1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #include "ARMISelLowering.h"
16 #include "ARMCallingConv.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMMachineFunctionInfo.h"
19 #include "ARMPerfectShuffle.h"
20 #include "ARMSubtarget.h"
21 #include "ARMTargetMachine.h"
22 #include "ARMTargetObjectFile.h"
23 #include "MCTargetDesc/ARMAddressingModes.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/CodeGen/CallingConvLower.h"
27 #include "llvm/CodeGen/IntrinsicLowering.h"
28 #include "llvm/CodeGen/MachineBasicBlock.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineJumpTableInfo.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/SelectionDAG.h"
36 #include "llvm/IR/CallingConv.h"
37 #include "llvm/IR/Constants.h"
38 #include "llvm/IR/Function.h"
39 #include "llvm/IR/GlobalValue.h"
40 #include "llvm/IR/IRBuilder.h"
41 #include "llvm/IR/Instruction.h"
42 #include "llvm/IR/Instructions.h"
43 #include "llvm/IR/Intrinsics.h"
44 #include "llvm/IR/Type.h"
45 #include "llvm/MC/MCSectionMachO.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Target/TargetOptions.h"
54 #define DEBUG_TYPE "arm-isel"
56 STATISTIC(NumTailCalls, "Number of tail calls");
57 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
58 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
61 EnableARMLongCalls("arm-long-calls", cl::Hidden,
62 cl::desc("Generate calls via indirect call instructions"),
66 ARMInterworking("arm-interworking", cl::Hidden,
67 cl::desc("Enable / disable ARM interworking (for debugging only)"),
71 class ARMCCState : public CCState {
73 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
74 SmallVectorImpl<CCValAssign> &locs, LLVMContext &C,
76 : CCState(CC, isVarArg, MF, locs, C) {
77 assert(((PC == Call) || (PC == Prologue)) &&
78 "ARMCCState users must specify whether their context is call"
79 "or prologue generation.");
85 // The APCS parameter registers.
86 static const MCPhysReg GPRArgRegs[] = {
87 ARM::R0, ARM::R1, ARM::R2, ARM::R3
90 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
91 MVT PromotedBitwiseVT) {
92 if (VT != PromotedLdStVT) {
93 setOperationAction(ISD::LOAD, VT, Promote);
94 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
96 setOperationAction(ISD::STORE, VT, Promote);
97 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
100 MVT ElemTy = VT.getVectorElementType();
101 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
102 setOperationAction(ISD::SETCC, VT, Custom);
103 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
104 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
105 if (ElemTy == MVT::i32) {
106 setOperationAction(ISD::SINT_TO_FP, VT, Custom);
107 setOperationAction(ISD::UINT_TO_FP, VT, Custom);
108 setOperationAction(ISD::FP_TO_SINT, VT, Custom);
109 setOperationAction(ISD::FP_TO_UINT, VT, Custom);
111 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
112 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
113 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
114 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
116 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
117 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
118 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);
119 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
120 setOperationAction(ISD::SELECT, VT, Expand);
121 setOperationAction(ISD::SELECT_CC, VT, Expand);
122 setOperationAction(ISD::VSELECT, VT, Expand);
123 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
124 if (VT.isInteger()) {
125 setOperationAction(ISD::SHL, VT, Custom);
126 setOperationAction(ISD::SRA, VT, Custom);
127 setOperationAction(ISD::SRL, VT, Custom);
130 // Promote all bit-wise operations.
131 if (VT.isInteger() && VT != PromotedBitwiseVT) {
132 setOperationAction(ISD::AND, VT, Promote);
133 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
134 setOperationAction(ISD::OR, VT, Promote);
135 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
136 setOperationAction(ISD::XOR, VT, Promote);
137 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
140 // Neon does not support vector divide/remainder operations.
141 setOperationAction(ISD::SDIV, VT, Expand);
142 setOperationAction(ISD::UDIV, VT, Expand);
143 setOperationAction(ISD::FDIV, VT, Expand);
144 setOperationAction(ISD::SREM, VT, Expand);
145 setOperationAction(ISD::UREM, VT, Expand);
146 setOperationAction(ISD::FREM, VT, Expand);
149 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
150 addRegisterClass(VT, &ARM::DPRRegClass);
151 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
154 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
155 addRegisterClass(VT, &ARM::DPairRegClass);
156 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
159 ARMTargetLowering::ARMTargetLowering(const TargetMachine &TM)
160 : TargetLowering(TM) {
161 Subtarget = &TM.getSubtarget<ARMSubtarget>();
162 RegInfo = TM.getSubtargetImpl()->getRegisterInfo();
163 Itins = TM.getSubtargetImpl()->getInstrItineraryData();
165 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
167 if (Subtarget->isTargetMachO()) {
168 // Uses VFP for Thumb libfuncs if available.
169 if (Subtarget->isThumb() && Subtarget->hasVFP2() &&
170 Subtarget->hasARMOps() && !TM.Options.UseSoftFloat) {
171 // Single-precision floating-point arithmetic.
172 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
173 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
174 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
175 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
177 // Double-precision floating-point arithmetic.
178 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
179 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
180 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
181 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
183 // Single-precision comparisons.
184 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
185 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
186 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
187 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
188 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
189 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
190 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
191 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
193 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
194 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
195 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
196 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
197 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
198 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
199 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
200 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
202 // Double-precision comparisons.
203 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
204 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
205 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
206 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
207 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
208 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
209 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
210 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
212 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
213 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
214 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
215 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
216 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
217 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
218 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
219 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
221 // Floating-point to integer conversions.
222 // i64 conversions are done via library routines even when generating VFP
223 // instructions, so use the same ones.
224 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
225 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
226 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
227 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
229 // Conversions between floating types.
230 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
231 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
233 // Integer to floating-point conversions.
234 // i64 conversions are done via library routines even when generating VFP
235 // instructions, so use the same ones.
236 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
237 // e.g., __floatunsidf vs. __floatunssidfvfp.
238 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
239 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
240 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
241 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
245 // These libcalls are not available in 32-bit.
246 setLibcallName(RTLIB::SHL_I128, nullptr);
247 setLibcallName(RTLIB::SRL_I128, nullptr);
248 setLibcallName(RTLIB::SRA_I128, nullptr);
250 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetMachO() &&
251 !Subtarget->isTargetWindows()) {
252 static const struct {
253 const RTLIB::Libcall Op;
254 const char * const Name;
255 const CallingConv::ID CC;
256 const ISD::CondCode Cond;
258 // Double-precision floating-point arithmetic helper functions
259 // RTABI chapter 4.1.2, Table 2
260 { RTLIB::ADD_F64, "__aeabi_dadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
261 { RTLIB::DIV_F64, "__aeabi_ddiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
262 { RTLIB::MUL_F64, "__aeabi_dmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
263 { RTLIB::SUB_F64, "__aeabi_dsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
265 // Double-precision floating-point comparison helper functions
266 // RTABI chapter 4.1.2, Table 3
267 { RTLIB::OEQ_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
268 { RTLIB::UNE_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
269 { RTLIB::OLT_F64, "__aeabi_dcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
270 { RTLIB::OLE_F64, "__aeabi_dcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
271 { RTLIB::OGE_F64, "__aeabi_dcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
272 { RTLIB::OGT_F64, "__aeabi_dcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
273 { RTLIB::UO_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
274 { RTLIB::O_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
276 // Single-precision floating-point arithmetic helper functions
277 // RTABI chapter 4.1.2, Table 4
278 { RTLIB::ADD_F32, "__aeabi_fadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
279 { RTLIB::DIV_F32, "__aeabi_fdiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
280 { RTLIB::MUL_F32, "__aeabi_fmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
281 { RTLIB::SUB_F32, "__aeabi_fsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
283 // Single-precision floating-point comparison helper functions
284 // RTABI chapter 4.1.2, Table 5
285 { RTLIB::OEQ_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
286 { RTLIB::UNE_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
287 { RTLIB::OLT_F32, "__aeabi_fcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
288 { RTLIB::OLE_F32, "__aeabi_fcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
289 { RTLIB::OGE_F32, "__aeabi_fcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
290 { RTLIB::OGT_F32, "__aeabi_fcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
291 { RTLIB::UO_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
292 { RTLIB::O_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
294 // Floating-point to integer conversions.
295 // RTABI chapter 4.1.2, Table 6
296 { RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
297 { RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
298 { RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
299 { RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
300 { RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
301 { RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
302 { RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
303 { RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
305 // Conversions between floating types.
306 // RTABI chapter 4.1.2, Table 7
307 { RTLIB::FPROUND_F64_F32, "__aeabi_d2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
308 { RTLIB::FPROUND_F64_F16, "__aeabi_d2h", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
309 { RTLIB::FPEXT_F32_F64, "__aeabi_f2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
311 // Integer to floating-point conversions.
312 // RTABI chapter 4.1.2, Table 8
313 { RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
314 { RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
315 { RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
316 { RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
317 { RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
318 { RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
319 { RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
320 { RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
322 // Long long helper functions
323 // RTABI chapter 4.2, Table 9
324 { RTLIB::MUL_I64, "__aeabi_lmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
325 { RTLIB::SHL_I64, "__aeabi_llsl", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
326 { RTLIB::SRL_I64, "__aeabi_llsr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
327 { RTLIB::SRA_I64, "__aeabi_lasr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
329 // Integer division functions
330 // RTABI chapter 4.3.1
331 { RTLIB::SDIV_I8, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
332 { RTLIB::SDIV_I16, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
333 { RTLIB::SDIV_I32, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
334 { RTLIB::SDIV_I64, "__aeabi_ldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
335 { RTLIB::UDIV_I8, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
336 { RTLIB::UDIV_I16, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
337 { RTLIB::UDIV_I32, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
338 { RTLIB::UDIV_I64, "__aeabi_uldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
341 // RTABI chapter 4.3.4
342 { RTLIB::MEMCPY, "__aeabi_memcpy", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
343 { RTLIB::MEMMOVE, "__aeabi_memmove", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
344 { RTLIB::MEMSET, "__aeabi_memset", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
347 for (const auto &LC : LibraryCalls) {
348 setLibcallName(LC.Op, LC.Name);
349 setLibcallCallingConv(LC.Op, LC.CC);
350 if (LC.Cond != ISD::SETCC_INVALID)
351 setCmpLibcallCC(LC.Op, LC.Cond);
355 if (Subtarget->isTargetWindows()) {
356 static const struct {
357 const RTLIB::Libcall Op;
358 const char * const Name;
359 const CallingConv::ID CC;
361 { RTLIB::FPTOSINT_F32_I64, "__stoi64", CallingConv::ARM_AAPCS_VFP },
362 { RTLIB::FPTOSINT_F64_I64, "__dtoi64", CallingConv::ARM_AAPCS_VFP },
363 { RTLIB::FPTOUINT_F32_I64, "__stou64", CallingConv::ARM_AAPCS_VFP },
364 { RTLIB::FPTOUINT_F64_I64, "__dtou64", CallingConv::ARM_AAPCS_VFP },
365 { RTLIB::SINTTOFP_I64_F32, "__i64tos", CallingConv::ARM_AAPCS_VFP },
366 { RTLIB::SINTTOFP_I64_F64, "__i64tod", CallingConv::ARM_AAPCS_VFP },
367 { RTLIB::UINTTOFP_I64_F32, "__u64tos", CallingConv::ARM_AAPCS_VFP },
368 { RTLIB::UINTTOFP_I64_F64, "__u64tod", CallingConv::ARM_AAPCS_VFP },
371 for (const auto &LC : LibraryCalls) {
372 setLibcallName(LC.Op, LC.Name);
373 setLibcallCallingConv(LC.Op, LC.CC);
377 // Use divmod compiler-rt calls for iOS 5.0 and later.
378 if (Subtarget->getTargetTriple().isiOS() &&
379 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
380 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
381 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
384 // The half <-> float conversion functions are always soft-float, but are
385 // needed for some targets which use a hard-float calling convention by
387 if (Subtarget->isAAPCS_ABI()) {
388 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_AAPCS);
389 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_AAPCS);
390 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_AAPCS);
392 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_APCS);
393 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_APCS);
394 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_APCS);
397 if (Subtarget->isThumb1Only())
398 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
400 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
401 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
402 !Subtarget->isThumb1Only()) {
403 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
404 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
407 for (MVT VT : MVT::vector_valuetypes()) {
408 for (MVT InnerVT : MVT::vector_valuetypes()) {
409 setTruncStoreAction(VT, InnerVT, Expand);
410 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
411 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
412 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
415 setOperationAction(ISD::MULHS, VT, Expand);
416 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
417 setOperationAction(ISD::MULHU, VT, Expand);
418 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
420 setOperationAction(ISD::BSWAP, VT, Expand);
423 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
424 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
426 if (Subtarget->hasNEON()) {
427 addDRTypeForNEON(MVT::v2f32);
428 addDRTypeForNEON(MVT::v8i8);
429 addDRTypeForNEON(MVT::v4i16);
430 addDRTypeForNEON(MVT::v2i32);
431 addDRTypeForNEON(MVT::v1i64);
433 addQRTypeForNEON(MVT::v4f32);
434 addQRTypeForNEON(MVT::v2f64);
435 addQRTypeForNEON(MVT::v16i8);
436 addQRTypeForNEON(MVT::v8i16);
437 addQRTypeForNEON(MVT::v4i32);
438 addQRTypeForNEON(MVT::v2i64);
440 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
441 // neither Neon nor VFP support any arithmetic operations on it.
442 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
443 // supported for v4f32.
444 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
445 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
446 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
447 // FIXME: Code duplication: FDIV and FREM are expanded always, see
448 // ARMTargetLowering::addTypeForNEON method for details.
449 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
450 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
451 // FIXME: Create unittest.
452 // In another words, find a way when "copysign" appears in DAG with vector
454 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
455 // FIXME: Code duplication: SETCC has custom operation action, see
456 // ARMTargetLowering::addTypeForNEON method for details.
457 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
458 // FIXME: Create unittest for FNEG and for FABS.
459 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
460 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
461 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
462 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
463 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
464 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
465 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
466 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
467 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
468 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
469 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
470 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
471 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
472 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
473 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
474 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
475 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
476 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
477 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
479 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
480 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
481 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
482 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
483 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
484 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
485 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
486 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
487 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
488 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
489 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
490 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
491 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
492 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
493 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
495 // Mark v2f32 intrinsics.
496 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
497 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
498 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
499 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
500 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
501 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
502 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
503 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
504 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
505 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
506 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
507 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
508 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
509 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
510 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
512 // Neon does not support some operations on v1i64 and v2i64 types.
513 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
514 // Custom handling for some quad-vector types to detect VMULL.
515 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
516 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
517 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
518 // Custom handling for some vector types to avoid expensive expansions
519 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
520 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
521 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
522 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
523 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
524 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
525 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
526 // a destination type that is wider than the source, and nor does
527 // it have a FP_TO_[SU]INT instruction with a narrower destination than
529 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
530 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
531 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
532 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
534 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
535 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
537 // NEON does not have single instruction CTPOP for vectors with element
538 // types wider than 8-bits. However, custom lowering can leverage the
539 // v8i8/v16i8 vcnt instruction.
540 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
541 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
542 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
543 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
545 // NEON only has FMA instructions as of VFP4.
546 if (!Subtarget->hasVFP4()) {
547 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
548 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
551 setTargetDAGCombine(ISD::INTRINSIC_VOID);
552 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
553 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
554 setTargetDAGCombine(ISD::SHL);
555 setTargetDAGCombine(ISD::SRL);
556 setTargetDAGCombine(ISD::SRA);
557 setTargetDAGCombine(ISD::SIGN_EXTEND);
558 setTargetDAGCombine(ISD::ZERO_EXTEND);
559 setTargetDAGCombine(ISD::ANY_EXTEND);
560 setTargetDAGCombine(ISD::SELECT_CC);
561 setTargetDAGCombine(ISD::BUILD_VECTOR);
562 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
563 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
564 setTargetDAGCombine(ISD::STORE);
565 setTargetDAGCombine(ISD::FP_TO_SINT);
566 setTargetDAGCombine(ISD::FP_TO_UINT);
567 setTargetDAGCombine(ISD::FDIV);
569 // It is legal to extload from v4i8 to v4i16 or v4i32.
570 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
571 MVT::v4i16, MVT::v2i16,
573 for (unsigned i = 0; i < 6; ++i) {
574 for (MVT VT : MVT::integer_vector_valuetypes()) {
575 setLoadExtAction(ISD::EXTLOAD, VT, Tys[i], Legal);
576 setLoadExtAction(ISD::ZEXTLOAD, VT, Tys[i], Legal);
577 setLoadExtAction(ISD::SEXTLOAD, VT, Tys[i], Legal);
582 // ARM and Thumb2 support UMLAL/SMLAL.
583 if (!Subtarget->isThumb1Only())
584 setTargetDAGCombine(ISD::ADDC);
586 if (Subtarget->isFPOnlySP()) {
587 // When targetting a floating-point unit with only single-precision
588 // operations, f64 is legal for the few double-precision instructions which
589 // are present However, no double-precision operations other than moves,
590 // loads and stores are provided by the hardware.
591 setOperationAction(ISD::FADD, MVT::f64, Expand);
592 setOperationAction(ISD::FSUB, MVT::f64, Expand);
593 setOperationAction(ISD::FMUL, MVT::f64, Expand);
594 setOperationAction(ISD::FMA, MVT::f64, Expand);
595 setOperationAction(ISD::FDIV, MVT::f64, Expand);
596 setOperationAction(ISD::FREM, MVT::f64, Expand);
597 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
598 setOperationAction(ISD::FGETSIGN, MVT::f64, Expand);
599 setOperationAction(ISD::FNEG, MVT::f64, Expand);
600 setOperationAction(ISD::FABS, MVT::f64, Expand);
601 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
602 setOperationAction(ISD::FSIN, MVT::f64, Expand);
603 setOperationAction(ISD::FCOS, MVT::f64, Expand);
604 setOperationAction(ISD::FPOWI, MVT::f64, Expand);
605 setOperationAction(ISD::FPOW, MVT::f64, Expand);
606 setOperationAction(ISD::FLOG, MVT::f64, Expand);
607 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
608 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
609 setOperationAction(ISD::FEXP, MVT::f64, Expand);
610 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
611 setOperationAction(ISD::FCEIL, MVT::f64, Expand);
612 setOperationAction(ISD::FTRUNC, MVT::f64, Expand);
613 setOperationAction(ISD::FRINT, MVT::f64, Expand);
614 setOperationAction(ISD::FNEARBYINT, MVT::f64, Expand);
615 setOperationAction(ISD::FFLOOR, MVT::f64, Expand);
616 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
617 setOperationAction(ISD::FP_EXTEND, MVT::f64, Custom);
620 computeRegisterProperties();
622 // ARM does not have floating-point extending loads.
623 for (MVT VT : MVT::fp_valuetypes()) {
624 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
625 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
628 // ... or truncating stores
629 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
630 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
631 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
633 // ARM does not have i1 sign extending load.
634 for (MVT VT : MVT::integer_valuetypes())
635 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
637 // ARM supports all 4 flavors of integer indexed load / store.
638 if (!Subtarget->isThumb1Only()) {
639 for (unsigned im = (unsigned)ISD::PRE_INC;
640 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
641 setIndexedLoadAction(im, MVT::i1, Legal);
642 setIndexedLoadAction(im, MVT::i8, Legal);
643 setIndexedLoadAction(im, MVT::i16, Legal);
644 setIndexedLoadAction(im, MVT::i32, Legal);
645 setIndexedStoreAction(im, MVT::i1, Legal);
646 setIndexedStoreAction(im, MVT::i8, Legal);
647 setIndexedStoreAction(im, MVT::i16, Legal);
648 setIndexedStoreAction(im, MVT::i32, Legal);
652 setOperationAction(ISD::SADDO, MVT::i32, Custom);
653 setOperationAction(ISD::UADDO, MVT::i32, Custom);
654 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
655 setOperationAction(ISD::USUBO, MVT::i32, Custom);
657 // i64 operation support.
658 setOperationAction(ISD::MUL, MVT::i64, Expand);
659 setOperationAction(ISD::MULHU, MVT::i32, Expand);
660 if (Subtarget->isThumb1Only()) {
661 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
662 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
664 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
665 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
666 setOperationAction(ISD::MULHS, MVT::i32, Expand);
668 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
669 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
670 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
671 setOperationAction(ISD::SRL, MVT::i64, Custom);
672 setOperationAction(ISD::SRA, MVT::i64, Custom);
674 if (!Subtarget->isThumb1Only()) {
675 // FIXME: We should do this for Thumb1 as well.
676 setOperationAction(ISD::ADDC, MVT::i32, Custom);
677 setOperationAction(ISD::ADDE, MVT::i32, Custom);
678 setOperationAction(ISD::SUBC, MVT::i32, Custom);
679 setOperationAction(ISD::SUBE, MVT::i32, Custom);
682 // ARM does not have ROTL.
683 setOperationAction(ISD::ROTL, MVT::i32, Expand);
684 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
685 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
686 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
687 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
689 // These just redirect to CTTZ and CTLZ on ARM.
690 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
691 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
693 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
695 // Only ARMv6 has BSWAP.
696 if (!Subtarget->hasV6Ops())
697 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
699 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
700 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
701 // These are expanded into libcalls if the cpu doesn't have HW divider.
702 setOperationAction(ISD::SDIV, MVT::i32, Expand);
703 setOperationAction(ISD::UDIV, MVT::i32, Expand);
706 // FIXME: Also set divmod for SREM on EABI
707 setOperationAction(ISD::SREM, MVT::i32, Expand);
708 setOperationAction(ISD::UREM, MVT::i32, Expand);
709 // Register based DivRem for AEABI (RTABI 4.2)
710 if (Subtarget->isTargetAEABI()) {
711 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod");
712 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod");
713 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod");
714 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod");
715 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod");
716 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod");
717 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod");
718 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod");
720 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS);
721 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS);
722 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS);
723 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS);
724 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS);
725 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS);
726 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS);
727 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS);
729 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
730 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
732 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
733 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
736 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
737 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
738 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
739 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
740 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
742 setOperationAction(ISD::TRAP, MVT::Other, Legal);
744 // Use the default implementation.
745 setOperationAction(ISD::VASTART, MVT::Other, Custom);
746 setOperationAction(ISD::VAARG, MVT::Other, Expand);
747 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
748 setOperationAction(ISD::VAEND, MVT::Other, Expand);
749 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
750 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
752 if (!Subtarget->isTargetMachO()) {
753 // Non-MachO platforms may return values in these registers via the
754 // personality function.
755 setExceptionPointerRegister(ARM::R0);
756 setExceptionSelectorRegister(ARM::R1);
759 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
760 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
762 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
764 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
765 // the default expansion. If we are targeting a single threaded system,
766 // then set them all for expand so we can lower them later into their
768 if (TM.Options.ThreadModel == ThreadModel::Single)
769 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
770 else if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only()) {
771 // ATOMIC_FENCE needs custom lowering; the others should have been expanded
772 // to ldrex/strex loops already.
773 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
775 // On v8, we have particularly efficient implementations of atomic fences
776 // if they can be combined with nearby atomic loads and stores.
777 if (!Subtarget->hasV8Ops()) {
778 // Automatically insert fences (dmb ish) around ATOMIC_SWAP etc.
779 setInsertFencesForAtomic(true);
782 // If there's anything we can use as a barrier, go through custom lowering
784 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
785 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
787 // Set them all for expansion, which will force libcalls.
788 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
789 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
790 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
791 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
792 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
793 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
794 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
795 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
796 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
797 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
798 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
799 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
800 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
801 // Unordered/Monotonic case.
802 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
803 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
806 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
808 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
809 if (!Subtarget->hasV6Ops()) {
810 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
811 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
813 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
815 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
816 !Subtarget->isThumb1Only()) {
817 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
818 // iff target supports vfp2.
819 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
820 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
823 // We want to custom lower some of our intrinsics.
824 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
825 if (Subtarget->isTargetDarwin()) {
826 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
827 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
828 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
831 setOperationAction(ISD::SETCC, MVT::i32, Expand);
832 setOperationAction(ISD::SETCC, MVT::f32, Expand);
833 setOperationAction(ISD::SETCC, MVT::f64, Expand);
834 setOperationAction(ISD::SELECT, MVT::i32, Custom);
835 setOperationAction(ISD::SELECT, MVT::f32, Custom);
836 setOperationAction(ISD::SELECT, MVT::f64, Custom);
837 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
838 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
839 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
841 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
842 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
843 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
844 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
845 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
847 // We don't support sin/cos/fmod/copysign/pow
848 setOperationAction(ISD::FSIN, MVT::f64, Expand);
849 setOperationAction(ISD::FSIN, MVT::f32, Expand);
850 setOperationAction(ISD::FCOS, MVT::f32, Expand);
851 setOperationAction(ISD::FCOS, MVT::f64, Expand);
852 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
853 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
854 setOperationAction(ISD::FREM, MVT::f64, Expand);
855 setOperationAction(ISD::FREM, MVT::f32, Expand);
856 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
857 !Subtarget->isThumb1Only()) {
858 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
859 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
861 setOperationAction(ISD::FPOW, MVT::f64, Expand);
862 setOperationAction(ISD::FPOW, MVT::f32, Expand);
864 if (!Subtarget->hasVFP4()) {
865 setOperationAction(ISD::FMA, MVT::f64, Expand);
866 setOperationAction(ISD::FMA, MVT::f32, Expand);
869 // Various VFP goodness
870 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
871 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
872 if (Subtarget->hasVFP2()) {
873 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
874 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
875 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
876 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
879 // FP-ARMv8 adds f64 <-> f16 conversion. Before that it should be expanded.
880 if (!Subtarget->hasFPARMv8() || Subtarget->isFPOnlySP()) {
881 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
882 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
885 // fp16 is a special v7 extension that adds f16 <-> f32 conversions.
886 if (!Subtarget->hasFP16()) {
887 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
888 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
892 // Combine sin / cos into one node or libcall if possible.
893 if (Subtarget->hasSinCos()) {
894 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
895 setLibcallName(RTLIB::SINCOS_F64, "sincos");
896 if (Subtarget->getTargetTriple().isiOS()) {
897 // For iOS, we don't want to the normal expansion of a libcall to
898 // sincos. We want to issue a libcall to __sincos_stret.
899 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
900 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
904 // FP-ARMv8 implements a lot of rounding-like FP operations.
905 if (Subtarget->hasFPARMv8()) {
906 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
907 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
908 setOperationAction(ISD::FROUND, MVT::f32, Legal);
909 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
910 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
911 setOperationAction(ISD::FRINT, MVT::f32, Legal);
912 if (!Subtarget->isFPOnlySP()) {
913 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
914 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
915 setOperationAction(ISD::FROUND, MVT::f64, Legal);
916 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
917 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
918 setOperationAction(ISD::FRINT, MVT::f64, Legal);
921 // We have target-specific dag combine patterns for the following nodes:
922 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
923 setTargetDAGCombine(ISD::ADD);
924 setTargetDAGCombine(ISD::SUB);
925 setTargetDAGCombine(ISD::MUL);
926 setTargetDAGCombine(ISD::AND);
927 setTargetDAGCombine(ISD::OR);
928 setTargetDAGCombine(ISD::XOR);
930 if (Subtarget->hasV6Ops())
931 setTargetDAGCombine(ISD::SRL);
933 setStackPointerRegisterToSaveRestore(ARM::SP);
935 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
936 !Subtarget->hasVFP2())
937 setSchedulingPreference(Sched::RegPressure);
939 setSchedulingPreference(Sched::Hybrid);
941 //// temporary - rewrite interface to use type
942 MaxStoresPerMemset = 8;
943 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
944 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
945 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
946 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
947 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
949 // On ARM arguments smaller than 4 bytes are extended, so all arguments
950 // are at least 4 bytes aligned.
951 setMinStackArgumentAlignment(4);
953 // Prefer likely predicted branches to selects on out-of-order cores.
954 PredictableSelectIsExpensive = Subtarget->isLikeA9();
956 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
959 // FIXME: It might make sense to define the representative register class as the
960 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
961 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
962 // SPR's representative would be DPR_VFP2. This should work well if register
963 // pressure tracking were modified such that a register use would increment the
964 // pressure of the register class's representative and all of it's super
965 // classes' representatives transitively. We have not implemented this because
966 // of the difficulty prior to coalescing of modeling operand register classes
967 // due to the common occurrence of cross class copies and subregister insertions
969 std::pair<const TargetRegisterClass*, uint8_t>
970 ARMTargetLowering::findRepresentativeClass(MVT VT) const{
971 const TargetRegisterClass *RRC = nullptr;
973 switch (VT.SimpleTy) {
975 return TargetLowering::findRepresentativeClass(VT);
976 // Use DPR as representative register class for all floating point
977 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
978 // the cost is 1 for both f32 and f64.
979 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
980 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
981 RRC = &ARM::DPRRegClass;
982 // When NEON is used for SP, only half of the register file is available
983 // because operations that define both SP and DP results will be constrained
984 // to the VFP2 class (D0-D15). We currently model this constraint prior to
985 // coalescing by double-counting the SP regs. See the FIXME above.
986 if (Subtarget->useNEONForSinglePrecisionFP())
989 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
990 case MVT::v4f32: case MVT::v2f64:
991 RRC = &ARM::DPRRegClass;
995 RRC = &ARM::DPRRegClass;
999 RRC = &ARM::DPRRegClass;
1003 return std::make_pair(RRC, Cost);
1006 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
1008 default: return nullptr;
1009 case ARMISD::Wrapper: return "ARMISD::Wrapper";
1010 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
1011 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
1012 case ARMISD::CALL: return "ARMISD::CALL";
1013 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
1014 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
1015 case ARMISD::tCALL: return "ARMISD::tCALL";
1016 case ARMISD::BRCOND: return "ARMISD::BRCOND";
1017 case ARMISD::BR_JT: return "ARMISD::BR_JT";
1018 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
1019 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
1020 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
1021 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
1022 case ARMISD::CMP: return "ARMISD::CMP";
1023 case ARMISD::CMN: return "ARMISD::CMN";
1024 case ARMISD::CMPZ: return "ARMISD::CMPZ";
1025 case ARMISD::CMPFP: return "ARMISD::CMPFP";
1026 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
1027 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
1028 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
1030 case ARMISD::CMOV: return "ARMISD::CMOV";
1032 case ARMISD::RBIT: return "ARMISD::RBIT";
1034 case ARMISD::FTOSI: return "ARMISD::FTOSI";
1035 case ARMISD::FTOUI: return "ARMISD::FTOUI";
1036 case ARMISD::SITOF: return "ARMISD::SITOF";
1037 case ARMISD::UITOF: return "ARMISD::UITOF";
1039 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
1040 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
1041 case ARMISD::RRX: return "ARMISD::RRX";
1043 case ARMISD::ADDC: return "ARMISD::ADDC";
1044 case ARMISD::ADDE: return "ARMISD::ADDE";
1045 case ARMISD::SUBC: return "ARMISD::SUBC";
1046 case ARMISD::SUBE: return "ARMISD::SUBE";
1048 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
1049 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
1051 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
1052 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
1054 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1056 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1058 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1060 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1062 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1064 case ARMISD::WIN__CHKSTK: return "ARMISD:::WIN__CHKSTK";
1066 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1067 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1068 case ARMISD::VCGE: return "ARMISD::VCGE";
1069 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1070 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1071 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1072 case ARMISD::VCGT: return "ARMISD::VCGT";
1073 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1074 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1075 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1076 case ARMISD::VTST: return "ARMISD::VTST";
1078 case ARMISD::VSHL: return "ARMISD::VSHL";
1079 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1080 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1081 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1082 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1083 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1084 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1085 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1086 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1087 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1088 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1089 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1090 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1091 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1092 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1093 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1094 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1095 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1096 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1097 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1098 case ARMISD::VDUP: return "ARMISD::VDUP";
1099 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1100 case ARMISD::VEXT: return "ARMISD::VEXT";
1101 case ARMISD::VREV64: return "ARMISD::VREV64";
1102 case ARMISD::VREV32: return "ARMISD::VREV32";
1103 case ARMISD::VREV16: return "ARMISD::VREV16";
1104 case ARMISD::VZIP: return "ARMISD::VZIP";
1105 case ARMISD::VUZP: return "ARMISD::VUZP";
1106 case ARMISD::VTRN: return "ARMISD::VTRN";
1107 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1108 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1109 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1110 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1111 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1112 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1113 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1114 case ARMISD::FMAX: return "ARMISD::FMAX";
1115 case ARMISD::FMIN: return "ARMISD::FMIN";
1116 case ARMISD::VMAXNM: return "ARMISD::VMAX";
1117 case ARMISD::VMINNM: return "ARMISD::VMIN";
1118 case ARMISD::BFI: return "ARMISD::BFI";
1119 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1120 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1121 case ARMISD::VBSL: return "ARMISD::VBSL";
1122 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1123 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1124 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1125 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1126 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1127 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1128 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1129 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1130 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1131 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1132 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1133 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1134 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1135 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1136 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1137 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1138 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1139 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1140 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1141 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1145 EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1146 if (!VT.isVector()) return getPointerTy();
1147 return VT.changeVectorElementTypeToInteger();
1150 /// getRegClassFor - Return the register class that should be used for the
1151 /// specified value type.
1152 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1153 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1154 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1155 // load / store 4 to 8 consecutive D registers.
1156 if (Subtarget->hasNEON()) {
1157 if (VT == MVT::v4i64)
1158 return &ARM::QQPRRegClass;
1159 if (VT == MVT::v8i64)
1160 return &ARM::QQQQPRRegClass;
1162 return TargetLowering::getRegClassFor(VT);
1165 // Create a fast isel object.
1167 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1168 const TargetLibraryInfo *libInfo) const {
1169 return ARM::createFastISel(funcInfo, libInfo);
1172 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1173 /// be used for loads / stores from the global.
1174 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1175 return (Subtarget->isThumb1Only() ? 127 : 4095);
1178 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1179 unsigned NumVals = N->getNumValues();
1181 return Sched::RegPressure;
1183 for (unsigned i = 0; i != NumVals; ++i) {
1184 EVT VT = N->getValueType(i);
1185 if (VT == MVT::Glue || VT == MVT::Other)
1187 if (VT.isFloatingPoint() || VT.isVector())
1191 if (!N->isMachineOpcode())
1192 return Sched::RegPressure;
1194 // Load are scheduled for latency even if there instruction itinerary
1195 // is not available.
1196 const TargetInstrInfo *TII =
1197 getTargetMachine().getSubtargetImpl()->getInstrInfo();
1198 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1200 if (MCID.getNumDefs() == 0)
1201 return Sched::RegPressure;
1202 if (!Itins->isEmpty() &&
1203 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1206 return Sched::RegPressure;
1209 //===----------------------------------------------------------------------===//
1211 //===----------------------------------------------------------------------===//
1213 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1214 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1216 default: llvm_unreachable("Unknown condition code!");
1217 case ISD::SETNE: return ARMCC::NE;
1218 case ISD::SETEQ: return ARMCC::EQ;
1219 case ISD::SETGT: return ARMCC::GT;
1220 case ISD::SETGE: return ARMCC::GE;
1221 case ISD::SETLT: return ARMCC::LT;
1222 case ISD::SETLE: return ARMCC::LE;
1223 case ISD::SETUGT: return ARMCC::HI;
1224 case ISD::SETUGE: return ARMCC::HS;
1225 case ISD::SETULT: return ARMCC::LO;
1226 case ISD::SETULE: return ARMCC::LS;
1230 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1231 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1232 ARMCC::CondCodes &CondCode2) {
1233 CondCode2 = ARMCC::AL;
1235 default: llvm_unreachable("Unknown FP condition!");
1237 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1239 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1241 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1242 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1243 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1244 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1245 case ISD::SETO: CondCode = ARMCC::VC; break;
1246 case ISD::SETUO: CondCode = ARMCC::VS; break;
1247 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1248 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1249 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1251 case ISD::SETULT: CondCode = ARMCC::LT; break;
1253 case ISD::SETULE: CondCode = ARMCC::LE; break;
1255 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1259 //===----------------------------------------------------------------------===//
1260 // Calling Convention Implementation
1261 //===----------------------------------------------------------------------===//
1263 #include "ARMGenCallingConv.inc"
1265 /// getEffectiveCallingConv - Get the effective calling convention, taking into
1266 /// account presence of floating point hardware and calling convention
1267 /// limitations, such as support for variadic functions.
1269 ARMTargetLowering::getEffectiveCallingConv(CallingConv::ID CC,
1270 bool isVarArg) const {
1273 llvm_unreachable("Unsupported calling convention");
1274 case CallingConv::ARM_AAPCS:
1275 case CallingConv::ARM_APCS:
1276 case CallingConv::GHC:
1278 case CallingConv::ARM_AAPCS_VFP:
1279 return isVarArg ? CallingConv::ARM_AAPCS : CallingConv::ARM_AAPCS_VFP;
1280 case CallingConv::C:
1281 if (!Subtarget->isAAPCS_ABI())
1282 return CallingConv::ARM_APCS;
1283 else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() &&
1284 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1286 return CallingConv::ARM_AAPCS_VFP;
1288 return CallingConv::ARM_AAPCS;
1289 case CallingConv::Fast:
1290 if (!Subtarget->isAAPCS_ABI()) {
1291 if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1292 return CallingConv::Fast;
1293 return CallingConv::ARM_APCS;
1294 } else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1295 return CallingConv::ARM_AAPCS_VFP;
1297 return CallingConv::ARM_AAPCS;
1301 /// CCAssignFnForNode - Selects the correct CCAssignFn for the given
1302 /// CallingConvention.
1303 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1305 bool isVarArg) const {
1306 switch (getEffectiveCallingConv(CC, isVarArg)) {
1308 llvm_unreachable("Unsupported calling convention");
1309 case CallingConv::ARM_APCS:
1310 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1311 case CallingConv::ARM_AAPCS:
1312 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1313 case CallingConv::ARM_AAPCS_VFP:
1314 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1315 case CallingConv::Fast:
1316 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1317 case CallingConv::GHC:
1318 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1322 /// LowerCallResult - Lower the result values of a call into the
1323 /// appropriate copies out of appropriate physical registers.
1325 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1326 CallingConv::ID CallConv, bool isVarArg,
1327 const SmallVectorImpl<ISD::InputArg> &Ins,
1328 SDLoc dl, SelectionDAG &DAG,
1329 SmallVectorImpl<SDValue> &InVals,
1330 bool isThisReturn, SDValue ThisVal) const {
1332 // Assign locations to each value returned by this call.
1333 SmallVector<CCValAssign, 16> RVLocs;
1334 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1335 *DAG.getContext(), Call);
1336 CCInfo.AnalyzeCallResult(Ins,
1337 CCAssignFnForNode(CallConv, /* Return*/ true,
1340 // Copy all of the result registers out of their specified physreg.
1341 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1342 CCValAssign VA = RVLocs[i];
1344 // Pass 'this' value directly from the argument to return value, to avoid
1345 // reg unit interference
1346 if (i == 0 && isThisReturn) {
1347 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1348 "unexpected return calling convention register assignment");
1349 InVals.push_back(ThisVal);
1354 if (VA.needsCustom()) {
1355 // Handle f64 or half of a v2f64.
1356 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1358 Chain = Lo.getValue(1);
1359 InFlag = Lo.getValue(2);
1360 VA = RVLocs[++i]; // skip ahead to next loc
1361 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1363 Chain = Hi.getValue(1);
1364 InFlag = Hi.getValue(2);
1365 if (!Subtarget->isLittle())
1367 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1369 if (VA.getLocVT() == MVT::v2f64) {
1370 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1371 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1372 DAG.getConstant(0, MVT::i32));
1374 VA = RVLocs[++i]; // skip ahead to next loc
1375 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1376 Chain = Lo.getValue(1);
1377 InFlag = Lo.getValue(2);
1378 VA = RVLocs[++i]; // skip ahead to next loc
1379 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1380 Chain = Hi.getValue(1);
1381 InFlag = Hi.getValue(2);
1382 if (!Subtarget->isLittle())
1384 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1385 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1386 DAG.getConstant(1, MVT::i32));
1389 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1391 Chain = Val.getValue(1);
1392 InFlag = Val.getValue(2);
1395 switch (VA.getLocInfo()) {
1396 default: llvm_unreachable("Unknown loc info!");
1397 case CCValAssign::Full: break;
1398 case CCValAssign::BCvt:
1399 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1403 InVals.push_back(Val);
1409 /// LowerMemOpCallTo - Store the argument to the stack.
1411 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1412 SDValue StackPtr, SDValue Arg,
1413 SDLoc dl, SelectionDAG &DAG,
1414 const CCValAssign &VA,
1415 ISD::ArgFlagsTy Flags) const {
1416 unsigned LocMemOffset = VA.getLocMemOffset();
1417 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1418 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1419 return DAG.getStore(Chain, dl, Arg, PtrOff,
1420 MachinePointerInfo::getStack(LocMemOffset),
1424 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1425 SDValue Chain, SDValue &Arg,
1426 RegsToPassVector &RegsToPass,
1427 CCValAssign &VA, CCValAssign &NextVA,
1429 SmallVectorImpl<SDValue> &MemOpChains,
1430 ISD::ArgFlagsTy Flags) const {
1432 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1433 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1434 unsigned id = Subtarget->isLittle() ? 0 : 1;
1435 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd.getValue(id)));
1437 if (NextVA.isRegLoc())
1438 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1-id)));
1440 assert(NextVA.isMemLoc());
1441 if (!StackPtr.getNode())
1442 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1444 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1-id),
1450 /// LowerCall - Lowering a call into a callseq_start <-
1451 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1454 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1455 SmallVectorImpl<SDValue> &InVals) const {
1456 SelectionDAG &DAG = CLI.DAG;
1458 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1459 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1460 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1461 SDValue Chain = CLI.Chain;
1462 SDValue Callee = CLI.Callee;
1463 bool &isTailCall = CLI.IsTailCall;
1464 CallingConv::ID CallConv = CLI.CallConv;
1465 bool doesNotRet = CLI.DoesNotReturn;
1466 bool isVarArg = CLI.IsVarArg;
1468 MachineFunction &MF = DAG.getMachineFunction();
1469 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1470 bool isThisReturn = false;
1471 bool isSibCall = false;
1473 // Disable tail calls if they're not supported.
1474 if (!Subtarget->supportsTailCall() || MF.getTarget().Options.DisableTailCalls)
1478 // Check if it's really possible to do a tail call.
1479 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1480 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1481 Outs, OutVals, Ins, DAG);
1482 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
1483 report_fatal_error("failed to perform tail call elimination on a call "
1484 "site marked musttail");
1485 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1486 // detected sibcalls.
1493 // Analyze operands of the call, assigning locations to each operand.
1494 SmallVector<CCValAssign, 16> ArgLocs;
1495 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1496 *DAG.getContext(), Call);
1497 CCInfo.AnalyzeCallOperands(Outs,
1498 CCAssignFnForNode(CallConv, /* Return*/ false,
1501 // Get a count of how many bytes are to be pushed on the stack.
1502 unsigned NumBytes = CCInfo.getNextStackOffset();
1504 // For tail calls, memory operands are available in our caller's stack.
1508 // Adjust the stack pointer for the new arguments...
1509 // These operations are automatically eliminated by the prolog/epilog pass
1511 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
1514 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1516 RegsToPassVector RegsToPass;
1517 SmallVector<SDValue, 8> MemOpChains;
1519 // Walk the register/memloc assignments, inserting copies/loads. In the case
1520 // of tail call optimization, arguments are handled later.
1521 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1523 ++i, ++realArgIdx) {
1524 CCValAssign &VA = ArgLocs[i];
1525 SDValue Arg = OutVals[realArgIdx];
1526 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1527 bool isByVal = Flags.isByVal();
1529 // Promote the value if needed.
1530 switch (VA.getLocInfo()) {
1531 default: llvm_unreachable("Unknown loc info!");
1532 case CCValAssign::Full: break;
1533 case CCValAssign::SExt:
1534 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1536 case CCValAssign::ZExt:
1537 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1539 case CCValAssign::AExt:
1540 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1542 case CCValAssign::BCvt:
1543 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1547 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1548 if (VA.needsCustom()) {
1549 if (VA.getLocVT() == MVT::v2f64) {
1550 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1551 DAG.getConstant(0, MVT::i32));
1552 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1553 DAG.getConstant(1, MVT::i32));
1555 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1556 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1558 VA = ArgLocs[++i]; // skip ahead to next loc
1559 if (VA.isRegLoc()) {
1560 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1561 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1563 assert(VA.isMemLoc());
1565 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1566 dl, DAG, VA, Flags));
1569 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1570 StackPtr, MemOpChains, Flags);
1572 } else if (VA.isRegLoc()) {
1573 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1574 assert(VA.getLocVT() == MVT::i32 &&
1575 "unexpected calling convention register assignment");
1576 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1577 "unexpected use of 'returned'");
1578 isThisReturn = true;
1580 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1581 } else if (isByVal) {
1582 assert(VA.isMemLoc());
1583 unsigned offset = 0;
1585 // True if this byval aggregate will be split between registers
1587 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1588 unsigned CurByValIdx = CCInfo.getInRegsParamsProcessed();
1590 if (CurByValIdx < ByValArgsCount) {
1592 unsigned RegBegin, RegEnd;
1593 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1595 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1597 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1598 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1599 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1600 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1601 MachinePointerInfo(),
1602 false, false, false,
1603 DAG.InferPtrAlignment(AddArg));
1604 MemOpChains.push_back(Load.getValue(1));
1605 RegsToPass.push_back(std::make_pair(j, Load));
1608 // If parameter size outsides register area, "offset" value
1609 // helps us to calculate stack slot for remained part properly.
1610 offset = RegEnd - RegBegin;
1612 CCInfo.nextInRegsParam();
1615 if (Flags.getByValSize() > 4*offset) {
1616 unsigned LocMemOffset = VA.getLocMemOffset();
1617 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1618 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1620 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1621 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1622 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1624 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1626 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1627 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1628 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1631 } else if (!isSibCall) {
1632 assert(VA.isMemLoc());
1634 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1635 dl, DAG, VA, Flags));
1639 if (!MemOpChains.empty())
1640 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1642 // Build a sequence of copy-to-reg nodes chained together with token chain
1643 // and flag operands which copy the outgoing args into the appropriate regs.
1645 // Tail call byval lowering might overwrite argument registers so in case of
1646 // tail call optimization the copies to registers are lowered later.
1648 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1649 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1650 RegsToPass[i].second, InFlag);
1651 InFlag = Chain.getValue(1);
1654 // For tail calls lower the arguments to the 'real' stack slot.
1656 // Force all the incoming stack arguments to be loaded from the stack
1657 // before any new outgoing arguments are stored to the stack, because the
1658 // outgoing stack slots may alias the incoming argument stack slots, and
1659 // the alias isn't otherwise explicit. This is slightly more conservative
1660 // than necessary, because it means that each store effectively depends
1661 // on every argument instead of just those arguments it would clobber.
1663 // Do not flag preceding copytoreg stuff together with the following stuff.
1665 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1666 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1667 RegsToPass[i].second, InFlag);
1668 InFlag = Chain.getValue(1);
1673 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1674 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1675 // node so that legalize doesn't hack it.
1676 bool isDirect = false;
1677 bool isARMFunc = false;
1678 bool isLocalARMFunc = false;
1679 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1681 if (EnableARMLongCalls) {
1682 assert((Subtarget->isTargetWindows() ||
1683 getTargetMachine().getRelocationModel() == Reloc::Static) &&
1684 "long-calls with non-static relocation model!");
1685 // Handle a global address or an external symbol. If it's not one of
1686 // those, the target's already in a register, so we don't need to do
1688 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1689 const GlobalValue *GV = G->getGlobal();
1690 // Create a constant pool entry for the callee address
1691 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1692 ARMConstantPoolValue *CPV =
1693 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1695 // Get the address of the callee into a register
1696 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1697 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1698 Callee = DAG.getLoad(getPointerTy(), dl,
1699 DAG.getEntryNode(), CPAddr,
1700 MachinePointerInfo::getConstantPool(),
1701 false, false, false, 0);
1702 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1703 const char *Sym = S->getSymbol();
1705 // Create a constant pool entry for the callee address
1706 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1707 ARMConstantPoolValue *CPV =
1708 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1709 ARMPCLabelIndex, 0);
1710 // Get the address of the callee into a register
1711 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1712 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1713 Callee = DAG.getLoad(getPointerTy(), dl,
1714 DAG.getEntryNode(), CPAddr,
1715 MachinePointerInfo::getConstantPool(),
1716 false, false, false, 0);
1718 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1719 const GlobalValue *GV = G->getGlobal();
1721 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1722 bool isStub = (isExt && Subtarget->isTargetMachO()) &&
1723 getTargetMachine().getRelocationModel() != Reloc::Static;
1724 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1725 // ARM call to a local ARM function is predicable.
1726 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1727 // tBX takes a register source operand.
1728 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1729 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1730 Callee = DAG.getNode(ARMISD::WrapperPIC, dl, getPointerTy(),
1731 DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
1732 0, ARMII::MO_NONLAZY));
1733 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
1734 MachinePointerInfo::getGOT(), false, false, true, 0);
1735 } else if (Subtarget->isTargetCOFF()) {
1736 assert(Subtarget->isTargetWindows() &&
1737 "Windows is the only supported COFF target");
1738 unsigned TargetFlags = GV->hasDLLImportStorageClass()
1739 ? ARMII::MO_DLLIMPORT
1740 : ARMII::MO_NO_FLAG;
1741 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), /*Offset=*/0,
1743 if (GV->hasDLLImportStorageClass())
1744 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
1745 DAG.getNode(ARMISD::Wrapper, dl, getPointerTy(),
1746 Callee), MachinePointerInfo::getGOT(),
1747 false, false, false, 0);
1749 // On ELF targets for PIC code, direct calls should go through the PLT
1750 unsigned OpFlags = 0;
1751 if (Subtarget->isTargetELF() &&
1752 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1753 OpFlags = ARMII::MO_PLT;
1754 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1756 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1758 bool isStub = Subtarget->isTargetMachO() &&
1759 getTargetMachine().getRelocationModel() != Reloc::Static;
1760 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1761 // tBX takes a register source operand.
1762 const char *Sym = S->getSymbol();
1763 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1764 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1765 ARMConstantPoolValue *CPV =
1766 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1767 ARMPCLabelIndex, 4);
1768 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1769 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1770 Callee = DAG.getLoad(getPointerTy(), dl,
1771 DAG.getEntryNode(), CPAddr,
1772 MachinePointerInfo::getConstantPool(),
1773 false, false, false, 0);
1774 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1775 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1776 getPointerTy(), Callee, PICLabel);
1778 unsigned OpFlags = 0;
1779 // On ELF targets for PIC code, direct calls should go through the PLT
1780 if (Subtarget->isTargetELF() &&
1781 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1782 OpFlags = ARMII::MO_PLT;
1783 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1787 // FIXME: handle tail calls differently.
1789 bool HasMinSizeAttr = MF.getFunction()->getAttributes().hasAttribute(
1790 AttributeSet::FunctionIndex, Attribute::MinSize);
1791 if (Subtarget->isThumb()) {
1792 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1793 CallOpc = ARMISD::CALL_NOLINK;
1795 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1797 if (!isDirect && !Subtarget->hasV5TOps())
1798 CallOpc = ARMISD::CALL_NOLINK;
1799 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1800 // Emit regular call when code size is the priority
1802 // "mov lr, pc; b _foo" to avoid confusing the RSP
1803 CallOpc = ARMISD::CALL_NOLINK;
1805 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1808 std::vector<SDValue> Ops;
1809 Ops.push_back(Chain);
1810 Ops.push_back(Callee);
1812 // Add argument registers to the end of the list so that they are known live
1814 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1815 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1816 RegsToPass[i].second.getValueType()));
1818 // Add a register mask operand representing the call-preserved registers.
1820 const uint32_t *Mask;
1821 const TargetRegisterInfo *TRI =
1822 getTargetMachine().getSubtargetImpl()->getRegisterInfo();
1823 const ARMBaseRegisterInfo *ARI = static_cast<const ARMBaseRegisterInfo*>(TRI);
1825 // For 'this' returns, use the R0-preserving mask if applicable
1826 Mask = ARI->getThisReturnPreservedMask(CallConv);
1828 // Set isThisReturn to false if the calling convention is not one that
1829 // allows 'returned' to be modeled in this way, so LowerCallResult does
1830 // not try to pass 'this' straight through
1831 isThisReturn = false;
1832 Mask = ARI->getCallPreservedMask(CallConv);
1835 Mask = ARI->getCallPreservedMask(CallConv);
1837 assert(Mask && "Missing call preserved mask for calling convention");
1838 Ops.push_back(DAG.getRegisterMask(Mask));
1841 if (InFlag.getNode())
1842 Ops.push_back(InFlag);
1844 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1846 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, Ops);
1848 // Returns a chain and a flag for retval copy to use.
1849 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
1850 InFlag = Chain.getValue(1);
1852 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1853 DAG.getIntPtrConstant(0, true), InFlag, dl);
1855 InFlag = Chain.getValue(1);
1857 // Handle result values, copying them out of physregs into vregs that we
1859 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1860 InVals, isThisReturn,
1861 isThisReturn ? OutVals[0] : SDValue());
1864 /// HandleByVal - Every parameter *after* a byval parameter is passed
1865 /// on the stack. Remember the next parameter register to allocate,
1866 /// and then confiscate the rest of the parameter registers to insure
1869 ARMTargetLowering::HandleByVal(
1870 CCState *State, unsigned &size, unsigned Align) const {
1871 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1872 assert((State->getCallOrPrologue() == Prologue ||
1873 State->getCallOrPrologue() == Call) &&
1874 "unhandled ParmContext");
1876 if ((ARM::R0 <= reg) && (reg <= ARM::R3)) {
1877 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1878 unsigned AlignInRegs = Align / 4;
1879 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1880 for (unsigned i = 0; i < Waste; ++i)
1881 reg = State->AllocateReg(GPRArgRegs, 4);
1884 unsigned excess = 4 * (ARM::R4 - reg);
1886 // Special case when NSAA != SP and parameter size greater than size of
1887 // all remained GPR regs. In that case we can't split parameter, we must
1888 // send it to stack. We also must set NCRN to R4, so waste all
1889 // remained registers.
1890 const unsigned NSAAOffset = State->getNextStackOffset();
1891 if (Subtarget->isAAPCS_ABI() && NSAAOffset != 0 && size > excess) {
1892 while (State->AllocateReg(GPRArgRegs, 4))
1897 // First register for byval parameter is the first register that wasn't
1898 // allocated before this method call, so it would be "reg".
1899 // If parameter is small enough to be saved in range [reg, r4), then
1900 // the end (first after last) register would be reg + param-size-in-regs,
1901 // else parameter would be splitted between registers and stack,
1902 // end register would be r4 in this case.
1903 unsigned ByValRegBegin = reg;
1904 unsigned ByValRegEnd = (size < excess) ? reg + size/4 : (unsigned)ARM::R4;
1905 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1906 // Note, first register is allocated in the beginning of function already,
1907 // allocate remained amount of registers we need.
1908 for (unsigned i = reg+1; i != ByValRegEnd; ++i)
1909 State->AllocateReg(GPRArgRegs, 4);
1910 // A byval parameter that is split between registers and memory needs its
1911 // size truncated here.
1912 // In the case where the entire structure fits in registers, we set the
1913 // size in memory to zero.
1922 /// MatchingStackOffset - Return true if the given stack call argument is
1923 /// already available in the same position (relatively) of the caller's
1924 /// incoming argument stack.
1926 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1927 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1928 const TargetInstrInfo *TII) {
1929 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1931 if (Arg.getOpcode() == ISD::CopyFromReg) {
1932 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1933 if (!TargetRegisterInfo::isVirtualRegister(VR))
1935 MachineInstr *Def = MRI->getVRegDef(VR);
1938 if (!Flags.isByVal()) {
1939 if (!TII->isLoadFromStackSlot(Def, FI))
1944 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1945 if (Flags.isByVal())
1946 // ByVal argument is passed in as a pointer but it's now being
1947 // dereferenced. e.g.
1948 // define @foo(%struct.X* %A) {
1949 // tail call @bar(%struct.X* byval %A)
1952 SDValue Ptr = Ld->getBasePtr();
1953 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1956 FI = FINode->getIndex();
1960 assert(FI != INT_MAX);
1961 if (!MFI->isFixedObjectIndex(FI))
1963 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1966 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1967 /// for tail call optimization. Targets which want to do tail call
1968 /// optimization should implement this function.
1970 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1971 CallingConv::ID CalleeCC,
1973 bool isCalleeStructRet,
1974 bool isCallerStructRet,
1975 const SmallVectorImpl<ISD::OutputArg> &Outs,
1976 const SmallVectorImpl<SDValue> &OutVals,
1977 const SmallVectorImpl<ISD::InputArg> &Ins,
1978 SelectionDAG& DAG) const {
1979 const Function *CallerF = DAG.getMachineFunction().getFunction();
1980 CallingConv::ID CallerCC = CallerF->getCallingConv();
1981 bool CCMatch = CallerCC == CalleeCC;
1983 // Look for obvious safe cases to perform tail call optimization that do not
1984 // require ABI changes. This is what gcc calls sibcall.
1986 // Do not sibcall optimize vararg calls unless the call site is not passing
1988 if (isVarArg && !Outs.empty())
1991 // Exception-handling functions need a special set of instructions to indicate
1992 // a return to the hardware. Tail-calling another function would probably
1994 if (CallerF->hasFnAttribute("interrupt"))
1997 // Also avoid sibcall optimization if either caller or callee uses struct
1998 // return semantics.
1999 if (isCalleeStructRet || isCallerStructRet)
2002 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
2003 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
2004 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
2005 // support in the assembler and linker to be used. This would need to be
2006 // fixed to fully support tail calls in Thumb1.
2008 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
2009 // LR. This means if we need to reload LR, it takes an extra instructions,
2010 // which outweighs the value of the tail call; but here we don't know yet
2011 // whether LR is going to be used. Probably the right approach is to
2012 // generate the tail call here and turn it back into CALL/RET in
2013 // emitEpilogue if LR is used.
2015 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
2016 // but we need to make sure there are enough registers; the only valid
2017 // registers are the 4 used for parameters. We don't currently do this
2019 if (Subtarget->isThumb1Only())
2022 // Externally-defined functions with weak linkage should not be
2023 // tail-called on ARM when the OS does not support dynamic
2024 // pre-emption of symbols, as the AAELF spec requires normal calls
2025 // to undefined weak functions to be replaced with a NOP or jump to the
2026 // next instruction. The behaviour of branch instructions in this
2027 // situation (as used for tail calls) is implementation-defined, so we
2028 // cannot rely on the linker replacing the tail call with a return.
2029 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2030 const GlobalValue *GV = G->getGlobal();
2031 const Triple TT(getTargetMachine().getTargetTriple());
2032 if (GV->hasExternalWeakLinkage() &&
2033 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2037 // If the calling conventions do not match, then we'd better make sure the
2038 // results are returned in the same way as what the caller expects.
2040 SmallVector<CCValAssign, 16> RVLocs1;
2041 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2042 *DAG.getContext(), Call);
2043 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
2045 SmallVector<CCValAssign, 16> RVLocs2;
2046 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2047 *DAG.getContext(), Call);
2048 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
2050 if (RVLocs1.size() != RVLocs2.size())
2052 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2053 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2055 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2057 if (RVLocs1[i].isRegLoc()) {
2058 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2061 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2067 // If Caller's vararg or byval argument has been split between registers and
2068 // stack, do not perform tail call, since part of the argument is in caller's
2070 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
2071 getInfo<ARMFunctionInfo>();
2072 if (AFI_Caller->getArgRegsSaveSize())
2075 // If the callee takes no arguments then go on to check the results of the
2077 if (!Outs.empty()) {
2078 // Check if stack adjustment is needed. For now, do not do this if any
2079 // argument is passed on the stack.
2080 SmallVector<CCValAssign, 16> ArgLocs;
2081 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2082 *DAG.getContext(), Call);
2083 CCInfo.AnalyzeCallOperands(Outs,
2084 CCAssignFnForNode(CalleeCC, false, isVarArg));
2085 if (CCInfo.getNextStackOffset()) {
2086 MachineFunction &MF = DAG.getMachineFunction();
2088 // Check if the arguments are already laid out in the right way as
2089 // the caller's fixed stack objects.
2090 MachineFrameInfo *MFI = MF.getFrameInfo();
2091 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2092 const TargetInstrInfo *TII =
2093 getTargetMachine().getSubtargetImpl()->getInstrInfo();
2094 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2096 ++i, ++realArgIdx) {
2097 CCValAssign &VA = ArgLocs[i];
2098 EVT RegVT = VA.getLocVT();
2099 SDValue Arg = OutVals[realArgIdx];
2100 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2101 if (VA.getLocInfo() == CCValAssign::Indirect)
2103 if (VA.needsCustom()) {
2104 // f64 and vector types are split into multiple registers or
2105 // register/stack-slot combinations. The types will not match
2106 // the registers; give up on memory f64 refs until we figure
2107 // out what to do about this.
2110 if (!ArgLocs[++i].isRegLoc())
2112 if (RegVT == MVT::v2f64) {
2113 if (!ArgLocs[++i].isRegLoc())
2115 if (!ArgLocs[++i].isRegLoc())
2118 } else if (!VA.isRegLoc()) {
2119 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2131 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2132 MachineFunction &MF, bool isVarArg,
2133 const SmallVectorImpl<ISD::OutputArg> &Outs,
2134 LLVMContext &Context) const {
2135 SmallVector<CCValAssign, 16> RVLocs;
2136 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2137 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2141 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2142 SDLoc DL, SelectionDAG &DAG) {
2143 const MachineFunction &MF = DAG.getMachineFunction();
2144 const Function *F = MF.getFunction();
2146 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2148 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2149 // version of the "preferred return address". These offsets affect the return
2150 // instruction if this is a return from PL1 without hypervisor extensions.
2151 // IRQ/FIQ: +4 "subs pc, lr, #4"
2152 // SWI: 0 "subs pc, lr, #0"
2153 // ABORT: +4 "subs pc, lr, #4"
2154 // UNDEF: +4/+2 "subs pc, lr, #0"
2155 // UNDEF varies depending on where the exception came from ARM or Thumb
2156 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2159 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2162 else if (IntKind == "SWI" || IntKind == "UNDEF")
2165 report_fatal_error("Unsupported interrupt attribute. If present, value "
2166 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2168 RetOps.insert(RetOps.begin() + 1, DAG.getConstant(LROffset, MVT::i32, false));
2170 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other, RetOps);
2174 ARMTargetLowering::LowerReturn(SDValue Chain,
2175 CallingConv::ID CallConv, bool isVarArg,
2176 const SmallVectorImpl<ISD::OutputArg> &Outs,
2177 const SmallVectorImpl<SDValue> &OutVals,
2178 SDLoc dl, SelectionDAG &DAG) const {
2180 // CCValAssign - represent the assignment of the return value to a location.
2181 SmallVector<CCValAssign, 16> RVLocs;
2183 // CCState - Info about the registers and stack slots.
2184 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2185 *DAG.getContext(), Call);
2187 // Analyze outgoing return values.
2188 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2192 SmallVector<SDValue, 4> RetOps;
2193 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2194 bool isLittleEndian = Subtarget->isLittle();
2196 MachineFunction &MF = DAG.getMachineFunction();
2197 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2198 AFI->setReturnRegsCount(RVLocs.size());
2200 // Copy the result values into the output registers.
2201 for (unsigned i = 0, realRVLocIdx = 0;
2203 ++i, ++realRVLocIdx) {
2204 CCValAssign &VA = RVLocs[i];
2205 assert(VA.isRegLoc() && "Can only return in registers!");
2207 SDValue Arg = OutVals[realRVLocIdx];
2209 switch (VA.getLocInfo()) {
2210 default: llvm_unreachable("Unknown loc info!");
2211 case CCValAssign::Full: break;
2212 case CCValAssign::BCvt:
2213 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2217 if (VA.needsCustom()) {
2218 if (VA.getLocVT() == MVT::v2f64) {
2219 // Extract the first half and return it in two registers.
2220 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2221 DAG.getConstant(0, MVT::i32));
2222 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2223 DAG.getVTList(MVT::i32, MVT::i32), Half);
2225 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2226 HalfGPRs.getValue(isLittleEndian ? 0 : 1),
2228 Flag = Chain.getValue(1);
2229 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2230 VA = RVLocs[++i]; // skip ahead to next loc
2231 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2232 HalfGPRs.getValue(isLittleEndian ? 1 : 0),
2234 Flag = Chain.getValue(1);
2235 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2236 VA = RVLocs[++i]; // skip ahead to next loc
2238 // Extract the 2nd half and fall through to handle it as an f64 value.
2239 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2240 DAG.getConstant(1, MVT::i32));
2242 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2244 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2245 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2246 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2247 fmrrd.getValue(isLittleEndian ? 0 : 1),
2249 Flag = Chain.getValue(1);
2250 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2251 VA = RVLocs[++i]; // skip ahead to next loc
2252 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2253 fmrrd.getValue(isLittleEndian ? 1 : 0),
2256 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2258 // Guarantee that all emitted copies are
2259 // stuck together, avoiding something bad.
2260 Flag = Chain.getValue(1);
2261 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2264 // Update chain and glue.
2267 RetOps.push_back(Flag);
2269 // CPUs which aren't M-class use a special sequence to return from
2270 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2271 // though we use "subs pc, lr, #N").
2273 // M-class CPUs actually use a normal return sequence with a special
2274 // (hardware-provided) value in LR, so the normal code path works.
2275 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2276 !Subtarget->isMClass()) {
2277 if (Subtarget->isThumb1Only())
2278 report_fatal_error("interrupt attribute is not supported in Thumb1");
2279 return LowerInterruptReturn(RetOps, dl, DAG);
2282 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, RetOps);
2285 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2286 if (N->getNumValues() != 1)
2288 if (!N->hasNUsesOfValue(1, 0))
2291 SDValue TCChain = Chain;
2292 SDNode *Copy = *N->use_begin();
2293 if (Copy->getOpcode() == ISD::CopyToReg) {
2294 // If the copy has a glue operand, we conservatively assume it isn't safe to
2295 // perform a tail call.
2296 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2298 TCChain = Copy->getOperand(0);
2299 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2300 SDNode *VMov = Copy;
2301 // f64 returned in a pair of GPRs.
2302 SmallPtrSet<SDNode*, 2> Copies;
2303 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2305 if (UI->getOpcode() != ISD::CopyToReg)
2309 if (Copies.size() > 2)
2312 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2314 SDValue UseChain = UI->getOperand(0);
2315 if (Copies.count(UseChain.getNode()))
2319 // We are at the top of this chain.
2320 // If the copy has a glue operand, we conservatively assume it
2321 // isn't safe to perform a tail call.
2322 if (UI->getOperand(UI->getNumOperands()-1).getValueType() == MVT::Glue)
2328 } else if (Copy->getOpcode() == ISD::BITCAST) {
2329 // f32 returned in a single GPR.
2330 if (!Copy->hasOneUse())
2332 Copy = *Copy->use_begin();
2333 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2335 // If the copy has a glue operand, we conservatively assume it isn't safe to
2336 // perform a tail call.
2337 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2339 TCChain = Copy->getOperand(0);
2344 bool HasRet = false;
2345 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2347 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2348 UI->getOpcode() != ARMISD::INTRET_FLAG)
2360 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2361 if (!Subtarget->supportsTailCall())
2364 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2367 return !Subtarget->isThumb1Only();
2370 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2371 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2372 // one of the above mentioned nodes. It has to be wrapped because otherwise
2373 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2374 // be used to form addressing mode. These wrapped nodes will be selected
2376 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2377 EVT PtrVT = Op.getValueType();
2378 // FIXME there is no actual debug info here
2380 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2382 if (CP->isMachineConstantPoolEntry())
2383 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2384 CP->getAlignment());
2386 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2387 CP->getAlignment());
2388 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2391 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2392 return MachineJumpTableInfo::EK_Inline;
2395 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2396 SelectionDAG &DAG) const {
2397 MachineFunction &MF = DAG.getMachineFunction();
2398 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2399 unsigned ARMPCLabelIndex = 0;
2401 EVT PtrVT = getPointerTy();
2402 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2403 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2405 if (RelocM == Reloc::Static) {
2406 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2408 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2409 ARMPCLabelIndex = AFI->createPICLabelUId();
2410 ARMConstantPoolValue *CPV =
2411 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2412 ARMCP::CPBlockAddress, PCAdj);
2413 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2415 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2416 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2417 MachinePointerInfo::getConstantPool(),
2418 false, false, false, 0);
2419 if (RelocM == Reloc::Static)
2421 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2422 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2425 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2427 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2428 SelectionDAG &DAG) const {
2430 EVT PtrVT = getPointerTy();
2431 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2432 MachineFunction &MF = DAG.getMachineFunction();
2433 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2434 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2435 ARMConstantPoolValue *CPV =
2436 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2437 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2438 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2439 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2440 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2441 MachinePointerInfo::getConstantPool(),
2442 false, false, false, 0);
2443 SDValue Chain = Argument.getValue(1);
2445 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2446 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2448 // call __tls_get_addr.
2451 Entry.Node = Argument;
2452 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2453 Args.push_back(Entry);
2455 // FIXME: is there useful debug info available here?
2456 TargetLowering::CallLoweringInfo CLI(DAG);
2457 CLI.setDebugLoc(dl).setChain(Chain)
2458 .setCallee(CallingConv::C, Type::getInt32Ty(*DAG.getContext()),
2459 DAG.getExternalSymbol("__tls_get_addr", PtrVT), std::move(Args),
2462 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2463 return CallResult.first;
2466 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2467 // "local exec" model.
2469 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2471 TLSModel::Model model) const {
2472 const GlobalValue *GV = GA->getGlobal();
2475 SDValue Chain = DAG.getEntryNode();
2476 EVT PtrVT = getPointerTy();
2477 // Get the Thread Pointer
2478 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2480 if (model == TLSModel::InitialExec) {
2481 MachineFunction &MF = DAG.getMachineFunction();
2482 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2483 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2484 // Initial exec model.
2485 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2486 ARMConstantPoolValue *CPV =
2487 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2488 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2490 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2491 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2492 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2493 MachinePointerInfo::getConstantPool(),
2494 false, false, false, 0);
2495 Chain = Offset.getValue(1);
2497 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2498 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2500 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2501 MachinePointerInfo::getConstantPool(),
2502 false, false, false, 0);
2505 assert(model == TLSModel::LocalExec);
2506 ARMConstantPoolValue *CPV =
2507 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2508 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2509 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2510 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2511 MachinePointerInfo::getConstantPool(),
2512 false, false, false, 0);
2515 // The address of the thread local variable is the add of the thread
2516 // pointer with the offset of the variable.
2517 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2521 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2522 // TODO: implement the "local dynamic" model
2523 assert(Subtarget->isTargetELF() &&
2524 "TLS not implemented for non-ELF targets");
2525 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2527 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2530 case TLSModel::GeneralDynamic:
2531 case TLSModel::LocalDynamic:
2532 return LowerToTLSGeneralDynamicModel(GA, DAG);
2533 case TLSModel::InitialExec:
2534 case TLSModel::LocalExec:
2535 return LowerToTLSExecModels(GA, DAG, model);
2537 llvm_unreachable("bogus TLS model");
2540 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2541 SelectionDAG &DAG) const {
2542 EVT PtrVT = getPointerTy();
2544 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2545 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2546 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2547 ARMConstantPoolValue *CPV =
2548 ARMConstantPoolConstant::Create(GV,
2549 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2550 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2551 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2552 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2554 MachinePointerInfo::getConstantPool(),
2555 false, false, false, 0);
2556 SDValue Chain = Result.getValue(1);
2557 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2558 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2560 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2561 MachinePointerInfo::getGOT(),
2562 false, false, false, 0);
2566 // If we have T2 ops, we can materialize the address directly via movt/movw
2567 // pair. This is always cheaper.
2568 if (Subtarget->useMovt(DAG.getMachineFunction())) {
2570 // FIXME: Once remat is capable of dealing with instructions with register
2571 // operands, expand this into two nodes.
2572 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2573 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2575 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2576 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2577 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2578 MachinePointerInfo::getConstantPool(),
2579 false, false, false, 0);
2583 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2584 SelectionDAG &DAG) const {
2585 EVT PtrVT = getPointerTy();
2587 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2588 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2590 if (Subtarget->useMovt(DAG.getMachineFunction()))
2593 // FIXME: Once remat is capable of dealing with instructions with register
2594 // operands, expand this into multiple nodes
2596 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2598 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2599 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2601 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2602 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2603 MachinePointerInfo::getGOT(), false, false, false, 0);
2607 SDValue ARMTargetLowering::LowerGlobalAddressWindows(SDValue Op,
2608 SelectionDAG &DAG) const {
2609 assert(Subtarget->isTargetWindows() && "non-Windows COFF is not supported");
2610 assert(Subtarget->useMovt(DAG.getMachineFunction()) &&
2611 "Windows on ARM expects to use movw/movt");
2613 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2614 const ARMII::TOF TargetFlags =
2615 (GV->hasDLLImportStorageClass() ? ARMII::MO_DLLIMPORT : ARMII::MO_NO_FLAG);
2616 EVT PtrVT = getPointerTy();
2622 // FIXME: Once remat is capable of dealing with instructions with register
2623 // operands, expand this into two nodes.
2624 Result = DAG.getNode(ARMISD::Wrapper, DL, PtrVT,
2625 DAG.getTargetGlobalAddress(GV, DL, PtrVT, /*Offset=*/0,
2627 if (GV->hasDLLImportStorageClass())
2628 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
2629 MachinePointerInfo::getGOT(), false, false, false, 0);
2633 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2634 SelectionDAG &DAG) const {
2635 assert(Subtarget->isTargetELF() &&
2636 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2637 MachineFunction &MF = DAG.getMachineFunction();
2638 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2639 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2640 EVT PtrVT = getPointerTy();
2642 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2643 ARMConstantPoolValue *CPV =
2644 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2645 ARMPCLabelIndex, PCAdj);
2646 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2647 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2648 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2649 MachinePointerInfo::getConstantPool(),
2650 false, false, false, 0);
2651 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2652 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2656 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2658 SDValue Val = DAG.getConstant(0, MVT::i32);
2659 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2660 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2661 Op.getOperand(1), Val);
2665 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2667 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2668 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2672 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2673 const ARMSubtarget *Subtarget) const {
2674 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2677 default: return SDValue(); // Don't custom lower most intrinsics.
2678 case Intrinsic::arm_rbit: {
2679 assert(Op.getOperand(1).getValueType() == MVT::i32 &&
2680 "RBIT intrinsic must have i32 type!");
2681 return DAG.getNode(ARMISD::RBIT, dl, MVT::i32, Op.getOperand(1));
2683 case Intrinsic::arm_thread_pointer: {
2684 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2685 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2687 case Intrinsic::eh_sjlj_lsda: {
2688 MachineFunction &MF = DAG.getMachineFunction();
2689 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2690 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2691 EVT PtrVT = getPointerTy();
2692 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2694 unsigned PCAdj = (RelocM != Reloc::PIC_)
2695 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2696 ARMConstantPoolValue *CPV =
2697 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2698 ARMCP::CPLSDA, PCAdj);
2699 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2700 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2702 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2703 MachinePointerInfo::getConstantPool(),
2704 false, false, false, 0);
2706 if (RelocM == Reloc::PIC_) {
2707 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2708 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2712 case Intrinsic::arm_neon_vmulls:
2713 case Intrinsic::arm_neon_vmullu: {
2714 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2715 ? ARMISD::VMULLs : ARMISD::VMULLu;
2716 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2717 Op.getOperand(1), Op.getOperand(2));
2722 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2723 const ARMSubtarget *Subtarget) {
2724 // FIXME: handle "fence singlethread" more efficiently.
2726 if (!Subtarget->hasDataBarrier()) {
2727 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2728 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2730 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2731 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2732 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2733 DAG.getConstant(0, MVT::i32));
2736 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2737 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2738 ARM_MB::MemBOpt Domain = ARM_MB::ISH;
2739 if (Subtarget->isMClass()) {
2740 // Only a full system barrier exists in the M-class architectures.
2741 Domain = ARM_MB::SY;
2742 } else if (Subtarget->isSwift() && Ord == Release) {
2743 // Swift happens to implement ISHST barriers in a way that's compatible with
2744 // Release semantics but weaker than ISH so we'd be fools not to use
2745 // it. Beware: other processors probably don't!
2746 Domain = ARM_MB::ISHST;
2749 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2750 DAG.getConstant(Intrinsic::arm_dmb, MVT::i32),
2751 DAG.getConstant(Domain, MVT::i32));
2754 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2755 const ARMSubtarget *Subtarget) {
2756 // ARM pre v5TE and Thumb1 does not have preload instructions.
2757 if (!(Subtarget->isThumb2() ||
2758 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2759 // Just preserve the chain.
2760 return Op.getOperand(0);
2763 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2765 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2766 // ARMv7 with MP extension has PLDW.
2767 return Op.getOperand(0);
2769 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2770 if (Subtarget->isThumb()) {
2772 isRead = ~isRead & 1;
2773 isData = ~isData & 1;
2776 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2777 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2778 DAG.getConstant(isData, MVT::i32));
2781 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2782 MachineFunction &MF = DAG.getMachineFunction();
2783 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2785 // vastart just stores the address of the VarArgsFrameIndex slot into the
2786 // memory location argument.
2788 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2789 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2790 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2791 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2792 MachinePointerInfo(SV), false, false, 0);
2796 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2797 SDValue &Root, SelectionDAG &DAG,
2799 MachineFunction &MF = DAG.getMachineFunction();
2800 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2802 const TargetRegisterClass *RC;
2803 if (AFI->isThumb1OnlyFunction())
2804 RC = &ARM::tGPRRegClass;
2806 RC = &ARM::GPRRegClass;
2808 // Transform the arguments stored in physical registers into virtual ones.
2809 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2810 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2813 if (NextVA.isMemLoc()) {
2814 MachineFrameInfo *MFI = MF.getFrameInfo();
2815 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2817 // Create load node to retrieve arguments from the stack.
2818 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2819 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2820 MachinePointerInfo::getFixedStack(FI),
2821 false, false, false, 0);
2823 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2824 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2826 if (!Subtarget->isLittle())
2827 std::swap (ArgValue, ArgValue2);
2828 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2832 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2833 unsigned InRegsParamRecordIdx,
2835 unsigned &ArgRegsSize,
2836 unsigned &ArgRegsSaveSize)
2839 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2840 unsigned RBegin, REnd;
2841 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2842 NumGPRs = REnd - RBegin;
2844 unsigned int firstUnalloced;
2845 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2846 sizeof(GPRArgRegs) /
2847 sizeof(GPRArgRegs[0]));
2848 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2851 unsigned Align = MF.getTarget()
2853 ->getFrameLowering()
2854 ->getStackAlignment();
2855 ArgRegsSize = NumGPRs * 4;
2857 // If parameter is split between stack and GPRs...
2858 if (NumGPRs && Align > 4 &&
2859 (ArgRegsSize < ArgSize ||
2860 InRegsParamRecordIdx >= CCInfo.getInRegsParamsCount())) {
2861 // Add padding for part of param recovered from GPRs. For example,
2862 // if Align == 8, its last byte must be at address K*8 - 1.
2863 // We need to do it, since remained (stack) part of parameter has
2864 // stack alignment, and we need to "attach" "GPRs head" without gaps
2867 // |---- 8 bytes block ----| |---- 8 bytes block ----| |---- 8 bytes...
2868 // [ [padding] [GPRs head] ] [ Tail passed via stack ....
2870 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2872 OffsetToAlignment(ArgRegsSize + AFI->getArgRegsSaveSize(), Align);
2873 ArgRegsSaveSize = ArgRegsSize + Padding;
2875 // We don't need to extend regs save size for byval parameters if they
2876 // are passed via GPRs only.
2877 ArgRegsSaveSize = ArgRegsSize;
2880 // The remaining GPRs hold either the beginning of variable-argument
2881 // data, or the beginning of an aggregate passed by value (usually
2882 // byval). Either way, we allocate stack slots adjacent to the data
2883 // provided by our caller, and store the unallocated registers there.
2884 // If this is a variadic function, the va_list pointer will begin with
2885 // these values; otherwise, this reassembles a (byval) structure that
2886 // was split between registers and memory.
2887 // Return: The frame index registers were stored into.
2889 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2890 SDLoc dl, SDValue &Chain,
2891 const Value *OrigArg,
2892 unsigned InRegsParamRecordIdx,
2893 unsigned OffsetFromOrigArg,
2897 unsigned ByValStoreOffset,
2898 unsigned TotalArgRegsSaveSize) const {
2900 // Currently, two use-cases possible:
2901 // Case #1. Non-var-args function, and we meet first byval parameter.
2902 // Setup first unallocated register as first byval register;
2903 // eat all remained registers
2904 // (these two actions are performed by HandleByVal method).
2905 // Then, here, we initialize stack frame with
2906 // "store-reg" instructions.
2907 // Case #2. Var-args function, that doesn't contain byval parameters.
2908 // The same: eat all remained unallocated registers,
2909 // initialize stack frame.
2911 MachineFunction &MF = DAG.getMachineFunction();
2912 MachineFrameInfo *MFI = MF.getFrameInfo();
2913 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2914 unsigned firstRegToSaveIndex, lastRegToSaveIndex;
2915 unsigned RBegin, REnd;
2916 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2917 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2918 firstRegToSaveIndex = RBegin - ARM::R0;
2919 lastRegToSaveIndex = REnd - ARM::R0;
2921 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2922 (GPRArgRegs, array_lengthof(GPRArgRegs));
2923 lastRegToSaveIndex = 4;
2926 unsigned ArgRegsSize, ArgRegsSaveSize;
2927 computeRegArea(CCInfo, MF, InRegsParamRecordIdx, ArgSize,
2928 ArgRegsSize, ArgRegsSaveSize);
2930 // Store any by-val regs to their spots on the stack so that they may be
2931 // loaded by deferencing the result of formal parameter pointer or va_next.
2932 // Note: once stack area for byval/varargs registers
2933 // was initialized, it can't be initialized again.
2934 if (ArgRegsSaveSize) {
2935 unsigned Padding = ArgRegsSaveSize - ArgRegsSize;
2938 assert(AFI->getStoredByValParamsPadding() == 0 &&
2939 "The only parameter may be padded.");
2940 AFI->setStoredByValParamsPadding(Padding);
2943 int FrameIndex = MFI->CreateFixedObject(ArgRegsSaveSize,
2946 (int64_t)TotalArgRegsSaveSize,
2948 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());
2950 MFI->CreateFixedObject(Padding,
2951 ArgOffset + ByValStoreOffset -
2952 (int64_t)ArgRegsSaveSize,
2956 SmallVector<SDValue, 4> MemOps;
2957 for (unsigned i = 0; firstRegToSaveIndex < lastRegToSaveIndex;
2958 ++firstRegToSaveIndex, ++i) {
2959 const TargetRegisterClass *RC;
2960 if (AFI->isThumb1OnlyFunction())
2961 RC = &ARM::tGPRRegClass;
2963 RC = &ARM::GPRRegClass;
2965 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2966 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2968 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2969 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2971 MemOps.push_back(Store);
2972 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2973 DAG.getConstant(4, getPointerTy()));
2976 AFI->setArgRegsSaveSize(ArgRegsSaveSize + AFI->getArgRegsSaveSize());
2978 if (!MemOps.empty())
2979 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2983 // We cannot allocate a zero-byte object for the first variadic argument,
2984 // so just make up a size.
2987 // This will point to the next argument passed via stack.
2988 return MFI->CreateFixedObject(
2989 ArgSize, ArgOffset, !ForceMutable);
2993 // Setup stack frame, the va_list pointer will start from.
2995 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2996 SDLoc dl, SDValue &Chain,
2998 unsigned TotalArgRegsSaveSize,
2999 bool ForceMutable) const {
3000 MachineFunction &MF = DAG.getMachineFunction();
3001 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3003 // Try to store any remaining integer argument regs
3004 // to their spots on the stack so that they may be loaded by deferencing
3005 // the result of va_next.
3006 // If there is no regs to be stored, just point address after last
3007 // argument passed via stack.
3009 StoreByValRegs(CCInfo, DAG, dl, Chain, nullptr,
3010 CCInfo.getInRegsParamsCount(), 0, ArgOffset, 0, ForceMutable,
3011 0, TotalArgRegsSaveSize);
3013 AFI->setVarArgsFrameIndex(FrameIndex);
3017 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
3018 CallingConv::ID CallConv, bool isVarArg,
3019 const SmallVectorImpl<ISD::InputArg>
3021 SDLoc dl, SelectionDAG &DAG,
3022 SmallVectorImpl<SDValue> &InVals)
3024 MachineFunction &MF = DAG.getMachineFunction();
3025 MachineFrameInfo *MFI = MF.getFrameInfo();
3027 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3029 // Assign locations to all of the incoming arguments.
3030 SmallVector<CCValAssign, 16> ArgLocs;
3031 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
3032 *DAG.getContext(), Prologue);
3033 CCInfo.AnalyzeFormalArguments(Ins,
3034 CCAssignFnForNode(CallConv, /* Return*/ false,
3037 SmallVector<SDValue, 16> ArgValues;
3038 int lastInsIndex = -1;
3040 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
3041 unsigned CurArgIdx = 0;
3043 // Initially ArgRegsSaveSize is zero.
3044 // Then we increase this value each time we meet byval parameter.
3045 // We also increase this value in case of varargs function.
3046 AFI->setArgRegsSaveSize(0);
3048 unsigned ByValStoreOffset = 0;
3049 unsigned TotalArgRegsSaveSize = 0;
3050 unsigned ArgRegsSaveSizeMaxAlign = 4;
3052 // Calculate the amount of stack space that we need to allocate to store
3053 // byval and variadic arguments that are passed in registers.
3054 // We need to know this before we allocate the first byval or variadic
3055 // argument, as they will be allocated a stack slot below the CFA (Canonical
3056 // Frame Address, the stack pointer at entry to the function).
3057 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3058 CCValAssign &VA = ArgLocs[i];
3059 if (VA.isMemLoc()) {
3060 int index = VA.getValNo();
3061 if (index != lastInsIndex) {
3062 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3063 if (Flags.isByVal()) {
3064 unsigned ExtraArgRegsSize;
3065 unsigned ExtraArgRegsSaveSize;
3066 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsProcessed(),
3067 Flags.getByValSize(),
3068 ExtraArgRegsSize, ExtraArgRegsSaveSize);
3070 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
3071 if (Flags.getByValAlign() > ArgRegsSaveSizeMaxAlign)
3072 ArgRegsSaveSizeMaxAlign = Flags.getByValAlign();
3073 CCInfo.nextInRegsParam();
3075 lastInsIndex = index;
3079 CCInfo.rewindByValRegsInfo();
3081 if (isVarArg && MFI->hasVAStart()) {
3082 unsigned ExtraArgRegsSize;
3083 unsigned ExtraArgRegsSaveSize;
3084 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsCount(), 0,
3085 ExtraArgRegsSize, ExtraArgRegsSaveSize);
3086 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
3088 // If the arg regs save area contains N-byte aligned values, the
3089 // bottom of it must be at least N-byte aligned.
3090 TotalArgRegsSaveSize = RoundUpToAlignment(TotalArgRegsSaveSize, ArgRegsSaveSizeMaxAlign);
3091 TotalArgRegsSaveSize = std::min(TotalArgRegsSaveSize, 16U);
3093 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3094 CCValAssign &VA = ArgLocs[i];
3095 if (Ins[VA.getValNo()].isOrigArg()) {
3096 std::advance(CurOrigArg,
3097 Ins[VA.getValNo()].getOrigArgIndex() - CurArgIdx);
3098 CurArgIdx = Ins[VA.getValNo()].getOrigArgIndex();
3100 // Arguments stored in registers.
3101 if (VA.isRegLoc()) {
3102 EVT RegVT = VA.getLocVT();
3104 if (VA.needsCustom()) {
3105 // f64 and vector types are split up into multiple registers or
3106 // combinations of registers and stack slots.
3107 if (VA.getLocVT() == MVT::v2f64) {
3108 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
3110 VA = ArgLocs[++i]; // skip ahead to next loc
3112 if (VA.isMemLoc()) {
3113 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
3114 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3115 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
3116 MachinePointerInfo::getFixedStack(FI),
3117 false, false, false, 0);
3119 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
3122 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
3123 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3124 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
3125 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3126 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
3128 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3131 const TargetRegisterClass *RC;
3133 if (RegVT == MVT::f32)
3134 RC = &ARM::SPRRegClass;
3135 else if (RegVT == MVT::f64)
3136 RC = &ARM::DPRRegClass;
3137 else if (RegVT == MVT::v2f64)
3138 RC = &ARM::QPRRegClass;
3139 else if (RegVT == MVT::i32)
3140 RC = AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass
3141 : &ARM::GPRRegClass;
3143 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3145 // Transform the arguments in physical registers into virtual ones.
3146 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3147 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3150 // If this is an 8 or 16-bit value, it is really passed promoted
3151 // to 32 bits. Insert an assert[sz]ext to capture this, then
3152 // truncate to the right size.
3153 switch (VA.getLocInfo()) {
3154 default: llvm_unreachable("Unknown loc info!");
3155 case CCValAssign::Full: break;
3156 case CCValAssign::BCvt:
3157 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3159 case CCValAssign::SExt:
3160 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3161 DAG.getValueType(VA.getValVT()));
3162 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3164 case CCValAssign::ZExt:
3165 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3166 DAG.getValueType(VA.getValVT()));
3167 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3171 InVals.push_back(ArgValue);
3173 } else { // VA.isRegLoc()
3176 assert(VA.isMemLoc());
3177 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3179 int index = VA.getValNo();
3181 // Some Ins[] entries become multiple ArgLoc[] entries.
3182 // Process them only once.
3183 if (index != lastInsIndex)
3185 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3186 // FIXME: For now, all byval parameter objects are marked mutable.
3187 // This can be changed with more analysis.
3188 // In case of tail call optimization mark all arguments mutable.
3189 // Since they could be overwritten by lowering of arguments in case of
3191 if (Flags.isByVal()) {
3192 assert(Ins[index].isOrigArg() &&
3193 "Byval arguments cannot be implicit");
3194 unsigned CurByValIndex = CCInfo.getInRegsParamsProcessed();
3196 ByValStoreOffset = RoundUpToAlignment(ByValStoreOffset, Flags.getByValAlign());
3197 int FrameIndex = StoreByValRegs(
3198 CCInfo, DAG, dl, Chain, CurOrigArg,
3200 Ins[VA.getValNo()].PartOffset,
3201 VA.getLocMemOffset(),
3202 Flags.getByValSize(),
3203 true /*force mutable frames*/,
3205 TotalArgRegsSaveSize);
3206 ByValStoreOffset += Flags.getByValSize();
3207 ByValStoreOffset = std::min(ByValStoreOffset, 16U);
3208 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
3209 CCInfo.nextInRegsParam();
3211 unsigned FIOffset = VA.getLocMemOffset();
3212 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3215 // Create load nodes to retrieve arguments from the stack.
3216 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3217 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3218 MachinePointerInfo::getFixedStack(FI),
3219 false, false, false, 0));
3221 lastInsIndex = index;
3227 if (isVarArg && MFI->hasVAStart())
3228 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3229 CCInfo.getNextStackOffset(),
3230 TotalArgRegsSaveSize);
3232 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3237 /// isFloatingPointZero - Return true if this is +0.0.
3238 static bool isFloatingPointZero(SDValue Op) {
3239 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3240 return CFP->getValueAPF().isPosZero();
3241 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3242 // Maybe this has already been legalized into the constant pool?
3243 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3244 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3245 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3246 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3247 return CFP->getValueAPF().isPosZero();
3249 } else if (Op->getOpcode() == ISD::BITCAST &&
3250 Op->getValueType(0) == MVT::f64) {
3251 // Handle (ISD::BITCAST (ARMISD::VMOVIMM (ISD::TargetConstant 0)) MVT::f64)
3252 // created by LowerConstantFP().
3253 SDValue BitcastOp = Op->getOperand(0);
3254 if (BitcastOp->getOpcode() == ARMISD::VMOVIMM) {
3255 SDValue MoveOp = BitcastOp->getOperand(0);
3256 if (MoveOp->getOpcode() == ISD::TargetConstant &&
3257 cast<ConstantSDNode>(MoveOp)->getZExtValue() == 0) {
3265 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3266 /// the given operands.
3268 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3269 SDValue &ARMcc, SelectionDAG &DAG,
3271 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3272 unsigned C = RHSC->getZExtValue();
3273 if (!isLegalICmpImmediate(C)) {
3274 // Constant does not fit, try adjusting it by one?
3279 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3280 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3281 RHS = DAG.getConstant(C-1, MVT::i32);
3286 if (C != 0 && isLegalICmpImmediate(C-1)) {
3287 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3288 RHS = DAG.getConstant(C-1, MVT::i32);
3293 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3294 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3295 RHS = DAG.getConstant(C+1, MVT::i32);
3300 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3301 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3302 RHS = DAG.getConstant(C+1, MVT::i32);
3309 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3310 ARMISD::NodeType CompareType;
3313 CompareType = ARMISD::CMP;
3318 CompareType = ARMISD::CMPZ;
3321 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3322 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3325 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3327 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3329 assert(!Subtarget->isFPOnlySP() || RHS.getValueType() != MVT::f64);
3331 if (!isFloatingPointZero(RHS))
3332 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3334 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3335 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3338 /// duplicateCmp - Glue values can have only one use, so this function
3339 /// duplicates a comparison node.
3341 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3342 unsigned Opc = Cmp.getOpcode();
3344 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3345 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3347 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3348 Cmp = Cmp.getOperand(0);
3349 Opc = Cmp.getOpcode();
3350 if (Opc == ARMISD::CMPFP)
3351 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3353 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3354 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3356 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3359 std::pair<SDValue, SDValue>
3360 ARMTargetLowering::getARMXALUOOp(SDValue Op, SelectionDAG &DAG,
3361 SDValue &ARMcc) const {
3362 assert(Op.getValueType() == MVT::i32 && "Unsupported value type");
3364 SDValue Value, OverflowCmp;
3365 SDValue LHS = Op.getOperand(0);
3366 SDValue RHS = Op.getOperand(1);
3369 // FIXME: We are currently always generating CMPs because we don't support
3370 // generating CMN through the backend. This is not as good as the natural
3371 // CMP case because it causes a register dependency and cannot be folded
3374 switch (Op.getOpcode()) {
3376 llvm_unreachable("Unknown overflow instruction!");
3378 ARMcc = DAG.getConstant(ARMCC::VC, MVT::i32);
3379 Value = DAG.getNode(ISD::ADD, SDLoc(Op), Op.getValueType(), LHS, RHS);
3380 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, Value, LHS);
3383 ARMcc = DAG.getConstant(ARMCC::HS, MVT::i32);
3384 Value = DAG.getNode(ISD::ADD, SDLoc(Op), Op.getValueType(), LHS, RHS);
3385 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, Value, LHS);
3388 ARMcc = DAG.getConstant(ARMCC::VC, MVT::i32);
3389 Value = DAG.getNode(ISD::SUB, SDLoc(Op), Op.getValueType(), LHS, RHS);
3390 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, LHS, RHS);
3393 ARMcc = DAG.getConstant(ARMCC::HS, MVT::i32);
3394 Value = DAG.getNode(ISD::SUB, SDLoc(Op), Op.getValueType(), LHS, RHS);
3395 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, LHS, RHS);
3399 return std::make_pair(Value, OverflowCmp);
3404 ARMTargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
3405 // Let legalize expand this if it isn't a legal type yet.
3406 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
3409 SDValue Value, OverflowCmp;
3411 std::tie(Value, OverflowCmp) = getARMXALUOOp(Op, DAG, ARMcc);
3412 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3413 // We use 0 and 1 as false and true values.
3414 SDValue TVal = DAG.getConstant(1, MVT::i32);
3415 SDValue FVal = DAG.getConstant(0, MVT::i32);
3416 EVT VT = Op.getValueType();
3418 SDValue Overflow = DAG.getNode(ARMISD::CMOV, SDLoc(Op), VT, TVal, FVal,
3419 ARMcc, CCR, OverflowCmp);
3421 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
3422 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
3426 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3427 SDValue Cond = Op.getOperand(0);
3428 SDValue SelectTrue = Op.getOperand(1);
3429 SDValue SelectFalse = Op.getOperand(2);
3431 unsigned Opc = Cond.getOpcode();
3433 if (Cond.getResNo() == 1 &&
3434 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3435 Opc == ISD::USUBO)) {
3436 if (!DAG.getTargetLoweringInfo().isTypeLegal(Cond->getValueType(0)))
3439 SDValue Value, OverflowCmp;
3441 std::tie(Value, OverflowCmp) = getARMXALUOOp(Cond, DAG, ARMcc);
3442 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3443 EVT VT = Op.getValueType();
3445 return getCMOV(SDLoc(Op), VT, SelectTrue, SelectFalse, ARMcc, CCR,
3451 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3452 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3454 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3455 const ConstantSDNode *CMOVTrue =
3456 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3457 const ConstantSDNode *CMOVFalse =
3458 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3460 if (CMOVTrue && CMOVFalse) {
3461 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3462 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3466 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3468 False = SelectFalse;
3469 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3474 if (True.getNode() && False.getNode()) {
3475 EVT VT = Op.getValueType();
3476 SDValue ARMcc = Cond.getOperand(2);
3477 SDValue CCR = Cond.getOperand(3);
3478 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3479 assert(True.getValueType() == VT);
3480 return getCMOV(dl, VT, True, False, ARMcc, CCR, Cmp, DAG);
3485 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3486 // undefined bits before doing a full-word comparison with zero.
3487 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3488 DAG.getConstant(1, Cond.getValueType()));
3490 return DAG.getSelectCC(dl, Cond,
3491 DAG.getConstant(0, Cond.getValueType()),
3492 SelectTrue, SelectFalse, ISD::SETNE);
3495 static ISD::CondCode getInverseCCForVSEL(ISD::CondCode CC) {
3496 if (CC == ISD::SETNE)
3498 return ISD::getSetCCInverse(CC, true);
3501 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3502 bool &swpCmpOps, bool &swpVselOps) {
3503 // Start by selecting the GE condition code for opcodes that return true for
3505 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3507 CondCode = ARMCC::GE;
3509 // and GT for opcodes that return false for 'equality'.
3510 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3512 CondCode = ARMCC::GT;
3514 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3515 // to swap the compare operands.
3516 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3520 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3521 // If we have an unordered opcode, we need to swap the operands to the VSEL
3522 // instruction (effectively negating the condition).
3524 // This also has the effect of swapping which one of 'less' or 'greater'
3525 // returns true, so we also swap the compare operands. It also switches
3526 // whether we return true for 'equality', so we compensate by picking the
3527 // opposite condition code to our original choice.
3528 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3529 CC == ISD::SETUGT) {
3530 swpCmpOps = !swpCmpOps;
3531 swpVselOps = !swpVselOps;
3532 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3535 // 'ordered' is 'anything but unordered', so use the VS condition code and
3536 // swap the VSEL operands.
3537 if (CC == ISD::SETO) {
3538 CondCode = ARMCC::VS;
3542 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3543 // code and swap the VSEL operands.
3544 if (CC == ISD::SETUNE) {
3545 CondCode = ARMCC::EQ;
3550 SDValue ARMTargetLowering::getCMOV(SDLoc dl, EVT VT, SDValue FalseVal,
3551 SDValue TrueVal, SDValue ARMcc, SDValue CCR,
3552 SDValue Cmp, SelectionDAG &DAG) const {
3553 if (Subtarget->isFPOnlySP() && VT == MVT::f64) {
3554 FalseVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3555 DAG.getVTList(MVT::i32, MVT::i32), FalseVal);
3556 TrueVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3557 DAG.getVTList(MVT::i32, MVT::i32), TrueVal);
3559 SDValue TrueLow = TrueVal.getValue(0);
3560 SDValue TrueHigh = TrueVal.getValue(1);
3561 SDValue FalseLow = FalseVal.getValue(0);
3562 SDValue FalseHigh = FalseVal.getValue(1);
3564 SDValue Low = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseLow, TrueLow,
3566 SDValue High = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseHigh, TrueHigh,
3567 ARMcc, CCR, duplicateCmp(Cmp, DAG));
3569 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Low, High);
3571 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3576 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3577 EVT VT = Op.getValueType();
3578 SDValue LHS = Op.getOperand(0);
3579 SDValue RHS = Op.getOperand(1);
3580 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3581 SDValue TrueVal = Op.getOperand(2);
3582 SDValue FalseVal = Op.getOperand(3);
3585 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3586 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3589 // If softenSetCCOperands only returned one value, we should compare it to
3591 if (!RHS.getNode()) {
3592 RHS = DAG.getConstant(0, LHS.getValueType());
3597 if (LHS.getValueType() == MVT::i32) {
3598 // Try to generate VSEL on ARMv8.
3599 // The VSEL instruction can't use all the usual ARM condition
3600 // codes: it only has two bits to select the condition code, so it's
3601 // constrained to use only GE, GT, VS and EQ.
3603 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3604 // swap the operands of the previous compare instruction (effectively
3605 // inverting the compare condition, swapping 'less' and 'greater') and
3606 // sometimes need to swap the operands to the VSEL (which inverts the
3607 // condition in the sense of firing whenever the previous condition didn't)
3608 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3609 TrueVal.getValueType() == MVT::f64)) {
3610 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3611 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3612 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3613 CC = getInverseCCForVSEL(CC);
3614 std::swap(TrueVal, FalseVal);
3619 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3620 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3621 return getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3624 ARMCC::CondCodes CondCode, CondCode2;
3625 FPCCToARMCC(CC, CondCode, CondCode2);
3627 // Try to generate VSEL on ARMv8.
3628 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3629 TrueVal.getValueType() == MVT::f64)) {
3630 // We can select VMAXNM/VMINNM from a compare followed by a select with the
3631 // same operands, as follows:
3632 // c = fcmp [ogt, olt, ugt, ult] a, b
3634 // We only do this in unsafe-fp-math, because signed zeros and NaNs are
3635 // handled differently than the original code sequence.
3636 if (getTargetMachine().Options.UnsafeFPMath) {
3637 if (LHS == TrueVal && RHS == FalseVal) {
3638 if (CC == ISD::SETOGT || CC == ISD::SETUGT)
3639 return DAG.getNode(ARMISD::VMAXNM, dl, VT, TrueVal, FalseVal);
3640 if (CC == ISD::SETOLT || CC == ISD::SETULT)
3641 return DAG.getNode(ARMISD::VMINNM, dl, VT, TrueVal, FalseVal);
3642 } else if (LHS == FalseVal && RHS == TrueVal) {
3643 if (CC == ISD::SETOLT || CC == ISD::SETULT)
3644 return DAG.getNode(ARMISD::VMAXNM, dl, VT, TrueVal, FalseVal);
3645 if (CC == ISD::SETOGT || CC == ISD::SETUGT)
3646 return DAG.getNode(ARMISD::VMINNM, dl, VT, TrueVal, FalseVal);
3650 bool swpCmpOps = false;
3651 bool swpVselOps = false;
3652 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3654 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3655 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3657 std::swap(LHS, RHS);
3659 std::swap(TrueVal, FalseVal);
3663 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3664 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3665 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3666 SDValue Result = getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3667 if (CondCode2 != ARMCC::AL) {
3668 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
3669 // FIXME: Needs another CMP because flag can have but one use.
3670 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3671 Result = getCMOV(dl, VT, Result, TrueVal, ARMcc2, CCR, Cmp2, DAG);
3676 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3677 /// to morph to an integer compare sequence.
3678 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3679 const ARMSubtarget *Subtarget) {
3680 SDNode *N = Op.getNode();
3681 if (!N->hasOneUse())
3682 // Otherwise it requires moving the value from fp to integer registers.
3684 if (!N->getNumValues())
3686 EVT VT = Op.getValueType();
3687 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3688 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3689 // vmrs are very slow, e.g. cortex-a8.
3692 if (isFloatingPointZero(Op)) {
3696 return ISD::isNormalLoad(N);
3699 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3700 if (isFloatingPointZero(Op))
3701 return DAG.getConstant(0, MVT::i32);
3703 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3704 return DAG.getLoad(MVT::i32, SDLoc(Op),
3705 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3706 Ld->isVolatile(), Ld->isNonTemporal(),
3707 Ld->isInvariant(), Ld->getAlignment());
3709 llvm_unreachable("Unknown VFP cmp argument!");
3712 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3713 SDValue &RetVal1, SDValue &RetVal2) {
3714 if (isFloatingPointZero(Op)) {
3715 RetVal1 = DAG.getConstant(0, MVT::i32);
3716 RetVal2 = DAG.getConstant(0, MVT::i32);
3720 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3721 SDValue Ptr = Ld->getBasePtr();
3722 RetVal1 = DAG.getLoad(MVT::i32, SDLoc(Op),
3723 Ld->getChain(), Ptr,
3724 Ld->getPointerInfo(),
3725 Ld->isVolatile(), Ld->isNonTemporal(),
3726 Ld->isInvariant(), Ld->getAlignment());
3728 EVT PtrType = Ptr.getValueType();
3729 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3730 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(Op),
3731 PtrType, Ptr, DAG.getConstant(4, PtrType));
3732 RetVal2 = DAG.getLoad(MVT::i32, SDLoc(Op),
3733 Ld->getChain(), NewPtr,
3734 Ld->getPointerInfo().getWithOffset(4),
3735 Ld->isVolatile(), Ld->isNonTemporal(),
3736 Ld->isInvariant(), NewAlign);
3740 llvm_unreachable("Unknown VFP cmp argument!");
3743 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3744 /// f32 and even f64 comparisons to integer ones.
3746 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3747 SDValue Chain = Op.getOperand(0);
3748 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3749 SDValue LHS = Op.getOperand(2);
3750 SDValue RHS = Op.getOperand(3);
3751 SDValue Dest = Op.getOperand(4);
3754 bool LHSSeenZero = false;
3755 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3756 bool RHSSeenZero = false;
3757 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3758 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3759 // If unsafe fp math optimization is enabled and there are no other uses of
3760 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3761 // to an integer comparison.
3762 if (CC == ISD::SETOEQ)
3764 else if (CC == ISD::SETUNE)
3767 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3769 if (LHS.getValueType() == MVT::f32) {
3770 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3771 bitcastf32Toi32(LHS, DAG), Mask);
3772 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3773 bitcastf32Toi32(RHS, DAG), Mask);
3774 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3775 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3776 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3777 Chain, Dest, ARMcc, CCR, Cmp);
3782 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3783 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3784 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3785 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3786 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3787 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3788 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3789 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3790 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops);
3796 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3797 SDValue Chain = Op.getOperand(0);
3798 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3799 SDValue LHS = Op.getOperand(2);
3800 SDValue RHS = Op.getOperand(3);
3801 SDValue Dest = Op.getOperand(4);
3804 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3805 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3808 // If softenSetCCOperands only returned one value, we should compare it to
3810 if (!RHS.getNode()) {
3811 RHS = DAG.getConstant(0, LHS.getValueType());
3816 if (LHS.getValueType() == MVT::i32) {
3818 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3819 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3820 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3821 Chain, Dest, ARMcc, CCR, Cmp);
3824 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3826 if (getTargetMachine().Options.UnsafeFPMath &&
3827 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3828 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3829 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3830 if (Result.getNode())
3834 ARMCC::CondCodes CondCode, CondCode2;
3835 FPCCToARMCC(CC, CondCode, CondCode2);
3837 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3838 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3839 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3840 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3841 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3842 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3843 if (CondCode2 != ARMCC::AL) {
3844 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3845 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3846 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3851 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3852 SDValue Chain = Op.getOperand(0);
3853 SDValue Table = Op.getOperand(1);
3854 SDValue Index = Op.getOperand(2);
3857 EVT PTy = getPointerTy();
3858 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3859 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3860 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3861 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3862 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3863 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3864 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3865 if (Subtarget->isThumb2()) {
3866 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3867 // which does another jump to the destination. This also makes it easier
3868 // to translate it to TBB / TBH later.
3869 // FIXME: This might not work if the function is extremely large.
3870 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3871 Addr, Op.getOperand(2), JTI, UId);
3873 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3874 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3875 MachinePointerInfo::getJumpTable(),
3876 false, false, false, 0);
3877 Chain = Addr.getValue(1);
3878 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3879 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3881 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3882 MachinePointerInfo::getJumpTable(),
3883 false, false, false, 0);
3884 Chain = Addr.getValue(1);
3885 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3889 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3890 EVT VT = Op.getValueType();
3893 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3894 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3896 return DAG.UnrollVectorOp(Op.getNode());
3899 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3900 "Invalid type for custom lowering!");
3901 if (VT != MVT::v4i16)
3902 return DAG.UnrollVectorOp(Op.getNode());
3904 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3905 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3908 SDValue ARMTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const {
3909 EVT VT = Op.getValueType();
3911 return LowerVectorFP_TO_INT(Op, DAG);
3913 if (Subtarget->isFPOnlySP() && Op.getOperand(0).getValueType() == MVT::f64) {
3915 if (Op.getOpcode() == ISD::FP_TO_SINT)
3916 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(),
3919 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(),
3921 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3922 /*isSigned*/ false, SDLoc(Op)).first;
3928 switch (Op.getOpcode()) {
3929 default: llvm_unreachable("Invalid opcode!");
3930 case ISD::FP_TO_SINT:
3931 Opc = ARMISD::FTOSI;
3933 case ISD::FP_TO_UINT:
3934 Opc = ARMISD::FTOUI;
3937 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3938 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3941 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3942 EVT VT = Op.getValueType();
3945 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3946 if (VT.getVectorElementType() == MVT::f32)
3948 return DAG.UnrollVectorOp(Op.getNode());
3951 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3952 "Invalid type for custom lowering!");
3953 if (VT != MVT::v4f32)
3954 return DAG.UnrollVectorOp(Op.getNode());
3958 switch (Op.getOpcode()) {
3959 default: llvm_unreachable("Invalid opcode!");
3960 case ISD::SINT_TO_FP:
3961 CastOpc = ISD::SIGN_EXTEND;
3962 Opc = ISD::SINT_TO_FP;
3964 case ISD::UINT_TO_FP:
3965 CastOpc = ISD::ZERO_EXTEND;
3966 Opc = ISD::UINT_TO_FP;
3970 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3971 return DAG.getNode(Opc, dl, VT, Op);
3974 SDValue ARMTargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const {
3975 EVT VT = Op.getValueType();
3977 return LowerVectorINT_TO_FP(Op, DAG);
3979 if (Subtarget->isFPOnlySP() && Op.getValueType() == MVT::f64) {
3981 if (Op.getOpcode() == ISD::SINT_TO_FP)
3982 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(),
3985 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(),
3987 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3988 /*isSigned*/ false, SDLoc(Op)).first;
3994 switch (Op.getOpcode()) {
3995 default: llvm_unreachable("Invalid opcode!");
3996 case ISD::SINT_TO_FP:
3997 Opc = ARMISD::SITOF;
3999 case ISD::UINT_TO_FP:
4000 Opc = ARMISD::UITOF;
4004 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
4005 return DAG.getNode(Opc, dl, VT, Op);
4008 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
4009 // Implement fcopysign with a fabs and a conditional fneg.
4010 SDValue Tmp0 = Op.getOperand(0);
4011 SDValue Tmp1 = Op.getOperand(1);
4013 EVT VT = Op.getValueType();
4014 EVT SrcVT = Tmp1.getValueType();
4015 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
4016 Tmp0.getOpcode() == ARMISD::VMOVDRR;
4017 bool UseNEON = !InGPR && Subtarget->hasNEON();
4020 // Use VBSL to copy the sign bit.
4021 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
4022 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
4023 DAG.getTargetConstant(EncodedVal, MVT::i32));
4024 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
4026 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
4027 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
4028 DAG.getConstant(32, MVT::i32));
4029 else /*if (VT == MVT::f32)*/
4030 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
4031 if (SrcVT == MVT::f32) {
4032 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
4034 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
4035 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
4036 DAG.getConstant(32, MVT::i32));
4037 } else if (VT == MVT::f32)
4038 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
4039 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
4040 DAG.getConstant(32, MVT::i32));
4041 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
4042 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
4044 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
4046 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
4047 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
4048 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
4050 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
4051 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
4052 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
4053 if (VT == MVT::f32) {
4054 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
4055 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
4056 DAG.getConstant(0, MVT::i32));
4058 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
4064 // Bitcast operand 1 to i32.
4065 if (SrcVT == MVT::f64)
4066 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4068 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
4070 // Or in the signbit with integer operations.
4071 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
4072 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
4073 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
4074 if (VT == MVT::f32) {
4075 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
4076 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
4077 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4078 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
4081 // f64: Or the high part with signbit and then combine two parts.
4082 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4084 SDValue Lo = Tmp0.getValue(0);
4085 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
4086 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
4087 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
4090 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
4091 MachineFunction &MF = DAG.getMachineFunction();
4092 MachineFrameInfo *MFI = MF.getFrameInfo();
4093 MFI->setReturnAddressIsTaken(true);
4095 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
4098 EVT VT = Op.getValueType();
4100 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4102 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4103 SDValue Offset = DAG.getConstant(4, MVT::i32);
4104 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
4105 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
4106 MachinePointerInfo(), false, false, false, 0);
4109 // Return LR, which contains the return address. Mark it an implicit live-in.
4110 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
4111 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
4114 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
4115 const ARMBaseRegisterInfo &ARI =
4116 *static_cast<const ARMBaseRegisterInfo*>(RegInfo);
4117 MachineFunction &MF = DAG.getMachineFunction();
4118 MachineFrameInfo *MFI = MF.getFrameInfo();
4119 MFI->setFrameAddressIsTaken(true);
4121 EVT VT = Op.getValueType();
4122 SDLoc dl(Op); // FIXME probably not meaningful
4123 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4124 unsigned FrameReg = ARI.getFrameRegister(MF);
4125 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
4127 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
4128 MachinePointerInfo(),
4129 false, false, false, 0);
4133 // FIXME? Maybe this could be a TableGen attribute on some registers and
4134 // this table could be generated automatically from RegInfo.
4135 unsigned ARMTargetLowering::getRegisterByName(const char* RegName,
4137 unsigned Reg = StringSwitch<unsigned>(RegName)
4138 .Case("sp", ARM::SP)
4142 report_fatal_error("Invalid register name global variable");
4145 /// ExpandBITCAST - If the target supports VFP, this function is called to
4146 /// expand a bit convert where either the source or destination type is i64 to
4147 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
4148 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
4149 /// vectors), since the legalizer won't know what to do with that.
4150 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
4151 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4153 SDValue Op = N->getOperand(0);
4155 // This function is only supposed to be called for i64 types, either as the
4156 // source or destination of the bit convert.
4157 EVT SrcVT = Op.getValueType();
4158 EVT DstVT = N->getValueType(0);
4159 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
4160 "ExpandBITCAST called for non-i64 type");
4162 // Turn i64->f64 into VMOVDRR.
4163 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
4164 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4165 DAG.getConstant(0, MVT::i32));
4166 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4167 DAG.getConstant(1, MVT::i32));
4168 return DAG.getNode(ISD::BITCAST, dl, DstVT,
4169 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
4172 // Turn f64->i64 into VMOVRRD.
4173 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
4175 if (TLI.isBigEndian() && SrcVT.isVector() &&
4176 SrcVT.getVectorNumElements() > 1)
4177 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4178 DAG.getVTList(MVT::i32, MVT::i32),
4179 DAG.getNode(ARMISD::VREV64, dl, SrcVT, Op));
4181 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4182 DAG.getVTList(MVT::i32, MVT::i32), Op);
4183 // Merge the pieces into a single i64 value.
4184 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
4190 /// getZeroVector - Returns a vector of specified type with all zero elements.
4191 /// Zero vectors are used to represent vector negation and in those cases
4192 /// will be implemented with the NEON VNEG instruction. However, VNEG does
4193 /// not support i64 elements, so sometimes the zero vectors will need to be
4194 /// explicitly constructed. Regardless, use a canonical VMOV to create the
4196 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
4197 assert(VT.isVector() && "Expected a vector type");
4198 // The canonical modified immediate encoding of a zero vector is....0!
4199 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
4200 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
4201 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
4202 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4205 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4206 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4207 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
4208 SelectionDAG &DAG) const {
4209 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4210 EVT VT = Op.getValueType();
4211 unsigned VTBits = VT.getSizeInBits();
4213 SDValue ShOpLo = Op.getOperand(0);
4214 SDValue ShOpHi = Op.getOperand(1);
4215 SDValue ShAmt = Op.getOperand(2);
4217 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4219 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4221 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4222 DAG.getConstant(VTBits, MVT::i32), ShAmt);
4223 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4224 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4225 DAG.getConstant(VTBits, MVT::i32));
4226 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4227 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4228 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4230 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4231 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
4233 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4234 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
4237 SDValue Ops[2] = { Lo, Hi };
4238 return DAG.getMergeValues(Ops, dl);
4241 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4242 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4243 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
4244 SelectionDAG &DAG) const {
4245 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4246 EVT VT = Op.getValueType();
4247 unsigned VTBits = VT.getSizeInBits();
4249 SDValue ShOpLo = Op.getOperand(0);
4250 SDValue ShOpHi = Op.getOperand(1);
4251 SDValue ShAmt = Op.getOperand(2);
4254 assert(Op.getOpcode() == ISD::SHL_PARTS);
4255 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4256 DAG.getConstant(VTBits, MVT::i32), ShAmt);
4257 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4258 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4259 DAG.getConstant(VTBits, MVT::i32));
4260 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4261 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4263 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4264 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4265 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
4267 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4268 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
4271 SDValue Ops[2] = { Lo, Hi };
4272 return DAG.getMergeValues(Ops, dl);
4275 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4276 SelectionDAG &DAG) const {
4277 // The rounding mode is in bits 23:22 of the FPSCR.
4278 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
4279 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
4280 // so that the shift + and get folded into a bitfield extract.
4282 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
4283 DAG.getConstant(Intrinsic::arm_get_fpscr,
4285 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
4286 DAG.getConstant(1U << 22, MVT::i32));
4287 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
4288 DAG.getConstant(22, MVT::i32));
4289 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
4290 DAG.getConstant(3, MVT::i32));
4293 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
4294 const ARMSubtarget *ST) {
4295 EVT VT = N->getValueType(0);
4298 if (!ST->hasV6T2Ops())
4301 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
4302 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
4305 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
4306 /// for each 16-bit element from operand, repeated. The basic idea is to
4307 /// leverage vcnt to get the 8-bit counts, gather and add the results.
4309 /// Trace for v4i16:
4310 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4311 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
4312 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
4313 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
4314 /// [b0 b1 b2 b3 b4 b5 b6 b7]
4315 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
4316 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
4317 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
4318 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
4319 EVT VT = N->getValueType(0);
4322 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4323 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
4324 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
4325 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
4326 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
4327 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
4330 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
4331 /// bit-count for each 16-bit element from the operand. We need slightly
4332 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
4333 /// 64/128-bit registers.
4335 /// Trace for v4i16:
4336 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4337 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4338 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4339 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4340 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4341 EVT VT = N->getValueType(0);
4344 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4345 if (VT.is64BitVector()) {
4346 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4347 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4348 DAG.getIntPtrConstant(0));
4350 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4351 BitCounts, DAG.getIntPtrConstant(0));
4352 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4356 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4357 /// bit-count for each 32-bit element from the operand. The idea here is
4358 /// to split the vector into 16-bit elements, leverage the 16-bit count
4359 /// routine, and then combine the results.
4361 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4362 /// input = [v0 v1 ] (vi: 32-bit elements)
4363 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4364 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4365 /// vrev: N0 = [k1 k0 k3 k2 ]
4367 /// N1 =+[k1 k0 k3 k2 ]
4369 /// N2 =+[k1 k3 k0 k2 ]
4371 /// Extended =+[k1 k3 k0 k2 ]
4373 /// Extracted=+[k1 k3 ]
4375 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4376 EVT VT = N->getValueType(0);
4379 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4381 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4382 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4383 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4384 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4385 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4387 if (VT.is64BitVector()) {
4388 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4389 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4390 DAG.getIntPtrConstant(0));
4392 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4393 DAG.getIntPtrConstant(0));
4394 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4398 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4399 const ARMSubtarget *ST) {
4400 EVT VT = N->getValueType(0);
4402 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4403 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4404 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4405 "Unexpected type for custom ctpop lowering");
4407 if (VT.getVectorElementType() == MVT::i32)
4408 return lowerCTPOP32BitElements(N, DAG);
4410 return lowerCTPOP16BitElements(N, DAG);
4413 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4414 const ARMSubtarget *ST) {
4415 EVT VT = N->getValueType(0);
4421 // Lower vector shifts on NEON to use VSHL.
4422 assert(ST->hasNEON() && "unexpected vector shift");
4424 // Left shifts translate directly to the vshiftu intrinsic.
4425 if (N->getOpcode() == ISD::SHL)
4426 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4427 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
4428 N->getOperand(0), N->getOperand(1));
4430 assert((N->getOpcode() == ISD::SRA ||
4431 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4433 // NEON uses the same intrinsics for both left and right shifts. For
4434 // right shifts, the shift amounts are negative, so negate the vector of
4436 EVT ShiftVT = N->getOperand(1).getValueType();
4437 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4438 getZeroVector(ShiftVT, DAG, dl),
4440 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4441 Intrinsic::arm_neon_vshifts :
4442 Intrinsic::arm_neon_vshiftu);
4443 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4444 DAG.getConstant(vshiftInt, MVT::i32),
4445 N->getOperand(0), NegatedCount);
4448 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4449 const ARMSubtarget *ST) {
4450 EVT VT = N->getValueType(0);
4453 // We can get here for a node like i32 = ISD::SHL i32, i64
4457 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4458 "Unknown shift to lower!");
4460 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4461 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4462 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4465 // If we are in thumb mode, we don't have RRX.
4466 if (ST->isThumb1Only()) return SDValue();
4468 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4469 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4470 DAG.getConstant(0, MVT::i32));
4471 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4472 DAG.getConstant(1, MVT::i32));
4474 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4475 // captures the result into a carry flag.
4476 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4477 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), Hi);
4479 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4480 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4482 // Merge the pieces into a single i64 value.
4483 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4486 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4487 SDValue TmpOp0, TmpOp1;
4488 bool Invert = false;
4492 SDValue Op0 = Op.getOperand(0);
4493 SDValue Op1 = Op.getOperand(1);
4494 SDValue CC = Op.getOperand(2);
4495 EVT CmpVT = Op0.getValueType().changeVectorElementTypeToInteger();
4496 EVT VT = Op.getValueType();
4497 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4500 if (Op1.getValueType().isFloatingPoint()) {
4501 switch (SetCCOpcode) {
4502 default: llvm_unreachable("Illegal FP comparison");
4504 case ISD::SETNE: Invert = true; // Fallthrough
4506 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4508 case ISD::SETLT: Swap = true; // Fallthrough
4510 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4512 case ISD::SETLE: Swap = true; // Fallthrough
4514 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4515 case ISD::SETUGE: Swap = true; // Fallthrough
4516 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4517 case ISD::SETUGT: Swap = true; // Fallthrough
4518 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4519 case ISD::SETUEQ: Invert = true; // Fallthrough
4521 // Expand this to (OLT | OGT).
4525 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4526 Op1 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp0, TmpOp1);
4528 case ISD::SETUO: Invert = true; // Fallthrough
4530 // Expand this to (OLT | OGE).
4534 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4535 Op1 = DAG.getNode(ARMISD::VCGE, dl, CmpVT, TmpOp0, TmpOp1);
4539 // Integer comparisons.
4540 switch (SetCCOpcode) {
4541 default: llvm_unreachable("Illegal integer comparison");
4542 case ISD::SETNE: Invert = true;
4543 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4544 case ISD::SETLT: Swap = true;
4545 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4546 case ISD::SETLE: Swap = true;
4547 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4548 case ISD::SETULT: Swap = true;
4549 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4550 case ISD::SETULE: Swap = true;
4551 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4554 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4555 if (Opc == ARMISD::VCEQ) {
4558 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4560 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4563 // Ignore bitconvert.
4564 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4565 AndOp = AndOp.getOperand(0);
4567 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4569 Op0 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(0));
4570 Op1 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(1));
4577 std::swap(Op0, Op1);
4579 // If one of the operands is a constant vector zero, attempt to fold the
4580 // comparison to a specialized compare-against-zero form.
4582 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4584 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4585 if (Opc == ARMISD::VCGE)
4586 Opc = ARMISD::VCLEZ;
4587 else if (Opc == ARMISD::VCGT)
4588 Opc = ARMISD::VCLTZ;
4593 if (SingleOp.getNode()) {
4596 Result = DAG.getNode(ARMISD::VCEQZ, dl, CmpVT, SingleOp); break;
4598 Result = DAG.getNode(ARMISD::VCGEZ, dl, CmpVT, SingleOp); break;
4600 Result = DAG.getNode(ARMISD::VCLEZ, dl, CmpVT, SingleOp); break;
4602 Result = DAG.getNode(ARMISD::VCGTZ, dl, CmpVT, SingleOp); break;
4604 Result = DAG.getNode(ARMISD::VCLTZ, dl, CmpVT, SingleOp); break;
4606 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4609 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4612 Result = DAG.getSExtOrTrunc(Result, dl, VT);
4615 Result = DAG.getNOT(dl, Result, VT);
4620 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4621 /// valid vector constant for a NEON instruction with a "modified immediate"
4622 /// operand (e.g., VMOV). If so, return the encoded value.
4623 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4624 unsigned SplatBitSize, SelectionDAG &DAG,
4625 EVT &VT, bool is128Bits, NEONModImmType type) {
4626 unsigned OpCmode, Imm;
4628 // SplatBitSize is set to the smallest size that splats the vector, so a
4629 // zero vector will always have SplatBitSize == 8. However, NEON modified
4630 // immediate instructions others than VMOV do not support the 8-bit encoding
4631 // of a zero vector, and the default encoding of zero is supposed to be the
4636 switch (SplatBitSize) {
4638 if (type != VMOVModImm)
4640 // Any 1-byte value is OK. Op=0, Cmode=1110.
4641 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4644 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4648 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4649 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4650 if ((SplatBits & ~0xff) == 0) {
4651 // Value = 0x00nn: Op=x, Cmode=100x.
4656 if ((SplatBits & ~0xff00) == 0) {
4657 // Value = 0xnn00: Op=x, Cmode=101x.
4659 Imm = SplatBits >> 8;
4665 // NEON's 32-bit VMOV supports splat values where:
4666 // * only one byte is nonzero, or
4667 // * the least significant byte is 0xff and the second byte is nonzero, or
4668 // * the least significant 2 bytes are 0xff and the third is nonzero.
4669 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4670 if ((SplatBits & ~0xff) == 0) {
4671 // Value = 0x000000nn: Op=x, Cmode=000x.
4676 if ((SplatBits & ~0xff00) == 0) {
4677 // Value = 0x0000nn00: Op=x, Cmode=001x.
4679 Imm = SplatBits >> 8;
4682 if ((SplatBits & ~0xff0000) == 0) {
4683 // Value = 0x00nn0000: Op=x, Cmode=010x.
4685 Imm = SplatBits >> 16;
4688 if ((SplatBits & ~0xff000000) == 0) {
4689 // Value = 0xnn000000: Op=x, Cmode=011x.
4691 Imm = SplatBits >> 24;
4695 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4696 if (type == OtherModImm) return SDValue();
4698 if ((SplatBits & ~0xffff) == 0 &&
4699 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4700 // Value = 0x0000nnff: Op=x, Cmode=1100.
4702 Imm = SplatBits >> 8;
4706 if ((SplatBits & ~0xffffff) == 0 &&
4707 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4708 // Value = 0x00nnffff: Op=x, Cmode=1101.
4710 Imm = SplatBits >> 16;
4714 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4715 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4716 // VMOV.I32. A (very) minor optimization would be to replicate the value
4717 // and fall through here to test for a valid 64-bit splat. But, then the
4718 // caller would also need to check and handle the change in size.
4722 if (type != VMOVModImm)
4724 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4725 uint64_t BitMask = 0xff;
4727 unsigned ImmMask = 1;
4729 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4730 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4733 } else if ((SplatBits & BitMask) != 0) {
4740 if (DAG.getTargetLoweringInfo().isBigEndian())
4741 // swap higher and lower 32 bit word
4742 Imm = ((Imm & 0xf) << 4) | ((Imm & 0xf0) >> 4);
4744 // Op=1, Cmode=1110.
4746 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4751 llvm_unreachable("unexpected size for isNEONModifiedImm");
4754 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4755 return DAG.getTargetConstant(EncodedVal, MVT::i32);
4758 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4759 const ARMSubtarget *ST) const {
4763 bool IsDouble = Op.getValueType() == MVT::f64;
4764 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4766 // Use the default (constant pool) lowering for double constants when we have
4768 if (IsDouble && Subtarget->isFPOnlySP())
4771 // Try splatting with a VMOV.f32...
4772 APFloat FPVal = CFP->getValueAPF();
4773 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4776 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4777 // We have code in place to select a valid ConstantFP already, no need to
4782 // It's a float and we are trying to use NEON operations where
4783 // possible. Lower it to a splat followed by an extract.
4785 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
4786 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4788 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4789 DAG.getConstant(0, MVT::i32));
4792 // The rest of our options are NEON only, make sure that's allowed before
4794 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4798 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4800 // It wouldn't really be worth bothering for doubles except for one very
4801 // important value, which does happen to match: 0.0. So make sure we don't do
4803 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4806 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4807 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4809 if (NewVal != SDValue()) {
4811 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4814 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4816 // It's a float: cast and extract a vector element.
4817 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4819 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4820 DAG.getConstant(0, MVT::i32));
4823 // Finally, try a VMVN.i32
4824 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4826 if (NewVal != SDValue()) {
4828 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4831 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4833 // It's a float: cast and extract a vector element.
4834 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4836 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4837 DAG.getConstant(0, MVT::i32));
4843 // check if an VEXT instruction can handle the shuffle mask when the
4844 // vector sources of the shuffle are the same.
4845 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4846 unsigned NumElts = VT.getVectorNumElements();
4848 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4854 // If this is a VEXT shuffle, the immediate value is the index of the first
4855 // element. The other shuffle indices must be the successive elements after
4857 unsigned ExpectedElt = Imm;
4858 for (unsigned i = 1; i < NumElts; ++i) {
4859 // Increment the expected index. If it wraps around, just follow it
4860 // back to index zero and keep going.
4862 if (ExpectedElt == NumElts)
4865 if (M[i] < 0) continue; // ignore UNDEF indices
4866 if (ExpectedElt != static_cast<unsigned>(M[i]))
4874 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4875 bool &ReverseVEXT, unsigned &Imm) {
4876 unsigned NumElts = VT.getVectorNumElements();
4877 ReverseVEXT = false;
4879 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4885 // If this is a VEXT shuffle, the immediate value is the index of the first
4886 // element. The other shuffle indices must be the successive elements after
4888 unsigned ExpectedElt = Imm;
4889 for (unsigned i = 1; i < NumElts; ++i) {
4890 // Increment the expected index. If it wraps around, it may still be
4891 // a VEXT but the source vectors must be swapped.
4893 if (ExpectedElt == NumElts * 2) {
4898 if (M[i] < 0) continue; // ignore UNDEF indices
4899 if (ExpectedElt != static_cast<unsigned>(M[i]))
4903 // Adjust the index value if the source operands will be swapped.
4910 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4911 /// instruction with the specified blocksize. (The order of the elements
4912 /// within each block of the vector is reversed.)
4913 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4914 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4915 "Only possible block sizes for VREV are: 16, 32, 64");
4917 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4921 unsigned NumElts = VT.getVectorNumElements();
4922 unsigned BlockElts = M[0] + 1;
4923 // If the first shuffle index is UNDEF, be optimistic.
4925 BlockElts = BlockSize / EltSz;
4927 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4930 for (unsigned i = 0; i < NumElts; ++i) {
4931 if (M[i] < 0) continue; // ignore UNDEF indices
4932 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4939 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4940 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4941 // range, then 0 is placed into the resulting vector. So pretty much any mask
4942 // of 8 elements can work here.
4943 return VT == MVT::v8i8 && M.size() == 8;
4946 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4947 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4951 unsigned NumElts = VT.getVectorNumElements();
4952 WhichResult = (M[0] == 0 ? 0 : 1);
4953 for (unsigned i = 0; i < NumElts; i += 2) {
4954 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4955 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4961 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4962 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4963 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4964 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4965 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4969 unsigned NumElts = VT.getVectorNumElements();
4970 WhichResult = (M[0] == 0 ? 0 : 1);
4971 for (unsigned i = 0; i < NumElts; i += 2) {
4972 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4973 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4979 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4980 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4984 unsigned NumElts = VT.getVectorNumElements();
4985 WhichResult = (M[0] == 0 ? 0 : 1);
4986 for (unsigned i = 0; i != NumElts; ++i) {
4987 if (M[i] < 0) continue; // ignore UNDEF indices
4988 if ((unsigned) M[i] != 2 * i + WhichResult)
4992 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4993 if (VT.is64BitVector() && EltSz == 32)
4999 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
5000 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5001 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
5002 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5003 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5007 unsigned Half = VT.getVectorNumElements() / 2;
5008 WhichResult = (M[0] == 0 ? 0 : 1);
5009 for (unsigned j = 0; j != 2; ++j) {
5010 unsigned Idx = WhichResult;
5011 for (unsigned i = 0; i != Half; ++i) {
5012 int MIdx = M[i + j * Half];
5013 if (MIdx >= 0 && (unsigned) MIdx != Idx)
5019 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5020 if (VT.is64BitVector() && EltSz == 32)
5026 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5027 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5031 unsigned NumElts = VT.getVectorNumElements();
5032 WhichResult = (M[0] == 0 ? 0 : 1);
5033 unsigned Idx = WhichResult * NumElts / 2;
5034 for (unsigned i = 0; i != NumElts; i += 2) {
5035 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
5036 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
5041 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5042 if (VT.is64BitVector() && EltSz == 32)
5048 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
5049 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5050 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
5051 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5052 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5056 unsigned NumElts = VT.getVectorNumElements();
5057 WhichResult = (M[0] == 0 ? 0 : 1);
5058 unsigned Idx = WhichResult * NumElts / 2;
5059 for (unsigned i = 0; i != NumElts; i += 2) {
5060 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
5061 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
5066 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5067 if (VT.is64BitVector() && EltSz == 32)
5073 /// \return true if this is a reverse operation on an vector.
5074 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
5075 unsigned NumElts = VT.getVectorNumElements();
5076 // Make sure the mask has the right size.
5077 if (NumElts != M.size())
5080 // Look for <15, ..., 3, -1, 1, 0>.
5081 for (unsigned i = 0; i != NumElts; ++i)
5082 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
5088 // If N is an integer constant that can be moved into a register in one
5089 // instruction, return an SDValue of such a constant (will become a MOV
5090 // instruction). Otherwise return null.
5091 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
5092 const ARMSubtarget *ST, SDLoc dl) {
5094 if (!isa<ConstantSDNode>(N))
5096 Val = cast<ConstantSDNode>(N)->getZExtValue();
5098 if (ST->isThumb1Only()) {
5099 if (Val <= 255 || ~Val <= 255)
5100 return DAG.getConstant(Val, MVT::i32);
5102 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
5103 return DAG.getConstant(Val, MVT::i32);
5108 // If this is a case we can't handle, return null and let the default
5109 // expansion code take care of it.
5110 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
5111 const ARMSubtarget *ST) const {
5112 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5114 EVT VT = Op.getValueType();
5116 APInt SplatBits, SplatUndef;
5117 unsigned SplatBitSize;
5119 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5120 if (SplatBitSize <= 64) {
5121 // Check if an immediate VMOV works.
5123 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
5124 SplatUndef.getZExtValue(), SplatBitSize,
5125 DAG, VmovVT, VT.is128BitVector(),
5127 if (Val.getNode()) {
5128 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
5129 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5132 // Try an immediate VMVN.
5133 uint64_t NegatedImm = (~SplatBits).getZExtValue();
5134 Val = isNEONModifiedImm(NegatedImm,
5135 SplatUndef.getZExtValue(), SplatBitSize,
5136 DAG, VmovVT, VT.is128BitVector(),
5138 if (Val.getNode()) {
5139 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
5140 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5143 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
5144 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
5145 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
5147 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
5148 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
5154 // Scan through the operands to see if only one value is used.
5156 // As an optimisation, even if more than one value is used it may be more
5157 // profitable to splat with one value then change some lanes.
5159 // Heuristically we decide to do this if the vector has a "dominant" value,
5160 // defined as splatted to more than half of the lanes.
5161 unsigned NumElts = VT.getVectorNumElements();
5162 bool isOnlyLowElement = true;
5163 bool usesOnlyOneValue = true;
5164 bool hasDominantValue = false;
5165 bool isConstant = true;
5167 // Map of the number of times a particular SDValue appears in the
5169 DenseMap<SDValue, unsigned> ValueCounts;
5171 for (unsigned i = 0; i < NumElts; ++i) {
5172 SDValue V = Op.getOperand(i);
5173 if (V.getOpcode() == ISD::UNDEF)
5176 isOnlyLowElement = false;
5177 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5180 ValueCounts.insert(std::make_pair(V, 0));
5181 unsigned &Count = ValueCounts[V];
5183 // Is this value dominant? (takes up more than half of the lanes)
5184 if (++Count > (NumElts / 2)) {
5185 hasDominantValue = true;
5189 if (ValueCounts.size() != 1)
5190 usesOnlyOneValue = false;
5191 if (!Value.getNode() && ValueCounts.size() > 0)
5192 Value = ValueCounts.begin()->first;
5194 if (ValueCounts.size() == 0)
5195 return DAG.getUNDEF(VT);
5197 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
5198 // Keep going if we are hitting this case.
5199 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
5200 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5202 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5204 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
5205 // i32 and try again.
5206 if (hasDominantValue && EltSize <= 32) {
5210 // If we are VDUPing a value that comes directly from a vector, that will
5211 // cause an unnecessary move to and from a GPR, where instead we could
5212 // just use VDUPLANE. We can only do this if the lane being extracted
5213 // is at a constant index, as the VDUP from lane instructions only have
5214 // constant-index forms.
5215 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5216 isa<ConstantSDNode>(Value->getOperand(1))) {
5217 // We need to create a new undef vector to use for the VDUPLANE if the
5218 // size of the vector from which we get the value is different than the
5219 // size of the vector that we need to create. We will insert the element
5220 // such that the register coalescer will remove unnecessary copies.
5221 if (VT != Value->getOperand(0).getValueType()) {
5222 ConstantSDNode *constIndex;
5223 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
5224 assert(constIndex && "The index is not a constant!");
5225 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
5226 VT.getVectorNumElements();
5227 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5228 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
5229 Value, DAG.getConstant(index, MVT::i32)),
5230 DAG.getConstant(index, MVT::i32));
5232 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5233 Value->getOperand(0), Value->getOperand(1));
5235 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
5237 if (!usesOnlyOneValue) {
5238 // The dominant value was splatted as 'N', but we now have to insert
5239 // all differing elements.
5240 for (unsigned I = 0; I < NumElts; ++I) {
5241 if (Op.getOperand(I) == Value)
5243 SmallVector<SDValue, 3> Ops;
5245 Ops.push_back(Op.getOperand(I));
5246 Ops.push_back(DAG.getConstant(I, MVT::i32));
5247 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ops);
5252 if (VT.getVectorElementType().isFloatingPoint()) {
5253 SmallVector<SDValue, 8> Ops;
5254 for (unsigned i = 0; i < NumElts; ++i)
5255 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
5257 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
5258 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5259 Val = LowerBUILD_VECTOR(Val, DAG, ST);
5261 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5263 if (usesOnlyOneValue) {
5264 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
5265 if (isConstant && Val.getNode())
5266 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
5270 // If all elements are constants and the case above didn't get hit, fall back
5271 // to the default expansion, which will generate a load from the constant
5276 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5278 SDValue shuffle = ReconstructShuffle(Op, DAG);
5279 if (shuffle != SDValue())
5283 // Vectors with 32- or 64-bit elements can be built by directly assigning
5284 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
5285 // will be legalized.
5286 if (EltSize >= 32) {
5287 // Do the expansion with floating-point types, since that is what the VFP
5288 // registers are defined to use, and since i64 is not legal.
5289 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5290 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5291 SmallVector<SDValue, 8> Ops;
5292 for (unsigned i = 0; i < NumElts; ++i)
5293 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
5294 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5295 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5298 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5299 // know the default expansion would otherwise fall back on something even
5300 // worse. For a vector with one or two non-undef values, that's
5301 // scalar_to_vector for the elements followed by a shuffle (provided the
5302 // shuffle is valid for the target) and materialization element by element
5303 // on the stack followed by a load for everything else.
5304 if (!isConstant && !usesOnlyOneValue) {
5305 SDValue Vec = DAG.getUNDEF(VT);
5306 for (unsigned i = 0 ; i < NumElts; ++i) {
5307 SDValue V = Op.getOperand(i);
5308 if (V.getOpcode() == ISD::UNDEF)
5310 SDValue LaneIdx = DAG.getConstant(i, MVT::i32);
5311 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5319 // Gather data to see if the operation can be modelled as a
5320 // shuffle in combination with VEXTs.
5321 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
5322 SelectionDAG &DAG) const {
5324 EVT VT = Op.getValueType();
5325 unsigned NumElts = VT.getVectorNumElements();
5327 SmallVector<SDValue, 2> SourceVecs;
5328 SmallVector<unsigned, 2> MinElts;
5329 SmallVector<unsigned, 2> MaxElts;
5331 for (unsigned i = 0; i < NumElts; ++i) {
5332 SDValue V = Op.getOperand(i);
5333 if (V.getOpcode() == ISD::UNDEF)
5335 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
5336 // A shuffle can only come from building a vector from various
5337 // elements of other vectors.
5339 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
5340 VT.getVectorElementType()) {
5341 // This code doesn't know how to handle shuffles where the vector
5342 // element types do not match (this happens because type legalization
5343 // promotes the return type of EXTRACT_VECTOR_ELT).
5344 // FIXME: It might be appropriate to extend this code to handle
5345 // mismatched types.
5349 // Record this extraction against the appropriate vector if possible...
5350 SDValue SourceVec = V.getOperand(0);
5351 // If the element number isn't a constant, we can't effectively
5352 // analyze what's going on.
5353 if (!isa<ConstantSDNode>(V.getOperand(1)))
5355 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5356 bool FoundSource = false;
5357 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
5358 if (SourceVecs[j] == SourceVec) {
5359 if (MinElts[j] > EltNo)
5361 if (MaxElts[j] < EltNo)
5368 // Or record a new source if not...
5370 SourceVecs.push_back(SourceVec);
5371 MinElts.push_back(EltNo);
5372 MaxElts.push_back(EltNo);
5376 // Currently only do something sane when at most two source vectors
5378 if (SourceVecs.size() > 2)
5381 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
5382 int VEXTOffsets[2] = {0, 0};
5384 // This loop extracts the usage patterns of the source vectors
5385 // and prepares appropriate SDValues for a shuffle if possible.
5386 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
5387 if (SourceVecs[i].getValueType() == VT) {
5388 // No VEXT necessary
5389 ShuffleSrcs[i] = SourceVecs[i];
5392 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
5393 // It probably isn't worth padding out a smaller vector just to
5394 // break it down again in a shuffle.
5398 // Since only 64-bit and 128-bit vectors are legal on ARM and
5399 // we've eliminated the other cases...
5400 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
5401 "unexpected vector sizes in ReconstructShuffle");
5403 if (MaxElts[i] - MinElts[i] >= NumElts) {
5404 // Span too large for a VEXT to cope
5408 if (MinElts[i] >= NumElts) {
5409 // The extraction can just take the second half
5410 VEXTOffsets[i] = NumElts;
5411 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5413 DAG.getIntPtrConstant(NumElts));
5414 } else if (MaxElts[i] < NumElts) {
5415 // The extraction can just take the first half
5417 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5419 DAG.getIntPtrConstant(0));
5421 // An actual VEXT is needed
5422 VEXTOffsets[i] = MinElts[i];
5423 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5425 DAG.getIntPtrConstant(0));
5426 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5428 DAG.getIntPtrConstant(NumElts));
5429 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
5430 DAG.getConstant(VEXTOffsets[i], MVT::i32));
5434 SmallVector<int, 8> Mask;
5436 for (unsigned i = 0; i < NumElts; ++i) {
5437 SDValue Entry = Op.getOperand(i);
5438 if (Entry.getOpcode() == ISD::UNDEF) {
5443 SDValue ExtractVec = Entry.getOperand(0);
5444 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
5445 .getOperand(1))->getSExtValue();
5446 if (ExtractVec == SourceVecs[0]) {
5447 Mask.push_back(ExtractElt - VEXTOffsets[0]);
5449 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
5453 // Final check before we try to produce nonsense...
5454 if (isShuffleMaskLegal(Mask, VT))
5455 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
5461 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5462 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5463 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5464 /// are assumed to be legal.
5466 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5468 if (VT.getVectorNumElements() == 4 &&
5469 (VT.is128BitVector() || VT.is64BitVector())) {
5470 unsigned PFIndexes[4];
5471 for (unsigned i = 0; i != 4; ++i) {
5475 PFIndexes[i] = M[i];
5478 // Compute the index in the perfect shuffle table.
5479 unsigned PFTableIndex =
5480 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5481 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5482 unsigned Cost = (PFEntry >> 30);
5489 unsigned Imm, WhichResult;
5491 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5492 return (EltSize >= 32 ||
5493 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5494 isVREVMask(M, VT, 64) ||
5495 isVREVMask(M, VT, 32) ||
5496 isVREVMask(M, VT, 16) ||
5497 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5498 isVTBLMask(M, VT) ||
5499 isVTRNMask(M, VT, WhichResult) ||
5500 isVUZPMask(M, VT, WhichResult) ||
5501 isVZIPMask(M, VT, WhichResult) ||
5502 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
5503 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
5504 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
5505 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5508 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5509 /// the specified operations to build the shuffle.
5510 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5511 SDValue RHS, SelectionDAG &DAG,
5513 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5514 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5515 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5518 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5527 OP_VUZPL, // VUZP, left result
5528 OP_VUZPR, // VUZP, right result
5529 OP_VZIPL, // VZIP, left result
5530 OP_VZIPR, // VZIP, right result
5531 OP_VTRNL, // VTRN, left result
5532 OP_VTRNR // VTRN, right result
5535 if (OpNum == OP_COPY) {
5536 if (LHSID == (1*9+2)*9+3) return LHS;
5537 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5541 SDValue OpLHS, OpRHS;
5542 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5543 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5544 EVT VT = OpLHS.getValueType();
5547 default: llvm_unreachable("Unknown shuffle opcode!");
5549 // VREV divides the vector in half and swaps within the half.
5550 if (VT.getVectorElementType() == MVT::i32 ||
5551 VT.getVectorElementType() == MVT::f32)
5552 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5553 // vrev <4 x i16> -> VREV32
5554 if (VT.getVectorElementType() == MVT::i16)
5555 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5556 // vrev <4 x i8> -> VREV16
5557 assert(VT.getVectorElementType() == MVT::i8);
5558 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5563 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5564 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
5568 return DAG.getNode(ARMISD::VEXT, dl, VT,
5570 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
5573 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5574 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5577 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5578 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5581 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5582 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5586 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5587 ArrayRef<int> ShuffleMask,
5588 SelectionDAG &DAG) {
5589 // Check to see if we can use the VTBL instruction.
5590 SDValue V1 = Op.getOperand(0);
5591 SDValue V2 = Op.getOperand(1);
5594 SmallVector<SDValue, 8> VTBLMask;
5595 for (ArrayRef<int>::iterator
5596 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5597 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
5599 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5600 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5601 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5603 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5604 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5607 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5608 SelectionDAG &DAG) {
5610 SDValue OpLHS = Op.getOperand(0);
5611 EVT VT = OpLHS.getValueType();
5613 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5614 "Expect an v8i16/v16i8 type");
5615 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5616 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5617 // extract the first 8 bytes into the top double word and the last 8 bytes
5618 // into the bottom double word. The v8i16 case is similar.
5619 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5620 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5621 DAG.getConstant(ExtractNum, MVT::i32));
5624 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5625 SDValue V1 = Op.getOperand(0);
5626 SDValue V2 = Op.getOperand(1);
5628 EVT VT = Op.getValueType();
5629 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5631 // Convert shuffles that are directly supported on NEON to target-specific
5632 // DAG nodes, instead of keeping them as shuffles and matching them again
5633 // during code selection. This is more efficient and avoids the possibility
5634 // of inconsistencies between legalization and selection.
5635 // FIXME: floating-point vectors should be canonicalized to integer vectors
5636 // of the same time so that they get CSEd properly.
5637 ArrayRef<int> ShuffleMask = SVN->getMask();
5639 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5640 if (EltSize <= 32) {
5641 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5642 int Lane = SVN->getSplatIndex();
5643 // If this is undef splat, generate it via "just" vdup, if possible.
5644 if (Lane == -1) Lane = 0;
5646 // Test if V1 is a SCALAR_TO_VECTOR.
5647 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5648 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5650 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5651 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5653 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5654 !isa<ConstantSDNode>(V1.getOperand(0))) {
5655 bool IsScalarToVector = true;
5656 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5657 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5658 IsScalarToVector = false;
5661 if (IsScalarToVector)
5662 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5664 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5665 DAG.getConstant(Lane, MVT::i32));
5670 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5673 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5674 DAG.getConstant(Imm, MVT::i32));
5677 if (isVREVMask(ShuffleMask, VT, 64))
5678 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5679 if (isVREVMask(ShuffleMask, VT, 32))
5680 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5681 if (isVREVMask(ShuffleMask, VT, 16))
5682 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5684 if (V2->getOpcode() == ISD::UNDEF &&
5685 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5686 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5687 DAG.getConstant(Imm, MVT::i32));
5690 // Check for Neon shuffles that modify both input vectors in place.
5691 // If both results are used, i.e., if there are two shuffles with the same
5692 // source operands and with masks corresponding to both results of one of
5693 // these operations, DAG memoization will ensure that a single node is
5694 // used for both shuffles.
5695 unsigned WhichResult;
5696 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5697 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5698 V1, V2).getValue(WhichResult);
5699 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5700 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5701 V1, V2).getValue(WhichResult);
5702 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5703 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5704 V1, V2).getValue(WhichResult);
5706 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5707 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5708 V1, V1).getValue(WhichResult);
5709 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5710 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5711 V1, V1).getValue(WhichResult);
5712 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5713 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5714 V1, V1).getValue(WhichResult);
5717 // If the shuffle is not directly supported and it has 4 elements, use
5718 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5719 unsigned NumElts = VT.getVectorNumElements();
5721 unsigned PFIndexes[4];
5722 for (unsigned i = 0; i != 4; ++i) {
5723 if (ShuffleMask[i] < 0)
5726 PFIndexes[i] = ShuffleMask[i];
5729 // Compute the index in the perfect shuffle table.
5730 unsigned PFTableIndex =
5731 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5732 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5733 unsigned Cost = (PFEntry >> 30);
5736 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5739 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5740 if (EltSize >= 32) {
5741 // Do the expansion with floating-point types, since that is what the VFP
5742 // registers are defined to use, and since i64 is not legal.
5743 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5744 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5745 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5746 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5747 SmallVector<SDValue, 8> Ops;
5748 for (unsigned i = 0; i < NumElts; ++i) {
5749 if (ShuffleMask[i] < 0)
5750 Ops.push_back(DAG.getUNDEF(EltVT));
5752 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5753 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5754 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5757 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5758 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5761 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5762 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5764 if (VT == MVT::v8i8) {
5765 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5766 if (NewOp.getNode())
5773 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5774 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5775 SDValue Lane = Op.getOperand(2);
5776 if (!isa<ConstantSDNode>(Lane))
5782 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5783 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5784 SDValue Lane = Op.getOperand(1);
5785 if (!isa<ConstantSDNode>(Lane))
5788 SDValue Vec = Op.getOperand(0);
5789 if (Op.getValueType() == MVT::i32 &&
5790 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5792 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5798 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5799 // The only time a CONCAT_VECTORS operation can have legal types is when
5800 // two 64-bit vectors are concatenated to a 128-bit vector.
5801 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5802 "unexpected CONCAT_VECTORS");
5804 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5805 SDValue Op0 = Op.getOperand(0);
5806 SDValue Op1 = Op.getOperand(1);
5807 if (Op0.getOpcode() != ISD::UNDEF)
5808 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5809 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5810 DAG.getIntPtrConstant(0));
5811 if (Op1.getOpcode() != ISD::UNDEF)
5812 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5813 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5814 DAG.getIntPtrConstant(1));
5815 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5818 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5819 /// element has been zero/sign-extended, depending on the isSigned parameter,
5820 /// from an integer type half its size.
5821 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5823 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5824 EVT VT = N->getValueType(0);
5825 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5826 SDNode *BVN = N->getOperand(0).getNode();
5827 if (BVN->getValueType(0) != MVT::v4i32 ||
5828 BVN->getOpcode() != ISD::BUILD_VECTOR)
5830 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5831 unsigned HiElt = 1 - LoElt;
5832 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5833 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5834 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5835 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5836 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5839 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5840 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5843 if (Hi0->isNullValue() && Hi1->isNullValue())
5849 if (N->getOpcode() != ISD::BUILD_VECTOR)
5852 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5853 SDNode *Elt = N->getOperand(i).getNode();
5854 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5855 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5856 unsigned HalfSize = EltSize / 2;
5858 if (!isIntN(HalfSize, C->getSExtValue()))
5861 if (!isUIntN(HalfSize, C->getZExtValue()))
5872 /// isSignExtended - Check if a node is a vector value that is sign-extended
5873 /// or a constant BUILD_VECTOR with sign-extended elements.
5874 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5875 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5877 if (isExtendedBUILD_VECTOR(N, DAG, true))
5882 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5883 /// or a constant BUILD_VECTOR with zero-extended elements.
5884 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5885 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5887 if (isExtendedBUILD_VECTOR(N, DAG, false))
5892 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
5893 if (OrigVT.getSizeInBits() >= 64)
5896 assert(OrigVT.isSimple() && "Expecting a simple value type");
5898 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
5899 switch (OrigSimpleTy) {
5900 default: llvm_unreachable("Unexpected Vector Type");
5909 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5910 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5911 /// We insert the required extension here to get the vector to fill a D register.
5912 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5915 unsigned ExtOpcode) {
5916 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5917 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5918 // 64-bits we need to insert a new extension so that it will be 64-bits.
5919 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5920 if (OrigTy.getSizeInBits() >= 64)
5923 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5924 EVT NewVT = getExtensionTo64Bits(OrigTy);
5926 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
5929 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5930 /// does not do any sign/zero extension. If the original vector is less
5931 /// than 64 bits, an appropriate extension will be added after the load to
5932 /// reach a total size of 64 bits. We have to add the extension separately
5933 /// because ARM does not have a sign/zero extending load for vectors.
5934 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5935 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
5937 // The load already has the right type.
5938 if (ExtendedTy == LD->getMemoryVT())
5939 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
5940 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5941 LD->isNonTemporal(), LD->isInvariant(),
5942 LD->getAlignment());
5944 // We need to create a zextload/sextload. We cannot just create a load
5945 // followed by a zext/zext node because LowerMUL is also run during normal
5946 // operation legalization where we can't create illegal types.
5947 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
5948 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
5949 LD->getMemoryVT(), LD->isVolatile(), LD->isInvariant(),
5950 LD->isNonTemporal(), LD->getAlignment());
5953 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5954 /// extending load, or BUILD_VECTOR with extended elements, return the
5955 /// unextended value. The unextended vector should be 64 bits so that it can
5956 /// be used as an operand to a VMULL instruction. If the original vector size
5957 /// before extension is less than 64 bits we add a an extension to resize
5958 /// the vector to 64 bits.
5959 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5960 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5961 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5962 N->getOperand(0)->getValueType(0),
5966 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5967 return SkipLoadExtensionForVMULL(LD, DAG);
5969 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5970 // have been legalized as a BITCAST from v4i32.
5971 if (N->getOpcode() == ISD::BITCAST) {
5972 SDNode *BVN = N->getOperand(0).getNode();
5973 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5974 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5975 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5976 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
5977 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5979 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5980 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5981 EVT VT = N->getValueType(0);
5982 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5983 unsigned NumElts = VT.getVectorNumElements();
5984 MVT TruncVT = MVT::getIntegerVT(EltSize);
5985 SmallVector<SDValue, 8> Ops;
5986 for (unsigned i = 0; i != NumElts; ++i) {
5987 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5988 const APInt &CInt = C->getAPIntValue();
5989 // Element types smaller than 32 bits are not legal, so use i32 elements.
5990 // The values are implicitly truncated so sext vs. zext doesn't matter.
5991 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
5993 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
5994 MVT::getVectorVT(TruncVT, NumElts), Ops);
5997 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5998 unsigned Opcode = N->getOpcode();
5999 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
6000 SDNode *N0 = N->getOperand(0).getNode();
6001 SDNode *N1 = N->getOperand(1).getNode();
6002 return N0->hasOneUse() && N1->hasOneUse() &&
6003 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
6008 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
6009 unsigned Opcode = N->getOpcode();
6010 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
6011 SDNode *N0 = N->getOperand(0).getNode();
6012 SDNode *N1 = N->getOperand(1).getNode();
6013 return N0->hasOneUse() && N1->hasOneUse() &&
6014 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
6019 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
6020 // Multiplications are only custom-lowered for 128-bit vectors so that
6021 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
6022 EVT VT = Op.getValueType();
6023 assert(VT.is128BitVector() && VT.isInteger() &&
6024 "unexpected type for custom-lowering ISD::MUL");
6025 SDNode *N0 = Op.getOperand(0).getNode();
6026 SDNode *N1 = Op.getOperand(1).getNode();
6027 unsigned NewOpc = 0;
6029 bool isN0SExt = isSignExtended(N0, DAG);
6030 bool isN1SExt = isSignExtended(N1, DAG);
6031 if (isN0SExt && isN1SExt)
6032 NewOpc = ARMISD::VMULLs;
6034 bool isN0ZExt = isZeroExtended(N0, DAG);
6035 bool isN1ZExt = isZeroExtended(N1, DAG);
6036 if (isN0ZExt && isN1ZExt)
6037 NewOpc = ARMISD::VMULLu;
6038 else if (isN1SExt || isN1ZExt) {
6039 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
6040 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
6041 if (isN1SExt && isAddSubSExt(N0, DAG)) {
6042 NewOpc = ARMISD::VMULLs;
6044 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
6045 NewOpc = ARMISD::VMULLu;
6047 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
6049 NewOpc = ARMISD::VMULLu;
6055 if (VT == MVT::v2i64)
6056 // Fall through to expand this. It is not legal.
6059 // Other vector multiplications are legal.
6064 // Legalize to a VMULL instruction.
6067 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
6069 Op0 = SkipExtensionForVMULL(N0, DAG);
6070 assert(Op0.getValueType().is64BitVector() &&
6071 Op1.getValueType().is64BitVector() &&
6072 "unexpected types for extended operands to VMULL");
6073 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
6076 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
6077 // isel lowering to take advantage of no-stall back to back vmul + vmla.
6084 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
6085 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
6086 EVT Op1VT = Op1.getValueType();
6087 return DAG.getNode(N0->getOpcode(), DL, VT,
6088 DAG.getNode(NewOpc, DL, VT,
6089 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
6090 DAG.getNode(NewOpc, DL, VT,
6091 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
6095 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
6097 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
6098 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
6099 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
6100 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
6101 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
6102 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
6103 // Get reciprocal estimate.
6104 // float4 recip = vrecpeq_f32(yf);
6105 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6106 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
6107 // Because char has a smaller range than uchar, we can actually get away
6108 // without any newton steps. This requires that we use a weird bias
6109 // of 0xb000, however (again, this has been exhaustively tested).
6110 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
6111 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
6112 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
6113 Y = DAG.getConstant(0xb000, MVT::i32);
6114 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
6115 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
6116 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
6117 // Convert back to short.
6118 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
6119 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
6124 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
6126 // Convert to float.
6127 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
6128 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
6129 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
6130 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
6131 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6132 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6134 // Use reciprocal estimate and one refinement step.
6135 // float4 recip = vrecpeq_f32(yf);
6136 // recip *= vrecpsq_f32(yf, recip);
6137 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6138 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
6139 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6140 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
6142 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6143 // Because short has a smaller range than ushort, we can actually get away
6144 // with only a single newton step. This requires that we use a weird bias
6145 // of 89, however (again, this has been exhaustively tested).
6146 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
6147 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6148 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6149 N1 = DAG.getConstant(0x89, MVT::i32);
6150 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6151 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6152 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6153 // Convert back to integer and return.
6154 // return vmovn_s32(vcvt_s32_f32(result));
6155 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6156 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6160 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
6161 EVT VT = Op.getValueType();
6162 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6163 "unexpected type for custom-lowering ISD::SDIV");
6166 SDValue N0 = Op.getOperand(0);
6167 SDValue N1 = Op.getOperand(1);
6170 if (VT == MVT::v8i8) {
6171 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
6172 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
6174 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6175 DAG.getIntPtrConstant(4));
6176 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6177 DAG.getIntPtrConstant(4));
6178 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6179 DAG.getIntPtrConstant(0));
6180 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6181 DAG.getIntPtrConstant(0));
6183 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
6184 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
6186 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6187 N0 = LowerCONCAT_VECTORS(N0, DAG);
6189 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
6192 return LowerSDIV_v4i16(N0, N1, dl, DAG);
6195 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
6196 EVT VT = Op.getValueType();
6197 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6198 "unexpected type for custom-lowering ISD::UDIV");
6201 SDValue N0 = Op.getOperand(0);
6202 SDValue N1 = Op.getOperand(1);
6205 if (VT == MVT::v8i8) {
6206 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
6207 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
6209 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6210 DAG.getIntPtrConstant(4));
6211 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6212 DAG.getIntPtrConstant(4));
6213 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6214 DAG.getIntPtrConstant(0));
6215 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6216 DAG.getIntPtrConstant(0));
6218 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
6219 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
6221 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6222 N0 = LowerCONCAT_VECTORS(N0, DAG);
6224 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
6225 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
6230 // v4i16 sdiv ... Convert to float.
6231 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
6232 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
6233 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
6234 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
6235 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6236 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6238 // Use reciprocal estimate and two refinement steps.
6239 // float4 recip = vrecpeq_f32(yf);
6240 // recip *= vrecpsq_f32(yf, recip);
6241 // recip *= vrecpsq_f32(yf, recip);
6242 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6243 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
6244 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6245 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
6247 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6248 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6249 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
6251 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6252 // Simply multiplying by the reciprocal estimate can leave us a few ulps
6253 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
6254 // and that it will never cause us to return an answer too large).
6255 // float4 result = as_float4(as_int4(xf*recip) + 2);
6256 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6257 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6258 N1 = DAG.getConstant(2, MVT::i32);
6259 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6260 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6261 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6262 // Convert back to integer and return.
6263 // return vmovn_u32(vcvt_s32_f32(result));
6264 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6265 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6269 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
6270 EVT VT = Op.getNode()->getValueType(0);
6271 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
6274 bool ExtraOp = false;
6275 switch (Op.getOpcode()) {
6276 default: llvm_unreachable("Invalid code");
6277 case ISD::ADDC: Opc = ARMISD::ADDC; break;
6278 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
6279 case ISD::SUBC: Opc = ARMISD::SUBC; break;
6280 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
6284 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6286 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6287 Op.getOperand(1), Op.getOperand(2));
6290 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
6291 assert(Subtarget->isTargetDarwin());
6293 // For iOS, we want to call an alternative entry point: __sincos_stret,
6294 // return values are passed via sret.
6296 SDValue Arg = Op.getOperand(0);
6297 EVT ArgVT = Arg.getValueType();
6298 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
6300 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
6301 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6303 // Pair of floats / doubles used to pass the result.
6304 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
6306 // Create stack object for sret.
6307 const uint64_t ByteSize = TLI.getDataLayout()->getTypeAllocSize(RetTy);
6308 const unsigned StackAlign = TLI.getDataLayout()->getPrefTypeAlignment(RetTy);
6309 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
6310 SDValue SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy());
6316 Entry.Ty = RetTy->getPointerTo();
6317 Entry.isSExt = false;
6318 Entry.isZExt = false;
6319 Entry.isSRet = true;
6320 Args.push_back(Entry);
6324 Entry.isSExt = false;
6325 Entry.isZExt = false;
6326 Args.push_back(Entry);
6328 const char *LibcallName = (ArgVT == MVT::f64)
6329 ? "__sincos_stret" : "__sincosf_stret";
6330 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
6332 TargetLowering::CallLoweringInfo CLI(DAG);
6333 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
6334 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()), Callee,
6336 .setDiscardResult();
6338 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
6340 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
6341 MachinePointerInfo(), false, false, false, 0);
6343 // Address of cos field.
6344 SDValue Add = DAG.getNode(ISD::ADD, dl, getPointerTy(), SRet,
6345 DAG.getIntPtrConstant(ArgVT.getStoreSize()));
6346 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
6347 MachinePointerInfo(), false, false, false, 0);
6349 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
6350 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
6351 LoadSin.getValue(0), LoadCos.getValue(0));
6354 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
6355 // Monotonic load/store is legal for all targets
6356 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
6359 // Acquire/Release load/store is not legal for targets without a
6360 // dmb or equivalent available.
6364 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6365 SmallVectorImpl<SDValue> &Results,
6367 const ARMSubtarget *Subtarget) {
6369 SDValue Cycles32, OutChain;
6371 if (Subtarget->hasPerfMon()) {
6372 // Under Power Management extensions, the cycle-count is:
6373 // mrc p15, #0, <Rt>, c9, c13, #0
6374 SDValue Ops[] = { N->getOperand(0), // Chain
6375 DAG.getConstant(Intrinsic::arm_mrc, MVT::i32),
6376 DAG.getConstant(15, MVT::i32),
6377 DAG.getConstant(0, MVT::i32),
6378 DAG.getConstant(9, MVT::i32),
6379 DAG.getConstant(13, MVT::i32),
6380 DAG.getConstant(0, MVT::i32)
6383 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6384 DAG.getVTList(MVT::i32, MVT::Other), Ops);
6385 OutChain = Cycles32.getValue(1);
6387 // Intrinsic is defined to return 0 on unsupported platforms. Technically
6388 // there are older ARM CPUs that have implementation-specific ways of
6389 // obtaining this information (FIXME!).
6390 Cycles32 = DAG.getConstant(0, MVT::i32);
6391 OutChain = DAG.getEntryNode();
6395 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
6396 Cycles32, DAG.getConstant(0, MVT::i32));
6397 Results.push_back(Cycles64);
6398 Results.push_back(OutChain);
6401 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6402 switch (Op.getOpcode()) {
6403 default: llvm_unreachable("Don't know how to custom lower this!");
6404 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6405 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6406 case ISD::GlobalAddress:
6407 switch (Subtarget->getTargetTriple().getObjectFormat()) {
6408 default: llvm_unreachable("unknown object format");
6410 return LowerGlobalAddressWindows(Op, DAG);
6412 return LowerGlobalAddressELF(Op, DAG);
6414 return LowerGlobalAddressDarwin(Op, DAG);
6416 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6417 case ISD::SELECT: return LowerSELECT(Op, DAG);
6418 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6419 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6420 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6421 case ISD::VASTART: return LowerVASTART(Op, DAG);
6422 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6423 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6424 case ISD::SINT_TO_FP:
6425 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6426 case ISD::FP_TO_SINT:
6427 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6428 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6429 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6430 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6431 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6432 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6433 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6434 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6436 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6439 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6440 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6441 case ISD::SRL_PARTS:
6442 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6443 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6444 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6445 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6446 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6447 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6448 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6449 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6450 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6451 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6452 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6453 case ISD::MUL: return LowerMUL(Op, DAG);
6454 case ISD::SDIV: return LowerSDIV(Op, DAG);
6455 case ISD::UDIV: return LowerUDIV(Op, DAG);
6459 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6464 return LowerXALUO(Op, DAG);
6465 case ISD::ATOMIC_LOAD:
6466 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6467 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6469 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6470 case ISD::DYNAMIC_STACKALLOC:
6471 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
6472 return LowerDYNAMIC_STACKALLOC(Op, DAG);
6473 llvm_unreachable("Don't know how to custom lower this!");
6474 case ISD::FP_ROUND: return LowerFP_ROUND(Op, DAG);
6475 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
6479 /// ReplaceNodeResults - Replace the results of node with an illegal result
6480 /// type with new values built out of custom code.
6481 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6482 SmallVectorImpl<SDValue>&Results,
6483 SelectionDAG &DAG) const {
6485 switch (N->getOpcode()) {
6487 llvm_unreachable("Don't know how to custom expand this!");
6489 Res = ExpandBITCAST(N, DAG);
6493 Res = Expand64BitShift(N, DAG, Subtarget);
6495 case ISD::READCYCLECOUNTER:
6496 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6500 Results.push_back(Res);
6503 //===----------------------------------------------------------------------===//
6504 // ARM Scheduler Hooks
6505 //===----------------------------------------------------------------------===//
6507 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6508 /// registers the function context.
6509 void ARMTargetLowering::
6510 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6511 MachineBasicBlock *DispatchBB, int FI) const {
6512 const TargetInstrInfo *TII =
6513 getTargetMachine().getSubtargetImpl()->getInstrInfo();
6514 DebugLoc dl = MI->getDebugLoc();
6515 MachineFunction *MF = MBB->getParent();
6516 MachineRegisterInfo *MRI = &MF->getRegInfo();
6517 MachineConstantPool *MCP = MF->getConstantPool();
6518 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6519 const Function *F = MF->getFunction();
6521 bool isThumb = Subtarget->isThumb();
6522 bool isThumb2 = Subtarget->isThumb2();
6524 unsigned PCLabelId = AFI->createPICLabelUId();
6525 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6526 ARMConstantPoolValue *CPV =
6527 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6528 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6530 const TargetRegisterClass *TRC = isThumb ? &ARM::tGPRRegClass
6531 : &ARM::GPRRegClass;
6533 // Grab constant pool and fixed stack memory operands.
6534 MachineMemOperand *CPMMO =
6535 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6536 MachineMemOperand::MOLoad, 4, 4);
6538 MachineMemOperand *FIMMOSt =
6539 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6540 MachineMemOperand::MOStore, 4, 4);
6542 // Load the address of the dispatch MBB into the jump buffer.
6544 // Incoming value: jbuf
6545 // ldr.n r5, LCPI1_1
6548 // str r5, [$jbuf, #+4] ; &jbuf[1]
6549 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6550 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6551 .addConstantPoolIndex(CPI)
6552 .addMemOperand(CPMMO));
6553 // Set the low bit because of thumb mode.
6554 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6556 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6557 .addReg(NewVReg1, RegState::Kill)
6559 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6560 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6561 .addReg(NewVReg2, RegState::Kill)
6563 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6564 .addReg(NewVReg3, RegState::Kill)
6566 .addImm(36) // &jbuf[1] :: pc
6567 .addMemOperand(FIMMOSt));
6568 } else if (isThumb) {
6569 // Incoming value: jbuf
6570 // ldr.n r1, LCPI1_4
6574 // add r2, $jbuf, #+4 ; &jbuf[1]
6576 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6577 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6578 .addConstantPoolIndex(CPI)
6579 .addMemOperand(CPMMO));
6580 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6581 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6582 .addReg(NewVReg1, RegState::Kill)
6584 // Set the low bit because of thumb mode.
6585 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6586 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6587 .addReg(ARM::CPSR, RegState::Define)
6589 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6590 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6591 .addReg(ARM::CPSR, RegState::Define)
6592 .addReg(NewVReg2, RegState::Kill)
6593 .addReg(NewVReg3, RegState::Kill));
6594 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6595 BuildMI(*MBB, MI, dl, TII->get(ARM::tADDframe), NewVReg5)
6597 .addImm(36); // &jbuf[1] :: pc
6598 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6599 .addReg(NewVReg4, RegState::Kill)
6600 .addReg(NewVReg5, RegState::Kill)
6602 .addMemOperand(FIMMOSt));
6604 // Incoming value: jbuf
6607 // str r1, [$jbuf, #+4] ; &jbuf[1]
6608 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6609 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6610 .addConstantPoolIndex(CPI)
6612 .addMemOperand(CPMMO));
6613 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6614 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6615 .addReg(NewVReg1, RegState::Kill)
6616 .addImm(PCLabelId));
6617 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6618 .addReg(NewVReg2, RegState::Kill)
6620 .addImm(36) // &jbuf[1] :: pc
6621 .addMemOperand(FIMMOSt));
6625 MachineBasicBlock *ARMTargetLowering::
6626 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
6627 const TargetInstrInfo *TII =
6628 getTargetMachine().getSubtargetImpl()->getInstrInfo();
6629 DebugLoc dl = MI->getDebugLoc();
6630 MachineFunction *MF = MBB->getParent();
6631 MachineRegisterInfo *MRI = &MF->getRegInfo();
6632 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6633 MachineFrameInfo *MFI = MF->getFrameInfo();
6634 int FI = MFI->getFunctionContextIndex();
6636 const TargetRegisterClass *TRC = Subtarget->isThumb() ? &ARM::tGPRRegClass
6637 : &ARM::GPRnopcRegClass;
6639 // Get a mapping of the call site numbers to all of the landing pads they're
6641 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6642 unsigned MaxCSNum = 0;
6643 MachineModuleInfo &MMI = MF->getMMI();
6644 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6646 if (!BB->isLandingPad()) continue;
6648 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6650 for (MachineBasicBlock::iterator
6651 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6652 if (!II->isEHLabel()) continue;
6654 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6655 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6657 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6658 for (SmallVectorImpl<unsigned>::iterator
6659 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6660 CSI != CSE; ++CSI) {
6661 CallSiteNumToLPad[*CSI].push_back(BB);
6662 MaxCSNum = std::max(MaxCSNum, *CSI);
6668 // Get an ordered list of the machine basic blocks for the jump table.
6669 std::vector<MachineBasicBlock*> LPadList;
6670 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6671 LPadList.reserve(CallSiteNumToLPad.size());
6672 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6673 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6674 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6675 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6676 LPadList.push_back(*II);
6677 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6681 assert(!LPadList.empty() &&
6682 "No landing pad destinations for the dispatch jump table!");
6684 // Create the jump table and associated information.
6685 MachineJumpTableInfo *JTI =
6686 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6687 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6688 unsigned UId = AFI->createJumpTableUId();
6689 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6691 // Create the MBBs for the dispatch code.
6693 // Shove the dispatch's address into the return slot in the function context.
6694 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6695 DispatchBB->setIsLandingPad();
6697 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6698 unsigned trap_opcode;
6699 if (Subtarget->isThumb())
6700 trap_opcode = ARM::tTRAP;
6702 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6704 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6705 DispatchBB->addSuccessor(TrapBB);
6707 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6708 DispatchBB->addSuccessor(DispContBB);
6711 MF->insert(MF->end(), DispatchBB);
6712 MF->insert(MF->end(), DispContBB);
6713 MF->insert(MF->end(), TrapBB);
6715 // Insert code into the entry block that creates and registers the function
6717 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6719 MachineMemOperand *FIMMOLd =
6720 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6721 MachineMemOperand::MOLoad |
6722 MachineMemOperand::MOVolatile, 4, 4);
6724 MachineInstrBuilder MIB;
6725 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6727 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6728 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6730 // Add a register mask with no preserved registers. This results in all
6731 // registers being marked as clobbered.
6732 MIB.addRegMask(RI.getNoPreservedMask());
6734 unsigned NumLPads = LPadList.size();
6735 if (Subtarget->isThumb2()) {
6736 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6737 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6740 .addMemOperand(FIMMOLd));
6742 if (NumLPads < 256) {
6743 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6745 .addImm(LPadList.size()));
6747 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6748 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6749 .addImm(NumLPads & 0xFFFF));
6751 unsigned VReg2 = VReg1;
6752 if ((NumLPads & 0xFFFF0000) != 0) {
6753 VReg2 = MRI->createVirtualRegister(TRC);
6754 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6756 .addImm(NumLPads >> 16));
6759 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6764 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6769 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6770 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6771 .addJumpTableIndex(MJTI)
6774 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6777 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6778 .addReg(NewVReg3, RegState::Kill)
6780 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6782 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6783 .addReg(NewVReg4, RegState::Kill)
6785 .addJumpTableIndex(MJTI)
6787 } else if (Subtarget->isThumb()) {
6788 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6789 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6792 .addMemOperand(FIMMOLd));
6794 if (NumLPads < 256) {
6795 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6799 MachineConstantPool *ConstantPool = MF->getConstantPool();
6800 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6801 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6803 // MachineConstantPool wants an explicit alignment.
6804 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6806 Align = getDataLayout()->getTypeAllocSize(C->getType());
6807 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6809 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6810 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6811 .addReg(VReg1, RegState::Define)
6812 .addConstantPoolIndex(Idx));
6813 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6818 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6823 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6824 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6825 .addReg(ARM::CPSR, RegState::Define)
6829 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6830 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6831 .addJumpTableIndex(MJTI)
6834 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6835 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6836 .addReg(ARM::CPSR, RegState::Define)
6837 .addReg(NewVReg2, RegState::Kill)
6840 MachineMemOperand *JTMMOLd =
6841 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6842 MachineMemOperand::MOLoad, 4, 4);
6844 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6845 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6846 .addReg(NewVReg4, RegState::Kill)
6848 .addMemOperand(JTMMOLd));
6850 unsigned NewVReg6 = NewVReg5;
6851 if (RelocM == Reloc::PIC_) {
6852 NewVReg6 = MRI->createVirtualRegister(TRC);
6853 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6854 .addReg(ARM::CPSR, RegState::Define)
6855 .addReg(NewVReg5, RegState::Kill)
6859 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6860 .addReg(NewVReg6, RegState::Kill)
6861 .addJumpTableIndex(MJTI)
6864 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6865 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6868 .addMemOperand(FIMMOLd));
6870 if (NumLPads < 256) {
6871 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6874 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6875 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6876 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6877 .addImm(NumLPads & 0xFFFF));
6879 unsigned VReg2 = VReg1;
6880 if ((NumLPads & 0xFFFF0000) != 0) {
6881 VReg2 = MRI->createVirtualRegister(TRC);
6882 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6884 .addImm(NumLPads >> 16));
6887 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6891 MachineConstantPool *ConstantPool = MF->getConstantPool();
6892 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6893 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6895 // MachineConstantPool wants an explicit alignment.
6896 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6898 Align = getDataLayout()->getTypeAllocSize(C->getType());
6899 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6901 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6902 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6903 .addReg(VReg1, RegState::Define)
6904 .addConstantPoolIndex(Idx)
6906 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6908 .addReg(VReg1, RegState::Kill));
6911 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6916 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6918 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6920 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6921 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6922 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6923 .addJumpTableIndex(MJTI)
6926 MachineMemOperand *JTMMOLd =
6927 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6928 MachineMemOperand::MOLoad, 4, 4);
6929 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6931 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6932 .addReg(NewVReg3, RegState::Kill)
6935 .addMemOperand(JTMMOLd));
6937 if (RelocM == Reloc::PIC_) {
6938 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6939 .addReg(NewVReg5, RegState::Kill)
6941 .addJumpTableIndex(MJTI)
6944 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
6945 .addReg(NewVReg5, RegState::Kill)
6946 .addJumpTableIndex(MJTI)
6951 // Add the jump table entries as successors to the MBB.
6952 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6953 for (std::vector<MachineBasicBlock*>::iterator
6954 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6955 MachineBasicBlock *CurMBB = *I;
6956 if (SeenMBBs.insert(CurMBB).second)
6957 DispContBB->addSuccessor(CurMBB);
6960 // N.B. the order the invoke BBs are processed in doesn't matter here.
6961 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
6962 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6963 for (MachineBasicBlock *BB : InvokeBBs) {
6965 // Remove the landing pad successor from the invoke block and replace it
6966 // with the new dispatch block.
6967 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6969 while (!Successors.empty()) {
6970 MachineBasicBlock *SMBB = Successors.pop_back_val();
6971 if (SMBB->isLandingPad()) {
6972 BB->removeSuccessor(SMBB);
6973 MBBLPads.push_back(SMBB);
6977 BB->addSuccessor(DispatchBB);
6979 // Find the invoke call and mark all of the callee-saved registers as
6980 // 'implicit defined' so that they're spilled. This prevents code from
6981 // moving instructions to before the EH block, where they will never be
6983 for (MachineBasicBlock::reverse_iterator
6984 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6985 if (!II->isCall()) continue;
6987 DenseMap<unsigned, bool> DefRegs;
6988 for (MachineInstr::mop_iterator
6989 OI = II->operands_begin(), OE = II->operands_end();
6991 if (!OI->isReg()) continue;
6992 DefRegs[OI->getReg()] = true;
6995 MachineInstrBuilder MIB(*MF, &*II);
6997 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6998 unsigned Reg = SavedRegs[i];
6999 if (Subtarget->isThumb2() &&
7000 !ARM::tGPRRegClass.contains(Reg) &&
7001 !ARM::hGPRRegClass.contains(Reg))
7003 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
7005 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
7008 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
7015 // Mark all former landing pads as non-landing pads. The dispatch is the only
7017 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7018 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
7019 (*I)->setIsLandingPad(false);
7021 // The instruction is gone now.
7022 MI->eraseFromParent();
7028 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
7029 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
7030 E = MBB->succ_end(); I != E; ++I)
7033 llvm_unreachable("Expecting a BB with two successors!");
7036 /// Return the load opcode for a given load size. If load size >= 8,
7037 /// neon opcode will be returned.
7038 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
7040 return LdSize == 16 ? ARM::VLD1q32wb_fixed
7041 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
7043 return LdSize == 4 ? ARM::tLDRi
7044 : LdSize == 2 ? ARM::tLDRHi
7045 : LdSize == 1 ? ARM::tLDRBi : 0;
7047 return LdSize == 4 ? ARM::t2LDR_POST
7048 : LdSize == 2 ? ARM::t2LDRH_POST
7049 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
7050 return LdSize == 4 ? ARM::LDR_POST_IMM
7051 : LdSize == 2 ? ARM::LDRH_POST
7052 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
7055 /// Return the store opcode for a given store size. If store size >= 8,
7056 /// neon opcode will be returned.
7057 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
7059 return StSize == 16 ? ARM::VST1q32wb_fixed
7060 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
7062 return StSize == 4 ? ARM::tSTRi
7063 : StSize == 2 ? ARM::tSTRHi
7064 : StSize == 1 ? ARM::tSTRBi : 0;
7066 return StSize == 4 ? ARM::t2STR_POST
7067 : StSize == 2 ? ARM::t2STRH_POST
7068 : StSize == 1 ? ARM::t2STRB_POST : 0;
7069 return StSize == 4 ? ARM::STR_POST_IMM
7070 : StSize == 2 ? ARM::STRH_POST
7071 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
7074 /// Emit a post-increment load operation with given size. The instructions
7075 /// will be added to BB at Pos.
7076 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
7077 const TargetInstrInfo *TII, DebugLoc dl,
7078 unsigned LdSize, unsigned Data, unsigned AddrIn,
7079 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7080 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
7081 assert(LdOpc != 0 && "Should have a load opcode");
7083 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7084 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7086 } else if (IsThumb1) {
7087 // load + update AddrIn
7088 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7089 .addReg(AddrIn).addImm(0));
7090 MachineInstrBuilder MIB =
7091 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7092 MIB = AddDefaultT1CC(MIB);
7093 MIB.addReg(AddrIn).addImm(LdSize);
7094 AddDefaultPred(MIB);
7095 } else if (IsThumb2) {
7096 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7097 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7100 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7101 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7102 .addReg(0).addImm(LdSize));
7106 /// Emit a post-increment store operation with given size. The instructions
7107 /// will be added to BB at Pos.
7108 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
7109 const TargetInstrInfo *TII, DebugLoc dl,
7110 unsigned StSize, unsigned Data, unsigned AddrIn,
7111 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7112 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
7113 assert(StOpc != 0 && "Should have a store opcode");
7115 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7116 .addReg(AddrIn).addImm(0).addReg(Data));
7117 } else if (IsThumb1) {
7118 // store + update AddrIn
7119 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
7120 .addReg(AddrIn).addImm(0));
7121 MachineInstrBuilder MIB =
7122 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7123 MIB = AddDefaultT1CC(MIB);
7124 MIB.addReg(AddrIn).addImm(StSize);
7125 AddDefaultPred(MIB);
7126 } else if (IsThumb2) {
7127 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7128 .addReg(Data).addReg(AddrIn).addImm(StSize));
7130 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7131 .addReg(Data).addReg(AddrIn).addReg(0)
7137 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
7138 MachineBasicBlock *BB) const {
7139 // This pseudo instruction has 3 operands: dst, src, size
7140 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
7141 // Otherwise, we will generate unrolled scalar copies.
7142 const TargetInstrInfo *TII =
7143 getTargetMachine().getSubtargetImpl()->getInstrInfo();
7144 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7145 MachineFunction::iterator It = BB;
7148 unsigned dest = MI->getOperand(0).getReg();
7149 unsigned src = MI->getOperand(1).getReg();
7150 unsigned SizeVal = MI->getOperand(2).getImm();
7151 unsigned Align = MI->getOperand(3).getImm();
7152 DebugLoc dl = MI->getDebugLoc();
7154 MachineFunction *MF = BB->getParent();
7155 MachineRegisterInfo &MRI = MF->getRegInfo();
7156 unsigned UnitSize = 0;
7157 const TargetRegisterClass *TRC = nullptr;
7158 const TargetRegisterClass *VecTRC = nullptr;
7160 bool IsThumb1 = Subtarget->isThumb1Only();
7161 bool IsThumb2 = Subtarget->isThumb2();
7165 } else if (Align & 2) {
7168 // Check whether we can use NEON instructions.
7169 if (!MF->getFunction()->getAttributes().
7170 hasAttribute(AttributeSet::FunctionIndex,
7171 Attribute::NoImplicitFloat) &&
7172 Subtarget->hasNEON()) {
7173 if ((Align % 16 == 0) && SizeVal >= 16)
7175 else if ((Align % 8 == 0) && SizeVal >= 8)
7178 // Can't use NEON instructions.
7183 // Select the correct opcode and register class for unit size load/store
7184 bool IsNeon = UnitSize >= 8;
7185 TRC = (IsThumb1 || IsThumb2) ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
7187 VecTRC = UnitSize == 16 ? &ARM::DPairRegClass
7188 : UnitSize == 8 ? &ARM::DPRRegClass
7191 unsigned BytesLeft = SizeVal % UnitSize;
7192 unsigned LoopSize = SizeVal - BytesLeft;
7194 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7195 // Use LDR and STR to copy.
7196 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7197 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7198 unsigned srcIn = src;
7199 unsigned destIn = dest;
7200 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7201 unsigned srcOut = MRI.createVirtualRegister(TRC);
7202 unsigned destOut = MRI.createVirtualRegister(TRC);
7203 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7204 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7205 IsThumb1, IsThumb2);
7206 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7207 IsThumb1, IsThumb2);
7212 // Handle the leftover bytes with LDRB and STRB.
7213 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7214 // [destOut] = STRB_POST(scratch, destIn, 1)
7215 for (unsigned i = 0; i < BytesLeft; i++) {
7216 unsigned srcOut = MRI.createVirtualRegister(TRC);
7217 unsigned destOut = MRI.createVirtualRegister(TRC);
7218 unsigned scratch = MRI.createVirtualRegister(TRC);
7219 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7220 IsThumb1, IsThumb2);
7221 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7222 IsThumb1, IsThumb2);
7226 MI->eraseFromParent(); // The instruction is gone now.
7230 // Expand the pseudo op to a loop.
7233 // movw varEnd, # --> with thumb2
7235 // ldrcp varEnd, idx --> without thumb2
7236 // fallthrough --> loopMBB
7238 // PHI varPhi, varEnd, varLoop
7239 // PHI srcPhi, src, srcLoop
7240 // PHI destPhi, dst, destLoop
7241 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7242 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7243 // subs varLoop, varPhi, #UnitSize
7245 // fallthrough --> exitMBB
7247 // epilogue to handle left-over bytes
7248 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7249 // [destOut] = STRB_POST(scratch, destLoop, 1)
7250 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7251 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7252 MF->insert(It, loopMBB);
7253 MF->insert(It, exitMBB);
7255 // Transfer the remainder of BB and its successor edges to exitMBB.
7256 exitMBB->splice(exitMBB->begin(), BB,
7257 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7258 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7260 // Load an immediate to varEnd.
7261 unsigned varEnd = MRI.createVirtualRegister(TRC);
7263 unsigned Vtmp = varEnd;
7264 if ((LoopSize & 0xFFFF0000) != 0)
7265 Vtmp = MRI.createVirtualRegister(TRC);
7266 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), Vtmp)
7267 .addImm(LoopSize & 0xFFFF));
7269 if ((LoopSize & 0xFFFF0000) != 0)
7270 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
7271 .addReg(Vtmp).addImm(LoopSize >> 16));
7273 MachineConstantPool *ConstantPool = MF->getConstantPool();
7274 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7275 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7277 // MachineConstantPool wants an explicit alignment.
7278 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7280 Align = getDataLayout()->getTypeAllocSize(C->getType());
7281 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7284 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7285 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7287 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7288 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7290 BB->addSuccessor(loopMBB);
7292 // Generate the loop body:
7293 // varPhi = PHI(varLoop, varEnd)
7294 // srcPhi = PHI(srcLoop, src)
7295 // destPhi = PHI(destLoop, dst)
7296 MachineBasicBlock *entryBB = BB;
7298 unsigned varLoop = MRI.createVirtualRegister(TRC);
7299 unsigned varPhi = MRI.createVirtualRegister(TRC);
7300 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7301 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7302 unsigned destLoop = MRI.createVirtualRegister(TRC);
7303 unsigned destPhi = MRI.createVirtualRegister(TRC);
7305 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7306 .addReg(varLoop).addMBB(loopMBB)
7307 .addReg(varEnd).addMBB(entryBB);
7308 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7309 .addReg(srcLoop).addMBB(loopMBB)
7310 .addReg(src).addMBB(entryBB);
7311 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7312 .addReg(destLoop).addMBB(loopMBB)
7313 .addReg(dest).addMBB(entryBB);
7315 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7316 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7317 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7318 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7319 IsThumb1, IsThumb2);
7320 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7321 IsThumb1, IsThumb2);
7323 // Decrement loop variable by UnitSize.
7325 MachineInstrBuilder MIB =
7326 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7327 MIB = AddDefaultT1CC(MIB);
7328 MIB.addReg(varPhi).addImm(UnitSize);
7329 AddDefaultPred(MIB);
7331 MachineInstrBuilder MIB =
7332 BuildMI(*BB, BB->end(), dl,
7333 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7334 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7335 MIB->getOperand(5).setReg(ARM::CPSR);
7336 MIB->getOperand(5).setIsDef(true);
7338 BuildMI(*BB, BB->end(), dl,
7339 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7340 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7342 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7343 BB->addSuccessor(loopMBB);
7344 BB->addSuccessor(exitMBB);
7346 // Add epilogue to handle BytesLeft.
7348 MachineInstr *StartOfExit = exitMBB->begin();
7350 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7351 // [destOut] = STRB_POST(scratch, destLoop, 1)
7352 unsigned srcIn = srcLoop;
7353 unsigned destIn = destLoop;
7354 for (unsigned i = 0; i < BytesLeft; i++) {
7355 unsigned srcOut = MRI.createVirtualRegister(TRC);
7356 unsigned destOut = MRI.createVirtualRegister(TRC);
7357 unsigned scratch = MRI.createVirtualRegister(TRC);
7358 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7359 IsThumb1, IsThumb2);
7360 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7361 IsThumb1, IsThumb2);
7366 MI->eraseFromParent(); // The instruction is gone now.
7371 ARMTargetLowering::EmitLowered__chkstk(MachineInstr *MI,
7372 MachineBasicBlock *MBB) const {
7373 const TargetMachine &TM = getTargetMachine();
7374 const TargetInstrInfo &TII = *TM.getSubtargetImpl()->getInstrInfo();
7375 DebugLoc DL = MI->getDebugLoc();
7377 assert(Subtarget->isTargetWindows() &&
7378 "__chkstk is only supported on Windows");
7379 assert(Subtarget->isThumb2() && "Windows on ARM requires Thumb-2 mode");
7381 // __chkstk takes the number of words to allocate on the stack in R4, and
7382 // returns the stack adjustment in number of bytes in R4. This will not
7383 // clober any other registers (other than the obvious lr).
7385 // Although, technically, IP should be considered a register which may be
7386 // clobbered, the call itself will not touch it. Windows on ARM is a pure
7387 // thumb-2 environment, so there is no interworking required. As a result, we
7388 // do not expect a veneer to be emitted by the linker, clobbering IP.
7390 // Each module receives its own copy of __chkstk, so no import thunk is
7391 // required, again, ensuring that IP is not clobbered.
7393 // Finally, although some linkers may theoretically provide a trampoline for
7394 // out of range calls (which is quite common due to a 32M range limitation of
7395 // branches for Thumb), we can generate the long-call version via
7396 // -mcmodel=large, alleviating the need for the trampoline which may clobber
7399 switch (TM.getCodeModel()) {
7400 case CodeModel::Small:
7401 case CodeModel::Medium:
7402 case CodeModel::Default:
7403 case CodeModel::Kernel:
7404 BuildMI(*MBB, MI, DL, TII.get(ARM::tBL))
7405 .addImm((unsigned)ARMCC::AL).addReg(0)
7406 .addExternalSymbol("__chkstk")
7407 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7408 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7409 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7411 case CodeModel::Large:
7412 case CodeModel::JITDefault: {
7413 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
7414 unsigned Reg = MRI.createVirtualRegister(&ARM::rGPRRegClass);
7416 BuildMI(*MBB, MI, DL, TII.get(ARM::t2MOVi32imm), Reg)
7417 .addExternalSymbol("__chkstk");
7418 BuildMI(*MBB, MI, DL, TII.get(ARM::tBLXr))
7419 .addImm((unsigned)ARMCC::AL).addReg(0)
7420 .addReg(Reg, RegState::Kill)
7421 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7422 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7423 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7428 AddDefaultCC(AddDefaultPred(BuildMI(*MBB, MI, DL, TII.get(ARM::t2SUBrr),
7430 .addReg(ARM::SP).addReg(ARM::R4)));
7432 MI->eraseFromParent();
7437 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7438 MachineBasicBlock *BB) const {
7439 const TargetInstrInfo *TII =
7440 getTargetMachine().getSubtargetImpl()->getInstrInfo();
7441 DebugLoc dl = MI->getDebugLoc();
7442 bool isThumb2 = Subtarget->isThumb2();
7443 switch (MI->getOpcode()) {
7446 llvm_unreachable("Unexpected instr type to insert");
7448 // The Thumb2 pre-indexed stores have the same MI operands, they just
7449 // define them differently in the .td files from the isel patterns, so
7450 // they need pseudos.
7451 case ARM::t2STR_preidx:
7452 MI->setDesc(TII->get(ARM::t2STR_PRE));
7454 case ARM::t2STRB_preidx:
7455 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7457 case ARM::t2STRH_preidx:
7458 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7461 case ARM::STRi_preidx:
7462 case ARM::STRBi_preidx: {
7463 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7464 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7465 // Decode the offset.
7466 unsigned Offset = MI->getOperand(4).getImm();
7467 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7468 Offset = ARM_AM::getAM2Offset(Offset);
7472 MachineMemOperand *MMO = *MI->memoperands_begin();
7473 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7474 .addOperand(MI->getOperand(0)) // Rn_wb
7475 .addOperand(MI->getOperand(1)) // Rt
7476 .addOperand(MI->getOperand(2)) // Rn
7477 .addImm(Offset) // offset (skip GPR==zero_reg)
7478 .addOperand(MI->getOperand(5)) // pred
7479 .addOperand(MI->getOperand(6))
7480 .addMemOperand(MMO);
7481 MI->eraseFromParent();
7484 case ARM::STRr_preidx:
7485 case ARM::STRBr_preidx:
7486 case ARM::STRH_preidx: {
7488 switch (MI->getOpcode()) {
7489 default: llvm_unreachable("unexpected opcode!");
7490 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7491 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7492 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7494 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7495 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7496 MIB.addOperand(MI->getOperand(i));
7497 MI->eraseFromParent();
7501 case ARM::tMOVCCr_pseudo: {
7502 // To "insert" a SELECT_CC instruction, we actually have to insert the
7503 // diamond control-flow pattern. The incoming instruction knows the
7504 // destination vreg to set, the condition code register to branch on, the
7505 // true/false values to select between, and a branch opcode to use.
7506 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7507 MachineFunction::iterator It = BB;
7513 // cmpTY ccX, r1, r2
7515 // fallthrough --> copy0MBB
7516 MachineBasicBlock *thisMBB = BB;
7517 MachineFunction *F = BB->getParent();
7518 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7519 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7520 F->insert(It, copy0MBB);
7521 F->insert(It, sinkMBB);
7523 // Transfer the remainder of BB and its successor edges to sinkMBB.
7524 sinkMBB->splice(sinkMBB->begin(), BB,
7525 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7526 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7528 BB->addSuccessor(copy0MBB);
7529 BB->addSuccessor(sinkMBB);
7531 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7532 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7535 // %FalseValue = ...
7536 // # fallthrough to sinkMBB
7539 // Update machine-CFG edges
7540 BB->addSuccessor(sinkMBB);
7543 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7546 BuildMI(*BB, BB->begin(), dl,
7547 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7548 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7549 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7551 MI->eraseFromParent(); // The pseudo instruction is gone now.
7556 case ARM::BCCZi64: {
7557 // If there is an unconditional branch to the other successor, remove it.
7558 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
7560 // Compare both parts that make up the double comparison separately for
7562 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7564 unsigned LHS1 = MI->getOperand(1).getReg();
7565 unsigned LHS2 = MI->getOperand(2).getReg();
7567 AddDefaultPred(BuildMI(BB, dl,
7568 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7569 .addReg(LHS1).addImm(0));
7570 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7571 .addReg(LHS2).addImm(0)
7572 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7574 unsigned RHS1 = MI->getOperand(3).getReg();
7575 unsigned RHS2 = MI->getOperand(4).getReg();
7576 AddDefaultPred(BuildMI(BB, dl,
7577 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7578 .addReg(LHS1).addReg(RHS1));
7579 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7580 .addReg(LHS2).addReg(RHS2)
7581 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7584 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7585 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7586 if (MI->getOperand(0).getImm() == ARMCC::NE)
7587 std::swap(destMBB, exitMBB);
7589 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7590 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7592 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7594 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7596 MI->eraseFromParent(); // The pseudo instruction is gone now.
7600 case ARM::Int_eh_sjlj_setjmp:
7601 case ARM::Int_eh_sjlj_setjmp_nofp:
7602 case ARM::tInt_eh_sjlj_setjmp:
7603 case ARM::t2Int_eh_sjlj_setjmp:
7604 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7605 EmitSjLjDispatchBlock(MI, BB);
7610 // To insert an ABS instruction, we have to insert the
7611 // diamond control-flow pattern. The incoming instruction knows the
7612 // source vreg to test against 0, the destination vreg to set,
7613 // the condition code register to branch on, the
7614 // true/false values to select between, and a branch opcode to use.
7619 // BCC (branch to SinkBB if V0 >= 0)
7620 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7621 // SinkBB: V1 = PHI(V2, V3)
7622 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7623 MachineFunction::iterator BBI = BB;
7625 MachineFunction *Fn = BB->getParent();
7626 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7627 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7628 Fn->insert(BBI, RSBBB);
7629 Fn->insert(BBI, SinkBB);
7631 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7632 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7633 bool isThumb2 = Subtarget->isThumb2();
7634 MachineRegisterInfo &MRI = Fn->getRegInfo();
7635 // In Thumb mode S must not be specified if source register is the SP or
7636 // PC and if destination register is the SP, so restrict register class
7637 unsigned NewRsbDstReg =
7638 MRI.createVirtualRegister(isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRRegClass);
7640 // Transfer the remainder of BB and its successor edges to sinkMBB.
7641 SinkBB->splice(SinkBB->begin(), BB,
7642 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7643 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7645 BB->addSuccessor(RSBBB);
7646 BB->addSuccessor(SinkBB);
7648 // fall through to SinkMBB
7649 RSBBB->addSuccessor(SinkBB);
7651 // insert a cmp at the end of BB
7652 AddDefaultPred(BuildMI(BB, dl,
7653 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7654 .addReg(ABSSrcReg).addImm(0));
7656 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7658 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7659 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7661 // insert rsbri in RSBBB
7662 // Note: BCC and rsbri will be converted into predicated rsbmi
7663 // by if-conversion pass
7664 BuildMI(*RSBBB, RSBBB->begin(), dl,
7665 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7666 .addReg(ABSSrcReg, RegState::Kill)
7667 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7669 // insert PHI in SinkBB,
7670 // reuse ABSDstReg to not change uses of ABS instruction
7671 BuildMI(*SinkBB, SinkBB->begin(), dl,
7672 TII->get(ARM::PHI), ABSDstReg)
7673 .addReg(NewRsbDstReg).addMBB(RSBBB)
7674 .addReg(ABSSrcReg).addMBB(BB);
7676 // remove ABS instruction
7677 MI->eraseFromParent();
7679 // return last added BB
7682 case ARM::COPY_STRUCT_BYVAL_I32:
7684 return EmitStructByval(MI, BB);
7685 case ARM::WIN__CHKSTK:
7686 return EmitLowered__chkstk(MI, BB);
7690 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7691 SDNode *Node) const {
7692 const MCInstrDesc *MCID = &MI->getDesc();
7693 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7694 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7695 // operand is still set to noreg. If needed, set the optional operand's
7696 // register to CPSR, and remove the redundant implicit def.
7698 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7700 // Rename pseudo opcodes.
7701 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7703 const ARMBaseInstrInfo *TII = static_cast<const ARMBaseInstrInfo *>(
7704 getTargetMachine().getSubtargetImpl()->getInstrInfo());
7705 MCID = &TII->get(NewOpc);
7707 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7708 "converted opcode should be the same except for cc_out");
7712 // Add the optional cc_out operand
7713 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7715 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7717 // Any ARM instruction that sets the 's' bit should specify an optional
7718 // "cc_out" operand in the last operand position.
7719 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7720 assert(!NewOpc && "Optional cc_out operand required");
7723 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7724 // since we already have an optional CPSR def.
7725 bool definesCPSR = false;
7726 bool deadCPSR = false;
7727 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7729 const MachineOperand &MO = MI->getOperand(i);
7730 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7734 MI->RemoveOperand(i);
7739 assert(!NewOpc && "Optional cc_out operand required");
7742 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7744 assert(!MI->getOperand(ccOutIdx).getReg() &&
7745 "expect uninitialized optional cc_out operand");
7749 // If this instruction was defined with an optional CPSR def and its dag node
7750 // had a live implicit CPSR def, then activate the optional CPSR def.
7751 MachineOperand &MO = MI->getOperand(ccOutIdx);
7752 MO.setReg(ARM::CPSR);
7756 //===----------------------------------------------------------------------===//
7757 // ARM Optimization Hooks
7758 //===----------------------------------------------------------------------===//
7760 // Helper function that checks if N is a null or all ones constant.
7761 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7762 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7765 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7768 // Return true if N is conditionally 0 or all ones.
7769 // Detects these expressions where cc is an i1 value:
7771 // (select cc 0, y) [AllOnes=0]
7772 // (select cc y, 0) [AllOnes=0]
7773 // (zext cc) [AllOnes=0]
7774 // (sext cc) [AllOnes=0/1]
7775 // (select cc -1, y) [AllOnes=1]
7776 // (select cc y, -1) [AllOnes=1]
7778 // Invert is set when N is the null/all ones constant when CC is false.
7779 // OtherOp is set to the alternative value of N.
7780 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7781 SDValue &CC, bool &Invert,
7783 SelectionDAG &DAG) {
7784 switch (N->getOpcode()) {
7785 default: return false;
7787 CC = N->getOperand(0);
7788 SDValue N1 = N->getOperand(1);
7789 SDValue N2 = N->getOperand(2);
7790 if (isZeroOrAllOnes(N1, AllOnes)) {
7795 if (isZeroOrAllOnes(N2, AllOnes)) {
7802 case ISD::ZERO_EXTEND:
7803 // (zext cc) can never be the all ones value.
7807 case ISD::SIGN_EXTEND: {
7808 EVT VT = N->getValueType(0);
7809 CC = N->getOperand(0);
7810 if (CC.getValueType() != MVT::i1)
7814 // When looking for an AllOnes constant, N is an sext, and the 'other'
7816 OtherOp = DAG.getConstant(0, VT);
7817 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7818 // When looking for a 0 constant, N can be zext or sext.
7819 OtherOp = DAG.getConstant(1, VT);
7821 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7827 // Combine a constant select operand into its use:
7829 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7830 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7831 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7832 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7833 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7835 // The transform is rejected if the select doesn't have a constant operand that
7836 // is null, or all ones when AllOnes is set.
7838 // Also recognize sext/zext from i1:
7840 // (add (zext cc), x) -> (select cc (add x, 1), x)
7841 // (add (sext cc), x) -> (select cc (add x, -1), x)
7843 // These transformations eventually create predicated instructions.
7845 // @param N The node to transform.
7846 // @param Slct The N operand that is a select.
7847 // @param OtherOp The other N operand (x above).
7848 // @param DCI Context.
7849 // @param AllOnes Require the select constant to be all ones instead of null.
7850 // @returns The new node, or SDValue() on failure.
7852 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7853 TargetLowering::DAGCombinerInfo &DCI,
7854 bool AllOnes = false) {
7855 SelectionDAG &DAG = DCI.DAG;
7856 EVT VT = N->getValueType(0);
7857 SDValue NonConstantVal;
7860 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7861 NonConstantVal, DAG))
7864 // Slct is now know to be the desired identity constant when CC is true.
7865 SDValue TrueVal = OtherOp;
7866 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
7867 OtherOp, NonConstantVal);
7868 // Unless SwapSelectOps says CC should be false.
7870 std::swap(TrueVal, FalseVal);
7872 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
7873 CCOp, TrueVal, FalseVal);
7876 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7878 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7879 TargetLowering::DAGCombinerInfo &DCI) {
7880 SDValue N0 = N->getOperand(0);
7881 SDValue N1 = N->getOperand(1);
7882 if (N0.getNode()->hasOneUse()) {
7883 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7884 if (Result.getNode())
7887 if (N1.getNode()->hasOneUse()) {
7888 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7889 if (Result.getNode())
7895 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7896 // (only after legalization).
7897 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7898 TargetLowering::DAGCombinerInfo &DCI,
7899 const ARMSubtarget *Subtarget) {
7901 // Only perform optimization if after legalize, and if NEON is available. We
7902 // also expected both operands to be BUILD_VECTORs.
7903 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7904 || N0.getOpcode() != ISD::BUILD_VECTOR
7905 || N1.getOpcode() != ISD::BUILD_VECTOR)
7908 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7909 EVT VT = N->getValueType(0);
7910 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7913 // Check that the vector operands are of the right form.
7914 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7915 // operands, where N is the size of the formed vector.
7916 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7917 // index such that we have a pair wise add pattern.
7919 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7920 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7922 SDValue Vec = N0->getOperand(0)->getOperand(0);
7923 SDNode *V = Vec.getNode();
7924 unsigned nextIndex = 0;
7926 // For each operands to the ADD which are BUILD_VECTORs,
7927 // check to see if each of their operands are an EXTRACT_VECTOR with
7928 // the same vector and appropriate index.
7929 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7930 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7931 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7933 SDValue ExtVec0 = N0->getOperand(i);
7934 SDValue ExtVec1 = N1->getOperand(i);
7936 // First operand is the vector, verify its the same.
7937 if (V != ExtVec0->getOperand(0).getNode() ||
7938 V != ExtVec1->getOperand(0).getNode())
7941 // Second is the constant, verify its correct.
7942 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7943 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7945 // For the constant, we want to see all the even or all the odd.
7946 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7947 || C1->getZExtValue() != nextIndex+1)
7956 // Create VPADDL node.
7957 SelectionDAG &DAG = DCI.DAG;
7958 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7960 // Build operand list.
7961 SmallVector<SDValue, 8> Ops;
7962 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
7963 TLI.getPointerTy()));
7965 // Input is the vector.
7968 // Get widened type and narrowed type.
7970 unsigned numElem = VT.getVectorNumElements();
7972 EVT inputLaneType = Vec.getValueType().getVectorElementType();
7973 switch (inputLaneType.getSimpleVT().SimpleTy) {
7974 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7975 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7976 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7978 llvm_unreachable("Invalid vector element type for padd optimization.");
7981 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), widenType, Ops);
7982 unsigned ExtOp = VT.bitsGT(tmp.getValueType()) ? ISD::ANY_EXTEND : ISD::TRUNCATE;
7983 return DAG.getNode(ExtOp, SDLoc(N), VT, tmp);
7986 static SDValue findMUL_LOHI(SDValue V) {
7987 if (V->getOpcode() == ISD::UMUL_LOHI ||
7988 V->getOpcode() == ISD::SMUL_LOHI)
7993 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7994 TargetLowering::DAGCombinerInfo &DCI,
7995 const ARMSubtarget *Subtarget) {
7997 if (Subtarget->isThumb1Only()) return SDValue();
7999 // Only perform the checks after legalize when the pattern is available.
8000 if (DCI.isBeforeLegalize()) return SDValue();
8002 // Look for multiply add opportunities.
8003 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
8004 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
8005 // a glue link from the first add to the second add.
8006 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
8007 // a S/UMLAL instruction.
8010 // \ / \ [no multiline comment]
8016 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
8017 SDValue AddcOp0 = AddcNode->getOperand(0);
8018 SDValue AddcOp1 = AddcNode->getOperand(1);
8020 // Check if the two operands are from the same mul_lohi node.
8021 if (AddcOp0.getNode() == AddcOp1.getNode())
8024 assert(AddcNode->getNumValues() == 2 &&
8025 AddcNode->getValueType(0) == MVT::i32 &&
8026 "Expect ADDC with two result values. First: i32");
8028 // Check that we have a glued ADDC node.
8029 if (AddcNode->getValueType(1) != MVT::Glue)
8032 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
8033 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
8034 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
8035 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
8036 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
8039 // Look for the glued ADDE.
8040 SDNode* AddeNode = AddcNode->getGluedUser();
8044 // Make sure it is really an ADDE.
8045 if (AddeNode->getOpcode() != ISD::ADDE)
8048 assert(AddeNode->getNumOperands() == 3 &&
8049 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
8050 "ADDE node has the wrong inputs");
8052 // Check for the triangle shape.
8053 SDValue AddeOp0 = AddeNode->getOperand(0);
8054 SDValue AddeOp1 = AddeNode->getOperand(1);
8056 // Make sure that the ADDE operands are not coming from the same node.
8057 if (AddeOp0.getNode() == AddeOp1.getNode())
8060 // Find the MUL_LOHI node walking up ADDE's operands.
8061 bool IsLeftOperandMUL = false;
8062 SDValue MULOp = findMUL_LOHI(AddeOp0);
8063 if (MULOp == SDValue())
8064 MULOp = findMUL_LOHI(AddeOp1);
8066 IsLeftOperandMUL = true;
8067 if (MULOp == SDValue())
8070 // Figure out the right opcode.
8071 unsigned Opc = MULOp->getOpcode();
8072 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
8074 // Figure out the high and low input values to the MLAL node.
8075 SDValue* HiMul = &MULOp;
8076 SDValue* HiAdd = nullptr;
8077 SDValue* LoMul = nullptr;
8078 SDValue* LowAdd = nullptr;
8080 if (IsLeftOperandMUL)
8086 if (AddcOp0->getOpcode() == Opc) {
8090 if (AddcOp1->getOpcode() == Opc) {
8098 if (LoMul->getNode() != HiMul->getNode())
8101 // Create the merged node.
8102 SelectionDAG &DAG = DCI.DAG;
8104 // Build operand list.
8105 SmallVector<SDValue, 8> Ops;
8106 Ops.push_back(LoMul->getOperand(0));
8107 Ops.push_back(LoMul->getOperand(1));
8108 Ops.push_back(*LowAdd);
8109 Ops.push_back(*HiAdd);
8111 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
8112 DAG.getVTList(MVT::i32, MVT::i32), Ops);
8114 // Replace the ADDs' nodes uses by the MLA node's values.
8115 SDValue HiMLALResult(MLALNode.getNode(), 1);
8116 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
8118 SDValue LoMLALResult(MLALNode.getNode(), 0);
8119 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
8121 // Return original node to notify the driver to stop replacing.
8122 SDValue resNode(AddcNode, 0);
8126 /// PerformADDCCombine - Target-specific dag combine transform from
8127 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
8128 static SDValue PerformADDCCombine(SDNode *N,
8129 TargetLowering::DAGCombinerInfo &DCI,
8130 const ARMSubtarget *Subtarget) {
8132 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
8136 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
8137 /// operands N0 and N1. This is a helper for PerformADDCombine that is
8138 /// called with the default operands, and if that fails, with commuted
8140 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8141 TargetLowering::DAGCombinerInfo &DCI,
8142 const ARMSubtarget *Subtarget){
8144 // Attempt to create vpaddl for this add.
8145 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8146 if (Result.getNode())
8149 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8150 if (N0.getNode()->hasOneUse()) {
8151 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8152 if (Result.getNode()) return Result;
8157 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8159 static SDValue PerformADDCombine(SDNode *N,
8160 TargetLowering::DAGCombinerInfo &DCI,
8161 const ARMSubtarget *Subtarget) {
8162 SDValue N0 = N->getOperand(0);
8163 SDValue N1 = N->getOperand(1);
8165 // First try with the default operand order.
8166 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8167 if (Result.getNode())
8170 // If that didn't work, try again with the operands commuted.
8171 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8174 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8176 static SDValue PerformSUBCombine(SDNode *N,
8177 TargetLowering::DAGCombinerInfo &DCI) {
8178 SDValue N0 = N->getOperand(0);
8179 SDValue N1 = N->getOperand(1);
8181 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8182 if (N1.getNode()->hasOneUse()) {
8183 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8184 if (Result.getNode()) return Result;
8190 /// PerformVMULCombine
8191 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8192 /// special multiplier accumulator forwarding.
8198 // However, for (A + B) * (A + B),
8205 static SDValue PerformVMULCombine(SDNode *N,
8206 TargetLowering::DAGCombinerInfo &DCI,
8207 const ARMSubtarget *Subtarget) {
8208 if (!Subtarget->hasVMLxForwarding())
8211 SelectionDAG &DAG = DCI.DAG;
8212 SDValue N0 = N->getOperand(0);
8213 SDValue N1 = N->getOperand(1);
8214 unsigned Opcode = N0.getOpcode();
8215 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8216 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8217 Opcode = N1.getOpcode();
8218 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8219 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8227 EVT VT = N->getValueType(0);
8229 SDValue N00 = N0->getOperand(0);
8230 SDValue N01 = N0->getOperand(1);
8231 return DAG.getNode(Opcode, DL, VT,
8232 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8233 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8236 static SDValue PerformMULCombine(SDNode *N,
8237 TargetLowering::DAGCombinerInfo &DCI,
8238 const ARMSubtarget *Subtarget) {
8239 SelectionDAG &DAG = DCI.DAG;
8241 if (Subtarget->isThumb1Only())
8244 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8247 EVT VT = N->getValueType(0);
8248 if (VT.is64BitVector() || VT.is128BitVector())
8249 return PerformVMULCombine(N, DCI, Subtarget);
8253 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8257 int64_t MulAmt = C->getSExtValue();
8258 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8260 ShiftAmt = ShiftAmt & (32 - 1);
8261 SDValue V = N->getOperand(0);
8265 MulAmt >>= ShiftAmt;
8268 if (isPowerOf2_32(MulAmt - 1)) {
8269 // (mul x, 2^N + 1) => (add (shl x, N), x)
8270 Res = DAG.getNode(ISD::ADD, DL, VT,
8272 DAG.getNode(ISD::SHL, DL, VT,
8274 DAG.getConstant(Log2_32(MulAmt - 1),
8276 } else if (isPowerOf2_32(MulAmt + 1)) {
8277 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8278 Res = DAG.getNode(ISD::SUB, DL, VT,
8279 DAG.getNode(ISD::SHL, DL, VT,
8281 DAG.getConstant(Log2_32(MulAmt + 1),
8287 uint64_t MulAmtAbs = -MulAmt;
8288 if (isPowerOf2_32(MulAmtAbs + 1)) {
8289 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8290 Res = DAG.getNode(ISD::SUB, DL, VT,
8292 DAG.getNode(ISD::SHL, DL, VT,
8294 DAG.getConstant(Log2_32(MulAmtAbs + 1),
8296 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8297 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8298 Res = DAG.getNode(ISD::ADD, DL, VT,
8300 DAG.getNode(ISD::SHL, DL, VT,
8302 DAG.getConstant(Log2_32(MulAmtAbs-1),
8304 Res = DAG.getNode(ISD::SUB, DL, VT,
8305 DAG.getConstant(0, MVT::i32),Res);
8312 Res = DAG.getNode(ISD::SHL, DL, VT,
8313 Res, DAG.getConstant(ShiftAmt, MVT::i32));
8315 // Do not add new nodes to DAG combiner worklist.
8316 DCI.CombineTo(N, Res, false);
8320 static SDValue PerformANDCombine(SDNode *N,
8321 TargetLowering::DAGCombinerInfo &DCI,
8322 const ARMSubtarget *Subtarget) {
8324 // Attempt to use immediate-form VBIC
8325 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8327 EVT VT = N->getValueType(0);
8328 SelectionDAG &DAG = DCI.DAG;
8330 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8333 APInt SplatBits, SplatUndef;
8334 unsigned SplatBitSize;
8337 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8338 if (SplatBitSize <= 64) {
8340 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8341 SplatUndef.getZExtValue(), SplatBitSize,
8342 DAG, VbicVT, VT.is128BitVector(),
8344 if (Val.getNode()) {
8346 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8347 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8348 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8353 if (!Subtarget->isThumb1Only()) {
8354 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8355 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8356 if (Result.getNode())
8363 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8364 static SDValue PerformORCombine(SDNode *N,
8365 TargetLowering::DAGCombinerInfo &DCI,
8366 const ARMSubtarget *Subtarget) {
8367 // Attempt to use immediate-form VORR
8368 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8370 EVT VT = N->getValueType(0);
8371 SelectionDAG &DAG = DCI.DAG;
8373 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8376 APInt SplatBits, SplatUndef;
8377 unsigned SplatBitSize;
8379 if (BVN && Subtarget->hasNEON() &&
8380 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8381 if (SplatBitSize <= 64) {
8383 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8384 SplatUndef.getZExtValue(), SplatBitSize,
8385 DAG, VorrVT, VT.is128BitVector(),
8387 if (Val.getNode()) {
8389 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8390 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8391 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8396 if (!Subtarget->isThumb1Only()) {
8397 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8398 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8399 if (Result.getNode())
8403 // The code below optimizes (or (and X, Y), Z).
8404 // The AND operand needs to have a single user to make these optimizations
8406 SDValue N0 = N->getOperand(0);
8407 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8409 SDValue N1 = N->getOperand(1);
8411 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8412 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8413 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8415 unsigned SplatBitSize;
8418 APInt SplatBits0, SplatBits1;
8419 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8420 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8421 // Ensure that the second operand of both ands are constants
8422 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8423 HasAnyUndefs) && !HasAnyUndefs) {
8424 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8425 HasAnyUndefs) && !HasAnyUndefs) {
8426 // Ensure that the bit width of the constants are the same and that
8427 // the splat arguments are logical inverses as per the pattern we
8428 // are trying to simplify.
8429 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8430 SplatBits0 == ~SplatBits1) {
8431 // Canonicalize the vector type to make instruction selection
8433 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8434 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8438 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8444 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8447 // BFI is only available on V6T2+
8448 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8452 // 1) or (and A, mask), val => ARMbfi A, val, mask
8453 // iff (val & mask) == val
8455 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8456 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8457 // && mask == ~mask2
8458 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8459 // && ~mask == mask2
8460 // (i.e., copy a bitfield value into another bitfield of the same width)
8465 SDValue N00 = N0.getOperand(0);
8467 // The value and the mask need to be constants so we can verify this is
8468 // actually a bitfield set. If the mask is 0xffff, we can do better
8469 // via a movt instruction, so don't use BFI in that case.
8470 SDValue MaskOp = N0.getOperand(1);
8471 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8474 unsigned Mask = MaskC->getZExtValue();
8478 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8479 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8481 unsigned Val = N1C->getZExtValue();
8482 if ((Val & ~Mask) != Val)
8485 if (ARM::isBitFieldInvertedMask(Mask)) {
8486 Val >>= countTrailingZeros(~Mask);
8488 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8489 DAG.getConstant(Val, MVT::i32),
8490 DAG.getConstant(Mask, MVT::i32));
8492 // Do not add new nodes to DAG combiner worklist.
8493 DCI.CombineTo(N, Res, false);
8496 } else if (N1.getOpcode() == ISD::AND) {
8497 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8498 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8501 unsigned Mask2 = N11C->getZExtValue();
8503 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8505 if (ARM::isBitFieldInvertedMask(Mask) &&
8507 // The pack halfword instruction works better for masks that fit it,
8508 // so use that when it's available.
8509 if (Subtarget->hasT2ExtractPack() &&
8510 (Mask == 0xffff || Mask == 0xffff0000))
8513 unsigned amt = countTrailingZeros(Mask2);
8514 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8515 DAG.getConstant(amt, MVT::i32));
8516 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8517 DAG.getConstant(Mask, MVT::i32));
8518 // Do not add new nodes to DAG combiner worklist.
8519 DCI.CombineTo(N, Res, false);
8521 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8523 // The pack halfword instruction works better for masks that fit it,
8524 // so use that when it's available.
8525 if (Subtarget->hasT2ExtractPack() &&
8526 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8529 unsigned lsb = countTrailingZeros(Mask);
8530 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8531 DAG.getConstant(lsb, MVT::i32));
8532 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8533 DAG.getConstant(Mask2, MVT::i32));
8534 // Do not add new nodes to DAG combiner worklist.
8535 DCI.CombineTo(N, Res, false);
8540 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8541 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8542 ARM::isBitFieldInvertedMask(~Mask)) {
8543 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8544 // where lsb(mask) == #shamt and masked bits of B are known zero.
8545 SDValue ShAmt = N00.getOperand(1);
8546 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8547 unsigned LSB = countTrailingZeros(Mask);
8551 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8552 DAG.getConstant(~Mask, MVT::i32));
8554 // Do not add new nodes to DAG combiner worklist.
8555 DCI.CombineTo(N, Res, false);
8561 static SDValue PerformXORCombine(SDNode *N,
8562 TargetLowering::DAGCombinerInfo &DCI,
8563 const ARMSubtarget *Subtarget) {
8564 EVT VT = N->getValueType(0);
8565 SelectionDAG &DAG = DCI.DAG;
8567 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8570 if (!Subtarget->isThumb1Only()) {
8571 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8572 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8573 if (Result.getNode())
8580 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8581 /// the bits being cleared by the AND are not demanded by the BFI.
8582 static SDValue PerformBFICombine(SDNode *N,
8583 TargetLowering::DAGCombinerInfo &DCI) {
8584 SDValue N1 = N->getOperand(1);
8585 if (N1.getOpcode() == ISD::AND) {
8586 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8589 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8590 unsigned LSB = countTrailingZeros(~InvMask);
8591 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8593 static_cast<unsigned>(std::numeric_limits<unsigned>::digits) &&
8594 "undefined behavior");
8595 unsigned Mask = (1u << Width) - 1;
8596 unsigned Mask2 = N11C->getZExtValue();
8597 if ((Mask & (~Mask2)) == 0)
8598 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8599 N->getOperand(0), N1.getOperand(0),
8605 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8606 /// ARMISD::VMOVRRD.
8607 static SDValue PerformVMOVRRDCombine(SDNode *N,
8608 TargetLowering::DAGCombinerInfo &DCI,
8609 const ARMSubtarget *Subtarget) {
8610 // vmovrrd(vmovdrr x, y) -> x,y
8611 SDValue InDouble = N->getOperand(0);
8612 if (InDouble.getOpcode() == ARMISD::VMOVDRR && !Subtarget->isFPOnlySP())
8613 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8615 // vmovrrd(load f64) -> (load i32), (load i32)
8616 SDNode *InNode = InDouble.getNode();
8617 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8618 InNode->getValueType(0) == MVT::f64 &&
8619 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8620 !cast<LoadSDNode>(InNode)->isVolatile()) {
8621 // TODO: Should this be done for non-FrameIndex operands?
8622 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8624 SelectionDAG &DAG = DCI.DAG;
8626 SDValue BasePtr = LD->getBasePtr();
8627 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8628 LD->getPointerInfo(), LD->isVolatile(),
8629 LD->isNonTemporal(), LD->isInvariant(),
8630 LD->getAlignment());
8632 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8633 DAG.getConstant(4, MVT::i32));
8634 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8635 LD->getPointerInfo(), LD->isVolatile(),
8636 LD->isNonTemporal(), LD->isInvariant(),
8637 std::min(4U, LD->getAlignment() / 2));
8639 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8640 if (DCI.DAG.getTargetLoweringInfo().isBigEndian())
8641 std::swap (NewLD1, NewLD2);
8642 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8649 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8650 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8651 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8652 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8653 SDValue Op0 = N->getOperand(0);
8654 SDValue Op1 = N->getOperand(1);
8655 if (Op0.getOpcode() == ISD::BITCAST)
8656 Op0 = Op0.getOperand(0);
8657 if (Op1.getOpcode() == ISD::BITCAST)
8658 Op1 = Op1.getOperand(0);
8659 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8660 Op0.getNode() == Op1.getNode() &&
8661 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8662 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8663 N->getValueType(0), Op0.getOperand(0));
8667 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8668 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8669 /// i64 vector to have f64 elements, since the value can then be loaded
8670 /// directly into a VFP register.
8671 static bool hasNormalLoadOperand(SDNode *N) {
8672 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8673 for (unsigned i = 0; i < NumElts; ++i) {
8674 SDNode *Elt = N->getOperand(i).getNode();
8675 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8681 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8682 /// ISD::BUILD_VECTOR.
8683 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8684 TargetLowering::DAGCombinerInfo &DCI,
8685 const ARMSubtarget *Subtarget) {
8686 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8687 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8688 // into a pair of GPRs, which is fine when the value is used as a scalar,
8689 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8690 SelectionDAG &DAG = DCI.DAG;
8691 if (N->getNumOperands() == 2) {
8692 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8697 // Load i64 elements as f64 values so that type legalization does not split
8698 // them up into i32 values.
8699 EVT VT = N->getValueType(0);
8700 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8703 SmallVector<SDValue, 8> Ops;
8704 unsigned NumElts = VT.getVectorNumElements();
8705 for (unsigned i = 0; i < NumElts; ++i) {
8706 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8708 // Make the DAGCombiner fold the bitcast.
8709 DCI.AddToWorklist(V.getNode());
8711 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8712 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops);
8713 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8716 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
8718 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8719 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
8720 // At that time, we may have inserted bitcasts from integer to float.
8721 // If these bitcasts have survived DAGCombine, change the lowering of this
8722 // BUILD_VECTOR in something more vector friendly, i.e., that does not
8723 // force to use floating point types.
8725 // Make sure we can change the type of the vector.
8726 // This is possible iff:
8727 // 1. The vector is only used in a bitcast to a integer type. I.e.,
8728 // 1.1. Vector is used only once.
8729 // 1.2. Use is a bit convert to an integer type.
8730 // 2. The size of its operands are 32-bits (64-bits are not legal).
8731 EVT VT = N->getValueType(0);
8732 EVT EltVT = VT.getVectorElementType();
8734 // Check 1.1. and 2.
8735 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
8738 // By construction, the input type must be float.
8739 assert(EltVT == MVT::f32 && "Unexpected type!");
8742 SDNode *Use = *N->use_begin();
8743 if (Use->getOpcode() != ISD::BITCAST ||
8744 Use->getValueType(0).isFloatingPoint())
8747 // Check profitability.
8748 // Model is, if more than half of the relevant operands are bitcast from
8749 // i32, turn the build_vector into a sequence of insert_vector_elt.
8750 // Relevant operands are everything that is not statically
8751 // (i.e., at compile time) bitcasted.
8752 unsigned NumOfBitCastedElts = 0;
8753 unsigned NumElts = VT.getVectorNumElements();
8754 unsigned NumOfRelevantElts = NumElts;
8755 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
8756 SDValue Elt = N->getOperand(Idx);
8757 if (Elt->getOpcode() == ISD::BITCAST) {
8758 // Assume only bit cast to i32 will go away.
8759 if (Elt->getOperand(0).getValueType() == MVT::i32)
8760 ++NumOfBitCastedElts;
8761 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
8762 // Constants are statically casted, thus do not count them as
8763 // relevant operands.
8764 --NumOfRelevantElts;
8767 // Check if more than half of the elements require a non-free bitcast.
8768 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
8771 SelectionDAG &DAG = DCI.DAG;
8772 // Create the new vector type.
8773 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
8774 // Check if the type is legal.
8775 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8776 if (!TLI.isTypeLegal(VecVT))
8780 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
8781 // => BITCAST INSERT_VECTOR_ELT
8782 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
8784 SDValue Vec = DAG.getUNDEF(VecVT);
8786 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
8787 SDValue V = N->getOperand(Idx);
8788 if (V.getOpcode() == ISD::UNDEF)
8790 if (V.getOpcode() == ISD::BITCAST &&
8791 V->getOperand(0).getValueType() == MVT::i32)
8792 // Fold obvious case.
8793 V = V.getOperand(0);
8795 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
8796 // Make the DAGCombiner fold the bitcasts.
8797 DCI.AddToWorklist(V.getNode());
8799 SDValue LaneIdx = DAG.getConstant(Idx, MVT::i32);
8800 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
8802 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
8803 // Make the DAGCombiner fold the bitcasts.
8804 DCI.AddToWorklist(Vec.getNode());
8808 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8809 /// ISD::INSERT_VECTOR_ELT.
8810 static SDValue PerformInsertEltCombine(SDNode *N,
8811 TargetLowering::DAGCombinerInfo &DCI) {
8812 // Bitcast an i64 load inserted into a vector to f64.
8813 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8814 EVT VT = N->getValueType(0);
8815 SDNode *Elt = N->getOperand(1).getNode();
8816 if (VT.getVectorElementType() != MVT::i64 ||
8817 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8820 SelectionDAG &DAG = DCI.DAG;
8822 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8823 VT.getVectorNumElements());
8824 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8825 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8826 // Make the DAGCombiner fold the bitcasts.
8827 DCI.AddToWorklist(Vec.getNode());
8828 DCI.AddToWorklist(V.getNode());
8829 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8830 Vec, V, N->getOperand(2));
8831 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8834 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8835 /// ISD::VECTOR_SHUFFLE.
8836 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8837 // The LLVM shufflevector instruction does not require the shuffle mask
8838 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8839 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8840 // operands do not match the mask length, they are extended by concatenating
8841 // them with undef vectors. That is probably the right thing for other
8842 // targets, but for NEON it is better to concatenate two double-register
8843 // size vector operands into a single quad-register size vector. Do that
8844 // transformation here:
8845 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8846 // shuffle(concat(v1, v2), undef)
8847 SDValue Op0 = N->getOperand(0);
8848 SDValue Op1 = N->getOperand(1);
8849 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8850 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8851 Op0.getNumOperands() != 2 ||
8852 Op1.getNumOperands() != 2)
8854 SDValue Concat0Op1 = Op0.getOperand(1);
8855 SDValue Concat1Op1 = Op1.getOperand(1);
8856 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8857 Concat1Op1.getOpcode() != ISD::UNDEF)
8859 // Skip the transformation if any of the types are illegal.
8860 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8861 EVT VT = N->getValueType(0);
8862 if (!TLI.isTypeLegal(VT) ||
8863 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8864 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8867 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
8868 Op0.getOperand(0), Op1.getOperand(0));
8869 // Translate the shuffle mask.
8870 SmallVector<int, 16> NewMask;
8871 unsigned NumElts = VT.getVectorNumElements();
8872 unsigned HalfElts = NumElts/2;
8873 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8874 for (unsigned n = 0; n < NumElts; ++n) {
8875 int MaskElt = SVN->getMaskElt(n);
8877 if (MaskElt < (int)HalfElts)
8879 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8880 NewElt = HalfElts + MaskElt - NumElts;
8881 NewMask.push_back(NewElt);
8883 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
8884 DAG.getUNDEF(VT), NewMask.data());
8887 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
8888 /// NEON load/store intrinsics to merge base address updates.
8889 static SDValue CombineBaseUpdate(SDNode *N,
8890 TargetLowering::DAGCombinerInfo &DCI) {
8891 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8894 SelectionDAG &DAG = DCI.DAG;
8895 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8896 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8897 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
8898 SDValue Addr = N->getOperand(AddrOpIdx);
8900 // Search for a use of the address operand that is an increment.
8901 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8902 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8904 if (User->getOpcode() != ISD::ADD ||
8905 UI.getUse().getResNo() != Addr.getResNo())
8908 // Check that the add is independent of the load/store. Otherwise, folding
8909 // it would create a cycle.
8910 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8913 // Find the new opcode for the updating load/store.
8915 bool isLaneOp = false;
8916 unsigned NewOpc = 0;
8917 unsigned NumVecs = 0;
8919 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8921 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8922 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8924 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8926 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8928 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8930 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8931 NumVecs = 2; isLaneOp = true; break;
8932 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8933 NumVecs = 3; isLaneOp = true; break;
8934 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8935 NumVecs = 4; isLaneOp = true; break;
8936 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8937 NumVecs = 1; isLoad = false; break;
8938 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8939 NumVecs = 2; isLoad = false; break;
8940 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8941 NumVecs = 3; isLoad = false; break;
8942 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8943 NumVecs = 4; isLoad = false; break;
8944 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8945 NumVecs = 2; isLoad = false; isLaneOp = true; break;
8946 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8947 NumVecs = 3; isLoad = false; isLaneOp = true; break;
8948 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8949 NumVecs = 4; isLoad = false; isLaneOp = true; break;
8953 switch (N->getOpcode()) {
8954 default: llvm_unreachable("unexpected opcode for Neon base update");
8955 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8956 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8957 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8961 // Find the size of memory referenced by the load/store.
8964 VecTy = N->getValueType(0);
8966 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8967 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8969 NumBytes /= VecTy.getVectorNumElements();
8971 // If the increment is a constant, it must match the memory ref size.
8972 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8973 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8974 uint64_t IncVal = CInc->getZExtValue();
8975 if (IncVal != NumBytes)
8977 } else if (NumBytes >= 3 * 16) {
8978 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8979 // separate instructions that make it harder to use a non-constant update.
8983 // Create the new updating load/store node.
8985 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
8987 for (n = 0; n < NumResultVecs; ++n)
8989 Tys[n++] = MVT::i32;
8990 Tys[n] = MVT::Other;
8991 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs+2));
8992 SmallVector<SDValue, 8> Ops;
8993 Ops.push_back(N->getOperand(0)); // incoming chain
8994 Ops.push_back(N->getOperand(AddrOpIdx));
8996 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
8997 Ops.push_back(N->getOperand(i));
8999 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
9000 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys,
9001 Ops, MemInt->getMemoryVT(),
9002 MemInt->getMemOperand());
9005 std::vector<SDValue> NewResults;
9006 for (unsigned i = 0; i < NumResultVecs; ++i) {
9007 NewResults.push_back(SDValue(UpdN.getNode(), i));
9009 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
9010 DCI.CombineTo(N, NewResults);
9011 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9018 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
9019 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
9020 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
9022 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9023 SelectionDAG &DAG = DCI.DAG;
9024 EVT VT = N->getValueType(0);
9025 // vldN-dup instructions only support 64-bit vectors for N > 1.
9026 if (!VT.is64BitVector())
9029 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
9030 SDNode *VLD = N->getOperand(0).getNode();
9031 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
9033 unsigned NumVecs = 0;
9034 unsigned NewOpc = 0;
9035 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
9036 if (IntNo == Intrinsic::arm_neon_vld2lane) {
9038 NewOpc = ARMISD::VLD2DUP;
9039 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
9041 NewOpc = ARMISD::VLD3DUP;
9042 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
9044 NewOpc = ARMISD::VLD4DUP;
9049 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
9050 // numbers match the load.
9051 unsigned VLDLaneNo =
9052 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
9053 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9055 // Ignore uses of the chain result.
9056 if (UI.getUse().getResNo() == NumVecs)
9059 if (User->getOpcode() != ARMISD::VDUPLANE ||
9060 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9064 // Create the vldN-dup node.
9067 for (n = 0; n < NumVecs; ++n)
9069 Tys[n] = MVT::Other;
9070 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumVecs+1));
9071 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9072 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9073 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9074 Ops, VLDMemInt->getMemoryVT(),
9075 VLDMemInt->getMemOperand());
9078 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9080 unsigned ResNo = UI.getUse().getResNo();
9081 // Ignore uses of the chain result.
9082 if (ResNo == NumVecs)
9085 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9088 // Now the vldN-lane intrinsic is dead except for its chain result.
9089 // Update uses of the chain.
9090 std::vector<SDValue> VLDDupResults;
9091 for (unsigned n = 0; n < NumVecs; ++n)
9092 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9093 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9094 DCI.CombineTo(VLD, VLDDupResults);
9099 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9100 /// ARMISD::VDUPLANE.
9101 static SDValue PerformVDUPLANECombine(SDNode *N,
9102 TargetLowering::DAGCombinerInfo &DCI) {
9103 SDValue Op = N->getOperand(0);
9105 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9106 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9107 if (CombineVLDDUP(N, DCI))
9108 return SDValue(N, 0);
9110 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9111 // redundant. Ignore bit_converts for now; element sizes are checked below.
9112 while (Op.getOpcode() == ISD::BITCAST)
9113 Op = Op.getOperand(0);
9114 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9117 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9118 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9119 // The canonical VMOV for a zero vector uses a 32-bit element size.
9120 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9122 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9124 EVT VT = N->getValueType(0);
9125 if (EltSize > VT.getVectorElementType().getSizeInBits())
9128 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9131 /// PerformSTORECombine - Target-specific dag combine xforms for
9133 static SDValue PerformSTORECombine(SDNode *N,
9134 TargetLowering::DAGCombinerInfo &DCI) {
9135 StoreSDNode *St = cast<StoreSDNode>(N);
9136 if (St->isVolatile())
9139 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
9140 // pack all of the elements in one place. Next, store to memory in fewer
9142 SDValue StVal = St->getValue();
9143 EVT VT = StVal.getValueType();
9144 if (St->isTruncatingStore() && VT.isVector()) {
9145 SelectionDAG &DAG = DCI.DAG;
9146 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9147 EVT StVT = St->getMemoryVT();
9148 unsigned NumElems = VT.getVectorNumElements();
9149 assert(StVT != VT && "Cannot truncate to the same type");
9150 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
9151 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
9153 // From, To sizes and ElemCount must be pow of two
9154 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
9156 // We are going to use the original vector elt for storing.
9157 // Accumulated smaller vector elements must be a multiple of the store size.
9158 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
9160 unsigned SizeRatio = FromEltSz / ToEltSz;
9161 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
9163 // Create a type on which we perform the shuffle.
9164 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
9165 NumElems*SizeRatio);
9166 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
9169 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
9170 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
9171 for (unsigned i = 0; i < NumElems; ++i)
9172 ShuffleVec[i] = TLI.isBigEndian() ? (i+1) * SizeRatio - 1 : i * SizeRatio;
9174 // Can't shuffle using an illegal type.
9175 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
9177 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
9178 DAG.getUNDEF(WideVec.getValueType()),
9180 // At this point all of the data is stored at the bottom of the
9181 // register. We now need to save it to mem.
9183 // Find the largest store unit
9184 MVT StoreType = MVT::i8;
9185 for (MVT Tp : MVT::integer_valuetypes()) {
9186 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
9189 // Didn't find a legal store type.
9190 if (!TLI.isTypeLegal(StoreType))
9193 // Bitcast the original vector into a vector of store-size units
9194 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
9195 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
9196 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
9197 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
9198 SmallVector<SDValue, 8> Chains;
9199 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
9200 TLI.getPointerTy());
9201 SDValue BasePtr = St->getBasePtr();
9203 // Perform one or more big stores into memory.
9204 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
9205 for (unsigned I = 0; I < E; I++) {
9206 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
9207 StoreType, ShuffWide,
9208 DAG.getIntPtrConstant(I));
9209 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
9210 St->getPointerInfo(), St->isVolatile(),
9211 St->isNonTemporal(), St->getAlignment());
9212 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
9214 Chains.push_back(Ch);
9216 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
9219 if (!ISD::isNormalStore(St))
9222 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
9223 // ARM stores of arguments in the same cache line.
9224 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
9225 StVal.getNode()->hasOneUse()) {
9226 SelectionDAG &DAG = DCI.DAG;
9227 bool isBigEndian = DAG.getTargetLoweringInfo().isBigEndian();
9229 SDValue BasePtr = St->getBasePtr();
9230 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
9231 StVal.getNode()->getOperand(isBigEndian ? 1 : 0 ),
9232 BasePtr, St->getPointerInfo(), St->isVolatile(),
9233 St->isNonTemporal(), St->getAlignment());
9235 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9236 DAG.getConstant(4, MVT::i32));
9237 return DAG.getStore(NewST1.getValue(0), DL,
9238 StVal.getNode()->getOperand(isBigEndian ? 0 : 1),
9239 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
9240 St->isNonTemporal(),
9241 std::min(4U, St->getAlignment() / 2));
9244 if (StVal.getValueType() == MVT::i64 &&
9245 StVal.getNode()->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9247 // Bitcast an i64 store extracted from a vector to f64.
9248 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9249 SelectionDAG &DAG = DCI.DAG;
9251 SDValue IntVec = StVal.getOperand(0);
9252 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9253 IntVec.getValueType().getVectorNumElements());
9254 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
9255 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
9256 Vec, StVal.getOperand(1));
9258 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
9259 // Make the DAGCombiner fold the bitcasts.
9260 DCI.AddToWorklist(Vec.getNode());
9261 DCI.AddToWorklist(ExtElt.getNode());
9262 DCI.AddToWorklist(V.getNode());
9263 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
9264 St->getPointerInfo(), St->isVolatile(),
9265 St->isNonTemporal(), St->getAlignment(),
9272 // isConstVecPow2 - Return true if each vector element is a power of 2, all
9273 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
9274 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
9278 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
9280 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
9285 APFloat APF = C->getValueAPF();
9286 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
9287 != APFloat::opOK || !isExact)
9290 c0 = (I == 0) ? cN : c0;
9291 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
9298 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9299 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9300 /// when the VMUL has a constant operand that is a power of 2.
9302 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9303 /// vmul.f32 d16, d17, d16
9304 /// vcvt.s32.f32 d16, d16
9306 /// vcvt.s32.f32 d16, d16, #3
9307 static SDValue PerformVCVTCombine(SDNode *N,
9308 TargetLowering::DAGCombinerInfo &DCI,
9309 const ARMSubtarget *Subtarget) {
9310 SelectionDAG &DAG = DCI.DAG;
9311 SDValue Op = N->getOperand(0);
9313 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
9314 Op.getOpcode() != ISD::FMUL)
9318 SDValue N0 = Op->getOperand(0);
9319 SDValue ConstVec = Op->getOperand(1);
9320 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9322 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9323 !isConstVecPow2(ConstVec, isSigned, C))
9326 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9327 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9328 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9329 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32 ||
9331 // These instructions only exist converting from f32 to i32. We can handle
9332 // smaller integers by generating an extra truncate, but larger ones would
9333 // be lossy. We also can't handle more then 4 lanes, since these intructions
9334 // only support v2i32/v4i32 types.
9338 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9339 Intrinsic::arm_neon_vcvtfp2fxu;
9340 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9341 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9342 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
9343 DAG.getConstant(Log2_64(C), MVT::i32));
9345 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9346 FixConv = DAG.getNode(ISD::TRUNCATE, SDLoc(N), N->getValueType(0), FixConv);
9351 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9352 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9353 /// when the VDIV has a constant operand that is a power of 2.
9355 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9356 /// vcvt.f32.s32 d16, d16
9357 /// vdiv.f32 d16, d17, d16
9359 /// vcvt.f32.s32 d16, d16, #3
9360 static SDValue PerformVDIVCombine(SDNode *N,
9361 TargetLowering::DAGCombinerInfo &DCI,
9362 const ARMSubtarget *Subtarget) {
9363 SelectionDAG &DAG = DCI.DAG;
9364 SDValue Op = N->getOperand(0);
9365 unsigned OpOpcode = Op.getNode()->getOpcode();
9367 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
9368 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9372 SDValue ConstVec = N->getOperand(1);
9373 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9375 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9376 !isConstVecPow2(ConstVec, isSigned, C))
9379 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9380 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9381 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9382 // These instructions only exist converting from i32 to f32. We can handle
9383 // smaller integers by generating an extra extend, but larger ones would
9388 SDValue ConvInput = Op.getOperand(0);
9389 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9390 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9391 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9392 SDLoc(N), NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9395 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9396 Intrinsic::arm_neon_vcvtfxu2fp;
9397 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9399 DAG.getConstant(IntrinsicOpcode, MVT::i32),
9400 ConvInput, DAG.getConstant(Log2_64(C), MVT::i32));
9403 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9404 /// operand of a vector shift operation, where all the elements of the
9405 /// build_vector must have the same constant integer value.
9406 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9407 // Ignore bit_converts.
9408 while (Op.getOpcode() == ISD::BITCAST)
9409 Op = Op.getOperand(0);
9410 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9411 APInt SplatBits, SplatUndef;
9412 unsigned SplatBitSize;
9414 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9415 HasAnyUndefs, ElementBits) ||
9416 SplatBitSize > ElementBits)
9418 Cnt = SplatBits.getSExtValue();
9422 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9423 /// operand of a vector shift left operation. That value must be in the range:
9424 /// 0 <= Value < ElementBits for a left shift; or
9425 /// 0 <= Value <= ElementBits for a long left shift.
9426 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9427 assert(VT.isVector() && "vector shift count is not a vector type");
9428 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9429 if (! getVShiftImm(Op, ElementBits, Cnt))
9431 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9434 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9435 /// operand of a vector shift right operation. For a shift opcode, the value
9436 /// is positive, but for an intrinsic the value count must be negative. The
9437 /// absolute value must be in the range:
9438 /// 1 <= |Value| <= ElementBits for a right shift; or
9439 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9440 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9442 assert(VT.isVector() && "vector shift count is not a vector type");
9443 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9444 if (! getVShiftImm(Op, ElementBits, Cnt))
9448 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9451 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9452 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9453 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9456 // Don't do anything for most intrinsics.
9459 // Vector shifts: check for immediate versions and lower them.
9460 // Note: This is done during DAG combining instead of DAG legalizing because
9461 // the build_vectors for 64-bit vector element shift counts are generally
9462 // not legal, and it is hard to see their values after they get legalized to
9463 // loads from a constant pool.
9464 case Intrinsic::arm_neon_vshifts:
9465 case Intrinsic::arm_neon_vshiftu:
9466 case Intrinsic::arm_neon_vrshifts:
9467 case Intrinsic::arm_neon_vrshiftu:
9468 case Intrinsic::arm_neon_vrshiftn:
9469 case Intrinsic::arm_neon_vqshifts:
9470 case Intrinsic::arm_neon_vqshiftu:
9471 case Intrinsic::arm_neon_vqshiftsu:
9472 case Intrinsic::arm_neon_vqshiftns:
9473 case Intrinsic::arm_neon_vqshiftnu:
9474 case Intrinsic::arm_neon_vqshiftnsu:
9475 case Intrinsic::arm_neon_vqrshiftns:
9476 case Intrinsic::arm_neon_vqrshiftnu:
9477 case Intrinsic::arm_neon_vqrshiftnsu: {
9478 EVT VT = N->getOperand(1).getValueType();
9480 unsigned VShiftOpc = 0;
9483 case Intrinsic::arm_neon_vshifts:
9484 case Intrinsic::arm_neon_vshiftu:
9485 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9486 VShiftOpc = ARMISD::VSHL;
9489 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9490 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9491 ARMISD::VSHRs : ARMISD::VSHRu);
9496 case Intrinsic::arm_neon_vrshifts:
9497 case Intrinsic::arm_neon_vrshiftu:
9498 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9502 case Intrinsic::arm_neon_vqshifts:
9503 case Intrinsic::arm_neon_vqshiftu:
9504 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9508 case Intrinsic::arm_neon_vqshiftsu:
9509 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9511 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9513 case Intrinsic::arm_neon_vrshiftn:
9514 case Intrinsic::arm_neon_vqshiftns:
9515 case Intrinsic::arm_neon_vqshiftnu:
9516 case Intrinsic::arm_neon_vqshiftnsu:
9517 case Intrinsic::arm_neon_vqrshiftns:
9518 case Intrinsic::arm_neon_vqrshiftnu:
9519 case Intrinsic::arm_neon_vqrshiftnsu:
9520 // Narrowing shifts require an immediate right shift.
9521 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9523 llvm_unreachable("invalid shift count for narrowing vector shift "
9527 llvm_unreachable("unhandled vector shift");
9531 case Intrinsic::arm_neon_vshifts:
9532 case Intrinsic::arm_neon_vshiftu:
9533 // Opcode already set above.
9535 case Intrinsic::arm_neon_vrshifts:
9536 VShiftOpc = ARMISD::VRSHRs; break;
9537 case Intrinsic::arm_neon_vrshiftu:
9538 VShiftOpc = ARMISD::VRSHRu; break;
9539 case Intrinsic::arm_neon_vrshiftn:
9540 VShiftOpc = ARMISD::VRSHRN; break;
9541 case Intrinsic::arm_neon_vqshifts:
9542 VShiftOpc = ARMISD::VQSHLs; break;
9543 case Intrinsic::arm_neon_vqshiftu:
9544 VShiftOpc = ARMISD::VQSHLu; break;
9545 case Intrinsic::arm_neon_vqshiftsu:
9546 VShiftOpc = ARMISD::VQSHLsu; break;
9547 case Intrinsic::arm_neon_vqshiftns:
9548 VShiftOpc = ARMISD::VQSHRNs; break;
9549 case Intrinsic::arm_neon_vqshiftnu:
9550 VShiftOpc = ARMISD::VQSHRNu; break;
9551 case Intrinsic::arm_neon_vqshiftnsu:
9552 VShiftOpc = ARMISD::VQSHRNsu; break;
9553 case Intrinsic::arm_neon_vqrshiftns:
9554 VShiftOpc = ARMISD::VQRSHRNs; break;
9555 case Intrinsic::arm_neon_vqrshiftnu:
9556 VShiftOpc = ARMISD::VQRSHRNu; break;
9557 case Intrinsic::arm_neon_vqrshiftnsu:
9558 VShiftOpc = ARMISD::VQRSHRNsu; break;
9561 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9562 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
9565 case Intrinsic::arm_neon_vshiftins: {
9566 EVT VT = N->getOperand(1).getValueType();
9568 unsigned VShiftOpc = 0;
9570 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9571 VShiftOpc = ARMISD::VSLI;
9572 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9573 VShiftOpc = ARMISD::VSRI;
9575 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9578 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9579 N->getOperand(1), N->getOperand(2),
9580 DAG.getConstant(Cnt, MVT::i32));
9583 case Intrinsic::arm_neon_vqrshifts:
9584 case Intrinsic::arm_neon_vqrshiftu:
9585 // No immediate versions of these to check for.
9592 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9593 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9594 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9595 /// vector element shift counts are generally not legal, and it is hard to see
9596 /// their values after they get legalized to loads from a constant pool.
9597 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9598 const ARMSubtarget *ST) {
9599 EVT VT = N->getValueType(0);
9600 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9601 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9602 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9603 SDValue N1 = N->getOperand(1);
9604 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9605 SDValue N0 = N->getOperand(0);
9606 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9607 DAG.MaskedValueIsZero(N0.getOperand(0),
9608 APInt::getHighBitsSet(32, 16)))
9609 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
9613 // Nothing to be done for scalar shifts.
9614 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9615 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9618 assert(ST->hasNEON() && "unexpected vector shift");
9621 switch (N->getOpcode()) {
9622 default: llvm_unreachable("unexpected shift opcode");
9625 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
9626 return DAG.getNode(ARMISD::VSHL, SDLoc(N), VT, N->getOperand(0),
9627 DAG.getConstant(Cnt, MVT::i32));
9632 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9633 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9634 ARMISD::VSHRs : ARMISD::VSHRu);
9635 return DAG.getNode(VShiftOpc, SDLoc(N), VT, N->getOperand(0),
9636 DAG.getConstant(Cnt, MVT::i32));
9642 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9643 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9644 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9645 const ARMSubtarget *ST) {
9646 SDValue N0 = N->getOperand(0);
9648 // Check for sign- and zero-extensions of vector extract operations of 8-
9649 // and 16-bit vector elements. NEON supports these directly. They are
9650 // handled during DAG combining because type legalization will promote them
9651 // to 32-bit types and it is messy to recognize the operations after that.
9652 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9653 SDValue Vec = N0.getOperand(0);
9654 SDValue Lane = N0.getOperand(1);
9655 EVT VT = N->getValueType(0);
9656 EVT EltVT = N0.getValueType();
9657 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9659 if (VT == MVT::i32 &&
9660 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9661 TLI.isTypeLegal(Vec.getValueType()) &&
9662 isa<ConstantSDNode>(Lane)) {
9665 switch (N->getOpcode()) {
9666 default: llvm_unreachable("unexpected opcode");
9667 case ISD::SIGN_EXTEND:
9668 Opc = ARMISD::VGETLANEs;
9670 case ISD::ZERO_EXTEND:
9671 case ISD::ANY_EXTEND:
9672 Opc = ARMISD::VGETLANEu;
9675 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
9682 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9683 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9684 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9685 const ARMSubtarget *ST) {
9686 // If the target supports NEON, try to use vmax/vmin instructions for f32
9687 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9688 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9689 // a NaN; only do the transformation when it matches that behavior.
9691 // For now only do this when using NEON for FP operations; if using VFP, it
9692 // is not obvious that the benefit outweighs the cost of switching to the
9694 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9695 N->getValueType(0) != MVT::f32)
9698 SDValue CondLHS = N->getOperand(0);
9699 SDValue CondRHS = N->getOperand(1);
9700 SDValue LHS = N->getOperand(2);
9701 SDValue RHS = N->getOperand(3);
9702 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9704 unsigned Opcode = 0;
9706 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9707 IsReversed = false; // x CC y ? x : y
9708 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9709 IsReversed = true ; // x CC y ? y : x
9723 // If LHS is NaN, an ordered comparison will be false and the result will
9724 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9725 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9726 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9727 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9729 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9730 // will return -0, so vmin can only be used for unsafe math or if one of
9731 // the operands is known to be nonzero.
9732 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9733 !DAG.getTarget().Options.UnsafeFPMath &&
9734 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9736 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9745 // If LHS is NaN, an ordered comparison will be false and the result will
9746 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9747 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9748 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9749 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9751 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9752 // will return +0, so vmax can only be used for unsafe math or if one of
9753 // the operands is known to be nonzero.
9754 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9755 !DAG.getTarget().Options.UnsafeFPMath &&
9756 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9758 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9764 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
9767 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9769 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9770 SDValue Cmp = N->getOperand(4);
9771 if (Cmp.getOpcode() != ARMISD::CMPZ)
9772 // Only looking at EQ and NE cases.
9775 EVT VT = N->getValueType(0);
9777 SDValue LHS = Cmp.getOperand(0);
9778 SDValue RHS = Cmp.getOperand(1);
9779 SDValue FalseVal = N->getOperand(0);
9780 SDValue TrueVal = N->getOperand(1);
9781 SDValue ARMcc = N->getOperand(2);
9782 ARMCC::CondCodes CC =
9783 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9801 /// FIXME: Turn this into a target neutral optimization?
9803 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9804 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9805 N->getOperand(3), Cmp);
9806 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9808 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9809 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9810 N->getOperand(3), NewCmp);
9813 if (Res.getNode()) {
9814 APInt KnownZero, KnownOne;
9815 DAG.computeKnownBits(SDValue(N,0), KnownZero, KnownOne);
9816 // Capture demanded bits information that would be otherwise lost.
9817 if (KnownZero == 0xfffffffe)
9818 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9819 DAG.getValueType(MVT::i1));
9820 else if (KnownZero == 0xffffff00)
9821 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9822 DAG.getValueType(MVT::i8));
9823 else if (KnownZero == 0xffff0000)
9824 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9825 DAG.getValueType(MVT::i16));
9831 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9832 DAGCombinerInfo &DCI) const {
9833 switch (N->getOpcode()) {
9835 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9836 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9837 case ISD::SUB: return PerformSUBCombine(N, DCI);
9838 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9839 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9840 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9841 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9842 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9843 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI, Subtarget);
9844 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9845 case ISD::STORE: return PerformSTORECombine(N, DCI);
9846 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI, Subtarget);
9847 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9848 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9849 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9850 case ISD::FP_TO_SINT:
9851 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9852 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9853 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9856 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9857 case ISD::SIGN_EXTEND:
9858 case ISD::ZERO_EXTEND:
9859 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9860 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9861 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9862 case ARMISD::VLD2DUP:
9863 case ARMISD::VLD3DUP:
9864 case ARMISD::VLD4DUP:
9865 return CombineBaseUpdate(N, DCI);
9866 case ARMISD::BUILD_VECTOR:
9867 return PerformARMBUILD_VECTORCombine(N, DCI);
9868 case ISD::INTRINSIC_VOID:
9869 case ISD::INTRINSIC_W_CHAIN:
9870 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9871 case Intrinsic::arm_neon_vld1:
9872 case Intrinsic::arm_neon_vld2:
9873 case Intrinsic::arm_neon_vld3:
9874 case Intrinsic::arm_neon_vld4:
9875 case Intrinsic::arm_neon_vld2lane:
9876 case Intrinsic::arm_neon_vld3lane:
9877 case Intrinsic::arm_neon_vld4lane:
9878 case Intrinsic::arm_neon_vst1:
9879 case Intrinsic::arm_neon_vst2:
9880 case Intrinsic::arm_neon_vst3:
9881 case Intrinsic::arm_neon_vst4:
9882 case Intrinsic::arm_neon_vst2lane:
9883 case Intrinsic::arm_neon_vst3lane:
9884 case Intrinsic::arm_neon_vst4lane:
9885 return CombineBaseUpdate(N, DCI);
9893 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9895 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9898 bool ARMTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
9902 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9903 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9905 switch (VT.getSimpleVT().SimpleTy) {
9911 // Unaligned access can use (for example) LRDB, LRDH, LDR
9912 if (AllowsUnaligned) {
9914 *Fast = Subtarget->hasV7Ops();
9921 // For any little-endian targets with neon, we can support unaligned ld/st
9922 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9923 // A big-endian target may also explicitly support unaligned accesses
9924 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
9934 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
9935 unsigned AlignCheck) {
9936 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
9937 (DstAlign == 0 || DstAlign % AlignCheck == 0));
9940 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
9941 unsigned DstAlign, unsigned SrcAlign,
9942 bool IsMemset, bool ZeroMemset,
9944 MachineFunction &MF) const {
9945 const Function *F = MF.getFunction();
9947 // See if we can use NEON instructions for this...
9948 if ((!IsMemset || ZeroMemset) &&
9949 Subtarget->hasNEON() &&
9950 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
9951 Attribute::NoImplicitFloat)) {
9954 (memOpAlign(SrcAlign, DstAlign, 16) ||
9955 (allowsMisalignedMemoryAccesses(MVT::v2f64, 0, 1, &Fast) && Fast))) {
9957 } else if (Size >= 8 &&
9958 (memOpAlign(SrcAlign, DstAlign, 8) ||
9959 (allowsMisalignedMemoryAccesses(MVT::f64, 0, 1, &Fast) &&
9965 // Lowering to i32/i16 if the size permits.
9971 // Let the target-independent logic figure it out.
9975 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
9976 if (Val.getOpcode() != ISD::LOAD)
9979 EVT VT1 = Val.getValueType();
9980 if (!VT1.isSimple() || !VT1.isInteger() ||
9981 !VT2.isSimple() || !VT2.isInteger())
9984 switch (VT1.getSimpleVT().SimpleTy) {
9989 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
9996 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
9997 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10000 if (!isTypeLegal(EVT::getEVT(Ty1)))
10003 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
10005 // Assuming the caller doesn't have a zeroext or signext return parameter,
10006 // truncation all the way down to i1 is valid.
10011 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
10015 unsigned Scale = 1;
10016 switch (VT.getSimpleVT().SimpleTy) {
10017 default: return false;
10032 if ((V & (Scale - 1)) != 0)
10035 return V == (V & ((1LL << 5) - 1));
10038 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10039 const ARMSubtarget *Subtarget) {
10040 bool isNeg = false;
10046 switch (VT.getSimpleVT().SimpleTy) {
10047 default: return false;
10052 // + imm12 or - imm8
10054 return V == (V & ((1LL << 8) - 1));
10055 return V == (V & ((1LL << 12) - 1));
10058 // Same as ARM mode. FIXME: NEON?
10059 if (!Subtarget->hasVFP2())
10064 return V == (V & ((1LL << 8) - 1));
10068 /// isLegalAddressImmediate - Return true if the integer value can be used
10069 /// as the offset of the target addressing mode for load / store of the
10071 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10072 const ARMSubtarget *Subtarget) {
10076 if (!VT.isSimple())
10079 if (Subtarget->isThumb1Only())
10080 return isLegalT1AddressImmediate(V, VT);
10081 else if (Subtarget->isThumb2())
10082 return isLegalT2AddressImmediate(V, VT, Subtarget);
10087 switch (VT.getSimpleVT().SimpleTy) {
10088 default: return false;
10093 return V == (V & ((1LL << 12) - 1));
10096 return V == (V & ((1LL << 8) - 1));
10099 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10104 return V == (V & ((1LL << 8) - 1));
10108 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10110 int Scale = AM.Scale;
10114 switch (VT.getSimpleVT().SimpleTy) {
10115 default: return false;
10123 Scale = Scale & ~1;
10124 return Scale == 2 || Scale == 4 || Scale == 8;
10127 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10131 // Note, we allow "void" uses (basically, uses that aren't loads or
10132 // stores), because arm allows folding a scale into many arithmetic
10133 // operations. This should be made more precise and revisited later.
10135 // Allow r << imm, but the imm has to be a multiple of two.
10136 if (Scale & 1) return false;
10137 return isPowerOf2_32(Scale);
10141 /// isLegalAddressingMode - Return true if the addressing mode represented
10142 /// by AM is legal for this target, for a load/store of the specified type.
10143 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
10145 EVT VT = getValueType(Ty, true);
10146 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10149 // Can never fold addr of global into load/store.
10153 switch (AM.Scale) {
10154 case 0: // no scale reg, must be "r+i" or "r", or "i".
10157 if (Subtarget->isThumb1Only())
10161 // ARM doesn't support any R+R*scale+imm addr modes.
10165 if (!VT.isSimple())
10168 if (Subtarget->isThumb2())
10169 return isLegalT2ScaledAddressingMode(AM, VT);
10171 int Scale = AM.Scale;
10172 switch (VT.getSimpleVT().SimpleTy) {
10173 default: return false;
10177 if (Scale < 0) Scale = -Scale;
10181 return isPowerOf2_32(Scale & ~1);
10185 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10190 // Note, we allow "void" uses (basically, uses that aren't loads or
10191 // stores), because arm allows folding a scale into many arithmetic
10192 // operations. This should be made more precise and revisited later.
10194 // Allow r << imm, but the imm has to be a multiple of two.
10195 if (Scale & 1) return false;
10196 return isPowerOf2_32(Scale);
10202 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10203 /// icmp immediate, that is the target has icmp instructions which can compare
10204 /// a register against the immediate without having to materialize the
10205 /// immediate into a register.
10206 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10207 // Thumb2 and ARM modes can use cmn for negative immediates.
10208 if (!Subtarget->isThumb())
10209 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
10210 if (Subtarget->isThumb2())
10211 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
10212 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10213 return Imm >= 0 && Imm <= 255;
10216 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10217 /// *or sub* immediate, that is the target has add or sub instructions which can
10218 /// add a register with the immediate without having to materialize the
10219 /// immediate into a register.
10220 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10221 // Same encoding for add/sub, just flip the sign.
10222 int64_t AbsImm = llvm::abs64(Imm);
10223 if (!Subtarget->isThumb())
10224 return ARM_AM::getSOImmVal(AbsImm) != -1;
10225 if (Subtarget->isThumb2())
10226 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10227 // Thumb1 only has 8-bit unsigned immediate.
10228 return AbsImm >= 0 && AbsImm <= 255;
10231 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10232 bool isSEXTLoad, SDValue &Base,
10233 SDValue &Offset, bool &isInc,
10234 SelectionDAG &DAG) {
10235 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10238 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10239 // AddressingMode 3
10240 Base = Ptr->getOperand(0);
10241 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10242 int RHSC = (int)RHS->getZExtValue();
10243 if (RHSC < 0 && RHSC > -256) {
10244 assert(Ptr->getOpcode() == ISD::ADD);
10246 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10250 isInc = (Ptr->getOpcode() == ISD::ADD);
10251 Offset = Ptr->getOperand(1);
10253 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10254 // AddressingMode 2
10255 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10256 int RHSC = (int)RHS->getZExtValue();
10257 if (RHSC < 0 && RHSC > -0x1000) {
10258 assert(Ptr->getOpcode() == ISD::ADD);
10260 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10261 Base = Ptr->getOperand(0);
10266 if (Ptr->getOpcode() == ISD::ADD) {
10268 ARM_AM::ShiftOpc ShOpcVal=
10269 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10270 if (ShOpcVal != ARM_AM::no_shift) {
10271 Base = Ptr->getOperand(1);
10272 Offset = Ptr->getOperand(0);
10274 Base = Ptr->getOperand(0);
10275 Offset = Ptr->getOperand(1);
10280 isInc = (Ptr->getOpcode() == ISD::ADD);
10281 Base = Ptr->getOperand(0);
10282 Offset = Ptr->getOperand(1);
10286 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10290 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10291 bool isSEXTLoad, SDValue &Base,
10292 SDValue &Offset, bool &isInc,
10293 SelectionDAG &DAG) {
10294 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10297 Base = Ptr->getOperand(0);
10298 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10299 int RHSC = (int)RHS->getZExtValue();
10300 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10301 assert(Ptr->getOpcode() == ISD::ADD);
10303 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10305 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10306 isInc = Ptr->getOpcode() == ISD::ADD;
10307 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
10315 /// getPreIndexedAddressParts - returns true by value, base pointer and
10316 /// offset pointer and addressing mode by reference if the node's address
10317 /// can be legally represented as pre-indexed load / store address.
10319 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10321 ISD::MemIndexedMode &AM,
10322 SelectionDAG &DAG) const {
10323 if (Subtarget->isThumb1Only())
10328 bool isSEXTLoad = false;
10329 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10330 Ptr = LD->getBasePtr();
10331 VT = LD->getMemoryVT();
10332 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10333 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10334 Ptr = ST->getBasePtr();
10335 VT = ST->getMemoryVT();
10340 bool isLegal = false;
10341 if (Subtarget->isThumb2())
10342 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10343 Offset, isInc, DAG);
10345 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10346 Offset, isInc, DAG);
10350 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10354 /// getPostIndexedAddressParts - returns true by value, base pointer and
10355 /// offset pointer and addressing mode by reference if this node can be
10356 /// combined with a load / store to form a post-indexed load / store.
10357 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10360 ISD::MemIndexedMode &AM,
10361 SelectionDAG &DAG) const {
10362 if (Subtarget->isThumb1Only())
10367 bool isSEXTLoad = false;
10368 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10369 VT = LD->getMemoryVT();
10370 Ptr = LD->getBasePtr();
10371 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10372 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10373 VT = ST->getMemoryVT();
10374 Ptr = ST->getBasePtr();
10379 bool isLegal = false;
10380 if (Subtarget->isThumb2())
10381 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10384 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10390 // Swap base ptr and offset to catch more post-index load / store when
10391 // it's legal. In Thumb2 mode, offset must be an immediate.
10392 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10393 !Subtarget->isThumb2())
10394 std::swap(Base, Offset);
10396 // Post-indexed load / store update the base pointer.
10401 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10405 void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
10408 const SelectionDAG &DAG,
10409 unsigned Depth) const {
10410 unsigned BitWidth = KnownOne.getBitWidth();
10411 KnownZero = KnownOne = APInt(BitWidth, 0);
10412 switch (Op.getOpcode()) {
10418 // These nodes' second result is a boolean
10419 if (Op.getResNo() == 0)
10421 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10423 case ARMISD::CMOV: {
10424 // Bits are known zero/one if known on the LHS and RHS.
10425 DAG.computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10426 if (KnownZero == 0 && KnownOne == 0) return;
10428 APInt KnownZeroRHS, KnownOneRHS;
10429 DAG.computeKnownBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10430 KnownZero &= KnownZeroRHS;
10431 KnownOne &= KnownOneRHS;
10434 case ISD::INTRINSIC_W_CHAIN: {
10435 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
10436 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
10439 case Intrinsic::arm_ldaex:
10440 case Intrinsic::arm_ldrex: {
10441 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
10442 unsigned MemBits = VT.getScalarType().getSizeInBits();
10443 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
10451 //===----------------------------------------------------------------------===//
10452 // ARM Inline Assembly Support
10453 //===----------------------------------------------------------------------===//
10455 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10456 // Looking for "rev" which is V6+.
10457 if (!Subtarget->hasV6Ops())
10460 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10461 std::string AsmStr = IA->getAsmString();
10462 SmallVector<StringRef, 4> AsmPieces;
10463 SplitString(AsmStr, AsmPieces, ";\n");
10465 switch (AsmPieces.size()) {
10466 default: return false;
10468 AsmStr = AsmPieces[0];
10470 SplitString(AsmStr, AsmPieces, " \t,");
10473 if (AsmPieces.size() == 3 &&
10474 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10475 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10476 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10477 if (Ty && Ty->getBitWidth() == 32)
10478 return IntrinsicLowering::LowerToByteSwap(CI);
10486 /// getConstraintType - Given a constraint letter, return the type of
10487 /// constraint it is for this target.
10488 ARMTargetLowering::ConstraintType
10489 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10490 if (Constraint.size() == 1) {
10491 switch (Constraint[0]) {
10493 case 'l': return C_RegisterClass;
10494 case 'w': return C_RegisterClass;
10495 case 'h': return C_RegisterClass;
10496 case 'x': return C_RegisterClass;
10497 case 't': return C_RegisterClass;
10498 case 'j': return C_Other; // Constant for movw.
10499 // An address with a single base register. Due to the way we
10500 // currently handle addresses it is the same as an 'r' memory constraint.
10501 case 'Q': return C_Memory;
10503 } else if (Constraint.size() == 2) {
10504 switch (Constraint[0]) {
10506 // All 'U+' constraints are addresses.
10507 case 'U': return C_Memory;
10510 return TargetLowering::getConstraintType(Constraint);
10513 /// Examine constraint type and operand type and determine a weight value.
10514 /// This object must already have been set up with the operand type
10515 /// and the current alternative constraint selected.
10516 TargetLowering::ConstraintWeight
10517 ARMTargetLowering::getSingleConstraintMatchWeight(
10518 AsmOperandInfo &info, const char *constraint) const {
10519 ConstraintWeight weight = CW_Invalid;
10520 Value *CallOperandVal = info.CallOperandVal;
10521 // If we don't have a value, we can't do a match,
10522 // but allow it at the lowest weight.
10523 if (!CallOperandVal)
10525 Type *type = CallOperandVal->getType();
10526 // Look at the constraint type.
10527 switch (*constraint) {
10529 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10532 if (type->isIntegerTy()) {
10533 if (Subtarget->isThumb())
10534 weight = CW_SpecificReg;
10536 weight = CW_Register;
10540 if (type->isFloatingPointTy())
10541 weight = CW_Register;
10547 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10549 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10551 if (Constraint.size() == 1) {
10552 // GCC ARM Constraint Letters
10553 switch (Constraint[0]) {
10554 case 'l': // Low regs or general regs.
10555 if (Subtarget->isThumb())
10556 return RCPair(0U, &ARM::tGPRRegClass);
10557 return RCPair(0U, &ARM::GPRRegClass);
10558 case 'h': // High regs or no regs.
10559 if (Subtarget->isThumb())
10560 return RCPair(0U, &ARM::hGPRRegClass);
10563 if (Subtarget->isThumb1Only())
10564 return RCPair(0U, &ARM::tGPRRegClass);
10565 return RCPair(0U, &ARM::GPRRegClass);
10567 if (VT == MVT::Other)
10569 if (VT == MVT::f32)
10570 return RCPair(0U, &ARM::SPRRegClass);
10571 if (VT.getSizeInBits() == 64)
10572 return RCPair(0U, &ARM::DPRRegClass);
10573 if (VT.getSizeInBits() == 128)
10574 return RCPair(0U, &ARM::QPRRegClass);
10577 if (VT == MVT::Other)
10579 if (VT == MVT::f32)
10580 return RCPair(0U, &ARM::SPR_8RegClass);
10581 if (VT.getSizeInBits() == 64)
10582 return RCPair(0U, &ARM::DPR_8RegClass);
10583 if (VT.getSizeInBits() == 128)
10584 return RCPair(0U, &ARM::QPR_8RegClass);
10587 if (VT == MVT::f32)
10588 return RCPair(0U, &ARM::SPRRegClass);
10592 if (StringRef("{cc}").equals_lower(Constraint))
10593 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10595 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10598 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10599 /// vector. If it is invalid, don't add anything to Ops.
10600 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10601 std::string &Constraint,
10602 std::vector<SDValue>&Ops,
10603 SelectionDAG &DAG) const {
10606 // Currently only support length 1 constraints.
10607 if (Constraint.length() != 1) return;
10609 char ConstraintLetter = Constraint[0];
10610 switch (ConstraintLetter) {
10613 case 'I': case 'J': case 'K': case 'L':
10614 case 'M': case 'N': case 'O':
10615 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10619 int64_t CVal64 = C->getSExtValue();
10620 int CVal = (int) CVal64;
10621 // None of these constraints allow values larger than 32 bits. Check
10622 // that the value fits in an int.
10623 if (CVal != CVal64)
10626 switch (ConstraintLetter) {
10628 // Constant suitable for movw, must be between 0 and
10630 if (Subtarget->hasV6T2Ops())
10631 if (CVal >= 0 && CVal <= 65535)
10635 if (Subtarget->isThumb1Only()) {
10636 // This must be a constant between 0 and 255, for ADD
10638 if (CVal >= 0 && CVal <= 255)
10640 } else if (Subtarget->isThumb2()) {
10641 // A constant that can be used as an immediate value in a
10642 // data-processing instruction.
10643 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10646 // A constant that can be used as an immediate value in a
10647 // data-processing instruction.
10648 if (ARM_AM::getSOImmVal(CVal) != -1)
10654 if (Subtarget->isThumb()) { // FIXME thumb2
10655 // This must be a constant between -255 and -1, for negated ADD
10656 // immediates. This can be used in GCC with an "n" modifier that
10657 // prints the negated value, for use with SUB instructions. It is
10658 // not useful otherwise but is implemented for compatibility.
10659 if (CVal >= -255 && CVal <= -1)
10662 // This must be a constant between -4095 and 4095. It is not clear
10663 // what this constraint is intended for. Implemented for
10664 // compatibility with GCC.
10665 if (CVal >= -4095 && CVal <= 4095)
10671 if (Subtarget->isThumb1Only()) {
10672 // A 32-bit value where only one byte has a nonzero value. Exclude
10673 // zero to match GCC. This constraint is used by GCC internally for
10674 // constants that can be loaded with a move/shift combination.
10675 // It is not useful otherwise but is implemented for compatibility.
10676 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10678 } else if (Subtarget->isThumb2()) {
10679 // A constant whose bitwise inverse can be used as an immediate
10680 // value in a data-processing instruction. This can be used in GCC
10681 // with a "B" modifier that prints the inverted value, for use with
10682 // BIC and MVN instructions. It is not useful otherwise but is
10683 // implemented for compatibility.
10684 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10687 // A constant whose bitwise inverse can be used as an immediate
10688 // value in a data-processing instruction. This can be used in GCC
10689 // with a "B" modifier that prints the inverted value, for use with
10690 // BIC and MVN instructions. It is not useful otherwise but is
10691 // implemented for compatibility.
10692 if (ARM_AM::getSOImmVal(~CVal) != -1)
10698 if (Subtarget->isThumb1Only()) {
10699 // This must be a constant between -7 and 7,
10700 // for 3-operand ADD/SUB immediate instructions.
10701 if (CVal >= -7 && CVal < 7)
10703 } else if (Subtarget->isThumb2()) {
10704 // A constant whose negation can be used as an immediate value in a
10705 // data-processing instruction. This can be used in GCC with an "n"
10706 // modifier that prints the negated value, for use with SUB
10707 // instructions. It is not useful otherwise but is implemented for
10709 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10712 // A constant whose negation can be used as an immediate value in a
10713 // data-processing instruction. This can be used in GCC with an "n"
10714 // modifier that prints the negated value, for use with SUB
10715 // instructions. It is not useful otherwise but is implemented for
10717 if (ARM_AM::getSOImmVal(-CVal) != -1)
10723 if (Subtarget->isThumb()) { // FIXME thumb2
10724 // This must be a multiple of 4 between 0 and 1020, for
10725 // ADD sp + immediate.
10726 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10729 // A power of two or a constant between 0 and 32. This is used in
10730 // GCC for the shift amount on shifted register operands, but it is
10731 // useful in general for any shift amounts.
10732 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10738 if (Subtarget->isThumb()) { // FIXME thumb2
10739 // This must be a constant between 0 and 31, for shift amounts.
10740 if (CVal >= 0 && CVal <= 31)
10746 if (Subtarget->isThumb()) { // FIXME thumb2
10747 // This must be a multiple of 4 between -508 and 508, for
10748 // ADD/SUB sp = sp + immediate.
10749 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10754 Result = DAG.getTargetConstant(CVal, Op.getValueType());
10758 if (Result.getNode()) {
10759 Ops.push_back(Result);
10762 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10765 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
10766 assert(Subtarget->isTargetAEABI() && "Register-based DivRem lowering only");
10767 unsigned Opcode = Op->getOpcode();
10768 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
10769 "Invalid opcode for Div/Rem lowering");
10770 bool isSigned = (Opcode == ISD::SDIVREM);
10771 EVT VT = Op->getValueType(0);
10772 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
10775 switch (VT.getSimpleVT().SimpleTy) {
10776 default: llvm_unreachable("Unexpected request for libcall!");
10777 case MVT::i8: LC = isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
10778 case MVT::i16: LC = isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
10779 case MVT::i32: LC = isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
10780 case MVT::i64: LC = isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
10783 SDValue InChain = DAG.getEntryNode();
10785 TargetLowering::ArgListTy Args;
10786 TargetLowering::ArgListEntry Entry;
10787 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
10788 EVT ArgVT = Op->getOperand(i).getValueType();
10789 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
10790 Entry.Node = Op->getOperand(i);
10792 Entry.isSExt = isSigned;
10793 Entry.isZExt = !isSigned;
10794 Args.push_back(Entry);
10797 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
10800 Type *RetTy = (Type*)StructType::get(Ty, Ty, nullptr);
10803 TargetLowering::CallLoweringInfo CLI(DAG);
10804 CLI.setDebugLoc(dl).setChain(InChain)
10805 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
10806 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
10808 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
10809 return CallInfo.first;
10813 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
10814 assert(Subtarget->isTargetWindows() && "unsupported target platform");
10818 SDValue Chain = Op.getOperand(0);
10819 SDValue Size = Op.getOperand(1);
10821 SDValue Words = DAG.getNode(ISD::SRL, DL, MVT::i32, Size,
10822 DAG.getConstant(2, MVT::i32));
10825 Chain = DAG.getCopyToReg(Chain, DL, ARM::R4, Words, Flag);
10826 Flag = Chain.getValue(1);
10828 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
10829 Chain = DAG.getNode(ARMISD::WIN__CHKSTK, DL, NodeTys, Chain, Flag);
10831 SDValue NewSP = DAG.getCopyFromReg(Chain, DL, ARM::SP, MVT::i32);
10832 Chain = NewSP.getValue(1);
10834 SDValue Ops[2] = { NewSP, Chain };
10835 return DAG.getMergeValues(Ops, DL);
10838 SDValue ARMTargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
10839 assert(Op.getValueType() == MVT::f64 && Subtarget->isFPOnlySP() &&
10840 "Unexpected type for custom-lowering FP_EXTEND");
10843 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
10845 SDValue SrcVal = Op.getOperand(0);
10846 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
10847 /*isSigned*/ false, SDLoc(Op)).first;
10850 SDValue ARMTargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
10851 assert(Op.getOperand(0).getValueType() == MVT::f64 &&
10852 Subtarget->isFPOnlySP() &&
10853 "Unexpected type for custom-lowering FP_ROUND");
10856 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
10858 SDValue SrcVal = Op.getOperand(0);
10859 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
10860 /*isSigned*/ false, SDLoc(Op)).first;
10864 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
10865 // The ARM target isn't yet aware of offsets.
10869 bool ARM::isBitFieldInvertedMask(unsigned v) {
10870 if (v == 0xffffffff)
10873 // there can be 1's on either or both "outsides", all the "inside"
10874 // bits must be 0's
10875 unsigned TO = CountTrailingOnes_32(v);
10876 unsigned LO = CountLeadingOnes_32(v);
10877 v = (v >> TO) << TO;
10878 v = (v << LO) >> LO;
10882 /// isFPImmLegal - Returns true if the target can instruction select the
10883 /// specified FP immediate natively. If false, the legalizer will
10884 /// materialize the FP immediate as a load from a constant pool.
10885 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
10886 if (!Subtarget->hasVFP3())
10888 if (VT == MVT::f32)
10889 return ARM_AM::getFP32Imm(Imm) != -1;
10890 if (VT == MVT::f64 && !Subtarget->isFPOnlySP())
10891 return ARM_AM::getFP64Imm(Imm) != -1;
10895 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
10896 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
10897 /// specified in the intrinsic calls.
10898 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
10900 unsigned Intrinsic) const {
10901 switch (Intrinsic) {
10902 case Intrinsic::arm_neon_vld1:
10903 case Intrinsic::arm_neon_vld2:
10904 case Intrinsic::arm_neon_vld3:
10905 case Intrinsic::arm_neon_vld4:
10906 case Intrinsic::arm_neon_vld2lane:
10907 case Intrinsic::arm_neon_vld3lane:
10908 case Intrinsic::arm_neon_vld4lane: {
10909 Info.opc = ISD::INTRINSIC_W_CHAIN;
10910 // Conservatively set memVT to the entire set of vectors loaded.
10911 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
10912 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10913 Info.ptrVal = I.getArgOperand(0);
10915 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10916 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10917 Info.vol = false; // volatile loads with NEON intrinsics not supported
10918 Info.readMem = true;
10919 Info.writeMem = false;
10922 case Intrinsic::arm_neon_vst1:
10923 case Intrinsic::arm_neon_vst2:
10924 case Intrinsic::arm_neon_vst3:
10925 case Intrinsic::arm_neon_vst4:
10926 case Intrinsic::arm_neon_vst2lane:
10927 case Intrinsic::arm_neon_vst3lane:
10928 case Intrinsic::arm_neon_vst4lane: {
10929 Info.opc = ISD::INTRINSIC_VOID;
10930 // Conservatively set memVT to the entire set of vectors stored.
10931 unsigned NumElts = 0;
10932 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
10933 Type *ArgTy = I.getArgOperand(ArgI)->getType();
10934 if (!ArgTy->isVectorTy())
10936 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
10938 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10939 Info.ptrVal = I.getArgOperand(0);
10941 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10942 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10943 Info.vol = false; // volatile stores with NEON intrinsics not supported
10944 Info.readMem = false;
10945 Info.writeMem = true;
10948 case Intrinsic::arm_ldaex:
10949 case Intrinsic::arm_ldrex: {
10950 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
10951 Info.opc = ISD::INTRINSIC_W_CHAIN;
10952 Info.memVT = MVT::getVT(PtrTy->getElementType());
10953 Info.ptrVal = I.getArgOperand(0);
10955 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
10957 Info.readMem = true;
10958 Info.writeMem = false;
10961 case Intrinsic::arm_stlex:
10962 case Intrinsic::arm_strex: {
10963 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
10964 Info.opc = ISD::INTRINSIC_W_CHAIN;
10965 Info.memVT = MVT::getVT(PtrTy->getElementType());
10966 Info.ptrVal = I.getArgOperand(1);
10968 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
10970 Info.readMem = false;
10971 Info.writeMem = true;
10974 case Intrinsic::arm_stlexd:
10975 case Intrinsic::arm_strexd: {
10976 Info.opc = ISD::INTRINSIC_W_CHAIN;
10977 Info.memVT = MVT::i64;
10978 Info.ptrVal = I.getArgOperand(2);
10982 Info.readMem = false;
10983 Info.writeMem = true;
10986 case Intrinsic::arm_ldaexd:
10987 case Intrinsic::arm_ldrexd: {
10988 Info.opc = ISD::INTRINSIC_W_CHAIN;
10989 Info.memVT = MVT::i64;
10990 Info.ptrVal = I.getArgOperand(0);
10994 Info.readMem = true;
10995 Info.writeMem = false;
11005 /// \brief Returns true if it is beneficial to convert a load of a constant
11006 /// to just the constant itself.
11007 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
11009 assert(Ty->isIntegerTy());
11011 unsigned Bits = Ty->getPrimitiveSizeInBits();
11012 if (Bits == 0 || Bits > 32)
11017 bool ARMTargetLowering::hasLoadLinkedStoreConditional() const { return true; }
11019 Instruction* ARMTargetLowering::makeDMB(IRBuilder<> &Builder,
11020 ARM_MB::MemBOpt Domain) const {
11021 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11023 // First, if the target has no DMB, see what fallback we can use.
11024 if (!Subtarget->hasDataBarrier()) {
11025 // Some ARMv6 cpus can support data barriers with an mcr instruction.
11026 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
11028 if (Subtarget->hasV6Ops() && !Subtarget->isThumb()) {
11029 Function *MCR = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_mcr);
11030 Value* args[6] = {Builder.getInt32(15), Builder.getInt32(0),
11031 Builder.getInt32(0), Builder.getInt32(7),
11032 Builder.getInt32(10), Builder.getInt32(5)};
11033 return Builder.CreateCall(MCR, args);
11035 // Instead of using barriers, atomic accesses on these subtargets use
11037 llvm_unreachable("makeDMB on a target so old that it has no barriers");
11040 Function *DMB = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_dmb);
11041 // Only a full system barrier exists in the M-class architectures.
11042 Domain = Subtarget->isMClass() ? ARM_MB::SY : Domain;
11043 Constant *CDomain = Builder.getInt32(Domain);
11044 return Builder.CreateCall(DMB, CDomain);
11048 // Based on http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html
11049 Instruction* ARMTargetLowering::emitLeadingFence(IRBuilder<> &Builder,
11050 AtomicOrdering Ord, bool IsStore,
11051 bool IsLoad) const {
11052 if (!getInsertFencesForAtomic())
11058 llvm_unreachable("Invalid fence: unordered/non-atomic");
11061 return nullptr; // Nothing to do
11062 case SequentiallyConsistent:
11064 return nullptr; // Nothing to do
11067 case AcquireRelease:
11068 if (Subtarget->isSwift())
11069 return makeDMB(Builder, ARM_MB::ISHST);
11070 // FIXME: add a comment with a link to documentation justifying this.
11072 return makeDMB(Builder, ARM_MB::ISH);
11074 llvm_unreachable("Unknown fence ordering in emitLeadingFence");
11077 Instruction* ARMTargetLowering::emitTrailingFence(IRBuilder<> &Builder,
11078 AtomicOrdering Ord, bool IsStore,
11079 bool IsLoad) const {
11080 if (!getInsertFencesForAtomic())
11086 llvm_unreachable("Invalid fence: unordered/not-atomic");
11089 return nullptr; // Nothing to do
11091 case AcquireRelease:
11092 case SequentiallyConsistent:
11093 return makeDMB(Builder, ARM_MB::ISH);
11095 llvm_unreachable("Unknown fence ordering in emitTrailingFence");
11098 // Loads and stores less than 64-bits are already atomic; ones above that
11099 // are doomed anyway, so defer to the default libcall and blame the OS when
11100 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11101 // anything for those.
11102 bool ARMTargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
11103 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
11104 return (Size == 64) && !Subtarget->isMClass();
11107 // Loads and stores less than 64-bits are already atomic; ones above that
11108 // are doomed anyway, so defer to the default libcall and blame the OS when
11109 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11110 // anything for those.
11111 // FIXME: ldrd and strd are atomic if the CPU has LPAE (e.g. A15 has that
11112 // guarantee, see DDI0406C ARM architecture reference manual,
11113 // sections A8.8.72-74 LDRD)
11114 bool ARMTargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
11115 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
11116 return (Size == 64) && !Subtarget->isMClass();
11119 // For the real atomic operations, we have ldrex/strex up to 32 bits,
11120 // and up to 64 bits on the non-M profiles
11121 bool ARMTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
11122 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
11123 return Size <= (Subtarget->isMClass() ? 32U : 64U);
11126 // This has so far only been implemented for MachO.
11127 bool ARMTargetLowering::useLoadStackGuardNode() const {
11128 return Subtarget->isTargetMachO();
11131 bool ARMTargetLowering::canCombineStoreAndExtract(Type *VectorTy, Value *Idx,
11132 unsigned &Cost) const {
11133 // If we do not have NEON, vector types are not natively supported.
11134 if (!Subtarget->hasNEON())
11137 // Floating point values and vector values map to the same register file.
11138 // Therefore, althought we could do a store extract of a vector type, this is
11139 // better to leave at float as we have more freedom in the addressing mode for
11141 if (VectorTy->isFPOrFPVectorTy())
11144 // If the index is unknown at compile time, this is very expensive to lower
11145 // and it is not possible to combine the store with the extract.
11146 if (!isa<ConstantInt>(Idx))
11149 assert(VectorTy->isVectorTy() && "VectorTy is not a vector type");
11150 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth();
11151 // We can do a store + vector extract on any vector that fits perfectly in a D
11153 if (BitWidth == 64 || BitWidth == 128) {
11160 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
11161 AtomicOrdering Ord) const {
11162 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11163 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
11164 bool IsAcquire = isAtLeastAcquire(Ord);
11166 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
11167 // intrinsic must return {i32, i32} and we have to recombine them into a
11168 // single i64 here.
11169 if (ValTy->getPrimitiveSizeInBits() == 64) {
11170 Intrinsic::ID Int =
11171 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
11172 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int);
11174 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11175 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
11177 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
11178 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
11179 if (!Subtarget->isLittle())
11180 std::swap (Lo, Hi);
11181 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
11182 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
11183 return Builder.CreateOr(
11184 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
11187 Type *Tys[] = { Addr->getType() };
11188 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
11189 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int, Tys);
11191 return Builder.CreateTruncOrBitCast(
11192 Builder.CreateCall(Ldrex, Addr),
11193 cast<PointerType>(Addr->getType())->getElementType());
11196 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
11198 AtomicOrdering Ord) const {
11199 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11200 bool IsRelease = isAtLeastRelease(Ord);
11202 // Since the intrinsics must have legal type, the i64 intrinsics take two
11203 // parameters: "i32, i32". We must marshal Val into the appropriate form
11204 // before the call.
11205 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
11206 Intrinsic::ID Int =
11207 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
11208 Function *Strex = Intrinsic::getDeclaration(M, Int);
11209 Type *Int32Ty = Type::getInt32Ty(M->getContext());
11211 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
11212 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
11213 if (!Subtarget->isLittle())
11214 std::swap (Lo, Hi);
11215 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11216 return Builder.CreateCall3(Strex, Lo, Hi, Addr);
11219 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
11220 Type *Tys[] = { Addr->getType() };
11221 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
11223 return Builder.CreateCall2(
11224 Strex, Builder.CreateZExtOrBitCast(
11225 Val, Strex->getFunctionType()->getParamType(0)),
11237 static bool isHomogeneousAggregate(Type *Ty, HABaseType &Base,
11238 uint64_t &Members) {
11239 if (const StructType *ST = dyn_cast<StructType>(Ty)) {
11240 for (unsigned i = 0; i < ST->getNumElements(); ++i) {
11241 uint64_t SubMembers = 0;
11242 if (!isHomogeneousAggregate(ST->getElementType(i), Base, SubMembers))
11244 Members += SubMembers;
11246 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
11247 uint64_t SubMembers = 0;
11248 if (!isHomogeneousAggregate(AT->getElementType(), Base, SubMembers))
11250 Members += SubMembers * AT->getNumElements();
11251 } else if (Ty->isFloatTy()) {
11252 if (Base != HA_UNKNOWN && Base != HA_FLOAT)
11256 } else if (Ty->isDoubleTy()) {
11257 if (Base != HA_UNKNOWN && Base != HA_DOUBLE)
11261 } else if (const VectorType *VT = dyn_cast<VectorType>(Ty)) {
11268 return VT->getBitWidth() == 64;
11270 return VT->getBitWidth() == 128;
11272 switch (VT->getBitWidth()) {
11285 return (Members > 0 && Members <= 4);
11288 /// \brief Return true if a type is an AAPCS-VFP homogeneous aggregate or one of
11289 /// [N x i32] or [N x i64]. This allows front-ends to skip emitting padding when
11290 /// passing according to AAPCS rules.
11291 bool ARMTargetLowering::functionArgumentNeedsConsecutiveRegisters(
11292 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
11293 if (getEffectiveCallingConv(CallConv, isVarArg) !=
11294 CallingConv::ARM_AAPCS_VFP)
11297 HABaseType Base = HA_UNKNOWN;
11298 uint64_t Members = 0;
11299 bool IsHA = isHomogeneousAggregate(Ty, Base, Members);
11300 DEBUG(dbgs() << "isHA: " << IsHA << " "; Ty->dump());
11302 bool IsIntArray = Ty->isArrayTy() && Ty->getArrayElementType()->isIntegerTy();
11303 return IsHA || IsIntArray;