;
13880 EVT VT = N->getValueType(0);
13881 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
13882 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
13883 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
13884 SDValue N1 = N->getOperand(1);
13885 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
13886 SDValue N0 = N->getOperand(0);
13887 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
13888 DAG.MaskedValueIsZero(N0.getOperand(0),
13889 APInt::getHighBitsSet(32, 16)))
13890 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
13894 if (ST->isThumb1Only() && N->getOpcode() == ISD::SHL && VT == MVT::i32 &&
13895 N->getOperand(0)->getOpcode() == ISD::AND &&
13896 N->getOperand(0)->hasOneUse()) {
13897 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
13899 // Look for the pattern (shl (and x, AndMask), ShiftAmt). This doesn't
13900 // usually show up because instcombine prefers to canonicalize it to
13901 // (and (shl x, ShiftAmt) (shl AndMask, ShiftAmt)), but the shift can come
13902 // out of GEP lowering in some cases.
13903 SDValue N0 = N->getOperand(0);
13904 ConstantSDNode *ShiftAmtNode = dyn_cast<ConstantSDNode>(N->getOperand(1));
13907 uint32_t ShiftAmt = static_cast<uint32_t>(ShiftAmtNode->getZExtValue());
13908 ConstantSDNode *AndMaskNode = dyn_cast<ConstantSDNode>(N0->getOperand(1));
13911 uint32_t AndMask = static_cast<uint32_t>(AndMaskNode->getZExtValue());
13912 // Don't transform uxtb/uxth.
13913 if (AndMask == 255 || AndMask == 65535)
13915 if (isMask_32(AndMask)) {
13916 uint32_t MaskedBits = countLeadingZeros(AndMask);
13917 if (MaskedBits > ShiftAmt) {
13919 SDValue SHL = DAG.getNode(ISD::SHL, DL, MVT::i32, N0->getOperand(0),
13920 DAG.getConstant(MaskedBits, DL, MVT::i32));
13921 return DAG.getNode(
13922 ISD::SRL, DL, MVT::i32, SHL,
13923 DAG.getConstant(MaskedBits - ShiftAmt, DL, MVT::i32));
13928 // Nothing to be done for scalar shifts.
13929 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13930 if (!VT.isVector() || !TLI.isTypeLegal(VT))
13932 if (ST->hasMVEIntegerOps() && VT == MVT::v2i64)
13937 switch (N->getOpcode()) {
13938 default: llvm_unreachable("unexpected shift opcode");
13941 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt)) {
13943 return DAG.getNode(ARMISD::VSHLIMM, dl, VT, N->getOperand(0),
13944 DAG.getConstant(Cnt, dl, MVT::i32));
13950 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
13951 unsigned VShiftOpc =
13952 (N->getOpcode() == ISD::SRA ? ARMISD::VSHRsIMM : ARMISD::VSHRuIMM);
13954 return DAG.getNode(VShiftOpc, dl, VT, N->getOperand(0),
13955 DAG.getConstant(Cnt, dl, MVT::i32));
13961 // Look for a sign/zero extend of a larger than legal load. This can be split
13962 // into two extending loads, which are simpler to deal with than an arbitrary
13964 static SDValue PerformSplittingToWideningLoad(SDNode *N, SelectionDAG &DAG) {
13965 SDValue N0 = N->getOperand(0);
13966 if (N0.getOpcode() != ISD::LOAD)
13968 LoadSDNode *LD = cast<LoadSDNode>(N0.getNode());
13969 if (!LD->isSimple() || !N0.hasOneUse() || LD->isIndexed() ||
13970 LD->getExtensionType() != ISD::NON_EXTLOAD)
13972 EVT FromVT = LD->getValueType(0);
13973 EVT ToVT = N->getValueType(0);
13974 if (!ToVT.isVector())
13976 assert(FromVT.getVectorNumElements() == ToVT.getVectorNumElements());
13977 EVT ToEltVT = ToVT.getVectorElementType();
13978 EVT FromEltVT = FromVT.getVectorElementType();
13980 unsigned NumElements = 0;
13981 if (ToEltVT == MVT::i32 && (FromEltVT == MVT::i16 || FromEltVT == MVT::i8))
13983 if (ToEltVT == MVT::i16 && FromEltVT == MVT::i8)
13985 if (NumElements == 0 ||
13986 FromVT.getVectorNumElements() == NumElements ||
13987 FromVT.getVectorNumElements() % NumElements != 0 ||
13988 !isPowerOf2_32(NumElements))
13992 // Details about the old load
13993 SDValue Ch = LD->getChain();
13994 SDValue BasePtr = LD->getBasePtr();
13995 unsigned Alignment = LD->getOriginalAlignment();
13996 MachineMemOperand::Flags MMOFlags = LD->getMemOperand()->getFlags();
13997 AAMDNodes AAInfo = LD->getAAInfo();
13999 ISD::LoadExtType NewExtType =
14000 N->getOpcode() == ISD::SIGN_EXTEND ? ISD::SEXTLOAD : ISD::ZEXTLOAD;
14001 SDValue Offset = DAG.getUNDEF(BasePtr.getValueType());
14002 EVT NewFromVT = FromVT.getHalfNumVectorElementsVT(*DAG.getContext());
14003 EVT NewToVT = ToVT.getHalfNumVectorElementsVT(*DAG.getContext());
14004 unsigned NewOffset = NewFromVT.getSizeInBits() / 8;
14005 SDValue NewPtr = DAG.getObjectPtrOffset(DL, BasePtr, NewOffset);
14007 // Split the load in half, each side of which is extended separately. This
14008 // is good enough, as legalisation will take it from there. They are either
14009 // already legal or they will be split further into something that is
14012 DAG.getLoad(ISD::UNINDEXED, NewExtType, NewToVT, DL, Ch, BasePtr, Offset,
14013 LD->getPointerInfo(), NewFromVT, Alignment, MMOFlags, AAInfo);
14015 DAG.getLoad(ISD::UNINDEXED, NewExtType, NewToVT, DL, Ch, NewPtr, Offset,
14016 LD->getPointerInfo().getWithOffset(NewOffset), NewFromVT,
14017 Alignment, MMOFlags, AAInfo);
14019 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
14020 SDValue(NewLoad1.getNode(), 1),
14021 SDValue(NewLoad2.getNode(), 1));
14022 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewChain);
14023 return DAG.getNode(ISD::CONCAT_VECTORS, DL, ToVT, NewLoad1, NewLoad2);
14026 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
14027 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
14028 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
14029 const ARMSubtarget *ST) {
14030 SDValue N0 = N->getOperand(0);
14032 // Check for sign- and zero-extensions of vector extract operations of 8- and
14033 // 16-bit vector elements. NEON and MVE support these directly. They are
14034 // handled during DAG combining because type legalization will promote them
14035 // to 32-bit types and it is messy to recognize the operations after that.
14036 if ((ST->hasNEON() || ST->hasMVEIntegerOps()) &&
14037 N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
14038 SDValue Vec = N0.getOperand(0);
14039 SDValue Lane = N0.getOperand(1);
14040 EVT VT = N->getValueType(0);
14041 EVT EltVT = N0.getValueType();
14042 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14044 if (VT == MVT::i32 &&
14045 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
14046 TLI.isTypeLegal(Vec.getValueType()) &&
14047 isa<ConstantSDNode>(Lane)) {
14050 switch (N->getOpcode()) {
14051 default: llvm_unreachable("unexpected opcode");
14052 case ISD::SIGN_EXTEND:
14053 Opc = ARMISD::VGETLANEs;
14055 case ISD::ZERO_EXTEND:
14056 case ISD::ANY_EXTEND:
14057 Opc = ARMISD::VGETLANEu;
14060 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
14064 if (ST->hasMVEIntegerOps())
14065 if (SDValue NewLoad = PerformSplittingToWideningLoad(N, DAG))
14071 static const APInt *isPowerOf2Constant(SDValue V) {
14072 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
14075 const APInt *CV = &C->getAPIntValue();
14076 return CV->isPowerOf2() ? CV : nullptr;
14079 SDValue ARMTargetLowering::PerformCMOVToBFICombine(SDNode *CMOV, SelectionDAG &DAG) const {
14080 // If we have a CMOV, OR and AND combination such as:
14085 // * CN is a single bit;
14086 // * All bits covered by CM are known zero in y
14088 // Then we can convert this into a sequence of BFI instructions. This will
14089 // always be a win if CM is a single bit, will always be no worse than the
14090 // TST&OR sequence if CM is two bits, and for thumb will be no worse if CM is
14091 // three bits (due to the extra IT instruction).
14093 SDValue Op0 = CMOV->getOperand(0);
14094 SDValue Op1 = CMOV->getOperand(1);
14095 auto CCNode = cast<ConstantSDNode>(CMOV->getOperand(2));
14096 auto CC = CCNode->getAPIntValue().getLimitedValue();
14097 SDValue CmpZ = CMOV->getOperand(4);
14099 // The compare must be against zero.
14100 if (!isNullConstant(CmpZ->getOperand(1)))
14103 assert(CmpZ->getOpcode() == ARMISD::CMPZ);
14104 SDValue And = CmpZ->getOperand(0);
14105 if (And->getOpcode() != ISD::AND)
14107 const APInt *AndC = isPowerOf2Constant(And->getOperand(1));
14110 SDValue X = And->getOperand(0);
14112 if (CC == ARMCC::EQ) {
14113 // We're performing an "equal to zero" compare. Swap the operands so we
14114 // canonicalize on a "not equal to zero" compare.
14115 std::swap(Op0, Op1);
14117 assert(CC == ARMCC::NE && "How can a CMPZ node not be EQ or NE?");
14120 if (Op1->getOpcode() != ISD::OR)
14123 ConstantSDNode *OrC = dyn_cast<ConstantSDNode>(Op1->getOperand(1));
14126 SDValue Y = Op1->getOperand(0);
14131 // Now, is it profitable to continue?
14132 APInt OrCI = OrC->getAPIntValue();
14133 unsigned Heuristic = Subtarget->isThumb() ? 3 : 2;
14134 if (OrCI.countPopulation() > Heuristic)
14137 // Lastly, can we determine that the bits defined by OrCI
14139 KnownBits Known = DAG.computeKnownBits(Y);
14140 if ((OrCI & Known.Zero) != OrCI)
14143 // OK, we can do the combine.
14146 EVT VT = X.getValueType();
14147 unsigned BitInX = AndC->logBase2();
14150 // We must shift X first.
14151 X = DAG.getNode(ISD::SRL, dl, VT, X,
14152 DAG.getConstant(BitInX, dl, VT));
14155 for (unsigned BitInY = 0, NumActiveBits = OrCI.getActiveBits();
14156 BitInY < NumActiveBits; ++BitInY) {
14157 if (OrCI[BitInY] == 0)
14159 APInt Mask(VT.getSizeInBits(), 0);
14160 Mask.setBit(BitInY);
14161 V = DAG.getNode(ARMISD::BFI, dl, VT, V, X,
14162 // Confusingly, the operand is an *inverted* mask.
14163 DAG.getConstant(~Mask, dl, VT));
14169 // Given N, the value controlling the conditional branch, search for the loop
14170 // intrinsic, returning it, along with how the value is used. We need to handle
14171 // patterns such as the following:
14172 // (brcond (xor (setcc (loop.decrement), 0, ne), 1), exit)
14173 // (brcond (setcc (loop.decrement), 0, eq), exit)
14174 // (brcond (setcc (loop.decrement), 0, ne), header)
14175 static SDValue SearchLoopIntrinsic(SDValue N, ISD::CondCode &CC, int &Imm,
14177 switch (N->getOpcode()) {
14181 if (!isa<ConstantSDNode>(N.getOperand(1)))
14183 if (!cast<ConstantSDNode>(N.getOperand(1))->isOne())
14186 return SearchLoopIntrinsic(N.getOperand(0), CC, Imm, Negate);
14189 auto *Const = dyn_cast<ConstantSDNode>(N.getOperand(1));
14192 if (Const->isNullValue())
14194 else if (Const->isOne())
14198 CC = cast<CondCodeSDNode>(N.getOperand(2))->get();
14199 return SearchLoopIntrinsic(N->getOperand(0), CC, Imm, Negate);
14201 case ISD::INTRINSIC_W_CHAIN: {
14202 unsigned IntOp = cast<ConstantSDNode>(N.getOperand(1))->getZExtValue();
14203 if (IntOp != Intrinsic::test_set_loop_iterations &&
14204 IntOp != Intrinsic::loop_decrement_reg)
14212 static SDValue PerformHWLoopCombine(SDNode *N,
14213 TargetLowering::DAGCombinerInfo &DCI,
14214 const ARMSubtarget *ST) {
14216 // The hwloop intrinsics that we're interested are used for control-flow,
14217 // either for entering or exiting the loop:
14218 // - test.set.loop.iterations will test whether its operand is zero. If it
14219 // is zero, the proceeding branch should not enter the loop.
14220 // - loop.decrement.reg also tests whether its operand is zero. If it is
14221 // zero, the proceeding branch should not branch back to the beginning of
14223 // So here, we need to check that how the brcond is using the result of each
14224 // of the intrinsics to ensure that we're branching to the right place at the
14230 bool Negate = false;
14231 SDValue Chain = N->getOperand(0);
14234 if (N->getOpcode() == ISD::BRCOND) {
14236 Cond = N->getOperand(1);
14237 Dest = N->getOperand(2);
14239 assert(N->getOpcode() == ISD::BR_CC && "Expected BRCOND or BR_CC!");
14240 CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
14241 Cond = N->getOperand(2);
14242 Dest = N->getOperand(4);
14243 if (auto *Const = dyn_cast<ConstantSDNode>(N->getOperand(3))) {
14244 if (!Const->isOne() && !Const->isNullValue())
14246 Imm = Const->getZExtValue();
14251 SDValue Int = SearchLoopIntrinsic(Cond, CC, Imm, Negate);
14256 CC = ISD::getSetCCInverse(CC, /* Integer inverse */ MVT::i32);
14258 auto IsTrueIfZero = [](ISD::CondCode CC, int Imm) {
14259 return (CC == ISD::SETEQ && Imm == 0) ||
14260 (CC == ISD::SETNE && Imm == 1) ||
14261 (CC == ISD::SETLT && Imm == 1) ||
14262 (CC == ISD::SETULT && Imm == 1);
14265 auto IsFalseIfZero = [](ISD::CondCode CC, int Imm) {
14266 return (CC == ISD::SETEQ && Imm == 1) ||
14267 (CC == ISD::SETNE && Imm == 0) ||
14268 (CC == ISD::SETGT && Imm == 0) ||
14269 (CC == ISD::SETUGT && Imm == 0) ||
14270 (CC == ISD::SETGE && Imm == 1) ||
14271 (CC == ISD::SETUGE && Imm == 1);
14274 assert((IsTrueIfZero(CC, Imm) || IsFalseIfZero(CC, Imm)) &&
14275 "unsupported condition");
14278 SelectionDAG &DAG = DCI.DAG;
14279 SDValue Elements = Int.getOperand(2);
14280 unsigned IntOp = cast<ConstantSDNode>(Int->getOperand(1))->getZExtValue();
14281 assert((N->hasOneUse() && N->use_begin()->getOpcode() == ISD::BR)
14282 && "expected single br user");
14283 SDNode *Br = *N->use_begin();
14284 SDValue OtherTarget = Br->getOperand(1);
14286 // Update the unconditional branch to branch to the given Dest.
14287 auto UpdateUncondBr = [](SDNode *Br, SDValue Dest, SelectionDAG &DAG) {
14288 SDValue NewBrOps[] = { Br->getOperand(0), Dest };
14289 SDValue NewBr = DAG.getNode(ISD::BR, SDLoc(Br), MVT::Other, NewBrOps);
14290 DAG.ReplaceAllUsesOfValueWith(SDValue(Br, 0), NewBr);
14293 if (IntOp == Intrinsic::test_set_loop_iterations) {
14295 // We expect this 'instruction' to branch when the counter is zero.
14296 if (IsTrueIfZero(CC, Imm)) {
14297 SDValue Ops[] = { Chain, Elements, Dest };
14298 Res = DAG.getNode(ARMISD::WLS, dl, MVT::Other, Ops);
14300 // The logic is the reverse of what we need for WLS, so find the other
14301 // basic block target: the target of the proceeding br.
14302 UpdateUncondBr(Br, Dest, DAG);
14304 SDValue Ops[] = { Chain, Elements, OtherTarget };
14305 Res = DAG.getNode(ARMISD::WLS, dl, MVT::Other, Ops);
14307 DAG.ReplaceAllUsesOfValueWith(Int.getValue(1), Int.getOperand(0));
14310 SDValue Size = DAG.getTargetConstant(
14311 cast<ConstantSDNode>(Int.getOperand(3))->getZExtValue(), dl, MVT::i32);
14312 SDValue Args[] = { Int.getOperand(0), Elements, Size, };
14313 SDValue LoopDec = DAG.getNode(ARMISD::LOOP_DEC, dl,
14314 DAG.getVTList(MVT::i32, MVT::Other), Args);
14315 DAG.ReplaceAllUsesWith(Int.getNode(), LoopDec.getNode());
14317 // We expect this instruction to branch when the count is not zero.
14318 SDValue Target = IsFalseIfZero(CC, Imm) ? Dest : OtherTarget;
14320 // Update the unconditional branch to target the loop preheader if we've
14321 // found the condition has been reversed.
14322 if (Target == OtherTarget)
14323 UpdateUncondBr(Br, Dest, DAG);
14325 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
14326 SDValue(LoopDec.getNode(), 1), Chain);
14328 SDValue EndArgs[] = { Chain, SDValue(LoopDec.getNode(), 0), Target };
14329 return DAG.getNode(ARMISD::LE, dl, MVT::Other, EndArgs);
14334 /// PerformBRCONDCombine - Target-specific DAG combining for ARMISD::BRCOND.
14336 ARMTargetLowering::PerformBRCONDCombine(SDNode *N, SelectionDAG &DAG) const {
14337 SDValue Cmp = N->getOperand(4);
14338 if (Cmp.getOpcode() != ARMISD::CMPZ)
14339 // Only looking at NE cases.
14342 EVT VT = N->getValueType(0);
14344 SDValue LHS = Cmp.getOperand(0);
14345 SDValue RHS = Cmp.getOperand(1);
14346 SDValue Chain = N->getOperand(0);
14347 SDValue BB = N->getOperand(1);
14348 SDValue ARMcc = N->getOperand(2);
14349 ARMCC::CondCodes CC =
14350 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
14352 // (brcond Chain BB ne CPSR (cmpz (and (cmov 0 1 CC CPSR Cmp) 1) 0))
14353 // -> (brcond Chain BB CC CPSR Cmp)
14354 if (CC == ARMCC::NE && LHS.getOpcode() == ISD::AND && LHS->hasOneUse() &&
14355 LHS->getOperand(0)->getOpcode() == ARMISD::CMOV &&
14356 LHS->getOperand(0)->hasOneUse()) {
14357 auto *LHS00C = dyn_cast<ConstantSDNode>(LHS->getOperand(0)->getOperand(0));
14358 auto *LHS01C = dyn_cast<ConstantSDNode>(LHS->getOperand(0)->getOperand(1));
14359 auto *LHS1C = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
14360 auto *RHSC = dyn_cast<ConstantSDNode>(RHS);
14361 if ((LHS00C && LHS00C->getZExtValue() == 0) &&
14362 (LHS01C && LHS01C->getZExtValue() == 1) &&
14363 (LHS1C && LHS1C->getZExtValue() == 1) &&
14364 (RHSC && RHSC->getZExtValue() == 0)) {
14365 return DAG.getNode(
14366 ARMISD::BRCOND, dl, VT, Chain, BB, LHS->getOperand(0)->getOperand(2),
14367 LHS->getOperand(0)->getOperand(3), LHS->getOperand(0)->getOperand(4));
14374 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
14376 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
14377 SDValue Cmp = N->getOperand(4);
14378 if (Cmp.getOpcode() != ARMISD::CMPZ)
14379 // Only looking at EQ and NE cases.
14382 EVT VT = N->getValueType(0);
14384 SDValue LHS = Cmp.getOperand(0);
14385 SDValue RHS = Cmp.getOperand(1);
14386 SDValue FalseVal = N->getOperand(0);
14387 SDValue TrueVal = N->getOperand(1);
14388 SDValue ARMcc = N->getOperand(2);
14389 ARMCC::CondCodes CC =
14390 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
14392 // BFI is only available on V6T2+.
14393 if (!Subtarget->isThumb1Only() && Subtarget->hasV6T2Ops()) {
14394 SDValue R = PerformCMOVToBFICombine(N, DAG);
14415 /// FIXME: Turn this into a target neutral optimization?
14417 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
14418 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
14419 N->getOperand(3), Cmp);
14420 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
14422 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
14423 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
14424 N->getOperand(3), NewCmp);
14427 // (cmov F T ne CPSR (cmpz (cmov 0 1 CC CPSR Cmp) 0))
14428 // -> (cmov F T CC CPSR Cmp)
14429 if (CC == ARMCC::NE && LHS.getOpcode() == ARMISD::CMOV && LHS->hasOneUse()) {
14430 auto *LHS0C = dyn_cast<ConstantSDNode>(LHS->getOperand(0));
14431 auto *LHS1C = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
14432 auto *RHSC = dyn_cast<ConstantSDNode>(RHS);
14433 if ((LHS0C && LHS0C->getZExtValue() == 0) &&
14434 (LHS1C && LHS1C->getZExtValue() == 1) &&
14435 (RHSC && RHSC->getZExtValue() == 0)) {
14436 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
14437 LHS->getOperand(2), LHS->getOperand(3),
14438 LHS->getOperand(4));
14442 if (!VT.isInteger())
14445 // Materialize a boolean comparison for integers so we can avoid branching.
14446 if (isNullConstant(FalseVal)) {
14447 if (CC == ARMCC::EQ && isOneConstant(TrueVal)) {
14448 if (!Subtarget->isThumb1Only() && Subtarget->hasV5TOps()) {
14449 // If x == y then x - y == 0 and ARM's CLZ will return 32, shifting it
14450 // right 5 bits will make that 32 be 1, otherwise it will be 0.
14451 // CMOV 0, 1, ==, (CMPZ x, y) -> SRL (CTLZ (SUB x, y)), 5
14452 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, LHS, RHS);
14453 Res = DAG.getNode(ISD::SRL, dl, VT, DAG.getNode(ISD::CTLZ, dl, VT, Sub),
14454 DAG.getConstant(5, dl, MVT::i32));
14456 // CMOV 0, 1, ==, (CMPZ x, y) ->
14457 // (ADDCARRY (SUB x, y), t:0, t:1)
14458 // where t = (SUBCARRY 0, (SUB x, y), 0)
14460 // The SUBCARRY computes 0 - (x - y) and this will give a borrow when
14461 // x != y. In other words, a carry C == 1 when x == y, C == 0
14463 // The final ADDCARRY computes
14464 // x - y + (0 - (x - y)) + C == C
14465 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, LHS, RHS);
14466 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
14467 SDValue Neg = DAG.getNode(ISD::USUBO, dl, VTs, FalseVal, Sub);
14468 // ISD::SUBCARRY returns a borrow but we want the carry here
14471 DAG.getNode(ISD::SUB, dl, MVT::i32,
14472 DAG.getConstant(1, dl, MVT::i32), Neg.getValue(1));
14473 Res = DAG.getNode(ISD::ADDCARRY, dl, VTs, Sub, Neg, Carry);
14475 } else if (CC == ARMCC::NE && !isNullConstant(RHS) &&
14476 (!Subtarget->isThumb1Only() || isPowerOf2Constant(TrueVal))) {
14477 // This seems pointless but will allow us to combine it further below.
14478 // CMOV 0, z, !=, (CMPZ x, y) -> CMOV (SUBS x, y), z, !=, (SUBS x, y):1
14480 DAG.getNode(ARMISD::SUBS, dl, DAG.getVTList(VT, MVT::i32), LHS, RHS);
14481 SDValue CPSRGlue = DAG.getCopyToReg(DAG.getEntryNode(), dl, ARM::CPSR,
14482 Sub.getValue(1), SDValue());
14483 Res = DAG.getNode(ARMISD::CMOV, dl, VT, Sub, TrueVal, ARMcc,
14484 N->getOperand(3), CPSRGlue.getValue(1));
14487 } else if (isNullConstant(TrueVal)) {
14488 if (CC == ARMCC::EQ && !isNullConstant(RHS) &&
14489 (!Subtarget->isThumb1Only() || isPowerOf2Constant(FalseVal))) {
14490 // This seems pointless but will allow us to combine it further below
14491 // Note that we change == for != as this is the dual for the case above.
14492 // CMOV z, 0, ==, (CMPZ x, y) -> CMOV (SUBS x, y), z, !=, (SUBS x, y):1
14494 DAG.getNode(ARMISD::SUBS, dl, DAG.getVTList(VT, MVT::i32), LHS, RHS);
14495 SDValue CPSRGlue = DAG.getCopyToReg(DAG.getEntryNode(), dl, ARM::CPSR,
14496 Sub.getValue(1), SDValue());
14497 Res = DAG.getNode(ARMISD::CMOV, dl, VT, Sub, FalseVal,
14498 DAG.getConstant(ARMCC::NE, dl, MVT::i32),
14499 N->getOperand(3), CPSRGlue.getValue(1));
14504 // On Thumb1, the DAG above may be further combined if z is a power of 2
14506 // CMOV (SUBS x, y), z, !=, (SUBS x, y):1 ->
14507 // t1 = (USUBO (SUB x, y), 1)
14508 // t2 = (SUBCARRY (SUB x, y), t1:0, t1:1)
14509 // Result = if K != 0 then (SHL t2:0, K) else t2:0
14511 // This also handles the special case of comparing against zero; it's
14512 // essentially, the same pattern, except there's no SUBS:
14513 // CMOV x, z, !=, (CMPZ x, 0) ->
14514 // t1 = (USUBO x, 1)
14515 // t2 = (SUBCARRY x, t1:0, t1:1)
14516 // Result = if K != 0 then (SHL t2:0, K) else t2:0
14517 const APInt *TrueConst;
14518 if (Subtarget->isThumb1Only() && CC == ARMCC::NE &&
14519 ((FalseVal.getOpcode() == ARMISD::SUBS &&
14520 FalseVal.getOperand(0) == LHS && FalseVal.getOperand(1) == RHS) ||
14521 (FalseVal == LHS && isNullConstant(RHS))) &&
14522 (TrueConst = isPowerOf2Constant(TrueVal))) {
14523 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
14524 unsigned ShiftAmount = TrueConst->logBase2();
14526 TrueVal = DAG.getConstant(1, dl, VT);
14527 SDValue Subc = DAG.getNode(ISD::USUBO, dl, VTs, FalseVal, TrueVal);
14528 Res = DAG.getNode(ISD::SUBCARRY, dl, VTs, FalseVal, Subc, Subc.getValue(1));
14531 Res = DAG.getNode(ISD::SHL, dl, VT, Res,
14532 DAG.getConstant(ShiftAmount, dl, MVT::i32));
14535 if (Res.getNode()) {
14536 KnownBits Known = DAG.computeKnownBits(SDValue(N,0));
14537 // Capture demanded bits information that would be otherwise lost.
14538 if (Known.Zero == 0xfffffffe)
14539 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
14540 DAG.getValueType(MVT::i1));
14541 else if (Known.Zero == 0xffffff00)
14542 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
14543 DAG.getValueType(MVT::i8));
14544 else if (Known.Zero == 0xffff0000)
14545 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
14546 DAG.getValueType(MVT::i16));
14552 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
14553 DAGCombinerInfo &DCI) const {
14554 switch (N->getOpcode()) {
14556 case ISD::ABS: return PerformABSCombine(N, DCI, Subtarget);
14557 case ARMISD::ADDE: return PerformADDECombine(N, DCI, Subtarget);
14558 case ARMISD::UMLAL: return PerformUMLALCombine(N, DCI.DAG, Subtarget);
14559 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
14560 case ISD::SUB: return PerformSUBCombine(N, DCI, Subtarget);
14561 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
14562 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
14563 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
14564 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
14566 case ISD::BR_CC: return PerformHWLoopCombine(N, DCI, Subtarget);
14568 case ARMISD::SUBC: return PerformAddcSubcCombine(N, DCI, Subtarget);
14569 case ARMISD::SUBE: return PerformAddeSubeCombine(N, DCI, Subtarget);
14570 case ARMISD::BFI: return PerformBFICombine(N, DCI);
14571 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI, Subtarget);
14572 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
14573 case ISD::STORE: return PerformSTORECombine(N, DCI, Subtarget);
14574 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI, Subtarget);
14575 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
14576 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
14577 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
14578 case ARMISD::VDUP: return PerformVDUPCombine(N, DCI, Subtarget);
14579 case ISD::FP_TO_SINT:
14580 case ISD::FP_TO_UINT:
14581 return PerformVCVTCombine(N, DCI.DAG, Subtarget);
14583 return PerformVDIVCombine(N, DCI.DAG, Subtarget);
14584 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
14588 return PerformShiftCombine(N, DCI, Subtarget);
14589 case ISD::SIGN_EXTEND:
14590 case ISD::ZERO_EXTEND:
14591 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
14592 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
14593 case ARMISD::BRCOND: return PerformBRCONDCombine(N, DCI.DAG);
14594 case ISD::LOAD: return PerformLOADCombine(N, DCI);
14595 case ARMISD::VLD1DUP:
14596 case ARMISD::VLD2DUP:
14597 case ARMISD::VLD3DUP:
14598 case ARMISD::VLD4DUP:
14599 return PerformVLDCombine(N, DCI);
14600 case ARMISD::BUILD_VECTOR:
14601 return PerformARMBUILD_VECTORCombine(N, DCI);
14602 case ARMISD::PREDICATE_CAST:
14603 return PerformPREDICATE_CASTCombine(N, DCI);
14605 return PerformVCMPCombine(N, DCI, Subtarget);
14606 case ARMISD::SMULWB: {
14607 unsigned BitWidth = N->getValueType(0).getSizeInBits();
14608 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, 16);
14609 if (SimplifyDemandedBits(N->getOperand(1), DemandedMask, DCI))
14613 case ARMISD::SMULWT: {
14614 unsigned BitWidth = N->getValueType(0).getSizeInBits();
14615 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 16);
14616 if (SimplifyDemandedBits(N->getOperand(1), DemandedMask, DCI))
14620 case ARMISD::SMLALBB:
14621 case ARMISD::QADD16b:
14622 case ARMISD::QSUB16b: {
14623 unsigned BitWidth = N->getValueType(0).getSizeInBits();
14624 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, 16);
14625 if ((SimplifyDemandedBits(N->getOperand(0), DemandedMask, DCI)) ||
14626 (SimplifyDemandedBits(N->getOperand(1), DemandedMask, DCI)))
14630 case ARMISD::SMLALBT: {
14631 unsigned LowWidth = N->getOperand(0).getValueType().getSizeInBits();
14632 APInt LowMask = APInt::getLowBitsSet(LowWidth, 16);
14633 unsigned HighWidth = N->getOperand(1).getValueType().getSizeInBits();
14634 APInt HighMask = APInt::getHighBitsSet(HighWidth, 16);
14635 if ((SimplifyDemandedBits(N->getOperand(0), LowMask, DCI)) ||
14636 (SimplifyDemandedBits(N->getOperand(1), HighMask, DCI)))
14640 case ARMISD::SMLALTB: {
14641 unsigned HighWidth = N->getOperand(0).getValueType().getSizeInBits();
14642 APInt HighMask = APInt::getHighBitsSet(HighWidth, 16);
14643 unsigned LowWidth = N->getOperand(1).getValueType().getSizeInBits();
14644 APInt LowMask = APInt::getLowBitsSet(LowWidth, 16);
14645 if ((SimplifyDemandedBits(N->getOperand(0), HighMask, DCI)) ||
14646 (SimplifyDemandedBits(N->getOperand(1), LowMask, DCI)))
14650 case ARMISD::SMLALTT: {
14651 unsigned BitWidth = N->getValueType(0).getSizeInBits();
14652 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 16);
14653 if ((SimplifyDemandedBits(N->getOperand(0), DemandedMask, DCI)) ||
14654 (SimplifyDemandedBits(N->getOperand(1), DemandedMask, DCI)))
14658 case ARMISD::QADD8b:
14659 case ARMISD::QSUB8b: {
14660 unsigned BitWidth = N->getValueType(0).getSizeInBits();
14661 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, 8);
14662 if ((SimplifyDemandedBits(N->getOperand(0), DemandedMask, DCI)) ||
14663 (SimplifyDemandedBits(N->getOperand(1), DemandedMask, DCI)))
14667 case ISD::INTRINSIC_VOID:
14668 case ISD::INTRINSIC_W_CHAIN:
14669 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
14670 case Intrinsic::arm_neon_vld1:
14671 case Intrinsic::arm_neon_vld1x2:
14672 case Intrinsic::arm_neon_vld1x3:
14673 case Intrinsic::arm_neon_vld1x4:
14674 case Intrinsic::arm_neon_vld2:
14675 case Intrinsic::arm_neon_vld3:
14676 case Intrinsic::arm_neon_vld4:
14677 case Intrinsic::arm_neon_vld2lane:
14678 case Intrinsic::arm_neon_vld3lane:
14679 case Intrinsic::arm_neon_vld4lane:
14680 case Intrinsic::arm_neon_vld2dup:
14681 case Intrinsic::arm_neon_vld3dup:
14682 case Intrinsic::arm_neon_vld4dup:
14683 case Intrinsic::arm_neon_vst1:
14684 case Intrinsic::arm_neon_vst1x2:
14685 case Intrinsic::arm_neon_vst1x3:
14686 case Intrinsic::arm_neon_vst1x4:
14687 case Intrinsic::arm_neon_vst2:
14688 case Intrinsic::arm_neon_vst3:
14689 case Intrinsic::arm_neon_vst4:
14690 case Intrinsic::arm_neon_vst2lane:
14691 case Intrinsic::arm_neon_vst3lane:
14692 case Intrinsic::arm_neon_vst4lane:
14693 return PerformVLDCombine(N, DCI);
14701 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
14703 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
14706 bool ARMTargetLowering::allowsMisalignedMemoryAccesses(EVT VT, unsigned,
14707 unsigned Alignment,
14708 MachineMemOperand::Flags,
14709 bool *Fast) const {
14710 // Depends what it gets converted into if the type is weird.
14711 if (!VT.isSimple())
14714 // The AllowsUnaligned flag models the SCTLR.A setting in ARM cpus
14715 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
14716 auto Ty = VT.getSimpleVT().SimpleTy;
14718 if (Ty == MVT::i8 || Ty == MVT::i16 || Ty == MVT::i32) {
14719 // Unaligned access can use (for example) LRDB, LRDH, LDR
14720 if (AllowsUnaligned) {
14722 *Fast = Subtarget->hasV7Ops();
14727 if (Ty == MVT::f64 || Ty == MVT::v2f64) {
14728 // For any little-endian targets with neon, we can support unaligned ld/st
14729 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
14730 // A big-endian target may also explicitly support unaligned accesses
14731 if (Subtarget->hasNEON() && (AllowsUnaligned || Subtarget->isLittle())) {
14738 if (!Subtarget->hasMVEIntegerOps())
14741 // These are for predicates
14742 if ((Ty == MVT::v16i1 || Ty == MVT::v8i1 || Ty == MVT::v4i1)) {
14748 // These are for truncated stores/narrowing loads. They are fine so long as
14749 // the alignment is at least the size of the item being loaded
14750 if ((Ty == MVT::v4i8 || Ty == MVT::v8i8 || Ty == MVT::v4i16) &&
14751 Alignment >= VT.getScalarSizeInBits() / 8) {
14757 // In little-endian MVE, the store instructions VSTRB.U8, VSTRH.U16 and
14758 // VSTRW.U32 all store the vector register in exactly the same format, and
14759 // differ only in the range of their immediate offset field and the required
14760 // alignment. So there is always a store that can be used, regardless of
14763 // For big endian, that is not the case. But can still emit a (VSTRB.U8;
14764 // VREV64.8) pair and get the same effect. This will likely be better than
14765 // aligning the vector through the stack.
14766 if (Ty == MVT::v16i8 || Ty == MVT::v8i16 || Ty == MVT::v8f16 ||
14767 Ty == MVT::v4i32 || Ty == MVT::v4f32 || Ty == MVT::v2i64 ||
14768 Ty == MVT::v2f64) {
14777 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
14778 unsigned AlignCheck) {
14779 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
14780 (DstAlign == 0 || DstAlign % AlignCheck == 0));
14783 EVT ARMTargetLowering::getOptimalMemOpType(
14784 uint64_t Size, unsigned DstAlign, unsigned SrcAlign, bool IsMemset,
14785 bool ZeroMemset, bool MemcpyStrSrc,
14786 const AttributeList &FuncAttributes) const {
14787 // See if we can use NEON instructions for this...
14788 if ((!IsMemset || ZeroMemset) && Subtarget->hasNEON() &&
14789 !FuncAttributes.hasFnAttribute(Attribute::NoImplicitFloat)) {
14792 (memOpAlign(SrcAlign, DstAlign, 16) ||
14793 (allowsMisalignedMemoryAccesses(MVT::v2f64, 0, 1,
14794 MachineMemOperand::MONone, &Fast) &&
14797 } else if (Size >= 8 &&
14798 (memOpAlign(SrcAlign, DstAlign, 8) ||
14799 (allowsMisalignedMemoryAccesses(
14800 MVT::f64, 0, 1, MachineMemOperand::MONone, &Fast) &&
14806 // Let the target-independent logic figure it out.
14810 // 64-bit integers are split into their high and low parts and held in two
14811 // different registers, so the trunc is free since the low register can just
14813 bool ARMTargetLowering::isTruncateFree(Type *SrcTy, Type *DstTy) const {
14814 if (!SrcTy->isIntegerTy() || !DstTy->isIntegerTy())
14816 unsigned SrcBits = SrcTy->getPrimitiveSizeInBits();
14817 unsigned DestBits = DstTy->getPrimitiveSizeInBits();
14818 return (SrcBits == 64 && DestBits == 32);
14821 bool ARMTargetLowering::isTruncateFree(EVT SrcVT, EVT DstVT) const {
14822 if (SrcVT.isVector() || DstVT.isVector() || !SrcVT.isInteger() ||
14823 !DstVT.isInteger())
14825 unsigned SrcBits = SrcVT.getSizeInBits();
14826 unsigned DestBits = DstVT.getSizeInBits();
14827 return (SrcBits == 64 && DestBits == 32);
14830 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
14831 if (Val.getOpcode() != ISD::LOAD)
14834 EVT VT1 = Val.getValueType();
14835 if (!VT1.isSimple() || !VT1.isInteger() ||
14836 !VT2.isSimple() || !VT2.isInteger())
14839 switch (VT1.getSimpleVT().SimpleTy) {
14844 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
14851 bool ARMTargetLowering::isFNegFree(EVT VT) const {
14852 if (!VT.isSimple())
14855 // There are quite a few FP16 instructions (e.g. VNMLA, VNMLS, etc.) that
14856 // negate values directly (fneg is free). So, we don't want to let the DAG
14857 // combiner rewrite fneg into xors and some other instructions. For f16 and
14858 // FullFP16 argument passing, some bitcast nodes may be introduced,
14859 // triggering this DAG combine rewrite, so we are avoiding that with this.
14860 switch (VT.getSimpleVT().SimpleTy) {
14863 return Subtarget->hasFullFP16();
14869 /// Check if Ext1 and Ext2 are extends of the same type, doubling the bitwidth
14870 /// of the vector elements.
14871 static bool areExtractExts(Value *Ext1, Value *Ext2) {
14872 auto areExtDoubled = [](Instruction *Ext) {
14873 return Ext->getType()->getScalarSizeInBits() ==
14874 2 * Ext->getOperand(0)->getType()->getScalarSizeInBits();
14877 if (!match(Ext1, m_ZExtOrSExt(m_Value())) ||
14878 !match(Ext2, m_ZExtOrSExt(m_Value())) ||
14879 !areExtDoubled(cast<Instruction>(Ext1)) ||
14880 !areExtDoubled(cast<Instruction>(Ext2)))
14886 /// Check if sinking \p I's operands to I's basic block is profitable, because
14887 /// the operands can be folded into a target instruction, e.g.
14888 /// sext/zext can be folded into vsubl.
14889 bool ARMTargetLowering::shouldSinkOperands(Instruction *I,
14890 SmallVectorImpl<Use *> &Ops) const {
14891 if (!I->getType()->isVectorTy())
14894 if (Subtarget->hasNEON()) {
14895 switch (I->getOpcode()) {
14896 case Instruction::Sub:
14897 case Instruction::Add: {
14898 if (!areExtractExts(I->getOperand(0), I->getOperand(1)))
14900 Ops.push_back(&I->getOperandUse(0));
14901 Ops.push_back(&I->getOperandUse(1));
14909 if (!Subtarget->hasMVEIntegerOps())
14912 auto IsSinker = [](Instruction *I, int Operand) {
14913 switch (I->getOpcode()) {
14914 case Instruction::Add:
14915 case Instruction::Mul:
14916 case Instruction::ICmp:
14918 case Instruction::Sub:
14919 case Instruction::Shl:
14920 case Instruction::LShr:
14921 case Instruction::AShr:
14922 return Operand == 1;
14929 if (!isa<ShuffleVectorInst>(I->getOperand(Op)))
14931 if (!IsSinker(I, Op))
14933 if (!match(I->getOperand(Op),
14934 m_ShuffleVector(m_InsertElement(m_Undef(), m_Value(), m_ZeroInt()),
14935 m_Undef(), m_Zero()))) {
14938 Instruction *Shuffle = cast<Instruction>(I->getOperand(Op));
14939 // All uses of the shuffle should be sunk to avoid duplicating it across gpr
14940 // and vector registers
14941 for (Use &U : Shuffle->uses()) {
14942 Instruction *Insn = cast<Instruction>(U.getUser());
14943 if (!IsSinker(Insn, U.getOperandNo()))
14946 Ops.push_back(&Shuffle->getOperandUse(0));
14947 Ops.push_back(&I->getOperandUse(Op));
14951 bool ARMTargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const {
14952 EVT VT = ExtVal.getValueType();
14954 if (!isTypeLegal(VT))
14957 if (auto *Ld = dyn_cast<MaskedLoadSDNode>(ExtVal.getOperand(0))) {
14958 if (Ld->isExpandingLoad())
14962 // Don't create a loadext if we can fold the extension into a wide/long
14964 // If there's more than one user instruction, the loadext is desirable no
14965 // matter what. There can be two uses by the same instruction.
14966 if (ExtVal->use_empty() ||
14967 !ExtVal->use_begin()->isOnlyUserOf(ExtVal.getNode()))
14970 SDNode *U = *ExtVal->use_begin();
14971 if ((U->getOpcode() == ISD::ADD || U->getOpcode() == ISD::SUB ||
14972 U->getOpcode() == ISD::SHL || U->getOpcode() == ARMISD::VSHLIMM))
14978 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
14979 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
14982 if (!isTypeLegal(EVT::getEVT(Ty1)))
14985 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
14987 // Assuming the caller doesn't have a zeroext or signext return parameter,
14988 // truncation all the way down to i1 is valid.
14992 int ARMTargetLowering::getScalingFactorCost(const DataLayout &DL,
14993 const AddrMode &AM, Type *Ty,
14994 unsigned AS) const {
14995 if (isLegalAddressingMode(DL, AM, Ty, AS)) {
14996 if (Subtarget->hasFPAO())
14997 return AM.Scale < 0 ? 1 : 0; // positive offsets execute faster
15003 /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
15004 /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
15005 /// expanded to FMAs when this method returns true, otherwise fmuladd is
15006 /// expanded to fmul + fadd.
15008 /// ARM supports both fused and unfused multiply-add operations; we already
15009 /// lower a pair of fmul and fadd to the latter so it's not clear that there
15010 /// would be a gain or that the gain would be worthwhile enough to risk
15011 /// correctness bugs.
15013 /// For MVE, we set this to true as it helps simplify the need for some
15014 /// patterns (and we don't have the non-fused floating point instruction).
15015 bool ARMTargetLowering::isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
15017 if (!VT.isSimple())
15020 switch (VT.getSimpleVT().SimpleTy) {
15023 return Subtarget->hasMVEFloatOps();
15025 return Subtarget->useFPVFMx16();
15027 return Subtarget->useFPVFMx();
15029 return Subtarget->useFPVFMx64();
15037 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
15041 unsigned Scale = 1;
15042 switch (VT.getSimpleVT().SimpleTy) {
15052 // On thumb1 we load most things (i32, i64, floats, etc) with a LDR
15058 if ((V & (Scale - 1)) != 0)
15060 return isUInt<5>(V / Scale);
15063 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
15064 const ARMSubtarget *Subtarget) {
15065 if (!VT.isInteger() && !VT.isFloatingPoint())
15067 if (VT.isVector() && Subtarget->hasNEON())
15069 if (VT.isVector() && VT.isFloatingPoint() && Subtarget->hasMVEIntegerOps() &&
15070 !Subtarget->hasMVEFloatOps())
15073 bool IsNeg = false;
15079 unsigned NumBytes = std::max((unsigned)VT.getSizeInBits() / 8, 1U);
15081 // MVE: size * imm7
15082 if (VT.isVector() && Subtarget->hasMVEIntegerOps()) {
15083 switch (VT.getSimpleVT().getVectorElementType().SimpleTy) {
15086 return isShiftedUInt<7,2>(V);
15089 return isShiftedUInt<7,1>(V);
15091 return isUInt<7>(V);
15097 // half VLDR: 2 * imm8
15098 if (VT.isFloatingPoint() && NumBytes == 2 && Subtarget->hasFPRegs16())
15099 return isShiftedUInt<8, 1>(V);
15100 // VLDR and LDRD: 4 * imm8
15101 if ((VT.isFloatingPoint() && Subtarget->hasVFP2Base()) || NumBytes == 8)
15102 return isShiftedUInt<8, 2>(V);
15104 if (NumBytes == 1 || NumBytes == 2 || NumBytes == 4) {
15105 // + imm12 or - imm8
15107 return isUInt<8>(V);
15108 return isUInt<12>(V);
15114 /// isLegalAddressImmediate - Return true if the integer value can be used
15115 /// as the offset of the target addressing mode for load / store of the
15117 static bool isLegalAddressImmediate(int64_t V, EVT VT,
15118 const ARMSubtarget *Subtarget) {
15122 if (!VT.isSimple())
15125 if (Subtarget->isThumb1Only())
15126 return isLegalT1AddressImmediate(V, VT);
15127 else if (Subtarget->isThumb2())
15128 return isLegalT2AddressImmediate(V, VT, Subtarget);
15133 switch (VT.getSimpleVT().SimpleTy) {
15134 default: return false;
15139 return isUInt<12>(V);
15142 return isUInt<8>(V);
15145 if (!Subtarget->hasVFP2Base()) // FIXME: NEON?
15147 return isShiftedUInt<8, 2>(V);
15151 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
15153 int Scale = AM.Scale;
15157 switch (VT.getSimpleVT().SimpleTy) {
15158 default: return false;
15166 Scale = Scale & ~1;
15167 return Scale == 2 || Scale == 4 || Scale == 8;
15169 // FIXME: What are we trying to model here? ldrd doesn't have an r + r
15170 // version in Thumb mode.
15174 // r * 2 (this can be lowered to r + r).
15175 if (!AM.HasBaseReg && Scale == 2)
15179 // Note, we allow "void" uses (basically, uses that aren't loads or
15180 // stores), because arm allows folding a scale into many arithmetic
15181 // operations. This should be made more precise and revisited later.
15183 // Allow r << imm, but the imm has to be a multiple of two.
15184 if (Scale & 1) return false;
15185 return isPowerOf2_32(Scale);
15189 bool ARMTargetLowering::isLegalT1ScaledAddressingMode(const AddrMode &AM,
15191 const int Scale = AM.Scale;
15193 // Negative scales are not supported in Thumb1.
15197 // Thumb1 addressing modes do not support register scaling excepting the
15198 // following cases:
15199 // 1. Scale == 1 means no scaling.
15200 // 2. Scale == 2 this can be lowered to r + r if there is no base register.
15201 return (Scale == 1) || (!AM.HasBaseReg && Scale == 2);
15204 /// isLegalAddressingMode - Return true if the addressing mode represented
15205 /// by AM is legal for this target, for a load/store of the specified type.
15206 bool ARMTargetLowering::isLegalAddressingMode(const DataLayout &DL,
15207 const AddrMode &AM, Type *Ty,
15208 unsigned AS, Instruction *I) const {
15209 EVT VT = getValueType(DL, Ty, true);
15210 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
15213 // Can never fold addr of global into load/store.
15217 switch (AM.Scale) {
15218 case 0: // no scale reg, must be "r+i" or "r", or "i".
15221 // ARM doesn't support any R+R*scale+imm addr modes.
15225 if (!VT.isSimple())
15228 if (Subtarget->isThumb1Only())
15229 return isLegalT1ScaledAddressingMode(AM, VT);
15231 if (Subtarget->isThumb2())
15232 return isLegalT2ScaledAddressingMode(AM, VT);
15234 int Scale = AM.Scale;
15235 switch (VT.getSimpleVT().SimpleTy) {
15236 default: return false;
15240 if (Scale < 0) Scale = -Scale;
15244 return isPowerOf2_32(Scale & ~1);
15248 if (Scale == 1 || (AM.HasBaseReg && Scale == -1))
15250 // r * 2 (this can be lowered to r + r).
15251 if (!AM.HasBaseReg && Scale == 2)
15256 // Note, we allow "void" uses (basically, uses that aren't loads or
15257 // stores), because arm allows folding a scale into many arithmetic
15258 // operations. This should be made more precise and revisited later.
15260 // Allow r << imm, but the imm has to be a multiple of two.
15261 if (Scale & 1) return false;
15262 return isPowerOf2_32(Scale);
15268 /// isLegalICmpImmediate - Return true if the specified immediate is legal
15269 /// icmp immediate, that is the target has icmp instructions which can compare
15270 /// a register against the immediate without having to materialize the
15271 /// immediate into a register.
15272 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
15273 // Thumb2 and ARM modes can use cmn for negative immediates.
15274 if (!Subtarget->isThumb())
15275 return ARM_AM::getSOImmVal((uint32_t)Imm) != -1 ||
15276 ARM_AM::getSOImmVal(-(uint32_t)Imm) != -1;
15277 if (Subtarget->isThumb2())
15278 return ARM_AM::getT2SOImmVal((uint32_t)Imm) != -1 ||
15279 ARM_AM::getT2SOImmVal(-(uint32_t)Imm) != -1;
15280 // Thumb1 doesn't have cmn, and only 8-bit immediates.
15281 return Imm >= 0 && Imm <= 255;
15284 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
15285 /// *or sub* immediate, that is the target has add or sub instructions which can
15286 /// add a register with the immediate without having to materialize the
15287 /// immediate into a register.
15288 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
15289 // Same encoding for add/sub, just flip the sign.
15290 int64_t AbsImm = std::abs(Imm);
15291 if (!Subtarget->isThumb())
15292 return ARM_AM::getSOImmVal(AbsImm) != -1;
15293 if (Subtarget->isThumb2())
15294 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
15295 // Thumb1 only has 8-bit unsigned immediate.
15296 return AbsImm >= 0 && AbsImm <= 255;
15299 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
15300 bool isSEXTLoad, SDValue &Base,
15301 SDValue &Offset, bool &isInc,
15302 SelectionDAG &DAG) {
15303 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
15306 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
15307 // AddressingMode 3
15308 Base = Ptr->getOperand(0);
15309 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
15310 int RHSC = (int)RHS->getZExtValue();
15311 if (RHSC < 0 && RHSC > -256) {
15312 assert(Ptr->getOpcode() == ISD::ADD);
15314 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
15318 isInc = (Ptr->getOpcode() == ISD::ADD);
15319 Offset = Ptr->getOperand(1);
15321 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
15322 // AddressingMode 2
15323 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
15324 int RHSC = (int)RHS->getZExtValue();
15325 if (RHSC < 0 && RHSC > -0x1000) {
15326 assert(Ptr->getOpcode() == ISD::ADD);
15328 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
15329 Base = Ptr->getOperand(0);
15334 if (Ptr->getOpcode() == ISD::ADD) {
15336 ARM_AM::ShiftOpc ShOpcVal=
15337 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
15338 if (ShOpcVal != ARM_AM::no_shift) {
15339 Base = Ptr->getOperand(1);
15340 Offset = Ptr->getOperand(0);
15342 Base = Ptr->getOperand(0);
15343 Offset = Ptr->getOperand(1);
15348 isInc = (Ptr->getOpcode() == ISD::ADD);
15349 Base = Ptr->getOperand(0);
15350 Offset = Ptr->getOperand(1);
15354 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
15358 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
15359 bool isSEXTLoad, SDValue &Base,
15360 SDValue &Offset, bool &isInc,
15361 SelectionDAG &DAG) {
15362 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
15365 Base = Ptr->getOperand(0);
15366 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
15367 int RHSC = (int)RHS->getZExtValue();
15368 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
15369 assert(Ptr->getOpcode() == ISD::ADD);
15371 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
15373 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
15374 isInc = Ptr->getOpcode() == ISD::ADD;
15375 Offset = DAG.getConstant(RHSC, SDLoc(Ptr), RHS->getValueType(0));
15383 static bool getMVEIndexedAddressParts(SDNode *Ptr, EVT VT, unsigned Align,
15384 bool isSEXTLoad, bool IsMasked, bool isLE,
15385 SDValue &Base, SDValue &Offset,
15386 bool &isInc, SelectionDAG &DAG) {
15387 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
15389 if (!isa<ConstantSDNode>(Ptr->getOperand(1)))
15392 // We allow LE non-masked loads to change the type (for example use a vldrb.8
15393 // as opposed to a vldrw.32). This can allow extra addressing modes or
15394 // alignments for what is otherwise an equivalent instruction.
15395 bool CanChangeType = isLE && !IsMasked;
15397 ConstantSDNode *RHS = cast<ConstantSDNode>(Ptr->getOperand(1));
15398 int RHSC = (int)RHS->getZExtValue();
15400 auto IsInRange = [&](int RHSC, int Limit, int Scale) {
15401 if (RHSC < 0 && RHSC > -Limit * Scale && RHSC % Scale == 0) {
15402 assert(Ptr->getOpcode() == ISD::ADD);
15404 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
15406 } else if (RHSC > 0 && RHSC < Limit * Scale && RHSC % Scale == 0) {
15407 isInc = Ptr->getOpcode() == ISD::ADD;
15408 Offset = DAG.getConstant(RHSC, SDLoc(Ptr), RHS->getValueType(0));
15414 // Try to find a matching instruction based on s/zext, Alignment, Offset and
15415 // (in BE/masked) type.
15416 Base = Ptr->getOperand(0);
15417 if (VT == MVT::v4i16) {
15418 if (Align >= 2 && IsInRange(RHSC, 0x80, 2))
15420 } else if (VT == MVT::v4i8 || VT == MVT::v8i8) {
15421 if (IsInRange(RHSC, 0x80, 1))
15423 } else if (Align >= 4 &&
15424 (CanChangeType || VT == MVT::v4i32 || VT == MVT::v4f32) &&
15425 IsInRange(RHSC, 0x80, 4))
15427 else if (Align >= 2 &&
15428 (CanChangeType || VT == MVT::v8i16 || VT == MVT::v8f16) &&
15429 IsInRange(RHSC, 0x80, 2))
15431 else if ((CanChangeType || VT == MVT::v16i8) && IsInRange(RHSC, 0x80, 1))
15436 /// getPreIndexedAddressParts - returns true by value, base pointer and
15437 /// offset pointer and addressing mode by reference if the node's address
15438 /// can be legally represented as pre-indexed load / store address.
15440 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
15442 ISD::MemIndexedMode &AM,
15443 SelectionDAG &DAG) const {
15444 if (Subtarget->isThumb1Only())
15450 bool isSEXTLoad = false;
15451 bool IsMasked = false;
15452 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
15453 Ptr = LD->getBasePtr();
15454 VT = LD->getMemoryVT();
15455 Align = LD->getAlignment();
15456 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
15457 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
15458 Ptr = ST->getBasePtr();
15459 VT = ST->getMemoryVT();
15460 Align = ST->getAlignment();
15461 } else if (MaskedLoadSDNode *LD = dyn_cast<MaskedLoadSDNode>(N)) {
15462 Ptr = LD->getBasePtr();
15463 VT = LD->getMemoryVT();
15464 Align = LD->getAlignment();
15465 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
15467 } else if (MaskedStoreSDNode *ST = dyn_cast<MaskedStoreSDNode>(N)) {
15468 Ptr = ST->getBasePtr();
15469 VT = ST->getMemoryVT();
15470 Align = ST->getAlignment();
15476 bool isLegal = false;
15478 isLegal = Subtarget->hasMVEIntegerOps() &&
15479 getMVEIndexedAddressParts(Ptr.getNode(), VT, Align, isSEXTLoad,
15480 IsMasked, Subtarget->isLittle(), Base,
15481 Offset, isInc, DAG);
15483 if (Subtarget->isThumb2())
15484 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
15485 Offset, isInc, DAG);
15487 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
15488 Offset, isInc, DAG);
15493 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
15497 /// getPostIndexedAddressParts - returns true by value, base pointer and
15498 /// offset pointer and addressing mode by reference if this node can be
15499 /// combined with a load / store to form a post-indexed load / store.
15500 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
15503 ISD::MemIndexedMode &AM,
15504 SelectionDAG &DAG) const {
15508 bool isSEXTLoad = false, isNonExt;
15509 bool IsMasked = false;
15510 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
15511 VT = LD->getMemoryVT();
15512 Ptr = LD->getBasePtr();
15513 Align = LD->getAlignment();
15514 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
15515 isNonExt = LD->getExtensionType() == ISD::NON_EXTLOAD;
15516 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
15517 VT = ST->getMemoryVT();
15518 Ptr = ST->getBasePtr();
15519 Align = ST->getAlignment();
15520 isNonExt = !ST->isTruncatingStore();
15521 } else if (MaskedLoadSDNode *LD = dyn_cast<MaskedLoadSDNode>(N)) {
15522 VT = LD->getMemoryVT();
15523 Ptr = LD->getBasePtr();
15524 Align = LD->getAlignment();
15525 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
15526 isNonExt = LD->getExtensionType() == ISD::NON_EXTLOAD;
15528 } else if (MaskedStoreSDNode *ST = dyn_cast<MaskedStoreSDNode>(N)) {
15529 VT = ST->getMemoryVT();
15530 Ptr = ST->getBasePtr();
15531 Align = ST->getAlignment();
15532 isNonExt = !ST->isTruncatingStore();
15537 if (Subtarget->isThumb1Only()) {
15538 // Thumb-1 can do a limited post-inc load or store as an updating LDM. It
15539 // must be non-extending/truncating, i32, with an offset of 4.
15540 assert(Op->getValueType(0) == MVT::i32 && "Non-i32 post-inc op?!");
15541 if (Op->getOpcode() != ISD::ADD || !isNonExt)
15543 auto *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1));
15544 if (!RHS || RHS->getZExtValue() != 4)
15547 Offset = Op->getOperand(1);
15548 Base = Op->getOperand(0);
15549 AM = ISD::POST_INC;
15554 bool isLegal = false;
15556 isLegal = Subtarget->hasMVEIntegerOps() &&
15557 getMVEIndexedAddressParts(Op, VT, Align, isSEXTLoad, IsMasked,
15558 Subtarget->isLittle(), Base, Offset,
15561 if (Subtarget->isThumb2())
15562 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
15565 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
15572 // Swap base ptr and offset to catch more post-index load / store when
15573 // it's legal. In Thumb2 mode, offset must be an immediate.
15574 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
15575 !Subtarget->isThumb2())
15576 std::swap(Base, Offset);
15578 // Post-indexed load / store update the base pointer.
15583 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
15587 void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
15589 const APInt &DemandedElts,
15590 const SelectionDAG &DAG,
15591 unsigned Depth) const {
15592 unsigned BitWidth = Known.getBitWidth();
15594 switch (Op.getOpcode()) {
15600 // Special cases when we convert a carry to a boolean.
15601 if (Op.getResNo() == 0) {
15602 SDValue LHS = Op.getOperand(0);
15603 SDValue RHS = Op.getOperand(1);
15604 // (ADDE 0, 0, C) will give us a single bit.
15605 if (Op->getOpcode() == ARMISD::ADDE && isNullConstant(LHS) &&
15606 isNullConstant(RHS)) {
15607 Known.Zero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
15612 case ARMISD::CMOV: {
15613 // Bits are known zero/one if known on the LHS and RHS.
15614 Known = DAG.computeKnownBits(Op.getOperand(0), Depth+1);
15615 if (Known.isUnknown())
15618 KnownBits KnownRHS = DAG.computeKnownBits(Op.getOperand(1), Depth+1);
15619 Known.Zero &= KnownRHS.Zero;
15620 Known.One &= KnownRHS.One;
15623 case ISD::INTRINSIC_W_CHAIN: {
15624 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
15625 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
15628 case Intrinsic::arm_ldaex:
15629 case Intrinsic::arm_ldrex: {
15630 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
15631 unsigned MemBits = VT.getScalarSizeInBits();
15632 Known.Zero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
15637 case ARMISD::BFI: {
15638 // Conservatively, we can recurse down the first operand
15639 // and just mask out all affected bits.
15640 Known = DAG.computeKnownBits(Op.getOperand(0), Depth + 1);
15642 // The operand to BFI is already a mask suitable for removing the bits it
15644 ConstantSDNode *CI = cast<ConstantSDNode>(Op.getOperand(2));
15645 const APInt &Mask = CI->getAPIntValue();
15646 Known.Zero &= Mask;
15650 case ARMISD::VGETLANEs:
15651 case ARMISD::VGETLANEu: {
15652 const SDValue &SrcSV = Op.getOperand(0);
15653 EVT VecVT = SrcSV.getValueType();
15654 assert(VecVT.isVector() && "VGETLANE expected a vector type");
15655 const unsigned NumSrcElts = VecVT.getVectorNumElements();
15656 ConstantSDNode *Pos = cast<ConstantSDNode>(Op.getOperand(1).getNode());
15657 assert(Pos->getAPIntValue().ult(NumSrcElts) &&
15658 "VGETLANE index out of bounds");
15659 unsigned Idx = Pos->getZExtValue();
15660 APInt DemandedElt = APInt::getOneBitSet(NumSrcElts, Idx);
15661 Known = DAG.computeKnownBits(SrcSV, DemandedElt, Depth + 1);
15663 EVT VT = Op.getValueType();
15664 const unsigned DstSz = VT.getScalarSizeInBits();
15665 const unsigned SrcSz = VecVT.getVectorElementType().getSizeInBits();
15667 assert(SrcSz == Known.getBitWidth());
15668 assert(DstSz > SrcSz);
15669 if (Op.getOpcode() == ARMISD::VGETLANEs)
15670 Known = Known.sext(DstSz);
15672 Known = Known.zext(DstSz, true /* extended bits are known zero */);
15674 assert(DstSz == Known.getBitWidth());
15681 ARMTargetLowering::targetShrinkDemandedConstant(SDValue Op,
15682 const APInt &DemandedAPInt,
15683 TargetLoweringOpt &TLO) const {
15684 // Delay optimization, so we don't have to deal with illegal types, or block
15689 // Only optimize AND for now.
15690 if (Op.getOpcode() != ISD::AND)
15693 EVT VT = Op.getValueType();
15699 assert(VT == MVT::i32 && "Unexpected integer type");
15701 // Make sure the RHS really is a constant.
15702 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
15706 unsigned Mask = C->getZExtValue();
15708 unsigned Demanded = DemandedAPInt.getZExtValue();
15709 unsigned ShrunkMask = Mask & Demanded;
15710 unsigned ExpandedMask = Mask | ~Demanded;
15712 // If the mask is all zeros, let the target-independent code replace the
15713 // result with zero.
15714 if (ShrunkMask == 0)
15717 // If the mask is all ones, erase the AND. (Currently, the target-independent
15718 // code won't do this, so we have to do it explicitly to avoid an infinite
15719 // loop in obscure cases.)
15720 if (ExpandedMask == ~0U)
15721 return TLO.CombineTo(Op, Op.getOperand(0));
15723 auto IsLegalMask = [ShrunkMask, ExpandedMask](unsigned Mask) -> bool {
15724 return (ShrunkMask & Mask) == ShrunkMask && (~ExpandedMask & Mask) == 0;
15726 auto UseMask = [Mask, Op, VT, &TLO](unsigned NewMask) -> bool {
15727 if (NewMask == Mask)
15730 SDValue NewC = TLO.DAG.getConstant(NewMask, DL, VT);
15731 SDValue NewOp = TLO.DAG.getNode(ISD::AND, DL, VT, Op.getOperand(0), NewC);
15732 return TLO.CombineTo(Op, NewOp);
15735 // Prefer uxtb mask.
15736 if (IsLegalMask(0xFF))
15737 return UseMask(0xFF);
15739 // Prefer uxth mask.
15740 if (IsLegalMask(0xFFFF))
15741 return UseMask(0xFFFF);
15743 // [1, 255] is Thumb1 movs+ands, legal immediate for ARM/Thumb2.
15744 // FIXME: Prefer a contiguous sequence of bits for other optimizations.
15745 if (ShrunkMask < 256)
15746 return UseMask(ShrunkMask);
15748 // [-256, -2] is Thumb1 movs+bics, legal immediate for ARM/Thumb2.
15749 // FIXME: Prefer a contiguous sequence of bits for other optimizations.
15750 if ((int)ExpandedMask <= -2 && (int)ExpandedMask >= -256)
15751 return UseMask(ExpandedMask);
15753 // Potential improvements:
15755 // We could try to recognize lsls+lsrs or lsrs+lsls pairs here.
15756 // We could try to prefer Thumb1 immediates which can be lowered to a
15757 // two-instruction sequence.
15758 // We could try to recognize more legal ARM/Thumb2 immediates here.
15764 //===----------------------------------------------------------------------===//
15765 // ARM Inline Assembly Support
15766 //===----------------------------------------------------------------------===//
15768 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
15769 // Looking for "rev" which is V6+.
15770 if (!Subtarget->hasV6Ops())
15773 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
15774 std::string AsmStr = IA->getAsmString();
15775 SmallVector<StringRef, 4> AsmPieces;
15776 SplitString(AsmStr, AsmPieces, ";\n");
15778 switch (AsmPieces.size()) {
15779 default: return false;
15781 AsmStr = AsmPieces[0];
15783 SplitString(AsmStr, AsmPieces, " \t,");
15786 if (AsmPieces.size() == 3 &&
15787 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
15788 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
15789 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
15790 if (Ty && Ty->getBitWidth() == 32)
15791 return IntrinsicLowering::LowerToByteSwap(CI);
15799 const char *ARMTargetLowering::LowerXConstraint(EVT ConstraintVT) const {
15800 // At this point, we have to lower this constraint to something else, so we
15801 // lower it to an "r" or "w". However, by doing this we will force the result
15802 // to be in register, while the X constraint is much more permissive.
15804 // Although we are correct (we are free to emit anything, without
15805 // constraints), we might break use cases that would expect us to be more
15806 // efficient and emit something else.
15807 if (!Subtarget->hasVFP2Base())
15809 if (ConstraintVT.isFloatingPoint())
15811 if (ConstraintVT.isVector() && Subtarget->hasNEON() &&
15812 (ConstraintVT.getSizeInBits() == 64 ||
15813 ConstraintVT.getSizeInBits() == 128))
15819 /// getConstraintType - Given a constraint letter, return the type of
15820 /// constraint it is for this target.
15821 ARMTargetLowering::ConstraintType
15822 ARMTargetLowering::getConstraintType(StringRef Constraint) const {
15823 unsigned S = Constraint.size();
15825 switch (Constraint[0]) {
15827 case 'l': return C_RegisterClass;
15828 case 'w': return C_RegisterClass;
15829 case 'h': return C_RegisterClass;
15830 case 'x': return C_RegisterClass;
15831 case 't': return C_RegisterClass;
15832 case 'j': return C_Immediate; // Constant for movw.
15833 // An address with a single base register. Due to the way we
15834 // currently handle addresses it is the same as an 'r' memory constraint.
15835 case 'Q': return C_Memory;
15837 } else if (S == 2) {
15838 switch (Constraint[0]) {
15840 case 'T': return C_RegisterClass;
15841 // All 'U+' constraints are addresses.
15842 case 'U': return C_Memory;
15845 return TargetLowering::getConstraintType(Constraint);
15848 /// Examine constraint type and operand type and determine a weight value.
15849 /// This object must already have been set up with the operand type
15850 /// and the current alternative constraint selected.
15851 TargetLowering::ConstraintWeight
15852 ARMTargetLowering::getSingleConstraintMatchWeight(
15853 AsmOperandInfo &info, const char *constraint) const {
15854 ConstraintWeight weight = CW_Invalid;
15855 Value *CallOperandVal = info.CallOperandVal;
15856 // If we don't have a value, we can't do a match,
15857 // but allow it at the lowest weight.
15858 if (!CallOperandVal)
15860 Type *type = CallOperandVal->getType();
15861 // Look at the constraint type.
15862 switch (*constraint) {
15864 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
15867 if (type->isIntegerTy()) {
15868 if (Subtarget->isThumb())
15869 weight = CW_SpecificReg;
15871 weight = CW_Register;
15875 if (type->isFloatingPointTy())
15876 weight = CW_Register;
15882 using RCPair = std::pair<unsigned, const TargetRegisterClass *>;
15884 RCPair ARMTargetLowering::getRegForInlineAsmConstraint(
15885 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
15886 switch (Constraint.size()) {
15888 // GCC ARM Constraint Letters
15889 switch (Constraint[0]) {
15890 case 'l': // Low regs or general regs.
15891 if (Subtarget->isThumb())
15892 return RCPair(0U, &ARM::tGPRRegClass);
15893 return RCPair(0U, &ARM::GPRRegClass);
15894 case 'h': // High regs or no regs.
15895 if (Subtarget->isThumb())
15896 return RCPair(0U, &ARM::hGPRRegClass);
15899 if (Subtarget->isThumb1Only())
15900 return RCPair(0U, &ARM::tGPRRegClass);
15901 return RCPair(0U, &ARM::GPRRegClass);
15903 if (VT == MVT::Other)
15905 if (VT == MVT::f32)
15906 return RCPair(0U, &ARM::SPRRegClass);
15907 if (VT.getSizeInBits() == 64)
15908 return RCPair(0U, &ARM::DPRRegClass);
15909 if (VT.getSizeInBits() == 128)
15910 return RCPair(0U, &ARM::QPRRegClass);
15913 if (VT == MVT::Other)
15915 if (VT == MVT::f32)
15916 return RCPair(0U, &ARM::SPR_8RegClass);
15917 if (VT.getSizeInBits() == 64)
15918 return RCPair(0U, &ARM::DPR_8RegClass);
15919 if (VT.getSizeInBits() == 128)
15920 return RCPair(0U, &ARM::QPR_8RegClass);
15923 if (VT == MVT::Other)
15925 if (VT == MVT::f32 || VT == MVT::i32)
15926 return RCPair(0U, &ARM::SPRRegClass);
15927 if (VT.getSizeInBits() == 64)
15928 return RCPair(0U, &ARM::DPR_VFP2RegClass);
15929 if (VT.getSizeInBits() == 128)
15930 return RCPair(0U, &ARM::QPR_VFP2RegClass);
15936 if (Constraint[0] == 'T') {
15937 switch (Constraint[1]) {
15941 return RCPair(0U, &ARM::tGPREvenRegClass);
15943 return RCPair(0U, &ARM::tGPROddRegClass);
15952 if (StringRef("{cc}").equals_lower(Constraint))
15953 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
15955 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
15958 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
15959 /// vector. If it is invalid, don't add anything to Ops.
15960 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
15961 std::string &Constraint,
15962 std::vector<SDValue>&Ops,
15963 SelectionDAG &DAG) const {
15966 // Currently only support length 1 constraints.
15967 if (Constraint.length() != 1) return;
15969 char ConstraintLetter = Constraint[0];
15970 switch (ConstraintLetter) {
15973 case 'I': case 'J': case 'K': case 'L':
15974 case 'M': case 'N': case 'O':
15975 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
15979 int64_t CVal64 = C->getSExtValue();
15980 int CVal = (int) CVal64;
15981 // None of these constraints allow values larger than 32 bits. Check
15982 // that the value fits in an int.
15983 if (CVal != CVal64)
15986 switch (ConstraintLetter) {
15988 // Constant suitable for movw, must be between 0 and
15990 if (Subtarget->hasV6T2Ops() || (Subtarget->hasV8MBaselineOps()))
15991 if (CVal >= 0 && CVal <= 65535)
15995 if (Subtarget->isThumb1Only()) {
15996 // This must be a constant between 0 and 255, for ADD
15998 if (CVal >= 0 && CVal <= 255)
16000 } else if (Subtarget->isThumb2()) {
16001 // A constant that can be used as an immediate value in a
16002 // data-processing instruction.
16003 if (ARM_AM::getT2SOImmVal(CVal) != -1)
16006 // A constant that can be used as an immediate value in a
16007 // data-processing instruction.
16008 if (ARM_AM::getSOImmVal(CVal) != -1)
16014 if (Subtarget->isThumb1Only()) {
16015 // This must be a constant between -255 and -1, for negated ADD
16016 // immediates. This can be used in GCC with an "n" modifier that
16017 // prints the negated value, for use with SUB instructions. It is
16018 // not useful otherwise but is implemented for compatibility.
16019 if (CVal >= -255 && CVal <= -1)
16022 // This must be a constant between -4095 and 4095. It is not clear
16023 // what this constraint is intended for. Implemented for
16024 // compatibility with GCC.
16025 if (CVal >= -4095 && CVal <= 4095)
16031 if (Subtarget->isThumb1Only()) {
16032 // A 32-bit value where only one byte has a nonzero value. Exclude
16033 // zero to match GCC. This constraint is used by GCC internally for
16034 // constants that can be loaded with a move/shift combination.
16035 // It is not useful otherwise but is implemented for compatibility.
16036 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
16038 } else if (Subtarget->isThumb2()) {
16039 // A constant whose bitwise inverse can be used as an immediate
16040 // value in a data-processing instruction. This can be used in GCC
16041 // with a "B" modifier that prints the inverted value, for use with
16042 // BIC and MVN instructions. It is not useful otherwise but is
16043 // implemented for compatibility.
16044 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
16047 // A constant whose bitwise inverse can be used as an immediate
16048 // value in a data-processing instruction. This can be used in GCC
16049 // with a "B" modifier that prints the inverted value, for use with
16050 // BIC and MVN instructions. It is not useful otherwise but is
16051 // implemented for compatibility.
16052 if (ARM_AM::getSOImmVal(~CVal) != -1)
16058 if (Subtarget->isThumb1Only()) {
16059 // This must be a constant between -7 and 7,
16060 // for 3-operand ADD/SUB immediate instructions.
16061 if (CVal >= -7 && CVal < 7)
16063 } else if (Subtarget->isThumb2()) {
16064 // A constant whose negation can be used as an immediate value in a
16065 // data-processing instruction. This can be used in GCC with an "n"
16066 // modifier that prints the negated value, for use with SUB
16067 // instructions. It is not useful otherwise but is implemented for
16069 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
16072 // A constant whose negation can be used as an immediate value in a
16073 // data-processing instruction. This can be used in GCC with an "n"
16074 // modifier that prints the negated value, for use with SUB
16075 // instructions. It is not useful otherwise but is implemented for
16077 if (ARM_AM::getSOImmVal(-CVal) != -1)
16083 if (Subtarget->isThumb1Only()) {
16084 // This must be a multiple of 4 between 0 and 1020, for
16085 // ADD sp + immediate.
16086 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
16089 // A power of two or a constant between 0 and 32. This is used in
16090 // GCC for the shift amount on shifted register operands, but it is
16091 // useful in general for any shift amounts.
16092 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
16098 if (Subtarget->isThumb1Only()) {
16099 // This must be a constant between 0 and 31, for shift amounts.
16100 if (CVal >= 0 && CVal <= 31)
16106 if (Subtarget->isThumb1Only()) {
16107 // This must be a multiple of 4 between -508 and 508, for
16108 // ADD/SUB sp = sp + immediate.
16109 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
16114 Result = DAG.getTargetConstant(CVal, SDLoc(Op), Op.getValueType());
16118 if (Result.getNode()) {
16119 Ops.push_back(Result);
16122 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
16125 static RTLIB::Libcall getDivRemLibcall(
16126 const SDNode *N, MVT::SimpleValueType SVT) {
16127 assert((N->getOpcode() == ISD::SDIVREM || N->getOpcode() == ISD::UDIVREM ||
16128 N->getOpcode() == ISD::SREM || N->getOpcode() == ISD::UREM) &&
16129 "Unhandled Opcode in getDivRemLibcall");
16130 bool isSigned = N->getOpcode() == ISD::SDIVREM ||
16131 N->getOpcode() == ISD::SREM;
16134 default: llvm_unreachable("Unexpected request for libcall!");
16135 case MVT::i8: LC = isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
16136 case MVT::i16: LC = isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
16137 case MVT::i32: LC = isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
16138 case MVT::i64: LC = isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
16143 static TargetLowering::ArgListTy getDivRemArgList(
16144 const SDNode *N, LLVMContext *Context, const ARMSubtarget *Subtarget) {
16145 assert((N->getOpcode() == ISD::SDIVREM || N->getOpcode() == ISD::UDIVREM ||
16146 N->getOpcode() == ISD::SREM || N->getOpcode() == ISD::UREM) &&
16147 "Unhandled Opcode in getDivRemArgList");
16148 bool isSigned = N->getOpcode() == ISD::SDIVREM ||
16149 N->getOpcode() == ISD::SREM;
16150 TargetLowering::ArgListTy Args;
16151 TargetLowering::ArgListEntry Entry;
16152 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
16153 EVT ArgVT = N->getOperand(i).getValueType();
16154 Type *ArgTy = ArgVT.getTypeForEVT(*Context);
16155 Entry.Node = N->getOperand(i);
16157 Entry.IsSExt = isSigned;
16158 Entry.IsZExt = !isSigned;
16159 Args.push_back(Entry);
16161 if (Subtarget->isTargetWindows() && Args.size() >= 2)
16162 std::swap(Args[0], Args[1]);
16166 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
16167 assert((Subtarget->isTargetAEABI() || Subtarget->isTargetAndroid() ||
16168 Subtarget->isTargetGNUAEABI() || Subtarget->isTargetMuslAEABI() ||
16169 Subtarget->isTargetWindows()) &&
16170 "Register-based DivRem lowering only");
16171 unsigned Opcode = Op->getOpcode();
16172 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
16173 "Invalid opcode for Div/Rem lowering");
16174 bool isSigned = (Opcode == ISD::SDIVREM);
16175 EVT VT = Op->getValueType(0);
16176 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
16179 // If the target has hardware divide, use divide + multiply + subtract:
16181 // rem = a - b * div
16182 // return {div, rem}
16183 // This should be lowered into UDIV/SDIV + MLS later on.
16184 bool hasDivide = Subtarget->isThumb() ? Subtarget->hasDivideInThumbMode()
16185 : Subtarget->hasDivideInARMMode();
16186 if (hasDivide && Op->getValueType(0).isSimple() &&
16187 Op->getSimpleValueType(0) == MVT::i32) {
16188 unsigned DivOpcode = isSigned ? ISD::SDIV : ISD::UDIV;
16189 const SDValue Dividend = Op->getOperand(0);
16190 const SDValue Divisor = Op->getOperand(1);
16191 SDValue Div = DAG.getNode(DivOpcode, dl, VT, Dividend, Divisor);
16192 SDValue Mul = DAG.getNode(ISD::MUL, dl, VT, Div, Divisor);
16193 SDValue Rem = DAG.getNode(ISD::SUB, dl, VT, Dividend, Mul);
16195 SDValue Values[2] = {Div, Rem};
16196 return DAG.getNode(ISD::MERGE_VALUES, dl, DAG.getVTList(VT, VT), Values);
16199 RTLIB::Libcall LC = getDivRemLibcall(Op.getNode(),
16200 VT.getSimpleVT().SimpleTy);
16201 SDValue InChain = DAG.getEntryNode();
16203 TargetLowering::ArgListTy Args = getDivRemArgList(Op.getNode(),
16207 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
16208 getPointerTy(DAG.getDataLayout()));
16210 Type *RetTy = StructType::get(Ty, Ty);
16212 if (Subtarget->isTargetWindows())
16213 InChain = WinDBZCheckDenominator(DAG, Op.getNode(), InChain);
16215 TargetLowering::CallLoweringInfo CLI(DAG);
16216 CLI.setDebugLoc(dl).setChain(InChain)
16217 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args))
16218 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
16220 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
16221 return CallInfo.first;
16224 // Lowers REM using divmod helpers
16225 // see RTABI section 4.2/4.3
16226 SDValue ARMTargetLowering::LowerREM(SDNode *N, SelectionDAG &DAG) const {
16227 // Build return types (div and rem)
16228 std::vector<Type*> RetTyParams;
16229 Type *RetTyElement;
16231 switch (N->getValueType(0).getSimpleVT().SimpleTy) {
16232 default: llvm_unreachable("Unexpected request for libcall!");
16233 case MVT::i8: RetTyElement = Type::getInt8Ty(*DAG.getContext()); break;
16234 case MVT::i16: RetTyElement = Type::getInt16Ty(*DAG.getContext()); break;
16235 case MVT::i32: RetTyElement = Type::getInt32Ty(*DAG.getContext()); break;
16236 case MVT::i64: RetTyElement = Type::getInt64Ty(*DAG.getContext()); break;
16239 RetTyParams.push_back(RetTyElement);
16240 RetTyParams.push_back(RetTyElement);
16241 ArrayRef<Type*> ret = ArrayRef<Type*>(RetTyParams);
16242 Type *RetTy = StructType::get(*DAG.getContext(), ret);
16244 RTLIB::Libcall LC = getDivRemLibcall(N, N->getValueType(0).getSimpleVT().
16246 SDValue InChain = DAG.getEntryNode();
16247 TargetLowering::ArgListTy Args = getDivRemArgList(N, DAG.getContext(),
16249 bool isSigned = N->getOpcode() == ISD::SREM;
16250 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
16251 getPointerTy(DAG.getDataLayout()));
16253 if (Subtarget->isTargetWindows())
16254 InChain = WinDBZCheckDenominator(DAG, N, InChain);
16257 CallLoweringInfo CLI(DAG);
16258 CLI.setChain(InChain)
16259 .setCallee(CallingConv::ARM_AAPCS, RetTy, Callee, std::move(Args))
16260 .setSExtResult(isSigned).setZExtResult(!isSigned).setDebugLoc(SDLoc(N));
16261 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
16263 // Return second (rem) result operand (first contains div)
16264 SDNode *ResNode = CallResult.first.getNode();
16265 assert(ResNode->getNumOperands() == 2 && "divmod should return two operands");
16266 return ResNode->getOperand(1);
16270 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
16271 assert(Subtarget->isTargetWindows() && "unsupported target platform");
16275 SDValue Chain = Op.getOperand(0);
16276 SDValue Size = Op.getOperand(1);
16278 if (DAG.getMachineFunction().getFunction().hasFnAttribute(
16279 "no-stack-arg-probe")) {
16280 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
16281 SDValue SP = DAG.getCopyFromReg(Chain, DL, ARM::SP, MVT::i32);
16282 Chain = SP.getValue(1);
16283 SP = DAG.getNode(ISD::SUB, DL, MVT::i32, SP, Size);
16285 SP = DAG.getNode(ISD::AND, DL, MVT::i32, SP.getValue(0),
16286 DAG.getConstant(-(uint64_t)Align, DL, MVT::i32));
16287 Chain = DAG.getCopyToReg(Chain, DL, ARM::SP, SP);
16288 SDValue Ops[2] = { SP, Chain };
16289 return DAG.getMergeValues(Ops, DL);
16292 SDValue Words = DAG.getNode(ISD::SRL, DL, MVT::i32, Size,
16293 DAG.getConstant(2, DL, MVT::i32));
16296 Chain = DAG.getCopyToReg(Chain, DL, ARM::R4, Words, Flag);
16297 Flag = Chain.getValue(1);
16299 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
16300 Chain = DAG.getNode(ARMISD::WIN__CHKSTK, DL, NodeTys, Chain, Flag);
16302 SDValue NewSP = DAG.getCopyFromReg(Chain, DL, ARM::SP, MVT::i32);
16303 Chain = NewSP.getValue(1);
16305 SDValue Ops[2] = { NewSP, Chain };
16306 return DAG.getMergeValues(Ops, DL);
16309 SDValue ARMTargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
16310 bool IsStrict = Op->isStrictFPOpcode();
16311 SDValue SrcVal = Op.getOperand(IsStrict ? 1 : 0);
16312 const unsigned DstSz = Op.getValueType().getSizeInBits();
16313 const unsigned SrcSz = SrcVal.getValueType().getSizeInBits();
16314 assert(DstSz > SrcSz && DstSz <= 64 && SrcSz >= 16 &&
16315 "Unexpected type for custom-lowering FP_EXTEND");
16317 assert((!Subtarget->hasFP64() || !Subtarget->hasFPARMv8Base()) &&
16318 "With both FP DP and 16, any FP conversion is legal!");
16320 assert(!(DstSz == 32 && Subtarget->hasFP16()) &&
16321 "With FP16, 16 to 32 conversion is legal!");
16323 // Either we are converting from 16 -> 64, without FP16 and/or
16324 // FP.double-precision or without Armv8-fp. So we must do it in two
16326 // Or we are converting from 32 -> 64 without fp.double-precision or 16 -> 32
16327 // without FP16. So we must do a function call.
16330 MakeLibCallOptions CallOptions;
16331 SDValue Chain = IsStrict ? Op.getOperand(0) : SDValue();
16332 for (unsigned Sz = SrcSz; Sz <= 32 && Sz < DstSz; Sz *= 2) {
16333 bool Supported = (Sz == 16 ? Subtarget->hasFP16() : Subtarget->hasFP64());
16334 MVT SrcVT = (Sz == 16 ? MVT::f16 : MVT::f32);
16335 MVT DstVT = (Sz == 16 ? MVT::f32 : MVT::f64);
16338 SrcVal = DAG.getNode(ISD::STRICT_FP_EXTEND, Loc,
16339 {DstVT, MVT::Other}, {Chain, SrcVal});
16340 Chain = SrcVal.getValue(1);
16342 SrcVal = DAG.getNode(ISD::FP_EXTEND, Loc, DstVT, SrcVal);
16345 LC = RTLIB::getFPEXT(SrcVT, DstVT);
16346 assert(LC != RTLIB::UNKNOWN_LIBCALL &&
16347 "Unexpected type for custom-lowering FP_EXTEND");
16348 std::tie(SrcVal, Chain) = makeLibCall(DAG, LC, DstVT, SrcVal, CallOptions,
16353 return IsStrict ? DAG.getMergeValues({SrcVal, Chain}, Loc) : SrcVal;
16356 SDValue ARMTargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
16357 bool IsStrict = Op->isStrictFPOpcode();
16359 SDValue SrcVal = Op.getOperand(IsStrict ? 1 : 0);
16360 EVT SrcVT = SrcVal.getValueType();
16361 EVT DstVT = Op.getValueType();
16362 const unsigned DstSz = Op.getValueType().getSizeInBits();
16363 const unsigned SrcSz = SrcVT.getSizeInBits();
16365 assert(DstSz < SrcSz && SrcSz <= 64 && DstSz >= 16 &&
16366 "Unexpected type for custom-lowering FP_ROUND");
16368 assert((!Subtarget->hasFP64() || !Subtarget->hasFPARMv8Base()) &&
16369 "With both FP DP and 16, any FP conversion is legal!");
16373 // Instruction from 32 -> 16 if hasFP16 is valid
16374 if (SrcSz == 32 && Subtarget->hasFP16())
16377 // Lib call from 32 -> 16 / 64 -> [32, 16]
16378 RTLIB::Libcall LC = RTLIB::getFPROUND(SrcVT, DstVT);
16379 assert(LC != RTLIB::UNKNOWN_LIBCALL &&
16380 "Unexpected type for custom-lowering FP_ROUND");
16381 MakeLibCallOptions CallOptions;
16382 SDValue Chain = IsStrict ? Op.getOperand(0) : SDValue();
16384 std::tie(Result, Chain) = makeLibCall(DAG, LC, DstVT, SrcVal, CallOptions,
16386 return IsStrict ? DAG.getMergeValues({Result, Chain}, Loc) : Result;
16389 void ARMTargetLowering::lowerABS(SDNode *N, SmallVectorImpl<SDValue> &Results,
16390 SelectionDAG &DAG) const {
16391 assert(N->getValueType(0) == MVT::i64 && "Unexpected type (!= i64) on ABS.");
16392 MVT HalfT = MVT::i32;
16394 SDValue Hi, Lo, Tmp;
16396 if (!isOperationLegalOrCustom(ISD::ADDCARRY, HalfT) ||
16397 !isOperationLegalOrCustom(ISD::UADDO, HalfT))
16400 unsigned OpTypeBits = HalfT.getScalarSizeInBits();
16401 SDVTList VTList = DAG.getVTList(HalfT, MVT::i1);
16403 Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(0),
16404 DAG.getConstant(0, dl, HalfT));
16405 Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(0),
16406 DAG.getConstant(1, dl, HalfT));
16408 Tmp = DAG.getNode(ISD::SRA, dl, HalfT, Hi,
16409 DAG.getConstant(OpTypeBits - 1, dl,
16410 getShiftAmountTy(HalfT, DAG.getDataLayout())));
16411 Lo = DAG.getNode(ISD::UADDO, dl, VTList, Tmp, Lo);
16412 Hi = DAG.getNode(ISD::ADDCARRY, dl, VTList, Tmp, Hi,
16413 SDValue(Lo.getNode(), 1));
16414 Hi = DAG.getNode(ISD::XOR, dl, HalfT, Tmp, Hi);
16415 Lo = DAG.getNode(ISD::XOR, dl, HalfT, Tmp, Lo);
16417 Results.push_back(Lo);
16418 Results.push_back(Hi);
16422 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
16423 // The ARM target isn't yet aware of offsets.
16427 bool ARM::isBitFieldInvertedMask(unsigned v) {
16428 if (v == 0xffffffff)
16431 // there can be 1's on either or both "outsides", all the "inside"
16432 // bits must be 0's
16433 return isShiftedMask_32(~v);
16436 /// isFPImmLegal - Returns true if the target can instruction select the
16437 /// specified FP immediate natively. If false, the legalizer will
16438 /// materialize the FP immediate as a load from a constant pool.
16439 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT,
16440 bool ForCodeSize) const {
16441 if (!Subtarget->hasVFP3Base())
16443 if (VT == MVT::f16 && Subtarget->hasFullFP16())
16444 return ARM_AM::getFP16Imm(Imm) != -1;
16445 if (VT == MVT::f32)
16446 return ARM_AM::getFP32Imm(Imm) != -1;
16447 if (VT == MVT::f64 && Subtarget->hasFP64())
16448 return ARM_AM::getFP64Imm(Imm) != -1;
16452 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
16453 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
16454 /// specified in the intrinsic calls.
16455 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
16457 MachineFunction &MF,
16458 unsigned Intrinsic) const {
16459 switch (Intrinsic) {
16460 case Intrinsic::arm_neon_vld1:
16461 case Intrinsic::arm_neon_vld2:
16462 case Intrinsic::arm_neon_vld3:
16463 case Intrinsic::arm_neon_vld4:
16464 case Intrinsic::arm_neon_vld2lane:
16465 case Intrinsic::arm_neon_vld3lane:
16466 case Intrinsic::arm_neon_vld4lane:
16467 case Intrinsic::arm_neon_vld2dup:
16468 case Intrinsic::arm_neon_vld3dup:
16469 case Intrinsic::arm_neon_vld4dup: {
16470 Info.opc = ISD::INTRINSIC_W_CHAIN;
16471 // Conservatively set memVT to the entire set of vectors loaded.
16472 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
16473 uint64_t NumElts = DL.getTypeSizeInBits(I.getType()) / 64;
16474 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
16475 Info.ptrVal = I.getArgOperand(0);
16477 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
16478 Info.align = MaybeAlign(cast<ConstantInt>(AlignArg)->getZExtValue());
16479 // volatile loads with NEON intrinsics not supported
16480 Info.flags = MachineMemOperand::MOLoad;
16483 case Intrinsic::arm_neon_vld1x2:
16484 case Intrinsic::arm_neon_vld1x3:
16485 case Intrinsic::arm_neon_vld1x4: {
16486 Info.opc = ISD::INTRINSIC_W_CHAIN;
16487 // Conservatively set memVT to the entire set of vectors loaded.
16488 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
16489 uint64_t NumElts = DL.getTypeSizeInBits(I.getType()) / 64;
16490 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
16491 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
16493 Info.align.reset();
16494 // volatile loads with NEON intrinsics not supported
16495 Info.flags = MachineMemOperand::MOLoad;
16498 case Intrinsic::arm_neon_vst1:
16499 case Intrinsic::arm_neon_vst2:
16500 case Intrinsic::arm_neon_vst3:
16501 case Intrinsic::arm_neon_vst4:
16502 case Intrinsic::arm_neon_vst2lane:
16503 case Intrinsic::arm_neon_vst3lane:
16504 case Intrinsic::arm_neon_vst4lane: {
16505 Info.opc = ISD::INTRINSIC_VOID;
16506 // Conservatively set memVT to the entire set of vectors stored.
16507 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
16508 unsigned NumElts = 0;
16509 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
16510 Type *ArgTy = I.getArgOperand(ArgI)->getType();
16511 if (!ArgTy->isVectorTy())
16513 NumElts += DL.getTypeSizeInBits(ArgTy) / 64;
16515 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
16516 Info.ptrVal = I.getArgOperand(0);
16518 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
16519 Info.align = MaybeAlign(cast<ConstantInt>(AlignArg)->getZExtValue());
16520 // volatile stores with NEON intrinsics not supported
16521 Info.flags = MachineMemOperand::MOStore;
16524 case Intrinsic::arm_neon_vst1x2:
16525 case Intrinsic::arm_neon_vst1x3:
16526 case Intrinsic::arm_neon_vst1x4: {
16527 Info.opc = ISD::INTRINSIC_VOID;
16528 // Conservatively set memVT to the entire set of vectors stored.
16529 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
16530 unsigned NumElts = 0;
16531 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
16532 Type *ArgTy = I.getArgOperand(ArgI)->getType();
16533 if (!ArgTy->isVectorTy())
16535 NumElts += DL.getTypeSizeInBits(ArgTy) / 64;
16537 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
16538 Info.ptrVal = I.getArgOperand(0);
16540 Info.align.reset();
16541 // volatile stores with NEON intrinsics not supported
16542 Info.flags = MachineMemOperand::MOStore;
16545 case Intrinsic::arm_ldaex:
16546 case Intrinsic::arm_ldrex: {
16547 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
16548 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
16549 Info.opc = ISD::INTRINSIC_W_CHAIN;
16550 Info.memVT = MVT::getVT(PtrTy->getElementType());
16551 Info.ptrVal = I.getArgOperand(0);
16553 Info.align = MaybeAlign(DL.getABITypeAlignment(PtrTy->getElementType()));
16554 Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MOVolatile;
16557 case Intrinsic::arm_stlex:
16558 case Intrinsic::arm_strex: {
16559 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
16560 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
16561 Info.opc = ISD::INTRINSIC_W_CHAIN;
16562 Info.memVT = MVT::getVT(PtrTy->getElementType());
16563 Info.ptrVal = I.getArgOperand(1);
16565 Info.align = MaybeAlign(DL.getABITypeAlignment(PtrTy->getElementType()));
16566 Info.flags = MachineMemOperand::MOStore | MachineMemOperand::MOVolatile;
16569 case Intrinsic::arm_stlexd:
16570 case Intrinsic::arm_strexd:
16571 Info.opc = ISD::INTRINSIC_W_CHAIN;
16572 Info.memVT = MVT::i64;
16573 Info.ptrVal = I.getArgOperand(2);
16575 Info.align = Align(8);
16576 Info.flags = MachineMemOperand::MOStore | MachineMemOperand::MOVolatile;
16579 case Intrinsic::arm_ldaexd:
16580 case Intrinsic::arm_ldrexd:
16581 Info.opc = ISD::INTRINSIC_W_CHAIN;
16582 Info.memVT = MVT::i64;
16583 Info.ptrVal = I.getArgOperand(0);
16585 Info.align = Align(8);
16586 Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MOVolatile;
16596 /// Returns true if it is beneficial to convert a load of a constant
16597 /// to just the constant itself.
16598 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
16600 assert(Ty->isIntegerTy());
16602 unsigned Bits = Ty->getPrimitiveSizeInBits();
16603 if (Bits == 0 || Bits > 32)
16608 bool ARMTargetLowering::isExtractSubvectorCheap(EVT ResVT, EVT SrcVT,
16609 unsigned Index) const {
16610 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
16613 return (Index == 0 || Index == ResVT.getVectorNumElements());
16616 Instruction* ARMTargetLowering::makeDMB(IRBuilder<> &Builder,
16617 ARM_MB::MemBOpt Domain) const {
16618 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
16620 // First, if the target has no DMB, see what fallback we can use.
16621 if (!Subtarget->hasDataBarrier()) {
16622 // Some ARMv6 cpus can support data barriers with an mcr instruction.
16623 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
16625 if (Subtarget->hasV6Ops() && !Subtarget->isThumb()) {
16626 Function *MCR = Intrinsic::getDeclaration(M, Intrinsic::arm_mcr);
16627 Value* args[6] = {Builder.getInt32(15), Builder.getInt32(0),
16628 Builder.getInt32(0), Builder.getInt32(7),
16629 Builder.getInt32(10), Builder.getInt32(5)};
16630 return Builder.CreateCall(MCR, args);
16632 // Instead of using barriers, atomic accesses on these subtargets use
16634 llvm_unreachable("makeDMB on a target so old that it has no barriers");
16637 Function *DMB = Intrinsic::getDeclaration(M, Intrinsic::arm_dmb);
16638 // Only a full system barrier exists in the M-class architectures.
16639 Domain = Subtarget->isMClass() ? ARM_MB::SY : Domain;
16640 Constant *CDomain = Builder.getInt32(Domain);
16641 return Builder.CreateCall(DMB, CDomain);
16645 // Based on http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html
16646 Instruction *ARMTargetLowering::emitLeadingFence(IRBuilder<> &Builder,
16648 AtomicOrdering Ord) const {
16650 case AtomicOrdering::NotAtomic:
16651 case AtomicOrdering::Unordered:
16652 llvm_unreachable("Invalid fence: unordered/non-atomic");
16653 case AtomicOrdering::Monotonic:
16654 case AtomicOrdering::Acquire:
16655 return nullptr; // Nothing to do
16656 case AtomicOrdering::SequentiallyConsistent:
16657 if (!Inst->hasAtomicStore())
16658 return nullptr; // Nothing to do
16660 case AtomicOrdering::Release:
16661 case AtomicOrdering::AcquireRelease:
16662 if (Subtarget->preferISHSTBarriers())
16663 return makeDMB(Builder, ARM_MB::ISHST);
16664 // FIXME: add a comment with a link to documentation justifying this.
16666 return makeDMB(Builder, ARM_MB::ISH);
16668 llvm_unreachable("Unknown fence ordering in emitLeadingFence");
16671 Instruction *ARMTargetLowering::emitTrailingFence(IRBuilder<> &Builder,
16673 AtomicOrdering Ord) const {
16675 case AtomicOrdering::NotAtomic:
16676 case AtomicOrdering::Unordered:
16677 llvm_unreachable("Invalid fence: unordered/not-atomic");
16678 case AtomicOrdering::Monotonic:
16679 case AtomicOrdering::Release:
16680 return nullptr; // Nothing to do
16681 case AtomicOrdering::Acquire:
16682 case AtomicOrdering::AcquireRelease:
16683 case AtomicOrdering::SequentiallyConsistent:
16684 return makeDMB(Builder, ARM_MB::ISH);
16686 llvm_unreachable("Unknown fence ordering in emitTrailingFence");
16689 // Loads and stores less than 64-bits are already atomic; ones above that
16690 // are doomed anyway, so defer to the default libcall and blame the OS when
16691 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
16692 // anything for those.
16693 bool ARMTargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
16694 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
16695 return (Size == 64) && !Subtarget->isMClass();
16698 // Loads and stores less than 64-bits are already atomic; ones above that
16699 // are doomed anyway, so defer to the default libcall and blame the OS when
16700 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
16701 // anything for those.
16702 // FIXME: ldrd and strd are atomic if the CPU has LPAE (e.g. A15 has that
16703 // guarantee, see DDI0406C ARM architecture reference manual,
16704 // sections A8.8.72-74 LDRD)
16705 TargetLowering::AtomicExpansionKind
16706 ARMTargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
16707 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
16708 return ((Size == 64) && !Subtarget->isMClass()) ? AtomicExpansionKind::LLOnly
16709 : AtomicExpansionKind::None;
16712 // For the real atomic operations, we have ldrex/strex up to 32 bits,
16713 // and up to 64 bits on the non-M profiles
16714 TargetLowering::AtomicExpansionKind
16715 ARMTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
16716 if (AI->isFloatingPointOperation())
16717 return AtomicExpansionKind::CmpXChg;
16719 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
16720 bool hasAtomicRMW = !Subtarget->isThumb() || Subtarget->hasV8MBaselineOps();
16721 return (Size <= (Subtarget->isMClass() ? 32U : 64U) && hasAtomicRMW)
16722 ? AtomicExpansionKind::LLSC
16723 : AtomicExpansionKind::None;
16726 TargetLowering::AtomicExpansionKind
16727 ARMTargetLowering::shouldExpandAtomicCmpXchgInIR(AtomicCmpXchgInst *AI) const {
16728 // At -O0, fast-regalloc cannot cope with the live vregs necessary to
16729 // implement cmpxchg without spilling. If the address being exchanged is also
16730 // on the stack and close enough to the spill slot, this can lead to a
16731 // situation where the monitor always gets cleared and the atomic operation
16732 // can never succeed. So at -O0 we need a late-expanded pseudo-inst instead.
16733 bool HasAtomicCmpXchg =
16734 !Subtarget->isThumb() || Subtarget->hasV8MBaselineOps();
16735 if (getTargetMachine().getOptLevel() != 0 && HasAtomicCmpXchg)
16736 return AtomicExpansionKind::LLSC;
16737 return AtomicExpansionKind::None;
16740 bool ARMTargetLowering::shouldInsertFencesForAtomic(
16741 const Instruction *I) const {
16742 return InsertFencesForAtomic;
16745 // This has so far only been implemented for MachO.
16746 bool ARMTargetLowering::useLoadStackGuardNode() const {
16747 return Subtarget->isTargetMachO();
16750 void ARMTargetLowering::insertSSPDeclarations(Module &M) const {
16751 if (!Subtarget->getTargetTriple().isWindowsMSVCEnvironment())
16752 return TargetLowering::insertSSPDeclarations(M);
16754 // MSVC CRT has a global variable holding security cookie.
16755 M.getOrInsertGlobal("__security_cookie",
16756 Type::getInt8PtrTy(M.getContext()));
16758 // MSVC CRT has a function to validate security cookie.
16759 FunctionCallee SecurityCheckCookie = M.getOrInsertFunction(
16760 "__security_check_cookie", Type::getVoidTy(M.getContext()),
16761 Type::getInt8PtrTy(M.getContext()));
16762 if (Function *F = dyn_cast<Function>(SecurityCheckCookie.getCallee()))
16763 F->addAttribute(1, Attribute::AttrKind::InReg);
16766 Value *ARMTargetLowering::getSDagStackGuard(const Module &M) const {
16767 // MSVC CRT has a global variable holding security cookie.
16768 if (Subtarget->getTargetTriple().isWindowsMSVCEnvironment())
16769 return M.getGlobalVariable("__security_cookie");
16770 return TargetLowering::getSDagStackGuard(M);
16773 Function *ARMTargetLowering::getSSPStackGuardCheck(const Module &M) const {
16774 // MSVC CRT has a function to validate security cookie.
16775 if (Subtarget->getTargetTriple().isWindowsMSVCEnvironment())
16776 return M.getFunction("__security_check_cookie");
16777 return TargetLowering::getSSPStackGuardCheck(M);
16780 bool ARMTargetLowering::canCombineStoreAndExtract(Type *VectorTy, Value *Idx,
16781 unsigned &Cost) const {
16782 // If we do not have NEON, vector types are not natively supported.
16783 if (!Subtarget->hasNEON())
16786 // Floating point values and vector values map to the same register file.
16787 // Therefore, although we could do a store extract of a vector type, this is
16788 // better to leave at float as we have more freedom in the addressing mode for
16790 if (VectorTy->isFPOrFPVectorTy())
16793 // If the index is unknown at compile time, this is very expensive to lower
16794 // and it is not possible to combine the store with the extract.
16795 if (!isa<ConstantInt>(Idx))
16798 assert(VectorTy->isVectorTy() && "VectorTy is not a vector type");
16799 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth();
16800 // We can do a store + vector extract on any vector that fits perfectly in a D
16802 if (BitWidth == 64 || BitWidth == 128) {
16809 bool ARMTargetLowering::isCheapToSpeculateCttz() const {
16810 return Subtarget->hasV6T2Ops();
16813 bool ARMTargetLowering::isCheapToSpeculateCtlz() const {
16814 return Subtarget->hasV6T2Ops();
16817 bool ARMTargetLowering::shouldExpandShift(SelectionDAG &DAG, SDNode *N) const {
16818 return !Subtarget->hasMinSize();
16821 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
16822 AtomicOrdering Ord) const {
16823 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
16824 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
16825 bool IsAcquire = isAcquireOrStronger(Ord);
16827 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
16828 // intrinsic must return {i32, i32} and we have to recombine them into a
16829 // single i64 here.
16830 if (ValTy->getPrimitiveSizeInBits() == 64) {
16831 Intrinsic::ID Int =
16832 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
16833 Function *Ldrex = Intrinsic::getDeclaration(M, Int);
16835 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
16836 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
16838 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
16839 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
16840 if (!Subtarget->isLittle())
16841 std::swap (Lo, Hi);
16842 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
16843 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
16844 return Builder.CreateOr(
16845 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
16848 Type *Tys[] = { Addr->getType() };
16849 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
16850 Function *Ldrex = Intrinsic::getDeclaration(M, Int, Tys);
16852 return Builder.CreateTruncOrBitCast(
16853 Builder.CreateCall(Ldrex, Addr),
16854 cast<PointerType>(Addr->getType())->getElementType());
16857 void ARMTargetLowering::emitAtomicCmpXchgNoStoreLLBalance(
16858 IRBuilder<> &Builder) const {
16859 if (!Subtarget->hasV7Ops())
16861 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
16862 Builder.CreateCall(Intrinsic::getDeclaration(M, Intrinsic::arm_clrex));
16865 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
16867 AtomicOrdering Ord) const {
16868 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
16869 bool IsRelease = isReleaseOrStronger(Ord);
16871 // Since the intrinsics must have legal type, the i64 intrinsics take two
16872 // parameters: "i32, i32". We must marshal Val into the appropriate form
16873 // before the call.
16874 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
16875 Intrinsic::ID Int =
16876 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
16877 Function *Strex = Intrinsic::getDeclaration(M, Int);
16878 Type *Int32Ty = Type::getInt32Ty(M->getContext());
16880 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
16881 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
16882 if (!Subtarget->isLittle())
16884 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
16885 return Builder.CreateCall(Strex, {Lo, Hi, Addr});
16888 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
16889 Type *Tys[] = { Addr->getType() };
16890 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
16892 return Builder.CreateCall(
16893 Strex, {Builder.CreateZExtOrBitCast(
16894 Val, Strex->getFunctionType()->getParamType(0)),
16899 bool ARMTargetLowering::alignLoopsWithOptSize() const {
16900 return Subtarget->isMClass();
16903 /// A helper function for determining the number of interleaved accesses we
16904 /// will generate when lowering accesses of the given type.
16906 ARMTargetLowering::getNumInterleavedAccesses(VectorType *VecTy,
16907 const DataLayout &DL) const {
16908 return (DL.getTypeSizeInBits(VecTy) + 127) / 128;
16911 bool ARMTargetLowering::isLegalInterleavedAccessType(
16912 unsigned Factor, VectorType *VecTy, const DataLayout &DL) const {
16914 unsigned VecSize = DL.getTypeSizeInBits(VecTy);
16915 unsigned ElSize = DL.getTypeSizeInBits(VecTy->getElementType());
16917 if (!Subtarget->hasNEON() && !Subtarget->hasMVEIntegerOps())
16920 // Ensure the vector doesn't have f16 elements. Even though we could do an
16921 // i16 vldN, we can't hold the f16 vectors and will end up converting via
16923 if (Subtarget->hasNEON() && VecTy->getElementType()->isHalfTy())
16925 if (Subtarget->hasMVEIntegerOps() && Factor == 3)
16928 // Ensure the number of vector elements is greater than 1.
16929 if (VecTy->getNumElements() < 2)
16932 // Ensure the element type is legal.
16933 if (ElSize != 8 && ElSize != 16 && ElSize != 32)
16936 // Ensure the total vector size is 64 or a multiple of 128. Types larger than
16937 // 128 will be split into multiple interleaved accesses.
16938 if (Subtarget->hasNEON() && VecSize == 64)
16940 return VecSize % 128 == 0;
16943 unsigned ARMTargetLowering::getMaxSupportedInterleaveFactor() const {
16944 if (Subtarget->hasNEON())
16946 if (Subtarget->hasMVEIntegerOps())
16947 return MVEMaxSupportedInterleaveFactor;
16948 return TargetLoweringBase::getMaxSupportedInterleaveFactor();
16951 /// Lower an interleaved load into a vldN intrinsic.
16953 /// E.g. Lower an interleaved load (Factor = 2):
16954 /// %wide.vec = load <8 x i32>, <8 x i32>* %ptr, align 4
16955 /// %v0 = shuffle %wide.vec, undef, <0, 2, 4, 6> ; Extract even elements
16956 /// %v1 = shuffle %wide.vec, undef, <1, 3, 5, 7> ; Extract odd elements
16959 /// %vld2 = { <4 x i32>, <4 x i32> } call llvm.arm.neon.vld2(%ptr, 4)
16960 /// %vec0 = extractelement { <4 x i32>, <4 x i32> } %vld2, i32 0
16961 /// %vec1 = extractelement { <4 x i32>, <4 x i32> } %vld2, i32 1
16962 bool ARMTargetLowering::lowerInterleavedLoad(
16963 LoadInst *LI, ArrayRef<ShuffleVectorInst *> Shuffles,
16964 ArrayRef<unsigned> Indices, unsigned Factor) const {
16965 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
16966 "Invalid interleave factor");
16967 assert(!Shuffles.empty() && "Empty shufflevector input");
16968 assert(Shuffles.size() == Indices.size() &&
16969 "Unmatched number of shufflevectors and indices");
16971 VectorType *VecTy = Shuffles[0]->getType();
16972 Type *EltTy = VecTy->getVectorElementType();
16974 const DataLayout &DL = LI->getModule()->getDataLayout();
16976 // Skip if we do not have NEON and skip illegal vector types. We can
16977 // "legalize" wide vector types into multiple interleaved accesses as long as
16978 // the vector types are divisible by 128.
16979 if (!isLegalInterleavedAccessType(Factor, VecTy, DL))
16982 unsigned NumLoads = getNumInterleavedAccesses(VecTy, DL);
16984 // A pointer vector can not be the return type of the ldN intrinsics. Need to
16985 // load integer vectors first and then convert to pointer vectors.
16986 if (EltTy->isPointerTy())
16988 VectorType::get(DL.getIntPtrType(EltTy), VecTy->getVectorNumElements());
16990 IRBuilder<> Builder(LI);
16992 // The base address of the load.
16993 Value *BaseAddr = LI->getPointerOperand();
16995 if (NumLoads > 1) {
16996 // If we're going to generate more than one load, reset the sub-vector type
16997 // to something legal.
16998 VecTy = VectorType::get(VecTy->getVectorElementType(),
16999 VecTy->getVectorNumElements() / NumLoads);
17001 // We will compute the pointer operand of each load from the original base
17002 // address using GEPs. Cast the base address to a pointer to the scalar
17004 BaseAddr = Builder.CreateBitCast(
17005 BaseAddr, VecTy->getVectorElementType()->getPointerTo(
17006 LI->getPointerAddressSpace()));
17009 assert(isTypeLegal(EVT::getEVT(VecTy)) && "Illegal vldN vector type!");
17011 auto createLoadIntrinsic = [&](Value *BaseAddr) {
17012 if (Subtarget->hasNEON()) {
17013 Type *Int8Ptr = Builder.getInt8PtrTy(LI->getPointerAddressSpace());
17014 Type *Tys[] = {VecTy, Int8Ptr};
17015 static const Intrinsic::ID LoadInts[3] = {Intrinsic::arm_neon_vld2,
17016 Intrinsic::arm_neon_vld3,
17017 Intrinsic::arm_neon_vld4};
17018 Function *VldnFunc =
17019 Intrinsic::getDeclaration(LI->getModule(), LoadInts[Factor - 2], Tys);
17021 SmallVector<Value *, 2> Ops;
17022 Ops.push_back(Builder.CreateBitCast(BaseAddr, Int8Ptr));
17023 Ops.push_back(Builder.getInt32(LI->getAlignment()));
17025 return Builder.CreateCall(VldnFunc, Ops, "vldN");
17027 assert((Factor == 2 || Factor == 4) &&
17028 "expected interleave factor of 2 or 4 for MVE");
17029 Intrinsic::ID LoadInts =
17030 Factor == 2 ? Intrinsic::arm_mve_vld2q : Intrinsic::arm_mve_vld4q;
17031 Type *VecEltTy = VecTy->getVectorElementType()->getPointerTo(
17032 LI->getPointerAddressSpace());
17033 Type *Tys[] = {VecTy, VecEltTy};
17034 Function *VldnFunc =
17035 Intrinsic::getDeclaration(LI->getModule(), LoadInts, Tys);
17037 SmallVector<Value *, 2> Ops;
17038 Ops.push_back(Builder.CreateBitCast(BaseAddr, VecEltTy));
17039 return Builder.CreateCall(VldnFunc, Ops, "vldN");
17043 // Holds sub-vectors extracted from the load intrinsic return values. The
17044 // sub-vectors are associated with the shufflevector instructions they will
17046 DenseMap<ShuffleVectorInst *, SmallVector<Value *, 4>> SubVecs;
17048 for (unsigned LoadCount = 0; LoadCount < NumLoads; ++LoadCount) {
17049 // If we're generating more than one load, compute the base address of
17050 // subsequent loads as an offset from the previous.
17053 Builder.CreateConstGEP1_32(VecTy->getVectorElementType(), BaseAddr,
17054 VecTy->getVectorNumElements() * Factor);
17056 CallInst *VldN = createLoadIntrinsic(BaseAddr);
17058 // Replace uses of each shufflevector with the corresponding vector loaded
17060 for (unsigned i = 0; i < Shuffles.size(); i++) {
17061 ShuffleVectorInst *SV = Shuffles[i];
17062 unsigned Index = Indices[i];
17064 Value *SubVec = Builder.CreateExtractValue(VldN, Index);
17066 // Convert the integer vector to pointer vector if the element is pointer.
17067 if (EltTy->isPointerTy())
17068 SubVec = Builder.CreateIntToPtr(
17069 SubVec, VectorType::get(SV->getType()->getVectorElementType(),
17070 VecTy->getVectorNumElements()));
17072 SubVecs[SV].push_back(SubVec);
17076 // Replace uses of the shufflevector instructions with the sub-vectors
17077 // returned by the load intrinsic. If a shufflevector instruction is
17078 // associated with more than one sub-vector, those sub-vectors will be
17079 // concatenated into a single wide vector.
17080 for (ShuffleVectorInst *SVI : Shuffles) {
17081 auto &SubVec = SubVecs[SVI];
17083 SubVec.size() > 1 ? concatenateVectors(Builder, SubVec) : SubVec[0];
17084 SVI->replaceAllUsesWith(WideVec);
17090 /// Lower an interleaved store into a vstN intrinsic.
17092 /// E.g. Lower an interleaved store (Factor = 3):
17093 /// %i.vec = shuffle <8 x i32> %v0, <8 x i32> %v1,
17094 /// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11>
17095 /// store <12 x i32> %i.vec, <12 x i32>* %ptr, align 4
17098 /// %sub.v0 = shuffle <8 x i32> %v0, <8 x i32> v1, <0, 1, 2, 3>
17099 /// %sub.v1 = shuffle <8 x i32> %v0, <8 x i32> v1, <4, 5, 6, 7>
17100 /// %sub.v2 = shuffle <8 x i32> %v0, <8 x i32> v1, <8, 9, 10, 11>
17101 /// call void llvm.arm.neon.vst3(%ptr, %sub.v0, %sub.v1, %sub.v2, 4)
17103 /// Note that the new shufflevectors will be removed and we'll only generate one
17104 /// vst3 instruction in CodeGen.
17106 /// Example for a more general valid mask (Factor 3). Lower:
17107 /// %i.vec = shuffle <32 x i32> %v0, <32 x i32> %v1,
17108 /// <4, 32, 16, 5, 33, 17, 6, 34, 18, 7, 35, 19>
17109 /// store <12 x i32> %i.vec, <12 x i32>* %ptr
17112 /// %sub.v0 = shuffle <32 x i32> %v0, <32 x i32> v1, <4, 5, 6, 7>
17113 /// %sub.v1 = shuffle <32 x i32> %v0, <32 x i32> v1, <32, 33, 34, 35>
17114 /// %sub.v2 = shuffle <32 x i32> %v0, <32 x i32> v1, <16, 17, 18, 19>
17115 /// call void llvm.arm.neon.vst3(%ptr, %sub.v0, %sub.v1, %sub.v2, 4)
17116 bool ARMTargetLowering::lowerInterleavedStore(StoreInst *SI,
17117 ShuffleVectorInst *SVI,
17118 unsigned Factor) const {
17119 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
17120 "Invalid interleave factor");
17122 VectorType *VecTy = SVI->getType();
17123 assert(VecTy->getVectorNumElements() % Factor == 0 &&
17124 "Invalid interleaved store");
17126 unsigned LaneLen = VecTy->getVectorNumElements() / Factor;
17127 Type *EltTy = VecTy->getVectorElementType();
17128 VectorType *SubVecTy = VectorType::get(EltTy, LaneLen);
17130 const DataLayout &DL = SI->getModule()->getDataLayout();
17132 // Skip if we do not have NEON and skip illegal vector types. We can
17133 // "legalize" wide vector types into multiple interleaved accesses as long as
17134 // the vector types are divisible by 128.
17135 if (!isLegalInterleavedAccessType(Factor, SubVecTy, DL))
17138 unsigned NumStores = getNumInterleavedAccesses(SubVecTy, DL);
17140 Value *Op0 = SVI->getOperand(0);
17141 Value *Op1 = SVI->getOperand(1);
17142 IRBuilder<> Builder(SI);
17144 // StN intrinsics don't support pointer vectors as arguments. Convert pointer
17145 // vectors to integer vectors.
17146 if (EltTy->isPointerTy()) {
17147 Type *IntTy = DL.getIntPtrType(EltTy);
17149 // Convert to the corresponding integer vector.
17151 VectorType::get(IntTy, Op0->getType()->getVectorNumElements());
17152 Op0 = Builder.CreatePtrToInt(Op0, IntVecTy);
17153 Op1 = Builder.CreatePtrToInt(Op1, IntVecTy);
17155 SubVecTy = VectorType::get(IntTy, LaneLen);
17158 // The base address of the store.
17159 Value *BaseAddr = SI->getPointerOperand();
17161 if (NumStores > 1) {
17162 // If we're going to generate more than one store, reset the lane length
17163 // and sub-vector type to something legal.
17164 LaneLen /= NumStores;
17165 SubVecTy = VectorType::get(SubVecTy->getVectorElementType(), LaneLen);
17167 // We will compute the pointer operand of each store from the original base
17168 // address using GEPs. Cast the base address to a pointer to the scalar
17170 BaseAddr = Builder.CreateBitCast(
17171 BaseAddr, SubVecTy->getVectorElementType()->getPointerTo(
17172 SI->getPointerAddressSpace()));
17175 assert(isTypeLegal(EVT::getEVT(SubVecTy)) && "Illegal vstN vector type!");
17177 auto Mask = SVI->getShuffleMask();
17179 auto createStoreIntrinsic = [&](Value *BaseAddr,
17180 SmallVectorImpl<Value *> &Shuffles) {
17181 if (Subtarget->hasNEON()) {
17182 static const Intrinsic::ID StoreInts[3] = {Intrinsic::arm_neon_vst2,
17183 Intrinsic::arm_neon_vst3,
17184 Intrinsic::arm_neon_vst4};
17185 Type *Int8Ptr = Builder.getInt8PtrTy(SI->getPointerAddressSpace());
17186 Type *Tys[] = {Int8Ptr, SubVecTy};
17188 Function *VstNFunc = Intrinsic::getDeclaration(
17189 SI->getModule(), StoreInts[Factor - 2], Tys);
17191 SmallVector<Value *, 6> Ops;
17192 Ops.push_back(Builder.CreateBitCast(BaseAddr, Int8Ptr));
17193 for (auto S : Shuffles)
17195 Ops.push_back(Builder.getInt32(SI->getAlignment()));
17196 Builder.CreateCall(VstNFunc, Ops);
17198 assert((Factor == 2 || Factor == 4) &&
17199 "expected interleave factor of 2 or 4 for MVE");
17200 Intrinsic::ID StoreInts =
17201 Factor == 2 ? Intrinsic::arm_mve_vst2q : Intrinsic::arm_mve_vst4q;
17202 Type *EltPtrTy = SubVecTy->getVectorElementType()->getPointerTo(
17203 SI->getPointerAddressSpace());
17204 Type *Tys[] = {EltPtrTy, SubVecTy};
17205 Function *VstNFunc =
17206 Intrinsic::getDeclaration(SI->getModule(), StoreInts, Tys);
17208 SmallVector<Value *, 6> Ops;
17209 Ops.push_back(Builder.CreateBitCast(BaseAddr, EltPtrTy));
17210 for (auto S : Shuffles)
17212 for (unsigned F = 0; F < Factor; F++) {
17213 Ops.push_back(Builder.getInt32(F));
17214 Builder.CreateCall(VstNFunc, Ops);
17220 for (unsigned StoreCount = 0; StoreCount < NumStores; ++StoreCount) {
17221 // If we generating more than one store, we compute the base address of
17222 // subsequent stores as an offset from the previous.
17223 if (StoreCount > 0)
17224 BaseAddr = Builder.CreateConstGEP1_32(SubVecTy->getVectorElementType(),
17225 BaseAddr, LaneLen * Factor);
17227 SmallVector<Value *, 4> Shuffles;
17229 // Split the shufflevector operands into sub vectors for the new vstN call.
17230 for (unsigned i = 0; i < Factor; i++) {
17231 unsigned IdxI = StoreCount * LaneLen * Factor + i;
17232 if (Mask[IdxI] >= 0) {
17233 Shuffles.push_back(Builder.CreateShuffleVector(
17234 Op0, Op1, createSequentialMask(Builder, Mask[IdxI], LaneLen, 0)));
17236 unsigned StartMask = 0;
17237 for (unsigned j = 1; j < LaneLen; j++) {
17238 unsigned IdxJ = StoreCount * LaneLen * Factor + j;
17239 if (Mask[IdxJ * Factor + IdxI] >= 0) {
17240 StartMask = Mask[IdxJ * Factor + IdxI] - IdxJ;
17244 // Note: If all elements in a chunk are undefs, StartMask=0!
17245 // Note: Filling undef gaps with random elements is ok, since
17246 // those elements were being written anyway (with undefs).
17247 // In the case of all undefs we're defaulting to using elems from 0
17248 // Note: StartMask cannot be negative, it's checked in
17249 // isReInterleaveMask
17250 Shuffles.push_back(Builder.CreateShuffleVector(
17251 Op0, Op1, createSequentialMask(Builder, StartMask, LaneLen, 0)));
17255 createStoreIntrinsic(BaseAddr, Shuffles);
17268 static bool isHomogeneousAggregate(Type *Ty, HABaseType &Base,
17269 uint64_t &Members) {
17270 if (auto *ST = dyn_cast<StructType>(Ty)) {
17271 for (unsigned i = 0; i < ST->getNumElements(); ++i) {
17272 uint64_t SubMembers = 0;
17273 if (!isHomogeneousAggregate(ST->getElementType(i), Base, SubMembers))
17275 Members += SubMembers;
17277 } else if (auto *AT = dyn_cast<ArrayType>(Ty)) {
17278 uint64_t SubMembers = 0;
17279 if (!isHomogeneousAggregate(AT->getElementType(), Base, SubMembers))
17281 Members += SubMembers * AT->getNumElements();
17282 } else if (Ty->isFloatTy()) {
17283 if (Base != HA_UNKNOWN && Base != HA_FLOAT)
17287 } else if (Ty->isDoubleTy()) {
17288 if (Base != HA_UNKNOWN && Base != HA_DOUBLE)
17292 } else if (auto *VT = dyn_cast<VectorType>(Ty)) {
17299 return VT->getBitWidth() == 64;
17301 return VT->getBitWidth() == 128;
17303 switch (VT->getBitWidth()) {
17316 return (Members > 0 && Members <= 4);
17319 /// Return the correct alignment for the current calling convention.
17320 Align ARMTargetLowering::getABIAlignmentForCallingConv(Type *ArgTy,
17321 DataLayout DL) const {
17322 const Align ABITypeAlign(DL.getABITypeAlignment(ArgTy));
17323 if (!ArgTy->isVectorTy())
17324 return ABITypeAlign;
17326 // Avoid over-aligning vector parameters. It would require realigning the
17327 // stack and waste space for no real benefit.
17328 return std::min(ABITypeAlign, DL.getStackAlignment());
17331 /// Return true if a type is an AAPCS-VFP homogeneous aggregate or one of
17332 /// [N x i32] or [N x i64]. This allows front-ends to skip emitting padding when
17333 /// passing according to AAPCS rules.
17334 bool ARMTargetLowering::functionArgumentNeedsConsecutiveRegisters(
17335 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
17336 if (getEffectiveCallingConv(CallConv, isVarArg) !=
17337 CallingConv::ARM_AAPCS_VFP)
17340 HABaseType Base = HA_UNKNOWN;
17341 uint64_t Members = 0;
17342 bool IsHA = isHomogeneousAggregate(Ty, Base, Members);
17343 LLVM_DEBUG(dbgs() << "isHA: " << IsHA << " "; Ty->dump());
17345 bool IsIntArray = Ty->isArrayTy() && Ty->getArrayElementType()->isIntegerTy();
17346 return IsHA || IsIntArray;
17349 unsigned ARMTargetLowering::getExceptionPointerRegister(
17350 const Constant *PersonalityFn) const {
17351 // Platforms which do not use SjLj EH may return values in these registers
17352 // via the personality function.
17353 return Subtarget->useSjLjEH() ? ARM::NoRegister : ARM::R0;
17356 unsigned ARMTargetLowering::getExceptionSelectorRegister(
17357 const Constant *PersonalityFn) const {
17358 // Platforms which do not use SjLj EH may return values in these registers
17359 // via the personality function.
17360 return Subtarget->useSjLjEH() ? ARM::NoRegister : ARM::R1;
17363 void ARMTargetLowering::initializeSplitCSR(MachineBasicBlock *Entry) const {
17364 // Update IsSplitCSR in ARMFunctionInfo.
17365 ARMFunctionInfo *AFI = Entry->getParent()->getInfo<ARMFunctionInfo>();
17366 AFI->setIsSplitCSR(true);
17369 void ARMTargetLowering::insertCopiesSplitCSR(
17370 MachineBasicBlock *Entry,
17371 const SmallVectorImpl<MachineBasicBlock *> &Exits) const {
17372 const ARMBaseRegisterInfo *TRI = Subtarget->getRegisterInfo();
17373 const MCPhysReg *IStart = TRI->getCalleeSavedRegsViaCopy(Entry->getParent());
17377 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
17378 MachineRegisterInfo *MRI = &Entry->getParent()->getRegInfo();
17379 MachineBasicBlock::iterator MBBI = Entry->begin();
17380 for (const MCPhysReg *I = IStart; *I; ++I) {
17381 const TargetRegisterClass *RC = nullptr;
17382 if (ARM::GPRRegClass.contains(*I))
17383 RC = &ARM::GPRRegClass;
17384 else if (ARM::DPRRegClass.contains(*I))
17385 RC = &ARM::DPRRegClass;
17387 llvm_unreachable("Unexpected register class in CSRsViaCopy!");
17389 Register NewVR = MRI->createVirtualRegister(RC);
17390 // Create copy from CSR to a virtual register.
17391 // FIXME: this currently does not emit CFI pseudo-instructions, it works
17392 // fine for CXX_FAST_TLS since the C++-style TLS access functions should be
17393 // nounwind. If we want to generalize this later, we may need to emit
17394 // CFI pseudo-instructions.
17395 assert(Entry->getParent()->getFunction().hasFnAttribute(
17396 Attribute::NoUnwind) &&
17397 "Function should be nounwind in insertCopiesSplitCSR!");
17398 Entry->addLiveIn(*I);
17399 BuildMI(*Entry, MBBI, DebugLoc(), TII->get(TargetOpcode::COPY), NewVR)
17402 // Insert the copy-back instructions right before the terminator.
17403 for (auto *Exit : Exits)
17404 BuildMI(*Exit, Exit->getFirstTerminator(), DebugLoc(),
17405 TII->get(TargetOpcode::COPY), *I)
17410 void ARMTargetLowering::finalizeLowering(MachineFunction &MF) const {
17411 MF.getFrameInfo().computeMaxCallFrameSize(MF);
17412 TargetLoweringBase::finalizeLowering(MF);