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/Intrinsics.h"
29 #include "llvm/IR/Type.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetOptions.h"
37 #define DEBUG_TYPE "aarch64-lower"
39 STATISTIC(NumTailCalls, "Number of tail calls");
40 STATISTIC(NumShiftInserts, "Number of vector shift inserts");
49 static cl::opt<AlignMode>
50 Align(cl::desc("Load/store alignment support"),
51 cl::Hidden, cl::init(NoStrictAlign),
53 clEnumValN(StrictAlign, "aarch64-strict-align",
54 "Disallow all unaligned memory accesses"),
55 clEnumValN(NoStrictAlign, "aarch64-no-strict-align",
56 "Allow unaligned memory accesses"),
59 // Place holder until extr generation is tested fully.
61 EnableAArch64ExtrGeneration("aarch64-extr-generation", cl::Hidden,
62 cl::desc("Allow AArch64 (or (shift)(shift))->extract"),
66 EnableAArch64SlrGeneration("aarch64-shift-insert-generation", cl::Hidden,
67 cl::desc("Allow AArch64 SLI/SRI formation"),
70 // FIXME: The necessary dtprel relocations don't seem to be supported
71 // well in the GNU bfd and gold linkers at the moment. Therefore, by
72 // default, for now, fall back to GeneralDynamic code generation.
73 cl::opt<bool> EnableAArch64ELFLocalDynamicTLSGeneration(
74 "aarch64-elf-ldtls-generation", cl::Hidden,
75 cl::desc("Allow AArch64 Local Dynamic TLS code generation"),
78 AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
79 const AArch64Subtarget &STI)
80 : TargetLowering(TM), Subtarget(&STI) {
82 // AArch64 doesn't have comparisons which set GPRs or setcc instructions, so
83 // we have to make something up. Arbitrarily, choose ZeroOrOne.
84 setBooleanContents(ZeroOrOneBooleanContent);
85 // When comparing vectors the result sets the different elements in the
86 // vector to all-one or all-zero.
87 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
89 // Set up the register classes.
90 addRegisterClass(MVT::i32, &AArch64::GPR32allRegClass);
91 addRegisterClass(MVT::i64, &AArch64::GPR64allRegClass);
93 if (Subtarget->hasFPARMv8()) {
94 addRegisterClass(MVT::f16, &AArch64::FPR16RegClass);
95 addRegisterClass(MVT::f32, &AArch64::FPR32RegClass);
96 addRegisterClass(MVT::f64, &AArch64::FPR64RegClass);
97 addRegisterClass(MVT::f128, &AArch64::FPR128RegClass);
100 if (Subtarget->hasNEON()) {
101 addRegisterClass(MVT::v16i8, &AArch64::FPR8RegClass);
102 addRegisterClass(MVT::v8i16, &AArch64::FPR16RegClass);
103 // Someone set us up the NEON.
104 addDRTypeForNEON(MVT::v2f32);
105 addDRTypeForNEON(MVT::v8i8);
106 addDRTypeForNEON(MVT::v4i16);
107 addDRTypeForNEON(MVT::v2i32);
108 addDRTypeForNEON(MVT::v1i64);
109 addDRTypeForNEON(MVT::v1f64);
110 addDRTypeForNEON(MVT::v4f16);
112 addQRTypeForNEON(MVT::v4f32);
113 addQRTypeForNEON(MVT::v2f64);
114 addQRTypeForNEON(MVT::v16i8);
115 addQRTypeForNEON(MVT::v8i16);
116 addQRTypeForNEON(MVT::v4i32);
117 addQRTypeForNEON(MVT::v2i64);
118 addQRTypeForNEON(MVT::v8f16);
121 // Compute derived properties from the register classes
122 computeRegisterProperties(Subtarget->getRegisterInfo());
124 // Provide all sorts of operation actions
125 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
126 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
127 setOperationAction(ISD::SETCC, MVT::i32, Custom);
128 setOperationAction(ISD::SETCC, MVT::i64, Custom);
129 setOperationAction(ISD::SETCC, MVT::f32, Custom);
130 setOperationAction(ISD::SETCC, MVT::f64, Custom);
131 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
132 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
133 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
134 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
135 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
136 setOperationAction(ISD::SELECT, MVT::i32, Custom);
137 setOperationAction(ISD::SELECT, MVT::i64, Custom);
138 setOperationAction(ISD::SELECT, MVT::f32, Custom);
139 setOperationAction(ISD::SELECT, MVT::f64, Custom);
140 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
141 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
142 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
143 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
144 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
145 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
147 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
148 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
149 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
151 setOperationAction(ISD::FREM, MVT::f32, Expand);
152 setOperationAction(ISD::FREM, MVT::f64, Expand);
153 setOperationAction(ISD::FREM, MVT::f80, Expand);
155 // Custom lowering hooks are needed for XOR
156 // to fold it into CSINC/CSINV.
157 setOperationAction(ISD::XOR, MVT::i32, Custom);
158 setOperationAction(ISD::XOR, MVT::i64, Custom);
160 // Virtually no operation on f128 is legal, but LLVM can't expand them when
161 // there's a valid register class, so we need custom operations in most cases.
162 setOperationAction(ISD::FABS, MVT::f128, Expand);
163 setOperationAction(ISD::FADD, MVT::f128, Custom);
164 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
165 setOperationAction(ISD::FCOS, MVT::f128, Expand);
166 setOperationAction(ISD::FDIV, MVT::f128, Custom);
167 setOperationAction(ISD::FMA, MVT::f128, Expand);
168 setOperationAction(ISD::FMUL, MVT::f128, Custom);
169 setOperationAction(ISD::FNEG, MVT::f128, Expand);
170 setOperationAction(ISD::FPOW, MVT::f128, Expand);
171 setOperationAction(ISD::FREM, MVT::f128, Expand);
172 setOperationAction(ISD::FRINT, MVT::f128, Expand);
173 setOperationAction(ISD::FSIN, MVT::f128, Expand);
174 setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
175 setOperationAction(ISD::FSQRT, MVT::f128, Expand);
176 setOperationAction(ISD::FSUB, MVT::f128, Custom);
177 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
178 setOperationAction(ISD::SETCC, MVT::f128, Custom);
179 setOperationAction(ISD::BR_CC, MVT::f128, Custom);
180 setOperationAction(ISD::SELECT, MVT::f128, Custom);
181 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
182 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);
184 // Lowering for many of the conversions is actually specified by the non-f128
185 // type. The LowerXXX function will be trivial when f128 isn't involved.
186 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
187 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
188 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
189 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
190 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
191 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
192 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
193 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
194 setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
195 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
196 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
197 setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
198 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
199 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
201 // Variable arguments.
202 setOperationAction(ISD::VASTART, MVT::Other, Custom);
203 setOperationAction(ISD::VAARG, MVT::Other, Custom);
204 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
205 setOperationAction(ISD::VAEND, MVT::Other, Expand);
207 // Variable-sized objects.
208 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
209 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
210 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
212 // Exception handling.
213 // FIXME: These are guesses. Has this been defined yet?
214 setExceptionPointerRegister(AArch64::X0);
215 setExceptionSelectorRegister(AArch64::X1);
217 // Constant pool entries
218 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
221 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
223 // Add/Sub overflow ops with MVT::Glues are lowered to NZCV dependences.
224 setOperationAction(ISD::ADDC, MVT::i32, Custom);
225 setOperationAction(ISD::ADDE, MVT::i32, Custom);
226 setOperationAction(ISD::SUBC, MVT::i32, Custom);
227 setOperationAction(ISD::SUBE, MVT::i32, Custom);
228 setOperationAction(ISD::ADDC, MVT::i64, Custom);
229 setOperationAction(ISD::ADDE, MVT::i64, Custom);
230 setOperationAction(ISD::SUBC, MVT::i64, Custom);
231 setOperationAction(ISD::SUBE, MVT::i64, Custom);
233 // AArch64 lacks both left-rotate and popcount instructions.
234 setOperationAction(ISD::ROTL, MVT::i32, Expand);
235 setOperationAction(ISD::ROTL, MVT::i64, Expand);
237 // AArch64 doesn't have {U|S}MUL_LOHI.
238 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
239 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
242 // Expand the undefined-at-zero variants to cttz/ctlz to their defined-at-zero
243 // counterparts, which AArch64 supports directly.
244 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
245 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
246 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
247 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
249 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
250 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
252 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
253 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
254 setOperationAction(ISD::SREM, MVT::i32, Expand);
255 setOperationAction(ISD::SREM, MVT::i64, Expand);
256 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
257 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
258 setOperationAction(ISD::UREM, MVT::i32, Expand);
259 setOperationAction(ISD::UREM, MVT::i64, Expand);
261 // Custom lower Add/Sub/Mul with overflow.
262 setOperationAction(ISD::SADDO, MVT::i32, Custom);
263 setOperationAction(ISD::SADDO, MVT::i64, Custom);
264 setOperationAction(ISD::UADDO, MVT::i32, Custom);
265 setOperationAction(ISD::UADDO, MVT::i64, Custom);
266 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
267 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
268 setOperationAction(ISD::USUBO, MVT::i32, Custom);
269 setOperationAction(ISD::USUBO, MVT::i64, Custom);
270 setOperationAction(ISD::SMULO, MVT::i32, Custom);
271 setOperationAction(ISD::SMULO, MVT::i64, Custom);
272 setOperationAction(ISD::UMULO, MVT::i32, Custom);
273 setOperationAction(ISD::UMULO, MVT::i64, Custom);
275 setOperationAction(ISD::FSIN, MVT::f32, Expand);
276 setOperationAction(ISD::FSIN, MVT::f64, Expand);
277 setOperationAction(ISD::FCOS, MVT::f32, Expand);
278 setOperationAction(ISD::FCOS, MVT::f64, Expand);
279 setOperationAction(ISD::FPOW, MVT::f32, Expand);
280 setOperationAction(ISD::FPOW, MVT::f64, Expand);
281 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
282 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
284 // f16 is a storage-only type, always promote it to f32.
285 setOperationAction(ISD::SETCC, MVT::f16, Promote);
286 setOperationAction(ISD::BR_CC, MVT::f16, Promote);
287 setOperationAction(ISD::SELECT_CC, MVT::f16, Promote);
288 setOperationAction(ISD::SELECT, MVT::f16, Promote);
289 setOperationAction(ISD::FADD, MVT::f16, Promote);
290 setOperationAction(ISD::FSUB, MVT::f16, Promote);
291 setOperationAction(ISD::FMUL, MVT::f16, Promote);
292 setOperationAction(ISD::FDIV, MVT::f16, Promote);
293 setOperationAction(ISD::FREM, MVT::f16, Promote);
294 setOperationAction(ISD::FMA, MVT::f16, Promote);
295 setOperationAction(ISD::FNEG, MVT::f16, Promote);
296 setOperationAction(ISD::FABS, MVT::f16, Promote);
297 setOperationAction(ISD::FCEIL, MVT::f16, Promote);
298 setOperationAction(ISD::FCOPYSIGN, MVT::f16, Promote);
299 setOperationAction(ISD::FCOS, MVT::f16, Promote);
300 setOperationAction(ISD::FFLOOR, MVT::f16, Promote);
301 setOperationAction(ISD::FNEARBYINT, MVT::f16, Promote);
302 setOperationAction(ISD::FPOW, MVT::f16, Promote);
303 setOperationAction(ISD::FPOWI, MVT::f16, Promote);
304 setOperationAction(ISD::FRINT, MVT::f16, Promote);
305 setOperationAction(ISD::FSIN, MVT::f16, Promote);
306 setOperationAction(ISD::FSINCOS, MVT::f16, Promote);
307 setOperationAction(ISD::FSQRT, MVT::f16, Promote);
308 setOperationAction(ISD::FEXP, MVT::f16, Promote);
309 setOperationAction(ISD::FEXP2, MVT::f16, Promote);
310 setOperationAction(ISD::FLOG, MVT::f16, Promote);
311 setOperationAction(ISD::FLOG2, MVT::f16, Promote);
312 setOperationAction(ISD::FLOG10, MVT::f16, Promote);
313 setOperationAction(ISD::FROUND, MVT::f16, Promote);
314 setOperationAction(ISD::FTRUNC, MVT::f16, Promote);
315 setOperationAction(ISD::FMINNUM, MVT::f16, Promote);
316 setOperationAction(ISD::FMAXNUM, MVT::f16, Promote);
318 // v4f16 is also a storage-only type, so promote it to v4f32 when that is
320 setOperationAction(ISD::FADD, MVT::v4f16, Promote);
321 setOperationAction(ISD::FSUB, MVT::v4f16, Promote);
322 setOperationAction(ISD::FMUL, MVT::v4f16, Promote);
323 setOperationAction(ISD::FDIV, MVT::v4f16, Promote);
324 setOperationAction(ISD::FP_EXTEND, MVT::v4f16, Promote);
325 setOperationAction(ISD::FP_ROUND, MVT::v4f16, Promote);
326 AddPromotedToType(ISD::FADD, MVT::v4f16, MVT::v4f32);
327 AddPromotedToType(ISD::FSUB, MVT::v4f16, MVT::v4f32);
328 AddPromotedToType(ISD::FMUL, MVT::v4f16, MVT::v4f32);
329 AddPromotedToType(ISD::FDIV, MVT::v4f16, MVT::v4f32);
330 AddPromotedToType(ISD::FP_EXTEND, MVT::v4f16, MVT::v4f32);
331 AddPromotedToType(ISD::FP_ROUND, MVT::v4f16, MVT::v4f32);
333 // Expand all other v4f16 operations.
334 // FIXME: We could generate better code by promoting some operations to
336 setOperationAction(ISD::FABS, MVT::v4f16, Expand);
337 setOperationAction(ISD::FCEIL, MVT::v4f16, Expand);
338 setOperationAction(ISD::FCOPYSIGN, MVT::v4f16, Expand);
339 setOperationAction(ISD::FCOS, MVT::v4f16, Expand);
340 setOperationAction(ISD::FFLOOR, MVT::v4f16, Expand);
341 setOperationAction(ISD::FMA, MVT::v4f16, Expand);
342 setOperationAction(ISD::FNEARBYINT, MVT::v4f16, Expand);
343 setOperationAction(ISD::FNEG, MVT::v4f16, Expand);
344 setOperationAction(ISD::FPOW, MVT::v4f16, Expand);
345 setOperationAction(ISD::FPOWI, MVT::v4f16, Expand);
346 setOperationAction(ISD::FREM, MVT::v4f16, Expand);
347 setOperationAction(ISD::FROUND, MVT::v4f16, Expand);
348 setOperationAction(ISD::FRINT, MVT::v4f16, Expand);
349 setOperationAction(ISD::FSIN, MVT::v4f16, Expand);
350 setOperationAction(ISD::FSINCOS, MVT::v4f16, Expand);
351 setOperationAction(ISD::FSQRT, MVT::v4f16, Expand);
352 setOperationAction(ISD::FTRUNC, MVT::v4f16, Expand);
353 setOperationAction(ISD::SETCC, MVT::v4f16, Expand);
354 setOperationAction(ISD::BR_CC, MVT::v4f16, Expand);
355 setOperationAction(ISD::SELECT, MVT::v4f16, Expand);
356 setOperationAction(ISD::SELECT_CC, MVT::v4f16, Expand);
357 setOperationAction(ISD::FEXP, MVT::v4f16, Expand);
358 setOperationAction(ISD::FEXP2, MVT::v4f16, Expand);
359 setOperationAction(ISD::FLOG, MVT::v4f16, Expand);
360 setOperationAction(ISD::FLOG2, MVT::v4f16, Expand);
361 setOperationAction(ISD::FLOG10, MVT::v4f16, Expand);
364 // v8f16 is also a storage-only type, so expand it.
365 setOperationAction(ISD::FABS, MVT::v8f16, Expand);
366 setOperationAction(ISD::FADD, MVT::v8f16, Expand);
367 setOperationAction(ISD::FCEIL, MVT::v8f16, Expand);
368 setOperationAction(ISD::FCOPYSIGN, MVT::v8f16, Expand);
369 setOperationAction(ISD::FCOS, MVT::v8f16, Expand);
370 setOperationAction(ISD::FDIV, MVT::v8f16, Expand);
371 setOperationAction(ISD::FFLOOR, MVT::v8f16, Expand);
372 setOperationAction(ISD::FMA, MVT::v8f16, Expand);
373 setOperationAction(ISD::FMUL, MVT::v8f16, Expand);
374 setOperationAction(ISD::FNEARBYINT, MVT::v8f16, Expand);
375 setOperationAction(ISD::FNEG, MVT::v8f16, Expand);
376 setOperationAction(ISD::FPOW, MVT::v8f16, Expand);
377 setOperationAction(ISD::FPOWI, MVT::v8f16, Expand);
378 setOperationAction(ISD::FREM, MVT::v8f16, Expand);
379 setOperationAction(ISD::FROUND, MVT::v8f16, Expand);
380 setOperationAction(ISD::FRINT, MVT::v8f16, Expand);
381 setOperationAction(ISD::FSIN, MVT::v8f16, Expand);
382 setOperationAction(ISD::FSINCOS, MVT::v8f16, Expand);
383 setOperationAction(ISD::FSQRT, MVT::v8f16, Expand);
384 setOperationAction(ISD::FSUB, MVT::v8f16, Expand);
385 setOperationAction(ISD::FTRUNC, MVT::v8f16, Expand);
386 setOperationAction(ISD::SETCC, MVT::v8f16, Expand);
387 setOperationAction(ISD::BR_CC, MVT::v8f16, Expand);
388 setOperationAction(ISD::SELECT, MVT::v8f16, Expand);
389 setOperationAction(ISD::SELECT_CC, MVT::v8f16, Expand);
390 setOperationAction(ISD::FP_EXTEND, MVT::v8f16, Expand);
391 setOperationAction(ISD::FEXP, MVT::v8f16, Expand);
392 setOperationAction(ISD::FEXP2, MVT::v8f16, Expand);
393 setOperationAction(ISD::FLOG, MVT::v8f16, Expand);
394 setOperationAction(ISD::FLOG2, MVT::v8f16, Expand);
395 setOperationAction(ISD::FLOG10, MVT::v8f16, Expand);
397 // AArch64 has implementations of a lot of rounding-like FP operations.
398 for (MVT Ty : {MVT::f32, MVT::f64}) {
399 setOperationAction(ISD::FFLOOR, Ty, Legal);
400 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
401 setOperationAction(ISD::FCEIL, Ty, Legal);
402 setOperationAction(ISD::FRINT, Ty, Legal);
403 setOperationAction(ISD::FTRUNC, Ty, Legal);
404 setOperationAction(ISD::FROUND, Ty, Legal);
407 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
409 if (Subtarget->isTargetMachO()) {
410 // For iOS, we don't want to the normal expansion of a libcall to
411 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
413 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
414 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
416 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
417 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
420 // Make floating-point constants legal for the large code model, so they don't
421 // become loads from the constant pool.
422 if (Subtarget->isTargetMachO() && TM.getCodeModel() == CodeModel::Large) {
423 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
424 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
427 // AArch64 does not have floating-point extending loads, i1 sign-extending
428 // load, floating-point truncating stores, or v2i32->v2i16 truncating store.
429 for (MVT VT : MVT::fp_valuetypes()) {
430 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
431 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
432 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f64, Expand);
433 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f80, Expand);
435 for (MVT VT : MVT::integer_valuetypes())
436 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Expand);
438 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
439 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
440 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
441 setTruncStoreAction(MVT::f128, MVT::f80, Expand);
442 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
443 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
444 setTruncStoreAction(MVT::f128, MVT::f16, Expand);
446 setOperationAction(ISD::BITCAST, MVT::i16, Custom);
447 setOperationAction(ISD::BITCAST, MVT::f16, Custom);
449 // Indexed loads and stores are supported.
450 for (unsigned im = (unsigned)ISD::PRE_INC;
451 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
452 setIndexedLoadAction(im, MVT::i8, Legal);
453 setIndexedLoadAction(im, MVT::i16, Legal);
454 setIndexedLoadAction(im, MVT::i32, Legal);
455 setIndexedLoadAction(im, MVT::i64, Legal);
456 setIndexedLoadAction(im, MVT::f64, Legal);
457 setIndexedLoadAction(im, MVT::f32, Legal);
458 setIndexedStoreAction(im, MVT::i8, Legal);
459 setIndexedStoreAction(im, MVT::i16, Legal);
460 setIndexedStoreAction(im, MVT::i32, Legal);
461 setIndexedStoreAction(im, MVT::i64, Legal);
462 setIndexedStoreAction(im, MVT::f64, Legal);
463 setIndexedStoreAction(im, MVT::f32, Legal);
467 setOperationAction(ISD::TRAP, MVT::Other, Legal);
469 // We combine OR nodes for bitfield operations.
470 setTargetDAGCombine(ISD::OR);
472 // Vector add and sub nodes may conceal a high-half opportunity.
473 // Also, try to fold ADD into CSINC/CSINV..
474 setTargetDAGCombine(ISD::ADD);
475 setTargetDAGCombine(ISD::SUB);
477 setTargetDAGCombine(ISD::XOR);
478 setTargetDAGCombine(ISD::SINT_TO_FP);
479 setTargetDAGCombine(ISD::UINT_TO_FP);
481 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
483 setTargetDAGCombine(ISD::ANY_EXTEND);
484 setTargetDAGCombine(ISD::ZERO_EXTEND);
485 setTargetDAGCombine(ISD::SIGN_EXTEND);
486 setTargetDAGCombine(ISD::BITCAST);
487 setTargetDAGCombine(ISD::CONCAT_VECTORS);
488 setTargetDAGCombine(ISD::STORE);
490 setTargetDAGCombine(ISD::MUL);
492 setTargetDAGCombine(ISD::SELECT);
493 setTargetDAGCombine(ISD::VSELECT);
495 setTargetDAGCombine(ISD::INTRINSIC_VOID);
496 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
497 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
499 MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8;
500 MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4;
501 MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 4;
503 setStackPointerRegisterToSaveRestore(AArch64::SP);
505 setSchedulingPreference(Sched::Hybrid);
508 MaskAndBranchFoldingIsLegal = true;
509 EnableExtLdPromotion = true;
511 setMinFunctionAlignment(2);
513 RequireStrictAlign = (Align == StrictAlign);
515 setHasExtractBitsInsn(true);
517 if (Subtarget->hasNEON()) {
518 // FIXME: v1f64 shouldn't be legal if we can avoid it, because it leads to
519 // silliness like this:
520 setOperationAction(ISD::FABS, MVT::v1f64, Expand);
521 setOperationAction(ISD::FADD, MVT::v1f64, Expand);
522 setOperationAction(ISD::FCEIL, MVT::v1f64, Expand);
523 setOperationAction(ISD::FCOPYSIGN, MVT::v1f64, Expand);
524 setOperationAction(ISD::FCOS, MVT::v1f64, Expand);
525 setOperationAction(ISD::FDIV, MVT::v1f64, Expand);
526 setOperationAction(ISD::FFLOOR, MVT::v1f64, Expand);
527 setOperationAction(ISD::FMA, MVT::v1f64, Expand);
528 setOperationAction(ISD::FMUL, MVT::v1f64, Expand);
529 setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Expand);
530 setOperationAction(ISD::FNEG, MVT::v1f64, Expand);
531 setOperationAction(ISD::FPOW, MVT::v1f64, Expand);
532 setOperationAction(ISD::FREM, MVT::v1f64, Expand);
533 setOperationAction(ISD::FROUND, MVT::v1f64, Expand);
534 setOperationAction(ISD::FRINT, MVT::v1f64, Expand);
535 setOperationAction(ISD::FSIN, MVT::v1f64, Expand);
536 setOperationAction(ISD::FSINCOS, MVT::v1f64, Expand);
537 setOperationAction(ISD::FSQRT, MVT::v1f64, Expand);
538 setOperationAction(ISD::FSUB, MVT::v1f64, Expand);
539 setOperationAction(ISD::FTRUNC, MVT::v1f64, Expand);
540 setOperationAction(ISD::SETCC, MVT::v1f64, Expand);
541 setOperationAction(ISD::BR_CC, MVT::v1f64, Expand);
542 setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
543 setOperationAction(ISD::SELECT_CC, MVT::v1f64, Expand);
544 setOperationAction(ISD::FP_EXTEND, MVT::v1f64, Expand);
546 setOperationAction(ISD::FP_TO_SINT, MVT::v1i64, Expand);
547 setOperationAction(ISD::FP_TO_UINT, MVT::v1i64, Expand);
548 setOperationAction(ISD::SINT_TO_FP, MVT::v1i64, Expand);
549 setOperationAction(ISD::UINT_TO_FP, MVT::v1i64, Expand);
550 setOperationAction(ISD::FP_ROUND, MVT::v1f64, Expand);
552 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
554 // AArch64 doesn't have a direct vector ->f32 conversion instructions for
555 // elements smaller than i32, so promote the input to i32 first.
556 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Promote);
557 setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Promote);
558 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Promote);
559 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Promote);
560 // Similarly, there is no direct i32 -> f64 vector conversion instruction.
561 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
562 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
563 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
564 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
566 // AArch64 doesn't have MUL.2d:
567 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
568 // Custom handling for some quad-vector types to detect MULL.
569 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
570 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
571 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
573 setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal);
574 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
575 // Likewise, narrowing and extending vector loads/stores aren't handled
577 for (MVT VT : MVT::vector_valuetypes()) {
578 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
580 setOperationAction(ISD::MULHS, VT, Expand);
581 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
582 setOperationAction(ISD::MULHU, VT, Expand);
583 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
585 setOperationAction(ISD::BSWAP, VT, Expand);
587 for (MVT InnerVT : MVT::vector_valuetypes()) {
588 setTruncStoreAction(VT, InnerVT, Expand);
589 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
590 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
591 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
595 // AArch64 has implementations of a lot of rounding-like FP operations.
596 for (MVT Ty : {MVT::v2f32, MVT::v4f32, MVT::v2f64}) {
597 setOperationAction(ISD::FFLOOR, Ty, Legal);
598 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
599 setOperationAction(ISD::FCEIL, Ty, Legal);
600 setOperationAction(ISD::FRINT, Ty, Legal);
601 setOperationAction(ISD::FTRUNC, Ty, Legal);
602 setOperationAction(ISD::FROUND, Ty, Legal);
606 // Prefer likely predicted branches to selects on out-of-order cores.
607 if (Subtarget->isCortexA57())
608 PredictableSelectIsExpensive = true;
611 void AArch64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) {
612 if (VT == MVT::v2f32 || VT == MVT::v4f16) {
613 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
614 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i32);
616 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
617 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i32);
618 } else if (VT == MVT::v2f64 || VT == MVT::v4f32 || VT == MVT::v8f16) {
619 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
620 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i64);
622 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
623 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i64);
626 // Mark vector float intrinsics as expand.
627 if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) {
628 setOperationAction(ISD::FSIN, VT.getSimpleVT(), Expand);
629 setOperationAction(ISD::FCOS, VT.getSimpleVT(), Expand);
630 setOperationAction(ISD::FPOWI, VT.getSimpleVT(), Expand);
631 setOperationAction(ISD::FPOW, VT.getSimpleVT(), Expand);
632 setOperationAction(ISD::FLOG, VT.getSimpleVT(), Expand);
633 setOperationAction(ISD::FLOG2, VT.getSimpleVT(), Expand);
634 setOperationAction(ISD::FLOG10, VT.getSimpleVT(), Expand);
635 setOperationAction(ISD::FEXP, VT.getSimpleVT(), Expand);
636 setOperationAction(ISD::FEXP2, VT.getSimpleVT(), Expand);
639 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
640 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
641 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
642 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
643 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Custom);
644 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
645 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
646 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
647 setOperationAction(ISD::AND, VT.getSimpleVT(), Custom);
648 setOperationAction(ISD::OR, VT.getSimpleVT(), Custom);
649 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
650 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
652 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
653 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
654 setOperationAction(ISD::VSELECT, VT.getSimpleVT(), Expand);
655 for (MVT InnerVT : MVT::all_valuetypes())
656 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT.getSimpleVT(), Expand);
658 // CNT supports only B element sizes.
659 if (VT != MVT::v8i8 && VT != MVT::v16i8)
660 setOperationAction(ISD::CTPOP, VT.getSimpleVT(), Expand);
662 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
663 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
664 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
665 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
666 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
668 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
669 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
671 if (Subtarget->isLittleEndian()) {
672 for (unsigned im = (unsigned)ISD::PRE_INC;
673 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
674 setIndexedLoadAction(im, VT.getSimpleVT(), Legal);
675 setIndexedStoreAction(im, VT.getSimpleVT(), Legal);
680 void AArch64TargetLowering::addDRTypeForNEON(MVT VT) {
681 addRegisterClass(VT, &AArch64::FPR64RegClass);
682 addTypeForNEON(VT, MVT::v2i32);
685 void AArch64TargetLowering::addQRTypeForNEON(MVT VT) {
686 addRegisterClass(VT, &AArch64::FPR128RegClass);
687 addTypeForNEON(VT, MVT::v4i32);
690 EVT AArch64TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
693 return VT.changeVectorElementTypeToInteger();
696 /// computeKnownBitsForTargetNode - Determine which of the bits specified in
697 /// Mask are known to be either zero or one and return them in the
698 /// KnownZero/KnownOne bitsets.
699 void AArch64TargetLowering::computeKnownBitsForTargetNode(
700 const SDValue Op, APInt &KnownZero, APInt &KnownOne,
701 const SelectionDAG &DAG, unsigned Depth) const {
702 switch (Op.getOpcode()) {
705 case AArch64ISD::CSEL: {
706 APInt KnownZero2, KnownOne2;
707 DAG.computeKnownBits(Op->getOperand(0), KnownZero, KnownOne, Depth + 1);
708 DAG.computeKnownBits(Op->getOperand(1), KnownZero2, KnownOne2, Depth + 1);
709 KnownZero &= KnownZero2;
710 KnownOne &= KnownOne2;
713 case ISD::INTRINSIC_W_CHAIN: {
714 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
715 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
718 case Intrinsic::aarch64_ldaxr:
719 case Intrinsic::aarch64_ldxr: {
720 unsigned BitWidth = KnownOne.getBitWidth();
721 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
722 unsigned MemBits = VT.getScalarType().getSizeInBits();
723 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
729 case ISD::INTRINSIC_WO_CHAIN:
730 case ISD::INTRINSIC_VOID: {
731 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
735 case Intrinsic::aarch64_neon_umaxv:
736 case Intrinsic::aarch64_neon_uminv: {
737 // Figure out the datatype of the vector operand. The UMINV instruction
738 // will zero extend the result, so we can mark as known zero all the
739 // bits larger than the element datatype. 32-bit or larget doesn't need
740 // this as those are legal types and will be handled by isel directly.
741 MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
742 unsigned BitWidth = KnownZero.getBitWidth();
743 if (VT == MVT::v8i8 || VT == MVT::v16i8) {
744 assert(BitWidth >= 8 && "Unexpected width!");
745 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8);
747 } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
748 assert(BitWidth >= 16 && "Unexpected width!");
749 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
759 MVT AArch64TargetLowering::getScalarShiftAmountTy(EVT LHSTy) const {
764 AArch64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
765 const TargetLibraryInfo *libInfo) const {
766 return AArch64::createFastISel(funcInfo, libInfo);
769 const char *AArch64TargetLowering::getTargetNodeName(unsigned Opcode) const {
773 case AArch64ISD::CALL: return "AArch64ISD::CALL";
774 case AArch64ISD::ADRP: return "AArch64ISD::ADRP";
775 case AArch64ISD::ADDlow: return "AArch64ISD::ADDlow";
776 case AArch64ISD::LOADgot: return "AArch64ISD::LOADgot";
777 case AArch64ISD::RET_FLAG: return "AArch64ISD::RET_FLAG";
778 case AArch64ISD::BRCOND: return "AArch64ISD::BRCOND";
779 case AArch64ISD::CSEL: return "AArch64ISD::CSEL";
780 case AArch64ISD::FCSEL: return "AArch64ISD::FCSEL";
781 case AArch64ISD::CSINV: return "AArch64ISD::CSINV";
782 case AArch64ISD::CSNEG: return "AArch64ISD::CSNEG";
783 case AArch64ISD::CSINC: return "AArch64ISD::CSINC";
784 case AArch64ISD::THREAD_POINTER: return "AArch64ISD::THREAD_POINTER";
785 case AArch64ISD::TLSDESC_CALLSEQ: return "AArch64ISD::TLSDESC_CALLSEQ";
786 case AArch64ISD::ADC: return "AArch64ISD::ADC";
787 case AArch64ISD::SBC: return "AArch64ISD::SBC";
788 case AArch64ISD::ADDS: return "AArch64ISD::ADDS";
789 case AArch64ISD::SUBS: return "AArch64ISD::SUBS";
790 case AArch64ISD::ADCS: return "AArch64ISD::ADCS";
791 case AArch64ISD::SBCS: return "AArch64ISD::SBCS";
792 case AArch64ISD::ANDS: return "AArch64ISD::ANDS";
793 case AArch64ISD::FCMP: return "AArch64ISD::FCMP";
794 case AArch64ISD::FMIN: return "AArch64ISD::FMIN";
795 case AArch64ISD::FMAX: return "AArch64ISD::FMAX";
796 case AArch64ISD::DUP: return "AArch64ISD::DUP";
797 case AArch64ISD::DUPLANE8: return "AArch64ISD::DUPLANE8";
798 case AArch64ISD::DUPLANE16: return "AArch64ISD::DUPLANE16";
799 case AArch64ISD::DUPLANE32: return "AArch64ISD::DUPLANE32";
800 case AArch64ISD::DUPLANE64: return "AArch64ISD::DUPLANE64";
801 case AArch64ISD::MOVI: return "AArch64ISD::MOVI";
802 case AArch64ISD::MOVIshift: return "AArch64ISD::MOVIshift";
803 case AArch64ISD::MOVIedit: return "AArch64ISD::MOVIedit";
804 case AArch64ISD::MOVImsl: return "AArch64ISD::MOVImsl";
805 case AArch64ISD::FMOV: return "AArch64ISD::FMOV";
806 case AArch64ISD::MVNIshift: return "AArch64ISD::MVNIshift";
807 case AArch64ISD::MVNImsl: return "AArch64ISD::MVNImsl";
808 case AArch64ISD::BICi: return "AArch64ISD::BICi";
809 case AArch64ISD::ORRi: return "AArch64ISD::ORRi";
810 case AArch64ISD::BSL: return "AArch64ISD::BSL";
811 case AArch64ISD::NEG: return "AArch64ISD::NEG";
812 case AArch64ISD::EXTR: return "AArch64ISD::EXTR";
813 case AArch64ISD::ZIP1: return "AArch64ISD::ZIP1";
814 case AArch64ISD::ZIP2: return "AArch64ISD::ZIP2";
815 case AArch64ISD::UZP1: return "AArch64ISD::UZP1";
816 case AArch64ISD::UZP2: return "AArch64ISD::UZP2";
817 case AArch64ISD::TRN1: return "AArch64ISD::TRN1";
818 case AArch64ISD::TRN2: return "AArch64ISD::TRN2";
819 case AArch64ISD::REV16: return "AArch64ISD::REV16";
820 case AArch64ISD::REV32: return "AArch64ISD::REV32";
821 case AArch64ISD::REV64: return "AArch64ISD::REV64";
822 case AArch64ISD::EXT: return "AArch64ISD::EXT";
823 case AArch64ISD::VSHL: return "AArch64ISD::VSHL";
824 case AArch64ISD::VLSHR: return "AArch64ISD::VLSHR";
825 case AArch64ISD::VASHR: return "AArch64ISD::VASHR";
826 case AArch64ISD::CMEQ: return "AArch64ISD::CMEQ";
827 case AArch64ISD::CMGE: return "AArch64ISD::CMGE";
828 case AArch64ISD::CMGT: return "AArch64ISD::CMGT";
829 case AArch64ISD::CMHI: return "AArch64ISD::CMHI";
830 case AArch64ISD::CMHS: return "AArch64ISD::CMHS";
831 case AArch64ISD::FCMEQ: return "AArch64ISD::FCMEQ";
832 case AArch64ISD::FCMGE: return "AArch64ISD::FCMGE";
833 case AArch64ISD::FCMGT: return "AArch64ISD::FCMGT";
834 case AArch64ISD::CMEQz: return "AArch64ISD::CMEQz";
835 case AArch64ISD::CMGEz: return "AArch64ISD::CMGEz";
836 case AArch64ISD::CMGTz: return "AArch64ISD::CMGTz";
837 case AArch64ISD::CMLEz: return "AArch64ISD::CMLEz";
838 case AArch64ISD::CMLTz: return "AArch64ISD::CMLTz";
839 case AArch64ISD::FCMEQz: return "AArch64ISD::FCMEQz";
840 case AArch64ISD::FCMGEz: return "AArch64ISD::FCMGEz";
841 case AArch64ISD::FCMGTz: return "AArch64ISD::FCMGTz";
842 case AArch64ISD::FCMLEz: return "AArch64ISD::FCMLEz";
843 case AArch64ISD::FCMLTz: return "AArch64ISD::FCMLTz";
844 case AArch64ISD::SADDV: return "AArch64ISD::SADDV";
845 case AArch64ISD::UADDV: return "AArch64ISD::UADDV";
846 case AArch64ISD::SMINV: return "AArch64ISD::SMINV";
847 case AArch64ISD::UMINV: return "AArch64ISD::UMINV";
848 case AArch64ISD::SMAXV: return "AArch64ISD::SMAXV";
849 case AArch64ISD::UMAXV: return "AArch64ISD::UMAXV";
850 case AArch64ISD::NOT: return "AArch64ISD::NOT";
851 case AArch64ISD::BIT: return "AArch64ISD::BIT";
852 case AArch64ISD::CBZ: return "AArch64ISD::CBZ";
853 case AArch64ISD::CBNZ: return "AArch64ISD::CBNZ";
854 case AArch64ISD::TBZ: return "AArch64ISD::TBZ";
855 case AArch64ISD::TBNZ: return "AArch64ISD::TBNZ";
856 case AArch64ISD::TC_RETURN: return "AArch64ISD::TC_RETURN";
857 case AArch64ISD::SITOF: return "AArch64ISD::SITOF";
858 case AArch64ISD::UITOF: return "AArch64ISD::UITOF";
859 case AArch64ISD::NVCAST: return "AArch64ISD::NVCAST";
860 case AArch64ISD::SQSHL_I: return "AArch64ISD::SQSHL_I";
861 case AArch64ISD::UQSHL_I: return "AArch64ISD::UQSHL_I";
862 case AArch64ISD::SRSHR_I: return "AArch64ISD::SRSHR_I";
863 case AArch64ISD::URSHR_I: return "AArch64ISD::URSHR_I";
864 case AArch64ISD::SQSHLU_I: return "AArch64ISD::SQSHLU_I";
865 case AArch64ISD::WrapperLarge: return "AArch64ISD::WrapperLarge";
866 case AArch64ISD::LD2post: return "AArch64ISD::LD2post";
867 case AArch64ISD::LD3post: return "AArch64ISD::LD3post";
868 case AArch64ISD::LD4post: return "AArch64ISD::LD4post";
869 case AArch64ISD::ST2post: return "AArch64ISD::ST2post";
870 case AArch64ISD::ST3post: return "AArch64ISD::ST3post";
871 case AArch64ISD::ST4post: return "AArch64ISD::ST4post";
872 case AArch64ISD::LD1x2post: return "AArch64ISD::LD1x2post";
873 case AArch64ISD::LD1x3post: return "AArch64ISD::LD1x3post";
874 case AArch64ISD::LD1x4post: return "AArch64ISD::LD1x4post";
875 case AArch64ISD::ST1x2post: return "AArch64ISD::ST1x2post";
876 case AArch64ISD::ST1x3post: return "AArch64ISD::ST1x3post";
877 case AArch64ISD::ST1x4post: return "AArch64ISD::ST1x4post";
878 case AArch64ISD::LD1DUPpost: return "AArch64ISD::LD1DUPpost";
879 case AArch64ISD::LD2DUPpost: return "AArch64ISD::LD2DUPpost";
880 case AArch64ISD::LD3DUPpost: return "AArch64ISD::LD3DUPpost";
881 case AArch64ISD::LD4DUPpost: return "AArch64ISD::LD4DUPpost";
882 case AArch64ISD::LD1LANEpost: return "AArch64ISD::LD1LANEpost";
883 case AArch64ISD::LD2LANEpost: return "AArch64ISD::LD2LANEpost";
884 case AArch64ISD::LD3LANEpost: return "AArch64ISD::LD3LANEpost";
885 case AArch64ISD::LD4LANEpost: return "AArch64ISD::LD4LANEpost";
886 case AArch64ISD::ST2LANEpost: return "AArch64ISD::ST2LANEpost";
887 case AArch64ISD::ST3LANEpost: return "AArch64ISD::ST3LANEpost";
888 case AArch64ISD::ST4LANEpost: return "AArch64ISD::ST4LANEpost";
889 case AArch64ISD::SMULL: return "AArch64ISD::SMULL";
890 case AArch64ISD::UMULL: return "AArch64ISD::UMULL";
895 AArch64TargetLowering::EmitF128CSEL(MachineInstr *MI,
896 MachineBasicBlock *MBB) const {
897 // We materialise the F128CSEL pseudo-instruction as some control flow and a
901 // [... previous instrs leading to comparison ...]
907 // Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB]
909 MachineFunction *MF = MBB->getParent();
910 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
911 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
912 DebugLoc DL = MI->getDebugLoc();
913 MachineFunction::iterator It = MBB;
916 unsigned DestReg = MI->getOperand(0).getReg();
917 unsigned IfTrueReg = MI->getOperand(1).getReg();
918 unsigned IfFalseReg = MI->getOperand(2).getReg();
919 unsigned CondCode = MI->getOperand(3).getImm();
920 bool NZCVKilled = MI->getOperand(4).isKill();
922 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
923 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
924 MF->insert(It, TrueBB);
925 MF->insert(It, EndBB);
927 // Transfer rest of current basic-block to EndBB
928 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
930 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
932 BuildMI(MBB, DL, TII->get(AArch64::Bcc)).addImm(CondCode).addMBB(TrueBB);
933 BuildMI(MBB, DL, TII->get(AArch64::B)).addMBB(EndBB);
934 MBB->addSuccessor(TrueBB);
935 MBB->addSuccessor(EndBB);
937 // TrueBB falls through to the end.
938 TrueBB->addSuccessor(EndBB);
941 TrueBB->addLiveIn(AArch64::NZCV);
942 EndBB->addLiveIn(AArch64::NZCV);
945 BuildMI(*EndBB, EndBB->begin(), DL, TII->get(AArch64::PHI), DestReg)
951 MI->eraseFromParent();
956 AArch64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
957 MachineBasicBlock *BB) const {
958 switch (MI->getOpcode()) {
963 llvm_unreachable("Unexpected instruction for custom inserter!");
965 case AArch64::F128CSEL:
966 return EmitF128CSEL(MI, BB);
968 case TargetOpcode::STACKMAP:
969 case TargetOpcode::PATCHPOINT:
970 return emitPatchPoint(MI, BB);
974 //===----------------------------------------------------------------------===//
975 // AArch64 Lowering private implementation.
976 //===----------------------------------------------------------------------===//
978 //===----------------------------------------------------------------------===//
980 //===----------------------------------------------------------------------===//
982 /// changeIntCCToAArch64CC - Convert a DAG integer condition code to an AArch64
984 static AArch64CC::CondCode changeIntCCToAArch64CC(ISD::CondCode CC) {
987 llvm_unreachable("Unknown condition code!");
989 return AArch64CC::NE;
991 return AArch64CC::EQ;
993 return AArch64CC::GT;
995 return AArch64CC::GE;
997 return AArch64CC::LT;
999 return AArch64CC::LE;
1001 return AArch64CC::HI;
1003 return AArch64CC::HS;
1005 return AArch64CC::LO;
1007 return AArch64CC::LS;
1011 /// changeFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64 CC.
1012 static void changeFPCCToAArch64CC(ISD::CondCode CC,
1013 AArch64CC::CondCode &CondCode,
1014 AArch64CC::CondCode &CondCode2) {
1015 CondCode2 = AArch64CC::AL;
1018 llvm_unreachable("Unknown FP condition!");
1021 CondCode = AArch64CC::EQ;
1025 CondCode = AArch64CC::GT;
1029 CondCode = AArch64CC::GE;
1032 CondCode = AArch64CC::MI;
1035 CondCode = AArch64CC::LS;
1038 CondCode = AArch64CC::MI;
1039 CondCode2 = AArch64CC::GT;
1042 CondCode = AArch64CC::VC;
1045 CondCode = AArch64CC::VS;
1048 CondCode = AArch64CC::EQ;
1049 CondCode2 = AArch64CC::VS;
1052 CondCode = AArch64CC::HI;
1055 CondCode = AArch64CC::PL;
1059 CondCode = AArch64CC::LT;
1063 CondCode = AArch64CC::LE;
1067 CondCode = AArch64CC::NE;
1072 /// changeVectorFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64
1073 /// CC usable with the vector instructions. Fewer operations are available
1074 /// without a real NZCV register, so we have to use less efficient combinations
1075 /// to get the same effect.
1076 static void changeVectorFPCCToAArch64CC(ISD::CondCode CC,
1077 AArch64CC::CondCode &CondCode,
1078 AArch64CC::CondCode &CondCode2,
1083 // Mostly the scalar mappings work fine.
1084 changeFPCCToAArch64CC(CC, CondCode, CondCode2);
1087 Invert = true; // Fallthrough
1089 CondCode = AArch64CC::MI;
1090 CondCode2 = AArch64CC::GE;
1097 // All of the compare-mask comparisons are ordered, but we can switch
1098 // between the two by a double inversion. E.g. ULE == !OGT.
1100 changeFPCCToAArch64CC(getSetCCInverse(CC, false), CondCode, CondCode2);
1105 static bool isLegalArithImmed(uint64_t C) {
1106 // Matches AArch64DAGToDAGISel::SelectArithImmed().
1107 return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
1110 static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1111 SDLoc dl, SelectionDAG &DAG) {
1112 EVT VT = LHS.getValueType();
1114 if (VT.isFloatingPoint())
1115 return DAG.getNode(AArch64ISD::FCMP, dl, VT, LHS, RHS);
1117 // The CMP instruction is just an alias for SUBS, and representing it as
1118 // SUBS means that it's possible to get CSE with subtract operations.
1119 // A later phase can perform the optimization of setting the destination
1120 // register to WZR/XZR if it ends up being unused.
1121 unsigned Opcode = AArch64ISD::SUBS;
1123 if (RHS.getOpcode() == ISD::SUB && isa<ConstantSDNode>(RHS.getOperand(0)) &&
1124 cast<ConstantSDNode>(RHS.getOperand(0))->getZExtValue() == 0 &&
1125 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1126 // We'd like to combine a (CMP op1, (sub 0, op2) into a CMN instruction on
1127 // the grounds that "op1 - (-op2) == op1 + op2". However, the C and V flags
1128 // can be set differently by this operation. It comes down to whether
1129 // "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are then
1130 // everything is fine. If not then the optimization is wrong. Thus general
1131 // comparisons are only valid if op2 != 0.
1133 // So, finally, the only LLVM-native comparisons that don't mention C and V
1134 // are SETEQ and SETNE. They're the only ones we can safely use CMN for in
1135 // the absence of information about op2.
1136 Opcode = AArch64ISD::ADDS;
1137 RHS = RHS.getOperand(1);
1138 } else if (LHS.getOpcode() == ISD::AND && isa<ConstantSDNode>(RHS) &&
1139 cast<ConstantSDNode>(RHS)->getZExtValue() == 0 &&
1140 !isUnsignedIntSetCC(CC)) {
1141 // Similarly, (CMP (and X, Y), 0) can be implemented with a TST
1142 // (a.k.a. ANDS) except that the flags are only guaranteed to work for one
1143 // of the signed comparisons.
1144 Opcode = AArch64ISD::ANDS;
1145 RHS = LHS.getOperand(1);
1146 LHS = LHS.getOperand(0);
1149 return DAG.getNode(Opcode, dl, DAG.getVTList(VT, MVT::i32), LHS, RHS)
1153 static SDValue getAArch64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1154 SDValue &AArch64cc, SelectionDAG &DAG, SDLoc dl) {
1156 AArch64CC::CondCode AArch64CC;
1157 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1158 EVT VT = RHS.getValueType();
1159 uint64_t C = RHSC->getZExtValue();
1160 if (!isLegalArithImmed(C)) {
1161 // Constant does not fit, try adjusting it by one?
1167 if ((VT == MVT::i32 && C != 0x80000000 &&
1168 isLegalArithImmed((uint32_t)(C - 1))) ||
1169 (VT == MVT::i64 && C != 0x80000000ULL &&
1170 isLegalArithImmed(C - 1ULL))) {
1171 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1172 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1173 RHS = DAG.getConstant(C, VT);
1178 if ((VT == MVT::i32 && C != 0 &&
1179 isLegalArithImmed((uint32_t)(C - 1))) ||
1180 (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) {
1181 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1182 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1183 RHS = DAG.getConstant(C, VT);
1188 if ((VT == MVT::i32 && C != INT32_MAX &&
1189 isLegalArithImmed((uint32_t)(C + 1))) ||
1190 (VT == MVT::i64 && C != INT64_MAX &&
1191 isLegalArithImmed(C + 1ULL))) {
1192 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1193 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1194 RHS = DAG.getConstant(C, VT);
1199 if ((VT == MVT::i32 && C != UINT32_MAX &&
1200 isLegalArithImmed((uint32_t)(C + 1))) ||
1201 (VT == MVT::i64 && C != UINT64_MAX &&
1202 isLegalArithImmed(C + 1ULL))) {
1203 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1204 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1205 RHS = DAG.getConstant(C, VT);
1211 // The imm operand of ADDS is an unsigned immediate, in the range 0 to 4095.
1212 // For the i8 operand, the largest immediate is 255, so this can be easily
1213 // encoded in the compare instruction. For the i16 operand, however, the
1214 // largest immediate cannot be encoded in the compare.
1215 // Therefore, use a sign extending load and cmn to avoid materializing the -1
1216 // constant. For example,
1218 // ldrh w0, [x0, #0]
1221 // ldrsh w0, [x0, #0]
1223 // Fundamental, we're relying on the property that (zext LHS) == (zext RHS)
1224 // if and only if (sext LHS) == (sext RHS). The checks are in place to ensure
1225 // both the LHS and RHS are truely zero extended and to make sure the
1226 // transformation is profitable.
1227 if ((CC == ISD::SETEQ || CC == ISD::SETNE) && isa<ConstantSDNode>(RHS)) {
1228 if ((cast<ConstantSDNode>(RHS)->getZExtValue() >> 16 == 0) &&
1229 isa<LoadSDNode>(LHS)) {
1230 if (cast<LoadSDNode>(LHS)->getExtensionType() == ISD::ZEXTLOAD &&
1231 cast<LoadSDNode>(LHS)->getMemoryVT() == MVT::i16 &&
1232 LHS.getNode()->hasNUsesOfValue(1, 0)) {
1233 int16_t ValueofRHS = cast<ConstantSDNode>(RHS)->getZExtValue();
1234 if (ValueofRHS < 0 && isLegalArithImmed(-ValueofRHS)) {
1236 DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, LHS.getValueType(), LHS,
1237 DAG.getValueType(MVT::i16));
1238 Cmp = emitComparison(SExt,
1239 DAG.getConstant(ValueofRHS, RHS.getValueType()),
1241 AArch64CC = changeIntCCToAArch64CC(CC);
1242 AArch64cc = DAG.getConstant(AArch64CC, MVT::i32);
1248 Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
1249 AArch64CC = changeIntCCToAArch64CC(CC);
1250 AArch64cc = DAG.getConstant(AArch64CC, MVT::i32);
1254 static std::pair<SDValue, SDValue>
1255 getAArch64XALUOOp(AArch64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
1256 assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
1257 "Unsupported value type");
1258 SDValue Value, Overflow;
1260 SDValue LHS = Op.getOperand(0);
1261 SDValue RHS = Op.getOperand(1);
1263 switch (Op.getOpcode()) {
1265 llvm_unreachable("Unknown overflow instruction!");
1267 Opc = AArch64ISD::ADDS;
1271 Opc = AArch64ISD::ADDS;
1275 Opc = AArch64ISD::SUBS;
1279 Opc = AArch64ISD::SUBS;
1282 // Multiply needs a little bit extra work.
1286 bool IsSigned = Op.getOpcode() == ISD::SMULO;
1287 if (Op.getValueType() == MVT::i32) {
1288 unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1289 // For a 32 bit multiply with overflow check we want the instruction
1290 // selector to generate a widening multiply (SMADDL/UMADDL). For that we
1291 // need to generate the following pattern:
1292 // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b))
1293 LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
1294 RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
1295 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1296 SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul,
1297 DAG.getConstant(0, MVT::i64));
1298 // On AArch64 the upper 32 bits are always zero extended for a 32 bit
1299 // operation. We need to clear out the upper 32 bits, because we used a
1300 // widening multiply that wrote all 64 bits. In the end this should be a
1302 Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add);
1304 // The signed overflow check requires more than just a simple check for
1305 // any bit set in the upper 32 bits of the result. These bits could be
1306 // just the sign bits of a negative number. To perform the overflow
1307 // check we have to arithmetic shift right the 32nd bit of the result by
1308 // 31 bits. Then we compare the result to the upper 32 bits.
1309 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add,
1310 DAG.getConstant(32, MVT::i64));
1311 UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits);
1312 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value,
1313 DAG.getConstant(31, MVT::i64));
1314 // It is important that LowerBits is last, otherwise the arithmetic
1315 // shift will not be folded into the compare (SUBS).
1316 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32);
1317 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1320 // The overflow check for unsigned multiply is easy. We only need to
1321 // check if any of the upper 32 bits are set. This can be done with a
1322 // CMP (shifted register). For that we need to generate the following
1324 // (i64 AArch64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32)
1325 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
1326 DAG.getConstant(32, MVT::i64));
1327 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1329 DAG.getNode(AArch64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1330 UpperBits).getValue(1);
1334 assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
1335 // For the 64 bit multiply
1336 Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1338 SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
1339 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
1340 DAG.getConstant(63, MVT::i64));
1341 // It is important that LowerBits is last, otherwise the arithmetic
1342 // shift will not be folded into the compare (SUBS).
1343 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1344 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1347 SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
1348 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1350 DAG.getNode(AArch64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1351 UpperBits).getValue(1);
1358 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);
1360 // Emit the AArch64 operation with overflow check.
1361 Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
1362 Overflow = Value.getValue(1);
1364 return std::make_pair(Value, Overflow);
1367 SDValue AArch64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG,
1368 RTLIB::Libcall Call) const {
1369 SmallVector<SDValue, 2> Ops(Op->op_begin(), Op->op_end());
1370 return makeLibCall(DAG, Call, MVT::f128, &Ops[0], Ops.size(), false,
1374 static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) {
1375 SDValue Sel = Op.getOperand(0);
1376 SDValue Other = Op.getOperand(1);
1378 // If neither operand is a SELECT_CC, give up.
1379 if (Sel.getOpcode() != ISD::SELECT_CC)
1380 std::swap(Sel, Other);
1381 if (Sel.getOpcode() != ISD::SELECT_CC)
1384 // The folding we want to perform is:
1385 // (xor x, (select_cc a, b, cc, 0, -1) )
1387 // (csel x, (xor x, -1), cc ...)
1389 // The latter will get matched to a CSINV instruction.
1391 ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
1392 SDValue LHS = Sel.getOperand(0);
1393 SDValue RHS = Sel.getOperand(1);
1394 SDValue TVal = Sel.getOperand(2);
1395 SDValue FVal = Sel.getOperand(3);
1398 // FIXME: This could be generalized to non-integer comparisons.
1399 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
1402 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
1403 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
1405 // The the values aren't constants, this isn't the pattern we're looking for.
1406 if (!CFVal || !CTVal)
1409 // We can commute the SELECT_CC by inverting the condition. This
1410 // might be needed to make this fit into a CSINV pattern.
1411 if (CTVal->isAllOnesValue() && CFVal->isNullValue()) {
1412 std::swap(TVal, FVal);
1413 std::swap(CTVal, CFVal);
1414 CC = ISD::getSetCCInverse(CC, true);
1417 // If the constants line up, perform the transform!
1418 if (CTVal->isNullValue() && CFVal->isAllOnesValue()) {
1420 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
1423 TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other,
1424 DAG.getConstant(-1ULL, Other.getValueType()));
1426 return DAG.getNode(AArch64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal,
1433 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
1434 EVT VT = Op.getValueType();
1436 // Let legalize expand this if it isn't a legal type yet.
1437 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
1440 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
1443 bool ExtraOp = false;
1444 switch (Op.getOpcode()) {
1446 llvm_unreachable("Invalid code");
1448 Opc = AArch64ISD::ADDS;
1451 Opc = AArch64ISD::SUBS;
1454 Opc = AArch64ISD::ADCS;
1458 Opc = AArch64ISD::SBCS;
1464 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1));
1465 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1),
1469 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
1470 // Let legalize expand this if it isn't a legal type yet.
1471 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
1474 AArch64CC::CondCode CC;
1475 // The actual operation that sets the overflow or carry flag.
1476 SDValue Value, Overflow;
1477 std::tie(Value, Overflow) = getAArch64XALUOOp(CC, Op, DAG);
1479 // We use 0 and 1 as false and true values.
1480 SDValue TVal = DAG.getConstant(1, MVT::i32);
1481 SDValue FVal = DAG.getConstant(0, MVT::i32);
1483 // We use an inverted condition, because the conditional select is inverted
1484 // too. This will allow it to be selected to a single instruction:
1485 // CSINC Wd, WZR, WZR, invert(cond).
1486 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), MVT::i32);
1487 Overflow = DAG.getNode(AArch64ISD::CSEL, SDLoc(Op), MVT::i32, FVal, TVal,
1490 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
1491 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
1494 // Prefetch operands are:
1495 // 1: Address to prefetch
1497 // 3: int locality (0 = no locality ... 3 = extreme locality)
1498 // 4: bool isDataCache
1499 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
1501 unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
1502 unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
1503 unsigned IsData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
1505 bool IsStream = !Locality;
1506 // When the locality number is set
1508 // The front-end should have filtered out the out-of-range values
1509 assert(Locality <= 3 && "Prefetch locality out-of-range");
1510 // The locality degree is the opposite of the cache speed.
1511 // Put the number the other way around.
1512 // The encoding starts at 0 for level 1
1513 Locality = 3 - Locality;
1516 // built the mask value encoding the expected behavior.
1517 unsigned PrfOp = (IsWrite << 4) | // Load/Store bit
1518 (!IsData << 3) | // IsDataCache bit
1519 (Locality << 1) | // Cache level bits
1520 (unsigned)IsStream; // Stream bit
1521 return DAG.getNode(AArch64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
1522 DAG.getConstant(PrfOp, MVT::i32), Op.getOperand(1));
1525 SDValue AArch64TargetLowering::LowerFP_EXTEND(SDValue Op,
1526 SelectionDAG &DAG) const {
1527 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
1530 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
1532 return LowerF128Call(Op, DAG, LC);
1535 SDValue AArch64TargetLowering::LowerFP_ROUND(SDValue Op,
1536 SelectionDAG &DAG) const {
1537 if (Op.getOperand(0).getValueType() != MVT::f128) {
1538 // It's legal except when f128 is involved
1543 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
1545 // FP_ROUND node has a second operand indicating whether it is known to be
1546 // precise. That doesn't take part in the LibCall so we can't directly use
1548 SDValue SrcVal = Op.getOperand(0);
1549 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
1550 /*isSigned*/ false, SDLoc(Op)).first;
1553 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1554 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1555 // Any additional optimization in this function should be recorded
1556 // in the cost tables.
1557 EVT InVT = Op.getOperand(0).getValueType();
1558 EVT VT = Op.getValueType();
1560 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1563 DAG.getNode(Op.getOpcode(), dl, InVT.changeVectorElementTypeToInteger(),
1565 return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
1568 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1571 MVT::getVectorVT(MVT::getFloatingPointVT(VT.getScalarSizeInBits()),
1572 VT.getVectorNumElements());
1573 SDValue Ext = DAG.getNode(ISD::FP_EXTEND, dl, ExtVT, Op.getOperand(0));
1574 return DAG.getNode(Op.getOpcode(), dl, VT, Ext);
1577 // Type changing conversions are illegal.
1581 SDValue AArch64TargetLowering::LowerFP_TO_INT(SDValue Op,
1582 SelectionDAG &DAG) const {
1583 if (Op.getOperand(0).getValueType().isVector())
1584 return LowerVectorFP_TO_INT(Op, DAG);
1586 // f16 conversions are promoted to f32.
1587 if (Op.getOperand(0).getValueType() == MVT::f16) {
1590 Op.getOpcode(), dl, Op.getValueType(),
1591 DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, Op.getOperand(0)));
1594 if (Op.getOperand(0).getValueType() != MVT::f128) {
1595 // It's legal except when f128 is involved
1600 if (Op.getOpcode() == ISD::FP_TO_SINT)
1601 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
1603 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
1605 SmallVector<SDValue, 2> Ops(Op->op_begin(), Op->op_end());
1606 return makeLibCall(DAG, LC, Op.getValueType(), &Ops[0], Ops.size(), false,
1610 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1611 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1612 // Any additional optimization in this function should be recorded
1613 // in the cost tables.
1614 EVT VT = Op.getValueType();
1616 SDValue In = Op.getOperand(0);
1617 EVT InVT = In.getValueType();
1619 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1621 MVT::getVectorVT(MVT::getFloatingPointVT(InVT.getScalarSizeInBits()),
1622 InVT.getVectorNumElements());
1623 In = DAG.getNode(Op.getOpcode(), dl, CastVT, In);
1624 return DAG.getNode(ISD::FP_ROUND, dl, VT, In, DAG.getIntPtrConstant(0));
1627 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1629 Op.getOpcode() == ISD::SINT_TO_FP ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1630 EVT CastVT = VT.changeVectorElementTypeToInteger();
1631 In = DAG.getNode(CastOpc, dl, CastVT, In);
1632 return DAG.getNode(Op.getOpcode(), dl, VT, In);
1638 SDValue AArch64TargetLowering::LowerINT_TO_FP(SDValue Op,
1639 SelectionDAG &DAG) const {
1640 if (Op.getValueType().isVector())
1641 return LowerVectorINT_TO_FP(Op, DAG);
1643 // f16 conversions are promoted to f32.
1644 if (Op.getValueType() == MVT::f16) {
1647 ISD::FP_ROUND, dl, MVT::f16,
1648 DAG.getNode(Op.getOpcode(), dl, MVT::f32, Op.getOperand(0)),
1649 DAG.getIntPtrConstant(0));
1652 // i128 conversions are libcalls.
1653 if (Op.getOperand(0).getValueType() == MVT::i128)
1656 // Other conversions are legal, unless it's to the completely software-based
1658 if (Op.getValueType() != MVT::f128)
1662 if (Op.getOpcode() == ISD::SINT_TO_FP)
1663 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1665 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1667 return LowerF128Call(Op, DAG, LC);
1670 SDValue AArch64TargetLowering::LowerFSINCOS(SDValue Op,
1671 SelectionDAG &DAG) const {
1672 // For iOS, we want to call an alternative entry point: __sincos_stret,
1673 // which returns the values in two S / D registers.
1675 SDValue Arg = Op.getOperand(0);
1676 EVT ArgVT = Arg.getValueType();
1677 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1684 Entry.isSExt = false;
1685 Entry.isZExt = false;
1686 Args.push_back(Entry);
1688 const char *LibcallName =
1689 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
1690 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
1692 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
1693 TargetLowering::CallLoweringInfo CLI(DAG);
1694 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
1695 .setCallee(CallingConv::Fast, RetTy, Callee, std::move(Args), 0);
1697 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1698 return CallResult.first;
1701 static SDValue LowerBITCAST(SDValue Op, SelectionDAG &DAG) {
1702 if (Op.getValueType() != MVT::f16)
1705 assert(Op.getOperand(0).getValueType() == MVT::i16);
1708 Op = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op.getOperand(0));
1709 Op = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Op);
1711 DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL, MVT::f16, Op,
1712 DAG.getTargetConstant(AArch64::hsub, MVT::i32)),
1716 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
1717 if (OrigVT.getSizeInBits() >= 64)
1720 assert(OrigVT.isSimple() && "Expecting a simple value type");
1722 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
1723 switch (OrigSimpleTy) {
1724 default: llvm_unreachable("Unexpected Vector Type");
1733 static SDValue addRequiredExtensionForVectorMULL(SDValue N, SelectionDAG &DAG,
1736 unsigned ExtOpcode) {
1737 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
1738 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
1739 // 64-bits we need to insert a new extension so that it will be 64-bits.
1740 assert(ExtTy.is128BitVector() && "Unexpected extension size");
1741 if (OrigTy.getSizeInBits() >= 64)
1744 // Must extend size to at least 64 bits to be used as an operand for VMULL.
1745 EVT NewVT = getExtensionTo64Bits(OrigTy);
1747 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
1750 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
1752 EVT VT = N->getValueType(0);
1754 if (N->getOpcode() != ISD::BUILD_VECTOR)
1757 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1758 SDNode *Elt = N->getOperand(i).getNode();
1759 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
1760 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1761 unsigned HalfSize = EltSize / 2;
1763 if (!isIntN(HalfSize, C->getSExtValue()))
1766 if (!isUIntN(HalfSize, C->getZExtValue()))
1777 static SDValue skipExtensionForVectorMULL(SDNode *N, SelectionDAG &DAG) {
1778 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
1779 return addRequiredExtensionForVectorMULL(N->getOperand(0), DAG,
1780 N->getOperand(0)->getValueType(0),
1784 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
1785 EVT VT = N->getValueType(0);
1786 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
1787 unsigned NumElts = VT.getVectorNumElements();
1788 MVT TruncVT = MVT::getIntegerVT(EltSize);
1789 SmallVector<SDValue, 8> Ops;
1790 for (unsigned i = 0; i != NumElts; ++i) {
1791 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
1792 const APInt &CInt = C->getAPIntValue();
1793 // Element types smaller than 32 bits are not legal, so use i32 elements.
1794 // The values are implicitly truncated so sext vs. zext doesn't matter.
1795 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
1797 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
1798 MVT::getVectorVT(TruncVT, NumElts), Ops);
1801 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
1802 if (N->getOpcode() == ISD::SIGN_EXTEND)
1804 if (isExtendedBUILD_VECTOR(N, DAG, true))
1809 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
1810 if (N->getOpcode() == ISD::ZERO_EXTEND)
1812 if (isExtendedBUILD_VECTOR(N, DAG, false))
1817 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
1818 unsigned Opcode = N->getOpcode();
1819 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
1820 SDNode *N0 = N->getOperand(0).getNode();
1821 SDNode *N1 = N->getOperand(1).getNode();
1822 return N0->hasOneUse() && N1->hasOneUse() &&
1823 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
1828 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
1829 unsigned Opcode = N->getOpcode();
1830 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
1831 SDNode *N0 = N->getOperand(0).getNode();
1832 SDNode *N1 = N->getOperand(1).getNode();
1833 return N0->hasOneUse() && N1->hasOneUse() &&
1834 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
1839 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
1840 // Multiplications are only custom-lowered for 128-bit vectors so that
1841 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
1842 EVT VT = Op.getValueType();
1843 assert(VT.is128BitVector() && VT.isInteger() &&
1844 "unexpected type for custom-lowering ISD::MUL");
1845 SDNode *N0 = Op.getOperand(0).getNode();
1846 SDNode *N1 = Op.getOperand(1).getNode();
1847 unsigned NewOpc = 0;
1849 bool isN0SExt = isSignExtended(N0, DAG);
1850 bool isN1SExt = isSignExtended(N1, DAG);
1851 if (isN0SExt && isN1SExt)
1852 NewOpc = AArch64ISD::SMULL;
1854 bool isN0ZExt = isZeroExtended(N0, DAG);
1855 bool isN1ZExt = isZeroExtended(N1, DAG);
1856 if (isN0ZExt && isN1ZExt)
1857 NewOpc = AArch64ISD::UMULL;
1858 else if (isN1SExt || isN1ZExt) {
1859 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
1860 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
1861 if (isN1SExt && isAddSubSExt(N0, DAG)) {
1862 NewOpc = AArch64ISD::SMULL;
1864 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
1865 NewOpc = AArch64ISD::UMULL;
1867 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
1869 NewOpc = AArch64ISD::UMULL;
1875 if (VT == MVT::v2i64)
1876 // Fall through to expand this. It is not legal.
1879 // Other vector multiplications are legal.
1884 // Legalize to a S/UMULL instruction
1887 SDValue Op1 = skipExtensionForVectorMULL(N1, DAG);
1889 Op0 = skipExtensionForVectorMULL(N0, DAG);
1890 assert(Op0.getValueType().is64BitVector() &&
1891 Op1.getValueType().is64BitVector() &&
1892 "unexpected types for extended operands to VMULL");
1893 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
1895 // Optimizing (zext A + zext B) * C, to (S/UMULL A, C) + (S/UMULL B, C) during
1896 // isel lowering to take advantage of no-stall back to back s/umul + s/umla.
1897 // This is true for CPUs with accumulate forwarding such as Cortex-A53/A57
1898 SDValue N00 = skipExtensionForVectorMULL(N0->getOperand(0).getNode(), DAG);
1899 SDValue N01 = skipExtensionForVectorMULL(N0->getOperand(1).getNode(), DAG);
1900 EVT Op1VT = Op1.getValueType();
1901 return DAG.getNode(N0->getOpcode(), DL, VT,
1902 DAG.getNode(NewOpc, DL, VT,
1903 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
1904 DAG.getNode(NewOpc, DL, VT,
1905 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
1908 SDValue AArch64TargetLowering::LowerOperation(SDValue Op,
1909 SelectionDAG &DAG) const {
1910 switch (Op.getOpcode()) {
1912 llvm_unreachable("unimplemented operand");
1915 return LowerBITCAST(Op, DAG);
1916 case ISD::GlobalAddress:
1917 return LowerGlobalAddress(Op, DAG);
1918 case ISD::GlobalTLSAddress:
1919 return LowerGlobalTLSAddress(Op, DAG);
1921 return LowerSETCC(Op, DAG);
1923 return LowerBR_CC(Op, DAG);
1925 return LowerSELECT(Op, DAG);
1926 case ISD::SELECT_CC:
1927 return LowerSELECT_CC(Op, DAG);
1928 case ISD::JumpTable:
1929 return LowerJumpTable(Op, DAG);
1930 case ISD::ConstantPool:
1931 return LowerConstantPool(Op, DAG);
1932 case ISD::BlockAddress:
1933 return LowerBlockAddress(Op, DAG);
1935 return LowerVASTART(Op, DAG);
1937 return LowerVACOPY(Op, DAG);
1939 return LowerVAARG(Op, DAG);
1944 return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
1951 return LowerXALUO(Op, DAG);
1953 return LowerF128Call(Op, DAG, RTLIB::ADD_F128);
1955 return LowerF128Call(Op, DAG, RTLIB::SUB_F128);
1957 return LowerF128Call(Op, DAG, RTLIB::MUL_F128);
1959 return LowerF128Call(Op, DAG, RTLIB::DIV_F128);
1961 return LowerFP_ROUND(Op, DAG);
1962 case ISD::FP_EXTEND:
1963 return LowerFP_EXTEND(Op, DAG);
1964 case ISD::FRAMEADDR:
1965 return LowerFRAMEADDR(Op, DAG);
1966 case ISD::RETURNADDR:
1967 return LowerRETURNADDR(Op, DAG);
1968 case ISD::INSERT_VECTOR_ELT:
1969 return LowerINSERT_VECTOR_ELT(Op, DAG);
1970 case ISD::EXTRACT_VECTOR_ELT:
1971 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
1972 case ISD::BUILD_VECTOR:
1973 return LowerBUILD_VECTOR(Op, DAG);
1974 case ISD::VECTOR_SHUFFLE:
1975 return LowerVECTOR_SHUFFLE(Op, DAG);
1976 case ISD::EXTRACT_SUBVECTOR:
1977 return LowerEXTRACT_SUBVECTOR(Op, DAG);
1981 return LowerVectorSRA_SRL_SHL(Op, DAG);
1982 case ISD::SHL_PARTS:
1983 return LowerShiftLeftParts(Op, DAG);
1984 case ISD::SRL_PARTS:
1985 case ISD::SRA_PARTS:
1986 return LowerShiftRightParts(Op, DAG);
1988 return LowerCTPOP(Op, DAG);
1989 case ISD::FCOPYSIGN:
1990 return LowerFCOPYSIGN(Op, DAG);
1992 return LowerVectorAND(Op, DAG);
1994 return LowerVectorOR(Op, DAG);
1996 return LowerXOR(Op, DAG);
1998 return LowerPREFETCH(Op, DAG);
1999 case ISD::SINT_TO_FP:
2000 case ISD::UINT_TO_FP:
2001 return LowerINT_TO_FP(Op, DAG);
2002 case ISD::FP_TO_SINT:
2003 case ISD::FP_TO_UINT:
2004 return LowerFP_TO_INT(Op, DAG);
2006 return LowerFSINCOS(Op, DAG);
2008 return LowerMUL(Op, DAG);
2012 /// getFunctionAlignment - Return the Log2 alignment of this function.
2013 unsigned AArch64TargetLowering::getFunctionAlignment(const Function *F) const {
2017 //===----------------------------------------------------------------------===//
2018 // Calling Convention Implementation
2019 //===----------------------------------------------------------------------===//
2021 #include "AArch64GenCallingConv.inc"
2023 /// Selects the correct CCAssignFn for a given CallingConvention value.
2024 CCAssignFn *AArch64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
2025 bool IsVarArg) const {
2028 llvm_unreachable("Unsupported calling convention.");
2029 case CallingConv::WebKit_JS:
2030 return CC_AArch64_WebKit_JS;
2031 case CallingConv::GHC:
2032 return CC_AArch64_GHC;
2033 case CallingConv::C:
2034 case CallingConv::Fast:
2035 if (!Subtarget->isTargetDarwin())
2036 return CC_AArch64_AAPCS;
2037 return IsVarArg ? CC_AArch64_DarwinPCS_VarArg : CC_AArch64_DarwinPCS;
2041 SDValue AArch64TargetLowering::LowerFormalArguments(
2042 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2043 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2044 SmallVectorImpl<SDValue> &InVals) const {
2045 MachineFunction &MF = DAG.getMachineFunction();
2046 MachineFrameInfo *MFI = MF.getFrameInfo();
2048 // Assign locations to all of the incoming arguments.
2049 SmallVector<CCValAssign, 16> ArgLocs;
2050 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
2053 // At this point, Ins[].VT may already be promoted to i32. To correctly
2054 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2055 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
2056 // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
2057 // we use a special version of AnalyzeFormalArguments to pass in ValVT and
2059 unsigned NumArgs = Ins.size();
2060 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2061 unsigned CurArgIdx = 0;
2062 for (unsigned i = 0; i != NumArgs; ++i) {
2063 MVT ValVT = Ins[i].VT;
2064 if (Ins[i].isOrigArg()) {
2065 std::advance(CurOrigArg, Ins[i].getOrigArgIndex() - CurArgIdx);
2066 CurArgIdx = Ins[i].getOrigArgIndex();
2068 // Get type of the original argument.
2069 EVT ActualVT = getValueType(CurOrigArg->getType(), /*AllowUnknown*/ true);
2070 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
2071 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2072 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2074 else if (ActualMVT == MVT::i16)
2077 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2079 AssignFn(i, ValVT, ValVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
2080 assert(!Res && "Call operand has unhandled type");
2083 assert(ArgLocs.size() == Ins.size());
2084 SmallVector<SDValue, 16> ArgValues;
2085 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2086 CCValAssign &VA = ArgLocs[i];
2088 if (Ins[i].Flags.isByVal()) {
2089 // Byval is used for HFAs in the PCS, but the system should work in a
2090 // non-compliant manner for larger structs.
2091 EVT PtrTy = getPointerTy();
2092 int Size = Ins[i].Flags.getByValSize();
2093 unsigned NumRegs = (Size + 7) / 8;
2095 // FIXME: This works on big-endian for composite byvals, which are the common
2096 // case. It should also work for fundamental types too.
2098 MFI->CreateFixedObject(8 * NumRegs, VA.getLocMemOffset(), false);
2099 SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrTy);
2100 InVals.push_back(FrameIdxN);
2105 if (VA.isRegLoc()) {
2106 // Arguments stored in registers.
2107 EVT RegVT = VA.getLocVT();
2110 const TargetRegisterClass *RC;
2112 if (RegVT == MVT::i32)
2113 RC = &AArch64::GPR32RegClass;
2114 else if (RegVT == MVT::i64)
2115 RC = &AArch64::GPR64RegClass;
2116 else if (RegVT == MVT::f16)
2117 RC = &AArch64::FPR16RegClass;
2118 else if (RegVT == MVT::f32)
2119 RC = &AArch64::FPR32RegClass;
2120 else if (RegVT == MVT::f64 || RegVT.is64BitVector())
2121 RC = &AArch64::FPR64RegClass;
2122 else if (RegVT == MVT::f128 || RegVT.is128BitVector())
2123 RC = &AArch64::FPR128RegClass;
2125 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2127 // Transform the arguments in physical registers into virtual ones.
2128 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2129 ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2131 // If this is an 8, 16 or 32-bit value, it is really passed promoted
2132 // to 64 bits. Insert an assert[sz]ext to capture this, then
2133 // truncate to the right size.
2134 switch (VA.getLocInfo()) {
2136 llvm_unreachable("Unknown loc info!");
2137 case CCValAssign::Full:
2139 case CCValAssign::BCvt:
2140 ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue);
2142 case CCValAssign::AExt:
2143 case CCValAssign::SExt:
2144 case CCValAssign::ZExt:
2145 // SelectionDAGBuilder will insert appropriate AssertZExt & AssertSExt
2146 // nodes after our lowering.
2147 assert(RegVT == Ins[i].VT && "incorrect register location selected");
2151 InVals.push_back(ArgValue);
2153 } else { // VA.isRegLoc()
2154 assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
2155 unsigned ArgOffset = VA.getLocMemOffset();
2156 unsigned ArgSize = VA.getValVT().getSizeInBits() / 8;
2158 uint32_t BEAlign = 0;
2159 if (!Subtarget->isLittleEndian() && ArgSize < 8 &&
2160 !Ins[i].Flags.isInConsecutiveRegs())
2161 BEAlign = 8 - ArgSize;
2163 int FI = MFI->CreateFixedObject(ArgSize, ArgOffset + BEAlign, true);
2165 // Create load nodes to retrieve arguments from the stack.
2166 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2169 // For NON_EXTLOAD, generic code in getLoad assert(ValVT == MemVT)
2170 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
2171 MVT MemVT = VA.getValVT();
2173 switch (VA.getLocInfo()) {
2176 case CCValAssign::BCvt:
2177 MemVT = VA.getLocVT();
2179 case CCValAssign::SExt:
2180 ExtType = ISD::SEXTLOAD;
2182 case CCValAssign::ZExt:
2183 ExtType = ISD::ZEXTLOAD;
2185 case CCValAssign::AExt:
2186 ExtType = ISD::EXTLOAD;
2190 ArgValue = DAG.getExtLoad(ExtType, DL, VA.getLocVT(), Chain, FIN,
2191 MachinePointerInfo::getFixedStack(FI),
2192 MemVT, false, false, false, 0);
2194 InVals.push_back(ArgValue);
2200 if (!Subtarget->isTargetDarwin()) {
2201 // The AAPCS variadic function ABI is identical to the non-variadic
2202 // one. As a result there may be more arguments in registers and we should
2203 // save them for future reference.
2204 saveVarArgRegisters(CCInfo, DAG, DL, Chain);
2207 AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>();
2208 // This will point to the next argument passed via stack.
2209 unsigned StackOffset = CCInfo.getNextStackOffset();
2210 // We currently pass all varargs at 8-byte alignment.
2211 StackOffset = ((StackOffset + 7) & ~7);
2212 AFI->setVarArgsStackIndex(MFI->CreateFixedObject(4, StackOffset, true));
2215 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2216 unsigned StackArgSize = CCInfo.getNextStackOffset();
2217 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2218 if (DoesCalleeRestoreStack(CallConv, TailCallOpt)) {
2219 // This is a non-standard ABI so by fiat I say we're allowed to make full
2220 // use of the stack area to be popped, which must be aligned to 16 bytes in
2222 StackArgSize = RoundUpToAlignment(StackArgSize, 16);
2224 // If we're expected to restore the stack (e.g. fastcc) then we'll be adding
2225 // a multiple of 16.
2226 FuncInfo->setArgumentStackToRestore(StackArgSize);
2228 // This realignment carries over to the available bytes below. Our own
2229 // callers will guarantee the space is free by giving an aligned value to
2232 // Even if we're not expected to free up the space, it's useful to know how
2233 // much is there while considering tail calls (because we can reuse it).
2234 FuncInfo->setBytesInStackArgArea(StackArgSize);
2239 void AArch64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
2240 SelectionDAG &DAG, SDLoc DL,
2241 SDValue &Chain) const {
2242 MachineFunction &MF = DAG.getMachineFunction();
2243 MachineFrameInfo *MFI = MF.getFrameInfo();
2244 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2246 SmallVector<SDValue, 8> MemOps;
2248 static const MCPhysReg GPRArgRegs[] = { AArch64::X0, AArch64::X1, AArch64::X2,
2249 AArch64::X3, AArch64::X4, AArch64::X5,
2250 AArch64::X6, AArch64::X7 };
2251 static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs);
2252 unsigned FirstVariadicGPR = CCInfo.getFirstUnallocated(GPRArgRegs);
2254 unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
2256 if (GPRSaveSize != 0) {
2257 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
2259 SDValue FIN = DAG.getFrameIndex(GPRIdx, getPointerTy());
2261 for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
2262 unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &AArch64::GPR64RegClass);
2263 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
2265 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2266 MachinePointerInfo::getStack(i * 8), false, false, 0);
2267 MemOps.push_back(Store);
2268 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
2269 DAG.getConstant(8, getPointerTy()));
2272 FuncInfo->setVarArgsGPRIndex(GPRIdx);
2273 FuncInfo->setVarArgsGPRSize(GPRSaveSize);
2275 if (Subtarget->hasFPARMv8()) {
2276 static const MCPhysReg FPRArgRegs[] = {
2277 AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3,
2278 AArch64::Q4, AArch64::Q5, AArch64::Q6, AArch64::Q7};
2279 static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs);
2280 unsigned FirstVariadicFPR = CCInfo.getFirstUnallocated(FPRArgRegs);
2282 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
2284 if (FPRSaveSize != 0) {
2285 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
2287 SDValue FIN = DAG.getFrameIndex(FPRIdx, getPointerTy());
2289 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
2290 unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &AArch64::FPR128RegClass);
2291 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f128);
2294 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2295 MachinePointerInfo::getStack(i * 16), false, false, 0);
2296 MemOps.push_back(Store);
2297 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
2298 DAG.getConstant(16, getPointerTy()));
2301 FuncInfo->setVarArgsFPRIndex(FPRIdx);
2302 FuncInfo->setVarArgsFPRSize(FPRSaveSize);
2305 if (!MemOps.empty()) {
2306 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
2310 /// LowerCallResult - Lower the result values of a call into the
2311 /// appropriate copies out of appropriate physical registers.
2312 SDValue AArch64TargetLowering::LowerCallResult(
2313 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
2314 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2315 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
2316 SDValue ThisVal) const {
2317 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2318 ? RetCC_AArch64_WebKit_JS
2319 : RetCC_AArch64_AAPCS;
2320 // Assign locations to each value returned by this call.
2321 SmallVector<CCValAssign, 16> RVLocs;
2322 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2324 CCInfo.AnalyzeCallResult(Ins, RetCC);
2326 // Copy all of the result registers out of their specified physreg.
2327 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2328 CCValAssign VA = RVLocs[i];
2330 // Pass 'this' value directly from the argument to return value, to avoid
2331 // reg unit interference
2332 if (i == 0 && isThisReturn) {
2333 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 &&
2334 "unexpected return calling convention register assignment");
2335 InVals.push_back(ThisVal);
2340 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag);
2341 Chain = Val.getValue(1);
2342 InFlag = Val.getValue(2);
2344 switch (VA.getLocInfo()) {
2346 llvm_unreachable("Unknown loc info!");
2347 case CCValAssign::Full:
2349 case CCValAssign::BCvt:
2350 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
2354 InVals.push_back(Val);
2360 bool AArch64TargetLowering::isEligibleForTailCallOptimization(
2361 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
2362 bool isCalleeStructRet, bool isCallerStructRet,
2363 const SmallVectorImpl<ISD::OutputArg> &Outs,
2364 const SmallVectorImpl<SDValue> &OutVals,
2365 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
2366 // For CallingConv::C this function knows whether the ABI needs
2367 // changing. That's not true for other conventions so they will have to opt in
2369 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
2372 const MachineFunction &MF = DAG.getMachineFunction();
2373 const Function *CallerF = MF.getFunction();
2374 CallingConv::ID CallerCC = CallerF->getCallingConv();
2375 bool CCMatch = CallerCC == CalleeCC;
2377 // Byval parameters hand the function a pointer directly into the stack area
2378 // we want to reuse during a tail call. Working around this *is* possible (see
2379 // X86) but less efficient and uglier in LowerCall.
2380 for (Function::const_arg_iterator i = CallerF->arg_begin(),
2381 e = CallerF->arg_end();
2383 if (i->hasByValAttr())
2386 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
2387 if (IsTailCallConvention(CalleeCC) && CCMatch)
2392 // Externally-defined functions with weak linkage should not be
2393 // tail-called on AArch64 when the OS does not support dynamic
2394 // pre-emption of symbols, as the AAELF spec requires normal calls
2395 // to undefined weak functions to be replaced with a NOP or jump to the
2396 // next instruction. The behaviour of branch instructions in this
2397 // situation (as used for tail calls) is implementation-defined, so we
2398 // cannot rely on the linker replacing the tail call with a return.
2399 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2400 const GlobalValue *GV = G->getGlobal();
2401 const Triple TT(getTargetMachine().getTargetTriple());
2402 if (GV->hasExternalWeakLinkage() &&
2403 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2407 // Now we search for cases where we can use a tail call without changing the
2408 // ABI. Sibcall is used in some places (particularly gcc) to refer to this
2411 // I want anyone implementing a new calling convention to think long and hard
2412 // about this assert.
2413 assert((!isVarArg || CalleeCC == CallingConv::C) &&
2414 "Unexpected variadic calling convention");
2416 if (isVarArg && !Outs.empty()) {
2417 // At least two cases here: if caller is fastcc then we can't have any
2418 // memory arguments (we'd be expected to clean up the stack afterwards). If
2419 // caller is C then we could potentially use its argument area.
2421 // FIXME: for now we take the most conservative of these in both cases:
2422 // disallow all variadic memory operands.
2423 SmallVector<CCValAssign, 16> ArgLocs;
2424 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2427 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, true));
2428 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
2429 if (!ArgLocs[i].isRegLoc())
2433 // If the calling conventions do not match, then we'd better make sure the
2434 // results are returned in the same way as what the caller expects.
2436 SmallVector<CCValAssign, 16> RVLocs1;
2437 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2439 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForCall(CalleeCC, isVarArg));
2441 SmallVector<CCValAssign, 16> RVLocs2;
2442 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2444 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForCall(CallerCC, isVarArg));
2446 if (RVLocs1.size() != RVLocs2.size())
2448 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2449 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2451 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2453 if (RVLocs1[i].isRegLoc()) {
2454 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2457 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2463 // Nothing more to check if the callee is taking no arguments
2467 SmallVector<CCValAssign, 16> ArgLocs;
2468 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2471 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, isVarArg));
2473 const AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2475 // If the stack arguments for this call would fit into our own save area then
2476 // the call can be made tail.
2477 return CCInfo.getNextStackOffset() <= FuncInfo->getBytesInStackArgArea();
2480 SDValue AArch64TargetLowering::addTokenForArgument(SDValue Chain,
2482 MachineFrameInfo *MFI,
2483 int ClobberedFI) const {
2484 SmallVector<SDValue, 8> ArgChains;
2485 int64_t FirstByte = MFI->getObjectOffset(ClobberedFI);
2486 int64_t LastByte = FirstByte + MFI->getObjectSize(ClobberedFI) - 1;
2488 // Include the original chain at the beginning of the list. When this is
2489 // used by target LowerCall hooks, this helps legalize find the
2490 // CALLSEQ_BEGIN node.
2491 ArgChains.push_back(Chain);
2493 // Add a chain value for each stack argument corresponding
2494 for (SDNode::use_iterator U = DAG.getEntryNode().getNode()->use_begin(),
2495 UE = DAG.getEntryNode().getNode()->use_end();
2497 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
2498 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
2499 if (FI->getIndex() < 0) {
2500 int64_t InFirstByte = MFI->getObjectOffset(FI->getIndex());
2501 int64_t InLastByte = InFirstByte;
2502 InLastByte += MFI->getObjectSize(FI->getIndex()) - 1;
2504 if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
2505 (FirstByte <= InFirstByte && InFirstByte <= LastByte))
2506 ArgChains.push_back(SDValue(L, 1));
2509 // Build a tokenfactor for all the chains.
2510 return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
2513 bool AArch64TargetLowering::DoesCalleeRestoreStack(CallingConv::ID CallCC,
2514 bool TailCallOpt) const {
2515 return CallCC == CallingConv::Fast && TailCallOpt;
2518 bool AArch64TargetLowering::IsTailCallConvention(CallingConv::ID CallCC) const {
2519 return CallCC == CallingConv::Fast;
2522 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
2523 /// and add input and output parameter nodes.
2525 AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
2526 SmallVectorImpl<SDValue> &InVals) const {
2527 SelectionDAG &DAG = CLI.DAG;
2529 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2530 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2531 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2532 SDValue Chain = CLI.Chain;
2533 SDValue Callee = CLI.Callee;
2534 bool &IsTailCall = CLI.IsTailCall;
2535 CallingConv::ID CallConv = CLI.CallConv;
2536 bool IsVarArg = CLI.IsVarArg;
2538 MachineFunction &MF = DAG.getMachineFunction();
2539 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
2540 bool IsThisReturn = false;
2542 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2543 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2544 bool IsSibCall = false;
2547 // Check if it's really possible to do a tail call.
2548 IsTailCall = isEligibleForTailCallOptimization(
2549 Callee, CallConv, IsVarArg, IsStructRet,
2550 MF.getFunction()->hasStructRetAttr(), Outs, OutVals, Ins, DAG);
2551 if (!IsTailCall && CLI.CS && CLI.CS->isMustTailCall())
2552 report_fatal_error("failed to perform tail call elimination on a call "
2553 "site marked musttail");
2555 // A sibling call is one where we're under the usual C ABI and not planning
2556 // to change that but can still do a tail call:
2557 if (!TailCallOpt && IsTailCall)
2564 // Analyze operands of the call, assigning locations to each operand.
2565 SmallVector<CCValAssign, 16> ArgLocs;
2566 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
2570 // Handle fixed and variable vector arguments differently.
2571 // Variable vector arguments always go into memory.
2572 unsigned NumArgs = Outs.size();
2574 for (unsigned i = 0; i != NumArgs; ++i) {
2575 MVT ArgVT = Outs[i].VT;
2576 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2577 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
2578 /*IsVarArg=*/ !Outs[i].IsFixed);
2579 bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
2580 assert(!Res && "Call operand has unhandled type");
2584 // At this point, Outs[].VT may already be promoted to i32. To correctly
2585 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2586 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
2587 // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here
2588 // we use a special version of AnalyzeCallOperands to pass in ValVT and
2590 unsigned NumArgs = Outs.size();
2591 for (unsigned i = 0; i != NumArgs; ++i) {
2592 MVT ValVT = Outs[i].VT;
2593 // Get type of the original argument.
2594 EVT ActualVT = getValueType(CLI.getArgs()[Outs[i].OrigArgIndex].Ty,
2595 /*AllowUnknown*/ true);
2596 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT;
2597 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2598 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2599 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2601 else if (ActualMVT == MVT::i16)
2604 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2605 bool Res = AssignFn(i, ValVT, ValVT, CCValAssign::Full, ArgFlags, CCInfo);
2606 assert(!Res && "Call operand has unhandled type");
2611 // Get a count of how many bytes are to be pushed on the stack.
2612 unsigned NumBytes = CCInfo.getNextStackOffset();
2615 // Since we're not changing the ABI to make this a tail call, the memory
2616 // operands are already available in the caller's incoming argument space.
2620 // FPDiff is the byte offset of the call's argument area from the callee's.
2621 // Stores to callee stack arguments will be placed in FixedStackSlots offset
2622 // by this amount for a tail call. In a sibling call it must be 0 because the
2623 // caller will deallocate the entire stack and the callee still expects its
2624 // arguments to begin at SP+0. Completely unused for non-tail calls.
2627 if (IsTailCall && !IsSibCall) {
2628 unsigned NumReusableBytes = FuncInfo->getBytesInStackArgArea();
2630 // Since callee will pop argument stack as a tail call, we must keep the
2631 // popped size 16-byte aligned.
2632 NumBytes = RoundUpToAlignment(NumBytes, 16);
2634 // FPDiff will be negative if this tail call requires more space than we
2635 // would automatically have in our incoming argument space. Positive if we
2636 // can actually shrink the stack.
2637 FPDiff = NumReusableBytes - NumBytes;
2639 // The stack pointer must be 16-byte aligned at all times it's used for a
2640 // memory operation, which in practice means at *all* times and in
2641 // particular across call boundaries. Therefore our own arguments started at
2642 // a 16-byte aligned SP and the delta applied for the tail call should
2643 // satisfy the same constraint.
2644 assert(FPDiff % 16 == 0 && "unaligned stack on tail call");
2647 // Adjust the stack pointer for the new arguments...
2648 // These operations are automatically eliminated by the prolog/epilog pass
2651 DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true), DL);
2653 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, AArch64::SP, getPointerTy());
2655 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2656 SmallVector<SDValue, 8> MemOpChains;
2658 // Walk the register/memloc assignments, inserting copies/loads.
2659 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2660 ++i, ++realArgIdx) {
2661 CCValAssign &VA = ArgLocs[i];
2662 SDValue Arg = OutVals[realArgIdx];
2663 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2665 // Promote the value if needed.
2666 switch (VA.getLocInfo()) {
2668 llvm_unreachable("Unknown loc info!");
2669 case CCValAssign::Full:
2671 case CCValAssign::SExt:
2672 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
2674 case CCValAssign::ZExt:
2675 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2677 case CCValAssign::AExt:
2678 if (Outs[realArgIdx].ArgVT == MVT::i1) {
2679 // AAPCS requires i1 to be zero-extended to 8-bits by the caller.
2680 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
2681 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i8, Arg);
2683 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
2685 case CCValAssign::BCvt:
2686 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2688 case CCValAssign::FPExt:
2689 Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
2693 if (VA.isRegLoc()) {
2694 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i64) {
2695 assert(VA.getLocVT() == MVT::i64 &&
2696 "unexpected calling convention register assignment");
2697 assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
2698 "unexpected use of 'returned'");
2699 IsThisReturn = true;
2701 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2703 assert(VA.isMemLoc());
2706 MachinePointerInfo DstInfo;
2708 // FIXME: This works on big-endian for composite byvals, which are the
2709 // common case. It should also work for fundamental types too.
2710 uint32_t BEAlign = 0;
2711 unsigned OpSize = Flags.isByVal() ? Flags.getByValSize() * 8
2712 : VA.getValVT().getSizeInBits();
2713 OpSize = (OpSize + 7) / 8;
2714 if (!Subtarget->isLittleEndian() && !Flags.isByVal() &&
2715 !Flags.isInConsecutiveRegs()) {
2717 BEAlign = 8 - OpSize;
2719 unsigned LocMemOffset = VA.getLocMemOffset();
2720 int32_t Offset = LocMemOffset + BEAlign;
2721 SDValue PtrOff = DAG.getIntPtrConstant(Offset);
2722 PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff);
2725 Offset = Offset + FPDiff;
2726 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2728 DstAddr = DAG.getFrameIndex(FI, getPointerTy());
2729 DstInfo = MachinePointerInfo::getFixedStack(FI);
2731 // Make sure any stack arguments overlapping with where we're storing
2732 // are loaded before this eventual operation. Otherwise they'll be
2734 Chain = addTokenForArgument(Chain, DAG, MF.getFrameInfo(), FI);
2736 SDValue PtrOff = DAG.getIntPtrConstant(Offset);
2738 DstAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff);
2739 DstInfo = MachinePointerInfo::getStack(LocMemOffset);
2742 if (Outs[i].Flags.isByVal()) {
2744 DAG.getConstant(Outs[i].Flags.getByValSize(), MVT::i64);
2745 SDValue Cpy = DAG.getMemcpy(
2746 Chain, DL, DstAddr, Arg, SizeNode, Outs[i].Flags.getByValAlign(),
2748 /*AlwaysInline = */ false, DstInfo, MachinePointerInfo());
2750 MemOpChains.push_back(Cpy);
2752 // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
2753 // promoted to a legal register type i32, we should truncate Arg back to
2755 if (VA.getValVT() == MVT::i1 || VA.getValVT() == MVT::i8 ||
2756 VA.getValVT() == MVT::i16)
2757 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Arg);
2760 DAG.getStore(Chain, DL, Arg, DstAddr, DstInfo, false, false, 0);
2761 MemOpChains.push_back(Store);
2766 if (!MemOpChains.empty())
2767 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
2769 // Build a sequence of copy-to-reg nodes chained together with token chain
2770 // and flag operands which copy the outgoing args into the appropriate regs.
2772 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2773 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[i].first,
2774 RegsToPass[i].second, InFlag);
2775 InFlag = Chain.getValue(1);
2778 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2779 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2780 // node so that legalize doesn't hack it.
2781 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
2782 Subtarget->isTargetMachO()) {
2783 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2784 const GlobalValue *GV = G->getGlobal();
2785 bool InternalLinkage = GV->hasInternalLinkage();
2786 if (InternalLinkage)
2787 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2789 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0,
2791 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, getPointerTy(), Callee);
2793 } else if (ExternalSymbolSDNode *S =
2794 dyn_cast<ExternalSymbolSDNode>(Callee)) {
2795 const char *Sym = S->getSymbol();
2797 DAG.getTargetExternalSymbol(Sym, getPointerTy(), AArch64II::MO_GOT);
2798 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, getPointerTy(), Callee);
2800 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2801 const GlobalValue *GV = G->getGlobal();
2802 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2803 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2804 const char *Sym = S->getSymbol();
2805 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), 0);
2808 // We don't usually want to end the call-sequence here because we would tidy
2809 // the frame up *after* the call, however in the ABI-changing tail-call case
2810 // we've carefully laid out the parameters so that when sp is reset they'll be
2811 // in the correct location.
2812 if (IsTailCall && !IsSibCall) {
2813 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2814 DAG.getIntPtrConstant(0, true), InFlag, DL);
2815 InFlag = Chain.getValue(1);
2818 std::vector<SDValue> Ops;
2819 Ops.push_back(Chain);
2820 Ops.push_back(Callee);
2823 // Each tail call may have to adjust the stack by a different amount, so
2824 // this information must travel along with the operation for eventual
2825 // consumption by emitEpilogue.
2826 Ops.push_back(DAG.getTargetConstant(FPDiff, MVT::i32));
2829 // Add argument registers to the end of the list so that they are known live
2831 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2832 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2833 RegsToPass[i].second.getValueType()));
2835 // Add a register mask operand representing the call-preserved registers.
2836 const uint32_t *Mask;
2837 const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();
2839 // For 'this' returns, use the X0-preserving mask if applicable
2840 Mask = TRI->getThisReturnPreservedMask(MF, CallConv);
2842 IsThisReturn = false;
2843 Mask = TRI->getCallPreservedMask(MF, CallConv);
2846 Mask = TRI->getCallPreservedMask(MF, CallConv);
2848 assert(Mask && "Missing call preserved mask for calling convention");
2849 Ops.push_back(DAG.getRegisterMask(Mask));
2851 if (InFlag.getNode())
2852 Ops.push_back(InFlag);
2854 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2856 // If we're doing a tall call, use a TC_RETURN here rather than an
2857 // actual call instruction.
2859 return DAG.getNode(AArch64ISD::TC_RETURN, DL, NodeTys, Ops);
2861 // Returns a chain and a flag for retval copy to use.
2862 Chain = DAG.getNode(AArch64ISD::CALL, DL, NodeTys, Ops);
2863 InFlag = Chain.getValue(1);
2865 uint64_t CalleePopBytes = DoesCalleeRestoreStack(CallConv, TailCallOpt)
2866 ? RoundUpToAlignment(NumBytes, 16)
2869 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2870 DAG.getIntPtrConstant(CalleePopBytes, true),
2873 InFlag = Chain.getValue(1);
2875 // Handle result values, copying them out of physregs into vregs that we
2877 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
2878 InVals, IsThisReturn,
2879 IsThisReturn ? OutVals[0] : SDValue());
2882 bool AArch64TargetLowering::CanLowerReturn(
2883 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2884 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2885 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2886 ? RetCC_AArch64_WebKit_JS
2887 : RetCC_AArch64_AAPCS;
2888 SmallVector<CCValAssign, 16> RVLocs;
2889 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2890 return CCInfo.CheckReturn(Outs, RetCC);
2894 AArch64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2896 const SmallVectorImpl<ISD::OutputArg> &Outs,
2897 const SmallVectorImpl<SDValue> &OutVals,
2898 SDLoc DL, SelectionDAG &DAG) const {
2899 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2900 ? RetCC_AArch64_WebKit_JS
2901 : RetCC_AArch64_AAPCS;
2902 SmallVector<CCValAssign, 16> RVLocs;
2903 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2905 CCInfo.AnalyzeReturn(Outs, RetCC);
2907 // Copy the result values into the output registers.
2909 SmallVector<SDValue, 4> RetOps(1, Chain);
2910 for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size();
2911 ++i, ++realRVLocIdx) {
2912 CCValAssign &VA = RVLocs[i];
2913 assert(VA.isRegLoc() && "Can only return in registers!");
2914 SDValue Arg = OutVals[realRVLocIdx];
2916 switch (VA.getLocInfo()) {
2918 llvm_unreachable("Unknown loc info!");
2919 case CCValAssign::Full:
2920 if (Outs[i].ArgVT == MVT::i1) {
2921 // AAPCS requires i1 to be zero-extended to i8 by the producer of the
2922 // value. This is strictly redundant on Darwin (which uses "zeroext
2923 // i1"), but will be optimised out before ISel.
2924 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
2925 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2928 case CCValAssign::BCvt:
2929 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2933 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
2934 Flag = Chain.getValue(1);
2935 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2938 RetOps[0] = Chain; // Update chain.
2940 // Add the flag if we have it.
2942 RetOps.push_back(Flag);
2944 return DAG.getNode(AArch64ISD::RET_FLAG, DL, MVT::Other, RetOps);
2947 //===----------------------------------------------------------------------===//
2948 // Other Lowering Code
2949 //===----------------------------------------------------------------------===//
2951 SDValue AArch64TargetLowering::LowerGlobalAddress(SDValue Op,
2952 SelectionDAG &DAG) const {
2953 EVT PtrVT = getPointerTy();
2955 const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
2956 const GlobalValue *GV = GN->getGlobal();
2957 unsigned char OpFlags =
2958 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
2960 assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
2961 "unexpected offset in global node");
2963 // This also catched the large code model case for Darwin.
2964 if ((OpFlags & AArch64II::MO_GOT) != 0) {
2965 SDValue GotAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
2966 // FIXME: Once remat is capable of dealing with instructions with register
2967 // operands, expand this into two nodes instead of using a wrapper node.
2968 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
2971 if ((OpFlags & AArch64II::MO_CONSTPOOL) != 0) {
2972 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
2973 "use of MO_CONSTPOOL only supported on small model");
2974 SDValue Hi = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, AArch64II::MO_PAGE);
2975 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
2976 unsigned char LoFlags = AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
2977 SDValue Lo = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, LoFlags);
2978 SDValue PoolAddr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
2979 SDValue GlobalAddr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), PoolAddr,
2980 MachinePointerInfo::getConstantPool(),
2981 /*isVolatile=*/ false,
2982 /*isNonTemporal=*/ true,
2983 /*isInvariant=*/ true, 8);
2984 if (GN->getOffset() != 0)
2985 return DAG.getNode(ISD::ADD, DL, PtrVT, GlobalAddr,
2986 DAG.getConstant(GN->getOffset(), PtrVT));
2990 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2991 const unsigned char MO_NC = AArch64II::MO_NC;
2993 AArch64ISD::WrapperLarge, DL, PtrVT,
2994 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G3),
2995 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
2996 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
2997 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
2999 // Use ADRP/ADD or ADRP/LDR for everything else: the small model on ELF and
3000 // the only correct model on Darwin.
3001 SDValue Hi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
3002 OpFlags | AArch64II::MO_PAGE);
3003 unsigned char LoFlags = OpFlags | AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
3004 SDValue Lo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, LoFlags);
3006 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3007 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3011 /// \brief Convert a TLS address reference into the correct sequence of loads
3012 /// and calls to compute the variable's address (for Darwin, currently) and
3013 /// return an SDValue containing the final node.
3015 /// Darwin only has one TLS scheme which must be capable of dealing with the
3016 /// fully general situation, in the worst case. This means:
3017 /// + "extern __thread" declaration.
3018 /// + Defined in a possibly unknown dynamic library.
3020 /// The general system is that each __thread variable has a [3 x i64] descriptor
3021 /// which contains information used by the runtime to calculate the address. The
3022 /// only part of this the compiler needs to know about is the first xword, which
3023 /// contains a function pointer that must be called with the address of the
3024 /// entire descriptor in "x0".
3026 /// Since this descriptor may be in a different unit, in general even the
3027 /// descriptor must be accessed via an indirect load. The "ideal" code sequence
3029 /// adrp x0, _var@TLVPPAGE
3030 /// ldr x0, [x0, _var@TLVPPAGEOFF] ; x0 now contains address of descriptor
3031 /// ldr x1, [x0] ; x1 contains 1st entry of descriptor,
3032 /// ; the function pointer
3033 /// blr x1 ; Uses descriptor address in x0
3034 /// ; Address of _var is now in x0.
3036 /// If the address of _var's descriptor *is* known to the linker, then it can
3037 /// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
3038 /// a slight efficiency gain.
3040 AArch64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
3041 SelectionDAG &DAG) const {
3042 assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");
3045 MVT PtrVT = getPointerTy();
3046 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3049 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3050 SDValue DescAddr = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TLVPAddr);
3052 // The first entry in the descriptor is a function pointer that we must call
3053 // to obtain the address of the variable.
3054 SDValue Chain = DAG.getEntryNode();
3055 SDValue FuncTLVGet =
3056 DAG.getLoad(MVT::i64, DL, Chain, DescAddr, MachinePointerInfo::getGOT(),
3057 false, true, true, 8);
3058 Chain = FuncTLVGet.getValue(1);
3060 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3061 MFI->setAdjustsStack(true);
3063 // TLS calls preserve all registers except those that absolutely must be
3064 // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
3066 const uint32_t *Mask =
3067 Subtarget->getRegisterInfo()->getTLSCallPreservedMask();
3069 // Finally, we can make the call. This is just a degenerate version of a
3070 // normal AArch64 call node: x0 takes the address of the descriptor, and
3071 // returns the address of the variable in this thread.
3072 Chain = DAG.getCopyToReg(Chain, DL, AArch64::X0, DescAddr, SDValue());
3074 DAG.getNode(AArch64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
3075 Chain, FuncTLVGet, DAG.getRegister(AArch64::X0, MVT::i64),
3076 DAG.getRegisterMask(Mask), Chain.getValue(1));
3077 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Chain.getValue(1));
3080 /// When accessing thread-local variables under either the general-dynamic or
3081 /// local-dynamic system, we make a "TLS-descriptor" call. The variable will
3082 /// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
3083 /// is a function pointer to carry out the resolution.
3085 /// The sequence is:
3086 /// adrp x0, :tlsdesc:var
3087 /// ldr x1, [x0, #:tlsdesc_lo12:var]
3088 /// add x0, x0, #:tlsdesc_lo12:var
3089 /// .tlsdesccall var
3091 /// (TPIDR_EL0 offset now in x0)
3093 /// The above sequence must be produced unscheduled, to enable the linker to
3094 /// optimize/relax this sequence.
3095 /// Therefore, a pseudo-instruction (TLSDESC_CALLSEQ) is used to represent the
3096 /// above sequence, and expanded really late in the compilation flow, to ensure
3097 /// the sequence is produced as per above.
3098 SDValue AArch64TargetLowering::LowerELFTLSDescCallSeq(SDValue SymAddr, SDLoc DL,
3099 SelectionDAG &DAG) const {
3100 EVT PtrVT = getPointerTy();
3102 SDValue Chain = DAG.getEntryNode();
3103 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3105 SmallVector<SDValue, 2> Ops;
3106 Ops.push_back(Chain);
3107 Ops.push_back(SymAddr);
3109 Chain = DAG.getNode(AArch64ISD::TLSDESC_CALLSEQ, DL, NodeTys, Ops);
3110 SDValue Glue = Chain.getValue(1);
3112 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Glue);
3116 AArch64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
3117 SelectionDAG &DAG) const {
3118 assert(Subtarget->isTargetELF() && "This function expects an ELF target");
3119 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
3120 "ELF TLS only supported in small memory model");
3121 // Different choices can be made for the maximum size of the TLS area for a
3122 // module. For the small address model, the default TLS size is 16MiB and the
3123 // maximum TLS size is 4GiB.
3124 // FIXME: add -mtls-size command line option and make it control the 16MiB
3125 // vs. 4GiB code sequence generation.
3126 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3128 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
3129 if (!EnableAArch64ELFLocalDynamicTLSGeneration) {
3130 if (Model == TLSModel::LocalDynamic)
3131 Model = TLSModel::GeneralDynamic;
3135 EVT PtrVT = getPointerTy();
3137 const GlobalValue *GV = GA->getGlobal();
3139 SDValue ThreadBase = DAG.getNode(AArch64ISD::THREAD_POINTER, DL, PtrVT);
3141 if (Model == TLSModel::LocalExec) {
3142 SDValue HiVar = DAG.getTargetGlobalAddress(
3143 GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_HI12);
3144 SDValue LoVar = DAG.getTargetGlobalAddress(
3146 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3148 SDValue TPWithOff_lo =
3149 SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, ThreadBase,
3150 HiVar, DAG.getTargetConstant(0, MVT::i32)),
3153 SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPWithOff_lo,
3154 LoVar, DAG.getTargetConstant(0, MVT::i32)),
3157 } else if (Model == TLSModel::InitialExec) {
3158 TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3159 TPOff = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TPOff);
3160 } else if (Model == TLSModel::LocalDynamic) {
3161 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
3162 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
3163 // the beginning of the module's TLS region, followed by a DTPREL offset
3166 // These accesses will need deduplicating if there's more than one.
3167 AArch64FunctionInfo *MFI =
3168 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
3169 MFI->incNumLocalDynamicTLSAccesses();
3171 // The call needs a relocation too for linker relaxation. It doesn't make
3172 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
3174 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
3177 // Now we can calculate the offset from TPIDR_EL0 to this module's
3178 // thread-local area.
3179 TPOff = LowerELFTLSDescCallSeq(SymAddr, DL, DAG);
3181 // Now use :dtprel_whatever: operations to calculate this variable's offset
3182 // in its thread-storage area.
3183 SDValue HiVar = DAG.getTargetGlobalAddress(
3184 GV, DL, MVT::i64, 0, AArch64II::MO_TLS | AArch64II::MO_HI12);
3185 SDValue LoVar = DAG.getTargetGlobalAddress(
3186 GV, DL, MVT::i64, 0,
3187 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3189 TPOff = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPOff, HiVar,
3190 DAG.getTargetConstant(0, MVT::i32)),
3192 TPOff = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPOff, LoVar,
3193 DAG.getTargetConstant(0, MVT::i32)),
3195 } else if (Model == TLSModel::GeneralDynamic) {
3196 // The call needs a relocation too for linker relaxation. It doesn't make
3197 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
3200 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3202 // Finally we can make a call to calculate the offset from tpidr_el0.
3203 TPOff = LowerELFTLSDescCallSeq(SymAddr, DL, DAG);
3205 llvm_unreachable("Unsupported ELF TLS access model");
3207 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
3210 SDValue AArch64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
3211 SelectionDAG &DAG) const {
3212 if (Subtarget->isTargetDarwin())
3213 return LowerDarwinGlobalTLSAddress(Op, DAG);
3214 else if (Subtarget->isTargetELF())
3215 return LowerELFGlobalTLSAddress(Op, DAG);
3217 llvm_unreachable("Unexpected platform trying to use TLS");
3219 SDValue AArch64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3220 SDValue Chain = Op.getOperand(0);
3221 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3222 SDValue LHS = Op.getOperand(2);
3223 SDValue RHS = Op.getOperand(3);
3224 SDValue Dest = Op.getOperand(4);
3227 // Handle f128 first, since lowering it will result in comparing the return
3228 // value of a libcall against zero, which is just what the rest of LowerBR_CC
3229 // is expecting to deal with.
3230 if (LHS.getValueType() == MVT::f128) {
3231 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3233 // If softenSetCCOperands returned a scalar, we need to compare the result
3234 // against zero to select between true and false values.
3235 if (!RHS.getNode()) {
3236 RHS = DAG.getConstant(0, LHS.getValueType());
3241 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
3243 unsigned Opc = LHS.getOpcode();
3244 if (LHS.getResNo() == 1 && isa<ConstantSDNode>(RHS) &&
3245 cast<ConstantSDNode>(RHS)->isOne() &&
3246 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3247 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3248 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
3249 "Unexpected condition code.");
3250 // Only lower legal XALUO ops.
3251 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
3254 // The actual operation with overflow check.
3255 AArch64CC::CondCode OFCC;
3256 SDValue Value, Overflow;
3257 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, LHS.getValue(0), DAG);
3259 if (CC == ISD::SETNE)
3260 OFCC = getInvertedCondCode(OFCC);
3261 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
3263 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3267 if (LHS.getValueType().isInteger()) {
3268 assert((LHS.getValueType() == RHS.getValueType()) &&
3269 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3271 // If the RHS of the comparison is zero, we can potentially fold this
3272 // to a specialized branch.
3273 const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
3274 if (RHSC && RHSC->getZExtValue() == 0) {
3275 if (CC == ISD::SETEQ) {
3276 // See if we can use a TBZ to fold in an AND as well.
3277 // TBZ has a smaller branch displacement than CBZ. If the offset is
3278 // out of bounds, a late MI-layer pass rewrites branches.
3279 // 403.gcc is an example that hits this case.
3280 if (LHS.getOpcode() == ISD::AND &&
3281 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3282 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3283 SDValue Test = LHS.getOperand(0);
3284 uint64_t Mask = LHS.getConstantOperandVal(1);
3285 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, Test,
3286 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
3289 return DAG.getNode(AArch64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest);
3290 } else if (CC == ISD::SETNE) {
3291 // See if we can use a TBZ to fold in an AND as well.
3292 // TBZ has a smaller branch displacement than CBZ. If the offset is
3293 // out of bounds, a late MI-layer pass rewrites branches.
3294 // 403.gcc is an example that hits this case.
3295 if (LHS.getOpcode() == ISD::AND &&
3296 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3297 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3298 SDValue Test = LHS.getOperand(0);
3299 uint64_t Mask = LHS.getConstantOperandVal(1);
3300 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, Test,
3301 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
3304 return DAG.getNode(AArch64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest);
3305 } else if (CC == ISD::SETLT && LHS.getOpcode() != ISD::AND) {
3306 // Don't combine AND since emitComparison converts the AND to an ANDS
3307 // (a.k.a. TST) and the test in the test bit and branch instruction
3308 // becomes redundant. This would also increase register pressure.
3309 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
3310 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, LHS,
3311 DAG.getConstant(Mask, MVT::i64), Dest);
3314 if (RHSC && RHSC->getSExtValue() == -1 && CC == ISD::SETGT &&
3315 LHS.getOpcode() != ISD::AND) {
3316 // Don't combine AND since emitComparison converts the AND to an ANDS
3317 // (a.k.a. TST) and the test in the test bit and branch instruction
3318 // becomes redundant. This would also increase register pressure.
3319 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
3320 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, LHS,
3321 DAG.getConstant(Mask, MVT::i64), Dest);
3325 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3326 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3330 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3332 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
3333 // clean. Some of them require two branches to implement.
3334 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3335 AArch64CC::CondCode CC1, CC2;
3336 changeFPCCToAArch64CC(CC, CC1, CC2);
3337 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3339 DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp);
3340 if (CC2 != AArch64CC::AL) {
3341 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3342 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val,
3349 SDValue AArch64TargetLowering::LowerFCOPYSIGN(SDValue Op,
3350 SelectionDAG &DAG) const {
3351 EVT VT = Op.getValueType();
3354 SDValue In1 = Op.getOperand(0);
3355 SDValue In2 = Op.getOperand(1);
3356 EVT SrcVT = In2.getValueType();
3358 if (SrcVT == MVT::f32 && VT == MVT::f64)
3359 In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2);
3360 else if (SrcVT == MVT::f64 && VT == MVT::f32)
3361 In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, DAG.getIntPtrConstant(0));
3363 // FIXME: Src type is different, bail out for now. Can VT really be a
3371 SDValue VecVal1, VecVal2;
3372 if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) {
3375 EltMask = 0x80000000ULL;
3377 if (!VT.isVector()) {
3378 VecVal1 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3379 DAG.getUNDEF(VecVT), In1);
3380 VecVal2 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3381 DAG.getUNDEF(VecVT), In2);
3383 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3384 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3386 } else if (VT == MVT::f64 || VT == MVT::v2f64) {
3390 // We want to materialize a mask with the the high bit set, but the AdvSIMD
3391 // immediate moves cannot materialize that in a single instruction for
3392 // 64-bit elements. Instead, materialize zero and then negate it.
3395 if (!VT.isVector()) {
3396 VecVal1 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3397 DAG.getUNDEF(VecVT), In1);
3398 VecVal2 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3399 DAG.getUNDEF(VecVT), In2);
3401 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3402 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3405 llvm_unreachable("Invalid type for copysign!");
3408 SDValue BuildVec = DAG.getConstant(EltMask, VecVT);
3410 // If we couldn't materialize the mask above, then the mask vector will be
3411 // the zero vector, and we need to negate it here.
3412 if (VT == MVT::f64 || VT == MVT::v2f64) {
3413 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec);
3414 BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec);
3415 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec);
3419 DAG.getNode(AArch64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec);
3422 return DAG.getTargetExtractSubreg(AArch64::ssub, DL, VT, Sel);
3423 else if (VT == MVT::f64)
3424 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, VT, Sel);
3426 return DAG.getNode(ISD::BITCAST, DL, VT, Sel);
3429 SDValue AArch64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
3430 if (DAG.getMachineFunction().getFunction()->hasFnAttribute(
3431 Attribute::NoImplicitFloat))
3434 if (!Subtarget->hasNEON())
3437 // While there is no integer popcount instruction, it can
3438 // be more efficiently lowered to the following sequence that uses
3439 // AdvSIMD registers/instructions as long as the copies to/from
3440 // the AdvSIMD registers are cheap.
3441 // FMOV D0, X0 // copy 64-bit int to vector, high bits zero'd
3442 // CNT V0.8B, V0.8B // 8xbyte pop-counts
3443 // ADDV B0, V0.8B // sum 8xbyte pop-counts
3444 // UMOV X0, V0.B[0] // copy byte result back to integer reg
3445 SDValue Val = Op.getOperand(0);
3447 EVT VT = Op.getValueType();
3450 Val = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Val);
3451 Val = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);
3453 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, Val);
3454 SDValue UaddLV = DAG.getNode(
3455 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
3456 DAG.getConstant(Intrinsic::aarch64_neon_uaddlv, MVT::i32), CtPop);
3459 UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV);
3463 SDValue AArch64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
3465 if (Op.getValueType().isVector())
3466 return LowerVSETCC(Op, DAG);
3468 SDValue LHS = Op.getOperand(0);
3469 SDValue RHS = Op.getOperand(1);
3470 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
3473 // We chose ZeroOrOneBooleanContents, so use zero and one.
3474 EVT VT = Op.getValueType();
3475 SDValue TVal = DAG.getConstant(1, VT);
3476 SDValue FVal = DAG.getConstant(0, VT);
3478 // Handle f128 first, since one possible outcome is a normal integer
3479 // comparison which gets picked up by the next if statement.
3480 if (LHS.getValueType() == MVT::f128) {
3481 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3483 // If softenSetCCOperands returned a scalar, use it.
3484 if (!RHS.getNode()) {
3485 assert(LHS.getValueType() == Op.getValueType() &&
3486 "Unexpected setcc expansion!");
3491 if (LHS.getValueType().isInteger()) {
3494 getAArch64Cmp(LHS, RHS, ISD::getSetCCInverse(CC, true), CCVal, DAG, dl);
3496 // Note that we inverted the condition above, so we reverse the order of
3497 // the true and false operands here. This will allow the setcc to be
3498 // matched to a single CSINC instruction.
3499 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp);
3502 // Now we know we're dealing with FP values.
3503 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3505 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3506 // and do the comparison.
3507 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3509 AArch64CC::CondCode CC1, CC2;
3510 changeFPCCToAArch64CC(CC, CC1, CC2);
3511 if (CC2 == AArch64CC::AL) {
3512 changeFPCCToAArch64CC(ISD::getSetCCInverse(CC, false), CC1, CC2);
3513 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3515 // Note that we inverted the condition above, so we reverse the order of
3516 // the true and false operands here. This will allow the setcc to be
3517 // matched to a single CSINC instruction.
3518 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp);
3520 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't
3521 // totally clean. Some of them require two CSELs to implement. As is in
3522 // this case, we emit the first CSEL and then emit a second using the output
3523 // of the first as the RHS. We're effectively OR'ing the two CC's together.
3525 // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
3526 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3528 DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3530 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3531 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3535 /// A SELECT_CC operation is really some kind of max or min if both values being
3536 /// compared are, in some sense, equal to the results in either case. However,
3537 /// it is permissible to compare f32 values and produce directly extended f64
3540 /// Extending the comparison operands would also be allowed, but is less likely
3541 /// to happen in practice since their use is right here. Note that truncate
3542 /// operations would *not* be semantically equivalent.
3543 static bool selectCCOpsAreFMaxCompatible(SDValue Cmp, SDValue Result) {
3547 ConstantFPSDNode *CCmp = dyn_cast<ConstantFPSDNode>(Cmp);
3548 ConstantFPSDNode *CResult = dyn_cast<ConstantFPSDNode>(Result);
3549 if (CCmp && CResult && Cmp.getValueType() == MVT::f32 &&
3550 Result.getValueType() == MVT::f64) {
3552 APFloat CmpVal = CCmp->getValueAPF();
3553 CmpVal.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &Lossy);
3554 return CResult->getValueAPF().bitwiseIsEqual(CmpVal);
3557 return Result->getOpcode() == ISD::FP_EXTEND && Result->getOperand(0) == Cmp;
3560 SDValue AArch64TargetLowering::LowerSELECT_CC(ISD::CondCode CC, SDValue LHS,
3561 SDValue RHS, SDValue TVal,
3562 SDValue FVal, SDLoc dl,
3563 SelectionDAG &DAG) const {
3564 // Handle f128 first, because it will result in a comparison of some RTLIB
3565 // call result against zero.
3566 if (LHS.getValueType() == MVT::f128) {
3567 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3569 // If softenSetCCOperands returned a scalar, we need to compare the result
3570 // against zero to select between true and false values.
3571 if (!RHS.getNode()) {
3572 RHS = DAG.getConstant(0, LHS.getValueType());
3577 // Handle integers first.
3578 if (LHS.getValueType().isInteger()) {
3579 assert((LHS.getValueType() == RHS.getValueType()) &&
3580 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3582 unsigned Opcode = AArch64ISD::CSEL;
3584 // If both the TVal and the FVal are constants, see if we can swap them in
3585 // order to for a CSINV or CSINC out of them.
3586 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
3587 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
3589 if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) {
3590 std::swap(TVal, FVal);
3591 std::swap(CTVal, CFVal);
3592 CC = ISD::getSetCCInverse(CC, true);
3593 } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) {
3594 std::swap(TVal, FVal);
3595 std::swap(CTVal, CFVal);
3596 CC = ISD::getSetCCInverse(CC, true);
3597 } else if (TVal.getOpcode() == ISD::XOR) {
3598 // If TVal is a NOT we want to swap TVal and FVal so that we can match
3599 // with a CSINV rather than a CSEL.
3600 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(1));
3602 if (CVal && CVal->isAllOnesValue()) {
3603 std::swap(TVal, FVal);
3604 std::swap(CTVal, CFVal);
3605 CC = ISD::getSetCCInverse(CC, true);
3607 } else if (TVal.getOpcode() == ISD::SUB) {
3608 // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
3609 // that we can match with a CSNEG rather than a CSEL.
3610 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(0));
3612 if (CVal && CVal->isNullValue()) {
3613 std::swap(TVal, FVal);
3614 std::swap(CTVal, CFVal);
3615 CC = ISD::getSetCCInverse(CC, true);
3617 } else if (CTVal && CFVal) {
3618 const int64_t TrueVal = CTVal->getSExtValue();
3619 const int64_t FalseVal = CFVal->getSExtValue();
3622 // If both TVal and FVal are constants, see if FVal is the
3623 // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
3624 // instead of a CSEL in that case.
3625 if (TrueVal == ~FalseVal) {
3626 Opcode = AArch64ISD::CSINV;
3627 } else if (TrueVal == -FalseVal) {
3628 Opcode = AArch64ISD::CSNEG;
3629 } else if (TVal.getValueType() == MVT::i32) {
3630 // If our operands are only 32-bit wide, make sure we use 32-bit
3631 // arithmetic for the check whether we can use CSINC. This ensures that
3632 // the addition in the check will wrap around properly in case there is
3633 // an overflow (which would not be the case if we do the check with
3634 // 64-bit arithmetic).
3635 const uint32_t TrueVal32 = CTVal->getZExtValue();
3636 const uint32_t FalseVal32 = CFVal->getZExtValue();
3638 if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
3639 Opcode = AArch64ISD::CSINC;
3641 if (TrueVal32 > FalseVal32) {
3645 // 64-bit check whether we can use CSINC.
3646 } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) {
3647 Opcode = AArch64ISD::CSINC;
3649 if (TrueVal > FalseVal) {
3654 // Swap TVal and FVal if necessary.
3656 std::swap(TVal, FVal);
3657 std::swap(CTVal, CFVal);
3658 CC = ISD::getSetCCInverse(CC, true);
3661 if (Opcode != AArch64ISD::CSEL) {
3662 // Drop FVal since we can get its value by simply inverting/negating
3669 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3671 EVT VT = TVal.getValueType();
3672 return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp);
3675 // Now we know we're dealing with FP values.
3676 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3677 assert(LHS.getValueType() == RHS.getValueType());
3678 EVT VT = TVal.getValueType();
3680 // Try to match this select into a max/min operation, which have dedicated
3681 // opcode in the instruction set.
3682 // FIXME: This is not correct in the presence of NaNs, so we only enable this
3684 if (getTargetMachine().Options.NoNaNsFPMath) {
3685 SDValue MinMaxLHS = TVal, MinMaxRHS = FVal;
3686 if (selectCCOpsAreFMaxCompatible(LHS, MinMaxRHS) &&
3687 selectCCOpsAreFMaxCompatible(RHS, MinMaxLHS)) {
3688 CC = ISD::getSetCCSwappedOperands(CC);
3689 std::swap(MinMaxLHS, MinMaxRHS);
3692 if (selectCCOpsAreFMaxCompatible(LHS, MinMaxLHS) &&
3693 selectCCOpsAreFMaxCompatible(RHS, MinMaxRHS)) {
3703 return DAG.getNode(AArch64ISD::FMAX, dl, VT, MinMaxLHS, MinMaxRHS);
3711 return DAG.getNode(AArch64ISD::FMIN, dl, VT, MinMaxLHS, MinMaxRHS);
3717 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3718 // and do the comparison.
3719 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3721 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
3722 // clean. Some of them require two CSELs to implement.
3723 AArch64CC::CondCode CC1, CC2;
3724 changeFPCCToAArch64CC(CC, CC1, CC2);
3725 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3726 SDValue CS1 = DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3728 // If we need a second CSEL, emit it, using the output of the first as the
3729 // RHS. We're effectively OR'ing the two CC's together.
3730 if (CC2 != AArch64CC::AL) {
3731 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3732 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3735 // Otherwise, return the output of the first CSEL.
3739 SDValue AArch64TargetLowering::LowerSELECT_CC(SDValue Op,
3740 SelectionDAG &DAG) const {
3741 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3742 SDValue LHS = Op.getOperand(0);
3743 SDValue RHS = Op.getOperand(1);
3744 SDValue TVal = Op.getOperand(2);
3745 SDValue FVal = Op.getOperand(3);
3747 return LowerSELECT_CC(CC, LHS, RHS, TVal, FVal, DL, DAG);
3750 SDValue AArch64TargetLowering::LowerSELECT(SDValue Op,
3751 SelectionDAG &DAG) const {
3752 SDValue CCVal = Op->getOperand(0);
3753 SDValue TVal = Op->getOperand(1);
3754 SDValue FVal = Op->getOperand(2);
3757 unsigned Opc = CCVal.getOpcode();
3758 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
3760 if (CCVal.getResNo() == 1 &&
3761 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3762 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3763 // Only lower legal XALUO ops.
3764 if (!DAG.getTargetLoweringInfo().isTypeLegal(CCVal->getValueType(0)))
3767 AArch64CC::CondCode OFCC;
3768 SDValue Value, Overflow;
3769 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, CCVal.getValue(0), DAG);
3770 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
3772 return DAG.getNode(AArch64ISD::CSEL, DL, Op.getValueType(), TVal, FVal,
3776 // Lower it the same way as we would lower a SELECT_CC node.
3779 if (CCVal.getOpcode() == ISD::SETCC) {
3780 LHS = CCVal.getOperand(0);
3781 RHS = CCVal.getOperand(1);
3782 CC = cast<CondCodeSDNode>(CCVal->getOperand(2))->get();
3785 RHS = DAG.getConstant(0, CCVal.getValueType());
3788 return LowerSELECT_CC(CC, LHS, RHS, TVal, FVal, DL, DAG);
3791 SDValue AArch64TargetLowering::LowerJumpTable(SDValue Op,
3792 SelectionDAG &DAG) const {
3793 // Jump table entries as PC relative offsets. No additional tweaking
3794 // is necessary here. Just get the address of the jump table.
3795 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3796 EVT PtrVT = getPointerTy();
3799 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3800 !Subtarget->isTargetMachO()) {
3801 const unsigned char MO_NC = AArch64II::MO_NC;
3803 AArch64ISD::WrapperLarge, DL, PtrVT,
3804 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G3),
3805 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G2 | MO_NC),
3806 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G1 | MO_NC),
3807 DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3808 AArch64II::MO_G0 | MO_NC));
3812 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_PAGE);
3813 SDValue Lo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3814 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3815 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3816 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3819 SDValue AArch64TargetLowering::LowerConstantPool(SDValue Op,
3820 SelectionDAG &DAG) const {
3821 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3822 EVT PtrVT = getPointerTy();
3825 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
3826 // Use the GOT for the large code model on iOS.
3827 if (Subtarget->isTargetMachO()) {
3828 SDValue GotAddr = DAG.getTargetConstantPool(
3829 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3831 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
3834 const unsigned char MO_NC = AArch64II::MO_NC;
3836 AArch64ISD::WrapperLarge, DL, PtrVT,
3837 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3838 CP->getOffset(), AArch64II::MO_G3),
3839 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3840 CP->getOffset(), AArch64II::MO_G2 | MO_NC),
3841 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3842 CP->getOffset(), AArch64II::MO_G1 | MO_NC),
3843 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3844 CP->getOffset(), AArch64II::MO_G0 | MO_NC));
3846 // Use ADRP/ADD or ADRP/LDR for everything else: the small memory model on
3847 // ELF, the only valid one on Darwin.
3849 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3850 CP->getOffset(), AArch64II::MO_PAGE);
3851 SDValue Lo = DAG.getTargetConstantPool(
3852 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3853 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3855 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3856 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3860 SDValue AArch64TargetLowering::LowerBlockAddress(SDValue Op,
3861 SelectionDAG &DAG) const {
3862 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
3863 EVT PtrVT = getPointerTy();
3865 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3866 !Subtarget->isTargetMachO()) {
3867 const unsigned char MO_NC = AArch64II::MO_NC;
3869 AArch64ISD::WrapperLarge, DL, PtrVT,
3870 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G3),
3871 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
3872 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
3873 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
3875 SDValue Hi = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGE);
3876 SDValue Lo = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGEOFF |
3878 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3879 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3883 SDValue AArch64TargetLowering::LowerDarwin_VASTART(SDValue Op,
3884 SelectionDAG &DAG) const {
3885 AArch64FunctionInfo *FuncInfo =
3886 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
3890 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3891 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3892 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
3893 MachinePointerInfo(SV), false, false, 0);
3896 SDValue AArch64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
3897 SelectionDAG &DAG) const {
3898 // The layout of the va_list struct is specified in the AArch64 Procedure Call
3899 // Standard, section B.3.
3900 MachineFunction &MF = DAG.getMachineFunction();
3901 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
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
3911 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3912 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
3913 MachinePointerInfo(SV), false, false, 8));
3915 // void *__gr_top at offset 8
3916 int GPRSize = FuncInfo->getVarArgsGPRSize();
3918 SDValue GRTop, GRTopAddr;
3920 GRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3921 DAG.getConstant(8, getPointerTy()));
3923 GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), getPointerTy());
3924 GRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), GRTop,
3925 DAG.getConstant(GPRSize, getPointerTy()));
3927 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
3928 MachinePointerInfo(SV, 8), false, false, 8));
3931 // void *__vr_top at offset 16
3932 int FPRSize = FuncInfo->getVarArgsFPRSize();
3934 SDValue VRTop, VRTopAddr;
3935 VRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3936 DAG.getConstant(16, getPointerTy()));
3938 VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), getPointerTy());
3939 VRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), VRTop,
3940 DAG.getConstant(FPRSize, getPointerTy()));
3942 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
3943 MachinePointerInfo(SV, 16), false, false, 8));
3946 // int __gr_offs at offset 24
3947 SDValue GROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3948 DAG.getConstant(24, getPointerTy()));
3949 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-GPRSize, MVT::i32),
3950 GROffsAddr, MachinePointerInfo(SV, 24), false,
3953 // int __vr_offs at offset 28
3954 SDValue VROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3955 DAG.getConstant(28, getPointerTy()));
3956 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-FPRSize, MVT::i32),
3957 VROffsAddr, MachinePointerInfo(SV, 28), false,
3960 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
3963 SDValue AArch64TargetLowering::LowerVASTART(SDValue Op,
3964 SelectionDAG &DAG) const {
3965 return Subtarget->isTargetDarwin() ? LowerDarwin_VASTART(Op, DAG)
3966 : LowerAAPCS_VASTART(Op, DAG);
3969 SDValue AArch64TargetLowering::LowerVACOPY(SDValue Op,
3970 SelectionDAG &DAG) const {
3971 // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
3973 unsigned VaListSize = Subtarget->isTargetDarwin() ? 8 : 32;
3974 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
3975 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
3977 return DAG.getMemcpy(Op.getOperand(0), SDLoc(Op), Op.getOperand(1),
3978 Op.getOperand(2), DAG.getConstant(VaListSize, MVT::i32),
3979 8, false, false, MachinePointerInfo(DestSV),
3980 MachinePointerInfo(SrcSV));
3983 SDValue AArch64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
3984 assert(Subtarget->isTargetDarwin() &&
3985 "automatic va_arg instruction only works on Darwin");
3987 const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3988 EVT VT = Op.getValueType();
3990 SDValue Chain = Op.getOperand(0);
3991 SDValue Addr = Op.getOperand(1);
3992 unsigned Align = Op.getConstantOperandVal(3);
3994 SDValue VAList = DAG.getLoad(getPointerTy(), DL, Chain, Addr,
3995 MachinePointerInfo(V), false, false, false, 0);
3996 Chain = VAList.getValue(1);
3999 assert(((Align & (Align - 1)) == 0) && "Expected Align to be a power of 2");
4000 VAList = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
4001 DAG.getConstant(Align - 1, getPointerTy()));
4002 VAList = DAG.getNode(ISD::AND, DL, getPointerTy(), VAList,
4003 DAG.getConstant(-(int64_t)Align, getPointerTy()));
4006 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
4007 uint64_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
4009 // Scalar integer and FP values smaller than 64 bits are implicitly extended
4010 // up to 64 bits. At the very least, we have to increase the striding of the
4011 // vaargs list to match this, and for FP values we need to introduce
4012 // FP_ROUND nodes as well.
4013 if (VT.isInteger() && !VT.isVector())
4015 bool NeedFPTrunc = false;
4016 if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
4021 // Increment the pointer, VAList, to the next vaarg
4022 SDValue VANext = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
4023 DAG.getConstant(ArgSize, getPointerTy()));
4024 // Store the incremented VAList to the legalized pointer
4025 SDValue APStore = DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V),
4028 // Load the actual argument out of the pointer VAList
4030 // Load the value as an f64.
4031 SDValue WideFP = DAG.getLoad(MVT::f64, DL, APStore, VAList,
4032 MachinePointerInfo(), false, false, false, 0);
4033 // Round the value down to an f32.
4034 SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
4035 DAG.getIntPtrConstant(1));
4036 SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
4037 // Merge the rounded value with the chain output of the load.
4038 return DAG.getMergeValues(Ops, DL);
4041 return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo(), false,
4045 SDValue AArch64TargetLowering::LowerFRAMEADDR(SDValue Op,
4046 SelectionDAG &DAG) const {
4047 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4048 MFI->setFrameAddressIsTaken(true);
4050 EVT VT = Op.getValueType();
4052 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4054 DAG.getCopyFromReg(DAG.getEntryNode(), DL, AArch64::FP, VT);
4056 FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
4057 MachinePointerInfo(), false, false, false, 0);
4061 // FIXME? Maybe this could be a TableGen attribute on some registers and
4062 // this table could be generated automatically from RegInfo.
4063 unsigned AArch64TargetLowering::getRegisterByName(const char* RegName,
4065 unsigned Reg = StringSwitch<unsigned>(RegName)
4066 .Case("sp", AArch64::SP)
4070 report_fatal_error("Invalid register name global variable");
4073 SDValue AArch64TargetLowering::LowerRETURNADDR(SDValue Op,
4074 SelectionDAG &DAG) const {
4075 MachineFunction &MF = DAG.getMachineFunction();
4076 MachineFrameInfo *MFI = MF.getFrameInfo();
4077 MFI->setReturnAddressIsTaken(true);
4079 EVT VT = Op.getValueType();
4081 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4083 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4084 SDValue Offset = DAG.getConstant(8, getPointerTy());
4085 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
4086 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
4087 MachinePointerInfo(), false, false, false, 0);
4090 // Return LR, which contains the return address. Mark it an implicit live-in.
4091 unsigned Reg = MF.addLiveIn(AArch64::LR, &AArch64::GPR64RegClass);
4092 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
4095 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4096 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4097 SDValue AArch64TargetLowering::LowerShiftRightParts(SDValue Op,
4098 SelectionDAG &DAG) const {
4099 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4100 EVT VT = Op.getValueType();
4101 unsigned VTBits = VT.getSizeInBits();
4103 SDValue ShOpLo = Op.getOperand(0);
4104 SDValue ShOpHi = Op.getOperand(1);
4105 SDValue ShAmt = Op.getOperand(2);
4107 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4109 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4111 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4112 DAG.getConstant(VTBits, MVT::i64), ShAmt);
4113 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4114 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4115 DAG.getConstant(VTBits, MVT::i64));
4116 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4118 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
4119 ISD::SETGE, dl, DAG);
4120 SDValue CCVal = DAG.getConstant(AArch64CC::GE, MVT::i32);
4122 SDValue FalseValLo = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4123 SDValue TrueValLo = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4125 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
4127 // AArch64 shifts larger than the register width are wrapped rather than
4128 // clamped, so we can't just emit "hi >> x".
4129 SDValue FalseValHi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4130 SDValue TrueValHi = Opc == ISD::SRA
4131 ? DAG.getNode(Opc, dl, VT, ShOpHi,
4132 DAG.getConstant(VTBits - 1, MVT::i64))
4133 : DAG.getConstant(0, VT);
4135 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValHi, FalseValHi, CCVal, Cmp);
4137 SDValue Ops[2] = { Lo, Hi };
4138 return DAG.getMergeValues(Ops, dl);
4141 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4142 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4143 SDValue AArch64TargetLowering::LowerShiftLeftParts(SDValue Op,
4144 SelectionDAG &DAG) const {
4145 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4146 EVT VT = Op.getValueType();
4147 unsigned VTBits = VT.getSizeInBits();
4149 SDValue ShOpLo = Op.getOperand(0);
4150 SDValue ShOpHi = Op.getOperand(1);
4151 SDValue ShAmt = Op.getOperand(2);
4154 assert(Op.getOpcode() == ISD::SHL_PARTS);
4155 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4156 DAG.getConstant(VTBits, MVT::i64), ShAmt);
4157 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4158 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4159 DAG.getConstant(VTBits, MVT::i64));
4160 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4161 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4163 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4165 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
4166 ISD::SETGE, dl, DAG);
4167 SDValue CCVal = DAG.getConstant(AArch64CC::GE, MVT::i32);
4169 DAG.getNode(AArch64ISD::CSEL, dl, VT, Tmp3, FalseVal, CCVal, Cmp);
4171 // AArch64 shifts of larger than register sizes are wrapped rather than
4172 // clamped, so we can't just emit "lo << a" if a is too big.
4173 SDValue TrueValLo = DAG.getConstant(0, VT);
4174 SDValue FalseValLo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4176 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
4178 SDValue Ops[2] = { Lo, Hi };
4179 return DAG.getMergeValues(Ops, dl);
4182 bool AArch64TargetLowering::isOffsetFoldingLegal(
4183 const GlobalAddressSDNode *GA) const {
4184 // The AArch64 target doesn't support folding offsets into global addresses.
4188 bool AArch64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4189 // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
4190 // FIXME: We should be able to handle f128 as well with a clever lowering.
4191 if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32))
4195 return AArch64_AM::getFP64Imm(Imm) != -1;
4196 else if (VT == MVT::f32)
4197 return AArch64_AM::getFP32Imm(Imm) != -1;
4201 //===----------------------------------------------------------------------===//
4202 // AArch64 Optimization Hooks
4203 //===----------------------------------------------------------------------===//
4205 //===----------------------------------------------------------------------===//
4206 // AArch64 Inline Assembly Support
4207 //===----------------------------------------------------------------------===//
4209 // Table of Constraints
4210 // TODO: This is the current set of constraints supported by ARM for the
4211 // compiler, not all of them may make sense, e.g. S may be difficult to support.
4213 // r - A general register
4214 // w - An FP/SIMD register of some size in the range v0-v31
4215 // x - An FP/SIMD register of some size in the range v0-v15
4216 // I - Constant that can be used with an ADD instruction
4217 // J - Constant that can be used with a SUB instruction
4218 // K - Constant that can be used with a 32-bit logical instruction
4219 // L - Constant that can be used with a 64-bit logical instruction
4220 // M - Constant that can be used as a 32-bit MOV immediate
4221 // N - Constant that can be used as a 64-bit MOV immediate
4222 // Q - A memory reference with base register and no offset
4223 // S - A symbolic address
4224 // Y - Floating point constant zero
4225 // Z - Integer constant zero
4227 // Note that general register operands will be output using their 64-bit x
4228 // register name, whatever the size of the variable, unless the asm operand
4229 // is prefixed by the %w modifier. Floating-point and SIMD register operands
4230 // will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
4233 /// getConstraintType - Given a constraint letter, return the type of
4234 /// constraint it is for this target.
4235 AArch64TargetLowering::ConstraintType
4236 AArch64TargetLowering::getConstraintType(const std::string &Constraint) const {
4237 if (Constraint.size() == 1) {
4238 switch (Constraint[0]) {
4245 return C_RegisterClass;
4246 // An address with a single base register. Due to the way we
4247 // currently handle addresses it is the same as 'r'.
4252 return TargetLowering::getConstraintType(Constraint);
4255 /// Examine constraint type and operand type and determine a weight value.
4256 /// This object must already have been set up with the operand type
4257 /// and the current alternative constraint selected.
4258 TargetLowering::ConstraintWeight
4259 AArch64TargetLowering::getSingleConstraintMatchWeight(
4260 AsmOperandInfo &info, const char *constraint) const {
4261 ConstraintWeight weight = CW_Invalid;
4262 Value *CallOperandVal = info.CallOperandVal;
4263 // If we don't have a value, we can't do a match,
4264 // but allow it at the lowest weight.
4265 if (!CallOperandVal)
4267 Type *type = CallOperandVal->getType();
4268 // Look at the constraint type.
4269 switch (*constraint) {
4271 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
4275 if (type->isFloatingPointTy() || type->isVectorTy())
4276 weight = CW_Register;
4279 weight = CW_Constant;
4285 std::pair<unsigned, const TargetRegisterClass *>
4286 AArch64TargetLowering::getRegForInlineAsmConstraint(
4287 const TargetRegisterInfo *TRI, const std::string &Constraint,
4289 if (Constraint.size() == 1) {
4290 switch (Constraint[0]) {
4292 if (VT.getSizeInBits() == 64)
4293 return std::make_pair(0U, &AArch64::GPR64commonRegClass);
4294 return std::make_pair(0U, &AArch64::GPR32commonRegClass);
4297 return std::make_pair(0U, &AArch64::FPR32RegClass);
4298 if (VT.getSizeInBits() == 64)
4299 return std::make_pair(0U, &AArch64::FPR64RegClass);
4300 if (VT.getSizeInBits() == 128)
4301 return std::make_pair(0U, &AArch64::FPR128RegClass);
4303 // The instructions that this constraint is designed for can
4304 // only take 128-bit registers so just use that regclass.
4306 if (VT.getSizeInBits() == 128)
4307 return std::make_pair(0U, &AArch64::FPR128_loRegClass);
4311 if (StringRef("{cc}").equals_lower(Constraint))
4312 return std::make_pair(unsigned(AArch64::NZCV), &AArch64::CCRRegClass);
4314 // Use the default implementation in TargetLowering to convert the register
4315 // constraint into a member of a register class.
4316 std::pair<unsigned, const TargetRegisterClass *> Res;
4317 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
4319 // Not found as a standard register?
4321 unsigned Size = Constraint.size();
4322 if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
4323 tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
4324 const std::string Reg =
4325 std::string(&Constraint[2], &Constraint[Size - 1]);
4326 int RegNo = atoi(Reg.c_str());
4327 if (RegNo >= 0 && RegNo <= 31) {
4328 // v0 - v31 are aliases of q0 - q31.
4329 // By default we'll emit v0-v31 for this unless there's a modifier where
4330 // we'll emit the correct register as well.
4331 Res.first = AArch64::FPR128RegClass.getRegister(RegNo);
4332 Res.second = &AArch64::FPR128RegClass;
4340 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4341 /// vector. If it is invalid, don't add anything to Ops.
4342 void AArch64TargetLowering::LowerAsmOperandForConstraint(
4343 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
4344 SelectionDAG &DAG) const {
4347 // Currently only support length 1 constraints.
4348 if (Constraint.length() != 1)
4351 char ConstraintLetter = Constraint[0];
4352 switch (ConstraintLetter) {
4356 // This set of constraints deal with valid constants for various instructions.
4357 // Validate and return a target constant for them if we can.
4359 // 'z' maps to xzr or wzr so it needs an input of 0.
4360 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4361 if (!C || C->getZExtValue() != 0)
4364 if (Op.getValueType() == MVT::i64)
4365 Result = DAG.getRegister(AArch64::XZR, MVT::i64);
4367 Result = DAG.getRegister(AArch64::WZR, MVT::i32);
4377 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4381 // Grab the value and do some validation.
4382 uint64_t CVal = C->getZExtValue();
4383 switch (ConstraintLetter) {
4384 // The I constraint applies only to simple ADD or SUB immediate operands:
4385 // i.e. 0 to 4095 with optional shift by 12
4386 // The J constraint applies only to ADD or SUB immediates that would be
4387 // valid when negated, i.e. if [an add pattern] were to be output as a SUB
4388 // instruction [or vice versa], in other words -1 to -4095 with optional
4389 // left shift by 12.
4391 if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
4395 uint64_t NVal = -C->getSExtValue();
4396 if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal)) {
4397 CVal = C->getSExtValue();
4402 // The K and L constraints apply *only* to logical immediates, including
4403 // what used to be the MOVI alias for ORR (though the MOVI alias has now
4404 // been removed and MOV should be used). So these constraints have to
4405 // distinguish between bit patterns that are valid 32-bit or 64-bit
4406 // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
4407 // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
4410 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4414 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4417 // The M and N constraints are a superset of K and L respectively, for use
4418 // with the MOV (immediate) alias. As well as the logical immediates they
4419 // also match 32 or 64-bit immediates that can be loaded either using a
4420 // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
4421 // (M) or 64-bit 0x1234000000000000 (N) etc.
4422 // As a note some of this code is liberally stolen from the asm parser.
4424 if (!isUInt<32>(CVal))
4426 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4428 if ((CVal & 0xFFFF) == CVal)
4430 if ((CVal & 0xFFFF0000ULL) == CVal)
4432 uint64_t NCVal = ~(uint32_t)CVal;
4433 if ((NCVal & 0xFFFFULL) == NCVal)
4435 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4440 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4442 if ((CVal & 0xFFFFULL) == CVal)
4444 if ((CVal & 0xFFFF0000ULL) == CVal)
4446 if ((CVal & 0xFFFF00000000ULL) == CVal)
4448 if ((CVal & 0xFFFF000000000000ULL) == CVal)
4450 uint64_t NCVal = ~CVal;
4451 if ((NCVal & 0xFFFFULL) == NCVal)
4453 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4455 if ((NCVal & 0xFFFF00000000ULL) == NCVal)
4457 if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
4465 // All assembler immediates are 64-bit integers.
4466 Result = DAG.getTargetConstant(CVal, MVT::i64);
4470 if (Result.getNode()) {
4471 Ops.push_back(Result);
4475 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
4478 //===----------------------------------------------------------------------===//
4479 // AArch64 Advanced SIMD Support
4480 //===----------------------------------------------------------------------===//
4482 /// WidenVector - Given a value in the V64 register class, produce the
4483 /// equivalent value in the V128 register class.
4484 static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
4485 EVT VT = V64Reg.getValueType();
4486 unsigned NarrowSize = VT.getVectorNumElements();
4487 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4488 MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
4491 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
4492 V64Reg, DAG.getConstant(0, MVT::i32));
4495 /// getExtFactor - Determine the adjustment factor for the position when
4496 /// generating an "extract from vector registers" instruction.
4497 static unsigned getExtFactor(SDValue &V) {
4498 EVT EltType = V.getValueType().getVectorElementType();
4499 return EltType.getSizeInBits() / 8;
4502 /// NarrowVector - Given a value in the V128 register class, produce the
4503 /// equivalent value in the V64 register class.
4504 static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) {
4505 EVT VT = V128Reg.getValueType();
4506 unsigned WideSize = VT.getVectorNumElements();
4507 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4508 MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2);
4511 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, NarrowTy, V128Reg);
4514 // Gather data to see if the operation can be modelled as a
4515 // shuffle in combination with VEXTs.
4516 SDValue AArch64TargetLowering::ReconstructShuffle(SDValue Op,
4517 SelectionDAG &DAG) const {
4518 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
4520 EVT VT = Op.getValueType();
4521 unsigned NumElts = VT.getVectorNumElements();
4523 struct ShuffleSourceInfo {
4528 // We may insert some combination of BITCASTs and VEXT nodes to force Vec to
4529 // be compatible with the shuffle we intend to construct. As a result
4530 // ShuffleVec will be some sliding window into the original Vec.
4533 // Code should guarantee that element i in Vec starts at element "WindowBase
4534 // + i * WindowScale in ShuffleVec".
4538 bool operator ==(SDValue OtherVec) { return Vec == OtherVec; }
4539 ShuffleSourceInfo(SDValue Vec)
4540 : Vec(Vec), MinElt(UINT_MAX), MaxElt(0), ShuffleVec(Vec), WindowBase(0),
4544 // First gather all vectors used as an immediate source for this BUILD_VECTOR
4546 SmallVector<ShuffleSourceInfo, 2> Sources;
4547 for (unsigned i = 0; i < NumElts; ++i) {
4548 SDValue V = Op.getOperand(i);
4549 if (V.getOpcode() == ISD::UNDEF)
4551 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4552 // A shuffle can only come from building a vector from various
4553 // elements of other vectors.
4557 // Add this element source to the list if it's not already there.
4558 SDValue SourceVec = V.getOperand(0);
4559 auto Source = std::find(Sources.begin(), Sources.end(), SourceVec);
4560 if (Source == Sources.end())
4561 Source = Sources.insert(Sources.end(), ShuffleSourceInfo(SourceVec));
4563 // Update the minimum and maximum lane number seen.
4564 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4565 Source->MinElt = std::min(Source->MinElt, EltNo);
4566 Source->MaxElt = std::max(Source->MaxElt, EltNo);
4569 // Currently only do something sane when at most two source vectors
4571 if (Sources.size() > 2)
4574 // Find out the smallest element size among result and two sources, and use
4575 // it as element size to build the shuffle_vector.
4576 EVT SmallestEltTy = VT.getVectorElementType();
4577 for (auto &Source : Sources) {
4578 EVT SrcEltTy = Source.Vec.getValueType().getVectorElementType();
4579 if (SrcEltTy.bitsLT(SmallestEltTy)) {
4580 SmallestEltTy = SrcEltTy;
4583 unsigned ResMultiplier =
4584 VT.getVectorElementType().getSizeInBits() / SmallestEltTy.getSizeInBits();
4585 NumElts = VT.getSizeInBits() / SmallestEltTy.getSizeInBits();
4586 EVT ShuffleVT = EVT::getVectorVT(*DAG.getContext(), SmallestEltTy, NumElts);
4588 // If the source vector is too wide or too narrow, we may nevertheless be able
4589 // to construct a compatible shuffle either by concatenating it with UNDEF or
4590 // extracting a suitable range of elements.
4591 for (auto &Src : Sources) {
4592 EVT SrcVT = Src.ShuffleVec.getValueType();
4594 if (SrcVT.getSizeInBits() == VT.getSizeInBits())
4597 // This stage of the search produces a source with the same element type as
4598 // the original, but with a total width matching the BUILD_VECTOR output.
4599 EVT EltVT = SrcVT.getVectorElementType();
4600 unsigned NumSrcElts = VT.getSizeInBits() / EltVT.getSizeInBits();
4601 EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumSrcElts);
4603 if (SrcVT.getSizeInBits() < VT.getSizeInBits()) {
4604 assert(2 * SrcVT.getSizeInBits() == VT.getSizeInBits());
4605 // We can pad out the smaller vector for free, so if it's part of a
4608 DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, Src.ShuffleVec,
4609 DAG.getUNDEF(Src.ShuffleVec.getValueType()));
4613 assert(SrcVT.getSizeInBits() == 2 * VT.getSizeInBits());
4615 if (Src.MaxElt - Src.MinElt >= NumSrcElts) {
4616 // Span too large for a VEXT to cope
4620 if (Src.MinElt >= NumSrcElts) {
4621 // The extraction can just take the second half
4623 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4624 DAG.getConstant(NumSrcElts, MVT::i64));
4625 Src.WindowBase = -NumSrcElts;
4626 } else if (Src.MaxElt < NumSrcElts) {
4627 // The extraction can just take the first half
4629 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4630 DAG.getConstant(0, MVT::i64));
4632 // An actual VEXT is needed
4634 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4635 DAG.getConstant(0, MVT::i64));
4637 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4638 DAG.getConstant(NumSrcElts, MVT::i64));
4639 unsigned Imm = Src.MinElt * getExtFactor(VEXTSrc1);
4641 Src.ShuffleVec = DAG.getNode(AArch64ISD::EXT, dl, DestVT, VEXTSrc1,
4642 VEXTSrc2, DAG.getConstant(Imm, MVT::i32));
4643 Src.WindowBase = -Src.MinElt;
4647 // Another possible incompatibility occurs from the vector element types. We
4648 // can fix this by bitcasting the source vectors to the same type we intend
4650 for (auto &Src : Sources) {
4651 EVT SrcEltTy = Src.ShuffleVec.getValueType().getVectorElementType();
4652 if (SrcEltTy == SmallestEltTy)
4654 assert(ShuffleVT.getVectorElementType() == SmallestEltTy);
4655 Src.ShuffleVec = DAG.getNode(ISD::BITCAST, dl, ShuffleVT, Src.ShuffleVec);
4656 Src.WindowScale = SrcEltTy.getSizeInBits() / SmallestEltTy.getSizeInBits();
4657 Src.WindowBase *= Src.WindowScale;
4660 // Final sanity check before we try to actually produce a shuffle.
4662 for (auto Src : Sources)
4663 assert(Src.ShuffleVec.getValueType() == ShuffleVT);
4666 // The stars all align, our next step is to produce the mask for the shuffle.
4667 SmallVector<int, 8> Mask(ShuffleVT.getVectorNumElements(), -1);
4668 int BitsPerShuffleLane = ShuffleVT.getVectorElementType().getSizeInBits();
4669 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
4670 SDValue Entry = Op.getOperand(i);
4671 if (Entry.getOpcode() == ISD::UNDEF)
4674 auto Src = std::find(Sources.begin(), Sources.end(), Entry.getOperand(0));
4675 int EltNo = cast<ConstantSDNode>(Entry.getOperand(1))->getSExtValue();
4677 // EXTRACT_VECTOR_ELT performs an implicit any_ext; BUILD_VECTOR an implicit
4678 // trunc. So only std::min(SrcBits, DestBits) actually get defined in this
4680 EVT OrigEltTy = Entry.getOperand(0).getValueType().getVectorElementType();
4681 int BitsDefined = std::min(OrigEltTy.getSizeInBits(),
4682 VT.getVectorElementType().getSizeInBits());
4683 int LanesDefined = BitsDefined / BitsPerShuffleLane;
4685 // This source is expected to fill ResMultiplier lanes of the final shuffle,
4686 // starting at the appropriate offset.
4687 int *LaneMask = &Mask[i * ResMultiplier];
4689 int ExtractBase = EltNo * Src->WindowScale + Src->WindowBase;
4690 ExtractBase += NumElts * (Src - Sources.begin());
4691 for (int j = 0; j < LanesDefined; ++j)
4692 LaneMask[j] = ExtractBase + j;
4695 // Final check before we try to produce nonsense...
4696 if (!isShuffleMaskLegal(Mask, ShuffleVT))
4699 SDValue ShuffleOps[] = { DAG.getUNDEF(ShuffleVT), DAG.getUNDEF(ShuffleVT) };
4700 for (unsigned i = 0; i < Sources.size(); ++i)
4701 ShuffleOps[i] = Sources[i].ShuffleVec;
4703 SDValue Shuffle = DAG.getVectorShuffle(ShuffleVT, dl, ShuffleOps[0],
4704 ShuffleOps[1], &Mask[0]);
4705 return DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
4708 // check if an EXT instruction can handle the shuffle mask when the
4709 // vector sources of the shuffle are the same.
4710 static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4711 unsigned NumElts = VT.getVectorNumElements();
4713 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4719 // If this is a VEXT shuffle, the immediate value is the index of the first
4720 // element. The other shuffle indices must be the successive elements after
4722 unsigned ExpectedElt = Imm;
4723 for (unsigned i = 1; i < NumElts; ++i) {
4724 // Increment the expected index. If it wraps around, just follow it
4725 // back to index zero and keep going.
4727 if (ExpectedElt == NumElts)
4731 continue; // ignore UNDEF indices
4732 if (ExpectedElt != static_cast<unsigned>(M[i]))
4739 // check if an EXT instruction can handle the shuffle mask when the
4740 // vector sources of the shuffle are different.
4741 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
4743 // Look for the first non-undef element.
4744 const int *FirstRealElt = std::find_if(M.begin(), M.end(),
4745 [](int Elt) {return Elt >= 0;});
4747 // Benefit form APInt to handle overflow when calculating expected element.
4748 unsigned NumElts = VT.getVectorNumElements();
4749 unsigned MaskBits = APInt(32, NumElts * 2).logBase2();
4750 APInt ExpectedElt = APInt(MaskBits, *FirstRealElt + 1);
4751 // The following shuffle indices must be the successive elements after the
4752 // first real element.
4753 const int *FirstWrongElt = std::find_if(FirstRealElt + 1, M.end(),
4754 [&](int Elt) {return Elt != ExpectedElt++ && Elt != -1;});
4755 if (FirstWrongElt != M.end())
4758 // The index of an EXT is the first element if it is not UNDEF.
4759 // Watch out for the beginning UNDEFs. The EXT index should be the expected
4760 // value of the first element. E.g.
4761 // <-1, -1, 3, ...> is treated as <1, 2, 3, ...>.
4762 // <-1, -1, 0, 1, ...> is treated as <2*NumElts-2, 2*NumElts-1, 0, 1, ...>.
4763 // ExpectedElt is the last mask index plus 1.
4764 Imm = ExpectedElt.getZExtValue();
4766 // There are two difference cases requiring to reverse input vectors.
4767 // For example, for vector <4 x i32> we have the following cases,
4768 // Case 1: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, -1, 0>)
4769 // Case 2: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, 7, 0>)
4770 // For both cases, we finally use mask <5, 6, 7, 0>, which requires
4771 // to reverse two input vectors.
4780 /// isREVMask - Check if a vector shuffle corresponds to a REV
4781 /// instruction with the specified blocksize. (The order of the elements
4782 /// within each block of the vector is reversed.)
4783 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4784 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
4785 "Only possible block sizes for REV are: 16, 32, 64");
4787 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4791 unsigned NumElts = VT.getVectorNumElements();
4792 unsigned BlockElts = M[0] + 1;
4793 // If the first shuffle index is UNDEF, be optimistic.
4795 BlockElts = BlockSize / EltSz;
4797 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4800 for (unsigned i = 0; i < NumElts; ++i) {
4802 continue; // ignore UNDEF indices
4803 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
4810 static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4811 unsigned NumElts = VT.getVectorNumElements();
4812 WhichResult = (M[0] == 0 ? 0 : 1);
4813 unsigned Idx = WhichResult * NumElts / 2;
4814 for (unsigned i = 0; i != NumElts; i += 2) {
4815 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4816 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
4824 static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4825 unsigned NumElts = VT.getVectorNumElements();
4826 WhichResult = (M[0] == 0 ? 0 : 1);
4827 for (unsigned i = 0; i != NumElts; ++i) {
4829 continue; // ignore UNDEF indices
4830 if ((unsigned)M[i] != 2 * i + WhichResult)
4837 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4838 unsigned NumElts = VT.getVectorNumElements();
4839 WhichResult = (M[0] == 0 ? 0 : 1);
4840 for (unsigned i = 0; i < NumElts; i += 2) {
4841 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4842 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + NumElts + WhichResult))
4848 /// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of
4849 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4850 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4851 static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4852 unsigned NumElts = VT.getVectorNumElements();
4853 WhichResult = (M[0] == 0 ? 0 : 1);
4854 unsigned Idx = WhichResult * NumElts / 2;
4855 for (unsigned i = 0; i != NumElts; i += 2) {
4856 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4857 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx))
4865 /// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of
4866 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4867 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4868 static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4869 unsigned Half = VT.getVectorNumElements() / 2;
4870 WhichResult = (M[0] == 0 ? 0 : 1);
4871 for (unsigned j = 0; j != 2; ++j) {
4872 unsigned Idx = WhichResult;
4873 for (unsigned i = 0; i != Half; ++i) {
4874 int MIdx = M[i + j * Half];
4875 if (MIdx >= 0 && (unsigned)MIdx != Idx)
4884 /// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of
4885 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4886 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4887 static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4888 unsigned NumElts = VT.getVectorNumElements();
4889 WhichResult = (M[0] == 0 ? 0 : 1);
4890 for (unsigned i = 0; i < NumElts; i += 2) {
4891 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4892 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult))
4898 static bool isINSMask(ArrayRef<int> M, int NumInputElements,
4899 bool &DstIsLeft, int &Anomaly) {
4900 if (M.size() != static_cast<size_t>(NumInputElements))
4903 int NumLHSMatch = 0, NumRHSMatch = 0;
4904 int LastLHSMismatch = -1, LastRHSMismatch = -1;
4906 for (int i = 0; i < NumInputElements; ++i) {
4916 LastLHSMismatch = i;
4918 if (M[i] == i + NumInputElements)
4921 LastRHSMismatch = i;
4924 if (NumLHSMatch == NumInputElements - 1) {
4926 Anomaly = LastLHSMismatch;
4928 } else if (NumRHSMatch == NumInputElements - 1) {
4930 Anomaly = LastRHSMismatch;
4937 static bool isConcatMask(ArrayRef<int> Mask, EVT VT, bool SplitLHS) {
4938 if (VT.getSizeInBits() != 128)
4941 unsigned NumElts = VT.getVectorNumElements();
4943 for (int I = 0, E = NumElts / 2; I != E; I++) {
4948 int Offset = NumElts / 2;
4949 for (int I = NumElts / 2, E = NumElts; I != E; I++) {
4950 if (Mask[I] != I + SplitLHS * Offset)
4957 static SDValue tryFormConcatFromShuffle(SDValue Op, SelectionDAG &DAG) {
4959 EVT VT = Op.getValueType();
4960 SDValue V0 = Op.getOperand(0);
4961 SDValue V1 = Op.getOperand(1);
4962 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op)->getMask();
4964 if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() ||
4965 VT.getVectorElementType() != V1.getValueType().getVectorElementType())
4968 bool SplitV0 = V0.getValueType().getSizeInBits() == 128;
4970 if (!isConcatMask(Mask, VT, SplitV0))
4973 EVT CastVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
4974 VT.getVectorNumElements() / 2);
4976 V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0,
4977 DAG.getConstant(0, MVT::i64));
4979 if (V1.getValueType().getSizeInBits() == 128) {
4980 V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1,
4981 DAG.getConstant(0, MVT::i64));
4983 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1);
4986 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4987 /// the specified operations to build the shuffle.
4988 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4989 SDValue RHS, SelectionDAG &DAG,
4991 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4992 unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1);
4993 unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1);
4996 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5005 OP_VUZPL, // VUZP, left result
5006 OP_VUZPR, // VUZP, right result
5007 OP_VZIPL, // VZIP, left result
5008 OP_VZIPR, // VZIP, right result
5009 OP_VTRNL, // VTRN, left result
5010 OP_VTRNR // VTRN, right result
5013 if (OpNum == OP_COPY) {
5014 if (LHSID == (1 * 9 + 2) * 9 + 3)
5016 assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!");
5020 SDValue OpLHS, OpRHS;
5021 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5022 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5023 EVT VT = OpLHS.getValueType();
5027 llvm_unreachable("Unknown shuffle opcode!");
5029 // VREV divides the vector in half and swaps within the half.
5030 if (VT.getVectorElementType() == MVT::i32 ||
5031 VT.getVectorElementType() == MVT::f32)
5032 return DAG.getNode(AArch64ISD::REV64, dl, VT, OpLHS);
5033 // vrev <4 x i16> -> REV32
5034 if (VT.getVectorElementType() == MVT::i16 ||
5035 VT.getVectorElementType() == MVT::f16)
5036 return DAG.getNode(AArch64ISD::REV32, dl, VT, OpLHS);
5037 // vrev <4 x i8> -> REV16
5038 assert(VT.getVectorElementType() == MVT::i8);
5039 return DAG.getNode(AArch64ISD::REV16, dl, VT, OpLHS);
5044 EVT EltTy = VT.getVectorElementType();
5046 if (EltTy == MVT::i8)
5047 Opcode = AArch64ISD::DUPLANE8;
5048 else if (EltTy == MVT::i16)
5049 Opcode = AArch64ISD::DUPLANE16;
5050 else if (EltTy == MVT::i32 || EltTy == MVT::f32)
5051 Opcode = AArch64ISD::DUPLANE32;
5052 else if (EltTy == MVT::i64 || EltTy == MVT::f64)
5053 Opcode = AArch64ISD::DUPLANE64;
5055 llvm_unreachable("Invalid vector element type?");
5057 if (VT.getSizeInBits() == 64)
5058 OpLHS = WidenVector(OpLHS, DAG);
5059 SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, MVT::i64);
5060 return DAG.getNode(Opcode, dl, VT, OpLHS, Lane);
5065 unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
5066 return DAG.getNode(AArch64ISD::EXT, dl, VT, OpLHS, OpRHS,
5067 DAG.getConstant(Imm, MVT::i32));
5070 return DAG.getNode(AArch64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS,
5073 return DAG.getNode(AArch64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS,
5076 return DAG.getNode(AArch64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS,
5079 return DAG.getNode(AArch64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS,
5082 return DAG.getNode(AArch64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS,
5085 return DAG.getNode(AArch64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS,
5090 static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask,
5091 SelectionDAG &DAG) {
5092 // Check to see if we can use the TBL instruction.
5093 SDValue V1 = Op.getOperand(0);
5094 SDValue V2 = Op.getOperand(1);
5097 EVT EltVT = Op.getValueType().getVectorElementType();
5098 unsigned BytesPerElt = EltVT.getSizeInBits() / 8;
5100 SmallVector<SDValue, 8> TBLMask;
5101 for (int Val : ShuffleMask) {
5102 for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
5103 unsigned Offset = Byte + Val * BytesPerElt;
5104 TBLMask.push_back(DAG.getConstant(Offset, MVT::i32));
5108 MVT IndexVT = MVT::v8i8;
5109 unsigned IndexLen = 8;
5110 if (Op.getValueType().getSizeInBits() == 128) {
5111 IndexVT = MVT::v16i8;
5115 SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
5116 SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);
5119 if (V2.getNode()->getOpcode() == ISD::UNDEF) {
5121 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
5122 Shuffle = DAG.getNode(
5123 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5124 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, MVT::i32), V1Cst,
5125 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5126 makeArrayRef(TBLMask.data(), IndexLen)));
5128 if (IndexLen == 8) {
5129 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
5130 Shuffle = DAG.getNode(
5131 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5132 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, MVT::i32), V1Cst,
5133 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5134 makeArrayRef(TBLMask.data(), IndexLen)));
5136 // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
5137 // cannot currently represent the register constraints on the input
5139 // Shuffle = DAG.getNode(AArch64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
5140 // DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5141 // &TBLMask[0], IndexLen));
5142 Shuffle = DAG.getNode(
5143 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5144 DAG.getConstant(Intrinsic::aarch64_neon_tbl2, MVT::i32), V1Cst, V2Cst,
5145 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5146 makeArrayRef(TBLMask.data(), IndexLen)));
5149 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
5152 static unsigned getDUPLANEOp(EVT EltType) {
5153 if (EltType == MVT::i8)
5154 return AArch64ISD::DUPLANE8;
5155 if (EltType == MVT::i16 || EltType == MVT::f16)
5156 return AArch64ISD::DUPLANE16;
5157 if (EltType == MVT::i32 || EltType == MVT::f32)
5158 return AArch64ISD::DUPLANE32;
5159 if (EltType == MVT::i64 || EltType == MVT::f64)
5160 return AArch64ISD::DUPLANE64;
5162 llvm_unreachable("Invalid vector element type?");
5165 SDValue AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
5166 SelectionDAG &DAG) const {
5168 EVT VT = Op.getValueType();
5170 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5172 // Convert shuffles that are directly supported on NEON to target-specific
5173 // DAG nodes, instead of keeping them as shuffles and matching them again
5174 // during code selection. This is more efficient and avoids the possibility
5175 // of inconsistencies between legalization and selection.
5176 ArrayRef<int> ShuffleMask = SVN->getMask();
5178 SDValue V1 = Op.getOperand(0);
5179 SDValue V2 = Op.getOperand(1);
5181 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0],
5182 V1.getValueType().getSimpleVT())) {
5183 int Lane = SVN->getSplatIndex();
5184 // If this is undef splat, generate it via "just" vdup, if possible.
5188 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
5189 return DAG.getNode(AArch64ISD::DUP, dl, V1.getValueType(),
5191 // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
5192 // constant. If so, we can just reference the lane's definition directly.
5193 if (V1.getOpcode() == ISD::BUILD_VECTOR &&
5194 !isa<ConstantSDNode>(V1.getOperand(Lane)))
5195 return DAG.getNode(AArch64ISD::DUP, dl, VT, V1.getOperand(Lane));
5197 // Otherwise, duplicate from the lane of the input vector.
5198 unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
5200 // SelectionDAGBuilder may have "helpfully" already extracted or conatenated
5201 // to make a vector of the same size as this SHUFFLE. We can ignore the
5202 // extract entirely, and canonicalise the concat using WidenVector.
5203 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
5204 Lane += cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
5205 V1 = V1.getOperand(0);
5206 } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
5207 unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
5208 Lane -= Idx * VT.getVectorNumElements() / 2;
5209 V1 = WidenVector(V1.getOperand(Idx), DAG);
5210 } else if (VT.getSizeInBits() == 64)
5211 V1 = WidenVector(V1, DAG);
5213 return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, MVT::i64));
5216 if (isREVMask(ShuffleMask, VT, 64))
5217 return DAG.getNode(AArch64ISD::REV64, dl, V1.getValueType(), V1, V2);
5218 if (isREVMask(ShuffleMask, VT, 32))
5219 return DAG.getNode(AArch64ISD::REV32, dl, V1.getValueType(), V1, V2);
5220 if (isREVMask(ShuffleMask, VT, 16))
5221 return DAG.getNode(AArch64ISD::REV16, dl, V1.getValueType(), V1, V2);
5223 bool ReverseEXT = false;
5225 if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
5228 Imm *= getExtFactor(V1);
5229 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V2,
5230 DAG.getConstant(Imm, MVT::i32));
5231 } else if (V2->getOpcode() == ISD::UNDEF &&
5232 isSingletonEXTMask(ShuffleMask, VT, Imm)) {
5233 Imm *= getExtFactor(V1);
5234 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V1,
5235 DAG.getConstant(Imm, MVT::i32));
5238 unsigned WhichResult;
5239 if (isZIPMask(ShuffleMask, VT, WhichResult)) {
5240 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
5241 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5243 if (isUZPMask(ShuffleMask, VT, WhichResult)) {
5244 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
5245 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5247 if (isTRNMask(ShuffleMask, VT, WhichResult)) {
5248 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
5249 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5252 if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5253 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
5254 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5256 if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5257 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
5258 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5260 if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5261 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
5262 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5265 SDValue Concat = tryFormConcatFromShuffle(Op, DAG);
5266 if (Concat.getNode())
5271 int NumInputElements = V1.getValueType().getVectorNumElements();
5272 if (isINSMask(ShuffleMask, NumInputElements, DstIsLeft, Anomaly)) {
5273 SDValue DstVec = DstIsLeft ? V1 : V2;
5274 SDValue DstLaneV = DAG.getConstant(Anomaly, MVT::i64);
5276 SDValue SrcVec = V1;
5277 int SrcLane = ShuffleMask[Anomaly];
5278 if (SrcLane >= NumInputElements) {
5280 SrcLane -= VT.getVectorNumElements();
5282 SDValue SrcLaneV = DAG.getConstant(SrcLane, MVT::i64);
5284 EVT ScalarVT = VT.getVectorElementType();
5286 if (ScalarVT.getSizeInBits() < 32 && ScalarVT.isInteger())
5287 ScalarVT = MVT::i32;
5290 ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5291 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ScalarVT, SrcVec, SrcLaneV),
5295 // If the shuffle is not directly supported and it has 4 elements, use
5296 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5297 unsigned NumElts = VT.getVectorNumElements();
5299 unsigned PFIndexes[4];
5300 for (unsigned i = 0; i != 4; ++i) {
5301 if (ShuffleMask[i] < 0)
5304 PFIndexes[i] = ShuffleMask[i];
5307 // Compute the index in the perfect shuffle table.
5308 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
5309 PFIndexes[2] * 9 + PFIndexes[3];
5310 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5311 unsigned Cost = (PFEntry >> 30);
5314 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5317 return GenerateTBL(Op, ShuffleMask, DAG);
5320 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
5322 EVT VT = BVN->getValueType(0);
5323 APInt SplatBits, SplatUndef;
5324 unsigned SplatBitSize;
5326 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5327 unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;
5329 for (unsigned i = 0; i < NumSplats; ++i) {
5330 CnstBits <<= SplatBitSize;
5331 UndefBits <<= SplatBitSize;
5332 CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits());
5333 UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits());
5342 SDValue AArch64TargetLowering::LowerVectorAND(SDValue Op,
5343 SelectionDAG &DAG) const {
5344 BuildVectorSDNode *BVN =
5345 dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5346 SDValue LHS = Op.getOperand(0);
5348 EVT VT = Op.getValueType();
5353 APInt CnstBits(VT.getSizeInBits(), 0);
5354 APInt UndefBits(VT.getSizeInBits(), 0);
5355 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5356 // We only have BIC vector immediate instruction, which is and-not.
5357 CnstBits = ~CnstBits;
5359 // We make use of a little bit of goto ickiness in order to avoid having to
5360 // duplicate the immediate matching logic for the undef toggled case.
5361 bool SecondTry = false;
5364 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5365 CnstBits = CnstBits.zextOrTrunc(64);
5366 uint64_t CnstVal = CnstBits.getZExtValue();
5368 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5369 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5370 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5371 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5372 DAG.getConstant(CnstVal, MVT::i32),
5373 DAG.getConstant(0, MVT::i32));
5374 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5377 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5378 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5379 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5380 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5381 DAG.getConstant(CnstVal, MVT::i32),
5382 DAG.getConstant(8, MVT::i32));
5383 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5386 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5387 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5388 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5389 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5390 DAG.getConstant(CnstVal, MVT::i32),
5391 DAG.getConstant(16, MVT::i32));
5392 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5395 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5396 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5397 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5398 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5399 DAG.getConstant(CnstVal, MVT::i32),
5400 DAG.getConstant(24, MVT::i32));
5401 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5404 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5405 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5406 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5407 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5408 DAG.getConstant(CnstVal, MVT::i32),
5409 DAG.getConstant(0, MVT::i32));
5410 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5413 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5414 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5415 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5416 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5417 DAG.getConstant(CnstVal, MVT::i32),
5418 DAG.getConstant(8, MVT::i32));
5419 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5426 CnstBits = ~UndefBits;
5430 // We can always fall back to a non-immediate AND.
5435 // Specialized code to quickly find if PotentialBVec is a BuildVector that
5436 // consists of only the same constant int value, returned in reference arg
5438 static bool isAllConstantBuildVector(const SDValue &PotentialBVec,
5439 uint64_t &ConstVal) {
5440 BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec);
5443 ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0));
5446 EVT VT = Bvec->getValueType(0);
5447 unsigned NumElts = VT.getVectorNumElements();
5448 for (unsigned i = 1; i < NumElts; ++i)
5449 if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt)
5451 ConstVal = FirstElt->getZExtValue();
5455 static unsigned getIntrinsicID(const SDNode *N) {
5456 unsigned Opcode = N->getOpcode();
5459 return Intrinsic::not_intrinsic;
5460 case ISD::INTRINSIC_WO_CHAIN: {
5461 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
5462 if (IID < Intrinsic::num_intrinsics)
5464 return Intrinsic::not_intrinsic;
5469 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
5470 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
5471 // BUILD_VECTORs with constant element C1, C2 is a constant, and C1 == ~C2.
5472 // Also, logical shift right -> sri, with the same structure.
5473 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
5474 EVT VT = N->getValueType(0);
5481 // Is the first op an AND?
5482 const SDValue And = N->getOperand(0);
5483 if (And.getOpcode() != ISD::AND)
5486 // Is the second op an shl or lshr?
5487 SDValue Shift = N->getOperand(1);
5488 // This will have been turned into: AArch64ISD::VSHL vector, #shift
5489 // or AArch64ISD::VLSHR vector, #shift
5490 unsigned ShiftOpc = Shift.getOpcode();
5491 if ((ShiftOpc != AArch64ISD::VSHL && ShiftOpc != AArch64ISD::VLSHR))
5493 bool IsShiftRight = ShiftOpc == AArch64ISD::VLSHR;
5495 // Is the shift amount constant?
5496 ConstantSDNode *C2node = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
5500 // Is the and mask vector all constant?
5502 if (!isAllConstantBuildVector(And.getOperand(1), C1))
5505 // Is C1 == ~C2, taking into account how much one can shift elements of a
5507 uint64_t C2 = C2node->getZExtValue();
5508 unsigned ElemSizeInBits = VT.getVectorElementType().getSizeInBits();
5509 if (C2 > ElemSizeInBits)
5511 unsigned ElemMask = (1 << ElemSizeInBits) - 1;
5512 if ((C1 & ElemMask) != (~C2 & ElemMask))
5515 SDValue X = And.getOperand(0);
5516 SDValue Y = Shift.getOperand(0);
5519 IsShiftRight ? Intrinsic::aarch64_neon_vsri : Intrinsic::aarch64_neon_vsli;
5521 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5522 DAG.getConstant(Intrin, MVT::i32), X, Y, Shift.getOperand(1));
5524 DEBUG(dbgs() << "aarch64-lower: transformed: \n");
5525 DEBUG(N->dump(&DAG));
5526 DEBUG(dbgs() << "into: \n");
5527 DEBUG(ResultSLI->dump(&DAG));
5533 SDValue AArch64TargetLowering::LowerVectorOR(SDValue Op,
5534 SelectionDAG &DAG) const {
5535 // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
5536 if (EnableAArch64SlrGeneration) {
5537 SDValue Res = tryLowerToSLI(Op.getNode(), DAG);
5542 BuildVectorSDNode *BVN =
5543 dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode());
5544 SDValue LHS = Op.getOperand(1);
5546 EVT VT = Op.getValueType();
5548 // OR commutes, so try swapping the operands.
5550 LHS = Op.getOperand(0);
5551 BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5556 APInt CnstBits(VT.getSizeInBits(), 0);
5557 APInt UndefBits(VT.getSizeInBits(), 0);
5558 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5559 // We make use of a little bit of goto ickiness in order to avoid having to
5560 // duplicate the immediate matching logic for the undef toggled case.
5561 bool SecondTry = false;
5564 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5565 CnstBits = CnstBits.zextOrTrunc(64);
5566 uint64_t CnstVal = CnstBits.getZExtValue();
5568 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5569 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5570 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5571 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5572 DAG.getConstant(CnstVal, MVT::i32),
5573 DAG.getConstant(0, MVT::i32));
5574 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5577 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5578 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5579 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5580 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5581 DAG.getConstant(CnstVal, MVT::i32),
5582 DAG.getConstant(8, MVT::i32));
5583 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5586 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5587 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5588 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5589 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5590 DAG.getConstant(CnstVal, MVT::i32),
5591 DAG.getConstant(16, MVT::i32));
5592 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5595 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5596 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5597 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5598 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5599 DAG.getConstant(CnstVal, MVT::i32),
5600 DAG.getConstant(24, MVT::i32));
5601 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5604 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5605 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5606 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5607 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5608 DAG.getConstant(CnstVal, MVT::i32),
5609 DAG.getConstant(0, MVT::i32));
5610 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5613 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5614 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5615 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5616 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5617 DAG.getConstant(CnstVal, MVT::i32),
5618 DAG.getConstant(8, MVT::i32));
5619 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5626 CnstBits = UndefBits;
5630 // We can always fall back to a non-immediate OR.
5635 // Normalize the operands of BUILD_VECTOR. The value of constant operands will
5636 // be truncated to fit element width.
5637 static SDValue NormalizeBuildVector(SDValue Op,
5638 SelectionDAG &DAG) {
5639 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
5641 EVT VT = Op.getValueType();
5642 EVT EltTy= VT.getVectorElementType();
5644 if (EltTy.isFloatingPoint() || EltTy.getSizeInBits() > 16)
5647 SmallVector<SDValue, 16> Ops;
5648 for (unsigned I = 0, E = VT.getVectorNumElements(); I != E; ++I) {
5649 SDValue Lane = Op.getOperand(I);
5650 if (Lane.getOpcode() == ISD::Constant) {
5651 APInt LowBits(EltTy.getSizeInBits(),
5652 cast<ConstantSDNode>(Lane)->getZExtValue());
5653 Lane = DAG.getConstant(LowBits.getZExtValue(), MVT::i32);
5655 Ops.push_back(Lane);
5657 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5660 SDValue AArch64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
5661 SelectionDAG &DAG) const {
5663 EVT VT = Op.getValueType();
5664 Op = NormalizeBuildVector(Op, DAG);
5665 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5667 APInt CnstBits(VT.getSizeInBits(), 0);
5668 APInt UndefBits(VT.getSizeInBits(), 0);
5669 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5670 // We make use of a little bit of goto ickiness in order to avoid having to
5671 // duplicate the immediate matching logic for the undef toggled case.
5672 bool SecondTry = false;
5675 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5676 CnstBits = CnstBits.zextOrTrunc(64);
5677 uint64_t CnstVal = CnstBits.getZExtValue();
5679 // Certain magic vector constants (used to express things like NOT
5680 // and NEG) are passed through unmodified. This allows codegen patterns
5681 // for these operations to match. Special-purpose patterns will lower
5682 // these immediates to MOVIs if it proves necessary.
5683 if (VT.isInteger() && (CnstVal == 0 || CnstVal == ~0ULL))
5686 // The many faces of MOVI...
5687 if (AArch64_AM::isAdvSIMDModImmType10(CnstVal)) {
5688 CnstVal = AArch64_AM::encodeAdvSIMDModImmType10(CnstVal);
5689 if (VT.getSizeInBits() == 128) {
5690 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::v2i64,
5691 DAG.getConstant(CnstVal, MVT::i32));
5692 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5695 // Support the V64 version via subregister insertion.
5696 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::f64,
5697 DAG.getConstant(CnstVal, MVT::i32));
5698 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5701 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5702 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5703 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5704 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5705 DAG.getConstant(CnstVal, MVT::i32),
5706 DAG.getConstant(0, MVT::i32));
5707 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5710 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5711 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5712 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5713 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5714 DAG.getConstant(CnstVal, MVT::i32),
5715 DAG.getConstant(8, MVT::i32));
5716 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5719 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5720 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5721 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5722 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5723 DAG.getConstant(CnstVal, MVT::i32),
5724 DAG.getConstant(16, MVT::i32));
5725 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5728 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5729 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5730 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5731 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5732 DAG.getConstant(CnstVal, MVT::i32),
5733 DAG.getConstant(24, MVT::i32));
5734 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5737 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5738 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5739 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5740 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5741 DAG.getConstant(CnstVal, MVT::i32),
5742 DAG.getConstant(0, MVT::i32));
5743 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5746 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5747 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5748 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5749 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5750 DAG.getConstant(CnstVal, MVT::i32),
5751 DAG.getConstant(8, MVT::i32));
5752 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5755 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
5756 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
5757 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5758 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
5759 DAG.getConstant(CnstVal, MVT::i32),
5760 DAG.getConstant(264, MVT::i32));
5761 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5764 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
5765 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
5766 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5767 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
5768 DAG.getConstant(CnstVal, MVT::i32),
5769 DAG.getConstant(272, MVT::i32));
5770 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5773 if (AArch64_AM::isAdvSIMDModImmType9(CnstVal)) {
5774 CnstVal = AArch64_AM::encodeAdvSIMDModImmType9(CnstVal);
5775 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;
5776 SDValue Mov = DAG.getNode(AArch64ISD::MOVI, dl, MovTy,
5777 DAG.getConstant(CnstVal, MVT::i32));
5778 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5781 // The few faces of FMOV...
5782 if (AArch64_AM::isAdvSIMDModImmType11(CnstVal)) {
5783 CnstVal = AArch64_AM::encodeAdvSIMDModImmType11(CnstVal);
5784 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4f32 : MVT::v2f32;
5785 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MovTy,
5786 DAG.getConstant(CnstVal, MVT::i32));
5787 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5790 if (AArch64_AM::isAdvSIMDModImmType12(CnstVal) &&
5791 VT.getSizeInBits() == 128) {
5792 CnstVal = AArch64_AM::encodeAdvSIMDModImmType12(CnstVal);
5793 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MVT::v2f64,
5794 DAG.getConstant(CnstVal, MVT::i32));
5795 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5798 // The many faces of MVNI...
5800 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5801 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5802 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5803 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5804 DAG.getConstant(CnstVal, MVT::i32),
5805 DAG.getConstant(0, MVT::i32));
5806 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5809 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5810 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5811 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5812 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5813 DAG.getConstant(CnstVal, MVT::i32),
5814 DAG.getConstant(8, MVT::i32));
5815 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5818 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5819 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5820 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5821 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5822 DAG.getConstant(CnstVal, MVT::i32),
5823 DAG.getConstant(16, MVT::i32));
5824 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5827 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5828 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5829 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5830 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5831 DAG.getConstant(CnstVal, MVT::i32),
5832 DAG.getConstant(24, MVT::i32));
5833 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5836 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5837 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5838 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5839 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5840 DAG.getConstant(CnstVal, MVT::i32),
5841 DAG.getConstant(0, MVT::i32));
5842 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5845 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5846 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5847 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5848 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5849 DAG.getConstant(CnstVal, MVT::i32),
5850 DAG.getConstant(8, MVT::i32));
5851 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5854 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
5855 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
5856 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5857 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
5858 DAG.getConstant(CnstVal, MVT::i32),
5859 DAG.getConstant(264, MVT::i32));
5860 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5863 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
5864 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
5865 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5866 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
5867 DAG.getConstant(CnstVal, MVT::i32),
5868 DAG.getConstant(272, MVT::i32));
5869 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5876 CnstBits = UndefBits;
5881 // Scan through the operands to find some interesting properties we can
5883 // 1) If only one value is used, we can use a DUP, or
5884 // 2) if only the low element is not undef, we can just insert that, or
5885 // 3) if only one constant value is used (w/ some non-constant lanes),
5886 // we can splat the constant value into the whole vector then fill
5887 // in the non-constant lanes.
5888 // 4) FIXME: If different constant values are used, but we can intelligently
5889 // select the values we'll be overwriting for the non-constant
5890 // lanes such that we can directly materialize the vector
5891 // some other way (MOVI, e.g.), we can be sneaky.
5892 unsigned NumElts = VT.getVectorNumElements();
5893 bool isOnlyLowElement = true;
5894 bool usesOnlyOneValue = true;
5895 bool usesOnlyOneConstantValue = true;
5896 bool isConstant = true;
5897 unsigned NumConstantLanes = 0;
5899 SDValue ConstantValue;
5900 for (unsigned i = 0; i < NumElts; ++i) {
5901 SDValue V = Op.getOperand(i);
5902 if (V.getOpcode() == ISD::UNDEF)
5905 isOnlyLowElement = false;
5906 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5909 if (isa<ConstantSDNode>(V) || isa<ConstantFPSDNode>(V)) {
5911 if (!ConstantValue.getNode())
5913 else if (ConstantValue != V)
5914 usesOnlyOneConstantValue = false;
5917 if (!Value.getNode())
5919 else if (V != Value)
5920 usesOnlyOneValue = false;
5923 if (!Value.getNode())
5924 return DAG.getUNDEF(VT);
5926 if (isOnlyLowElement)
5927 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5929 // Use DUP for non-constant splats. For f32 constant splats, reduce to
5930 // i32 and try again.
5931 if (usesOnlyOneValue) {
5933 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5934 Value.getValueType() != VT)
5935 return DAG.getNode(AArch64ISD::DUP, dl, VT, Value);
5937 // This is actually a DUPLANExx operation, which keeps everything vectory.
5939 // DUPLANE works on 128-bit vectors, widen it if necessary.
5940 SDValue Lane = Value.getOperand(1);
5941 Value = Value.getOperand(0);
5942 if (Value.getValueType().getSizeInBits() == 64)
5943 Value = WidenVector(Value, DAG);
5945 unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
5946 return DAG.getNode(Opcode, dl, VT, Value, Lane);
5949 if (VT.getVectorElementType().isFloatingPoint()) {
5950 SmallVector<SDValue, 8> Ops;
5951 EVT EltTy = VT.getVectorElementType();
5952 assert ((EltTy == MVT::f16 || EltTy == MVT::f32 || EltTy == MVT::f64) &&
5953 "Unsupported floating-point vector type");
5954 MVT NewType = MVT::getIntegerVT(EltTy.getSizeInBits());
5955 for (unsigned i = 0; i < NumElts; ++i)
5956 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, NewType, Op.getOperand(i)));
5957 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts);
5958 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5959 Val = LowerBUILD_VECTOR(Val, DAG);
5961 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5965 // If there was only one constant value used and for more than one lane,
5966 // start by splatting that value, then replace the non-constant lanes. This
5967 // is better than the default, which will perform a separate initialization
5969 if (NumConstantLanes > 0 && usesOnlyOneConstantValue) {
5970 SDValue Val = DAG.getNode(AArch64ISD::DUP, dl, VT, ConstantValue);
5971 // Now insert the non-constant lanes.
5972 for (unsigned i = 0; i < NumElts; ++i) {
5973 SDValue V = Op.getOperand(i);
5974 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5975 if (!isa<ConstantSDNode>(V) && !isa<ConstantFPSDNode>(V)) {
5976 // Note that type legalization likely mucked about with the VT of the
5977 // source operand, so we may have to convert it here before inserting.
5978 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, V, LaneIdx);
5984 // If all elements are constants and the case above didn't get hit, fall back
5985 // to the default expansion, which will generate a load from the constant
5990 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5992 SDValue shuffle = ReconstructShuffle(Op, DAG);
5993 if (shuffle != SDValue())
5997 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5998 // know the default expansion would otherwise fall back on something even
5999 // worse. For a vector with one or two non-undef values, that's
6000 // scalar_to_vector for the elements followed by a shuffle (provided the
6001 // shuffle is valid for the target) and materialization element by element
6002 // on the stack followed by a load for everything else.
6003 if (!isConstant && !usesOnlyOneValue) {
6004 SDValue Vec = DAG.getUNDEF(VT);
6005 SDValue Op0 = Op.getOperand(0);
6006 unsigned ElemSize = VT.getVectorElementType().getSizeInBits();
6008 // For 32 and 64 bit types, use INSERT_SUBREG for lane zero to
6009 // a) Avoid a RMW dependency on the full vector register, and
6010 // b) Allow the register coalescer to fold away the copy if the
6011 // value is already in an S or D register.
6012 if (Op0.getOpcode() != ISD::UNDEF && (ElemSize == 32 || ElemSize == 64)) {
6013 unsigned SubIdx = ElemSize == 32 ? AArch64::ssub : AArch64::dsub;
6015 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0,
6016 DAG.getTargetConstant(SubIdx, MVT::i32));
6017 Vec = SDValue(N, 0);
6020 for (; i < NumElts; ++i) {
6021 SDValue V = Op.getOperand(i);
6022 if (V.getOpcode() == ISD::UNDEF)
6024 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
6025 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
6030 // Just use the default expansion. We failed to find a better alternative.
6034 SDValue AArch64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
6035 SelectionDAG &DAG) const {
6036 assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!");
6038 // Check for non-constant or out of range lane.
6039 EVT VT = Op.getOperand(0).getValueType();
6040 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(2));
6041 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
6045 // Insertion/extraction are legal for V128 types.
6046 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
6047 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
6051 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
6052 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
6055 // For V64 types, we perform insertion by expanding the value
6056 // to a V128 type and perform the insertion on that.
6058 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
6059 EVT WideTy = WideVec.getValueType();
6061 SDValue Node = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideTy, WideVec,
6062 Op.getOperand(1), Op.getOperand(2));
6063 // Re-narrow the resultant vector.
6064 return NarrowVector(Node, DAG);
6068 AArch64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
6069 SelectionDAG &DAG) const {
6070 assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!");
6072 // Check for non-constant or out of range lane.
6073 EVT VT = Op.getOperand(0).getValueType();
6074 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6075 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
6079 // Insertion/extraction are legal for V128 types.
6080 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
6081 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
6085 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
6086 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
6089 // For V64 types, we perform extraction by expanding the value
6090 // to a V128 type and perform the extraction on that.
6092 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
6093 EVT WideTy = WideVec.getValueType();
6095 EVT ExtrTy = WideTy.getVectorElementType();
6096 if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8)
6099 // For extractions, we just return the result directly.
6100 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec,
6104 SDValue AArch64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
6105 SelectionDAG &DAG) const {
6106 EVT VT = Op.getOperand(0).getValueType();
6112 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6115 unsigned Val = Cst->getZExtValue();
6117 unsigned Size = Op.getValueType().getSizeInBits();
6121 return DAG.getTargetExtractSubreg(AArch64::bsub, dl, Op.getValueType(),
6124 return DAG.getTargetExtractSubreg(AArch64::hsub, dl, Op.getValueType(),
6127 return DAG.getTargetExtractSubreg(AArch64::ssub, dl, Op.getValueType(),
6130 return DAG.getTargetExtractSubreg(AArch64::dsub, dl, Op.getValueType(),
6133 llvm_unreachable("Unexpected vector type in extract_subvector!");
6136 // If this is extracting the upper 64-bits of a 128-bit vector, we match
6138 if (Size == 64 && Val * VT.getVectorElementType().getSizeInBits() == 64)
6144 bool AArch64TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
6146 if (VT.getVectorNumElements() == 4 &&
6147 (VT.is128BitVector() || VT.is64BitVector())) {
6148 unsigned PFIndexes[4];
6149 for (unsigned i = 0; i != 4; ++i) {
6153 PFIndexes[i] = M[i];
6156 // Compute the index in the perfect shuffle table.
6157 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
6158 PFIndexes[2] * 9 + PFIndexes[3];
6159 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
6160 unsigned Cost = (PFEntry >> 30);
6168 unsigned DummyUnsigned;
6170 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || isREVMask(M, VT, 64) ||
6171 isREVMask(M, VT, 32) || isREVMask(M, VT, 16) ||
6172 isEXTMask(M, VT, DummyBool, DummyUnsigned) ||
6173 // isTBLMask(M, VT) || // FIXME: Port TBL support from ARM.
6174 isTRNMask(M, VT, DummyUnsigned) || isUZPMask(M, VT, DummyUnsigned) ||
6175 isZIPMask(M, VT, DummyUnsigned) ||
6176 isTRN_v_undef_Mask(M, VT, DummyUnsigned) ||
6177 isUZP_v_undef_Mask(M, VT, DummyUnsigned) ||
6178 isZIP_v_undef_Mask(M, VT, DummyUnsigned) ||
6179 isINSMask(M, VT.getVectorNumElements(), DummyBool, DummyInt) ||
6180 isConcatMask(M, VT, VT.getSizeInBits() == 128));
6183 /// getVShiftImm - Check if this is a valid build_vector for the immediate
6184 /// operand of a vector shift operation, where all the elements of the
6185 /// build_vector must have the same constant integer value.
6186 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
6187 // Ignore bit_converts.
6188 while (Op.getOpcode() == ISD::BITCAST)
6189 Op = Op.getOperand(0);
6190 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
6191 APInt SplatBits, SplatUndef;
6192 unsigned SplatBitSize;
6194 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
6195 HasAnyUndefs, ElementBits) ||
6196 SplatBitSize > ElementBits)
6198 Cnt = SplatBits.getSExtValue();
6202 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
6203 /// operand of a vector shift left operation. That value must be in the range:
6204 /// 0 <= Value < ElementBits for a left shift; or
6205 /// 0 <= Value <= ElementBits for a long left shift.
6206 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
6207 assert(VT.isVector() && "vector shift count is not a vector type");
6208 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
6209 if (!getVShiftImm(Op, ElementBits, Cnt))
6211 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
6214 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
6215 /// operand of a vector shift right operation. For a shift opcode, the value
6216 /// is positive, but for an intrinsic the value count must be negative. The
6217 /// absolute value must be in the range:
6218 /// 1 <= |Value| <= ElementBits for a right shift; or
6219 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
6220 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
6222 assert(VT.isVector() && "vector shift count is not a vector type");
6223 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
6224 if (!getVShiftImm(Op, ElementBits, Cnt))
6228 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
6231 SDValue AArch64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
6232 SelectionDAG &DAG) const {
6233 EVT VT = Op.getValueType();
6237 if (!Op.getOperand(1).getValueType().isVector())
6239 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
6241 switch (Op.getOpcode()) {
6243 llvm_unreachable("unexpected shift opcode");
6246 if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
6247 return DAG.getNode(AArch64ISD::VSHL, SDLoc(Op), VT, Op.getOperand(0),
6248 DAG.getConstant(Cnt, MVT::i32));
6249 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
6250 DAG.getConstant(Intrinsic::aarch64_neon_ushl, MVT::i32),
6251 Op.getOperand(0), Op.getOperand(1));
6254 // Right shift immediate
6255 if (isVShiftRImm(Op.getOperand(1), VT, false, false, Cnt) &&
6258 (Op.getOpcode() == ISD::SRA) ? AArch64ISD::VASHR : AArch64ISD::VLSHR;
6259 return DAG.getNode(Opc, SDLoc(Op), VT, Op.getOperand(0),
6260 DAG.getConstant(Cnt, MVT::i32));
6263 // Right shift register. Note, there is not a shift right register
6264 // instruction, but the shift left register instruction takes a signed
6265 // value, where negative numbers specify a right shift.
6266 unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::aarch64_neon_sshl
6267 : Intrinsic::aarch64_neon_ushl;
6268 // negate the shift amount
6269 SDValue NegShift = DAG.getNode(AArch64ISD::NEG, DL, VT, Op.getOperand(1));
6270 SDValue NegShiftLeft =
6271 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
6272 DAG.getConstant(Opc, MVT::i32), Op.getOperand(0), NegShift);
6273 return NegShiftLeft;
6279 static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS,
6280 AArch64CC::CondCode CC, bool NoNans, EVT VT,
6281 SDLoc dl, SelectionDAG &DAG) {
6282 EVT SrcVT = LHS.getValueType();
6283 assert(VT.getSizeInBits() == SrcVT.getSizeInBits() &&
6284 "function only supposed to emit natural comparisons");
6286 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
6287 APInt CnstBits(VT.getSizeInBits(), 0);
6288 APInt UndefBits(VT.getSizeInBits(), 0);
6289 bool IsCnst = BVN && resolveBuildVector(BVN, CnstBits, UndefBits);
6290 bool IsZero = IsCnst && (CnstBits == 0);
6292 if (SrcVT.getVectorElementType().isFloatingPoint()) {
6296 case AArch64CC::NE: {
6299 Fcmeq = DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
6301 Fcmeq = DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
6302 return DAG.getNode(AArch64ISD::NOT, dl, VT, Fcmeq);
6306 return DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
6307 return DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
6310 return DAG.getNode(AArch64ISD::FCMGEz, dl, VT, LHS);
6311 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, LHS, RHS);
6314 return DAG.getNode(AArch64ISD::FCMGTz, dl, VT, LHS);
6315 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, LHS, RHS);
6318 return DAG.getNode(AArch64ISD::FCMLEz, dl, VT, LHS);
6319 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, RHS, LHS);
6323 // If we ignore NaNs then we can use to the MI implementation.
6327 return DAG.getNode(AArch64ISD::FCMLTz, dl, VT, LHS);
6328 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, RHS, LHS);
6335 case AArch64CC::NE: {
6338 Cmeq = DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
6340 Cmeq = DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
6341 return DAG.getNode(AArch64ISD::NOT, dl, VT, Cmeq);
6345 return DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
6346 return DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
6349 return DAG.getNode(AArch64ISD::CMGEz, dl, VT, LHS);
6350 return DAG.getNode(AArch64ISD::CMGE, dl, VT, LHS, RHS);
6353 return DAG.getNode(AArch64ISD::CMGTz, dl, VT, LHS);
6354 return DAG.getNode(AArch64ISD::CMGT, dl, VT, LHS, RHS);
6357 return DAG.getNode(AArch64ISD::CMLEz, dl, VT, LHS);
6358 return DAG.getNode(AArch64ISD::CMGE, dl, VT, RHS, LHS);
6360 return DAG.getNode(AArch64ISD::CMHS, dl, VT, RHS, LHS);
6362 return DAG.getNode(AArch64ISD::CMHI, dl, VT, RHS, LHS);
6365 return DAG.getNode(AArch64ISD::CMLTz, dl, VT, LHS);
6366 return DAG.getNode(AArch64ISD::CMGT, dl, VT, RHS, LHS);
6368 return DAG.getNode(AArch64ISD::CMHI, dl, VT, LHS, RHS);
6370 return DAG.getNode(AArch64ISD::CMHS, dl, VT, LHS, RHS);
6374 SDValue AArch64TargetLowering::LowerVSETCC(SDValue Op,
6375 SelectionDAG &DAG) const {
6376 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6377 SDValue LHS = Op.getOperand(0);
6378 SDValue RHS = Op.getOperand(1);
6379 EVT CmpVT = LHS.getValueType().changeVectorElementTypeToInteger();
6382 if (LHS.getValueType().getVectorElementType().isInteger()) {
6383 assert(LHS.getValueType() == RHS.getValueType());
6384 AArch64CC::CondCode AArch64CC = changeIntCCToAArch64CC(CC);
6386 EmitVectorComparison(LHS, RHS, AArch64CC, false, CmpVT, dl, DAG);
6387 return DAG.getSExtOrTrunc(Cmp, dl, Op.getValueType());
6390 assert(LHS.getValueType().getVectorElementType() == MVT::f32 ||
6391 LHS.getValueType().getVectorElementType() == MVT::f64);
6393 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
6394 // clean. Some of them require two branches to implement.
6395 AArch64CC::CondCode CC1, CC2;
6397 changeVectorFPCCToAArch64CC(CC, CC1, CC2, ShouldInvert);
6399 bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath;
6401 EmitVectorComparison(LHS, RHS, CC1, NoNaNs, CmpVT, dl, DAG);
6405 if (CC2 != AArch64CC::AL) {
6407 EmitVectorComparison(LHS, RHS, CC2, NoNaNs, CmpVT, dl, DAG);
6408 if (!Cmp2.getNode())
6411 Cmp = DAG.getNode(ISD::OR, dl, CmpVT, Cmp, Cmp2);
6414 Cmp = DAG.getSExtOrTrunc(Cmp, dl, Op.getValueType());
6417 return Cmp = DAG.getNOT(dl, Cmp, Cmp.getValueType());
6422 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
6423 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
6424 /// specified in the intrinsic calls.
6425 bool AArch64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
6427 unsigned Intrinsic) const {
6428 switch (Intrinsic) {
6429 case Intrinsic::aarch64_neon_ld2:
6430 case Intrinsic::aarch64_neon_ld3:
6431 case Intrinsic::aarch64_neon_ld4:
6432 case Intrinsic::aarch64_neon_ld1x2:
6433 case Intrinsic::aarch64_neon_ld1x3:
6434 case Intrinsic::aarch64_neon_ld1x4:
6435 case Intrinsic::aarch64_neon_ld2lane:
6436 case Intrinsic::aarch64_neon_ld3lane:
6437 case Intrinsic::aarch64_neon_ld4lane:
6438 case Intrinsic::aarch64_neon_ld2r:
6439 case Intrinsic::aarch64_neon_ld3r:
6440 case Intrinsic::aarch64_neon_ld4r: {
6441 Info.opc = ISD::INTRINSIC_W_CHAIN;
6442 // Conservatively set memVT to the entire set of vectors loaded.
6443 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
6444 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6445 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6448 Info.vol = false; // volatile loads with NEON intrinsics not supported
6449 Info.readMem = true;
6450 Info.writeMem = false;
6453 case Intrinsic::aarch64_neon_st2:
6454 case Intrinsic::aarch64_neon_st3:
6455 case Intrinsic::aarch64_neon_st4:
6456 case Intrinsic::aarch64_neon_st1x2:
6457 case Intrinsic::aarch64_neon_st1x3:
6458 case Intrinsic::aarch64_neon_st1x4:
6459 case Intrinsic::aarch64_neon_st2lane:
6460 case Intrinsic::aarch64_neon_st3lane:
6461 case Intrinsic::aarch64_neon_st4lane: {
6462 Info.opc = ISD::INTRINSIC_VOID;
6463 // Conservatively set memVT to the entire set of vectors stored.
6464 unsigned NumElts = 0;
6465 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
6466 Type *ArgTy = I.getArgOperand(ArgI)->getType();
6467 if (!ArgTy->isVectorTy())
6469 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
6471 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6472 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6475 Info.vol = false; // volatile stores with NEON intrinsics not supported
6476 Info.readMem = false;
6477 Info.writeMem = true;
6480 case Intrinsic::aarch64_ldaxr:
6481 case Intrinsic::aarch64_ldxr: {
6482 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
6483 Info.opc = ISD::INTRINSIC_W_CHAIN;
6484 Info.memVT = MVT::getVT(PtrTy->getElementType());
6485 Info.ptrVal = I.getArgOperand(0);
6487 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
6489 Info.readMem = true;
6490 Info.writeMem = false;
6493 case Intrinsic::aarch64_stlxr:
6494 case Intrinsic::aarch64_stxr: {
6495 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
6496 Info.opc = ISD::INTRINSIC_W_CHAIN;
6497 Info.memVT = MVT::getVT(PtrTy->getElementType());
6498 Info.ptrVal = I.getArgOperand(1);
6500 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
6502 Info.readMem = false;
6503 Info.writeMem = true;
6506 case Intrinsic::aarch64_ldaxp:
6507 case Intrinsic::aarch64_ldxp: {
6508 Info.opc = ISD::INTRINSIC_W_CHAIN;
6509 Info.memVT = MVT::i128;
6510 Info.ptrVal = I.getArgOperand(0);
6514 Info.readMem = true;
6515 Info.writeMem = false;
6518 case Intrinsic::aarch64_stlxp:
6519 case Intrinsic::aarch64_stxp: {
6520 Info.opc = ISD::INTRINSIC_W_CHAIN;
6521 Info.memVT = MVT::i128;
6522 Info.ptrVal = I.getArgOperand(2);
6526 Info.readMem = false;
6527 Info.writeMem = true;
6537 // Truncations from 64-bit GPR to 32-bit GPR is free.
6538 bool AArch64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
6539 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6541 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6542 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6543 return NumBits1 > NumBits2;
6545 bool AArch64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
6546 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6548 unsigned NumBits1 = VT1.getSizeInBits();
6549 unsigned NumBits2 = VT2.getSizeInBits();
6550 return NumBits1 > NumBits2;
6553 /// Check if it is profitable to hoist instruction in then/else to if.
6554 /// Not profitable if I and it's user can form a FMA instruction
6555 /// because we prefer FMSUB/FMADD.
6556 bool AArch64TargetLowering::isProfitableToHoist(Instruction *I) const {
6557 if (I->getOpcode() != Instruction::FMul)
6560 if (I->getNumUses() != 1)
6563 Instruction *User = I->user_back();
6566 !(User->getOpcode() == Instruction::FSub ||
6567 User->getOpcode() == Instruction::FAdd))
6570 const TargetOptions &Options = getTargetMachine().Options;
6571 EVT VT = getValueType(User->getOperand(0)->getType());
6573 if (isFMAFasterThanFMulAndFAdd(VT) &&
6574 isOperationLegalOrCustom(ISD::FMA, VT) &&
6575 (Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath))
6581 // All 32-bit GPR operations implicitly zero the high-half of the corresponding
6583 bool AArch64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
6584 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6586 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6587 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6588 return NumBits1 == 32 && NumBits2 == 64;
6590 bool AArch64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
6591 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6593 unsigned NumBits1 = VT1.getSizeInBits();
6594 unsigned NumBits2 = VT2.getSizeInBits();
6595 return NumBits1 == 32 && NumBits2 == 64;
6598 bool AArch64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
6599 EVT VT1 = Val.getValueType();
6600 if (isZExtFree(VT1, VT2)) {
6604 if (Val.getOpcode() != ISD::LOAD)
6607 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
6608 return (VT1.isSimple() && !VT1.isVector() && VT1.isInteger() &&
6609 VT2.isSimple() && !VT2.isVector() && VT2.isInteger() &&
6610 VT1.getSizeInBits() <= 32);
6613 bool AArch64TargetLowering::isExtFreeImpl(const Instruction *Ext) const {
6614 if (isa<FPExtInst>(Ext))
6617 // Vector types are next free.
6618 if (Ext->getType()->isVectorTy())
6621 for (const Use &U : Ext->uses()) {
6622 // The extension is free if we can fold it with a left shift in an
6623 // addressing mode or an arithmetic operation: add, sub, and cmp.
6625 // Is there a shift?
6626 const Instruction *Instr = cast<Instruction>(U.getUser());
6628 // Is this a constant shift?
6629 switch (Instr->getOpcode()) {
6630 case Instruction::Shl:
6631 if (!isa<ConstantInt>(Instr->getOperand(1)))
6634 case Instruction::GetElementPtr: {
6635 gep_type_iterator GTI = gep_type_begin(Instr);
6636 std::advance(GTI, U.getOperandNo());
6638 // This extension will end up with a shift because of the scaling factor.
6639 // 8-bit sized types have a scaling factor of 1, thus a shift amount of 0.
6640 // Get the shift amount based on the scaling factor:
6641 // log2(sizeof(IdxTy)) - log2(8).
6643 countTrailingZeros(getDataLayout()->getTypeStoreSizeInBits(IdxTy)) - 3;
6644 // Is the constant foldable in the shift of the addressing mode?
6645 // I.e., shift amount is between 1 and 4 inclusive.
6646 if (ShiftAmt == 0 || ShiftAmt > 4)
6650 case Instruction::Trunc:
6651 // Check if this is a noop.
6652 // trunc(sext ty1 to ty2) to ty1.
6653 if (Instr->getType() == Ext->getOperand(0)->getType())
6660 // At this point we can use the bfm family, so this extension is free
6666 bool AArch64TargetLowering::hasPairedLoad(Type *LoadedType,
6667 unsigned &RequiredAligment) const {
6668 if (!LoadedType->isIntegerTy() && !LoadedType->isFloatTy())
6670 // Cyclone supports unaligned accesses.
6671 RequiredAligment = 0;
6672 unsigned NumBits = LoadedType->getPrimitiveSizeInBits();
6673 return NumBits == 32 || NumBits == 64;
6676 bool AArch64TargetLowering::hasPairedLoad(EVT LoadedType,
6677 unsigned &RequiredAligment) const {
6678 if (!LoadedType.isSimple() ||
6679 (!LoadedType.isInteger() && !LoadedType.isFloatingPoint()))
6681 // Cyclone supports unaligned accesses.
6682 RequiredAligment = 0;
6683 unsigned NumBits = LoadedType.getSizeInBits();
6684 return NumBits == 32 || NumBits == 64;
6687 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
6688 unsigned AlignCheck) {
6689 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
6690 (DstAlign == 0 || DstAlign % AlignCheck == 0));
6693 EVT AArch64TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
6694 unsigned SrcAlign, bool IsMemset,
6697 MachineFunction &MF) const {
6698 // Don't use AdvSIMD to implement 16-byte memset. It would have taken one
6699 // instruction to materialize the v2i64 zero and one store (with restrictive
6700 // addressing mode). Just do two i64 store of zero-registers.
6702 const Function *F = MF.getFunction();
6703 if (Subtarget->hasFPARMv8() && !IsMemset && Size >= 16 &&
6704 !F->hasFnAttribute(Attribute::NoImplicitFloat) &&
6705 (memOpAlign(SrcAlign, DstAlign, 16) ||
6706 (allowsMisalignedMemoryAccesses(MVT::f128, 0, 1, &Fast) && Fast)))
6710 (memOpAlign(SrcAlign, DstAlign, 8) ||
6711 (allowsMisalignedMemoryAccesses(MVT::i64, 0, 1, &Fast) && Fast)))
6715 (memOpAlign(SrcAlign, DstAlign, 4) ||
6716 (allowsMisalignedMemoryAccesses(MVT::i32, 0, 1, &Fast) && Fast)))
6722 // 12-bit optionally shifted immediates are legal for adds.
6723 bool AArch64TargetLowering::isLegalAddImmediate(int64_t Immed) const {
6724 if ((Immed >> 12) == 0 || ((Immed & 0xfff) == 0 && Immed >> 24 == 0))
6729 // Integer comparisons are implemented with ADDS/SUBS, so the range of valid
6730 // immediates is the same as for an add or a sub.
6731 bool AArch64TargetLowering::isLegalICmpImmediate(int64_t Immed) const {
6734 return isLegalAddImmediate(Immed);
6737 /// isLegalAddressingMode - Return true if the addressing mode represented
6738 /// by AM is legal for this target, for a load/store of the specified type.
6739 bool AArch64TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6741 // AArch64 has five basic addressing modes:
6743 // reg + 9-bit signed offset
6744 // reg + SIZE_IN_BYTES * 12-bit unsigned offset
6746 // reg + SIZE_IN_BYTES * reg
6748 // No global is ever allowed as a base.
6752 // No reg+reg+imm addressing.
6753 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
6756 // check reg + imm case:
6757 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
6758 uint64_t NumBytes = 0;
6759 if (Ty->isSized()) {
6760 uint64_t NumBits = getDataLayout()->getTypeSizeInBits(Ty);
6761 NumBytes = NumBits / 8;
6762 if (!isPowerOf2_64(NumBits))
6767 int64_t Offset = AM.BaseOffs;
6769 // 9-bit signed offset
6770 if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
6773 // 12-bit unsigned offset
6774 unsigned shift = Log2_64(NumBytes);
6775 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
6776 // Must be a multiple of NumBytes (NumBytes is a power of 2)
6777 (Offset >> shift) << shift == Offset)
6782 // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2
6784 if (!AM.Scale || AM.Scale == 1 ||
6785 (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
6790 int AArch64TargetLowering::getScalingFactorCost(const AddrMode &AM,
6792 // Scaling factors are not free at all.
6793 // Operands | Rt Latency
6794 // -------------------------------------------
6796 // -------------------------------------------
6797 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
6798 // Rt, [Xn, Wm, <extend> #imm] |
6799 if (isLegalAddressingMode(AM, Ty))
6800 // Scale represents reg2 * scale, thus account for 1 if
6801 // it is not equal to 0 or 1.
6802 return AM.Scale != 0 && AM.Scale != 1;
6806 bool AArch64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
6807 VT = VT.getScalarType();
6812 switch (VT.getSimpleVT().SimpleTy) {
6824 AArch64TargetLowering::getScratchRegisters(CallingConv::ID) const {
6825 // LR is a callee-save register, but we must treat it as clobbered by any call
6826 // site. Hence we include LR in the scratch registers, which are in turn added
6827 // as implicit-defs for stackmaps and patchpoints.
6828 static const MCPhysReg ScratchRegs[] = {
6829 AArch64::X16, AArch64::X17, AArch64::LR, 0
6835 AArch64TargetLowering::isDesirableToCommuteWithShift(const SDNode *N) const {
6836 EVT VT = N->getValueType(0);
6837 // If N is unsigned bit extraction: ((x >> C) & mask), then do not combine
6838 // it with shift to let it be lowered to UBFX.
6839 if (N->getOpcode() == ISD::AND && (VT == MVT::i32 || VT == MVT::i64) &&
6840 isa<ConstantSDNode>(N->getOperand(1))) {
6841 uint64_t TruncMask = N->getConstantOperandVal(1);
6842 if (isMask_64(TruncMask) &&
6843 N->getOperand(0).getOpcode() == ISD::SRL &&
6844 isa<ConstantSDNode>(N->getOperand(0)->getOperand(1)))
6850 bool AArch64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
6852 assert(Ty->isIntegerTy());
6854 unsigned BitSize = Ty->getPrimitiveSizeInBits();
6858 int64_t Val = Imm.getSExtValue();
6859 if (Val == 0 || AArch64_AM::isLogicalImmediate(Val, BitSize))
6862 if ((int64_t)Val < 0)
6865 Val &= (1LL << 32) - 1;
6867 unsigned LZ = countLeadingZeros((uint64_t)Val);
6868 unsigned Shift = (63 - LZ) / 16;
6869 // MOVZ is free so return true for one or fewer MOVK.
6873 // Generate SUBS and CSEL for integer abs.
6874 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
6875 EVT VT = N->getValueType(0);
6877 SDValue N0 = N->getOperand(0);
6878 SDValue N1 = N->getOperand(1);
6881 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
6882 // and change it to SUB and CSEL.
6883 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
6884 N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
6885 N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0))
6886 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
6887 if (Y1C->getAPIntValue() == VT.getSizeInBits() - 1) {
6888 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
6890 // Generate SUBS & CSEL.
6892 DAG.getNode(AArch64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32),
6893 N0.getOperand(0), DAG.getConstant(0, VT));
6894 return DAG.getNode(AArch64ISD::CSEL, DL, VT, N0.getOperand(0), Neg,
6895 DAG.getConstant(AArch64CC::PL, MVT::i32),
6896 SDValue(Cmp.getNode(), 1));
6901 // performXorCombine - Attempts to handle integer ABS.
6902 static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
6903 TargetLowering::DAGCombinerInfo &DCI,
6904 const AArch64Subtarget *Subtarget) {
6905 if (DCI.isBeforeLegalizeOps())
6908 return performIntegerAbsCombine(N, DAG);
6912 AArch64TargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor,
6914 std::vector<SDNode *> *Created) const {
6915 // fold (sdiv X, pow2)
6916 EVT VT = N->getValueType(0);
6917 if ((VT != MVT::i32 && VT != MVT::i64) ||
6918 !(Divisor.isPowerOf2() || (-Divisor).isPowerOf2()))
6922 SDValue N0 = N->getOperand(0);
6923 unsigned Lg2 = Divisor.countTrailingZeros();
6924 SDValue Zero = DAG.getConstant(0, VT);
6925 SDValue Pow2MinusOne = DAG.getConstant((1ULL << Lg2) - 1, VT);
6927 // Add (N0 < 0) ? Pow2 - 1 : 0;
6929 SDValue Cmp = getAArch64Cmp(N0, Zero, ISD::SETLT, CCVal, DAG, DL);
6930 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0, Pow2MinusOne);
6931 SDValue CSel = DAG.getNode(AArch64ISD::CSEL, DL, VT, Add, N0, CCVal, Cmp);
6934 Created->push_back(Cmp.getNode());
6935 Created->push_back(Add.getNode());
6936 Created->push_back(CSel.getNode());
6941 DAG.getNode(ISD::SRA, DL, VT, CSel, DAG.getConstant(Lg2, MVT::i64));
6943 // If we're dividing by a positive value, we're done. Otherwise, we must
6944 // negate the result.
6945 if (Divisor.isNonNegative())
6949 Created->push_back(SRA.getNode());
6950 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT), SRA);
6953 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG,
6954 TargetLowering::DAGCombinerInfo &DCI,
6955 const AArch64Subtarget *Subtarget) {
6956 if (DCI.isBeforeLegalizeOps())
6959 // Multiplication of a power of two plus/minus one can be done more
6960 // cheaply as as shift+add/sub. For now, this is true unilaterally. If
6961 // future CPUs have a cheaper MADD instruction, this may need to be
6962 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
6963 // 64-bit is 5 cycles, so this is always a win.
6964 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
6965 APInt Value = C->getAPIntValue();
6966 EVT VT = N->getValueType(0);
6967 if (Value.isNonNegative()) {
6968 // (mul x, 2^N + 1) => (add (shl x, N), x)
6969 APInt VM1 = Value - 1;
6970 if (VM1.isPowerOf2()) {
6971 SDValue ShiftedVal =
6972 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6973 DAG.getConstant(VM1.logBase2(), MVT::i64));
6974 return DAG.getNode(ISD::ADD, SDLoc(N), VT, ShiftedVal,
6977 // (mul x, 2^N - 1) => (sub (shl x, N), x)
6978 APInt VP1 = Value + 1;
6979 if (VP1.isPowerOf2()) {
6980 SDValue ShiftedVal =
6981 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6982 DAG.getConstant(VP1.logBase2(), MVT::i64));
6983 return DAG.getNode(ISD::SUB, SDLoc(N), VT, ShiftedVal,
6987 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
6988 APInt VNP1 = -Value + 1;
6989 if (VNP1.isPowerOf2()) {
6990 SDValue ShiftedVal =
6991 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6992 DAG.getConstant(VNP1.logBase2(), MVT::i64));
6993 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N->getOperand(0),
6996 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
6997 APInt VNM1 = -Value - 1;
6998 if (VNM1.isPowerOf2()) {
6999 SDValue ShiftedVal =
7000 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
7001 DAG.getConstant(VNM1.logBase2(), MVT::i64));
7003 DAG.getNode(ISD::ADD, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
7004 return DAG.getNode(ISD::SUB, SDLoc(N), VT, DAG.getConstant(0, VT), Add);
7011 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
7012 SelectionDAG &DAG) {
7013 // Take advantage of vector comparisons producing 0 or -1 in each lane to
7014 // optimize away operation when it's from a constant.
7016 // The general transformation is:
7017 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
7018 // AND(VECTOR_CMP(x,y), constant2)
7019 // constant2 = UNARYOP(constant)
7021 // Early exit if this isn't a vector operation, the operand of the
7022 // unary operation isn't a bitwise AND, or if the sizes of the operations
7024 EVT VT = N->getValueType(0);
7025 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
7026 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
7027 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
7030 // Now check that the other operand of the AND is a constant. We could
7031 // make the transformation for non-constant splats as well, but it's unclear
7032 // that would be a benefit as it would not eliminate any operations, just
7033 // perform one more step in scalar code before moving to the vector unit.
7034 if (BuildVectorSDNode *BV =
7035 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
7036 // Bail out if the vector isn't a constant.
7037 if (!BV->isConstant())
7040 // Everything checks out. Build up the new and improved node.
7042 EVT IntVT = BV->getValueType(0);
7043 // Create a new constant of the appropriate type for the transformed
7045 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
7046 // The AND node needs bitcasts to/from an integer vector type around it.
7047 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
7048 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
7049 N->getOperand(0)->getOperand(0), MaskConst);
7050 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
7057 static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG,
7058 const AArch64Subtarget *Subtarget) {
7059 // First try to optimize away the conversion when it's conditionally from
7060 // a constant. Vectors only.
7061 SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
7062 if (Res != SDValue())
7065 EVT VT = N->getValueType(0);
7066 if (VT != MVT::f32 && VT != MVT::f64)
7069 // Only optimize when the source and destination types have the same width.
7070 if (VT.getSizeInBits() != N->getOperand(0).getValueType().getSizeInBits())
7073 // If the result of an integer load is only used by an integer-to-float
7074 // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
7075 // This eliminates an "integer-to-vector-move UOP and improve throughput.
7076 SDValue N0 = N->getOperand(0);
7077 if (Subtarget->hasNEON() && ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7078 // Do not change the width of a volatile load.
7079 !cast<LoadSDNode>(N0)->isVolatile()) {
7080 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7081 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
7082 LN0->getPointerInfo(), LN0->isVolatile(),
7083 LN0->isNonTemporal(), LN0->isInvariant(),
7084 LN0->getAlignment());
7086 // Make sure successors of the original load stay after it by updating them
7087 // to use the new Chain.
7088 DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));
7091 (N->getOpcode() == ISD::SINT_TO_FP) ? AArch64ISD::SITOF : AArch64ISD::UITOF;
7092 return DAG.getNode(Opcode, SDLoc(N), VT, Load);
7098 /// An EXTR instruction is made up of two shifts, ORed together. This helper
7099 /// searches for and classifies those shifts.
7100 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
7102 if (N.getOpcode() == ISD::SHL)
7104 else if (N.getOpcode() == ISD::SRL)
7109 if (!isa<ConstantSDNode>(N.getOperand(1)))
7112 ShiftAmount = N->getConstantOperandVal(1);
7113 Src = N->getOperand(0);
7117 /// EXTR instruction extracts a contiguous chunk of bits from two existing
7118 /// registers viewed as a high/low pair. This function looks for the pattern:
7119 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
7120 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
7122 static SDValue tryCombineToEXTR(SDNode *N,
7123 TargetLowering::DAGCombinerInfo &DCI) {
7124 SelectionDAG &DAG = DCI.DAG;
7126 EVT VT = N->getValueType(0);
7128 assert(N->getOpcode() == ISD::OR && "Unexpected root");
7130 if (VT != MVT::i32 && VT != MVT::i64)
7134 uint32_t ShiftLHS = 0;
7136 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
7140 uint32_t ShiftRHS = 0;
7142 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
7145 // If they're both trying to come from the high part of the register, they're
7146 // not really an EXTR.
7147 if (LHSFromHi == RHSFromHi)
7150 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
7154 std::swap(LHS, RHS);
7155 std::swap(ShiftLHS, ShiftRHS);
7158 return DAG.getNode(AArch64ISD::EXTR, DL, VT, LHS, RHS,
7159 DAG.getConstant(ShiftRHS, MVT::i64));
7162 static SDValue tryCombineToBSL(SDNode *N,
7163 TargetLowering::DAGCombinerInfo &DCI) {
7164 EVT VT = N->getValueType(0);
7165 SelectionDAG &DAG = DCI.DAG;
7171 SDValue N0 = N->getOperand(0);
7172 if (N0.getOpcode() != ISD::AND)
7175 SDValue N1 = N->getOperand(1);
7176 if (N1.getOpcode() != ISD::AND)
7179 // We only have to look for constant vectors here since the general, variable
7180 // case can be handled in TableGen.
7181 unsigned Bits = VT.getVectorElementType().getSizeInBits();
7182 uint64_t BitMask = Bits == 64 ? -1ULL : ((1ULL << Bits) - 1);
7183 for (int i = 1; i >= 0; --i)
7184 for (int j = 1; j >= 0; --j) {
7185 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(i));
7186 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(j));
7190 bool FoundMatch = true;
7191 for (unsigned k = 0; k < VT.getVectorNumElements(); ++k) {
7192 ConstantSDNode *CN0 = dyn_cast<ConstantSDNode>(BVN0->getOperand(k));
7193 ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(BVN1->getOperand(k));
7195 CN0->getZExtValue() != (BitMask & ~CN1->getZExtValue())) {
7202 return DAG.getNode(AArch64ISD::BSL, DL, VT, SDValue(BVN0, 0),
7203 N0->getOperand(1 - i), N1->getOperand(1 - j));
7209 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
7210 const AArch64Subtarget *Subtarget) {
7211 // Attempt to form an EXTR from (or (shl VAL1, #N), (srl VAL2, #RegWidth-N))
7212 if (!EnableAArch64ExtrGeneration)
7214 SelectionDAG &DAG = DCI.DAG;
7215 EVT VT = N->getValueType(0);
7217 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7220 SDValue Res = tryCombineToEXTR(N, DCI);
7224 Res = tryCombineToBSL(N, DCI);
7231 static SDValue performBitcastCombine(SDNode *N,
7232 TargetLowering::DAGCombinerInfo &DCI,
7233 SelectionDAG &DAG) {
7234 // Wait 'til after everything is legalized to try this. That way we have
7235 // legal vector types and such.
7236 if (DCI.isBeforeLegalizeOps())
7239 // Remove extraneous bitcasts around an extract_subvector.
7241 // (v4i16 (bitconvert
7242 // (extract_subvector (v2i64 (bitconvert (v8i16 ...)), (i64 1)))))
7244 // (extract_subvector ((v8i16 ...), (i64 4)))
7246 // Only interested in 64-bit vectors as the ultimate result.
7247 EVT VT = N->getValueType(0);
7250 if (VT.getSimpleVT().getSizeInBits() != 64)
7252 // Is the operand an extract_subvector starting at the beginning or halfway
7253 // point of the vector? A low half may also come through as an
7254 // EXTRACT_SUBREG, so look for that, too.
7255 SDValue Op0 = N->getOperand(0);
7256 if (Op0->getOpcode() != ISD::EXTRACT_SUBVECTOR &&
7257 !(Op0->isMachineOpcode() &&
7258 Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG))
7260 uint64_t idx = cast<ConstantSDNode>(Op0->getOperand(1))->getZExtValue();
7261 if (Op0->getOpcode() == ISD::EXTRACT_SUBVECTOR) {
7262 if (Op0->getValueType(0).getVectorNumElements() != idx && idx != 0)
7264 } else if (Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG) {
7265 if (idx != AArch64::dsub)
7267 // The dsub reference is equivalent to a lane zero subvector reference.
7270 // Look through the bitcast of the input to the extract.
7271 if (Op0->getOperand(0)->getOpcode() != ISD::BITCAST)
7273 SDValue Source = Op0->getOperand(0)->getOperand(0);
7274 // If the source type has twice the number of elements as our destination
7275 // type, we know this is an extract of the high or low half of the vector.
7276 EVT SVT = Source->getValueType(0);
7277 if (SVT.getVectorNumElements() != VT.getVectorNumElements() * 2)
7280 DEBUG(dbgs() << "aarch64-lower: bitcast extract_subvector simplification\n");
7282 // Create the simplified form to just extract the low or high half of the
7283 // vector directly rather than bothering with the bitcasts.
7285 unsigned NumElements = VT.getVectorNumElements();
7287 SDValue HalfIdx = DAG.getConstant(NumElements, MVT::i64);
7288 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Source, HalfIdx);
7290 SDValue SubReg = DAG.getTargetConstant(AArch64::dsub, MVT::i32);
7291 return SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl, VT,
7297 static SDValue performConcatVectorsCombine(SDNode *N,
7298 TargetLowering::DAGCombinerInfo &DCI,
7299 SelectionDAG &DAG) {
7301 EVT VT = N->getValueType(0);
7302 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
7304 // Optimize concat_vectors of truncated vectors, where the intermediate
7305 // type is illegal, to avoid said illegality, e.g.,
7306 // (v4i16 (concat_vectors (v2i16 (truncate (v2i64))),
7307 // (v2i16 (truncate (v2i64)))))
7309 // (v4i16 (truncate (vector_shuffle (v4i32 (bitcast (v2i64))),
7310 // (v4i32 (bitcast (v2i64))),
7312 // This isn't really target-specific, but ISD::TRUNCATE legality isn't keyed
7313 // on both input and result type, so we might generate worse code.
7314 // On AArch64 we know it's fine for v2i64->v4i16 and v4i32->v8i8.
7315 if (N->getNumOperands() == 2 &&
7316 N0->getOpcode() == ISD::TRUNCATE &&
7317 N1->getOpcode() == ISD::TRUNCATE) {
7318 SDValue N00 = N0->getOperand(0);
7319 SDValue N10 = N1->getOperand(0);
7320 EVT N00VT = N00.getValueType();
7322 if (N00VT == N10.getValueType() &&
7323 (N00VT == MVT::v2i64 || N00VT == MVT::v4i32) &&
7324 N00VT.getScalarSizeInBits() == 4 * VT.getScalarSizeInBits()) {
7325 MVT MidVT = (N00VT == MVT::v2i64 ? MVT::v4i32 : MVT::v8i16);
7326 SmallVector<int, 8> Mask(MidVT.getVectorNumElements());
7327 for (size_t i = 0; i < Mask.size(); ++i)
7329 return DAG.getNode(ISD::TRUNCATE, dl, VT,
7330 DAG.getVectorShuffle(
7332 DAG.getNode(ISD::BITCAST, dl, MidVT, N00),
7333 DAG.getNode(ISD::BITCAST, dl, MidVT, N10), Mask));
7337 // Wait 'til after everything is legalized to try this. That way we have
7338 // legal vector types and such.
7339 if (DCI.isBeforeLegalizeOps())
7342 // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector
7343 // splat. The indexed instructions are going to be expecting a DUPLANE64, so
7344 // canonicalise to that.
7345 if (N0 == N1 && VT.getVectorNumElements() == 2) {
7346 assert(VT.getVectorElementType().getSizeInBits() == 64);
7347 return DAG.getNode(AArch64ISD::DUPLANE64, dl, VT, WidenVector(N0, DAG),
7348 DAG.getConstant(0, MVT::i64));
7351 // Canonicalise concat_vectors so that the right-hand vector has as few
7352 // bit-casts as possible before its real operation. The primary matching
7353 // destination for these operations will be the narrowing "2" instructions,
7354 // which depend on the operation being performed on this right-hand vector.
7356 // (concat_vectors LHS, (v1i64 (bitconvert (v4i16 RHS))))
7358 // (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS))
7360 if (N1->getOpcode() != ISD::BITCAST)
7362 SDValue RHS = N1->getOperand(0);
7363 MVT RHSTy = RHS.getValueType().getSimpleVT();
7364 // If the RHS is not a vector, this is not the pattern we're looking for.
7365 if (!RHSTy.isVector())
7368 DEBUG(dbgs() << "aarch64-lower: concat_vectors bitcast simplification\n");
7370 MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(),
7371 RHSTy.getVectorNumElements() * 2);
7372 return DAG.getNode(ISD::BITCAST, dl, VT,
7373 DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatTy,
7374 DAG.getNode(ISD::BITCAST, dl, RHSTy, N0),
7378 static SDValue tryCombineFixedPointConvert(SDNode *N,
7379 TargetLowering::DAGCombinerInfo &DCI,
7380 SelectionDAG &DAG) {
7381 // Wait 'til after everything is legalized to try this. That way we have
7382 // legal vector types and such.
7383 if (DCI.isBeforeLegalizeOps())
7385 // Transform a scalar conversion of a value from a lane extract into a
7386 // lane extract of a vector conversion. E.g., from foo1 to foo2:
7387 // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); }
7388 // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; }
7390 // The second form interacts better with instruction selection and the
7391 // register allocator to avoid cross-class register copies that aren't
7392 // coalescable due to a lane reference.
7394 // Check the operand and see if it originates from a lane extract.
7395 SDValue Op1 = N->getOperand(1);
7396 if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7397 // Yep, no additional predication needed. Perform the transform.
7398 SDValue IID = N->getOperand(0);
7399 SDValue Shift = N->getOperand(2);
7400 SDValue Vec = Op1.getOperand(0);
7401 SDValue Lane = Op1.getOperand(1);
7402 EVT ResTy = N->getValueType(0);
7406 // The vector width should be 128 bits by the time we get here, even
7407 // if it started as 64 bits (the extract_vector handling will have
7409 assert(Vec.getValueType().getSizeInBits() == 128 &&
7410 "unexpected vector size on extract_vector_elt!");
7411 if (Vec.getValueType() == MVT::v4i32)
7412 VecResTy = MVT::v4f32;
7413 else if (Vec.getValueType() == MVT::v2i64)
7414 VecResTy = MVT::v2f64;
7416 llvm_unreachable("unexpected vector type!");
7419 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift);
7420 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane);
7425 // AArch64 high-vector "long" operations are formed by performing the non-high
7426 // version on an extract_subvector of each operand which gets the high half:
7428 // (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS))
7430 // However, there are cases which don't have an extract_high explicitly, but
7431 // have another operation that can be made compatible with one for free. For
7434 // (dupv64 scalar) --> (extract_high (dup128 scalar))
7436 // This routine does the actual conversion of such DUPs, once outer routines
7437 // have determined that everything else is in order.
7438 static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) {
7439 // We can handle most types of duplicate, but the lane ones have an extra
7440 // operand saying *which* lane, so we need to know.
7442 switch (N.getOpcode()) {
7443 case AArch64ISD::DUP:
7446 case AArch64ISD::DUPLANE8:
7447 case AArch64ISD::DUPLANE16:
7448 case AArch64ISD::DUPLANE32:
7449 case AArch64ISD::DUPLANE64:
7456 MVT NarrowTy = N.getSimpleValueType();
7457 if (!NarrowTy.is64BitVector())
7460 MVT ElementTy = NarrowTy.getVectorElementType();
7461 unsigned NumElems = NarrowTy.getVectorNumElements();
7462 MVT NewDUPVT = MVT::getVectorVT(ElementTy, NumElems * 2);
7466 NewDUP = DAG.getNode(N.getOpcode(), SDLoc(N), NewDUPVT, N.getOperand(0),
7469 NewDUP = DAG.getNode(AArch64ISD::DUP, SDLoc(N), NewDUPVT, N.getOperand(0));
7471 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N.getNode()), NarrowTy,
7472 NewDUP, DAG.getConstant(NumElems, MVT::i64));
7475 static bool isEssentiallyExtractSubvector(SDValue N) {
7476 if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR)
7479 return N.getOpcode() == ISD::BITCAST &&
7480 N.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR;
7483 /// \brief Helper structure to keep track of ISD::SET_CC operands.
7484 struct GenericSetCCInfo {
7485 const SDValue *Opnd0;
7486 const SDValue *Opnd1;
7490 /// \brief Helper structure to keep track of a SET_CC lowered into AArch64 code.
7491 struct AArch64SetCCInfo {
7493 AArch64CC::CondCode CC;
7496 /// \brief Helper structure to keep track of SetCC information.
7498 GenericSetCCInfo Generic;
7499 AArch64SetCCInfo AArch64;
7502 /// \brief Helper structure to be able to read SetCC information. If set to
7503 /// true, IsAArch64 field, Info is a AArch64SetCCInfo, otherwise Info is a
7504 /// GenericSetCCInfo.
7505 struct SetCCInfoAndKind {
7510 /// \brief Check whether or not \p Op is a SET_CC operation, either a generic or
7512 /// AArch64 lowered one.
7513 /// \p SetCCInfo is filled accordingly.
7514 /// \post SetCCInfo is meanginfull only when this function returns true.
7515 /// \return True when Op is a kind of SET_CC operation.
7516 static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) {
7517 // If this is a setcc, this is straight forward.
7518 if (Op.getOpcode() == ISD::SETCC) {
7519 SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0);
7520 SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1);
7521 SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
7522 SetCCInfo.IsAArch64 = false;
7525 // Otherwise, check if this is a matching csel instruction.
7529 if (Op.getOpcode() != AArch64ISD::CSEL)
7531 // Set the information about the operands.
7532 // TODO: we want the operands of the Cmp not the csel
7533 SetCCInfo.Info.AArch64.Cmp = &Op.getOperand(3);
7534 SetCCInfo.IsAArch64 = true;
7535 SetCCInfo.Info.AArch64.CC = static_cast<AArch64CC::CondCode>(
7536 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
7538 // Check that the operands matches the constraints:
7539 // (1) Both operands must be constants.
7540 // (2) One must be 1 and the other must be 0.
7541 ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0));
7542 ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7545 if (!TValue || !FValue)
7549 if (!TValue->isOne()) {
7550 // Update the comparison when we are interested in !cc.
7551 std::swap(TValue, FValue);
7552 SetCCInfo.Info.AArch64.CC =
7553 AArch64CC::getInvertedCondCode(SetCCInfo.Info.AArch64.CC);
7555 return TValue->isOne() && FValue->isNullValue();
7558 // Returns true if Op is setcc or zext of setcc.
7559 static bool isSetCCOrZExtSetCC(const SDValue& Op, SetCCInfoAndKind &Info) {
7560 if (isSetCC(Op, Info))
7562 return ((Op.getOpcode() == ISD::ZERO_EXTEND) &&
7563 isSetCC(Op->getOperand(0), Info));
7566 // The folding we want to perform is:
7567 // (add x, [zext] (setcc cc ...) )
7569 // (csel x, (add x, 1), !cc ...)
7571 // The latter will get matched to a CSINC instruction.
7572 static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) {
7573 assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!");
7574 SDValue LHS = Op->getOperand(0);
7575 SDValue RHS = Op->getOperand(1);
7576 SetCCInfoAndKind InfoAndKind;
7578 // If neither operand is a SET_CC, give up.
7579 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind)) {
7580 std::swap(LHS, RHS);
7581 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind))
7585 // FIXME: This could be generatized to work for FP comparisons.
7586 EVT CmpVT = InfoAndKind.IsAArch64
7587 ? InfoAndKind.Info.AArch64.Cmp->getOperand(0).getValueType()
7588 : InfoAndKind.Info.Generic.Opnd0->getValueType();
7589 if (CmpVT != MVT::i32 && CmpVT != MVT::i64)
7595 if (InfoAndKind.IsAArch64) {
7596 CCVal = DAG.getConstant(
7597 AArch64CC::getInvertedCondCode(InfoAndKind.Info.AArch64.CC), MVT::i32);
7598 Cmp = *InfoAndKind.Info.AArch64.Cmp;
7600 Cmp = getAArch64Cmp(*InfoAndKind.Info.Generic.Opnd0,
7601 *InfoAndKind.Info.Generic.Opnd1,
7602 ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, true),
7605 EVT VT = Op->getValueType(0);
7606 LHS = DAG.getNode(ISD::ADD, dl, VT, RHS, DAG.getConstant(1, VT));
7607 return DAG.getNode(AArch64ISD::CSEL, dl, VT, RHS, LHS, CCVal, Cmp);
7610 // The basic add/sub long vector instructions have variants with "2" on the end
7611 // which act on the high-half of their inputs. They are normally matched by
7614 // (add (zeroext (extract_high LHS)),
7615 // (zeroext (extract_high RHS)))
7616 // -> uaddl2 vD, vN, vM
7618 // However, if one of the extracts is something like a duplicate, this
7619 // instruction can still be used profitably. This function puts the DAG into a
7620 // more appropriate form for those patterns to trigger.
7621 static SDValue performAddSubLongCombine(SDNode *N,
7622 TargetLowering::DAGCombinerInfo &DCI,
7623 SelectionDAG &DAG) {
7624 if (DCI.isBeforeLegalizeOps())
7627 MVT VT = N->getSimpleValueType(0);
7628 if (!VT.is128BitVector()) {
7629 if (N->getOpcode() == ISD::ADD)
7630 return performSetccAddFolding(N, DAG);
7634 // Make sure both branches are extended in the same way.
7635 SDValue LHS = N->getOperand(0);
7636 SDValue RHS = N->getOperand(1);
7637 if ((LHS.getOpcode() != ISD::ZERO_EXTEND &&
7638 LHS.getOpcode() != ISD::SIGN_EXTEND) ||
7639 LHS.getOpcode() != RHS.getOpcode())
7642 unsigned ExtType = LHS.getOpcode();
7644 // It's not worth doing if at least one of the inputs isn't already an
7645 // extract, but we don't know which it'll be so we have to try both.
7646 if (isEssentiallyExtractSubvector(LHS.getOperand(0))) {
7647 RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG);
7651 RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS);
7652 } else if (isEssentiallyExtractSubvector(RHS.getOperand(0))) {
7653 LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG);
7657 LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS);
7660 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS);
7663 // Massage DAGs which we can use the high-half "long" operations on into
7664 // something isel will recognize better. E.g.
7666 // (aarch64_neon_umull (extract_high vec) (dupv64 scalar)) -->
7667 // (aarch64_neon_umull (extract_high (v2i64 vec)))
7668 // (extract_high (v2i64 (dup128 scalar)))))
7670 static SDValue tryCombineLongOpWithDup(unsigned IID, SDNode *N,
7671 TargetLowering::DAGCombinerInfo &DCI,
7672 SelectionDAG &DAG) {
7673 if (DCI.isBeforeLegalizeOps())
7676 SDValue LHS = N->getOperand(1);
7677 SDValue RHS = N->getOperand(2);
7678 assert(LHS.getValueType().is64BitVector() &&
7679 RHS.getValueType().is64BitVector() &&
7680 "unexpected shape for long operation");
7682 // Either node could be a DUP, but it's not worth doing both of them (you'd
7683 // just as well use the non-high version) so look for a corresponding extract
7684 // operation on the other "wing".
7685 if (isEssentiallyExtractSubvector(LHS)) {
7686 RHS = tryExtendDUPToExtractHigh(RHS, DAG);
7689 } else if (isEssentiallyExtractSubvector(RHS)) {
7690 LHS = tryExtendDUPToExtractHigh(LHS, DAG);
7695 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0),
7696 N->getOperand(0), LHS, RHS);
7699 static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
7700 MVT ElemTy = N->getSimpleValueType(0).getScalarType();
7701 unsigned ElemBits = ElemTy.getSizeInBits();
7703 int64_t ShiftAmount;
7704 if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) {
7705 APInt SplatValue, SplatUndef;
7706 unsigned SplatBitSize;
7708 if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
7709 HasAnyUndefs, ElemBits) ||
7710 SplatBitSize != ElemBits)
7713 ShiftAmount = SplatValue.getSExtValue();
7714 } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) {
7715 ShiftAmount = CVN->getSExtValue();
7723 llvm_unreachable("Unknown shift intrinsic");
7724 case Intrinsic::aarch64_neon_sqshl:
7725 Opcode = AArch64ISD::SQSHL_I;
7726 IsRightShift = false;
7728 case Intrinsic::aarch64_neon_uqshl:
7729 Opcode = AArch64ISD::UQSHL_I;
7730 IsRightShift = false;
7732 case Intrinsic::aarch64_neon_srshl:
7733 Opcode = AArch64ISD::SRSHR_I;
7734 IsRightShift = true;
7736 case Intrinsic::aarch64_neon_urshl:
7737 Opcode = AArch64ISD::URSHR_I;
7738 IsRightShift = true;
7740 case Intrinsic::aarch64_neon_sqshlu:
7741 Opcode = AArch64ISD::SQSHLU_I;
7742 IsRightShift = false;
7746 if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits)
7747 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
7748 DAG.getConstant(-ShiftAmount, MVT::i32));
7749 else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount < ElemBits)
7750 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
7751 DAG.getConstant(ShiftAmount, MVT::i32));
7756 // The CRC32[BH] instructions ignore the high bits of their data operand. Since
7757 // the intrinsics must be legal and take an i32, this means there's almost
7758 // certainly going to be a zext in the DAG which we can eliminate.
7759 static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) {
7760 SDValue AndN = N->getOperand(2);
7761 if (AndN.getOpcode() != ISD::AND)
7764 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1));
7765 if (!CMask || CMask->getZExtValue() != Mask)
7768 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32,
7769 N->getOperand(0), N->getOperand(1), AndN.getOperand(0));
7772 static SDValue combineAcrossLanesIntrinsic(unsigned Opc, SDNode *N,
7773 SelectionDAG &DAG) {
7774 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), N->getValueType(0),
7775 DAG.getNode(Opc, SDLoc(N),
7776 N->getOperand(1).getSimpleValueType(),
7778 DAG.getConstant(0, MVT::i64));
7781 static SDValue performIntrinsicCombine(SDNode *N,
7782 TargetLowering::DAGCombinerInfo &DCI,
7783 const AArch64Subtarget *Subtarget) {
7784 SelectionDAG &DAG = DCI.DAG;
7785 unsigned IID = getIntrinsicID(N);
7789 case Intrinsic::aarch64_neon_vcvtfxs2fp:
7790 case Intrinsic::aarch64_neon_vcvtfxu2fp:
7791 return tryCombineFixedPointConvert(N, DCI, DAG);
7793 case Intrinsic::aarch64_neon_saddv:
7794 return combineAcrossLanesIntrinsic(AArch64ISD::SADDV, N, DAG);
7795 case Intrinsic::aarch64_neon_uaddv:
7796 return combineAcrossLanesIntrinsic(AArch64ISD::UADDV, N, DAG);
7797 case Intrinsic::aarch64_neon_sminv:
7798 return combineAcrossLanesIntrinsic(AArch64ISD::SMINV, N, DAG);
7799 case Intrinsic::aarch64_neon_uminv:
7800 return combineAcrossLanesIntrinsic(AArch64ISD::UMINV, N, DAG);
7801 case Intrinsic::aarch64_neon_smaxv:
7802 return combineAcrossLanesIntrinsic(AArch64ISD::SMAXV, N, DAG);
7803 case Intrinsic::aarch64_neon_umaxv:
7804 return combineAcrossLanesIntrinsic(AArch64ISD::UMAXV, N, DAG);
7805 case Intrinsic::aarch64_neon_fmax:
7806 return DAG.getNode(AArch64ISD::FMAX, SDLoc(N), N->getValueType(0),
7807 N->getOperand(1), N->getOperand(2));
7808 case Intrinsic::aarch64_neon_fmin:
7809 return DAG.getNode(AArch64ISD::FMIN, SDLoc(N), N->getValueType(0),
7810 N->getOperand(1), N->getOperand(2));
7811 case Intrinsic::aarch64_neon_smull:
7812 case Intrinsic::aarch64_neon_umull:
7813 case Intrinsic::aarch64_neon_pmull:
7814 case Intrinsic::aarch64_neon_sqdmull:
7815 return tryCombineLongOpWithDup(IID, N, DCI, DAG);
7816 case Intrinsic::aarch64_neon_sqshl:
7817 case Intrinsic::aarch64_neon_uqshl:
7818 case Intrinsic::aarch64_neon_sqshlu:
7819 case Intrinsic::aarch64_neon_srshl:
7820 case Intrinsic::aarch64_neon_urshl:
7821 return tryCombineShiftImm(IID, N, DAG);
7822 case Intrinsic::aarch64_crc32b:
7823 case Intrinsic::aarch64_crc32cb:
7824 return tryCombineCRC32(0xff, N, DAG);
7825 case Intrinsic::aarch64_crc32h:
7826 case Intrinsic::aarch64_crc32ch:
7827 return tryCombineCRC32(0xffff, N, DAG);
7832 static SDValue performExtendCombine(SDNode *N,
7833 TargetLowering::DAGCombinerInfo &DCI,
7834 SelectionDAG &DAG) {
7835 // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then
7836 // we can convert that DUP into another extract_high (of a bigger DUP), which
7837 // helps the backend to decide that an sabdl2 would be useful, saving a real
7838 // extract_high operation.
7839 if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND &&
7840 N->getOperand(0).getOpcode() == ISD::INTRINSIC_WO_CHAIN) {
7841 SDNode *ABDNode = N->getOperand(0).getNode();
7842 unsigned IID = getIntrinsicID(ABDNode);
7843 if (IID == Intrinsic::aarch64_neon_sabd ||
7844 IID == Intrinsic::aarch64_neon_uabd) {
7845 SDValue NewABD = tryCombineLongOpWithDup(IID, ABDNode, DCI, DAG);
7846 if (!NewABD.getNode())
7849 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0),
7854 // This is effectively a custom type legalization for AArch64.
7856 // Type legalization will split an extend of a small, legal, type to a larger
7857 // illegal type by first splitting the destination type, often creating
7858 // illegal source types, which then get legalized in isel-confusing ways,
7859 // leading to really terrible codegen. E.g.,
7860 // %result = v8i32 sext v8i8 %value
7862 // %losrc = extract_subreg %value, ...
7863 // %hisrc = extract_subreg %value, ...
7864 // %lo = v4i32 sext v4i8 %losrc
7865 // %hi = v4i32 sext v4i8 %hisrc
7866 // Things go rapidly downhill from there.
7868 // For AArch64, the [sz]ext vector instructions can only go up one element
7869 // size, so we can, e.g., extend from i8 to i16, but to go from i8 to i32
7870 // take two instructions.
7872 // This implies that the most efficient way to do the extend from v8i8
7873 // to two v4i32 values is to first extend the v8i8 to v8i16, then do
7874 // the normal splitting to happen for the v8i16->v8i32.
7876 // This is pre-legalization to catch some cases where the default
7877 // type legalization will create ill-tempered code.
7878 if (!DCI.isBeforeLegalizeOps())
7881 // We're only interested in cleaning things up for non-legal vector types
7882 // here. If both the source and destination are legal, things will just
7883 // work naturally without any fiddling.
7884 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7885 EVT ResVT = N->getValueType(0);
7886 if (!ResVT.isVector() || TLI.isTypeLegal(ResVT))
7888 // If the vector type isn't a simple VT, it's beyond the scope of what
7889 // we're worried about here. Let legalization do its thing and hope for
7891 SDValue Src = N->getOperand(0);
7892 EVT SrcVT = Src->getValueType(0);
7893 if (!ResVT.isSimple() || !SrcVT.isSimple())
7896 // If the source VT is a 64-bit vector, we can play games and get the
7897 // better results we want.
7898 if (SrcVT.getSizeInBits() != 64)
7901 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
7902 unsigned ElementCount = SrcVT.getVectorNumElements();
7903 SrcVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize * 2), ElementCount);
7905 Src = DAG.getNode(N->getOpcode(), DL, SrcVT, Src);
7907 // Now split the rest of the operation into two halves, each with a 64
7911 unsigned NumElements = ResVT.getVectorNumElements();
7912 assert(!(NumElements & 1) && "Splitting vector, but not in half!");
7913 LoVT = HiVT = EVT::getVectorVT(*DAG.getContext(),
7914 ResVT.getVectorElementType(), NumElements / 2);
7916 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
7917 LoVT.getVectorNumElements());
7918 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7919 DAG.getConstant(0, MVT::i64));
7920 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7921 DAG.getConstant(InNVT.getVectorNumElements(), MVT::i64));
7922 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, Lo);
7923 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, Hi);
7925 // Now combine the parts back together so we still have a single result
7926 // like the combiner expects.
7927 return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
7930 /// Replace a splat of a scalar to a vector store by scalar stores of the scalar
7931 /// value. The load store optimizer pass will merge them to store pair stores.
7932 /// This has better performance than a splat of the scalar followed by a split
7933 /// vector store. Even if the stores are not merged it is four stores vs a dup,
7934 /// followed by an ext.b and two stores.
7935 static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) {
7936 SDValue StVal = St->getValue();
7937 EVT VT = StVal.getValueType();
7939 // Don't replace floating point stores, they possibly won't be transformed to
7940 // stp because of the store pair suppress pass.
7941 if (VT.isFloatingPoint())
7944 // Check for insert vector elements.
7945 if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT)
7948 // We can express a splat as store pair(s) for 2 or 4 elements.
7949 unsigned NumVecElts = VT.getVectorNumElements();
7950 if (NumVecElts != 4 && NumVecElts != 2)
7952 SDValue SplatVal = StVal.getOperand(1);
7953 unsigned RemainInsertElts = NumVecElts - 1;
7955 // Check that this is a splat.
7956 while (--RemainInsertElts) {
7957 SDValue NextInsertElt = StVal.getOperand(0);
7958 if (NextInsertElt.getOpcode() != ISD::INSERT_VECTOR_ELT)
7960 if (NextInsertElt.getOperand(1) != SplatVal)
7962 StVal = NextInsertElt;
7964 unsigned OrigAlignment = St->getAlignment();
7965 unsigned EltOffset = NumVecElts == 4 ? 4 : 8;
7966 unsigned Alignment = std::min(OrigAlignment, EltOffset);
7968 // Create scalar stores. This is at least as good as the code sequence for a
7969 // split unaligned store wich is a dup.s, ext.b, and two stores.
7970 // Most of the time the three stores should be replaced by store pair
7971 // instructions (stp).
7973 SDValue BasePtr = St->getBasePtr();
7975 DAG.getStore(St->getChain(), DL, SplatVal, BasePtr, St->getPointerInfo(),
7976 St->isVolatile(), St->isNonTemporal(), St->getAlignment());
7978 unsigned Offset = EltOffset;
7979 while (--NumVecElts) {
7980 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7981 DAG.getConstant(Offset, MVT::i64));
7982 NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr,
7983 St->getPointerInfo(), St->isVolatile(),
7984 St->isNonTemporal(), Alignment);
7985 Offset += EltOffset;
7990 static SDValue performSTORECombine(SDNode *N,
7991 TargetLowering::DAGCombinerInfo &DCI,
7993 const AArch64Subtarget *Subtarget) {
7994 if (!DCI.isBeforeLegalize())
7997 StoreSDNode *S = cast<StoreSDNode>(N);
7998 if (S->isVolatile())
8001 // Cyclone has bad performance on unaligned 16B stores when crossing line and
8002 // page boundaries. We want to split such stores.
8003 if (!Subtarget->isCyclone())
8006 // Don't split at Oz.
8007 MachineFunction &MF = DAG.getMachineFunction();
8008 bool IsMinSize = MF.getFunction()->hasFnAttribute(Attribute::MinSize);
8012 SDValue StVal = S->getValue();
8013 EVT VT = StVal.getValueType();
8015 // Don't split v2i64 vectors. Memcpy lowering produces those and splitting
8016 // those up regresses performance on micro-benchmarks and olden/bh.
8017 if (!VT.isVector() || VT.getVectorNumElements() < 2 || VT == MVT::v2i64)
8020 // Split unaligned 16B stores. They are terrible for performance.
8021 // Don't split stores with alignment of 1 or 2. Code that uses clang vector
8022 // extensions can use this to mark that it does not want splitting to happen
8023 // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of
8024 // eliminating alignment hazards is only 1 in 8 for alignment of 2.
8025 if (VT.getSizeInBits() != 128 || S->getAlignment() >= 16 ||
8026 S->getAlignment() <= 2)
8029 // If we get a splat of a scalar convert this vector store to a store of
8030 // scalars. They will be merged into store pairs thereby removing two
8032 SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S);
8033 if (ReplacedSplat != SDValue())
8034 return ReplacedSplat;
8037 unsigned NumElts = VT.getVectorNumElements() / 2;
8038 // Split VT into two.
8040 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), NumElts);
8041 SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
8042 DAG.getConstant(0, MVT::i64));
8043 SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
8044 DAG.getConstant(NumElts, MVT::i64));
8045 SDValue BasePtr = S->getBasePtr();
8047 DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
8048 S->isVolatile(), S->isNonTemporal(), S->getAlignment());
8049 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
8050 DAG.getConstant(8, MVT::i64));
8051 return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr,
8052 S->getPointerInfo(), S->isVolatile(), S->isNonTemporal(),
8056 /// Target-specific DAG combine function for post-increment LD1 (lane) and
8057 /// post-increment LD1R.
8058 static SDValue performPostLD1Combine(SDNode *N,
8059 TargetLowering::DAGCombinerInfo &DCI,
8061 if (DCI.isBeforeLegalizeOps())
8064 SelectionDAG &DAG = DCI.DAG;
8065 EVT VT = N->getValueType(0);
8067 unsigned LoadIdx = IsLaneOp ? 1 : 0;
8068 SDNode *LD = N->getOperand(LoadIdx).getNode();
8069 // If it is not LOAD, can not do such combine.
8070 if (LD->getOpcode() != ISD::LOAD)
8073 LoadSDNode *LoadSDN = cast<LoadSDNode>(LD);
8074 EVT MemVT = LoadSDN->getMemoryVT();
8075 // Check if memory operand is the same type as the vector element.
8076 if (MemVT != VT.getVectorElementType())
8079 // Check if there are other uses. If so, do not combine as it will introduce
8081 for (SDNode::use_iterator UI = LD->use_begin(), UE = LD->use_end(); UI != UE;
8083 if (UI.getUse().getResNo() == 1) // Ignore uses of the chain result.
8089 SDValue Addr = LD->getOperand(1);
8090 SDValue Vector = N->getOperand(0);
8091 // Search for a use of the address operand that is an increment.
8092 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(), UE =
8093 Addr.getNode()->use_end(); UI != UE; ++UI) {
8095 if (User->getOpcode() != ISD::ADD
8096 || UI.getUse().getResNo() != Addr.getResNo())
8099 // Check that the add is independent of the load. Otherwise, folding it
8100 // would create a cycle.
8101 if (User->isPredecessorOf(LD) || LD->isPredecessorOf(User))
8103 // Also check that add is not used in the vector operand. This would also
8105 if (User->isPredecessorOf(Vector.getNode()))
8108 // If the increment is a constant, it must match the memory ref size.
8109 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8110 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8111 uint32_t IncVal = CInc->getZExtValue();
8112 unsigned NumBytes = VT.getScalarSizeInBits() / 8;
8113 if (IncVal != NumBytes)
8115 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
8118 SmallVector<SDValue, 8> Ops;
8119 Ops.push_back(LD->getOperand(0)); // Chain
8121 Ops.push_back(Vector); // The vector to be inserted
8122 Ops.push_back(N->getOperand(2)); // The lane to be inserted in the vector
8124 Ops.push_back(Addr);
8127 EVT Tys[3] = { VT, MVT::i64, MVT::Other };
8128 SDVTList SDTys = DAG.getVTList(Tys);
8129 unsigned NewOp = IsLaneOp ? AArch64ISD::LD1LANEpost : AArch64ISD::LD1DUPpost;
8130 SDValue UpdN = DAG.getMemIntrinsicNode(NewOp, SDLoc(N), SDTys, Ops,
8132 LoadSDN->getMemOperand());
8135 SmallVector<SDValue, 2> NewResults;
8136 NewResults.push_back(SDValue(LD, 0)); // The result of load
8137 NewResults.push_back(SDValue(UpdN.getNode(), 2)); // Chain
8138 DCI.CombineTo(LD, NewResults);
8139 DCI.CombineTo(N, SDValue(UpdN.getNode(), 0)); // Dup/Inserted Result
8140 DCI.CombineTo(User, SDValue(UpdN.getNode(), 1)); // Write back register
8147 /// Target-specific DAG combine function for NEON load/store intrinsics
8148 /// to merge base address updates.
8149 static SDValue performNEONPostLDSTCombine(SDNode *N,
8150 TargetLowering::DAGCombinerInfo &DCI,
8151 SelectionDAG &DAG) {
8152 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8155 unsigned AddrOpIdx = N->getNumOperands() - 1;
8156 SDValue Addr = N->getOperand(AddrOpIdx);
8158 // Search for a use of the address operand that is an increment.
8159 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8160 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8162 if (User->getOpcode() != ISD::ADD ||
8163 UI.getUse().getResNo() != Addr.getResNo())
8166 // Check that the add is independent of the load/store. Otherwise, folding
8167 // it would create a cycle.
8168 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8171 // Find the new opcode for the updating load/store.
8172 bool IsStore = false;
8173 bool IsLaneOp = false;
8174 bool IsDupOp = false;
8175 unsigned NewOpc = 0;
8176 unsigned NumVecs = 0;
8177 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8179 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8180 case Intrinsic::aarch64_neon_ld2: NewOpc = AArch64ISD::LD2post;
8182 case Intrinsic::aarch64_neon_ld3: NewOpc = AArch64ISD::LD3post;
8184 case Intrinsic::aarch64_neon_ld4: NewOpc = AArch64ISD::LD4post;
8186 case Intrinsic::aarch64_neon_st2: NewOpc = AArch64ISD::ST2post;
8187 NumVecs = 2; IsStore = true; break;
8188 case Intrinsic::aarch64_neon_st3: NewOpc = AArch64ISD::ST3post;
8189 NumVecs = 3; IsStore = true; break;
8190 case Intrinsic::aarch64_neon_st4: NewOpc = AArch64ISD::ST4post;
8191 NumVecs = 4; IsStore = true; break;
8192 case Intrinsic::aarch64_neon_ld1x2: NewOpc = AArch64ISD::LD1x2post;
8194 case Intrinsic::aarch64_neon_ld1x3: NewOpc = AArch64ISD::LD1x3post;
8196 case Intrinsic::aarch64_neon_ld1x4: NewOpc = AArch64ISD::LD1x4post;
8198 case Intrinsic::aarch64_neon_st1x2: NewOpc = AArch64ISD::ST1x2post;
8199 NumVecs = 2; IsStore = true; break;
8200 case Intrinsic::aarch64_neon_st1x3: NewOpc = AArch64ISD::ST1x3post;
8201 NumVecs = 3; IsStore = true; break;
8202 case Intrinsic::aarch64_neon_st1x4: NewOpc = AArch64ISD::ST1x4post;
8203 NumVecs = 4; IsStore = true; break;
8204 case Intrinsic::aarch64_neon_ld2r: NewOpc = AArch64ISD::LD2DUPpost;
8205 NumVecs = 2; IsDupOp = true; break;
8206 case Intrinsic::aarch64_neon_ld3r: NewOpc = AArch64ISD::LD3DUPpost;
8207 NumVecs = 3; IsDupOp = true; break;
8208 case Intrinsic::aarch64_neon_ld4r: NewOpc = AArch64ISD::LD4DUPpost;
8209 NumVecs = 4; IsDupOp = true; break;
8210 case Intrinsic::aarch64_neon_ld2lane: NewOpc = AArch64ISD::LD2LANEpost;
8211 NumVecs = 2; IsLaneOp = true; break;
8212 case Intrinsic::aarch64_neon_ld3lane: NewOpc = AArch64ISD::LD3LANEpost;
8213 NumVecs = 3; IsLaneOp = true; break;
8214 case Intrinsic::aarch64_neon_ld4lane: NewOpc = AArch64ISD::LD4LANEpost;
8215 NumVecs = 4; IsLaneOp = true; break;
8216 case Intrinsic::aarch64_neon_st2lane: NewOpc = AArch64ISD::ST2LANEpost;
8217 NumVecs = 2; IsStore = true; IsLaneOp = true; break;
8218 case Intrinsic::aarch64_neon_st3lane: NewOpc = AArch64ISD::ST3LANEpost;
8219 NumVecs = 3; IsStore = true; IsLaneOp = true; break;
8220 case Intrinsic::aarch64_neon_st4lane: NewOpc = AArch64ISD::ST4LANEpost;
8221 NumVecs = 4; IsStore = true; IsLaneOp = true; break;
8226 VecTy = N->getOperand(2).getValueType();
8228 VecTy = N->getValueType(0);
8230 // If the increment is a constant, it must match the memory ref size.
8231 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8232 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8233 uint32_t IncVal = CInc->getZExtValue();
8234 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8235 if (IsLaneOp || IsDupOp)
8236 NumBytes /= VecTy.getVectorNumElements();
8237 if (IncVal != NumBytes)
8239 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
8241 SmallVector<SDValue, 8> Ops;
8242 Ops.push_back(N->getOperand(0)); // Incoming chain
8243 // Load lane and store have vector list as input.
8244 if (IsLaneOp || IsStore)
8245 for (unsigned i = 2; i < AddrOpIdx; ++i)
8246 Ops.push_back(N->getOperand(i));
8247 Ops.push_back(Addr); // Base register
8252 unsigned NumResultVecs = (IsStore ? 0 : NumVecs);
8254 for (n = 0; n < NumResultVecs; ++n)
8256 Tys[n++] = MVT::i64; // Type of write back register
8257 Tys[n] = MVT::Other; // Type of the chain
8258 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs + 2));
8260 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8261 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys, Ops,
8262 MemInt->getMemoryVT(),
8263 MemInt->getMemOperand());
8266 std::vector<SDValue> NewResults;
8267 for (unsigned i = 0; i < NumResultVecs; ++i) {
8268 NewResults.push_back(SDValue(UpdN.getNode(), i));
8270 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs + 1));
8271 DCI.CombineTo(N, NewResults);
8272 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8279 // Checks to see if the value is the prescribed width and returns information
8280 // about its extension mode.
8282 bool checkValueWidth(SDValue V, unsigned width, ISD::LoadExtType &ExtType) {
8283 ExtType = ISD::NON_EXTLOAD;
8284 switch(V.getNode()->getOpcode()) {
8288 LoadSDNode *LoadNode = cast<LoadSDNode>(V.getNode());
8289 if ((LoadNode->getMemoryVT() == MVT::i8 && width == 8)
8290 || (LoadNode->getMemoryVT() == MVT::i16 && width == 16)) {
8291 ExtType = LoadNode->getExtensionType();
8296 case ISD::AssertSext: {
8297 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
8298 if ((TypeNode->getVT() == MVT::i8 && width == 8)
8299 || (TypeNode->getVT() == MVT::i16 && width == 16)) {
8300 ExtType = ISD::SEXTLOAD;
8305 case ISD::AssertZext: {
8306 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
8307 if ((TypeNode->getVT() == MVT::i8 && width == 8)
8308 || (TypeNode->getVT() == MVT::i16 && width == 16)) {
8309 ExtType = ISD::ZEXTLOAD;
8315 case ISD::TargetConstant: {
8316 if (std::abs(cast<ConstantSDNode>(V.getNode())->getSExtValue()) <
8326 // This function does a whole lot of voodoo to determine if the tests are
8327 // equivalent without and with a mask. Essentially what happens is that given a
8330 // +-------------+ +-------------+ +-------------+ +-------------+
8331 // | Input | | AddConstant | | CompConstant| | CC |
8332 // +-------------+ +-------------+ +-------------+ +-------------+
8334 // V V | +----------+
8335 // +-------------+ +----+ | |
8336 // | ADD | |0xff| | |
8337 // +-------------+ +----+ | |
8340 // +-------------+ | |
8342 // +-------------+ | |
8351 // The AND node may be safely removed for some combinations of inputs. In
8352 // particular we need to take into account the extension type of the Input,
8353 // the exact values of AddConstant, CompConstant, and CC, along with the nominal
8354 // width of the input (this can work for any width inputs, the above graph is
8355 // specific to 8 bits.
8357 // The specific equations were worked out by generating output tables for each
8358 // AArch64CC value in terms of and AddConstant (w1), CompConstant(w2). The
8359 // problem was simplified by working with 4 bit inputs, which means we only
8360 // needed to reason about 24 distinct bit patterns: 8 patterns unique to zero
8361 // extension (8,15), 8 patterns unique to sign extensions (-8,-1), and 8
8362 // patterns present in both extensions (0,7). For every distinct set of
8363 // AddConstant and CompConstants bit patterns we can consider the masked and
8364 // unmasked versions to be equivalent if the result of this function is true for
8365 // all 16 distinct bit patterns of for the current extension type of Input (w0).
8368 // and w10, w8, #0x0f
8370 // cset w9, AArch64CC
8372 // cset w11, AArch64CC
8377 // Since the above function shows when the outputs are equivalent it defines
8378 // when it is safe to remove the AND. Unfortunately it only runs on AArch64 and
8379 // would be expensive to run during compiles. The equations below were written
8380 // in a test harness that confirmed they gave equivalent outputs to the above
8381 // for all inputs function, so they can be used determine if the removal is
8384 // isEquivalentMaskless() is the code for testing if the AND can be removed
8385 // factored out of the DAG recognition as the DAG can take several forms.
8388 bool isEquivalentMaskless(unsigned CC, unsigned width,
8389 ISD::LoadExtType ExtType, signed AddConstant,
8390 signed CompConstant) {
8391 // By being careful about our equations and only writing the in term
8392 // symbolic values and well known constants (0, 1, -1, MaxUInt) we can
8393 // make them generally applicable to all bit widths.
8394 signed MaxUInt = (1 << width);
8396 // For the purposes of these comparisons sign extending the type is
8397 // equivalent to zero extending the add and displacing it by half the integer
8398 // width. Provided we are careful and make sure our equations are valid over
8399 // the whole range we can just adjust the input and avoid writing equations
8400 // for sign extended inputs.
8401 if (ExtType == ISD::SEXTLOAD)
8402 AddConstant -= (1 << (width-1));
8406 case AArch64CC::GT: {
8407 if ((AddConstant == 0) ||
8408 (CompConstant == MaxUInt - 1 && AddConstant < 0) ||
8409 (AddConstant >= 0 && CompConstant < 0) ||
8410 (AddConstant <= 0 && CompConstant <= 0 && CompConstant < AddConstant))
8414 case AArch64CC::GE: {
8415 if ((AddConstant == 0) ||
8416 (AddConstant >= 0 && CompConstant <= 0) ||
8417 (AddConstant <= 0 && CompConstant <= 0 && CompConstant <= AddConstant))
8421 case AArch64CC::LS: {
8422 if ((AddConstant >= 0 && CompConstant < 0) ||
8423 (AddConstant <= 0 && CompConstant >= -1 &&
8424 CompConstant < AddConstant + MaxUInt))
8428 case AArch64CC::MI: {
8429 if ((AddConstant == 0) ||
8430 (AddConstant > 0 && CompConstant <= 0) ||
8431 (AddConstant < 0 && CompConstant <= AddConstant))
8435 case AArch64CC::HS: {
8436 if ((AddConstant >= 0 && CompConstant <= 0) ||
8437 (AddConstant <= 0 && CompConstant >= 0 &&
8438 CompConstant <= AddConstant + MaxUInt))
8442 case AArch64CC::NE: {
8443 if ((AddConstant > 0 && CompConstant < 0) ||
8444 (AddConstant < 0 && CompConstant >= 0 &&
8445 CompConstant < AddConstant + MaxUInt) ||
8446 (AddConstant >= 0 && CompConstant >= 0 &&
8447 CompConstant >= AddConstant) ||
8448 (AddConstant <= 0 && CompConstant < 0 && CompConstant < AddConstant))
8457 case AArch64CC::Invalid:
8465 SDValue performCONDCombine(SDNode *N,
8466 TargetLowering::DAGCombinerInfo &DCI,
8467 SelectionDAG &DAG, unsigned CCIndex,
8468 unsigned CmpIndex) {
8469 unsigned CC = cast<ConstantSDNode>(N->getOperand(CCIndex))->getSExtValue();
8470 SDNode *SubsNode = N->getOperand(CmpIndex).getNode();
8471 unsigned CondOpcode = SubsNode->getOpcode();
8473 if (CondOpcode != AArch64ISD::SUBS)
8476 // There is a SUBS feeding this condition. Is it fed by a mask we can
8479 SDNode *AndNode = SubsNode->getOperand(0).getNode();
8480 unsigned MaskBits = 0;
8482 if (AndNode->getOpcode() != ISD::AND)
8485 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(AndNode->getOperand(1))) {
8486 uint32_t CNV = CN->getZExtValue();
8489 else if (CNV == 65535)
8496 SDValue AddValue = AndNode->getOperand(0);
8498 if (AddValue.getOpcode() != ISD::ADD)
8501 // The basic dag structure is correct, grab the inputs and validate them.
8503 SDValue AddInputValue1 = AddValue.getNode()->getOperand(0);
8504 SDValue AddInputValue2 = AddValue.getNode()->getOperand(1);
8505 SDValue SubsInputValue = SubsNode->getOperand(1);
8507 // The mask is present and the provenance of all the values is a smaller type,
8508 // lets see if the mask is superfluous.
8510 if (!isa<ConstantSDNode>(AddInputValue2.getNode()) ||
8511 !isa<ConstantSDNode>(SubsInputValue.getNode()))
8514 ISD::LoadExtType ExtType;
8516 if (!checkValueWidth(SubsInputValue, MaskBits, ExtType) ||
8517 !checkValueWidth(AddInputValue2, MaskBits, ExtType) ||
8518 !checkValueWidth(AddInputValue1, MaskBits, ExtType) )
8521 if(!isEquivalentMaskless(CC, MaskBits, ExtType,
8522 cast<ConstantSDNode>(AddInputValue2.getNode())->getSExtValue(),
8523 cast<ConstantSDNode>(SubsInputValue.getNode())->getSExtValue()))
8526 // The AND is not necessary, remove it.
8528 SDVTList VTs = DAG.getVTList(SubsNode->getValueType(0),
8529 SubsNode->getValueType(1));
8530 SDValue Ops[] = { AddValue, SubsNode->getOperand(1) };
8532 SDValue NewValue = DAG.getNode(CondOpcode, SDLoc(SubsNode), VTs, Ops);
8533 DAG.ReplaceAllUsesWith(SubsNode, NewValue.getNode());
8535 return SDValue(N, 0);
8538 // Optimize compare with zero and branch.
8539 static SDValue performBRCONDCombine(SDNode *N,
8540 TargetLowering::DAGCombinerInfo &DCI,
8541 SelectionDAG &DAG) {
8542 SDValue NV = performCONDCombine(N, DCI, DAG, 2, 3);
8545 SDValue Chain = N->getOperand(0);
8546 SDValue Dest = N->getOperand(1);
8547 SDValue CCVal = N->getOperand(2);
8548 SDValue Cmp = N->getOperand(3);
8550 assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!");
8551 unsigned CC = cast<ConstantSDNode>(CCVal)->getZExtValue();
8552 if (CC != AArch64CC::EQ && CC != AArch64CC::NE)
8555 unsigned CmpOpc = Cmp.getOpcode();
8556 if (CmpOpc != AArch64ISD::ADDS && CmpOpc != AArch64ISD::SUBS)
8559 // Only attempt folding if there is only one use of the flag and no use of the
8561 if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1))
8564 SDValue LHS = Cmp.getOperand(0);
8565 SDValue RHS = Cmp.getOperand(1);
8567 assert(LHS.getValueType() == RHS.getValueType() &&
8568 "Expected the value type to be the same for both operands!");
8569 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
8572 if (isa<ConstantSDNode>(LHS) && cast<ConstantSDNode>(LHS)->isNullValue())
8573 std::swap(LHS, RHS);
8575 if (!isa<ConstantSDNode>(RHS) || !cast<ConstantSDNode>(RHS)->isNullValue())
8578 if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA ||
8579 LHS.getOpcode() == ISD::SRL)
8582 // Fold the compare into the branch instruction.
8584 if (CC == AArch64CC::EQ)
8585 BR = DAG.getNode(AArch64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
8587 BR = DAG.getNode(AArch64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
8589 // Do not add new nodes to DAG combiner worklist.
8590 DCI.CombineTo(N, BR, false);
8595 // vselect (v1i1 setcc) ->
8596 // vselect (v1iXX setcc) (XX is the size of the compared operand type)
8597 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
8598 // condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
8600 static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) {
8601 SDValue N0 = N->getOperand(0);
8602 EVT CCVT = N0.getValueType();
8604 if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
8605 CCVT.getVectorElementType() != MVT::i1)
8608 EVT ResVT = N->getValueType(0);
8609 EVT CmpVT = N0.getOperand(0).getValueType();
8610 // Only combine when the result type is of the same size as the compared
8612 if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
8615 SDValue IfTrue = N->getOperand(1);
8616 SDValue IfFalse = N->getOperand(2);
8618 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
8619 N0.getOperand(0), N0.getOperand(1),
8620 cast<CondCodeSDNode>(N0.getOperand(2))->get());
8621 return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
8625 /// A vector select: "(select vL, vR, (setcc LHS, RHS))" is best performed with
8626 /// the compare-mask instructions rather than going via NZCV, even if LHS and
8627 /// RHS are really scalar. This replaces any scalar setcc in the above pattern
8628 /// with a vector one followed by a DUP shuffle on the result.
8629 static SDValue performSelectCombine(SDNode *N, SelectionDAG &DAG) {
8630 SDValue N0 = N->getOperand(0);
8631 EVT ResVT = N->getValueType(0);
8633 if (N0.getOpcode() != ISD::SETCC || N0.getValueType() != MVT::i1)
8636 // If NumMaskElts == 0, the comparison is larger than select result. The
8637 // largest real NEON comparison is 64-bits per lane, which means the result is
8638 // at most 32-bits and an illegal vector. Just bail out for now.
8639 EVT SrcVT = N0.getOperand(0).getValueType();
8641 // Don't try to do this optimization when the setcc itself has i1 operands.
8642 // There are no legal vectors of i1, so this would be pointless.
8643 if (SrcVT == MVT::i1)
8646 int NumMaskElts = ResVT.getSizeInBits() / SrcVT.getSizeInBits();
8647 if (!ResVT.isVector() || NumMaskElts == 0)
8650 SrcVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumMaskElts);
8651 EVT CCVT = SrcVT.changeVectorElementTypeToInteger();
8653 // First perform a vector comparison, where lane 0 is the one we're interested
8657 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(0));
8659 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(1));
8660 SDValue SetCC = DAG.getNode(ISD::SETCC, DL, CCVT, LHS, RHS, N0.getOperand(2));
8662 // Now duplicate the comparison mask we want across all other lanes.
8663 SmallVector<int, 8> DUPMask(CCVT.getVectorNumElements(), 0);
8664 SDValue Mask = DAG.getVectorShuffle(CCVT, DL, SetCC, SetCC, DUPMask.data());
8665 Mask = DAG.getNode(ISD::BITCAST, DL,
8666 ResVT.changeVectorElementTypeToInteger(), Mask);
8668 return DAG.getSelect(DL, ResVT, Mask, N->getOperand(1), N->getOperand(2));
8671 SDValue AArch64TargetLowering::PerformDAGCombine(SDNode *N,
8672 DAGCombinerInfo &DCI) const {
8673 SelectionDAG &DAG = DCI.DAG;
8674 switch (N->getOpcode()) {
8679 return performAddSubLongCombine(N, DCI, DAG);
8681 return performXorCombine(N, DAG, DCI, Subtarget);
8683 return performMulCombine(N, DAG, DCI, Subtarget);
8684 case ISD::SINT_TO_FP:
8685 case ISD::UINT_TO_FP:
8686 return performIntToFpCombine(N, DAG, Subtarget);
8688 return performORCombine(N, DCI, Subtarget);
8689 case ISD::INTRINSIC_WO_CHAIN:
8690 return performIntrinsicCombine(N, DCI, Subtarget);
8691 case ISD::ANY_EXTEND:
8692 case ISD::ZERO_EXTEND:
8693 case ISD::SIGN_EXTEND:
8694 return performExtendCombine(N, DCI, DAG);
8696 return performBitcastCombine(N, DCI, DAG);
8697 case ISD::CONCAT_VECTORS:
8698 return performConcatVectorsCombine(N, DCI, DAG);
8700 return performSelectCombine(N, DAG);
8702 return performVSelectCombine(N, DCI.DAG);
8704 return performSTORECombine(N, DCI, DAG, Subtarget);
8705 case AArch64ISD::BRCOND:
8706 return performBRCONDCombine(N, DCI, DAG);
8707 case AArch64ISD::CSEL:
8708 return performCONDCombine(N, DCI, DAG, 2, 3);
8709 case AArch64ISD::DUP:
8710 return performPostLD1Combine(N, DCI, false);
8711 case ISD::INSERT_VECTOR_ELT:
8712 return performPostLD1Combine(N, DCI, true);
8713 case ISD::INTRINSIC_VOID:
8714 case ISD::INTRINSIC_W_CHAIN:
8715 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
8716 case Intrinsic::aarch64_neon_ld2:
8717 case Intrinsic::aarch64_neon_ld3:
8718 case Intrinsic::aarch64_neon_ld4:
8719 case Intrinsic::aarch64_neon_ld1x2:
8720 case Intrinsic::aarch64_neon_ld1x3:
8721 case Intrinsic::aarch64_neon_ld1x4:
8722 case Intrinsic::aarch64_neon_ld2lane:
8723 case Intrinsic::aarch64_neon_ld3lane:
8724 case Intrinsic::aarch64_neon_ld4lane:
8725 case Intrinsic::aarch64_neon_ld2r:
8726 case Intrinsic::aarch64_neon_ld3r:
8727 case Intrinsic::aarch64_neon_ld4r:
8728 case Intrinsic::aarch64_neon_st2:
8729 case Intrinsic::aarch64_neon_st3:
8730 case Intrinsic::aarch64_neon_st4:
8731 case Intrinsic::aarch64_neon_st1x2:
8732 case Intrinsic::aarch64_neon_st1x3:
8733 case Intrinsic::aarch64_neon_st1x4:
8734 case Intrinsic::aarch64_neon_st2lane:
8735 case Intrinsic::aarch64_neon_st3lane:
8736 case Intrinsic::aarch64_neon_st4lane:
8737 return performNEONPostLDSTCombine(N, DCI, DAG);
8745 // Check if the return value is used as only a return value, as otherwise
8746 // we can't perform a tail-call. In particular, we need to check for
8747 // target ISD nodes that are returns and any other "odd" constructs
8748 // that the generic analysis code won't necessarily catch.
8749 bool AArch64TargetLowering::isUsedByReturnOnly(SDNode *N,
8750 SDValue &Chain) const {
8751 if (N->getNumValues() != 1)
8753 if (!N->hasNUsesOfValue(1, 0))
8756 SDValue TCChain = Chain;
8757 SDNode *Copy = *N->use_begin();
8758 if (Copy->getOpcode() == ISD::CopyToReg) {
8759 // If the copy has a glue operand, we conservatively assume it isn't safe to
8760 // perform a tail call.
8761 if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() ==
8764 TCChain = Copy->getOperand(0);
8765 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
8768 bool HasRet = false;
8769 for (SDNode *Node : Copy->uses()) {
8770 if (Node->getOpcode() != AArch64ISD::RET_FLAG)
8782 // Return whether the an instruction can potentially be optimized to a tail
8783 // call. This will cause the optimizers to attempt to move, or duplicate,
8784 // return instructions to help enable tail call optimizations for this
8786 bool AArch64TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
8787 if (!CI->isTailCall())
8793 bool AArch64TargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base,
8795 ISD::MemIndexedMode &AM,
8797 SelectionDAG &DAG) const {
8798 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
8801 Base = Op->getOperand(0);
8802 // All of the indexed addressing mode instructions take a signed
8803 // 9 bit immediate offset.
8804 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
8805 int64_t RHSC = (int64_t)RHS->getZExtValue();
8806 if (RHSC >= 256 || RHSC <= -256)
8808 IsInc = (Op->getOpcode() == ISD::ADD);
8809 Offset = Op->getOperand(1);
8815 bool AArch64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
8817 ISD::MemIndexedMode &AM,
8818 SelectionDAG &DAG) const {
8821 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8822 VT = LD->getMemoryVT();
8823 Ptr = LD->getBasePtr();
8824 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8825 VT = ST->getMemoryVT();
8826 Ptr = ST->getBasePtr();
8831 if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, IsInc, DAG))
8833 AM = IsInc ? ISD::PRE_INC : ISD::PRE_DEC;
8837 bool AArch64TargetLowering::getPostIndexedAddressParts(
8838 SDNode *N, SDNode *Op, SDValue &Base, SDValue &Offset,
8839 ISD::MemIndexedMode &AM, SelectionDAG &DAG) const {
8842 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8843 VT = LD->getMemoryVT();
8844 Ptr = LD->getBasePtr();
8845 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8846 VT = ST->getMemoryVT();
8847 Ptr = ST->getBasePtr();
8852 if (!getIndexedAddressParts(Op, Base, Offset, AM, IsInc, DAG))
8854 // Post-indexing updates the base, so it's not a valid transform
8855 // if that's not the same as the load's pointer.
8858 AM = IsInc ? ISD::POST_INC : ISD::POST_DEC;
8862 static void ReplaceBITCASTResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
8863 SelectionDAG &DAG) {
8865 SDValue Op = N->getOperand(0);
8867 if (N->getValueType(0) != MVT::i16 || Op.getValueType() != MVT::f16)
8871 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, DL, MVT::f32,
8872 DAG.getUNDEF(MVT::i32), Op,
8873 DAG.getTargetConstant(AArch64::hsub, MVT::i32)),
8875 Op = DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op);
8876 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, Op));
8879 void AArch64TargetLowering::ReplaceNodeResults(
8880 SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
8881 switch (N->getOpcode()) {
8883 llvm_unreachable("Don't know how to custom expand this");
8885 ReplaceBITCASTResults(N, Results, DAG);
8887 case ISD::FP_TO_UINT:
8888 case ISD::FP_TO_SINT:
8889 assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion");
8890 // Let normal code take care of it by not adding anything to Results.
8895 bool AArch64TargetLowering::useLoadStackGuardNode() const {
8899 bool AArch64TargetLowering::combineRepeatedFPDivisors(unsigned NumUsers) const {
8900 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
8901 // reciprocal if there are three or more FDIVs.
8902 return NumUsers > 2;
8905 TargetLoweringBase::LegalizeTypeAction
8906 AArch64TargetLowering::getPreferredVectorAction(EVT VT) const {
8907 MVT SVT = VT.getSimpleVT();
8908 // During type legalization, we prefer to widen v1i8, v1i16, v1i32 to v8i8,
8909 // v4i16, v2i32 instead of to promote.
8910 if (SVT == MVT::v1i8 || SVT == MVT::v1i16 || SVT == MVT::v1i32
8911 || SVT == MVT::v1f32)
8912 return TypeWidenVector;
8914 return TargetLoweringBase::getPreferredVectorAction(VT);
8917 // Loads and stores less than 128-bits are already atomic; ones above that
8918 // are doomed anyway, so defer to the default libcall and blame the OS when
8920 bool AArch64TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
8921 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
8925 // Loads and stores less than 128-bits are already atomic; ones above that
8926 // are doomed anyway, so defer to the default libcall and blame the OS when
8928 bool AArch64TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
8929 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
8933 // For the real atomic operations, we have ldxr/stxr up to 128 bits,
8934 TargetLoweringBase::AtomicRMWExpansionKind
8935 AArch64TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
8936 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
8937 return Size <= 128 ? AtomicRMWExpansionKind::LLSC
8938 : AtomicRMWExpansionKind::None;
8941 bool AArch64TargetLowering::hasLoadLinkedStoreConditional() const {
8945 Value *AArch64TargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
8946 AtomicOrdering Ord) const {
8947 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
8948 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
8949 bool IsAcquire = isAtLeastAcquire(Ord);
8951 // Since i128 isn't legal and intrinsics don't get type-lowered, the ldrexd
8952 // intrinsic must return {i64, i64} and we have to recombine them into a
8953 // single i128 here.
8954 if (ValTy->getPrimitiveSizeInBits() == 128) {
8956 IsAcquire ? Intrinsic::aarch64_ldaxp : Intrinsic::aarch64_ldxp;
8957 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int);
8959 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
8960 Value *LoHi = Builder.CreateCall(Ldxr, Addr, "lohi");
8962 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
8963 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
8964 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
8965 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
8966 return Builder.CreateOr(
8967 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 64)), "val64");
8970 Type *Tys[] = { Addr->getType() };
8972 IsAcquire ? Intrinsic::aarch64_ldaxr : Intrinsic::aarch64_ldxr;
8973 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int, Tys);
8975 return Builder.CreateTruncOrBitCast(
8976 Builder.CreateCall(Ldxr, Addr),
8977 cast<PointerType>(Addr->getType())->getElementType());
8980 Value *AArch64TargetLowering::emitStoreConditional(IRBuilder<> &Builder,
8981 Value *Val, Value *Addr,
8982 AtomicOrdering Ord) const {
8983 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
8984 bool IsRelease = isAtLeastRelease(Ord);
8986 // Since the intrinsics must have legal type, the i128 intrinsics take two
8987 // parameters: "i64, i64". We must marshal Val into the appropriate form
8989 if (Val->getType()->getPrimitiveSizeInBits() == 128) {
8991 IsRelease ? Intrinsic::aarch64_stlxp : Intrinsic::aarch64_stxp;
8992 Function *Stxr = Intrinsic::getDeclaration(M, Int);
8993 Type *Int64Ty = Type::getInt64Ty(M->getContext());
8995 Value *Lo = Builder.CreateTrunc(Val, Int64Ty, "lo");
8996 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 64), Int64Ty, "hi");
8997 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
8998 return Builder.CreateCall3(Stxr, Lo, Hi, Addr);
9002 IsRelease ? Intrinsic::aarch64_stlxr : Intrinsic::aarch64_stxr;
9003 Type *Tys[] = { Addr->getType() };
9004 Function *Stxr = Intrinsic::getDeclaration(M, Int, Tys);
9006 return Builder.CreateCall2(
9007 Stxr, Builder.CreateZExtOrBitCast(
9008 Val, Stxr->getFunctionType()->getParamType(0)),
9012 bool AArch64TargetLowering::functionArgumentNeedsConsecutiveRegisters(
9013 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
9014 return Ty->isArrayTy();