1 //===-- AArch64ISelLowering.cpp - AArch64 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 implements the AArch64TargetLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "AArch64ISelLowering.h"
15 #include "AArch64CallingConvention.h"
16 #include "AArch64MachineFunctionInfo.h"
17 #include "AArch64PerfectShuffle.h"
18 #include "AArch64Subtarget.h"
19 #include "AArch64TargetMachine.h"
20 #include "AArch64TargetObjectFile.h"
21 #include "MCTargetDesc/AArch64AddressingModes.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/CodeGen/CallingConvLower.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/GetElementPtrTypeIterator.h"
29 #include "llvm/IR/Intrinsics.h"
30 #include "llvm/IR/Type.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Target/TargetOptions.h"
38 #define DEBUG_TYPE "aarch64-lower"
40 STATISTIC(NumTailCalls, "Number of tail calls");
41 STATISTIC(NumShiftInserts, "Number of vector shift inserts");
43 // Place holder until extr generation is tested fully.
45 EnableAArch64ExtrGeneration("aarch64-extr-generation", cl::Hidden,
46 cl::desc("Allow AArch64 (or (shift)(shift))->extract"),
50 EnableAArch64SlrGeneration("aarch64-shift-insert-generation", cl::Hidden,
51 cl::desc("Allow AArch64 SLI/SRI formation"),
54 // FIXME: The necessary dtprel relocations don't seem to be supported
55 // well in the GNU bfd and gold linkers at the moment. Therefore, by
56 // default, for now, fall back to GeneralDynamic code generation.
57 cl::opt<bool> EnableAArch64ELFLocalDynamicTLSGeneration(
58 "aarch64-elf-ldtls-generation", cl::Hidden,
59 cl::desc("Allow AArch64 Local Dynamic TLS code generation"),
62 /// Value type used for condition codes.
63 static const MVT MVT_CC = MVT::i32;
65 AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
66 const AArch64Subtarget &STI)
67 : TargetLowering(TM), Subtarget(&STI) {
69 // AArch64 doesn't have comparisons which set GPRs or setcc instructions, so
70 // we have to make something up. Arbitrarily, choose ZeroOrOne.
71 setBooleanContents(ZeroOrOneBooleanContent);
72 // When comparing vectors the result sets the different elements in the
73 // vector to all-one or all-zero.
74 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
76 // Set up the register classes.
77 addRegisterClass(MVT::i32, &AArch64::GPR32allRegClass);
78 addRegisterClass(MVT::i64, &AArch64::GPR64allRegClass);
80 if (Subtarget->hasFPARMv8()) {
81 addRegisterClass(MVT::f16, &AArch64::FPR16RegClass);
82 addRegisterClass(MVT::f32, &AArch64::FPR32RegClass);
83 addRegisterClass(MVT::f64, &AArch64::FPR64RegClass);
84 addRegisterClass(MVT::f128, &AArch64::FPR128RegClass);
87 if (Subtarget->hasNEON()) {
88 addRegisterClass(MVT::v16i8, &AArch64::FPR8RegClass);
89 addRegisterClass(MVT::v8i16, &AArch64::FPR16RegClass);
90 // Someone set us up the NEON.
91 addDRTypeForNEON(MVT::v2f32);
92 addDRTypeForNEON(MVT::v8i8);
93 addDRTypeForNEON(MVT::v4i16);
94 addDRTypeForNEON(MVT::v2i32);
95 addDRTypeForNEON(MVT::v1i64);
96 addDRTypeForNEON(MVT::v1f64);
97 addDRTypeForNEON(MVT::v4f16);
99 addQRTypeForNEON(MVT::v4f32);
100 addQRTypeForNEON(MVT::v2f64);
101 addQRTypeForNEON(MVT::v16i8);
102 addQRTypeForNEON(MVT::v8i16);
103 addQRTypeForNEON(MVT::v4i32);
104 addQRTypeForNEON(MVT::v2i64);
105 addQRTypeForNEON(MVT::v8f16);
108 // Compute derived properties from the register classes
109 computeRegisterProperties(Subtarget->getRegisterInfo());
111 // Provide all sorts of operation actions
112 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
113 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
114 setOperationAction(ISD::SETCC, MVT::i32, Custom);
115 setOperationAction(ISD::SETCC, MVT::i64, Custom);
116 setOperationAction(ISD::SETCC, MVT::f32, Custom);
117 setOperationAction(ISD::SETCC, MVT::f64, Custom);
118 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
119 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
120 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
121 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
122 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
123 setOperationAction(ISD::SELECT, MVT::i32, Custom);
124 setOperationAction(ISD::SELECT, MVT::i64, Custom);
125 setOperationAction(ISD::SELECT, MVT::f32, Custom);
126 setOperationAction(ISD::SELECT, MVT::f64, Custom);
127 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
128 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
129 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
130 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
131 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
132 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
134 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
135 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
136 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
138 setOperationAction(ISD::FREM, MVT::f32, Expand);
139 setOperationAction(ISD::FREM, MVT::f64, Expand);
140 setOperationAction(ISD::FREM, MVT::f80, Expand);
142 // Custom lowering hooks are needed for XOR
143 // to fold it into CSINC/CSINV.
144 setOperationAction(ISD::XOR, MVT::i32, Custom);
145 setOperationAction(ISD::XOR, MVT::i64, Custom);
147 // Virtually no operation on f128 is legal, but LLVM can't expand them when
148 // there's a valid register class, so we need custom operations in most cases.
149 setOperationAction(ISD::FABS, MVT::f128, Expand);
150 setOperationAction(ISD::FADD, MVT::f128, Custom);
151 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
152 setOperationAction(ISD::FCOS, MVT::f128, Expand);
153 setOperationAction(ISD::FDIV, MVT::f128, Custom);
154 setOperationAction(ISD::FMA, MVT::f128, Expand);
155 setOperationAction(ISD::FMUL, MVT::f128, Custom);
156 setOperationAction(ISD::FNEG, MVT::f128, Expand);
157 setOperationAction(ISD::FPOW, MVT::f128, Expand);
158 setOperationAction(ISD::FREM, MVT::f128, Expand);
159 setOperationAction(ISD::FRINT, MVT::f128, Expand);
160 setOperationAction(ISD::FSIN, MVT::f128, Expand);
161 setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
162 setOperationAction(ISD::FSQRT, MVT::f128, Expand);
163 setOperationAction(ISD::FSUB, MVT::f128, Custom);
164 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
165 setOperationAction(ISD::SETCC, MVT::f128, Custom);
166 setOperationAction(ISD::BR_CC, MVT::f128, Custom);
167 setOperationAction(ISD::SELECT, MVT::f128, Custom);
168 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
169 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);
171 // Lowering for many of the conversions is actually specified by the non-f128
172 // type. The LowerXXX function will be trivial when f128 isn't involved.
173 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
174 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
175 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
176 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
177 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
178 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
179 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
180 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
181 setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
182 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
183 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
184 setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
185 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
186 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
188 // Variable arguments.
189 setOperationAction(ISD::VASTART, MVT::Other, Custom);
190 setOperationAction(ISD::VAARG, MVT::Other, Custom);
191 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
192 setOperationAction(ISD::VAEND, MVT::Other, Expand);
194 // Variable-sized objects.
195 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
196 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
197 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
199 // Constant pool entries
200 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
203 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
205 // Add/Sub overflow ops with MVT::Glues are lowered to NZCV dependences.
206 setOperationAction(ISD::ADDC, MVT::i32, Custom);
207 setOperationAction(ISD::ADDE, MVT::i32, Custom);
208 setOperationAction(ISD::SUBC, MVT::i32, Custom);
209 setOperationAction(ISD::SUBE, MVT::i32, Custom);
210 setOperationAction(ISD::ADDC, MVT::i64, Custom);
211 setOperationAction(ISD::ADDE, MVT::i64, Custom);
212 setOperationAction(ISD::SUBC, MVT::i64, Custom);
213 setOperationAction(ISD::SUBE, MVT::i64, Custom);
215 // AArch64 lacks both left-rotate and popcount instructions.
216 setOperationAction(ISD::ROTL, MVT::i32, Expand);
217 setOperationAction(ISD::ROTL, MVT::i64, Expand);
218 for (MVT VT : MVT::vector_valuetypes()) {
219 setOperationAction(ISD::ROTL, VT, Expand);
220 setOperationAction(ISD::ROTR, VT, Expand);
223 // AArch64 doesn't have {U|S}MUL_LOHI.
224 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
225 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
228 // Expand the undefined-at-zero variants to cttz/ctlz to their defined-at-zero
229 // counterparts, which AArch64 supports directly.
230 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
231 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
232 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
233 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
235 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
236 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
238 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
239 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
240 for (MVT VT : MVT::vector_valuetypes()) {
241 setOperationAction(ISD::SDIVREM, VT, Expand);
242 setOperationAction(ISD::UDIVREM, VT, Expand);
244 setOperationAction(ISD::SREM, MVT::i32, Expand);
245 setOperationAction(ISD::SREM, MVT::i64, Expand);
246 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
247 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
248 setOperationAction(ISD::UREM, MVT::i32, Expand);
249 setOperationAction(ISD::UREM, MVT::i64, Expand);
251 // Custom lower Add/Sub/Mul with overflow.
252 setOperationAction(ISD::SADDO, MVT::i32, Custom);
253 setOperationAction(ISD::SADDO, MVT::i64, Custom);
254 setOperationAction(ISD::UADDO, MVT::i32, Custom);
255 setOperationAction(ISD::UADDO, MVT::i64, Custom);
256 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
257 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
258 setOperationAction(ISD::USUBO, MVT::i32, Custom);
259 setOperationAction(ISD::USUBO, MVT::i64, Custom);
260 setOperationAction(ISD::SMULO, MVT::i32, Custom);
261 setOperationAction(ISD::SMULO, MVT::i64, Custom);
262 setOperationAction(ISD::UMULO, MVT::i32, Custom);
263 setOperationAction(ISD::UMULO, MVT::i64, Custom);
265 setOperationAction(ISD::FSIN, MVT::f32, Expand);
266 setOperationAction(ISD::FSIN, MVT::f64, Expand);
267 setOperationAction(ISD::FCOS, MVT::f32, Expand);
268 setOperationAction(ISD::FCOS, MVT::f64, Expand);
269 setOperationAction(ISD::FPOW, MVT::f32, Expand);
270 setOperationAction(ISD::FPOW, MVT::f64, Expand);
271 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
272 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
274 // f16 is a storage-only type, always promote it to f32.
275 setOperationAction(ISD::SETCC, MVT::f16, Promote);
276 setOperationAction(ISD::BR_CC, MVT::f16, Promote);
277 setOperationAction(ISD::SELECT_CC, MVT::f16, Promote);
278 setOperationAction(ISD::SELECT, MVT::f16, Promote);
279 setOperationAction(ISD::FADD, MVT::f16, Promote);
280 setOperationAction(ISD::FSUB, MVT::f16, Promote);
281 setOperationAction(ISD::FMUL, MVT::f16, Promote);
282 setOperationAction(ISD::FDIV, MVT::f16, Promote);
283 setOperationAction(ISD::FREM, MVT::f16, Promote);
284 setOperationAction(ISD::FMA, MVT::f16, Promote);
285 setOperationAction(ISD::FNEG, MVT::f16, Promote);
286 setOperationAction(ISD::FABS, MVT::f16, Promote);
287 setOperationAction(ISD::FCEIL, MVT::f16, Promote);
288 setOperationAction(ISD::FCOPYSIGN, MVT::f16, Promote);
289 setOperationAction(ISD::FCOS, MVT::f16, Promote);
290 setOperationAction(ISD::FFLOOR, MVT::f16, Promote);
291 setOperationAction(ISD::FNEARBYINT, MVT::f16, Promote);
292 setOperationAction(ISD::FPOW, MVT::f16, Promote);
293 setOperationAction(ISD::FPOWI, MVT::f16, Promote);
294 setOperationAction(ISD::FRINT, MVT::f16, Promote);
295 setOperationAction(ISD::FSIN, MVT::f16, Promote);
296 setOperationAction(ISD::FSINCOS, MVT::f16, Promote);
297 setOperationAction(ISD::FSQRT, MVT::f16, Promote);
298 setOperationAction(ISD::FEXP, MVT::f16, Promote);
299 setOperationAction(ISD::FEXP2, MVT::f16, Promote);
300 setOperationAction(ISD::FLOG, MVT::f16, Promote);
301 setOperationAction(ISD::FLOG2, MVT::f16, Promote);
302 setOperationAction(ISD::FLOG10, MVT::f16, Promote);
303 setOperationAction(ISD::FROUND, MVT::f16, Promote);
304 setOperationAction(ISD::FTRUNC, MVT::f16, Promote);
305 setOperationAction(ISD::FMINNUM, MVT::f16, Promote);
306 setOperationAction(ISD::FMAXNUM, MVT::f16, Promote);
307 setOperationAction(ISD::FMINNAN, MVT::f16, Promote);
308 setOperationAction(ISD::FMAXNAN, MVT::f16, Promote);
310 // v4f16 is also a storage-only type, so promote it to v4f32 when that is
312 setOperationAction(ISD::FADD, MVT::v4f16, Promote);
313 setOperationAction(ISD::FSUB, MVT::v4f16, Promote);
314 setOperationAction(ISD::FMUL, MVT::v4f16, Promote);
315 setOperationAction(ISD::FDIV, MVT::v4f16, Promote);
316 setOperationAction(ISD::FP_EXTEND, MVT::v4f16, Promote);
317 setOperationAction(ISD::FP_ROUND, MVT::v4f16, Promote);
318 AddPromotedToType(ISD::FADD, MVT::v4f16, MVT::v4f32);
319 AddPromotedToType(ISD::FSUB, MVT::v4f16, MVT::v4f32);
320 AddPromotedToType(ISD::FMUL, MVT::v4f16, MVT::v4f32);
321 AddPromotedToType(ISD::FDIV, MVT::v4f16, MVT::v4f32);
322 AddPromotedToType(ISD::FP_EXTEND, MVT::v4f16, MVT::v4f32);
323 AddPromotedToType(ISD::FP_ROUND, MVT::v4f16, MVT::v4f32);
325 // Expand all other v4f16 operations.
326 // FIXME: We could generate better code by promoting some operations to
328 setOperationAction(ISD::FABS, MVT::v4f16, Expand);
329 setOperationAction(ISD::FCEIL, MVT::v4f16, Expand);
330 setOperationAction(ISD::FCOPYSIGN, MVT::v4f16, Expand);
331 setOperationAction(ISD::FCOS, MVT::v4f16, Expand);
332 setOperationAction(ISD::FFLOOR, MVT::v4f16, Expand);
333 setOperationAction(ISD::FMA, MVT::v4f16, Expand);
334 setOperationAction(ISD::FNEARBYINT, MVT::v4f16, Expand);
335 setOperationAction(ISD::FNEG, MVT::v4f16, Expand);
336 setOperationAction(ISD::FPOW, MVT::v4f16, Expand);
337 setOperationAction(ISD::FPOWI, MVT::v4f16, Expand);
338 setOperationAction(ISD::FREM, MVT::v4f16, Expand);
339 setOperationAction(ISD::FROUND, MVT::v4f16, Expand);
340 setOperationAction(ISD::FRINT, MVT::v4f16, Expand);
341 setOperationAction(ISD::FSIN, MVT::v4f16, Expand);
342 setOperationAction(ISD::FSINCOS, MVT::v4f16, Expand);
343 setOperationAction(ISD::FSQRT, MVT::v4f16, Expand);
344 setOperationAction(ISD::FTRUNC, MVT::v4f16, Expand);
345 setOperationAction(ISD::SETCC, MVT::v4f16, Expand);
346 setOperationAction(ISD::BR_CC, MVT::v4f16, Expand);
347 setOperationAction(ISD::SELECT, MVT::v4f16, Expand);
348 setOperationAction(ISD::SELECT_CC, MVT::v4f16, Expand);
349 setOperationAction(ISD::FEXP, MVT::v4f16, Expand);
350 setOperationAction(ISD::FEXP2, MVT::v4f16, Expand);
351 setOperationAction(ISD::FLOG, MVT::v4f16, Expand);
352 setOperationAction(ISD::FLOG2, MVT::v4f16, Expand);
353 setOperationAction(ISD::FLOG10, MVT::v4f16, Expand);
356 // v8f16 is also a storage-only type, so expand it.
357 setOperationAction(ISD::FABS, MVT::v8f16, Expand);
358 setOperationAction(ISD::FADD, MVT::v8f16, Expand);
359 setOperationAction(ISD::FCEIL, MVT::v8f16, Expand);
360 setOperationAction(ISD::FCOPYSIGN, MVT::v8f16, Expand);
361 setOperationAction(ISD::FCOS, MVT::v8f16, Expand);
362 setOperationAction(ISD::FDIV, MVT::v8f16, Expand);
363 setOperationAction(ISD::FFLOOR, MVT::v8f16, Expand);
364 setOperationAction(ISD::FMA, MVT::v8f16, Expand);
365 setOperationAction(ISD::FMUL, MVT::v8f16, Expand);
366 setOperationAction(ISD::FNEARBYINT, MVT::v8f16, Expand);
367 setOperationAction(ISD::FNEG, MVT::v8f16, Expand);
368 setOperationAction(ISD::FPOW, MVT::v8f16, Expand);
369 setOperationAction(ISD::FPOWI, MVT::v8f16, Expand);
370 setOperationAction(ISD::FREM, MVT::v8f16, Expand);
371 setOperationAction(ISD::FROUND, MVT::v8f16, Expand);
372 setOperationAction(ISD::FRINT, MVT::v8f16, Expand);
373 setOperationAction(ISD::FSIN, MVT::v8f16, Expand);
374 setOperationAction(ISD::FSINCOS, MVT::v8f16, Expand);
375 setOperationAction(ISD::FSQRT, MVT::v8f16, Expand);
376 setOperationAction(ISD::FSUB, MVT::v8f16, Expand);
377 setOperationAction(ISD::FTRUNC, MVT::v8f16, Expand);
378 setOperationAction(ISD::SETCC, MVT::v8f16, Expand);
379 setOperationAction(ISD::BR_CC, MVT::v8f16, Expand);
380 setOperationAction(ISD::SELECT, MVT::v8f16, Expand);
381 setOperationAction(ISD::SELECT_CC, MVT::v8f16, Expand);
382 setOperationAction(ISD::FP_EXTEND, MVT::v8f16, Expand);
383 setOperationAction(ISD::FEXP, MVT::v8f16, Expand);
384 setOperationAction(ISD::FEXP2, MVT::v8f16, Expand);
385 setOperationAction(ISD::FLOG, MVT::v8f16, Expand);
386 setOperationAction(ISD::FLOG2, MVT::v8f16, Expand);
387 setOperationAction(ISD::FLOG10, MVT::v8f16, Expand);
389 // AArch64 has implementations of a lot of rounding-like FP operations.
390 for (MVT Ty : {MVT::f32, MVT::f64}) {
391 setOperationAction(ISD::FFLOOR, Ty, Legal);
392 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
393 setOperationAction(ISD::FCEIL, Ty, Legal);
394 setOperationAction(ISD::FRINT, Ty, Legal);
395 setOperationAction(ISD::FTRUNC, Ty, Legal);
396 setOperationAction(ISD::FROUND, Ty, Legal);
397 setOperationAction(ISD::FMINNUM, Ty, Legal);
398 setOperationAction(ISD::FMAXNUM, Ty, Legal);
399 setOperationAction(ISD::FMINNAN, Ty, Legal);
400 setOperationAction(ISD::FMAXNAN, Ty, Legal);
403 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
405 // Lower READCYCLECOUNTER using an mrs from PMCCNTR_EL0.
406 // This requires the Performance Monitors extension.
407 if (Subtarget->hasPerfMon())
408 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Legal);
410 if (Subtarget->isTargetMachO()) {
411 // For iOS, we don't want to the normal expansion of a libcall to
412 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
414 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
415 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
417 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
418 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
421 // Make floating-point constants legal for the large code model, so they don't
422 // become loads from the constant pool.
423 if (Subtarget->isTargetMachO() && TM.getCodeModel() == CodeModel::Large) {
424 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
425 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
428 // AArch64 does not have floating-point extending loads, i1 sign-extending
429 // load, floating-point truncating stores, or v2i32->v2i16 truncating store.
430 for (MVT VT : MVT::fp_valuetypes()) {
431 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
432 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
433 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f64, Expand);
434 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f80, Expand);
436 for (MVT VT : MVT::integer_valuetypes())
437 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Expand);
439 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
440 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
441 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
442 setTruncStoreAction(MVT::f128, MVT::f80, Expand);
443 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
444 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
445 setTruncStoreAction(MVT::f128, MVT::f16, Expand);
447 setOperationAction(ISD::BITCAST, MVT::i16, Custom);
448 setOperationAction(ISD::BITCAST, MVT::f16, Custom);
450 // Indexed loads and stores are supported.
451 for (unsigned im = (unsigned)ISD::PRE_INC;
452 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
453 setIndexedLoadAction(im, MVT::i8, Legal);
454 setIndexedLoadAction(im, MVT::i16, Legal);
455 setIndexedLoadAction(im, MVT::i32, Legal);
456 setIndexedLoadAction(im, MVT::i64, Legal);
457 setIndexedLoadAction(im, MVT::f64, Legal);
458 setIndexedLoadAction(im, MVT::f32, Legal);
459 setIndexedLoadAction(im, MVT::f16, Legal);
460 setIndexedStoreAction(im, MVT::i8, Legal);
461 setIndexedStoreAction(im, MVT::i16, Legal);
462 setIndexedStoreAction(im, MVT::i32, Legal);
463 setIndexedStoreAction(im, MVT::i64, Legal);
464 setIndexedStoreAction(im, MVT::f64, Legal);
465 setIndexedStoreAction(im, MVT::f32, Legal);
466 setIndexedStoreAction(im, MVT::f16, Legal);
470 setOperationAction(ISD::TRAP, MVT::Other, Legal);
472 // We combine OR nodes for bitfield operations.
473 setTargetDAGCombine(ISD::OR);
475 // Vector add and sub nodes may conceal a high-half opportunity.
476 // Also, try to fold ADD into CSINC/CSINV..
477 setTargetDAGCombine(ISD::ADD);
478 setTargetDAGCombine(ISD::SUB);
480 setTargetDAGCombine(ISD::XOR);
481 setTargetDAGCombine(ISD::SINT_TO_FP);
482 setTargetDAGCombine(ISD::UINT_TO_FP);
484 setTargetDAGCombine(ISD::FP_TO_SINT);
485 setTargetDAGCombine(ISD::FP_TO_UINT);
486 setTargetDAGCombine(ISD::FDIV);
488 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
490 setTargetDAGCombine(ISD::ANY_EXTEND);
491 setTargetDAGCombine(ISD::ZERO_EXTEND);
492 setTargetDAGCombine(ISD::SIGN_EXTEND);
493 setTargetDAGCombine(ISD::BITCAST);
494 setTargetDAGCombine(ISD::CONCAT_VECTORS);
495 setTargetDAGCombine(ISD::STORE);
496 if (Subtarget->supportsAddressTopByteIgnored())
497 setTargetDAGCombine(ISD::LOAD);
499 setTargetDAGCombine(ISD::MUL);
501 setTargetDAGCombine(ISD::SELECT);
502 setTargetDAGCombine(ISD::VSELECT);
504 setTargetDAGCombine(ISD::INTRINSIC_VOID);
505 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
506 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
507 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
509 MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8;
510 MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4;
511 MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 4;
513 setStackPointerRegisterToSaveRestore(AArch64::SP);
515 setSchedulingPreference(Sched::Hybrid);
518 MaskAndBranchFoldingIsLegal = true;
519 EnableExtLdPromotion = true;
521 setMinFunctionAlignment(2);
523 setHasExtractBitsInsn(true);
525 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
527 if (Subtarget->hasNEON()) {
528 // FIXME: v1f64 shouldn't be legal if we can avoid it, because it leads to
529 // silliness like this:
530 setOperationAction(ISD::FABS, MVT::v1f64, Expand);
531 setOperationAction(ISD::FADD, MVT::v1f64, Expand);
532 setOperationAction(ISD::FCEIL, MVT::v1f64, Expand);
533 setOperationAction(ISD::FCOPYSIGN, MVT::v1f64, Expand);
534 setOperationAction(ISD::FCOS, MVT::v1f64, Expand);
535 setOperationAction(ISD::FDIV, MVT::v1f64, Expand);
536 setOperationAction(ISD::FFLOOR, MVT::v1f64, Expand);
537 setOperationAction(ISD::FMA, MVT::v1f64, Expand);
538 setOperationAction(ISD::FMUL, MVT::v1f64, Expand);
539 setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Expand);
540 setOperationAction(ISD::FNEG, MVT::v1f64, Expand);
541 setOperationAction(ISD::FPOW, MVT::v1f64, Expand);
542 setOperationAction(ISD::FREM, MVT::v1f64, Expand);
543 setOperationAction(ISD::FROUND, MVT::v1f64, Expand);
544 setOperationAction(ISD::FRINT, MVT::v1f64, Expand);
545 setOperationAction(ISD::FSIN, MVT::v1f64, Expand);
546 setOperationAction(ISD::FSINCOS, MVT::v1f64, Expand);
547 setOperationAction(ISD::FSQRT, MVT::v1f64, Expand);
548 setOperationAction(ISD::FSUB, MVT::v1f64, Expand);
549 setOperationAction(ISD::FTRUNC, MVT::v1f64, Expand);
550 setOperationAction(ISD::SETCC, MVT::v1f64, Expand);
551 setOperationAction(ISD::BR_CC, MVT::v1f64, Expand);
552 setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
553 setOperationAction(ISD::SELECT_CC, MVT::v1f64, Expand);
554 setOperationAction(ISD::FP_EXTEND, MVT::v1f64, Expand);
556 setOperationAction(ISD::FP_TO_SINT, MVT::v1i64, Expand);
557 setOperationAction(ISD::FP_TO_UINT, MVT::v1i64, Expand);
558 setOperationAction(ISD::SINT_TO_FP, MVT::v1i64, Expand);
559 setOperationAction(ISD::UINT_TO_FP, MVT::v1i64, Expand);
560 setOperationAction(ISD::FP_ROUND, MVT::v1f64, Expand);
562 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
564 // AArch64 doesn't have a direct vector ->f32 conversion instructions for
565 // elements smaller than i32, so promote the input to i32 first.
566 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Promote);
567 setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Promote);
568 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Promote);
569 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Promote);
570 // i8 and i16 vector elements also need promotion to i32 for v8i8 or v8i16
571 // -> v8f16 conversions.
572 setOperationAction(ISD::SINT_TO_FP, MVT::v8i8, Promote);
573 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Promote);
574 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
575 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Promote);
576 // Similarly, there is no direct i32 -> f64 vector conversion instruction.
577 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
578 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
579 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
580 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
581 // Or, direct i32 -> f16 vector conversion. Set it so custom, so the
582 // conversion happens in two steps: v4i32 -> v4f32 -> v4f16
583 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Custom);
584 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
586 // AArch64 doesn't have MUL.2d:
587 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
588 // Custom handling for some quad-vector types to detect MULL.
589 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
590 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
591 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
593 setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal);
594 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
595 // Likewise, narrowing and extending vector loads/stores aren't handled
597 for (MVT VT : MVT::vector_valuetypes()) {
598 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
600 setOperationAction(ISD::MULHS, VT, Expand);
601 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
602 setOperationAction(ISD::MULHU, VT, Expand);
603 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
605 setOperationAction(ISD::BSWAP, VT, Expand);
607 for (MVT InnerVT : MVT::vector_valuetypes()) {
608 setTruncStoreAction(VT, InnerVT, Expand);
609 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
610 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
611 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
615 // AArch64 has implementations of a lot of rounding-like FP operations.
616 for (MVT Ty : {MVT::v2f32, MVT::v4f32, MVT::v2f64}) {
617 setOperationAction(ISD::FFLOOR, Ty, Legal);
618 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
619 setOperationAction(ISD::FCEIL, Ty, Legal);
620 setOperationAction(ISD::FRINT, Ty, Legal);
621 setOperationAction(ISD::FTRUNC, Ty, Legal);
622 setOperationAction(ISD::FROUND, Ty, Legal);
626 // Prefer likely predicted branches to selects on out-of-order cores.
627 if (Subtarget->isCortexA57())
628 PredictableSelectIsExpensive = true;
631 void AArch64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) {
632 if (VT == MVT::v2f32 || VT == MVT::v4f16) {
633 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
634 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i32);
636 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
637 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i32);
638 } else if (VT == MVT::v2f64 || VT == MVT::v4f32 || VT == MVT::v8f16) {
639 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
640 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i64);
642 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
643 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i64);
646 // Mark vector float intrinsics as expand.
647 if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) {
648 setOperationAction(ISD::FSIN, VT.getSimpleVT(), Expand);
649 setOperationAction(ISD::FCOS, VT.getSimpleVT(), Expand);
650 setOperationAction(ISD::FPOWI, VT.getSimpleVT(), Expand);
651 setOperationAction(ISD::FPOW, VT.getSimpleVT(), Expand);
652 setOperationAction(ISD::FLOG, VT.getSimpleVT(), Expand);
653 setOperationAction(ISD::FLOG2, VT.getSimpleVT(), Expand);
654 setOperationAction(ISD::FLOG10, VT.getSimpleVT(), Expand);
655 setOperationAction(ISD::FEXP, VT.getSimpleVT(), Expand);
656 setOperationAction(ISD::FEXP2, VT.getSimpleVT(), Expand);
658 // But we do support custom-lowering for FCOPYSIGN.
659 setOperationAction(ISD::FCOPYSIGN, VT.getSimpleVT(), Custom);
662 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
663 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
664 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
665 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
666 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Custom);
667 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
668 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
669 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
670 setOperationAction(ISD::AND, VT.getSimpleVT(), Custom);
671 setOperationAction(ISD::OR, VT.getSimpleVT(), Custom);
672 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
673 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
675 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
676 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
677 setOperationAction(ISD::VSELECT, VT.getSimpleVT(), Expand);
678 for (MVT InnerVT : MVT::all_valuetypes())
679 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT.getSimpleVT(), Expand);
681 // CNT supports only B element sizes.
682 if (VT != MVT::v8i8 && VT != MVT::v16i8)
683 setOperationAction(ISD::CTPOP, VT.getSimpleVT(), Expand);
685 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
686 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
687 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
688 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
689 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
691 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
692 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
694 // [SU][MIN|MAX] are available for all NEON types apart from i64.
695 if (!VT.isFloatingPoint() &&
696 VT.getSimpleVT() != MVT::v2i64 && VT.getSimpleVT() != MVT::v1i64)
697 for (unsigned Opcode : {ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX})
698 setOperationAction(Opcode, VT.getSimpleVT(), Legal);
700 // F[MIN|MAX][NUM|NAN] are available for all FP NEON types (not f16 though!).
701 if (VT.isFloatingPoint() && VT.getVectorElementType() != MVT::f16)
702 for (unsigned Opcode : {ISD::FMINNAN, ISD::FMAXNAN,
703 ISD::FMINNUM, ISD::FMAXNUM})
704 setOperationAction(Opcode, VT.getSimpleVT(), Legal);
706 if (Subtarget->isLittleEndian()) {
707 for (unsigned im = (unsigned)ISD::PRE_INC;
708 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
709 setIndexedLoadAction(im, VT.getSimpleVT(), Legal);
710 setIndexedStoreAction(im, VT.getSimpleVT(), Legal);
715 void AArch64TargetLowering::addDRTypeForNEON(MVT VT) {
716 addRegisterClass(VT, &AArch64::FPR64RegClass);
717 addTypeForNEON(VT, MVT::v2i32);
720 void AArch64TargetLowering::addQRTypeForNEON(MVT VT) {
721 addRegisterClass(VT, &AArch64::FPR128RegClass);
722 addTypeForNEON(VT, MVT::v4i32);
725 EVT AArch64TargetLowering::getSetCCResultType(const DataLayout &, LLVMContext &,
729 return VT.changeVectorElementTypeToInteger();
732 /// computeKnownBitsForTargetNode - Determine which of the bits specified in
733 /// Mask are known to be either zero or one and return them in the
734 /// KnownZero/KnownOne bitsets.
735 void AArch64TargetLowering::computeKnownBitsForTargetNode(
736 const SDValue Op, APInt &KnownZero, APInt &KnownOne,
737 const SelectionDAG &DAG, unsigned Depth) const {
738 switch (Op.getOpcode()) {
741 case AArch64ISD::CSEL: {
742 APInt KnownZero2, KnownOne2;
743 DAG.computeKnownBits(Op->getOperand(0), KnownZero, KnownOne, Depth + 1);
744 DAG.computeKnownBits(Op->getOperand(1), KnownZero2, KnownOne2, Depth + 1);
745 KnownZero &= KnownZero2;
746 KnownOne &= KnownOne2;
749 case ISD::INTRINSIC_W_CHAIN: {
750 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
751 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
754 case Intrinsic::aarch64_ldaxr:
755 case Intrinsic::aarch64_ldxr: {
756 unsigned BitWidth = KnownOne.getBitWidth();
757 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
758 unsigned MemBits = VT.getScalarType().getSizeInBits();
759 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
765 case ISD::INTRINSIC_WO_CHAIN:
766 case ISD::INTRINSIC_VOID: {
767 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
771 case Intrinsic::aarch64_neon_umaxv:
772 case Intrinsic::aarch64_neon_uminv: {
773 // Figure out the datatype of the vector operand. The UMINV instruction
774 // will zero extend the result, so we can mark as known zero all the
775 // bits larger than the element datatype. 32-bit or larget doesn't need
776 // this as those are legal types and will be handled by isel directly.
777 MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
778 unsigned BitWidth = KnownZero.getBitWidth();
779 if (VT == MVT::v8i8 || VT == MVT::v16i8) {
780 assert(BitWidth >= 8 && "Unexpected width!");
781 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8);
783 } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
784 assert(BitWidth >= 16 && "Unexpected width!");
785 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
795 MVT AArch64TargetLowering::getScalarShiftAmountTy(const DataLayout &DL,
800 bool AArch64TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
804 if (Subtarget->requiresStrictAlign())
807 // FIXME: This is mostly true for Cyclone, but not necessarily others.
809 // FIXME: Define an attribute for slow unaligned accesses instead of
810 // relying on the CPU type as a proxy.
811 // On Cyclone, unaligned 128-bit stores are slow.
812 *Fast = !Subtarget->isCyclone() || VT.getStoreSize() != 16 ||
813 // See comments in performSTORECombine() for more details about
816 // Code that uses clang vector extensions can mark that it
817 // wants unaligned accesses to be treated as fast by
818 // underspecifying alignment to be 1 or 2.
821 // Disregard v2i64. Memcpy lowering produces those and splitting
822 // them regresses performance on micro-benchmarks and olden/bh.
829 AArch64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
830 const TargetLibraryInfo *libInfo) const {
831 return AArch64::createFastISel(funcInfo, libInfo);
834 const char *AArch64TargetLowering::getTargetNodeName(unsigned Opcode) const {
835 switch ((AArch64ISD::NodeType)Opcode) {
836 case AArch64ISD::FIRST_NUMBER: break;
837 case AArch64ISD::CALL: return "AArch64ISD::CALL";
838 case AArch64ISD::ADRP: return "AArch64ISD::ADRP";
839 case AArch64ISD::ADDlow: return "AArch64ISD::ADDlow";
840 case AArch64ISD::LOADgot: return "AArch64ISD::LOADgot";
841 case AArch64ISD::RET_FLAG: return "AArch64ISD::RET_FLAG";
842 case AArch64ISD::BRCOND: return "AArch64ISD::BRCOND";
843 case AArch64ISD::CSEL: return "AArch64ISD::CSEL";
844 case AArch64ISD::FCSEL: return "AArch64ISD::FCSEL";
845 case AArch64ISD::CSINV: return "AArch64ISD::CSINV";
846 case AArch64ISD::CSNEG: return "AArch64ISD::CSNEG";
847 case AArch64ISD::CSINC: return "AArch64ISD::CSINC";
848 case AArch64ISD::THREAD_POINTER: return "AArch64ISD::THREAD_POINTER";
849 case AArch64ISD::TLSDESC_CALLSEQ: return "AArch64ISD::TLSDESC_CALLSEQ";
850 case AArch64ISD::ADC: return "AArch64ISD::ADC";
851 case AArch64ISD::SBC: return "AArch64ISD::SBC";
852 case AArch64ISD::ADDS: return "AArch64ISD::ADDS";
853 case AArch64ISD::SUBS: return "AArch64ISD::SUBS";
854 case AArch64ISD::ADCS: return "AArch64ISD::ADCS";
855 case AArch64ISD::SBCS: return "AArch64ISD::SBCS";
856 case AArch64ISD::ANDS: return "AArch64ISD::ANDS";
857 case AArch64ISD::CCMP: return "AArch64ISD::CCMP";
858 case AArch64ISD::CCMN: return "AArch64ISD::CCMN";
859 case AArch64ISD::FCCMP: return "AArch64ISD::FCCMP";
860 case AArch64ISD::FCMP: return "AArch64ISD::FCMP";
861 case AArch64ISD::DUP: return "AArch64ISD::DUP";
862 case AArch64ISD::DUPLANE8: return "AArch64ISD::DUPLANE8";
863 case AArch64ISD::DUPLANE16: return "AArch64ISD::DUPLANE16";
864 case AArch64ISD::DUPLANE32: return "AArch64ISD::DUPLANE32";
865 case AArch64ISD::DUPLANE64: return "AArch64ISD::DUPLANE64";
866 case AArch64ISD::MOVI: return "AArch64ISD::MOVI";
867 case AArch64ISD::MOVIshift: return "AArch64ISD::MOVIshift";
868 case AArch64ISD::MOVIedit: return "AArch64ISD::MOVIedit";
869 case AArch64ISD::MOVImsl: return "AArch64ISD::MOVImsl";
870 case AArch64ISD::FMOV: return "AArch64ISD::FMOV";
871 case AArch64ISD::MVNIshift: return "AArch64ISD::MVNIshift";
872 case AArch64ISD::MVNImsl: return "AArch64ISD::MVNImsl";
873 case AArch64ISD::BICi: return "AArch64ISD::BICi";
874 case AArch64ISD::ORRi: return "AArch64ISD::ORRi";
875 case AArch64ISD::BSL: return "AArch64ISD::BSL";
876 case AArch64ISD::NEG: return "AArch64ISD::NEG";
877 case AArch64ISD::EXTR: return "AArch64ISD::EXTR";
878 case AArch64ISD::ZIP1: return "AArch64ISD::ZIP1";
879 case AArch64ISD::ZIP2: return "AArch64ISD::ZIP2";
880 case AArch64ISD::UZP1: return "AArch64ISD::UZP1";
881 case AArch64ISD::UZP2: return "AArch64ISD::UZP2";
882 case AArch64ISD::TRN1: return "AArch64ISD::TRN1";
883 case AArch64ISD::TRN2: return "AArch64ISD::TRN2";
884 case AArch64ISD::REV16: return "AArch64ISD::REV16";
885 case AArch64ISD::REV32: return "AArch64ISD::REV32";
886 case AArch64ISD::REV64: return "AArch64ISD::REV64";
887 case AArch64ISD::EXT: return "AArch64ISD::EXT";
888 case AArch64ISD::VSHL: return "AArch64ISD::VSHL";
889 case AArch64ISD::VLSHR: return "AArch64ISD::VLSHR";
890 case AArch64ISD::VASHR: return "AArch64ISD::VASHR";
891 case AArch64ISD::CMEQ: return "AArch64ISD::CMEQ";
892 case AArch64ISD::CMGE: return "AArch64ISD::CMGE";
893 case AArch64ISD::CMGT: return "AArch64ISD::CMGT";
894 case AArch64ISD::CMHI: return "AArch64ISD::CMHI";
895 case AArch64ISD::CMHS: return "AArch64ISD::CMHS";
896 case AArch64ISD::FCMEQ: return "AArch64ISD::FCMEQ";
897 case AArch64ISD::FCMGE: return "AArch64ISD::FCMGE";
898 case AArch64ISD::FCMGT: return "AArch64ISD::FCMGT";
899 case AArch64ISD::CMEQz: return "AArch64ISD::CMEQz";
900 case AArch64ISD::CMGEz: return "AArch64ISD::CMGEz";
901 case AArch64ISD::CMGTz: return "AArch64ISD::CMGTz";
902 case AArch64ISD::CMLEz: return "AArch64ISD::CMLEz";
903 case AArch64ISD::CMLTz: return "AArch64ISD::CMLTz";
904 case AArch64ISD::FCMEQz: return "AArch64ISD::FCMEQz";
905 case AArch64ISD::FCMGEz: return "AArch64ISD::FCMGEz";
906 case AArch64ISD::FCMGTz: return "AArch64ISD::FCMGTz";
907 case AArch64ISD::FCMLEz: return "AArch64ISD::FCMLEz";
908 case AArch64ISD::FCMLTz: return "AArch64ISD::FCMLTz";
909 case AArch64ISD::SADDV: return "AArch64ISD::SADDV";
910 case AArch64ISD::UADDV: return "AArch64ISD::UADDV";
911 case AArch64ISD::SMINV: return "AArch64ISD::SMINV";
912 case AArch64ISD::UMINV: return "AArch64ISD::UMINV";
913 case AArch64ISD::SMAXV: return "AArch64ISD::SMAXV";
914 case AArch64ISD::UMAXV: return "AArch64ISD::UMAXV";
915 case AArch64ISD::NOT: return "AArch64ISD::NOT";
916 case AArch64ISD::BIT: return "AArch64ISD::BIT";
917 case AArch64ISD::CBZ: return "AArch64ISD::CBZ";
918 case AArch64ISD::CBNZ: return "AArch64ISD::CBNZ";
919 case AArch64ISD::TBZ: return "AArch64ISD::TBZ";
920 case AArch64ISD::TBNZ: return "AArch64ISD::TBNZ";
921 case AArch64ISD::TC_RETURN: return "AArch64ISD::TC_RETURN";
922 case AArch64ISD::PREFETCH: return "AArch64ISD::PREFETCH";
923 case AArch64ISD::SITOF: return "AArch64ISD::SITOF";
924 case AArch64ISD::UITOF: return "AArch64ISD::UITOF";
925 case AArch64ISD::NVCAST: return "AArch64ISD::NVCAST";
926 case AArch64ISD::SQSHL_I: return "AArch64ISD::SQSHL_I";
927 case AArch64ISD::UQSHL_I: return "AArch64ISD::UQSHL_I";
928 case AArch64ISD::SRSHR_I: return "AArch64ISD::SRSHR_I";
929 case AArch64ISD::URSHR_I: return "AArch64ISD::URSHR_I";
930 case AArch64ISD::SQSHLU_I: return "AArch64ISD::SQSHLU_I";
931 case AArch64ISD::WrapperLarge: return "AArch64ISD::WrapperLarge";
932 case AArch64ISD::LD2post: return "AArch64ISD::LD2post";
933 case AArch64ISD::LD3post: return "AArch64ISD::LD3post";
934 case AArch64ISD::LD4post: return "AArch64ISD::LD4post";
935 case AArch64ISD::ST2post: return "AArch64ISD::ST2post";
936 case AArch64ISD::ST3post: return "AArch64ISD::ST3post";
937 case AArch64ISD::ST4post: return "AArch64ISD::ST4post";
938 case AArch64ISD::LD1x2post: return "AArch64ISD::LD1x2post";
939 case AArch64ISD::LD1x3post: return "AArch64ISD::LD1x3post";
940 case AArch64ISD::LD1x4post: return "AArch64ISD::LD1x4post";
941 case AArch64ISD::ST1x2post: return "AArch64ISD::ST1x2post";
942 case AArch64ISD::ST1x3post: return "AArch64ISD::ST1x3post";
943 case AArch64ISD::ST1x4post: return "AArch64ISD::ST1x4post";
944 case AArch64ISD::LD1DUPpost: return "AArch64ISD::LD1DUPpost";
945 case AArch64ISD::LD2DUPpost: return "AArch64ISD::LD2DUPpost";
946 case AArch64ISD::LD3DUPpost: return "AArch64ISD::LD3DUPpost";
947 case AArch64ISD::LD4DUPpost: return "AArch64ISD::LD4DUPpost";
948 case AArch64ISD::LD1LANEpost: return "AArch64ISD::LD1LANEpost";
949 case AArch64ISD::LD2LANEpost: return "AArch64ISD::LD2LANEpost";
950 case AArch64ISD::LD3LANEpost: return "AArch64ISD::LD3LANEpost";
951 case AArch64ISD::LD4LANEpost: return "AArch64ISD::LD4LANEpost";
952 case AArch64ISD::ST2LANEpost: return "AArch64ISD::ST2LANEpost";
953 case AArch64ISD::ST3LANEpost: return "AArch64ISD::ST3LANEpost";
954 case AArch64ISD::ST4LANEpost: return "AArch64ISD::ST4LANEpost";
955 case AArch64ISD::SMULL: return "AArch64ISD::SMULL";
956 case AArch64ISD::UMULL: return "AArch64ISD::UMULL";
962 AArch64TargetLowering::EmitF128CSEL(MachineInstr *MI,
963 MachineBasicBlock *MBB) const {
964 // We materialise the F128CSEL pseudo-instruction as some control flow and a
968 // [... previous instrs leading to comparison ...]
974 // Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB]
976 MachineFunction *MF = MBB->getParent();
977 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
978 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
979 DebugLoc DL = MI->getDebugLoc();
980 MachineFunction::iterator It = ++MBB->getIterator();
982 unsigned DestReg = MI->getOperand(0).getReg();
983 unsigned IfTrueReg = MI->getOperand(1).getReg();
984 unsigned IfFalseReg = MI->getOperand(2).getReg();
985 unsigned CondCode = MI->getOperand(3).getImm();
986 bool NZCVKilled = MI->getOperand(4).isKill();
988 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
989 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
990 MF->insert(It, TrueBB);
991 MF->insert(It, EndBB);
993 // Transfer rest of current basic-block to EndBB
994 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
996 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
998 BuildMI(MBB, DL, TII->get(AArch64::Bcc)).addImm(CondCode).addMBB(TrueBB);
999 BuildMI(MBB, DL, TII->get(AArch64::B)).addMBB(EndBB);
1000 MBB->addSuccessor(TrueBB);
1001 MBB->addSuccessor(EndBB);
1003 // TrueBB falls through to the end.
1004 TrueBB->addSuccessor(EndBB);
1007 TrueBB->addLiveIn(AArch64::NZCV);
1008 EndBB->addLiveIn(AArch64::NZCV);
1011 BuildMI(*EndBB, EndBB->begin(), DL, TII->get(AArch64::PHI), DestReg)
1017 MI->eraseFromParent();
1022 AArch64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
1023 MachineBasicBlock *BB) const {
1024 switch (MI->getOpcode()) {
1029 llvm_unreachable("Unexpected instruction for custom inserter!");
1031 case AArch64::F128CSEL:
1032 return EmitF128CSEL(MI, BB);
1034 case TargetOpcode::STACKMAP:
1035 case TargetOpcode::PATCHPOINT:
1036 return emitPatchPoint(MI, BB);
1040 //===----------------------------------------------------------------------===//
1041 // AArch64 Lowering private implementation.
1042 //===----------------------------------------------------------------------===//
1044 //===----------------------------------------------------------------------===//
1046 //===----------------------------------------------------------------------===//
1048 /// changeIntCCToAArch64CC - Convert a DAG integer condition code to an AArch64
1050 static AArch64CC::CondCode changeIntCCToAArch64CC(ISD::CondCode CC) {
1053 llvm_unreachable("Unknown condition code!");
1055 return AArch64CC::NE;
1057 return AArch64CC::EQ;
1059 return AArch64CC::GT;
1061 return AArch64CC::GE;
1063 return AArch64CC::LT;
1065 return AArch64CC::LE;
1067 return AArch64CC::HI;
1069 return AArch64CC::HS;
1071 return AArch64CC::LO;
1073 return AArch64CC::LS;
1077 /// changeFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64 CC.
1078 static void changeFPCCToAArch64CC(ISD::CondCode CC,
1079 AArch64CC::CondCode &CondCode,
1080 AArch64CC::CondCode &CondCode2) {
1081 CondCode2 = AArch64CC::AL;
1084 llvm_unreachable("Unknown FP condition!");
1087 CondCode = AArch64CC::EQ;
1091 CondCode = AArch64CC::GT;
1095 CondCode = AArch64CC::GE;
1098 CondCode = AArch64CC::MI;
1101 CondCode = AArch64CC::LS;
1104 CondCode = AArch64CC::MI;
1105 CondCode2 = AArch64CC::GT;
1108 CondCode = AArch64CC::VC;
1111 CondCode = AArch64CC::VS;
1114 CondCode = AArch64CC::EQ;
1115 CondCode2 = AArch64CC::VS;
1118 CondCode = AArch64CC::HI;
1121 CondCode = AArch64CC::PL;
1125 CondCode = AArch64CC::LT;
1129 CondCode = AArch64CC::LE;
1133 CondCode = AArch64CC::NE;
1138 /// changeVectorFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64
1139 /// CC usable with the vector instructions. Fewer operations are available
1140 /// without a real NZCV register, so we have to use less efficient combinations
1141 /// to get the same effect.
1142 static void changeVectorFPCCToAArch64CC(ISD::CondCode CC,
1143 AArch64CC::CondCode &CondCode,
1144 AArch64CC::CondCode &CondCode2,
1149 // Mostly the scalar mappings work fine.
1150 changeFPCCToAArch64CC(CC, CondCode, CondCode2);
1153 Invert = true; // Fallthrough
1155 CondCode = AArch64CC::MI;
1156 CondCode2 = AArch64CC::GE;
1163 // All of the compare-mask comparisons are ordered, but we can switch
1164 // between the two by a double inversion. E.g. ULE == !OGT.
1166 changeFPCCToAArch64CC(getSetCCInverse(CC, false), CondCode, CondCode2);
1171 static bool isLegalArithImmed(uint64_t C) {
1172 // Matches AArch64DAGToDAGISel::SelectArithImmed().
1173 return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
1176 static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1177 SDLoc dl, SelectionDAG &DAG) {
1178 EVT VT = LHS.getValueType();
1180 if (VT.isFloatingPoint()) {
1181 assert(VT != MVT::f128);
1182 if (VT == MVT::f16) {
1183 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, LHS);
1184 RHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, RHS);
1186 return DAG.getNode(AArch64ISD::FCMP, dl, VT, LHS, RHS);
1189 // The CMP instruction is just an alias for SUBS, and representing it as
1190 // SUBS means that it's possible to get CSE with subtract operations.
1191 // A later phase can perform the optimization of setting the destination
1192 // register to WZR/XZR if it ends up being unused.
1193 unsigned Opcode = AArch64ISD::SUBS;
1195 if (RHS.getOpcode() == ISD::SUB && isNullConstant(RHS.getOperand(0)) &&
1196 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1197 // We'd like to combine a (CMP op1, (sub 0, op2) into a CMN instruction on
1198 // the grounds that "op1 - (-op2) == op1 + op2". However, the C and V flags
1199 // can be set differently by this operation. It comes down to whether
1200 // "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are then
1201 // everything is fine. If not then the optimization is wrong. Thus general
1202 // comparisons are only valid if op2 != 0.
1204 // So, finally, the only LLVM-native comparisons that don't mention C and V
1205 // are SETEQ and SETNE. They're the only ones we can safely use CMN for in
1206 // the absence of information about op2.
1207 Opcode = AArch64ISD::ADDS;
1208 RHS = RHS.getOperand(1);
1209 } else if (LHS.getOpcode() == ISD::AND && isNullConstant(RHS) &&
1210 !isUnsignedIntSetCC(CC)) {
1211 // Similarly, (CMP (and X, Y), 0) can be implemented with a TST
1212 // (a.k.a. ANDS) except that the flags are only guaranteed to work for one
1213 // of the signed comparisons.
1214 Opcode = AArch64ISD::ANDS;
1215 RHS = LHS.getOperand(1);
1216 LHS = LHS.getOperand(0);
1219 return DAG.getNode(Opcode, dl, DAG.getVTList(VT, MVT_CC), LHS, RHS)
1223 /// \defgroup AArch64CCMP CMP;CCMP matching
1225 /// These functions deal with the formation of CMP;CCMP;... sequences.
1226 /// The CCMP/CCMN/FCCMP/FCCMPE instructions allow the conditional execution of
1227 /// a comparison. They set the NZCV flags to a predefined value if their
1228 /// predicate is false. This allows to express arbitrary conjunctions, for
1229 /// example "cmp 0 (and (setCA (cmp A)) (setCB (cmp B))))"
1232 /// ccmp B, inv(CB), CA
1233 /// check for CB flags
1235 /// In general we can create code for arbitrary "... (and (and A B) C)"
1236 /// sequences. We can also implement some "or" expressions, because "(or A B)"
1237 /// is equivalent to "not (and (not A) (not B))" and we can implement some
1238 /// negation operations:
1239 /// We can negate the results of a single comparison by inverting the flags
1240 /// used when the predicate fails and inverting the flags tested in the next
1241 /// instruction; We can also negate the results of the whole previous
1242 /// conditional compare sequence by inverting the flags tested in the next
1243 /// instruction. However there is no way to negate the result of a partial
1246 /// Therefore on encountering an "or" expression we can negate the subtree on
1247 /// one side and have to be able to push the negate to the leafs of the subtree
1248 /// on the other side (see also the comments in code). As complete example:
1249 /// "or (or (setCA (cmp A)) (setCB (cmp B)))
1250 /// (and (setCC (cmp C)) (setCD (cmp D)))"
1251 /// is transformed to
1252 /// "not (and (not (and (setCC (cmp C)) (setCC (cmp D))))
1253 /// (and (not (setCA (cmp A)) (not (setCB (cmp B))))))"
1254 /// and implemented as:
1256 /// ccmp D, inv(CD), CC
1257 /// ccmp A, CA, inv(CD)
1258 /// ccmp B, CB, inv(CA)
1259 /// check for CB flags
1260 /// A counterexample is "or (and A B) (and C D)" which cannot be implemented
1261 /// by conditional compare sequences.
1264 /// Create a conditional comparison; Use CCMP, CCMN or FCCMP as appropriate.
1265 static SDValue emitConditionalComparison(SDValue LHS, SDValue RHS,
1266 ISD::CondCode CC, SDValue CCOp,
1267 SDValue Condition, unsigned NZCV,
1268 SDLoc DL, SelectionDAG &DAG) {
1269 unsigned Opcode = 0;
1270 if (LHS.getValueType().isFloatingPoint()) {
1271 assert(LHS.getValueType() != MVT::f128);
1272 if (LHS.getValueType() == MVT::f16) {
1273 LHS = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, LHS);
1274 RHS = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, RHS);
1276 Opcode = AArch64ISD::FCCMP;
1277 } else if (RHS.getOpcode() == ISD::SUB) {
1278 SDValue SubOp0 = RHS.getOperand(0);
1279 if (isNullConstant(SubOp0) && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1280 // See emitComparison() on why we can only do this for SETEQ and SETNE.
1281 Opcode = AArch64ISD::CCMN;
1282 RHS = RHS.getOperand(1);
1286 Opcode = AArch64ISD::CCMP;
1288 SDValue NZCVOp = DAG.getConstant(NZCV, DL, MVT::i32);
1289 return DAG.getNode(Opcode, DL, MVT_CC, LHS, RHS, NZCVOp, Condition, CCOp);
1292 /// Returns true if @p Val is a tree of AND/OR/SETCC operations.
1293 /// CanPushNegate is set to true if we can push a negate operation through
1294 /// the tree in a was that we are left with AND operations and negate operations
1295 /// at the leafs only. i.e. "not (or (or x y) z)" can be changed to
1296 /// "and (and (not x) (not y)) (not z)"; "not (or (and x y) z)" cannot be
1297 /// brought into such a form.
1298 static bool isConjunctionDisjunctionTree(const SDValue Val, bool &CanPushNegate,
1299 unsigned Depth = 0) {
1300 if (!Val.hasOneUse())
1302 unsigned Opcode = Val->getOpcode();
1303 if (Opcode == ISD::SETCC) {
1304 if (Val->getOperand(0).getValueType() == MVT::f128)
1306 CanPushNegate = true;
1309 // Protect against stack overflow.
1312 if (Opcode == ISD::AND || Opcode == ISD::OR) {
1313 SDValue O0 = Val->getOperand(0);
1314 SDValue O1 = Val->getOperand(1);
1315 bool CanPushNegateL;
1316 if (!isConjunctionDisjunctionTree(O0, CanPushNegateL, Depth+1))
1318 bool CanPushNegateR;
1319 if (!isConjunctionDisjunctionTree(O1, CanPushNegateR, Depth+1))
1321 // We cannot push a negate through an AND operation (it would become an OR),
1322 // we can however change a (not (or x y)) to (and (not x) (not y)) if we can
1323 // push the negate through the x/y subtrees.
1324 CanPushNegate = (Opcode == ISD::OR) && CanPushNegateL && CanPushNegateR;
1330 /// Emit conjunction or disjunction tree with the CMP/FCMP followed by a chain
1331 /// of CCMP/CFCMP ops. See @ref AArch64CCMP.
1332 /// Tries to transform the given i1 producing node @p Val to a series compare
1333 /// and conditional compare operations. @returns an NZCV flags producing node
1334 /// and sets @p OutCC to the flags that should be tested or returns SDValue() if
1335 /// transformation was not possible.
1336 /// On recursive invocations @p PushNegate may be set to true to have negation
1337 /// effects pushed to the tree leafs; @p Predicate is an NZCV flag predicate
1338 /// for the comparisons in the current subtree; @p Depth limits the search
1339 /// depth to avoid stack overflow.
1340 static SDValue emitConjunctionDisjunctionTree(SelectionDAG &DAG, SDValue Val,
1341 AArch64CC::CondCode &OutCC, bool PushNegate = false,
1342 SDValue CCOp = SDValue(), AArch64CC::CondCode Predicate = AArch64CC::AL,
1343 unsigned Depth = 0) {
1344 // We're at a tree leaf, produce a conditional comparison operation.
1345 unsigned Opcode = Val->getOpcode();
1346 if (Opcode == ISD::SETCC) {
1347 SDValue LHS = Val->getOperand(0);
1348 SDValue RHS = Val->getOperand(1);
1349 ISD::CondCode CC = cast<CondCodeSDNode>(Val->getOperand(2))->get();
1350 bool isInteger = LHS.getValueType().isInteger();
1352 CC = getSetCCInverse(CC, isInteger);
1354 // Determine OutCC and handle FP special case.
1356 OutCC = changeIntCCToAArch64CC(CC);
1358 assert(LHS.getValueType().isFloatingPoint());
1359 AArch64CC::CondCode ExtraCC;
1360 changeFPCCToAArch64CC(CC, OutCC, ExtraCC);
1361 // Surpisingly some floating point conditions can't be tested with a
1362 // single condition code. Construct an additional comparison in this case.
1363 // See comment below on how we deal with OR conditions.
1364 if (ExtraCC != AArch64CC::AL) {
1366 if (!CCOp.getNode())
1367 ExtraCmp = emitComparison(LHS, RHS, CC, DL, DAG);
1369 SDValue ConditionOp = DAG.getConstant(Predicate, DL, MVT_CC);
1370 // Note that we want the inverse of ExtraCC, so NZCV is not inversed.
1371 unsigned NZCV = AArch64CC::getNZCVToSatisfyCondCode(ExtraCC);
1372 ExtraCmp = emitConditionalComparison(LHS, RHS, CC, CCOp, ConditionOp,
1376 Predicate = AArch64CC::getInvertedCondCode(ExtraCC);
1377 OutCC = AArch64CC::getInvertedCondCode(OutCC);
1381 // Produce a normal comparison if we are first in the chain
1382 if (!CCOp.getNode())
1383 return emitComparison(LHS, RHS, CC, DL, DAG);
1384 // Otherwise produce a ccmp.
1385 SDValue ConditionOp = DAG.getConstant(Predicate, DL, MVT_CC);
1386 AArch64CC::CondCode InvOutCC = AArch64CC::getInvertedCondCode(OutCC);
1387 unsigned NZCV = AArch64CC::getNZCVToSatisfyCondCode(InvOutCC);
1388 return emitConditionalComparison(LHS, RHS, CC, CCOp, ConditionOp, NZCV, DL,
1390 } else if ((Opcode != ISD::AND && Opcode != ISD::OR) || !Val->hasOneUse())
1393 assert((Opcode == ISD::OR || !PushNegate)
1394 && "Can only push negate through OR operation");
1396 // Check if both sides can be transformed.
1397 SDValue LHS = Val->getOperand(0);
1398 SDValue RHS = Val->getOperand(1);
1399 bool CanPushNegateL;
1400 if (!isConjunctionDisjunctionTree(LHS, CanPushNegateL, Depth+1))
1402 bool CanPushNegateR;
1403 if (!isConjunctionDisjunctionTree(RHS, CanPushNegateR, Depth+1))
1406 // Do we need to negate our operands?
1407 bool NegateOperands = Opcode == ISD::OR;
1408 // We can negate the results of all previous operations by inverting the
1409 // predicate flags giving us a free negation for one side. For the other side
1410 // we need to be able to push the negation to the leafs of the tree.
1411 if (NegateOperands) {
1412 if (!CanPushNegateL && !CanPushNegateR)
1414 // Order the side where we can push the negate through to LHS.
1415 if (!CanPushNegateL && CanPushNegateR)
1416 std::swap(LHS, RHS);
1418 bool NeedsNegOutL = LHS->getOpcode() == ISD::OR;
1419 bool NeedsNegOutR = RHS->getOpcode() == ISD::OR;
1420 if (NeedsNegOutL && NeedsNegOutR)
1422 // Order the side where we need to negate the output flags to RHS so it
1423 // gets emitted first.
1425 std::swap(LHS, RHS);
1428 // Emit RHS. If we want to negate the tree we only need to push a negate
1429 // through if we are already in a PushNegate case, otherwise we can negate
1430 // the "flags to test" afterwards.
1431 AArch64CC::CondCode RHSCC;
1432 SDValue CmpR = emitConjunctionDisjunctionTree(DAG, RHS, RHSCC, PushNegate,
1433 CCOp, Predicate, Depth+1);
1434 if (NegateOperands && !PushNegate)
1435 RHSCC = AArch64CC::getInvertedCondCode(RHSCC);
1436 // Emit LHS. We must push the negate through if we need to negate it.
1437 SDValue CmpL = emitConjunctionDisjunctionTree(DAG, LHS, OutCC, NegateOperands,
1438 CmpR, RHSCC, Depth+1);
1439 // If we transformed an OR to and AND then we have to negate the result
1440 // (or absorb a PushNegate resulting in a double negation).
1441 if (Opcode == ISD::OR && !PushNegate)
1442 OutCC = AArch64CC::getInvertedCondCode(OutCC);
1448 static SDValue getAArch64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1449 SDValue &AArch64cc, SelectionDAG &DAG, SDLoc dl) {
1450 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1451 EVT VT = RHS.getValueType();
1452 uint64_t C = RHSC->getZExtValue();
1453 if (!isLegalArithImmed(C)) {
1454 // Constant does not fit, try adjusting it by one?
1460 if ((VT == MVT::i32 && C != 0x80000000 &&
1461 isLegalArithImmed((uint32_t)(C - 1))) ||
1462 (VT == MVT::i64 && C != 0x80000000ULL &&
1463 isLegalArithImmed(C - 1ULL))) {
1464 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1465 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1466 RHS = DAG.getConstant(C, dl, VT);
1471 if ((VT == MVT::i32 && C != 0 &&
1472 isLegalArithImmed((uint32_t)(C - 1))) ||
1473 (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) {
1474 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1475 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1476 RHS = DAG.getConstant(C, dl, VT);
1481 if ((VT == MVT::i32 && C != INT32_MAX &&
1482 isLegalArithImmed((uint32_t)(C + 1))) ||
1483 (VT == MVT::i64 && C != INT64_MAX &&
1484 isLegalArithImmed(C + 1ULL))) {
1485 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1486 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1487 RHS = DAG.getConstant(C, dl, VT);
1492 if ((VT == MVT::i32 && C != UINT32_MAX &&
1493 isLegalArithImmed((uint32_t)(C + 1))) ||
1494 (VT == MVT::i64 && C != UINT64_MAX &&
1495 isLegalArithImmed(C + 1ULL))) {
1496 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1497 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1498 RHS = DAG.getConstant(C, dl, VT);
1505 AArch64CC::CondCode AArch64CC;
1506 if ((CC == ISD::SETEQ || CC == ISD::SETNE) && isa<ConstantSDNode>(RHS)) {
1507 const ConstantSDNode *RHSC = cast<ConstantSDNode>(RHS);
1509 // The imm operand of ADDS is an unsigned immediate, in the range 0 to 4095.
1510 // For the i8 operand, the largest immediate is 255, so this can be easily
1511 // encoded in the compare instruction. For the i16 operand, however, the
1512 // largest immediate cannot be encoded in the compare.
1513 // Therefore, use a sign extending load and cmn to avoid materializing the
1514 // -1 constant. For example,
1516 // ldrh w0, [x0, #0]
1519 // ldrsh w0, [x0, #0]
1521 // Fundamental, we're relying on the property that (zext LHS) == (zext RHS)
1522 // if and only if (sext LHS) == (sext RHS). The checks are in place to
1523 // ensure both the LHS and RHS are truly zero extended and to make sure the
1524 // transformation is profitable.
1525 if ((RHSC->getZExtValue() >> 16 == 0) && isa<LoadSDNode>(LHS) &&
1526 cast<LoadSDNode>(LHS)->getExtensionType() == ISD::ZEXTLOAD &&
1527 cast<LoadSDNode>(LHS)->getMemoryVT() == MVT::i16 &&
1528 LHS.getNode()->hasNUsesOfValue(1, 0)) {
1529 int16_t ValueofRHS = cast<ConstantSDNode>(RHS)->getZExtValue();
1530 if (ValueofRHS < 0 && isLegalArithImmed(-ValueofRHS)) {
1532 DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, LHS.getValueType(), LHS,
1533 DAG.getValueType(MVT::i16));
1534 Cmp = emitComparison(SExt, DAG.getConstant(ValueofRHS, dl,
1535 RHS.getValueType()),
1537 AArch64CC = changeIntCCToAArch64CC(CC);
1541 if (!Cmp && (RHSC->isNullValue() || RHSC->isOne())) {
1542 if ((Cmp = emitConjunctionDisjunctionTree(DAG, LHS, AArch64CC))) {
1543 if ((CC == ISD::SETNE) ^ RHSC->isNullValue())
1544 AArch64CC = AArch64CC::getInvertedCondCode(AArch64CC);
1550 Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
1551 AArch64CC = changeIntCCToAArch64CC(CC);
1553 AArch64cc = DAG.getConstant(AArch64CC, dl, MVT_CC);
1557 static std::pair<SDValue, SDValue>
1558 getAArch64XALUOOp(AArch64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
1559 assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
1560 "Unsupported value type");
1561 SDValue Value, Overflow;
1563 SDValue LHS = Op.getOperand(0);
1564 SDValue RHS = Op.getOperand(1);
1566 switch (Op.getOpcode()) {
1568 llvm_unreachable("Unknown overflow instruction!");
1570 Opc = AArch64ISD::ADDS;
1574 Opc = AArch64ISD::ADDS;
1578 Opc = AArch64ISD::SUBS;
1582 Opc = AArch64ISD::SUBS;
1585 // Multiply needs a little bit extra work.
1589 bool IsSigned = Op.getOpcode() == ISD::SMULO;
1590 if (Op.getValueType() == MVT::i32) {
1591 unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1592 // For a 32 bit multiply with overflow check we want the instruction
1593 // selector to generate a widening multiply (SMADDL/UMADDL). For that we
1594 // need to generate the following pattern:
1595 // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b))
1596 LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
1597 RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
1598 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1599 SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul,
1600 DAG.getConstant(0, DL, MVT::i64));
1601 // On AArch64 the upper 32 bits are always zero extended for a 32 bit
1602 // operation. We need to clear out the upper 32 bits, because we used a
1603 // widening multiply that wrote all 64 bits. In the end this should be a
1605 Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add);
1607 // The signed overflow check requires more than just a simple check for
1608 // any bit set in the upper 32 bits of the result. These bits could be
1609 // just the sign bits of a negative number. To perform the overflow
1610 // check we have to arithmetic shift right the 32nd bit of the result by
1611 // 31 bits. Then we compare the result to the upper 32 bits.
1612 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add,
1613 DAG.getConstant(32, DL, MVT::i64));
1614 UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits);
1615 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value,
1616 DAG.getConstant(31, DL, MVT::i64));
1617 // It is important that LowerBits is last, otherwise the arithmetic
1618 // shift will not be folded into the compare (SUBS).
1619 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32);
1620 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1623 // The overflow check for unsigned multiply is easy. We only need to
1624 // check if any of the upper 32 bits are set. This can be done with a
1625 // CMP (shifted register). For that we need to generate the following
1627 // (i64 AArch64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32)
1628 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
1629 DAG.getConstant(32, DL, MVT::i64));
1630 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1632 DAG.getNode(AArch64ISD::SUBS, DL, VTs,
1633 DAG.getConstant(0, DL, MVT::i64),
1634 UpperBits).getValue(1);
1638 assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
1639 // For the 64 bit multiply
1640 Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1642 SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
1643 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
1644 DAG.getConstant(63, DL, MVT::i64));
1645 // It is important that LowerBits is last, otherwise the arithmetic
1646 // shift will not be folded into the compare (SUBS).
1647 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1648 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1651 SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
1652 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1654 DAG.getNode(AArch64ISD::SUBS, DL, VTs,
1655 DAG.getConstant(0, DL, MVT::i64),
1656 UpperBits).getValue(1);
1663 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);
1665 // Emit the AArch64 operation with overflow check.
1666 Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
1667 Overflow = Value.getValue(1);
1669 return std::make_pair(Value, Overflow);
1672 SDValue AArch64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG,
1673 RTLIB::Libcall Call) const {
1674 SmallVector<SDValue, 2> Ops(Op->op_begin(), Op->op_end());
1675 return makeLibCall(DAG, Call, MVT::f128, Ops, false, SDLoc(Op)).first;
1678 static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) {
1679 SDValue Sel = Op.getOperand(0);
1680 SDValue Other = Op.getOperand(1);
1682 // If neither operand is a SELECT_CC, give up.
1683 if (Sel.getOpcode() != ISD::SELECT_CC)
1684 std::swap(Sel, Other);
1685 if (Sel.getOpcode() != ISD::SELECT_CC)
1688 // The folding we want to perform is:
1689 // (xor x, (select_cc a, b, cc, 0, -1) )
1691 // (csel x, (xor x, -1), cc ...)
1693 // The latter will get matched to a CSINV instruction.
1695 ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
1696 SDValue LHS = Sel.getOperand(0);
1697 SDValue RHS = Sel.getOperand(1);
1698 SDValue TVal = Sel.getOperand(2);
1699 SDValue FVal = Sel.getOperand(3);
1702 // FIXME: This could be generalized to non-integer comparisons.
1703 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
1706 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
1707 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
1709 // The values aren't constants, this isn't the pattern we're looking for.
1710 if (!CFVal || !CTVal)
1713 // We can commute the SELECT_CC by inverting the condition. This
1714 // might be needed to make this fit into a CSINV pattern.
1715 if (CTVal->isAllOnesValue() && CFVal->isNullValue()) {
1716 std::swap(TVal, FVal);
1717 std::swap(CTVal, CFVal);
1718 CC = ISD::getSetCCInverse(CC, true);
1721 // If the constants line up, perform the transform!
1722 if (CTVal->isNullValue() && CFVal->isAllOnesValue()) {
1724 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
1727 TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other,
1728 DAG.getConstant(-1ULL, dl, Other.getValueType()));
1730 return DAG.getNode(AArch64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal,
1737 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
1738 EVT VT = Op.getValueType();
1740 // Let legalize expand this if it isn't a legal type yet.
1741 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
1744 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
1747 bool ExtraOp = false;
1748 switch (Op.getOpcode()) {
1750 llvm_unreachable("Invalid code");
1752 Opc = AArch64ISD::ADDS;
1755 Opc = AArch64ISD::SUBS;
1758 Opc = AArch64ISD::ADCS;
1762 Opc = AArch64ISD::SBCS;
1768 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1));
1769 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1),
1773 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
1774 // Let legalize expand this if it isn't a legal type yet.
1775 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
1779 AArch64CC::CondCode CC;
1780 // The actual operation that sets the overflow or carry flag.
1781 SDValue Value, Overflow;
1782 std::tie(Value, Overflow) = getAArch64XALUOOp(CC, Op, DAG);
1784 // We use 0 and 1 as false and true values.
1785 SDValue TVal = DAG.getConstant(1, dl, MVT::i32);
1786 SDValue FVal = DAG.getConstant(0, dl, MVT::i32);
1788 // We use an inverted condition, because the conditional select is inverted
1789 // too. This will allow it to be selected to a single instruction:
1790 // CSINC Wd, WZR, WZR, invert(cond).
1791 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), dl, MVT::i32);
1792 Overflow = DAG.getNode(AArch64ISD::CSEL, dl, MVT::i32, FVal, TVal,
1795 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
1796 return DAG.getNode(ISD::MERGE_VALUES, dl, VTs, Value, Overflow);
1799 // Prefetch operands are:
1800 // 1: Address to prefetch
1802 // 3: int locality (0 = no locality ... 3 = extreme locality)
1803 // 4: bool isDataCache
1804 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
1806 unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
1807 unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
1808 unsigned IsData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
1810 bool IsStream = !Locality;
1811 // When the locality number is set
1813 // The front-end should have filtered out the out-of-range values
1814 assert(Locality <= 3 && "Prefetch locality out-of-range");
1815 // The locality degree is the opposite of the cache speed.
1816 // Put the number the other way around.
1817 // The encoding starts at 0 for level 1
1818 Locality = 3 - Locality;
1821 // built the mask value encoding the expected behavior.
1822 unsigned PrfOp = (IsWrite << 4) | // Load/Store bit
1823 (!IsData << 3) | // IsDataCache bit
1824 (Locality << 1) | // Cache level bits
1825 (unsigned)IsStream; // Stream bit
1826 return DAG.getNode(AArch64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
1827 DAG.getConstant(PrfOp, DL, MVT::i32), Op.getOperand(1));
1830 SDValue AArch64TargetLowering::LowerFP_EXTEND(SDValue Op,
1831 SelectionDAG &DAG) const {
1832 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
1835 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
1837 return LowerF128Call(Op, DAG, LC);
1840 SDValue AArch64TargetLowering::LowerFP_ROUND(SDValue Op,
1841 SelectionDAG &DAG) const {
1842 if (Op.getOperand(0).getValueType() != MVT::f128) {
1843 // It's legal except when f128 is involved
1848 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
1850 // FP_ROUND node has a second operand indicating whether it is known to be
1851 // precise. That doesn't take part in the LibCall so we can't directly use
1853 SDValue SrcVal = Op.getOperand(0);
1854 return makeLibCall(DAG, LC, Op.getValueType(), SrcVal, /*isSigned*/ false,
1858 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1859 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1860 // Any additional optimization in this function should be recorded
1861 // in the cost tables.
1862 EVT InVT = Op.getOperand(0).getValueType();
1863 EVT VT = Op.getValueType();
1864 unsigned NumElts = InVT.getVectorNumElements();
1866 // f16 vectors are promoted to f32 before a conversion.
1867 if (InVT.getVectorElementType() == MVT::f16) {
1868 MVT NewVT = MVT::getVectorVT(MVT::f32, NumElts);
1871 Op.getOpcode(), dl, Op.getValueType(),
1872 DAG.getNode(ISD::FP_EXTEND, dl, NewVT, Op.getOperand(0)));
1875 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1878 DAG.getNode(Op.getOpcode(), dl, InVT.changeVectorElementTypeToInteger(),
1880 return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
1883 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1886 MVT::getVectorVT(MVT::getFloatingPointVT(VT.getScalarSizeInBits()),
1887 VT.getVectorNumElements());
1888 SDValue Ext = DAG.getNode(ISD::FP_EXTEND, dl, ExtVT, Op.getOperand(0));
1889 return DAG.getNode(Op.getOpcode(), dl, VT, Ext);
1892 // Type changing conversions are illegal.
1896 SDValue AArch64TargetLowering::LowerFP_TO_INT(SDValue Op,
1897 SelectionDAG &DAG) const {
1898 if (Op.getOperand(0).getValueType().isVector())
1899 return LowerVectorFP_TO_INT(Op, DAG);
1901 // f16 conversions are promoted to f32.
1902 if (Op.getOperand(0).getValueType() == MVT::f16) {
1905 Op.getOpcode(), dl, Op.getValueType(),
1906 DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, Op.getOperand(0)));
1909 if (Op.getOperand(0).getValueType() != MVT::f128) {
1910 // It's legal except when f128 is involved
1915 if (Op.getOpcode() == ISD::FP_TO_SINT)
1916 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
1918 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
1920 SmallVector<SDValue, 2> Ops(Op->op_begin(), Op->op_end());
1921 return makeLibCall(DAG, LC, Op.getValueType(), Ops, false, SDLoc(Op)).first;
1924 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1925 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1926 // Any additional optimization in this function should be recorded
1927 // in the cost tables.
1928 EVT VT = Op.getValueType();
1930 SDValue In = Op.getOperand(0);
1931 EVT InVT = In.getValueType();
1933 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1935 MVT::getVectorVT(MVT::getFloatingPointVT(InVT.getScalarSizeInBits()),
1936 InVT.getVectorNumElements());
1937 In = DAG.getNode(Op.getOpcode(), dl, CastVT, In);
1938 return DAG.getNode(ISD::FP_ROUND, dl, VT, In, DAG.getIntPtrConstant(0, dl));
1941 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1943 Op.getOpcode() == ISD::SINT_TO_FP ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1944 EVT CastVT = VT.changeVectorElementTypeToInteger();
1945 In = DAG.getNode(CastOpc, dl, CastVT, In);
1946 return DAG.getNode(Op.getOpcode(), dl, VT, In);
1952 SDValue AArch64TargetLowering::LowerINT_TO_FP(SDValue Op,
1953 SelectionDAG &DAG) const {
1954 if (Op.getValueType().isVector())
1955 return LowerVectorINT_TO_FP(Op, DAG);
1957 // f16 conversions are promoted to f32.
1958 if (Op.getValueType() == MVT::f16) {
1961 ISD::FP_ROUND, dl, MVT::f16,
1962 DAG.getNode(Op.getOpcode(), dl, MVT::f32, Op.getOperand(0)),
1963 DAG.getIntPtrConstant(0, dl));
1966 // i128 conversions are libcalls.
1967 if (Op.getOperand(0).getValueType() == MVT::i128)
1970 // Other conversions are legal, unless it's to the completely software-based
1972 if (Op.getValueType() != MVT::f128)
1976 if (Op.getOpcode() == ISD::SINT_TO_FP)
1977 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1979 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1981 return LowerF128Call(Op, DAG, LC);
1984 SDValue AArch64TargetLowering::LowerFSINCOS(SDValue Op,
1985 SelectionDAG &DAG) const {
1986 // For iOS, we want to call an alternative entry point: __sincos_stret,
1987 // which returns the values in two S / D registers.
1989 SDValue Arg = Op.getOperand(0);
1990 EVT ArgVT = Arg.getValueType();
1991 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1998 Entry.isSExt = false;
1999 Entry.isZExt = false;
2000 Args.push_back(Entry);
2002 const char *LibcallName =
2003 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
2005 DAG.getExternalSymbol(LibcallName, getPointerTy(DAG.getDataLayout()));
2007 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
2008 TargetLowering::CallLoweringInfo CLI(DAG);
2009 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
2010 .setCallee(CallingConv::Fast, RetTy, Callee, std::move(Args), 0);
2012 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2013 return CallResult.first;
2016 static SDValue LowerBITCAST(SDValue Op, SelectionDAG &DAG) {
2017 if (Op.getValueType() != MVT::f16)
2020 assert(Op.getOperand(0).getValueType() == MVT::i16);
2023 Op = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op.getOperand(0));
2024 Op = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Op);
2026 DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL, MVT::f16, Op,
2027 DAG.getTargetConstant(AArch64::hsub, DL, MVT::i32)),
2031 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
2032 if (OrigVT.getSizeInBits() >= 64)
2035 assert(OrigVT.isSimple() && "Expecting a simple value type");
2037 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
2038 switch (OrigSimpleTy) {
2039 default: llvm_unreachable("Unexpected Vector Type");
2048 static SDValue addRequiredExtensionForVectorMULL(SDValue N, SelectionDAG &DAG,
2051 unsigned ExtOpcode) {
2052 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
2053 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
2054 // 64-bits we need to insert a new extension so that it will be 64-bits.
2055 assert(ExtTy.is128BitVector() && "Unexpected extension size");
2056 if (OrigTy.getSizeInBits() >= 64)
2059 // Must extend size to at least 64 bits to be used as an operand for VMULL.
2060 EVT NewVT = getExtensionTo64Bits(OrigTy);
2062 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
2065 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
2067 EVT VT = N->getValueType(0);
2069 if (N->getOpcode() != ISD::BUILD_VECTOR)
2072 for (const SDValue &Elt : N->op_values()) {
2073 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
2074 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
2075 unsigned HalfSize = EltSize / 2;
2077 if (!isIntN(HalfSize, C->getSExtValue()))
2080 if (!isUIntN(HalfSize, C->getZExtValue()))
2091 static SDValue skipExtensionForVectorMULL(SDNode *N, SelectionDAG &DAG) {
2092 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
2093 return addRequiredExtensionForVectorMULL(N->getOperand(0), DAG,
2094 N->getOperand(0)->getValueType(0),
2098 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
2099 EVT VT = N->getValueType(0);
2101 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
2102 unsigned NumElts = VT.getVectorNumElements();
2103 MVT TruncVT = MVT::getIntegerVT(EltSize);
2104 SmallVector<SDValue, 8> Ops;
2105 for (unsigned i = 0; i != NumElts; ++i) {
2106 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
2107 const APInt &CInt = C->getAPIntValue();
2108 // Element types smaller than 32 bits are not legal, so use i32 elements.
2109 // The values are implicitly truncated so sext vs. zext doesn't matter.
2110 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), dl, MVT::i32));
2112 return DAG.getNode(ISD::BUILD_VECTOR, dl,
2113 MVT::getVectorVT(TruncVT, NumElts), Ops);
2116 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
2117 if (N->getOpcode() == ISD::SIGN_EXTEND)
2119 if (isExtendedBUILD_VECTOR(N, DAG, true))
2124 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
2125 if (N->getOpcode() == ISD::ZERO_EXTEND)
2127 if (isExtendedBUILD_VECTOR(N, DAG, false))
2132 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
2133 unsigned Opcode = N->getOpcode();
2134 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
2135 SDNode *N0 = N->getOperand(0).getNode();
2136 SDNode *N1 = N->getOperand(1).getNode();
2137 return N0->hasOneUse() && N1->hasOneUse() &&
2138 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
2143 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
2144 unsigned Opcode = N->getOpcode();
2145 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
2146 SDNode *N0 = N->getOperand(0).getNode();
2147 SDNode *N1 = N->getOperand(1).getNode();
2148 return N0->hasOneUse() && N1->hasOneUse() &&
2149 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
2154 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
2155 // Multiplications are only custom-lowered for 128-bit vectors so that
2156 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
2157 EVT VT = Op.getValueType();
2158 assert(VT.is128BitVector() && VT.isInteger() &&
2159 "unexpected type for custom-lowering ISD::MUL");
2160 SDNode *N0 = Op.getOperand(0).getNode();
2161 SDNode *N1 = Op.getOperand(1).getNode();
2162 unsigned NewOpc = 0;
2164 bool isN0SExt = isSignExtended(N0, DAG);
2165 bool isN1SExt = isSignExtended(N1, DAG);
2166 if (isN0SExt && isN1SExt)
2167 NewOpc = AArch64ISD::SMULL;
2169 bool isN0ZExt = isZeroExtended(N0, DAG);
2170 bool isN1ZExt = isZeroExtended(N1, DAG);
2171 if (isN0ZExt && isN1ZExt)
2172 NewOpc = AArch64ISD::UMULL;
2173 else if (isN1SExt || isN1ZExt) {
2174 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
2175 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
2176 if (isN1SExt && isAddSubSExt(N0, DAG)) {
2177 NewOpc = AArch64ISD::SMULL;
2179 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
2180 NewOpc = AArch64ISD::UMULL;
2182 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
2184 NewOpc = AArch64ISD::UMULL;
2190 if (VT == MVT::v2i64)
2191 // Fall through to expand this. It is not legal.
2194 // Other vector multiplications are legal.
2199 // Legalize to a S/UMULL instruction
2202 SDValue Op1 = skipExtensionForVectorMULL(N1, DAG);
2204 Op0 = skipExtensionForVectorMULL(N0, DAG);
2205 assert(Op0.getValueType().is64BitVector() &&
2206 Op1.getValueType().is64BitVector() &&
2207 "unexpected types for extended operands to VMULL");
2208 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
2210 // Optimizing (zext A + zext B) * C, to (S/UMULL A, C) + (S/UMULL B, C) during
2211 // isel lowering to take advantage of no-stall back to back s/umul + s/umla.
2212 // This is true for CPUs with accumulate forwarding such as Cortex-A53/A57
2213 SDValue N00 = skipExtensionForVectorMULL(N0->getOperand(0).getNode(), DAG);
2214 SDValue N01 = skipExtensionForVectorMULL(N0->getOperand(1).getNode(), DAG);
2215 EVT Op1VT = Op1.getValueType();
2216 return DAG.getNode(N0->getOpcode(), DL, VT,
2217 DAG.getNode(NewOpc, DL, VT,
2218 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
2219 DAG.getNode(NewOpc, DL, VT,
2220 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
2223 SDValue AArch64TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
2224 SelectionDAG &DAG) const {
2225 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2228 default: return SDValue(); // Don't custom lower most intrinsics.
2229 case Intrinsic::aarch64_thread_pointer: {
2230 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2231 return DAG.getNode(AArch64ISD::THREAD_POINTER, dl, PtrVT);
2233 case Intrinsic::aarch64_neon_smax:
2234 return DAG.getNode(ISD::SMAX, dl, Op.getValueType(),
2235 Op.getOperand(1), Op.getOperand(2));
2236 case Intrinsic::aarch64_neon_umax:
2237 return DAG.getNode(ISD::UMAX, dl, Op.getValueType(),
2238 Op.getOperand(1), Op.getOperand(2));
2239 case Intrinsic::aarch64_neon_smin:
2240 return DAG.getNode(ISD::SMIN, dl, Op.getValueType(),
2241 Op.getOperand(1), Op.getOperand(2));
2242 case Intrinsic::aarch64_neon_umin:
2243 return DAG.getNode(ISD::UMIN, dl, Op.getValueType(),
2244 Op.getOperand(1), Op.getOperand(2));
2248 SDValue AArch64TargetLowering::LowerOperation(SDValue Op,
2249 SelectionDAG &DAG) const {
2250 switch (Op.getOpcode()) {
2252 llvm_unreachable("unimplemented operand");
2255 return LowerBITCAST(Op, DAG);
2256 case ISD::GlobalAddress:
2257 return LowerGlobalAddress(Op, DAG);
2258 case ISD::GlobalTLSAddress:
2259 return LowerGlobalTLSAddress(Op, DAG);
2261 return LowerSETCC(Op, DAG);
2263 return LowerBR_CC(Op, DAG);
2265 return LowerSELECT(Op, DAG);
2266 case ISD::SELECT_CC:
2267 return LowerSELECT_CC(Op, DAG);
2268 case ISD::JumpTable:
2269 return LowerJumpTable(Op, DAG);
2270 case ISD::ConstantPool:
2271 return LowerConstantPool(Op, DAG);
2272 case ISD::BlockAddress:
2273 return LowerBlockAddress(Op, DAG);
2275 return LowerVASTART(Op, DAG);
2277 return LowerVACOPY(Op, DAG);
2279 return LowerVAARG(Op, DAG);
2284 return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
2291 return LowerXALUO(Op, DAG);
2293 return LowerF128Call(Op, DAG, RTLIB::ADD_F128);
2295 return LowerF128Call(Op, DAG, RTLIB::SUB_F128);
2297 return LowerF128Call(Op, DAG, RTLIB::MUL_F128);
2299 return LowerF128Call(Op, DAG, RTLIB::DIV_F128);
2301 return LowerFP_ROUND(Op, DAG);
2302 case ISD::FP_EXTEND:
2303 return LowerFP_EXTEND(Op, DAG);
2304 case ISD::FRAMEADDR:
2305 return LowerFRAMEADDR(Op, DAG);
2306 case ISD::RETURNADDR:
2307 return LowerRETURNADDR(Op, DAG);
2308 case ISD::INSERT_VECTOR_ELT:
2309 return LowerINSERT_VECTOR_ELT(Op, DAG);
2310 case ISD::EXTRACT_VECTOR_ELT:
2311 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2312 case ISD::BUILD_VECTOR:
2313 return LowerBUILD_VECTOR(Op, DAG);
2314 case ISD::VECTOR_SHUFFLE:
2315 return LowerVECTOR_SHUFFLE(Op, DAG);
2316 case ISD::EXTRACT_SUBVECTOR:
2317 return LowerEXTRACT_SUBVECTOR(Op, DAG);
2321 return LowerVectorSRA_SRL_SHL(Op, DAG);
2322 case ISD::SHL_PARTS:
2323 return LowerShiftLeftParts(Op, DAG);
2324 case ISD::SRL_PARTS:
2325 case ISD::SRA_PARTS:
2326 return LowerShiftRightParts(Op, DAG);
2328 return LowerCTPOP(Op, DAG);
2329 case ISD::FCOPYSIGN:
2330 return LowerFCOPYSIGN(Op, DAG);
2332 return LowerVectorAND(Op, DAG);
2334 return LowerVectorOR(Op, DAG);
2336 return LowerXOR(Op, DAG);
2338 return LowerPREFETCH(Op, DAG);
2339 case ISD::SINT_TO_FP:
2340 case ISD::UINT_TO_FP:
2341 return LowerINT_TO_FP(Op, DAG);
2342 case ISD::FP_TO_SINT:
2343 case ISD::FP_TO_UINT:
2344 return LowerFP_TO_INT(Op, DAG);
2346 return LowerFSINCOS(Op, DAG);
2348 return LowerMUL(Op, DAG);
2349 case ISD::INTRINSIC_WO_CHAIN:
2350 return LowerINTRINSIC_WO_CHAIN(Op, DAG);
2354 //===----------------------------------------------------------------------===//
2355 // Calling Convention Implementation
2356 //===----------------------------------------------------------------------===//
2358 #include "AArch64GenCallingConv.inc"
2360 /// Selects the correct CCAssignFn for a given CallingConvention value.
2361 CCAssignFn *AArch64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
2362 bool IsVarArg) const {
2365 llvm_unreachable("Unsupported calling convention.");
2366 case CallingConv::WebKit_JS:
2367 return CC_AArch64_WebKit_JS;
2368 case CallingConv::GHC:
2369 return CC_AArch64_GHC;
2370 case CallingConv::C:
2371 case CallingConv::Fast:
2372 if (!Subtarget->isTargetDarwin())
2373 return CC_AArch64_AAPCS;
2374 return IsVarArg ? CC_AArch64_DarwinPCS_VarArg : CC_AArch64_DarwinPCS;
2378 SDValue AArch64TargetLowering::LowerFormalArguments(
2379 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2380 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2381 SmallVectorImpl<SDValue> &InVals) const {
2382 MachineFunction &MF = DAG.getMachineFunction();
2383 MachineFrameInfo *MFI = MF.getFrameInfo();
2385 // Assign locations to all of the incoming arguments.
2386 SmallVector<CCValAssign, 16> ArgLocs;
2387 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
2390 // At this point, Ins[].VT may already be promoted to i32. To correctly
2391 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2392 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
2393 // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
2394 // we use a special version of AnalyzeFormalArguments to pass in ValVT and
2396 unsigned NumArgs = Ins.size();
2397 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2398 unsigned CurArgIdx = 0;
2399 for (unsigned i = 0; i != NumArgs; ++i) {
2400 MVT ValVT = Ins[i].VT;
2401 if (Ins[i].isOrigArg()) {
2402 std::advance(CurOrigArg, Ins[i].getOrigArgIndex() - CurArgIdx);
2403 CurArgIdx = Ins[i].getOrigArgIndex();
2405 // Get type of the original argument.
2406 EVT ActualVT = getValueType(DAG.getDataLayout(), CurOrigArg->getType(),
2407 /*AllowUnknown*/ true);
2408 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
2409 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2410 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2412 else if (ActualMVT == MVT::i16)
2415 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2417 AssignFn(i, ValVT, ValVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
2418 assert(!Res && "Call operand has unhandled type");
2421 assert(ArgLocs.size() == Ins.size());
2422 SmallVector<SDValue, 16> ArgValues;
2423 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2424 CCValAssign &VA = ArgLocs[i];
2426 if (Ins[i].Flags.isByVal()) {
2427 // Byval is used for HFAs in the PCS, but the system should work in a
2428 // non-compliant manner for larger structs.
2429 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2430 int Size = Ins[i].Flags.getByValSize();
2431 unsigned NumRegs = (Size + 7) / 8;
2433 // FIXME: This works on big-endian for composite byvals, which are the common
2434 // case. It should also work for fundamental types too.
2436 MFI->CreateFixedObject(8 * NumRegs, VA.getLocMemOffset(), false);
2437 SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrVT);
2438 InVals.push_back(FrameIdxN);
2443 if (VA.isRegLoc()) {
2444 // Arguments stored in registers.
2445 EVT RegVT = VA.getLocVT();
2448 const TargetRegisterClass *RC;
2450 if (RegVT == MVT::i32)
2451 RC = &AArch64::GPR32RegClass;
2452 else if (RegVT == MVT::i64)
2453 RC = &AArch64::GPR64RegClass;
2454 else if (RegVT == MVT::f16)
2455 RC = &AArch64::FPR16RegClass;
2456 else if (RegVT == MVT::f32)
2457 RC = &AArch64::FPR32RegClass;
2458 else if (RegVT == MVT::f64 || RegVT.is64BitVector())
2459 RC = &AArch64::FPR64RegClass;
2460 else if (RegVT == MVT::f128 || RegVT.is128BitVector())
2461 RC = &AArch64::FPR128RegClass;
2463 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2465 // Transform the arguments in physical registers into virtual ones.
2466 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2467 ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2469 // If this is an 8, 16 or 32-bit value, it is really passed promoted
2470 // to 64 bits. Insert an assert[sz]ext to capture this, then
2471 // truncate to the right size.
2472 switch (VA.getLocInfo()) {
2474 llvm_unreachable("Unknown loc info!");
2475 case CCValAssign::Full:
2477 case CCValAssign::BCvt:
2478 ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue);
2480 case CCValAssign::AExt:
2481 case CCValAssign::SExt:
2482 case CCValAssign::ZExt:
2483 // SelectionDAGBuilder will insert appropriate AssertZExt & AssertSExt
2484 // nodes after our lowering.
2485 assert(RegVT == Ins[i].VT && "incorrect register location selected");
2489 InVals.push_back(ArgValue);
2491 } else { // VA.isRegLoc()
2492 assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
2493 unsigned ArgOffset = VA.getLocMemOffset();
2494 unsigned ArgSize = VA.getValVT().getSizeInBits() / 8;
2496 uint32_t BEAlign = 0;
2497 if (!Subtarget->isLittleEndian() && ArgSize < 8 &&
2498 !Ins[i].Flags.isInConsecutiveRegs())
2499 BEAlign = 8 - ArgSize;
2501 int FI = MFI->CreateFixedObject(ArgSize, ArgOffset + BEAlign, true);
2503 // Create load nodes to retrieve arguments from the stack.
2504 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2507 // For NON_EXTLOAD, generic code in getLoad assert(ValVT == MemVT)
2508 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
2509 MVT MemVT = VA.getValVT();
2511 switch (VA.getLocInfo()) {
2514 case CCValAssign::BCvt:
2515 MemVT = VA.getLocVT();
2517 case CCValAssign::SExt:
2518 ExtType = ISD::SEXTLOAD;
2520 case CCValAssign::ZExt:
2521 ExtType = ISD::ZEXTLOAD;
2523 case CCValAssign::AExt:
2524 ExtType = ISD::EXTLOAD;
2528 ArgValue = DAG.getExtLoad(
2529 ExtType, DL, VA.getLocVT(), Chain, FIN,
2530 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
2531 MemVT, false, false, false, 0);
2533 InVals.push_back(ArgValue);
2539 if (!Subtarget->isTargetDarwin()) {
2540 // The AAPCS variadic function ABI is identical to the non-variadic
2541 // one. As a result there may be more arguments in registers and we should
2542 // save them for future reference.
2543 saveVarArgRegisters(CCInfo, DAG, DL, Chain);
2546 AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>();
2547 // This will point to the next argument passed via stack.
2548 unsigned StackOffset = CCInfo.getNextStackOffset();
2549 // We currently pass all varargs at 8-byte alignment.
2550 StackOffset = ((StackOffset + 7) & ~7);
2551 AFI->setVarArgsStackIndex(MFI->CreateFixedObject(4, StackOffset, true));
2554 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2555 unsigned StackArgSize = CCInfo.getNextStackOffset();
2556 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2557 if (DoesCalleeRestoreStack(CallConv, TailCallOpt)) {
2558 // This is a non-standard ABI so by fiat I say we're allowed to make full
2559 // use of the stack area to be popped, which must be aligned to 16 bytes in
2561 StackArgSize = RoundUpToAlignment(StackArgSize, 16);
2563 // If we're expected to restore the stack (e.g. fastcc) then we'll be adding
2564 // a multiple of 16.
2565 FuncInfo->setArgumentStackToRestore(StackArgSize);
2567 // This realignment carries over to the available bytes below. Our own
2568 // callers will guarantee the space is free by giving an aligned value to
2571 // Even if we're not expected to free up the space, it's useful to know how
2572 // much is there while considering tail calls (because we can reuse it).
2573 FuncInfo->setBytesInStackArgArea(StackArgSize);
2578 void AArch64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
2579 SelectionDAG &DAG, SDLoc DL,
2580 SDValue &Chain) const {
2581 MachineFunction &MF = DAG.getMachineFunction();
2582 MachineFrameInfo *MFI = MF.getFrameInfo();
2583 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2584 auto PtrVT = getPointerTy(DAG.getDataLayout());
2586 SmallVector<SDValue, 8> MemOps;
2588 static const MCPhysReg GPRArgRegs[] = { AArch64::X0, AArch64::X1, AArch64::X2,
2589 AArch64::X3, AArch64::X4, AArch64::X5,
2590 AArch64::X6, AArch64::X7 };
2591 static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs);
2592 unsigned FirstVariadicGPR = CCInfo.getFirstUnallocated(GPRArgRegs);
2594 unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
2596 if (GPRSaveSize != 0) {
2597 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
2599 SDValue FIN = DAG.getFrameIndex(GPRIdx, PtrVT);
2601 for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
2602 unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &AArch64::GPR64RegClass);
2603 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
2604 SDValue Store = DAG.getStore(
2605 Val.getValue(1), DL, Val, FIN,
2606 MachinePointerInfo::getStack(DAG.getMachineFunction(), i * 8), false,
2608 MemOps.push_back(Store);
2610 DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getConstant(8, DL, PtrVT));
2613 FuncInfo->setVarArgsGPRIndex(GPRIdx);
2614 FuncInfo->setVarArgsGPRSize(GPRSaveSize);
2616 if (Subtarget->hasFPARMv8()) {
2617 static const MCPhysReg FPRArgRegs[] = {
2618 AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3,
2619 AArch64::Q4, AArch64::Q5, AArch64::Q6, AArch64::Q7};
2620 static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs);
2621 unsigned FirstVariadicFPR = CCInfo.getFirstUnallocated(FPRArgRegs);
2623 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
2625 if (FPRSaveSize != 0) {
2626 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
2628 SDValue FIN = DAG.getFrameIndex(FPRIdx, PtrVT);
2630 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
2631 unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &AArch64::FPR128RegClass);
2632 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f128);
2634 SDValue Store = DAG.getStore(
2635 Val.getValue(1), DL, Val, FIN,
2636 MachinePointerInfo::getStack(DAG.getMachineFunction(), i * 16),
2638 MemOps.push_back(Store);
2639 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN,
2640 DAG.getConstant(16, DL, PtrVT));
2643 FuncInfo->setVarArgsFPRIndex(FPRIdx);
2644 FuncInfo->setVarArgsFPRSize(FPRSaveSize);
2647 if (!MemOps.empty()) {
2648 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
2652 /// LowerCallResult - Lower the result values of a call into the
2653 /// appropriate copies out of appropriate physical registers.
2654 SDValue AArch64TargetLowering::LowerCallResult(
2655 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
2656 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2657 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
2658 SDValue ThisVal) const {
2659 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2660 ? RetCC_AArch64_WebKit_JS
2661 : RetCC_AArch64_AAPCS;
2662 // Assign locations to each value returned by this call.
2663 SmallVector<CCValAssign, 16> RVLocs;
2664 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2666 CCInfo.AnalyzeCallResult(Ins, RetCC);
2668 // Copy all of the result registers out of their specified physreg.
2669 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2670 CCValAssign VA = RVLocs[i];
2672 // Pass 'this' value directly from the argument to return value, to avoid
2673 // reg unit interference
2674 if (i == 0 && isThisReturn) {
2675 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 &&
2676 "unexpected return calling convention register assignment");
2677 InVals.push_back(ThisVal);
2682 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag);
2683 Chain = Val.getValue(1);
2684 InFlag = Val.getValue(2);
2686 switch (VA.getLocInfo()) {
2688 llvm_unreachable("Unknown loc info!");
2689 case CCValAssign::Full:
2691 case CCValAssign::BCvt:
2692 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
2696 InVals.push_back(Val);
2702 bool AArch64TargetLowering::isEligibleForTailCallOptimization(
2703 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
2704 bool isCalleeStructRet, bool isCallerStructRet,
2705 const SmallVectorImpl<ISD::OutputArg> &Outs,
2706 const SmallVectorImpl<SDValue> &OutVals,
2707 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
2708 // For CallingConv::C this function knows whether the ABI needs
2709 // changing. That's not true for other conventions so they will have to opt in
2711 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
2714 const MachineFunction &MF = DAG.getMachineFunction();
2715 const Function *CallerF = MF.getFunction();
2716 CallingConv::ID CallerCC = CallerF->getCallingConv();
2717 bool CCMatch = CallerCC == CalleeCC;
2719 // Byval parameters hand the function a pointer directly into the stack area
2720 // we want to reuse during a tail call. Working around this *is* possible (see
2721 // X86) but less efficient and uglier in LowerCall.
2722 for (Function::const_arg_iterator i = CallerF->arg_begin(),
2723 e = CallerF->arg_end();
2725 if (i->hasByValAttr())
2728 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
2729 if (IsTailCallConvention(CalleeCC) && CCMatch)
2734 // Externally-defined functions with weak linkage should not be
2735 // tail-called on AArch64 when the OS does not support dynamic
2736 // pre-emption of symbols, as the AAELF spec requires normal calls
2737 // to undefined weak functions to be replaced with a NOP or jump to the
2738 // next instruction. The behaviour of branch instructions in this
2739 // situation (as used for tail calls) is implementation-defined, so we
2740 // cannot rely on the linker replacing the tail call with a return.
2741 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2742 const GlobalValue *GV = G->getGlobal();
2743 const Triple &TT = getTargetMachine().getTargetTriple();
2744 if (GV->hasExternalWeakLinkage() &&
2745 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2749 // Now we search for cases where we can use a tail call without changing the
2750 // ABI. Sibcall is used in some places (particularly gcc) to refer to this
2753 // I want anyone implementing a new calling convention to think long and hard
2754 // about this assert.
2755 assert((!isVarArg || CalleeCC == CallingConv::C) &&
2756 "Unexpected variadic calling convention");
2758 if (isVarArg && !Outs.empty()) {
2759 // At least two cases here: if caller is fastcc then we can't have any
2760 // memory arguments (we'd be expected to clean up the stack afterwards). If
2761 // caller is C then we could potentially use its argument area.
2763 // FIXME: for now we take the most conservative of these in both cases:
2764 // disallow all variadic memory operands.
2765 SmallVector<CCValAssign, 16> ArgLocs;
2766 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2769 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, true));
2770 for (const CCValAssign &ArgLoc : ArgLocs)
2771 if (!ArgLoc.isRegLoc())
2775 // If the calling conventions do not match, then we'd better make sure the
2776 // results are returned in the same way as what the caller expects.
2778 SmallVector<CCValAssign, 16> RVLocs1;
2779 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2781 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForCall(CalleeCC, isVarArg));
2783 SmallVector<CCValAssign, 16> RVLocs2;
2784 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2786 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForCall(CallerCC, isVarArg));
2788 if (RVLocs1.size() != RVLocs2.size())
2790 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2791 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2793 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2795 if (RVLocs1[i].isRegLoc()) {
2796 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2799 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2805 // Nothing more to check if the callee is taking no arguments
2809 SmallVector<CCValAssign, 16> ArgLocs;
2810 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2813 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, isVarArg));
2815 const AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2817 // If the stack arguments for this call would fit into our own save area then
2818 // the call can be made tail.
2819 return CCInfo.getNextStackOffset() <= FuncInfo->getBytesInStackArgArea();
2822 SDValue AArch64TargetLowering::addTokenForArgument(SDValue Chain,
2824 MachineFrameInfo *MFI,
2825 int ClobberedFI) const {
2826 SmallVector<SDValue, 8> ArgChains;
2827 int64_t FirstByte = MFI->getObjectOffset(ClobberedFI);
2828 int64_t LastByte = FirstByte + MFI->getObjectSize(ClobberedFI) - 1;
2830 // Include the original chain at the beginning of the list. When this is
2831 // used by target LowerCall hooks, this helps legalize find the
2832 // CALLSEQ_BEGIN node.
2833 ArgChains.push_back(Chain);
2835 // Add a chain value for each stack argument corresponding
2836 for (SDNode::use_iterator U = DAG.getEntryNode().getNode()->use_begin(),
2837 UE = DAG.getEntryNode().getNode()->use_end();
2839 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
2840 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
2841 if (FI->getIndex() < 0) {
2842 int64_t InFirstByte = MFI->getObjectOffset(FI->getIndex());
2843 int64_t InLastByte = InFirstByte;
2844 InLastByte += MFI->getObjectSize(FI->getIndex()) - 1;
2846 if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
2847 (FirstByte <= InFirstByte && InFirstByte <= LastByte))
2848 ArgChains.push_back(SDValue(L, 1));
2851 // Build a tokenfactor for all the chains.
2852 return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
2855 bool AArch64TargetLowering::DoesCalleeRestoreStack(CallingConv::ID CallCC,
2856 bool TailCallOpt) const {
2857 return CallCC == CallingConv::Fast && TailCallOpt;
2860 bool AArch64TargetLowering::IsTailCallConvention(CallingConv::ID CallCC) const {
2861 return CallCC == CallingConv::Fast;
2864 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
2865 /// and add input and output parameter nodes.
2867 AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
2868 SmallVectorImpl<SDValue> &InVals) const {
2869 SelectionDAG &DAG = CLI.DAG;
2871 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2872 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2873 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2874 SDValue Chain = CLI.Chain;
2875 SDValue Callee = CLI.Callee;
2876 bool &IsTailCall = CLI.IsTailCall;
2877 CallingConv::ID CallConv = CLI.CallConv;
2878 bool IsVarArg = CLI.IsVarArg;
2880 MachineFunction &MF = DAG.getMachineFunction();
2881 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
2882 bool IsThisReturn = false;
2884 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2885 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2886 bool IsSibCall = false;
2889 // Check if it's really possible to do a tail call.
2890 IsTailCall = isEligibleForTailCallOptimization(
2891 Callee, CallConv, IsVarArg, IsStructRet,
2892 MF.getFunction()->hasStructRetAttr(), Outs, OutVals, Ins, DAG);
2893 if (!IsTailCall && CLI.CS && CLI.CS->isMustTailCall())
2894 report_fatal_error("failed to perform tail call elimination on a call "
2895 "site marked musttail");
2897 // A sibling call is one where we're under the usual C ABI and not planning
2898 // to change that but can still do a tail call:
2899 if (!TailCallOpt && IsTailCall)
2906 // Analyze operands of the call, assigning locations to each operand.
2907 SmallVector<CCValAssign, 16> ArgLocs;
2908 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
2912 // Handle fixed and variable vector arguments differently.
2913 // Variable vector arguments always go into memory.
2914 unsigned NumArgs = Outs.size();
2916 for (unsigned i = 0; i != NumArgs; ++i) {
2917 MVT ArgVT = Outs[i].VT;
2918 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2919 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
2920 /*IsVarArg=*/ !Outs[i].IsFixed);
2921 bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
2922 assert(!Res && "Call operand has unhandled type");
2926 // At this point, Outs[].VT may already be promoted to i32. To correctly
2927 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2928 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
2929 // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here
2930 // we use a special version of AnalyzeCallOperands to pass in ValVT and
2932 unsigned NumArgs = Outs.size();
2933 for (unsigned i = 0; i != NumArgs; ++i) {
2934 MVT ValVT = Outs[i].VT;
2935 // Get type of the original argument.
2936 EVT ActualVT = getValueType(DAG.getDataLayout(),
2937 CLI.getArgs()[Outs[i].OrigArgIndex].Ty,
2938 /*AllowUnknown*/ true);
2939 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT;
2940 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2941 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2942 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2944 else if (ActualMVT == MVT::i16)
2947 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2948 bool Res = AssignFn(i, ValVT, ValVT, CCValAssign::Full, ArgFlags, CCInfo);
2949 assert(!Res && "Call operand has unhandled type");
2954 // Get a count of how many bytes are to be pushed on the stack.
2955 unsigned NumBytes = CCInfo.getNextStackOffset();
2958 // Since we're not changing the ABI to make this a tail call, the memory
2959 // operands are already available in the caller's incoming argument space.
2963 // FPDiff is the byte offset of the call's argument area from the callee's.
2964 // Stores to callee stack arguments will be placed in FixedStackSlots offset
2965 // by this amount for a tail call. In a sibling call it must be 0 because the
2966 // caller will deallocate the entire stack and the callee still expects its
2967 // arguments to begin at SP+0. Completely unused for non-tail calls.
2970 if (IsTailCall && !IsSibCall) {
2971 unsigned NumReusableBytes = FuncInfo->getBytesInStackArgArea();
2973 // Since callee will pop argument stack as a tail call, we must keep the
2974 // popped size 16-byte aligned.
2975 NumBytes = RoundUpToAlignment(NumBytes, 16);
2977 // FPDiff will be negative if this tail call requires more space than we
2978 // would automatically have in our incoming argument space. Positive if we
2979 // can actually shrink the stack.
2980 FPDiff = NumReusableBytes - NumBytes;
2982 // The stack pointer must be 16-byte aligned at all times it's used for a
2983 // memory operation, which in practice means at *all* times and in
2984 // particular across call boundaries. Therefore our own arguments started at
2985 // a 16-byte aligned SP and the delta applied for the tail call should
2986 // satisfy the same constraint.
2987 assert(FPDiff % 16 == 0 && "unaligned stack on tail call");
2990 // Adjust the stack pointer for the new arguments...
2991 // These operations are automatically eliminated by the prolog/epilog pass
2993 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, DL,
2997 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, AArch64::SP,
2998 getPointerTy(DAG.getDataLayout()));
3000 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3001 SmallVector<SDValue, 8> MemOpChains;
3002 auto PtrVT = getPointerTy(DAG.getDataLayout());
3004 // Walk the register/memloc assignments, inserting copies/loads.
3005 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
3006 ++i, ++realArgIdx) {
3007 CCValAssign &VA = ArgLocs[i];
3008 SDValue Arg = OutVals[realArgIdx];
3009 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
3011 // Promote the value if needed.
3012 switch (VA.getLocInfo()) {
3014 llvm_unreachable("Unknown loc info!");
3015 case CCValAssign::Full:
3017 case CCValAssign::SExt:
3018 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
3020 case CCValAssign::ZExt:
3021 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
3023 case CCValAssign::AExt:
3024 if (Outs[realArgIdx].ArgVT == MVT::i1) {
3025 // AAPCS requires i1 to be zero-extended to 8-bits by the caller.
3026 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
3027 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i8, Arg);
3029 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
3031 case CCValAssign::BCvt:
3032 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
3034 case CCValAssign::FPExt:
3035 Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
3039 if (VA.isRegLoc()) {
3040 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i64) {
3041 assert(VA.getLocVT() == MVT::i64 &&
3042 "unexpected calling convention register assignment");
3043 assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
3044 "unexpected use of 'returned'");
3045 IsThisReturn = true;
3047 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3049 assert(VA.isMemLoc());
3052 MachinePointerInfo DstInfo;
3054 // FIXME: This works on big-endian for composite byvals, which are the
3055 // common case. It should also work for fundamental types too.
3056 uint32_t BEAlign = 0;
3057 unsigned OpSize = Flags.isByVal() ? Flags.getByValSize() * 8
3058 : VA.getValVT().getSizeInBits();
3059 OpSize = (OpSize + 7) / 8;
3060 if (!Subtarget->isLittleEndian() && !Flags.isByVal() &&
3061 !Flags.isInConsecutiveRegs()) {
3063 BEAlign = 8 - OpSize;
3065 unsigned LocMemOffset = VA.getLocMemOffset();
3066 int32_t Offset = LocMemOffset + BEAlign;
3067 SDValue PtrOff = DAG.getIntPtrConstant(Offset, DL);
3068 PtrOff = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, PtrOff);
3071 Offset = Offset + FPDiff;
3072 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3074 DstAddr = DAG.getFrameIndex(FI, PtrVT);
3076 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI);
3078 // Make sure any stack arguments overlapping with where we're storing
3079 // are loaded before this eventual operation. Otherwise they'll be
3081 Chain = addTokenForArgument(Chain, DAG, MF.getFrameInfo(), FI);
3083 SDValue PtrOff = DAG.getIntPtrConstant(Offset, DL);
3085 DstAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, PtrOff);
3086 DstInfo = MachinePointerInfo::getStack(DAG.getMachineFunction(),
3090 if (Outs[i].Flags.isByVal()) {
3092 DAG.getConstant(Outs[i].Flags.getByValSize(), DL, MVT::i64);
3093 SDValue Cpy = DAG.getMemcpy(
3094 Chain, DL, DstAddr, Arg, SizeNode, Outs[i].Flags.getByValAlign(),
3095 /*isVol = */ false, /*AlwaysInline = */ false,
3096 /*isTailCall = */ false,
3097 DstInfo, MachinePointerInfo());
3099 MemOpChains.push_back(Cpy);
3101 // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
3102 // promoted to a legal register type i32, we should truncate Arg back to
3104 if (VA.getValVT() == MVT::i1 || VA.getValVT() == MVT::i8 ||
3105 VA.getValVT() == MVT::i16)
3106 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Arg);
3109 DAG.getStore(Chain, DL, Arg, DstAddr, DstInfo, false, false, 0);
3110 MemOpChains.push_back(Store);
3115 if (!MemOpChains.empty())
3116 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
3118 // Build a sequence of copy-to-reg nodes chained together with token chain
3119 // and flag operands which copy the outgoing args into the appropriate regs.
3121 for (auto &RegToPass : RegsToPass) {
3122 Chain = DAG.getCopyToReg(Chain, DL, RegToPass.first,
3123 RegToPass.second, InFlag);
3124 InFlag = Chain.getValue(1);
3127 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
3128 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
3129 // node so that legalize doesn't hack it.
3130 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3131 Subtarget->isTargetMachO()) {
3132 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3133 const GlobalValue *GV = G->getGlobal();
3134 bool InternalLinkage = GV->hasInternalLinkage();
3135 if (InternalLinkage)
3136 Callee = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 0);
3139 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_GOT);
3140 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, Callee);
3142 } else if (ExternalSymbolSDNode *S =
3143 dyn_cast<ExternalSymbolSDNode>(Callee)) {
3144 const char *Sym = S->getSymbol();
3145 Callee = DAG.getTargetExternalSymbol(Sym, PtrVT, AArch64II::MO_GOT);
3146 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, Callee);
3148 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3149 const GlobalValue *GV = G->getGlobal();
3150 Callee = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 0);
3151 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3152 const char *Sym = S->getSymbol();
3153 Callee = DAG.getTargetExternalSymbol(Sym, PtrVT, 0);
3156 // We don't usually want to end the call-sequence here because we would tidy
3157 // the frame up *after* the call, however in the ABI-changing tail-call case
3158 // we've carefully laid out the parameters so that when sp is reset they'll be
3159 // in the correct location.
3160 if (IsTailCall && !IsSibCall) {
3161 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, DL, true),
3162 DAG.getIntPtrConstant(0, DL, true), InFlag, DL);
3163 InFlag = Chain.getValue(1);
3166 std::vector<SDValue> Ops;
3167 Ops.push_back(Chain);
3168 Ops.push_back(Callee);
3171 // Each tail call may have to adjust the stack by a different amount, so
3172 // this information must travel along with the operation for eventual
3173 // consumption by emitEpilogue.
3174 Ops.push_back(DAG.getTargetConstant(FPDiff, DL, MVT::i32));
3177 // Add argument registers to the end of the list so that they are known live
3179 for (auto &RegToPass : RegsToPass)
3180 Ops.push_back(DAG.getRegister(RegToPass.first,
3181 RegToPass.second.getValueType()));
3183 // Add a register mask operand representing the call-preserved registers.
3184 const uint32_t *Mask;
3185 const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();
3187 // For 'this' returns, use the X0-preserving mask if applicable
3188 Mask = TRI->getThisReturnPreservedMask(MF, CallConv);
3190 IsThisReturn = false;
3191 Mask = TRI->getCallPreservedMask(MF, CallConv);
3194 Mask = TRI->getCallPreservedMask(MF, CallConv);
3196 assert(Mask && "Missing call preserved mask for calling convention");
3197 Ops.push_back(DAG.getRegisterMask(Mask));
3199 if (InFlag.getNode())
3200 Ops.push_back(InFlag);
3202 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3204 // If we're doing a tall call, use a TC_RETURN here rather than an
3205 // actual call instruction.
3207 MF.getFrameInfo()->setHasTailCall();
3208 return DAG.getNode(AArch64ISD::TC_RETURN, DL, NodeTys, Ops);
3211 // Returns a chain and a flag for retval copy to use.
3212 Chain = DAG.getNode(AArch64ISD::CALL, DL, NodeTys, Ops);
3213 InFlag = Chain.getValue(1);
3215 uint64_t CalleePopBytes = DoesCalleeRestoreStack(CallConv, TailCallOpt)
3216 ? RoundUpToAlignment(NumBytes, 16)
3219 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, DL, true),
3220 DAG.getIntPtrConstant(CalleePopBytes, DL, true),
3223 InFlag = Chain.getValue(1);
3225 // Handle result values, copying them out of physregs into vregs that we
3227 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
3228 InVals, IsThisReturn,
3229 IsThisReturn ? OutVals[0] : SDValue());
3232 bool AArch64TargetLowering::CanLowerReturn(
3233 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
3234 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
3235 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
3236 ? RetCC_AArch64_WebKit_JS
3237 : RetCC_AArch64_AAPCS;
3238 SmallVector<CCValAssign, 16> RVLocs;
3239 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
3240 return CCInfo.CheckReturn(Outs, RetCC);
3244 AArch64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
3246 const SmallVectorImpl<ISD::OutputArg> &Outs,
3247 const SmallVectorImpl<SDValue> &OutVals,
3248 SDLoc DL, SelectionDAG &DAG) const {
3249 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
3250 ? RetCC_AArch64_WebKit_JS
3251 : RetCC_AArch64_AAPCS;
3252 SmallVector<CCValAssign, 16> RVLocs;
3253 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
3255 CCInfo.AnalyzeReturn(Outs, RetCC);
3257 // Copy the result values into the output registers.
3259 SmallVector<SDValue, 4> RetOps(1, Chain);
3260 for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size();
3261 ++i, ++realRVLocIdx) {
3262 CCValAssign &VA = RVLocs[i];
3263 assert(VA.isRegLoc() && "Can only return in registers!");
3264 SDValue Arg = OutVals[realRVLocIdx];
3266 switch (VA.getLocInfo()) {
3268 llvm_unreachable("Unknown loc info!");
3269 case CCValAssign::Full:
3270 if (Outs[i].ArgVT == MVT::i1) {
3271 // AAPCS requires i1 to be zero-extended to i8 by the producer of the
3272 // value. This is strictly redundant on Darwin (which uses "zeroext
3273 // i1"), but will be optimised out before ISel.
3274 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
3275 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
3278 case CCValAssign::BCvt:
3279 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
3283 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
3284 Flag = Chain.getValue(1);
3285 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
3287 const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();
3288 const MCPhysReg *I =
3289 TRI->getCalleeSavedRegsViaCopy(&DAG.getMachineFunction());
3292 if (AArch64::GPR64RegClass.contains(*I))
3293 RetOps.push_back(DAG.getRegister(*I, MVT::i64));
3294 else if (AArch64::FPR64RegClass.contains(*I))
3295 RetOps.push_back(DAG.getRegister(*I, MVT::getFloatingPointVT(64)));
3297 llvm_unreachable("Unexpected register class in CSRsViaCopy!");
3301 RetOps[0] = Chain; // Update chain.
3303 // Add the flag if we have it.
3305 RetOps.push_back(Flag);
3307 return DAG.getNode(AArch64ISD::RET_FLAG, DL, MVT::Other, RetOps);
3310 //===----------------------------------------------------------------------===//
3311 // Other Lowering Code
3312 //===----------------------------------------------------------------------===//
3314 SDValue AArch64TargetLowering::LowerGlobalAddress(SDValue Op,
3315 SelectionDAG &DAG) const {
3316 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3318 const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
3319 const GlobalValue *GV = GN->getGlobal();
3320 unsigned char OpFlags =
3321 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
3323 assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
3324 "unexpected offset in global node");
3326 // This also catched the large code model case for Darwin.
3327 if ((OpFlags & AArch64II::MO_GOT) != 0) {
3328 SDValue GotAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
3329 // FIXME: Once remat is capable of dealing with instructions with register
3330 // operands, expand this into two nodes instead of using a wrapper node.
3331 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
3334 if ((OpFlags & AArch64II::MO_CONSTPOOL) != 0) {
3335 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
3336 "use of MO_CONSTPOOL only supported on small model");
3337 SDValue Hi = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, AArch64II::MO_PAGE);
3338 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3339 unsigned char LoFlags = AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
3340 SDValue Lo = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, LoFlags);
3341 SDValue PoolAddr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3342 SDValue GlobalAddr = DAG.getLoad(
3343 PtrVT, DL, DAG.getEntryNode(), PoolAddr,
3344 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
3345 /*isVolatile=*/false,
3346 /*isNonTemporal=*/true,
3347 /*isInvariant=*/true, 8);
3348 if (GN->getOffset() != 0)
3349 return DAG.getNode(ISD::ADD, DL, PtrVT, GlobalAddr,
3350 DAG.getConstant(GN->getOffset(), DL, PtrVT));
3354 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
3355 const unsigned char MO_NC = AArch64II::MO_NC;
3357 AArch64ISD::WrapperLarge, DL, PtrVT,
3358 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G3),
3359 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
3360 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
3361 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
3363 // Use ADRP/ADD or ADRP/LDR for everything else: the small model on ELF and
3364 // the only correct model on Darwin.
3365 SDValue Hi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
3366 OpFlags | AArch64II::MO_PAGE);
3367 unsigned char LoFlags = OpFlags | AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
3368 SDValue Lo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, LoFlags);
3370 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3371 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3375 /// \brief Convert a TLS address reference into the correct sequence of loads
3376 /// and calls to compute the variable's address (for Darwin, currently) and
3377 /// return an SDValue containing the final node.
3379 /// Darwin only has one TLS scheme which must be capable of dealing with the
3380 /// fully general situation, in the worst case. This means:
3381 /// + "extern __thread" declaration.
3382 /// + Defined in a possibly unknown dynamic library.
3384 /// The general system is that each __thread variable has a [3 x i64] descriptor
3385 /// which contains information used by the runtime to calculate the address. The
3386 /// only part of this the compiler needs to know about is the first xword, which
3387 /// contains a function pointer that must be called with the address of the
3388 /// entire descriptor in "x0".
3390 /// Since this descriptor may be in a different unit, in general even the
3391 /// descriptor must be accessed via an indirect load. The "ideal" code sequence
3393 /// adrp x0, _var@TLVPPAGE
3394 /// ldr x0, [x0, _var@TLVPPAGEOFF] ; x0 now contains address of descriptor
3395 /// ldr x1, [x0] ; x1 contains 1st entry of descriptor,
3396 /// ; the function pointer
3397 /// blr x1 ; Uses descriptor address in x0
3398 /// ; Address of _var is now in x0.
3400 /// If the address of _var's descriptor *is* known to the linker, then it can
3401 /// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
3402 /// a slight efficiency gain.
3404 AArch64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
3405 SelectionDAG &DAG) const {
3406 assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");
3409 MVT PtrVT = getPointerTy(DAG.getDataLayout());
3410 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3413 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3414 SDValue DescAddr = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TLVPAddr);
3416 // The first entry in the descriptor is a function pointer that we must call
3417 // to obtain the address of the variable.
3418 SDValue Chain = DAG.getEntryNode();
3419 SDValue FuncTLVGet =
3420 DAG.getLoad(MVT::i64, DL, Chain, DescAddr,
3421 MachinePointerInfo::getGOT(DAG.getMachineFunction()), false,
3423 Chain = FuncTLVGet.getValue(1);
3425 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3426 MFI->setAdjustsStack(true);
3428 // TLS calls preserve all registers except those that absolutely must be
3429 // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
3431 const uint32_t *Mask =
3432 Subtarget->getRegisterInfo()->getTLSCallPreservedMask();
3434 // Finally, we can make the call. This is just a degenerate version of a
3435 // normal AArch64 call node: x0 takes the address of the descriptor, and
3436 // returns the address of the variable in this thread.
3437 Chain = DAG.getCopyToReg(Chain, DL, AArch64::X0, DescAddr, SDValue());
3439 DAG.getNode(AArch64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
3440 Chain, FuncTLVGet, DAG.getRegister(AArch64::X0, MVT::i64),
3441 DAG.getRegisterMask(Mask), Chain.getValue(1));
3442 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Chain.getValue(1));
3445 /// When accessing thread-local variables under either the general-dynamic or
3446 /// local-dynamic system, we make a "TLS-descriptor" call. The variable will
3447 /// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
3448 /// is a function pointer to carry out the resolution.
3450 /// The sequence is:
3451 /// adrp x0, :tlsdesc:var
3452 /// ldr x1, [x0, #:tlsdesc_lo12:var]
3453 /// add x0, x0, #:tlsdesc_lo12:var
3454 /// .tlsdesccall var
3456 /// (TPIDR_EL0 offset now in x0)
3458 /// The above sequence must be produced unscheduled, to enable the linker to
3459 /// optimize/relax this sequence.
3460 /// Therefore, a pseudo-instruction (TLSDESC_CALLSEQ) is used to represent the
3461 /// above sequence, and expanded really late in the compilation flow, to ensure
3462 /// the sequence is produced as per above.
3463 SDValue AArch64TargetLowering::LowerELFTLSDescCallSeq(SDValue SymAddr, SDLoc DL,
3464 SelectionDAG &DAG) const {
3465 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3467 SDValue Chain = DAG.getEntryNode();
3468 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3470 SmallVector<SDValue, 2> Ops;
3471 Ops.push_back(Chain);
3472 Ops.push_back(SymAddr);
3474 Chain = DAG.getNode(AArch64ISD::TLSDESC_CALLSEQ, DL, NodeTys, Ops);
3475 SDValue Glue = Chain.getValue(1);
3477 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Glue);
3481 AArch64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
3482 SelectionDAG &DAG) const {
3483 assert(Subtarget->isTargetELF() && "This function expects an ELF target");
3484 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
3485 "ELF TLS only supported in small memory model");
3486 // Different choices can be made for the maximum size of the TLS area for a
3487 // module. For the small address model, the default TLS size is 16MiB and the
3488 // maximum TLS size is 4GiB.
3489 // FIXME: add -mtls-size command line option and make it control the 16MiB
3490 // vs. 4GiB code sequence generation.
3491 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3493 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
3495 if (DAG.getTarget().Options.EmulatedTLS)
3496 return LowerToTLSEmulatedModel(GA, DAG);
3498 if (!EnableAArch64ELFLocalDynamicTLSGeneration) {
3499 if (Model == TLSModel::LocalDynamic)
3500 Model = TLSModel::GeneralDynamic;
3504 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3506 const GlobalValue *GV = GA->getGlobal();
3508 SDValue ThreadBase = DAG.getNode(AArch64ISD::THREAD_POINTER, DL, PtrVT);
3510 if (Model == TLSModel::LocalExec) {
3511 SDValue HiVar = DAG.getTargetGlobalAddress(
3512 GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_HI12);
3513 SDValue LoVar = DAG.getTargetGlobalAddress(
3515 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3517 SDValue TPWithOff_lo =
3518 SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, ThreadBase,
3520 DAG.getTargetConstant(0, DL, MVT::i32)),
3523 SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPWithOff_lo,
3525 DAG.getTargetConstant(0, DL, MVT::i32)),
3528 } else if (Model == TLSModel::InitialExec) {
3529 TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3530 TPOff = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TPOff);
3531 } else if (Model == TLSModel::LocalDynamic) {
3532 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
3533 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
3534 // the beginning of the module's TLS region, followed by a DTPREL offset
3537 // These accesses will need deduplicating if there's more than one.
3538 AArch64FunctionInfo *MFI =
3539 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
3540 MFI->incNumLocalDynamicTLSAccesses();
3542 // The call needs a relocation too for linker relaxation. It doesn't make
3543 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
3545 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
3548 // Now we can calculate the offset from TPIDR_EL0 to this module's
3549 // thread-local area.
3550 TPOff = LowerELFTLSDescCallSeq(SymAddr, DL, DAG);
3552 // Now use :dtprel_whatever: operations to calculate this variable's offset
3553 // in its thread-storage area.
3554 SDValue HiVar = DAG.getTargetGlobalAddress(
3555 GV, DL, MVT::i64, 0, AArch64II::MO_TLS | AArch64II::MO_HI12);
3556 SDValue LoVar = DAG.getTargetGlobalAddress(
3557 GV, DL, MVT::i64, 0,
3558 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3560 TPOff = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPOff, HiVar,
3561 DAG.getTargetConstant(0, DL, MVT::i32)),
3563 TPOff = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPOff, LoVar,
3564 DAG.getTargetConstant(0, DL, MVT::i32)),
3566 } else if (Model == TLSModel::GeneralDynamic) {
3567 // The call needs a relocation too for linker relaxation. It doesn't make
3568 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
3571 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3573 // Finally we can make a call to calculate the offset from tpidr_el0.
3574 TPOff = LowerELFTLSDescCallSeq(SymAddr, DL, DAG);
3576 llvm_unreachable("Unsupported ELF TLS access model");
3578 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
3581 SDValue AArch64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
3582 SelectionDAG &DAG) const {
3583 if (Subtarget->isTargetDarwin())
3584 return LowerDarwinGlobalTLSAddress(Op, DAG);
3585 else if (Subtarget->isTargetELF())
3586 return LowerELFGlobalTLSAddress(Op, DAG);
3588 llvm_unreachable("Unexpected platform trying to use TLS");
3590 SDValue AArch64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3591 SDValue Chain = Op.getOperand(0);
3592 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3593 SDValue LHS = Op.getOperand(2);
3594 SDValue RHS = Op.getOperand(3);
3595 SDValue Dest = Op.getOperand(4);
3598 // Handle f128 first, since lowering it will result in comparing the return
3599 // value of a libcall against zero, which is just what the rest of LowerBR_CC
3600 // is expecting to deal with.
3601 if (LHS.getValueType() == MVT::f128) {
3602 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3604 // If softenSetCCOperands returned a scalar, we need to compare the result
3605 // against zero to select between true and false values.
3606 if (!RHS.getNode()) {
3607 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3612 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
3614 unsigned Opc = LHS.getOpcode();
3615 if (LHS.getResNo() == 1 && isOneConstant(RHS) &&
3616 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3617 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3618 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
3619 "Unexpected condition code.");
3620 // Only lower legal XALUO ops.
3621 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
3624 // The actual operation with overflow check.
3625 AArch64CC::CondCode OFCC;
3626 SDValue Value, Overflow;
3627 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, LHS.getValue(0), DAG);
3629 if (CC == ISD::SETNE)
3630 OFCC = getInvertedCondCode(OFCC);
3631 SDValue CCVal = DAG.getConstant(OFCC, dl, MVT::i32);
3633 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3637 if (LHS.getValueType().isInteger()) {
3638 assert((LHS.getValueType() == RHS.getValueType()) &&
3639 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3641 // If the RHS of the comparison is zero, we can potentially fold this
3642 // to a specialized branch.
3643 const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
3644 if (RHSC && RHSC->getZExtValue() == 0) {
3645 if (CC == ISD::SETEQ) {
3646 // See if we can use a TBZ to fold in an AND as well.
3647 // TBZ has a smaller branch displacement than CBZ. If the offset is
3648 // out of bounds, a late MI-layer pass rewrites branches.
3649 // 403.gcc is an example that hits this case.
3650 if (LHS.getOpcode() == ISD::AND &&
3651 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3652 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3653 SDValue Test = LHS.getOperand(0);
3654 uint64_t Mask = LHS.getConstantOperandVal(1);
3655 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, Test,
3656 DAG.getConstant(Log2_64(Mask), dl, MVT::i64),
3660 return DAG.getNode(AArch64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest);
3661 } else if (CC == ISD::SETNE) {
3662 // See if we can use a TBZ to fold in an AND as well.
3663 // TBZ has a smaller branch displacement than CBZ. If the offset is
3664 // out of bounds, a late MI-layer pass rewrites branches.
3665 // 403.gcc is an example that hits this case.
3666 if (LHS.getOpcode() == ISD::AND &&
3667 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3668 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3669 SDValue Test = LHS.getOperand(0);
3670 uint64_t Mask = LHS.getConstantOperandVal(1);
3671 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, Test,
3672 DAG.getConstant(Log2_64(Mask), dl, MVT::i64),
3676 return DAG.getNode(AArch64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest);
3677 } else if (CC == ISD::SETLT && LHS.getOpcode() != ISD::AND) {
3678 // Don't combine AND since emitComparison converts the AND to an ANDS
3679 // (a.k.a. TST) and the test in the test bit and branch instruction
3680 // becomes redundant. This would also increase register pressure.
3681 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
3682 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, LHS,
3683 DAG.getConstant(Mask, dl, MVT::i64), Dest);
3686 if (RHSC && RHSC->getSExtValue() == -1 && CC == ISD::SETGT &&
3687 LHS.getOpcode() != ISD::AND) {
3688 // Don't combine AND since emitComparison converts the AND to an ANDS
3689 // (a.k.a. TST) and the test in the test bit and branch instruction
3690 // becomes redundant. This would also increase register pressure.
3691 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
3692 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, LHS,
3693 DAG.getConstant(Mask, dl, MVT::i64), Dest);
3697 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3698 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3702 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3704 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
3705 // clean. Some of them require two branches to implement.
3706 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3707 AArch64CC::CondCode CC1, CC2;
3708 changeFPCCToAArch64CC(CC, CC1, CC2);
3709 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
3711 DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp);
3712 if (CC2 != AArch64CC::AL) {
3713 SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32);
3714 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val,
3721 SDValue AArch64TargetLowering::LowerFCOPYSIGN(SDValue Op,
3722 SelectionDAG &DAG) const {
3723 EVT VT = Op.getValueType();
3726 SDValue In1 = Op.getOperand(0);
3727 SDValue In2 = Op.getOperand(1);
3728 EVT SrcVT = In2.getValueType();
3730 if (SrcVT.bitsLT(VT))
3731 In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2);
3732 else if (SrcVT.bitsGT(VT))
3733 In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, DAG.getIntPtrConstant(0, DL));
3738 SDValue VecVal1, VecVal2;
3739 if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) {
3741 VecVT = (VT == MVT::v2f32 ? MVT::v2i32 : MVT::v4i32);
3742 EltMask = 0x80000000ULL;
3744 if (!VT.isVector()) {
3745 VecVal1 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3746 DAG.getUNDEF(VecVT), In1);
3747 VecVal2 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3748 DAG.getUNDEF(VecVT), In2);
3750 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3751 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3753 } else if (VT == MVT::f64 || VT == MVT::v2f64) {
3757 // We want to materialize a mask with the high bit set, but the AdvSIMD
3758 // immediate moves cannot materialize that in a single instruction for
3759 // 64-bit elements. Instead, materialize zero and then negate it.
3762 if (!VT.isVector()) {
3763 VecVal1 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3764 DAG.getUNDEF(VecVT), In1);
3765 VecVal2 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3766 DAG.getUNDEF(VecVT), In2);
3768 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3769 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3772 llvm_unreachable("Invalid type for copysign!");
3775 SDValue BuildVec = DAG.getConstant(EltMask, DL, VecVT);
3777 // If we couldn't materialize the mask above, then the mask vector will be
3778 // the zero vector, and we need to negate it here.
3779 if (VT == MVT::f64 || VT == MVT::v2f64) {
3780 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec);
3781 BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec);
3782 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec);
3786 DAG.getNode(AArch64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec);
3789 return DAG.getTargetExtractSubreg(AArch64::ssub, DL, VT, Sel);
3790 else if (VT == MVT::f64)
3791 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, VT, Sel);
3793 return DAG.getNode(ISD::BITCAST, DL, VT, Sel);
3796 SDValue AArch64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
3797 if (DAG.getMachineFunction().getFunction()->hasFnAttribute(
3798 Attribute::NoImplicitFloat))
3801 if (!Subtarget->hasNEON())
3804 // While there is no integer popcount instruction, it can
3805 // be more efficiently lowered to the following sequence that uses
3806 // AdvSIMD registers/instructions as long as the copies to/from
3807 // the AdvSIMD registers are cheap.
3808 // FMOV D0, X0 // copy 64-bit int to vector, high bits zero'd
3809 // CNT V0.8B, V0.8B // 8xbyte pop-counts
3810 // ADDV B0, V0.8B // sum 8xbyte pop-counts
3811 // UMOV X0, V0.B[0] // copy byte result back to integer reg
3812 SDValue Val = Op.getOperand(0);
3814 EVT VT = Op.getValueType();
3817 Val = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Val);
3818 Val = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);
3820 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, Val);
3821 SDValue UaddLV = DAG.getNode(
3822 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
3823 DAG.getConstant(Intrinsic::aarch64_neon_uaddlv, DL, MVT::i32), CtPop);
3826 UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV);
3830 SDValue AArch64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
3832 if (Op.getValueType().isVector())
3833 return LowerVSETCC(Op, DAG);
3835 SDValue LHS = Op.getOperand(0);
3836 SDValue RHS = Op.getOperand(1);
3837 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
3840 // We chose ZeroOrOneBooleanContents, so use zero and one.
3841 EVT VT = Op.getValueType();
3842 SDValue TVal = DAG.getConstant(1, dl, VT);
3843 SDValue FVal = DAG.getConstant(0, dl, VT);
3845 // Handle f128 first, since one possible outcome is a normal integer
3846 // comparison which gets picked up by the next if statement.
3847 if (LHS.getValueType() == MVT::f128) {
3848 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3850 // If softenSetCCOperands returned a scalar, use it.
3851 if (!RHS.getNode()) {
3852 assert(LHS.getValueType() == Op.getValueType() &&
3853 "Unexpected setcc expansion!");
3858 if (LHS.getValueType().isInteger()) {
3861 getAArch64Cmp(LHS, RHS, ISD::getSetCCInverse(CC, true), CCVal, DAG, dl);
3863 // Note that we inverted the condition above, so we reverse the order of
3864 // the true and false operands here. This will allow the setcc to be
3865 // matched to a single CSINC instruction.
3866 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp);
3869 // Now we know we're dealing with FP values.
3870 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3872 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3873 // and do the comparison.
3874 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3876 AArch64CC::CondCode CC1, CC2;
3877 changeFPCCToAArch64CC(CC, CC1, CC2);
3878 if (CC2 == AArch64CC::AL) {
3879 changeFPCCToAArch64CC(ISD::getSetCCInverse(CC, false), CC1, CC2);
3880 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
3882 // Note that we inverted the condition above, so we reverse the order of
3883 // the true and false operands here. This will allow the setcc to be
3884 // matched to a single CSINC instruction.
3885 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp);
3887 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't
3888 // totally clean. Some of them require two CSELs to implement. As is in
3889 // this case, we emit the first CSEL and then emit a second using the output
3890 // of the first as the RHS. We're effectively OR'ing the two CC's together.
3892 // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
3893 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
3895 DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3897 SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32);
3898 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3902 SDValue AArch64TargetLowering::LowerSELECT_CC(ISD::CondCode CC, SDValue LHS,
3903 SDValue RHS, SDValue TVal,
3904 SDValue FVal, SDLoc dl,
3905 SelectionDAG &DAG) const {
3906 // Handle f128 first, because it will result in a comparison of some RTLIB
3907 // call result against zero.
3908 if (LHS.getValueType() == MVT::f128) {
3909 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3911 // If softenSetCCOperands returned a scalar, we need to compare the result
3912 // against zero to select between true and false values.
3913 if (!RHS.getNode()) {
3914 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3919 // Also handle f16, for which we need to do a f32 comparison.
3920 if (LHS.getValueType() == MVT::f16) {
3921 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, LHS);
3922 RHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, RHS);
3925 // Next, handle integers.
3926 if (LHS.getValueType().isInteger()) {
3927 assert((LHS.getValueType() == RHS.getValueType()) &&
3928 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3930 unsigned Opcode = AArch64ISD::CSEL;
3932 // If both the TVal and the FVal are constants, see if we can swap them in
3933 // order to for a CSINV or CSINC out of them.
3934 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
3935 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
3937 if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) {
3938 std::swap(TVal, FVal);
3939 std::swap(CTVal, CFVal);
3940 CC = ISD::getSetCCInverse(CC, true);
3941 } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) {
3942 std::swap(TVal, FVal);
3943 std::swap(CTVal, CFVal);
3944 CC = ISD::getSetCCInverse(CC, true);
3945 } else if (TVal.getOpcode() == ISD::XOR) {
3946 // If TVal is a NOT we want to swap TVal and FVal so that we can match
3947 // with a CSINV rather than a CSEL.
3948 if (isAllOnesConstant(TVal.getOperand(1))) {
3949 std::swap(TVal, FVal);
3950 std::swap(CTVal, CFVal);
3951 CC = ISD::getSetCCInverse(CC, true);
3953 } else if (TVal.getOpcode() == ISD::SUB) {
3954 // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
3955 // that we can match with a CSNEG rather than a CSEL.
3956 if (isNullConstant(TVal.getOperand(0))) {
3957 std::swap(TVal, FVal);
3958 std::swap(CTVal, CFVal);
3959 CC = ISD::getSetCCInverse(CC, true);
3961 } else if (CTVal && CFVal) {
3962 const int64_t TrueVal = CTVal->getSExtValue();
3963 const int64_t FalseVal = CFVal->getSExtValue();
3966 // If both TVal and FVal are constants, see if FVal is the
3967 // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
3968 // instead of a CSEL in that case.
3969 if (TrueVal == ~FalseVal) {
3970 Opcode = AArch64ISD::CSINV;
3971 } else if (TrueVal == -FalseVal) {
3972 Opcode = AArch64ISD::CSNEG;
3973 } else if (TVal.getValueType() == MVT::i32) {
3974 // If our operands are only 32-bit wide, make sure we use 32-bit
3975 // arithmetic for the check whether we can use CSINC. This ensures that
3976 // the addition in the check will wrap around properly in case there is
3977 // an overflow (which would not be the case if we do the check with
3978 // 64-bit arithmetic).
3979 const uint32_t TrueVal32 = CTVal->getZExtValue();
3980 const uint32_t FalseVal32 = CFVal->getZExtValue();
3982 if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
3983 Opcode = AArch64ISD::CSINC;
3985 if (TrueVal32 > FalseVal32) {
3989 // 64-bit check whether we can use CSINC.
3990 } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) {
3991 Opcode = AArch64ISD::CSINC;
3993 if (TrueVal > FalseVal) {
3998 // Swap TVal and FVal if necessary.
4000 std::swap(TVal, FVal);
4001 std::swap(CTVal, CFVal);
4002 CC = ISD::getSetCCInverse(CC, true);
4005 if (Opcode != AArch64ISD::CSEL) {
4006 // Drop FVal since we can get its value by simply inverting/negating
4013 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
4015 EVT VT = TVal.getValueType();
4016 return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp);
4019 // Now we know we're dealing with FP values.
4020 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
4021 assert(LHS.getValueType() == RHS.getValueType());
4022 EVT VT = TVal.getValueType();
4023 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
4025 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
4026 // clean. Some of them require two CSELs to implement.
4027 AArch64CC::CondCode CC1, CC2;
4028 changeFPCCToAArch64CC(CC, CC1, CC2);
4029 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
4030 SDValue CS1 = DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
4032 // If we need a second CSEL, emit it, using the output of the first as the
4033 // RHS. We're effectively OR'ing the two CC's together.
4034 if (CC2 != AArch64CC::AL) {
4035 SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32);
4036 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
4039 // Otherwise, return the output of the first CSEL.
4043 SDValue AArch64TargetLowering::LowerSELECT_CC(SDValue Op,
4044 SelectionDAG &DAG) const {
4045 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
4046 SDValue LHS = Op.getOperand(0);
4047 SDValue RHS = Op.getOperand(1);
4048 SDValue TVal = Op.getOperand(2);
4049 SDValue FVal = Op.getOperand(3);
4051 return LowerSELECT_CC(CC, LHS, RHS, TVal, FVal, DL, DAG);
4054 SDValue AArch64TargetLowering::LowerSELECT(SDValue Op,
4055 SelectionDAG &DAG) const {
4056 SDValue CCVal = Op->getOperand(0);
4057 SDValue TVal = Op->getOperand(1);
4058 SDValue FVal = Op->getOperand(2);
4061 unsigned Opc = CCVal.getOpcode();
4062 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
4064 if (CCVal.getResNo() == 1 &&
4065 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
4066 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
4067 // Only lower legal XALUO ops.
4068 if (!DAG.getTargetLoweringInfo().isTypeLegal(CCVal->getValueType(0)))
4071 AArch64CC::CondCode OFCC;
4072 SDValue Value, Overflow;
4073 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, CCVal.getValue(0), DAG);
4074 SDValue CCVal = DAG.getConstant(OFCC, DL, MVT::i32);
4076 return DAG.getNode(AArch64ISD::CSEL, DL, Op.getValueType(), TVal, FVal,
4080 // Lower it the same way as we would lower a SELECT_CC node.
4083 if (CCVal.getOpcode() == ISD::SETCC) {
4084 LHS = CCVal.getOperand(0);
4085 RHS = CCVal.getOperand(1);
4086 CC = cast<CondCodeSDNode>(CCVal->getOperand(2))->get();
4089 RHS = DAG.getConstant(0, DL, CCVal.getValueType());
4092 return LowerSELECT_CC(CC, LHS, RHS, TVal, FVal, DL, DAG);
4095 SDValue AArch64TargetLowering::LowerJumpTable(SDValue Op,
4096 SelectionDAG &DAG) const {
4097 // Jump table entries as PC relative offsets. No additional tweaking
4098 // is necessary here. Just get the address of the jump table.
4099 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4100 EVT PtrVT = getPointerTy(DAG.getDataLayout());
4103 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
4104 !Subtarget->isTargetMachO()) {
4105 const unsigned char MO_NC = AArch64II::MO_NC;
4107 AArch64ISD::WrapperLarge, DL, PtrVT,
4108 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G3),
4109 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G2 | MO_NC),
4110 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G1 | MO_NC),
4111 DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
4112 AArch64II::MO_G0 | MO_NC));
4116 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_PAGE);
4117 SDValue Lo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
4118 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
4119 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
4120 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
4123 SDValue AArch64TargetLowering::LowerConstantPool(SDValue Op,
4124 SelectionDAG &DAG) const {
4125 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4126 EVT PtrVT = getPointerTy(DAG.getDataLayout());
4129 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
4130 // Use the GOT for the large code model on iOS.
4131 if (Subtarget->isTargetMachO()) {
4132 SDValue GotAddr = DAG.getTargetConstantPool(
4133 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
4135 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
4138 const unsigned char MO_NC = AArch64II::MO_NC;
4140 AArch64ISD::WrapperLarge, DL, PtrVT,
4141 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4142 CP->getOffset(), AArch64II::MO_G3),
4143 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4144 CP->getOffset(), AArch64II::MO_G2 | MO_NC),
4145 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4146 CP->getOffset(), AArch64II::MO_G1 | MO_NC),
4147 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4148 CP->getOffset(), AArch64II::MO_G0 | MO_NC));
4150 // Use ADRP/ADD or ADRP/LDR for everything else: the small memory model on
4151 // ELF, the only valid one on Darwin.
4153 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4154 CP->getOffset(), AArch64II::MO_PAGE);
4155 SDValue Lo = DAG.getTargetConstantPool(
4156 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
4157 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
4159 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
4160 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
4164 SDValue AArch64TargetLowering::LowerBlockAddress(SDValue Op,
4165 SelectionDAG &DAG) const {
4166 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
4167 EVT PtrVT = getPointerTy(DAG.getDataLayout());
4169 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
4170 !Subtarget->isTargetMachO()) {
4171 const unsigned char MO_NC = AArch64II::MO_NC;
4173 AArch64ISD::WrapperLarge, DL, PtrVT,
4174 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G3),
4175 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
4176 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
4177 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
4179 SDValue Hi = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGE);
4180 SDValue Lo = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGEOFF |
4182 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
4183 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
4187 SDValue AArch64TargetLowering::LowerDarwin_VASTART(SDValue Op,
4188 SelectionDAG &DAG) const {
4189 AArch64FunctionInfo *FuncInfo =
4190 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
4193 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(),
4194 getPointerTy(DAG.getDataLayout()));
4195 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
4196 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
4197 MachinePointerInfo(SV), false, false, 0);
4200 SDValue AArch64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
4201 SelectionDAG &DAG) const {
4202 // The layout of the va_list struct is specified in the AArch64 Procedure Call
4203 // Standard, section B.3.
4204 MachineFunction &MF = DAG.getMachineFunction();
4205 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
4206 auto PtrVT = getPointerTy(DAG.getDataLayout());
4209 SDValue Chain = Op.getOperand(0);
4210 SDValue VAList = Op.getOperand(1);
4211 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
4212 SmallVector<SDValue, 4> MemOps;
4214 // void *__stack at offset 0
4215 SDValue Stack = DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), PtrVT);
4216 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
4217 MachinePointerInfo(SV), false, false, 8));
4219 // void *__gr_top at offset 8
4220 int GPRSize = FuncInfo->getVarArgsGPRSize();
4222 SDValue GRTop, GRTopAddr;
4225 DAG.getNode(ISD::ADD, DL, PtrVT, VAList, DAG.getConstant(8, DL, PtrVT));
4227 GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), PtrVT);
4228 GRTop = DAG.getNode(ISD::ADD, DL, PtrVT, GRTop,
4229 DAG.getConstant(GPRSize, DL, PtrVT));
4231 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
4232 MachinePointerInfo(SV, 8), false, false, 8));
4235 // void *__vr_top at offset 16
4236 int FPRSize = FuncInfo->getVarArgsFPRSize();
4238 SDValue VRTop, VRTopAddr;
4239 VRTopAddr = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
4240 DAG.getConstant(16, DL, PtrVT));
4242 VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), PtrVT);
4243 VRTop = DAG.getNode(ISD::ADD, DL, PtrVT, VRTop,
4244 DAG.getConstant(FPRSize, DL, PtrVT));
4246 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
4247 MachinePointerInfo(SV, 16), false, false, 8));
4250 // int __gr_offs at offset 24
4251 SDValue GROffsAddr =
4252 DAG.getNode(ISD::ADD, DL, PtrVT, VAList, DAG.getConstant(24, DL, PtrVT));
4253 MemOps.push_back(DAG.getStore(Chain, DL,
4254 DAG.getConstant(-GPRSize, DL, MVT::i32),
4255 GROffsAddr, MachinePointerInfo(SV, 24), false,
4258 // int __vr_offs at offset 28
4259 SDValue VROffsAddr =
4260 DAG.getNode(ISD::ADD, DL, PtrVT, VAList, DAG.getConstant(28, DL, PtrVT));
4261 MemOps.push_back(DAG.getStore(Chain, DL,
4262 DAG.getConstant(-FPRSize, DL, MVT::i32),
4263 VROffsAddr, MachinePointerInfo(SV, 28), false,
4266 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
4269 SDValue AArch64TargetLowering::LowerVASTART(SDValue Op,
4270 SelectionDAG &DAG) const {
4271 return Subtarget->isTargetDarwin() ? LowerDarwin_VASTART(Op, DAG)
4272 : LowerAAPCS_VASTART(Op, DAG);
4275 SDValue AArch64TargetLowering::LowerVACOPY(SDValue Op,
4276 SelectionDAG &DAG) const {
4277 // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
4280 unsigned VaListSize = Subtarget->isTargetDarwin() ? 8 : 32;
4281 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
4282 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
4284 return DAG.getMemcpy(Op.getOperand(0), DL, Op.getOperand(1),
4286 DAG.getConstant(VaListSize, DL, MVT::i32),
4287 8, false, false, false, MachinePointerInfo(DestSV),
4288 MachinePointerInfo(SrcSV));
4291 SDValue AArch64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
4292 assert(Subtarget->isTargetDarwin() &&
4293 "automatic va_arg instruction only works on Darwin");
4295 const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
4296 EVT VT = Op.getValueType();
4298 SDValue Chain = Op.getOperand(0);
4299 SDValue Addr = Op.getOperand(1);
4300 unsigned Align = Op.getConstantOperandVal(3);
4301 auto PtrVT = getPointerTy(DAG.getDataLayout());
4303 SDValue VAList = DAG.getLoad(PtrVT, DL, Chain, Addr, MachinePointerInfo(V),
4304 false, false, false, 0);
4305 Chain = VAList.getValue(1);
4308 assert(((Align & (Align - 1)) == 0) && "Expected Align to be a power of 2");
4309 VAList = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
4310 DAG.getConstant(Align - 1, DL, PtrVT));
4311 VAList = DAG.getNode(ISD::AND, DL, PtrVT, VAList,
4312 DAG.getConstant(-(int64_t)Align, DL, PtrVT));
4315 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
4316 uint64_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
4318 // Scalar integer and FP values smaller than 64 bits are implicitly extended
4319 // up to 64 bits. At the very least, we have to increase the striding of the
4320 // vaargs list to match this, and for FP values we need to introduce
4321 // FP_ROUND nodes as well.
4322 if (VT.isInteger() && !VT.isVector())
4324 bool NeedFPTrunc = false;
4325 if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
4330 // Increment the pointer, VAList, to the next vaarg
4331 SDValue VANext = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
4332 DAG.getConstant(ArgSize, DL, PtrVT));
4333 // Store the incremented VAList to the legalized pointer
4334 SDValue APStore = DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V),
4337 // Load the actual argument out of the pointer VAList
4339 // Load the value as an f64.
4340 SDValue WideFP = DAG.getLoad(MVT::f64, DL, APStore, VAList,
4341 MachinePointerInfo(), false, false, false, 0);
4342 // Round the value down to an f32.
4343 SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
4344 DAG.getIntPtrConstant(1, DL));
4345 SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
4346 // Merge the rounded value with the chain output of the load.
4347 return DAG.getMergeValues(Ops, DL);
4350 return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo(), false,
4354 SDValue AArch64TargetLowering::LowerFRAMEADDR(SDValue Op,
4355 SelectionDAG &DAG) const {
4356 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4357 MFI->setFrameAddressIsTaken(true);
4359 EVT VT = Op.getValueType();
4361 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4363 DAG.getCopyFromReg(DAG.getEntryNode(), DL, AArch64::FP, VT);
4365 FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
4366 MachinePointerInfo(), false, false, false, 0);
4370 // FIXME? Maybe this could be a TableGen attribute on some registers and
4371 // this table could be generated automatically from RegInfo.
4372 unsigned AArch64TargetLowering::getRegisterByName(const char* RegName, EVT VT,
4373 SelectionDAG &DAG) const {
4374 unsigned Reg = StringSwitch<unsigned>(RegName)
4375 .Case("sp", AArch64::SP)
4379 report_fatal_error(Twine("Invalid register name \""
4380 + StringRef(RegName) + "\"."));
4383 SDValue AArch64TargetLowering::LowerRETURNADDR(SDValue Op,
4384 SelectionDAG &DAG) const {
4385 MachineFunction &MF = DAG.getMachineFunction();
4386 MachineFrameInfo *MFI = MF.getFrameInfo();
4387 MFI->setReturnAddressIsTaken(true);
4389 EVT VT = Op.getValueType();
4391 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4393 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4394 SDValue Offset = DAG.getConstant(8, DL, getPointerTy(DAG.getDataLayout()));
4395 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
4396 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
4397 MachinePointerInfo(), false, false, false, 0);
4400 // Return LR, which contains the return address. Mark it an implicit live-in.
4401 unsigned Reg = MF.addLiveIn(AArch64::LR, &AArch64::GPR64RegClass);
4402 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
4405 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4406 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4407 SDValue AArch64TargetLowering::LowerShiftRightParts(SDValue Op,
4408 SelectionDAG &DAG) const {
4409 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4410 EVT VT = Op.getValueType();
4411 unsigned VTBits = VT.getSizeInBits();
4413 SDValue ShOpLo = Op.getOperand(0);
4414 SDValue ShOpHi = Op.getOperand(1);
4415 SDValue ShAmt = Op.getOperand(2);
4416 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4418 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4420 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4421 DAG.getConstant(VTBits, dl, MVT::i64), ShAmt);
4422 SDValue HiBitsForLo = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4424 // Unfortunately, if ShAmt == 0, we just calculated "(SHL ShOpHi, 64)" which
4425 // is "undef". We wanted 0, so CSEL it directly.
4426 SDValue Cmp = emitComparison(ShAmt, DAG.getConstant(0, dl, MVT::i64),
4427 ISD::SETEQ, dl, DAG);
4428 SDValue CCVal = DAG.getConstant(AArch64CC::EQ, dl, MVT::i32);
4430 DAG.getNode(AArch64ISD::CSEL, dl, VT, DAG.getConstant(0, dl, MVT::i64),
4431 HiBitsForLo, CCVal, Cmp);
4433 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4434 DAG.getConstant(VTBits, dl, MVT::i64));
4436 SDValue LoBitsForLo = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4437 SDValue LoForNormalShift =
4438 DAG.getNode(ISD::OR, dl, VT, LoBitsForLo, HiBitsForLo);
4440 Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, dl, MVT::i64), ISD::SETGE,
4442 CCVal = DAG.getConstant(AArch64CC::GE, dl, MVT::i32);
4443 SDValue LoForBigShift = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4444 SDValue Lo = DAG.getNode(AArch64ISD::CSEL, dl, VT, LoForBigShift,
4445 LoForNormalShift, CCVal, Cmp);
4447 // AArch64 shifts larger than the register width are wrapped rather than
4448 // clamped, so we can't just emit "hi >> x".
4449 SDValue HiForNormalShift = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4450 SDValue HiForBigShift =
4452 ? DAG.getNode(Opc, dl, VT, ShOpHi,
4453 DAG.getConstant(VTBits - 1, dl, MVT::i64))
4454 : DAG.getConstant(0, dl, VT);
4455 SDValue Hi = DAG.getNode(AArch64ISD::CSEL, dl, VT, HiForBigShift,
4456 HiForNormalShift, CCVal, Cmp);
4458 SDValue Ops[2] = { Lo, Hi };
4459 return DAG.getMergeValues(Ops, dl);
4463 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4464 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4465 SDValue AArch64TargetLowering::LowerShiftLeftParts(SDValue Op,
4466 SelectionDAG &DAG) const {
4467 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4468 EVT VT = Op.getValueType();
4469 unsigned VTBits = VT.getSizeInBits();
4471 SDValue ShOpLo = Op.getOperand(0);
4472 SDValue ShOpHi = Op.getOperand(1);
4473 SDValue ShAmt = Op.getOperand(2);
4475 assert(Op.getOpcode() == ISD::SHL_PARTS);
4476 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4477 DAG.getConstant(VTBits, dl, MVT::i64), ShAmt);
4478 SDValue LoBitsForHi = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4480 // Unfortunately, if ShAmt == 0, we just calculated "(SRL ShOpLo, 64)" which
4481 // is "undef". We wanted 0, so CSEL it directly.
4482 SDValue Cmp = emitComparison(ShAmt, DAG.getConstant(0, dl, MVT::i64),
4483 ISD::SETEQ, dl, DAG);
4484 SDValue CCVal = DAG.getConstant(AArch64CC::EQ, dl, MVT::i32);
4486 DAG.getNode(AArch64ISD::CSEL, dl, VT, DAG.getConstant(0, dl, MVT::i64),
4487 LoBitsForHi, CCVal, Cmp);
4489 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4490 DAG.getConstant(VTBits, dl, MVT::i64));
4491 SDValue HiBitsForHi = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4492 SDValue HiForNormalShift =
4493 DAG.getNode(ISD::OR, dl, VT, LoBitsForHi, HiBitsForHi);
4495 SDValue HiForBigShift = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4497 Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, dl, MVT::i64), ISD::SETGE,
4499 CCVal = DAG.getConstant(AArch64CC::GE, dl, MVT::i32);
4500 SDValue Hi = DAG.getNode(AArch64ISD::CSEL, dl, VT, HiForBigShift,
4501 HiForNormalShift, CCVal, Cmp);
4503 // AArch64 shifts of larger than register sizes are wrapped rather than
4504 // clamped, so we can't just emit "lo << a" if a is too big.
4505 SDValue LoForBigShift = DAG.getConstant(0, dl, VT);
4506 SDValue LoForNormalShift = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4507 SDValue Lo = DAG.getNode(AArch64ISD::CSEL, dl, VT, LoForBigShift,
4508 LoForNormalShift, CCVal, Cmp);
4510 SDValue Ops[2] = { Lo, Hi };
4511 return DAG.getMergeValues(Ops, dl);
4514 bool AArch64TargetLowering::isOffsetFoldingLegal(
4515 const GlobalAddressSDNode *GA) const {
4516 // The AArch64 target doesn't support folding offsets into global addresses.
4520 bool AArch64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4521 // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
4522 // FIXME: We should be able to handle f128 as well with a clever lowering.
4523 if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32))
4527 return AArch64_AM::getFP64Imm(Imm) != -1;
4528 else if (VT == MVT::f32)
4529 return AArch64_AM::getFP32Imm(Imm) != -1;
4533 //===----------------------------------------------------------------------===//
4534 // AArch64 Optimization Hooks
4535 //===----------------------------------------------------------------------===//
4537 //===----------------------------------------------------------------------===//
4538 // AArch64 Inline Assembly Support
4539 //===----------------------------------------------------------------------===//
4541 // Table of Constraints
4542 // TODO: This is the current set of constraints supported by ARM for the
4543 // compiler, not all of them may make sense, e.g. S may be difficult to support.
4545 // r - A general register
4546 // w - An FP/SIMD register of some size in the range v0-v31
4547 // x - An FP/SIMD register of some size in the range v0-v15
4548 // I - Constant that can be used with an ADD instruction
4549 // J - Constant that can be used with a SUB instruction
4550 // K - Constant that can be used with a 32-bit logical instruction
4551 // L - Constant that can be used with a 64-bit logical instruction
4552 // M - Constant that can be used as a 32-bit MOV immediate
4553 // N - Constant that can be used as a 64-bit MOV immediate
4554 // Q - A memory reference with base register and no offset
4555 // S - A symbolic address
4556 // Y - Floating point constant zero
4557 // Z - Integer constant zero
4559 // Note that general register operands will be output using their 64-bit x
4560 // register name, whatever the size of the variable, unless the asm operand
4561 // is prefixed by the %w modifier. Floating-point and SIMD register operands
4562 // will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
4565 /// getConstraintType - Given a constraint letter, return the type of
4566 /// constraint it is for this target.
4567 AArch64TargetLowering::ConstraintType
4568 AArch64TargetLowering::getConstraintType(StringRef Constraint) const {
4569 if (Constraint.size() == 1) {
4570 switch (Constraint[0]) {
4577 return C_RegisterClass;
4578 // An address with a single base register. Due to the way we
4579 // currently handle addresses it is the same as 'r'.
4584 return TargetLowering::getConstraintType(Constraint);
4587 /// Examine constraint type and operand type and determine a weight value.
4588 /// This object must already have been set up with the operand type
4589 /// and the current alternative constraint selected.
4590 TargetLowering::ConstraintWeight
4591 AArch64TargetLowering::getSingleConstraintMatchWeight(
4592 AsmOperandInfo &info, const char *constraint) const {
4593 ConstraintWeight weight = CW_Invalid;
4594 Value *CallOperandVal = info.CallOperandVal;
4595 // If we don't have a value, we can't do a match,
4596 // but allow it at the lowest weight.
4597 if (!CallOperandVal)
4599 Type *type = CallOperandVal->getType();
4600 // Look at the constraint type.
4601 switch (*constraint) {
4603 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
4607 if (type->isFloatingPointTy() || type->isVectorTy())
4608 weight = CW_Register;
4611 weight = CW_Constant;
4617 std::pair<unsigned, const TargetRegisterClass *>
4618 AArch64TargetLowering::getRegForInlineAsmConstraint(
4619 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
4620 if (Constraint.size() == 1) {
4621 switch (Constraint[0]) {
4623 if (VT.getSizeInBits() == 64)
4624 return std::make_pair(0U, &AArch64::GPR64commonRegClass);
4625 return std::make_pair(0U, &AArch64::GPR32commonRegClass);
4628 return std::make_pair(0U, &AArch64::FPR32RegClass);
4629 if (VT.getSizeInBits() == 64)
4630 return std::make_pair(0U, &AArch64::FPR64RegClass);
4631 if (VT.getSizeInBits() == 128)
4632 return std::make_pair(0U, &AArch64::FPR128RegClass);
4634 // The instructions that this constraint is designed for can
4635 // only take 128-bit registers so just use that regclass.
4637 if (VT.getSizeInBits() == 128)
4638 return std::make_pair(0U, &AArch64::FPR128_loRegClass);
4642 if (StringRef("{cc}").equals_lower(Constraint))
4643 return std::make_pair(unsigned(AArch64::NZCV), &AArch64::CCRRegClass);
4645 // Use the default implementation in TargetLowering to convert the register
4646 // constraint into a member of a register class.
4647 std::pair<unsigned, const TargetRegisterClass *> Res;
4648 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
4650 // Not found as a standard register?
4652 unsigned Size = Constraint.size();
4653 if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
4654 tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
4656 bool Failed = Constraint.slice(2, Size - 1).getAsInteger(10, RegNo);
4657 if (!Failed && RegNo >= 0 && RegNo <= 31) {
4658 // v0 - v31 are aliases of q0 - q31.
4659 // By default we'll emit v0-v31 for this unless there's a modifier where
4660 // we'll emit the correct register as well.
4661 Res.first = AArch64::FPR128RegClass.getRegister(RegNo);
4662 Res.second = &AArch64::FPR128RegClass;
4670 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4671 /// vector. If it is invalid, don't add anything to Ops.
4672 void AArch64TargetLowering::LowerAsmOperandForConstraint(
4673 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
4674 SelectionDAG &DAG) const {
4677 // Currently only support length 1 constraints.
4678 if (Constraint.length() != 1)
4681 char ConstraintLetter = Constraint[0];
4682 switch (ConstraintLetter) {
4686 // This set of constraints deal with valid constants for various instructions.
4687 // Validate and return a target constant for them if we can.
4689 // 'z' maps to xzr or wzr so it needs an input of 0.
4690 if (!isNullConstant(Op))
4693 if (Op.getValueType() == MVT::i64)
4694 Result = DAG.getRegister(AArch64::XZR, MVT::i64);
4696 Result = DAG.getRegister(AArch64::WZR, MVT::i32);
4706 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4710 // Grab the value and do some validation.
4711 uint64_t CVal = C->getZExtValue();
4712 switch (ConstraintLetter) {
4713 // The I constraint applies only to simple ADD or SUB immediate operands:
4714 // i.e. 0 to 4095 with optional shift by 12
4715 // The J constraint applies only to ADD or SUB immediates that would be
4716 // valid when negated, i.e. if [an add pattern] were to be output as a SUB
4717 // instruction [or vice versa], in other words -1 to -4095 with optional
4718 // left shift by 12.
4720 if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
4724 uint64_t NVal = -C->getSExtValue();
4725 if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal)) {
4726 CVal = C->getSExtValue();
4731 // The K and L constraints apply *only* to logical immediates, including
4732 // what used to be the MOVI alias for ORR (though the MOVI alias has now
4733 // been removed and MOV should be used). So these constraints have to
4734 // distinguish between bit patterns that are valid 32-bit or 64-bit
4735 // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
4736 // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
4739 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4743 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4746 // The M and N constraints are a superset of K and L respectively, for use
4747 // with the MOV (immediate) alias. As well as the logical immediates they
4748 // also match 32 or 64-bit immediates that can be loaded either using a
4749 // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
4750 // (M) or 64-bit 0x1234000000000000 (N) etc.
4751 // As a note some of this code is liberally stolen from the asm parser.
4753 if (!isUInt<32>(CVal))
4755 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4757 if ((CVal & 0xFFFF) == CVal)
4759 if ((CVal & 0xFFFF0000ULL) == CVal)
4761 uint64_t NCVal = ~(uint32_t)CVal;
4762 if ((NCVal & 0xFFFFULL) == NCVal)
4764 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4769 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4771 if ((CVal & 0xFFFFULL) == CVal)
4773 if ((CVal & 0xFFFF0000ULL) == CVal)
4775 if ((CVal & 0xFFFF00000000ULL) == CVal)
4777 if ((CVal & 0xFFFF000000000000ULL) == CVal)
4779 uint64_t NCVal = ~CVal;
4780 if ((NCVal & 0xFFFFULL) == NCVal)
4782 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4784 if ((NCVal & 0xFFFF00000000ULL) == NCVal)
4786 if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
4794 // All assembler immediates are 64-bit integers.
4795 Result = DAG.getTargetConstant(CVal, SDLoc(Op), MVT::i64);
4799 if (Result.getNode()) {
4800 Ops.push_back(Result);
4804 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
4807 //===----------------------------------------------------------------------===//
4808 // AArch64 Advanced SIMD Support
4809 //===----------------------------------------------------------------------===//
4811 /// WidenVector - Given a value in the V64 register class, produce the
4812 /// equivalent value in the V128 register class.
4813 static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
4814 EVT VT = V64Reg.getValueType();
4815 unsigned NarrowSize = VT.getVectorNumElements();
4816 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4817 MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
4820 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
4821 V64Reg, DAG.getConstant(0, DL, MVT::i32));
4824 /// getExtFactor - Determine the adjustment factor for the position when
4825 /// generating an "extract from vector registers" instruction.
4826 static unsigned getExtFactor(SDValue &V) {
4827 EVT EltType = V.getValueType().getVectorElementType();
4828 return EltType.getSizeInBits() / 8;
4831 /// NarrowVector - Given a value in the V128 register class, produce the
4832 /// equivalent value in the V64 register class.
4833 static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) {
4834 EVT VT = V128Reg.getValueType();
4835 unsigned WideSize = VT.getVectorNumElements();
4836 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4837 MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2);
4840 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, NarrowTy, V128Reg);
4843 // Gather data to see if the operation can be modelled as a
4844 // shuffle in combination with VEXTs.
4845 SDValue AArch64TargetLowering::ReconstructShuffle(SDValue Op,
4846 SelectionDAG &DAG) const {
4847 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
4849 EVT VT = Op.getValueType();
4850 unsigned NumElts = VT.getVectorNumElements();
4852 struct ShuffleSourceInfo {
4857 // We may insert some combination of BITCASTs and VEXT nodes to force Vec to
4858 // be compatible with the shuffle we intend to construct. As a result
4859 // ShuffleVec will be some sliding window into the original Vec.
4862 // Code should guarantee that element i in Vec starts at element "WindowBase
4863 // + i * WindowScale in ShuffleVec".
4867 bool operator ==(SDValue OtherVec) { return Vec == OtherVec; }
4868 ShuffleSourceInfo(SDValue Vec)
4869 : Vec(Vec), MinElt(UINT_MAX), MaxElt(0), ShuffleVec(Vec), WindowBase(0),
4873 // First gather all vectors used as an immediate source for this BUILD_VECTOR
4875 SmallVector<ShuffleSourceInfo, 2> Sources;
4876 for (unsigned i = 0; i < NumElts; ++i) {
4877 SDValue V = Op.getOperand(i);
4878 if (V.getOpcode() == ISD::UNDEF)
4880 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4881 // A shuffle can only come from building a vector from various
4882 // elements of other vectors.
4886 // Add this element source to the list if it's not already there.
4887 SDValue SourceVec = V.getOperand(0);
4888 auto Source = std::find(Sources.begin(), Sources.end(), SourceVec);
4889 if (Source == Sources.end())
4890 Source = Sources.insert(Sources.end(), ShuffleSourceInfo(SourceVec));
4892 // Update the minimum and maximum lane number seen.
4893 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4894 Source->MinElt = std::min(Source->MinElt, EltNo);
4895 Source->MaxElt = std::max(Source->MaxElt, EltNo);
4898 // Currently only do something sane when at most two source vectors
4900 if (Sources.size() > 2)
4903 // Find out the smallest element size among result and two sources, and use
4904 // it as element size to build the shuffle_vector.
4905 EVT SmallestEltTy = VT.getVectorElementType();
4906 for (auto &Source : Sources) {
4907 EVT SrcEltTy = Source.Vec.getValueType().getVectorElementType();
4908 if (SrcEltTy.bitsLT(SmallestEltTy)) {
4909 SmallestEltTy = SrcEltTy;
4912 unsigned ResMultiplier =
4913 VT.getVectorElementType().getSizeInBits() / SmallestEltTy.getSizeInBits();
4914 NumElts = VT.getSizeInBits() / SmallestEltTy.getSizeInBits();
4915 EVT ShuffleVT = EVT::getVectorVT(*DAG.getContext(), SmallestEltTy, NumElts);
4917 // If the source vector is too wide or too narrow, we may nevertheless be able
4918 // to construct a compatible shuffle either by concatenating it with UNDEF or
4919 // extracting a suitable range of elements.
4920 for (auto &Src : Sources) {
4921 EVT SrcVT = Src.ShuffleVec.getValueType();
4923 if (SrcVT.getSizeInBits() == VT.getSizeInBits())
4926 // This stage of the search produces a source with the same element type as
4927 // the original, but with a total width matching the BUILD_VECTOR output.
4928 EVT EltVT = SrcVT.getVectorElementType();
4929 unsigned NumSrcElts = VT.getSizeInBits() / EltVT.getSizeInBits();
4930 EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumSrcElts);
4932 if (SrcVT.getSizeInBits() < VT.getSizeInBits()) {
4933 assert(2 * SrcVT.getSizeInBits() == VT.getSizeInBits());
4934 // We can pad out the smaller vector for free, so if it's part of a
4937 DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, Src.ShuffleVec,
4938 DAG.getUNDEF(Src.ShuffleVec.getValueType()));
4942 assert(SrcVT.getSizeInBits() == 2 * VT.getSizeInBits());
4944 if (Src.MaxElt - Src.MinElt >= NumSrcElts) {
4945 // Span too large for a VEXT to cope
4949 if (Src.MinElt >= NumSrcElts) {
4950 // The extraction can just take the second half
4952 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4953 DAG.getConstant(NumSrcElts, dl, MVT::i64));
4954 Src.WindowBase = -NumSrcElts;
4955 } else if (Src.MaxElt < NumSrcElts) {
4956 // The extraction can just take the first half
4958 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4959 DAG.getConstant(0, dl, MVT::i64));
4961 // An actual VEXT is needed
4963 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4964 DAG.getConstant(0, dl, MVT::i64));
4966 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4967 DAG.getConstant(NumSrcElts, dl, MVT::i64));
4968 unsigned Imm = Src.MinElt * getExtFactor(VEXTSrc1);
4970 Src.ShuffleVec = DAG.getNode(AArch64ISD::EXT, dl, DestVT, VEXTSrc1,
4972 DAG.getConstant(Imm, dl, MVT::i32));
4973 Src.WindowBase = -Src.MinElt;
4977 // Another possible incompatibility occurs from the vector element types. We
4978 // can fix this by bitcasting the source vectors to the same type we intend
4980 for (auto &Src : Sources) {
4981 EVT SrcEltTy = Src.ShuffleVec.getValueType().getVectorElementType();
4982 if (SrcEltTy == SmallestEltTy)
4984 assert(ShuffleVT.getVectorElementType() == SmallestEltTy);
4985 Src.ShuffleVec = DAG.getNode(ISD::BITCAST, dl, ShuffleVT, Src.ShuffleVec);
4986 Src.WindowScale = SrcEltTy.getSizeInBits() / SmallestEltTy.getSizeInBits();
4987 Src.WindowBase *= Src.WindowScale;
4990 // Final sanity check before we try to actually produce a shuffle.
4992 for (auto Src : Sources)
4993 assert(Src.ShuffleVec.getValueType() == ShuffleVT);
4996 // The stars all align, our next step is to produce the mask for the shuffle.
4997 SmallVector<int, 8> Mask(ShuffleVT.getVectorNumElements(), -1);
4998 int BitsPerShuffleLane = ShuffleVT.getVectorElementType().getSizeInBits();
4999 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
5000 SDValue Entry = Op.getOperand(i);
5001 if (Entry.getOpcode() == ISD::UNDEF)
5004 auto Src = std::find(Sources.begin(), Sources.end(), Entry.getOperand(0));
5005 int EltNo = cast<ConstantSDNode>(Entry.getOperand(1))->getSExtValue();
5007 // EXTRACT_VECTOR_ELT performs an implicit any_ext; BUILD_VECTOR an implicit
5008 // trunc. So only std::min(SrcBits, DestBits) actually get defined in this
5010 EVT OrigEltTy = Entry.getOperand(0).getValueType().getVectorElementType();
5011 int BitsDefined = std::min(OrigEltTy.getSizeInBits(),
5012 VT.getVectorElementType().getSizeInBits());
5013 int LanesDefined = BitsDefined / BitsPerShuffleLane;
5015 // This source is expected to fill ResMultiplier lanes of the final shuffle,
5016 // starting at the appropriate offset.
5017 int *LaneMask = &Mask[i * ResMultiplier];
5019 int ExtractBase = EltNo * Src->WindowScale + Src->WindowBase;
5020 ExtractBase += NumElts * (Src - Sources.begin());
5021 for (int j = 0; j < LanesDefined; ++j)
5022 LaneMask[j] = ExtractBase + j;
5025 // Final check before we try to produce nonsense...
5026 if (!isShuffleMaskLegal(Mask, ShuffleVT))
5029 SDValue ShuffleOps[] = { DAG.getUNDEF(ShuffleVT), DAG.getUNDEF(ShuffleVT) };
5030 for (unsigned i = 0; i < Sources.size(); ++i)
5031 ShuffleOps[i] = Sources[i].ShuffleVec;
5033 SDValue Shuffle = DAG.getVectorShuffle(ShuffleVT, dl, ShuffleOps[0],
5034 ShuffleOps[1], &Mask[0]);
5035 return DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
5038 // check if an EXT instruction can handle the shuffle mask when the
5039 // vector sources of the shuffle are the same.
5040 static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
5041 unsigned NumElts = VT.getVectorNumElements();
5043 // Assume that the first shuffle index is not UNDEF. Fail if it is.
5049 // If this is a VEXT shuffle, the immediate value is the index of the first
5050 // element. The other shuffle indices must be the successive elements after
5052 unsigned ExpectedElt = Imm;
5053 for (unsigned i = 1; i < NumElts; ++i) {
5054 // Increment the expected index. If it wraps around, just follow it
5055 // back to index zero and keep going.
5057 if (ExpectedElt == NumElts)
5061 continue; // ignore UNDEF indices
5062 if (ExpectedElt != static_cast<unsigned>(M[i]))
5069 // check if an EXT instruction can handle the shuffle mask when the
5070 // vector sources of the shuffle are different.
5071 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
5073 // Look for the first non-undef element.
5074 const int *FirstRealElt = std::find_if(M.begin(), M.end(),
5075 [](int Elt) {return Elt >= 0;});
5077 // Benefit form APInt to handle overflow when calculating expected element.
5078 unsigned NumElts = VT.getVectorNumElements();
5079 unsigned MaskBits = APInt(32, NumElts * 2).logBase2();
5080 APInt ExpectedElt = APInt(MaskBits, *FirstRealElt + 1);
5081 // The following shuffle indices must be the successive elements after the
5082 // first real element.
5083 const int *FirstWrongElt = std::find_if(FirstRealElt + 1, M.end(),
5084 [&](int Elt) {return Elt != ExpectedElt++ && Elt != -1;});
5085 if (FirstWrongElt != M.end())
5088 // The index of an EXT is the first element if it is not UNDEF.
5089 // Watch out for the beginning UNDEFs. The EXT index should be the expected
5090 // value of the first element. E.g.
5091 // <-1, -1, 3, ...> is treated as <1, 2, 3, ...>.
5092 // <-1, -1, 0, 1, ...> is treated as <2*NumElts-2, 2*NumElts-1, 0, 1, ...>.
5093 // ExpectedElt is the last mask index plus 1.
5094 Imm = ExpectedElt.getZExtValue();
5096 // There are two difference cases requiring to reverse input vectors.
5097 // For example, for vector <4 x i32> we have the following cases,
5098 // Case 1: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, -1, 0>)
5099 // Case 2: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, 7, 0>)
5100 // For both cases, we finally use mask <5, 6, 7, 0>, which requires
5101 // to reverse two input vectors.
5110 /// isREVMask - Check if a vector shuffle corresponds to a REV
5111 /// instruction with the specified blocksize. (The order of the elements
5112 /// within each block of the vector is reversed.)
5113 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
5114 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
5115 "Only possible block sizes for REV are: 16, 32, 64");
5117 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5121 unsigned NumElts = VT.getVectorNumElements();
5122 unsigned BlockElts = M[0] + 1;
5123 // If the first shuffle index is UNDEF, be optimistic.
5125 BlockElts = BlockSize / EltSz;
5127 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
5130 for (unsigned i = 0; i < NumElts; ++i) {
5132 continue; // ignore UNDEF indices
5133 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
5140 static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5141 unsigned NumElts = VT.getVectorNumElements();
5142 WhichResult = (M[0] == 0 ? 0 : 1);
5143 unsigned Idx = WhichResult * NumElts / 2;
5144 for (unsigned i = 0; i != NumElts; i += 2) {
5145 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
5146 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
5154 static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5155 unsigned NumElts = VT.getVectorNumElements();
5156 WhichResult = (M[0] == 0 ? 0 : 1);
5157 for (unsigned i = 0; i != NumElts; ++i) {
5159 continue; // ignore UNDEF indices
5160 if ((unsigned)M[i] != 2 * i + WhichResult)
5167 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5168 unsigned NumElts = VT.getVectorNumElements();
5169 WhichResult = (M[0] == 0 ? 0 : 1);
5170 for (unsigned i = 0; i < NumElts; i += 2) {
5171 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
5172 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + NumElts + WhichResult))
5178 /// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of
5179 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5180 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
5181 static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5182 unsigned NumElts = VT.getVectorNumElements();
5183 WhichResult = (M[0] == 0 ? 0 : 1);
5184 unsigned Idx = WhichResult * NumElts / 2;
5185 for (unsigned i = 0; i != NumElts; i += 2) {
5186 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
5187 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx))
5195 /// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of
5196 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5197 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
5198 static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5199 unsigned Half = VT.getVectorNumElements() / 2;
5200 WhichResult = (M[0] == 0 ? 0 : 1);
5201 for (unsigned j = 0; j != 2; ++j) {
5202 unsigned Idx = WhichResult;
5203 for (unsigned i = 0; i != Half; ++i) {
5204 int MIdx = M[i + j * Half];
5205 if (MIdx >= 0 && (unsigned)MIdx != Idx)
5214 /// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of
5215 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5216 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
5217 static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5218 unsigned NumElts = VT.getVectorNumElements();
5219 WhichResult = (M[0] == 0 ? 0 : 1);
5220 for (unsigned i = 0; i < NumElts; i += 2) {
5221 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
5222 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult))
5228 static bool isINSMask(ArrayRef<int> M, int NumInputElements,
5229 bool &DstIsLeft, int &Anomaly) {
5230 if (M.size() != static_cast<size_t>(NumInputElements))
5233 int NumLHSMatch = 0, NumRHSMatch = 0;
5234 int LastLHSMismatch = -1, LastRHSMismatch = -1;
5236 for (int i = 0; i < NumInputElements; ++i) {
5246 LastLHSMismatch = i;
5248 if (M[i] == i + NumInputElements)
5251 LastRHSMismatch = i;
5254 if (NumLHSMatch == NumInputElements - 1) {
5256 Anomaly = LastLHSMismatch;
5258 } else if (NumRHSMatch == NumInputElements - 1) {
5260 Anomaly = LastRHSMismatch;
5267 static bool isConcatMask(ArrayRef<int> Mask, EVT VT, bool SplitLHS) {
5268 if (VT.getSizeInBits() != 128)
5271 unsigned NumElts = VT.getVectorNumElements();
5273 for (int I = 0, E = NumElts / 2; I != E; I++) {
5278 int Offset = NumElts / 2;
5279 for (int I = NumElts / 2, E = NumElts; I != E; I++) {
5280 if (Mask[I] != I + SplitLHS * Offset)
5287 static SDValue tryFormConcatFromShuffle(SDValue Op, SelectionDAG &DAG) {
5289 EVT VT = Op.getValueType();
5290 SDValue V0 = Op.getOperand(0);
5291 SDValue V1 = Op.getOperand(1);
5292 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op)->getMask();
5294 if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() ||
5295 VT.getVectorElementType() != V1.getValueType().getVectorElementType())
5298 bool SplitV0 = V0.getValueType().getSizeInBits() == 128;
5300 if (!isConcatMask(Mask, VT, SplitV0))
5303 EVT CastVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
5304 VT.getVectorNumElements() / 2);
5306 V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0,
5307 DAG.getConstant(0, DL, MVT::i64));
5309 if (V1.getValueType().getSizeInBits() == 128) {
5310 V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1,
5311 DAG.getConstant(0, DL, MVT::i64));
5313 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1);
5316 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5317 /// the specified operations to build the shuffle.
5318 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5319 SDValue RHS, SelectionDAG &DAG,
5321 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5322 unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1);
5323 unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1);
5326 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5335 OP_VUZPL, // VUZP, left result
5336 OP_VUZPR, // VUZP, right result
5337 OP_VZIPL, // VZIP, left result
5338 OP_VZIPR, // VZIP, right result
5339 OP_VTRNL, // VTRN, left result
5340 OP_VTRNR // VTRN, right result
5343 if (OpNum == OP_COPY) {
5344 if (LHSID == (1 * 9 + 2) * 9 + 3)
5346 assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!");
5350 SDValue OpLHS, OpRHS;
5351 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5352 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5353 EVT VT = OpLHS.getValueType();
5357 llvm_unreachable("Unknown shuffle opcode!");
5359 // VREV divides the vector in half and swaps within the half.
5360 if (VT.getVectorElementType() == MVT::i32 ||
5361 VT.getVectorElementType() == MVT::f32)
5362 return DAG.getNode(AArch64ISD::REV64, dl, VT, OpLHS);
5363 // vrev <4 x i16> -> REV32
5364 if (VT.getVectorElementType() == MVT::i16 ||
5365 VT.getVectorElementType() == MVT::f16)
5366 return DAG.getNode(AArch64ISD::REV32, dl, VT, OpLHS);
5367 // vrev <4 x i8> -> REV16
5368 assert(VT.getVectorElementType() == MVT::i8);
5369 return DAG.getNode(AArch64ISD::REV16, dl, VT, OpLHS);
5374 EVT EltTy = VT.getVectorElementType();
5376 if (EltTy == MVT::i8)
5377 Opcode = AArch64ISD::DUPLANE8;
5378 else if (EltTy == MVT::i16 || EltTy == MVT::f16)
5379 Opcode = AArch64ISD::DUPLANE16;
5380 else if (EltTy == MVT::i32 || EltTy == MVT::f32)
5381 Opcode = AArch64ISD::DUPLANE32;
5382 else if (EltTy == MVT::i64 || EltTy == MVT::f64)
5383 Opcode = AArch64ISD::DUPLANE64;
5385 llvm_unreachable("Invalid vector element type?");
5387 if (VT.getSizeInBits() == 64)
5388 OpLHS = WidenVector(OpLHS, DAG);
5389 SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, dl, MVT::i64);
5390 return DAG.getNode(Opcode, dl, VT, OpLHS, Lane);
5395 unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
5396 return DAG.getNode(AArch64ISD::EXT, dl, VT, OpLHS, OpRHS,
5397 DAG.getConstant(Imm, dl, MVT::i32));
5400 return DAG.getNode(AArch64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS,
5403 return DAG.getNode(AArch64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS,
5406 return DAG.getNode(AArch64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS,
5409 return DAG.getNode(AArch64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS,
5412 return DAG.getNode(AArch64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS,
5415 return DAG.getNode(AArch64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS,
5420 static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask,
5421 SelectionDAG &DAG) {
5422 // Check to see if we can use the TBL instruction.
5423 SDValue V1 = Op.getOperand(0);
5424 SDValue V2 = Op.getOperand(1);
5427 EVT EltVT = Op.getValueType().getVectorElementType();
5428 unsigned BytesPerElt = EltVT.getSizeInBits() / 8;
5430 SmallVector<SDValue, 8> TBLMask;
5431 for (int Val : ShuffleMask) {
5432 for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
5433 unsigned Offset = Byte + Val * BytesPerElt;
5434 TBLMask.push_back(DAG.getConstant(Offset, DL, MVT::i32));
5438 MVT IndexVT = MVT::v8i8;
5439 unsigned IndexLen = 8;
5440 if (Op.getValueType().getSizeInBits() == 128) {
5441 IndexVT = MVT::v16i8;
5445 SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
5446 SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);
5449 if (V2.getNode()->getOpcode() == ISD::UNDEF) {
5451 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
5452 Shuffle = DAG.getNode(
5453 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5454 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, DL, MVT::i32), V1Cst,
5455 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5456 makeArrayRef(TBLMask.data(), IndexLen)));
5458 if (IndexLen == 8) {
5459 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
5460 Shuffle = DAG.getNode(
5461 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5462 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, DL, MVT::i32), V1Cst,
5463 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5464 makeArrayRef(TBLMask.data(), IndexLen)));
5466 // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
5467 // cannot currently represent the register constraints on the input
5469 // Shuffle = DAG.getNode(AArch64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
5470 // DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5471 // &TBLMask[0], IndexLen));
5472 Shuffle = DAG.getNode(
5473 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5474 DAG.getConstant(Intrinsic::aarch64_neon_tbl2, DL, MVT::i32),
5476 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5477 makeArrayRef(TBLMask.data(), IndexLen)));
5480 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
5483 static unsigned getDUPLANEOp(EVT EltType) {
5484 if (EltType == MVT::i8)
5485 return AArch64ISD::DUPLANE8;
5486 if (EltType == MVT::i16 || EltType == MVT::f16)
5487 return AArch64ISD::DUPLANE16;
5488 if (EltType == MVT::i32 || EltType == MVT::f32)
5489 return AArch64ISD::DUPLANE32;
5490 if (EltType == MVT::i64 || EltType == MVT::f64)
5491 return AArch64ISD::DUPLANE64;
5493 llvm_unreachable("Invalid vector element type?");
5496 SDValue AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
5497 SelectionDAG &DAG) const {
5499 EVT VT = Op.getValueType();
5501 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5503 // Convert shuffles that are directly supported on NEON to target-specific
5504 // DAG nodes, instead of keeping them as shuffles and matching them again
5505 // during code selection. This is more efficient and avoids the possibility
5506 // of inconsistencies between legalization and selection.
5507 ArrayRef<int> ShuffleMask = SVN->getMask();
5509 SDValue V1 = Op.getOperand(0);
5510 SDValue V2 = Op.getOperand(1);
5512 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0],
5513 V1.getValueType().getSimpleVT())) {
5514 int Lane = SVN->getSplatIndex();
5515 // If this is undef splat, generate it via "just" vdup, if possible.
5519 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
5520 return DAG.getNode(AArch64ISD::DUP, dl, V1.getValueType(),
5522 // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
5523 // constant. If so, we can just reference the lane's definition directly.
5524 if (V1.getOpcode() == ISD::BUILD_VECTOR &&
5525 !isa<ConstantSDNode>(V1.getOperand(Lane)))
5526 return DAG.getNode(AArch64ISD::DUP, dl, VT, V1.getOperand(Lane));
5528 // Otherwise, duplicate from the lane of the input vector.
5529 unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
5531 // SelectionDAGBuilder may have "helpfully" already extracted or conatenated
5532 // to make a vector of the same size as this SHUFFLE. We can ignore the
5533 // extract entirely, and canonicalise the concat using WidenVector.
5534 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
5535 Lane += cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
5536 V1 = V1.getOperand(0);
5537 } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
5538 unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
5539 Lane -= Idx * VT.getVectorNumElements() / 2;
5540 V1 = WidenVector(V1.getOperand(Idx), DAG);
5541 } else if (VT.getSizeInBits() == 64)
5542 V1 = WidenVector(V1, DAG);
5544 return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, dl, MVT::i64));
5547 if (isREVMask(ShuffleMask, VT, 64))
5548 return DAG.getNode(AArch64ISD::REV64, dl, V1.getValueType(), V1, V2);
5549 if (isREVMask(ShuffleMask, VT, 32))
5550 return DAG.getNode(AArch64ISD::REV32, dl, V1.getValueType(), V1, V2);
5551 if (isREVMask(ShuffleMask, VT, 16))
5552 return DAG.getNode(AArch64ISD::REV16, dl, V1.getValueType(), V1, V2);
5554 bool ReverseEXT = false;
5556 if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
5559 Imm *= getExtFactor(V1);
5560 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V2,
5561 DAG.getConstant(Imm, dl, MVT::i32));
5562 } else if (V2->getOpcode() == ISD::UNDEF &&
5563 isSingletonEXTMask(ShuffleMask, VT, Imm)) {
5564 Imm *= getExtFactor(V1);
5565 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V1,
5566 DAG.getConstant(Imm, dl, MVT::i32));
5569 unsigned WhichResult;
5570 if (isZIPMask(ShuffleMask, VT, WhichResult)) {
5571 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
5572 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5574 if (isUZPMask(ShuffleMask, VT, WhichResult)) {
5575 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
5576 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5578 if (isTRNMask(ShuffleMask, VT, WhichResult)) {
5579 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
5580 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5583 if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5584 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
5585 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5587 if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5588 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
5589 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5591 if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5592 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
5593 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5596 SDValue Concat = tryFormConcatFromShuffle(Op, DAG);
5597 if (Concat.getNode())
5602 int NumInputElements = V1.getValueType().getVectorNumElements();
5603 if (isINSMask(ShuffleMask, NumInputElements, DstIsLeft, Anomaly)) {
5604 SDValue DstVec = DstIsLeft ? V1 : V2;
5605 SDValue DstLaneV = DAG.getConstant(Anomaly, dl, MVT::i64);
5607 SDValue SrcVec = V1;
5608 int SrcLane = ShuffleMask[Anomaly];
5609 if (SrcLane >= NumInputElements) {
5611 SrcLane -= VT.getVectorNumElements();
5613 SDValue SrcLaneV = DAG.getConstant(SrcLane, dl, MVT::i64);
5615 EVT ScalarVT = VT.getVectorElementType();
5617 if (ScalarVT.getSizeInBits() < 32 && ScalarVT.isInteger())
5618 ScalarVT = MVT::i32;
5621 ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5622 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ScalarVT, SrcVec, SrcLaneV),
5626 // If the shuffle is not directly supported and it has 4 elements, use
5627 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5628 unsigned NumElts = VT.getVectorNumElements();
5630 unsigned PFIndexes[4];
5631 for (unsigned i = 0; i != 4; ++i) {
5632 if (ShuffleMask[i] < 0)
5635 PFIndexes[i] = ShuffleMask[i];
5638 // Compute the index in the perfect shuffle table.
5639 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
5640 PFIndexes[2] * 9 + PFIndexes[3];
5641 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5642 unsigned Cost = (PFEntry >> 30);
5645 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5648 return GenerateTBL(Op, ShuffleMask, DAG);
5651 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
5653 EVT VT = BVN->getValueType(0);
5654 APInt SplatBits, SplatUndef;
5655 unsigned SplatBitSize;
5657 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5658 unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;
5660 for (unsigned i = 0; i < NumSplats; ++i) {
5661 CnstBits <<= SplatBitSize;
5662 UndefBits <<= SplatBitSize;
5663 CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits());
5664 UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits());
5673 SDValue AArch64TargetLowering::LowerVectorAND(SDValue Op,
5674 SelectionDAG &DAG) const {
5675 BuildVectorSDNode *BVN =
5676 dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5677 SDValue LHS = Op.getOperand(0);
5679 EVT VT = Op.getValueType();
5684 APInt CnstBits(VT.getSizeInBits(), 0);
5685 APInt UndefBits(VT.getSizeInBits(), 0);
5686 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5687 // We only have BIC vector immediate instruction, which is and-not.
5688 CnstBits = ~CnstBits;
5690 // We make use of a little bit of goto ickiness in order to avoid having to
5691 // duplicate the immediate matching logic for the undef toggled case.
5692 bool SecondTry = false;
5695 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5696 CnstBits = CnstBits.zextOrTrunc(64);
5697 uint64_t CnstVal = CnstBits.getZExtValue();
5699 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5700 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5701 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5702 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5703 DAG.getConstant(CnstVal, dl, MVT::i32),
5704 DAG.getConstant(0, dl, MVT::i32));
5705 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5708 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5709 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5710 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5711 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5712 DAG.getConstant(CnstVal, dl, MVT::i32),
5713 DAG.getConstant(8, dl, MVT::i32));
5714 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5717 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5718 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5719 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5720 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5721 DAG.getConstant(CnstVal, dl, MVT::i32),
5722 DAG.getConstant(16, dl, MVT::i32));
5723 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5726 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5727 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5728 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5729 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5730 DAG.getConstant(CnstVal, dl, MVT::i32),
5731 DAG.getConstant(24, dl, MVT::i32));
5732 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5735 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5736 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5737 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5738 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5739 DAG.getConstant(CnstVal, dl, MVT::i32),
5740 DAG.getConstant(0, dl, MVT::i32));
5741 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5744 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5745 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5746 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5747 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5748 DAG.getConstant(CnstVal, dl, MVT::i32),
5749 DAG.getConstant(8, dl, MVT::i32));
5750 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5757 CnstBits = ~UndefBits;
5761 // We can always fall back to a non-immediate AND.
5766 // Specialized code to quickly find if PotentialBVec is a BuildVector that
5767 // consists of only the same constant int value, returned in reference arg
5769 static bool isAllConstantBuildVector(const SDValue &PotentialBVec,
5770 uint64_t &ConstVal) {
5771 BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec);
5774 ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0));
5777 EVT VT = Bvec->getValueType(0);
5778 unsigned NumElts = VT.getVectorNumElements();
5779 for (unsigned i = 1; i < NumElts; ++i)
5780 if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt)
5782 ConstVal = FirstElt->getZExtValue();
5786 static unsigned getIntrinsicID(const SDNode *N) {
5787 unsigned Opcode = N->getOpcode();
5790 return Intrinsic::not_intrinsic;
5791 case ISD::INTRINSIC_WO_CHAIN: {
5792 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
5793 if (IID < Intrinsic::num_intrinsics)
5795 return Intrinsic::not_intrinsic;
5800 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
5801 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
5802 // BUILD_VECTORs with constant element C1, C2 is a constant, and C1 == ~C2.
5803 // Also, logical shift right -> sri, with the same structure.
5804 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
5805 EVT VT = N->getValueType(0);
5812 // Is the first op an AND?
5813 const SDValue And = N->getOperand(0);
5814 if (And.getOpcode() != ISD::AND)
5817 // Is the second op an shl or lshr?
5818 SDValue Shift = N->getOperand(1);
5819 // This will have been turned into: AArch64ISD::VSHL vector, #shift
5820 // or AArch64ISD::VLSHR vector, #shift
5821 unsigned ShiftOpc = Shift.getOpcode();
5822 if ((ShiftOpc != AArch64ISD::VSHL && ShiftOpc != AArch64ISD::VLSHR))
5824 bool IsShiftRight = ShiftOpc == AArch64ISD::VLSHR;
5826 // Is the shift amount constant?
5827 ConstantSDNode *C2node = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
5831 // Is the and mask vector all constant?
5833 if (!isAllConstantBuildVector(And.getOperand(1), C1))
5836 // Is C1 == ~C2, taking into account how much one can shift elements of a
5838 uint64_t C2 = C2node->getZExtValue();
5839 unsigned ElemSizeInBits = VT.getVectorElementType().getSizeInBits();
5840 if (C2 > ElemSizeInBits)
5842 unsigned ElemMask = (1 << ElemSizeInBits) - 1;
5843 if ((C1 & ElemMask) != (~C2 & ElemMask))
5846 SDValue X = And.getOperand(0);
5847 SDValue Y = Shift.getOperand(0);
5850 IsShiftRight ? Intrinsic::aarch64_neon_vsri : Intrinsic::aarch64_neon_vsli;
5852 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5853 DAG.getConstant(Intrin, DL, MVT::i32), X, Y,
5854 Shift.getOperand(1));
5856 DEBUG(dbgs() << "aarch64-lower: transformed: \n");
5857 DEBUG(N->dump(&DAG));
5858 DEBUG(dbgs() << "into: \n");
5859 DEBUG(ResultSLI->dump(&DAG));
5865 SDValue AArch64TargetLowering::LowerVectorOR(SDValue Op,
5866 SelectionDAG &DAG) const {
5867 // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
5868 if (EnableAArch64SlrGeneration) {
5869 SDValue Res = tryLowerToSLI(Op.getNode(), DAG);
5874 BuildVectorSDNode *BVN =
5875 dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode());
5876 SDValue LHS = Op.getOperand(1);
5878 EVT VT = Op.getValueType();
5880 // OR commutes, so try swapping the operands.
5882 LHS = Op.getOperand(0);
5883 BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5888 APInt CnstBits(VT.getSizeInBits(), 0);
5889 APInt UndefBits(VT.getSizeInBits(), 0);
5890 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5891 // We make use of a little bit of goto ickiness in order to avoid having to
5892 // duplicate the immediate matching logic for the undef toggled case.
5893 bool SecondTry = false;
5896 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5897 CnstBits = CnstBits.zextOrTrunc(64);
5898 uint64_t CnstVal = CnstBits.getZExtValue();
5900 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5901 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5902 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5903 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5904 DAG.getConstant(CnstVal, dl, MVT::i32),
5905 DAG.getConstant(0, dl, MVT::i32));
5906 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5909 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5910 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5911 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5912 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5913 DAG.getConstant(CnstVal, dl, MVT::i32),
5914 DAG.getConstant(8, dl, MVT::i32));
5915 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5918 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5919 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5920 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5921 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5922 DAG.getConstant(CnstVal, dl, MVT::i32),
5923 DAG.getConstant(16, dl, MVT::i32));
5924 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5927 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5928 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5929 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5930 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5931 DAG.getConstant(CnstVal, dl, MVT::i32),
5932 DAG.getConstant(24, dl, MVT::i32));
5933 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5936 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5937 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5938 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5939 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5940 DAG.getConstant(CnstVal, dl, MVT::i32),
5941 DAG.getConstant(0, dl, MVT::i32));
5942 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5945 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5946 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5947 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5948 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5949 DAG.getConstant(CnstVal, dl, MVT::i32),
5950 DAG.getConstant(8, dl, MVT::i32));
5951 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5958 CnstBits = UndefBits;
5962 // We can always fall back to a non-immediate OR.
5967 // Normalize the operands of BUILD_VECTOR. The value of constant operands will
5968 // be truncated to fit element width.
5969 static SDValue NormalizeBuildVector(SDValue Op,
5970 SelectionDAG &DAG) {
5971 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
5973 EVT VT = Op.getValueType();
5974 EVT EltTy= VT.getVectorElementType();
5976 if (EltTy.isFloatingPoint() || EltTy.getSizeInBits() > 16)
5979 SmallVector<SDValue, 16> Ops;
5980 for (SDValue Lane : Op->ops()) {
5981 if (auto *CstLane = dyn_cast<ConstantSDNode>(Lane)) {
5982 APInt LowBits(EltTy.getSizeInBits(),
5983 CstLane->getZExtValue());
5984 Lane = DAG.getConstant(LowBits.getZExtValue(), dl, MVT::i32);
5986 Ops.push_back(Lane);
5988 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5991 SDValue AArch64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
5992 SelectionDAG &DAG) const {
5994 EVT VT = Op.getValueType();
5995 Op = NormalizeBuildVector(Op, DAG);
5996 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5998 APInt CnstBits(VT.getSizeInBits(), 0);
5999 APInt UndefBits(VT.getSizeInBits(), 0);
6000 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
6001 // We make use of a little bit of goto ickiness in order to avoid having to
6002 // duplicate the immediate matching logic for the undef toggled case.
6003 bool SecondTry = false;
6006 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
6007 CnstBits = CnstBits.zextOrTrunc(64);
6008 uint64_t CnstVal = CnstBits.getZExtValue();
6010 // Certain magic vector constants (used to express things like NOT
6011 // and NEG) are passed through unmodified. This allows codegen patterns
6012 // for these operations to match. Special-purpose patterns will lower
6013 // these immediates to MOVIs if it proves necessary.
6014 if (VT.isInteger() && (CnstVal == 0 || CnstVal == ~0ULL))
6017 // The many faces of MOVI...
6018 if (AArch64_AM::isAdvSIMDModImmType10(CnstVal)) {
6019 CnstVal = AArch64_AM::encodeAdvSIMDModImmType10(CnstVal);
6020 if (VT.getSizeInBits() == 128) {
6021 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::v2i64,
6022 DAG.getConstant(CnstVal, dl, MVT::i32));
6023 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6026 // Support the V64 version via subregister insertion.
6027 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::f64,
6028 DAG.getConstant(CnstVal, dl, MVT::i32));
6029 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6032 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
6033 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
6034 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6035 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
6036 DAG.getConstant(CnstVal, dl, MVT::i32),
6037 DAG.getConstant(0, dl, MVT::i32));
6038 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6041 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
6042 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
6043 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6044 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
6045 DAG.getConstant(CnstVal, dl, MVT::i32),
6046 DAG.getConstant(8, dl, MVT::i32));
6047 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6050 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
6051 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
6052 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6053 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
6054 DAG.getConstant(CnstVal, dl, MVT::i32),
6055 DAG.getConstant(16, dl, MVT::i32));
6056 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6059 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
6060 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
6061 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6062 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
6063 DAG.getConstant(CnstVal, dl, MVT::i32),
6064 DAG.getConstant(24, dl, MVT::i32));
6065 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6068 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
6069 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
6070 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
6071 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
6072 DAG.getConstant(CnstVal, dl, MVT::i32),
6073 DAG.getConstant(0, dl, MVT::i32));
6074 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6077 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
6078 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
6079 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
6080 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
6081 DAG.getConstant(CnstVal, dl, MVT::i32),
6082 DAG.getConstant(8, dl, MVT::i32));
6083 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6086 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
6087 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
6088 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6089 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
6090 DAG.getConstant(CnstVal, dl, MVT::i32),
6091 DAG.getConstant(264, dl, MVT::i32));
6092 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6095 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
6096 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
6097 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6098 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
6099 DAG.getConstant(CnstVal, dl, MVT::i32),
6100 DAG.getConstant(272, dl, MVT::i32));
6101 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6104 if (AArch64_AM::isAdvSIMDModImmType9(CnstVal)) {
6105 CnstVal = AArch64_AM::encodeAdvSIMDModImmType9(CnstVal);
6106 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;
6107 SDValue Mov = DAG.getNode(AArch64ISD::MOVI, dl, MovTy,
6108 DAG.getConstant(CnstVal, dl, MVT::i32));
6109 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6112 // The few faces of FMOV...
6113 if (AArch64_AM::isAdvSIMDModImmType11(CnstVal)) {
6114 CnstVal = AArch64_AM::encodeAdvSIMDModImmType11(CnstVal);
6115 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4f32 : MVT::v2f32;
6116 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MovTy,
6117 DAG.getConstant(CnstVal, dl, MVT::i32));
6118 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6121 if (AArch64_AM::isAdvSIMDModImmType12(CnstVal) &&
6122 VT.getSizeInBits() == 128) {
6123 CnstVal = AArch64_AM::encodeAdvSIMDModImmType12(CnstVal);
6124 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MVT::v2f64,
6125 DAG.getConstant(CnstVal, dl, MVT::i32));
6126 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6129 // The many faces of MVNI...
6131 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
6132 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
6133 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6134 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6135 DAG.getConstant(CnstVal, dl, MVT::i32),
6136 DAG.getConstant(0, dl, MVT::i32));
6137 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6140 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
6141 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
6142 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6143 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6144 DAG.getConstant(CnstVal, dl, MVT::i32),
6145 DAG.getConstant(8, dl, MVT::i32));
6146 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6149 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
6150 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
6151 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6152 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6153 DAG.getConstant(CnstVal, dl, MVT::i32),
6154 DAG.getConstant(16, dl, MVT::i32));
6155 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6158 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
6159 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
6160 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6161 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6162 DAG.getConstant(CnstVal, dl, MVT::i32),
6163 DAG.getConstant(24, dl, MVT::i32));
6164 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6167 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
6168 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
6169 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
6170 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6171 DAG.getConstant(CnstVal, dl, MVT::i32),
6172 DAG.getConstant(0, dl, MVT::i32));
6173 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6176 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
6177 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
6178 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
6179 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6180 DAG.getConstant(CnstVal, dl, MVT::i32),
6181 DAG.getConstant(8, dl, MVT::i32));
6182 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6185 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
6186 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
6187 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6188 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
6189 DAG.getConstant(CnstVal, dl, MVT::i32),
6190 DAG.getConstant(264, dl, MVT::i32));
6191 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6194 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
6195 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
6196 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6197 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
6198 DAG.getConstant(CnstVal, dl, MVT::i32),
6199 DAG.getConstant(272, dl, MVT::i32));
6200 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6207 CnstBits = UndefBits;
6212 // Scan through the operands to find some interesting properties we can
6214 // 1) If only one value is used, we can use a DUP, or
6215 // 2) if only the low element is not undef, we can just insert that, or
6216 // 3) if only one constant value is used (w/ some non-constant lanes),
6217 // we can splat the constant value into the whole vector then fill
6218 // in the non-constant lanes.
6219 // 4) FIXME: If different constant values are used, but we can intelligently
6220 // select the values we'll be overwriting for the non-constant
6221 // lanes such that we can directly materialize the vector
6222 // some other way (MOVI, e.g.), we can be sneaky.
6223 unsigned NumElts = VT.getVectorNumElements();
6224 bool isOnlyLowElement = true;
6225 bool usesOnlyOneValue = true;
6226 bool usesOnlyOneConstantValue = true;
6227 bool isConstant = true;
6228 unsigned NumConstantLanes = 0;
6230 SDValue ConstantValue;
6231 for (unsigned i = 0; i < NumElts; ++i) {
6232 SDValue V = Op.getOperand(i);
6233 if (V.getOpcode() == ISD::UNDEF)
6236 isOnlyLowElement = false;
6237 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
6240 if (isa<ConstantSDNode>(V) || isa<ConstantFPSDNode>(V)) {
6242 if (!ConstantValue.getNode())
6244 else if (ConstantValue != V)
6245 usesOnlyOneConstantValue = false;
6248 if (!Value.getNode())
6250 else if (V != Value)
6251 usesOnlyOneValue = false;
6254 if (!Value.getNode())
6255 return DAG.getUNDEF(VT);
6257 if (isOnlyLowElement)
6258 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
6260 // Use DUP for non-constant splats. For f32 constant splats, reduce to
6261 // i32 and try again.
6262 if (usesOnlyOneValue) {
6264 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6265 Value.getValueType() != VT)
6266 return DAG.getNode(AArch64ISD::DUP, dl, VT, Value);
6268 // This is actually a DUPLANExx operation, which keeps everything vectory.
6270 // DUPLANE works on 128-bit vectors, widen it if necessary.
6271 SDValue Lane = Value.getOperand(1);
6272 Value = Value.getOperand(0);
6273 if (Value.getValueType().getSizeInBits() == 64)
6274 Value = WidenVector(Value, DAG);
6276 unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
6277 return DAG.getNode(Opcode, dl, VT, Value, Lane);
6280 if (VT.getVectorElementType().isFloatingPoint()) {
6281 SmallVector<SDValue, 8> Ops;
6282 EVT EltTy = VT.getVectorElementType();
6283 assert ((EltTy == MVT::f16 || EltTy == MVT::f32 || EltTy == MVT::f64) &&
6284 "Unsupported floating-point vector type");
6285 MVT NewType = MVT::getIntegerVT(EltTy.getSizeInBits());
6286 for (unsigned i = 0; i < NumElts; ++i)
6287 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, NewType, Op.getOperand(i)));
6288 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts);
6289 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
6290 Val = LowerBUILD_VECTOR(Val, DAG);
6292 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
6296 // If there was only one constant value used and for more than one lane,
6297 // start by splatting that value, then replace the non-constant lanes. This
6298 // is better than the default, which will perform a separate initialization
6300 if (NumConstantLanes > 0 && usesOnlyOneConstantValue) {
6301 SDValue Val = DAG.getNode(AArch64ISD::DUP, dl, VT, ConstantValue);
6302 // Now insert the non-constant lanes.
6303 for (unsigned i = 0; i < NumElts; ++i) {
6304 SDValue V = Op.getOperand(i);
6305 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i64);
6306 if (!isa<ConstantSDNode>(V) && !isa<ConstantFPSDNode>(V)) {
6307 // Note that type legalization likely mucked about with the VT of the
6308 // source operand, so we may have to convert it here before inserting.
6309 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, V, LaneIdx);
6315 // If all elements are constants and the case above didn't get hit, fall back
6316 // to the default expansion, which will generate a load from the constant
6321 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
6323 if (SDValue shuffle = ReconstructShuffle(Op, DAG))
6327 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
6328 // know the default expansion would otherwise fall back on something even
6329 // worse. For a vector with one or two non-undef values, that's
6330 // scalar_to_vector for the elements followed by a shuffle (provided the
6331 // shuffle is valid for the target) and materialization element by element
6332 // on the stack followed by a load for everything else.
6333 if (!isConstant && !usesOnlyOneValue) {
6334 SDValue Vec = DAG.getUNDEF(VT);
6335 SDValue Op0 = Op.getOperand(0);
6336 unsigned ElemSize = VT.getVectorElementType().getSizeInBits();
6338 // For 32 and 64 bit types, use INSERT_SUBREG for lane zero to
6339 // a) Avoid a RMW dependency on the full vector register, and
6340 // b) Allow the register coalescer to fold away the copy if the
6341 // value is already in an S or D register.
6342 // Do not do this for UNDEF/LOAD nodes because we have better patterns
6343 // for those avoiding the SCALAR_TO_VECTOR/BUILD_VECTOR.
6344 if (Op0.getOpcode() != ISD::UNDEF && Op0.getOpcode() != ISD::LOAD &&
6345 (ElemSize == 32 || ElemSize == 64)) {
6346 unsigned SubIdx = ElemSize == 32 ? AArch64::ssub : AArch64::dsub;
6348 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0,
6349 DAG.getTargetConstant(SubIdx, dl, MVT::i32));
6350 Vec = SDValue(N, 0);
6353 for (; i < NumElts; ++i) {
6354 SDValue V = Op.getOperand(i);
6355 if (V.getOpcode() == ISD::UNDEF)
6357 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i64);
6358 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
6363 // Just use the default expansion. We failed to find a better alternative.
6367 SDValue AArch64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
6368 SelectionDAG &DAG) const {
6369 assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!");
6371 // Check for non-constant or out of range lane.
6372 EVT VT = Op.getOperand(0).getValueType();
6373 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(2));
6374 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
6378 // Insertion/extraction are legal for V128 types.
6379 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
6380 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
6384 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
6385 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
6388 // For V64 types, we perform insertion by expanding the value
6389 // to a V128 type and perform the insertion on that.
6391 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
6392 EVT WideTy = WideVec.getValueType();
6394 SDValue Node = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideTy, WideVec,
6395 Op.getOperand(1), Op.getOperand(2));
6396 // Re-narrow the resultant vector.
6397 return NarrowVector(Node, DAG);
6401 AArch64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
6402 SelectionDAG &DAG) const {
6403 assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!");
6405 // Check for non-constant or out of range lane.
6406 EVT VT = Op.getOperand(0).getValueType();
6407 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6408 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
6412 // Insertion/extraction are legal for V128 types.
6413 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
6414 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
6418 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
6419 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
6422 // For V64 types, we perform extraction by expanding the value
6423 // to a V128 type and perform the extraction on that.
6425 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
6426 EVT WideTy = WideVec.getValueType();
6428 EVT ExtrTy = WideTy.getVectorElementType();
6429 if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8)
6432 // For extractions, we just return the result directly.
6433 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec,
6437 SDValue AArch64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
6438 SelectionDAG &DAG) const {
6439 EVT VT = Op.getOperand(0).getValueType();
6445 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6448 unsigned Val = Cst->getZExtValue();
6450 unsigned Size = Op.getValueType().getSizeInBits();
6452 // This will get lowered to an appropriate EXTRACT_SUBREG in ISel.
6456 // If this is extracting the upper 64-bits of a 128-bit vector, we match
6458 if (Size == 64 && Val * VT.getVectorElementType().getSizeInBits() == 64)
6464 bool AArch64TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
6466 if (VT.getVectorNumElements() == 4 &&
6467 (VT.is128BitVector() || VT.is64BitVector())) {
6468 unsigned PFIndexes[4];
6469 for (unsigned i = 0; i != 4; ++i) {
6473 PFIndexes[i] = M[i];
6476 // Compute the index in the perfect shuffle table.
6477 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
6478 PFIndexes[2] * 9 + PFIndexes[3];
6479 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
6480 unsigned Cost = (PFEntry >> 30);
6488 unsigned DummyUnsigned;
6490 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || isREVMask(M, VT, 64) ||
6491 isREVMask(M, VT, 32) || isREVMask(M, VT, 16) ||
6492 isEXTMask(M, VT, DummyBool, DummyUnsigned) ||
6493 // isTBLMask(M, VT) || // FIXME: Port TBL support from ARM.
6494 isTRNMask(M, VT, DummyUnsigned) || isUZPMask(M, VT, DummyUnsigned) ||
6495 isZIPMask(M, VT, DummyUnsigned) ||
6496 isTRN_v_undef_Mask(M, VT, DummyUnsigned) ||
6497 isUZP_v_undef_Mask(M, VT, DummyUnsigned) ||
6498 isZIP_v_undef_Mask(M, VT, DummyUnsigned) ||
6499 isINSMask(M, VT.getVectorNumElements(), DummyBool, DummyInt) ||
6500 isConcatMask(M, VT, VT.getSizeInBits() == 128));
6503 /// getVShiftImm - Check if this is a valid build_vector for the immediate
6504 /// operand of a vector shift operation, where all the elements of the
6505 /// build_vector must have the same constant integer value.
6506 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
6507 // Ignore bit_converts.
6508 while (Op.getOpcode() == ISD::BITCAST)
6509 Op = Op.getOperand(0);
6510 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
6511 APInt SplatBits, SplatUndef;
6512 unsigned SplatBitSize;
6514 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
6515 HasAnyUndefs, ElementBits) ||
6516 SplatBitSize > ElementBits)
6518 Cnt = SplatBits.getSExtValue();
6522 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
6523 /// operand of a vector shift left operation. That value must be in the range:
6524 /// 0 <= Value < ElementBits for a left shift; or
6525 /// 0 <= Value <= ElementBits for a long left shift.
6526 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
6527 assert(VT.isVector() && "vector shift count is not a vector type");
6528 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
6529 if (!getVShiftImm(Op, ElementBits, Cnt))
6531 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
6534 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
6535 /// operand of a vector shift right operation. The value must be in the range:
6536 /// 1 <= Value <= ElementBits for a right shift; or
6537 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, int64_t &Cnt) {
6538 assert(VT.isVector() && "vector shift count is not a vector type");
6539 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
6540 if (!getVShiftImm(Op, ElementBits, Cnt))
6542 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
6545 SDValue AArch64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
6546 SelectionDAG &DAG) const {
6547 EVT VT = Op.getValueType();
6551 if (!Op.getOperand(1).getValueType().isVector())
6553 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
6555 switch (Op.getOpcode()) {
6557 llvm_unreachable("unexpected shift opcode");
6560 if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
6561 return DAG.getNode(AArch64ISD::VSHL, DL, VT, Op.getOperand(0),
6562 DAG.getConstant(Cnt, DL, MVT::i32));
6563 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
6564 DAG.getConstant(Intrinsic::aarch64_neon_ushl, DL,
6566 Op.getOperand(0), Op.getOperand(1));
6569 // Right shift immediate
6570 if (isVShiftRImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize) {
6572 (Op.getOpcode() == ISD::SRA) ? AArch64ISD::VASHR : AArch64ISD::VLSHR;
6573 return DAG.getNode(Opc, DL, VT, Op.getOperand(0),
6574 DAG.getConstant(Cnt, DL, MVT::i32));
6577 // Right shift register. Note, there is not a shift right register
6578 // instruction, but the shift left register instruction takes a signed
6579 // value, where negative numbers specify a right shift.
6580 unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::aarch64_neon_sshl
6581 : Intrinsic::aarch64_neon_ushl;
6582 // negate the shift amount
6583 SDValue NegShift = DAG.getNode(AArch64ISD::NEG, DL, VT, Op.getOperand(1));
6584 SDValue NegShiftLeft =
6585 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
6586 DAG.getConstant(Opc, DL, MVT::i32), Op.getOperand(0),
6588 return NegShiftLeft;
6594 static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS,
6595 AArch64CC::CondCode CC, bool NoNans, EVT VT,
6596 SDLoc dl, SelectionDAG &DAG) {
6597 EVT SrcVT = LHS.getValueType();
6598 assert(VT.getSizeInBits() == SrcVT.getSizeInBits() &&
6599 "function only supposed to emit natural comparisons");
6601 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
6602 APInt CnstBits(VT.getSizeInBits(), 0);
6603 APInt UndefBits(VT.getSizeInBits(), 0);
6604 bool IsCnst = BVN && resolveBuildVector(BVN, CnstBits, UndefBits);
6605 bool IsZero = IsCnst && (CnstBits == 0);
6607 if (SrcVT.getVectorElementType().isFloatingPoint()) {
6611 case AArch64CC::NE: {
6614 Fcmeq = DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
6616 Fcmeq = DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
6617 return DAG.getNode(AArch64ISD::NOT, dl, VT, Fcmeq);
6621 return DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
6622 return DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
6625 return DAG.getNode(AArch64ISD::FCMGEz, dl, VT, LHS);
6626 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, LHS, RHS);
6629 return DAG.getNode(AArch64ISD::FCMGTz, dl, VT, LHS);
6630 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, LHS, RHS);
6633 return DAG.getNode(AArch64ISD::FCMLEz, dl, VT, LHS);
6634 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, RHS, LHS);
6638 // If we ignore NaNs then we can use to the MI implementation.
6642 return DAG.getNode(AArch64ISD::FCMLTz, dl, VT, LHS);
6643 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, RHS, LHS);
6650 case AArch64CC::NE: {
6653 Cmeq = DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
6655 Cmeq = DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
6656 return DAG.getNode(AArch64ISD::NOT, dl, VT, Cmeq);
6660 return DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
6661 return DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
6664 return DAG.getNode(AArch64ISD::CMGEz, dl, VT, LHS);
6665 return DAG.getNode(AArch64ISD::CMGE, dl, VT, LHS, RHS);
6668 return DAG.getNode(AArch64ISD::CMGTz, dl, VT, LHS);
6669 return DAG.getNode(AArch64ISD::CMGT, dl, VT, LHS, RHS);
6672 return DAG.getNode(AArch64ISD::CMLEz, dl, VT, LHS);
6673 return DAG.getNode(AArch64ISD::CMGE, dl, VT, RHS, LHS);
6675 return DAG.getNode(AArch64ISD::CMHS, dl, VT, RHS, LHS);
6677 return DAG.getNode(AArch64ISD::CMHI, dl, VT, RHS, LHS);
6680 return DAG.getNode(AArch64ISD::CMLTz, dl, VT, LHS);
6681 return DAG.getNode(AArch64ISD::CMGT, dl, VT, RHS, LHS);
6683 return DAG.getNode(AArch64ISD::CMHI, dl, VT, LHS, RHS);
6685 return DAG.getNode(AArch64ISD::CMHS, dl, VT, LHS, RHS);
6689 SDValue AArch64TargetLowering::LowerVSETCC(SDValue Op,
6690 SelectionDAG &DAG) const {
6691 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6692 SDValue LHS = Op.getOperand(0);
6693 SDValue RHS = Op.getOperand(1);
6694 EVT CmpVT = LHS.getValueType().changeVectorElementTypeToInteger();
6697 if (LHS.getValueType().getVectorElementType().isInteger()) {
6698 assert(LHS.getValueType() == RHS.getValueType());
6699 AArch64CC::CondCode AArch64CC = changeIntCCToAArch64CC(CC);
6701 EmitVectorComparison(LHS, RHS, AArch64CC, false, CmpVT, dl, DAG);
6702 return DAG.getSExtOrTrunc(Cmp, dl, Op.getValueType());
6705 if (LHS.getValueType().getVectorElementType() == MVT::f16)
6708 assert(LHS.getValueType().getVectorElementType() == MVT::f32 ||
6709 LHS.getValueType().getVectorElementType() == MVT::f64);
6711 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
6712 // clean. Some of them require two branches to implement.
6713 AArch64CC::CondCode CC1, CC2;
6715 changeVectorFPCCToAArch64CC(CC, CC1, CC2, ShouldInvert);
6717 bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath;
6719 EmitVectorComparison(LHS, RHS, CC1, NoNaNs, CmpVT, dl, DAG);
6723 if (CC2 != AArch64CC::AL) {
6725 EmitVectorComparison(LHS, RHS, CC2, NoNaNs, CmpVT, dl, DAG);
6726 if (!Cmp2.getNode())
6729 Cmp = DAG.getNode(ISD::OR, dl, CmpVT, Cmp, Cmp2);
6732 Cmp = DAG.getSExtOrTrunc(Cmp, dl, Op.getValueType());
6735 return Cmp = DAG.getNOT(dl, Cmp, Cmp.getValueType());
6740 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
6741 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
6742 /// specified in the intrinsic calls.
6743 bool AArch64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
6745 unsigned Intrinsic) const {
6746 auto &DL = I.getModule()->getDataLayout();
6747 switch (Intrinsic) {
6748 case Intrinsic::aarch64_neon_ld2:
6749 case Intrinsic::aarch64_neon_ld3:
6750 case Intrinsic::aarch64_neon_ld4:
6751 case Intrinsic::aarch64_neon_ld1x2:
6752 case Intrinsic::aarch64_neon_ld1x3:
6753 case Intrinsic::aarch64_neon_ld1x4:
6754 case Intrinsic::aarch64_neon_ld2lane:
6755 case Intrinsic::aarch64_neon_ld3lane:
6756 case Intrinsic::aarch64_neon_ld4lane:
6757 case Intrinsic::aarch64_neon_ld2r:
6758 case Intrinsic::aarch64_neon_ld3r:
6759 case Intrinsic::aarch64_neon_ld4r: {
6760 Info.opc = ISD::INTRINSIC_W_CHAIN;
6761 // Conservatively set memVT to the entire set of vectors loaded.
6762 uint64_t NumElts = DL.getTypeSizeInBits(I.getType()) / 64;
6763 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6764 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6767 Info.vol = false; // volatile loads with NEON intrinsics not supported
6768 Info.readMem = true;
6769 Info.writeMem = false;
6772 case Intrinsic::aarch64_neon_st2:
6773 case Intrinsic::aarch64_neon_st3:
6774 case Intrinsic::aarch64_neon_st4:
6775 case Intrinsic::aarch64_neon_st1x2:
6776 case Intrinsic::aarch64_neon_st1x3:
6777 case Intrinsic::aarch64_neon_st1x4:
6778 case Intrinsic::aarch64_neon_st2lane:
6779 case Intrinsic::aarch64_neon_st3lane:
6780 case Intrinsic::aarch64_neon_st4lane: {
6781 Info.opc = ISD::INTRINSIC_VOID;
6782 // Conservatively set memVT to the entire set of vectors stored.
6783 unsigned NumElts = 0;
6784 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
6785 Type *ArgTy = I.getArgOperand(ArgI)->getType();
6786 if (!ArgTy->isVectorTy())
6788 NumElts += DL.getTypeSizeInBits(ArgTy) / 64;
6790 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6791 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6794 Info.vol = false; // volatile stores with NEON intrinsics not supported
6795 Info.readMem = false;
6796 Info.writeMem = true;
6799 case Intrinsic::aarch64_ldaxr:
6800 case Intrinsic::aarch64_ldxr: {
6801 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
6802 Info.opc = ISD::INTRINSIC_W_CHAIN;
6803 Info.memVT = MVT::getVT(PtrTy->getElementType());
6804 Info.ptrVal = I.getArgOperand(0);
6806 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
6808 Info.readMem = true;
6809 Info.writeMem = false;
6812 case Intrinsic::aarch64_stlxr:
6813 case Intrinsic::aarch64_stxr: {
6814 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
6815 Info.opc = ISD::INTRINSIC_W_CHAIN;
6816 Info.memVT = MVT::getVT(PtrTy->getElementType());
6817 Info.ptrVal = I.getArgOperand(1);
6819 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
6821 Info.readMem = false;
6822 Info.writeMem = true;
6825 case Intrinsic::aarch64_ldaxp:
6826 case Intrinsic::aarch64_ldxp: {
6827 Info.opc = ISD::INTRINSIC_W_CHAIN;
6828 Info.memVT = MVT::i128;
6829 Info.ptrVal = I.getArgOperand(0);
6833 Info.readMem = true;
6834 Info.writeMem = false;
6837 case Intrinsic::aarch64_stlxp:
6838 case Intrinsic::aarch64_stxp: {
6839 Info.opc = ISD::INTRINSIC_W_CHAIN;
6840 Info.memVT = MVT::i128;
6841 Info.ptrVal = I.getArgOperand(2);
6845 Info.readMem = false;
6846 Info.writeMem = true;
6856 // Truncations from 64-bit GPR to 32-bit GPR is free.
6857 bool AArch64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
6858 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6860 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6861 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6862 return NumBits1 > NumBits2;
6864 bool AArch64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
6865 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6867 unsigned NumBits1 = VT1.getSizeInBits();
6868 unsigned NumBits2 = VT2.getSizeInBits();
6869 return NumBits1 > NumBits2;
6872 /// Check if it is profitable to hoist instruction in then/else to if.
6873 /// Not profitable if I and it's user can form a FMA instruction
6874 /// because we prefer FMSUB/FMADD.
6875 bool AArch64TargetLowering::isProfitableToHoist(Instruction *I) const {
6876 if (I->getOpcode() != Instruction::FMul)
6879 if (I->getNumUses() != 1)
6882 Instruction *User = I->user_back();
6885 !(User->getOpcode() == Instruction::FSub ||
6886 User->getOpcode() == Instruction::FAdd))
6889 const TargetOptions &Options = getTargetMachine().Options;
6890 const DataLayout &DL = I->getModule()->getDataLayout();
6891 EVT VT = getValueType(DL, User->getOperand(0)->getType());
6893 if (isFMAFasterThanFMulAndFAdd(VT) &&
6894 isOperationLegalOrCustom(ISD::FMA, VT) &&
6895 (Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath))
6901 // All 32-bit GPR operations implicitly zero the high-half of the corresponding
6903 bool AArch64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
6904 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6906 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6907 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6908 return NumBits1 == 32 && NumBits2 == 64;
6910 bool AArch64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
6911 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6913 unsigned NumBits1 = VT1.getSizeInBits();
6914 unsigned NumBits2 = VT2.getSizeInBits();
6915 return NumBits1 == 32 && NumBits2 == 64;
6918 bool AArch64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
6919 EVT VT1 = Val.getValueType();
6920 if (isZExtFree(VT1, VT2)) {
6924 if (Val.getOpcode() != ISD::LOAD)
6927 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
6928 return (VT1.isSimple() && !VT1.isVector() && VT1.isInteger() &&
6929 VT2.isSimple() && !VT2.isVector() && VT2.isInteger() &&
6930 VT1.getSizeInBits() <= 32);
6933 bool AArch64TargetLowering::isExtFreeImpl(const Instruction *Ext) const {
6934 if (isa<FPExtInst>(Ext))
6937 // Vector types are next free.
6938 if (Ext->getType()->isVectorTy())
6941 for (const Use &U : Ext->uses()) {
6942 // The extension is free if we can fold it with a left shift in an
6943 // addressing mode or an arithmetic operation: add, sub, and cmp.
6945 // Is there a shift?
6946 const Instruction *Instr = cast<Instruction>(U.getUser());
6948 // Is this a constant shift?
6949 switch (Instr->getOpcode()) {
6950 case Instruction::Shl:
6951 if (!isa<ConstantInt>(Instr->getOperand(1)))
6954 case Instruction::GetElementPtr: {
6955 gep_type_iterator GTI = gep_type_begin(Instr);
6956 auto &DL = Ext->getModule()->getDataLayout();
6957 std::advance(GTI, U.getOperandNo());
6959 // This extension will end up with a shift because of the scaling factor.
6960 // 8-bit sized types have a scaling factor of 1, thus a shift amount of 0.
6961 // Get the shift amount based on the scaling factor:
6962 // log2(sizeof(IdxTy)) - log2(8).
6964 countTrailingZeros(DL.getTypeStoreSizeInBits(IdxTy)) - 3;
6965 // Is the constant foldable in the shift of the addressing mode?
6966 // I.e., shift amount is between 1 and 4 inclusive.
6967 if (ShiftAmt == 0 || ShiftAmt > 4)
6971 case Instruction::Trunc:
6972 // Check if this is a noop.
6973 // trunc(sext ty1 to ty2) to ty1.
6974 if (Instr->getType() == Ext->getOperand(0)->getType())
6981 // At this point we can use the bfm family, so this extension is free
6987 bool AArch64TargetLowering::hasPairedLoad(Type *LoadedType,
6988 unsigned &RequiredAligment) const {
6989 if (!LoadedType->isIntegerTy() && !LoadedType->isFloatTy())
6991 // Cyclone supports unaligned accesses.
6992 RequiredAligment = 0;
6993 unsigned NumBits = LoadedType->getPrimitiveSizeInBits();
6994 return NumBits == 32 || NumBits == 64;
6997 bool AArch64TargetLowering::hasPairedLoad(EVT LoadedType,
6998 unsigned &RequiredAligment) const {
6999 if (!LoadedType.isSimple() ||
7000 (!LoadedType.isInteger() && !LoadedType.isFloatingPoint()))
7002 // Cyclone supports unaligned accesses.
7003 RequiredAligment = 0;
7004 unsigned NumBits = LoadedType.getSizeInBits();
7005 return NumBits == 32 || NumBits == 64;
7008 /// \brief Lower an interleaved load into a ldN intrinsic.
7010 /// E.g. Lower an interleaved load (Factor = 2):
7011 /// %wide.vec = load <8 x i32>, <8 x i32>* %ptr
7012 /// %v0 = shuffle %wide.vec, undef, <0, 2, 4, 6> ; Extract even elements
7013 /// %v1 = shuffle %wide.vec, undef, <1, 3, 5, 7> ; Extract odd elements
7016 /// %ld2 = { <4 x i32>, <4 x i32> } call llvm.aarch64.neon.ld2(%ptr)
7017 /// %vec0 = extractelement { <4 x i32>, <4 x i32> } %ld2, i32 0
7018 /// %vec1 = extractelement { <4 x i32>, <4 x i32> } %ld2, i32 1
7019 bool AArch64TargetLowering::lowerInterleavedLoad(
7020 LoadInst *LI, ArrayRef<ShuffleVectorInst *> Shuffles,
7021 ArrayRef<unsigned> Indices, unsigned Factor) const {
7022 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
7023 "Invalid interleave factor");
7024 assert(!Shuffles.empty() && "Empty shufflevector input");
7025 assert(Shuffles.size() == Indices.size() &&
7026 "Unmatched number of shufflevectors and indices");
7028 const DataLayout &DL = LI->getModule()->getDataLayout();
7030 VectorType *VecTy = Shuffles[0]->getType();
7031 unsigned VecSize = DL.getTypeSizeInBits(VecTy);
7033 // Skip if we do not have NEON and skip illegal vector types.
7034 if (!Subtarget->hasNEON() || (VecSize != 64 && VecSize != 128))
7037 // A pointer vector can not be the return type of the ldN intrinsics. Need to
7038 // load integer vectors first and then convert to pointer vectors.
7039 Type *EltTy = VecTy->getVectorElementType();
7040 if (EltTy->isPointerTy())
7042 VectorType::get(DL.getIntPtrType(EltTy), VecTy->getVectorNumElements());
7044 Type *PtrTy = VecTy->getPointerTo(LI->getPointerAddressSpace());
7045 Type *Tys[2] = {VecTy, PtrTy};
7046 static const Intrinsic::ID LoadInts[3] = {Intrinsic::aarch64_neon_ld2,
7047 Intrinsic::aarch64_neon_ld3,
7048 Intrinsic::aarch64_neon_ld4};
7050 Intrinsic::getDeclaration(LI->getModule(), LoadInts[Factor - 2], Tys);
7052 IRBuilder<> Builder(LI);
7053 Value *Ptr = Builder.CreateBitCast(LI->getPointerOperand(), PtrTy);
7055 CallInst *LdN = Builder.CreateCall(LdNFunc, Ptr, "ldN");
7057 // Replace uses of each shufflevector with the corresponding vector loaded
7059 for (unsigned i = 0; i < Shuffles.size(); i++) {
7060 ShuffleVectorInst *SVI = Shuffles[i];
7061 unsigned Index = Indices[i];
7063 Value *SubVec = Builder.CreateExtractValue(LdN, Index);
7065 // Convert the integer vector to pointer vector if the element is pointer.
7066 if (EltTy->isPointerTy())
7067 SubVec = Builder.CreateIntToPtr(SubVec, SVI->getType());
7069 SVI->replaceAllUsesWith(SubVec);
7075 /// \brief Get a mask consisting of sequential integers starting from \p Start.
7077 /// I.e. <Start, Start + 1, ..., Start + NumElts - 1>
7078 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned Start,
7080 SmallVector<Constant *, 16> Mask;
7081 for (unsigned i = 0; i < NumElts; i++)
7082 Mask.push_back(Builder.getInt32(Start + i));
7084 return ConstantVector::get(Mask);
7087 /// \brief Lower an interleaved store into a stN intrinsic.
7089 /// E.g. Lower an interleaved store (Factor = 3):
7090 /// %i.vec = shuffle <8 x i32> %v0, <8 x i32> %v1,
7091 /// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11>
7092 /// store <12 x i32> %i.vec, <12 x i32>* %ptr
7095 /// %sub.v0 = shuffle <8 x i32> %v0, <8 x i32> v1, <0, 1, 2, 3>
7096 /// %sub.v1 = shuffle <8 x i32> %v0, <8 x i32> v1, <4, 5, 6, 7>
7097 /// %sub.v2 = shuffle <8 x i32> %v0, <8 x i32> v1, <8, 9, 10, 11>
7098 /// call void llvm.aarch64.neon.st3(%sub.v0, %sub.v1, %sub.v2, %ptr)
7100 /// Note that the new shufflevectors will be removed and we'll only generate one
7101 /// st3 instruction in CodeGen.
7102 bool AArch64TargetLowering::lowerInterleavedStore(StoreInst *SI,
7103 ShuffleVectorInst *SVI,
7104 unsigned Factor) const {
7105 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
7106 "Invalid interleave factor");
7108 VectorType *VecTy = SVI->getType();
7109 assert(VecTy->getVectorNumElements() % Factor == 0 &&
7110 "Invalid interleaved store");
7112 unsigned NumSubElts = VecTy->getVectorNumElements() / Factor;
7113 Type *EltTy = VecTy->getVectorElementType();
7114 VectorType *SubVecTy = VectorType::get(EltTy, NumSubElts);
7116 const DataLayout &DL = SI->getModule()->getDataLayout();
7117 unsigned SubVecSize = DL.getTypeSizeInBits(SubVecTy);
7119 // Skip if we do not have NEON and skip illegal vector types.
7120 if (!Subtarget->hasNEON() || (SubVecSize != 64 && SubVecSize != 128))
7123 Value *Op0 = SVI->getOperand(0);
7124 Value *Op1 = SVI->getOperand(1);
7125 IRBuilder<> Builder(SI);
7127 // StN intrinsics don't support pointer vectors as arguments. Convert pointer
7128 // vectors to integer vectors.
7129 if (EltTy->isPointerTy()) {
7130 Type *IntTy = DL.getIntPtrType(EltTy);
7131 unsigned NumOpElts =
7132 dyn_cast<VectorType>(Op0->getType())->getVectorNumElements();
7134 // Convert to the corresponding integer vector.
7135 Type *IntVecTy = VectorType::get(IntTy, NumOpElts);
7136 Op0 = Builder.CreatePtrToInt(Op0, IntVecTy);
7137 Op1 = Builder.CreatePtrToInt(Op1, IntVecTy);
7139 SubVecTy = VectorType::get(IntTy, NumSubElts);
7142 Type *PtrTy = SubVecTy->getPointerTo(SI->getPointerAddressSpace());
7143 Type *Tys[2] = {SubVecTy, PtrTy};
7144 static const Intrinsic::ID StoreInts[3] = {Intrinsic::aarch64_neon_st2,
7145 Intrinsic::aarch64_neon_st3,
7146 Intrinsic::aarch64_neon_st4};
7148 Intrinsic::getDeclaration(SI->getModule(), StoreInts[Factor - 2], Tys);
7150 SmallVector<Value *, 5> Ops;
7152 // Split the shufflevector operands into sub vectors for the new stN call.
7153 for (unsigned i = 0; i < Factor; i++)
7154 Ops.push_back(Builder.CreateShuffleVector(
7155 Op0, Op1, getSequentialMask(Builder, NumSubElts * i, NumSubElts)));
7157 Ops.push_back(Builder.CreateBitCast(SI->getPointerOperand(), PtrTy));
7158 Builder.CreateCall(StNFunc, Ops);
7162 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
7163 unsigned AlignCheck) {
7164 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
7165 (DstAlign == 0 || DstAlign % AlignCheck == 0));
7168 EVT AArch64TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
7169 unsigned SrcAlign, bool IsMemset,
7172 MachineFunction &MF) const {
7173 // Don't use AdvSIMD to implement 16-byte memset. It would have taken one
7174 // instruction to materialize the v2i64 zero and one store (with restrictive
7175 // addressing mode). Just do two i64 store of zero-registers.
7177 const Function *F = MF.getFunction();
7178 if (Subtarget->hasFPARMv8() && !IsMemset && Size >= 16 &&
7179 !F->hasFnAttribute(Attribute::NoImplicitFloat) &&
7180 (memOpAlign(SrcAlign, DstAlign, 16) ||
7181 (allowsMisalignedMemoryAccesses(MVT::f128, 0, 1, &Fast) && Fast)))
7185 (memOpAlign(SrcAlign, DstAlign, 8) ||
7186 (allowsMisalignedMemoryAccesses(MVT::i64, 0, 1, &Fast) && Fast)))
7190 (memOpAlign(SrcAlign, DstAlign, 4) ||
7191 (allowsMisalignedMemoryAccesses(MVT::i32, 0, 1, &Fast) && Fast)))
7197 // 12-bit optionally shifted immediates are legal for adds.
7198 bool AArch64TargetLowering::isLegalAddImmediate(int64_t Immed) const {
7199 if ((Immed >> 12) == 0 || ((Immed & 0xfff) == 0 && Immed >> 24 == 0))
7204 // Integer comparisons are implemented with ADDS/SUBS, so the range of valid
7205 // immediates is the same as for an add or a sub.
7206 bool AArch64TargetLowering::isLegalICmpImmediate(int64_t Immed) const {
7209 return isLegalAddImmediate(Immed);
7212 /// isLegalAddressingMode - Return true if the addressing mode represented
7213 /// by AM is legal for this target, for a load/store of the specified type.
7214 bool AArch64TargetLowering::isLegalAddressingMode(const DataLayout &DL,
7215 const AddrMode &AM, Type *Ty,
7216 unsigned AS) const {
7217 // AArch64 has five basic addressing modes:
7219 // reg + 9-bit signed offset
7220 // reg + SIZE_IN_BYTES * 12-bit unsigned offset
7222 // reg + SIZE_IN_BYTES * reg
7224 // No global is ever allowed as a base.
7228 // No reg+reg+imm addressing.
7229 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
7232 // check reg + imm case:
7233 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
7234 uint64_t NumBytes = 0;
7235 if (Ty->isSized()) {
7236 uint64_t NumBits = DL.getTypeSizeInBits(Ty);
7237 NumBytes = NumBits / 8;
7238 if (!isPowerOf2_64(NumBits))
7243 int64_t Offset = AM.BaseOffs;
7245 // 9-bit signed offset
7246 if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
7249 // 12-bit unsigned offset
7250 unsigned shift = Log2_64(NumBytes);
7251 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
7252 // Must be a multiple of NumBytes (NumBytes is a power of 2)
7253 (Offset >> shift) << shift == Offset)
7258 // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2
7260 if (!AM.Scale || AM.Scale == 1 ||
7261 (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
7266 int AArch64TargetLowering::getScalingFactorCost(const DataLayout &DL,
7267 const AddrMode &AM, Type *Ty,
7268 unsigned AS) const {
7269 // Scaling factors are not free at all.
7270 // Operands | Rt Latency
7271 // -------------------------------------------
7273 // -------------------------------------------
7274 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
7275 // Rt, [Xn, Wm, <extend> #imm] |
7276 if (isLegalAddressingMode(DL, AM, Ty, AS))
7277 // Scale represents reg2 * scale, thus account for 1 if
7278 // it is not equal to 0 or 1.
7279 return AM.Scale != 0 && AM.Scale != 1;
7283 bool AArch64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
7284 VT = VT.getScalarType();
7289 switch (VT.getSimpleVT().SimpleTy) {
7301 AArch64TargetLowering::getScratchRegisters(CallingConv::ID) const {
7302 // LR is a callee-save register, but we must treat it as clobbered by any call
7303 // site. Hence we include LR in the scratch registers, which are in turn added
7304 // as implicit-defs for stackmaps and patchpoints.
7305 static const MCPhysReg ScratchRegs[] = {
7306 AArch64::X16, AArch64::X17, AArch64::LR, 0
7312 AArch64TargetLowering::isDesirableToCommuteWithShift(const SDNode *N) const {
7313 EVT VT = N->getValueType(0);
7314 // If N is unsigned bit extraction: ((x >> C) & mask), then do not combine
7315 // it with shift to let it be lowered to UBFX.
7316 if (N->getOpcode() == ISD::AND && (VT == MVT::i32 || VT == MVT::i64) &&
7317 isa<ConstantSDNode>(N->getOperand(1))) {
7318 uint64_t TruncMask = N->getConstantOperandVal(1);
7319 if (isMask_64(TruncMask) &&
7320 N->getOperand(0).getOpcode() == ISD::SRL &&
7321 isa<ConstantSDNode>(N->getOperand(0)->getOperand(1)))
7327 bool AArch64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
7329 assert(Ty->isIntegerTy());
7331 unsigned BitSize = Ty->getPrimitiveSizeInBits();
7335 int64_t Val = Imm.getSExtValue();
7336 if (Val == 0 || AArch64_AM::isLogicalImmediate(Val, BitSize))
7339 if ((int64_t)Val < 0)
7342 Val &= (1LL << 32) - 1;
7344 unsigned LZ = countLeadingZeros((uint64_t)Val);
7345 unsigned Shift = (63 - LZ) / 16;
7346 // MOVZ is free so return true for one or fewer MOVK.
7350 // Generate SUBS and CSEL for integer abs.
7351 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
7352 EVT VT = N->getValueType(0);
7354 SDValue N0 = N->getOperand(0);
7355 SDValue N1 = N->getOperand(1);
7358 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
7359 // and change it to SUB and CSEL.
7360 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
7361 N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
7362 N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0))
7363 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
7364 if (Y1C->getAPIntValue() == VT.getSizeInBits() - 1) {
7365 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT),
7367 // Generate SUBS & CSEL.
7369 DAG.getNode(AArch64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32),
7370 N0.getOperand(0), DAG.getConstant(0, DL, VT));
7371 return DAG.getNode(AArch64ISD::CSEL, DL, VT, N0.getOperand(0), Neg,
7372 DAG.getConstant(AArch64CC::PL, DL, MVT::i32),
7373 SDValue(Cmp.getNode(), 1));
7378 // performXorCombine - Attempts to handle integer ABS.
7379 static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
7380 TargetLowering::DAGCombinerInfo &DCI,
7381 const AArch64Subtarget *Subtarget) {
7382 if (DCI.isBeforeLegalizeOps())
7385 return performIntegerAbsCombine(N, DAG);
7389 AArch64TargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor,
7391 std::vector<SDNode *> *Created) const {
7392 // fold (sdiv X, pow2)
7393 EVT VT = N->getValueType(0);
7394 if ((VT != MVT::i32 && VT != MVT::i64) ||
7395 !(Divisor.isPowerOf2() || (-Divisor).isPowerOf2()))
7399 SDValue N0 = N->getOperand(0);
7400 unsigned Lg2 = Divisor.countTrailingZeros();
7401 SDValue Zero = DAG.getConstant(0, DL, VT);
7402 SDValue Pow2MinusOne = DAG.getConstant((1ULL << Lg2) - 1, DL, VT);
7404 // Add (N0 < 0) ? Pow2 - 1 : 0;
7406 SDValue Cmp = getAArch64Cmp(N0, Zero, ISD::SETLT, CCVal, DAG, DL);
7407 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0, Pow2MinusOne);
7408 SDValue CSel = DAG.getNode(AArch64ISD::CSEL, DL, VT, Add, N0, CCVal, Cmp);
7411 Created->push_back(Cmp.getNode());
7412 Created->push_back(Add.getNode());
7413 Created->push_back(CSel.getNode());
7418 DAG.getNode(ISD::SRA, DL, VT, CSel, DAG.getConstant(Lg2, DL, MVT::i64));
7420 // If we're dividing by a positive value, we're done. Otherwise, we must
7421 // negate the result.
7422 if (Divisor.isNonNegative())
7426 Created->push_back(SRA.getNode());
7427 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), SRA);
7430 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG,
7431 TargetLowering::DAGCombinerInfo &DCI,
7432 const AArch64Subtarget *Subtarget) {
7433 if (DCI.isBeforeLegalizeOps())
7436 // Multiplication of a power of two plus/minus one can be done more
7437 // cheaply as as shift+add/sub. For now, this is true unilaterally. If
7438 // future CPUs have a cheaper MADD instruction, this may need to be
7439 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
7440 // 64-bit is 5 cycles, so this is always a win.
7441 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
7442 APInt Value = C->getAPIntValue();
7443 EVT VT = N->getValueType(0);
7445 if (Value.isNonNegative()) {
7446 // (mul x, 2^N + 1) => (add (shl x, N), x)
7447 APInt VM1 = Value - 1;
7448 if (VM1.isPowerOf2()) {
7449 SDValue ShiftedVal =
7450 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7451 DAG.getConstant(VM1.logBase2(), DL, MVT::i64));
7452 return DAG.getNode(ISD::ADD, DL, VT, ShiftedVal,
7455 // (mul x, 2^N - 1) => (sub (shl x, N), x)
7456 APInt VP1 = Value + 1;
7457 if (VP1.isPowerOf2()) {
7458 SDValue ShiftedVal =
7459 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7460 DAG.getConstant(VP1.logBase2(), DL, MVT::i64));
7461 return DAG.getNode(ISD::SUB, DL, VT, ShiftedVal,
7465 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
7466 APInt VNP1 = -Value + 1;
7467 if (VNP1.isPowerOf2()) {
7468 SDValue ShiftedVal =
7469 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7470 DAG.getConstant(VNP1.logBase2(), DL, MVT::i64));
7471 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0),
7474 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
7475 APInt VNM1 = -Value - 1;
7476 if (VNM1.isPowerOf2()) {
7477 SDValue ShiftedVal =
7478 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7479 DAG.getConstant(VNM1.logBase2(), DL, MVT::i64));
7481 DAG.getNode(ISD::ADD, DL, VT, ShiftedVal, N->getOperand(0));
7482 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Add);
7489 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
7490 SelectionDAG &DAG) {
7491 // Take advantage of vector comparisons producing 0 or -1 in each lane to
7492 // optimize away operation when it's from a constant.
7494 // The general transformation is:
7495 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
7496 // AND(VECTOR_CMP(x,y), constant2)
7497 // constant2 = UNARYOP(constant)
7499 // Early exit if this isn't a vector operation, the operand of the
7500 // unary operation isn't a bitwise AND, or if the sizes of the operations
7502 EVT VT = N->getValueType(0);
7503 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
7504 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
7505 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
7508 // Now check that the other operand of the AND is a constant. We could
7509 // make the transformation for non-constant splats as well, but it's unclear
7510 // that would be a benefit as it would not eliminate any operations, just
7511 // perform one more step in scalar code before moving to the vector unit.
7512 if (BuildVectorSDNode *BV =
7513 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
7514 // Bail out if the vector isn't a constant.
7515 if (!BV->isConstant())
7518 // Everything checks out. Build up the new and improved node.
7520 EVT IntVT = BV->getValueType(0);
7521 // Create a new constant of the appropriate type for the transformed
7523 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
7524 // The AND node needs bitcasts to/from an integer vector type around it.
7525 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
7526 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
7527 N->getOperand(0)->getOperand(0), MaskConst);
7528 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
7535 static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG,
7536 const AArch64Subtarget *Subtarget) {
7537 // First try to optimize away the conversion when it's conditionally from
7538 // a constant. Vectors only.
7539 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
7542 EVT VT = N->getValueType(0);
7543 if (VT != MVT::f32 && VT != MVT::f64)
7546 // Only optimize when the source and destination types have the same width.
7547 if (VT.getSizeInBits() != N->getOperand(0).getValueType().getSizeInBits())
7550 // If the result of an integer load is only used by an integer-to-float
7551 // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
7552 // This eliminates an "integer-to-vector-move" UOP and improves throughput.
7553 SDValue N0 = N->getOperand(0);
7554 if (Subtarget->hasNEON() && ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7555 // Do not change the width of a volatile load.
7556 !cast<LoadSDNode>(N0)->isVolatile()) {
7557 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7558 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
7559 LN0->getPointerInfo(), LN0->isVolatile(),
7560 LN0->isNonTemporal(), LN0->isInvariant(),
7561 LN0->getAlignment());
7563 // Make sure successors of the original load stay after it by updating them
7564 // to use the new Chain.
7565 DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));
7568 (N->getOpcode() == ISD::SINT_TO_FP) ? AArch64ISD::SITOF : AArch64ISD::UITOF;
7569 return DAG.getNode(Opcode, SDLoc(N), VT, Load);
7575 /// Fold a floating-point multiply by power of two into floating-point to
7576 /// fixed-point conversion.
7577 static SDValue performFpToIntCombine(SDNode *N, SelectionDAG &DAG,
7578 TargetLowering::DAGCombinerInfo &DCI,
7579 const AArch64Subtarget *Subtarget) {
7580 if (!Subtarget->hasNEON())
7583 SDValue Op = N->getOperand(0);
7584 if (!Op.getValueType().isVector() || !Op.getValueType().isSimple() ||
7585 Op.getOpcode() != ISD::FMUL)
7588 SDValue ConstVec = Op->getOperand(1);
7589 if (!isa<BuildVectorSDNode>(ConstVec))
7592 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
7593 uint32_t FloatBits = FloatTy.getSizeInBits();
7594 if (FloatBits != 32 && FloatBits != 64)
7597 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
7598 uint32_t IntBits = IntTy.getSizeInBits();
7599 if (IntBits != 16 && IntBits != 32 && IntBits != 64)
7602 // Avoid conversions where iN is larger than the float (e.g., float -> i64).
7603 if (IntBits > FloatBits)
7606 BitVector UndefElements;
7607 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(ConstVec);
7608 int32_t Bits = IntBits == 64 ? 64 : 32;
7609 int32_t C = BV->getConstantFPSplatPow2ToLog2Int(&UndefElements, Bits + 1);
7610 if (C == -1 || C == 0 || C > Bits)
7614 unsigned NumLanes = Op.getValueType().getVectorNumElements();
7619 ResTy = FloatBits == 32 ? MVT::v2i32 : MVT::v2i64;
7622 ResTy = FloatBits == 32 ? MVT::v4i32 : MVT::v4i64;
7626 if (ResTy == MVT::v4i64 && DCI.isBeforeLegalizeOps())
7629 assert((ResTy != MVT::v4i64 || DCI.isBeforeLegalizeOps()) &&
7630 "Illegal vector type after legalization");
7633 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
7634 unsigned IntrinsicOpcode = IsSigned ? Intrinsic::aarch64_neon_vcvtfp2fxs
7635 : Intrinsic::aarch64_neon_vcvtfp2fxu;
7637 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, ResTy,
7638 DAG.getConstant(IntrinsicOpcode, DL, MVT::i32),
7639 Op->getOperand(0), DAG.getConstant(C, DL, MVT::i32));
7640 // We can handle smaller integers by generating an extra trunc.
7641 if (IntBits < FloatBits)
7642 FixConv = DAG.getNode(ISD::TRUNCATE, DL, N->getValueType(0), FixConv);
7647 /// Fold a floating-point divide by power of two into fixed-point to
7648 /// floating-point conversion.
7649 static SDValue performFDivCombine(SDNode *N, SelectionDAG &DAG,
7650 const AArch64Subtarget *Subtarget) {
7651 if (!Subtarget->hasNEON())
7654 SDValue Op = N->getOperand(0);
7655 unsigned Opc = Op->getOpcode();
7656 if (!Op.getValueType().isVector() ||
7657 (Opc != ISD::SINT_TO_FP && Opc != ISD::UINT_TO_FP))
7660 SDValue ConstVec = N->getOperand(1);
7661 if (!isa<BuildVectorSDNode>(ConstVec))
7664 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
7665 int32_t IntBits = IntTy.getSizeInBits();
7666 if (IntBits != 16 && IntBits != 32 && IntBits != 64)
7669 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
7670 int32_t FloatBits = FloatTy.getSizeInBits();
7671 if (FloatBits != 32 && FloatBits != 64)
7674 // Avoid conversions where iN is larger than the float (e.g., i64 -> float).
7675 if (IntBits > FloatBits)
7678 BitVector UndefElements;
7679 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(ConstVec);
7680 int32_t C = BV->getConstantFPSplatPow2ToLog2Int(&UndefElements, FloatBits + 1);
7681 if (C == -1 || C == 0 || C > FloatBits)
7685 unsigned NumLanes = Op.getValueType().getVectorNumElements();
7690 ResTy = FloatBits == 32 ? MVT::v2i32 : MVT::v2i64;
7698 SDValue ConvInput = Op.getOperand(0);
7699 bool IsSigned = Opc == ISD::SINT_TO_FP;
7700 if (IntBits < FloatBits)
7701 ConvInput = DAG.getNode(IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND, DL,
7704 unsigned IntrinsicOpcode = IsSigned ? Intrinsic::aarch64_neon_vcvtfxs2fp
7705 : Intrinsic::aarch64_neon_vcvtfxu2fp;
7706 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, Op.getValueType(),
7707 DAG.getConstant(IntrinsicOpcode, DL, MVT::i32), ConvInput,
7708 DAG.getConstant(C, DL, MVT::i32));
7711 /// An EXTR instruction is made up of two shifts, ORed together. This helper
7712 /// searches for and classifies those shifts.
7713 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
7715 if (N.getOpcode() == ISD::SHL)
7717 else if (N.getOpcode() == ISD::SRL)
7722 if (!isa<ConstantSDNode>(N.getOperand(1)))
7725 ShiftAmount = N->getConstantOperandVal(1);
7726 Src = N->getOperand(0);
7730 /// EXTR instruction extracts a contiguous chunk of bits from two existing
7731 /// registers viewed as a high/low pair. This function looks for the pattern:
7732 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
7733 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
7735 static SDValue tryCombineToEXTR(SDNode *N,
7736 TargetLowering::DAGCombinerInfo &DCI) {
7737 SelectionDAG &DAG = DCI.DAG;
7739 EVT VT = N->getValueType(0);
7741 assert(N->getOpcode() == ISD::OR && "Unexpected root");
7743 if (VT != MVT::i32 && VT != MVT::i64)
7747 uint32_t ShiftLHS = 0;
7749 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
7753 uint32_t ShiftRHS = 0;
7755 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
7758 // If they're both trying to come from the high part of the register, they're
7759 // not really an EXTR.
7760 if (LHSFromHi == RHSFromHi)
7763 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
7767 std::swap(LHS, RHS);
7768 std::swap(ShiftLHS, ShiftRHS);
7771 return DAG.getNode(AArch64ISD::EXTR, DL, VT, LHS, RHS,
7772 DAG.getConstant(ShiftRHS, DL, MVT::i64));
7775 static SDValue tryCombineToBSL(SDNode *N,
7776 TargetLowering::DAGCombinerInfo &DCI) {
7777 EVT VT = N->getValueType(0);
7778 SelectionDAG &DAG = DCI.DAG;
7784 SDValue N0 = N->getOperand(0);
7785 if (N0.getOpcode() != ISD::AND)
7788 SDValue N1 = N->getOperand(1);
7789 if (N1.getOpcode() != ISD::AND)
7792 // We only have to look for constant vectors here since the general, variable
7793 // case can be handled in TableGen.
7794 unsigned Bits = VT.getVectorElementType().getSizeInBits();
7795 uint64_t BitMask = Bits == 64 ? -1ULL : ((1ULL << Bits) - 1);
7796 for (int i = 1; i >= 0; --i)
7797 for (int j = 1; j >= 0; --j) {
7798 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(i));
7799 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(j));
7803 bool FoundMatch = true;
7804 for (unsigned k = 0; k < VT.getVectorNumElements(); ++k) {
7805 ConstantSDNode *CN0 = dyn_cast<ConstantSDNode>(BVN0->getOperand(k));
7806 ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(BVN1->getOperand(k));
7808 CN0->getZExtValue() != (BitMask & ~CN1->getZExtValue())) {
7815 return DAG.getNode(AArch64ISD::BSL, DL, VT, SDValue(BVN0, 0),
7816 N0->getOperand(1 - i), N1->getOperand(1 - j));
7822 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
7823 const AArch64Subtarget *Subtarget) {
7824 // Attempt to form an EXTR from (or (shl VAL1, #N), (srl VAL2, #RegWidth-N))
7825 if (!EnableAArch64ExtrGeneration)
7827 SelectionDAG &DAG = DCI.DAG;
7828 EVT VT = N->getValueType(0);
7830 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7833 SDValue Res = tryCombineToEXTR(N, DCI);
7837 Res = tryCombineToBSL(N, DCI);
7844 static SDValue performBitcastCombine(SDNode *N,
7845 TargetLowering::DAGCombinerInfo &DCI,
7846 SelectionDAG &DAG) {
7847 // Wait 'til after everything is legalized to try this. That way we have
7848 // legal vector types and such.
7849 if (DCI.isBeforeLegalizeOps())
7852 // Remove extraneous bitcasts around an extract_subvector.
7854 // (v4i16 (bitconvert
7855 // (extract_subvector (v2i64 (bitconvert (v8i16 ...)), (i64 1)))))
7857 // (extract_subvector ((v8i16 ...), (i64 4)))
7859 // Only interested in 64-bit vectors as the ultimate result.
7860 EVT VT = N->getValueType(0);
7863 if (VT.getSimpleVT().getSizeInBits() != 64)
7865 // Is the operand an extract_subvector starting at the beginning or halfway
7866 // point of the vector? A low half may also come through as an
7867 // EXTRACT_SUBREG, so look for that, too.
7868 SDValue Op0 = N->getOperand(0);
7869 if (Op0->getOpcode() != ISD::EXTRACT_SUBVECTOR &&
7870 !(Op0->isMachineOpcode() &&
7871 Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG))
7873 uint64_t idx = cast<ConstantSDNode>(Op0->getOperand(1))->getZExtValue();
7874 if (Op0->getOpcode() == ISD::EXTRACT_SUBVECTOR) {
7875 if (Op0->getValueType(0).getVectorNumElements() != idx && idx != 0)
7877 } else if (Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG) {
7878 if (idx != AArch64::dsub)
7880 // The dsub reference is equivalent to a lane zero subvector reference.
7883 // Look through the bitcast of the input to the extract.
7884 if (Op0->getOperand(0)->getOpcode() != ISD::BITCAST)
7886 SDValue Source = Op0->getOperand(0)->getOperand(0);
7887 // If the source type has twice the number of elements as our destination
7888 // type, we know this is an extract of the high or low half of the vector.
7889 EVT SVT = Source->getValueType(0);
7890 if (SVT.getVectorNumElements() != VT.getVectorNumElements() * 2)
7893 DEBUG(dbgs() << "aarch64-lower: bitcast extract_subvector simplification\n");
7895 // Create the simplified form to just extract the low or high half of the
7896 // vector directly rather than bothering with the bitcasts.
7898 unsigned NumElements = VT.getVectorNumElements();
7900 SDValue HalfIdx = DAG.getConstant(NumElements, dl, MVT::i64);
7901 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Source, HalfIdx);
7903 SDValue SubReg = DAG.getTargetConstant(AArch64::dsub, dl, MVT::i32);
7904 return SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl, VT,
7910 static SDValue performConcatVectorsCombine(SDNode *N,
7911 TargetLowering::DAGCombinerInfo &DCI,
7912 SelectionDAG &DAG) {
7914 EVT VT = N->getValueType(0);
7915 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
7917 // Optimize concat_vectors of truncated vectors, where the intermediate
7918 // type is illegal, to avoid said illegality, e.g.,
7919 // (v4i16 (concat_vectors (v2i16 (truncate (v2i64))),
7920 // (v2i16 (truncate (v2i64)))))
7922 // (v4i16 (truncate (vector_shuffle (v4i32 (bitcast (v2i64))),
7923 // (v4i32 (bitcast (v2i64))),
7925 // This isn't really target-specific, but ISD::TRUNCATE legality isn't keyed
7926 // on both input and result type, so we might generate worse code.
7927 // On AArch64 we know it's fine for v2i64->v4i16 and v4i32->v8i8.
7928 if (N->getNumOperands() == 2 &&
7929 N0->getOpcode() == ISD::TRUNCATE &&
7930 N1->getOpcode() == ISD::TRUNCATE) {
7931 SDValue N00 = N0->getOperand(0);
7932 SDValue N10 = N1->getOperand(0);
7933 EVT N00VT = N00.getValueType();
7935 if (N00VT == N10.getValueType() &&
7936 (N00VT == MVT::v2i64 || N00VT == MVT::v4i32) &&
7937 N00VT.getScalarSizeInBits() == 4 * VT.getScalarSizeInBits()) {
7938 MVT MidVT = (N00VT == MVT::v2i64 ? MVT::v4i32 : MVT::v8i16);
7939 SmallVector<int, 8> Mask(MidVT.getVectorNumElements());
7940 for (size_t i = 0; i < Mask.size(); ++i)
7942 return DAG.getNode(ISD::TRUNCATE, dl, VT,
7943 DAG.getVectorShuffle(
7945 DAG.getNode(ISD::BITCAST, dl, MidVT, N00),
7946 DAG.getNode(ISD::BITCAST, dl, MidVT, N10), Mask));
7950 // Wait 'til after everything is legalized to try this. That way we have
7951 // legal vector types and such.
7952 if (DCI.isBeforeLegalizeOps())
7955 // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector
7956 // splat. The indexed instructions are going to be expecting a DUPLANE64, so
7957 // canonicalise to that.
7958 if (N0 == N1 && VT.getVectorNumElements() == 2) {
7959 assert(VT.getVectorElementType().getSizeInBits() == 64);
7960 return DAG.getNode(AArch64ISD::DUPLANE64, dl, VT, WidenVector(N0, DAG),
7961 DAG.getConstant(0, dl, MVT::i64));
7964 // Canonicalise concat_vectors so that the right-hand vector has as few
7965 // bit-casts as possible before its real operation. The primary matching
7966 // destination for these operations will be the narrowing "2" instructions,
7967 // which depend on the operation being performed on this right-hand vector.
7969 // (concat_vectors LHS, (v1i64 (bitconvert (v4i16 RHS))))
7971 // (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS))
7973 if (N1->getOpcode() != ISD::BITCAST)
7975 SDValue RHS = N1->getOperand(0);
7976 MVT RHSTy = RHS.getValueType().getSimpleVT();
7977 // If the RHS is not a vector, this is not the pattern we're looking for.
7978 if (!RHSTy.isVector())
7981 DEBUG(dbgs() << "aarch64-lower: concat_vectors bitcast simplification\n");
7983 MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(),
7984 RHSTy.getVectorNumElements() * 2);
7985 return DAG.getNode(ISD::BITCAST, dl, VT,
7986 DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatTy,
7987 DAG.getNode(ISD::BITCAST, dl, RHSTy, N0),
7991 static SDValue tryCombineFixedPointConvert(SDNode *N,
7992 TargetLowering::DAGCombinerInfo &DCI,
7993 SelectionDAG &DAG) {
7994 // Wait 'til after everything is legalized to try this. That way we have
7995 // legal vector types and such.
7996 if (DCI.isBeforeLegalizeOps())
7998 // Transform a scalar conversion of a value from a lane extract into a
7999 // lane extract of a vector conversion. E.g., from foo1 to foo2:
8000 // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); }
8001 // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; }
8003 // The second form interacts better with instruction selection and the
8004 // register allocator to avoid cross-class register copies that aren't
8005 // coalescable due to a lane reference.
8007 // Check the operand and see if it originates from a lane extract.
8008 SDValue Op1 = N->getOperand(1);
8009 if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8010 // Yep, no additional predication needed. Perform the transform.
8011 SDValue IID = N->getOperand(0);
8012 SDValue Shift = N->getOperand(2);
8013 SDValue Vec = Op1.getOperand(0);
8014 SDValue Lane = Op1.getOperand(1);
8015 EVT ResTy = N->getValueType(0);
8019 // The vector width should be 128 bits by the time we get here, even
8020 // if it started as 64 bits (the extract_vector handling will have
8022 assert(Vec.getValueType().getSizeInBits() == 128 &&
8023 "unexpected vector size on extract_vector_elt!");
8024 if (Vec.getValueType() == MVT::v4i32)
8025 VecResTy = MVT::v4f32;
8026 else if (Vec.getValueType() == MVT::v2i64)
8027 VecResTy = MVT::v2f64;
8029 llvm_unreachable("unexpected vector type!");
8032 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift);
8033 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane);
8038 // AArch64 high-vector "long" operations are formed by performing the non-high
8039 // version on an extract_subvector of each operand which gets the high half:
8041 // (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS))
8043 // However, there are cases which don't have an extract_high explicitly, but
8044 // have another operation that can be made compatible with one for free. For
8047 // (dupv64 scalar) --> (extract_high (dup128 scalar))
8049 // This routine does the actual conversion of such DUPs, once outer routines
8050 // have determined that everything else is in order.
8051 // It also supports immediate DUP-like nodes (MOVI/MVNi), which we can fold
8053 static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) {
8054 switch (N.getOpcode()) {
8055 case AArch64ISD::DUP:
8056 case AArch64ISD::DUPLANE8:
8057 case AArch64ISD::DUPLANE16:
8058 case AArch64ISD::DUPLANE32:
8059 case AArch64ISD::DUPLANE64:
8060 case AArch64ISD::MOVI:
8061 case AArch64ISD::MOVIshift:
8062 case AArch64ISD::MOVIedit:
8063 case AArch64ISD::MOVImsl:
8064 case AArch64ISD::MVNIshift:
8065 case AArch64ISD::MVNImsl:
8068 // FMOV could be supported, but isn't very useful, as it would only occur
8069 // if you passed a bitcast' floating point immediate to an eligible long
8070 // integer op (addl, smull, ...).
8074 MVT NarrowTy = N.getSimpleValueType();
8075 if (!NarrowTy.is64BitVector())
8078 MVT ElementTy = NarrowTy.getVectorElementType();
8079 unsigned NumElems = NarrowTy.getVectorNumElements();
8080 MVT NewVT = MVT::getVectorVT(ElementTy, NumElems * 2);
8083 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NarrowTy,
8084 DAG.getNode(N->getOpcode(), dl, NewVT, N->ops()),
8085 DAG.getConstant(NumElems, dl, MVT::i64));
8088 static bool isEssentiallyExtractSubvector(SDValue N) {
8089 if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR)
8092 return N.getOpcode() == ISD::BITCAST &&
8093 N.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR;
8096 /// \brief Helper structure to keep track of ISD::SET_CC operands.
8097 struct GenericSetCCInfo {
8098 const SDValue *Opnd0;
8099 const SDValue *Opnd1;
8103 /// \brief Helper structure to keep track of a SET_CC lowered into AArch64 code.
8104 struct AArch64SetCCInfo {
8106 AArch64CC::CondCode CC;
8109 /// \brief Helper structure to keep track of SetCC information.
8111 GenericSetCCInfo Generic;
8112 AArch64SetCCInfo AArch64;
8115 /// \brief Helper structure to be able to read SetCC information. If set to
8116 /// true, IsAArch64 field, Info is a AArch64SetCCInfo, otherwise Info is a
8117 /// GenericSetCCInfo.
8118 struct SetCCInfoAndKind {
8123 /// \brief Check whether or not \p Op is a SET_CC operation, either a generic or
8125 /// AArch64 lowered one.
8126 /// \p SetCCInfo is filled accordingly.
8127 /// \post SetCCInfo is meanginfull only when this function returns true.
8128 /// \return True when Op is a kind of SET_CC operation.
8129 static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) {
8130 // If this is a setcc, this is straight forward.
8131 if (Op.getOpcode() == ISD::SETCC) {
8132 SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0);
8133 SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1);
8134 SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
8135 SetCCInfo.IsAArch64 = false;
8138 // Otherwise, check if this is a matching csel instruction.
8142 if (Op.getOpcode() != AArch64ISD::CSEL)
8144 // Set the information about the operands.
8145 // TODO: we want the operands of the Cmp not the csel
8146 SetCCInfo.Info.AArch64.Cmp = &Op.getOperand(3);
8147 SetCCInfo.IsAArch64 = true;
8148 SetCCInfo.Info.AArch64.CC = static_cast<AArch64CC::CondCode>(
8149 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
8151 // Check that the operands matches the constraints:
8152 // (1) Both operands must be constants.
8153 // (2) One must be 1 and the other must be 0.
8154 ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0));
8155 ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1));
8158 if (!TValue || !FValue)
8162 if (!TValue->isOne()) {
8163 // Update the comparison when we are interested in !cc.
8164 std::swap(TValue, FValue);
8165 SetCCInfo.Info.AArch64.CC =
8166 AArch64CC::getInvertedCondCode(SetCCInfo.Info.AArch64.CC);
8168 return TValue->isOne() && FValue->isNullValue();
8171 // Returns true if Op is setcc or zext of setcc.
8172 static bool isSetCCOrZExtSetCC(const SDValue& Op, SetCCInfoAndKind &Info) {
8173 if (isSetCC(Op, Info))
8175 return ((Op.getOpcode() == ISD::ZERO_EXTEND) &&
8176 isSetCC(Op->getOperand(0), Info));
8179 // The folding we want to perform is:
8180 // (add x, [zext] (setcc cc ...) )
8182 // (csel x, (add x, 1), !cc ...)
8184 // The latter will get matched to a CSINC instruction.
8185 static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) {
8186 assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!");
8187 SDValue LHS = Op->getOperand(0);
8188 SDValue RHS = Op->getOperand(1);
8189 SetCCInfoAndKind InfoAndKind;
8191 // If neither operand is a SET_CC, give up.
8192 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind)) {
8193 std::swap(LHS, RHS);
8194 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind))
8198 // FIXME: This could be generatized to work for FP comparisons.
8199 EVT CmpVT = InfoAndKind.IsAArch64
8200 ? InfoAndKind.Info.AArch64.Cmp->getOperand(0).getValueType()
8201 : InfoAndKind.Info.Generic.Opnd0->getValueType();
8202 if (CmpVT != MVT::i32 && CmpVT != MVT::i64)
8208 if (InfoAndKind.IsAArch64) {
8209 CCVal = DAG.getConstant(
8210 AArch64CC::getInvertedCondCode(InfoAndKind.Info.AArch64.CC), dl,
8212 Cmp = *InfoAndKind.Info.AArch64.Cmp;
8214 Cmp = getAArch64Cmp(*InfoAndKind.Info.Generic.Opnd0,
8215 *InfoAndKind.Info.Generic.Opnd1,
8216 ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, true),
8219 EVT VT = Op->getValueType(0);
8220 LHS = DAG.getNode(ISD::ADD, dl, VT, RHS, DAG.getConstant(1, dl, VT));
8221 return DAG.getNode(AArch64ISD::CSEL, dl, VT, RHS, LHS, CCVal, Cmp);
8224 // The basic add/sub long vector instructions have variants with "2" on the end
8225 // which act on the high-half of their inputs. They are normally matched by
8228 // (add (zeroext (extract_high LHS)),
8229 // (zeroext (extract_high RHS)))
8230 // -> uaddl2 vD, vN, vM
8232 // However, if one of the extracts is something like a duplicate, this
8233 // instruction can still be used profitably. This function puts the DAG into a
8234 // more appropriate form for those patterns to trigger.
8235 static SDValue performAddSubLongCombine(SDNode *N,
8236 TargetLowering::DAGCombinerInfo &DCI,
8237 SelectionDAG &DAG) {
8238 if (DCI.isBeforeLegalizeOps())
8241 MVT VT = N->getSimpleValueType(0);
8242 if (!VT.is128BitVector()) {
8243 if (N->getOpcode() == ISD::ADD)
8244 return performSetccAddFolding(N, DAG);
8248 // Make sure both branches are extended in the same way.
8249 SDValue LHS = N->getOperand(0);
8250 SDValue RHS = N->getOperand(1);
8251 if ((LHS.getOpcode() != ISD::ZERO_EXTEND &&
8252 LHS.getOpcode() != ISD::SIGN_EXTEND) ||
8253 LHS.getOpcode() != RHS.getOpcode())
8256 unsigned ExtType = LHS.getOpcode();
8258 // It's not worth doing if at least one of the inputs isn't already an
8259 // extract, but we don't know which it'll be so we have to try both.
8260 if (isEssentiallyExtractSubvector(LHS.getOperand(0))) {
8261 RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG);
8265 RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS);
8266 } else if (isEssentiallyExtractSubvector(RHS.getOperand(0))) {
8267 LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG);
8271 LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS);
8274 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS);
8277 // Massage DAGs which we can use the high-half "long" operations on into
8278 // something isel will recognize better. E.g.
8280 // (aarch64_neon_umull (extract_high vec) (dupv64 scalar)) -->
8281 // (aarch64_neon_umull (extract_high (v2i64 vec)))
8282 // (extract_high (v2i64 (dup128 scalar)))))
8284 static SDValue tryCombineLongOpWithDup(unsigned IID, SDNode *N,
8285 TargetLowering::DAGCombinerInfo &DCI,
8286 SelectionDAG &DAG) {
8287 if (DCI.isBeforeLegalizeOps())
8290 SDValue LHS = N->getOperand(1);
8291 SDValue RHS = N->getOperand(2);
8292 assert(LHS.getValueType().is64BitVector() &&
8293 RHS.getValueType().is64BitVector() &&
8294 "unexpected shape for long operation");
8296 // Either node could be a DUP, but it's not worth doing both of them (you'd
8297 // just as well use the non-high version) so look for a corresponding extract
8298 // operation on the other "wing".
8299 if (isEssentiallyExtractSubvector(LHS)) {
8300 RHS = tryExtendDUPToExtractHigh(RHS, DAG);
8303 } else if (isEssentiallyExtractSubvector(RHS)) {
8304 LHS = tryExtendDUPToExtractHigh(LHS, DAG);
8309 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0),
8310 N->getOperand(0), LHS, RHS);
8313 static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
8314 MVT ElemTy = N->getSimpleValueType(0).getScalarType();
8315 unsigned ElemBits = ElemTy.getSizeInBits();
8317 int64_t ShiftAmount;
8318 if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) {
8319 APInt SplatValue, SplatUndef;
8320 unsigned SplatBitSize;
8322 if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
8323 HasAnyUndefs, ElemBits) ||
8324 SplatBitSize != ElemBits)
8327 ShiftAmount = SplatValue.getSExtValue();
8328 } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) {
8329 ShiftAmount = CVN->getSExtValue();
8337 llvm_unreachable("Unknown shift intrinsic");
8338 case Intrinsic::aarch64_neon_sqshl:
8339 Opcode = AArch64ISD::SQSHL_I;
8340 IsRightShift = false;
8342 case Intrinsic::aarch64_neon_uqshl:
8343 Opcode = AArch64ISD::UQSHL_I;
8344 IsRightShift = false;
8346 case Intrinsic::aarch64_neon_srshl:
8347 Opcode = AArch64ISD::SRSHR_I;
8348 IsRightShift = true;
8350 case Intrinsic::aarch64_neon_urshl:
8351 Opcode = AArch64ISD::URSHR_I;
8352 IsRightShift = true;
8354 case Intrinsic::aarch64_neon_sqshlu:
8355 Opcode = AArch64ISD::SQSHLU_I;
8356 IsRightShift = false;
8360 if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits) {
8362 return DAG.getNode(Opcode, dl, N->getValueType(0), N->getOperand(1),
8363 DAG.getConstant(-ShiftAmount, dl, MVT::i32));
8364 } else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount < ElemBits) {
8366 return DAG.getNode(Opcode, dl, N->getValueType(0), N->getOperand(1),
8367 DAG.getConstant(ShiftAmount, dl, MVT::i32));
8373 // The CRC32[BH] instructions ignore the high bits of their data operand. Since
8374 // the intrinsics must be legal and take an i32, this means there's almost
8375 // certainly going to be a zext in the DAG which we can eliminate.
8376 static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) {
8377 SDValue AndN = N->getOperand(2);
8378 if (AndN.getOpcode() != ISD::AND)
8381 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1));
8382 if (!CMask || CMask->getZExtValue() != Mask)
8385 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32,
8386 N->getOperand(0), N->getOperand(1), AndN.getOperand(0));
8389 static SDValue combineAcrossLanesIntrinsic(unsigned Opc, SDNode *N,
8390 SelectionDAG &DAG) {
8392 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0),
8393 DAG.getNode(Opc, dl,
8394 N->getOperand(1).getSimpleValueType(),
8396 DAG.getConstant(0, dl, MVT::i64));
8399 static SDValue performIntrinsicCombine(SDNode *N,
8400 TargetLowering::DAGCombinerInfo &DCI,
8401 const AArch64Subtarget *Subtarget) {
8402 SelectionDAG &DAG = DCI.DAG;
8403 unsigned IID = getIntrinsicID(N);
8407 case Intrinsic::aarch64_neon_vcvtfxs2fp:
8408 case Intrinsic::aarch64_neon_vcvtfxu2fp:
8409 return tryCombineFixedPointConvert(N, DCI, DAG);
8410 case Intrinsic::aarch64_neon_saddv:
8411 return combineAcrossLanesIntrinsic(AArch64ISD::SADDV, N, DAG);
8412 case Intrinsic::aarch64_neon_uaddv:
8413 return combineAcrossLanesIntrinsic(AArch64ISD::UADDV, N, DAG);
8414 case Intrinsic::aarch64_neon_sminv:
8415 return combineAcrossLanesIntrinsic(AArch64ISD::SMINV, N, DAG);
8416 case Intrinsic::aarch64_neon_uminv:
8417 return combineAcrossLanesIntrinsic(AArch64ISD::UMINV, N, DAG);
8418 case Intrinsic::aarch64_neon_smaxv:
8419 return combineAcrossLanesIntrinsic(AArch64ISD::SMAXV, N, DAG);
8420 case Intrinsic::aarch64_neon_umaxv:
8421 return combineAcrossLanesIntrinsic(AArch64ISD::UMAXV, N, DAG);
8422 case Intrinsic::aarch64_neon_fmax:
8423 return DAG.getNode(ISD::FMAXNAN, SDLoc(N), N->getValueType(0),
8424 N->getOperand(1), N->getOperand(2));
8425 case Intrinsic::aarch64_neon_fmin:
8426 return DAG.getNode(ISD::FMINNAN, SDLoc(N), N->getValueType(0),
8427 N->getOperand(1), N->getOperand(2));
8428 case Intrinsic::aarch64_neon_fmaxnm:
8429 return DAG.getNode(ISD::FMAXNUM, SDLoc(N), N->getValueType(0),
8430 N->getOperand(1), N->getOperand(2));
8431 case Intrinsic::aarch64_neon_fminnm:
8432 return DAG.getNode(ISD::FMINNUM, SDLoc(N), N->getValueType(0),
8433 N->getOperand(1), N->getOperand(2));
8434 case Intrinsic::aarch64_neon_smull:
8435 case Intrinsic::aarch64_neon_umull:
8436 case Intrinsic::aarch64_neon_pmull:
8437 case Intrinsic::aarch64_neon_sqdmull:
8438 return tryCombineLongOpWithDup(IID, N, DCI, DAG);
8439 case Intrinsic::aarch64_neon_sqshl:
8440 case Intrinsic::aarch64_neon_uqshl:
8441 case Intrinsic::aarch64_neon_sqshlu:
8442 case Intrinsic::aarch64_neon_srshl:
8443 case Intrinsic::aarch64_neon_urshl:
8444 return tryCombineShiftImm(IID, N, DAG);
8445 case Intrinsic::aarch64_crc32b:
8446 case Intrinsic::aarch64_crc32cb:
8447 return tryCombineCRC32(0xff, N, DAG);
8448 case Intrinsic::aarch64_crc32h:
8449 case Intrinsic::aarch64_crc32ch:
8450 return tryCombineCRC32(0xffff, N, DAG);
8455 static SDValue performExtendCombine(SDNode *N,
8456 TargetLowering::DAGCombinerInfo &DCI,
8457 SelectionDAG &DAG) {
8458 // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then
8459 // we can convert that DUP into another extract_high (of a bigger DUP), which
8460 // helps the backend to decide that an sabdl2 would be useful, saving a real
8461 // extract_high operation.
8462 if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND &&
8463 N->getOperand(0).getOpcode() == ISD::INTRINSIC_WO_CHAIN) {
8464 SDNode *ABDNode = N->getOperand(0).getNode();
8465 unsigned IID = getIntrinsicID(ABDNode);
8466 if (IID == Intrinsic::aarch64_neon_sabd ||
8467 IID == Intrinsic::aarch64_neon_uabd) {
8468 SDValue NewABD = tryCombineLongOpWithDup(IID, ABDNode, DCI, DAG);
8469 if (!NewABD.getNode())
8472 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0),
8477 // This is effectively a custom type legalization for AArch64.
8479 // Type legalization will split an extend of a small, legal, type to a larger
8480 // illegal type by first splitting the destination type, often creating
8481 // illegal source types, which then get legalized in isel-confusing ways,
8482 // leading to really terrible codegen. E.g.,
8483 // %result = v8i32 sext v8i8 %value
8485 // %losrc = extract_subreg %value, ...
8486 // %hisrc = extract_subreg %value, ...
8487 // %lo = v4i32 sext v4i8 %losrc
8488 // %hi = v4i32 sext v4i8 %hisrc
8489 // Things go rapidly downhill from there.
8491 // For AArch64, the [sz]ext vector instructions can only go up one element
8492 // size, so we can, e.g., extend from i8 to i16, but to go from i8 to i32
8493 // take two instructions.
8495 // This implies that the most efficient way to do the extend from v8i8
8496 // to two v4i32 values is to first extend the v8i8 to v8i16, then do
8497 // the normal splitting to happen for the v8i16->v8i32.
8499 // This is pre-legalization to catch some cases where the default
8500 // type legalization will create ill-tempered code.
8501 if (!DCI.isBeforeLegalizeOps())
8504 // We're only interested in cleaning things up for non-legal vector types
8505 // here. If both the source and destination are legal, things will just
8506 // work naturally without any fiddling.
8507 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8508 EVT ResVT = N->getValueType(0);
8509 if (!ResVT.isVector() || TLI.isTypeLegal(ResVT))
8511 // If the vector type isn't a simple VT, it's beyond the scope of what
8512 // we're worried about here. Let legalization do its thing and hope for
8514 SDValue Src = N->getOperand(0);
8515 EVT SrcVT = Src->getValueType(0);
8516 if (!ResVT.isSimple() || !SrcVT.isSimple())
8519 // If the source VT is a 64-bit vector, we can play games and get the
8520 // better results we want.
8521 if (SrcVT.getSizeInBits() != 64)
8524 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
8525 unsigned ElementCount = SrcVT.getVectorNumElements();
8526 SrcVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize * 2), ElementCount);
8528 Src = DAG.getNode(N->getOpcode(), DL, SrcVT, Src);
8530 // Now split the rest of the operation into two halves, each with a 64
8534 unsigned NumElements = ResVT.getVectorNumElements();
8535 assert(!(NumElements & 1) && "Splitting vector, but not in half!");
8536 LoVT = HiVT = EVT::getVectorVT(*DAG.getContext(),
8537 ResVT.getVectorElementType(), NumElements / 2);
8539 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
8540 LoVT.getVectorNumElements());
8541 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
8542 DAG.getConstant(0, DL, MVT::i64));
8543 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
8544 DAG.getConstant(InNVT.getVectorNumElements(), DL, MVT::i64));
8545 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, Lo);
8546 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, Hi);
8548 // Now combine the parts back together so we still have a single result
8549 // like the combiner expects.
8550 return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
8553 /// Replace a splat of a scalar to a vector store by scalar stores of the scalar
8554 /// value. The load store optimizer pass will merge them to store pair stores.
8555 /// This has better performance than a splat of the scalar followed by a split
8556 /// vector store. Even if the stores are not merged it is four stores vs a dup,
8557 /// followed by an ext.b and two stores.
8558 static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) {
8559 SDValue StVal = St->getValue();
8560 EVT VT = StVal.getValueType();
8562 // Don't replace floating point stores, they possibly won't be transformed to
8563 // stp because of the store pair suppress pass.
8564 if (VT.isFloatingPoint())
8567 // Check for insert vector elements.
8568 if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT)
8571 // We can express a splat as store pair(s) for 2 or 4 elements.
8572 unsigned NumVecElts = VT.getVectorNumElements();
8573 if (NumVecElts != 4 && NumVecElts != 2)
8575 SDValue SplatVal = StVal.getOperand(1);
8576 unsigned RemainInsertElts = NumVecElts - 1;
8578 // Check that this is a splat.
8579 while (--RemainInsertElts) {
8580 SDValue NextInsertElt = StVal.getOperand(0);
8581 if (NextInsertElt.getOpcode() != ISD::INSERT_VECTOR_ELT)
8583 if (NextInsertElt.getOperand(1) != SplatVal)
8585 StVal = NextInsertElt;
8587 unsigned OrigAlignment = St->getAlignment();
8588 unsigned EltOffset = NumVecElts == 4 ? 4 : 8;
8589 unsigned Alignment = std::min(OrigAlignment, EltOffset);
8591 // Create scalar stores. This is at least as good as the code sequence for a
8592 // split unaligned store which is a dup.s, ext.b, and two stores.
8593 // Most of the time the three stores should be replaced by store pair
8594 // instructions (stp).
8596 SDValue BasePtr = St->getBasePtr();
8598 DAG.getStore(St->getChain(), DL, SplatVal, BasePtr, St->getPointerInfo(),
8599 St->isVolatile(), St->isNonTemporal(), St->getAlignment());
8601 unsigned Offset = EltOffset;
8602 while (--NumVecElts) {
8603 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
8604 DAG.getConstant(Offset, DL, MVT::i64));
8605 NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr,
8606 St->getPointerInfo(), St->isVolatile(),
8607 St->isNonTemporal(), Alignment);
8608 Offset += EltOffset;
8613 static SDValue split16BStores(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
8615 const AArch64Subtarget *Subtarget) {
8616 if (!DCI.isBeforeLegalize())
8619 StoreSDNode *S = cast<StoreSDNode>(N);
8620 if (S->isVolatile())
8623 // FIXME: The logic for deciding if an unaligned store should be split should
8624 // be included in TLI.allowsMisalignedMemoryAccesses(), and there should be
8625 // a call to that function here.
8627 // Cyclone has bad performance on unaligned 16B stores when crossing line and
8628 // page boundaries. We want to split such stores.
8629 if (!Subtarget->isCyclone())
8632 // Don't split at -Oz.
8633 if (DAG.getMachineFunction().getFunction()->optForMinSize())
8636 SDValue StVal = S->getValue();
8637 EVT VT = StVal.getValueType();
8639 // Don't split v2i64 vectors. Memcpy lowering produces those and splitting
8640 // those up regresses performance on micro-benchmarks and olden/bh.
8641 if (!VT.isVector() || VT.getVectorNumElements() < 2 || VT == MVT::v2i64)
8644 // Split unaligned 16B stores. They are terrible for performance.
8645 // Don't split stores with alignment of 1 or 2. Code that uses clang vector
8646 // extensions can use this to mark that it does not want splitting to happen
8647 // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of
8648 // eliminating alignment hazards is only 1 in 8 for alignment of 2.
8649 if (VT.getSizeInBits() != 128 || S->getAlignment() >= 16 ||
8650 S->getAlignment() <= 2)
8653 // If we get a splat of a scalar convert this vector store to a store of
8654 // scalars. They will be merged into store pairs thereby removing two
8656 if (SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S))
8657 return ReplacedSplat;
8660 unsigned NumElts = VT.getVectorNumElements() / 2;
8661 // Split VT into two.
8663 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), NumElts);
8664 SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
8665 DAG.getConstant(0, DL, MVT::i64));
8666 SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
8667 DAG.getConstant(NumElts, DL, MVT::i64));
8668 SDValue BasePtr = S->getBasePtr();
8670 DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
8671 S->isVolatile(), S->isNonTemporal(), S->getAlignment());
8672 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
8673 DAG.getConstant(8, DL, MVT::i64));
8674 return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr,
8675 S->getPointerInfo(), S->isVolatile(), S->isNonTemporal(),
8679 /// Target-specific DAG combine function for post-increment LD1 (lane) and
8680 /// post-increment LD1R.
8681 static SDValue performPostLD1Combine(SDNode *N,
8682 TargetLowering::DAGCombinerInfo &DCI,
8684 if (DCI.isBeforeLegalizeOps())
8687 SelectionDAG &DAG = DCI.DAG;
8688 EVT VT = N->getValueType(0);
8690 unsigned LoadIdx = IsLaneOp ? 1 : 0;
8691 SDNode *LD = N->getOperand(LoadIdx).getNode();
8692 // If it is not LOAD, can not do such combine.
8693 if (LD->getOpcode() != ISD::LOAD)
8696 LoadSDNode *LoadSDN = cast<LoadSDNode>(LD);
8697 EVT MemVT = LoadSDN->getMemoryVT();
8698 // Check if memory operand is the same type as the vector element.
8699 if (MemVT != VT.getVectorElementType())
8702 // Check if there are other uses. If so, do not combine as it will introduce
8704 for (SDNode::use_iterator UI = LD->use_begin(), UE = LD->use_end(); UI != UE;
8706 if (UI.getUse().getResNo() == 1) // Ignore uses of the chain result.
8712 SDValue Addr = LD->getOperand(1);
8713 SDValue Vector = N->getOperand(0);
8714 // Search for a use of the address operand that is an increment.
8715 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(), UE =
8716 Addr.getNode()->use_end(); UI != UE; ++UI) {
8718 if (User->getOpcode() != ISD::ADD
8719 || UI.getUse().getResNo() != Addr.getResNo())
8722 // Check that the add is independent of the load. Otherwise, folding it
8723 // would create a cycle.
8724 if (User->isPredecessorOf(LD) || LD->isPredecessorOf(User))
8726 // Also check that add is not used in the vector operand. This would also
8728 if (User->isPredecessorOf(Vector.getNode()))
8731 // If the increment is a constant, it must match the memory ref size.
8732 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8733 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8734 uint32_t IncVal = CInc->getZExtValue();
8735 unsigned NumBytes = VT.getScalarSizeInBits() / 8;
8736 if (IncVal != NumBytes)
8738 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
8741 // Finally, check that the vector doesn't depend on the load.
8742 // Again, this would create a cycle.
8743 // The load depending on the vector is fine, as that's the case for the
8744 // LD1*post we'll eventually generate anyway.
8745 if (LoadSDN->isPredecessorOf(Vector.getNode()))
8748 SmallVector<SDValue, 8> Ops;
8749 Ops.push_back(LD->getOperand(0)); // Chain
8751 Ops.push_back(Vector); // The vector to be inserted
8752 Ops.push_back(N->getOperand(2)); // The lane to be inserted in the vector
8754 Ops.push_back(Addr);
8757 EVT Tys[3] = { VT, MVT::i64, MVT::Other };
8758 SDVTList SDTys = DAG.getVTList(Tys);
8759 unsigned NewOp = IsLaneOp ? AArch64ISD::LD1LANEpost : AArch64ISD::LD1DUPpost;
8760 SDValue UpdN = DAG.getMemIntrinsicNode(NewOp, SDLoc(N), SDTys, Ops,
8762 LoadSDN->getMemOperand());
8765 SmallVector<SDValue, 2> NewResults;
8766 NewResults.push_back(SDValue(LD, 0)); // The result of load
8767 NewResults.push_back(SDValue(UpdN.getNode(), 2)); // Chain
8768 DCI.CombineTo(LD, NewResults);
8769 DCI.CombineTo(N, SDValue(UpdN.getNode(), 0)); // Dup/Inserted Result
8770 DCI.CombineTo(User, SDValue(UpdN.getNode(), 1)); // Write back register
8777 /// Simplify \Addr given that the top byte of it is ignored by HW during
8778 /// address translation.
8779 static bool performTBISimplification(SDValue Addr,
8780 TargetLowering::DAGCombinerInfo &DCI,
8781 SelectionDAG &DAG) {
8782 APInt DemandedMask = APInt::getLowBitsSet(64, 56);
8783 APInt KnownZero, KnownOne;
8784 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
8785 DCI.isBeforeLegalizeOps());
8786 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8787 if (TLI.SimplifyDemandedBits(Addr, DemandedMask, KnownZero, KnownOne, TLO)) {
8788 DCI.CommitTargetLoweringOpt(TLO);
8794 static SDValue performSTORECombine(SDNode *N,
8795 TargetLowering::DAGCombinerInfo &DCI,
8797 const AArch64Subtarget *Subtarget) {
8798 SDValue Split = split16BStores(N, DCI, DAG, Subtarget);
8799 if (Split.getNode())
8802 if (Subtarget->supportsAddressTopByteIgnored() &&
8803 performTBISimplification(N->getOperand(2), DCI, DAG))
8804 return SDValue(N, 0);
8809 /// This function handles the log2-shuffle pattern produced by the
8810 /// LoopVectorizer for the across vector reduction. It consists of
8811 /// log2(NumVectorElements) steps and, in each step, 2^(s) elements
8812 /// are reduced, where s is an induction variable from 0 to
8813 /// log2(NumVectorElements).
8814 static SDValue tryMatchAcrossLaneShuffleForReduction(SDNode *N, SDValue OpV,
8816 SelectionDAG &DAG) {
8817 EVT VTy = OpV->getOperand(0).getValueType();
8818 if (!VTy.isVector())
8821 int NumVecElts = VTy.getVectorNumElements();
8822 if (Op == ISD::FMAXNUM || Op == ISD::FMINNUM) {
8823 if (NumVecElts != 4)
8826 if (NumVecElts != 4 && NumVecElts != 8 && NumVecElts != 16)
8830 int NumExpectedSteps = APInt(8, NumVecElts).logBase2();
8831 SDValue PreOp = OpV;
8832 // Iterate over each step of the across vector reduction.
8833 for (int CurStep = 0; CurStep != NumExpectedSteps; ++CurStep) {
8834 SDValue CurOp = PreOp.getOperand(0);
8835 SDValue Shuffle = PreOp.getOperand(1);
8836 if (Shuffle.getOpcode() != ISD::VECTOR_SHUFFLE) {
8837 // Try to swap the 1st and 2nd operand as add and min/max instructions
8839 CurOp = PreOp.getOperand(1);
8840 Shuffle = PreOp.getOperand(0);
8841 if (Shuffle.getOpcode() != ISD::VECTOR_SHUFFLE)
8845 // Check if the input vector is fed by the operator we want to handle,
8846 // except the last step; the very first input vector is not necessarily
8847 // the same operator we are handling.
8848 if (CurOp.getOpcode() != Op && (CurStep != (NumExpectedSteps - 1)))
8851 // Check if it forms one step of the across vector reduction.
8853 // %cur = add %1, %0
8854 // %shuffle = vector_shuffle %cur, <2, 3, u, u>
8855 // %pre = add %cur, %shuffle
8856 if (Shuffle.getOperand(0) != CurOp)
8859 int NumMaskElts = 1 << CurStep;
8860 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Shuffle)->getMask();
8861 // Check mask values in each step.
8862 // We expect the shuffle mask in each step follows a specific pattern
8863 // denoted here by the <M, U> form, where M is a sequence of integers
8864 // starting from NumMaskElts, increasing by 1, and the number integers
8865 // in M should be NumMaskElts. U is a sequence of UNDEFs and the number
8866 // of undef in U should be NumVecElts - NumMaskElts.
8867 // E.g., for <8 x i16>, mask values in each step should be :
8868 // step 0 : <1,u,u,u,u,u,u,u>
8869 // step 1 : <2,3,u,u,u,u,u,u>
8870 // step 2 : <4,5,6,7,u,u,u,u>
8871 for (int i = 0; i < NumVecElts; ++i)
8872 if ((i < NumMaskElts && Mask[i] != (NumMaskElts + i)) ||
8873 (i >= NumMaskElts && !(Mask[i] < 0)))
8879 bool IsIntrinsic = false;
8883 llvm_unreachable("Unexpected operator for across vector reduction");
8885 Opcode = AArch64ISD::UADDV;
8888 Opcode = AArch64ISD::SMAXV;
8891 Opcode = AArch64ISD::UMAXV;
8894 Opcode = AArch64ISD::SMINV;
8897 Opcode = AArch64ISD::UMINV;
8900 Opcode = Intrinsic::aarch64_neon_fmaxnmv;
8904 Opcode = Intrinsic::aarch64_neon_fminnmv;
8911 ? DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, N->getValueType(0),
8912 DAG.getConstant(Opcode, DL, MVT::i32), PreOp)
8914 ISD::EXTRACT_VECTOR_ELT, DL, N->getValueType(0),
8915 DAG.getNode(Opcode, DL, PreOp.getSimpleValueType(), PreOp),
8916 DAG.getConstant(0, DL, MVT::i64));
8919 /// Target-specific DAG combine for the across vector min/max reductions.
8920 /// This function specifically handles the final clean-up step of the vector
8921 /// min/max reductions produced by the LoopVectorizer. It is the log2-shuffle
8922 /// pattern, which narrows down and finds the final min/max value from all
8923 /// elements of the vector.
8924 /// For example, for a <16 x i8> vector :
8925 /// svn0 = vector_shuffle %0, undef<8,9,10,11,12,13,14,15,u,u,u,u,u,u,u,u>
8926 /// %smax0 = smax %arr, svn0
8927 /// %svn1 = vector_shuffle %smax0, undef<4,5,6,7,u,u,u,u,u,u,u,u,u,u,u,u>
8928 /// %smax1 = smax %smax0, %svn1
8929 /// %svn2 = vector_shuffle %smax1, undef<2,3,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
8930 /// %smax2 = smax %smax1, svn2
8931 /// %svn3 = vector_shuffle %smax2, undef<1,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
8932 /// %sc = setcc %smax2, %svn3, gt
8933 /// %n0 = extract_vector_elt %sc, #0
8934 /// %n1 = extract_vector_elt %smax2, #0
8935 /// %n2 = extract_vector_elt $smax2, #1
8936 /// %result = select %n0, %n1, n2
8939 /// %result = extract_vector_elt %1, 0
8941 performAcrossLaneMinMaxReductionCombine(SDNode *N, SelectionDAG &DAG,
8942 const AArch64Subtarget *Subtarget) {
8943 if (!Subtarget->hasNEON())
8946 SDValue N0 = N->getOperand(0);
8947 SDValue IfTrue = N->getOperand(1);
8948 SDValue IfFalse = N->getOperand(2);
8950 // Check if the SELECT merges up the final result of the min/max
8952 if (N0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
8953 IfTrue.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
8954 IfFalse.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8957 // Expect N0 is fed by SETCC.
8958 SDValue SetCC = N0.getOperand(0);
8959 EVT SetCCVT = SetCC.getValueType();
8960 if (SetCC.getOpcode() != ISD::SETCC || !SetCCVT.isVector() ||
8961 SetCCVT.getVectorElementType() != MVT::i1)
8964 SDValue VectorOp = SetCC.getOperand(0);
8965 unsigned Op = VectorOp->getOpcode();
8966 // Check if the input vector is fed by the operator we want to handle.
8967 if (Op != ISD::SMAX && Op != ISD::UMAX && Op != ISD::SMIN &&
8968 Op != ISD::UMIN && Op != ISD::FMAXNUM && Op != ISD::FMINNUM)
8971 EVT VTy = VectorOp.getValueType();
8972 if (!VTy.isVector())
8975 if (VTy.getSizeInBits() < 64)
8978 EVT EltTy = VTy.getVectorElementType();
8979 if (Op == ISD::FMAXNUM || Op == ISD::FMINNUM) {
8980 if (EltTy != MVT::f32)
8983 if (EltTy != MVT::i32 && EltTy != MVT::i16 && EltTy != MVT::i8)
8987 // Check if extracting from the same vector.
8989 // %sc = setcc %vector, %svn1, gt
8990 // %n0 = extract_vector_elt %sc, #0
8991 // %n1 = extract_vector_elt %vector, #0
8992 // %n2 = extract_vector_elt $vector, #1
8993 if (!(VectorOp == IfTrue->getOperand(0) &&
8994 VectorOp == IfFalse->getOperand(0)))
8997 // Check if the condition code is matched with the operator type.
8998 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC->getOperand(2))->get();
8999 if ((Op == ISD::SMAX && CC != ISD::SETGT && CC != ISD::SETGE) ||
9000 (Op == ISD::UMAX && CC != ISD::SETUGT && CC != ISD::SETUGE) ||
9001 (Op == ISD::SMIN && CC != ISD::SETLT && CC != ISD::SETLE) ||
9002 (Op == ISD::UMIN && CC != ISD::SETULT && CC != ISD::SETULE) ||
9003 (Op == ISD::FMAXNUM && CC != ISD::SETOGT && CC != ISD::SETOGE &&
9004 CC != ISD::SETUGT && CC != ISD::SETUGE && CC != ISD::SETGT &&
9005 CC != ISD::SETGE) ||
9006 (Op == ISD::FMINNUM && CC != ISD::SETOLT && CC != ISD::SETOLE &&
9007 CC != ISD::SETULT && CC != ISD::SETULE && CC != ISD::SETLT &&
9011 // Expect to check only lane 0 from the vector SETCC.
9012 if (!isNullConstant(N0.getOperand(1)))
9015 // Expect to extract the true value from lane 0.
9016 if (!isNullConstant(IfTrue.getOperand(1)))
9019 // Expect to extract the false value from lane 1.
9020 if (!isOneConstant(IfFalse.getOperand(1)))
9023 return tryMatchAcrossLaneShuffleForReduction(N, SetCC, Op, DAG);
9026 /// Target-specific DAG combine for the across vector add reduction.
9027 /// This function specifically handles the final clean-up step of the vector
9028 /// add reduction produced by the LoopVectorizer. It is the log2-shuffle
9029 /// pattern, which adds all elements of a vector together.
9030 /// For example, for a <4 x i32> vector :
9031 /// %1 = vector_shuffle %0, <2,3,u,u>
9033 /// %3 = vector_shuffle %2, <1,u,u,u>
9035 /// %result = extract_vector_elt %4, 0
9038 /// %result = extract_vector_elt %0, 0
9040 performAcrossLaneAddReductionCombine(SDNode *N, SelectionDAG &DAG,
9041 const AArch64Subtarget *Subtarget) {
9042 if (!Subtarget->hasNEON())
9044 SDValue N0 = N->getOperand(0);
9045 SDValue N1 = N->getOperand(1);
9047 // Check if the input vector is fed by the ADD.
9048 if (N0->getOpcode() != ISD::ADD)
9051 // The vector extract idx must constant zero because we only expect the final
9052 // result of the reduction is placed in lane 0.
9053 if (!isNullConstant(N1))
9056 EVT VTy = N0.getValueType();
9057 if (!VTy.isVector())
9060 EVT EltTy = VTy.getVectorElementType();
9061 if (EltTy != MVT::i32 && EltTy != MVT::i16 && EltTy != MVT::i8)
9064 if (VTy.getSizeInBits() < 64)
9067 return tryMatchAcrossLaneShuffleForReduction(N, N0, ISD::ADD, DAG);
9070 /// Target-specific DAG combine function for NEON load/store intrinsics
9071 /// to merge base address updates.
9072 static SDValue performNEONPostLDSTCombine(SDNode *N,
9073 TargetLowering::DAGCombinerInfo &DCI,
9074 SelectionDAG &DAG) {
9075 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9078 unsigned AddrOpIdx = N->getNumOperands() - 1;
9079 SDValue Addr = N->getOperand(AddrOpIdx);
9081 // Search for a use of the address operand that is an increment.
9082 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
9083 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
9085 if (User->getOpcode() != ISD::ADD ||
9086 UI.getUse().getResNo() != Addr.getResNo())
9089 // Check that the add is independent of the load/store. Otherwise, folding
9090 // it would create a cycle.
9091 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
9094 // Find the new opcode for the updating load/store.
9095 bool IsStore = false;
9096 bool IsLaneOp = false;
9097 bool IsDupOp = false;
9098 unsigned NewOpc = 0;
9099 unsigned NumVecs = 0;
9100 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
9102 default: llvm_unreachable("unexpected intrinsic for Neon base update");
9103 case Intrinsic::aarch64_neon_ld2: NewOpc = AArch64ISD::LD2post;
9105 case Intrinsic::aarch64_neon_ld3: NewOpc = AArch64ISD::LD3post;
9107 case Intrinsic::aarch64_neon_ld4: NewOpc = AArch64ISD::LD4post;
9109 case Intrinsic::aarch64_neon_st2: NewOpc = AArch64ISD::ST2post;
9110 NumVecs = 2; IsStore = true; break;
9111 case Intrinsic::aarch64_neon_st3: NewOpc = AArch64ISD::ST3post;
9112 NumVecs = 3; IsStore = true; break;
9113 case Intrinsic::aarch64_neon_st4: NewOpc = AArch64ISD::ST4post;
9114 NumVecs = 4; IsStore = true; break;
9115 case Intrinsic::aarch64_neon_ld1x2: NewOpc = AArch64ISD::LD1x2post;
9117 case Intrinsic::aarch64_neon_ld1x3: NewOpc = AArch64ISD::LD1x3post;
9119 case Intrinsic::aarch64_neon_ld1x4: NewOpc = AArch64ISD::LD1x4post;
9121 case Intrinsic::aarch64_neon_st1x2: NewOpc = AArch64ISD::ST1x2post;
9122 NumVecs = 2; IsStore = true; break;
9123 case Intrinsic::aarch64_neon_st1x3: NewOpc = AArch64ISD::ST1x3post;
9124 NumVecs = 3; IsStore = true; break;
9125 case Intrinsic::aarch64_neon_st1x4: NewOpc = AArch64ISD::ST1x4post;
9126 NumVecs = 4; IsStore = true; break;
9127 case Intrinsic::aarch64_neon_ld2r: NewOpc = AArch64ISD::LD2DUPpost;
9128 NumVecs = 2; IsDupOp = true; break;
9129 case Intrinsic::aarch64_neon_ld3r: NewOpc = AArch64ISD::LD3DUPpost;
9130 NumVecs = 3; IsDupOp = true; break;
9131 case Intrinsic::aarch64_neon_ld4r: NewOpc = AArch64ISD::LD4DUPpost;
9132 NumVecs = 4; IsDupOp = true; break;
9133 case Intrinsic::aarch64_neon_ld2lane: NewOpc = AArch64ISD::LD2LANEpost;
9134 NumVecs = 2; IsLaneOp = true; break;
9135 case Intrinsic::aarch64_neon_ld3lane: NewOpc = AArch64ISD::LD3LANEpost;
9136 NumVecs = 3; IsLaneOp = true; break;
9137 case Intrinsic::aarch64_neon_ld4lane: NewOpc = AArch64ISD::LD4LANEpost;
9138 NumVecs = 4; IsLaneOp = true; break;
9139 case Intrinsic::aarch64_neon_st2lane: NewOpc = AArch64ISD::ST2LANEpost;
9140 NumVecs = 2; IsStore = true; IsLaneOp = true; break;
9141 case Intrinsic::aarch64_neon_st3lane: NewOpc = AArch64ISD::ST3LANEpost;
9142 NumVecs = 3; IsStore = true; IsLaneOp = true; break;
9143 case Intrinsic::aarch64_neon_st4lane: NewOpc = AArch64ISD::ST4LANEpost;
9144 NumVecs = 4; IsStore = true; IsLaneOp = true; break;
9149 VecTy = N->getOperand(2).getValueType();
9151 VecTy = N->getValueType(0);
9153 // If the increment is a constant, it must match the memory ref size.
9154 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
9155 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
9156 uint32_t IncVal = CInc->getZExtValue();
9157 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
9158 if (IsLaneOp || IsDupOp)
9159 NumBytes /= VecTy.getVectorNumElements();
9160 if (IncVal != NumBytes)
9162 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
9164 SmallVector<SDValue, 8> Ops;
9165 Ops.push_back(N->getOperand(0)); // Incoming chain
9166 // Load lane and store have vector list as input.
9167 if (IsLaneOp || IsStore)
9168 for (unsigned i = 2; i < AddrOpIdx; ++i)
9169 Ops.push_back(N->getOperand(i));
9170 Ops.push_back(Addr); // Base register
9175 unsigned NumResultVecs = (IsStore ? 0 : NumVecs);
9177 for (n = 0; n < NumResultVecs; ++n)
9179 Tys[n++] = MVT::i64; // Type of write back register
9180 Tys[n] = MVT::Other; // Type of the chain
9181 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs + 2));
9183 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
9184 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys, Ops,
9185 MemInt->getMemoryVT(),
9186 MemInt->getMemOperand());
9189 std::vector<SDValue> NewResults;
9190 for (unsigned i = 0; i < NumResultVecs; ++i) {
9191 NewResults.push_back(SDValue(UpdN.getNode(), i));
9193 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs + 1));
9194 DCI.CombineTo(N, NewResults);
9195 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9202 // Checks to see if the value is the prescribed width and returns information
9203 // about its extension mode.
9205 bool checkValueWidth(SDValue V, unsigned width, ISD::LoadExtType &ExtType) {
9206 ExtType = ISD::NON_EXTLOAD;
9207 switch(V.getNode()->getOpcode()) {
9211 LoadSDNode *LoadNode = cast<LoadSDNode>(V.getNode());
9212 if ((LoadNode->getMemoryVT() == MVT::i8 && width == 8)
9213 || (LoadNode->getMemoryVT() == MVT::i16 && width == 16)) {
9214 ExtType = LoadNode->getExtensionType();
9219 case ISD::AssertSext: {
9220 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
9221 if ((TypeNode->getVT() == MVT::i8 && width == 8)
9222 || (TypeNode->getVT() == MVT::i16 && width == 16)) {
9223 ExtType = ISD::SEXTLOAD;
9228 case ISD::AssertZext: {
9229 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
9230 if ((TypeNode->getVT() == MVT::i8 && width == 8)
9231 || (TypeNode->getVT() == MVT::i16 && width == 16)) {
9232 ExtType = ISD::ZEXTLOAD;
9238 case ISD::TargetConstant: {
9239 if (std::abs(cast<ConstantSDNode>(V.getNode())->getSExtValue()) <
9249 // This function does a whole lot of voodoo to determine if the tests are
9250 // equivalent without and with a mask. Essentially what happens is that given a
9253 // +-------------+ +-------------+ +-------------+ +-------------+
9254 // | Input | | AddConstant | | CompConstant| | CC |
9255 // +-------------+ +-------------+ +-------------+ +-------------+
9257 // V V | +----------+
9258 // +-------------+ +----+ | |
9259 // | ADD | |0xff| | |
9260 // +-------------+ +----+ | |
9263 // +-------------+ | |
9265 // +-------------+ | |
9274 // The AND node may be safely removed for some combinations of inputs. In
9275 // particular we need to take into account the extension type of the Input,
9276 // the exact values of AddConstant, CompConstant, and CC, along with the nominal
9277 // width of the input (this can work for any width inputs, the above graph is
9278 // specific to 8 bits.
9280 // The specific equations were worked out by generating output tables for each
9281 // AArch64CC value in terms of and AddConstant (w1), CompConstant(w2). The
9282 // problem was simplified by working with 4 bit inputs, which means we only
9283 // needed to reason about 24 distinct bit patterns: 8 patterns unique to zero
9284 // extension (8,15), 8 patterns unique to sign extensions (-8,-1), and 8
9285 // patterns present in both extensions (0,7). For every distinct set of
9286 // AddConstant and CompConstants bit patterns we can consider the masked and
9287 // unmasked versions to be equivalent if the result of this function is true for
9288 // all 16 distinct bit patterns of for the current extension type of Input (w0).
9291 // and w10, w8, #0x0f
9293 // cset w9, AArch64CC
9295 // cset w11, AArch64CC
9300 // Since the above function shows when the outputs are equivalent it defines
9301 // when it is safe to remove the AND. Unfortunately it only runs on AArch64 and
9302 // would be expensive to run during compiles. The equations below were written
9303 // in a test harness that confirmed they gave equivalent outputs to the above
9304 // for all inputs function, so they can be used determine if the removal is
9307 // isEquivalentMaskless() is the code for testing if the AND can be removed
9308 // factored out of the DAG recognition as the DAG can take several forms.
9311 bool isEquivalentMaskless(unsigned CC, unsigned width,
9312 ISD::LoadExtType ExtType, signed AddConstant,
9313 signed CompConstant) {
9314 // By being careful about our equations and only writing the in term
9315 // symbolic values and well known constants (0, 1, -1, MaxUInt) we can
9316 // make them generally applicable to all bit widths.
9317 signed MaxUInt = (1 << width);
9319 // For the purposes of these comparisons sign extending the type is
9320 // equivalent to zero extending the add and displacing it by half the integer
9321 // width. Provided we are careful and make sure our equations are valid over
9322 // the whole range we can just adjust the input and avoid writing equations
9323 // for sign extended inputs.
9324 if (ExtType == ISD::SEXTLOAD)
9325 AddConstant -= (1 << (width-1));
9329 case AArch64CC::GT: {
9330 if ((AddConstant == 0) ||
9331 (CompConstant == MaxUInt - 1 && AddConstant < 0) ||
9332 (AddConstant >= 0 && CompConstant < 0) ||
9333 (AddConstant <= 0 && CompConstant <= 0 && CompConstant < AddConstant))
9337 case AArch64CC::GE: {
9338 if ((AddConstant == 0) ||
9339 (AddConstant >= 0 && CompConstant <= 0) ||
9340 (AddConstant <= 0 && CompConstant <= 0 && CompConstant <= AddConstant))
9344 case AArch64CC::LS: {
9345 if ((AddConstant >= 0 && CompConstant < 0) ||
9346 (AddConstant <= 0 && CompConstant >= -1 &&
9347 CompConstant < AddConstant + MaxUInt))
9351 case AArch64CC::MI: {
9352 if ((AddConstant == 0) ||
9353 (AddConstant > 0 && CompConstant <= 0) ||
9354 (AddConstant < 0 && CompConstant <= AddConstant))
9358 case AArch64CC::HS: {
9359 if ((AddConstant >= 0 && CompConstant <= 0) ||
9360 (AddConstant <= 0 && CompConstant >= 0 &&
9361 CompConstant <= AddConstant + MaxUInt))
9365 case AArch64CC::NE: {
9366 if ((AddConstant > 0 && CompConstant < 0) ||
9367 (AddConstant < 0 && CompConstant >= 0 &&
9368 CompConstant < AddConstant + MaxUInt) ||
9369 (AddConstant >= 0 && CompConstant >= 0 &&
9370 CompConstant >= AddConstant) ||
9371 (AddConstant <= 0 && CompConstant < 0 && CompConstant < AddConstant))
9380 case AArch64CC::Invalid:
9388 SDValue performCONDCombine(SDNode *N,
9389 TargetLowering::DAGCombinerInfo &DCI,
9390 SelectionDAG &DAG, unsigned CCIndex,
9391 unsigned CmpIndex) {
9392 unsigned CC = cast<ConstantSDNode>(N->getOperand(CCIndex))->getSExtValue();
9393 SDNode *SubsNode = N->getOperand(CmpIndex).getNode();
9394 unsigned CondOpcode = SubsNode->getOpcode();
9396 if (CondOpcode != AArch64ISD::SUBS)
9399 // There is a SUBS feeding this condition. Is it fed by a mask we can
9402 SDNode *AndNode = SubsNode->getOperand(0).getNode();
9403 unsigned MaskBits = 0;
9405 if (AndNode->getOpcode() != ISD::AND)
9408 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(AndNode->getOperand(1))) {
9409 uint32_t CNV = CN->getZExtValue();
9412 else if (CNV == 65535)
9419 SDValue AddValue = AndNode->getOperand(0);
9421 if (AddValue.getOpcode() != ISD::ADD)
9424 // The basic dag structure is correct, grab the inputs and validate them.
9426 SDValue AddInputValue1 = AddValue.getNode()->getOperand(0);
9427 SDValue AddInputValue2 = AddValue.getNode()->getOperand(1);
9428 SDValue SubsInputValue = SubsNode->getOperand(1);
9430 // The mask is present and the provenance of all the values is a smaller type,
9431 // lets see if the mask is superfluous.
9433 if (!isa<ConstantSDNode>(AddInputValue2.getNode()) ||
9434 !isa<ConstantSDNode>(SubsInputValue.getNode()))
9437 ISD::LoadExtType ExtType;
9439 if (!checkValueWidth(SubsInputValue, MaskBits, ExtType) ||
9440 !checkValueWidth(AddInputValue2, MaskBits, ExtType) ||
9441 !checkValueWidth(AddInputValue1, MaskBits, ExtType) )
9444 if(!isEquivalentMaskless(CC, MaskBits, ExtType,
9445 cast<ConstantSDNode>(AddInputValue2.getNode())->getSExtValue(),
9446 cast<ConstantSDNode>(SubsInputValue.getNode())->getSExtValue()))
9449 // The AND is not necessary, remove it.
9451 SDVTList VTs = DAG.getVTList(SubsNode->getValueType(0),
9452 SubsNode->getValueType(1));
9453 SDValue Ops[] = { AddValue, SubsNode->getOperand(1) };
9455 SDValue NewValue = DAG.getNode(CondOpcode, SDLoc(SubsNode), VTs, Ops);
9456 DAG.ReplaceAllUsesWith(SubsNode, NewValue.getNode());
9458 return SDValue(N, 0);
9461 // Optimize compare with zero and branch.
9462 static SDValue performBRCONDCombine(SDNode *N,
9463 TargetLowering::DAGCombinerInfo &DCI,
9464 SelectionDAG &DAG) {
9465 SDValue NV = performCONDCombine(N, DCI, DAG, 2, 3);
9468 SDValue Chain = N->getOperand(0);
9469 SDValue Dest = N->getOperand(1);
9470 SDValue CCVal = N->getOperand(2);
9471 SDValue Cmp = N->getOperand(3);
9473 assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!");
9474 unsigned CC = cast<ConstantSDNode>(CCVal)->getZExtValue();
9475 if (CC != AArch64CC::EQ && CC != AArch64CC::NE)
9478 unsigned CmpOpc = Cmp.getOpcode();
9479 if (CmpOpc != AArch64ISD::ADDS && CmpOpc != AArch64ISD::SUBS)
9482 // Only attempt folding if there is only one use of the flag and no use of the
9484 if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1))
9487 SDValue LHS = Cmp.getOperand(0);
9488 SDValue RHS = Cmp.getOperand(1);
9490 assert(LHS.getValueType() == RHS.getValueType() &&
9491 "Expected the value type to be the same for both operands!");
9492 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
9495 if (isNullConstant(LHS))
9496 std::swap(LHS, RHS);
9498 if (!isNullConstant(RHS))
9501 if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA ||
9502 LHS.getOpcode() == ISD::SRL)
9505 // Fold the compare into the branch instruction.
9507 if (CC == AArch64CC::EQ)
9508 BR = DAG.getNode(AArch64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
9510 BR = DAG.getNode(AArch64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
9512 // Do not add new nodes to DAG combiner worklist.
9513 DCI.CombineTo(N, BR, false);
9518 // Optimize some simple tbz/tbnz cases. Returns the new operand and bit to test
9519 // as well as whether the test should be inverted. This code is required to
9520 // catch these cases (as opposed to standard dag combines) because
9521 // AArch64ISD::TBZ is matched during legalization.
9522 static SDValue getTestBitOperand(SDValue Op, unsigned &Bit, bool &Invert,
9523 SelectionDAG &DAG) {
9525 if (!Op->hasOneUse())
9528 // We don't handle undef/constant-fold cases below, as they should have
9529 // already been taken care of (e.g. and of 0, test of undefined shifted bits,
9532 // (tbz (trunc x), b) -> (tbz x, b)
9533 // This case is just here to enable more of the below cases to be caught.
9534 if (Op->getOpcode() == ISD::TRUNCATE &&
9535 Bit < Op->getValueType(0).getSizeInBits()) {
9536 return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG);
9539 if (Op->getNumOperands() != 2)
9542 auto *C = dyn_cast<ConstantSDNode>(Op->getOperand(1));
9546 switch (Op->getOpcode()) {
9550 // (tbz (and x, m), b) -> (tbz x, b)
9552 if ((C->getZExtValue() >> Bit) & 1)
9553 return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG);
9556 // (tbz (shl x, c), b) -> (tbz x, b-c)
9558 if (C->getZExtValue() <= Bit &&
9559 (Bit - C->getZExtValue()) < Op->getValueType(0).getSizeInBits()) {
9560 Bit = Bit - C->getZExtValue();
9561 return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG);
9565 // (tbz (sra x, c), b) -> (tbz x, b+c) or (tbz x, msb) if b+c is > # bits in x
9567 Bit = Bit + C->getZExtValue();
9568 if (Bit >= Op->getValueType(0).getSizeInBits())
9569 Bit = Op->getValueType(0).getSizeInBits() - 1;
9570 return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG);
9572 // (tbz (srl x, c), b) -> (tbz x, b+c)
9574 if ((Bit + C->getZExtValue()) < Op->getValueType(0).getSizeInBits()) {
9575 Bit = Bit + C->getZExtValue();
9576 return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG);
9580 // (tbz (xor x, -1), b) -> (tbnz x, b)
9582 if ((C->getZExtValue() >> Bit) & 1)
9584 return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG);
9588 // Optimize test single bit zero/non-zero and branch.
9589 static SDValue performTBZCombine(SDNode *N,
9590 TargetLowering::DAGCombinerInfo &DCI,
9591 SelectionDAG &DAG) {
9592 unsigned Bit = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
9593 bool Invert = false;
9594 SDValue TestSrc = N->getOperand(1);
9595 SDValue NewTestSrc = getTestBitOperand(TestSrc, Bit, Invert, DAG);
9597 if (TestSrc == NewTestSrc)
9600 unsigned NewOpc = N->getOpcode();
9602 if (NewOpc == AArch64ISD::TBZ)
9603 NewOpc = AArch64ISD::TBNZ;
9605 assert(NewOpc == AArch64ISD::TBNZ);
9606 NewOpc = AArch64ISD::TBZ;
9611 return DAG.getNode(NewOpc, DL, MVT::Other, N->getOperand(0), NewTestSrc,
9612 DAG.getConstant(Bit, DL, MVT::i64), N->getOperand(3));
9615 // vselect (v1i1 setcc) ->
9616 // vselect (v1iXX setcc) (XX is the size of the compared operand type)
9617 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
9618 // condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
9620 static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) {
9621 SDValue N0 = N->getOperand(0);
9622 EVT CCVT = N0.getValueType();
9624 if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
9625 CCVT.getVectorElementType() != MVT::i1)
9628 EVT ResVT = N->getValueType(0);
9629 EVT CmpVT = N0.getOperand(0).getValueType();
9630 // Only combine when the result type is of the same size as the compared
9632 if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
9635 SDValue IfTrue = N->getOperand(1);
9636 SDValue IfFalse = N->getOperand(2);
9638 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
9639 N0.getOperand(0), N0.getOperand(1),
9640 cast<CondCodeSDNode>(N0.getOperand(2))->get());
9641 return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
9645 /// A vector select: "(select vL, vR, (setcc LHS, RHS))" is best performed with
9646 /// the compare-mask instructions rather than going via NZCV, even if LHS and
9647 /// RHS are really scalar. This replaces any scalar setcc in the above pattern
9648 /// with a vector one followed by a DUP shuffle on the result.
9649 static SDValue performSelectCombine(SDNode *N,
9650 TargetLowering::DAGCombinerInfo &DCI) {
9651 SelectionDAG &DAG = DCI.DAG;
9652 SDValue N0 = N->getOperand(0);
9653 EVT ResVT = N->getValueType(0);
9655 if (N0.getOpcode() != ISD::SETCC)
9658 // Make sure the SETCC result is either i1 (initial DAG), or i32, the lowered
9659 // scalar SetCCResultType. We also don't expect vectors, because we assume
9660 // that selects fed by vector SETCCs are canonicalized to VSELECT.
9661 assert((N0.getValueType() == MVT::i1 || N0.getValueType() == MVT::i32) &&
9662 "Scalar-SETCC feeding SELECT has unexpected result type!");
9664 // If NumMaskElts == 0, the comparison is larger than select result. The
9665 // largest real NEON comparison is 64-bits per lane, which means the result is
9666 // at most 32-bits and an illegal vector. Just bail out for now.
9667 EVT SrcVT = N0.getOperand(0).getValueType();
9669 // Don't try to do this optimization when the setcc itself has i1 operands.
9670 // There are no legal vectors of i1, so this would be pointless.
9671 if (SrcVT == MVT::i1)
9674 int NumMaskElts = ResVT.getSizeInBits() / SrcVT.getSizeInBits();
9675 if (!ResVT.isVector() || NumMaskElts == 0)
9678 SrcVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumMaskElts);
9679 EVT CCVT = SrcVT.changeVectorElementTypeToInteger();
9681 // Also bail out if the vector CCVT isn't the same size as ResVT.
9682 // This can happen if the SETCC operand size doesn't divide the ResVT size
9683 // (e.g., f64 vs v3f32).
9684 if (CCVT.getSizeInBits() != ResVT.getSizeInBits())
9687 // Make sure we didn't create illegal types, if we're not supposed to.
9688 assert(DCI.isBeforeLegalize() ||
9689 DAG.getTargetLoweringInfo().isTypeLegal(SrcVT));
9691 // First perform a vector comparison, where lane 0 is the one we're interested
9695 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(0));
9697 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(1));
9698 SDValue SetCC = DAG.getNode(ISD::SETCC, DL, CCVT, LHS, RHS, N0.getOperand(2));
9700 // Now duplicate the comparison mask we want across all other lanes.
9701 SmallVector<int, 8> DUPMask(CCVT.getVectorNumElements(), 0);
9702 SDValue Mask = DAG.getVectorShuffle(CCVT, DL, SetCC, SetCC, DUPMask.data());
9703 Mask = DAG.getNode(ISD::BITCAST, DL,
9704 ResVT.changeVectorElementTypeToInteger(), Mask);
9706 return DAG.getSelect(DL, ResVT, Mask, N->getOperand(1), N->getOperand(2));
9709 /// Get rid of unnecessary NVCASTs (that don't change the type).
9710 static SDValue performNVCASTCombine(SDNode *N) {
9711 if (N->getValueType(0) == N->getOperand(0).getValueType())
9712 return N->getOperand(0);
9717 SDValue AArch64TargetLowering::PerformDAGCombine(SDNode *N,
9718 DAGCombinerInfo &DCI) const {
9719 SelectionDAG &DAG = DCI.DAG;
9720 switch (N->getOpcode()) {
9725 return performAddSubLongCombine(N, DCI, DAG);
9727 return performXorCombine(N, DAG, DCI, Subtarget);
9729 return performMulCombine(N, DAG, DCI, Subtarget);
9730 case ISD::SINT_TO_FP:
9731 case ISD::UINT_TO_FP:
9732 return performIntToFpCombine(N, DAG, Subtarget);
9733 case ISD::FP_TO_SINT:
9734 case ISD::FP_TO_UINT:
9735 return performFpToIntCombine(N, DAG, DCI, Subtarget);
9737 return performFDivCombine(N, DAG, Subtarget);
9739 return performORCombine(N, DCI, Subtarget);
9740 case ISD::INTRINSIC_WO_CHAIN:
9741 return performIntrinsicCombine(N, DCI, Subtarget);
9742 case ISD::ANY_EXTEND:
9743 case ISD::ZERO_EXTEND:
9744 case ISD::SIGN_EXTEND:
9745 return performExtendCombine(N, DCI, DAG);
9747 return performBitcastCombine(N, DCI, DAG);
9748 case ISD::CONCAT_VECTORS:
9749 return performConcatVectorsCombine(N, DCI, DAG);
9751 SDValue RV = performSelectCombine(N, DCI);
9753 RV = performAcrossLaneMinMaxReductionCombine(N, DAG, Subtarget);
9757 return performVSelectCombine(N, DCI.DAG);
9759 if (performTBISimplification(N->getOperand(1), DCI, DAG))
9760 return SDValue(N, 0);
9763 return performSTORECombine(N, DCI, DAG, Subtarget);
9764 case AArch64ISD::BRCOND:
9765 return performBRCONDCombine(N, DCI, DAG);
9766 case AArch64ISD::TBNZ:
9767 case AArch64ISD::TBZ:
9768 return performTBZCombine(N, DCI, DAG);
9769 case AArch64ISD::CSEL:
9770 return performCONDCombine(N, DCI, DAG, 2, 3);
9771 case AArch64ISD::DUP:
9772 return performPostLD1Combine(N, DCI, false);
9773 case AArch64ISD::NVCAST:
9774 return performNVCASTCombine(N);
9775 case ISD::INSERT_VECTOR_ELT:
9776 return performPostLD1Combine(N, DCI, true);
9777 case ISD::EXTRACT_VECTOR_ELT:
9778 return performAcrossLaneAddReductionCombine(N, DAG, Subtarget);
9779 case ISD::INTRINSIC_VOID:
9780 case ISD::INTRINSIC_W_CHAIN:
9781 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9782 case Intrinsic::aarch64_neon_ld2:
9783 case Intrinsic::aarch64_neon_ld3:
9784 case Intrinsic::aarch64_neon_ld4:
9785 case Intrinsic::aarch64_neon_ld1x2:
9786 case Intrinsic::aarch64_neon_ld1x3:
9787 case Intrinsic::aarch64_neon_ld1x4:
9788 case Intrinsic::aarch64_neon_ld2lane:
9789 case Intrinsic::aarch64_neon_ld3lane:
9790 case Intrinsic::aarch64_neon_ld4lane:
9791 case Intrinsic::aarch64_neon_ld2r:
9792 case Intrinsic::aarch64_neon_ld3r:
9793 case Intrinsic::aarch64_neon_ld4r:
9794 case Intrinsic::aarch64_neon_st2:
9795 case Intrinsic::aarch64_neon_st3:
9796 case Intrinsic::aarch64_neon_st4:
9797 case Intrinsic::aarch64_neon_st1x2:
9798 case Intrinsic::aarch64_neon_st1x3:
9799 case Intrinsic::aarch64_neon_st1x4:
9800 case Intrinsic::aarch64_neon_st2lane:
9801 case Intrinsic::aarch64_neon_st3lane:
9802 case Intrinsic::aarch64_neon_st4lane:
9803 return performNEONPostLDSTCombine(N, DCI, DAG);
9811 // Check if the return value is used as only a return value, as otherwise
9812 // we can't perform a tail-call. In particular, we need to check for
9813 // target ISD nodes that are returns and any other "odd" constructs
9814 // that the generic analysis code won't necessarily catch.
9815 bool AArch64TargetLowering::isUsedByReturnOnly(SDNode *N,
9816 SDValue &Chain) const {
9817 if (N->getNumValues() != 1)
9819 if (!N->hasNUsesOfValue(1, 0))
9822 SDValue TCChain = Chain;
9823 SDNode *Copy = *N->use_begin();
9824 if (Copy->getOpcode() == ISD::CopyToReg) {
9825 // If the copy has a glue operand, we conservatively assume it isn't safe to
9826 // perform a tail call.
9827 if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() ==
9830 TCChain = Copy->getOperand(0);
9831 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
9834 bool HasRet = false;
9835 for (SDNode *Node : Copy->uses()) {
9836 if (Node->getOpcode() != AArch64ISD::RET_FLAG)
9848 // Return whether the an instruction can potentially be optimized to a tail
9849 // call. This will cause the optimizers to attempt to move, or duplicate,
9850 // return instructions to help enable tail call optimizations for this
9852 bool AArch64TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
9853 if (!CI->isTailCall())
9859 bool AArch64TargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base,
9861 ISD::MemIndexedMode &AM,
9863 SelectionDAG &DAG) const {
9864 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
9867 Base = Op->getOperand(0);
9868 // All of the indexed addressing mode instructions take a signed
9869 // 9 bit immediate offset.
9870 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
9871 int64_t RHSC = (int64_t)RHS->getZExtValue();
9872 if (RHSC >= 256 || RHSC <= -256)
9874 IsInc = (Op->getOpcode() == ISD::ADD);
9875 Offset = Op->getOperand(1);
9881 bool AArch64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
9883 ISD::MemIndexedMode &AM,
9884 SelectionDAG &DAG) const {
9887 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9888 VT = LD->getMemoryVT();
9889 Ptr = LD->getBasePtr();
9890 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9891 VT = ST->getMemoryVT();
9892 Ptr = ST->getBasePtr();
9897 if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, IsInc, DAG))
9899 AM = IsInc ? ISD::PRE_INC : ISD::PRE_DEC;
9903 bool AArch64TargetLowering::getPostIndexedAddressParts(
9904 SDNode *N, SDNode *Op, SDValue &Base, SDValue &Offset,
9905 ISD::MemIndexedMode &AM, SelectionDAG &DAG) const {
9908 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9909 VT = LD->getMemoryVT();
9910 Ptr = LD->getBasePtr();
9911 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9912 VT = ST->getMemoryVT();
9913 Ptr = ST->getBasePtr();
9918 if (!getIndexedAddressParts(Op, Base, Offset, AM, IsInc, DAG))
9920 // Post-indexing updates the base, so it's not a valid transform
9921 // if that's not the same as the load's pointer.
9924 AM = IsInc ? ISD::POST_INC : ISD::POST_DEC;
9928 static void ReplaceBITCASTResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
9929 SelectionDAG &DAG) {
9931 SDValue Op = N->getOperand(0);
9933 if (N->getValueType(0) != MVT::i16 || Op.getValueType() != MVT::f16)
9937 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, DL, MVT::f32,
9938 DAG.getUNDEF(MVT::i32), Op,
9939 DAG.getTargetConstant(AArch64::hsub, DL, MVT::i32)),
9941 Op = DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op);
9942 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, Op));
9945 static void ReplaceReductionResults(SDNode *N,
9946 SmallVectorImpl<SDValue> &Results,
9947 SelectionDAG &DAG, unsigned InterOp,
9948 unsigned AcrossOp) {
9952 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
9953 std::tie(Lo, Hi) = DAG.SplitVectorOperand(N, 0);
9954 SDValue InterVal = DAG.getNode(InterOp, dl, LoVT, Lo, Hi);
9955 SDValue SplitVal = DAG.getNode(AcrossOp, dl, LoVT, InterVal);
9956 Results.push_back(SplitVal);
9959 void AArch64TargetLowering::ReplaceNodeResults(
9960 SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
9961 switch (N->getOpcode()) {
9963 llvm_unreachable("Don't know how to custom expand this");
9965 ReplaceBITCASTResults(N, Results, DAG);
9967 case AArch64ISD::SADDV:
9968 ReplaceReductionResults(N, Results, DAG, ISD::ADD, AArch64ISD::SADDV);
9970 case AArch64ISD::UADDV:
9971 ReplaceReductionResults(N, Results, DAG, ISD::ADD, AArch64ISD::UADDV);
9973 case AArch64ISD::SMINV:
9974 ReplaceReductionResults(N, Results, DAG, ISD::SMIN, AArch64ISD::SMINV);
9976 case AArch64ISD::UMINV:
9977 ReplaceReductionResults(N, Results, DAG, ISD::UMIN, AArch64ISD::UMINV);
9979 case AArch64ISD::SMAXV:
9980 ReplaceReductionResults(N, Results, DAG, ISD::SMAX, AArch64ISD::SMAXV);
9982 case AArch64ISD::UMAXV:
9983 ReplaceReductionResults(N, Results, DAG, ISD::UMAX, AArch64ISD::UMAXV);
9985 case ISD::FP_TO_UINT:
9986 case ISD::FP_TO_SINT:
9987 assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion");
9988 // Let normal code take care of it by not adding anything to Results.
9993 bool AArch64TargetLowering::useLoadStackGuardNode() const {
9997 unsigned AArch64TargetLowering::combineRepeatedFPDivisors() const {
9998 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
9999 // reciprocal if there are three or more FDIVs.
10003 TargetLoweringBase::LegalizeTypeAction
10004 AArch64TargetLowering::getPreferredVectorAction(EVT VT) const {
10005 MVT SVT = VT.getSimpleVT();
10006 // During type legalization, we prefer to widen v1i8, v1i16, v1i32 to v8i8,
10007 // v4i16, v2i32 instead of to promote.
10008 if (SVT == MVT::v1i8 || SVT == MVT::v1i16 || SVT == MVT::v1i32
10009 || SVT == MVT::v1f32)
10010 return TypeWidenVector;
10012 return TargetLoweringBase::getPreferredVectorAction(VT);
10015 // Loads and stores less than 128-bits are already atomic; ones above that
10016 // are doomed anyway, so defer to the default libcall and blame the OS when
10017 // things go wrong.
10018 bool AArch64TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
10019 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
10020 return Size == 128;
10023 // Loads and stores less than 128-bits are already atomic; ones above that
10024 // are doomed anyway, so defer to the default libcall and blame the OS when
10025 // things go wrong.
10026 TargetLowering::AtomicExpansionKind
10027 AArch64TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
10028 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
10029 return Size == 128 ? AtomicExpansionKind::LLSC : AtomicExpansionKind::None;
10032 // For the real atomic operations, we have ldxr/stxr up to 128 bits,
10033 TargetLowering::AtomicExpansionKind
10034 AArch64TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
10035 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
10036 return Size <= 128 ? AtomicExpansionKind::LLSC : AtomicExpansionKind::None;
10039 bool AArch64TargetLowering::shouldExpandAtomicCmpXchgInIR(
10040 AtomicCmpXchgInst *AI) const {
10044 Value *AArch64TargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
10045 AtomicOrdering Ord) const {
10046 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
10047 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
10048 bool IsAcquire = isAtLeastAcquire(Ord);
10050 // Since i128 isn't legal and intrinsics don't get type-lowered, the ldrexd
10051 // intrinsic must return {i64, i64} and we have to recombine them into a
10052 // single i128 here.
10053 if (ValTy->getPrimitiveSizeInBits() == 128) {
10054 Intrinsic::ID Int =
10055 IsAcquire ? Intrinsic::aarch64_ldaxp : Intrinsic::aarch64_ldxp;
10056 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int);
10058 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
10059 Value *LoHi = Builder.CreateCall(Ldxr, Addr, "lohi");
10061 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
10062 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
10063 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
10064 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
10065 return Builder.CreateOr(
10066 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 64)), "val64");
10069 Type *Tys[] = { Addr->getType() };
10070 Intrinsic::ID Int =
10071 IsAcquire ? Intrinsic::aarch64_ldaxr : Intrinsic::aarch64_ldxr;
10072 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int, Tys);
10074 return Builder.CreateTruncOrBitCast(
10075 Builder.CreateCall(Ldxr, Addr),
10076 cast<PointerType>(Addr->getType())->getElementType());
10079 void AArch64TargetLowering::emitAtomicCmpXchgNoStoreLLBalance(
10080 IRBuilder<> &Builder) const {
10081 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
10082 Builder.CreateCall(
10083 llvm::Intrinsic::getDeclaration(M, Intrinsic::aarch64_clrex));
10086 Value *AArch64TargetLowering::emitStoreConditional(IRBuilder<> &Builder,
10087 Value *Val, Value *Addr,
10088 AtomicOrdering Ord) const {
10089 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
10090 bool IsRelease = isAtLeastRelease(Ord);
10092 // Since the intrinsics must have legal type, the i128 intrinsics take two
10093 // parameters: "i64, i64". We must marshal Val into the appropriate form
10094 // before the call.
10095 if (Val->getType()->getPrimitiveSizeInBits() == 128) {
10096 Intrinsic::ID Int =
10097 IsRelease ? Intrinsic::aarch64_stlxp : Intrinsic::aarch64_stxp;
10098 Function *Stxr = Intrinsic::getDeclaration(M, Int);
10099 Type *Int64Ty = Type::getInt64Ty(M->getContext());
10101 Value *Lo = Builder.CreateTrunc(Val, Int64Ty, "lo");
10102 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 64), Int64Ty, "hi");
10103 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
10104 return Builder.CreateCall(Stxr, {Lo, Hi, Addr});
10107 Intrinsic::ID Int =
10108 IsRelease ? Intrinsic::aarch64_stlxr : Intrinsic::aarch64_stxr;
10109 Type *Tys[] = { Addr->getType() };
10110 Function *Stxr = Intrinsic::getDeclaration(M, Int, Tys);
10112 return Builder.CreateCall(Stxr,
10113 {Builder.CreateZExtOrBitCast(
10114 Val, Stxr->getFunctionType()->getParamType(0)),
10118 bool AArch64TargetLowering::functionArgumentNeedsConsecutiveRegisters(
10119 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
10120 return Ty->isArrayTy();
10123 bool AArch64TargetLowering::shouldNormalizeToSelectSequence(LLVMContext &,
10128 Value *AArch64TargetLowering::getSafeStackPointerLocation(IRBuilder<> &IRB) const {
10129 if (!Subtarget->isTargetAndroid())
10130 return TargetLowering::getSafeStackPointerLocation(IRB);
10132 // Android provides a fixed TLS slot for the SafeStack pointer. See the
10133 // definition of TLS_SLOT_SAFESTACK in
10134 // https://android.googlesource.com/platform/bionic/+/master/libc/private/bionic_tls.h
10135 const unsigned TlsOffset = 0x48;
10136 Module *M = IRB.GetInsertBlock()->getParent()->getParent();
10137 Function *ThreadPointerFunc =
10138 Intrinsic::getDeclaration(M, Intrinsic::aarch64_thread_pointer);
10139 return IRB.CreatePointerCast(
10140 IRB.CreateConstGEP1_32(IRB.CreateCall(ThreadPointerFunc), TlsOffset),
10141 Type::getInt8PtrTy(IRB.getContext())->getPointerTo(0));
10144 void AArch64TargetLowering::initializeSplitCSR(MachineBasicBlock *Entry) const {
10145 // Update IsSplitCSR in AArch64unctionInfo.
10146 AArch64FunctionInfo *AFI = Entry->getParent()->getInfo<AArch64FunctionInfo>();
10147 AFI->setIsSplitCSR(true);
10150 void AArch64TargetLowering::insertCopiesSplitCSR(
10151 MachineBasicBlock *Entry,
10152 const SmallVectorImpl<MachineBasicBlock *> &Exits) const {
10153 const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();
10154 const MCPhysReg *IStart = TRI->getCalleeSavedRegsViaCopy(Entry->getParent());
10158 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
10159 MachineRegisterInfo *MRI = &Entry->getParent()->getRegInfo();
10160 MachineBasicBlock::iterator MBBI = Entry->begin();
10161 for (const MCPhysReg *I = IStart; *I; ++I) {
10162 const TargetRegisterClass *RC = nullptr;
10163 if (AArch64::GPR64RegClass.contains(*I))
10164 RC = &AArch64::GPR64RegClass;
10165 else if (AArch64::FPR64RegClass.contains(*I))
10166 RC = &AArch64::FPR64RegClass;
10168 llvm_unreachable("Unexpected register class in CSRsViaCopy!");
10170 unsigned NewVR = MRI->createVirtualRegister(RC);
10171 // Create copy from CSR to a virtual register.
10172 // FIXME: this currently does not emit CFI pseudo-instructions, it works
10173 // fine for CXX_FAST_TLS since the C++-style TLS access functions should be
10174 // nounwind. If we want to generalize this later, we may need to emit
10175 // CFI pseudo-instructions.
10176 assert(Entry->getParent()->getFunction()->hasFnAttribute(
10177 Attribute::NoUnwind) &&
10178 "Function should be nounwind in insertCopiesSplitCSR!");
10179 Entry->addLiveIn(*I);
10180 BuildMI(*Entry, MBBI, DebugLoc(), TII->get(TargetOpcode::COPY), NewVR)
10183 // Insert the copy-back instructions right before the terminator.
10184 for (auto *Exit : Exits)
10185 BuildMI(*Exit, Exit->getFirstTerminator(), DebugLoc(),
10186 TII->get(TargetOpcode::COPY), *I)