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");
50 static cl::opt<AlignMode>
51 Align(cl::desc("Load/store alignment support"),
52 cl::Hidden, cl::init(NoStrictAlign),
54 clEnumValN(StrictAlign, "aarch64-strict-align",
55 "Disallow all unaligned memory accesses"),
56 clEnumValN(NoStrictAlign, "aarch64-no-strict-align",
57 "Allow unaligned memory accesses"),
60 // Place holder until extr generation is tested fully.
62 EnableAArch64ExtrGeneration("aarch64-extr-generation", cl::Hidden,
63 cl::desc("Allow AArch64 (or (shift)(shift))->extract"),
67 EnableAArch64SlrGeneration("aarch64-shift-insert-generation", cl::Hidden,
68 cl::desc("Allow AArch64 SLI/SRI formation"),
71 // FIXME: The necessary dtprel relocations don't seem to be supported
72 // well in the GNU bfd and gold linkers at the moment. Therefore, by
73 // default, for now, fall back to GeneralDynamic code generation.
74 cl::opt<bool> EnableAArch64ELFLocalDynamicTLSGeneration(
75 "aarch64-elf-ldtls-generation", cl::Hidden,
76 cl::desc("Allow AArch64 Local Dynamic TLS code generation"),
79 AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
80 const AArch64Subtarget &STI)
81 : TargetLowering(TM), Subtarget(&STI) {
83 // AArch64 doesn't have comparisons which set GPRs or setcc instructions, so
84 // we have to make something up. Arbitrarily, choose ZeroOrOne.
85 setBooleanContents(ZeroOrOneBooleanContent);
86 // When comparing vectors the result sets the different elements in the
87 // vector to all-one or all-zero.
88 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
90 // Set up the register classes.
91 addRegisterClass(MVT::i32, &AArch64::GPR32allRegClass);
92 addRegisterClass(MVT::i64, &AArch64::GPR64allRegClass);
94 if (Subtarget->hasFPARMv8()) {
95 addRegisterClass(MVT::f16, &AArch64::FPR16RegClass);
96 addRegisterClass(MVT::f32, &AArch64::FPR32RegClass);
97 addRegisterClass(MVT::f64, &AArch64::FPR64RegClass);
98 addRegisterClass(MVT::f128, &AArch64::FPR128RegClass);
101 if (Subtarget->hasNEON()) {
102 addRegisterClass(MVT::v16i8, &AArch64::FPR8RegClass);
103 addRegisterClass(MVT::v8i16, &AArch64::FPR16RegClass);
104 // Someone set us up the NEON.
105 addDRTypeForNEON(MVT::v2f32);
106 addDRTypeForNEON(MVT::v8i8);
107 addDRTypeForNEON(MVT::v4i16);
108 addDRTypeForNEON(MVT::v2i32);
109 addDRTypeForNEON(MVT::v1i64);
110 addDRTypeForNEON(MVT::v1f64);
111 addDRTypeForNEON(MVT::v4f16);
113 addQRTypeForNEON(MVT::v4f32);
114 addQRTypeForNEON(MVT::v2f64);
115 addQRTypeForNEON(MVT::v16i8);
116 addQRTypeForNEON(MVT::v8i16);
117 addQRTypeForNEON(MVT::v4i32);
118 addQRTypeForNEON(MVT::v2i64);
119 addQRTypeForNEON(MVT::v8f16);
122 // Compute derived properties from the register classes
123 computeRegisterProperties(Subtarget->getRegisterInfo());
125 // Provide all sorts of operation actions
126 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
127 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
128 setOperationAction(ISD::SETCC, MVT::i32, Custom);
129 setOperationAction(ISD::SETCC, MVT::i64, Custom);
130 setOperationAction(ISD::SETCC, MVT::f32, Custom);
131 setOperationAction(ISD::SETCC, MVT::f64, Custom);
132 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
133 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
134 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
135 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
136 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
137 setOperationAction(ISD::SELECT, MVT::i32, Custom);
138 setOperationAction(ISD::SELECT, MVT::i64, Custom);
139 setOperationAction(ISD::SELECT, MVT::f32, Custom);
140 setOperationAction(ISD::SELECT, MVT::f64, Custom);
141 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
142 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
143 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
144 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
145 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
146 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
148 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
149 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
150 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
152 setOperationAction(ISD::FREM, MVT::f32, Expand);
153 setOperationAction(ISD::FREM, MVT::f64, Expand);
154 setOperationAction(ISD::FREM, MVT::f80, Expand);
156 // Custom lowering hooks are needed for XOR
157 // to fold it into CSINC/CSINV.
158 setOperationAction(ISD::XOR, MVT::i32, Custom);
159 setOperationAction(ISD::XOR, MVT::i64, Custom);
161 // Virtually no operation on f128 is legal, but LLVM can't expand them when
162 // there's a valid register class, so we need custom operations in most cases.
163 setOperationAction(ISD::FABS, MVT::f128, Expand);
164 setOperationAction(ISD::FADD, MVT::f128, Custom);
165 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
166 setOperationAction(ISD::FCOS, MVT::f128, Expand);
167 setOperationAction(ISD::FDIV, MVT::f128, Custom);
168 setOperationAction(ISD::FMA, MVT::f128, Expand);
169 setOperationAction(ISD::FMUL, MVT::f128, Custom);
170 setOperationAction(ISD::FNEG, MVT::f128, Expand);
171 setOperationAction(ISD::FPOW, MVT::f128, Expand);
172 setOperationAction(ISD::FREM, MVT::f128, Expand);
173 setOperationAction(ISD::FRINT, MVT::f128, Expand);
174 setOperationAction(ISD::FSIN, MVT::f128, Expand);
175 setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
176 setOperationAction(ISD::FSQRT, MVT::f128, Expand);
177 setOperationAction(ISD::FSUB, MVT::f128, Custom);
178 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
179 setOperationAction(ISD::SETCC, MVT::f128, Custom);
180 setOperationAction(ISD::BR_CC, MVT::f128, Custom);
181 setOperationAction(ISD::SELECT, MVT::f128, Custom);
182 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
183 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);
185 // Lowering for many of the conversions is actually specified by the non-f128
186 // type. The LowerXXX function will be trivial when f128 isn't involved.
187 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
188 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
189 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
190 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
191 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
192 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
193 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
194 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
195 setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
196 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
197 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
198 setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
199 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
200 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
202 // Variable arguments.
203 setOperationAction(ISD::VASTART, MVT::Other, Custom);
204 setOperationAction(ISD::VAARG, MVT::Other, Custom);
205 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
206 setOperationAction(ISD::VAEND, MVT::Other, Expand);
208 // Variable-sized objects.
209 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
210 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
211 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
213 // Exception handling.
214 // FIXME: These are guesses. Has this been defined yet?
215 setExceptionPointerRegister(AArch64::X0);
216 setExceptionSelectorRegister(AArch64::X1);
218 // Constant pool entries
219 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
222 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
224 // Add/Sub overflow ops with MVT::Glues are lowered to NZCV dependences.
225 setOperationAction(ISD::ADDC, MVT::i32, Custom);
226 setOperationAction(ISD::ADDE, MVT::i32, Custom);
227 setOperationAction(ISD::SUBC, MVT::i32, Custom);
228 setOperationAction(ISD::SUBE, MVT::i32, Custom);
229 setOperationAction(ISD::ADDC, MVT::i64, Custom);
230 setOperationAction(ISD::ADDE, MVT::i64, Custom);
231 setOperationAction(ISD::SUBC, MVT::i64, Custom);
232 setOperationAction(ISD::SUBE, MVT::i64, Custom);
234 // AArch64 lacks both left-rotate and popcount instructions.
235 setOperationAction(ISD::ROTL, MVT::i32, Expand);
236 setOperationAction(ISD::ROTL, MVT::i64, Expand);
238 // AArch64 doesn't have {U|S}MUL_LOHI.
239 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
240 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
243 // Expand the undefined-at-zero variants to cttz/ctlz to their defined-at-zero
244 // counterparts, which AArch64 supports directly.
245 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
246 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
247 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
248 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
250 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
251 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
253 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
254 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
255 setOperationAction(ISD::SREM, MVT::i32, Expand);
256 setOperationAction(ISD::SREM, MVT::i64, Expand);
257 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
258 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
259 setOperationAction(ISD::UREM, MVT::i32, Expand);
260 setOperationAction(ISD::UREM, MVT::i64, Expand);
262 // Custom lower Add/Sub/Mul with overflow.
263 setOperationAction(ISD::SADDO, MVT::i32, Custom);
264 setOperationAction(ISD::SADDO, MVT::i64, Custom);
265 setOperationAction(ISD::UADDO, MVT::i32, Custom);
266 setOperationAction(ISD::UADDO, MVT::i64, Custom);
267 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
268 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
269 setOperationAction(ISD::USUBO, MVT::i32, Custom);
270 setOperationAction(ISD::USUBO, MVT::i64, Custom);
271 setOperationAction(ISD::SMULO, MVT::i32, Custom);
272 setOperationAction(ISD::SMULO, MVT::i64, Custom);
273 setOperationAction(ISD::UMULO, MVT::i32, Custom);
274 setOperationAction(ISD::UMULO, MVT::i64, Custom);
276 setOperationAction(ISD::FSIN, MVT::f32, Expand);
277 setOperationAction(ISD::FSIN, MVT::f64, Expand);
278 setOperationAction(ISD::FCOS, MVT::f32, Expand);
279 setOperationAction(ISD::FCOS, MVT::f64, Expand);
280 setOperationAction(ISD::FPOW, MVT::f32, Expand);
281 setOperationAction(ISD::FPOW, MVT::f64, Expand);
282 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
283 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
285 // f16 is a storage-only type, always promote it to f32.
286 setOperationAction(ISD::SETCC, MVT::f16, Promote);
287 setOperationAction(ISD::BR_CC, MVT::f16, Promote);
288 setOperationAction(ISD::SELECT_CC, MVT::f16, Promote);
289 setOperationAction(ISD::SELECT, MVT::f16, Promote);
290 setOperationAction(ISD::FADD, MVT::f16, Promote);
291 setOperationAction(ISD::FSUB, MVT::f16, Promote);
292 setOperationAction(ISD::FMUL, MVT::f16, Promote);
293 setOperationAction(ISD::FDIV, MVT::f16, Promote);
294 setOperationAction(ISD::FREM, MVT::f16, Promote);
295 setOperationAction(ISD::FMA, MVT::f16, Promote);
296 setOperationAction(ISD::FNEG, MVT::f16, Promote);
297 setOperationAction(ISD::FABS, MVT::f16, Promote);
298 setOperationAction(ISD::FCEIL, MVT::f16, Promote);
299 setOperationAction(ISD::FCOPYSIGN, MVT::f16, Promote);
300 setOperationAction(ISD::FCOS, MVT::f16, Promote);
301 setOperationAction(ISD::FFLOOR, MVT::f16, Promote);
302 setOperationAction(ISD::FNEARBYINT, MVT::f16, Promote);
303 setOperationAction(ISD::FPOW, MVT::f16, Promote);
304 setOperationAction(ISD::FPOWI, MVT::f16, Promote);
305 setOperationAction(ISD::FRINT, MVT::f16, Promote);
306 setOperationAction(ISD::FSIN, MVT::f16, Promote);
307 setOperationAction(ISD::FSINCOS, MVT::f16, Promote);
308 setOperationAction(ISD::FSQRT, MVT::f16, Promote);
309 setOperationAction(ISD::FEXP, MVT::f16, Promote);
310 setOperationAction(ISD::FEXP2, MVT::f16, Promote);
311 setOperationAction(ISD::FLOG, MVT::f16, Promote);
312 setOperationAction(ISD::FLOG2, MVT::f16, Promote);
313 setOperationAction(ISD::FLOG10, MVT::f16, Promote);
314 setOperationAction(ISD::FROUND, MVT::f16, Promote);
315 setOperationAction(ISD::FTRUNC, MVT::f16, Promote);
316 setOperationAction(ISD::FMINNUM, MVT::f16, Promote);
317 setOperationAction(ISD::FMAXNUM, MVT::f16, Promote);
319 // v4f16 is also a storage-only type, so promote it to v4f32 when that is
321 setOperationAction(ISD::FADD, MVT::v4f16, Promote);
322 setOperationAction(ISD::FSUB, MVT::v4f16, Promote);
323 setOperationAction(ISD::FMUL, MVT::v4f16, Promote);
324 setOperationAction(ISD::FDIV, MVT::v4f16, Promote);
325 setOperationAction(ISD::FP_EXTEND, MVT::v4f16, Promote);
326 setOperationAction(ISD::FP_ROUND, MVT::v4f16, Promote);
327 AddPromotedToType(ISD::FADD, MVT::v4f16, MVT::v4f32);
328 AddPromotedToType(ISD::FSUB, MVT::v4f16, MVT::v4f32);
329 AddPromotedToType(ISD::FMUL, MVT::v4f16, MVT::v4f32);
330 AddPromotedToType(ISD::FDIV, MVT::v4f16, MVT::v4f32);
331 AddPromotedToType(ISD::FP_EXTEND, MVT::v4f16, MVT::v4f32);
332 AddPromotedToType(ISD::FP_ROUND, MVT::v4f16, MVT::v4f32);
334 // Expand all other v4f16 operations.
335 // FIXME: We could generate better code by promoting some operations to
337 setOperationAction(ISD::FABS, MVT::v4f16, Expand);
338 setOperationAction(ISD::FCEIL, MVT::v4f16, Expand);
339 setOperationAction(ISD::FCOPYSIGN, MVT::v4f16, Expand);
340 setOperationAction(ISD::FCOS, MVT::v4f16, Expand);
341 setOperationAction(ISD::FFLOOR, MVT::v4f16, Expand);
342 setOperationAction(ISD::FMA, MVT::v4f16, Expand);
343 setOperationAction(ISD::FNEARBYINT, MVT::v4f16, Expand);
344 setOperationAction(ISD::FNEG, MVT::v4f16, Expand);
345 setOperationAction(ISD::FPOW, MVT::v4f16, Expand);
346 setOperationAction(ISD::FPOWI, MVT::v4f16, Expand);
347 setOperationAction(ISD::FREM, MVT::v4f16, Expand);
348 setOperationAction(ISD::FROUND, MVT::v4f16, Expand);
349 setOperationAction(ISD::FRINT, MVT::v4f16, Expand);
350 setOperationAction(ISD::FSIN, MVT::v4f16, Expand);
351 setOperationAction(ISD::FSINCOS, MVT::v4f16, Expand);
352 setOperationAction(ISD::FSQRT, MVT::v4f16, Expand);
353 setOperationAction(ISD::FTRUNC, MVT::v4f16, Expand);
354 setOperationAction(ISD::SETCC, MVT::v4f16, Expand);
355 setOperationAction(ISD::BR_CC, MVT::v4f16, Expand);
356 setOperationAction(ISD::SELECT, MVT::v4f16, Expand);
357 setOperationAction(ISD::SELECT_CC, MVT::v4f16, Expand);
358 setOperationAction(ISD::FEXP, MVT::v4f16, Expand);
359 setOperationAction(ISD::FEXP2, MVT::v4f16, Expand);
360 setOperationAction(ISD::FLOG, MVT::v4f16, Expand);
361 setOperationAction(ISD::FLOG2, MVT::v4f16, Expand);
362 setOperationAction(ISD::FLOG10, MVT::v4f16, Expand);
365 // v8f16 is also a storage-only type, so expand it.
366 setOperationAction(ISD::FABS, MVT::v8f16, Expand);
367 setOperationAction(ISD::FADD, MVT::v8f16, Expand);
368 setOperationAction(ISD::FCEIL, MVT::v8f16, Expand);
369 setOperationAction(ISD::FCOPYSIGN, MVT::v8f16, Expand);
370 setOperationAction(ISD::FCOS, MVT::v8f16, Expand);
371 setOperationAction(ISD::FDIV, MVT::v8f16, Expand);
372 setOperationAction(ISD::FFLOOR, MVT::v8f16, Expand);
373 setOperationAction(ISD::FMA, MVT::v8f16, Expand);
374 setOperationAction(ISD::FMUL, MVT::v8f16, Expand);
375 setOperationAction(ISD::FNEARBYINT, MVT::v8f16, Expand);
376 setOperationAction(ISD::FNEG, MVT::v8f16, Expand);
377 setOperationAction(ISD::FPOW, MVT::v8f16, Expand);
378 setOperationAction(ISD::FPOWI, MVT::v8f16, Expand);
379 setOperationAction(ISD::FREM, MVT::v8f16, Expand);
380 setOperationAction(ISD::FROUND, MVT::v8f16, Expand);
381 setOperationAction(ISD::FRINT, MVT::v8f16, Expand);
382 setOperationAction(ISD::FSIN, MVT::v8f16, Expand);
383 setOperationAction(ISD::FSINCOS, MVT::v8f16, Expand);
384 setOperationAction(ISD::FSQRT, MVT::v8f16, Expand);
385 setOperationAction(ISD::FSUB, MVT::v8f16, Expand);
386 setOperationAction(ISD::FTRUNC, MVT::v8f16, Expand);
387 setOperationAction(ISD::SETCC, MVT::v8f16, Expand);
388 setOperationAction(ISD::BR_CC, MVT::v8f16, Expand);
389 setOperationAction(ISD::SELECT, MVT::v8f16, Expand);
390 setOperationAction(ISD::SELECT_CC, MVT::v8f16, Expand);
391 setOperationAction(ISD::FP_EXTEND, MVT::v8f16, Expand);
392 setOperationAction(ISD::FEXP, MVT::v8f16, Expand);
393 setOperationAction(ISD::FEXP2, MVT::v8f16, Expand);
394 setOperationAction(ISD::FLOG, MVT::v8f16, Expand);
395 setOperationAction(ISD::FLOG2, MVT::v8f16, Expand);
396 setOperationAction(ISD::FLOG10, MVT::v8f16, Expand);
398 // AArch64 has implementations of a lot of rounding-like FP operations.
399 for (MVT Ty : {MVT::f32, MVT::f64}) {
400 setOperationAction(ISD::FFLOOR, Ty, Legal);
401 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
402 setOperationAction(ISD::FCEIL, Ty, Legal);
403 setOperationAction(ISD::FRINT, Ty, Legal);
404 setOperationAction(ISD::FTRUNC, Ty, Legal);
405 setOperationAction(ISD::FROUND, Ty, Legal);
408 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
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 setIndexedStoreAction(im, MVT::i8, Legal);
460 setIndexedStoreAction(im, MVT::i16, Legal);
461 setIndexedStoreAction(im, MVT::i32, Legal);
462 setIndexedStoreAction(im, MVT::i64, Legal);
463 setIndexedStoreAction(im, MVT::f64, Legal);
464 setIndexedStoreAction(im, MVT::f32, Legal);
468 setOperationAction(ISD::TRAP, MVT::Other, Legal);
470 // We combine OR nodes for bitfield operations.
471 setTargetDAGCombine(ISD::OR);
473 // Vector add and sub nodes may conceal a high-half opportunity.
474 // Also, try to fold ADD into CSINC/CSINV..
475 setTargetDAGCombine(ISD::ADD);
476 setTargetDAGCombine(ISD::SUB);
478 setTargetDAGCombine(ISD::XOR);
479 setTargetDAGCombine(ISD::SINT_TO_FP);
480 setTargetDAGCombine(ISD::UINT_TO_FP);
482 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
484 setTargetDAGCombine(ISD::ANY_EXTEND);
485 setTargetDAGCombine(ISD::ZERO_EXTEND);
486 setTargetDAGCombine(ISD::SIGN_EXTEND);
487 setTargetDAGCombine(ISD::BITCAST);
488 setTargetDAGCombine(ISD::CONCAT_VECTORS);
489 setTargetDAGCombine(ISD::STORE);
491 setTargetDAGCombine(ISD::MUL);
493 setTargetDAGCombine(ISD::SELECT);
494 setTargetDAGCombine(ISD::VSELECT);
495 setTargetDAGCombine(ISD::SELECT_CC);
497 setTargetDAGCombine(ISD::INTRINSIC_VOID);
498 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
499 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
501 MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8;
502 MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4;
503 MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 4;
505 setStackPointerRegisterToSaveRestore(AArch64::SP);
507 setSchedulingPreference(Sched::Hybrid);
510 MaskAndBranchFoldingIsLegal = true;
511 EnableExtLdPromotion = true;
513 setMinFunctionAlignment(2);
515 RequireStrictAlign = (Align == StrictAlign);
517 setHasExtractBitsInsn(true);
519 if (Subtarget->hasNEON()) {
520 // FIXME: v1f64 shouldn't be legal if we can avoid it, because it leads to
521 // silliness like this:
522 setOperationAction(ISD::FABS, MVT::v1f64, Expand);
523 setOperationAction(ISD::FADD, MVT::v1f64, Expand);
524 setOperationAction(ISD::FCEIL, MVT::v1f64, Expand);
525 setOperationAction(ISD::FCOPYSIGN, MVT::v1f64, Expand);
526 setOperationAction(ISD::FCOS, MVT::v1f64, Expand);
527 setOperationAction(ISD::FDIV, MVT::v1f64, Expand);
528 setOperationAction(ISD::FFLOOR, MVT::v1f64, Expand);
529 setOperationAction(ISD::FMA, MVT::v1f64, Expand);
530 setOperationAction(ISD::FMUL, MVT::v1f64, Expand);
531 setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Expand);
532 setOperationAction(ISD::FNEG, MVT::v1f64, Expand);
533 setOperationAction(ISD::FPOW, MVT::v1f64, Expand);
534 setOperationAction(ISD::FREM, MVT::v1f64, Expand);
535 setOperationAction(ISD::FROUND, MVT::v1f64, Expand);
536 setOperationAction(ISD::FRINT, MVT::v1f64, Expand);
537 setOperationAction(ISD::FSIN, MVT::v1f64, Expand);
538 setOperationAction(ISD::FSINCOS, MVT::v1f64, Expand);
539 setOperationAction(ISD::FSQRT, MVT::v1f64, Expand);
540 setOperationAction(ISD::FSUB, MVT::v1f64, Expand);
541 setOperationAction(ISD::FTRUNC, MVT::v1f64, Expand);
542 setOperationAction(ISD::SETCC, MVT::v1f64, Expand);
543 setOperationAction(ISD::BR_CC, MVT::v1f64, Expand);
544 setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
545 setOperationAction(ISD::SELECT_CC, MVT::v1f64, Expand);
546 setOperationAction(ISD::FP_EXTEND, MVT::v1f64, Expand);
548 setOperationAction(ISD::FP_TO_SINT, MVT::v1i64, Expand);
549 setOperationAction(ISD::FP_TO_UINT, MVT::v1i64, Expand);
550 setOperationAction(ISD::SINT_TO_FP, MVT::v1i64, Expand);
551 setOperationAction(ISD::UINT_TO_FP, MVT::v1i64, Expand);
552 setOperationAction(ISD::FP_ROUND, MVT::v1f64, Expand);
554 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
556 // AArch64 doesn't have a direct vector ->f32 conversion instructions for
557 // elements smaller than i32, so promote the input to i32 first.
558 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Promote);
559 setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Promote);
560 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Promote);
561 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Promote);
562 // i8 and i16 vector elements also need promotion to i32 for v8i8 or v8i16
563 // -> v8f16 conversions.
564 setOperationAction(ISD::SINT_TO_FP, MVT::v8i8, Promote);
565 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Promote);
566 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
567 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Promote);
568 // Similarly, there is no direct i32 -> f64 vector conversion instruction.
569 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
570 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
571 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
572 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
573 // Or, direct i32 -> f16 vector conversion. Set it so custom, so the
574 // conversion happens in two steps: v4i32 -> v4f32 -> v4f16
575 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Custom);
576 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
578 // AArch64 doesn't have MUL.2d:
579 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
580 // Custom handling for some quad-vector types to detect MULL.
581 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
582 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
583 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
585 setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal);
586 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
587 // Likewise, narrowing and extending vector loads/stores aren't handled
589 for (MVT VT : MVT::vector_valuetypes()) {
590 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
592 setOperationAction(ISD::MULHS, VT, Expand);
593 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
594 setOperationAction(ISD::MULHU, VT, Expand);
595 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
597 setOperationAction(ISD::BSWAP, VT, Expand);
599 for (MVT InnerVT : MVT::vector_valuetypes()) {
600 setTruncStoreAction(VT, InnerVT, Expand);
601 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
602 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
603 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
607 // AArch64 has implementations of a lot of rounding-like FP operations.
608 for (MVT Ty : {MVT::v2f32, MVT::v4f32, MVT::v2f64}) {
609 setOperationAction(ISD::FFLOOR, Ty, Legal);
610 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
611 setOperationAction(ISD::FCEIL, Ty, Legal);
612 setOperationAction(ISD::FRINT, Ty, Legal);
613 setOperationAction(ISD::FTRUNC, Ty, Legal);
614 setOperationAction(ISD::FROUND, Ty, Legal);
618 // Prefer likely predicted branches to selects on out-of-order cores.
619 if (Subtarget->isCortexA57())
620 PredictableSelectIsExpensive = true;
623 void AArch64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) {
624 if (VT == MVT::v2f32 || VT == MVT::v4f16) {
625 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
626 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i32);
628 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
629 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i32);
630 } else if (VT == MVT::v2f64 || VT == MVT::v4f32 || VT == MVT::v8f16) {
631 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
632 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i64);
634 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
635 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i64);
638 // Mark vector float intrinsics as expand.
639 if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) {
640 setOperationAction(ISD::FSIN, VT.getSimpleVT(), Expand);
641 setOperationAction(ISD::FCOS, VT.getSimpleVT(), Expand);
642 setOperationAction(ISD::FPOWI, VT.getSimpleVT(), Expand);
643 setOperationAction(ISD::FPOW, VT.getSimpleVT(), Expand);
644 setOperationAction(ISD::FLOG, VT.getSimpleVT(), Expand);
645 setOperationAction(ISD::FLOG2, VT.getSimpleVT(), Expand);
646 setOperationAction(ISD::FLOG10, VT.getSimpleVT(), Expand);
647 setOperationAction(ISD::FEXP, VT.getSimpleVT(), Expand);
648 setOperationAction(ISD::FEXP2, VT.getSimpleVT(), Expand);
651 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
652 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
653 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
654 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
655 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Custom);
656 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
657 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
658 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
659 setOperationAction(ISD::AND, VT.getSimpleVT(), Custom);
660 setOperationAction(ISD::OR, VT.getSimpleVT(), Custom);
661 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
662 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
664 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
665 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
666 setOperationAction(ISD::VSELECT, VT.getSimpleVT(), Expand);
667 for (MVT InnerVT : MVT::all_valuetypes())
668 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT.getSimpleVT(), Expand);
670 // CNT supports only B element sizes.
671 if (VT != MVT::v8i8 && VT != MVT::v16i8)
672 setOperationAction(ISD::CTPOP, VT.getSimpleVT(), Expand);
674 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
675 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
676 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
677 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
678 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
680 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
681 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
683 // [SU][MIN|MAX] are available for all NEON types apart from i64.
684 if (!VT.isFloatingPoint() &&
685 VT.getSimpleVT() != MVT::v2i64 && VT.getSimpleVT() != MVT::v1i64)
686 for (unsigned Opcode : {ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX})
687 setOperationAction(Opcode, VT.getSimpleVT(), Legal);
689 if (Subtarget->isLittleEndian()) {
690 for (unsigned im = (unsigned)ISD::PRE_INC;
691 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
692 setIndexedLoadAction(im, VT.getSimpleVT(), Legal);
693 setIndexedStoreAction(im, VT.getSimpleVT(), Legal);
698 void AArch64TargetLowering::addDRTypeForNEON(MVT VT) {
699 addRegisterClass(VT, &AArch64::FPR64RegClass);
700 addTypeForNEON(VT, MVT::v2i32);
703 void AArch64TargetLowering::addQRTypeForNEON(MVT VT) {
704 addRegisterClass(VT, &AArch64::FPR128RegClass);
705 addTypeForNEON(VT, MVT::v4i32);
708 EVT AArch64TargetLowering::getSetCCResultType(const DataLayout &, LLVMContext &,
712 return VT.changeVectorElementTypeToInteger();
715 /// computeKnownBitsForTargetNode - Determine which of the bits specified in
716 /// Mask are known to be either zero or one and return them in the
717 /// KnownZero/KnownOne bitsets.
718 void AArch64TargetLowering::computeKnownBitsForTargetNode(
719 const SDValue Op, APInt &KnownZero, APInt &KnownOne,
720 const SelectionDAG &DAG, unsigned Depth) const {
721 switch (Op.getOpcode()) {
724 case AArch64ISD::CSEL: {
725 APInt KnownZero2, KnownOne2;
726 DAG.computeKnownBits(Op->getOperand(0), KnownZero, KnownOne, Depth + 1);
727 DAG.computeKnownBits(Op->getOperand(1), KnownZero2, KnownOne2, Depth + 1);
728 KnownZero &= KnownZero2;
729 KnownOne &= KnownOne2;
732 case ISD::INTRINSIC_W_CHAIN: {
733 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
734 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
737 case Intrinsic::aarch64_ldaxr:
738 case Intrinsic::aarch64_ldxr: {
739 unsigned BitWidth = KnownOne.getBitWidth();
740 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
741 unsigned MemBits = VT.getScalarType().getSizeInBits();
742 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
748 case ISD::INTRINSIC_WO_CHAIN:
749 case ISD::INTRINSIC_VOID: {
750 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
754 case Intrinsic::aarch64_neon_umaxv:
755 case Intrinsic::aarch64_neon_uminv: {
756 // Figure out the datatype of the vector operand. The UMINV instruction
757 // will zero extend the result, so we can mark as known zero all the
758 // bits larger than the element datatype. 32-bit or larget doesn't need
759 // this as those are legal types and will be handled by isel directly.
760 MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
761 unsigned BitWidth = KnownZero.getBitWidth();
762 if (VT == MVT::v8i8 || VT == MVT::v16i8) {
763 assert(BitWidth >= 8 && "Unexpected width!");
764 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8);
766 } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
767 assert(BitWidth >= 16 && "Unexpected width!");
768 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
778 MVT AArch64TargetLowering::getScalarShiftAmountTy(EVT LHSTy) const {
783 AArch64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
784 const TargetLibraryInfo *libInfo) const {
785 return AArch64::createFastISel(funcInfo, libInfo);
788 const char *AArch64TargetLowering::getTargetNodeName(unsigned Opcode) const {
789 switch ((AArch64ISD::NodeType)Opcode) {
790 case AArch64ISD::FIRST_NUMBER: break;
791 case AArch64ISD::CALL: return "AArch64ISD::CALL";
792 case AArch64ISD::ADRP: return "AArch64ISD::ADRP";
793 case AArch64ISD::ADDlow: return "AArch64ISD::ADDlow";
794 case AArch64ISD::LOADgot: return "AArch64ISD::LOADgot";
795 case AArch64ISD::RET_FLAG: return "AArch64ISD::RET_FLAG";
796 case AArch64ISD::BRCOND: return "AArch64ISD::BRCOND";
797 case AArch64ISD::CSEL: return "AArch64ISD::CSEL";
798 case AArch64ISD::FCSEL: return "AArch64ISD::FCSEL";
799 case AArch64ISD::CSINV: return "AArch64ISD::CSINV";
800 case AArch64ISD::CSNEG: return "AArch64ISD::CSNEG";
801 case AArch64ISD::CSINC: return "AArch64ISD::CSINC";
802 case AArch64ISD::THREAD_POINTER: return "AArch64ISD::THREAD_POINTER";
803 case AArch64ISD::TLSDESC_CALLSEQ: return "AArch64ISD::TLSDESC_CALLSEQ";
804 case AArch64ISD::ADC: return "AArch64ISD::ADC";
805 case AArch64ISD::SBC: return "AArch64ISD::SBC";
806 case AArch64ISD::ADDS: return "AArch64ISD::ADDS";
807 case AArch64ISD::SUBS: return "AArch64ISD::SUBS";
808 case AArch64ISD::ADCS: return "AArch64ISD::ADCS";
809 case AArch64ISD::SBCS: return "AArch64ISD::SBCS";
810 case AArch64ISD::ANDS: return "AArch64ISD::ANDS";
811 case AArch64ISD::FCMP: return "AArch64ISD::FCMP";
812 case AArch64ISD::FMIN: return "AArch64ISD::FMIN";
813 case AArch64ISD::FMAX: return "AArch64ISD::FMAX";
814 case AArch64ISD::DUP: return "AArch64ISD::DUP";
815 case AArch64ISD::DUPLANE8: return "AArch64ISD::DUPLANE8";
816 case AArch64ISD::DUPLANE16: return "AArch64ISD::DUPLANE16";
817 case AArch64ISD::DUPLANE32: return "AArch64ISD::DUPLANE32";
818 case AArch64ISD::DUPLANE64: return "AArch64ISD::DUPLANE64";
819 case AArch64ISD::MOVI: return "AArch64ISD::MOVI";
820 case AArch64ISD::MOVIshift: return "AArch64ISD::MOVIshift";
821 case AArch64ISD::MOVIedit: return "AArch64ISD::MOVIedit";
822 case AArch64ISD::MOVImsl: return "AArch64ISD::MOVImsl";
823 case AArch64ISD::FMOV: return "AArch64ISD::FMOV";
824 case AArch64ISD::MVNIshift: return "AArch64ISD::MVNIshift";
825 case AArch64ISD::MVNImsl: return "AArch64ISD::MVNImsl";
826 case AArch64ISD::BICi: return "AArch64ISD::BICi";
827 case AArch64ISD::ORRi: return "AArch64ISD::ORRi";
828 case AArch64ISD::BSL: return "AArch64ISD::BSL";
829 case AArch64ISD::NEG: return "AArch64ISD::NEG";
830 case AArch64ISD::EXTR: return "AArch64ISD::EXTR";
831 case AArch64ISD::ZIP1: return "AArch64ISD::ZIP1";
832 case AArch64ISD::ZIP2: return "AArch64ISD::ZIP2";
833 case AArch64ISD::UZP1: return "AArch64ISD::UZP1";
834 case AArch64ISD::UZP2: return "AArch64ISD::UZP2";
835 case AArch64ISD::TRN1: return "AArch64ISD::TRN1";
836 case AArch64ISD::TRN2: return "AArch64ISD::TRN2";
837 case AArch64ISD::REV16: return "AArch64ISD::REV16";
838 case AArch64ISD::REV32: return "AArch64ISD::REV32";
839 case AArch64ISD::REV64: return "AArch64ISD::REV64";
840 case AArch64ISD::EXT: return "AArch64ISD::EXT";
841 case AArch64ISD::VSHL: return "AArch64ISD::VSHL";
842 case AArch64ISD::VLSHR: return "AArch64ISD::VLSHR";
843 case AArch64ISD::VASHR: return "AArch64ISD::VASHR";
844 case AArch64ISD::CMEQ: return "AArch64ISD::CMEQ";
845 case AArch64ISD::CMGE: return "AArch64ISD::CMGE";
846 case AArch64ISD::CMGT: return "AArch64ISD::CMGT";
847 case AArch64ISD::CMHI: return "AArch64ISD::CMHI";
848 case AArch64ISD::CMHS: return "AArch64ISD::CMHS";
849 case AArch64ISD::FCMEQ: return "AArch64ISD::FCMEQ";
850 case AArch64ISD::FCMGE: return "AArch64ISD::FCMGE";
851 case AArch64ISD::FCMGT: return "AArch64ISD::FCMGT";
852 case AArch64ISD::CMEQz: return "AArch64ISD::CMEQz";
853 case AArch64ISD::CMGEz: return "AArch64ISD::CMGEz";
854 case AArch64ISD::CMGTz: return "AArch64ISD::CMGTz";
855 case AArch64ISD::CMLEz: return "AArch64ISD::CMLEz";
856 case AArch64ISD::CMLTz: return "AArch64ISD::CMLTz";
857 case AArch64ISD::FCMEQz: return "AArch64ISD::FCMEQz";
858 case AArch64ISD::FCMGEz: return "AArch64ISD::FCMGEz";
859 case AArch64ISD::FCMGTz: return "AArch64ISD::FCMGTz";
860 case AArch64ISD::FCMLEz: return "AArch64ISD::FCMLEz";
861 case AArch64ISD::FCMLTz: return "AArch64ISD::FCMLTz";
862 case AArch64ISD::SADDV: return "AArch64ISD::SADDV";
863 case AArch64ISD::UADDV: return "AArch64ISD::UADDV";
864 case AArch64ISD::SMINV: return "AArch64ISD::SMINV";
865 case AArch64ISD::UMINV: return "AArch64ISD::UMINV";
866 case AArch64ISD::SMAXV: return "AArch64ISD::SMAXV";
867 case AArch64ISD::UMAXV: return "AArch64ISD::UMAXV";
868 case AArch64ISD::NOT: return "AArch64ISD::NOT";
869 case AArch64ISD::BIT: return "AArch64ISD::BIT";
870 case AArch64ISD::CBZ: return "AArch64ISD::CBZ";
871 case AArch64ISD::CBNZ: return "AArch64ISD::CBNZ";
872 case AArch64ISD::TBZ: return "AArch64ISD::TBZ";
873 case AArch64ISD::TBNZ: return "AArch64ISD::TBNZ";
874 case AArch64ISD::TC_RETURN: return "AArch64ISD::TC_RETURN";
875 case AArch64ISD::PREFETCH: return "AArch64ISD::PREFETCH";
876 case AArch64ISD::SITOF: return "AArch64ISD::SITOF";
877 case AArch64ISD::UITOF: return "AArch64ISD::UITOF";
878 case AArch64ISD::NVCAST: return "AArch64ISD::NVCAST";
879 case AArch64ISD::SQSHL_I: return "AArch64ISD::SQSHL_I";
880 case AArch64ISD::UQSHL_I: return "AArch64ISD::UQSHL_I";
881 case AArch64ISD::SRSHR_I: return "AArch64ISD::SRSHR_I";
882 case AArch64ISD::URSHR_I: return "AArch64ISD::URSHR_I";
883 case AArch64ISD::SQSHLU_I: return "AArch64ISD::SQSHLU_I";
884 case AArch64ISD::WrapperLarge: return "AArch64ISD::WrapperLarge";
885 case AArch64ISD::LD2post: return "AArch64ISD::LD2post";
886 case AArch64ISD::LD3post: return "AArch64ISD::LD3post";
887 case AArch64ISD::LD4post: return "AArch64ISD::LD4post";
888 case AArch64ISD::ST2post: return "AArch64ISD::ST2post";
889 case AArch64ISD::ST3post: return "AArch64ISD::ST3post";
890 case AArch64ISD::ST4post: return "AArch64ISD::ST4post";
891 case AArch64ISD::LD1x2post: return "AArch64ISD::LD1x2post";
892 case AArch64ISD::LD1x3post: return "AArch64ISD::LD1x3post";
893 case AArch64ISD::LD1x4post: return "AArch64ISD::LD1x4post";
894 case AArch64ISD::ST1x2post: return "AArch64ISD::ST1x2post";
895 case AArch64ISD::ST1x3post: return "AArch64ISD::ST1x3post";
896 case AArch64ISD::ST1x4post: return "AArch64ISD::ST1x4post";
897 case AArch64ISD::LD1DUPpost: return "AArch64ISD::LD1DUPpost";
898 case AArch64ISD::LD2DUPpost: return "AArch64ISD::LD2DUPpost";
899 case AArch64ISD::LD3DUPpost: return "AArch64ISD::LD3DUPpost";
900 case AArch64ISD::LD4DUPpost: return "AArch64ISD::LD4DUPpost";
901 case AArch64ISD::LD1LANEpost: return "AArch64ISD::LD1LANEpost";
902 case AArch64ISD::LD2LANEpost: return "AArch64ISD::LD2LANEpost";
903 case AArch64ISD::LD3LANEpost: return "AArch64ISD::LD3LANEpost";
904 case AArch64ISD::LD4LANEpost: return "AArch64ISD::LD4LANEpost";
905 case AArch64ISD::ST2LANEpost: return "AArch64ISD::ST2LANEpost";
906 case AArch64ISD::ST3LANEpost: return "AArch64ISD::ST3LANEpost";
907 case AArch64ISD::ST4LANEpost: return "AArch64ISD::ST4LANEpost";
908 case AArch64ISD::SMULL: return "AArch64ISD::SMULL";
909 case AArch64ISD::UMULL: return "AArch64ISD::UMULL";
915 AArch64TargetLowering::EmitF128CSEL(MachineInstr *MI,
916 MachineBasicBlock *MBB) const {
917 // We materialise the F128CSEL pseudo-instruction as some control flow and a
921 // [... previous instrs leading to comparison ...]
927 // Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB]
929 MachineFunction *MF = MBB->getParent();
930 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
931 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
932 DebugLoc DL = MI->getDebugLoc();
933 MachineFunction::iterator It = MBB;
936 unsigned DestReg = MI->getOperand(0).getReg();
937 unsigned IfTrueReg = MI->getOperand(1).getReg();
938 unsigned IfFalseReg = MI->getOperand(2).getReg();
939 unsigned CondCode = MI->getOperand(3).getImm();
940 bool NZCVKilled = MI->getOperand(4).isKill();
942 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
943 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
944 MF->insert(It, TrueBB);
945 MF->insert(It, EndBB);
947 // Transfer rest of current basic-block to EndBB
948 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
950 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
952 BuildMI(MBB, DL, TII->get(AArch64::Bcc)).addImm(CondCode).addMBB(TrueBB);
953 BuildMI(MBB, DL, TII->get(AArch64::B)).addMBB(EndBB);
954 MBB->addSuccessor(TrueBB);
955 MBB->addSuccessor(EndBB);
957 // TrueBB falls through to the end.
958 TrueBB->addSuccessor(EndBB);
961 TrueBB->addLiveIn(AArch64::NZCV);
962 EndBB->addLiveIn(AArch64::NZCV);
965 BuildMI(*EndBB, EndBB->begin(), DL, TII->get(AArch64::PHI), DestReg)
971 MI->eraseFromParent();
976 AArch64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
977 MachineBasicBlock *BB) const {
978 switch (MI->getOpcode()) {
983 llvm_unreachable("Unexpected instruction for custom inserter!");
985 case AArch64::F128CSEL:
986 return EmitF128CSEL(MI, BB);
988 case TargetOpcode::STACKMAP:
989 case TargetOpcode::PATCHPOINT:
990 return emitPatchPoint(MI, BB);
994 //===----------------------------------------------------------------------===//
995 // AArch64 Lowering private implementation.
996 //===----------------------------------------------------------------------===//
998 //===----------------------------------------------------------------------===//
1000 //===----------------------------------------------------------------------===//
1002 /// changeIntCCToAArch64CC - Convert a DAG integer condition code to an AArch64
1004 static AArch64CC::CondCode changeIntCCToAArch64CC(ISD::CondCode CC) {
1007 llvm_unreachable("Unknown condition code!");
1009 return AArch64CC::NE;
1011 return AArch64CC::EQ;
1013 return AArch64CC::GT;
1015 return AArch64CC::GE;
1017 return AArch64CC::LT;
1019 return AArch64CC::LE;
1021 return AArch64CC::HI;
1023 return AArch64CC::HS;
1025 return AArch64CC::LO;
1027 return AArch64CC::LS;
1031 /// changeFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64 CC.
1032 static void changeFPCCToAArch64CC(ISD::CondCode CC,
1033 AArch64CC::CondCode &CondCode,
1034 AArch64CC::CondCode &CondCode2) {
1035 CondCode2 = AArch64CC::AL;
1038 llvm_unreachable("Unknown FP condition!");
1041 CondCode = AArch64CC::EQ;
1045 CondCode = AArch64CC::GT;
1049 CondCode = AArch64CC::GE;
1052 CondCode = AArch64CC::MI;
1055 CondCode = AArch64CC::LS;
1058 CondCode = AArch64CC::MI;
1059 CondCode2 = AArch64CC::GT;
1062 CondCode = AArch64CC::VC;
1065 CondCode = AArch64CC::VS;
1068 CondCode = AArch64CC::EQ;
1069 CondCode2 = AArch64CC::VS;
1072 CondCode = AArch64CC::HI;
1075 CondCode = AArch64CC::PL;
1079 CondCode = AArch64CC::LT;
1083 CondCode = AArch64CC::LE;
1087 CondCode = AArch64CC::NE;
1092 /// changeVectorFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64
1093 /// CC usable with the vector instructions. Fewer operations are available
1094 /// without a real NZCV register, so we have to use less efficient combinations
1095 /// to get the same effect.
1096 static void changeVectorFPCCToAArch64CC(ISD::CondCode CC,
1097 AArch64CC::CondCode &CondCode,
1098 AArch64CC::CondCode &CondCode2,
1103 // Mostly the scalar mappings work fine.
1104 changeFPCCToAArch64CC(CC, CondCode, CondCode2);
1107 Invert = true; // Fallthrough
1109 CondCode = AArch64CC::MI;
1110 CondCode2 = AArch64CC::GE;
1117 // All of the compare-mask comparisons are ordered, but we can switch
1118 // between the two by a double inversion. E.g. ULE == !OGT.
1120 changeFPCCToAArch64CC(getSetCCInverse(CC, false), CondCode, CondCode2);
1125 static bool isLegalArithImmed(uint64_t C) {
1126 // Matches AArch64DAGToDAGISel::SelectArithImmed().
1127 return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
1130 static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1131 SDLoc dl, SelectionDAG &DAG) {
1132 EVT VT = LHS.getValueType();
1134 if (VT.isFloatingPoint())
1135 return DAG.getNode(AArch64ISD::FCMP, dl, VT, LHS, RHS);
1137 // The CMP instruction is just an alias for SUBS, and representing it as
1138 // SUBS means that it's possible to get CSE with subtract operations.
1139 // A later phase can perform the optimization of setting the destination
1140 // register to WZR/XZR if it ends up being unused.
1141 unsigned Opcode = AArch64ISD::SUBS;
1143 if (RHS.getOpcode() == ISD::SUB && isa<ConstantSDNode>(RHS.getOperand(0)) &&
1144 cast<ConstantSDNode>(RHS.getOperand(0))->getZExtValue() == 0 &&
1145 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1146 // We'd like to combine a (CMP op1, (sub 0, op2) into a CMN instruction on
1147 // the grounds that "op1 - (-op2) == op1 + op2". However, the C and V flags
1148 // can be set differently by this operation. It comes down to whether
1149 // "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are then
1150 // everything is fine. If not then the optimization is wrong. Thus general
1151 // comparisons are only valid if op2 != 0.
1153 // So, finally, the only LLVM-native comparisons that don't mention C and V
1154 // are SETEQ and SETNE. They're the only ones we can safely use CMN for in
1155 // the absence of information about op2.
1156 Opcode = AArch64ISD::ADDS;
1157 RHS = RHS.getOperand(1);
1158 } else if (LHS.getOpcode() == ISD::AND && isa<ConstantSDNode>(RHS) &&
1159 cast<ConstantSDNode>(RHS)->getZExtValue() == 0 &&
1160 !isUnsignedIntSetCC(CC)) {
1161 // Similarly, (CMP (and X, Y), 0) can be implemented with a TST
1162 // (a.k.a. ANDS) except that the flags are only guaranteed to work for one
1163 // of the signed comparisons.
1164 Opcode = AArch64ISD::ANDS;
1165 RHS = LHS.getOperand(1);
1166 LHS = LHS.getOperand(0);
1169 return DAG.getNode(Opcode, dl, DAG.getVTList(VT, MVT::i32), LHS, RHS)
1173 static SDValue getAArch64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1174 SDValue &AArch64cc, SelectionDAG &DAG, SDLoc dl) {
1176 AArch64CC::CondCode AArch64CC;
1177 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1178 EVT VT = RHS.getValueType();
1179 uint64_t C = RHSC->getZExtValue();
1180 if (!isLegalArithImmed(C)) {
1181 // Constant does not fit, try adjusting it by one?
1187 if ((VT == MVT::i32 && C != 0x80000000 &&
1188 isLegalArithImmed((uint32_t)(C - 1))) ||
1189 (VT == MVT::i64 && C != 0x80000000ULL &&
1190 isLegalArithImmed(C - 1ULL))) {
1191 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1192 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1193 RHS = DAG.getConstant(C, dl, VT);
1198 if ((VT == MVT::i32 && C != 0 &&
1199 isLegalArithImmed((uint32_t)(C - 1))) ||
1200 (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) {
1201 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1202 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1203 RHS = DAG.getConstant(C, dl, VT);
1208 if ((VT == MVT::i32 && C != INT32_MAX &&
1209 isLegalArithImmed((uint32_t)(C + 1))) ||
1210 (VT == MVT::i64 && C != INT64_MAX &&
1211 isLegalArithImmed(C + 1ULL))) {
1212 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1213 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1214 RHS = DAG.getConstant(C, dl, VT);
1219 if ((VT == MVT::i32 && C != UINT32_MAX &&
1220 isLegalArithImmed((uint32_t)(C + 1))) ||
1221 (VT == MVT::i64 && C != UINT64_MAX &&
1222 isLegalArithImmed(C + 1ULL))) {
1223 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1224 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1225 RHS = DAG.getConstant(C, dl, VT);
1231 // The imm operand of ADDS is an unsigned immediate, in the range 0 to 4095.
1232 // For the i8 operand, the largest immediate is 255, so this can be easily
1233 // encoded in the compare instruction. For the i16 operand, however, the
1234 // largest immediate cannot be encoded in the compare.
1235 // Therefore, use a sign extending load and cmn to avoid materializing the -1
1236 // constant. For example,
1238 // ldrh w0, [x0, #0]
1241 // ldrsh w0, [x0, #0]
1243 // Fundamental, we're relying on the property that (zext LHS) == (zext RHS)
1244 // if and only if (sext LHS) == (sext RHS). The checks are in place to ensure
1245 // both the LHS and RHS are truely zero extended and to make sure the
1246 // transformation is profitable.
1247 if ((CC == ISD::SETEQ || CC == ISD::SETNE) && isa<ConstantSDNode>(RHS)) {
1248 if ((cast<ConstantSDNode>(RHS)->getZExtValue() >> 16 == 0) &&
1249 isa<LoadSDNode>(LHS)) {
1250 if (cast<LoadSDNode>(LHS)->getExtensionType() == ISD::ZEXTLOAD &&
1251 cast<LoadSDNode>(LHS)->getMemoryVT() == MVT::i16 &&
1252 LHS.getNode()->hasNUsesOfValue(1, 0)) {
1253 int16_t ValueofRHS = cast<ConstantSDNode>(RHS)->getZExtValue();
1254 if (ValueofRHS < 0 && isLegalArithImmed(-ValueofRHS)) {
1256 DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, LHS.getValueType(), LHS,
1257 DAG.getValueType(MVT::i16));
1258 Cmp = emitComparison(SExt,
1259 DAG.getConstant(ValueofRHS, dl,
1260 RHS.getValueType()),
1262 AArch64CC = changeIntCCToAArch64CC(CC);
1263 AArch64cc = DAG.getConstant(AArch64CC, dl, MVT::i32);
1269 Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
1270 AArch64CC = changeIntCCToAArch64CC(CC);
1271 AArch64cc = DAG.getConstant(AArch64CC, dl, MVT::i32);
1275 static std::pair<SDValue, SDValue>
1276 getAArch64XALUOOp(AArch64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
1277 assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
1278 "Unsupported value type");
1279 SDValue Value, Overflow;
1281 SDValue LHS = Op.getOperand(0);
1282 SDValue RHS = Op.getOperand(1);
1284 switch (Op.getOpcode()) {
1286 llvm_unreachable("Unknown overflow instruction!");
1288 Opc = AArch64ISD::ADDS;
1292 Opc = AArch64ISD::ADDS;
1296 Opc = AArch64ISD::SUBS;
1300 Opc = AArch64ISD::SUBS;
1303 // Multiply needs a little bit extra work.
1307 bool IsSigned = Op.getOpcode() == ISD::SMULO;
1308 if (Op.getValueType() == MVT::i32) {
1309 unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1310 // For a 32 bit multiply with overflow check we want the instruction
1311 // selector to generate a widening multiply (SMADDL/UMADDL). For that we
1312 // need to generate the following pattern:
1313 // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b))
1314 LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
1315 RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
1316 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1317 SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul,
1318 DAG.getConstant(0, DL, MVT::i64));
1319 // On AArch64 the upper 32 bits are always zero extended for a 32 bit
1320 // operation. We need to clear out the upper 32 bits, because we used a
1321 // widening multiply that wrote all 64 bits. In the end this should be a
1323 Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add);
1325 // The signed overflow check requires more than just a simple check for
1326 // any bit set in the upper 32 bits of the result. These bits could be
1327 // just the sign bits of a negative number. To perform the overflow
1328 // check we have to arithmetic shift right the 32nd bit of the result by
1329 // 31 bits. Then we compare the result to the upper 32 bits.
1330 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add,
1331 DAG.getConstant(32, DL, MVT::i64));
1332 UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits);
1333 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value,
1334 DAG.getConstant(31, DL, MVT::i64));
1335 // It is important that LowerBits is last, otherwise the arithmetic
1336 // shift will not be folded into the compare (SUBS).
1337 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32);
1338 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1341 // The overflow check for unsigned multiply is easy. We only need to
1342 // check if any of the upper 32 bits are set. This can be done with a
1343 // CMP (shifted register). For that we need to generate the following
1345 // (i64 AArch64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32)
1346 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
1347 DAG.getConstant(32, DL, MVT::i64));
1348 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1350 DAG.getNode(AArch64ISD::SUBS, DL, VTs,
1351 DAG.getConstant(0, DL, MVT::i64),
1352 UpperBits).getValue(1);
1356 assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
1357 // For the 64 bit multiply
1358 Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1360 SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
1361 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
1362 DAG.getConstant(63, DL, MVT::i64));
1363 // It is important that LowerBits is last, otherwise the arithmetic
1364 // shift will not be folded into the compare (SUBS).
1365 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1366 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1369 SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
1370 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1372 DAG.getNode(AArch64ISD::SUBS, DL, VTs,
1373 DAG.getConstant(0, DL, MVT::i64),
1374 UpperBits).getValue(1);
1381 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);
1383 // Emit the AArch64 operation with overflow check.
1384 Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
1385 Overflow = Value.getValue(1);
1387 return std::make_pair(Value, Overflow);
1390 SDValue AArch64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG,
1391 RTLIB::Libcall Call) const {
1392 SmallVector<SDValue, 2> Ops(Op->op_begin(), Op->op_end());
1393 return makeLibCall(DAG, Call, MVT::f128, &Ops[0], Ops.size(), false,
1397 static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) {
1398 SDValue Sel = Op.getOperand(0);
1399 SDValue Other = Op.getOperand(1);
1401 // If neither operand is a SELECT_CC, give up.
1402 if (Sel.getOpcode() != ISD::SELECT_CC)
1403 std::swap(Sel, Other);
1404 if (Sel.getOpcode() != ISD::SELECT_CC)
1407 // The folding we want to perform is:
1408 // (xor x, (select_cc a, b, cc, 0, -1) )
1410 // (csel x, (xor x, -1), cc ...)
1412 // The latter will get matched to a CSINV instruction.
1414 ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
1415 SDValue LHS = Sel.getOperand(0);
1416 SDValue RHS = Sel.getOperand(1);
1417 SDValue TVal = Sel.getOperand(2);
1418 SDValue FVal = Sel.getOperand(3);
1421 // FIXME: This could be generalized to non-integer comparisons.
1422 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
1425 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
1426 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
1428 // The values aren't constants, this isn't the pattern we're looking for.
1429 if (!CFVal || !CTVal)
1432 // We can commute the SELECT_CC by inverting the condition. This
1433 // might be needed to make this fit into a CSINV pattern.
1434 if (CTVal->isAllOnesValue() && CFVal->isNullValue()) {
1435 std::swap(TVal, FVal);
1436 std::swap(CTVal, CFVal);
1437 CC = ISD::getSetCCInverse(CC, true);
1440 // If the constants line up, perform the transform!
1441 if (CTVal->isNullValue() && CFVal->isAllOnesValue()) {
1443 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
1446 TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other,
1447 DAG.getConstant(-1ULL, dl, Other.getValueType()));
1449 return DAG.getNode(AArch64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal,
1456 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
1457 EVT VT = Op.getValueType();
1459 // Let legalize expand this if it isn't a legal type yet.
1460 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
1463 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
1466 bool ExtraOp = false;
1467 switch (Op.getOpcode()) {
1469 llvm_unreachable("Invalid code");
1471 Opc = AArch64ISD::ADDS;
1474 Opc = AArch64ISD::SUBS;
1477 Opc = AArch64ISD::ADCS;
1481 Opc = AArch64ISD::SBCS;
1487 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1));
1488 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1),
1492 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
1493 // Let legalize expand this if it isn't a legal type yet.
1494 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
1498 AArch64CC::CondCode CC;
1499 // The actual operation that sets the overflow or carry flag.
1500 SDValue Value, Overflow;
1501 std::tie(Value, Overflow) = getAArch64XALUOOp(CC, Op, DAG);
1503 // We use 0 and 1 as false and true values.
1504 SDValue TVal = DAG.getConstant(1, dl, MVT::i32);
1505 SDValue FVal = DAG.getConstant(0, dl, MVT::i32);
1507 // We use an inverted condition, because the conditional select is inverted
1508 // too. This will allow it to be selected to a single instruction:
1509 // CSINC Wd, WZR, WZR, invert(cond).
1510 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), dl, MVT::i32);
1511 Overflow = DAG.getNode(AArch64ISD::CSEL, dl, MVT::i32, FVal, TVal,
1514 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
1515 return DAG.getNode(ISD::MERGE_VALUES, dl, VTs, Value, Overflow);
1518 // Prefetch operands are:
1519 // 1: Address to prefetch
1521 // 3: int locality (0 = no locality ... 3 = extreme locality)
1522 // 4: bool isDataCache
1523 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
1525 unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
1526 unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
1527 unsigned IsData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
1529 bool IsStream = !Locality;
1530 // When the locality number is set
1532 // The front-end should have filtered out the out-of-range values
1533 assert(Locality <= 3 && "Prefetch locality out-of-range");
1534 // The locality degree is the opposite of the cache speed.
1535 // Put the number the other way around.
1536 // The encoding starts at 0 for level 1
1537 Locality = 3 - Locality;
1540 // built the mask value encoding the expected behavior.
1541 unsigned PrfOp = (IsWrite << 4) | // Load/Store bit
1542 (!IsData << 3) | // IsDataCache bit
1543 (Locality << 1) | // Cache level bits
1544 (unsigned)IsStream; // Stream bit
1545 return DAG.getNode(AArch64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
1546 DAG.getConstant(PrfOp, DL, MVT::i32), Op.getOperand(1));
1549 SDValue AArch64TargetLowering::LowerFP_EXTEND(SDValue Op,
1550 SelectionDAG &DAG) const {
1551 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
1554 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
1556 return LowerF128Call(Op, DAG, LC);
1559 SDValue AArch64TargetLowering::LowerFP_ROUND(SDValue Op,
1560 SelectionDAG &DAG) const {
1561 if (Op.getOperand(0).getValueType() != MVT::f128) {
1562 // It's legal except when f128 is involved
1567 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
1569 // FP_ROUND node has a second operand indicating whether it is known to be
1570 // precise. That doesn't take part in the LibCall so we can't directly use
1572 SDValue SrcVal = Op.getOperand(0);
1573 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
1574 /*isSigned*/ false, SDLoc(Op)).first;
1577 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1578 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1579 // Any additional optimization in this function should be recorded
1580 // in the cost tables.
1581 EVT InVT = Op.getOperand(0).getValueType();
1582 EVT VT = Op.getValueType();
1584 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1587 DAG.getNode(Op.getOpcode(), dl, InVT.changeVectorElementTypeToInteger(),
1589 return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
1592 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1595 MVT::getVectorVT(MVT::getFloatingPointVT(VT.getScalarSizeInBits()),
1596 VT.getVectorNumElements());
1597 SDValue Ext = DAG.getNode(ISD::FP_EXTEND, dl, ExtVT, Op.getOperand(0));
1598 return DAG.getNode(Op.getOpcode(), dl, VT, Ext);
1601 // Type changing conversions are illegal.
1605 SDValue AArch64TargetLowering::LowerFP_TO_INT(SDValue Op,
1606 SelectionDAG &DAG) const {
1607 if (Op.getOperand(0).getValueType().isVector())
1608 return LowerVectorFP_TO_INT(Op, DAG);
1610 // f16 conversions are promoted to f32.
1611 if (Op.getOperand(0).getValueType() == MVT::f16) {
1614 Op.getOpcode(), dl, Op.getValueType(),
1615 DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, Op.getOperand(0)));
1618 if (Op.getOperand(0).getValueType() != MVT::f128) {
1619 // It's legal except when f128 is involved
1624 if (Op.getOpcode() == ISD::FP_TO_SINT)
1625 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
1627 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
1629 SmallVector<SDValue, 2> Ops(Op->op_begin(), Op->op_end());
1630 return makeLibCall(DAG, LC, Op.getValueType(), &Ops[0], Ops.size(), false,
1634 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1635 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1636 // Any additional optimization in this function should be recorded
1637 // in the cost tables.
1638 EVT VT = Op.getValueType();
1640 SDValue In = Op.getOperand(0);
1641 EVT InVT = In.getValueType();
1643 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1645 MVT::getVectorVT(MVT::getFloatingPointVT(InVT.getScalarSizeInBits()),
1646 InVT.getVectorNumElements());
1647 In = DAG.getNode(Op.getOpcode(), dl, CastVT, In);
1648 return DAG.getNode(ISD::FP_ROUND, dl, VT, In, DAG.getIntPtrConstant(0, dl));
1651 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1653 Op.getOpcode() == ISD::SINT_TO_FP ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1654 EVT CastVT = VT.changeVectorElementTypeToInteger();
1655 In = DAG.getNode(CastOpc, dl, CastVT, In);
1656 return DAG.getNode(Op.getOpcode(), dl, VT, In);
1662 SDValue AArch64TargetLowering::LowerINT_TO_FP(SDValue Op,
1663 SelectionDAG &DAG) const {
1664 if (Op.getValueType().isVector())
1665 return LowerVectorINT_TO_FP(Op, DAG);
1667 // f16 conversions are promoted to f32.
1668 if (Op.getValueType() == MVT::f16) {
1671 ISD::FP_ROUND, dl, MVT::f16,
1672 DAG.getNode(Op.getOpcode(), dl, MVT::f32, Op.getOperand(0)),
1673 DAG.getIntPtrConstant(0, dl));
1676 // i128 conversions are libcalls.
1677 if (Op.getOperand(0).getValueType() == MVT::i128)
1680 // Other conversions are legal, unless it's to the completely software-based
1682 if (Op.getValueType() != MVT::f128)
1686 if (Op.getOpcode() == ISD::SINT_TO_FP)
1687 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1689 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1691 return LowerF128Call(Op, DAG, LC);
1694 SDValue AArch64TargetLowering::LowerFSINCOS(SDValue Op,
1695 SelectionDAG &DAG) const {
1696 // For iOS, we want to call an alternative entry point: __sincos_stret,
1697 // which returns the values in two S / D registers.
1699 SDValue Arg = Op.getOperand(0);
1700 EVT ArgVT = Arg.getValueType();
1701 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1708 Entry.isSExt = false;
1709 Entry.isZExt = false;
1710 Args.push_back(Entry);
1712 const char *LibcallName =
1713 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
1715 DAG.getExternalSymbol(LibcallName, getPointerTy(DAG.getDataLayout()));
1717 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
1718 TargetLowering::CallLoweringInfo CLI(DAG);
1719 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
1720 .setCallee(CallingConv::Fast, RetTy, Callee, std::move(Args), 0);
1722 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1723 return CallResult.first;
1726 static SDValue LowerBITCAST(SDValue Op, SelectionDAG &DAG) {
1727 if (Op.getValueType() != MVT::f16)
1730 assert(Op.getOperand(0).getValueType() == MVT::i16);
1733 Op = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op.getOperand(0));
1734 Op = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Op);
1736 DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL, MVT::f16, Op,
1737 DAG.getTargetConstant(AArch64::hsub, DL, MVT::i32)),
1741 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
1742 if (OrigVT.getSizeInBits() >= 64)
1745 assert(OrigVT.isSimple() && "Expecting a simple value type");
1747 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
1748 switch (OrigSimpleTy) {
1749 default: llvm_unreachable("Unexpected Vector Type");
1758 static SDValue addRequiredExtensionForVectorMULL(SDValue N, SelectionDAG &DAG,
1761 unsigned ExtOpcode) {
1762 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
1763 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
1764 // 64-bits we need to insert a new extension so that it will be 64-bits.
1765 assert(ExtTy.is128BitVector() && "Unexpected extension size");
1766 if (OrigTy.getSizeInBits() >= 64)
1769 // Must extend size to at least 64 bits to be used as an operand for VMULL.
1770 EVT NewVT = getExtensionTo64Bits(OrigTy);
1772 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
1775 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
1777 EVT VT = N->getValueType(0);
1779 if (N->getOpcode() != ISD::BUILD_VECTOR)
1782 for (const SDValue &Elt : N->op_values()) {
1783 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
1784 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1785 unsigned HalfSize = EltSize / 2;
1787 if (!isIntN(HalfSize, C->getSExtValue()))
1790 if (!isUIntN(HalfSize, C->getZExtValue()))
1801 static SDValue skipExtensionForVectorMULL(SDNode *N, SelectionDAG &DAG) {
1802 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
1803 return addRequiredExtensionForVectorMULL(N->getOperand(0), DAG,
1804 N->getOperand(0)->getValueType(0),
1808 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
1809 EVT VT = N->getValueType(0);
1811 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
1812 unsigned NumElts = VT.getVectorNumElements();
1813 MVT TruncVT = MVT::getIntegerVT(EltSize);
1814 SmallVector<SDValue, 8> Ops;
1815 for (unsigned i = 0; i != NumElts; ++i) {
1816 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
1817 const APInt &CInt = C->getAPIntValue();
1818 // Element types smaller than 32 bits are not legal, so use i32 elements.
1819 // The values are implicitly truncated so sext vs. zext doesn't matter.
1820 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), dl, MVT::i32));
1822 return DAG.getNode(ISD::BUILD_VECTOR, dl,
1823 MVT::getVectorVT(TruncVT, NumElts), Ops);
1826 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
1827 if (N->getOpcode() == ISD::SIGN_EXTEND)
1829 if (isExtendedBUILD_VECTOR(N, DAG, true))
1834 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
1835 if (N->getOpcode() == ISD::ZERO_EXTEND)
1837 if (isExtendedBUILD_VECTOR(N, DAG, false))
1842 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
1843 unsigned Opcode = N->getOpcode();
1844 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
1845 SDNode *N0 = N->getOperand(0).getNode();
1846 SDNode *N1 = N->getOperand(1).getNode();
1847 return N0->hasOneUse() && N1->hasOneUse() &&
1848 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
1853 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
1854 unsigned Opcode = N->getOpcode();
1855 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
1856 SDNode *N0 = N->getOperand(0).getNode();
1857 SDNode *N1 = N->getOperand(1).getNode();
1858 return N0->hasOneUse() && N1->hasOneUse() &&
1859 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
1864 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
1865 // Multiplications are only custom-lowered for 128-bit vectors so that
1866 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
1867 EVT VT = Op.getValueType();
1868 assert(VT.is128BitVector() && VT.isInteger() &&
1869 "unexpected type for custom-lowering ISD::MUL");
1870 SDNode *N0 = Op.getOperand(0).getNode();
1871 SDNode *N1 = Op.getOperand(1).getNode();
1872 unsigned NewOpc = 0;
1874 bool isN0SExt = isSignExtended(N0, DAG);
1875 bool isN1SExt = isSignExtended(N1, DAG);
1876 if (isN0SExt && isN1SExt)
1877 NewOpc = AArch64ISD::SMULL;
1879 bool isN0ZExt = isZeroExtended(N0, DAG);
1880 bool isN1ZExt = isZeroExtended(N1, DAG);
1881 if (isN0ZExt && isN1ZExt)
1882 NewOpc = AArch64ISD::UMULL;
1883 else if (isN1SExt || isN1ZExt) {
1884 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
1885 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
1886 if (isN1SExt && isAddSubSExt(N0, DAG)) {
1887 NewOpc = AArch64ISD::SMULL;
1889 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
1890 NewOpc = AArch64ISD::UMULL;
1892 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
1894 NewOpc = AArch64ISD::UMULL;
1900 if (VT == MVT::v2i64)
1901 // Fall through to expand this. It is not legal.
1904 // Other vector multiplications are legal.
1909 // Legalize to a S/UMULL instruction
1912 SDValue Op1 = skipExtensionForVectorMULL(N1, DAG);
1914 Op0 = skipExtensionForVectorMULL(N0, DAG);
1915 assert(Op0.getValueType().is64BitVector() &&
1916 Op1.getValueType().is64BitVector() &&
1917 "unexpected types for extended operands to VMULL");
1918 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
1920 // Optimizing (zext A + zext B) * C, to (S/UMULL A, C) + (S/UMULL B, C) during
1921 // isel lowering to take advantage of no-stall back to back s/umul + s/umla.
1922 // This is true for CPUs with accumulate forwarding such as Cortex-A53/A57
1923 SDValue N00 = skipExtensionForVectorMULL(N0->getOperand(0).getNode(), DAG);
1924 SDValue N01 = skipExtensionForVectorMULL(N0->getOperand(1).getNode(), DAG);
1925 EVT Op1VT = Op1.getValueType();
1926 return DAG.getNode(N0->getOpcode(), DL, VT,
1927 DAG.getNode(NewOpc, DL, VT,
1928 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
1929 DAG.getNode(NewOpc, DL, VT,
1930 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
1933 SDValue AArch64TargetLowering::LowerOperation(SDValue Op,
1934 SelectionDAG &DAG) const {
1935 switch (Op.getOpcode()) {
1937 llvm_unreachable("unimplemented operand");
1940 return LowerBITCAST(Op, DAG);
1941 case ISD::GlobalAddress:
1942 return LowerGlobalAddress(Op, DAG);
1943 case ISD::GlobalTLSAddress:
1944 return LowerGlobalTLSAddress(Op, DAG);
1946 return LowerSETCC(Op, DAG);
1948 return LowerBR_CC(Op, DAG);
1950 return LowerSELECT(Op, DAG);
1951 case ISD::SELECT_CC:
1952 return LowerSELECT_CC(Op, DAG);
1953 case ISD::JumpTable:
1954 return LowerJumpTable(Op, DAG);
1955 case ISD::ConstantPool:
1956 return LowerConstantPool(Op, DAG);
1957 case ISD::BlockAddress:
1958 return LowerBlockAddress(Op, DAG);
1960 return LowerVASTART(Op, DAG);
1962 return LowerVACOPY(Op, DAG);
1964 return LowerVAARG(Op, DAG);
1969 return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
1976 return LowerXALUO(Op, DAG);
1978 return LowerF128Call(Op, DAG, RTLIB::ADD_F128);
1980 return LowerF128Call(Op, DAG, RTLIB::SUB_F128);
1982 return LowerF128Call(Op, DAG, RTLIB::MUL_F128);
1984 return LowerF128Call(Op, DAG, RTLIB::DIV_F128);
1986 return LowerFP_ROUND(Op, DAG);
1987 case ISD::FP_EXTEND:
1988 return LowerFP_EXTEND(Op, DAG);
1989 case ISD::FRAMEADDR:
1990 return LowerFRAMEADDR(Op, DAG);
1991 case ISD::RETURNADDR:
1992 return LowerRETURNADDR(Op, DAG);
1993 case ISD::INSERT_VECTOR_ELT:
1994 return LowerINSERT_VECTOR_ELT(Op, DAG);
1995 case ISD::EXTRACT_VECTOR_ELT:
1996 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
1997 case ISD::BUILD_VECTOR:
1998 return LowerBUILD_VECTOR(Op, DAG);
1999 case ISD::VECTOR_SHUFFLE:
2000 return LowerVECTOR_SHUFFLE(Op, DAG);
2001 case ISD::EXTRACT_SUBVECTOR:
2002 return LowerEXTRACT_SUBVECTOR(Op, DAG);
2006 return LowerVectorSRA_SRL_SHL(Op, DAG);
2007 case ISD::SHL_PARTS:
2008 return LowerShiftLeftParts(Op, DAG);
2009 case ISD::SRL_PARTS:
2010 case ISD::SRA_PARTS:
2011 return LowerShiftRightParts(Op, DAG);
2013 return LowerCTPOP(Op, DAG);
2014 case ISD::FCOPYSIGN:
2015 return LowerFCOPYSIGN(Op, DAG);
2017 return LowerVectorAND(Op, DAG);
2019 return LowerVectorOR(Op, DAG);
2021 return LowerXOR(Op, DAG);
2023 return LowerPREFETCH(Op, DAG);
2024 case ISD::SINT_TO_FP:
2025 case ISD::UINT_TO_FP:
2026 return LowerINT_TO_FP(Op, DAG);
2027 case ISD::FP_TO_SINT:
2028 case ISD::FP_TO_UINT:
2029 return LowerFP_TO_INT(Op, DAG);
2031 return LowerFSINCOS(Op, DAG);
2033 return LowerMUL(Op, DAG);
2037 /// getFunctionAlignment - Return the Log2 alignment of this function.
2038 unsigned AArch64TargetLowering::getFunctionAlignment(const Function *F) const {
2042 //===----------------------------------------------------------------------===//
2043 // Calling Convention Implementation
2044 //===----------------------------------------------------------------------===//
2046 #include "AArch64GenCallingConv.inc"
2048 /// Selects the correct CCAssignFn for a given CallingConvention value.
2049 CCAssignFn *AArch64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
2050 bool IsVarArg) const {
2053 llvm_unreachable("Unsupported calling convention.");
2054 case CallingConv::WebKit_JS:
2055 return CC_AArch64_WebKit_JS;
2056 case CallingConv::GHC:
2057 return CC_AArch64_GHC;
2058 case CallingConv::C:
2059 case CallingConv::Fast:
2060 if (!Subtarget->isTargetDarwin())
2061 return CC_AArch64_AAPCS;
2062 return IsVarArg ? CC_AArch64_DarwinPCS_VarArg : CC_AArch64_DarwinPCS;
2066 SDValue AArch64TargetLowering::LowerFormalArguments(
2067 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2068 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2069 SmallVectorImpl<SDValue> &InVals) const {
2070 MachineFunction &MF = DAG.getMachineFunction();
2071 MachineFrameInfo *MFI = MF.getFrameInfo();
2073 // Assign locations to all of the incoming arguments.
2074 SmallVector<CCValAssign, 16> ArgLocs;
2075 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
2078 // At this point, Ins[].VT may already be promoted to i32. To correctly
2079 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2080 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
2081 // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
2082 // we use a special version of AnalyzeFormalArguments to pass in ValVT and
2084 unsigned NumArgs = Ins.size();
2085 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2086 unsigned CurArgIdx = 0;
2087 for (unsigned i = 0; i != NumArgs; ++i) {
2088 MVT ValVT = Ins[i].VT;
2089 if (Ins[i].isOrigArg()) {
2090 std::advance(CurOrigArg, Ins[i].getOrigArgIndex() - CurArgIdx);
2091 CurArgIdx = Ins[i].getOrigArgIndex();
2093 // Get type of the original argument.
2094 EVT ActualVT = getValueType(DAG.getDataLayout(), CurOrigArg->getType(),
2095 /*AllowUnknown*/ true);
2096 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
2097 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2098 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2100 else if (ActualMVT == MVT::i16)
2103 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2105 AssignFn(i, ValVT, ValVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
2106 assert(!Res && "Call operand has unhandled type");
2109 assert(ArgLocs.size() == Ins.size());
2110 SmallVector<SDValue, 16> ArgValues;
2111 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2112 CCValAssign &VA = ArgLocs[i];
2114 if (Ins[i].Flags.isByVal()) {
2115 // Byval is used for HFAs in the PCS, but the system should work in a
2116 // non-compliant manner for larger structs.
2117 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2118 int Size = Ins[i].Flags.getByValSize();
2119 unsigned NumRegs = (Size + 7) / 8;
2121 // FIXME: This works on big-endian for composite byvals, which are the common
2122 // case. It should also work for fundamental types too.
2124 MFI->CreateFixedObject(8 * NumRegs, VA.getLocMemOffset(), false);
2125 SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrVT);
2126 InVals.push_back(FrameIdxN);
2131 if (VA.isRegLoc()) {
2132 // Arguments stored in registers.
2133 EVT RegVT = VA.getLocVT();
2136 const TargetRegisterClass *RC;
2138 if (RegVT == MVT::i32)
2139 RC = &AArch64::GPR32RegClass;
2140 else if (RegVT == MVT::i64)
2141 RC = &AArch64::GPR64RegClass;
2142 else if (RegVT == MVT::f16)
2143 RC = &AArch64::FPR16RegClass;
2144 else if (RegVT == MVT::f32)
2145 RC = &AArch64::FPR32RegClass;
2146 else if (RegVT == MVT::f64 || RegVT.is64BitVector())
2147 RC = &AArch64::FPR64RegClass;
2148 else if (RegVT == MVT::f128 || RegVT.is128BitVector())
2149 RC = &AArch64::FPR128RegClass;
2151 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2153 // Transform the arguments in physical registers into virtual ones.
2154 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2155 ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2157 // If this is an 8, 16 or 32-bit value, it is really passed promoted
2158 // to 64 bits. Insert an assert[sz]ext to capture this, then
2159 // truncate to the right size.
2160 switch (VA.getLocInfo()) {
2162 llvm_unreachable("Unknown loc info!");
2163 case CCValAssign::Full:
2165 case CCValAssign::BCvt:
2166 ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue);
2168 case CCValAssign::AExt:
2169 case CCValAssign::SExt:
2170 case CCValAssign::ZExt:
2171 // SelectionDAGBuilder will insert appropriate AssertZExt & AssertSExt
2172 // nodes after our lowering.
2173 assert(RegVT == Ins[i].VT && "incorrect register location selected");
2177 InVals.push_back(ArgValue);
2179 } else { // VA.isRegLoc()
2180 assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
2181 unsigned ArgOffset = VA.getLocMemOffset();
2182 unsigned ArgSize = VA.getValVT().getSizeInBits() / 8;
2184 uint32_t BEAlign = 0;
2185 if (!Subtarget->isLittleEndian() && ArgSize < 8 &&
2186 !Ins[i].Flags.isInConsecutiveRegs())
2187 BEAlign = 8 - ArgSize;
2189 int FI = MFI->CreateFixedObject(ArgSize, ArgOffset + BEAlign, true);
2191 // Create load nodes to retrieve arguments from the stack.
2192 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2195 // For NON_EXTLOAD, generic code in getLoad assert(ValVT == MemVT)
2196 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
2197 MVT MemVT = VA.getValVT();
2199 switch (VA.getLocInfo()) {
2202 case CCValAssign::BCvt:
2203 MemVT = VA.getLocVT();
2205 case CCValAssign::SExt:
2206 ExtType = ISD::SEXTLOAD;
2208 case CCValAssign::ZExt:
2209 ExtType = ISD::ZEXTLOAD;
2211 case CCValAssign::AExt:
2212 ExtType = ISD::EXTLOAD;
2216 ArgValue = DAG.getExtLoad(ExtType, DL, VA.getLocVT(), Chain, FIN,
2217 MachinePointerInfo::getFixedStack(FI),
2218 MemVT, false, false, false, 0);
2220 InVals.push_back(ArgValue);
2226 if (!Subtarget->isTargetDarwin()) {
2227 // The AAPCS variadic function ABI is identical to the non-variadic
2228 // one. As a result there may be more arguments in registers and we should
2229 // save them for future reference.
2230 saveVarArgRegisters(CCInfo, DAG, DL, Chain);
2233 AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>();
2234 // This will point to the next argument passed via stack.
2235 unsigned StackOffset = CCInfo.getNextStackOffset();
2236 // We currently pass all varargs at 8-byte alignment.
2237 StackOffset = ((StackOffset + 7) & ~7);
2238 AFI->setVarArgsStackIndex(MFI->CreateFixedObject(4, StackOffset, true));
2241 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2242 unsigned StackArgSize = CCInfo.getNextStackOffset();
2243 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2244 if (DoesCalleeRestoreStack(CallConv, TailCallOpt)) {
2245 // This is a non-standard ABI so by fiat I say we're allowed to make full
2246 // use of the stack area to be popped, which must be aligned to 16 bytes in
2248 StackArgSize = RoundUpToAlignment(StackArgSize, 16);
2250 // If we're expected to restore the stack (e.g. fastcc) then we'll be adding
2251 // a multiple of 16.
2252 FuncInfo->setArgumentStackToRestore(StackArgSize);
2254 // This realignment carries over to the available bytes below. Our own
2255 // callers will guarantee the space is free by giving an aligned value to
2258 // Even if we're not expected to free up the space, it's useful to know how
2259 // much is there while considering tail calls (because we can reuse it).
2260 FuncInfo->setBytesInStackArgArea(StackArgSize);
2265 void AArch64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
2266 SelectionDAG &DAG, SDLoc DL,
2267 SDValue &Chain) const {
2268 MachineFunction &MF = DAG.getMachineFunction();
2269 MachineFrameInfo *MFI = MF.getFrameInfo();
2270 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2271 auto PtrVT = getPointerTy(DAG.getDataLayout());
2273 SmallVector<SDValue, 8> MemOps;
2275 static const MCPhysReg GPRArgRegs[] = { AArch64::X0, AArch64::X1, AArch64::X2,
2276 AArch64::X3, AArch64::X4, AArch64::X5,
2277 AArch64::X6, AArch64::X7 };
2278 static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs);
2279 unsigned FirstVariadicGPR = CCInfo.getFirstUnallocated(GPRArgRegs);
2281 unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
2283 if (GPRSaveSize != 0) {
2284 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
2286 SDValue FIN = DAG.getFrameIndex(GPRIdx, PtrVT);
2288 for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
2289 unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &AArch64::GPR64RegClass);
2290 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
2292 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2293 MachinePointerInfo::getStack(i * 8), false, false, 0);
2294 MemOps.push_back(Store);
2296 DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getConstant(8, DL, PtrVT));
2299 FuncInfo->setVarArgsGPRIndex(GPRIdx);
2300 FuncInfo->setVarArgsGPRSize(GPRSaveSize);
2302 if (Subtarget->hasFPARMv8()) {
2303 static const MCPhysReg FPRArgRegs[] = {
2304 AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3,
2305 AArch64::Q4, AArch64::Q5, AArch64::Q6, AArch64::Q7};
2306 static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs);
2307 unsigned FirstVariadicFPR = CCInfo.getFirstUnallocated(FPRArgRegs);
2309 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
2311 if (FPRSaveSize != 0) {
2312 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
2314 SDValue FIN = DAG.getFrameIndex(FPRIdx, PtrVT);
2316 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
2317 unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &AArch64::FPR128RegClass);
2318 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f128);
2321 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2322 MachinePointerInfo::getStack(i * 16), false, false, 0);
2323 MemOps.push_back(Store);
2324 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN,
2325 DAG.getConstant(16, DL, PtrVT));
2328 FuncInfo->setVarArgsFPRIndex(FPRIdx);
2329 FuncInfo->setVarArgsFPRSize(FPRSaveSize);
2332 if (!MemOps.empty()) {
2333 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
2337 /// LowerCallResult - Lower the result values of a call into the
2338 /// appropriate copies out of appropriate physical registers.
2339 SDValue AArch64TargetLowering::LowerCallResult(
2340 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
2341 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2342 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
2343 SDValue ThisVal) const {
2344 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2345 ? RetCC_AArch64_WebKit_JS
2346 : RetCC_AArch64_AAPCS;
2347 // Assign locations to each value returned by this call.
2348 SmallVector<CCValAssign, 16> RVLocs;
2349 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2351 CCInfo.AnalyzeCallResult(Ins, RetCC);
2353 // Copy all of the result registers out of their specified physreg.
2354 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2355 CCValAssign VA = RVLocs[i];
2357 // Pass 'this' value directly from the argument to return value, to avoid
2358 // reg unit interference
2359 if (i == 0 && isThisReturn) {
2360 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 &&
2361 "unexpected return calling convention register assignment");
2362 InVals.push_back(ThisVal);
2367 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag);
2368 Chain = Val.getValue(1);
2369 InFlag = Val.getValue(2);
2371 switch (VA.getLocInfo()) {
2373 llvm_unreachable("Unknown loc info!");
2374 case CCValAssign::Full:
2376 case CCValAssign::BCvt:
2377 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
2381 InVals.push_back(Val);
2387 bool AArch64TargetLowering::isEligibleForTailCallOptimization(
2388 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
2389 bool isCalleeStructRet, bool isCallerStructRet,
2390 const SmallVectorImpl<ISD::OutputArg> &Outs,
2391 const SmallVectorImpl<SDValue> &OutVals,
2392 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
2393 // For CallingConv::C this function knows whether the ABI needs
2394 // changing. That's not true for other conventions so they will have to opt in
2396 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
2399 const MachineFunction &MF = DAG.getMachineFunction();
2400 const Function *CallerF = MF.getFunction();
2401 CallingConv::ID CallerCC = CallerF->getCallingConv();
2402 bool CCMatch = CallerCC == CalleeCC;
2404 // Byval parameters hand the function a pointer directly into the stack area
2405 // we want to reuse during a tail call. Working around this *is* possible (see
2406 // X86) but less efficient and uglier in LowerCall.
2407 for (Function::const_arg_iterator i = CallerF->arg_begin(),
2408 e = CallerF->arg_end();
2410 if (i->hasByValAttr())
2413 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
2414 if (IsTailCallConvention(CalleeCC) && CCMatch)
2419 // Externally-defined functions with weak linkage should not be
2420 // tail-called on AArch64 when the OS does not support dynamic
2421 // pre-emption of symbols, as the AAELF spec requires normal calls
2422 // to undefined weak functions to be replaced with a NOP or jump to the
2423 // next instruction. The behaviour of branch instructions in this
2424 // situation (as used for tail calls) is implementation-defined, so we
2425 // cannot rely on the linker replacing the tail call with a return.
2426 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2427 const GlobalValue *GV = G->getGlobal();
2428 const Triple &TT = getTargetMachine().getTargetTriple();
2429 if (GV->hasExternalWeakLinkage() &&
2430 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2434 // Now we search for cases where we can use a tail call without changing the
2435 // ABI. Sibcall is used in some places (particularly gcc) to refer to this
2438 // I want anyone implementing a new calling convention to think long and hard
2439 // about this assert.
2440 assert((!isVarArg || CalleeCC == CallingConv::C) &&
2441 "Unexpected variadic calling convention");
2443 if (isVarArg && !Outs.empty()) {
2444 // At least two cases here: if caller is fastcc then we can't have any
2445 // memory arguments (we'd be expected to clean up the stack afterwards). If
2446 // caller is C then we could potentially use its argument area.
2448 // FIXME: for now we take the most conservative of these in both cases:
2449 // disallow all variadic memory operands.
2450 SmallVector<CCValAssign, 16> ArgLocs;
2451 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2454 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, true));
2455 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
2456 if (!ArgLocs[i].isRegLoc())
2460 // If the calling conventions do not match, then we'd better make sure the
2461 // results are returned in the same way as what the caller expects.
2463 SmallVector<CCValAssign, 16> RVLocs1;
2464 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2466 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForCall(CalleeCC, isVarArg));
2468 SmallVector<CCValAssign, 16> RVLocs2;
2469 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2471 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForCall(CallerCC, isVarArg));
2473 if (RVLocs1.size() != RVLocs2.size())
2475 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2476 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2478 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2480 if (RVLocs1[i].isRegLoc()) {
2481 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2484 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2490 // Nothing more to check if the callee is taking no arguments
2494 SmallVector<CCValAssign, 16> ArgLocs;
2495 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2498 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, isVarArg));
2500 const AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2502 // If the stack arguments for this call would fit into our own save area then
2503 // the call can be made tail.
2504 return CCInfo.getNextStackOffset() <= FuncInfo->getBytesInStackArgArea();
2507 SDValue AArch64TargetLowering::addTokenForArgument(SDValue Chain,
2509 MachineFrameInfo *MFI,
2510 int ClobberedFI) const {
2511 SmallVector<SDValue, 8> ArgChains;
2512 int64_t FirstByte = MFI->getObjectOffset(ClobberedFI);
2513 int64_t LastByte = FirstByte + MFI->getObjectSize(ClobberedFI) - 1;
2515 // Include the original chain at the beginning of the list. When this is
2516 // used by target LowerCall hooks, this helps legalize find the
2517 // CALLSEQ_BEGIN node.
2518 ArgChains.push_back(Chain);
2520 // Add a chain value for each stack argument corresponding
2521 for (SDNode::use_iterator U = DAG.getEntryNode().getNode()->use_begin(),
2522 UE = DAG.getEntryNode().getNode()->use_end();
2524 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
2525 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
2526 if (FI->getIndex() < 0) {
2527 int64_t InFirstByte = MFI->getObjectOffset(FI->getIndex());
2528 int64_t InLastByte = InFirstByte;
2529 InLastByte += MFI->getObjectSize(FI->getIndex()) - 1;
2531 if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
2532 (FirstByte <= InFirstByte && InFirstByte <= LastByte))
2533 ArgChains.push_back(SDValue(L, 1));
2536 // Build a tokenfactor for all the chains.
2537 return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
2540 bool AArch64TargetLowering::DoesCalleeRestoreStack(CallingConv::ID CallCC,
2541 bool TailCallOpt) const {
2542 return CallCC == CallingConv::Fast && TailCallOpt;
2545 bool AArch64TargetLowering::IsTailCallConvention(CallingConv::ID CallCC) const {
2546 return CallCC == CallingConv::Fast;
2549 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
2550 /// and add input and output parameter nodes.
2552 AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
2553 SmallVectorImpl<SDValue> &InVals) const {
2554 SelectionDAG &DAG = CLI.DAG;
2556 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2557 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2558 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2559 SDValue Chain = CLI.Chain;
2560 SDValue Callee = CLI.Callee;
2561 bool &IsTailCall = CLI.IsTailCall;
2562 CallingConv::ID CallConv = CLI.CallConv;
2563 bool IsVarArg = CLI.IsVarArg;
2565 MachineFunction &MF = DAG.getMachineFunction();
2566 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
2567 bool IsThisReturn = false;
2569 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2570 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2571 bool IsSibCall = false;
2574 // Check if it's really possible to do a tail call.
2575 IsTailCall = isEligibleForTailCallOptimization(
2576 Callee, CallConv, IsVarArg, IsStructRet,
2577 MF.getFunction()->hasStructRetAttr(), Outs, OutVals, Ins, DAG);
2578 if (!IsTailCall && CLI.CS && CLI.CS->isMustTailCall())
2579 report_fatal_error("failed to perform tail call elimination on a call "
2580 "site marked musttail");
2582 // A sibling call is one where we're under the usual C ABI and not planning
2583 // to change that but can still do a tail call:
2584 if (!TailCallOpt && IsTailCall)
2591 // Analyze operands of the call, assigning locations to each operand.
2592 SmallVector<CCValAssign, 16> ArgLocs;
2593 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
2597 // Handle fixed and variable vector arguments differently.
2598 // Variable vector arguments always go into memory.
2599 unsigned NumArgs = Outs.size();
2601 for (unsigned i = 0; i != NumArgs; ++i) {
2602 MVT ArgVT = Outs[i].VT;
2603 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2604 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
2605 /*IsVarArg=*/ !Outs[i].IsFixed);
2606 bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
2607 assert(!Res && "Call operand has unhandled type");
2611 // At this point, Outs[].VT may already be promoted to i32. To correctly
2612 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2613 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
2614 // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here
2615 // we use a special version of AnalyzeCallOperands to pass in ValVT and
2617 unsigned NumArgs = Outs.size();
2618 for (unsigned i = 0; i != NumArgs; ++i) {
2619 MVT ValVT = Outs[i].VT;
2620 // Get type of the original argument.
2621 EVT ActualVT = getValueType(DAG.getDataLayout(),
2622 CLI.getArgs()[Outs[i].OrigArgIndex].Ty,
2623 /*AllowUnknown*/ true);
2624 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT;
2625 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2626 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2627 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2629 else if (ActualMVT == MVT::i16)
2632 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2633 bool Res = AssignFn(i, ValVT, ValVT, CCValAssign::Full, ArgFlags, CCInfo);
2634 assert(!Res && "Call operand has unhandled type");
2639 // Get a count of how many bytes are to be pushed on the stack.
2640 unsigned NumBytes = CCInfo.getNextStackOffset();
2643 // Since we're not changing the ABI to make this a tail call, the memory
2644 // operands are already available in the caller's incoming argument space.
2648 // FPDiff is the byte offset of the call's argument area from the callee's.
2649 // Stores to callee stack arguments will be placed in FixedStackSlots offset
2650 // by this amount for a tail call. In a sibling call it must be 0 because the
2651 // caller will deallocate the entire stack and the callee still expects its
2652 // arguments to begin at SP+0. Completely unused for non-tail calls.
2655 if (IsTailCall && !IsSibCall) {
2656 unsigned NumReusableBytes = FuncInfo->getBytesInStackArgArea();
2658 // Since callee will pop argument stack as a tail call, we must keep the
2659 // popped size 16-byte aligned.
2660 NumBytes = RoundUpToAlignment(NumBytes, 16);
2662 // FPDiff will be negative if this tail call requires more space than we
2663 // would automatically have in our incoming argument space. Positive if we
2664 // can actually shrink the stack.
2665 FPDiff = NumReusableBytes - NumBytes;
2667 // The stack pointer must be 16-byte aligned at all times it's used for a
2668 // memory operation, which in practice means at *all* times and in
2669 // particular across call boundaries. Therefore our own arguments started at
2670 // a 16-byte aligned SP and the delta applied for the tail call should
2671 // satisfy the same constraint.
2672 assert(FPDiff % 16 == 0 && "unaligned stack on tail call");
2675 // Adjust the stack pointer for the new arguments...
2676 // These operations are automatically eliminated by the prolog/epilog pass
2678 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, DL,
2682 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, AArch64::SP,
2683 getPointerTy(DAG.getDataLayout()));
2685 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2686 SmallVector<SDValue, 8> MemOpChains;
2687 auto PtrVT = getPointerTy(DAG.getDataLayout());
2689 // Walk the register/memloc assignments, inserting copies/loads.
2690 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2691 ++i, ++realArgIdx) {
2692 CCValAssign &VA = ArgLocs[i];
2693 SDValue Arg = OutVals[realArgIdx];
2694 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2696 // Promote the value if needed.
2697 switch (VA.getLocInfo()) {
2699 llvm_unreachable("Unknown loc info!");
2700 case CCValAssign::Full:
2702 case CCValAssign::SExt:
2703 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
2705 case CCValAssign::ZExt:
2706 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2708 case CCValAssign::AExt:
2709 if (Outs[realArgIdx].ArgVT == MVT::i1) {
2710 // AAPCS requires i1 to be zero-extended to 8-bits by the caller.
2711 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
2712 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i8, Arg);
2714 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
2716 case CCValAssign::BCvt:
2717 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2719 case CCValAssign::FPExt:
2720 Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
2724 if (VA.isRegLoc()) {
2725 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i64) {
2726 assert(VA.getLocVT() == MVT::i64 &&
2727 "unexpected calling convention register assignment");
2728 assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
2729 "unexpected use of 'returned'");
2730 IsThisReturn = true;
2732 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2734 assert(VA.isMemLoc());
2737 MachinePointerInfo DstInfo;
2739 // FIXME: This works on big-endian for composite byvals, which are the
2740 // common case. It should also work for fundamental types too.
2741 uint32_t BEAlign = 0;
2742 unsigned OpSize = Flags.isByVal() ? Flags.getByValSize() * 8
2743 : VA.getValVT().getSizeInBits();
2744 OpSize = (OpSize + 7) / 8;
2745 if (!Subtarget->isLittleEndian() && !Flags.isByVal() &&
2746 !Flags.isInConsecutiveRegs()) {
2748 BEAlign = 8 - OpSize;
2750 unsigned LocMemOffset = VA.getLocMemOffset();
2751 int32_t Offset = LocMemOffset + BEAlign;
2752 SDValue PtrOff = DAG.getIntPtrConstant(Offset, DL);
2753 PtrOff = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, PtrOff);
2756 Offset = Offset + FPDiff;
2757 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2759 DstAddr = DAG.getFrameIndex(FI, PtrVT);
2760 DstInfo = MachinePointerInfo::getFixedStack(FI);
2762 // Make sure any stack arguments overlapping with where we're storing
2763 // are loaded before this eventual operation. Otherwise they'll be
2765 Chain = addTokenForArgument(Chain, DAG, MF.getFrameInfo(), FI);
2767 SDValue PtrOff = DAG.getIntPtrConstant(Offset, DL);
2769 DstAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, PtrOff);
2770 DstInfo = MachinePointerInfo::getStack(LocMemOffset);
2773 if (Outs[i].Flags.isByVal()) {
2775 DAG.getConstant(Outs[i].Flags.getByValSize(), DL, MVT::i64);
2776 SDValue Cpy = DAG.getMemcpy(
2777 Chain, DL, DstAddr, Arg, SizeNode, Outs[i].Flags.getByValAlign(),
2778 /*isVol = */ false, /*AlwaysInline = */ false,
2779 /*isTailCall = */ false,
2780 DstInfo, MachinePointerInfo());
2782 MemOpChains.push_back(Cpy);
2784 // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
2785 // promoted to a legal register type i32, we should truncate Arg back to
2787 if (VA.getValVT() == MVT::i1 || VA.getValVT() == MVT::i8 ||
2788 VA.getValVT() == MVT::i16)
2789 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Arg);
2792 DAG.getStore(Chain, DL, Arg, DstAddr, DstInfo, false, false, 0);
2793 MemOpChains.push_back(Store);
2798 if (!MemOpChains.empty())
2799 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
2801 // Build a sequence of copy-to-reg nodes chained together with token chain
2802 // and flag operands which copy the outgoing args into the appropriate regs.
2804 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2805 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[i].first,
2806 RegsToPass[i].second, InFlag);
2807 InFlag = Chain.getValue(1);
2810 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2811 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2812 // node so that legalize doesn't hack it.
2813 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
2814 Subtarget->isTargetMachO()) {
2815 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2816 const GlobalValue *GV = G->getGlobal();
2817 bool InternalLinkage = GV->hasInternalLinkage();
2818 if (InternalLinkage)
2819 Callee = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 0);
2822 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_GOT);
2823 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, Callee);
2825 } else if (ExternalSymbolSDNode *S =
2826 dyn_cast<ExternalSymbolSDNode>(Callee)) {
2827 const char *Sym = S->getSymbol();
2828 Callee = DAG.getTargetExternalSymbol(Sym, PtrVT, AArch64II::MO_GOT);
2829 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, Callee);
2831 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2832 const GlobalValue *GV = G->getGlobal();
2833 Callee = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 0);
2834 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2835 const char *Sym = S->getSymbol();
2836 Callee = DAG.getTargetExternalSymbol(Sym, PtrVT, 0);
2839 // We don't usually want to end the call-sequence here because we would tidy
2840 // the frame up *after* the call, however in the ABI-changing tail-call case
2841 // we've carefully laid out the parameters so that when sp is reset they'll be
2842 // in the correct location.
2843 if (IsTailCall && !IsSibCall) {
2844 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, DL, true),
2845 DAG.getIntPtrConstant(0, DL, true), InFlag, DL);
2846 InFlag = Chain.getValue(1);
2849 std::vector<SDValue> Ops;
2850 Ops.push_back(Chain);
2851 Ops.push_back(Callee);
2854 // Each tail call may have to adjust the stack by a different amount, so
2855 // this information must travel along with the operation for eventual
2856 // consumption by emitEpilogue.
2857 Ops.push_back(DAG.getTargetConstant(FPDiff, DL, MVT::i32));
2860 // Add argument registers to the end of the list so that they are known live
2862 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2863 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2864 RegsToPass[i].second.getValueType()));
2866 // Add a register mask operand representing the call-preserved registers.
2867 const uint32_t *Mask;
2868 const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();
2870 // For 'this' returns, use the X0-preserving mask if applicable
2871 Mask = TRI->getThisReturnPreservedMask(MF, CallConv);
2873 IsThisReturn = false;
2874 Mask = TRI->getCallPreservedMask(MF, CallConv);
2877 Mask = TRI->getCallPreservedMask(MF, CallConv);
2879 assert(Mask && "Missing call preserved mask for calling convention");
2880 Ops.push_back(DAG.getRegisterMask(Mask));
2882 if (InFlag.getNode())
2883 Ops.push_back(InFlag);
2885 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2887 // If we're doing a tall call, use a TC_RETURN here rather than an
2888 // actual call instruction.
2890 MF.getFrameInfo()->setHasTailCall();
2891 return DAG.getNode(AArch64ISD::TC_RETURN, DL, NodeTys, Ops);
2894 // Returns a chain and a flag for retval copy to use.
2895 Chain = DAG.getNode(AArch64ISD::CALL, DL, NodeTys, Ops);
2896 InFlag = Chain.getValue(1);
2898 uint64_t CalleePopBytes = DoesCalleeRestoreStack(CallConv, TailCallOpt)
2899 ? RoundUpToAlignment(NumBytes, 16)
2902 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, DL, true),
2903 DAG.getIntPtrConstant(CalleePopBytes, DL, true),
2906 InFlag = Chain.getValue(1);
2908 // Handle result values, copying them out of physregs into vregs that we
2910 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
2911 InVals, IsThisReturn,
2912 IsThisReturn ? OutVals[0] : SDValue());
2915 bool AArch64TargetLowering::CanLowerReturn(
2916 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2917 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2918 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2919 ? RetCC_AArch64_WebKit_JS
2920 : RetCC_AArch64_AAPCS;
2921 SmallVector<CCValAssign, 16> RVLocs;
2922 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2923 return CCInfo.CheckReturn(Outs, RetCC);
2927 AArch64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2929 const SmallVectorImpl<ISD::OutputArg> &Outs,
2930 const SmallVectorImpl<SDValue> &OutVals,
2931 SDLoc DL, SelectionDAG &DAG) const {
2932 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2933 ? RetCC_AArch64_WebKit_JS
2934 : RetCC_AArch64_AAPCS;
2935 SmallVector<CCValAssign, 16> RVLocs;
2936 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2938 CCInfo.AnalyzeReturn(Outs, RetCC);
2940 // Copy the result values into the output registers.
2942 SmallVector<SDValue, 4> RetOps(1, Chain);
2943 for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size();
2944 ++i, ++realRVLocIdx) {
2945 CCValAssign &VA = RVLocs[i];
2946 assert(VA.isRegLoc() && "Can only return in registers!");
2947 SDValue Arg = OutVals[realRVLocIdx];
2949 switch (VA.getLocInfo()) {
2951 llvm_unreachable("Unknown loc info!");
2952 case CCValAssign::Full:
2953 if (Outs[i].ArgVT == MVT::i1) {
2954 // AAPCS requires i1 to be zero-extended to i8 by the producer of the
2955 // value. This is strictly redundant on Darwin (which uses "zeroext
2956 // i1"), but will be optimised out before ISel.
2957 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
2958 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2961 case CCValAssign::BCvt:
2962 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2966 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
2967 Flag = Chain.getValue(1);
2968 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2971 RetOps[0] = Chain; // Update chain.
2973 // Add the flag if we have it.
2975 RetOps.push_back(Flag);
2977 return DAG.getNode(AArch64ISD::RET_FLAG, DL, MVT::Other, RetOps);
2980 //===----------------------------------------------------------------------===//
2981 // Other Lowering Code
2982 //===----------------------------------------------------------------------===//
2984 SDValue AArch64TargetLowering::LowerGlobalAddress(SDValue Op,
2985 SelectionDAG &DAG) const {
2986 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2988 const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
2989 const GlobalValue *GV = GN->getGlobal();
2990 unsigned char OpFlags =
2991 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
2993 assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
2994 "unexpected offset in global node");
2996 // This also catched the large code model case for Darwin.
2997 if ((OpFlags & AArch64II::MO_GOT) != 0) {
2998 SDValue GotAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
2999 // FIXME: Once remat is capable of dealing with instructions with register
3000 // operands, expand this into two nodes instead of using a wrapper node.
3001 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
3004 if ((OpFlags & AArch64II::MO_CONSTPOOL) != 0) {
3005 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
3006 "use of MO_CONSTPOOL only supported on small model");
3007 SDValue Hi = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, AArch64II::MO_PAGE);
3008 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3009 unsigned char LoFlags = AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
3010 SDValue Lo = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, LoFlags);
3011 SDValue PoolAddr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3012 SDValue GlobalAddr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), PoolAddr,
3013 MachinePointerInfo::getConstantPool(),
3014 /*isVolatile=*/ false,
3015 /*isNonTemporal=*/ true,
3016 /*isInvariant=*/ true, 8);
3017 if (GN->getOffset() != 0)
3018 return DAG.getNode(ISD::ADD, DL, PtrVT, GlobalAddr,
3019 DAG.getConstant(GN->getOffset(), DL, PtrVT));
3023 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
3024 const unsigned char MO_NC = AArch64II::MO_NC;
3026 AArch64ISD::WrapperLarge, DL, PtrVT,
3027 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G3),
3028 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
3029 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
3030 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
3032 // Use ADRP/ADD or ADRP/LDR for everything else: the small model on ELF and
3033 // the only correct model on Darwin.
3034 SDValue Hi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
3035 OpFlags | AArch64II::MO_PAGE);
3036 unsigned char LoFlags = OpFlags | AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
3037 SDValue Lo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, LoFlags);
3039 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3040 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3044 /// \brief Convert a TLS address reference into the correct sequence of loads
3045 /// and calls to compute the variable's address (for Darwin, currently) and
3046 /// return an SDValue containing the final node.
3048 /// Darwin only has one TLS scheme which must be capable of dealing with the
3049 /// fully general situation, in the worst case. This means:
3050 /// + "extern __thread" declaration.
3051 /// + Defined in a possibly unknown dynamic library.
3053 /// The general system is that each __thread variable has a [3 x i64] descriptor
3054 /// which contains information used by the runtime to calculate the address. The
3055 /// only part of this the compiler needs to know about is the first xword, which
3056 /// contains a function pointer that must be called with the address of the
3057 /// entire descriptor in "x0".
3059 /// Since this descriptor may be in a different unit, in general even the
3060 /// descriptor must be accessed via an indirect load. The "ideal" code sequence
3062 /// adrp x0, _var@TLVPPAGE
3063 /// ldr x0, [x0, _var@TLVPPAGEOFF] ; x0 now contains address of descriptor
3064 /// ldr x1, [x0] ; x1 contains 1st entry of descriptor,
3065 /// ; the function pointer
3066 /// blr x1 ; Uses descriptor address in x0
3067 /// ; Address of _var is now in x0.
3069 /// If the address of _var's descriptor *is* known to the linker, then it can
3070 /// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
3071 /// a slight efficiency gain.
3073 AArch64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
3074 SelectionDAG &DAG) const {
3075 assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");
3078 MVT PtrVT = getPointerTy(DAG.getDataLayout());
3079 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3082 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3083 SDValue DescAddr = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TLVPAddr);
3085 // The first entry in the descriptor is a function pointer that we must call
3086 // to obtain the address of the variable.
3087 SDValue Chain = DAG.getEntryNode();
3088 SDValue FuncTLVGet =
3089 DAG.getLoad(MVT::i64, DL, Chain, DescAddr, MachinePointerInfo::getGOT(),
3090 false, true, true, 8);
3091 Chain = FuncTLVGet.getValue(1);
3093 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3094 MFI->setAdjustsStack(true);
3096 // TLS calls preserve all registers except those that absolutely must be
3097 // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
3099 const uint32_t *Mask =
3100 Subtarget->getRegisterInfo()->getTLSCallPreservedMask();
3102 // Finally, we can make the call. This is just a degenerate version of a
3103 // normal AArch64 call node: x0 takes the address of the descriptor, and
3104 // returns the address of the variable in this thread.
3105 Chain = DAG.getCopyToReg(Chain, DL, AArch64::X0, DescAddr, SDValue());
3107 DAG.getNode(AArch64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
3108 Chain, FuncTLVGet, DAG.getRegister(AArch64::X0, MVT::i64),
3109 DAG.getRegisterMask(Mask), Chain.getValue(1));
3110 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Chain.getValue(1));
3113 /// When accessing thread-local variables under either the general-dynamic or
3114 /// local-dynamic system, we make a "TLS-descriptor" call. The variable will
3115 /// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
3116 /// is a function pointer to carry out the resolution.
3118 /// The sequence is:
3119 /// adrp x0, :tlsdesc:var
3120 /// ldr x1, [x0, #:tlsdesc_lo12:var]
3121 /// add x0, x0, #:tlsdesc_lo12:var
3122 /// .tlsdesccall var
3124 /// (TPIDR_EL0 offset now in x0)
3126 /// The above sequence must be produced unscheduled, to enable the linker to
3127 /// optimize/relax this sequence.
3128 /// Therefore, a pseudo-instruction (TLSDESC_CALLSEQ) is used to represent the
3129 /// above sequence, and expanded really late in the compilation flow, to ensure
3130 /// the sequence is produced as per above.
3131 SDValue AArch64TargetLowering::LowerELFTLSDescCallSeq(SDValue SymAddr, SDLoc DL,
3132 SelectionDAG &DAG) const {
3133 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3135 SDValue Chain = DAG.getEntryNode();
3136 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3138 SmallVector<SDValue, 2> Ops;
3139 Ops.push_back(Chain);
3140 Ops.push_back(SymAddr);
3142 Chain = DAG.getNode(AArch64ISD::TLSDESC_CALLSEQ, DL, NodeTys, Ops);
3143 SDValue Glue = Chain.getValue(1);
3145 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Glue);
3149 AArch64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
3150 SelectionDAG &DAG) const {
3151 assert(Subtarget->isTargetELF() && "This function expects an ELF target");
3152 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
3153 "ELF TLS only supported in small memory model");
3154 // Different choices can be made for the maximum size of the TLS area for a
3155 // module. For the small address model, the default TLS size is 16MiB and the
3156 // maximum TLS size is 4GiB.
3157 // FIXME: add -mtls-size command line option and make it control the 16MiB
3158 // vs. 4GiB code sequence generation.
3159 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3161 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
3162 if (!EnableAArch64ELFLocalDynamicTLSGeneration) {
3163 if (Model == TLSModel::LocalDynamic)
3164 Model = TLSModel::GeneralDynamic;
3168 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3170 const GlobalValue *GV = GA->getGlobal();
3172 SDValue ThreadBase = DAG.getNode(AArch64ISD::THREAD_POINTER, DL, PtrVT);
3174 if (Model == TLSModel::LocalExec) {
3175 SDValue HiVar = DAG.getTargetGlobalAddress(
3176 GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_HI12);
3177 SDValue LoVar = DAG.getTargetGlobalAddress(
3179 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3181 SDValue TPWithOff_lo =
3182 SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, ThreadBase,
3184 DAG.getTargetConstant(0, DL, MVT::i32)),
3187 SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPWithOff_lo,
3189 DAG.getTargetConstant(0, DL, MVT::i32)),
3192 } else if (Model == TLSModel::InitialExec) {
3193 TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3194 TPOff = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TPOff);
3195 } else if (Model == TLSModel::LocalDynamic) {
3196 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
3197 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
3198 // the beginning of the module's TLS region, followed by a DTPREL offset
3201 // These accesses will need deduplicating if there's more than one.
3202 AArch64FunctionInfo *MFI =
3203 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
3204 MFI->incNumLocalDynamicTLSAccesses();
3206 // The call needs a relocation too for linker relaxation. It doesn't make
3207 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
3209 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
3212 // Now we can calculate the offset from TPIDR_EL0 to this module's
3213 // thread-local area.
3214 TPOff = LowerELFTLSDescCallSeq(SymAddr, DL, DAG);
3216 // Now use :dtprel_whatever: operations to calculate this variable's offset
3217 // in its thread-storage area.
3218 SDValue HiVar = DAG.getTargetGlobalAddress(
3219 GV, DL, MVT::i64, 0, AArch64II::MO_TLS | AArch64II::MO_HI12);
3220 SDValue LoVar = DAG.getTargetGlobalAddress(
3221 GV, DL, MVT::i64, 0,
3222 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3224 TPOff = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPOff, HiVar,
3225 DAG.getTargetConstant(0, DL, MVT::i32)),
3227 TPOff = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPOff, LoVar,
3228 DAG.getTargetConstant(0, DL, MVT::i32)),
3230 } else if (Model == TLSModel::GeneralDynamic) {
3231 // The call needs a relocation too for linker relaxation. It doesn't make
3232 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
3235 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3237 // Finally we can make a call to calculate the offset from tpidr_el0.
3238 TPOff = LowerELFTLSDescCallSeq(SymAddr, DL, DAG);
3240 llvm_unreachable("Unsupported ELF TLS access model");
3242 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
3245 SDValue AArch64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
3246 SelectionDAG &DAG) const {
3247 if (Subtarget->isTargetDarwin())
3248 return LowerDarwinGlobalTLSAddress(Op, DAG);
3249 else if (Subtarget->isTargetELF())
3250 return LowerELFGlobalTLSAddress(Op, DAG);
3252 llvm_unreachable("Unexpected platform trying to use TLS");
3254 SDValue AArch64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3255 SDValue Chain = Op.getOperand(0);
3256 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3257 SDValue LHS = Op.getOperand(2);
3258 SDValue RHS = Op.getOperand(3);
3259 SDValue Dest = Op.getOperand(4);
3262 // Handle f128 first, since lowering it will result in comparing the return
3263 // value of a libcall against zero, which is just what the rest of LowerBR_CC
3264 // is expecting to deal with.
3265 if (LHS.getValueType() == MVT::f128) {
3266 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3268 // If softenSetCCOperands returned a scalar, we need to compare the result
3269 // against zero to select between true and false values.
3270 if (!RHS.getNode()) {
3271 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3276 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
3278 unsigned Opc = LHS.getOpcode();
3279 if (LHS.getResNo() == 1 && isa<ConstantSDNode>(RHS) &&
3280 cast<ConstantSDNode>(RHS)->isOne() &&
3281 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3282 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3283 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
3284 "Unexpected condition code.");
3285 // Only lower legal XALUO ops.
3286 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
3289 // The actual operation with overflow check.
3290 AArch64CC::CondCode OFCC;
3291 SDValue Value, Overflow;
3292 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, LHS.getValue(0), DAG);
3294 if (CC == ISD::SETNE)
3295 OFCC = getInvertedCondCode(OFCC);
3296 SDValue CCVal = DAG.getConstant(OFCC, dl, MVT::i32);
3298 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3302 if (LHS.getValueType().isInteger()) {
3303 assert((LHS.getValueType() == RHS.getValueType()) &&
3304 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3306 // If the RHS of the comparison is zero, we can potentially fold this
3307 // to a specialized branch.
3308 const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
3309 if (RHSC && RHSC->getZExtValue() == 0) {
3310 if (CC == ISD::SETEQ) {
3311 // See if we can use a TBZ to fold in an AND as well.
3312 // TBZ has a smaller branch displacement than CBZ. If the offset is
3313 // out of bounds, a late MI-layer pass rewrites branches.
3314 // 403.gcc is an example that hits this case.
3315 if (LHS.getOpcode() == ISD::AND &&
3316 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3317 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3318 SDValue Test = LHS.getOperand(0);
3319 uint64_t Mask = LHS.getConstantOperandVal(1);
3320 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, Test,
3321 DAG.getConstant(Log2_64(Mask), dl, MVT::i64),
3325 return DAG.getNode(AArch64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest);
3326 } else if (CC == ISD::SETNE) {
3327 // See if we can use a TBZ to fold in an AND as well.
3328 // TBZ has a smaller branch displacement than CBZ. If the offset is
3329 // out of bounds, a late MI-layer pass rewrites branches.
3330 // 403.gcc is an example that hits this case.
3331 if (LHS.getOpcode() == ISD::AND &&
3332 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3333 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3334 SDValue Test = LHS.getOperand(0);
3335 uint64_t Mask = LHS.getConstantOperandVal(1);
3336 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, Test,
3337 DAG.getConstant(Log2_64(Mask), dl, MVT::i64),
3341 return DAG.getNode(AArch64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest);
3342 } else if (CC == ISD::SETLT && LHS.getOpcode() != ISD::AND) {
3343 // Don't combine AND since emitComparison converts the AND to an ANDS
3344 // (a.k.a. TST) and the test in the test bit and branch instruction
3345 // becomes redundant. This would also increase register pressure.
3346 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
3347 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, LHS,
3348 DAG.getConstant(Mask, dl, MVT::i64), Dest);
3351 if (RHSC && RHSC->getSExtValue() == -1 && CC == ISD::SETGT &&
3352 LHS.getOpcode() != ISD::AND) {
3353 // Don't combine AND since emitComparison converts the AND to an ANDS
3354 // (a.k.a. TST) and the test in the test bit and branch instruction
3355 // becomes redundant. This would also increase register pressure.
3356 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
3357 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, LHS,
3358 DAG.getConstant(Mask, dl, MVT::i64), Dest);
3362 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3363 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3367 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3369 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
3370 // clean. Some of them require two branches to implement.
3371 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3372 AArch64CC::CondCode CC1, CC2;
3373 changeFPCCToAArch64CC(CC, CC1, CC2);
3374 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
3376 DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp);
3377 if (CC2 != AArch64CC::AL) {
3378 SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32);
3379 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val,
3386 SDValue AArch64TargetLowering::LowerFCOPYSIGN(SDValue Op,
3387 SelectionDAG &DAG) const {
3388 EVT VT = Op.getValueType();
3391 SDValue In1 = Op.getOperand(0);
3392 SDValue In2 = Op.getOperand(1);
3393 EVT SrcVT = In2.getValueType();
3395 if (SrcVT == MVT::f32 && VT == MVT::f64)
3396 In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2);
3397 else if (SrcVT == MVT::f64 && VT == MVT::f32)
3398 In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2,
3399 DAG.getIntPtrConstant(0, DL));
3401 // FIXME: Src type is different, bail out for now. Can VT really be a
3409 SDValue VecVal1, VecVal2;
3410 if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) {
3413 EltMask = 0x80000000ULL;
3415 if (!VT.isVector()) {
3416 VecVal1 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3417 DAG.getUNDEF(VecVT), In1);
3418 VecVal2 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3419 DAG.getUNDEF(VecVT), In2);
3421 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3422 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3424 } else if (VT == MVT::f64 || VT == MVT::v2f64) {
3428 // We want to materialize a mask with the high bit set, but the AdvSIMD
3429 // immediate moves cannot materialize that in a single instruction for
3430 // 64-bit elements. Instead, materialize zero and then negate it.
3433 if (!VT.isVector()) {
3434 VecVal1 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3435 DAG.getUNDEF(VecVT), In1);
3436 VecVal2 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3437 DAG.getUNDEF(VecVT), In2);
3439 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3440 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3443 llvm_unreachable("Invalid type for copysign!");
3446 SDValue BuildVec = DAG.getConstant(EltMask, DL, VecVT);
3448 // If we couldn't materialize the mask above, then the mask vector will be
3449 // the zero vector, and we need to negate it here.
3450 if (VT == MVT::f64 || VT == MVT::v2f64) {
3451 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec);
3452 BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec);
3453 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec);
3457 DAG.getNode(AArch64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec);
3460 return DAG.getTargetExtractSubreg(AArch64::ssub, DL, VT, Sel);
3461 else if (VT == MVT::f64)
3462 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, VT, Sel);
3464 return DAG.getNode(ISD::BITCAST, DL, VT, Sel);
3467 SDValue AArch64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
3468 if (DAG.getMachineFunction().getFunction()->hasFnAttribute(
3469 Attribute::NoImplicitFloat))
3472 if (!Subtarget->hasNEON())
3475 // While there is no integer popcount instruction, it can
3476 // be more efficiently lowered to the following sequence that uses
3477 // AdvSIMD registers/instructions as long as the copies to/from
3478 // the AdvSIMD registers are cheap.
3479 // FMOV D0, X0 // copy 64-bit int to vector, high bits zero'd
3480 // CNT V0.8B, V0.8B // 8xbyte pop-counts
3481 // ADDV B0, V0.8B // sum 8xbyte pop-counts
3482 // UMOV X0, V0.B[0] // copy byte result back to integer reg
3483 SDValue Val = Op.getOperand(0);
3485 EVT VT = Op.getValueType();
3488 Val = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Val);
3489 Val = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);
3491 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, Val);
3492 SDValue UaddLV = DAG.getNode(
3493 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
3494 DAG.getConstant(Intrinsic::aarch64_neon_uaddlv, DL, MVT::i32), CtPop);
3497 UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV);
3501 SDValue AArch64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
3503 if (Op.getValueType().isVector())
3504 return LowerVSETCC(Op, DAG);
3506 SDValue LHS = Op.getOperand(0);
3507 SDValue RHS = Op.getOperand(1);
3508 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
3511 // We chose ZeroOrOneBooleanContents, so use zero and one.
3512 EVT VT = Op.getValueType();
3513 SDValue TVal = DAG.getConstant(1, dl, VT);
3514 SDValue FVal = DAG.getConstant(0, dl, VT);
3516 // Handle f128 first, since one possible outcome is a normal integer
3517 // comparison which gets picked up by the next if statement.
3518 if (LHS.getValueType() == MVT::f128) {
3519 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3521 // If softenSetCCOperands returned a scalar, use it.
3522 if (!RHS.getNode()) {
3523 assert(LHS.getValueType() == Op.getValueType() &&
3524 "Unexpected setcc expansion!");
3529 if (LHS.getValueType().isInteger()) {
3532 getAArch64Cmp(LHS, RHS, ISD::getSetCCInverse(CC, true), CCVal, DAG, dl);
3534 // Note that we inverted the condition above, so we reverse the order of
3535 // the true and false operands here. This will allow the setcc to be
3536 // matched to a single CSINC instruction.
3537 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp);
3540 // Now we know we're dealing with FP values.
3541 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3543 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3544 // and do the comparison.
3545 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3547 AArch64CC::CondCode CC1, CC2;
3548 changeFPCCToAArch64CC(CC, CC1, CC2);
3549 if (CC2 == AArch64CC::AL) {
3550 changeFPCCToAArch64CC(ISD::getSetCCInverse(CC, false), CC1, CC2);
3551 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
3553 // Note that we inverted the condition above, so we reverse the order of
3554 // the true and false operands here. This will allow the setcc to be
3555 // matched to a single CSINC instruction.
3556 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp);
3558 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't
3559 // totally clean. Some of them require two CSELs to implement. As is in
3560 // this case, we emit the first CSEL and then emit a second using the output
3561 // of the first as the RHS. We're effectively OR'ing the two CC's together.
3563 // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
3564 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
3566 DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3568 SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32);
3569 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3573 /// A SELECT_CC operation is really some kind of max or min if both values being
3574 /// compared are, in some sense, equal to the results in either case. However,
3575 /// it is permissible to compare f32 values and produce directly extended f64
3578 /// Extending the comparison operands would also be allowed, but is less likely
3579 /// to happen in practice since their use is right here. Note that truncate
3580 /// operations would *not* be semantically equivalent.
3581 static bool selectCCOpsAreFMaxCompatible(SDValue Cmp, SDValue Result) {
3583 return (Cmp.getValueType() == MVT::f32 ||
3584 Cmp.getValueType() == MVT::f64);
3586 ConstantFPSDNode *CCmp = dyn_cast<ConstantFPSDNode>(Cmp);
3587 ConstantFPSDNode *CResult = dyn_cast<ConstantFPSDNode>(Result);
3588 if (CCmp && CResult && Cmp.getValueType() == MVT::f32 &&
3589 Result.getValueType() == MVT::f64) {
3591 APFloat CmpVal = CCmp->getValueAPF();
3592 CmpVal.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &Lossy);
3593 return CResult->getValueAPF().bitwiseIsEqual(CmpVal);
3596 return Result->getOpcode() == ISD::FP_EXTEND && Result->getOperand(0) == Cmp;
3599 SDValue AArch64TargetLowering::LowerSELECT_CC(ISD::CondCode CC, SDValue LHS,
3600 SDValue RHS, SDValue TVal,
3601 SDValue FVal, SDLoc dl,
3602 SelectionDAG &DAG) const {
3603 // Handle f128 first, because it will result in a comparison of some RTLIB
3604 // call result against zero.
3605 if (LHS.getValueType() == MVT::f128) {
3606 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3608 // If softenSetCCOperands returned a scalar, we need to compare the result
3609 // against zero to select between true and false values.
3610 if (!RHS.getNode()) {
3611 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3616 // Handle integers first.
3617 if (LHS.getValueType().isInteger()) {
3618 assert((LHS.getValueType() == RHS.getValueType()) &&
3619 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3621 unsigned Opcode = AArch64ISD::CSEL;
3623 // If both the TVal and the FVal are constants, see if we can swap them in
3624 // order to for a CSINV or CSINC out of them.
3625 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
3626 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
3628 if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) {
3629 std::swap(TVal, FVal);
3630 std::swap(CTVal, CFVal);
3631 CC = ISD::getSetCCInverse(CC, true);
3632 } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) {
3633 std::swap(TVal, FVal);
3634 std::swap(CTVal, CFVal);
3635 CC = ISD::getSetCCInverse(CC, true);
3636 } else if (TVal.getOpcode() == ISD::XOR) {
3637 // If TVal is a NOT we want to swap TVal and FVal so that we can match
3638 // with a CSINV rather than a CSEL.
3639 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(1));
3641 if (CVal && CVal->isAllOnesValue()) {
3642 std::swap(TVal, FVal);
3643 std::swap(CTVal, CFVal);
3644 CC = ISD::getSetCCInverse(CC, true);
3646 } else if (TVal.getOpcode() == ISD::SUB) {
3647 // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
3648 // that we can match with a CSNEG rather than a CSEL.
3649 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(0));
3651 if (CVal && CVal->isNullValue()) {
3652 std::swap(TVal, FVal);
3653 std::swap(CTVal, CFVal);
3654 CC = ISD::getSetCCInverse(CC, true);
3656 } else if (CTVal && CFVal) {
3657 const int64_t TrueVal = CTVal->getSExtValue();
3658 const int64_t FalseVal = CFVal->getSExtValue();
3661 // If both TVal and FVal are constants, see if FVal is the
3662 // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
3663 // instead of a CSEL in that case.
3664 if (TrueVal == ~FalseVal) {
3665 Opcode = AArch64ISD::CSINV;
3666 } else if (TrueVal == -FalseVal) {
3667 Opcode = AArch64ISD::CSNEG;
3668 } else if (TVal.getValueType() == MVT::i32) {
3669 // If our operands are only 32-bit wide, make sure we use 32-bit
3670 // arithmetic for the check whether we can use CSINC. This ensures that
3671 // the addition in the check will wrap around properly in case there is
3672 // an overflow (which would not be the case if we do the check with
3673 // 64-bit arithmetic).
3674 const uint32_t TrueVal32 = CTVal->getZExtValue();
3675 const uint32_t FalseVal32 = CFVal->getZExtValue();
3677 if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
3678 Opcode = AArch64ISD::CSINC;
3680 if (TrueVal32 > FalseVal32) {
3684 // 64-bit check whether we can use CSINC.
3685 } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) {
3686 Opcode = AArch64ISD::CSINC;
3688 if (TrueVal > FalseVal) {
3693 // Swap TVal and FVal if necessary.
3695 std::swap(TVal, FVal);
3696 std::swap(CTVal, CFVal);
3697 CC = ISD::getSetCCInverse(CC, true);
3700 if (Opcode != AArch64ISD::CSEL) {
3701 // Drop FVal since we can get its value by simply inverting/negating
3708 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3710 EVT VT = TVal.getValueType();
3711 return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp);
3714 // Now we know we're dealing with FP values.
3715 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3716 assert(LHS.getValueType() == RHS.getValueType());
3717 EVT VT = TVal.getValueType();
3718 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3720 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
3721 // clean. Some of them require two CSELs to implement.
3722 AArch64CC::CondCode CC1, CC2;
3723 changeFPCCToAArch64CC(CC, CC1, CC2);
3724 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
3725 SDValue CS1 = DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3727 // If we need a second CSEL, emit it, using the output of the first as the
3728 // RHS. We're effectively OR'ing the two CC's together.
3729 if (CC2 != AArch64CC::AL) {
3730 SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32);
3731 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3734 // Otherwise, return the output of the first CSEL.
3738 SDValue AArch64TargetLowering::LowerSELECT_CC(SDValue Op,
3739 SelectionDAG &DAG) const {
3740 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3741 SDValue LHS = Op.getOperand(0);
3742 SDValue RHS = Op.getOperand(1);
3743 SDValue TVal = Op.getOperand(2);
3744 SDValue FVal = Op.getOperand(3);
3746 return LowerSELECT_CC(CC, LHS, RHS, TVal, FVal, DL, DAG);
3749 SDValue AArch64TargetLowering::LowerSELECT(SDValue Op,
3750 SelectionDAG &DAG) const {
3751 SDValue CCVal = Op->getOperand(0);
3752 SDValue TVal = Op->getOperand(1);
3753 SDValue FVal = Op->getOperand(2);
3756 unsigned Opc = CCVal.getOpcode();
3757 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
3759 if (CCVal.getResNo() == 1 &&
3760 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3761 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3762 // Only lower legal XALUO ops.
3763 if (!DAG.getTargetLoweringInfo().isTypeLegal(CCVal->getValueType(0)))
3766 AArch64CC::CondCode OFCC;
3767 SDValue Value, Overflow;
3768 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, CCVal.getValue(0), DAG);
3769 SDValue CCVal = DAG.getConstant(OFCC, DL, MVT::i32);
3771 return DAG.getNode(AArch64ISD::CSEL, DL, Op.getValueType(), TVal, FVal,
3775 // Lower it the same way as we would lower a SELECT_CC node.
3778 if (CCVal.getOpcode() == ISD::SETCC) {
3779 LHS = CCVal.getOperand(0);
3780 RHS = CCVal.getOperand(1);
3781 CC = cast<CondCodeSDNode>(CCVal->getOperand(2))->get();
3784 RHS = DAG.getConstant(0, DL, CCVal.getValueType());
3787 return LowerSELECT_CC(CC, LHS, RHS, TVal, FVal, DL, DAG);
3790 SDValue AArch64TargetLowering::LowerJumpTable(SDValue Op,
3791 SelectionDAG &DAG) const {
3792 // Jump table entries as PC relative offsets. No additional tweaking
3793 // is necessary here. Just get the address of the jump table.
3794 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3795 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3798 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3799 !Subtarget->isTargetMachO()) {
3800 const unsigned char MO_NC = AArch64II::MO_NC;
3802 AArch64ISD::WrapperLarge, DL, PtrVT,
3803 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G3),
3804 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G2 | MO_NC),
3805 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G1 | MO_NC),
3806 DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3807 AArch64II::MO_G0 | MO_NC));
3811 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_PAGE);
3812 SDValue Lo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3813 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3814 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3815 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3818 SDValue AArch64TargetLowering::LowerConstantPool(SDValue Op,
3819 SelectionDAG &DAG) const {
3820 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3821 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3824 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
3825 // Use the GOT for the large code model on iOS.
3826 if (Subtarget->isTargetMachO()) {
3827 SDValue GotAddr = DAG.getTargetConstantPool(
3828 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3830 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
3833 const unsigned char MO_NC = AArch64II::MO_NC;
3835 AArch64ISD::WrapperLarge, DL, PtrVT,
3836 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3837 CP->getOffset(), AArch64II::MO_G3),
3838 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3839 CP->getOffset(), AArch64II::MO_G2 | MO_NC),
3840 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3841 CP->getOffset(), AArch64II::MO_G1 | MO_NC),
3842 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3843 CP->getOffset(), AArch64II::MO_G0 | MO_NC));
3845 // Use ADRP/ADD or ADRP/LDR for everything else: the small memory model on
3846 // ELF, the only valid one on Darwin.
3848 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3849 CP->getOffset(), AArch64II::MO_PAGE);
3850 SDValue Lo = DAG.getTargetConstantPool(
3851 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3852 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3854 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3855 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3859 SDValue AArch64TargetLowering::LowerBlockAddress(SDValue Op,
3860 SelectionDAG &DAG) const {
3861 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
3862 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3864 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3865 !Subtarget->isTargetMachO()) {
3866 const unsigned char MO_NC = AArch64II::MO_NC;
3868 AArch64ISD::WrapperLarge, DL, PtrVT,
3869 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G3),
3870 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
3871 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
3872 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
3874 SDValue Hi = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGE);
3875 SDValue Lo = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGEOFF |
3877 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3878 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3882 SDValue AArch64TargetLowering::LowerDarwin_VASTART(SDValue Op,
3883 SelectionDAG &DAG) const {
3884 AArch64FunctionInfo *FuncInfo =
3885 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
3888 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(),
3889 getPointerTy(DAG.getDataLayout()));
3890 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3891 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
3892 MachinePointerInfo(SV), false, false, 0);
3895 SDValue AArch64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
3896 SelectionDAG &DAG) const {
3897 // The layout of the va_list struct is specified in the AArch64 Procedure Call
3898 // Standard, section B.3.
3899 MachineFunction &MF = DAG.getMachineFunction();
3900 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
3901 auto PtrVT = getPointerTy(DAG.getDataLayout());
3904 SDValue Chain = Op.getOperand(0);
3905 SDValue VAList = Op.getOperand(1);
3906 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3907 SmallVector<SDValue, 4> MemOps;
3909 // void *__stack at offset 0
3910 SDValue Stack = DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), PtrVT);
3911 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
3912 MachinePointerInfo(SV), false, false, 8));
3914 // void *__gr_top at offset 8
3915 int GPRSize = FuncInfo->getVarArgsGPRSize();
3917 SDValue GRTop, GRTopAddr;
3920 DAG.getNode(ISD::ADD, DL, PtrVT, VAList, DAG.getConstant(8, DL, PtrVT));
3922 GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), PtrVT);
3923 GRTop = DAG.getNode(ISD::ADD, DL, PtrVT, GRTop,
3924 DAG.getConstant(GPRSize, DL, PtrVT));
3926 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
3927 MachinePointerInfo(SV, 8), false, false, 8));
3930 // void *__vr_top at offset 16
3931 int FPRSize = FuncInfo->getVarArgsFPRSize();
3933 SDValue VRTop, VRTopAddr;
3934 VRTopAddr = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
3935 DAG.getConstant(16, DL, PtrVT));
3937 VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), PtrVT);
3938 VRTop = DAG.getNode(ISD::ADD, DL, PtrVT, VRTop,
3939 DAG.getConstant(FPRSize, DL, PtrVT));
3941 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
3942 MachinePointerInfo(SV, 16), false, false, 8));
3945 // int __gr_offs at offset 24
3946 SDValue GROffsAddr =
3947 DAG.getNode(ISD::ADD, DL, PtrVT, VAList, DAG.getConstant(24, DL, PtrVT));
3948 MemOps.push_back(DAG.getStore(Chain, DL,
3949 DAG.getConstant(-GPRSize, DL, MVT::i32),
3950 GROffsAddr, MachinePointerInfo(SV, 24), false,
3953 // int __vr_offs at offset 28
3954 SDValue VROffsAddr =
3955 DAG.getNode(ISD::ADD, DL, PtrVT, VAList, DAG.getConstant(28, DL, PtrVT));
3956 MemOps.push_back(DAG.getStore(Chain, DL,
3957 DAG.getConstant(-FPRSize, DL, MVT::i32),
3958 VROffsAddr, MachinePointerInfo(SV, 28), false,
3961 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
3964 SDValue AArch64TargetLowering::LowerVASTART(SDValue Op,
3965 SelectionDAG &DAG) const {
3966 return Subtarget->isTargetDarwin() ? LowerDarwin_VASTART(Op, DAG)
3967 : LowerAAPCS_VASTART(Op, DAG);
3970 SDValue AArch64TargetLowering::LowerVACOPY(SDValue Op,
3971 SelectionDAG &DAG) const {
3972 // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
3975 unsigned VaListSize = Subtarget->isTargetDarwin() ? 8 : 32;
3976 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
3977 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
3979 return DAG.getMemcpy(Op.getOperand(0), DL, Op.getOperand(1),
3981 DAG.getConstant(VaListSize, DL, MVT::i32),
3982 8, false, false, false, MachinePointerInfo(DestSV),
3983 MachinePointerInfo(SrcSV));
3986 SDValue AArch64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
3987 assert(Subtarget->isTargetDarwin() &&
3988 "automatic va_arg instruction only works on Darwin");
3990 const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3991 EVT VT = Op.getValueType();
3993 SDValue Chain = Op.getOperand(0);
3994 SDValue Addr = Op.getOperand(1);
3995 unsigned Align = Op.getConstantOperandVal(3);
3996 auto PtrVT = getPointerTy(DAG.getDataLayout());
3998 SDValue VAList = DAG.getLoad(PtrVT, DL, Chain, Addr, MachinePointerInfo(V),
3999 false, false, false, 0);
4000 Chain = VAList.getValue(1);
4003 assert(((Align & (Align - 1)) == 0) && "Expected Align to be a power of 2");
4004 VAList = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
4005 DAG.getConstant(Align - 1, DL, PtrVT));
4006 VAList = DAG.getNode(ISD::AND, DL, PtrVT, VAList,
4007 DAG.getConstant(-(int64_t)Align, DL, PtrVT));
4010 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
4011 uint64_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
4013 // Scalar integer and FP values smaller than 64 bits are implicitly extended
4014 // up to 64 bits. At the very least, we have to increase the striding of the
4015 // vaargs list to match this, and for FP values we need to introduce
4016 // FP_ROUND nodes as well.
4017 if (VT.isInteger() && !VT.isVector())
4019 bool NeedFPTrunc = false;
4020 if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
4025 // Increment the pointer, VAList, to the next vaarg
4026 SDValue VANext = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
4027 DAG.getConstant(ArgSize, DL, PtrVT));
4028 // Store the incremented VAList to the legalized pointer
4029 SDValue APStore = DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V),
4032 // Load the actual argument out of the pointer VAList
4034 // Load the value as an f64.
4035 SDValue WideFP = DAG.getLoad(MVT::f64, DL, APStore, VAList,
4036 MachinePointerInfo(), false, false, false, 0);
4037 // Round the value down to an f32.
4038 SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
4039 DAG.getIntPtrConstant(1, DL));
4040 SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
4041 // Merge the rounded value with the chain output of the load.
4042 return DAG.getMergeValues(Ops, DL);
4045 return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo(), false,
4049 SDValue AArch64TargetLowering::LowerFRAMEADDR(SDValue Op,
4050 SelectionDAG &DAG) const {
4051 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4052 MFI->setFrameAddressIsTaken(true);
4054 EVT VT = Op.getValueType();
4056 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4058 DAG.getCopyFromReg(DAG.getEntryNode(), DL, AArch64::FP, VT);
4060 FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
4061 MachinePointerInfo(), false, false, false, 0);
4065 // FIXME? Maybe this could be a TableGen attribute on some registers and
4066 // this table could be generated automatically from RegInfo.
4067 unsigned AArch64TargetLowering::getRegisterByName(const char* RegName,
4069 unsigned Reg = StringSwitch<unsigned>(RegName)
4070 .Case("sp", AArch64::SP)
4074 report_fatal_error(Twine("Invalid register name \""
4075 + StringRef(RegName) + "\"."));
4078 SDValue AArch64TargetLowering::LowerRETURNADDR(SDValue Op,
4079 SelectionDAG &DAG) const {
4080 MachineFunction &MF = DAG.getMachineFunction();
4081 MachineFrameInfo *MFI = MF.getFrameInfo();
4082 MFI->setReturnAddressIsTaken(true);
4084 EVT VT = Op.getValueType();
4086 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4088 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4089 SDValue Offset = DAG.getConstant(8, DL, getPointerTy(DAG.getDataLayout()));
4090 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
4091 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
4092 MachinePointerInfo(), false, false, false, 0);
4095 // Return LR, which contains the return address. Mark it an implicit live-in.
4096 unsigned Reg = MF.addLiveIn(AArch64::LR, &AArch64::GPR64RegClass);
4097 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
4100 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4101 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4102 SDValue AArch64TargetLowering::LowerShiftRightParts(SDValue Op,
4103 SelectionDAG &DAG) const {
4104 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4105 EVT VT = Op.getValueType();
4106 unsigned VTBits = VT.getSizeInBits();
4108 SDValue ShOpLo = Op.getOperand(0);
4109 SDValue ShOpHi = Op.getOperand(1);
4110 SDValue ShAmt = Op.getOperand(2);
4112 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4114 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4116 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4117 DAG.getConstant(VTBits, dl, MVT::i64), ShAmt);
4118 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4119 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4120 DAG.getConstant(VTBits, dl, MVT::i64));
4121 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4123 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, dl, MVT::i64),
4124 ISD::SETGE, dl, DAG);
4125 SDValue CCVal = DAG.getConstant(AArch64CC::GE, dl, MVT::i32);
4127 SDValue FalseValLo = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4128 SDValue TrueValLo = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4130 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
4132 // AArch64 shifts larger than the register width are wrapped rather than
4133 // clamped, so we can't just emit "hi >> x".
4134 SDValue FalseValHi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4135 SDValue TrueValHi = Opc == ISD::SRA
4136 ? DAG.getNode(Opc, dl, VT, ShOpHi,
4137 DAG.getConstant(VTBits - 1, dl,
4139 : DAG.getConstant(0, dl, VT);
4141 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValHi, FalseValHi, CCVal, Cmp);
4143 SDValue Ops[2] = { Lo, Hi };
4144 return DAG.getMergeValues(Ops, dl);
4147 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4148 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4149 SDValue AArch64TargetLowering::LowerShiftLeftParts(SDValue Op,
4150 SelectionDAG &DAG) const {
4151 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4152 EVT VT = Op.getValueType();
4153 unsigned VTBits = VT.getSizeInBits();
4155 SDValue ShOpLo = Op.getOperand(0);
4156 SDValue ShOpHi = Op.getOperand(1);
4157 SDValue ShAmt = Op.getOperand(2);
4160 assert(Op.getOpcode() == ISD::SHL_PARTS);
4161 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4162 DAG.getConstant(VTBits, dl, MVT::i64), ShAmt);
4163 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4164 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4165 DAG.getConstant(VTBits, dl, MVT::i64));
4166 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4167 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4169 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4171 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, dl, MVT::i64),
4172 ISD::SETGE, dl, DAG);
4173 SDValue CCVal = DAG.getConstant(AArch64CC::GE, dl, MVT::i32);
4175 DAG.getNode(AArch64ISD::CSEL, dl, VT, Tmp3, FalseVal, CCVal, Cmp);
4177 // AArch64 shifts of larger than register sizes are wrapped rather than
4178 // clamped, so we can't just emit "lo << a" if a is too big.
4179 SDValue TrueValLo = DAG.getConstant(0, dl, VT);
4180 SDValue FalseValLo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4182 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
4184 SDValue Ops[2] = { Lo, Hi };
4185 return DAG.getMergeValues(Ops, dl);
4188 bool AArch64TargetLowering::isOffsetFoldingLegal(
4189 const GlobalAddressSDNode *GA) const {
4190 // The AArch64 target doesn't support folding offsets into global addresses.
4194 bool AArch64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4195 // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
4196 // FIXME: We should be able to handle f128 as well with a clever lowering.
4197 if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32))
4201 return AArch64_AM::getFP64Imm(Imm) != -1;
4202 else if (VT == MVT::f32)
4203 return AArch64_AM::getFP32Imm(Imm) != -1;
4207 //===----------------------------------------------------------------------===//
4208 // AArch64 Optimization Hooks
4209 //===----------------------------------------------------------------------===//
4211 //===----------------------------------------------------------------------===//
4212 // AArch64 Inline Assembly Support
4213 //===----------------------------------------------------------------------===//
4215 // Table of Constraints
4216 // TODO: This is the current set of constraints supported by ARM for the
4217 // compiler, not all of them may make sense, e.g. S may be difficult to support.
4219 // r - A general register
4220 // w - An FP/SIMD register of some size in the range v0-v31
4221 // x - An FP/SIMD register of some size in the range v0-v15
4222 // I - Constant that can be used with an ADD instruction
4223 // J - Constant that can be used with a SUB instruction
4224 // K - Constant that can be used with a 32-bit logical instruction
4225 // L - Constant that can be used with a 64-bit logical instruction
4226 // M - Constant that can be used as a 32-bit MOV immediate
4227 // N - Constant that can be used as a 64-bit MOV immediate
4228 // Q - A memory reference with base register and no offset
4229 // S - A symbolic address
4230 // Y - Floating point constant zero
4231 // Z - Integer constant zero
4233 // Note that general register operands will be output using their 64-bit x
4234 // register name, whatever the size of the variable, unless the asm operand
4235 // is prefixed by the %w modifier. Floating-point and SIMD register operands
4236 // will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
4239 /// getConstraintType - Given a constraint letter, return the type of
4240 /// constraint it is for this target.
4241 AArch64TargetLowering::ConstraintType
4242 AArch64TargetLowering::getConstraintType(StringRef Constraint) const {
4243 if (Constraint.size() == 1) {
4244 switch (Constraint[0]) {
4251 return C_RegisterClass;
4252 // An address with a single base register. Due to the way we
4253 // currently handle addresses it is the same as 'r'.
4258 return TargetLowering::getConstraintType(Constraint);
4261 /// Examine constraint type and operand type and determine a weight value.
4262 /// This object must already have been set up with the operand type
4263 /// and the current alternative constraint selected.
4264 TargetLowering::ConstraintWeight
4265 AArch64TargetLowering::getSingleConstraintMatchWeight(
4266 AsmOperandInfo &info, const char *constraint) const {
4267 ConstraintWeight weight = CW_Invalid;
4268 Value *CallOperandVal = info.CallOperandVal;
4269 // If we don't have a value, we can't do a match,
4270 // but allow it at the lowest weight.
4271 if (!CallOperandVal)
4273 Type *type = CallOperandVal->getType();
4274 // Look at the constraint type.
4275 switch (*constraint) {
4277 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
4281 if (type->isFloatingPointTy() || type->isVectorTy())
4282 weight = CW_Register;
4285 weight = CW_Constant;
4291 std::pair<unsigned, const TargetRegisterClass *>
4292 AArch64TargetLowering::getRegForInlineAsmConstraint(
4293 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
4294 if (Constraint.size() == 1) {
4295 switch (Constraint[0]) {
4297 if (VT.getSizeInBits() == 64)
4298 return std::make_pair(0U, &AArch64::GPR64commonRegClass);
4299 return std::make_pair(0U, &AArch64::GPR32commonRegClass);
4302 return std::make_pair(0U, &AArch64::FPR32RegClass);
4303 if (VT.getSizeInBits() == 64)
4304 return std::make_pair(0U, &AArch64::FPR64RegClass);
4305 if (VT.getSizeInBits() == 128)
4306 return std::make_pair(0U, &AArch64::FPR128RegClass);
4308 // The instructions that this constraint is designed for can
4309 // only take 128-bit registers so just use that regclass.
4311 if (VT.getSizeInBits() == 128)
4312 return std::make_pair(0U, &AArch64::FPR128_loRegClass);
4316 if (StringRef("{cc}").equals_lower(Constraint))
4317 return std::make_pair(unsigned(AArch64::NZCV), &AArch64::CCRRegClass);
4319 // Use the default implementation in TargetLowering to convert the register
4320 // constraint into a member of a register class.
4321 std::pair<unsigned, const TargetRegisterClass *> Res;
4322 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
4324 // Not found as a standard register?
4326 unsigned Size = Constraint.size();
4327 if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
4328 tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
4330 bool Failed = Constraint.slice(2, Size - 1).getAsInteger(10, RegNo);
4331 if (!Failed && RegNo >= 0 && RegNo <= 31) {
4332 // v0 - v31 are aliases of q0 - q31.
4333 // By default we'll emit v0-v31 for this unless there's a modifier where
4334 // we'll emit the correct register as well.
4335 Res.first = AArch64::FPR128RegClass.getRegister(RegNo);
4336 Res.second = &AArch64::FPR128RegClass;
4344 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4345 /// vector. If it is invalid, don't add anything to Ops.
4346 void AArch64TargetLowering::LowerAsmOperandForConstraint(
4347 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
4348 SelectionDAG &DAG) const {
4351 // Currently only support length 1 constraints.
4352 if (Constraint.length() != 1)
4355 char ConstraintLetter = Constraint[0];
4356 switch (ConstraintLetter) {
4360 // This set of constraints deal with valid constants for various instructions.
4361 // Validate and return a target constant for them if we can.
4363 // 'z' maps to xzr or wzr so it needs an input of 0.
4364 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4365 if (!C || C->getZExtValue() != 0)
4368 if (Op.getValueType() == MVT::i64)
4369 Result = DAG.getRegister(AArch64::XZR, MVT::i64);
4371 Result = DAG.getRegister(AArch64::WZR, MVT::i32);
4381 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4385 // Grab the value and do some validation.
4386 uint64_t CVal = C->getZExtValue();
4387 switch (ConstraintLetter) {
4388 // The I constraint applies only to simple ADD or SUB immediate operands:
4389 // i.e. 0 to 4095 with optional shift by 12
4390 // The J constraint applies only to ADD or SUB immediates that would be
4391 // valid when negated, i.e. if [an add pattern] were to be output as a SUB
4392 // instruction [or vice versa], in other words -1 to -4095 with optional
4393 // left shift by 12.
4395 if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
4399 uint64_t NVal = -C->getSExtValue();
4400 if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal)) {
4401 CVal = C->getSExtValue();
4406 // The K and L constraints apply *only* to logical immediates, including
4407 // what used to be the MOVI alias for ORR (though the MOVI alias has now
4408 // been removed and MOV should be used). So these constraints have to
4409 // distinguish between bit patterns that are valid 32-bit or 64-bit
4410 // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
4411 // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
4414 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4418 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4421 // The M and N constraints are a superset of K and L respectively, for use
4422 // with the MOV (immediate) alias. As well as the logical immediates they
4423 // also match 32 or 64-bit immediates that can be loaded either using a
4424 // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
4425 // (M) or 64-bit 0x1234000000000000 (N) etc.
4426 // As a note some of this code is liberally stolen from the asm parser.
4428 if (!isUInt<32>(CVal))
4430 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4432 if ((CVal & 0xFFFF) == CVal)
4434 if ((CVal & 0xFFFF0000ULL) == CVal)
4436 uint64_t NCVal = ~(uint32_t)CVal;
4437 if ((NCVal & 0xFFFFULL) == NCVal)
4439 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4444 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4446 if ((CVal & 0xFFFFULL) == CVal)
4448 if ((CVal & 0xFFFF0000ULL) == CVal)
4450 if ((CVal & 0xFFFF00000000ULL) == CVal)
4452 if ((CVal & 0xFFFF000000000000ULL) == CVal)
4454 uint64_t NCVal = ~CVal;
4455 if ((NCVal & 0xFFFFULL) == NCVal)
4457 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4459 if ((NCVal & 0xFFFF00000000ULL) == NCVal)
4461 if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
4469 // All assembler immediates are 64-bit integers.
4470 Result = DAG.getTargetConstant(CVal, SDLoc(Op), MVT::i64);
4474 if (Result.getNode()) {
4475 Ops.push_back(Result);
4479 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
4482 //===----------------------------------------------------------------------===//
4483 // AArch64 Advanced SIMD Support
4484 //===----------------------------------------------------------------------===//
4486 /// WidenVector - Given a value in the V64 register class, produce the
4487 /// equivalent value in the V128 register class.
4488 static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
4489 EVT VT = V64Reg.getValueType();
4490 unsigned NarrowSize = VT.getVectorNumElements();
4491 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4492 MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
4495 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
4496 V64Reg, DAG.getConstant(0, DL, MVT::i32));
4499 /// getExtFactor - Determine the adjustment factor for the position when
4500 /// generating an "extract from vector registers" instruction.
4501 static unsigned getExtFactor(SDValue &V) {
4502 EVT EltType = V.getValueType().getVectorElementType();
4503 return EltType.getSizeInBits() / 8;
4506 /// NarrowVector - Given a value in the V128 register class, produce the
4507 /// equivalent value in the V64 register class.
4508 static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) {
4509 EVT VT = V128Reg.getValueType();
4510 unsigned WideSize = VT.getVectorNumElements();
4511 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4512 MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2);
4515 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, NarrowTy, V128Reg);
4518 // Gather data to see if the operation can be modelled as a
4519 // shuffle in combination with VEXTs.
4520 SDValue AArch64TargetLowering::ReconstructShuffle(SDValue Op,
4521 SelectionDAG &DAG) const {
4522 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
4524 EVT VT = Op.getValueType();
4525 unsigned NumElts = VT.getVectorNumElements();
4527 struct ShuffleSourceInfo {
4532 // We may insert some combination of BITCASTs and VEXT nodes to force Vec to
4533 // be compatible with the shuffle we intend to construct. As a result
4534 // ShuffleVec will be some sliding window into the original Vec.
4537 // Code should guarantee that element i in Vec starts at element "WindowBase
4538 // + i * WindowScale in ShuffleVec".
4542 bool operator ==(SDValue OtherVec) { return Vec == OtherVec; }
4543 ShuffleSourceInfo(SDValue Vec)
4544 : Vec(Vec), MinElt(UINT_MAX), MaxElt(0), ShuffleVec(Vec), WindowBase(0),
4548 // First gather all vectors used as an immediate source for this BUILD_VECTOR
4550 SmallVector<ShuffleSourceInfo, 2> Sources;
4551 for (unsigned i = 0; i < NumElts; ++i) {
4552 SDValue V = Op.getOperand(i);
4553 if (V.getOpcode() == ISD::UNDEF)
4555 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4556 // A shuffle can only come from building a vector from various
4557 // elements of other vectors.
4561 // Add this element source to the list if it's not already there.
4562 SDValue SourceVec = V.getOperand(0);
4563 auto Source = std::find(Sources.begin(), Sources.end(), SourceVec);
4564 if (Source == Sources.end())
4565 Source = Sources.insert(Sources.end(), ShuffleSourceInfo(SourceVec));
4567 // Update the minimum and maximum lane number seen.
4568 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4569 Source->MinElt = std::min(Source->MinElt, EltNo);
4570 Source->MaxElt = std::max(Source->MaxElt, EltNo);
4573 // Currently only do something sane when at most two source vectors
4575 if (Sources.size() > 2)
4578 // Find out the smallest element size among result and two sources, and use
4579 // it as element size to build the shuffle_vector.
4580 EVT SmallestEltTy = VT.getVectorElementType();
4581 for (auto &Source : Sources) {
4582 EVT SrcEltTy = Source.Vec.getValueType().getVectorElementType();
4583 if (SrcEltTy.bitsLT(SmallestEltTy)) {
4584 SmallestEltTy = SrcEltTy;
4587 unsigned ResMultiplier =
4588 VT.getVectorElementType().getSizeInBits() / SmallestEltTy.getSizeInBits();
4589 NumElts = VT.getSizeInBits() / SmallestEltTy.getSizeInBits();
4590 EVT ShuffleVT = EVT::getVectorVT(*DAG.getContext(), SmallestEltTy, NumElts);
4592 // If the source vector is too wide or too narrow, we may nevertheless be able
4593 // to construct a compatible shuffle either by concatenating it with UNDEF or
4594 // extracting a suitable range of elements.
4595 for (auto &Src : Sources) {
4596 EVT SrcVT = Src.ShuffleVec.getValueType();
4598 if (SrcVT.getSizeInBits() == VT.getSizeInBits())
4601 // This stage of the search produces a source with the same element type as
4602 // the original, but with a total width matching the BUILD_VECTOR output.
4603 EVT EltVT = SrcVT.getVectorElementType();
4604 unsigned NumSrcElts = VT.getSizeInBits() / EltVT.getSizeInBits();
4605 EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumSrcElts);
4607 if (SrcVT.getSizeInBits() < VT.getSizeInBits()) {
4608 assert(2 * SrcVT.getSizeInBits() == VT.getSizeInBits());
4609 // We can pad out the smaller vector for free, so if it's part of a
4612 DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, Src.ShuffleVec,
4613 DAG.getUNDEF(Src.ShuffleVec.getValueType()));
4617 assert(SrcVT.getSizeInBits() == 2 * VT.getSizeInBits());
4619 if (Src.MaxElt - Src.MinElt >= NumSrcElts) {
4620 // Span too large for a VEXT to cope
4624 if (Src.MinElt >= NumSrcElts) {
4625 // The extraction can just take the second half
4627 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4628 DAG.getConstant(NumSrcElts, dl, MVT::i64));
4629 Src.WindowBase = -NumSrcElts;
4630 } else if (Src.MaxElt < NumSrcElts) {
4631 // The extraction can just take the first half
4633 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4634 DAG.getConstant(0, dl, MVT::i64));
4636 // An actual VEXT is needed
4638 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4639 DAG.getConstant(0, dl, MVT::i64));
4641 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4642 DAG.getConstant(NumSrcElts, dl, MVT::i64));
4643 unsigned Imm = Src.MinElt * getExtFactor(VEXTSrc1);
4645 Src.ShuffleVec = DAG.getNode(AArch64ISD::EXT, dl, DestVT, VEXTSrc1,
4647 DAG.getConstant(Imm, dl, MVT::i32));
4648 Src.WindowBase = -Src.MinElt;
4652 // Another possible incompatibility occurs from the vector element types. We
4653 // can fix this by bitcasting the source vectors to the same type we intend
4655 for (auto &Src : Sources) {
4656 EVT SrcEltTy = Src.ShuffleVec.getValueType().getVectorElementType();
4657 if (SrcEltTy == SmallestEltTy)
4659 assert(ShuffleVT.getVectorElementType() == SmallestEltTy);
4660 Src.ShuffleVec = DAG.getNode(ISD::BITCAST, dl, ShuffleVT, Src.ShuffleVec);
4661 Src.WindowScale = SrcEltTy.getSizeInBits() / SmallestEltTy.getSizeInBits();
4662 Src.WindowBase *= Src.WindowScale;
4665 // Final sanity check before we try to actually produce a shuffle.
4667 for (auto Src : Sources)
4668 assert(Src.ShuffleVec.getValueType() == ShuffleVT);
4671 // The stars all align, our next step is to produce the mask for the shuffle.
4672 SmallVector<int, 8> Mask(ShuffleVT.getVectorNumElements(), -1);
4673 int BitsPerShuffleLane = ShuffleVT.getVectorElementType().getSizeInBits();
4674 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
4675 SDValue Entry = Op.getOperand(i);
4676 if (Entry.getOpcode() == ISD::UNDEF)
4679 auto Src = std::find(Sources.begin(), Sources.end(), Entry.getOperand(0));
4680 int EltNo = cast<ConstantSDNode>(Entry.getOperand(1))->getSExtValue();
4682 // EXTRACT_VECTOR_ELT performs an implicit any_ext; BUILD_VECTOR an implicit
4683 // trunc. So only std::min(SrcBits, DestBits) actually get defined in this
4685 EVT OrigEltTy = Entry.getOperand(0).getValueType().getVectorElementType();
4686 int BitsDefined = std::min(OrigEltTy.getSizeInBits(),
4687 VT.getVectorElementType().getSizeInBits());
4688 int LanesDefined = BitsDefined / BitsPerShuffleLane;
4690 // This source is expected to fill ResMultiplier lanes of the final shuffle,
4691 // starting at the appropriate offset.
4692 int *LaneMask = &Mask[i * ResMultiplier];
4694 int ExtractBase = EltNo * Src->WindowScale + Src->WindowBase;
4695 ExtractBase += NumElts * (Src - Sources.begin());
4696 for (int j = 0; j < LanesDefined; ++j)
4697 LaneMask[j] = ExtractBase + j;
4700 // Final check before we try to produce nonsense...
4701 if (!isShuffleMaskLegal(Mask, ShuffleVT))
4704 SDValue ShuffleOps[] = { DAG.getUNDEF(ShuffleVT), DAG.getUNDEF(ShuffleVT) };
4705 for (unsigned i = 0; i < Sources.size(); ++i)
4706 ShuffleOps[i] = Sources[i].ShuffleVec;
4708 SDValue Shuffle = DAG.getVectorShuffle(ShuffleVT, dl, ShuffleOps[0],
4709 ShuffleOps[1], &Mask[0]);
4710 return DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
4713 // check if an EXT instruction can handle the shuffle mask when the
4714 // vector sources of the shuffle are the same.
4715 static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4716 unsigned NumElts = VT.getVectorNumElements();
4718 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4724 // If this is a VEXT shuffle, the immediate value is the index of the first
4725 // element. The other shuffle indices must be the successive elements after
4727 unsigned ExpectedElt = Imm;
4728 for (unsigned i = 1; i < NumElts; ++i) {
4729 // Increment the expected index. If it wraps around, just follow it
4730 // back to index zero and keep going.
4732 if (ExpectedElt == NumElts)
4736 continue; // ignore UNDEF indices
4737 if (ExpectedElt != static_cast<unsigned>(M[i]))
4744 // check if an EXT instruction can handle the shuffle mask when the
4745 // vector sources of the shuffle are different.
4746 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
4748 // Look for the first non-undef element.
4749 const int *FirstRealElt = std::find_if(M.begin(), M.end(),
4750 [](int Elt) {return Elt >= 0;});
4752 // Benefit form APInt to handle overflow when calculating expected element.
4753 unsigned NumElts = VT.getVectorNumElements();
4754 unsigned MaskBits = APInt(32, NumElts * 2).logBase2();
4755 APInt ExpectedElt = APInt(MaskBits, *FirstRealElt + 1);
4756 // The following shuffle indices must be the successive elements after the
4757 // first real element.
4758 const int *FirstWrongElt = std::find_if(FirstRealElt + 1, M.end(),
4759 [&](int Elt) {return Elt != ExpectedElt++ && Elt != -1;});
4760 if (FirstWrongElt != M.end())
4763 // The index of an EXT is the first element if it is not UNDEF.
4764 // Watch out for the beginning UNDEFs. The EXT index should be the expected
4765 // value of the first element. E.g.
4766 // <-1, -1, 3, ...> is treated as <1, 2, 3, ...>.
4767 // <-1, -1, 0, 1, ...> is treated as <2*NumElts-2, 2*NumElts-1, 0, 1, ...>.
4768 // ExpectedElt is the last mask index plus 1.
4769 Imm = ExpectedElt.getZExtValue();
4771 // There are two difference cases requiring to reverse input vectors.
4772 // For example, for vector <4 x i32> we have the following cases,
4773 // Case 1: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, -1, 0>)
4774 // Case 2: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, 7, 0>)
4775 // For both cases, we finally use mask <5, 6, 7, 0>, which requires
4776 // to reverse two input vectors.
4785 /// isREVMask - Check if a vector shuffle corresponds to a REV
4786 /// instruction with the specified blocksize. (The order of the elements
4787 /// within each block of the vector is reversed.)
4788 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4789 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
4790 "Only possible block sizes for REV are: 16, 32, 64");
4792 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4796 unsigned NumElts = VT.getVectorNumElements();
4797 unsigned BlockElts = M[0] + 1;
4798 // If the first shuffle index is UNDEF, be optimistic.
4800 BlockElts = BlockSize / EltSz;
4802 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4805 for (unsigned i = 0; i < NumElts; ++i) {
4807 continue; // ignore UNDEF indices
4808 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
4815 static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4816 unsigned NumElts = VT.getVectorNumElements();
4817 WhichResult = (M[0] == 0 ? 0 : 1);
4818 unsigned Idx = WhichResult * NumElts / 2;
4819 for (unsigned i = 0; i != NumElts; i += 2) {
4820 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4821 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
4829 static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4830 unsigned NumElts = VT.getVectorNumElements();
4831 WhichResult = (M[0] == 0 ? 0 : 1);
4832 for (unsigned i = 0; i != NumElts; ++i) {
4834 continue; // ignore UNDEF indices
4835 if ((unsigned)M[i] != 2 * i + WhichResult)
4842 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4843 unsigned NumElts = VT.getVectorNumElements();
4844 WhichResult = (M[0] == 0 ? 0 : 1);
4845 for (unsigned i = 0; i < NumElts; i += 2) {
4846 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4847 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + NumElts + WhichResult))
4853 /// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of
4854 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4855 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4856 static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4857 unsigned NumElts = VT.getVectorNumElements();
4858 WhichResult = (M[0] == 0 ? 0 : 1);
4859 unsigned Idx = WhichResult * NumElts / 2;
4860 for (unsigned i = 0; i != NumElts; i += 2) {
4861 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4862 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx))
4870 /// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of
4871 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4872 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4873 static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4874 unsigned Half = VT.getVectorNumElements() / 2;
4875 WhichResult = (M[0] == 0 ? 0 : 1);
4876 for (unsigned j = 0; j != 2; ++j) {
4877 unsigned Idx = WhichResult;
4878 for (unsigned i = 0; i != Half; ++i) {
4879 int MIdx = M[i + j * Half];
4880 if (MIdx >= 0 && (unsigned)MIdx != Idx)
4889 /// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of
4890 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4891 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4892 static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4893 unsigned NumElts = VT.getVectorNumElements();
4894 WhichResult = (M[0] == 0 ? 0 : 1);
4895 for (unsigned i = 0; i < NumElts; i += 2) {
4896 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4897 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult))
4903 static bool isINSMask(ArrayRef<int> M, int NumInputElements,
4904 bool &DstIsLeft, int &Anomaly) {
4905 if (M.size() != static_cast<size_t>(NumInputElements))
4908 int NumLHSMatch = 0, NumRHSMatch = 0;
4909 int LastLHSMismatch = -1, LastRHSMismatch = -1;
4911 for (int i = 0; i < NumInputElements; ++i) {
4921 LastLHSMismatch = i;
4923 if (M[i] == i + NumInputElements)
4926 LastRHSMismatch = i;
4929 if (NumLHSMatch == NumInputElements - 1) {
4931 Anomaly = LastLHSMismatch;
4933 } else if (NumRHSMatch == NumInputElements - 1) {
4935 Anomaly = LastRHSMismatch;
4942 static bool isConcatMask(ArrayRef<int> Mask, EVT VT, bool SplitLHS) {
4943 if (VT.getSizeInBits() != 128)
4946 unsigned NumElts = VT.getVectorNumElements();
4948 for (int I = 0, E = NumElts / 2; I != E; I++) {
4953 int Offset = NumElts / 2;
4954 for (int I = NumElts / 2, E = NumElts; I != E; I++) {
4955 if (Mask[I] != I + SplitLHS * Offset)
4962 static SDValue tryFormConcatFromShuffle(SDValue Op, SelectionDAG &DAG) {
4964 EVT VT = Op.getValueType();
4965 SDValue V0 = Op.getOperand(0);
4966 SDValue V1 = Op.getOperand(1);
4967 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op)->getMask();
4969 if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() ||
4970 VT.getVectorElementType() != V1.getValueType().getVectorElementType())
4973 bool SplitV0 = V0.getValueType().getSizeInBits() == 128;
4975 if (!isConcatMask(Mask, VT, SplitV0))
4978 EVT CastVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
4979 VT.getVectorNumElements() / 2);
4981 V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0,
4982 DAG.getConstant(0, DL, MVT::i64));
4984 if (V1.getValueType().getSizeInBits() == 128) {
4985 V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1,
4986 DAG.getConstant(0, DL, MVT::i64));
4988 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1);
4991 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4992 /// the specified operations to build the shuffle.
4993 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4994 SDValue RHS, SelectionDAG &DAG,
4996 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4997 unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1);
4998 unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1);
5001 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5010 OP_VUZPL, // VUZP, left result
5011 OP_VUZPR, // VUZP, right result
5012 OP_VZIPL, // VZIP, left result
5013 OP_VZIPR, // VZIP, right result
5014 OP_VTRNL, // VTRN, left result
5015 OP_VTRNR // VTRN, right result
5018 if (OpNum == OP_COPY) {
5019 if (LHSID == (1 * 9 + 2) * 9 + 3)
5021 assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!");
5025 SDValue OpLHS, OpRHS;
5026 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5027 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5028 EVT VT = OpLHS.getValueType();
5032 llvm_unreachable("Unknown shuffle opcode!");
5034 // VREV divides the vector in half and swaps within the half.
5035 if (VT.getVectorElementType() == MVT::i32 ||
5036 VT.getVectorElementType() == MVT::f32)
5037 return DAG.getNode(AArch64ISD::REV64, dl, VT, OpLHS);
5038 // vrev <4 x i16> -> REV32
5039 if (VT.getVectorElementType() == MVT::i16 ||
5040 VT.getVectorElementType() == MVT::f16)
5041 return DAG.getNode(AArch64ISD::REV32, dl, VT, OpLHS);
5042 // vrev <4 x i8> -> REV16
5043 assert(VT.getVectorElementType() == MVT::i8);
5044 return DAG.getNode(AArch64ISD::REV16, dl, VT, OpLHS);
5049 EVT EltTy = VT.getVectorElementType();
5051 if (EltTy == MVT::i8)
5052 Opcode = AArch64ISD::DUPLANE8;
5053 else if (EltTy == MVT::i16 || EltTy == MVT::f16)
5054 Opcode = AArch64ISD::DUPLANE16;
5055 else if (EltTy == MVT::i32 || EltTy == MVT::f32)
5056 Opcode = AArch64ISD::DUPLANE32;
5057 else if (EltTy == MVT::i64 || EltTy == MVT::f64)
5058 Opcode = AArch64ISD::DUPLANE64;
5060 llvm_unreachable("Invalid vector element type?");
5062 if (VT.getSizeInBits() == 64)
5063 OpLHS = WidenVector(OpLHS, DAG);
5064 SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, dl, MVT::i64);
5065 return DAG.getNode(Opcode, dl, VT, OpLHS, Lane);
5070 unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
5071 return DAG.getNode(AArch64ISD::EXT, dl, VT, OpLHS, OpRHS,
5072 DAG.getConstant(Imm, dl, MVT::i32));
5075 return DAG.getNode(AArch64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS,
5078 return DAG.getNode(AArch64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS,
5081 return DAG.getNode(AArch64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS,
5084 return DAG.getNode(AArch64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS,
5087 return DAG.getNode(AArch64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS,
5090 return DAG.getNode(AArch64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS,
5095 static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask,
5096 SelectionDAG &DAG) {
5097 // Check to see if we can use the TBL instruction.
5098 SDValue V1 = Op.getOperand(0);
5099 SDValue V2 = Op.getOperand(1);
5102 EVT EltVT = Op.getValueType().getVectorElementType();
5103 unsigned BytesPerElt = EltVT.getSizeInBits() / 8;
5105 SmallVector<SDValue, 8> TBLMask;
5106 for (int Val : ShuffleMask) {
5107 for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
5108 unsigned Offset = Byte + Val * BytesPerElt;
5109 TBLMask.push_back(DAG.getConstant(Offset, DL, MVT::i32));
5113 MVT IndexVT = MVT::v8i8;
5114 unsigned IndexLen = 8;
5115 if (Op.getValueType().getSizeInBits() == 128) {
5116 IndexVT = MVT::v16i8;
5120 SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
5121 SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);
5124 if (V2.getNode()->getOpcode() == ISD::UNDEF) {
5126 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
5127 Shuffle = DAG.getNode(
5128 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5129 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, DL, MVT::i32), V1Cst,
5130 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5131 makeArrayRef(TBLMask.data(), IndexLen)));
5133 if (IndexLen == 8) {
5134 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
5135 Shuffle = DAG.getNode(
5136 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5137 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, DL, MVT::i32), V1Cst,
5138 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5139 makeArrayRef(TBLMask.data(), IndexLen)));
5141 // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
5142 // cannot currently represent the register constraints on the input
5144 // Shuffle = DAG.getNode(AArch64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
5145 // DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5146 // &TBLMask[0], IndexLen));
5147 Shuffle = DAG.getNode(
5148 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5149 DAG.getConstant(Intrinsic::aarch64_neon_tbl2, DL, MVT::i32),
5151 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5152 makeArrayRef(TBLMask.data(), IndexLen)));
5155 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
5158 static unsigned getDUPLANEOp(EVT EltType) {
5159 if (EltType == MVT::i8)
5160 return AArch64ISD::DUPLANE8;
5161 if (EltType == MVT::i16 || EltType == MVT::f16)
5162 return AArch64ISD::DUPLANE16;
5163 if (EltType == MVT::i32 || EltType == MVT::f32)
5164 return AArch64ISD::DUPLANE32;
5165 if (EltType == MVT::i64 || EltType == MVT::f64)
5166 return AArch64ISD::DUPLANE64;
5168 llvm_unreachable("Invalid vector element type?");
5171 SDValue AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
5172 SelectionDAG &DAG) const {
5174 EVT VT = Op.getValueType();
5176 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5178 // Convert shuffles that are directly supported on NEON to target-specific
5179 // DAG nodes, instead of keeping them as shuffles and matching them again
5180 // during code selection. This is more efficient and avoids the possibility
5181 // of inconsistencies between legalization and selection.
5182 ArrayRef<int> ShuffleMask = SVN->getMask();
5184 SDValue V1 = Op.getOperand(0);
5185 SDValue V2 = Op.getOperand(1);
5187 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0],
5188 V1.getValueType().getSimpleVT())) {
5189 int Lane = SVN->getSplatIndex();
5190 // If this is undef splat, generate it via "just" vdup, if possible.
5194 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
5195 return DAG.getNode(AArch64ISD::DUP, dl, V1.getValueType(),
5197 // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
5198 // constant. If so, we can just reference the lane's definition directly.
5199 if (V1.getOpcode() == ISD::BUILD_VECTOR &&
5200 !isa<ConstantSDNode>(V1.getOperand(Lane)))
5201 return DAG.getNode(AArch64ISD::DUP, dl, VT, V1.getOperand(Lane));
5203 // Otherwise, duplicate from the lane of the input vector.
5204 unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
5206 // SelectionDAGBuilder may have "helpfully" already extracted or conatenated
5207 // to make a vector of the same size as this SHUFFLE. We can ignore the
5208 // extract entirely, and canonicalise the concat using WidenVector.
5209 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
5210 Lane += cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
5211 V1 = V1.getOperand(0);
5212 } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
5213 unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
5214 Lane -= Idx * VT.getVectorNumElements() / 2;
5215 V1 = WidenVector(V1.getOperand(Idx), DAG);
5216 } else if (VT.getSizeInBits() == 64)
5217 V1 = WidenVector(V1, DAG);
5219 return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, dl, MVT::i64));
5222 if (isREVMask(ShuffleMask, VT, 64))
5223 return DAG.getNode(AArch64ISD::REV64, dl, V1.getValueType(), V1, V2);
5224 if (isREVMask(ShuffleMask, VT, 32))
5225 return DAG.getNode(AArch64ISD::REV32, dl, V1.getValueType(), V1, V2);
5226 if (isREVMask(ShuffleMask, VT, 16))
5227 return DAG.getNode(AArch64ISD::REV16, dl, V1.getValueType(), V1, V2);
5229 bool ReverseEXT = false;
5231 if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
5234 Imm *= getExtFactor(V1);
5235 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V2,
5236 DAG.getConstant(Imm, dl, MVT::i32));
5237 } else if (V2->getOpcode() == ISD::UNDEF &&
5238 isSingletonEXTMask(ShuffleMask, VT, Imm)) {
5239 Imm *= getExtFactor(V1);
5240 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V1,
5241 DAG.getConstant(Imm, dl, MVT::i32));
5244 unsigned WhichResult;
5245 if (isZIPMask(ShuffleMask, VT, WhichResult)) {
5246 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
5247 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5249 if (isUZPMask(ShuffleMask, VT, WhichResult)) {
5250 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
5251 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5253 if (isTRNMask(ShuffleMask, VT, WhichResult)) {
5254 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
5255 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5258 if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5259 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
5260 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5262 if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5263 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
5264 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5266 if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5267 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
5268 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5271 SDValue Concat = tryFormConcatFromShuffle(Op, DAG);
5272 if (Concat.getNode())
5277 int NumInputElements = V1.getValueType().getVectorNumElements();
5278 if (isINSMask(ShuffleMask, NumInputElements, DstIsLeft, Anomaly)) {
5279 SDValue DstVec = DstIsLeft ? V1 : V2;
5280 SDValue DstLaneV = DAG.getConstant(Anomaly, dl, MVT::i64);
5282 SDValue SrcVec = V1;
5283 int SrcLane = ShuffleMask[Anomaly];
5284 if (SrcLane >= NumInputElements) {
5286 SrcLane -= VT.getVectorNumElements();
5288 SDValue SrcLaneV = DAG.getConstant(SrcLane, dl, MVT::i64);
5290 EVT ScalarVT = VT.getVectorElementType();
5292 if (ScalarVT.getSizeInBits() < 32 && ScalarVT.isInteger())
5293 ScalarVT = MVT::i32;
5296 ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5297 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ScalarVT, SrcVec, SrcLaneV),
5301 // If the shuffle is not directly supported and it has 4 elements, use
5302 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5303 unsigned NumElts = VT.getVectorNumElements();
5305 unsigned PFIndexes[4];
5306 for (unsigned i = 0; i != 4; ++i) {
5307 if (ShuffleMask[i] < 0)
5310 PFIndexes[i] = ShuffleMask[i];
5313 // Compute the index in the perfect shuffle table.
5314 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
5315 PFIndexes[2] * 9 + PFIndexes[3];
5316 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5317 unsigned Cost = (PFEntry >> 30);
5320 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5323 return GenerateTBL(Op, ShuffleMask, DAG);
5326 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
5328 EVT VT = BVN->getValueType(0);
5329 APInt SplatBits, SplatUndef;
5330 unsigned SplatBitSize;
5332 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5333 unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;
5335 for (unsigned i = 0; i < NumSplats; ++i) {
5336 CnstBits <<= SplatBitSize;
5337 UndefBits <<= SplatBitSize;
5338 CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits());
5339 UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits());
5348 SDValue AArch64TargetLowering::LowerVectorAND(SDValue Op,
5349 SelectionDAG &DAG) const {
5350 BuildVectorSDNode *BVN =
5351 dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5352 SDValue LHS = Op.getOperand(0);
5354 EVT VT = Op.getValueType();
5359 APInt CnstBits(VT.getSizeInBits(), 0);
5360 APInt UndefBits(VT.getSizeInBits(), 0);
5361 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5362 // We only have BIC vector immediate instruction, which is and-not.
5363 CnstBits = ~CnstBits;
5365 // We make use of a little bit of goto ickiness in order to avoid having to
5366 // duplicate the immediate matching logic for the undef toggled case.
5367 bool SecondTry = false;
5370 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5371 CnstBits = CnstBits.zextOrTrunc(64);
5372 uint64_t CnstVal = CnstBits.getZExtValue();
5374 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5375 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5376 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5377 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5378 DAG.getConstant(CnstVal, dl, MVT::i32),
5379 DAG.getConstant(0, dl, MVT::i32));
5380 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5383 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5384 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5385 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5386 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5387 DAG.getConstant(CnstVal, dl, MVT::i32),
5388 DAG.getConstant(8, dl, MVT::i32));
5389 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5392 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5393 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5394 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5395 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5396 DAG.getConstant(CnstVal, dl, MVT::i32),
5397 DAG.getConstant(16, dl, MVT::i32));
5398 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5401 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5402 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5403 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5404 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5405 DAG.getConstant(CnstVal, dl, MVT::i32),
5406 DAG.getConstant(24, dl, MVT::i32));
5407 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5410 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5411 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5412 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5413 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5414 DAG.getConstant(CnstVal, dl, MVT::i32),
5415 DAG.getConstant(0, dl, MVT::i32));
5416 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5419 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5420 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5421 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5422 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5423 DAG.getConstant(CnstVal, dl, MVT::i32),
5424 DAG.getConstant(8, dl, MVT::i32));
5425 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5432 CnstBits = ~UndefBits;
5436 // We can always fall back to a non-immediate AND.
5441 // Specialized code to quickly find if PotentialBVec is a BuildVector that
5442 // consists of only the same constant int value, returned in reference arg
5444 static bool isAllConstantBuildVector(const SDValue &PotentialBVec,
5445 uint64_t &ConstVal) {
5446 BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec);
5449 ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0));
5452 EVT VT = Bvec->getValueType(0);
5453 unsigned NumElts = VT.getVectorNumElements();
5454 for (unsigned i = 1; i < NumElts; ++i)
5455 if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt)
5457 ConstVal = FirstElt->getZExtValue();
5461 static unsigned getIntrinsicID(const SDNode *N) {
5462 unsigned Opcode = N->getOpcode();
5465 return Intrinsic::not_intrinsic;
5466 case ISD::INTRINSIC_WO_CHAIN: {
5467 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
5468 if (IID < Intrinsic::num_intrinsics)
5470 return Intrinsic::not_intrinsic;
5475 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
5476 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
5477 // BUILD_VECTORs with constant element C1, C2 is a constant, and C1 == ~C2.
5478 // Also, logical shift right -> sri, with the same structure.
5479 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
5480 EVT VT = N->getValueType(0);
5487 // Is the first op an AND?
5488 const SDValue And = N->getOperand(0);
5489 if (And.getOpcode() != ISD::AND)
5492 // Is the second op an shl or lshr?
5493 SDValue Shift = N->getOperand(1);
5494 // This will have been turned into: AArch64ISD::VSHL vector, #shift
5495 // or AArch64ISD::VLSHR vector, #shift
5496 unsigned ShiftOpc = Shift.getOpcode();
5497 if ((ShiftOpc != AArch64ISD::VSHL && ShiftOpc != AArch64ISD::VLSHR))
5499 bool IsShiftRight = ShiftOpc == AArch64ISD::VLSHR;
5501 // Is the shift amount constant?
5502 ConstantSDNode *C2node = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
5506 // Is the and mask vector all constant?
5508 if (!isAllConstantBuildVector(And.getOperand(1), C1))
5511 // Is C1 == ~C2, taking into account how much one can shift elements of a
5513 uint64_t C2 = C2node->getZExtValue();
5514 unsigned ElemSizeInBits = VT.getVectorElementType().getSizeInBits();
5515 if (C2 > ElemSizeInBits)
5517 unsigned ElemMask = (1 << ElemSizeInBits) - 1;
5518 if ((C1 & ElemMask) != (~C2 & ElemMask))
5521 SDValue X = And.getOperand(0);
5522 SDValue Y = Shift.getOperand(0);
5525 IsShiftRight ? Intrinsic::aarch64_neon_vsri : Intrinsic::aarch64_neon_vsli;
5527 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5528 DAG.getConstant(Intrin, DL, MVT::i32), X, Y,
5529 Shift.getOperand(1));
5531 DEBUG(dbgs() << "aarch64-lower: transformed: \n");
5532 DEBUG(N->dump(&DAG));
5533 DEBUG(dbgs() << "into: \n");
5534 DEBUG(ResultSLI->dump(&DAG));
5540 SDValue AArch64TargetLowering::LowerVectorOR(SDValue Op,
5541 SelectionDAG &DAG) const {
5542 // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
5543 if (EnableAArch64SlrGeneration) {
5544 SDValue Res = tryLowerToSLI(Op.getNode(), DAG);
5549 BuildVectorSDNode *BVN =
5550 dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode());
5551 SDValue LHS = Op.getOperand(1);
5553 EVT VT = Op.getValueType();
5555 // OR commutes, so try swapping the operands.
5557 LHS = Op.getOperand(0);
5558 BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5563 APInt CnstBits(VT.getSizeInBits(), 0);
5564 APInt UndefBits(VT.getSizeInBits(), 0);
5565 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5566 // We make use of a little bit of goto ickiness in order to avoid having to
5567 // duplicate the immediate matching logic for the undef toggled case.
5568 bool SecondTry = false;
5571 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5572 CnstBits = CnstBits.zextOrTrunc(64);
5573 uint64_t CnstVal = CnstBits.getZExtValue();
5575 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5576 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5577 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5578 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5579 DAG.getConstant(CnstVal, dl, MVT::i32),
5580 DAG.getConstant(0, dl, MVT::i32));
5581 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5584 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5585 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5586 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5587 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5588 DAG.getConstant(CnstVal, dl, MVT::i32),
5589 DAG.getConstant(8, dl, MVT::i32));
5590 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5593 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5594 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5595 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5596 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5597 DAG.getConstant(CnstVal, dl, MVT::i32),
5598 DAG.getConstant(16, dl, MVT::i32));
5599 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5602 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5603 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5604 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5605 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5606 DAG.getConstant(CnstVal, dl, MVT::i32),
5607 DAG.getConstant(24, dl, MVT::i32));
5608 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5611 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5612 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5613 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5614 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5615 DAG.getConstant(CnstVal, dl, MVT::i32),
5616 DAG.getConstant(0, dl, MVT::i32));
5617 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5620 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5621 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5622 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5623 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5624 DAG.getConstant(CnstVal, dl, MVT::i32),
5625 DAG.getConstant(8, dl, MVT::i32));
5626 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5633 CnstBits = UndefBits;
5637 // We can always fall back to a non-immediate OR.
5642 // Normalize the operands of BUILD_VECTOR. The value of constant operands will
5643 // be truncated to fit element width.
5644 static SDValue NormalizeBuildVector(SDValue Op,
5645 SelectionDAG &DAG) {
5646 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
5648 EVT VT = Op.getValueType();
5649 EVT EltTy= VT.getVectorElementType();
5651 if (EltTy.isFloatingPoint() || EltTy.getSizeInBits() > 16)
5654 SmallVector<SDValue, 16> Ops;
5655 for (unsigned I = 0, E = VT.getVectorNumElements(); I != E; ++I) {
5656 SDValue Lane = Op.getOperand(I);
5657 if (Lane.getOpcode() == ISD::Constant) {
5658 APInt LowBits(EltTy.getSizeInBits(),
5659 cast<ConstantSDNode>(Lane)->getZExtValue());
5660 Lane = DAG.getConstant(LowBits.getZExtValue(), dl, MVT::i32);
5662 Ops.push_back(Lane);
5664 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5667 SDValue AArch64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
5668 SelectionDAG &DAG) const {
5670 EVT VT = Op.getValueType();
5671 Op = NormalizeBuildVector(Op, DAG);
5672 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5674 APInt CnstBits(VT.getSizeInBits(), 0);
5675 APInt UndefBits(VT.getSizeInBits(), 0);
5676 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5677 // We make use of a little bit of goto ickiness in order to avoid having to
5678 // duplicate the immediate matching logic for the undef toggled case.
5679 bool SecondTry = false;
5682 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5683 CnstBits = CnstBits.zextOrTrunc(64);
5684 uint64_t CnstVal = CnstBits.getZExtValue();
5686 // Certain magic vector constants (used to express things like NOT
5687 // and NEG) are passed through unmodified. This allows codegen patterns
5688 // for these operations to match. Special-purpose patterns will lower
5689 // these immediates to MOVIs if it proves necessary.
5690 if (VT.isInteger() && (CnstVal == 0 || CnstVal == ~0ULL))
5693 // The many faces of MOVI...
5694 if (AArch64_AM::isAdvSIMDModImmType10(CnstVal)) {
5695 CnstVal = AArch64_AM::encodeAdvSIMDModImmType10(CnstVal);
5696 if (VT.getSizeInBits() == 128) {
5697 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::v2i64,
5698 DAG.getConstant(CnstVal, dl, MVT::i32));
5699 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5702 // Support the V64 version via subregister insertion.
5703 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::f64,
5704 DAG.getConstant(CnstVal, dl, MVT::i32));
5705 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5708 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5709 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5710 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5711 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5712 DAG.getConstant(CnstVal, dl, MVT::i32),
5713 DAG.getConstant(0, dl, MVT::i32));
5714 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5717 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5718 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5719 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5720 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5721 DAG.getConstant(CnstVal, dl, MVT::i32),
5722 DAG.getConstant(8, dl, MVT::i32));
5723 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5726 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5727 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5728 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5729 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5730 DAG.getConstant(CnstVal, dl, MVT::i32),
5731 DAG.getConstant(16, dl, MVT::i32));
5732 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5735 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5736 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5737 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5738 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5739 DAG.getConstant(CnstVal, dl, MVT::i32),
5740 DAG.getConstant(24, dl, MVT::i32));
5741 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5744 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5745 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5746 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5747 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5748 DAG.getConstant(CnstVal, dl, MVT::i32),
5749 DAG.getConstant(0, dl, MVT::i32));
5750 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5753 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5754 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5755 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5756 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5757 DAG.getConstant(CnstVal, dl, MVT::i32),
5758 DAG.getConstant(8, dl, MVT::i32));
5759 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5762 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
5763 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
5764 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5765 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
5766 DAG.getConstant(CnstVal, dl, MVT::i32),
5767 DAG.getConstant(264, dl, MVT::i32));
5768 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5771 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
5772 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
5773 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5774 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
5775 DAG.getConstant(CnstVal, dl, MVT::i32),
5776 DAG.getConstant(272, dl, MVT::i32));
5777 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5780 if (AArch64_AM::isAdvSIMDModImmType9(CnstVal)) {
5781 CnstVal = AArch64_AM::encodeAdvSIMDModImmType9(CnstVal);
5782 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;
5783 SDValue Mov = DAG.getNode(AArch64ISD::MOVI, dl, MovTy,
5784 DAG.getConstant(CnstVal, dl, MVT::i32));
5785 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5788 // The few faces of FMOV...
5789 if (AArch64_AM::isAdvSIMDModImmType11(CnstVal)) {
5790 CnstVal = AArch64_AM::encodeAdvSIMDModImmType11(CnstVal);
5791 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4f32 : MVT::v2f32;
5792 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MovTy,
5793 DAG.getConstant(CnstVal, dl, MVT::i32));
5794 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5797 if (AArch64_AM::isAdvSIMDModImmType12(CnstVal) &&
5798 VT.getSizeInBits() == 128) {
5799 CnstVal = AArch64_AM::encodeAdvSIMDModImmType12(CnstVal);
5800 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MVT::v2f64,
5801 DAG.getConstant(CnstVal, dl, MVT::i32));
5802 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5805 // The many faces of MVNI...
5807 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5808 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5809 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5810 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5811 DAG.getConstant(CnstVal, dl, MVT::i32),
5812 DAG.getConstant(0, dl, MVT::i32));
5813 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5816 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5817 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5818 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5819 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5820 DAG.getConstant(CnstVal, dl, MVT::i32),
5821 DAG.getConstant(8, dl, MVT::i32));
5822 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5825 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5826 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5827 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5828 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5829 DAG.getConstant(CnstVal, dl, MVT::i32),
5830 DAG.getConstant(16, dl, MVT::i32));
5831 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5834 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5835 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5836 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5837 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5838 DAG.getConstant(CnstVal, dl, MVT::i32),
5839 DAG.getConstant(24, dl, MVT::i32));
5840 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5843 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5844 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5845 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5846 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5847 DAG.getConstant(CnstVal, dl, MVT::i32),
5848 DAG.getConstant(0, dl, MVT::i32));
5849 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5852 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5853 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5854 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5855 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5856 DAG.getConstant(CnstVal, dl, MVT::i32),
5857 DAG.getConstant(8, dl, MVT::i32));
5858 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5861 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
5862 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
5863 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5864 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
5865 DAG.getConstant(CnstVal, dl, MVT::i32),
5866 DAG.getConstant(264, dl, MVT::i32));
5867 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5870 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
5871 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
5872 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5873 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
5874 DAG.getConstant(CnstVal, dl, MVT::i32),
5875 DAG.getConstant(272, dl, MVT::i32));
5876 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5883 CnstBits = UndefBits;
5888 // Scan through the operands to find some interesting properties we can
5890 // 1) If only one value is used, we can use a DUP, or
5891 // 2) if only the low element is not undef, we can just insert that, or
5892 // 3) if only one constant value is used (w/ some non-constant lanes),
5893 // we can splat the constant value into the whole vector then fill
5894 // in the non-constant lanes.
5895 // 4) FIXME: If different constant values are used, but we can intelligently
5896 // select the values we'll be overwriting for the non-constant
5897 // lanes such that we can directly materialize the vector
5898 // some other way (MOVI, e.g.), we can be sneaky.
5899 unsigned NumElts = VT.getVectorNumElements();
5900 bool isOnlyLowElement = true;
5901 bool usesOnlyOneValue = true;
5902 bool usesOnlyOneConstantValue = true;
5903 bool isConstant = true;
5904 unsigned NumConstantLanes = 0;
5906 SDValue ConstantValue;
5907 for (unsigned i = 0; i < NumElts; ++i) {
5908 SDValue V = Op.getOperand(i);
5909 if (V.getOpcode() == ISD::UNDEF)
5912 isOnlyLowElement = false;
5913 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5916 if (isa<ConstantSDNode>(V) || isa<ConstantFPSDNode>(V)) {
5918 if (!ConstantValue.getNode())
5920 else if (ConstantValue != V)
5921 usesOnlyOneConstantValue = false;
5924 if (!Value.getNode())
5926 else if (V != Value)
5927 usesOnlyOneValue = false;
5930 if (!Value.getNode())
5931 return DAG.getUNDEF(VT);
5933 if (isOnlyLowElement)
5934 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5936 // Use DUP for non-constant splats. For f32 constant splats, reduce to
5937 // i32 and try again.
5938 if (usesOnlyOneValue) {
5940 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5941 Value.getValueType() != VT)
5942 return DAG.getNode(AArch64ISD::DUP, dl, VT, Value);
5944 // This is actually a DUPLANExx operation, which keeps everything vectory.
5946 // DUPLANE works on 128-bit vectors, widen it if necessary.
5947 SDValue Lane = Value.getOperand(1);
5948 Value = Value.getOperand(0);
5949 if (Value.getValueType().getSizeInBits() == 64)
5950 Value = WidenVector(Value, DAG);
5952 unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
5953 return DAG.getNode(Opcode, dl, VT, Value, Lane);
5956 if (VT.getVectorElementType().isFloatingPoint()) {
5957 SmallVector<SDValue, 8> Ops;
5958 EVT EltTy = VT.getVectorElementType();
5959 assert ((EltTy == MVT::f16 || EltTy == MVT::f32 || EltTy == MVT::f64) &&
5960 "Unsupported floating-point vector type");
5961 MVT NewType = MVT::getIntegerVT(EltTy.getSizeInBits());
5962 for (unsigned i = 0; i < NumElts; ++i)
5963 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, NewType, Op.getOperand(i)));
5964 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts);
5965 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5966 Val = LowerBUILD_VECTOR(Val, DAG);
5968 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5972 // If there was only one constant value used and for more than one lane,
5973 // start by splatting that value, then replace the non-constant lanes. This
5974 // is better than the default, which will perform a separate initialization
5976 if (NumConstantLanes > 0 && usesOnlyOneConstantValue) {
5977 SDValue Val = DAG.getNode(AArch64ISD::DUP, dl, VT, ConstantValue);
5978 // Now insert the non-constant lanes.
5979 for (unsigned i = 0; i < NumElts; ++i) {
5980 SDValue V = Op.getOperand(i);
5981 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i64);
5982 if (!isa<ConstantSDNode>(V) && !isa<ConstantFPSDNode>(V)) {
5983 // Note that type legalization likely mucked about with the VT of the
5984 // source operand, so we may have to convert it here before inserting.
5985 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, V, LaneIdx);
5991 // If all elements are constants and the case above didn't get hit, fall back
5992 // to the default expansion, which will generate a load from the constant
5997 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5999 SDValue shuffle = ReconstructShuffle(Op, DAG);
6000 if (shuffle != SDValue())
6004 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
6005 // know the default expansion would otherwise fall back on something even
6006 // worse. For a vector with one or two non-undef values, that's
6007 // scalar_to_vector for the elements followed by a shuffle (provided the
6008 // shuffle is valid for the target) and materialization element by element
6009 // on the stack followed by a load for everything else.
6010 if (!isConstant && !usesOnlyOneValue) {
6011 SDValue Vec = DAG.getUNDEF(VT);
6012 SDValue Op0 = Op.getOperand(0);
6013 unsigned ElemSize = VT.getVectorElementType().getSizeInBits();
6015 // For 32 and 64 bit types, use INSERT_SUBREG for lane zero to
6016 // a) Avoid a RMW dependency on the full vector register, and
6017 // b) Allow the register coalescer to fold away the copy if the
6018 // value is already in an S or D register.
6019 if (Op0.getOpcode() != ISD::UNDEF && (ElemSize == 32 || ElemSize == 64)) {
6020 unsigned SubIdx = ElemSize == 32 ? AArch64::ssub : AArch64::dsub;
6022 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0,
6023 DAG.getTargetConstant(SubIdx, dl, MVT::i32));
6024 Vec = SDValue(N, 0);
6027 for (; i < NumElts; ++i) {
6028 SDValue V = Op.getOperand(i);
6029 if (V.getOpcode() == ISD::UNDEF)
6031 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i64);
6032 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
6037 // Just use the default expansion. We failed to find a better alternative.
6041 SDValue AArch64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
6042 SelectionDAG &DAG) const {
6043 assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!");
6045 // Check for non-constant or out of range lane.
6046 EVT VT = Op.getOperand(0).getValueType();
6047 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(2));
6048 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
6052 // Insertion/extraction are legal for V128 types.
6053 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
6054 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
6058 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
6059 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
6062 // For V64 types, we perform insertion by expanding the value
6063 // to a V128 type and perform the insertion on that.
6065 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
6066 EVT WideTy = WideVec.getValueType();
6068 SDValue Node = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideTy, WideVec,
6069 Op.getOperand(1), Op.getOperand(2));
6070 // Re-narrow the resultant vector.
6071 return NarrowVector(Node, DAG);
6075 AArch64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
6076 SelectionDAG &DAG) const {
6077 assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!");
6079 // Check for non-constant or out of range lane.
6080 EVT VT = Op.getOperand(0).getValueType();
6081 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6082 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
6086 // Insertion/extraction are legal for V128 types.
6087 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
6088 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
6092 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
6093 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
6096 // For V64 types, we perform extraction by expanding the value
6097 // to a V128 type and perform the extraction on that.
6099 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
6100 EVT WideTy = WideVec.getValueType();
6102 EVT ExtrTy = WideTy.getVectorElementType();
6103 if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8)
6106 // For extractions, we just return the result directly.
6107 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec,
6111 SDValue AArch64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
6112 SelectionDAG &DAG) const {
6113 EVT VT = Op.getOperand(0).getValueType();
6119 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6122 unsigned Val = Cst->getZExtValue();
6124 unsigned Size = Op.getValueType().getSizeInBits();
6128 return DAG.getTargetExtractSubreg(AArch64::bsub, dl, Op.getValueType(),
6131 return DAG.getTargetExtractSubreg(AArch64::hsub, dl, Op.getValueType(),
6134 return DAG.getTargetExtractSubreg(AArch64::ssub, dl, Op.getValueType(),
6137 return DAG.getTargetExtractSubreg(AArch64::dsub, dl, Op.getValueType(),
6140 llvm_unreachable("Unexpected vector type in extract_subvector!");
6143 // If this is extracting the upper 64-bits of a 128-bit vector, we match
6145 if (Size == 64 && Val * VT.getVectorElementType().getSizeInBits() == 64)
6151 bool AArch64TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
6153 if (VT.getVectorNumElements() == 4 &&
6154 (VT.is128BitVector() || VT.is64BitVector())) {
6155 unsigned PFIndexes[4];
6156 for (unsigned i = 0; i != 4; ++i) {
6160 PFIndexes[i] = M[i];
6163 // Compute the index in the perfect shuffle table.
6164 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
6165 PFIndexes[2] * 9 + PFIndexes[3];
6166 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
6167 unsigned Cost = (PFEntry >> 30);
6175 unsigned DummyUnsigned;
6177 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || isREVMask(M, VT, 64) ||
6178 isREVMask(M, VT, 32) || isREVMask(M, VT, 16) ||
6179 isEXTMask(M, VT, DummyBool, DummyUnsigned) ||
6180 // isTBLMask(M, VT) || // FIXME: Port TBL support from ARM.
6181 isTRNMask(M, VT, DummyUnsigned) || isUZPMask(M, VT, DummyUnsigned) ||
6182 isZIPMask(M, VT, DummyUnsigned) ||
6183 isTRN_v_undef_Mask(M, VT, DummyUnsigned) ||
6184 isUZP_v_undef_Mask(M, VT, DummyUnsigned) ||
6185 isZIP_v_undef_Mask(M, VT, DummyUnsigned) ||
6186 isINSMask(M, VT.getVectorNumElements(), DummyBool, DummyInt) ||
6187 isConcatMask(M, VT, VT.getSizeInBits() == 128));
6190 /// getVShiftImm - Check if this is a valid build_vector for the immediate
6191 /// operand of a vector shift operation, where all the elements of the
6192 /// build_vector must have the same constant integer value.
6193 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
6194 // Ignore bit_converts.
6195 while (Op.getOpcode() == ISD::BITCAST)
6196 Op = Op.getOperand(0);
6197 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
6198 APInt SplatBits, SplatUndef;
6199 unsigned SplatBitSize;
6201 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
6202 HasAnyUndefs, ElementBits) ||
6203 SplatBitSize > ElementBits)
6205 Cnt = SplatBits.getSExtValue();
6209 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
6210 /// operand of a vector shift left operation. That value must be in the range:
6211 /// 0 <= Value < ElementBits for a left shift; or
6212 /// 0 <= Value <= ElementBits for a long left shift.
6213 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
6214 assert(VT.isVector() && "vector shift count is not a vector type");
6215 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
6216 if (!getVShiftImm(Op, ElementBits, Cnt))
6218 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
6221 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
6222 /// operand of a vector shift right operation. For a shift opcode, the value
6223 /// is positive, but for an intrinsic the value count must be negative. The
6224 /// absolute value must be in the range:
6225 /// 1 <= |Value| <= ElementBits for a right shift; or
6226 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
6227 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
6229 assert(VT.isVector() && "vector shift count is not a vector type");
6230 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
6231 if (!getVShiftImm(Op, ElementBits, Cnt))
6235 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
6238 SDValue AArch64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
6239 SelectionDAG &DAG) const {
6240 EVT VT = Op.getValueType();
6244 if (!Op.getOperand(1).getValueType().isVector())
6246 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
6248 switch (Op.getOpcode()) {
6250 llvm_unreachable("unexpected shift opcode");
6253 if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
6254 return DAG.getNode(AArch64ISD::VSHL, DL, VT, Op.getOperand(0),
6255 DAG.getConstant(Cnt, DL, MVT::i32));
6256 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
6257 DAG.getConstant(Intrinsic::aarch64_neon_ushl, DL,
6259 Op.getOperand(0), Op.getOperand(1));
6262 // Right shift immediate
6263 if (isVShiftRImm(Op.getOperand(1), VT, false, false, Cnt) &&
6266 (Op.getOpcode() == ISD::SRA) ? AArch64ISD::VASHR : AArch64ISD::VLSHR;
6267 return DAG.getNode(Opc, DL, VT, Op.getOperand(0),
6268 DAG.getConstant(Cnt, DL, MVT::i32));
6271 // Right shift register. Note, there is not a shift right register
6272 // instruction, but the shift left register instruction takes a signed
6273 // value, where negative numbers specify a right shift.
6274 unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::aarch64_neon_sshl
6275 : Intrinsic::aarch64_neon_ushl;
6276 // negate the shift amount
6277 SDValue NegShift = DAG.getNode(AArch64ISD::NEG, DL, VT, Op.getOperand(1));
6278 SDValue NegShiftLeft =
6279 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
6280 DAG.getConstant(Opc, DL, MVT::i32), Op.getOperand(0),
6282 return NegShiftLeft;
6288 static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS,
6289 AArch64CC::CondCode CC, bool NoNans, EVT VT,
6290 SDLoc dl, SelectionDAG &DAG) {
6291 EVT SrcVT = LHS.getValueType();
6292 assert(VT.getSizeInBits() == SrcVT.getSizeInBits() &&
6293 "function only supposed to emit natural comparisons");
6295 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
6296 APInt CnstBits(VT.getSizeInBits(), 0);
6297 APInt UndefBits(VT.getSizeInBits(), 0);
6298 bool IsCnst = BVN && resolveBuildVector(BVN, CnstBits, UndefBits);
6299 bool IsZero = IsCnst && (CnstBits == 0);
6301 if (SrcVT.getVectorElementType().isFloatingPoint()) {
6305 case AArch64CC::NE: {
6308 Fcmeq = DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
6310 Fcmeq = DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
6311 return DAG.getNode(AArch64ISD::NOT, dl, VT, Fcmeq);
6315 return DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
6316 return DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
6319 return DAG.getNode(AArch64ISD::FCMGEz, dl, VT, LHS);
6320 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, LHS, RHS);
6323 return DAG.getNode(AArch64ISD::FCMGTz, dl, VT, LHS);
6324 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, LHS, RHS);
6327 return DAG.getNode(AArch64ISD::FCMLEz, dl, VT, LHS);
6328 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, RHS, LHS);
6332 // If we ignore NaNs then we can use to the MI implementation.
6336 return DAG.getNode(AArch64ISD::FCMLTz, dl, VT, LHS);
6337 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, RHS, LHS);
6344 case AArch64CC::NE: {
6347 Cmeq = DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
6349 Cmeq = DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
6350 return DAG.getNode(AArch64ISD::NOT, dl, VT, Cmeq);
6354 return DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
6355 return DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
6358 return DAG.getNode(AArch64ISD::CMGEz, dl, VT, LHS);
6359 return DAG.getNode(AArch64ISD::CMGE, dl, VT, LHS, RHS);
6362 return DAG.getNode(AArch64ISD::CMGTz, dl, VT, LHS);
6363 return DAG.getNode(AArch64ISD::CMGT, dl, VT, LHS, RHS);
6366 return DAG.getNode(AArch64ISD::CMLEz, dl, VT, LHS);
6367 return DAG.getNode(AArch64ISD::CMGE, dl, VT, RHS, LHS);
6369 return DAG.getNode(AArch64ISD::CMHS, dl, VT, RHS, LHS);
6371 return DAG.getNode(AArch64ISD::CMHI, dl, VT, RHS, LHS);
6374 return DAG.getNode(AArch64ISD::CMLTz, dl, VT, LHS);
6375 return DAG.getNode(AArch64ISD::CMGT, dl, VT, RHS, LHS);
6377 return DAG.getNode(AArch64ISD::CMHI, dl, VT, LHS, RHS);
6379 return DAG.getNode(AArch64ISD::CMHS, dl, VT, LHS, RHS);
6383 SDValue AArch64TargetLowering::LowerVSETCC(SDValue Op,
6384 SelectionDAG &DAG) const {
6385 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6386 SDValue LHS = Op.getOperand(0);
6387 SDValue RHS = Op.getOperand(1);
6388 EVT CmpVT = LHS.getValueType().changeVectorElementTypeToInteger();
6391 if (LHS.getValueType().getVectorElementType().isInteger()) {
6392 assert(LHS.getValueType() == RHS.getValueType());
6393 AArch64CC::CondCode AArch64CC = changeIntCCToAArch64CC(CC);
6395 EmitVectorComparison(LHS, RHS, AArch64CC, false, CmpVT, dl, DAG);
6396 return DAG.getSExtOrTrunc(Cmp, dl, Op.getValueType());
6399 assert(LHS.getValueType().getVectorElementType() == MVT::f32 ||
6400 LHS.getValueType().getVectorElementType() == MVT::f64);
6402 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
6403 // clean. Some of them require two branches to implement.
6404 AArch64CC::CondCode CC1, CC2;
6406 changeVectorFPCCToAArch64CC(CC, CC1, CC2, ShouldInvert);
6408 bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath;
6410 EmitVectorComparison(LHS, RHS, CC1, NoNaNs, CmpVT, dl, DAG);
6414 if (CC2 != AArch64CC::AL) {
6416 EmitVectorComparison(LHS, RHS, CC2, NoNaNs, CmpVT, dl, DAG);
6417 if (!Cmp2.getNode())
6420 Cmp = DAG.getNode(ISD::OR, dl, CmpVT, Cmp, Cmp2);
6423 Cmp = DAG.getSExtOrTrunc(Cmp, dl, Op.getValueType());
6426 return Cmp = DAG.getNOT(dl, Cmp, Cmp.getValueType());
6431 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
6432 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
6433 /// specified in the intrinsic calls.
6434 bool AArch64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
6436 unsigned Intrinsic) const {
6437 switch (Intrinsic) {
6438 case Intrinsic::aarch64_neon_ld2:
6439 case Intrinsic::aarch64_neon_ld3:
6440 case Intrinsic::aarch64_neon_ld4:
6441 case Intrinsic::aarch64_neon_ld1x2:
6442 case Intrinsic::aarch64_neon_ld1x3:
6443 case Intrinsic::aarch64_neon_ld1x4:
6444 case Intrinsic::aarch64_neon_ld2lane:
6445 case Intrinsic::aarch64_neon_ld3lane:
6446 case Intrinsic::aarch64_neon_ld4lane:
6447 case Intrinsic::aarch64_neon_ld2r:
6448 case Intrinsic::aarch64_neon_ld3r:
6449 case Intrinsic::aarch64_neon_ld4r: {
6450 Info.opc = ISD::INTRINSIC_W_CHAIN;
6451 // Conservatively set memVT to the entire set of vectors loaded.
6452 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
6453 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6454 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6457 Info.vol = false; // volatile loads with NEON intrinsics not supported
6458 Info.readMem = true;
6459 Info.writeMem = false;
6462 case Intrinsic::aarch64_neon_st2:
6463 case Intrinsic::aarch64_neon_st3:
6464 case Intrinsic::aarch64_neon_st4:
6465 case Intrinsic::aarch64_neon_st1x2:
6466 case Intrinsic::aarch64_neon_st1x3:
6467 case Intrinsic::aarch64_neon_st1x4:
6468 case Intrinsic::aarch64_neon_st2lane:
6469 case Intrinsic::aarch64_neon_st3lane:
6470 case Intrinsic::aarch64_neon_st4lane: {
6471 Info.opc = ISD::INTRINSIC_VOID;
6472 // Conservatively set memVT to the entire set of vectors stored.
6473 unsigned NumElts = 0;
6474 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
6475 Type *ArgTy = I.getArgOperand(ArgI)->getType();
6476 if (!ArgTy->isVectorTy())
6478 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
6480 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6481 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6484 Info.vol = false; // volatile stores with NEON intrinsics not supported
6485 Info.readMem = false;
6486 Info.writeMem = true;
6489 case Intrinsic::aarch64_ldaxr:
6490 case Intrinsic::aarch64_ldxr: {
6491 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
6492 Info.opc = ISD::INTRINSIC_W_CHAIN;
6493 Info.memVT = MVT::getVT(PtrTy->getElementType());
6494 Info.ptrVal = I.getArgOperand(0);
6496 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
6498 Info.readMem = true;
6499 Info.writeMem = false;
6502 case Intrinsic::aarch64_stlxr:
6503 case Intrinsic::aarch64_stxr: {
6504 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
6505 Info.opc = ISD::INTRINSIC_W_CHAIN;
6506 Info.memVT = MVT::getVT(PtrTy->getElementType());
6507 Info.ptrVal = I.getArgOperand(1);
6509 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
6511 Info.readMem = false;
6512 Info.writeMem = true;
6515 case Intrinsic::aarch64_ldaxp:
6516 case Intrinsic::aarch64_ldxp: {
6517 Info.opc = ISD::INTRINSIC_W_CHAIN;
6518 Info.memVT = MVT::i128;
6519 Info.ptrVal = I.getArgOperand(0);
6523 Info.readMem = true;
6524 Info.writeMem = false;
6527 case Intrinsic::aarch64_stlxp:
6528 case Intrinsic::aarch64_stxp: {
6529 Info.opc = ISD::INTRINSIC_W_CHAIN;
6530 Info.memVT = MVT::i128;
6531 Info.ptrVal = I.getArgOperand(2);
6535 Info.readMem = false;
6536 Info.writeMem = true;
6546 // Truncations from 64-bit GPR to 32-bit GPR is free.
6547 bool AArch64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
6548 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6550 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6551 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6552 return NumBits1 > NumBits2;
6554 bool AArch64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
6555 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6557 unsigned NumBits1 = VT1.getSizeInBits();
6558 unsigned NumBits2 = VT2.getSizeInBits();
6559 return NumBits1 > NumBits2;
6562 /// Check if it is profitable to hoist instruction in then/else to if.
6563 /// Not profitable if I and it's user can form a FMA instruction
6564 /// because we prefer FMSUB/FMADD.
6565 bool AArch64TargetLowering::isProfitableToHoist(Instruction *I) const {
6566 if (I->getOpcode() != Instruction::FMul)
6569 if (I->getNumUses() != 1)
6572 Instruction *User = I->user_back();
6575 !(User->getOpcode() == Instruction::FSub ||
6576 User->getOpcode() == Instruction::FAdd))
6579 const TargetOptions &Options = getTargetMachine().Options;
6580 const DataLayout &DL = I->getModule()->getDataLayout();
6581 EVT VT = getValueType(DL, User->getOperand(0)->getType());
6583 if (isFMAFasterThanFMulAndFAdd(VT) &&
6584 isOperationLegalOrCustom(ISD::FMA, VT) &&
6585 (Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath))
6591 // All 32-bit GPR operations implicitly zero the high-half of the corresponding
6593 bool AArch64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
6594 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6596 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6597 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6598 return NumBits1 == 32 && NumBits2 == 64;
6600 bool AArch64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
6601 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6603 unsigned NumBits1 = VT1.getSizeInBits();
6604 unsigned NumBits2 = VT2.getSizeInBits();
6605 return NumBits1 == 32 && NumBits2 == 64;
6608 bool AArch64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
6609 EVT VT1 = Val.getValueType();
6610 if (isZExtFree(VT1, VT2)) {
6614 if (Val.getOpcode() != ISD::LOAD)
6617 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
6618 return (VT1.isSimple() && !VT1.isVector() && VT1.isInteger() &&
6619 VT2.isSimple() && !VT2.isVector() && VT2.isInteger() &&
6620 VT1.getSizeInBits() <= 32);
6623 bool AArch64TargetLowering::isExtFreeImpl(const Instruction *Ext) const {
6624 if (isa<FPExtInst>(Ext))
6627 // Vector types are next free.
6628 if (Ext->getType()->isVectorTy())
6631 for (const Use &U : Ext->uses()) {
6632 // The extension is free if we can fold it with a left shift in an
6633 // addressing mode or an arithmetic operation: add, sub, and cmp.
6635 // Is there a shift?
6636 const Instruction *Instr = cast<Instruction>(U.getUser());
6638 // Is this a constant shift?
6639 switch (Instr->getOpcode()) {
6640 case Instruction::Shl:
6641 if (!isa<ConstantInt>(Instr->getOperand(1)))
6644 case Instruction::GetElementPtr: {
6645 gep_type_iterator GTI = gep_type_begin(Instr);
6646 std::advance(GTI, U.getOperandNo());
6648 // This extension will end up with a shift because of the scaling factor.
6649 // 8-bit sized types have a scaling factor of 1, thus a shift amount of 0.
6650 // Get the shift amount based on the scaling factor:
6651 // log2(sizeof(IdxTy)) - log2(8).
6653 countTrailingZeros(getDataLayout()->getTypeStoreSizeInBits(IdxTy)) - 3;
6654 // Is the constant foldable in the shift of the addressing mode?
6655 // I.e., shift amount is between 1 and 4 inclusive.
6656 if (ShiftAmt == 0 || ShiftAmt > 4)
6660 case Instruction::Trunc:
6661 // Check if this is a noop.
6662 // trunc(sext ty1 to ty2) to ty1.
6663 if (Instr->getType() == Ext->getOperand(0)->getType())
6670 // At this point we can use the bfm family, so this extension is free
6676 bool AArch64TargetLowering::hasPairedLoad(Type *LoadedType,
6677 unsigned &RequiredAligment) const {
6678 if (!LoadedType->isIntegerTy() && !LoadedType->isFloatTy())
6680 // Cyclone supports unaligned accesses.
6681 RequiredAligment = 0;
6682 unsigned NumBits = LoadedType->getPrimitiveSizeInBits();
6683 return NumBits == 32 || NumBits == 64;
6686 bool AArch64TargetLowering::hasPairedLoad(EVT LoadedType,
6687 unsigned &RequiredAligment) const {
6688 if (!LoadedType.isSimple() ||
6689 (!LoadedType.isInteger() && !LoadedType.isFloatingPoint()))
6691 // Cyclone supports unaligned accesses.
6692 RequiredAligment = 0;
6693 unsigned NumBits = LoadedType.getSizeInBits();
6694 return NumBits == 32 || NumBits == 64;
6697 /// \brief Lower an interleaved load into a ldN intrinsic.
6699 /// E.g. Lower an interleaved load (Factor = 2):
6700 /// %wide.vec = load <8 x i32>, <8 x i32>* %ptr
6701 /// %v0 = shuffle %wide.vec, undef, <0, 2, 4, 6> ; Extract even elements
6702 /// %v1 = shuffle %wide.vec, undef, <1, 3, 5, 7> ; Extract odd elements
6705 /// %ld2 = { <4 x i32>, <4 x i32> } call llvm.aarch64.neon.ld2(%ptr)
6706 /// %vec0 = extractelement { <4 x i32>, <4 x i32> } %ld2, i32 0
6707 /// %vec1 = extractelement { <4 x i32>, <4 x i32> } %ld2, i32 1
6708 bool AArch64TargetLowering::lowerInterleavedLoad(
6709 LoadInst *LI, ArrayRef<ShuffleVectorInst *> Shuffles,
6710 ArrayRef<unsigned> Indices, unsigned Factor) const {
6711 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
6712 "Invalid interleave factor");
6713 assert(!Shuffles.empty() && "Empty shufflevector input");
6714 assert(Shuffles.size() == Indices.size() &&
6715 "Unmatched number of shufflevectors and indices");
6717 const DataLayout *DL = getDataLayout();
6719 VectorType *VecTy = Shuffles[0]->getType();
6720 unsigned VecSize = DL->getTypeAllocSizeInBits(VecTy);
6722 // Skip illegal vector types.
6723 if (VecSize != 64 && VecSize != 128)
6726 // A pointer vector can not be the return type of the ldN intrinsics. Need to
6727 // load integer vectors first and then convert to pointer vectors.
6728 Type *EltTy = VecTy->getVectorElementType();
6729 if (EltTy->isPointerTy())
6730 VecTy = VectorType::get(DL->getIntPtrType(EltTy),
6731 VecTy->getVectorNumElements());
6733 Type *PtrTy = VecTy->getPointerTo(LI->getPointerAddressSpace());
6734 Type *Tys[2] = {VecTy, PtrTy};
6735 static const Intrinsic::ID LoadInts[3] = {Intrinsic::aarch64_neon_ld2,
6736 Intrinsic::aarch64_neon_ld3,
6737 Intrinsic::aarch64_neon_ld4};
6739 Intrinsic::getDeclaration(LI->getModule(), LoadInts[Factor - 2], Tys);
6741 IRBuilder<> Builder(LI);
6742 Value *Ptr = Builder.CreateBitCast(LI->getPointerOperand(), PtrTy);
6744 CallInst *LdN = Builder.CreateCall(LdNFunc, Ptr, "ldN");
6746 // Replace uses of each shufflevector with the corresponding vector loaded
6748 for (unsigned i = 0; i < Shuffles.size(); i++) {
6749 ShuffleVectorInst *SVI = Shuffles[i];
6750 unsigned Index = Indices[i];
6752 Value *SubVec = Builder.CreateExtractValue(LdN, Index);
6754 // Convert the integer vector to pointer vector if the element is pointer.
6755 if (EltTy->isPointerTy())
6756 SubVec = Builder.CreateIntToPtr(SubVec, SVI->getType());
6758 SVI->replaceAllUsesWith(SubVec);
6764 /// \brief Get a mask consisting of sequential integers starting from \p Start.
6766 /// I.e. <Start, Start + 1, ..., Start + NumElts - 1>
6767 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned Start,
6769 SmallVector<Constant *, 16> Mask;
6770 for (unsigned i = 0; i < NumElts; i++)
6771 Mask.push_back(Builder.getInt32(Start + i));
6773 return ConstantVector::get(Mask);
6776 /// \brief Lower an interleaved store into a stN intrinsic.
6778 /// E.g. Lower an interleaved store (Factor = 3):
6779 /// %i.vec = shuffle <8 x i32> %v0, <8 x i32> %v1,
6780 /// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11>
6781 /// store <12 x i32> %i.vec, <12 x i32>* %ptr
6784 /// %sub.v0 = shuffle <8 x i32> %v0, <8 x i32> v1, <0, 1, 2, 3>
6785 /// %sub.v1 = shuffle <8 x i32> %v0, <8 x i32> v1, <4, 5, 6, 7>
6786 /// %sub.v2 = shuffle <8 x i32> %v0, <8 x i32> v1, <8, 9, 10, 11>
6787 /// call void llvm.aarch64.neon.st3(%sub.v0, %sub.v1, %sub.v2, %ptr)
6789 /// Note that the new shufflevectors will be removed and we'll only generate one
6790 /// st3 instruction in CodeGen.
6791 bool AArch64TargetLowering::lowerInterleavedStore(StoreInst *SI,
6792 ShuffleVectorInst *SVI,
6793 unsigned Factor) const {
6794 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
6795 "Invalid interleave factor");
6797 VectorType *VecTy = SVI->getType();
6798 assert(VecTy->getVectorNumElements() % Factor == 0 &&
6799 "Invalid interleaved store");
6801 unsigned NumSubElts = VecTy->getVectorNumElements() / Factor;
6802 Type *EltTy = VecTy->getVectorElementType();
6803 VectorType *SubVecTy = VectorType::get(EltTy, NumSubElts);
6805 const DataLayout *DL = getDataLayout();
6806 unsigned SubVecSize = DL->getTypeAllocSizeInBits(SubVecTy);
6808 // Skip illegal vector types.
6809 if (SubVecSize != 64 && SubVecSize != 128)
6812 Value *Op0 = SVI->getOperand(0);
6813 Value *Op1 = SVI->getOperand(1);
6814 IRBuilder<> Builder(SI);
6816 // StN intrinsics don't support pointer vectors as arguments. Convert pointer
6817 // vectors to integer vectors.
6818 if (EltTy->isPointerTy()) {
6819 Type *IntTy = DL->getIntPtrType(EltTy);
6820 unsigned NumOpElts =
6821 dyn_cast<VectorType>(Op0->getType())->getVectorNumElements();
6823 // Convert to the corresponding integer vector.
6824 Type *IntVecTy = VectorType::get(IntTy, NumOpElts);
6825 Op0 = Builder.CreatePtrToInt(Op0, IntVecTy);
6826 Op1 = Builder.CreatePtrToInt(Op1, IntVecTy);
6828 SubVecTy = VectorType::get(IntTy, NumSubElts);
6831 Type *PtrTy = SubVecTy->getPointerTo(SI->getPointerAddressSpace());
6832 Type *Tys[2] = {SubVecTy, PtrTy};
6833 static const Intrinsic::ID StoreInts[3] = {Intrinsic::aarch64_neon_st2,
6834 Intrinsic::aarch64_neon_st3,
6835 Intrinsic::aarch64_neon_st4};
6837 Intrinsic::getDeclaration(SI->getModule(), StoreInts[Factor - 2], Tys);
6839 SmallVector<Value *, 5> Ops;
6841 // Split the shufflevector operands into sub vectors for the new stN call.
6842 for (unsigned i = 0; i < Factor; i++)
6843 Ops.push_back(Builder.CreateShuffleVector(
6844 Op0, Op1, getSequentialMask(Builder, NumSubElts * i, NumSubElts)));
6846 Ops.push_back(Builder.CreateBitCast(SI->getPointerOperand(), PtrTy));
6847 Builder.CreateCall(StNFunc, Ops);
6851 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
6852 unsigned AlignCheck) {
6853 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
6854 (DstAlign == 0 || DstAlign % AlignCheck == 0));
6857 EVT AArch64TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
6858 unsigned SrcAlign, bool IsMemset,
6861 MachineFunction &MF) const {
6862 // Don't use AdvSIMD to implement 16-byte memset. It would have taken one
6863 // instruction to materialize the v2i64 zero and one store (with restrictive
6864 // addressing mode). Just do two i64 store of zero-registers.
6866 const Function *F = MF.getFunction();
6867 if (Subtarget->hasFPARMv8() && !IsMemset && Size >= 16 &&
6868 !F->hasFnAttribute(Attribute::NoImplicitFloat) &&
6869 (memOpAlign(SrcAlign, DstAlign, 16) ||
6870 (allowsMisalignedMemoryAccesses(MVT::f128, 0, 1, &Fast) && Fast)))
6874 (memOpAlign(SrcAlign, DstAlign, 8) ||
6875 (allowsMisalignedMemoryAccesses(MVT::i64, 0, 1, &Fast) && Fast)))
6879 (memOpAlign(SrcAlign, DstAlign, 4) ||
6880 (allowsMisalignedMemoryAccesses(MVT::i32, 0, 1, &Fast) && Fast)))
6886 // 12-bit optionally shifted immediates are legal for adds.
6887 bool AArch64TargetLowering::isLegalAddImmediate(int64_t Immed) const {
6888 if ((Immed >> 12) == 0 || ((Immed & 0xfff) == 0 && Immed >> 24 == 0))
6893 // Integer comparisons are implemented with ADDS/SUBS, so the range of valid
6894 // immediates is the same as for an add or a sub.
6895 bool AArch64TargetLowering::isLegalICmpImmediate(int64_t Immed) const {
6898 return isLegalAddImmediate(Immed);
6901 /// isLegalAddressingMode - Return true if the addressing mode represented
6902 /// by AM is legal for this target, for a load/store of the specified type.
6903 bool AArch64TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6905 unsigned AS) const {
6906 // AArch64 has five basic addressing modes:
6908 // reg + 9-bit signed offset
6909 // reg + SIZE_IN_BYTES * 12-bit unsigned offset
6911 // reg + SIZE_IN_BYTES * reg
6913 // No global is ever allowed as a base.
6917 // No reg+reg+imm addressing.
6918 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
6921 // check reg + imm case:
6922 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
6923 uint64_t NumBytes = 0;
6924 if (Ty->isSized()) {
6925 uint64_t NumBits = getDataLayout()->getTypeSizeInBits(Ty);
6926 NumBytes = NumBits / 8;
6927 if (!isPowerOf2_64(NumBits))
6932 int64_t Offset = AM.BaseOffs;
6934 // 9-bit signed offset
6935 if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
6938 // 12-bit unsigned offset
6939 unsigned shift = Log2_64(NumBytes);
6940 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
6941 // Must be a multiple of NumBytes (NumBytes is a power of 2)
6942 (Offset >> shift) << shift == Offset)
6947 // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2
6949 if (!AM.Scale || AM.Scale == 1 ||
6950 (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
6955 int AArch64TargetLowering::getScalingFactorCost(const AddrMode &AM,
6957 unsigned AS) const {
6958 // Scaling factors are not free at all.
6959 // Operands | Rt Latency
6960 // -------------------------------------------
6962 // -------------------------------------------
6963 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
6964 // Rt, [Xn, Wm, <extend> #imm] |
6965 if (isLegalAddressingMode(AM, Ty, AS))
6966 // Scale represents reg2 * scale, thus account for 1 if
6967 // it is not equal to 0 or 1.
6968 return AM.Scale != 0 && AM.Scale != 1;
6972 bool AArch64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
6973 VT = VT.getScalarType();
6978 switch (VT.getSimpleVT().SimpleTy) {
6990 AArch64TargetLowering::getScratchRegisters(CallingConv::ID) const {
6991 // LR is a callee-save register, but we must treat it as clobbered by any call
6992 // site. Hence we include LR in the scratch registers, which are in turn added
6993 // as implicit-defs for stackmaps and patchpoints.
6994 static const MCPhysReg ScratchRegs[] = {
6995 AArch64::X16, AArch64::X17, AArch64::LR, 0
7001 AArch64TargetLowering::isDesirableToCommuteWithShift(const SDNode *N) const {
7002 EVT VT = N->getValueType(0);
7003 // If N is unsigned bit extraction: ((x >> C) & mask), then do not combine
7004 // it with shift to let it be lowered to UBFX.
7005 if (N->getOpcode() == ISD::AND && (VT == MVT::i32 || VT == MVT::i64) &&
7006 isa<ConstantSDNode>(N->getOperand(1))) {
7007 uint64_t TruncMask = N->getConstantOperandVal(1);
7008 if (isMask_64(TruncMask) &&
7009 N->getOperand(0).getOpcode() == ISD::SRL &&
7010 isa<ConstantSDNode>(N->getOperand(0)->getOperand(1)))
7016 bool AArch64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
7018 assert(Ty->isIntegerTy());
7020 unsigned BitSize = Ty->getPrimitiveSizeInBits();
7024 int64_t Val = Imm.getSExtValue();
7025 if (Val == 0 || AArch64_AM::isLogicalImmediate(Val, BitSize))
7028 if ((int64_t)Val < 0)
7031 Val &= (1LL << 32) - 1;
7033 unsigned LZ = countLeadingZeros((uint64_t)Val);
7034 unsigned Shift = (63 - LZ) / 16;
7035 // MOVZ is free so return true for one or fewer MOVK.
7039 // Generate SUBS and CSEL for integer abs.
7040 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
7041 EVT VT = N->getValueType(0);
7043 SDValue N0 = N->getOperand(0);
7044 SDValue N1 = N->getOperand(1);
7047 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
7048 // and change it to SUB and CSEL.
7049 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
7050 N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
7051 N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0))
7052 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
7053 if (Y1C->getAPIntValue() == VT.getSizeInBits() - 1) {
7054 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT),
7056 // Generate SUBS & CSEL.
7058 DAG.getNode(AArch64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32),
7059 N0.getOperand(0), DAG.getConstant(0, DL, VT));
7060 return DAG.getNode(AArch64ISD::CSEL, DL, VT, N0.getOperand(0), Neg,
7061 DAG.getConstant(AArch64CC::PL, DL, MVT::i32),
7062 SDValue(Cmp.getNode(), 1));
7067 // performXorCombine - Attempts to handle integer ABS.
7068 static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
7069 TargetLowering::DAGCombinerInfo &DCI,
7070 const AArch64Subtarget *Subtarget) {
7071 if (DCI.isBeforeLegalizeOps())
7074 return performIntegerAbsCombine(N, DAG);
7078 AArch64TargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor,
7080 std::vector<SDNode *> *Created) const {
7081 // fold (sdiv X, pow2)
7082 EVT VT = N->getValueType(0);
7083 if ((VT != MVT::i32 && VT != MVT::i64) ||
7084 !(Divisor.isPowerOf2() || (-Divisor).isPowerOf2()))
7088 SDValue N0 = N->getOperand(0);
7089 unsigned Lg2 = Divisor.countTrailingZeros();
7090 SDValue Zero = DAG.getConstant(0, DL, VT);
7091 SDValue Pow2MinusOne = DAG.getConstant((1ULL << Lg2) - 1, DL, VT);
7093 // Add (N0 < 0) ? Pow2 - 1 : 0;
7095 SDValue Cmp = getAArch64Cmp(N0, Zero, ISD::SETLT, CCVal, DAG, DL);
7096 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0, Pow2MinusOne);
7097 SDValue CSel = DAG.getNode(AArch64ISD::CSEL, DL, VT, Add, N0, CCVal, Cmp);
7100 Created->push_back(Cmp.getNode());
7101 Created->push_back(Add.getNode());
7102 Created->push_back(CSel.getNode());
7107 DAG.getNode(ISD::SRA, DL, VT, CSel, DAG.getConstant(Lg2, DL, MVT::i64));
7109 // If we're dividing by a positive value, we're done. Otherwise, we must
7110 // negate the result.
7111 if (Divisor.isNonNegative())
7115 Created->push_back(SRA.getNode());
7116 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), SRA);
7119 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG,
7120 TargetLowering::DAGCombinerInfo &DCI,
7121 const AArch64Subtarget *Subtarget) {
7122 if (DCI.isBeforeLegalizeOps())
7125 // Multiplication of a power of two plus/minus one can be done more
7126 // cheaply as as shift+add/sub. For now, this is true unilaterally. If
7127 // future CPUs have a cheaper MADD instruction, this may need to be
7128 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
7129 // 64-bit is 5 cycles, so this is always a win.
7130 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
7131 APInt Value = C->getAPIntValue();
7132 EVT VT = N->getValueType(0);
7134 if (Value.isNonNegative()) {
7135 // (mul x, 2^N + 1) => (add (shl x, N), x)
7136 APInt VM1 = Value - 1;
7137 if (VM1.isPowerOf2()) {
7138 SDValue ShiftedVal =
7139 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7140 DAG.getConstant(VM1.logBase2(), DL, MVT::i64));
7141 return DAG.getNode(ISD::ADD, DL, VT, ShiftedVal,
7144 // (mul x, 2^N - 1) => (sub (shl x, N), x)
7145 APInt VP1 = Value + 1;
7146 if (VP1.isPowerOf2()) {
7147 SDValue ShiftedVal =
7148 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7149 DAG.getConstant(VP1.logBase2(), DL, MVT::i64));
7150 return DAG.getNode(ISD::SUB, DL, VT, ShiftedVal,
7154 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
7155 APInt VNP1 = -Value + 1;
7156 if (VNP1.isPowerOf2()) {
7157 SDValue ShiftedVal =
7158 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7159 DAG.getConstant(VNP1.logBase2(), DL, MVT::i64));
7160 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0),
7163 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
7164 APInt VNM1 = -Value - 1;
7165 if (VNM1.isPowerOf2()) {
7166 SDValue ShiftedVal =
7167 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7168 DAG.getConstant(VNM1.logBase2(), DL, MVT::i64));
7170 DAG.getNode(ISD::ADD, DL, VT, ShiftedVal, N->getOperand(0));
7171 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Add);
7178 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
7179 SelectionDAG &DAG) {
7180 // Take advantage of vector comparisons producing 0 or -1 in each lane to
7181 // optimize away operation when it's from a constant.
7183 // The general transformation is:
7184 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
7185 // AND(VECTOR_CMP(x,y), constant2)
7186 // constant2 = UNARYOP(constant)
7188 // Early exit if this isn't a vector operation, the operand of the
7189 // unary operation isn't a bitwise AND, or if the sizes of the operations
7191 EVT VT = N->getValueType(0);
7192 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
7193 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
7194 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
7197 // Now check that the other operand of the AND is a constant. We could
7198 // make the transformation for non-constant splats as well, but it's unclear
7199 // that would be a benefit as it would not eliminate any operations, just
7200 // perform one more step in scalar code before moving to the vector unit.
7201 if (BuildVectorSDNode *BV =
7202 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
7203 // Bail out if the vector isn't a constant.
7204 if (!BV->isConstant())
7207 // Everything checks out. Build up the new and improved node.
7209 EVT IntVT = BV->getValueType(0);
7210 // Create a new constant of the appropriate type for the transformed
7212 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
7213 // The AND node needs bitcasts to/from an integer vector type around it.
7214 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
7215 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
7216 N->getOperand(0)->getOperand(0), MaskConst);
7217 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
7224 static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG,
7225 const AArch64Subtarget *Subtarget) {
7226 // First try to optimize away the conversion when it's conditionally from
7227 // a constant. Vectors only.
7228 SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
7229 if (Res != SDValue())
7232 EVT VT = N->getValueType(0);
7233 if (VT != MVT::f32 && VT != MVT::f64)
7236 // Only optimize when the source and destination types have the same width.
7237 if (VT.getSizeInBits() != N->getOperand(0).getValueType().getSizeInBits())
7240 // If the result of an integer load is only used by an integer-to-float
7241 // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
7242 // This eliminates an "integer-to-vector-move UOP and improve throughput.
7243 SDValue N0 = N->getOperand(0);
7244 if (Subtarget->hasNEON() && ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7245 // Do not change the width of a volatile load.
7246 !cast<LoadSDNode>(N0)->isVolatile()) {
7247 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7248 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
7249 LN0->getPointerInfo(), LN0->isVolatile(),
7250 LN0->isNonTemporal(), LN0->isInvariant(),
7251 LN0->getAlignment());
7253 // Make sure successors of the original load stay after it by updating them
7254 // to use the new Chain.
7255 DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));
7258 (N->getOpcode() == ISD::SINT_TO_FP) ? AArch64ISD::SITOF : AArch64ISD::UITOF;
7259 return DAG.getNode(Opcode, SDLoc(N), VT, Load);
7265 /// An EXTR instruction is made up of two shifts, ORed together. This helper
7266 /// searches for and classifies those shifts.
7267 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
7269 if (N.getOpcode() == ISD::SHL)
7271 else if (N.getOpcode() == ISD::SRL)
7276 if (!isa<ConstantSDNode>(N.getOperand(1)))
7279 ShiftAmount = N->getConstantOperandVal(1);
7280 Src = N->getOperand(0);
7284 /// EXTR instruction extracts a contiguous chunk of bits from two existing
7285 /// registers viewed as a high/low pair. This function looks for the pattern:
7286 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
7287 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
7289 static SDValue tryCombineToEXTR(SDNode *N,
7290 TargetLowering::DAGCombinerInfo &DCI) {
7291 SelectionDAG &DAG = DCI.DAG;
7293 EVT VT = N->getValueType(0);
7295 assert(N->getOpcode() == ISD::OR && "Unexpected root");
7297 if (VT != MVT::i32 && VT != MVT::i64)
7301 uint32_t ShiftLHS = 0;
7303 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
7307 uint32_t ShiftRHS = 0;
7309 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
7312 // If they're both trying to come from the high part of the register, they're
7313 // not really an EXTR.
7314 if (LHSFromHi == RHSFromHi)
7317 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
7321 std::swap(LHS, RHS);
7322 std::swap(ShiftLHS, ShiftRHS);
7325 return DAG.getNode(AArch64ISD::EXTR, DL, VT, LHS, RHS,
7326 DAG.getConstant(ShiftRHS, DL, MVT::i64));
7329 static SDValue tryCombineToBSL(SDNode *N,
7330 TargetLowering::DAGCombinerInfo &DCI) {
7331 EVT VT = N->getValueType(0);
7332 SelectionDAG &DAG = DCI.DAG;
7338 SDValue N0 = N->getOperand(0);
7339 if (N0.getOpcode() != ISD::AND)
7342 SDValue N1 = N->getOperand(1);
7343 if (N1.getOpcode() != ISD::AND)
7346 // We only have to look for constant vectors here since the general, variable
7347 // case can be handled in TableGen.
7348 unsigned Bits = VT.getVectorElementType().getSizeInBits();
7349 uint64_t BitMask = Bits == 64 ? -1ULL : ((1ULL << Bits) - 1);
7350 for (int i = 1; i >= 0; --i)
7351 for (int j = 1; j >= 0; --j) {
7352 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(i));
7353 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(j));
7357 bool FoundMatch = true;
7358 for (unsigned k = 0; k < VT.getVectorNumElements(); ++k) {
7359 ConstantSDNode *CN0 = dyn_cast<ConstantSDNode>(BVN0->getOperand(k));
7360 ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(BVN1->getOperand(k));
7362 CN0->getZExtValue() != (BitMask & ~CN1->getZExtValue())) {
7369 return DAG.getNode(AArch64ISD::BSL, DL, VT, SDValue(BVN0, 0),
7370 N0->getOperand(1 - i), N1->getOperand(1 - j));
7376 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
7377 const AArch64Subtarget *Subtarget) {
7378 // Attempt to form an EXTR from (or (shl VAL1, #N), (srl VAL2, #RegWidth-N))
7379 if (!EnableAArch64ExtrGeneration)
7381 SelectionDAG &DAG = DCI.DAG;
7382 EVT VT = N->getValueType(0);
7384 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7387 SDValue Res = tryCombineToEXTR(N, DCI);
7391 Res = tryCombineToBSL(N, DCI);
7398 static SDValue performBitcastCombine(SDNode *N,
7399 TargetLowering::DAGCombinerInfo &DCI,
7400 SelectionDAG &DAG) {
7401 // Wait 'til after everything is legalized to try this. That way we have
7402 // legal vector types and such.
7403 if (DCI.isBeforeLegalizeOps())
7406 // Remove extraneous bitcasts around an extract_subvector.
7408 // (v4i16 (bitconvert
7409 // (extract_subvector (v2i64 (bitconvert (v8i16 ...)), (i64 1)))))
7411 // (extract_subvector ((v8i16 ...), (i64 4)))
7413 // Only interested in 64-bit vectors as the ultimate result.
7414 EVT VT = N->getValueType(0);
7417 if (VT.getSimpleVT().getSizeInBits() != 64)
7419 // Is the operand an extract_subvector starting at the beginning or halfway
7420 // point of the vector? A low half may also come through as an
7421 // EXTRACT_SUBREG, so look for that, too.
7422 SDValue Op0 = N->getOperand(0);
7423 if (Op0->getOpcode() != ISD::EXTRACT_SUBVECTOR &&
7424 !(Op0->isMachineOpcode() &&
7425 Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG))
7427 uint64_t idx = cast<ConstantSDNode>(Op0->getOperand(1))->getZExtValue();
7428 if (Op0->getOpcode() == ISD::EXTRACT_SUBVECTOR) {
7429 if (Op0->getValueType(0).getVectorNumElements() != idx && idx != 0)
7431 } else if (Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG) {
7432 if (idx != AArch64::dsub)
7434 // The dsub reference is equivalent to a lane zero subvector reference.
7437 // Look through the bitcast of the input to the extract.
7438 if (Op0->getOperand(0)->getOpcode() != ISD::BITCAST)
7440 SDValue Source = Op0->getOperand(0)->getOperand(0);
7441 // If the source type has twice the number of elements as our destination
7442 // type, we know this is an extract of the high or low half of the vector.
7443 EVT SVT = Source->getValueType(0);
7444 if (SVT.getVectorNumElements() != VT.getVectorNumElements() * 2)
7447 DEBUG(dbgs() << "aarch64-lower: bitcast extract_subvector simplification\n");
7449 // Create the simplified form to just extract the low or high half of the
7450 // vector directly rather than bothering with the bitcasts.
7452 unsigned NumElements = VT.getVectorNumElements();
7454 SDValue HalfIdx = DAG.getConstant(NumElements, dl, MVT::i64);
7455 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Source, HalfIdx);
7457 SDValue SubReg = DAG.getTargetConstant(AArch64::dsub, dl, MVT::i32);
7458 return SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl, VT,
7464 static SDValue performConcatVectorsCombine(SDNode *N,
7465 TargetLowering::DAGCombinerInfo &DCI,
7466 SelectionDAG &DAG) {
7468 EVT VT = N->getValueType(0);
7469 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
7471 // Optimize concat_vectors of truncated vectors, where the intermediate
7472 // type is illegal, to avoid said illegality, e.g.,
7473 // (v4i16 (concat_vectors (v2i16 (truncate (v2i64))),
7474 // (v2i16 (truncate (v2i64)))))
7476 // (v4i16 (truncate (vector_shuffle (v4i32 (bitcast (v2i64))),
7477 // (v4i32 (bitcast (v2i64))),
7479 // This isn't really target-specific, but ISD::TRUNCATE legality isn't keyed
7480 // on both input and result type, so we might generate worse code.
7481 // On AArch64 we know it's fine for v2i64->v4i16 and v4i32->v8i8.
7482 if (N->getNumOperands() == 2 &&
7483 N0->getOpcode() == ISD::TRUNCATE &&
7484 N1->getOpcode() == ISD::TRUNCATE) {
7485 SDValue N00 = N0->getOperand(0);
7486 SDValue N10 = N1->getOperand(0);
7487 EVT N00VT = N00.getValueType();
7489 if (N00VT == N10.getValueType() &&
7490 (N00VT == MVT::v2i64 || N00VT == MVT::v4i32) &&
7491 N00VT.getScalarSizeInBits() == 4 * VT.getScalarSizeInBits()) {
7492 MVT MidVT = (N00VT == MVT::v2i64 ? MVT::v4i32 : MVT::v8i16);
7493 SmallVector<int, 8> Mask(MidVT.getVectorNumElements());
7494 for (size_t i = 0; i < Mask.size(); ++i)
7496 return DAG.getNode(ISD::TRUNCATE, dl, VT,
7497 DAG.getVectorShuffle(
7499 DAG.getNode(ISD::BITCAST, dl, MidVT, N00),
7500 DAG.getNode(ISD::BITCAST, dl, MidVT, N10), Mask));
7504 // Wait 'til after everything is legalized to try this. That way we have
7505 // legal vector types and such.
7506 if (DCI.isBeforeLegalizeOps())
7509 // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector
7510 // splat. The indexed instructions are going to be expecting a DUPLANE64, so
7511 // canonicalise to that.
7512 if (N0 == N1 && VT.getVectorNumElements() == 2) {
7513 assert(VT.getVectorElementType().getSizeInBits() == 64);
7514 return DAG.getNode(AArch64ISD::DUPLANE64, dl, VT, WidenVector(N0, DAG),
7515 DAG.getConstant(0, dl, MVT::i64));
7518 // Canonicalise concat_vectors so that the right-hand vector has as few
7519 // bit-casts as possible before its real operation. The primary matching
7520 // destination for these operations will be the narrowing "2" instructions,
7521 // which depend on the operation being performed on this right-hand vector.
7523 // (concat_vectors LHS, (v1i64 (bitconvert (v4i16 RHS))))
7525 // (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS))
7527 if (N1->getOpcode() != ISD::BITCAST)
7529 SDValue RHS = N1->getOperand(0);
7530 MVT RHSTy = RHS.getValueType().getSimpleVT();
7531 // If the RHS is not a vector, this is not the pattern we're looking for.
7532 if (!RHSTy.isVector())
7535 DEBUG(dbgs() << "aarch64-lower: concat_vectors bitcast simplification\n");
7537 MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(),
7538 RHSTy.getVectorNumElements() * 2);
7539 return DAG.getNode(ISD::BITCAST, dl, VT,
7540 DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatTy,
7541 DAG.getNode(ISD::BITCAST, dl, RHSTy, N0),
7545 static SDValue tryCombineFixedPointConvert(SDNode *N,
7546 TargetLowering::DAGCombinerInfo &DCI,
7547 SelectionDAG &DAG) {
7548 // Wait 'til after everything is legalized to try this. That way we have
7549 // legal vector types and such.
7550 if (DCI.isBeforeLegalizeOps())
7552 // Transform a scalar conversion of a value from a lane extract into a
7553 // lane extract of a vector conversion. E.g., from foo1 to foo2:
7554 // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); }
7555 // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; }
7557 // The second form interacts better with instruction selection and the
7558 // register allocator to avoid cross-class register copies that aren't
7559 // coalescable due to a lane reference.
7561 // Check the operand and see if it originates from a lane extract.
7562 SDValue Op1 = N->getOperand(1);
7563 if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7564 // Yep, no additional predication needed. Perform the transform.
7565 SDValue IID = N->getOperand(0);
7566 SDValue Shift = N->getOperand(2);
7567 SDValue Vec = Op1.getOperand(0);
7568 SDValue Lane = Op1.getOperand(1);
7569 EVT ResTy = N->getValueType(0);
7573 // The vector width should be 128 bits by the time we get here, even
7574 // if it started as 64 bits (the extract_vector handling will have
7576 assert(Vec.getValueType().getSizeInBits() == 128 &&
7577 "unexpected vector size on extract_vector_elt!");
7578 if (Vec.getValueType() == MVT::v4i32)
7579 VecResTy = MVT::v4f32;
7580 else if (Vec.getValueType() == MVT::v2i64)
7581 VecResTy = MVT::v2f64;
7583 llvm_unreachable("unexpected vector type!");
7586 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift);
7587 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane);
7592 // AArch64 high-vector "long" operations are formed by performing the non-high
7593 // version on an extract_subvector of each operand which gets the high half:
7595 // (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS))
7597 // However, there are cases which don't have an extract_high explicitly, but
7598 // have another operation that can be made compatible with one for free. For
7601 // (dupv64 scalar) --> (extract_high (dup128 scalar))
7603 // This routine does the actual conversion of such DUPs, once outer routines
7604 // have determined that everything else is in order.
7605 // It also supports immediate DUP-like nodes (MOVI/MVNi), which we can fold
7607 static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) {
7608 switch (N.getOpcode()) {
7609 case AArch64ISD::DUP:
7610 case AArch64ISD::DUPLANE8:
7611 case AArch64ISD::DUPLANE16:
7612 case AArch64ISD::DUPLANE32:
7613 case AArch64ISD::DUPLANE64:
7614 case AArch64ISD::MOVI:
7615 case AArch64ISD::MOVIshift:
7616 case AArch64ISD::MOVIedit:
7617 case AArch64ISD::MOVImsl:
7618 case AArch64ISD::MVNIshift:
7619 case AArch64ISD::MVNImsl:
7622 // FMOV could be supported, but isn't very useful, as it would only occur
7623 // if you passed a bitcast' floating point immediate to an eligible long
7624 // integer op (addl, smull, ...).
7628 MVT NarrowTy = N.getSimpleValueType();
7629 if (!NarrowTy.is64BitVector())
7632 MVT ElementTy = NarrowTy.getVectorElementType();
7633 unsigned NumElems = NarrowTy.getVectorNumElements();
7634 MVT NewVT = MVT::getVectorVT(ElementTy, NumElems * 2);
7637 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NarrowTy,
7638 DAG.getNode(N->getOpcode(), dl, NewVT, N->ops()),
7639 DAG.getConstant(NumElems, dl, MVT::i64));
7642 static bool isEssentiallyExtractSubvector(SDValue N) {
7643 if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR)
7646 return N.getOpcode() == ISD::BITCAST &&
7647 N.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR;
7650 /// \brief Helper structure to keep track of ISD::SET_CC operands.
7651 struct GenericSetCCInfo {
7652 const SDValue *Opnd0;
7653 const SDValue *Opnd1;
7657 /// \brief Helper structure to keep track of a SET_CC lowered into AArch64 code.
7658 struct AArch64SetCCInfo {
7660 AArch64CC::CondCode CC;
7663 /// \brief Helper structure to keep track of SetCC information.
7665 GenericSetCCInfo Generic;
7666 AArch64SetCCInfo AArch64;
7669 /// \brief Helper structure to be able to read SetCC information. If set to
7670 /// true, IsAArch64 field, Info is a AArch64SetCCInfo, otherwise Info is a
7671 /// GenericSetCCInfo.
7672 struct SetCCInfoAndKind {
7677 /// \brief Check whether or not \p Op is a SET_CC operation, either a generic or
7679 /// AArch64 lowered one.
7680 /// \p SetCCInfo is filled accordingly.
7681 /// \post SetCCInfo is meanginfull only when this function returns true.
7682 /// \return True when Op is a kind of SET_CC operation.
7683 static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) {
7684 // If this is a setcc, this is straight forward.
7685 if (Op.getOpcode() == ISD::SETCC) {
7686 SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0);
7687 SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1);
7688 SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
7689 SetCCInfo.IsAArch64 = false;
7692 // Otherwise, check if this is a matching csel instruction.
7696 if (Op.getOpcode() != AArch64ISD::CSEL)
7698 // Set the information about the operands.
7699 // TODO: we want the operands of the Cmp not the csel
7700 SetCCInfo.Info.AArch64.Cmp = &Op.getOperand(3);
7701 SetCCInfo.IsAArch64 = true;
7702 SetCCInfo.Info.AArch64.CC = static_cast<AArch64CC::CondCode>(
7703 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
7705 // Check that the operands matches the constraints:
7706 // (1) Both operands must be constants.
7707 // (2) One must be 1 and the other must be 0.
7708 ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0));
7709 ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7712 if (!TValue || !FValue)
7716 if (!TValue->isOne()) {
7717 // Update the comparison when we are interested in !cc.
7718 std::swap(TValue, FValue);
7719 SetCCInfo.Info.AArch64.CC =
7720 AArch64CC::getInvertedCondCode(SetCCInfo.Info.AArch64.CC);
7722 return TValue->isOne() && FValue->isNullValue();
7725 // Returns true if Op is setcc or zext of setcc.
7726 static bool isSetCCOrZExtSetCC(const SDValue& Op, SetCCInfoAndKind &Info) {
7727 if (isSetCC(Op, Info))
7729 return ((Op.getOpcode() == ISD::ZERO_EXTEND) &&
7730 isSetCC(Op->getOperand(0), Info));
7733 // The folding we want to perform is:
7734 // (add x, [zext] (setcc cc ...) )
7736 // (csel x, (add x, 1), !cc ...)
7738 // The latter will get matched to a CSINC instruction.
7739 static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) {
7740 assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!");
7741 SDValue LHS = Op->getOperand(0);
7742 SDValue RHS = Op->getOperand(1);
7743 SetCCInfoAndKind InfoAndKind;
7745 // If neither operand is a SET_CC, give up.
7746 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind)) {
7747 std::swap(LHS, RHS);
7748 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind))
7752 // FIXME: This could be generatized to work for FP comparisons.
7753 EVT CmpVT = InfoAndKind.IsAArch64
7754 ? InfoAndKind.Info.AArch64.Cmp->getOperand(0).getValueType()
7755 : InfoAndKind.Info.Generic.Opnd0->getValueType();
7756 if (CmpVT != MVT::i32 && CmpVT != MVT::i64)
7762 if (InfoAndKind.IsAArch64) {
7763 CCVal = DAG.getConstant(
7764 AArch64CC::getInvertedCondCode(InfoAndKind.Info.AArch64.CC), dl,
7766 Cmp = *InfoAndKind.Info.AArch64.Cmp;
7768 Cmp = getAArch64Cmp(*InfoAndKind.Info.Generic.Opnd0,
7769 *InfoAndKind.Info.Generic.Opnd1,
7770 ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, true),
7773 EVT VT = Op->getValueType(0);
7774 LHS = DAG.getNode(ISD::ADD, dl, VT, RHS, DAG.getConstant(1, dl, VT));
7775 return DAG.getNode(AArch64ISD::CSEL, dl, VT, RHS, LHS, CCVal, Cmp);
7778 // The basic add/sub long vector instructions have variants with "2" on the end
7779 // which act on the high-half of their inputs. They are normally matched by
7782 // (add (zeroext (extract_high LHS)),
7783 // (zeroext (extract_high RHS)))
7784 // -> uaddl2 vD, vN, vM
7786 // However, if one of the extracts is something like a duplicate, this
7787 // instruction can still be used profitably. This function puts the DAG into a
7788 // more appropriate form for those patterns to trigger.
7789 static SDValue performAddSubLongCombine(SDNode *N,
7790 TargetLowering::DAGCombinerInfo &DCI,
7791 SelectionDAG &DAG) {
7792 if (DCI.isBeforeLegalizeOps())
7795 MVT VT = N->getSimpleValueType(0);
7796 if (!VT.is128BitVector()) {
7797 if (N->getOpcode() == ISD::ADD)
7798 return performSetccAddFolding(N, DAG);
7802 // Make sure both branches are extended in the same way.
7803 SDValue LHS = N->getOperand(0);
7804 SDValue RHS = N->getOperand(1);
7805 if ((LHS.getOpcode() != ISD::ZERO_EXTEND &&
7806 LHS.getOpcode() != ISD::SIGN_EXTEND) ||
7807 LHS.getOpcode() != RHS.getOpcode())
7810 unsigned ExtType = LHS.getOpcode();
7812 // It's not worth doing if at least one of the inputs isn't already an
7813 // extract, but we don't know which it'll be so we have to try both.
7814 if (isEssentiallyExtractSubvector(LHS.getOperand(0))) {
7815 RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG);
7819 RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS);
7820 } else if (isEssentiallyExtractSubvector(RHS.getOperand(0))) {
7821 LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG);
7825 LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS);
7828 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS);
7831 // Massage DAGs which we can use the high-half "long" operations on into
7832 // something isel will recognize better. E.g.
7834 // (aarch64_neon_umull (extract_high vec) (dupv64 scalar)) -->
7835 // (aarch64_neon_umull (extract_high (v2i64 vec)))
7836 // (extract_high (v2i64 (dup128 scalar)))))
7838 static SDValue tryCombineLongOpWithDup(unsigned IID, SDNode *N,
7839 TargetLowering::DAGCombinerInfo &DCI,
7840 SelectionDAG &DAG) {
7841 if (DCI.isBeforeLegalizeOps())
7844 SDValue LHS = N->getOperand(1);
7845 SDValue RHS = N->getOperand(2);
7846 assert(LHS.getValueType().is64BitVector() &&
7847 RHS.getValueType().is64BitVector() &&
7848 "unexpected shape for long operation");
7850 // Either node could be a DUP, but it's not worth doing both of them (you'd
7851 // just as well use the non-high version) so look for a corresponding extract
7852 // operation on the other "wing".
7853 if (isEssentiallyExtractSubvector(LHS)) {
7854 RHS = tryExtendDUPToExtractHigh(RHS, DAG);
7857 } else if (isEssentiallyExtractSubvector(RHS)) {
7858 LHS = tryExtendDUPToExtractHigh(LHS, DAG);
7863 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0),
7864 N->getOperand(0), LHS, RHS);
7867 static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
7868 MVT ElemTy = N->getSimpleValueType(0).getScalarType();
7869 unsigned ElemBits = ElemTy.getSizeInBits();
7871 int64_t ShiftAmount;
7872 if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) {
7873 APInt SplatValue, SplatUndef;
7874 unsigned SplatBitSize;
7876 if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
7877 HasAnyUndefs, ElemBits) ||
7878 SplatBitSize != ElemBits)
7881 ShiftAmount = SplatValue.getSExtValue();
7882 } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) {
7883 ShiftAmount = CVN->getSExtValue();
7891 llvm_unreachable("Unknown shift intrinsic");
7892 case Intrinsic::aarch64_neon_sqshl:
7893 Opcode = AArch64ISD::SQSHL_I;
7894 IsRightShift = false;
7896 case Intrinsic::aarch64_neon_uqshl:
7897 Opcode = AArch64ISD::UQSHL_I;
7898 IsRightShift = false;
7900 case Intrinsic::aarch64_neon_srshl:
7901 Opcode = AArch64ISD::SRSHR_I;
7902 IsRightShift = true;
7904 case Intrinsic::aarch64_neon_urshl:
7905 Opcode = AArch64ISD::URSHR_I;
7906 IsRightShift = true;
7908 case Intrinsic::aarch64_neon_sqshlu:
7909 Opcode = AArch64ISD::SQSHLU_I;
7910 IsRightShift = false;
7914 if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits) {
7916 return DAG.getNode(Opcode, dl, N->getValueType(0), N->getOperand(1),
7917 DAG.getConstant(-ShiftAmount, dl, MVT::i32));
7918 } else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount < ElemBits) {
7920 return DAG.getNode(Opcode, dl, N->getValueType(0), N->getOperand(1),
7921 DAG.getConstant(ShiftAmount, dl, MVT::i32));
7927 // The CRC32[BH] instructions ignore the high bits of their data operand. Since
7928 // the intrinsics must be legal and take an i32, this means there's almost
7929 // certainly going to be a zext in the DAG which we can eliminate.
7930 static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) {
7931 SDValue AndN = N->getOperand(2);
7932 if (AndN.getOpcode() != ISD::AND)
7935 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1));
7936 if (!CMask || CMask->getZExtValue() != Mask)
7939 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32,
7940 N->getOperand(0), N->getOperand(1), AndN.getOperand(0));
7943 static SDValue combineAcrossLanesIntrinsic(unsigned Opc, SDNode *N,
7944 SelectionDAG &DAG) {
7946 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0),
7947 DAG.getNode(Opc, dl,
7948 N->getOperand(1).getSimpleValueType(),
7950 DAG.getConstant(0, dl, MVT::i64));
7953 static SDValue performIntrinsicCombine(SDNode *N,
7954 TargetLowering::DAGCombinerInfo &DCI,
7955 const AArch64Subtarget *Subtarget) {
7956 SelectionDAG &DAG = DCI.DAG;
7957 unsigned IID = getIntrinsicID(N);
7961 case Intrinsic::aarch64_neon_vcvtfxs2fp:
7962 case Intrinsic::aarch64_neon_vcvtfxu2fp:
7963 return tryCombineFixedPointConvert(N, DCI, DAG);
7965 case Intrinsic::aarch64_neon_saddv:
7966 return combineAcrossLanesIntrinsic(AArch64ISD::SADDV, N, DAG);
7967 case Intrinsic::aarch64_neon_uaddv:
7968 return combineAcrossLanesIntrinsic(AArch64ISD::UADDV, N, DAG);
7969 case Intrinsic::aarch64_neon_sminv:
7970 return combineAcrossLanesIntrinsic(AArch64ISD::SMINV, N, DAG);
7971 case Intrinsic::aarch64_neon_uminv:
7972 return combineAcrossLanesIntrinsic(AArch64ISD::UMINV, N, DAG);
7973 case Intrinsic::aarch64_neon_smaxv:
7974 return combineAcrossLanesIntrinsic(AArch64ISD::SMAXV, N, DAG);
7975 case Intrinsic::aarch64_neon_umaxv:
7976 return combineAcrossLanesIntrinsic(AArch64ISD::UMAXV, N, DAG);
7977 case Intrinsic::aarch64_neon_fmax:
7978 return DAG.getNode(AArch64ISD::FMAX, SDLoc(N), N->getValueType(0),
7979 N->getOperand(1), N->getOperand(2));
7980 case Intrinsic::aarch64_neon_fmin:
7981 return DAG.getNode(AArch64ISD::FMIN, SDLoc(N), N->getValueType(0),
7982 N->getOperand(1), N->getOperand(2));
7983 case Intrinsic::aarch64_neon_smull:
7984 case Intrinsic::aarch64_neon_umull:
7985 case Intrinsic::aarch64_neon_pmull:
7986 case Intrinsic::aarch64_neon_sqdmull:
7987 return tryCombineLongOpWithDup(IID, N, DCI, DAG);
7988 case Intrinsic::aarch64_neon_sqshl:
7989 case Intrinsic::aarch64_neon_uqshl:
7990 case Intrinsic::aarch64_neon_sqshlu:
7991 case Intrinsic::aarch64_neon_srshl:
7992 case Intrinsic::aarch64_neon_urshl:
7993 return tryCombineShiftImm(IID, N, DAG);
7994 case Intrinsic::aarch64_crc32b:
7995 case Intrinsic::aarch64_crc32cb:
7996 return tryCombineCRC32(0xff, N, DAG);
7997 case Intrinsic::aarch64_crc32h:
7998 case Intrinsic::aarch64_crc32ch:
7999 return tryCombineCRC32(0xffff, N, DAG);
8004 static SDValue performExtendCombine(SDNode *N,
8005 TargetLowering::DAGCombinerInfo &DCI,
8006 SelectionDAG &DAG) {
8007 // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then
8008 // we can convert that DUP into another extract_high (of a bigger DUP), which
8009 // helps the backend to decide that an sabdl2 would be useful, saving a real
8010 // extract_high operation.
8011 if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND &&
8012 N->getOperand(0).getOpcode() == ISD::INTRINSIC_WO_CHAIN) {
8013 SDNode *ABDNode = N->getOperand(0).getNode();
8014 unsigned IID = getIntrinsicID(ABDNode);
8015 if (IID == Intrinsic::aarch64_neon_sabd ||
8016 IID == Intrinsic::aarch64_neon_uabd) {
8017 SDValue NewABD = tryCombineLongOpWithDup(IID, ABDNode, DCI, DAG);
8018 if (!NewABD.getNode())
8021 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0),
8026 // This is effectively a custom type legalization for AArch64.
8028 // Type legalization will split an extend of a small, legal, type to a larger
8029 // illegal type by first splitting the destination type, often creating
8030 // illegal source types, which then get legalized in isel-confusing ways,
8031 // leading to really terrible codegen. E.g.,
8032 // %result = v8i32 sext v8i8 %value
8034 // %losrc = extract_subreg %value, ...
8035 // %hisrc = extract_subreg %value, ...
8036 // %lo = v4i32 sext v4i8 %losrc
8037 // %hi = v4i32 sext v4i8 %hisrc
8038 // Things go rapidly downhill from there.
8040 // For AArch64, the [sz]ext vector instructions can only go up one element
8041 // size, so we can, e.g., extend from i8 to i16, but to go from i8 to i32
8042 // take two instructions.
8044 // This implies that the most efficient way to do the extend from v8i8
8045 // to two v4i32 values is to first extend the v8i8 to v8i16, then do
8046 // the normal splitting to happen for the v8i16->v8i32.
8048 // This is pre-legalization to catch some cases where the default
8049 // type legalization will create ill-tempered code.
8050 if (!DCI.isBeforeLegalizeOps())
8053 // We're only interested in cleaning things up for non-legal vector types
8054 // here. If both the source and destination are legal, things will just
8055 // work naturally without any fiddling.
8056 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8057 EVT ResVT = N->getValueType(0);
8058 if (!ResVT.isVector() || TLI.isTypeLegal(ResVT))
8060 // If the vector type isn't a simple VT, it's beyond the scope of what
8061 // we're worried about here. Let legalization do its thing and hope for
8063 SDValue Src = N->getOperand(0);
8064 EVT SrcVT = Src->getValueType(0);
8065 if (!ResVT.isSimple() || !SrcVT.isSimple())
8068 // If the source VT is a 64-bit vector, we can play games and get the
8069 // better results we want.
8070 if (SrcVT.getSizeInBits() != 64)
8073 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
8074 unsigned ElementCount = SrcVT.getVectorNumElements();
8075 SrcVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize * 2), ElementCount);
8077 Src = DAG.getNode(N->getOpcode(), DL, SrcVT, Src);
8079 // Now split the rest of the operation into two halves, each with a 64
8083 unsigned NumElements = ResVT.getVectorNumElements();
8084 assert(!(NumElements & 1) && "Splitting vector, but not in half!");
8085 LoVT = HiVT = EVT::getVectorVT(*DAG.getContext(),
8086 ResVT.getVectorElementType(), NumElements / 2);
8088 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
8089 LoVT.getVectorNumElements());
8090 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
8091 DAG.getConstant(0, DL, MVT::i64));
8092 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
8093 DAG.getConstant(InNVT.getVectorNumElements(), DL, MVT::i64));
8094 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, Lo);
8095 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, Hi);
8097 // Now combine the parts back together so we still have a single result
8098 // like the combiner expects.
8099 return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
8102 /// Replace a splat of a scalar to a vector store by scalar stores of the scalar
8103 /// value. The load store optimizer pass will merge them to store pair stores.
8104 /// This has better performance than a splat of the scalar followed by a split
8105 /// vector store. Even if the stores are not merged it is four stores vs a dup,
8106 /// followed by an ext.b and two stores.
8107 static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) {
8108 SDValue StVal = St->getValue();
8109 EVT VT = StVal.getValueType();
8111 // Don't replace floating point stores, they possibly won't be transformed to
8112 // stp because of the store pair suppress pass.
8113 if (VT.isFloatingPoint())
8116 // Check for insert vector elements.
8117 if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT)
8120 // We can express a splat as store pair(s) for 2 or 4 elements.
8121 unsigned NumVecElts = VT.getVectorNumElements();
8122 if (NumVecElts != 4 && NumVecElts != 2)
8124 SDValue SplatVal = StVal.getOperand(1);
8125 unsigned RemainInsertElts = NumVecElts - 1;
8127 // Check that this is a splat.
8128 while (--RemainInsertElts) {
8129 SDValue NextInsertElt = StVal.getOperand(0);
8130 if (NextInsertElt.getOpcode() != ISD::INSERT_VECTOR_ELT)
8132 if (NextInsertElt.getOperand(1) != SplatVal)
8134 StVal = NextInsertElt;
8136 unsigned OrigAlignment = St->getAlignment();
8137 unsigned EltOffset = NumVecElts == 4 ? 4 : 8;
8138 unsigned Alignment = std::min(OrigAlignment, EltOffset);
8140 // Create scalar stores. This is at least as good as the code sequence for a
8141 // split unaligned store wich is a dup.s, ext.b, and two stores.
8142 // Most of the time the three stores should be replaced by store pair
8143 // instructions (stp).
8145 SDValue BasePtr = St->getBasePtr();
8147 DAG.getStore(St->getChain(), DL, SplatVal, BasePtr, St->getPointerInfo(),
8148 St->isVolatile(), St->isNonTemporal(), St->getAlignment());
8150 unsigned Offset = EltOffset;
8151 while (--NumVecElts) {
8152 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
8153 DAG.getConstant(Offset, DL, MVT::i64));
8154 NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr,
8155 St->getPointerInfo(), St->isVolatile(),
8156 St->isNonTemporal(), Alignment);
8157 Offset += EltOffset;
8162 static SDValue performSTORECombine(SDNode *N,
8163 TargetLowering::DAGCombinerInfo &DCI,
8165 const AArch64Subtarget *Subtarget) {
8166 if (!DCI.isBeforeLegalize())
8169 StoreSDNode *S = cast<StoreSDNode>(N);
8170 if (S->isVolatile())
8173 // Cyclone has bad performance on unaligned 16B stores when crossing line and
8174 // page boundaries. We want to split such stores.
8175 if (!Subtarget->isCyclone())
8178 // Don't split at Oz.
8179 MachineFunction &MF = DAG.getMachineFunction();
8180 bool IsMinSize = MF.getFunction()->hasFnAttribute(Attribute::MinSize);
8184 SDValue StVal = S->getValue();
8185 EVT VT = StVal.getValueType();
8187 // Don't split v2i64 vectors. Memcpy lowering produces those and splitting
8188 // those up regresses performance on micro-benchmarks and olden/bh.
8189 if (!VT.isVector() || VT.getVectorNumElements() < 2 || VT == MVT::v2i64)
8192 // Split unaligned 16B stores. They are terrible for performance.
8193 // Don't split stores with alignment of 1 or 2. Code that uses clang vector
8194 // extensions can use this to mark that it does not want splitting to happen
8195 // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of
8196 // eliminating alignment hazards is only 1 in 8 for alignment of 2.
8197 if (VT.getSizeInBits() != 128 || S->getAlignment() >= 16 ||
8198 S->getAlignment() <= 2)
8201 // If we get a splat of a scalar convert this vector store to a store of
8202 // scalars. They will be merged into store pairs thereby removing two
8204 SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S);
8205 if (ReplacedSplat != SDValue())
8206 return ReplacedSplat;
8209 unsigned NumElts = VT.getVectorNumElements() / 2;
8210 // Split VT into two.
8212 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), NumElts);
8213 SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
8214 DAG.getConstant(0, DL, MVT::i64));
8215 SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
8216 DAG.getConstant(NumElts, DL, MVT::i64));
8217 SDValue BasePtr = S->getBasePtr();
8219 DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
8220 S->isVolatile(), S->isNonTemporal(), S->getAlignment());
8221 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
8222 DAG.getConstant(8, DL, MVT::i64));
8223 return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr,
8224 S->getPointerInfo(), S->isVolatile(), S->isNonTemporal(),
8228 /// Target-specific DAG combine function for post-increment LD1 (lane) and
8229 /// post-increment LD1R.
8230 static SDValue performPostLD1Combine(SDNode *N,
8231 TargetLowering::DAGCombinerInfo &DCI,
8233 if (DCI.isBeforeLegalizeOps())
8236 SelectionDAG &DAG = DCI.DAG;
8237 EVT VT = N->getValueType(0);
8239 unsigned LoadIdx = IsLaneOp ? 1 : 0;
8240 SDNode *LD = N->getOperand(LoadIdx).getNode();
8241 // If it is not LOAD, can not do such combine.
8242 if (LD->getOpcode() != ISD::LOAD)
8245 LoadSDNode *LoadSDN = cast<LoadSDNode>(LD);
8246 EVT MemVT = LoadSDN->getMemoryVT();
8247 // Check if memory operand is the same type as the vector element.
8248 if (MemVT != VT.getVectorElementType())
8251 // Check if there are other uses. If so, do not combine as it will introduce
8253 for (SDNode::use_iterator UI = LD->use_begin(), UE = LD->use_end(); UI != UE;
8255 if (UI.getUse().getResNo() == 1) // Ignore uses of the chain result.
8261 SDValue Addr = LD->getOperand(1);
8262 SDValue Vector = N->getOperand(0);
8263 // Search for a use of the address operand that is an increment.
8264 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(), UE =
8265 Addr.getNode()->use_end(); UI != UE; ++UI) {
8267 if (User->getOpcode() != ISD::ADD
8268 || UI.getUse().getResNo() != Addr.getResNo())
8271 // Check that the add is independent of the load. Otherwise, folding it
8272 // would create a cycle.
8273 if (User->isPredecessorOf(LD) || LD->isPredecessorOf(User))
8275 // Also check that add is not used in the vector operand. This would also
8277 if (User->isPredecessorOf(Vector.getNode()))
8280 // If the increment is a constant, it must match the memory ref size.
8281 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8282 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8283 uint32_t IncVal = CInc->getZExtValue();
8284 unsigned NumBytes = VT.getScalarSizeInBits() / 8;
8285 if (IncVal != NumBytes)
8287 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
8290 // Finally, check that the vector doesn't depend on the load.
8291 // Again, this would create a cycle.
8292 // The load depending on the vector is fine, as that's the case for the
8293 // LD1*post we'll eventually generate anyway.
8294 if (LoadSDN->isPredecessorOf(Vector.getNode()))
8297 SmallVector<SDValue, 8> Ops;
8298 Ops.push_back(LD->getOperand(0)); // Chain
8300 Ops.push_back(Vector); // The vector to be inserted
8301 Ops.push_back(N->getOperand(2)); // The lane to be inserted in the vector
8303 Ops.push_back(Addr);
8306 EVT Tys[3] = { VT, MVT::i64, MVT::Other };
8307 SDVTList SDTys = DAG.getVTList(Tys);
8308 unsigned NewOp = IsLaneOp ? AArch64ISD::LD1LANEpost : AArch64ISD::LD1DUPpost;
8309 SDValue UpdN = DAG.getMemIntrinsicNode(NewOp, SDLoc(N), SDTys, Ops,
8311 LoadSDN->getMemOperand());
8314 SmallVector<SDValue, 2> NewResults;
8315 NewResults.push_back(SDValue(LD, 0)); // The result of load
8316 NewResults.push_back(SDValue(UpdN.getNode(), 2)); // Chain
8317 DCI.CombineTo(LD, NewResults);
8318 DCI.CombineTo(N, SDValue(UpdN.getNode(), 0)); // Dup/Inserted Result
8319 DCI.CombineTo(User, SDValue(UpdN.getNode(), 1)); // Write back register
8326 /// Target-specific DAG combine function for NEON load/store intrinsics
8327 /// to merge base address updates.
8328 static SDValue performNEONPostLDSTCombine(SDNode *N,
8329 TargetLowering::DAGCombinerInfo &DCI,
8330 SelectionDAG &DAG) {
8331 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8334 unsigned AddrOpIdx = N->getNumOperands() - 1;
8335 SDValue Addr = N->getOperand(AddrOpIdx);
8337 // Search for a use of the address operand that is an increment.
8338 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8339 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8341 if (User->getOpcode() != ISD::ADD ||
8342 UI.getUse().getResNo() != Addr.getResNo())
8345 // Check that the add is independent of the load/store. Otherwise, folding
8346 // it would create a cycle.
8347 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8350 // Find the new opcode for the updating load/store.
8351 bool IsStore = false;
8352 bool IsLaneOp = false;
8353 bool IsDupOp = false;
8354 unsigned NewOpc = 0;
8355 unsigned NumVecs = 0;
8356 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8358 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8359 case Intrinsic::aarch64_neon_ld2: NewOpc = AArch64ISD::LD2post;
8361 case Intrinsic::aarch64_neon_ld3: NewOpc = AArch64ISD::LD3post;
8363 case Intrinsic::aarch64_neon_ld4: NewOpc = AArch64ISD::LD4post;
8365 case Intrinsic::aarch64_neon_st2: NewOpc = AArch64ISD::ST2post;
8366 NumVecs = 2; IsStore = true; break;
8367 case Intrinsic::aarch64_neon_st3: NewOpc = AArch64ISD::ST3post;
8368 NumVecs = 3; IsStore = true; break;
8369 case Intrinsic::aarch64_neon_st4: NewOpc = AArch64ISD::ST4post;
8370 NumVecs = 4; IsStore = true; break;
8371 case Intrinsic::aarch64_neon_ld1x2: NewOpc = AArch64ISD::LD1x2post;
8373 case Intrinsic::aarch64_neon_ld1x3: NewOpc = AArch64ISD::LD1x3post;
8375 case Intrinsic::aarch64_neon_ld1x4: NewOpc = AArch64ISD::LD1x4post;
8377 case Intrinsic::aarch64_neon_st1x2: NewOpc = AArch64ISD::ST1x2post;
8378 NumVecs = 2; IsStore = true; break;
8379 case Intrinsic::aarch64_neon_st1x3: NewOpc = AArch64ISD::ST1x3post;
8380 NumVecs = 3; IsStore = true; break;
8381 case Intrinsic::aarch64_neon_st1x4: NewOpc = AArch64ISD::ST1x4post;
8382 NumVecs = 4; IsStore = true; break;
8383 case Intrinsic::aarch64_neon_ld2r: NewOpc = AArch64ISD::LD2DUPpost;
8384 NumVecs = 2; IsDupOp = true; break;
8385 case Intrinsic::aarch64_neon_ld3r: NewOpc = AArch64ISD::LD3DUPpost;
8386 NumVecs = 3; IsDupOp = true; break;
8387 case Intrinsic::aarch64_neon_ld4r: NewOpc = AArch64ISD::LD4DUPpost;
8388 NumVecs = 4; IsDupOp = true; break;
8389 case Intrinsic::aarch64_neon_ld2lane: NewOpc = AArch64ISD::LD2LANEpost;
8390 NumVecs = 2; IsLaneOp = true; break;
8391 case Intrinsic::aarch64_neon_ld3lane: NewOpc = AArch64ISD::LD3LANEpost;
8392 NumVecs = 3; IsLaneOp = true; break;
8393 case Intrinsic::aarch64_neon_ld4lane: NewOpc = AArch64ISD::LD4LANEpost;
8394 NumVecs = 4; IsLaneOp = true; break;
8395 case Intrinsic::aarch64_neon_st2lane: NewOpc = AArch64ISD::ST2LANEpost;
8396 NumVecs = 2; IsStore = true; IsLaneOp = true; break;
8397 case Intrinsic::aarch64_neon_st3lane: NewOpc = AArch64ISD::ST3LANEpost;
8398 NumVecs = 3; IsStore = true; IsLaneOp = true; break;
8399 case Intrinsic::aarch64_neon_st4lane: NewOpc = AArch64ISD::ST4LANEpost;
8400 NumVecs = 4; IsStore = true; IsLaneOp = true; break;
8405 VecTy = N->getOperand(2).getValueType();
8407 VecTy = N->getValueType(0);
8409 // If the increment is a constant, it must match the memory ref size.
8410 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8411 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8412 uint32_t IncVal = CInc->getZExtValue();
8413 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8414 if (IsLaneOp || IsDupOp)
8415 NumBytes /= VecTy.getVectorNumElements();
8416 if (IncVal != NumBytes)
8418 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
8420 SmallVector<SDValue, 8> Ops;
8421 Ops.push_back(N->getOperand(0)); // Incoming chain
8422 // Load lane and store have vector list as input.
8423 if (IsLaneOp || IsStore)
8424 for (unsigned i = 2; i < AddrOpIdx; ++i)
8425 Ops.push_back(N->getOperand(i));
8426 Ops.push_back(Addr); // Base register
8431 unsigned NumResultVecs = (IsStore ? 0 : NumVecs);
8433 for (n = 0; n < NumResultVecs; ++n)
8435 Tys[n++] = MVT::i64; // Type of write back register
8436 Tys[n] = MVT::Other; // Type of the chain
8437 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs + 2));
8439 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8440 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys, Ops,
8441 MemInt->getMemoryVT(),
8442 MemInt->getMemOperand());
8445 std::vector<SDValue> NewResults;
8446 for (unsigned i = 0; i < NumResultVecs; ++i) {
8447 NewResults.push_back(SDValue(UpdN.getNode(), i));
8449 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs + 1));
8450 DCI.CombineTo(N, NewResults);
8451 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8458 // Checks to see if the value is the prescribed width and returns information
8459 // about its extension mode.
8461 bool checkValueWidth(SDValue V, unsigned width, ISD::LoadExtType &ExtType) {
8462 ExtType = ISD::NON_EXTLOAD;
8463 switch(V.getNode()->getOpcode()) {
8467 LoadSDNode *LoadNode = cast<LoadSDNode>(V.getNode());
8468 if ((LoadNode->getMemoryVT() == MVT::i8 && width == 8)
8469 || (LoadNode->getMemoryVT() == MVT::i16 && width == 16)) {
8470 ExtType = LoadNode->getExtensionType();
8475 case ISD::AssertSext: {
8476 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
8477 if ((TypeNode->getVT() == MVT::i8 && width == 8)
8478 || (TypeNode->getVT() == MVT::i16 && width == 16)) {
8479 ExtType = ISD::SEXTLOAD;
8484 case ISD::AssertZext: {
8485 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
8486 if ((TypeNode->getVT() == MVT::i8 && width == 8)
8487 || (TypeNode->getVT() == MVT::i16 && width == 16)) {
8488 ExtType = ISD::ZEXTLOAD;
8494 case ISD::TargetConstant: {
8495 if (std::abs(cast<ConstantSDNode>(V.getNode())->getSExtValue()) <
8505 // This function does a whole lot of voodoo to determine if the tests are
8506 // equivalent without and with a mask. Essentially what happens is that given a
8509 // +-------------+ +-------------+ +-------------+ +-------------+
8510 // | Input | | AddConstant | | CompConstant| | CC |
8511 // +-------------+ +-------------+ +-------------+ +-------------+
8513 // V V | +----------+
8514 // +-------------+ +----+ | |
8515 // | ADD | |0xff| | |
8516 // +-------------+ +----+ | |
8519 // +-------------+ | |
8521 // +-------------+ | |
8530 // The AND node may be safely removed for some combinations of inputs. In
8531 // particular we need to take into account the extension type of the Input,
8532 // the exact values of AddConstant, CompConstant, and CC, along with the nominal
8533 // width of the input (this can work for any width inputs, the above graph is
8534 // specific to 8 bits.
8536 // The specific equations were worked out by generating output tables for each
8537 // AArch64CC value in terms of and AddConstant (w1), CompConstant(w2). The
8538 // problem was simplified by working with 4 bit inputs, which means we only
8539 // needed to reason about 24 distinct bit patterns: 8 patterns unique to zero
8540 // extension (8,15), 8 patterns unique to sign extensions (-8,-1), and 8
8541 // patterns present in both extensions (0,7). For every distinct set of
8542 // AddConstant and CompConstants bit patterns we can consider the masked and
8543 // unmasked versions to be equivalent if the result of this function is true for
8544 // all 16 distinct bit patterns of for the current extension type of Input (w0).
8547 // and w10, w8, #0x0f
8549 // cset w9, AArch64CC
8551 // cset w11, AArch64CC
8556 // Since the above function shows when the outputs are equivalent it defines
8557 // when it is safe to remove the AND. Unfortunately it only runs on AArch64 and
8558 // would be expensive to run during compiles. The equations below were written
8559 // in a test harness that confirmed they gave equivalent outputs to the above
8560 // for all inputs function, so they can be used determine if the removal is
8563 // isEquivalentMaskless() is the code for testing if the AND can be removed
8564 // factored out of the DAG recognition as the DAG can take several forms.
8567 bool isEquivalentMaskless(unsigned CC, unsigned width,
8568 ISD::LoadExtType ExtType, signed AddConstant,
8569 signed CompConstant) {
8570 // By being careful about our equations and only writing the in term
8571 // symbolic values and well known constants (0, 1, -1, MaxUInt) we can
8572 // make them generally applicable to all bit widths.
8573 signed MaxUInt = (1 << width);
8575 // For the purposes of these comparisons sign extending the type is
8576 // equivalent to zero extending the add and displacing it by half the integer
8577 // width. Provided we are careful and make sure our equations are valid over
8578 // the whole range we can just adjust the input and avoid writing equations
8579 // for sign extended inputs.
8580 if (ExtType == ISD::SEXTLOAD)
8581 AddConstant -= (1 << (width-1));
8585 case AArch64CC::GT: {
8586 if ((AddConstant == 0) ||
8587 (CompConstant == MaxUInt - 1 && AddConstant < 0) ||
8588 (AddConstant >= 0 && CompConstant < 0) ||
8589 (AddConstant <= 0 && CompConstant <= 0 && CompConstant < AddConstant))
8593 case AArch64CC::GE: {
8594 if ((AddConstant == 0) ||
8595 (AddConstant >= 0 && CompConstant <= 0) ||
8596 (AddConstant <= 0 && CompConstant <= 0 && CompConstant <= AddConstant))
8600 case AArch64CC::LS: {
8601 if ((AddConstant >= 0 && CompConstant < 0) ||
8602 (AddConstant <= 0 && CompConstant >= -1 &&
8603 CompConstant < AddConstant + MaxUInt))
8607 case AArch64CC::MI: {
8608 if ((AddConstant == 0) ||
8609 (AddConstant > 0 && CompConstant <= 0) ||
8610 (AddConstant < 0 && CompConstant <= AddConstant))
8614 case AArch64CC::HS: {
8615 if ((AddConstant >= 0 && CompConstant <= 0) ||
8616 (AddConstant <= 0 && CompConstant >= 0 &&
8617 CompConstant <= AddConstant + MaxUInt))
8621 case AArch64CC::NE: {
8622 if ((AddConstant > 0 && CompConstant < 0) ||
8623 (AddConstant < 0 && CompConstant >= 0 &&
8624 CompConstant < AddConstant + MaxUInt) ||
8625 (AddConstant >= 0 && CompConstant >= 0 &&
8626 CompConstant >= AddConstant) ||
8627 (AddConstant <= 0 && CompConstant < 0 && CompConstant < AddConstant))
8636 case AArch64CC::Invalid:
8644 SDValue performCONDCombine(SDNode *N,
8645 TargetLowering::DAGCombinerInfo &DCI,
8646 SelectionDAG &DAG, unsigned CCIndex,
8647 unsigned CmpIndex) {
8648 unsigned CC = cast<ConstantSDNode>(N->getOperand(CCIndex))->getSExtValue();
8649 SDNode *SubsNode = N->getOperand(CmpIndex).getNode();
8650 unsigned CondOpcode = SubsNode->getOpcode();
8652 if (CondOpcode != AArch64ISD::SUBS)
8655 // There is a SUBS feeding this condition. Is it fed by a mask we can
8658 SDNode *AndNode = SubsNode->getOperand(0).getNode();
8659 unsigned MaskBits = 0;
8661 if (AndNode->getOpcode() != ISD::AND)
8664 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(AndNode->getOperand(1))) {
8665 uint32_t CNV = CN->getZExtValue();
8668 else if (CNV == 65535)
8675 SDValue AddValue = AndNode->getOperand(0);
8677 if (AddValue.getOpcode() != ISD::ADD)
8680 // The basic dag structure is correct, grab the inputs and validate them.
8682 SDValue AddInputValue1 = AddValue.getNode()->getOperand(0);
8683 SDValue AddInputValue2 = AddValue.getNode()->getOperand(1);
8684 SDValue SubsInputValue = SubsNode->getOperand(1);
8686 // The mask is present and the provenance of all the values is a smaller type,
8687 // lets see if the mask is superfluous.
8689 if (!isa<ConstantSDNode>(AddInputValue2.getNode()) ||
8690 !isa<ConstantSDNode>(SubsInputValue.getNode()))
8693 ISD::LoadExtType ExtType;
8695 if (!checkValueWidth(SubsInputValue, MaskBits, ExtType) ||
8696 !checkValueWidth(AddInputValue2, MaskBits, ExtType) ||
8697 !checkValueWidth(AddInputValue1, MaskBits, ExtType) )
8700 if(!isEquivalentMaskless(CC, MaskBits, ExtType,
8701 cast<ConstantSDNode>(AddInputValue2.getNode())->getSExtValue(),
8702 cast<ConstantSDNode>(SubsInputValue.getNode())->getSExtValue()))
8705 // The AND is not necessary, remove it.
8707 SDVTList VTs = DAG.getVTList(SubsNode->getValueType(0),
8708 SubsNode->getValueType(1));
8709 SDValue Ops[] = { AddValue, SubsNode->getOperand(1) };
8711 SDValue NewValue = DAG.getNode(CondOpcode, SDLoc(SubsNode), VTs, Ops);
8712 DAG.ReplaceAllUsesWith(SubsNode, NewValue.getNode());
8714 return SDValue(N, 0);
8717 // Optimize compare with zero and branch.
8718 static SDValue performBRCONDCombine(SDNode *N,
8719 TargetLowering::DAGCombinerInfo &DCI,
8720 SelectionDAG &DAG) {
8721 SDValue NV = performCONDCombine(N, DCI, DAG, 2, 3);
8724 SDValue Chain = N->getOperand(0);
8725 SDValue Dest = N->getOperand(1);
8726 SDValue CCVal = N->getOperand(2);
8727 SDValue Cmp = N->getOperand(3);
8729 assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!");
8730 unsigned CC = cast<ConstantSDNode>(CCVal)->getZExtValue();
8731 if (CC != AArch64CC::EQ && CC != AArch64CC::NE)
8734 unsigned CmpOpc = Cmp.getOpcode();
8735 if (CmpOpc != AArch64ISD::ADDS && CmpOpc != AArch64ISD::SUBS)
8738 // Only attempt folding if there is only one use of the flag and no use of the
8740 if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1))
8743 SDValue LHS = Cmp.getOperand(0);
8744 SDValue RHS = Cmp.getOperand(1);
8746 assert(LHS.getValueType() == RHS.getValueType() &&
8747 "Expected the value type to be the same for both operands!");
8748 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
8751 if (isa<ConstantSDNode>(LHS) && cast<ConstantSDNode>(LHS)->isNullValue())
8752 std::swap(LHS, RHS);
8754 if (!isa<ConstantSDNode>(RHS) || !cast<ConstantSDNode>(RHS)->isNullValue())
8757 if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA ||
8758 LHS.getOpcode() == ISD::SRL)
8761 // Fold the compare into the branch instruction.
8763 if (CC == AArch64CC::EQ)
8764 BR = DAG.getNode(AArch64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
8766 BR = DAG.getNode(AArch64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
8768 // Do not add new nodes to DAG combiner worklist.
8769 DCI.CombineTo(N, BR, false);
8774 // vselect (v1i1 setcc) ->
8775 // vselect (v1iXX setcc) (XX is the size of the compared operand type)
8776 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
8777 // condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
8779 static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) {
8780 SDValue N0 = N->getOperand(0);
8781 EVT CCVT = N0.getValueType();
8783 if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
8784 CCVT.getVectorElementType() != MVT::i1)
8787 EVT ResVT = N->getValueType(0);
8788 EVT CmpVT = N0.getOperand(0).getValueType();
8789 // Only combine when the result type is of the same size as the compared
8791 if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
8794 SDValue IfTrue = N->getOperand(1);
8795 SDValue IfFalse = N->getOperand(2);
8797 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
8798 N0.getOperand(0), N0.getOperand(1),
8799 cast<CondCodeSDNode>(N0.getOperand(2))->get());
8800 return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
8804 /// A vector select: "(select vL, vR, (setcc LHS, RHS))" is best performed with
8805 /// the compare-mask instructions rather than going via NZCV, even if LHS and
8806 /// RHS are really scalar. This replaces any scalar setcc in the above pattern
8807 /// with a vector one followed by a DUP shuffle on the result.
8808 static SDValue performSelectCombine(SDNode *N,
8809 TargetLowering::DAGCombinerInfo &DCI) {
8810 SelectionDAG &DAG = DCI.DAG;
8811 SDValue N0 = N->getOperand(0);
8812 EVT ResVT = N->getValueType(0);
8814 if (N0.getOpcode() != ISD::SETCC)
8817 // Make sure the SETCC result is either i1 (initial DAG), or i32, the lowered
8818 // scalar SetCCResultType. We also don't expect vectors, because we assume
8819 // that selects fed by vector SETCCs are canonicalized to VSELECT.
8820 assert((N0.getValueType() == MVT::i1 || N0.getValueType() == MVT::i32) &&
8821 "Scalar-SETCC feeding SELECT has unexpected result type!");
8823 // If NumMaskElts == 0, the comparison is larger than select result. The
8824 // largest real NEON comparison is 64-bits per lane, which means the result is
8825 // at most 32-bits and an illegal vector. Just bail out for now.
8826 EVT SrcVT = N0.getOperand(0).getValueType();
8828 // Don't try to do this optimization when the setcc itself has i1 operands.
8829 // There are no legal vectors of i1, so this would be pointless.
8830 if (SrcVT == MVT::i1)
8833 int NumMaskElts = ResVT.getSizeInBits() / SrcVT.getSizeInBits();
8834 if (!ResVT.isVector() || NumMaskElts == 0)
8837 SrcVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumMaskElts);
8838 EVT CCVT = SrcVT.changeVectorElementTypeToInteger();
8840 // Also bail out if the vector CCVT isn't the same size as ResVT.
8841 // This can happen if the SETCC operand size doesn't divide the ResVT size
8842 // (e.g., f64 vs v3f32).
8843 if (CCVT.getSizeInBits() != ResVT.getSizeInBits())
8846 // Make sure we didn't create illegal types, if we're not supposed to.
8847 assert(DCI.isBeforeLegalize() ||
8848 DAG.getTargetLoweringInfo().isTypeLegal(SrcVT));
8850 // First perform a vector comparison, where lane 0 is the one we're interested
8854 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(0));
8856 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(1));
8857 SDValue SetCC = DAG.getNode(ISD::SETCC, DL, CCVT, LHS, RHS, N0.getOperand(2));
8859 // Now duplicate the comparison mask we want across all other lanes.
8860 SmallVector<int, 8> DUPMask(CCVT.getVectorNumElements(), 0);
8861 SDValue Mask = DAG.getVectorShuffle(CCVT, DL, SetCC, SetCC, DUPMask.data());
8862 Mask = DAG.getNode(ISD::BITCAST, DL,
8863 ResVT.changeVectorElementTypeToInteger(), Mask);
8865 return DAG.getSelect(DL, ResVT, Mask, N->getOperand(1), N->getOperand(2));
8868 /// performSelectCCCombine - Target-specific DAG combining for ISD::SELECT_CC
8869 /// to match FMIN/FMAX patterns.
8870 static SDValue performSelectCCCombine(SDNode *N, SelectionDAG &DAG) {
8871 // Try to use FMIN/FMAX instructions for FP selects like "x < y ? x : y".
8872 // Unless the NoNaNsFPMath option is set, be careful about NaNs:
8873 // vmax/vmin return NaN if either operand is a NaN;
8874 // only do the transformation when it matches that behavior.
8876 SDValue CondLHS = N->getOperand(0);
8877 SDValue CondRHS = N->getOperand(1);
8878 SDValue LHS = N->getOperand(2);
8879 SDValue RHS = N->getOperand(3);
8880 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
8884 if (selectCCOpsAreFMaxCompatible(CondLHS, LHS) &&
8885 selectCCOpsAreFMaxCompatible(CondRHS, RHS)) {
8886 IsReversed = false; // x CC y ? x : y
8887 } else if (selectCCOpsAreFMaxCompatible(CondRHS, LHS) &&
8888 selectCCOpsAreFMaxCompatible(CondLHS, RHS)) {
8889 IsReversed = true ; // x CC y ? y : x
8894 bool IsUnordered = false, IsOrEqual;
8905 IsOrEqual = (CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE);
8906 Opcode = IsReversed ? AArch64ISD::FMAX : AArch64ISD::FMIN;
8916 IsOrEqual = (CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE);
8917 Opcode = IsReversed ? AArch64ISD::FMIN : AArch64ISD::FMAX;
8921 // If LHS is NaN, an ordered comparison will be false and the result will be
8922 // the RHS, but FMIN(NaN, RHS) = FMAX(NaN, RHS) = NaN. Avoid this by checking
8923 // that LHS != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8924 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8927 // For xxx-or-equal comparisons, "+0 <= -0" and "-0 >= +0" will both be true,
8928 // but FMIN will return -0, and FMAX will return +0. So FMIN/FMAX can only be
8929 // used for unsafe math or if one of the operands is known to be nonzero.
8930 if (IsOrEqual && !DAG.getTarget().Options.UnsafeFPMath &&
8931 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8934 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
8937 /// Get rid of unnecessary NVCASTs (that don't change the type).
8938 static SDValue performNVCASTCombine(SDNode *N) {
8939 if (N->getValueType(0) == N->getOperand(0).getValueType())
8940 return N->getOperand(0);
8945 SDValue AArch64TargetLowering::PerformDAGCombine(SDNode *N,
8946 DAGCombinerInfo &DCI) const {
8947 SelectionDAG &DAG = DCI.DAG;
8948 switch (N->getOpcode()) {
8953 return performAddSubLongCombine(N, DCI, DAG);
8955 return performXorCombine(N, DAG, DCI, Subtarget);
8957 return performMulCombine(N, DAG, DCI, Subtarget);
8958 case ISD::SINT_TO_FP:
8959 case ISD::UINT_TO_FP:
8960 return performIntToFpCombine(N, DAG, Subtarget);
8962 return performORCombine(N, DCI, Subtarget);
8963 case ISD::INTRINSIC_WO_CHAIN:
8964 return performIntrinsicCombine(N, DCI, Subtarget);
8965 case ISD::ANY_EXTEND:
8966 case ISD::ZERO_EXTEND:
8967 case ISD::SIGN_EXTEND:
8968 return performExtendCombine(N, DCI, DAG);
8970 return performBitcastCombine(N, DCI, DAG);
8971 case ISD::CONCAT_VECTORS:
8972 return performConcatVectorsCombine(N, DCI, DAG);
8974 return performSelectCombine(N, DCI);
8976 return performVSelectCombine(N, DCI.DAG);
8977 case ISD::SELECT_CC:
8978 return performSelectCCCombine(N, DCI.DAG);
8980 return performSTORECombine(N, DCI, DAG, Subtarget);
8981 case AArch64ISD::BRCOND:
8982 return performBRCONDCombine(N, DCI, DAG);
8983 case AArch64ISD::CSEL:
8984 return performCONDCombine(N, DCI, DAG, 2, 3);
8985 case AArch64ISD::DUP:
8986 return performPostLD1Combine(N, DCI, false);
8987 case AArch64ISD::NVCAST:
8988 return performNVCASTCombine(N);
8989 case ISD::INSERT_VECTOR_ELT:
8990 return performPostLD1Combine(N, DCI, true);
8991 case ISD::INTRINSIC_VOID:
8992 case ISD::INTRINSIC_W_CHAIN:
8993 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
8994 case Intrinsic::aarch64_neon_ld2:
8995 case Intrinsic::aarch64_neon_ld3:
8996 case Intrinsic::aarch64_neon_ld4:
8997 case Intrinsic::aarch64_neon_ld1x2:
8998 case Intrinsic::aarch64_neon_ld1x3:
8999 case Intrinsic::aarch64_neon_ld1x4:
9000 case Intrinsic::aarch64_neon_ld2lane:
9001 case Intrinsic::aarch64_neon_ld3lane:
9002 case Intrinsic::aarch64_neon_ld4lane:
9003 case Intrinsic::aarch64_neon_ld2r:
9004 case Intrinsic::aarch64_neon_ld3r:
9005 case Intrinsic::aarch64_neon_ld4r:
9006 case Intrinsic::aarch64_neon_st2:
9007 case Intrinsic::aarch64_neon_st3:
9008 case Intrinsic::aarch64_neon_st4:
9009 case Intrinsic::aarch64_neon_st1x2:
9010 case Intrinsic::aarch64_neon_st1x3:
9011 case Intrinsic::aarch64_neon_st1x4:
9012 case Intrinsic::aarch64_neon_st2lane:
9013 case Intrinsic::aarch64_neon_st3lane:
9014 case Intrinsic::aarch64_neon_st4lane:
9015 return performNEONPostLDSTCombine(N, DCI, DAG);
9023 // Check if the return value is used as only a return value, as otherwise
9024 // we can't perform a tail-call. In particular, we need to check for
9025 // target ISD nodes that are returns and any other "odd" constructs
9026 // that the generic analysis code won't necessarily catch.
9027 bool AArch64TargetLowering::isUsedByReturnOnly(SDNode *N,
9028 SDValue &Chain) const {
9029 if (N->getNumValues() != 1)
9031 if (!N->hasNUsesOfValue(1, 0))
9034 SDValue TCChain = Chain;
9035 SDNode *Copy = *N->use_begin();
9036 if (Copy->getOpcode() == ISD::CopyToReg) {
9037 // If the copy has a glue operand, we conservatively assume it isn't safe to
9038 // perform a tail call.
9039 if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() ==
9042 TCChain = Copy->getOperand(0);
9043 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
9046 bool HasRet = false;
9047 for (SDNode *Node : Copy->uses()) {
9048 if (Node->getOpcode() != AArch64ISD::RET_FLAG)
9060 // Return whether the an instruction can potentially be optimized to a tail
9061 // call. This will cause the optimizers to attempt to move, or duplicate,
9062 // return instructions to help enable tail call optimizations for this
9064 bool AArch64TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
9065 if (!CI->isTailCall())
9071 bool AArch64TargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base,
9073 ISD::MemIndexedMode &AM,
9075 SelectionDAG &DAG) const {
9076 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
9079 Base = Op->getOperand(0);
9080 // All of the indexed addressing mode instructions take a signed
9081 // 9 bit immediate offset.
9082 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
9083 int64_t RHSC = (int64_t)RHS->getZExtValue();
9084 if (RHSC >= 256 || RHSC <= -256)
9086 IsInc = (Op->getOpcode() == ISD::ADD);
9087 Offset = Op->getOperand(1);
9093 bool AArch64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
9095 ISD::MemIndexedMode &AM,
9096 SelectionDAG &DAG) const {
9099 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9100 VT = LD->getMemoryVT();
9101 Ptr = LD->getBasePtr();
9102 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9103 VT = ST->getMemoryVT();
9104 Ptr = ST->getBasePtr();
9109 if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, IsInc, DAG))
9111 AM = IsInc ? ISD::PRE_INC : ISD::PRE_DEC;
9115 bool AArch64TargetLowering::getPostIndexedAddressParts(
9116 SDNode *N, SDNode *Op, SDValue &Base, SDValue &Offset,
9117 ISD::MemIndexedMode &AM, SelectionDAG &DAG) const {
9120 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9121 VT = LD->getMemoryVT();
9122 Ptr = LD->getBasePtr();
9123 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9124 VT = ST->getMemoryVT();
9125 Ptr = ST->getBasePtr();
9130 if (!getIndexedAddressParts(Op, Base, Offset, AM, IsInc, DAG))
9132 // Post-indexing updates the base, so it's not a valid transform
9133 // if that's not the same as the load's pointer.
9136 AM = IsInc ? ISD::POST_INC : ISD::POST_DEC;
9140 static void ReplaceBITCASTResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
9141 SelectionDAG &DAG) {
9143 SDValue Op = N->getOperand(0);
9145 if (N->getValueType(0) != MVT::i16 || Op.getValueType() != MVT::f16)
9149 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, DL, MVT::f32,
9150 DAG.getUNDEF(MVT::i32), Op,
9151 DAG.getTargetConstant(AArch64::hsub, DL, MVT::i32)),
9153 Op = DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op);
9154 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, Op));
9157 void AArch64TargetLowering::ReplaceNodeResults(
9158 SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
9159 switch (N->getOpcode()) {
9161 llvm_unreachable("Don't know how to custom expand this");
9163 ReplaceBITCASTResults(N, Results, DAG);
9165 case ISD::FP_TO_UINT:
9166 case ISD::FP_TO_SINT:
9167 assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion");
9168 // Let normal code take care of it by not adding anything to Results.
9173 bool AArch64TargetLowering::useLoadStackGuardNode() const {
9177 bool AArch64TargetLowering::combineRepeatedFPDivisors(unsigned NumUsers) const {
9178 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
9179 // reciprocal if there are three or more FDIVs.
9180 return NumUsers > 2;
9183 TargetLoweringBase::LegalizeTypeAction
9184 AArch64TargetLowering::getPreferredVectorAction(EVT VT) const {
9185 MVT SVT = VT.getSimpleVT();
9186 // During type legalization, we prefer to widen v1i8, v1i16, v1i32 to v8i8,
9187 // v4i16, v2i32 instead of to promote.
9188 if (SVT == MVT::v1i8 || SVT == MVT::v1i16 || SVT == MVT::v1i32
9189 || SVT == MVT::v1f32)
9190 return TypeWidenVector;
9192 return TargetLoweringBase::getPreferredVectorAction(VT);
9195 // Loads and stores less than 128-bits are already atomic; ones above that
9196 // are doomed anyway, so defer to the default libcall and blame the OS when
9198 bool AArch64TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
9199 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
9203 // Loads and stores less than 128-bits are already atomic; ones above that
9204 // are doomed anyway, so defer to the default libcall and blame the OS when
9206 bool AArch64TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
9207 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
9211 // For the real atomic operations, we have ldxr/stxr up to 128 bits,
9212 TargetLoweringBase::AtomicRMWExpansionKind
9213 AArch64TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
9214 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
9215 return Size <= 128 ? AtomicRMWExpansionKind::LLSC
9216 : AtomicRMWExpansionKind::None;
9219 bool AArch64TargetLowering::hasLoadLinkedStoreConditional() const {
9223 Value *AArch64TargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
9224 AtomicOrdering Ord) const {
9225 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
9226 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
9227 bool IsAcquire = isAtLeastAcquire(Ord);
9229 // Since i128 isn't legal and intrinsics don't get type-lowered, the ldrexd
9230 // intrinsic must return {i64, i64} and we have to recombine them into a
9231 // single i128 here.
9232 if (ValTy->getPrimitiveSizeInBits() == 128) {
9234 IsAcquire ? Intrinsic::aarch64_ldaxp : Intrinsic::aarch64_ldxp;
9235 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int);
9237 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
9238 Value *LoHi = Builder.CreateCall(Ldxr, Addr, "lohi");
9240 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
9241 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
9242 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
9243 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
9244 return Builder.CreateOr(
9245 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 64)), "val64");
9248 Type *Tys[] = { Addr->getType() };
9250 IsAcquire ? Intrinsic::aarch64_ldaxr : Intrinsic::aarch64_ldxr;
9251 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int, Tys);
9253 return Builder.CreateTruncOrBitCast(
9254 Builder.CreateCall(Ldxr, Addr),
9255 cast<PointerType>(Addr->getType())->getElementType());
9258 Value *AArch64TargetLowering::emitStoreConditional(IRBuilder<> &Builder,
9259 Value *Val, Value *Addr,
9260 AtomicOrdering Ord) const {
9261 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
9262 bool IsRelease = isAtLeastRelease(Ord);
9264 // Since the intrinsics must have legal type, the i128 intrinsics take two
9265 // parameters: "i64, i64". We must marshal Val into the appropriate form
9267 if (Val->getType()->getPrimitiveSizeInBits() == 128) {
9269 IsRelease ? Intrinsic::aarch64_stlxp : Intrinsic::aarch64_stxp;
9270 Function *Stxr = Intrinsic::getDeclaration(M, Int);
9271 Type *Int64Ty = Type::getInt64Ty(M->getContext());
9273 Value *Lo = Builder.CreateTrunc(Val, Int64Ty, "lo");
9274 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 64), Int64Ty, "hi");
9275 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
9276 return Builder.CreateCall(Stxr, {Lo, Hi, Addr});
9280 IsRelease ? Intrinsic::aarch64_stlxr : Intrinsic::aarch64_stxr;
9281 Type *Tys[] = { Addr->getType() };
9282 Function *Stxr = Intrinsic::getDeclaration(M, Int, Tys);
9284 return Builder.CreateCall(Stxr,
9285 {Builder.CreateZExtOrBitCast(
9286 Val, Stxr->getFunctionType()->getParamType(0)),
9290 bool AArch64TargetLowering::functionArgumentNeedsConsecutiveRegisters(
9291 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
9292 return Ty->isArrayTy();