1 //===-- AArch64ISelLowering.cpp - AArch64 DAG Lowering Implementation ----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the AArch64TargetLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "AArch64ISelLowering.h"
15 #include "AArch64CallingConvention.h"
16 #include "AArch64MachineFunctionInfo.h"
17 #include "AArch64PerfectShuffle.h"
18 #include "AArch64Subtarget.h"
19 #include "AArch64TargetMachine.h"
20 #include "AArch64TargetObjectFile.h"
21 #include "MCTargetDesc/AArch64AddressingModes.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/CodeGen/CallingConvLower.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/GetElementPtrTypeIterator.h"
29 #include "llvm/IR/Intrinsics.h"
30 #include "llvm/IR/Type.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Target/TargetOptions.h"
38 #define DEBUG_TYPE "aarch64-lower"
40 STATISTIC(NumTailCalls, "Number of tail calls");
41 STATISTIC(NumShiftInserts, "Number of vector shift inserts");
43 // Place holder until extr generation is tested fully.
45 EnableAArch64ExtrGeneration("aarch64-extr-generation", cl::Hidden,
46 cl::desc("Allow AArch64 (or (shift)(shift))->extract"),
50 EnableAArch64SlrGeneration("aarch64-shift-insert-generation", cl::Hidden,
51 cl::desc("Allow AArch64 SLI/SRI formation"),
54 // FIXME: The necessary dtprel relocations don't seem to be supported
55 // well in the GNU bfd and gold linkers at the moment. Therefore, by
56 // default, for now, fall back to GeneralDynamic code generation.
57 cl::opt<bool> EnableAArch64ELFLocalDynamicTLSGeneration(
58 "aarch64-elf-ldtls-generation", cl::Hidden,
59 cl::desc("Allow AArch64 Local Dynamic TLS code generation"),
62 /// Value type used for condition codes.
63 static const MVT MVT_CC = MVT::i32;
65 AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
66 const AArch64Subtarget &STI)
67 : TargetLowering(TM), Subtarget(&STI) {
69 // AArch64 doesn't have comparisons which set GPRs or setcc instructions, so
70 // we have to make something up. Arbitrarily, choose ZeroOrOne.
71 setBooleanContents(ZeroOrOneBooleanContent);
72 // When comparing vectors the result sets the different elements in the
73 // vector to all-one or all-zero.
74 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
76 // Set up the register classes.
77 addRegisterClass(MVT::i32, &AArch64::GPR32allRegClass);
78 addRegisterClass(MVT::i64, &AArch64::GPR64allRegClass);
80 if (Subtarget->hasFPARMv8()) {
81 addRegisterClass(MVT::f16, &AArch64::FPR16RegClass);
82 addRegisterClass(MVT::f32, &AArch64::FPR32RegClass);
83 addRegisterClass(MVT::f64, &AArch64::FPR64RegClass);
84 addRegisterClass(MVT::f128, &AArch64::FPR128RegClass);
87 if (Subtarget->hasNEON()) {
88 addRegisterClass(MVT::v16i8, &AArch64::FPR8RegClass);
89 addRegisterClass(MVT::v8i16, &AArch64::FPR16RegClass);
90 // Someone set us up the NEON.
91 addDRTypeForNEON(MVT::v2f32);
92 addDRTypeForNEON(MVT::v8i8);
93 addDRTypeForNEON(MVT::v4i16);
94 addDRTypeForNEON(MVT::v2i32);
95 addDRTypeForNEON(MVT::v1i64);
96 addDRTypeForNEON(MVT::v1f64);
97 addDRTypeForNEON(MVT::v4f16);
99 addQRTypeForNEON(MVT::v4f32);
100 addQRTypeForNEON(MVT::v2f64);
101 addQRTypeForNEON(MVT::v16i8);
102 addQRTypeForNEON(MVT::v8i16);
103 addQRTypeForNEON(MVT::v4i32);
104 addQRTypeForNEON(MVT::v2i64);
105 addQRTypeForNEON(MVT::v8f16);
108 // Compute derived properties from the register classes
109 computeRegisterProperties(Subtarget->getRegisterInfo());
111 // Provide all sorts of operation actions
112 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
113 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
114 setOperationAction(ISD::SETCC, MVT::i32, Custom);
115 setOperationAction(ISD::SETCC, MVT::i64, Custom);
116 setOperationAction(ISD::SETCC, MVT::f32, Custom);
117 setOperationAction(ISD::SETCC, MVT::f64, Custom);
118 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
119 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
120 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
121 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
122 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
123 setOperationAction(ISD::SELECT, MVT::i32, Custom);
124 setOperationAction(ISD::SELECT, MVT::i64, Custom);
125 setOperationAction(ISD::SELECT, MVT::f32, Custom);
126 setOperationAction(ISD::SELECT, MVT::f64, Custom);
127 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
128 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
129 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
130 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
131 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
132 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
134 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
135 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
136 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
138 setOperationAction(ISD::FREM, MVT::f32, Expand);
139 setOperationAction(ISD::FREM, MVT::f64, Expand);
140 setOperationAction(ISD::FREM, MVT::f80, Expand);
142 // Custom lowering hooks are needed for XOR
143 // to fold it into CSINC/CSINV.
144 setOperationAction(ISD::XOR, MVT::i32, Custom);
145 setOperationAction(ISD::XOR, MVT::i64, Custom);
147 // Virtually no operation on f128 is legal, but LLVM can't expand them when
148 // there's a valid register class, so we need custom operations in most cases.
149 setOperationAction(ISD::FABS, MVT::f128, Expand);
150 setOperationAction(ISD::FADD, MVT::f128, Custom);
151 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
152 setOperationAction(ISD::FCOS, MVT::f128, Expand);
153 setOperationAction(ISD::FDIV, MVT::f128, Custom);
154 setOperationAction(ISD::FMA, MVT::f128, Expand);
155 setOperationAction(ISD::FMUL, MVT::f128, Custom);
156 setOperationAction(ISD::FNEG, MVT::f128, Expand);
157 setOperationAction(ISD::FPOW, MVT::f128, Expand);
158 setOperationAction(ISD::FREM, MVT::f128, Expand);
159 setOperationAction(ISD::FRINT, MVT::f128, Expand);
160 setOperationAction(ISD::FSIN, MVT::f128, Expand);
161 setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
162 setOperationAction(ISD::FSQRT, MVT::f128, Expand);
163 setOperationAction(ISD::FSUB, MVT::f128, Custom);
164 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
165 setOperationAction(ISD::SETCC, MVT::f128, Custom);
166 setOperationAction(ISD::BR_CC, MVT::f128, Custom);
167 setOperationAction(ISD::SELECT, MVT::f128, Custom);
168 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
169 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);
171 // Lowering for many of the conversions is actually specified by the non-f128
172 // type. The LowerXXX function will be trivial when f128 isn't involved.
173 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
174 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
175 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
176 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
177 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
178 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
179 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
180 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
181 setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
182 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
183 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
184 setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
185 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
186 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
188 // Variable arguments.
189 setOperationAction(ISD::VASTART, MVT::Other, Custom);
190 setOperationAction(ISD::VAARG, MVT::Other, Custom);
191 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
192 setOperationAction(ISD::VAEND, MVT::Other, Expand);
194 // Variable-sized objects.
195 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
196 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
197 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
199 // Exception handling.
200 // FIXME: These are guesses. Has this been defined yet?
201 setExceptionPointerRegister(AArch64::X0);
202 setExceptionSelectorRegister(AArch64::X1);
204 // Constant pool entries
205 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
208 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
210 // Add/Sub overflow ops with MVT::Glues are lowered to NZCV dependences.
211 setOperationAction(ISD::ADDC, MVT::i32, Custom);
212 setOperationAction(ISD::ADDE, MVT::i32, Custom);
213 setOperationAction(ISD::SUBC, MVT::i32, Custom);
214 setOperationAction(ISD::SUBE, MVT::i32, Custom);
215 setOperationAction(ISD::ADDC, MVT::i64, Custom);
216 setOperationAction(ISD::ADDE, MVT::i64, Custom);
217 setOperationAction(ISD::SUBC, MVT::i64, Custom);
218 setOperationAction(ISD::SUBE, MVT::i64, Custom);
220 // AArch64 lacks both left-rotate and popcount instructions.
221 setOperationAction(ISD::ROTL, MVT::i32, Expand);
222 setOperationAction(ISD::ROTL, MVT::i64, Expand);
224 // AArch64 doesn't have {U|S}MUL_LOHI.
225 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
226 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
229 // Expand the undefined-at-zero variants to cttz/ctlz to their defined-at-zero
230 // counterparts, which AArch64 supports directly.
231 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
232 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
233 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
234 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
236 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
237 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
239 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
240 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
241 setOperationAction(ISD::SREM, MVT::i32, Expand);
242 setOperationAction(ISD::SREM, MVT::i64, Expand);
243 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
244 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
245 setOperationAction(ISD::UREM, MVT::i32, Expand);
246 setOperationAction(ISD::UREM, MVT::i64, Expand);
248 // Custom lower Add/Sub/Mul with overflow.
249 setOperationAction(ISD::SADDO, MVT::i32, Custom);
250 setOperationAction(ISD::SADDO, MVT::i64, Custom);
251 setOperationAction(ISD::UADDO, MVT::i32, Custom);
252 setOperationAction(ISD::UADDO, MVT::i64, Custom);
253 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
254 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
255 setOperationAction(ISD::USUBO, MVT::i32, Custom);
256 setOperationAction(ISD::USUBO, MVT::i64, Custom);
257 setOperationAction(ISD::SMULO, MVT::i32, Custom);
258 setOperationAction(ISD::SMULO, MVT::i64, Custom);
259 setOperationAction(ISD::UMULO, MVT::i32, Custom);
260 setOperationAction(ISD::UMULO, MVT::i64, Custom);
262 setOperationAction(ISD::FSIN, MVT::f32, Expand);
263 setOperationAction(ISD::FSIN, MVT::f64, Expand);
264 setOperationAction(ISD::FCOS, MVT::f32, Expand);
265 setOperationAction(ISD::FCOS, MVT::f64, Expand);
266 setOperationAction(ISD::FPOW, MVT::f32, Expand);
267 setOperationAction(ISD::FPOW, MVT::f64, Expand);
268 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
269 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
271 // f16 is a storage-only type, always promote it to f32.
272 setOperationAction(ISD::SETCC, MVT::f16, Promote);
273 setOperationAction(ISD::BR_CC, MVT::f16, Promote);
274 setOperationAction(ISD::SELECT_CC, MVT::f16, Promote);
275 setOperationAction(ISD::SELECT, MVT::f16, Promote);
276 setOperationAction(ISD::FADD, MVT::f16, Promote);
277 setOperationAction(ISD::FSUB, MVT::f16, Promote);
278 setOperationAction(ISD::FMUL, MVT::f16, Promote);
279 setOperationAction(ISD::FDIV, MVT::f16, Promote);
280 setOperationAction(ISD::FREM, MVT::f16, Promote);
281 setOperationAction(ISD::FMA, MVT::f16, Promote);
282 setOperationAction(ISD::FNEG, MVT::f16, Promote);
283 setOperationAction(ISD::FABS, MVT::f16, Promote);
284 setOperationAction(ISD::FCEIL, MVT::f16, Promote);
285 setOperationAction(ISD::FCOPYSIGN, MVT::f16, Promote);
286 setOperationAction(ISD::FCOS, MVT::f16, Promote);
287 setOperationAction(ISD::FFLOOR, MVT::f16, Promote);
288 setOperationAction(ISD::FNEARBYINT, MVT::f16, Promote);
289 setOperationAction(ISD::FPOW, MVT::f16, Promote);
290 setOperationAction(ISD::FPOWI, MVT::f16, Promote);
291 setOperationAction(ISD::FRINT, MVT::f16, Promote);
292 setOperationAction(ISD::FSIN, MVT::f16, Promote);
293 setOperationAction(ISD::FSINCOS, MVT::f16, Promote);
294 setOperationAction(ISD::FSQRT, MVT::f16, Promote);
295 setOperationAction(ISD::FEXP, MVT::f16, Promote);
296 setOperationAction(ISD::FEXP2, MVT::f16, Promote);
297 setOperationAction(ISD::FLOG, MVT::f16, Promote);
298 setOperationAction(ISD::FLOG2, MVT::f16, Promote);
299 setOperationAction(ISD::FLOG10, MVT::f16, Promote);
300 setOperationAction(ISD::FROUND, MVT::f16, Promote);
301 setOperationAction(ISD::FTRUNC, MVT::f16, Promote);
302 setOperationAction(ISD::FMINNUM, MVT::f16, Promote);
303 setOperationAction(ISD::FMAXNUM, MVT::f16, Promote);
305 // v4f16 is also a storage-only type, so promote it to v4f32 when that is
307 setOperationAction(ISD::FADD, MVT::v4f16, Promote);
308 setOperationAction(ISD::FSUB, MVT::v4f16, Promote);
309 setOperationAction(ISD::FMUL, MVT::v4f16, Promote);
310 setOperationAction(ISD::FDIV, MVT::v4f16, Promote);
311 setOperationAction(ISD::FP_EXTEND, MVT::v4f16, Promote);
312 setOperationAction(ISD::FP_ROUND, MVT::v4f16, Promote);
313 AddPromotedToType(ISD::FADD, MVT::v4f16, MVT::v4f32);
314 AddPromotedToType(ISD::FSUB, MVT::v4f16, MVT::v4f32);
315 AddPromotedToType(ISD::FMUL, MVT::v4f16, MVT::v4f32);
316 AddPromotedToType(ISD::FDIV, MVT::v4f16, MVT::v4f32);
317 AddPromotedToType(ISD::FP_EXTEND, MVT::v4f16, MVT::v4f32);
318 AddPromotedToType(ISD::FP_ROUND, MVT::v4f16, MVT::v4f32);
320 // Expand all other v4f16 operations.
321 // FIXME: We could generate better code by promoting some operations to
323 setOperationAction(ISD::FABS, MVT::v4f16, Expand);
324 setOperationAction(ISD::FCEIL, MVT::v4f16, Expand);
325 setOperationAction(ISD::FCOPYSIGN, MVT::v4f16, Expand);
326 setOperationAction(ISD::FCOS, MVT::v4f16, Expand);
327 setOperationAction(ISD::FFLOOR, MVT::v4f16, Expand);
328 setOperationAction(ISD::FMA, MVT::v4f16, Expand);
329 setOperationAction(ISD::FNEARBYINT, MVT::v4f16, Expand);
330 setOperationAction(ISD::FNEG, MVT::v4f16, Expand);
331 setOperationAction(ISD::FPOW, MVT::v4f16, Expand);
332 setOperationAction(ISD::FPOWI, MVT::v4f16, Expand);
333 setOperationAction(ISD::FREM, MVT::v4f16, Expand);
334 setOperationAction(ISD::FROUND, MVT::v4f16, Expand);
335 setOperationAction(ISD::FRINT, MVT::v4f16, Expand);
336 setOperationAction(ISD::FSIN, MVT::v4f16, Expand);
337 setOperationAction(ISD::FSINCOS, MVT::v4f16, Expand);
338 setOperationAction(ISD::FSQRT, MVT::v4f16, Expand);
339 setOperationAction(ISD::FTRUNC, MVT::v4f16, Expand);
340 setOperationAction(ISD::SETCC, MVT::v4f16, Expand);
341 setOperationAction(ISD::BR_CC, MVT::v4f16, Expand);
342 setOperationAction(ISD::SELECT, MVT::v4f16, Expand);
343 setOperationAction(ISD::SELECT_CC, MVT::v4f16, Expand);
344 setOperationAction(ISD::FEXP, MVT::v4f16, Expand);
345 setOperationAction(ISD::FEXP2, MVT::v4f16, Expand);
346 setOperationAction(ISD::FLOG, MVT::v4f16, Expand);
347 setOperationAction(ISD::FLOG2, MVT::v4f16, Expand);
348 setOperationAction(ISD::FLOG10, MVT::v4f16, Expand);
351 // v8f16 is also a storage-only type, so expand it.
352 setOperationAction(ISD::FABS, MVT::v8f16, Expand);
353 setOperationAction(ISD::FADD, MVT::v8f16, Expand);
354 setOperationAction(ISD::FCEIL, MVT::v8f16, Expand);
355 setOperationAction(ISD::FCOPYSIGN, MVT::v8f16, Expand);
356 setOperationAction(ISD::FCOS, MVT::v8f16, Expand);
357 setOperationAction(ISD::FDIV, MVT::v8f16, Expand);
358 setOperationAction(ISD::FFLOOR, MVT::v8f16, Expand);
359 setOperationAction(ISD::FMA, MVT::v8f16, Expand);
360 setOperationAction(ISD::FMUL, MVT::v8f16, Expand);
361 setOperationAction(ISD::FNEARBYINT, MVT::v8f16, Expand);
362 setOperationAction(ISD::FNEG, MVT::v8f16, Expand);
363 setOperationAction(ISD::FPOW, MVT::v8f16, Expand);
364 setOperationAction(ISD::FPOWI, MVT::v8f16, Expand);
365 setOperationAction(ISD::FREM, MVT::v8f16, Expand);
366 setOperationAction(ISD::FROUND, MVT::v8f16, Expand);
367 setOperationAction(ISD::FRINT, MVT::v8f16, Expand);
368 setOperationAction(ISD::FSIN, MVT::v8f16, Expand);
369 setOperationAction(ISD::FSINCOS, MVT::v8f16, Expand);
370 setOperationAction(ISD::FSQRT, MVT::v8f16, Expand);
371 setOperationAction(ISD::FSUB, MVT::v8f16, Expand);
372 setOperationAction(ISD::FTRUNC, MVT::v8f16, Expand);
373 setOperationAction(ISD::SETCC, MVT::v8f16, Expand);
374 setOperationAction(ISD::BR_CC, MVT::v8f16, Expand);
375 setOperationAction(ISD::SELECT, MVT::v8f16, Expand);
376 setOperationAction(ISD::SELECT_CC, MVT::v8f16, Expand);
377 setOperationAction(ISD::FP_EXTEND, MVT::v8f16, Expand);
378 setOperationAction(ISD::FEXP, MVT::v8f16, Expand);
379 setOperationAction(ISD::FEXP2, MVT::v8f16, Expand);
380 setOperationAction(ISD::FLOG, MVT::v8f16, Expand);
381 setOperationAction(ISD::FLOG2, MVT::v8f16, Expand);
382 setOperationAction(ISD::FLOG10, MVT::v8f16, Expand);
384 // AArch64 has implementations of a lot of rounding-like FP operations.
385 for (MVT Ty : {MVT::f32, MVT::f64}) {
386 setOperationAction(ISD::FFLOOR, Ty, Legal);
387 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
388 setOperationAction(ISD::FCEIL, Ty, Legal);
389 setOperationAction(ISD::FRINT, Ty, Legal);
390 setOperationAction(ISD::FTRUNC, Ty, Legal);
391 setOperationAction(ISD::FROUND, Ty, Legal);
392 setOperationAction(ISD::FMINNUM, Ty, Legal);
393 setOperationAction(ISD::FMAXNUM, Ty, Legal);
396 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
398 if (Subtarget->isTargetMachO()) {
399 // For iOS, we don't want to the normal expansion of a libcall to
400 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
402 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
403 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
405 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
406 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
409 // Make floating-point constants legal for the large code model, so they don't
410 // become loads from the constant pool.
411 if (Subtarget->isTargetMachO() && TM.getCodeModel() == CodeModel::Large) {
412 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
413 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
416 // AArch64 does not have floating-point extending loads, i1 sign-extending
417 // load, floating-point truncating stores, or v2i32->v2i16 truncating store.
418 for (MVT VT : MVT::fp_valuetypes()) {
419 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
420 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
421 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f64, Expand);
422 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f80, Expand);
424 for (MVT VT : MVT::integer_valuetypes())
425 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Expand);
427 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
428 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
429 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
430 setTruncStoreAction(MVT::f128, MVT::f80, Expand);
431 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
432 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
433 setTruncStoreAction(MVT::f128, MVT::f16, Expand);
435 setOperationAction(ISD::BITCAST, MVT::i16, Custom);
436 setOperationAction(ISD::BITCAST, MVT::f16, Custom);
438 // Indexed loads and stores are supported.
439 for (unsigned im = (unsigned)ISD::PRE_INC;
440 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
441 setIndexedLoadAction(im, MVT::i8, Legal);
442 setIndexedLoadAction(im, MVT::i16, Legal);
443 setIndexedLoadAction(im, MVT::i32, Legal);
444 setIndexedLoadAction(im, MVT::i64, Legal);
445 setIndexedLoadAction(im, MVT::f64, Legal);
446 setIndexedLoadAction(im, MVT::f32, Legal);
447 setIndexedLoadAction(im, MVT::f16, Legal);
448 setIndexedStoreAction(im, MVT::i8, Legal);
449 setIndexedStoreAction(im, MVT::i16, Legal);
450 setIndexedStoreAction(im, MVT::i32, Legal);
451 setIndexedStoreAction(im, MVT::i64, Legal);
452 setIndexedStoreAction(im, MVT::f64, Legal);
453 setIndexedStoreAction(im, MVT::f32, Legal);
454 setIndexedStoreAction(im, MVT::f16, Legal);
458 setOperationAction(ISD::TRAP, MVT::Other, Legal);
460 // We combine OR nodes for bitfield operations.
461 setTargetDAGCombine(ISD::OR);
463 // Vector add and sub nodes may conceal a high-half opportunity.
464 // Also, try to fold ADD into CSINC/CSINV..
465 setTargetDAGCombine(ISD::ADD);
466 setTargetDAGCombine(ISD::SUB);
468 setTargetDAGCombine(ISD::XOR);
469 setTargetDAGCombine(ISD::SINT_TO_FP);
470 setTargetDAGCombine(ISD::UINT_TO_FP);
472 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
474 setTargetDAGCombine(ISD::ANY_EXTEND);
475 setTargetDAGCombine(ISD::ZERO_EXTEND);
476 setTargetDAGCombine(ISD::SIGN_EXTEND);
477 setTargetDAGCombine(ISD::BITCAST);
478 setTargetDAGCombine(ISD::CONCAT_VECTORS);
479 setTargetDAGCombine(ISD::STORE);
481 setTargetDAGCombine(ISD::MUL);
483 setTargetDAGCombine(ISD::SELECT);
484 setTargetDAGCombine(ISD::VSELECT);
485 setTargetDAGCombine(ISD::SELECT_CC);
487 setTargetDAGCombine(ISD::INTRINSIC_VOID);
488 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
489 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
491 MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8;
492 MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4;
493 MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 4;
495 setStackPointerRegisterToSaveRestore(AArch64::SP);
497 setSchedulingPreference(Sched::Hybrid);
500 MaskAndBranchFoldingIsLegal = true;
501 EnableExtLdPromotion = true;
503 setMinFunctionAlignment(2);
505 setHasExtractBitsInsn(true);
507 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
509 if (Subtarget->hasNEON()) {
510 // FIXME: v1f64 shouldn't be legal if we can avoid it, because it leads to
511 // silliness like this:
512 setOperationAction(ISD::FABS, MVT::v1f64, Expand);
513 setOperationAction(ISD::FADD, MVT::v1f64, Expand);
514 setOperationAction(ISD::FCEIL, MVT::v1f64, Expand);
515 setOperationAction(ISD::FCOPYSIGN, MVT::v1f64, Expand);
516 setOperationAction(ISD::FCOS, MVT::v1f64, Expand);
517 setOperationAction(ISD::FDIV, MVT::v1f64, Expand);
518 setOperationAction(ISD::FFLOOR, MVT::v1f64, Expand);
519 setOperationAction(ISD::FMA, MVT::v1f64, Expand);
520 setOperationAction(ISD::FMUL, MVT::v1f64, Expand);
521 setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Expand);
522 setOperationAction(ISD::FNEG, MVT::v1f64, Expand);
523 setOperationAction(ISD::FPOW, MVT::v1f64, Expand);
524 setOperationAction(ISD::FREM, MVT::v1f64, Expand);
525 setOperationAction(ISD::FROUND, MVT::v1f64, Expand);
526 setOperationAction(ISD::FRINT, MVT::v1f64, Expand);
527 setOperationAction(ISD::FSIN, MVT::v1f64, Expand);
528 setOperationAction(ISD::FSINCOS, MVT::v1f64, Expand);
529 setOperationAction(ISD::FSQRT, MVT::v1f64, Expand);
530 setOperationAction(ISD::FSUB, MVT::v1f64, Expand);
531 setOperationAction(ISD::FTRUNC, MVT::v1f64, Expand);
532 setOperationAction(ISD::SETCC, MVT::v1f64, Expand);
533 setOperationAction(ISD::BR_CC, MVT::v1f64, Expand);
534 setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
535 setOperationAction(ISD::SELECT_CC, MVT::v1f64, Expand);
536 setOperationAction(ISD::FP_EXTEND, MVT::v1f64, Expand);
538 setOperationAction(ISD::FP_TO_SINT, MVT::v1i64, Expand);
539 setOperationAction(ISD::FP_TO_UINT, MVT::v1i64, Expand);
540 setOperationAction(ISD::SINT_TO_FP, MVT::v1i64, Expand);
541 setOperationAction(ISD::UINT_TO_FP, MVT::v1i64, Expand);
542 setOperationAction(ISD::FP_ROUND, MVT::v1f64, Expand);
544 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
546 // AArch64 doesn't have a direct vector ->f32 conversion instructions for
547 // elements smaller than i32, so promote the input to i32 first.
548 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Promote);
549 setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Promote);
550 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Promote);
551 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Promote);
552 // i8 and i16 vector elements also need promotion to i32 for v8i8 or v8i16
553 // -> v8f16 conversions.
554 setOperationAction(ISD::SINT_TO_FP, MVT::v8i8, Promote);
555 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Promote);
556 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
557 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Promote);
558 // Similarly, there is no direct i32 -> f64 vector conversion instruction.
559 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
560 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
561 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
562 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
563 // Or, direct i32 -> f16 vector conversion. Set it so custom, so the
564 // conversion happens in two steps: v4i32 -> v4f32 -> v4f16
565 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Custom);
566 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
568 // AArch64 doesn't have MUL.2d:
569 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
570 // Custom handling for some quad-vector types to detect MULL.
571 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
572 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
573 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
575 setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal);
576 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
577 // Likewise, narrowing and extending vector loads/stores aren't handled
579 for (MVT VT : MVT::vector_valuetypes()) {
580 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
582 setOperationAction(ISD::MULHS, VT, Expand);
583 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
584 setOperationAction(ISD::MULHU, VT, Expand);
585 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
587 setOperationAction(ISD::BSWAP, VT, Expand);
589 for (MVT InnerVT : MVT::vector_valuetypes()) {
590 setTruncStoreAction(VT, InnerVT, Expand);
591 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
592 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
593 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
597 // AArch64 has implementations of a lot of rounding-like FP operations.
598 for (MVT Ty : {MVT::v2f32, MVT::v4f32, MVT::v2f64}) {
599 setOperationAction(ISD::FFLOOR, Ty, Legal);
600 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
601 setOperationAction(ISD::FCEIL, Ty, Legal);
602 setOperationAction(ISD::FRINT, Ty, Legal);
603 setOperationAction(ISD::FTRUNC, Ty, Legal);
604 setOperationAction(ISD::FROUND, Ty, Legal);
608 // Prefer likely predicted branches to selects on out-of-order cores.
609 if (Subtarget->isCortexA57())
610 PredictableSelectIsExpensive = true;
613 void AArch64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) {
614 if (VT == MVT::v2f32 || VT == MVT::v4f16) {
615 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
616 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i32);
618 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
619 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i32);
620 } else if (VT == MVT::v2f64 || VT == MVT::v4f32 || VT == MVT::v8f16) {
621 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
622 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i64);
624 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
625 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i64);
628 // Mark vector float intrinsics as expand.
629 if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) {
630 setOperationAction(ISD::FSIN, VT.getSimpleVT(), Expand);
631 setOperationAction(ISD::FCOS, VT.getSimpleVT(), Expand);
632 setOperationAction(ISD::FPOWI, VT.getSimpleVT(), Expand);
633 setOperationAction(ISD::FPOW, VT.getSimpleVT(), Expand);
634 setOperationAction(ISD::FLOG, VT.getSimpleVT(), Expand);
635 setOperationAction(ISD::FLOG2, VT.getSimpleVT(), Expand);
636 setOperationAction(ISD::FLOG10, VT.getSimpleVT(), Expand);
637 setOperationAction(ISD::FEXP, VT.getSimpleVT(), Expand);
638 setOperationAction(ISD::FEXP2, VT.getSimpleVT(), Expand);
640 // But we do support custom-lowering for FCOPYSIGN.
641 setOperationAction(ISD::FCOPYSIGN, VT.getSimpleVT(), Custom);
644 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
645 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
646 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
647 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
648 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Custom);
649 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
650 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
651 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
652 setOperationAction(ISD::AND, VT.getSimpleVT(), Custom);
653 setOperationAction(ISD::OR, VT.getSimpleVT(), Custom);
654 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
655 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
657 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
658 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
659 setOperationAction(ISD::VSELECT, VT.getSimpleVT(), Expand);
660 for (MVT InnerVT : MVT::all_valuetypes())
661 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT.getSimpleVT(), Expand);
663 // CNT supports only B element sizes.
664 if (VT != MVT::v8i8 && VT != MVT::v16i8)
665 setOperationAction(ISD::CTPOP, VT.getSimpleVT(), Expand);
667 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
668 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
669 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
670 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
671 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
673 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
674 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
676 // [SU][MIN|MAX] and [SU]ABSDIFF are available for all NEON types apart from
678 if (!VT.isFloatingPoint() &&
679 VT.getSimpleVT() != MVT::v2i64 && VT.getSimpleVT() != MVT::v1i64)
680 for (unsigned Opcode : {ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX,
681 ISD::SABSDIFF, ISD::UABSDIFF})
682 setOperationAction(Opcode, VT.getSimpleVT(), Legal);
684 // F[MIN|MAX][NUM|NAN] are available for all FP NEON types.
685 if (VT.isFloatingPoint())
686 for (unsigned Opcode : {ISD::FMINNAN, ISD::FMAXNAN,
687 ISD::FMINNUM, ISD::FMAXNUM})
688 setOperationAction(Opcode, VT.getSimpleVT(), Legal);
690 if (Subtarget->isLittleEndian()) {
691 for (unsigned im = (unsigned)ISD::PRE_INC;
692 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
693 setIndexedLoadAction(im, VT.getSimpleVT(), Legal);
694 setIndexedStoreAction(im, VT.getSimpleVT(), Legal);
699 void AArch64TargetLowering::addDRTypeForNEON(MVT VT) {
700 addRegisterClass(VT, &AArch64::FPR64RegClass);
701 addTypeForNEON(VT, MVT::v2i32);
704 void AArch64TargetLowering::addQRTypeForNEON(MVT VT) {
705 addRegisterClass(VT, &AArch64::FPR128RegClass);
706 addTypeForNEON(VT, MVT::v4i32);
709 EVT AArch64TargetLowering::getSetCCResultType(const DataLayout &, LLVMContext &,
713 return VT.changeVectorElementTypeToInteger();
716 /// computeKnownBitsForTargetNode - Determine which of the bits specified in
717 /// Mask are known to be either zero or one and return them in the
718 /// KnownZero/KnownOne bitsets.
719 void AArch64TargetLowering::computeKnownBitsForTargetNode(
720 const SDValue Op, APInt &KnownZero, APInt &KnownOne,
721 const SelectionDAG &DAG, unsigned Depth) const {
722 switch (Op.getOpcode()) {
725 case AArch64ISD::CSEL: {
726 APInt KnownZero2, KnownOne2;
727 DAG.computeKnownBits(Op->getOperand(0), KnownZero, KnownOne, Depth + 1);
728 DAG.computeKnownBits(Op->getOperand(1), KnownZero2, KnownOne2, Depth + 1);
729 KnownZero &= KnownZero2;
730 KnownOne &= KnownOne2;
733 case ISD::INTRINSIC_W_CHAIN: {
734 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
735 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
738 case Intrinsic::aarch64_ldaxr:
739 case Intrinsic::aarch64_ldxr: {
740 unsigned BitWidth = KnownOne.getBitWidth();
741 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
742 unsigned MemBits = VT.getScalarType().getSizeInBits();
743 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
749 case ISD::INTRINSIC_WO_CHAIN:
750 case ISD::INTRINSIC_VOID: {
751 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
755 case Intrinsic::aarch64_neon_umaxv:
756 case Intrinsic::aarch64_neon_uminv: {
757 // Figure out the datatype of the vector operand. The UMINV instruction
758 // will zero extend the result, so we can mark as known zero all the
759 // bits larger than the element datatype. 32-bit or larget doesn't need
760 // this as those are legal types and will be handled by isel directly.
761 MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
762 unsigned BitWidth = KnownZero.getBitWidth();
763 if (VT == MVT::v8i8 || VT == MVT::v16i8) {
764 assert(BitWidth >= 8 && "Unexpected width!");
765 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8);
767 } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
768 assert(BitWidth >= 16 && "Unexpected width!");
769 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
779 MVT AArch64TargetLowering::getScalarShiftAmountTy(const DataLayout &DL,
784 bool AArch64TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
788 if (Subtarget->requiresStrictAlign())
790 // FIXME: True for Cyclone, but not necessary others.
797 AArch64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
798 const TargetLibraryInfo *libInfo) const {
799 return AArch64::createFastISel(funcInfo, libInfo);
802 const char *AArch64TargetLowering::getTargetNodeName(unsigned Opcode) const {
803 switch ((AArch64ISD::NodeType)Opcode) {
804 case AArch64ISD::FIRST_NUMBER: break;
805 case AArch64ISD::CALL: return "AArch64ISD::CALL";
806 case AArch64ISD::ADRP: return "AArch64ISD::ADRP";
807 case AArch64ISD::ADDlow: return "AArch64ISD::ADDlow";
808 case AArch64ISD::LOADgot: return "AArch64ISD::LOADgot";
809 case AArch64ISD::RET_FLAG: return "AArch64ISD::RET_FLAG";
810 case AArch64ISD::BRCOND: return "AArch64ISD::BRCOND";
811 case AArch64ISD::CSEL: return "AArch64ISD::CSEL";
812 case AArch64ISD::FCSEL: return "AArch64ISD::FCSEL";
813 case AArch64ISD::CSINV: return "AArch64ISD::CSINV";
814 case AArch64ISD::CSNEG: return "AArch64ISD::CSNEG";
815 case AArch64ISD::CSINC: return "AArch64ISD::CSINC";
816 case AArch64ISD::THREAD_POINTER: return "AArch64ISD::THREAD_POINTER";
817 case AArch64ISD::TLSDESC_CALLSEQ: return "AArch64ISD::TLSDESC_CALLSEQ";
818 case AArch64ISD::ADC: return "AArch64ISD::ADC";
819 case AArch64ISD::SBC: return "AArch64ISD::SBC";
820 case AArch64ISD::ADDS: return "AArch64ISD::ADDS";
821 case AArch64ISD::SUBS: return "AArch64ISD::SUBS";
822 case AArch64ISD::ADCS: return "AArch64ISD::ADCS";
823 case AArch64ISD::SBCS: return "AArch64ISD::SBCS";
824 case AArch64ISD::ANDS: return "AArch64ISD::ANDS";
825 case AArch64ISD::CCMP: return "AArch64ISD::CCMP";
826 case AArch64ISD::CCMN: return "AArch64ISD::CCMN";
827 case AArch64ISD::FCCMP: return "AArch64ISD::FCCMP";
828 case AArch64ISD::FCMP: return "AArch64ISD::FCMP";
829 case AArch64ISD::DUP: return "AArch64ISD::DUP";
830 case AArch64ISD::DUPLANE8: return "AArch64ISD::DUPLANE8";
831 case AArch64ISD::DUPLANE16: return "AArch64ISD::DUPLANE16";
832 case AArch64ISD::DUPLANE32: return "AArch64ISD::DUPLANE32";
833 case AArch64ISD::DUPLANE64: return "AArch64ISD::DUPLANE64";
834 case AArch64ISD::MOVI: return "AArch64ISD::MOVI";
835 case AArch64ISD::MOVIshift: return "AArch64ISD::MOVIshift";
836 case AArch64ISD::MOVIedit: return "AArch64ISD::MOVIedit";
837 case AArch64ISD::MOVImsl: return "AArch64ISD::MOVImsl";
838 case AArch64ISD::FMOV: return "AArch64ISD::FMOV";
839 case AArch64ISD::MVNIshift: return "AArch64ISD::MVNIshift";
840 case AArch64ISD::MVNImsl: return "AArch64ISD::MVNImsl";
841 case AArch64ISD::BICi: return "AArch64ISD::BICi";
842 case AArch64ISD::ORRi: return "AArch64ISD::ORRi";
843 case AArch64ISD::BSL: return "AArch64ISD::BSL";
844 case AArch64ISD::NEG: return "AArch64ISD::NEG";
845 case AArch64ISD::EXTR: return "AArch64ISD::EXTR";
846 case AArch64ISD::ZIP1: return "AArch64ISD::ZIP1";
847 case AArch64ISD::ZIP2: return "AArch64ISD::ZIP2";
848 case AArch64ISD::UZP1: return "AArch64ISD::UZP1";
849 case AArch64ISD::UZP2: return "AArch64ISD::UZP2";
850 case AArch64ISD::TRN1: return "AArch64ISD::TRN1";
851 case AArch64ISD::TRN2: return "AArch64ISD::TRN2";
852 case AArch64ISD::REV16: return "AArch64ISD::REV16";
853 case AArch64ISD::REV32: return "AArch64ISD::REV32";
854 case AArch64ISD::REV64: return "AArch64ISD::REV64";
855 case AArch64ISD::EXT: return "AArch64ISD::EXT";
856 case AArch64ISD::VSHL: return "AArch64ISD::VSHL";
857 case AArch64ISD::VLSHR: return "AArch64ISD::VLSHR";
858 case AArch64ISD::VASHR: return "AArch64ISD::VASHR";
859 case AArch64ISD::CMEQ: return "AArch64ISD::CMEQ";
860 case AArch64ISD::CMGE: return "AArch64ISD::CMGE";
861 case AArch64ISD::CMGT: return "AArch64ISD::CMGT";
862 case AArch64ISD::CMHI: return "AArch64ISD::CMHI";
863 case AArch64ISD::CMHS: return "AArch64ISD::CMHS";
864 case AArch64ISD::FCMEQ: return "AArch64ISD::FCMEQ";
865 case AArch64ISD::FCMGE: return "AArch64ISD::FCMGE";
866 case AArch64ISD::FCMGT: return "AArch64ISD::FCMGT";
867 case AArch64ISD::CMEQz: return "AArch64ISD::CMEQz";
868 case AArch64ISD::CMGEz: return "AArch64ISD::CMGEz";
869 case AArch64ISD::CMGTz: return "AArch64ISD::CMGTz";
870 case AArch64ISD::CMLEz: return "AArch64ISD::CMLEz";
871 case AArch64ISD::CMLTz: return "AArch64ISD::CMLTz";
872 case AArch64ISD::FCMEQz: return "AArch64ISD::FCMEQz";
873 case AArch64ISD::FCMGEz: return "AArch64ISD::FCMGEz";
874 case AArch64ISD::FCMGTz: return "AArch64ISD::FCMGTz";
875 case AArch64ISD::FCMLEz: return "AArch64ISD::FCMLEz";
876 case AArch64ISD::FCMLTz: return "AArch64ISD::FCMLTz";
877 case AArch64ISD::SADDV: return "AArch64ISD::SADDV";
878 case AArch64ISD::UADDV: return "AArch64ISD::UADDV";
879 case AArch64ISD::SMINV: return "AArch64ISD::SMINV";
880 case AArch64ISD::UMINV: return "AArch64ISD::UMINV";
881 case AArch64ISD::SMAXV: return "AArch64ISD::SMAXV";
882 case AArch64ISD::UMAXV: return "AArch64ISD::UMAXV";
883 case AArch64ISD::NOT: return "AArch64ISD::NOT";
884 case AArch64ISD::BIT: return "AArch64ISD::BIT";
885 case AArch64ISD::CBZ: return "AArch64ISD::CBZ";
886 case AArch64ISD::CBNZ: return "AArch64ISD::CBNZ";
887 case AArch64ISD::TBZ: return "AArch64ISD::TBZ";
888 case AArch64ISD::TBNZ: return "AArch64ISD::TBNZ";
889 case AArch64ISD::TC_RETURN: return "AArch64ISD::TC_RETURN";
890 case AArch64ISD::PREFETCH: return "AArch64ISD::PREFETCH";
891 case AArch64ISD::SITOF: return "AArch64ISD::SITOF";
892 case AArch64ISD::UITOF: return "AArch64ISD::UITOF";
893 case AArch64ISD::NVCAST: return "AArch64ISD::NVCAST";
894 case AArch64ISD::SQSHL_I: return "AArch64ISD::SQSHL_I";
895 case AArch64ISD::UQSHL_I: return "AArch64ISD::UQSHL_I";
896 case AArch64ISD::SRSHR_I: return "AArch64ISD::SRSHR_I";
897 case AArch64ISD::URSHR_I: return "AArch64ISD::URSHR_I";
898 case AArch64ISD::SQSHLU_I: return "AArch64ISD::SQSHLU_I";
899 case AArch64ISD::WrapperLarge: return "AArch64ISD::WrapperLarge";
900 case AArch64ISD::LD2post: return "AArch64ISD::LD2post";
901 case AArch64ISD::LD3post: return "AArch64ISD::LD3post";
902 case AArch64ISD::LD4post: return "AArch64ISD::LD4post";
903 case AArch64ISD::ST2post: return "AArch64ISD::ST2post";
904 case AArch64ISD::ST3post: return "AArch64ISD::ST3post";
905 case AArch64ISD::ST4post: return "AArch64ISD::ST4post";
906 case AArch64ISD::LD1x2post: return "AArch64ISD::LD1x2post";
907 case AArch64ISD::LD1x3post: return "AArch64ISD::LD1x3post";
908 case AArch64ISD::LD1x4post: return "AArch64ISD::LD1x4post";
909 case AArch64ISD::ST1x2post: return "AArch64ISD::ST1x2post";
910 case AArch64ISD::ST1x3post: return "AArch64ISD::ST1x3post";
911 case AArch64ISD::ST1x4post: return "AArch64ISD::ST1x4post";
912 case AArch64ISD::LD1DUPpost: return "AArch64ISD::LD1DUPpost";
913 case AArch64ISD::LD2DUPpost: return "AArch64ISD::LD2DUPpost";
914 case AArch64ISD::LD3DUPpost: return "AArch64ISD::LD3DUPpost";
915 case AArch64ISD::LD4DUPpost: return "AArch64ISD::LD4DUPpost";
916 case AArch64ISD::LD1LANEpost: return "AArch64ISD::LD1LANEpost";
917 case AArch64ISD::LD2LANEpost: return "AArch64ISD::LD2LANEpost";
918 case AArch64ISD::LD3LANEpost: return "AArch64ISD::LD3LANEpost";
919 case AArch64ISD::LD4LANEpost: return "AArch64ISD::LD4LANEpost";
920 case AArch64ISD::ST2LANEpost: return "AArch64ISD::ST2LANEpost";
921 case AArch64ISD::ST3LANEpost: return "AArch64ISD::ST3LANEpost";
922 case AArch64ISD::ST4LANEpost: return "AArch64ISD::ST4LANEpost";
923 case AArch64ISD::SMULL: return "AArch64ISD::SMULL";
924 case AArch64ISD::UMULL: return "AArch64ISD::UMULL";
930 AArch64TargetLowering::EmitF128CSEL(MachineInstr *MI,
931 MachineBasicBlock *MBB) const {
932 // We materialise the F128CSEL pseudo-instruction as some control flow and a
936 // [... previous instrs leading to comparison ...]
942 // Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB]
944 MachineFunction *MF = MBB->getParent();
945 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
946 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
947 DebugLoc DL = MI->getDebugLoc();
948 MachineFunction::iterator It = MBB;
951 unsigned DestReg = MI->getOperand(0).getReg();
952 unsigned IfTrueReg = MI->getOperand(1).getReg();
953 unsigned IfFalseReg = MI->getOperand(2).getReg();
954 unsigned CondCode = MI->getOperand(3).getImm();
955 bool NZCVKilled = MI->getOperand(4).isKill();
957 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
958 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
959 MF->insert(It, TrueBB);
960 MF->insert(It, EndBB);
962 // Transfer rest of current basic-block to EndBB
963 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
965 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
967 BuildMI(MBB, DL, TII->get(AArch64::Bcc)).addImm(CondCode).addMBB(TrueBB);
968 BuildMI(MBB, DL, TII->get(AArch64::B)).addMBB(EndBB);
969 MBB->addSuccessor(TrueBB);
970 MBB->addSuccessor(EndBB);
972 // TrueBB falls through to the end.
973 TrueBB->addSuccessor(EndBB);
976 TrueBB->addLiveIn(AArch64::NZCV);
977 EndBB->addLiveIn(AArch64::NZCV);
980 BuildMI(*EndBB, EndBB->begin(), DL, TII->get(AArch64::PHI), DestReg)
986 MI->eraseFromParent();
991 AArch64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
992 MachineBasicBlock *BB) const {
993 switch (MI->getOpcode()) {
998 llvm_unreachable("Unexpected instruction for custom inserter!");
1000 case AArch64::F128CSEL:
1001 return EmitF128CSEL(MI, BB);
1003 case TargetOpcode::STACKMAP:
1004 case TargetOpcode::PATCHPOINT:
1005 return emitPatchPoint(MI, BB);
1009 //===----------------------------------------------------------------------===//
1010 // AArch64 Lowering private implementation.
1011 //===----------------------------------------------------------------------===//
1013 //===----------------------------------------------------------------------===//
1015 //===----------------------------------------------------------------------===//
1017 /// changeIntCCToAArch64CC - Convert a DAG integer condition code to an AArch64
1019 static AArch64CC::CondCode changeIntCCToAArch64CC(ISD::CondCode CC) {
1022 llvm_unreachable("Unknown condition code!");
1024 return AArch64CC::NE;
1026 return AArch64CC::EQ;
1028 return AArch64CC::GT;
1030 return AArch64CC::GE;
1032 return AArch64CC::LT;
1034 return AArch64CC::LE;
1036 return AArch64CC::HI;
1038 return AArch64CC::HS;
1040 return AArch64CC::LO;
1042 return AArch64CC::LS;
1046 /// changeFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64 CC.
1047 static void changeFPCCToAArch64CC(ISD::CondCode CC,
1048 AArch64CC::CondCode &CondCode,
1049 AArch64CC::CondCode &CondCode2) {
1050 CondCode2 = AArch64CC::AL;
1053 llvm_unreachable("Unknown FP condition!");
1056 CondCode = AArch64CC::EQ;
1060 CondCode = AArch64CC::GT;
1064 CondCode = AArch64CC::GE;
1067 CondCode = AArch64CC::MI;
1070 CondCode = AArch64CC::LS;
1073 CondCode = AArch64CC::MI;
1074 CondCode2 = AArch64CC::GT;
1077 CondCode = AArch64CC::VC;
1080 CondCode = AArch64CC::VS;
1083 CondCode = AArch64CC::EQ;
1084 CondCode2 = AArch64CC::VS;
1087 CondCode = AArch64CC::HI;
1090 CondCode = AArch64CC::PL;
1094 CondCode = AArch64CC::LT;
1098 CondCode = AArch64CC::LE;
1102 CondCode = AArch64CC::NE;
1107 /// changeVectorFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64
1108 /// CC usable with the vector instructions. Fewer operations are available
1109 /// without a real NZCV register, so we have to use less efficient combinations
1110 /// to get the same effect.
1111 static void changeVectorFPCCToAArch64CC(ISD::CondCode CC,
1112 AArch64CC::CondCode &CondCode,
1113 AArch64CC::CondCode &CondCode2,
1118 // Mostly the scalar mappings work fine.
1119 changeFPCCToAArch64CC(CC, CondCode, CondCode2);
1122 Invert = true; // Fallthrough
1124 CondCode = AArch64CC::MI;
1125 CondCode2 = AArch64CC::GE;
1132 // All of the compare-mask comparisons are ordered, but we can switch
1133 // between the two by a double inversion. E.g. ULE == !OGT.
1135 changeFPCCToAArch64CC(getSetCCInverse(CC, false), CondCode, CondCode2);
1140 static bool isLegalArithImmed(uint64_t C) {
1141 // Matches AArch64DAGToDAGISel::SelectArithImmed().
1142 return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
1145 static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1146 SDLoc dl, SelectionDAG &DAG) {
1147 EVT VT = LHS.getValueType();
1149 if (VT.isFloatingPoint())
1150 return DAG.getNode(AArch64ISD::FCMP, dl, VT, LHS, RHS);
1152 // The CMP instruction is just an alias for SUBS, and representing it as
1153 // SUBS means that it's possible to get CSE with subtract operations.
1154 // A later phase can perform the optimization of setting the destination
1155 // register to WZR/XZR if it ends up being unused.
1156 unsigned Opcode = AArch64ISD::SUBS;
1158 if (RHS.getOpcode() == ISD::SUB && isa<ConstantSDNode>(RHS.getOperand(0)) &&
1159 cast<ConstantSDNode>(RHS.getOperand(0))->getZExtValue() == 0 &&
1160 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1161 // We'd like to combine a (CMP op1, (sub 0, op2) into a CMN instruction on
1162 // the grounds that "op1 - (-op2) == op1 + op2". However, the C and V flags
1163 // can be set differently by this operation. It comes down to whether
1164 // "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are then
1165 // everything is fine. If not then the optimization is wrong. Thus general
1166 // comparisons are only valid if op2 != 0.
1168 // So, finally, the only LLVM-native comparisons that don't mention C and V
1169 // are SETEQ and SETNE. They're the only ones we can safely use CMN for in
1170 // the absence of information about op2.
1171 Opcode = AArch64ISD::ADDS;
1172 RHS = RHS.getOperand(1);
1173 } else if (LHS.getOpcode() == ISD::AND && isa<ConstantSDNode>(RHS) &&
1174 cast<ConstantSDNode>(RHS)->getZExtValue() == 0 &&
1175 !isUnsignedIntSetCC(CC)) {
1176 // Similarly, (CMP (and X, Y), 0) can be implemented with a TST
1177 // (a.k.a. ANDS) except that the flags are only guaranteed to work for one
1178 // of the signed comparisons.
1179 Opcode = AArch64ISD::ANDS;
1180 RHS = LHS.getOperand(1);
1181 LHS = LHS.getOperand(0);
1184 return DAG.getNode(Opcode, dl, DAG.getVTList(VT, MVT_CC), LHS, RHS)
1188 /// \defgroup AArch64CCMP CMP;CCMP matching
1190 /// These functions deal with the formation of CMP;CCMP;... sequences.
1191 /// The CCMP/CCMN/FCCMP/FCCMPE instructions allow the conditional execution of
1192 /// a comparison. They set the NZCV flags to a predefined value if their
1193 /// predicate is false. This allows to express arbitrary conjunctions, for
1194 /// example "cmp 0 (and (setCA (cmp A)) (setCB (cmp B))))"
1197 /// ccmp B, inv(CB), CA
1198 /// check for CB flags
1200 /// In general we can create code for arbitrary "... (and (and A B) C)"
1201 /// sequences. We can also implement some "or" expressions, because "(or A B)"
1202 /// is equivalent to "not (and (not A) (not B))" and we can implement some
1203 /// negation operations:
1204 /// We can negate the results of a single comparison by inverting the flags
1205 /// used when the predicate fails and inverting the flags tested in the next
1206 /// instruction; We can also negate the results of the whole previous
1207 /// conditional compare sequence by inverting the flags tested in the next
1208 /// instruction. However there is no way to negate the result of a partial
1211 /// Therefore on encountering an "or" expression we can negate the subtree on
1212 /// one side and have to be able to push the negate to the leafs of the subtree
1213 /// on the other side (see also the comments in code). As complete example:
1214 /// "or (or (setCA (cmp A)) (setCB (cmp B)))
1215 /// (and (setCC (cmp C)) (setCD (cmp D)))"
1216 /// is transformed to
1217 /// "not (and (not (and (setCC (cmp C)) (setCC (cmp D))))
1218 /// (and (not (setCA (cmp A)) (not (setCB (cmp B))))))"
1219 /// and implemented as:
1221 /// ccmp D, inv(CD), CC
1222 /// ccmp A, CA, inv(CD)
1223 /// ccmp B, CB, inv(CA)
1224 /// check for CB flags
1225 /// A counterexample is "or (and A B) (and C D)" which cannot be implemented
1226 /// by conditional compare sequences.
1229 /// Create a conditional comparison; Use CCMP, CCMN or FCCMP as appropriate.
1230 static SDValue emitConditionalComparison(SDValue LHS, SDValue RHS,
1231 ISD::CondCode CC, SDValue CCOp,
1232 SDValue Condition, unsigned NZCV,
1233 SDLoc DL, SelectionDAG &DAG) {
1234 unsigned Opcode = 0;
1235 if (LHS.getValueType().isFloatingPoint())
1236 Opcode = AArch64ISD::FCCMP;
1237 else if (RHS.getOpcode() == ISD::SUB) {
1238 SDValue SubOp0 = RHS.getOperand(0);
1239 if (const ConstantSDNode *SubOp0C = dyn_cast<ConstantSDNode>(SubOp0))
1240 if (SubOp0C->isNullValue() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1241 // See emitComparison() on why we can only do this for SETEQ and SETNE.
1242 Opcode = AArch64ISD::CCMN;
1243 RHS = RHS.getOperand(1);
1247 Opcode = AArch64ISD::CCMP;
1249 SDValue NZCVOp = DAG.getConstant(NZCV, DL, MVT::i32);
1250 return DAG.getNode(Opcode, DL, MVT_CC, LHS, RHS, NZCVOp, Condition, CCOp);
1253 /// Returns true if @p Val is a tree of AND/OR/SETCC operations.
1254 /// CanPushNegate is set to true if we can push a negate operation through
1255 /// the tree in a was that we are left with AND operations and negate operations
1256 /// at the leafs only. i.e. "not (or (or x y) z)" can be changed to
1257 /// "and (and (not x) (not y)) (not z)"; "not (or (and x y) z)" cannot be
1258 /// brought into such a form.
1259 static bool isConjunctionDisjunctionTree(const SDValue Val, bool &CanPushNegate,
1260 unsigned Depth = 0) {
1261 if (!Val.hasOneUse())
1263 unsigned Opcode = Val->getOpcode();
1264 if (Opcode == ISD::SETCC) {
1265 CanPushNegate = true;
1268 // Protect against stack overflow.
1271 if (Opcode == ISD::AND || Opcode == ISD::OR) {
1272 SDValue O0 = Val->getOperand(0);
1273 SDValue O1 = Val->getOperand(1);
1274 bool CanPushNegateL;
1275 if (!isConjunctionDisjunctionTree(O0, CanPushNegateL, Depth+1))
1277 bool CanPushNegateR;
1278 if (!isConjunctionDisjunctionTree(O1, CanPushNegateR, Depth+1))
1280 // We cannot push a negate through an AND operation (it would become an OR),
1281 // we can however change a (not (or x y)) to (and (not x) (not y)) if we can
1282 // push the negate through the x/y subtrees.
1283 CanPushNegate = (Opcode == ISD::OR) && CanPushNegateL && CanPushNegateR;
1289 /// Emit conjunction or disjunction tree with the CMP/FCMP followed by a chain
1290 /// of CCMP/CFCMP ops. See @ref AArch64CCMP.
1291 /// Tries to transform the given i1 producing node @p Val to a series compare
1292 /// and conditional compare operations. @returns an NZCV flags producing node
1293 /// and sets @p OutCC to the flags that should be tested or returns SDValue() if
1294 /// transformation was not possible.
1295 /// On recursive invocations @p PushNegate may be set to true to have negation
1296 /// effects pushed to the tree leafs; @p Predicate is an NZCV flag predicate
1297 /// for the comparisons in the current subtree; @p Depth limits the search
1298 /// depth to avoid stack overflow.
1299 static SDValue emitConjunctionDisjunctionTree(SelectionDAG &DAG, SDValue Val,
1300 AArch64CC::CondCode &OutCC, bool PushNegate = false,
1301 SDValue CCOp = SDValue(), AArch64CC::CondCode Predicate = AArch64CC::AL,
1302 unsigned Depth = 0) {
1303 // We're at a tree leaf, produce a conditional comparison operation.
1304 unsigned Opcode = Val->getOpcode();
1305 if (Opcode == ISD::SETCC) {
1306 SDValue LHS = Val->getOperand(0);
1307 SDValue RHS = Val->getOperand(1);
1308 ISD::CondCode CC = cast<CondCodeSDNode>(Val->getOperand(2))->get();
1309 bool isInteger = LHS.getValueType().isInteger();
1311 CC = getSetCCInverse(CC, isInteger);
1313 // Determine OutCC and handle FP special case.
1315 OutCC = changeIntCCToAArch64CC(CC);
1317 assert(LHS.getValueType().isFloatingPoint());
1318 AArch64CC::CondCode ExtraCC;
1319 changeFPCCToAArch64CC(CC, OutCC, ExtraCC);
1320 // Surpisingly some floating point conditions can't be tested with a
1321 // single condition code. Construct an additional comparison in this case.
1322 // See comment below on how we deal with OR conditions.
1323 if (ExtraCC != AArch64CC::AL) {
1325 if (!CCOp.getNode())
1326 ExtraCmp = emitComparison(LHS, RHS, CC, DL, DAG);
1328 SDValue ConditionOp = DAG.getConstant(Predicate, DL, MVT_CC);
1329 // Note that we want the inverse of ExtraCC, so NZCV is not inversed.
1330 unsigned NZCV = AArch64CC::getNZCVToSatisfyCondCode(ExtraCC);
1331 ExtraCmp = emitConditionalComparison(LHS, RHS, CC, CCOp, ConditionOp,
1335 Predicate = AArch64CC::getInvertedCondCode(ExtraCC);
1336 OutCC = AArch64CC::getInvertedCondCode(OutCC);
1340 // Produce a normal comparison if we are first in the chain
1341 if (!CCOp.getNode())
1342 return emitComparison(LHS, RHS, CC, DL, DAG);
1343 // Otherwise produce a ccmp.
1344 SDValue ConditionOp = DAG.getConstant(Predicate, DL, MVT_CC);
1345 AArch64CC::CondCode InvOutCC = AArch64CC::getInvertedCondCode(OutCC);
1346 unsigned NZCV = AArch64CC::getNZCVToSatisfyCondCode(InvOutCC);
1347 return emitConditionalComparison(LHS, RHS, CC, CCOp, ConditionOp, NZCV, DL,
1349 } else if (Opcode != ISD::AND && Opcode != ISD::OR)
1352 assert((Opcode == ISD::OR || !PushNegate)
1353 && "Can only push negate through OR operation");
1355 // Check if both sides can be transformed.
1356 SDValue LHS = Val->getOperand(0);
1357 SDValue RHS = Val->getOperand(1);
1358 bool CanPushNegateL;
1359 if (!isConjunctionDisjunctionTree(LHS, CanPushNegateL, Depth+1))
1361 bool CanPushNegateR;
1362 if (!isConjunctionDisjunctionTree(RHS, CanPushNegateR, Depth+1))
1365 // Do we need to negate our operands?
1366 bool NegateOperands = Opcode == ISD::OR;
1367 // We can negate the results of all previous operations by inverting the
1368 // predicate flags giving us a free negation for one side. For the other side
1369 // we need to be able to push the negation to the leafs of the tree.
1370 if (NegateOperands) {
1371 if (!CanPushNegateL && !CanPushNegateR)
1373 // Order the side where we can push the negate through to LHS.
1374 if (!CanPushNegateL && CanPushNegateR) {
1375 std::swap(LHS, RHS);
1376 CanPushNegateL = true;
1380 // Emit RHS. If we want to negate the tree we only need to push a negate
1381 // through if we are already in a PushNegate case, otherwise we can negate
1382 // the "flags to test" afterwards.
1383 AArch64CC::CondCode RHSCC;
1384 SDValue CmpR = emitConjunctionDisjunctionTree(DAG, RHS, RHSCC, PushNegate,
1385 CCOp, Predicate, Depth+1);
1386 if (NegateOperands && !PushNegate)
1387 RHSCC = AArch64CC::getInvertedCondCode(RHSCC);
1388 // Emit LHS. We must push the negate through if we need to negate it.
1389 SDValue CmpL = emitConjunctionDisjunctionTree(DAG, LHS, OutCC, NegateOperands,
1390 CmpR, RHSCC, Depth+1);
1391 // If we transformed an OR to and AND then we have to negate the result
1392 // (or absorb a PushNegate resulting in a double negation).
1393 if (Opcode == ISD::OR && !PushNegate)
1394 OutCC = AArch64CC::getInvertedCondCode(OutCC);
1400 static SDValue getAArch64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1401 SDValue &AArch64cc, SelectionDAG &DAG, SDLoc dl) {
1402 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1403 EVT VT = RHS.getValueType();
1404 uint64_t C = RHSC->getZExtValue();
1405 if (!isLegalArithImmed(C)) {
1406 // Constant does not fit, try adjusting it by one?
1412 if ((VT == MVT::i32 && C != 0x80000000 &&
1413 isLegalArithImmed((uint32_t)(C - 1))) ||
1414 (VT == MVT::i64 && C != 0x80000000ULL &&
1415 isLegalArithImmed(C - 1ULL))) {
1416 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1417 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1418 RHS = DAG.getConstant(C, dl, VT);
1423 if ((VT == MVT::i32 && C != 0 &&
1424 isLegalArithImmed((uint32_t)(C - 1))) ||
1425 (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) {
1426 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1427 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1428 RHS = DAG.getConstant(C, dl, VT);
1433 if ((VT == MVT::i32 && C != INT32_MAX &&
1434 isLegalArithImmed((uint32_t)(C + 1))) ||
1435 (VT == MVT::i64 && C != INT64_MAX &&
1436 isLegalArithImmed(C + 1ULL))) {
1437 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1438 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1439 RHS = DAG.getConstant(C, dl, VT);
1444 if ((VT == MVT::i32 && C != UINT32_MAX &&
1445 isLegalArithImmed((uint32_t)(C + 1))) ||
1446 (VT == MVT::i64 && C != UINT64_MAX &&
1447 isLegalArithImmed(C + 1ULL))) {
1448 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1449 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1450 RHS = DAG.getConstant(C, dl, VT);
1457 AArch64CC::CondCode AArch64CC;
1458 if ((CC == ISD::SETEQ || CC == ISD::SETNE) && isa<ConstantSDNode>(RHS)) {
1459 const ConstantSDNode *RHSC = cast<ConstantSDNode>(RHS);
1461 // The imm operand of ADDS is an unsigned immediate, in the range 0 to 4095.
1462 // For the i8 operand, the largest immediate is 255, so this can be easily
1463 // encoded in the compare instruction. For the i16 operand, however, the
1464 // largest immediate cannot be encoded in the compare.
1465 // Therefore, use a sign extending load and cmn to avoid materializing the
1466 // -1 constant. For example,
1468 // ldrh w0, [x0, #0]
1471 // ldrsh w0, [x0, #0]
1473 // Fundamental, we're relying on the property that (zext LHS) == (zext RHS)
1474 // if and only if (sext LHS) == (sext RHS). The checks are in place to
1475 // ensure both the LHS and RHS are truly zero extended and to make sure the
1476 // transformation is profitable.
1477 if ((RHSC->getZExtValue() >> 16 == 0) && isa<LoadSDNode>(LHS) &&
1478 cast<LoadSDNode>(LHS)->getExtensionType() == ISD::ZEXTLOAD &&
1479 cast<LoadSDNode>(LHS)->getMemoryVT() == MVT::i16 &&
1480 LHS.getNode()->hasNUsesOfValue(1, 0)) {
1481 int16_t ValueofRHS = cast<ConstantSDNode>(RHS)->getZExtValue();
1482 if (ValueofRHS < 0 && isLegalArithImmed(-ValueofRHS)) {
1484 DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, LHS.getValueType(), LHS,
1485 DAG.getValueType(MVT::i16));
1486 Cmp = emitComparison(SExt, DAG.getConstant(ValueofRHS, dl,
1487 RHS.getValueType()),
1489 AArch64CC = changeIntCCToAArch64CC(CC);
1493 if (!Cmp && (RHSC->isNullValue() || RHSC->isOne())) {
1494 if ((Cmp = emitConjunctionDisjunctionTree(DAG, LHS, AArch64CC))) {
1495 if ((CC == ISD::SETNE) ^ RHSC->isNullValue())
1496 AArch64CC = AArch64CC::getInvertedCondCode(AArch64CC);
1502 Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
1503 AArch64CC = changeIntCCToAArch64CC(CC);
1505 AArch64cc = DAG.getConstant(AArch64CC, dl, MVT_CC);
1509 static std::pair<SDValue, SDValue>
1510 getAArch64XALUOOp(AArch64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
1511 assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
1512 "Unsupported value type");
1513 SDValue Value, Overflow;
1515 SDValue LHS = Op.getOperand(0);
1516 SDValue RHS = Op.getOperand(1);
1518 switch (Op.getOpcode()) {
1520 llvm_unreachable("Unknown overflow instruction!");
1522 Opc = AArch64ISD::ADDS;
1526 Opc = AArch64ISD::ADDS;
1530 Opc = AArch64ISD::SUBS;
1534 Opc = AArch64ISD::SUBS;
1537 // Multiply needs a little bit extra work.
1541 bool IsSigned = Op.getOpcode() == ISD::SMULO;
1542 if (Op.getValueType() == MVT::i32) {
1543 unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1544 // For a 32 bit multiply with overflow check we want the instruction
1545 // selector to generate a widening multiply (SMADDL/UMADDL). For that we
1546 // need to generate the following pattern:
1547 // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b))
1548 LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
1549 RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
1550 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1551 SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul,
1552 DAG.getConstant(0, DL, MVT::i64));
1553 // On AArch64 the upper 32 bits are always zero extended for a 32 bit
1554 // operation. We need to clear out the upper 32 bits, because we used a
1555 // widening multiply that wrote all 64 bits. In the end this should be a
1557 Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add);
1559 // The signed overflow check requires more than just a simple check for
1560 // any bit set in the upper 32 bits of the result. These bits could be
1561 // just the sign bits of a negative number. To perform the overflow
1562 // check we have to arithmetic shift right the 32nd bit of the result by
1563 // 31 bits. Then we compare the result to the upper 32 bits.
1564 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add,
1565 DAG.getConstant(32, DL, MVT::i64));
1566 UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits);
1567 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value,
1568 DAG.getConstant(31, DL, MVT::i64));
1569 // It is important that LowerBits is last, otherwise the arithmetic
1570 // shift will not be folded into the compare (SUBS).
1571 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32);
1572 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1575 // The overflow check for unsigned multiply is easy. We only need to
1576 // check if any of the upper 32 bits are set. This can be done with a
1577 // CMP (shifted register). For that we need to generate the following
1579 // (i64 AArch64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32)
1580 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
1581 DAG.getConstant(32, DL, MVT::i64));
1582 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1584 DAG.getNode(AArch64ISD::SUBS, DL, VTs,
1585 DAG.getConstant(0, DL, MVT::i64),
1586 UpperBits).getValue(1);
1590 assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
1591 // For the 64 bit multiply
1592 Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1594 SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
1595 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
1596 DAG.getConstant(63, DL, MVT::i64));
1597 // It is important that LowerBits is last, otherwise the arithmetic
1598 // shift will not be folded into the compare (SUBS).
1599 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1600 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1603 SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
1604 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1606 DAG.getNode(AArch64ISD::SUBS, DL, VTs,
1607 DAG.getConstant(0, DL, MVT::i64),
1608 UpperBits).getValue(1);
1615 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);
1617 // Emit the AArch64 operation with overflow check.
1618 Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
1619 Overflow = Value.getValue(1);
1621 return std::make_pair(Value, Overflow);
1624 SDValue AArch64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG,
1625 RTLIB::Libcall Call) const {
1626 SmallVector<SDValue, 2> Ops(Op->op_begin(), Op->op_end());
1627 return makeLibCall(DAG, Call, MVT::f128, &Ops[0], Ops.size(), false,
1631 static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) {
1632 SDValue Sel = Op.getOperand(0);
1633 SDValue Other = Op.getOperand(1);
1635 // If neither operand is a SELECT_CC, give up.
1636 if (Sel.getOpcode() != ISD::SELECT_CC)
1637 std::swap(Sel, Other);
1638 if (Sel.getOpcode() != ISD::SELECT_CC)
1641 // The folding we want to perform is:
1642 // (xor x, (select_cc a, b, cc, 0, -1) )
1644 // (csel x, (xor x, -1), cc ...)
1646 // The latter will get matched to a CSINV instruction.
1648 ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
1649 SDValue LHS = Sel.getOperand(0);
1650 SDValue RHS = Sel.getOperand(1);
1651 SDValue TVal = Sel.getOperand(2);
1652 SDValue FVal = Sel.getOperand(3);
1655 // FIXME: This could be generalized to non-integer comparisons.
1656 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
1659 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
1660 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
1662 // The values aren't constants, this isn't the pattern we're looking for.
1663 if (!CFVal || !CTVal)
1666 // We can commute the SELECT_CC by inverting the condition. This
1667 // might be needed to make this fit into a CSINV pattern.
1668 if (CTVal->isAllOnesValue() && CFVal->isNullValue()) {
1669 std::swap(TVal, FVal);
1670 std::swap(CTVal, CFVal);
1671 CC = ISD::getSetCCInverse(CC, true);
1674 // If the constants line up, perform the transform!
1675 if (CTVal->isNullValue() && CFVal->isAllOnesValue()) {
1677 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
1680 TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other,
1681 DAG.getConstant(-1ULL, dl, Other.getValueType()));
1683 return DAG.getNode(AArch64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal,
1690 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
1691 EVT VT = Op.getValueType();
1693 // Let legalize expand this if it isn't a legal type yet.
1694 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
1697 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
1700 bool ExtraOp = false;
1701 switch (Op.getOpcode()) {
1703 llvm_unreachable("Invalid code");
1705 Opc = AArch64ISD::ADDS;
1708 Opc = AArch64ISD::SUBS;
1711 Opc = AArch64ISD::ADCS;
1715 Opc = AArch64ISD::SBCS;
1721 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1));
1722 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1),
1726 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
1727 // Let legalize expand this if it isn't a legal type yet.
1728 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
1732 AArch64CC::CondCode CC;
1733 // The actual operation that sets the overflow or carry flag.
1734 SDValue Value, Overflow;
1735 std::tie(Value, Overflow) = getAArch64XALUOOp(CC, Op, DAG);
1737 // We use 0 and 1 as false and true values.
1738 SDValue TVal = DAG.getConstant(1, dl, MVT::i32);
1739 SDValue FVal = DAG.getConstant(0, dl, MVT::i32);
1741 // We use an inverted condition, because the conditional select is inverted
1742 // too. This will allow it to be selected to a single instruction:
1743 // CSINC Wd, WZR, WZR, invert(cond).
1744 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), dl, MVT::i32);
1745 Overflow = DAG.getNode(AArch64ISD::CSEL, dl, MVT::i32, FVal, TVal,
1748 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
1749 return DAG.getNode(ISD::MERGE_VALUES, dl, VTs, Value, Overflow);
1752 // Prefetch operands are:
1753 // 1: Address to prefetch
1755 // 3: int locality (0 = no locality ... 3 = extreme locality)
1756 // 4: bool isDataCache
1757 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
1759 unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
1760 unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
1761 unsigned IsData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
1763 bool IsStream = !Locality;
1764 // When the locality number is set
1766 // The front-end should have filtered out the out-of-range values
1767 assert(Locality <= 3 && "Prefetch locality out-of-range");
1768 // The locality degree is the opposite of the cache speed.
1769 // Put the number the other way around.
1770 // The encoding starts at 0 for level 1
1771 Locality = 3 - Locality;
1774 // built the mask value encoding the expected behavior.
1775 unsigned PrfOp = (IsWrite << 4) | // Load/Store bit
1776 (!IsData << 3) | // IsDataCache bit
1777 (Locality << 1) | // Cache level bits
1778 (unsigned)IsStream; // Stream bit
1779 return DAG.getNode(AArch64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
1780 DAG.getConstant(PrfOp, DL, MVT::i32), Op.getOperand(1));
1783 SDValue AArch64TargetLowering::LowerFP_EXTEND(SDValue Op,
1784 SelectionDAG &DAG) const {
1785 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
1788 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
1790 return LowerF128Call(Op, DAG, LC);
1793 SDValue AArch64TargetLowering::LowerFP_ROUND(SDValue Op,
1794 SelectionDAG &DAG) const {
1795 if (Op.getOperand(0).getValueType() != MVT::f128) {
1796 // It's legal except when f128 is involved
1801 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
1803 // FP_ROUND node has a second operand indicating whether it is known to be
1804 // precise. That doesn't take part in the LibCall so we can't directly use
1806 SDValue SrcVal = Op.getOperand(0);
1807 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
1808 /*isSigned*/ false, SDLoc(Op)).first;
1811 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1812 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1813 // Any additional optimization in this function should be recorded
1814 // in the cost tables.
1815 EVT InVT = Op.getOperand(0).getValueType();
1816 EVT VT = Op.getValueType();
1818 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1821 DAG.getNode(Op.getOpcode(), dl, InVT.changeVectorElementTypeToInteger(),
1823 return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
1826 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1829 MVT::getVectorVT(MVT::getFloatingPointVT(VT.getScalarSizeInBits()),
1830 VT.getVectorNumElements());
1831 SDValue Ext = DAG.getNode(ISD::FP_EXTEND, dl, ExtVT, Op.getOperand(0));
1832 return DAG.getNode(Op.getOpcode(), dl, VT, Ext);
1835 // Type changing conversions are illegal.
1839 SDValue AArch64TargetLowering::LowerFP_TO_INT(SDValue Op,
1840 SelectionDAG &DAG) const {
1841 if (Op.getOperand(0).getValueType().isVector())
1842 return LowerVectorFP_TO_INT(Op, DAG);
1844 // f16 conversions are promoted to f32.
1845 if (Op.getOperand(0).getValueType() == MVT::f16) {
1848 Op.getOpcode(), dl, Op.getValueType(),
1849 DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, Op.getOperand(0)));
1852 if (Op.getOperand(0).getValueType() != MVT::f128) {
1853 // It's legal except when f128 is involved
1858 if (Op.getOpcode() == ISD::FP_TO_SINT)
1859 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
1861 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
1863 SmallVector<SDValue, 2> Ops(Op->op_begin(), Op->op_end());
1864 return makeLibCall(DAG, LC, Op.getValueType(), &Ops[0], Ops.size(), false,
1868 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1869 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1870 // Any additional optimization in this function should be recorded
1871 // in the cost tables.
1872 EVT VT = Op.getValueType();
1874 SDValue In = Op.getOperand(0);
1875 EVT InVT = In.getValueType();
1877 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1879 MVT::getVectorVT(MVT::getFloatingPointVT(InVT.getScalarSizeInBits()),
1880 InVT.getVectorNumElements());
1881 In = DAG.getNode(Op.getOpcode(), dl, CastVT, In);
1882 return DAG.getNode(ISD::FP_ROUND, dl, VT, In, DAG.getIntPtrConstant(0, dl));
1885 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1887 Op.getOpcode() == ISD::SINT_TO_FP ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1888 EVT CastVT = VT.changeVectorElementTypeToInteger();
1889 In = DAG.getNode(CastOpc, dl, CastVT, In);
1890 return DAG.getNode(Op.getOpcode(), dl, VT, In);
1896 SDValue AArch64TargetLowering::LowerINT_TO_FP(SDValue Op,
1897 SelectionDAG &DAG) const {
1898 if (Op.getValueType().isVector())
1899 return LowerVectorINT_TO_FP(Op, DAG);
1901 // f16 conversions are promoted to f32.
1902 if (Op.getValueType() == MVT::f16) {
1905 ISD::FP_ROUND, dl, MVT::f16,
1906 DAG.getNode(Op.getOpcode(), dl, MVT::f32, Op.getOperand(0)),
1907 DAG.getIntPtrConstant(0, dl));
1910 // i128 conversions are libcalls.
1911 if (Op.getOperand(0).getValueType() == MVT::i128)
1914 // Other conversions are legal, unless it's to the completely software-based
1916 if (Op.getValueType() != MVT::f128)
1920 if (Op.getOpcode() == ISD::SINT_TO_FP)
1921 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1923 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1925 return LowerF128Call(Op, DAG, LC);
1928 SDValue AArch64TargetLowering::LowerFSINCOS(SDValue Op,
1929 SelectionDAG &DAG) const {
1930 // For iOS, we want to call an alternative entry point: __sincos_stret,
1931 // which returns the values in two S / D registers.
1933 SDValue Arg = Op.getOperand(0);
1934 EVT ArgVT = Arg.getValueType();
1935 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1942 Entry.isSExt = false;
1943 Entry.isZExt = false;
1944 Args.push_back(Entry);
1946 const char *LibcallName =
1947 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
1949 DAG.getExternalSymbol(LibcallName, getPointerTy(DAG.getDataLayout()));
1951 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
1952 TargetLowering::CallLoweringInfo CLI(DAG);
1953 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
1954 .setCallee(CallingConv::Fast, RetTy, Callee, std::move(Args), 0);
1956 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1957 return CallResult.first;
1960 static SDValue LowerBITCAST(SDValue Op, SelectionDAG &DAG) {
1961 if (Op.getValueType() != MVT::f16)
1964 assert(Op.getOperand(0).getValueType() == MVT::i16);
1967 Op = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op.getOperand(0));
1968 Op = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Op);
1970 DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL, MVT::f16, Op,
1971 DAG.getTargetConstant(AArch64::hsub, DL, MVT::i32)),
1975 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
1976 if (OrigVT.getSizeInBits() >= 64)
1979 assert(OrigVT.isSimple() && "Expecting a simple value type");
1981 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
1982 switch (OrigSimpleTy) {
1983 default: llvm_unreachable("Unexpected Vector Type");
1992 static SDValue addRequiredExtensionForVectorMULL(SDValue N, SelectionDAG &DAG,
1995 unsigned ExtOpcode) {
1996 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
1997 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
1998 // 64-bits we need to insert a new extension so that it will be 64-bits.
1999 assert(ExtTy.is128BitVector() && "Unexpected extension size");
2000 if (OrigTy.getSizeInBits() >= 64)
2003 // Must extend size to at least 64 bits to be used as an operand for VMULL.
2004 EVT NewVT = getExtensionTo64Bits(OrigTy);
2006 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
2009 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
2011 EVT VT = N->getValueType(0);
2013 if (N->getOpcode() != ISD::BUILD_VECTOR)
2016 for (const SDValue &Elt : N->op_values()) {
2017 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
2018 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
2019 unsigned HalfSize = EltSize / 2;
2021 if (!isIntN(HalfSize, C->getSExtValue()))
2024 if (!isUIntN(HalfSize, C->getZExtValue()))
2035 static SDValue skipExtensionForVectorMULL(SDNode *N, SelectionDAG &DAG) {
2036 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
2037 return addRequiredExtensionForVectorMULL(N->getOperand(0), DAG,
2038 N->getOperand(0)->getValueType(0),
2042 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
2043 EVT VT = N->getValueType(0);
2045 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
2046 unsigned NumElts = VT.getVectorNumElements();
2047 MVT TruncVT = MVT::getIntegerVT(EltSize);
2048 SmallVector<SDValue, 8> Ops;
2049 for (unsigned i = 0; i != NumElts; ++i) {
2050 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
2051 const APInt &CInt = C->getAPIntValue();
2052 // Element types smaller than 32 bits are not legal, so use i32 elements.
2053 // The values are implicitly truncated so sext vs. zext doesn't matter.
2054 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), dl, MVT::i32));
2056 return DAG.getNode(ISD::BUILD_VECTOR, dl,
2057 MVT::getVectorVT(TruncVT, NumElts), Ops);
2060 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
2061 if (N->getOpcode() == ISD::SIGN_EXTEND)
2063 if (isExtendedBUILD_VECTOR(N, DAG, true))
2068 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
2069 if (N->getOpcode() == ISD::ZERO_EXTEND)
2071 if (isExtendedBUILD_VECTOR(N, DAG, false))
2076 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
2077 unsigned Opcode = N->getOpcode();
2078 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
2079 SDNode *N0 = N->getOperand(0).getNode();
2080 SDNode *N1 = N->getOperand(1).getNode();
2081 return N0->hasOneUse() && N1->hasOneUse() &&
2082 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
2087 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
2088 unsigned Opcode = N->getOpcode();
2089 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
2090 SDNode *N0 = N->getOperand(0).getNode();
2091 SDNode *N1 = N->getOperand(1).getNode();
2092 return N0->hasOneUse() && N1->hasOneUse() &&
2093 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
2098 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
2099 // Multiplications are only custom-lowered for 128-bit vectors so that
2100 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
2101 EVT VT = Op.getValueType();
2102 assert(VT.is128BitVector() && VT.isInteger() &&
2103 "unexpected type for custom-lowering ISD::MUL");
2104 SDNode *N0 = Op.getOperand(0).getNode();
2105 SDNode *N1 = Op.getOperand(1).getNode();
2106 unsigned NewOpc = 0;
2108 bool isN0SExt = isSignExtended(N0, DAG);
2109 bool isN1SExt = isSignExtended(N1, DAG);
2110 if (isN0SExt && isN1SExt)
2111 NewOpc = AArch64ISD::SMULL;
2113 bool isN0ZExt = isZeroExtended(N0, DAG);
2114 bool isN1ZExt = isZeroExtended(N1, DAG);
2115 if (isN0ZExt && isN1ZExt)
2116 NewOpc = AArch64ISD::UMULL;
2117 else if (isN1SExt || isN1ZExt) {
2118 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
2119 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
2120 if (isN1SExt && isAddSubSExt(N0, DAG)) {
2121 NewOpc = AArch64ISD::SMULL;
2123 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
2124 NewOpc = AArch64ISD::UMULL;
2126 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
2128 NewOpc = AArch64ISD::UMULL;
2134 if (VT == MVT::v2i64)
2135 // Fall through to expand this. It is not legal.
2138 // Other vector multiplications are legal.
2143 // Legalize to a S/UMULL instruction
2146 SDValue Op1 = skipExtensionForVectorMULL(N1, DAG);
2148 Op0 = skipExtensionForVectorMULL(N0, DAG);
2149 assert(Op0.getValueType().is64BitVector() &&
2150 Op1.getValueType().is64BitVector() &&
2151 "unexpected types for extended operands to VMULL");
2152 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
2154 // Optimizing (zext A + zext B) * C, to (S/UMULL A, C) + (S/UMULL B, C) during
2155 // isel lowering to take advantage of no-stall back to back s/umul + s/umla.
2156 // This is true for CPUs with accumulate forwarding such as Cortex-A53/A57
2157 SDValue N00 = skipExtensionForVectorMULL(N0->getOperand(0).getNode(), DAG);
2158 SDValue N01 = skipExtensionForVectorMULL(N0->getOperand(1).getNode(), DAG);
2159 EVT Op1VT = Op1.getValueType();
2160 return DAG.getNode(N0->getOpcode(), DL, VT,
2161 DAG.getNode(NewOpc, DL, VT,
2162 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
2163 DAG.getNode(NewOpc, DL, VT,
2164 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
2167 SDValue AArch64TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
2168 SelectionDAG &DAG) const {
2169 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2172 default: return SDValue(); // Don't custom lower most intrinsics.
2173 case Intrinsic::aarch64_thread_pointer: {
2174 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2175 return DAG.getNode(AArch64ISD::THREAD_POINTER, dl, PtrVT);
2180 SDValue AArch64TargetLowering::LowerOperation(SDValue Op,
2181 SelectionDAG &DAG) const {
2182 switch (Op.getOpcode()) {
2184 llvm_unreachable("unimplemented operand");
2187 return LowerBITCAST(Op, DAG);
2188 case ISD::GlobalAddress:
2189 return LowerGlobalAddress(Op, DAG);
2190 case ISD::GlobalTLSAddress:
2191 return LowerGlobalTLSAddress(Op, DAG);
2193 return LowerSETCC(Op, DAG);
2195 return LowerBR_CC(Op, DAG);
2197 return LowerSELECT(Op, DAG);
2198 case ISD::SELECT_CC:
2199 return LowerSELECT_CC(Op, DAG);
2200 case ISD::JumpTable:
2201 return LowerJumpTable(Op, DAG);
2202 case ISD::ConstantPool:
2203 return LowerConstantPool(Op, DAG);
2204 case ISD::BlockAddress:
2205 return LowerBlockAddress(Op, DAG);
2207 return LowerVASTART(Op, DAG);
2209 return LowerVACOPY(Op, DAG);
2211 return LowerVAARG(Op, DAG);
2216 return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
2223 return LowerXALUO(Op, DAG);
2225 return LowerF128Call(Op, DAG, RTLIB::ADD_F128);
2227 return LowerF128Call(Op, DAG, RTLIB::SUB_F128);
2229 return LowerF128Call(Op, DAG, RTLIB::MUL_F128);
2231 return LowerF128Call(Op, DAG, RTLIB::DIV_F128);
2233 return LowerFP_ROUND(Op, DAG);
2234 case ISD::FP_EXTEND:
2235 return LowerFP_EXTEND(Op, DAG);
2236 case ISD::FRAMEADDR:
2237 return LowerFRAMEADDR(Op, DAG);
2238 case ISD::RETURNADDR:
2239 return LowerRETURNADDR(Op, DAG);
2240 case ISD::INSERT_VECTOR_ELT:
2241 return LowerINSERT_VECTOR_ELT(Op, DAG);
2242 case ISD::EXTRACT_VECTOR_ELT:
2243 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2244 case ISD::BUILD_VECTOR:
2245 return LowerBUILD_VECTOR(Op, DAG);
2246 case ISD::VECTOR_SHUFFLE:
2247 return LowerVECTOR_SHUFFLE(Op, DAG);
2248 case ISD::EXTRACT_SUBVECTOR:
2249 return LowerEXTRACT_SUBVECTOR(Op, DAG);
2253 return LowerVectorSRA_SRL_SHL(Op, DAG);
2254 case ISD::SHL_PARTS:
2255 return LowerShiftLeftParts(Op, DAG);
2256 case ISD::SRL_PARTS:
2257 case ISD::SRA_PARTS:
2258 return LowerShiftRightParts(Op, DAG);
2260 return LowerCTPOP(Op, DAG);
2261 case ISD::FCOPYSIGN:
2262 return LowerFCOPYSIGN(Op, DAG);
2264 return LowerVectorAND(Op, DAG);
2266 return LowerVectorOR(Op, DAG);
2268 return LowerXOR(Op, DAG);
2270 return LowerPREFETCH(Op, DAG);
2271 case ISD::SINT_TO_FP:
2272 case ISD::UINT_TO_FP:
2273 return LowerINT_TO_FP(Op, DAG);
2274 case ISD::FP_TO_SINT:
2275 case ISD::FP_TO_UINT:
2276 return LowerFP_TO_INT(Op, DAG);
2278 return LowerFSINCOS(Op, DAG);
2280 return LowerMUL(Op, DAG);
2281 case ISD::INTRINSIC_WO_CHAIN:
2282 return LowerINTRINSIC_WO_CHAIN(Op, DAG);
2286 /// getFunctionAlignment - Return the Log2 alignment of this function.
2287 unsigned AArch64TargetLowering::getFunctionAlignment(const Function *F) const {
2291 //===----------------------------------------------------------------------===//
2292 // Calling Convention Implementation
2293 //===----------------------------------------------------------------------===//
2295 #include "AArch64GenCallingConv.inc"
2297 /// Selects the correct CCAssignFn for a given CallingConvention value.
2298 CCAssignFn *AArch64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
2299 bool IsVarArg) const {
2302 llvm_unreachable("Unsupported calling convention.");
2303 case CallingConv::WebKit_JS:
2304 return CC_AArch64_WebKit_JS;
2305 case CallingConv::GHC:
2306 return CC_AArch64_GHC;
2307 case CallingConv::C:
2308 case CallingConv::Fast:
2309 if (!Subtarget->isTargetDarwin())
2310 return CC_AArch64_AAPCS;
2311 return IsVarArg ? CC_AArch64_DarwinPCS_VarArg : CC_AArch64_DarwinPCS;
2315 SDValue AArch64TargetLowering::LowerFormalArguments(
2316 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2317 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2318 SmallVectorImpl<SDValue> &InVals) const {
2319 MachineFunction &MF = DAG.getMachineFunction();
2320 MachineFrameInfo *MFI = MF.getFrameInfo();
2322 // Assign locations to all of the incoming arguments.
2323 SmallVector<CCValAssign, 16> ArgLocs;
2324 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
2327 // At this point, Ins[].VT may already be promoted to i32. To correctly
2328 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2329 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
2330 // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
2331 // we use a special version of AnalyzeFormalArguments to pass in ValVT and
2333 unsigned NumArgs = Ins.size();
2334 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2335 unsigned CurArgIdx = 0;
2336 for (unsigned i = 0; i != NumArgs; ++i) {
2337 MVT ValVT = Ins[i].VT;
2338 if (Ins[i].isOrigArg()) {
2339 std::advance(CurOrigArg, Ins[i].getOrigArgIndex() - CurArgIdx);
2340 CurArgIdx = Ins[i].getOrigArgIndex();
2342 // Get type of the original argument.
2343 EVT ActualVT = getValueType(DAG.getDataLayout(), CurOrigArg->getType(),
2344 /*AllowUnknown*/ true);
2345 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
2346 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2347 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2349 else if (ActualMVT == MVT::i16)
2352 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2354 AssignFn(i, ValVT, ValVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
2355 assert(!Res && "Call operand has unhandled type");
2358 assert(ArgLocs.size() == Ins.size());
2359 SmallVector<SDValue, 16> ArgValues;
2360 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2361 CCValAssign &VA = ArgLocs[i];
2363 if (Ins[i].Flags.isByVal()) {
2364 // Byval is used for HFAs in the PCS, but the system should work in a
2365 // non-compliant manner for larger structs.
2366 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2367 int Size = Ins[i].Flags.getByValSize();
2368 unsigned NumRegs = (Size + 7) / 8;
2370 // FIXME: This works on big-endian for composite byvals, which are the common
2371 // case. It should also work for fundamental types too.
2373 MFI->CreateFixedObject(8 * NumRegs, VA.getLocMemOffset(), false);
2374 SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrVT);
2375 InVals.push_back(FrameIdxN);
2380 if (VA.isRegLoc()) {
2381 // Arguments stored in registers.
2382 EVT RegVT = VA.getLocVT();
2385 const TargetRegisterClass *RC;
2387 if (RegVT == MVT::i32)
2388 RC = &AArch64::GPR32RegClass;
2389 else if (RegVT == MVT::i64)
2390 RC = &AArch64::GPR64RegClass;
2391 else if (RegVT == MVT::f16)
2392 RC = &AArch64::FPR16RegClass;
2393 else if (RegVT == MVT::f32)
2394 RC = &AArch64::FPR32RegClass;
2395 else if (RegVT == MVT::f64 || RegVT.is64BitVector())
2396 RC = &AArch64::FPR64RegClass;
2397 else if (RegVT == MVT::f128 || RegVT.is128BitVector())
2398 RC = &AArch64::FPR128RegClass;
2400 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2402 // Transform the arguments in physical registers into virtual ones.
2403 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2404 ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2406 // If this is an 8, 16 or 32-bit value, it is really passed promoted
2407 // to 64 bits. Insert an assert[sz]ext to capture this, then
2408 // truncate to the right size.
2409 switch (VA.getLocInfo()) {
2411 llvm_unreachable("Unknown loc info!");
2412 case CCValAssign::Full:
2414 case CCValAssign::BCvt:
2415 ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue);
2417 case CCValAssign::AExt:
2418 case CCValAssign::SExt:
2419 case CCValAssign::ZExt:
2420 // SelectionDAGBuilder will insert appropriate AssertZExt & AssertSExt
2421 // nodes after our lowering.
2422 assert(RegVT == Ins[i].VT && "incorrect register location selected");
2426 InVals.push_back(ArgValue);
2428 } else { // VA.isRegLoc()
2429 assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
2430 unsigned ArgOffset = VA.getLocMemOffset();
2431 unsigned ArgSize = VA.getValVT().getSizeInBits() / 8;
2433 uint32_t BEAlign = 0;
2434 if (!Subtarget->isLittleEndian() && ArgSize < 8 &&
2435 !Ins[i].Flags.isInConsecutiveRegs())
2436 BEAlign = 8 - ArgSize;
2438 int FI = MFI->CreateFixedObject(ArgSize, ArgOffset + BEAlign, true);
2440 // Create load nodes to retrieve arguments from the stack.
2441 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2444 // For NON_EXTLOAD, generic code in getLoad assert(ValVT == MemVT)
2445 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
2446 MVT MemVT = VA.getValVT();
2448 switch (VA.getLocInfo()) {
2451 case CCValAssign::BCvt:
2452 MemVT = VA.getLocVT();
2454 case CCValAssign::SExt:
2455 ExtType = ISD::SEXTLOAD;
2457 case CCValAssign::ZExt:
2458 ExtType = ISD::ZEXTLOAD;
2460 case CCValAssign::AExt:
2461 ExtType = ISD::EXTLOAD;
2465 ArgValue = DAG.getExtLoad(
2466 ExtType, DL, VA.getLocVT(), Chain, FIN,
2467 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
2468 MemVT, false, false, false, 0);
2470 InVals.push_back(ArgValue);
2476 if (!Subtarget->isTargetDarwin()) {
2477 // The AAPCS variadic function ABI is identical to the non-variadic
2478 // one. As a result there may be more arguments in registers and we should
2479 // save them for future reference.
2480 saveVarArgRegisters(CCInfo, DAG, DL, Chain);
2483 AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>();
2484 // This will point to the next argument passed via stack.
2485 unsigned StackOffset = CCInfo.getNextStackOffset();
2486 // We currently pass all varargs at 8-byte alignment.
2487 StackOffset = ((StackOffset + 7) & ~7);
2488 AFI->setVarArgsStackIndex(MFI->CreateFixedObject(4, StackOffset, true));
2491 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2492 unsigned StackArgSize = CCInfo.getNextStackOffset();
2493 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2494 if (DoesCalleeRestoreStack(CallConv, TailCallOpt)) {
2495 // This is a non-standard ABI so by fiat I say we're allowed to make full
2496 // use of the stack area to be popped, which must be aligned to 16 bytes in
2498 StackArgSize = RoundUpToAlignment(StackArgSize, 16);
2500 // If we're expected to restore the stack (e.g. fastcc) then we'll be adding
2501 // a multiple of 16.
2502 FuncInfo->setArgumentStackToRestore(StackArgSize);
2504 // This realignment carries over to the available bytes below. Our own
2505 // callers will guarantee the space is free by giving an aligned value to
2508 // Even if we're not expected to free up the space, it's useful to know how
2509 // much is there while considering tail calls (because we can reuse it).
2510 FuncInfo->setBytesInStackArgArea(StackArgSize);
2515 void AArch64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
2516 SelectionDAG &DAG, SDLoc DL,
2517 SDValue &Chain) const {
2518 MachineFunction &MF = DAG.getMachineFunction();
2519 MachineFrameInfo *MFI = MF.getFrameInfo();
2520 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2521 auto PtrVT = getPointerTy(DAG.getDataLayout());
2523 SmallVector<SDValue, 8> MemOps;
2525 static const MCPhysReg GPRArgRegs[] = { AArch64::X0, AArch64::X1, AArch64::X2,
2526 AArch64::X3, AArch64::X4, AArch64::X5,
2527 AArch64::X6, AArch64::X7 };
2528 static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs);
2529 unsigned FirstVariadicGPR = CCInfo.getFirstUnallocated(GPRArgRegs);
2531 unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
2533 if (GPRSaveSize != 0) {
2534 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
2536 SDValue FIN = DAG.getFrameIndex(GPRIdx, PtrVT);
2538 for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
2539 unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &AArch64::GPR64RegClass);
2540 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
2541 SDValue Store = DAG.getStore(
2542 Val.getValue(1), DL, Val, FIN,
2543 MachinePointerInfo::getStack(DAG.getMachineFunction(), i * 8), false,
2545 MemOps.push_back(Store);
2547 DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getConstant(8, DL, PtrVT));
2550 FuncInfo->setVarArgsGPRIndex(GPRIdx);
2551 FuncInfo->setVarArgsGPRSize(GPRSaveSize);
2553 if (Subtarget->hasFPARMv8()) {
2554 static const MCPhysReg FPRArgRegs[] = {
2555 AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3,
2556 AArch64::Q4, AArch64::Q5, AArch64::Q6, AArch64::Q7};
2557 static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs);
2558 unsigned FirstVariadicFPR = CCInfo.getFirstUnallocated(FPRArgRegs);
2560 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
2562 if (FPRSaveSize != 0) {
2563 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
2565 SDValue FIN = DAG.getFrameIndex(FPRIdx, PtrVT);
2567 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
2568 unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &AArch64::FPR128RegClass);
2569 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f128);
2571 SDValue Store = DAG.getStore(
2572 Val.getValue(1), DL, Val, FIN,
2573 MachinePointerInfo::getStack(DAG.getMachineFunction(), i * 16),
2575 MemOps.push_back(Store);
2576 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN,
2577 DAG.getConstant(16, DL, PtrVT));
2580 FuncInfo->setVarArgsFPRIndex(FPRIdx);
2581 FuncInfo->setVarArgsFPRSize(FPRSaveSize);
2584 if (!MemOps.empty()) {
2585 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
2589 /// LowerCallResult - Lower the result values of a call into the
2590 /// appropriate copies out of appropriate physical registers.
2591 SDValue AArch64TargetLowering::LowerCallResult(
2592 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
2593 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2594 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
2595 SDValue ThisVal) const {
2596 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2597 ? RetCC_AArch64_WebKit_JS
2598 : RetCC_AArch64_AAPCS;
2599 // Assign locations to each value returned by this call.
2600 SmallVector<CCValAssign, 16> RVLocs;
2601 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2603 CCInfo.AnalyzeCallResult(Ins, RetCC);
2605 // Copy all of the result registers out of their specified physreg.
2606 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2607 CCValAssign VA = RVLocs[i];
2609 // Pass 'this' value directly from the argument to return value, to avoid
2610 // reg unit interference
2611 if (i == 0 && isThisReturn) {
2612 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 &&
2613 "unexpected return calling convention register assignment");
2614 InVals.push_back(ThisVal);
2619 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag);
2620 Chain = Val.getValue(1);
2621 InFlag = Val.getValue(2);
2623 switch (VA.getLocInfo()) {
2625 llvm_unreachable("Unknown loc info!");
2626 case CCValAssign::Full:
2628 case CCValAssign::BCvt:
2629 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
2633 InVals.push_back(Val);
2639 bool AArch64TargetLowering::isEligibleForTailCallOptimization(
2640 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
2641 bool isCalleeStructRet, bool isCallerStructRet,
2642 const SmallVectorImpl<ISD::OutputArg> &Outs,
2643 const SmallVectorImpl<SDValue> &OutVals,
2644 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
2645 // For CallingConv::C this function knows whether the ABI needs
2646 // changing. That's not true for other conventions so they will have to opt in
2648 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
2651 const MachineFunction &MF = DAG.getMachineFunction();
2652 const Function *CallerF = MF.getFunction();
2653 CallingConv::ID CallerCC = CallerF->getCallingConv();
2654 bool CCMatch = CallerCC == CalleeCC;
2656 // Byval parameters hand the function a pointer directly into the stack area
2657 // we want to reuse during a tail call. Working around this *is* possible (see
2658 // X86) but less efficient and uglier in LowerCall.
2659 for (Function::const_arg_iterator i = CallerF->arg_begin(),
2660 e = CallerF->arg_end();
2662 if (i->hasByValAttr())
2665 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
2666 if (IsTailCallConvention(CalleeCC) && CCMatch)
2671 // Externally-defined functions with weak linkage should not be
2672 // tail-called on AArch64 when the OS does not support dynamic
2673 // pre-emption of symbols, as the AAELF spec requires normal calls
2674 // to undefined weak functions to be replaced with a NOP or jump to the
2675 // next instruction. The behaviour of branch instructions in this
2676 // situation (as used for tail calls) is implementation-defined, so we
2677 // cannot rely on the linker replacing the tail call with a return.
2678 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2679 const GlobalValue *GV = G->getGlobal();
2680 const Triple &TT = getTargetMachine().getTargetTriple();
2681 if (GV->hasExternalWeakLinkage() &&
2682 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2686 // Now we search for cases where we can use a tail call without changing the
2687 // ABI. Sibcall is used in some places (particularly gcc) to refer to this
2690 // I want anyone implementing a new calling convention to think long and hard
2691 // about this assert.
2692 assert((!isVarArg || CalleeCC == CallingConv::C) &&
2693 "Unexpected variadic calling convention");
2695 if (isVarArg && !Outs.empty()) {
2696 // At least two cases here: if caller is fastcc then we can't have any
2697 // memory arguments (we'd be expected to clean up the stack afterwards). If
2698 // caller is C then we could potentially use its argument area.
2700 // FIXME: for now we take the most conservative of these in both cases:
2701 // disallow all variadic memory operands.
2702 SmallVector<CCValAssign, 16> ArgLocs;
2703 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2706 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, true));
2707 for (const CCValAssign &ArgLoc : ArgLocs)
2708 if (!ArgLoc.isRegLoc())
2712 // If the calling conventions do not match, then we'd better make sure the
2713 // results are returned in the same way as what the caller expects.
2715 SmallVector<CCValAssign, 16> RVLocs1;
2716 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2718 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForCall(CalleeCC, isVarArg));
2720 SmallVector<CCValAssign, 16> RVLocs2;
2721 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2723 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForCall(CallerCC, isVarArg));
2725 if (RVLocs1.size() != RVLocs2.size())
2727 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2728 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2730 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2732 if (RVLocs1[i].isRegLoc()) {
2733 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2736 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2742 // Nothing more to check if the callee is taking no arguments
2746 SmallVector<CCValAssign, 16> ArgLocs;
2747 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2750 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, isVarArg));
2752 const AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2754 // If the stack arguments for this call would fit into our own save area then
2755 // the call can be made tail.
2756 return CCInfo.getNextStackOffset() <= FuncInfo->getBytesInStackArgArea();
2759 SDValue AArch64TargetLowering::addTokenForArgument(SDValue Chain,
2761 MachineFrameInfo *MFI,
2762 int ClobberedFI) const {
2763 SmallVector<SDValue, 8> ArgChains;
2764 int64_t FirstByte = MFI->getObjectOffset(ClobberedFI);
2765 int64_t LastByte = FirstByte + MFI->getObjectSize(ClobberedFI) - 1;
2767 // Include the original chain at the beginning of the list. When this is
2768 // used by target LowerCall hooks, this helps legalize find the
2769 // CALLSEQ_BEGIN node.
2770 ArgChains.push_back(Chain);
2772 // Add a chain value for each stack argument corresponding
2773 for (SDNode::use_iterator U = DAG.getEntryNode().getNode()->use_begin(),
2774 UE = DAG.getEntryNode().getNode()->use_end();
2776 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
2777 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
2778 if (FI->getIndex() < 0) {
2779 int64_t InFirstByte = MFI->getObjectOffset(FI->getIndex());
2780 int64_t InLastByte = InFirstByte;
2781 InLastByte += MFI->getObjectSize(FI->getIndex()) - 1;
2783 if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
2784 (FirstByte <= InFirstByte && InFirstByte <= LastByte))
2785 ArgChains.push_back(SDValue(L, 1));
2788 // Build a tokenfactor for all the chains.
2789 return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
2792 bool AArch64TargetLowering::DoesCalleeRestoreStack(CallingConv::ID CallCC,
2793 bool TailCallOpt) const {
2794 return CallCC == CallingConv::Fast && TailCallOpt;
2797 bool AArch64TargetLowering::IsTailCallConvention(CallingConv::ID CallCC) const {
2798 return CallCC == CallingConv::Fast;
2801 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
2802 /// and add input and output parameter nodes.
2804 AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
2805 SmallVectorImpl<SDValue> &InVals) const {
2806 SelectionDAG &DAG = CLI.DAG;
2808 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2809 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2810 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2811 SDValue Chain = CLI.Chain;
2812 SDValue Callee = CLI.Callee;
2813 bool &IsTailCall = CLI.IsTailCall;
2814 CallingConv::ID CallConv = CLI.CallConv;
2815 bool IsVarArg = CLI.IsVarArg;
2817 MachineFunction &MF = DAG.getMachineFunction();
2818 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
2819 bool IsThisReturn = false;
2821 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2822 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2823 bool IsSibCall = false;
2826 // Check if it's really possible to do a tail call.
2827 IsTailCall = isEligibleForTailCallOptimization(
2828 Callee, CallConv, IsVarArg, IsStructRet,
2829 MF.getFunction()->hasStructRetAttr(), Outs, OutVals, Ins, DAG);
2830 if (!IsTailCall && CLI.CS && CLI.CS->isMustTailCall())
2831 report_fatal_error("failed to perform tail call elimination on a call "
2832 "site marked musttail");
2834 // A sibling call is one where we're under the usual C ABI and not planning
2835 // to change that but can still do a tail call:
2836 if (!TailCallOpt && IsTailCall)
2843 // Analyze operands of the call, assigning locations to each operand.
2844 SmallVector<CCValAssign, 16> ArgLocs;
2845 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
2849 // Handle fixed and variable vector arguments differently.
2850 // Variable vector arguments always go into memory.
2851 unsigned NumArgs = Outs.size();
2853 for (unsigned i = 0; i != NumArgs; ++i) {
2854 MVT ArgVT = Outs[i].VT;
2855 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2856 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
2857 /*IsVarArg=*/ !Outs[i].IsFixed);
2858 bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
2859 assert(!Res && "Call operand has unhandled type");
2863 // At this point, Outs[].VT may already be promoted to i32. To correctly
2864 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2865 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
2866 // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here
2867 // we use a special version of AnalyzeCallOperands to pass in ValVT and
2869 unsigned NumArgs = Outs.size();
2870 for (unsigned i = 0; i != NumArgs; ++i) {
2871 MVT ValVT = Outs[i].VT;
2872 // Get type of the original argument.
2873 EVT ActualVT = getValueType(DAG.getDataLayout(),
2874 CLI.getArgs()[Outs[i].OrigArgIndex].Ty,
2875 /*AllowUnknown*/ true);
2876 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT;
2877 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2878 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2879 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2881 else if (ActualMVT == MVT::i16)
2884 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2885 bool Res = AssignFn(i, ValVT, ValVT, CCValAssign::Full, ArgFlags, CCInfo);
2886 assert(!Res && "Call operand has unhandled type");
2891 // Get a count of how many bytes are to be pushed on the stack.
2892 unsigned NumBytes = CCInfo.getNextStackOffset();
2895 // Since we're not changing the ABI to make this a tail call, the memory
2896 // operands are already available in the caller's incoming argument space.
2900 // FPDiff is the byte offset of the call's argument area from the callee's.
2901 // Stores to callee stack arguments will be placed in FixedStackSlots offset
2902 // by this amount for a tail call. In a sibling call it must be 0 because the
2903 // caller will deallocate the entire stack and the callee still expects its
2904 // arguments to begin at SP+0. Completely unused for non-tail calls.
2907 if (IsTailCall && !IsSibCall) {
2908 unsigned NumReusableBytes = FuncInfo->getBytesInStackArgArea();
2910 // Since callee will pop argument stack as a tail call, we must keep the
2911 // popped size 16-byte aligned.
2912 NumBytes = RoundUpToAlignment(NumBytes, 16);
2914 // FPDiff will be negative if this tail call requires more space than we
2915 // would automatically have in our incoming argument space. Positive if we
2916 // can actually shrink the stack.
2917 FPDiff = NumReusableBytes - NumBytes;
2919 // The stack pointer must be 16-byte aligned at all times it's used for a
2920 // memory operation, which in practice means at *all* times and in
2921 // particular across call boundaries. Therefore our own arguments started at
2922 // a 16-byte aligned SP and the delta applied for the tail call should
2923 // satisfy the same constraint.
2924 assert(FPDiff % 16 == 0 && "unaligned stack on tail call");
2927 // Adjust the stack pointer for the new arguments...
2928 // These operations are automatically eliminated by the prolog/epilog pass
2930 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, DL,
2934 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, AArch64::SP,
2935 getPointerTy(DAG.getDataLayout()));
2937 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2938 SmallVector<SDValue, 8> MemOpChains;
2939 auto PtrVT = getPointerTy(DAG.getDataLayout());
2941 // Walk the register/memloc assignments, inserting copies/loads.
2942 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2943 ++i, ++realArgIdx) {
2944 CCValAssign &VA = ArgLocs[i];
2945 SDValue Arg = OutVals[realArgIdx];
2946 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2948 // Promote the value if needed.
2949 switch (VA.getLocInfo()) {
2951 llvm_unreachable("Unknown loc info!");
2952 case CCValAssign::Full:
2954 case CCValAssign::SExt:
2955 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
2957 case CCValAssign::ZExt:
2958 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2960 case CCValAssign::AExt:
2961 if (Outs[realArgIdx].ArgVT == MVT::i1) {
2962 // AAPCS requires i1 to be zero-extended to 8-bits by the caller.
2963 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
2964 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i8, Arg);
2966 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
2968 case CCValAssign::BCvt:
2969 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2971 case CCValAssign::FPExt:
2972 Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
2976 if (VA.isRegLoc()) {
2977 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i64) {
2978 assert(VA.getLocVT() == MVT::i64 &&
2979 "unexpected calling convention register assignment");
2980 assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
2981 "unexpected use of 'returned'");
2982 IsThisReturn = true;
2984 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2986 assert(VA.isMemLoc());
2989 MachinePointerInfo DstInfo;
2991 // FIXME: This works on big-endian for composite byvals, which are the
2992 // common case. It should also work for fundamental types too.
2993 uint32_t BEAlign = 0;
2994 unsigned OpSize = Flags.isByVal() ? Flags.getByValSize() * 8
2995 : VA.getValVT().getSizeInBits();
2996 OpSize = (OpSize + 7) / 8;
2997 if (!Subtarget->isLittleEndian() && !Flags.isByVal() &&
2998 !Flags.isInConsecutiveRegs()) {
3000 BEAlign = 8 - OpSize;
3002 unsigned LocMemOffset = VA.getLocMemOffset();
3003 int32_t Offset = LocMemOffset + BEAlign;
3004 SDValue PtrOff = DAG.getIntPtrConstant(Offset, DL);
3005 PtrOff = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, PtrOff);
3008 Offset = Offset + FPDiff;
3009 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3011 DstAddr = DAG.getFrameIndex(FI, PtrVT);
3013 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI);
3015 // Make sure any stack arguments overlapping with where we're storing
3016 // are loaded before this eventual operation. Otherwise they'll be
3018 Chain = addTokenForArgument(Chain, DAG, MF.getFrameInfo(), FI);
3020 SDValue PtrOff = DAG.getIntPtrConstant(Offset, DL);
3022 DstAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, PtrOff);
3023 DstInfo = MachinePointerInfo::getStack(DAG.getMachineFunction(),
3027 if (Outs[i].Flags.isByVal()) {
3029 DAG.getConstant(Outs[i].Flags.getByValSize(), DL, MVT::i64);
3030 SDValue Cpy = DAG.getMemcpy(
3031 Chain, DL, DstAddr, Arg, SizeNode, Outs[i].Flags.getByValAlign(),
3032 /*isVol = */ false, /*AlwaysInline = */ false,
3033 /*isTailCall = */ false,
3034 DstInfo, MachinePointerInfo());
3036 MemOpChains.push_back(Cpy);
3038 // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
3039 // promoted to a legal register type i32, we should truncate Arg back to
3041 if (VA.getValVT() == MVT::i1 || VA.getValVT() == MVT::i8 ||
3042 VA.getValVT() == MVT::i16)
3043 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Arg);
3046 DAG.getStore(Chain, DL, Arg, DstAddr, DstInfo, false, false, 0);
3047 MemOpChains.push_back(Store);
3052 if (!MemOpChains.empty())
3053 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
3055 // Build a sequence of copy-to-reg nodes chained together with token chain
3056 // and flag operands which copy the outgoing args into the appropriate regs.
3058 for (auto &RegToPass : RegsToPass) {
3059 Chain = DAG.getCopyToReg(Chain, DL, RegToPass.first,
3060 RegToPass.second, InFlag);
3061 InFlag = Chain.getValue(1);
3064 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
3065 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
3066 // node so that legalize doesn't hack it.
3067 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3068 Subtarget->isTargetMachO()) {
3069 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3070 const GlobalValue *GV = G->getGlobal();
3071 bool InternalLinkage = GV->hasInternalLinkage();
3072 if (InternalLinkage)
3073 Callee = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 0);
3076 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_GOT);
3077 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, Callee);
3079 } else if (ExternalSymbolSDNode *S =
3080 dyn_cast<ExternalSymbolSDNode>(Callee)) {
3081 const char *Sym = S->getSymbol();
3082 Callee = DAG.getTargetExternalSymbol(Sym, PtrVT, AArch64II::MO_GOT);
3083 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, Callee);
3085 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3086 const GlobalValue *GV = G->getGlobal();
3087 Callee = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 0);
3088 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3089 const char *Sym = S->getSymbol();
3090 Callee = DAG.getTargetExternalSymbol(Sym, PtrVT, 0);
3093 // We don't usually want to end the call-sequence here because we would tidy
3094 // the frame up *after* the call, however in the ABI-changing tail-call case
3095 // we've carefully laid out the parameters so that when sp is reset they'll be
3096 // in the correct location.
3097 if (IsTailCall && !IsSibCall) {
3098 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, DL, true),
3099 DAG.getIntPtrConstant(0, DL, true), InFlag, DL);
3100 InFlag = Chain.getValue(1);
3103 std::vector<SDValue> Ops;
3104 Ops.push_back(Chain);
3105 Ops.push_back(Callee);
3108 // Each tail call may have to adjust the stack by a different amount, so
3109 // this information must travel along with the operation for eventual
3110 // consumption by emitEpilogue.
3111 Ops.push_back(DAG.getTargetConstant(FPDiff, DL, MVT::i32));
3114 // Add argument registers to the end of the list so that they are known live
3116 for (auto &RegToPass : RegsToPass)
3117 Ops.push_back(DAG.getRegister(RegToPass.first,
3118 RegToPass.second.getValueType()));
3120 // Add a register mask operand representing the call-preserved registers.
3121 const uint32_t *Mask;
3122 const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();
3124 // For 'this' returns, use the X0-preserving mask if applicable
3125 Mask = TRI->getThisReturnPreservedMask(MF, CallConv);
3127 IsThisReturn = false;
3128 Mask = TRI->getCallPreservedMask(MF, CallConv);
3131 Mask = TRI->getCallPreservedMask(MF, CallConv);
3133 assert(Mask && "Missing call preserved mask for calling convention");
3134 Ops.push_back(DAG.getRegisterMask(Mask));
3136 if (InFlag.getNode())
3137 Ops.push_back(InFlag);
3139 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3141 // If we're doing a tall call, use a TC_RETURN here rather than an
3142 // actual call instruction.
3144 MF.getFrameInfo()->setHasTailCall();
3145 return DAG.getNode(AArch64ISD::TC_RETURN, DL, NodeTys, Ops);
3148 // Returns a chain and a flag for retval copy to use.
3149 Chain = DAG.getNode(AArch64ISD::CALL, DL, NodeTys, Ops);
3150 InFlag = Chain.getValue(1);
3152 uint64_t CalleePopBytes = DoesCalleeRestoreStack(CallConv, TailCallOpt)
3153 ? RoundUpToAlignment(NumBytes, 16)
3156 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, DL, true),
3157 DAG.getIntPtrConstant(CalleePopBytes, DL, true),
3160 InFlag = Chain.getValue(1);
3162 // Handle result values, copying them out of physregs into vregs that we
3164 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
3165 InVals, IsThisReturn,
3166 IsThisReturn ? OutVals[0] : SDValue());
3169 bool AArch64TargetLowering::CanLowerReturn(
3170 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
3171 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
3172 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
3173 ? RetCC_AArch64_WebKit_JS
3174 : RetCC_AArch64_AAPCS;
3175 SmallVector<CCValAssign, 16> RVLocs;
3176 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
3177 return CCInfo.CheckReturn(Outs, RetCC);
3181 AArch64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
3183 const SmallVectorImpl<ISD::OutputArg> &Outs,
3184 const SmallVectorImpl<SDValue> &OutVals,
3185 SDLoc DL, SelectionDAG &DAG) const {
3186 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
3187 ? RetCC_AArch64_WebKit_JS
3188 : RetCC_AArch64_AAPCS;
3189 SmallVector<CCValAssign, 16> RVLocs;
3190 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
3192 CCInfo.AnalyzeReturn(Outs, RetCC);
3194 // Copy the result values into the output registers.
3196 SmallVector<SDValue, 4> RetOps(1, Chain);
3197 for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size();
3198 ++i, ++realRVLocIdx) {
3199 CCValAssign &VA = RVLocs[i];
3200 assert(VA.isRegLoc() && "Can only return in registers!");
3201 SDValue Arg = OutVals[realRVLocIdx];
3203 switch (VA.getLocInfo()) {
3205 llvm_unreachable("Unknown loc info!");
3206 case CCValAssign::Full:
3207 if (Outs[i].ArgVT == MVT::i1) {
3208 // AAPCS requires i1 to be zero-extended to i8 by the producer of the
3209 // value. This is strictly redundant on Darwin (which uses "zeroext
3210 // i1"), but will be optimised out before ISel.
3211 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
3212 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
3215 case CCValAssign::BCvt:
3216 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
3220 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
3221 Flag = Chain.getValue(1);
3222 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
3225 RetOps[0] = Chain; // Update chain.
3227 // Add the flag if we have it.
3229 RetOps.push_back(Flag);
3231 return DAG.getNode(AArch64ISD::RET_FLAG, DL, MVT::Other, RetOps);
3234 //===----------------------------------------------------------------------===//
3235 // Other Lowering Code
3236 //===----------------------------------------------------------------------===//
3238 SDValue AArch64TargetLowering::LowerGlobalAddress(SDValue Op,
3239 SelectionDAG &DAG) const {
3240 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3242 const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
3243 const GlobalValue *GV = GN->getGlobal();
3244 unsigned char OpFlags =
3245 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
3247 assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
3248 "unexpected offset in global node");
3250 // This also catched the large code model case for Darwin.
3251 if ((OpFlags & AArch64II::MO_GOT) != 0) {
3252 SDValue GotAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
3253 // FIXME: Once remat is capable of dealing with instructions with register
3254 // operands, expand this into two nodes instead of using a wrapper node.
3255 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
3258 if ((OpFlags & AArch64II::MO_CONSTPOOL) != 0) {
3259 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
3260 "use of MO_CONSTPOOL only supported on small model");
3261 SDValue Hi = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, AArch64II::MO_PAGE);
3262 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3263 unsigned char LoFlags = AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
3264 SDValue Lo = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, LoFlags);
3265 SDValue PoolAddr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3266 SDValue GlobalAddr = DAG.getLoad(
3267 PtrVT, DL, DAG.getEntryNode(), PoolAddr,
3268 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
3269 /*isVolatile=*/false,
3270 /*isNonTemporal=*/true,
3271 /*isInvariant=*/true, 8);
3272 if (GN->getOffset() != 0)
3273 return DAG.getNode(ISD::ADD, DL, PtrVT, GlobalAddr,
3274 DAG.getConstant(GN->getOffset(), DL, PtrVT));
3278 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
3279 const unsigned char MO_NC = AArch64II::MO_NC;
3281 AArch64ISD::WrapperLarge, DL, PtrVT,
3282 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G3),
3283 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
3284 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
3285 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
3287 // Use ADRP/ADD or ADRP/LDR for everything else: the small model on ELF and
3288 // the only correct model on Darwin.
3289 SDValue Hi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
3290 OpFlags | AArch64II::MO_PAGE);
3291 unsigned char LoFlags = OpFlags | AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
3292 SDValue Lo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, LoFlags);
3294 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3295 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3299 /// \brief Convert a TLS address reference into the correct sequence of loads
3300 /// and calls to compute the variable's address (for Darwin, currently) and
3301 /// return an SDValue containing the final node.
3303 /// Darwin only has one TLS scheme which must be capable of dealing with the
3304 /// fully general situation, in the worst case. This means:
3305 /// + "extern __thread" declaration.
3306 /// + Defined in a possibly unknown dynamic library.
3308 /// The general system is that each __thread variable has a [3 x i64] descriptor
3309 /// which contains information used by the runtime to calculate the address. The
3310 /// only part of this the compiler needs to know about is the first xword, which
3311 /// contains a function pointer that must be called with the address of the
3312 /// entire descriptor in "x0".
3314 /// Since this descriptor may be in a different unit, in general even the
3315 /// descriptor must be accessed via an indirect load. The "ideal" code sequence
3317 /// adrp x0, _var@TLVPPAGE
3318 /// ldr x0, [x0, _var@TLVPPAGEOFF] ; x0 now contains address of descriptor
3319 /// ldr x1, [x0] ; x1 contains 1st entry of descriptor,
3320 /// ; the function pointer
3321 /// blr x1 ; Uses descriptor address in x0
3322 /// ; Address of _var is now in x0.
3324 /// If the address of _var's descriptor *is* known to the linker, then it can
3325 /// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
3326 /// a slight efficiency gain.
3328 AArch64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
3329 SelectionDAG &DAG) const {
3330 assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");
3333 MVT PtrVT = getPointerTy(DAG.getDataLayout());
3334 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3337 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3338 SDValue DescAddr = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TLVPAddr);
3340 // The first entry in the descriptor is a function pointer that we must call
3341 // to obtain the address of the variable.
3342 SDValue Chain = DAG.getEntryNode();
3343 SDValue FuncTLVGet =
3344 DAG.getLoad(MVT::i64, DL, Chain, DescAddr,
3345 MachinePointerInfo::getGOT(DAG.getMachineFunction()), false,
3347 Chain = FuncTLVGet.getValue(1);
3349 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3350 MFI->setAdjustsStack(true);
3352 // TLS calls preserve all registers except those that absolutely must be
3353 // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
3355 const uint32_t *Mask =
3356 Subtarget->getRegisterInfo()->getTLSCallPreservedMask();
3358 // Finally, we can make the call. This is just a degenerate version of a
3359 // normal AArch64 call node: x0 takes the address of the descriptor, and
3360 // returns the address of the variable in this thread.
3361 Chain = DAG.getCopyToReg(Chain, DL, AArch64::X0, DescAddr, SDValue());
3363 DAG.getNode(AArch64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
3364 Chain, FuncTLVGet, DAG.getRegister(AArch64::X0, MVT::i64),
3365 DAG.getRegisterMask(Mask), Chain.getValue(1));
3366 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Chain.getValue(1));
3369 /// When accessing thread-local variables under either the general-dynamic or
3370 /// local-dynamic system, we make a "TLS-descriptor" call. The variable will
3371 /// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
3372 /// is a function pointer to carry out the resolution.
3374 /// The sequence is:
3375 /// adrp x0, :tlsdesc:var
3376 /// ldr x1, [x0, #:tlsdesc_lo12:var]
3377 /// add x0, x0, #:tlsdesc_lo12:var
3378 /// .tlsdesccall var
3380 /// (TPIDR_EL0 offset now in x0)
3382 /// The above sequence must be produced unscheduled, to enable the linker to
3383 /// optimize/relax this sequence.
3384 /// Therefore, a pseudo-instruction (TLSDESC_CALLSEQ) is used to represent the
3385 /// above sequence, and expanded really late in the compilation flow, to ensure
3386 /// the sequence is produced as per above.
3387 SDValue AArch64TargetLowering::LowerELFTLSDescCallSeq(SDValue SymAddr, SDLoc DL,
3388 SelectionDAG &DAG) const {
3389 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3391 SDValue Chain = DAG.getEntryNode();
3392 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3394 SmallVector<SDValue, 2> Ops;
3395 Ops.push_back(Chain);
3396 Ops.push_back(SymAddr);
3398 Chain = DAG.getNode(AArch64ISD::TLSDESC_CALLSEQ, DL, NodeTys, Ops);
3399 SDValue Glue = Chain.getValue(1);
3401 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Glue);
3405 AArch64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
3406 SelectionDAG &DAG) const {
3407 assert(Subtarget->isTargetELF() && "This function expects an ELF target");
3408 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
3409 "ELF TLS only supported in small memory model");
3410 // Different choices can be made for the maximum size of the TLS area for a
3411 // module. For the small address model, the default TLS size is 16MiB and the
3412 // maximum TLS size is 4GiB.
3413 // FIXME: add -mtls-size command line option and make it control the 16MiB
3414 // vs. 4GiB code sequence generation.
3415 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3417 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
3419 if (DAG.getTarget().Options.EmulatedTLS)
3420 return LowerToTLSEmulatedModel(GA, DAG);
3422 if (!EnableAArch64ELFLocalDynamicTLSGeneration) {
3423 if (Model == TLSModel::LocalDynamic)
3424 Model = TLSModel::GeneralDynamic;
3428 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3430 const GlobalValue *GV = GA->getGlobal();
3432 SDValue ThreadBase = DAG.getNode(AArch64ISD::THREAD_POINTER, DL, PtrVT);
3434 if (Model == TLSModel::LocalExec) {
3435 SDValue HiVar = DAG.getTargetGlobalAddress(
3436 GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_HI12);
3437 SDValue LoVar = DAG.getTargetGlobalAddress(
3439 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3441 SDValue TPWithOff_lo =
3442 SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, ThreadBase,
3444 DAG.getTargetConstant(0, DL, MVT::i32)),
3447 SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPWithOff_lo,
3449 DAG.getTargetConstant(0, DL, MVT::i32)),
3452 } else if (Model == TLSModel::InitialExec) {
3453 TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3454 TPOff = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TPOff);
3455 } else if (Model == TLSModel::LocalDynamic) {
3456 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
3457 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
3458 // the beginning of the module's TLS region, followed by a DTPREL offset
3461 // These accesses will need deduplicating if there's more than one.
3462 AArch64FunctionInfo *MFI =
3463 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
3464 MFI->incNumLocalDynamicTLSAccesses();
3466 // The call needs a relocation too for linker relaxation. It doesn't make
3467 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
3469 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
3472 // Now we can calculate the offset from TPIDR_EL0 to this module's
3473 // thread-local area.
3474 TPOff = LowerELFTLSDescCallSeq(SymAddr, DL, DAG);
3476 // Now use :dtprel_whatever: operations to calculate this variable's offset
3477 // in its thread-storage area.
3478 SDValue HiVar = DAG.getTargetGlobalAddress(
3479 GV, DL, MVT::i64, 0, AArch64II::MO_TLS | AArch64II::MO_HI12);
3480 SDValue LoVar = DAG.getTargetGlobalAddress(
3481 GV, DL, MVT::i64, 0,
3482 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3484 TPOff = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPOff, HiVar,
3485 DAG.getTargetConstant(0, DL, MVT::i32)),
3487 TPOff = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPOff, LoVar,
3488 DAG.getTargetConstant(0, DL, MVT::i32)),
3490 } else if (Model == TLSModel::GeneralDynamic) {
3491 // The call needs a relocation too for linker relaxation. It doesn't make
3492 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
3495 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3497 // Finally we can make a call to calculate the offset from tpidr_el0.
3498 TPOff = LowerELFTLSDescCallSeq(SymAddr, DL, DAG);
3500 llvm_unreachable("Unsupported ELF TLS access model");
3502 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
3505 SDValue AArch64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
3506 SelectionDAG &DAG) const {
3507 if (Subtarget->isTargetDarwin())
3508 return LowerDarwinGlobalTLSAddress(Op, DAG);
3509 else if (Subtarget->isTargetELF())
3510 return LowerELFGlobalTLSAddress(Op, DAG);
3512 llvm_unreachable("Unexpected platform trying to use TLS");
3514 SDValue AArch64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3515 SDValue Chain = Op.getOperand(0);
3516 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3517 SDValue LHS = Op.getOperand(2);
3518 SDValue RHS = Op.getOperand(3);
3519 SDValue Dest = Op.getOperand(4);
3522 // Handle f128 first, since lowering it will result in comparing the return
3523 // value of a libcall against zero, which is just what the rest of LowerBR_CC
3524 // is expecting to deal with.
3525 if (LHS.getValueType() == MVT::f128) {
3526 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3528 // If softenSetCCOperands returned a scalar, we need to compare the result
3529 // against zero to select between true and false values.
3530 if (!RHS.getNode()) {
3531 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3536 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
3538 unsigned Opc = LHS.getOpcode();
3539 if (LHS.getResNo() == 1 && isa<ConstantSDNode>(RHS) &&
3540 cast<ConstantSDNode>(RHS)->isOne() &&
3541 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3542 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3543 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
3544 "Unexpected condition code.");
3545 // Only lower legal XALUO ops.
3546 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
3549 // The actual operation with overflow check.
3550 AArch64CC::CondCode OFCC;
3551 SDValue Value, Overflow;
3552 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, LHS.getValue(0), DAG);
3554 if (CC == ISD::SETNE)
3555 OFCC = getInvertedCondCode(OFCC);
3556 SDValue CCVal = DAG.getConstant(OFCC, dl, MVT::i32);
3558 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3562 if (LHS.getValueType().isInteger()) {
3563 assert((LHS.getValueType() == RHS.getValueType()) &&
3564 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3566 // If the RHS of the comparison is zero, we can potentially fold this
3567 // to a specialized branch.
3568 const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
3569 if (RHSC && RHSC->getZExtValue() == 0) {
3570 if (CC == ISD::SETEQ) {
3571 // See if we can use a TBZ to fold in an AND as well.
3572 // TBZ has a smaller branch displacement than CBZ. If the offset is
3573 // out of bounds, a late MI-layer pass rewrites branches.
3574 // 403.gcc is an example that hits this case.
3575 if (LHS.getOpcode() == ISD::AND &&
3576 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3577 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3578 SDValue Test = LHS.getOperand(0);
3579 uint64_t Mask = LHS.getConstantOperandVal(1);
3580 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, Test,
3581 DAG.getConstant(Log2_64(Mask), dl, MVT::i64),
3585 return DAG.getNode(AArch64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest);
3586 } else if (CC == ISD::SETNE) {
3587 // See if we can use a TBZ to fold in an AND as well.
3588 // TBZ has a smaller branch displacement than CBZ. If the offset is
3589 // out of bounds, a late MI-layer pass rewrites branches.
3590 // 403.gcc is an example that hits this case.
3591 if (LHS.getOpcode() == ISD::AND &&
3592 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3593 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3594 SDValue Test = LHS.getOperand(0);
3595 uint64_t Mask = LHS.getConstantOperandVal(1);
3596 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, Test,
3597 DAG.getConstant(Log2_64(Mask), dl, MVT::i64),
3601 return DAG.getNode(AArch64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest);
3602 } else if (CC == ISD::SETLT && LHS.getOpcode() != ISD::AND) {
3603 // Don't combine AND since emitComparison converts the AND to an ANDS
3604 // (a.k.a. TST) and the test in the test bit and branch instruction
3605 // becomes redundant. This would also increase register pressure.
3606 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
3607 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, LHS,
3608 DAG.getConstant(Mask, dl, MVT::i64), Dest);
3611 if (RHSC && RHSC->getSExtValue() == -1 && CC == ISD::SETGT &&
3612 LHS.getOpcode() != ISD::AND) {
3613 // Don't combine AND since emitComparison converts the AND to an ANDS
3614 // (a.k.a. TST) and the test in the test bit and branch instruction
3615 // becomes redundant. This would also increase register pressure.
3616 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
3617 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, LHS,
3618 DAG.getConstant(Mask, dl, MVT::i64), Dest);
3622 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3623 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3627 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3629 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
3630 // clean. Some of them require two branches to implement.
3631 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3632 AArch64CC::CondCode CC1, CC2;
3633 changeFPCCToAArch64CC(CC, CC1, CC2);
3634 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
3636 DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp);
3637 if (CC2 != AArch64CC::AL) {
3638 SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32);
3639 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val,
3646 SDValue AArch64TargetLowering::LowerFCOPYSIGN(SDValue Op,
3647 SelectionDAG &DAG) const {
3648 EVT VT = Op.getValueType();
3651 SDValue In1 = Op.getOperand(0);
3652 SDValue In2 = Op.getOperand(1);
3653 EVT SrcVT = In2.getValueType();
3655 if (SrcVT.bitsLT(VT))
3656 In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2);
3657 else if (SrcVT.bitsGT(VT))
3658 In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, DAG.getIntPtrConstant(0, DL));
3663 SDValue VecVal1, VecVal2;
3664 if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) {
3666 VecVT = (VT == MVT::v2f32 ? MVT::v2i32 : MVT::v4i32);
3667 EltMask = 0x80000000ULL;
3669 if (!VT.isVector()) {
3670 VecVal1 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3671 DAG.getUNDEF(VecVT), In1);
3672 VecVal2 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3673 DAG.getUNDEF(VecVT), In2);
3675 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3676 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3678 } else if (VT == MVT::f64 || VT == MVT::v2f64) {
3682 // We want to materialize a mask with the high bit set, but the AdvSIMD
3683 // immediate moves cannot materialize that in a single instruction for
3684 // 64-bit elements. Instead, materialize zero and then negate it.
3687 if (!VT.isVector()) {
3688 VecVal1 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3689 DAG.getUNDEF(VecVT), In1);
3690 VecVal2 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3691 DAG.getUNDEF(VecVT), In2);
3693 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3694 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3697 llvm_unreachable("Invalid type for copysign!");
3700 SDValue BuildVec = DAG.getConstant(EltMask, DL, VecVT);
3702 // If we couldn't materialize the mask above, then the mask vector will be
3703 // the zero vector, and we need to negate it here.
3704 if (VT == MVT::f64 || VT == MVT::v2f64) {
3705 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec);
3706 BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec);
3707 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec);
3711 DAG.getNode(AArch64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec);
3714 return DAG.getTargetExtractSubreg(AArch64::ssub, DL, VT, Sel);
3715 else if (VT == MVT::f64)
3716 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, VT, Sel);
3718 return DAG.getNode(ISD::BITCAST, DL, VT, Sel);
3721 SDValue AArch64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
3722 if (DAG.getMachineFunction().getFunction()->hasFnAttribute(
3723 Attribute::NoImplicitFloat))
3726 if (!Subtarget->hasNEON())
3729 // While there is no integer popcount instruction, it can
3730 // be more efficiently lowered to the following sequence that uses
3731 // AdvSIMD registers/instructions as long as the copies to/from
3732 // the AdvSIMD registers are cheap.
3733 // FMOV D0, X0 // copy 64-bit int to vector, high bits zero'd
3734 // CNT V0.8B, V0.8B // 8xbyte pop-counts
3735 // ADDV B0, V0.8B // sum 8xbyte pop-counts
3736 // UMOV X0, V0.B[0] // copy byte result back to integer reg
3737 SDValue Val = Op.getOperand(0);
3739 EVT VT = Op.getValueType();
3742 Val = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Val);
3743 Val = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);
3745 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, Val);
3746 SDValue UaddLV = DAG.getNode(
3747 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
3748 DAG.getConstant(Intrinsic::aarch64_neon_uaddlv, DL, MVT::i32), CtPop);
3751 UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV);
3755 SDValue AArch64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
3757 if (Op.getValueType().isVector())
3758 return LowerVSETCC(Op, DAG);
3760 SDValue LHS = Op.getOperand(0);
3761 SDValue RHS = Op.getOperand(1);
3762 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
3765 // We chose ZeroOrOneBooleanContents, so use zero and one.
3766 EVT VT = Op.getValueType();
3767 SDValue TVal = DAG.getConstant(1, dl, VT);
3768 SDValue FVal = DAG.getConstant(0, dl, VT);
3770 // Handle f128 first, since one possible outcome is a normal integer
3771 // comparison which gets picked up by the next if statement.
3772 if (LHS.getValueType() == MVT::f128) {
3773 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3775 // If softenSetCCOperands returned a scalar, use it.
3776 if (!RHS.getNode()) {
3777 assert(LHS.getValueType() == Op.getValueType() &&
3778 "Unexpected setcc expansion!");
3783 if (LHS.getValueType().isInteger()) {
3786 getAArch64Cmp(LHS, RHS, ISD::getSetCCInverse(CC, true), CCVal, DAG, dl);
3788 // Note that we inverted the condition above, so we reverse the order of
3789 // the true and false operands here. This will allow the setcc to be
3790 // matched to a single CSINC instruction.
3791 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp);
3794 // Now we know we're dealing with FP values.
3795 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3797 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3798 // and do the comparison.
3799 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3801 AArch64CC::CondCode CC1, CC2;
3802 changeFPCCToAArch64CC(CC, CC1, CC2);
3803 if (CC2 == AArch64CC::AL) {
3804 changeFPCCToAArch64CC(ISD::getSetCCInverse(CC, false), CC1, CC2);
3805 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
3807 // Note that we inverted the condition above, so we reverse the order of
3808 // the true and false operands here. This will allow the setcc to be
3809 // matched to a single CSINC instruction.
3810 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp);
3812 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't
3813 // totally clean. Some of them require two CSELs to implement. As is in
3814 // this case, we emit the first CSEL and then emit a second using the output
3815 // of the first as the RHS. We're effectively OR'ing the two CC's together.
3817 // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
3818 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
3820 DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3822 SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32);
3823 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3827 /// A SELECT_CC operation is really some kind of max or min if both values being
3828 /// compared are, in some sense, equal to the results in either case. However,
3829 /// it is permissible to compare f32 values and produce directly extended f64
3832 /// Extending the comparison operands would also be allowed, but is less likely
3833 /// to happen in practice since their use is right here. Note that truncate
3834 /// operations would *not* be semantically equivalent.
3835 static bool selectCCOpsAreFMaxCompatible(SDValue Cmp, SDValue Result) {
3837 return (Cmp.getValueType() == MVT::f32 ||
3838 Cmp.getValueType() == MVT::f64);
3840 ConstantFPSDNode *CCmp = dyn_cast<ConstantFPSDNode>(Cmp);
3841 ConstantFPSDNode *CResult = dyn_cast<ConstantFPSDNode>(Result);
3842 if (CCmp && CResult && Cmp.getValueType() == MVT::f32 &&
3843 Result.getValueType() == MVT::f64) {
3845 APFloat CmpVal = CCmp->getValueAPF();
3846 CmpVal.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &Lossy);
3847 return CResult->getValueAPF().bitwiseIsEqual(CmpVal);
3850 return Result->getOpcode() == ISD::FP_EXTEND && Result->getOperand(0) == Cmp;
3853 SDValue AArch64TargetLowering::LowerSELECT_CC(ISD::CondCode CC, SDValue LHS,
3854 SDValue RHS, SDValue TVal,
3855 SDValue FVal, SDLoc dl,
3856 SelectionDAG &DAG) const {
3857 // Handle f128 first, because it will result in a comparison of some RTLIB
3858 // call result against zero.
3859 if (LHS.getValueType() == MVT::f128) {
3860 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3862 // If softenSetCCOperands returned a scalar, we need to compare the result
3863 // against zero to select between true and false values.
3864 if (!RHS.getNode()) {
3865 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3870 // Handle integers first.
3871 if (LHS.getValueType().isInteger()) {
3872 assert((LHS.getValueType() == RHS.getValueType()) &&
3873 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3875 unsigned Opcode = AArch64ISD::CSEL;
3877 // If both the TVal and the FVal are constants, see if we can swap them in
3878 // order to for a CSINV or CSINC out of them.
3879 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
3880 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
3882 if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) {
3883 std::swap(TVal, FVal);
3884 std::swap(CTVal, CFVal);
3885 CC = ISD::getSetCCInverse(CC, true);
3886 } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) {
3887 std::swap(TVal, FVal);
3888 std::swap(CTVal, CFVal);
3889 CC = ISD::getSetCCInverse(CC, true);
3890 } else if (TVal.getOpcode() == ISD::XOR) {
3891 // If TVal is a NOT we want to swap TVal and FVal so that we can match
3892 // with a CSINV rather than a CSEL.
3893 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(1));
3895 if (CVal && CVal->isAllOnesValue()) {
3896 std::swap(TVal, FVal);
3897 std::swap(CTVal, CFVal);
3898 CC = ISD::getSetCCInverse(CC, true);
3900 } else if (TVal.getOpcode() == ISD::SUB) {
3901 // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
3902 // that we can match with a CSNEG rather than a CSEL.
3903 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(0));
3905 if (CVal && CVal->isNullValue()) {
3906 std::swap(TVal, FVal);
3907 std::swap(CTVal, CFVal);
3908 CC = ISD::getSetCCInverse(CC, true);
3910 } else if (CTVal && CFVal) {
3911 const int64_t TrueVal = CTVal->getSExtValue();
3912 const int64_t FalseVal = CFVal->getSExtValue();
3915 // If both TVal and FVal are constants, see if FVal is the
3916 // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
3917 // instead of a CSEL in that case.
3918 if (TrueVal == ~FalseVal) {
3919 Opcode = AArch64ISD::CSINV;
3920 } else if (TrueVal == -FalseVal) {
3921 Opcode = AArch64ISD::CSNEG;
3922 } else if (TVal.getValueType() == MVT::i32) {
3923 // If our operands are only 32-bit wide, make sure we use 32-bit
3924 // arithmetic for the check whether we can use CSINC. This ensures that
3925 // the addition in the check will wrap around properly in case there is
3926 // an overflow (which would not be the case if we do the check with
3927 // 64-bit arithmetic).
3928 const uint32_t TrueVal32 = CTVal->getZExtValue();
3929 const uint32_t FalseVal32 = CFVal->getZExtValue();
3931 if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
3932 Opcode = AArch64ISD::CSINC;
3934 if (TrueVal32 > FalseVal32) {
3938 // 64-bit check whether we can use CSINC.
3939 } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) {
3940 Opcode = AArch64ISD::CSINC;
3942 if (TrueVal > FalseVal) {
3947 // Swap TVal and FVal if necessary.
3949 std::swap(TVal, FVal);
3950 std::swap(CTVal, CFVal);
3951 CC = ISD::getSetCCInverse(CC, true);
3954 if (Opcode != AArch64ISD::CSEL) {
3955 // Drop FVal since we can get its value by simply inverting/negating
3962 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3964 EVT VT = TVal.getValueType();
3965 return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp);
3968 // Now we know we're dealing with FP values.
3969 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3970 assert(LHS.getValueType() == RHS.getValueType());
3971 EVT VT = TVal.getValueType();
3972 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3974 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
3975 // clean. Some of them require two CSELs to implement.
3976 AArch64CC::CondCode CC1, CC2;
3977 changeFPCCToAArch64CC(CC, CC1, CC2);
3978 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
3979 SDValue CS1 = DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3981 // If we need a second CSEL, emit it, using the output of the first as the
3982 // RHS. We're effectively OR'ing the two CC's together.
3983 if (CC2 != AArch64CC::AL) {
3984 SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32);
3985 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3988 // Otherwise, return the output of the first CSEL.
3992 SDValue AArch64TargetLowering::LowerSELECT_CC(SDValue Op,
3993 SelectionDAG &DAG) const {
3994 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3995 SDValue LHS = Op.getOperand(0);
3996 SDValue RHS = Op.getOperand(1);
3997 SDValue TVal = Op.getOperand(2);
3998 SDValue FVal = Op.getOperand(3);
4000 return LowerSELECT_CC(CC, LHS, RHS, TVal, FVal, DL, DAG);
4003 SDValue AArch64TargetLowering::LowerSELECT(SDValue Op,
4004 SelectionDAG &DAG) const {
4005 SDValue CCVal = Op->getOperand(0);
4006 SDValue TVal = Op->getOperand(1);
4007 SDValue FVal = Op->getOperand(2);
4010 unsigned Opc = CCVal.getOpcode();
4011 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
4013 if (CCVal.getResNo() == 1 &&
4014 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
4015 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
4016 // Only lower legal XALUO ops.
4017 if (!DAG.getTargetLoweringInfo().isTypeLegal(CCVal->getValueType(0)))
4020 AArch64CC::CondCode OFCC;
4021 SDValue Value, Overflow;
4022 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, CCVal.getValue(0), DAG);
4023 SDValue CCVal = DAG.getConstant(OFCC, DL, MVT::i32);
4025 return DAG.getNode(AArch64ISD::CSEL, DL, Op.getValueType(), TVal, FVal,
4029 // Lower it the same way as we would lower a SELECT_CC node.
4032 if (CCVal.getOpcode() == ISD::SETCC) {
4033 LHS = CCVal.getOperand(0);
4034 RHS = CCVal.getOperand(1);
4035 CC = cast<CondCodeSDNode>(CCVal->getOperand(2))->get();
4038 RHS = DAG.getConstant(0, DL, CCVal.getValueType());
4041 return LowerSELECT_CC(CC, LHS, RHS, TVal, FVal, DL, DAG);
4044 SDValue AArch64TargetLowering::LowerJumpTable(SDValue Op,
4045 SelectionDAG &DAG) const {
4046 // Jump table entries as PC relative offsets. No additional tweaking
4047 // is necessary here. Just get the address of the jump table.
4048 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4049 EVT PtrVT = getPointerTy(DAG.getDataLayout());
4052 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
4053 !Subtarget->isTargetMachO()) {
4054 const unsigned char MO_NC = AArch64II::MO_NC;
4056 AArch64ISD::WrapperLarge, DL, PtrVT,
4057 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G3),
4058 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G2 | MO_NC),
4059 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G1 | MO_NC),
4060 DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
4061 AArch64II::MO_G0 | MO_NC));
4065 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_PAGE);
4066 SDValue Lo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
4067 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
4068 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
4069 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
4072 SDValue AArch64TargetLowering::LowerConstantPool(SDValue Op,
4073 SelectionDAG &DAG) const {
4074 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4075 EVT PtrVT = getPointerTy(DAG.getDataLayout());
4078 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
4079 // Use the GOT for the large code model on iOS.
4080 if (Subtarget->isTargetMachO()) {
4081 SDValue GotAddr = DAG.getTargetConstantPool(
4082 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
4084 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
4087 const unsigned char MO_NC = AArch64II::MO_NC;
4089 AArch64ISD::WrapperLarge, DL, PtrVT,
4090 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4091 CP->getOffset(), AArch64II::MO_G3),
4092 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4093 CP->getOffset(), AArch64II::MO_G2 | MO_NC),
4094 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4095 CP->getOffset(), AArch64II::MO_G1 | MO_NC),
4096 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4097 CP->getOffset(), AArch64II::MO_G0 | MO_NC));
4099 // Use ADRP/ADD or ADRP/LDR for everything else: the small memory model on
4100 // ELF, the only valid one on Darwin.
4102 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4103 CP->getOffset(), AArch64II::MO_PAGE);
4104 SDValue Lo = DAG.getTargetConstantPool(
4105 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
4106 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
4108 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
4109 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
4113 SDValue AArch64TargetLowering::LowerBlockAddress(SDValue Op,
4114 SelectionDAG &DAG) const {
4115 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
4116 EVT PtrVT = getPointerTy(DAG.getDataLayout());
4118 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
4119 !Subtarget->isTargetMachO()) {
4120 const unsigned char MO_NC = AArch64II::MO_NC;
4122 AArch64ISD::WrapperLarge, DL, PtrVT,
4123 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G3),
4124 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
4125 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
4126 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
4128 SDValue Hi = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGE);
4129 SDValue Lo = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGEOFF |
4131 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
4132 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
4136 SDValue AArch64TargetLowering::LowerDarwin_VASTART(SDValue Op,
4137 SelectionDAG &DAG) const {
4138 AArch64FunctionInfo *FuncInfo =
4139 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
4142 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(),
4143 getPointerTy(DAG.getDataLayout()));
4144 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
4145 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
4146 MachinePointerInfo(SV), false, false, 0);
4149 SDValue AArch64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
4150 SelectionDAG &DAG) const {
4151 // The layout of the va_list struct is specified in the AArch64 Procedure Call
4152 // Standard, section B.3.
4153 MachineFunction &MF = DAG.getMachineFunction();
4154 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
4155 auto PtrVT = getPointerTy(DAG.getDataLayout());
4158 SDValue Chain = Op.getOperand(0);
4159 SDValue VAList = Op.getOperand(1);
4160 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
4161 SmallVector<SDValue, 4> MemOps;
4163 // void *__stack at offset 0
4164 SDValue Stack = DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), PtrVT);
4165 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
4166 MachinePointerInfo(SV), false, false, 8));
4168 // void *__gr_top at offset 8
4169 int GPRSize = FuncInfo->getVarArgsGPRSize();
4171 SDValue GRTop, GRTopAddr;
4174 DAG.getNode(ISD::ADD, DL, PtrVT, VAList, DAG.getConstant(8, DL, PtrVT));
4176 GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), PtrVT);
4177 GRTop = DAG.getNode(ISD::ADD, DL, PtrVT, GRTop,
4178 DAG.getConstant(GPRSize, DL, PtrVT));
4180 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
4181 MachinePointerInfo(SV, 8), false, false, 8));
4184 // void *__vr_top at offset 16
4185 int FPRSize = FuncInfo->getVarArgsFPRSize();
4187 SDValue VRTop, VRTopAddr;
4188 VRTopAddr = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
4189 DAG.getConstant(16, DL, PtrVT));
4191 VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), PtrVT);
4192 VRTop = DAG.getNode(ISD::ADD, DL, PtrVT, VRTop,
4193 DAG.getConstant(FPRSize, DL, PtrVT));
4195 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
4196 MachinePointerInfo(SV, 16), false, false, 8));
4199 // int __gr_offs at offset 24
4200 SDValue GROffsAddr =
4201 DAG.getNode(ISD::ADD, DL, PtrVT, VAList, DAG.getConstant(24, DL, PtrVT));
4202 MemOps.push_back(DAG.getStore(Chain, DL,
4203 DAG.getConstant(-GPRSize, DL, MVT::i32),
4204 GROffsAddr, MachinePointerInfo(SV, 24), false,
4207 // int __vr_offs at offset 28
4208 SDValue VROffsAddr =
4209 DAG.getNode(ISD::ADD, DL, PtrVT, VAList, DAG.getConstant(28, DL, PtrVT));
4210 MemOps.push_back(DAG.getStore(Chain, DL,
4211 DAG.getConstant(-FPRSize, DL, MVT::i32),
4212 VROffsAddr, MachinePointerInfo(SV, 28), false,
4215 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
4218 SDValue AArch64TargetLowering::LowerVASTART(SDValue Op,
4219 SelectionDAG &DAG) const {
4220 return Subtarget->isTargetDarwin() ? LowerDarwin_VASTART(Op, DAG)
4221 : LowerAAPCS_VASTART(Op, DAG);
4224 SDValue AArch64TargetLowering::LowerVACOPY(SDValue Op,
4225 SelectionDAG &DAG) const {
4226 // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
4229 unsigned VaListSize = Subtarget->isTargetDarwin() ? 8 : 32;
4230 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
4231 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
4233 return DAG.getMemcpy(Op.getOperand(0), DL, Op.getOperand(1),
4235 DAG.getConstant(VaListSize, DL, MVT::i32),
4236 8, false, false, false, MachinePointerInfo(DestSV),
4237 MachinePointerInfo(SrcSV));
4240 SDValue AArch64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
4241 assert(Subtarget->isTargetDarwin() &&
4242 "automatic va_arg instruction only works on Darwin");
4244 const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
4245 EVT VT = Op.getValueType();
4247 SDValue Chain = Op.getOperand(0);
4248 SDValue Addr = Op.getOperand(1);
4249 unsigned Align = Op.getConstantOperandVal(3);
4250 auto PtrVT = getPointerTy(DAG.getDataLayout());
4252 SDValue VAList = DAG.getLoad(PtrVT, DL, Chain, Addr, MachinePointerInfo(V),
4253 false, false, false, 0);
4254 Chain = VAList.getValue(1);
4257 assert(((Align & (Align - 1)) == 0) && "Expected Align to be a power of 2");
4258 VAList = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
4259 DAG.getConstant(Align - 1, DL, PtrVT));
4260 VAList = DAG.getNode(ISD::AND, DL, PtrVT, VAList,
4261 DAG.getConstant(-(int64_t)Align, DL, PtrVT));
4264 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
4265 uint64_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
4267 // Scalar integer and FP values smaller than 64 bits are implicitly extended
4268 // up to 64 bits. At the very least, we have to increase the striding of the
4269 // vaargs list to match this, and for FP values we need to introduce
4270 // FP_ROUND nodes as well.
4271 if (VT.isInteger() && !VT.isVector())
4273 bool NeedFPTrunc = false;
4274 if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
4279 // Increment the pointer, VAList, to the next vaarg
4280 SDValue VANext = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
4281 DAG.getConstant(ArgSize, DL, PtrVT));
4282 // Store the incremented VAList to the legalized pointer
4283 SDValue APStore = DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V),
4286 // Load the actual argument out of the pointer VAList
4288 // Load the value as an f64.
4289 SDValue WideFP = DAG.getLoad(MVT::f64, DL, APStore, VAList,
4290 MachinePointerInfo(), false, false, false, 0);
4291 // Round the value down to an f32.
4292 SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
4293 DAG.getIntPtrConstant(1, DL));
4294 SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
4295 // Merge the rounded value with the chain output of the load.
4296 return DAG.getMergeValues(Ops, DL);
4299 return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo(), false,
4303 SDValue AArch64TargetLowering::LowerFRAMEADDR(SDValue Op,
4304 SelectionDAG &DAG) const {
4305 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4306 MFI->setFrameAddressIsTaken(true);
4308 EVT VT = Op.getValueType();
4310 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4312 DAG.getCopyFromReg(DAG.getEntryNode(), DL, AArch64::FP, VT);
4314 FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
4315 MachinePointerInfo(), false, false, false, 0);
4319 // FIXME? Maybe this could be a TableGen attribute on some registers and
4320 // this table could be generated automatically from RegInfo.
4321 unsigned AArch64TargetLowering::getRegisterByName(const char* RegName, EVT VT,
4322 SelectionDAG &DAG) const {
4323 unsigned Reg = StringSwitch<unsigned>(RegName)
4324 .Case("sp", AArch64::SP)
4328 report_fatal_error(Twine("Invalid register name \""
4329 + StringRef(RegName) + "\"."));
4332 SDValue AArch64TargetLowering::LowerRETURNADDR(SDValue Op,
4333 SelectionDAG &DAG) const {
4334 MachineFunction &MF = DAG.getMachineFunction();
4335 MachineFrameInfo *MFI = MF.getFrameInfo();
4336 MFI->setReturnAddressIsTaken(true);
4338 EVT VT = Op.getValueType();
4340 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4342 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4343 SDValue Offset = DAG.getConstant(8, DL, getPointerTy(DAG.getDataLayout()));
4344 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
4345 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
4346 MachinePointerInfo(), false, false, false, 0);
4349 // Return LR, which contains the return address. Mark it an implicit live-in.
4350 unsigned Reg = MF.addLiveIn(AArch64::LR, &AArch64::GPR64RegClass);
4351 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
4354 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4355 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4356 SDValue AArch64TargetLowering::LowerShiftRightParts(SDValue Op,
4357 SelectionDAG &DAG) const {
4358 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4359 EVT VT = Op.getValueType();
4360 unsigned VTBits = VT.getSizeInBits();
4362 SDValue ShOpLo = Op.getOperand(0);
4363 SDValue ShOpHi = Op.getOperand(1);
4364 SDValue ShAmt = Op.getOperand(2);
4366 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4368 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4370 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4371 DAG.getConstant(VTBits, dl, MVT::i64), ShAmt);
4372 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4373 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4374 DAG.getConstant(VTBits, dl, MVT::i64));
4375 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4377 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, dl, MVT::i64),
4378 ISD::SETGE, dl, DAG);
4379 SDValue CCVal = DAG.getConstant(AArch64CC::GE, dl, MVT::i32);
4381 SDValue FalseValLo = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4382 SDValue TrueValLo = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4384 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
4386 // AArch64 shifts larger than the register width are wrapped rather than
4387 // clamped, so we can't just emit "hi >> x".
4388 SDValue FalseValHi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4389 SDValue TrueValHi = Opc == ISD::SRA
4390 ? DAG.getNode(Opc, dl, VT, ShOpHi,
4391 DAG.getConstant(VTBits - 1, dl,
4393 : DAG.getConstant(0, dl, VT);
4395 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValHi, FalseValHi, CCVal, Cmp);
4397 SDValue Ops[2] = { Lo, Hi };
4398 return DAG.getMergeValues(Ops, dl);
4401 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4402 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4403 SDValue AArch64TargetLowering::LowerShiftLeftParts(SDValue Op,
4404 SelectionDAG &DAG) const {
4405 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4406 EVT VT = Op.getValueType();
4407 unsigned VTBits = VT.getSizeInBits();
4409 SDValue ShOpLo = Op.getOperand(0);
4410 SDValue ShOpHi = Op.getOperand(1);
4411 SDValue ShAmt = Op.getOperand(2);
4414 assert(Op.getOpcode() == ISD::SHL_PARTS);
4415 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4416 DAG.getConstant(VTBits, dl, MVT::i64), ShAmt);
4417 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4418 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4419 DAG.getConstant(VTBits, dl, MVT::i64));
4420 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4421 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4423 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4425 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, dl, MVT::i64),
4426 ISD::SETGE, dl, DAG);
4427 SDValue CCVal = DAG.getConstant(AArch64CC::GE, dl, MVT::i32);
4429 DAG.getNode(AArch64ISD::CSEL, dl, VT, Tmp3, FalseVal, CCVal, Cmp);
4431 // AArch64 shifts of larger than register sizes are wrapped rather than
4432 // clamped, so we can't just emit "lo << a" if a is too big.
4433 SDValue TrueValLo = DAG.getConstant(0, dl, VT);
4434 SDValue FalseValLo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4436 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
4438 SDValue Ops[2] = { Lo, Hi };
4439 return DAG.getMergeValues(Ops, dl);
4442 bool AArch64TargetLowering::isOffsetFoldingLegal(
4443 const GlobalAddressSDNode *GA) const {
4444 // The AArch64 target doesn't support folding offsets into global addresses.
4448 bool AArch64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4449 // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
4450 // FIXME: We should be able to handle f128 as well with a clever lowering.
4451 if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32))
4455 return AArch64_AM::getFP64Imm(Imm) != -1;
4456 else if (VT == MVT::f32)
4457 return AArch64_AM::getFP32Imm(Imm) != -1;
4461 //===----------------------------------------------------------------------===//
4462 // AArch64 Optimization Hooks
4463 //===----------------------------------------------------------------------===//
4465 //===----------------------------------------------------------------------===//
4466 // AArch64 Inline Assembly Support
4467 //===----------------------------------------------------------------------===//
4469 // Table of Constraints
4470 // TODO: This is the current set of constraints supported by ARM for the
4471 // compiler, not all of them may make sense, e.g. S may be difficult to support.
4473 // r - A general register
4474 // w - An FP/SIMD register of some size in the range v0-v31
4475 // x - An FP/SIMD register of some size in the range v0-v15
4476 // I - Constant that can be used with an ADD instruction
4477 // J - Constant that can be used with a SUB instruction
4478 // K - Constant that can be used with a 32-bit logical instruction
4479 // L - Constant that can be used with a 64-bit logical instruction
4480 // M - Constant that can be used as a 32-bit MOV immediate
4481 // N - Constant that can be used as a 64-bit MOV immediate
4482 // Q - A memory reference with base register and no offset
4483 // S - A symbolic address
4484 // Y - Floating point constant zero
4485 // Z - Integer constant zero
4487 // Note that general register operands will be output using their 64-bit x
4488 // register name, whatever the size of the variable, unless the asm operand
4489 // is prefixed by the %w modifier. Floating-point and SIMD register operands
4490 // will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
4493 /// getConstraintType - Given a constraint letter, return the type of
4494 /// constraint it is for this target.
4495 AArch64TargetLowering::ConstraintType
4496 AArch64TargetLowering::getConstraintType(StringRef Constraint) const {
4497 if (Constraint.size() == 1) {
4498 switch (Constraint[0]) {
4505 return C_RegisterClass;
4506 // An address with a single base register. Due to the way we
4507 // currently handle addresses it is the same as 'r'.
4512 return TargetLowering::getConstraintType(Constraint);
4515 /// Examine constraint type and operand type and determine a weight value.
4516 /// This object must already have been set up with the operand type
4517 /// and the current alternative constraint selected.
4518 TargetLowering::ConstraintWeight
4519 AArch64TargetLowering::getSingleConstraintMatchWeight(
4520 AsmOperandInfo &info, const char *constraint) const {
4521 ConstraintWeight weight = CW_Invalid;
4522 Value *CallOperandVal = info.CallOperandVal;
4523 // If we don't have a value, we can't do a match,
4524 // but allow it at the lowest weight.
4525 if (!CallOperandVal)
4527 Type *type = CallOperandVal->getType();
4528 // Look at the constraint type.
4529 switch (*constraint) {
4531 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
4535 if (type->isFloatingPointTy() || type->isVectorTy())
4536 weight = CW_Register;
4539 weight = CW_Constant;
4545 std::pair<unsigned, const TargetRegisterClass *>
4546 AArch64TargetLowering::getRegForInlineAsmConstraint(
4547 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
4548 if (Constraint.size() == 1) {
4549 switch (Constraint[0]) {
4551 if (VT.getSizeInBits() == 64)
4552 return std::make_pair(0U, &AArch64::GPR64commonRegClass);
4553 return std::make_pair(0U, &AArch64::GPR32commonRegClass);
4556 return std::make_pair(0U, &AArch64::FPR32RegClass);
4557 if (VT.getSizeInBits() == 64)
4558 return std::make_pair(0U, &AArch64::FPR64RegClass);
4559 if (VT.getSizeInBits() == 128)
4560 return std::make_pair(0U, &AArch64::FPR128RegClass);
4562 // The instructions that this constraint is designed for can
4563 // only take 128-bit registers so just use that regclass.
4565 if (VT.getSizeInBits() == 128)
4566 return std::make_pair(0U, &AArch64::FPR128_loRegClass);
4570 if (StringRef("{cc}").equals_lower(Constraint))
4571 return std::make_pair(unsigned(AArch64::NZCV), &AArch64::CCRRegClass);
4573 // Use the default implementation in TargetLowering to convert the register
4574 // constraint into a member of a register class.
4575 std::pair<unsigned, const TargetRegisterClass *> Res;
4576 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
4578 // Not found as a standard register?
4580 unsigned Size = Constraint.size();
4581 if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
4582 tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
4584 bool Failed = Constraint.slice(2, Size - 1).getAsInteger(10, RegNo);
4585 if (!Failed && RegNo >= 0 && RegNo <= 31) {
4586 // v0 - v31 are aliases of q0 - q31.
4587 // By default we'll emit v0-v31 for this unless there's a modifier where
4588 // we'll emit the correct register as well.
4589 Res.first = AArch64::FPR128RegClass.getRegister(RegNo);
4590 Res.second = &AArch64::FPR128RegClass;
4598 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4599 /// vector. If it is invalid, don't add anything to Ops.
4600 void AArch64TargetLowering::LowerAsmOperandForConstraint(
4601 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
4602 SelectionDAG &DAG) const {
4605 // Currently only support length 1 constraints.
4606 if (Constraint.length() != 1)
4609 char ConstraintLetter = Constraint[0];
4610 switch (ConstraintLetter) {
4614 // This set of constraints deal with valid constants for various instructions.
4615 // Validate and return a target constant for them if we can.
4617 // 'z' maps to xzr or wzr so it needs an input of 0.
4618 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4619 if (!C || C->getZExtValue() != 0)
4622 if (Op.getValueType() == MVT::i64)
4623 Result = DAG.getRegister(AArch64::XZR, MVT::i64);
4625 Result = DAG.getRegister(AArch64::WZR, MVT::i32);
4635 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4639 // Grab the value and do some validation.
4640 uint64_t CVal = C->getZExtValue();
4641 switch (ConstraintLetter) {
4642 // The I constraint applies only to simple ADD or SUB immediate operands:
4643 // i.e. 0 to 4095 with optional shift by 12
4644 // The J constraint applies only to ADD or SUB immediates that would be
4645 // valid when negated, i.e. if [an add pattern] were to be output as a SUB
4646 // instruction [or vice versa], in other words -1 to -4095 with optional
4647 // left shift by 12.
4649 if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
4653 uint64_t NVal = -C->getSExtValue();
4654 if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal)) {
4655 CVal = C->getSExtValue();
4660 // The K and L constraints apply *only* to logical immediates, including
4661 // what used to be the MOVI alias for ORR (though the MOVI alias has now
4662 // been removed and MOV should be used). So these constraints have to
4663 // distinguish between bit patterns that are valid 32-bit or 64-bit
4664 // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
4665 // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
4668 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4672 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4675 // The M and N constraints are a superset of K and L respectively, for use
4676 // with the MOV (immediate) alias. As well as the logical immediates they
4677 // also match 32 or 64-bit immediates that can be loaded either using a
4678 // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
4679 // (M) or 64-bit 0x1234000000000000 (N) etc.
4680 // As a note some of this code is liberally stolen from the asm parser.
4682 if (!isUInt<32>(CVal))
4684 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4686 if ((CVal & 0xFFFF) == CVal)
4688 if ((CVal & 0xFFFF0000ULL) == CVal)
4690 uint64_t NCVal = ~(uint32_t)CVal;
4691 if ((NCVal & 0xFFFFULL) == NCVal)
4693 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4698 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4700 if ((CVal & 0xFFFFULL) == CVal)
4702 if ((CVal & 0xFFFF0000ULL) == CVal)
4704 if ((CVal & 0xFFFF00000000ULL) == CVal)
4706 if ((CVal & 0xFFFF000000000000ULL) == CVal)
4708 uint64_t NCVal = ~CVal;
4709 if ((NCVal & 0xFFFFULL) == NCVal)
4711 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4713 if ((NCVal & 0xFFFF00000000ULL) == NCVal)
4715 if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
4723 // All assembler immediates are 64-bit integers.
4724 Result = DAG.getTargetConstant(CVal, SDLoc(Op), MVT::i64);
4728 if (Result.getNode()) {
4729 Ops.push_back(Result);
4733 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
4736 //===----------------------------------------------------------------------===//
4737 // AArch64 Advanced SIMD Support
4738 //===----------------------------------------------------------------------===//
4740 /// WidenVector - Given a value in the V64 register class, produce the
4741 /// equivalent value in the V128 register class.
4742 static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
4743 EVT VT = V64Reg.getValueType();
4744 unsigned NarrowSize = VT.getVectorNumElements();
4745 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4746 MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
4749 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
4750 V64Reg, DAG.getConstant(0, DL, MVT::i32));
4753 /// getExtFactor - Determine the adjustment factor for the position when
4754 /// generating an "extract from vector registers" instruction.
4755 static unsigned getExtFactor(SDValue &V) {
4756 EVT EltType = V.getValueType().getVectorElementType();
4757 return EltType.getSizeInBits() / 8;
4760 /// NarrowVector - Given a value in the V128 register class, produce the
4761 /// equivalent value in the V64 register class.
4762 static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) {
4763 EVT VT = V128Reg.getValueType();
4764 unsigned WideSize = VT.getVectorNumElements();
4765 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4766 MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2);
4769 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, NarrowTy, V128Reg);
4772 // Gather data to see if the operation can be modelled as a
4773 // shuffle in combination with VEXTs.
4774 SDValue AArch64TargetLowering::ReconstructShuffle(SDValue Op,
4775 SelectionDAG &DAG) const {
4776 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
4778 EVT VT = Op.getValueType();
4779 unsigned NumElts = VT.getVectorNumElements();
4781 struct ShuffleSourceInfo {
4786 // We may insert some combination of BITCASTs and VEXT nodes to force Vec to
4787 // be compatible with the shuffle we intend to construct. As a result
4788 // ShuffleVec will be some sliding window into the original Vec.
4791 // Code should guarantee that element i in Vec starts at element "WindowBase
4792 // + i * WindowScale in ShuffleVec".
4796 bool operator ==(SDValue OtherVec) { return Vec == OtherVec; }
4797 ShuffleSourceInfo(SDValue Vec)
4798 : Vec(Vec), MinElt(UINT_MAX), MaxElt(0), ShuffleVec(Vec), WindowBase(0),
4802 // First gather all vectors used as an immediate source for this BUILD_VECTOR
4804 SmallVector<ShuffleSourceInfo, 2> Sources;
4805 for (unsigned i = 0; i < NumElts; ++i) {
4806 SDValue V = Op.getOperand(i);
4807 if (V.getOpcode() == ISD::UNDEF)
4809 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4810 // A shuffle can only come from building a vector from various
4811 // elements of other vectors.
4815 // Add this element source to the list if it's not already there.
4816 SDValue SourceVec = V.getOperand(0);
4817 auto Source = std::find(Sources.begin(), Sources.end(), SourceVec);
4818 if (Source == Sources.end())
4819 Source = Sources.insert(Sources.end(), ShuffleSourceInfo(SourceVec));
4821 // Update the minimum and maximum lane number seen.
4822 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4823 Source->MinElt = std::min(Source->MinElt, EltNo);
4824 Source->MaxElt = std::max(Source->MaxElt, EltNo);
4827 // Currently only do something sane when at most two source vectors
4829 if (Sources.size() > 2)
4832 // Find out the smallest element size among result and two sources, and use
4833 // it as element size to build the shuffle_vector.
4834 EVT SmallestEltTy = VT.getVectorElementType();
4835 for (auto &Source : Sources) {
4836 EVT SrcEltTy = Source.Vec.getValueType().getVectorElementType();
4837 if (SrcEltTy.bitsLT(SmallestEltTy)) {
4838 SmallestEltTy = SrcEltTy;
4841 unsigned ResMultiplier =
4842 VT.getVectorElementType().getSizeInBits() / SmallestEltTy.getSizeInBits();
4843 NumElts = VT.getSizeInBits() / SmallestEltTy.getSizeInBits();
4844 EVT ShuffleVT = EVT::getVectorVT(*DAG.getContext(), SmallestEltTy, NumElts);
4846 // If the source vector is too wide or too narrow, we may nevertheless be able
4847 // to construct a compatible shuffle either by concatenating it with UNDEF or
4848 // extracting a suitable range of elements.
4849 for (auto &Src : Sources) {
4850 EVT SrcVT = Src.ShuffleVec.getValueType();
4852 if (SrcVT.getSizeInBits() == VT.getSizeInBits())
4855 // This stage of the search produces a source with the same element type as
4856 // the original, but with a total width matching the BUILD_VECTOR output.
4857 EVT EltVT = SrcVT.getVectorElementType();
4858 unsigned NumSrcElts = VT.getSizeInBits() / EltVT.getSizeInBits();
4859 EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumSrcElts);
4861 if (SrcVT.getSizeInBits() < VT.getSizeInBits()) {
4862 assert(2 * SrcVT.getSizeInBits() == VT.getSizeInBits());
4863 // We can pad out the smaller vector for free, so if it's part of a
4866 DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, Src.ShuffleVec,
4867 DAG.getUNDEF(Src.ShuffleVec.getValueType()));
4871 assert(SrcVT.getSizeInBits() == 2 * VT.getSizeInBits());
4873 if (Src.MaxElt - Src.MinElt >= NumSrcElts) {
4874 // Span too large for a VEXT to cope
4878 if (Src.MinElt >= NumSrcElts) {
4879 // The extraction can just take the second half
4881 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4882 DAG.getConstant(NumSrcElts, dl, MVT::i64));
4883 Src.WindowBase = -NumSrcElts;
4884 } else if (Src.MaxElt < NumSrcElts) {
4885 // The extraction can just take the first half
4887 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4888 DAG.getConstant(0, dl, MVT::i64));
4890 // An actual VEXT is needed
4892 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4893 DAG.getConstant(0, dl, MVT::i64));
4895 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4896 DAG.getConstant(NumSrcElts, dl, MVT::i64));
4897 unsigned Imm = Src.MinElt * getExtFactor(VEXTSrc1);
4899 Src.ShuffleVec = DAG.getNode(AArch64ISD::EXT, dl, DestVT, VEXTSrc1,
4901 DAG.getConstant(Imm, dl, MVT::i32));
4902 Src.WindowBase = -Src.MinElt;
4906 // Another possible incompatibility occurs from the vector element types. We
4907 // can fix this by bitcasting the source vectors to the same type we intend
4909 for (auto &Src : Sources) {
4910 EVT SrcEltTy = Src.ShuffleVec.getValueType().getVectorElementType();
4911 if (SrcEltTy == SmallestEltTy)
4913 assert(ShuffleVT.getVectorElementType() == SmallestEltTy);
4914 Src.ShuffleVec = DAG.getNode(ISD::BITCAST, dl, ShuffleVT, Src.ShuffleVec);
4915 Src.WindowScale = SrcEltTy.getSizeInBits() / SmallestEltTy.getSizeInBits();
4916 Src.WindowBase *= Src.WindowScale;
4919 // Final sanity check before we try to actually produce a shuffle.
4921 for (auto Src : Sources)
4922 assert(Src.ShuffleVec.getValueType() == ShuffleVT);
4925 // The stars all align, our next step is to produce the mask for the shuffle.
4926 SmallVector<int, 8> Mask(ShuffleVT.getVectorNumElements(), -1);
4927 int BitsPerShuffleLane = ShuffleVT.getVectorElementType().getSizeInBits();
4928 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
4929 SDValue Entry = Op.getOperand(i);
4930 if (Entry.getOpcode() == ISD::UNDEF)
4933 auto Src = std::find(Sources.begin(), Sources.end(), Entry.getOperand(0));
4934 int EltNo = cast<ConstantSDNode>(Entry.getOperand(1))->getSExtValue();
4936 // EXTRACT_VECTOR_ELT performs an implicit any_ext; BUILD_VECTOR an implicit
4937 // trunc. So only std::min(SrcBits, DestBits) actually get defined in this
4939 EVT OrigEltTy = Entry.getOperand(0).getValueType().getVectorElementType();
4940 int BitsDefined = std::min(OrigEltTy.getSizeInBits(),
4941 VT.getVectorElementType().getSizeInBits());
4942 int LanesDefined = BitsDefined / BitsPerShuffleLane;
4944 // This source is expected to fill ResMultiplier lanes of the final shuffle,
4945 // starting at the appropriate offset.
4946 int *LaneMask = &Mask[i * ResMultiplier];
4948 int ExtractBase = EltNo * Src->WindowScale + Src->WindowBase;
4949 ExtractBase += NumElts * (Src - Sources.begin());
4950 for (int j = 0; j < LanesDefined; ++j)
4951 LaneMask[j] = ExtractBase + j;
4954 // Final check before we try to produce nonsense...
4955 if (!isShuffleMaskLegal(Mask, ShuffleVT))
4958 SDValue ShuffleOps[] = { DAG.getUNDEF(ShuffleVT), DAG.getUNDEF(ShuffleVT) };
4959 for (unsigned i = 0; i < Sources.size(); ++i)
4960 ShuffleOps[i] = Sources[i].ShuffleVec;
4962 SDValue Shuffle = DAG.getVectorShuffle(ShuffleVT, dl, ShuffleOps[0],
4963 ShuffleOps[1], &Mask[0]);
4964 return DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
4967 // check if an EXT instruction can handle the shuffle mask when the
4968 // vector sources of the shuffle are the same.
4969 static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4970 unsigned NumElts = VT.getVectorNumElements();
4972 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4978 // If this is a VEXT shuffle, the immediate value is the index of the first
4979 // element. The other shuffle indices must be the successive elements after
4981 unsigned ExpectedElt = Imm;
4982 for (unsigned i = 1; i < NumElts; ++i) {
4983 // Increment the expected index. If it wraps around, just follow it
4984 // back to index zero and keep going.
4986 if (ExpectedElt == NumElts)
4990 continue; // ignore UNDEF indices
4991 if (ExpectedElt != static_cast<unsigned>(M[i]))
4998 // check if an EXT instruction can handle the shuffle mask when the
4999 // vector sources of the shuffle are different.
5000 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
5002 // Look for the first non-undef element.
5003 const int *FirstRealElt = std::find_if(M.begin(), M.end(),
5004 [](int Elt) {return Elt >= 0;});
5006 // Benefit form APInt to handle overflow when calculating expected element.
5007 unsigned NumElts = VT.getVectorNumElements();
5008 unsigned MaskBits = APInt(32, NumElts * 2).logBase2();
5009 APInt ExpectedElt = APInt(MaskBits, *FirstRealElt + 1);
5010 // The following shuffle indices must be the successive elements after the
5011 // first real element.
5012 const int *FirstWrongElt = std::find_if(FirstRealElt + 1, M.end(),
5013 [&](int Elt) {return Elt != ExpectedElt++ && Elt != -1;});
5014 if (FirstWrongElt != M.end())
5017 // The index of an EXT is the first element if it is not UNDEF.
5018 // Watch out for the beginning UNDEFs. The EXT index should be the expected
5019 // value of the first element. E.g.
5020 // <-1, -1, 3, ...> is treated as <1, 2, 3, ...>.
5021 // <-1, -1, 0, 1, ...> is treated as <2*NumElts-2, 2*NumElts-1, 0, 1, ...>.
5022 // ExpectedElt is the last mask index plus 1.
5023 Imm = ExpectedElt.getZExtValue();
5025 // There are two difference cases requiring to reverse input vectors.
5026 // For example, for vector <4 x i32> we have the following cases,
5027 // Case 1: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, -1, 0>)
5028 // Case 2: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, 7, 0>)
5029 // For both cases, we finally use mask <5, 6, 7, 0>, which requires
5030 // to reverse two input vectors.
5039 /// isREVMask - Check if a vector shuffle corresponds to a REV
5040 /// instruction with the specified blocksize. (The order of the elements
5041 /// within each block of the vector is reversed.)
5042 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
5043 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
5044 "Only possible block sizes for REV are: 16, 32, 64");
5046 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5050 unsigned NumElts = VT.getVectorNumElements();
5051 unsigned BlockElts = M[0] + 1;
5052 // If the first shuffle index is UNDEF, be optimistic.
5054 BlockElts = BlockSize / EltSz;
5056 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
5059 for (unsigned i = 0; i < NumElts; ++i) {
5061 continue; // ignore UNDEF indices
5062 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
5069 static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5070 unsigned NumElts = VT.getVectorNumElements();
5071 WhichResult = (M[0] == 0 ? 0 : 1);
5072 unsigned Idx = WhichResult * NumElts / 2;
5073 for (unsigned i = 0; i != NumElts; i += 2) {
5074 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
5075 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
5083 static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5084 unsigned NumElts = VT.getVectorNumElements();
5085 WhichResult = (M[0] == 0 ? 0 : 1);
5086 for (unsigned i = 0; i != NumElts; ++i) {
5088 continue; // ignore UNDEF indices
5089 if ((unsigned)M[i] != 2 * i + WhichResult)
5096 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5097 unsigned NumElts = VT.getVectorNumElements();
5098 WhichResult = (M[0] == 0 ? 0 : 1);
5099 for (unsigned i = 0; i < NumElts; i += 2) {
5100 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
5101 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + NumElts + WhichResult))
5107 /// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of
5108 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5109 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
5110 static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5111 unsigned NumElts = VT.getVectorNumElements();
5112 WhichResult = (M[0] == 0 ? 0 : 1);
5113 unsigned Idx = WhichResult * NumElts / 2;
5114 for (unsigned i = 0; i != NumElts; i += 2) {
5115 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
5116 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx))
5124 /// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of
5125 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5126 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
5127 static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5128 unsigned Half = VT.getVectorNumElements() / 2;
5129 WhichResult = (M[0] == 0 ? 0 : 1);
5130 for (unsigned j = 0; j != 2; ++j) {
5131 unsigned Idx = WhichResult;
5132 for (unsigned i = 0; i != Half; ++i) {
5133 int MIdx = M[i + j * Half];
5134 if (MIdx >= 0 && (unsigned)MIdx != Idx)
5143 /// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of
5144 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5145 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
5146 static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5147 unsigned NumElts = VT.getVectorNumElements();
5148 WhichResult = (M[0] == 0 ? 0 : 1);
5149 for (unsigned i = 0; i < NumElts; i += 2) {
5150 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
5151 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult))
5157 static bool isINSMask(ArrayRef<int> M, int NumInputElements,
5158 bool &DstIsLeft, int &Anomaly) {
5159 if (M.size() != static_cast<size_t>(NumInputElements))
5162 int NumLHSMatch = 0, NumRHSMatch = 0;
5163 int LastLHSMismatch = -1, LastRHSMismatch = -1;
5165 for (int i = 0; i < NumInputElements; ++i) {
5175 LastLHSMismatch = i;
5177 if (M[i] == i + NumInputElements)
5180 LastRHSMismatch = i;
5183 if (NumLHSMatch == NumInputElements - 1) {
5185 Anomaly = LastLHSMismatch;
5187 } else if (NumRHSMatch == NumInputElements - 1) {
5189 Anomaly = LastRHSMismatch;
5196 static bool isConcatMask(ArrayRef<int> Mask, EVT VT, bool SplitLHS) {
5197 if (VT.getSizeInBits() != 128)
5200 unsigned NumElts = VT.getVectorNumElements();
5202 for (int I = 0, E = NumElts / 2; I != E; I++) {
5207 int Offset = NumElts / 2;
5208 for (int I = NumElts / 2, E = NumElts; I != E; I++) {
5209 if (Mask[I] != I + SplitLHS * Offset)
5216 static SDValue tryFormConcatFromShuffle(SDValue Op, SelectionDAG &DAG) {
5218 EVT VT = Op.getValueType();
5219 SDValue V0 = Op.getOperand(0);
5220 SDValue V1 = Op.getOperand(1);
5221 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op)->getMask();
5223 if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() ||
5224 VT.getVectorElementType() != V1.getValueType().getVectorElementType())
5227 bool SplitV0 = V0.getValueType().getSizeInBits() == 128;
5229 if (!isConcatMask(Mask, VT, SplitV0))
5232 EVT CastVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
5233 VT.getVectorNumElements() / 2);
5235 V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0,
5236 DAG.getConstant(0, DL, MVT::i64));
5238 if (V1.getValueType().getSizeInBits() == 128) {
5239 V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1,
5240 DAG.getConstant(0, DL, MVT::i64));
5242 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1);
5245 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5246 /// the specified operations to build the shuffle.
5247 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5248 SDValue RHS, SelectionDAG &DAG,
5250 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5251 unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1);
5252 unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1);
5255 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5264 OP_VUZPL, // VUZP, left result
5265 OP_VUZPR, // VUZP, right result
5266 OP_VZIPL, // VZIP, left result
5267 OP_VZIPR, // VZIP, right result
5268 OP_VTRNL, // VTRN, left result
5269 OP_VTRNR // VTRN, right result
5272 if (OpNum == OP_COPY) {
5273 if (LHSID == (1 * 9 + 2) * 9 + 3)
5275 assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!");
5279 SDValue OpLHS, OpRHS;
5280 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5281 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5282 EVT VT = OpLHS.getValueType();
5286 llvm_unreachable("Unknown shuffle opcode!");
5288 // VREV divides the vector in half and swaps within the half.
5289 if (VT.getVectorElementType() == MVT::i32 ||
5290 VT.getVectorElementType() == MVT::f32)
5291 return DAG.getNode(AArch64ISD::REV64, dl, VT, OpLHS);
5292 // vrev <4 x i16> -> REV32
5293 if (VT.getVectorElementType() == MVT::i16 ||
5294 VT.getVectorElementType() == MVT::f16)
5295 return DAG.getNode(AArch64ISD::REV32, dl, VT, OpLHS);
5296 // vrev <4 x i8> -> REV16
5297 assert(VT.getVectorElementType() == MVT::i8);
5298 return DAG.getNode(AArch64ISD::REV16, dl, VT, OpLHS);
5303 EVT EltTy = VT.getVectorElementType();
5305 if (EltTy == MVT::i8)
5306 Opcode = AArch64ISD::DUPLANE8;
5307 else if (EltTy == MVT::i16 || EltTy == MVT::f16)
5308 Opcode = AArch64ISD::DUPLANE16;
5309 else if (EltTy == MVT::i32 || EltTy == MVT::f32)
5310 Opcode = AArch64ISD::DUPLANE32;
5311 else if (EltTy == MVT::i64 || EltTy == MVT::f64)
5312 Opcode = AArch64ISD::DUPLANE64;
5314 llvm_unreachable("Invalid vector element type?");
5316 if (VT.getSizeInBits() == 64)
5317 OpLHS = WidenVector(OpLHS, DAG);
5318 SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, dl, MVT::i64);
5319 return DAG.getNode(Opcode, dl, VT, OpLHS, Lane);
5324 unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
5325 return DAG.getNode(AArch64ISD::EXT, dl, VT, OpLHS, OpRHS,
5326 DAG.getConstant(Imm, dl, MVT::i32));
5329 return DAG.getNode(AArch64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS,
5332 return DAG.getNode(AArch64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS,
5335 return DAG.getNode(AArch64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS,
5338 return DAG.getNode(AArch64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS,
5341 return DAG.getNode(AArch64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS,
5344 return DAG.getNode(AArch64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS,
5349 static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask,
5350 SelectionDAG &DAG) {
5351 // Check to see if we can use the TBL instruction.
5352 SDValue V1 = Op.getOperand(0);
5353 SDValue V2 = Op.getOperand(1);
5356 EVT EltVT = Op.getValueType().getVectorElementType();
5357 unsigned BytesPerElt = EltVT.getSizeInBits() / 8;
5359 SmallVector<SDValue, 8> TBLMask;
5360 for (int Val : ShuffleMask) {
5361 for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
5362 unsigned Offset = Byte + Val * BytesPerElt;
5363 TBLMask.push_back(DAG.getConstant(Offset, DL, MVT::i32));
5367 MVT IndexVT = MVT::v8i8;
5368 unsigned IndexLen = 8;
5369 if (Op.getValueType().getSizeInBits() == 128) {
5370 IndexVT = MVT::v16i8;
5374 SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
5375 SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);
5378 if (V2.getNode()->getOpcode() == ISD::UNDEF) {
5380 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
5381 Shuffle = DAG.getNode(
5382 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5383 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, DL, MVT::i32), V1Cst,
5384 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5385 makeArrayRef(TBLMask.data(), IndexLen)));
5387 if (IndexLen == 8) {
5388 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
5389 Shuffle = DAG.getNode(
5390 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5391 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, DL, MVT::i32), V1Cst,
5392 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5393 makeArrayRef(TBLMask.data(), IndexLen)));
5395 // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
5396 // cannot currently represent the register constraints on the input
5398 // Shuffle = DAG.getNode(AArch64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
5399 // DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5400 // &TBLMask[0], IndexLen));
5401 Shuffle = DAG.getNode(
5402 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5403 DAG.getConstant(Intrinsic::aarch64_neon_tbl2, DL, MVT::i32),
5405 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5406 makeArrayRef(TBLMask.data(), IndexLen)));
5409 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
5412 static unsigned getDUPLANEOp(EVT EltType) {
5413 if (EltType == MVT::i8)
5414 return AArch64ISD::DUPLANE8;
5415 if (EltType == MVT::i16 || EltType == MVT::f16)
5416 return AArch64ISD::DUPLANE16;
5417 if (EltType == MVT::i32 || EltType == MVT::f32)
5418 return AArch64ISD::DUPLANE32;
5419 if (EltType == MVT::i64 || EltType == MVT::f64)
5420 return AArch64ISD::DUPLANE64;
5422 llvm_unreachable("Invalid vector element type?");
5425 SDValue AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
5426 SelectionDAG &DAG) const {
5428 EVT VT = Op.getValueType();
5430 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5432 // Convert shuffles that are directly supported on NEON to target-specific
5433 // DAG nodes, instead of keeping them as shuffles and matching them again
5434 // during code selection. This is more efficient and avoids the possibility
5435 // of inconsistencies between legalization and selection.
5436 ArrayRef<int> ShuffleMask = SVN->getMask();
5438 SDValue V1 = Op.getOperand(0);
5439 SDValue V2 = Op.getOperand(1);
5441 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0],
5442 V1.getValueType().getSimpleVT())) {
5443 int Lane = SVN->getSplatIndex();
5444 // If this is undef splat, generate it via "just" vdup, if possible.
5448 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
5449 return DAG.getNode(AArch64ISD::DUP, dl, V1.getValueType(),
5451 // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
5452 // constant. If so, we can just reference the lane's definition directly.
5453 if (V1.getOpcode() == ISD::BUILD_VECTOR &&
5454 !isa<ConstantSDNode>(V1.getOperand(Lane)))
5455 return DAG.getNode(AArch64ISD::DUP, dl, VT, V1.getOperand(Lane));
5457 // Otherwise, duplicate from the lane of the input vector.
5458 unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
5460 // SelectionDAGBuilder may have "helpfully" already extracted or conatenated
5461 // to make a vector of the same size as this SHUFFLE. We can ignore the
5462 // extract entirely, and canonicalise the concat using WidenVector.
5463 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
5464 Lane += cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
5465 V1 = V1.getOperand(0);
5466 } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
5467 unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
5468 Lane -= Idx * VT.getVectorNumElements() / 2;
5469 V1 = WidenVector(V1.getOperand(Idx), DAG);
5470 } else if (VT.getSizeInBits() == 64)
5471 V1 = WidenVector(V1, DAG);
5473 return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, dl, MVT::i64));
5476 if (isREVMask(ShuffleMask, VT, 64))
5477 return DAG.getNode(AArch64ISD::REV64, dl, V1.getValueType(), V1, V2);
5478 if (isREVMask(ShuffleMask, VT, 32))
5479 return DAG.getNode(AArch64ISD::REV32, dl, V1.getValueType(), V1, V2);
5480 if (isREVMask(ShuffleMask, VT, 16))
5481 return DAG.getNode(AArch64ISD::REV16, dl, V1.getValueType(), V1, V2);
5483 bool ReverseEXT = false;
5485 if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
5488 Imm *= getExtFactor(V1);
5489 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V2,
5490 DAG.getConstant(Imm, dl, MVT::i32));
5491 } else if (V2->getOpcode() == ISD::UNDEF &&
5492 isSingletonEXTMask(ShuffleMask, VT, Imm)) {
5493 Imm *= getExtFactor(V1);
5494 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V1,
5495 DAG.getConstant(Imm, dl, MVT::i32));
5498 unsigned WhichResult;
5499 if (isZIPMask(ShuffleMask, VT, WhichResult)) {
5500 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
5501 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5503 if (isUZPMask(ShuffleMask, VT, WhichResult)) {
5504 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
5505 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5507 if (isTRNMask(ShuffleMask, VT, WhichResult)) {
5508 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
5509 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5512 if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5513 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
5514 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5516 if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5517 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
5518 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5520 if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5521 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
5522 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5525 SDValue Concat = tryFormConcatFromShuffle(Op, DAG);
5526 if (Concat.getNode())
5531 int NumInputElements = V1.getValueType().getVectorNumElements();
5532 if (isINSMask(ShuffleMask, NumInputElements, DstIsLeft, Anomaly)) {
5533 SDValue DstVec = DstIsLeft ? V1 : V2;
5534 SDValue DstLaneV = DAG.getConstant(Anomaly, dl, MVT::i64);
5536 SDValue SrcVec = V1;
5537 int SrcLane = ShuffleMask[Anomaly];
5538 if (SrcLane >= NumInputElements) {
5540 SrcLane -= VT.getVectorNumElements();
5542 SDValue SrcLaneV = DAG.getConstant(SrcLane, dl, MVT::i64);
5544 EVT ScalarVT = VT.getVectorElementType();
5546 if (ScalarVT.getSizeInBits() < 32 && ScalarVT.isInteger())
5547 ScalarVT = MVT::i32;
5550 ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5551 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ScalarVT, SrcVec, SrcLaneV),
5555 // If the shuffle is not directly supported and it has 4 elements, use
5556 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5557 unsigned NumElts = VT.getVectorNumElements();
5559 unsigned PFIndexes[4];
5560 for (unsigned i = 0; i != 4; ++i) {
5561 if (ShuffleMask[i] < 0)
5564 PFIndexes[i] = ShuffleMask[i];
5567 // Compute the index in the perfect shuffle table.
5568 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
5569 PFIndexes[2] * 9 + PFIndexes[3];
5570 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5571 unsigned Cost = (PFEntry >> 30);
5574 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5577 return GenerateTBL(Op, ShuffleMask, DAG);
5580 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
5582 EVT VT = BVN->getValueType(0);
5583 APInt SplatBits, SplatUndef;
5584 unsigned SplatBitSize;
5586 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5587 unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;
5589 for (unsigned i = 0; i < NumSplats; ++i) {
5590 CnstBits <<= SplatBitSize;
5591 UndefBits <<= SplatBitSize;
5592 CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits());
5593 UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits());
5602 SDValue AArch64TargetLowering::LowerVectorAND(SDValue Op,
5603 SelectionDAG &DAG) const {
5604 BuildVectorSDNode *BVN =
5605 dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5606 SDValue LHS = Op.getOperand(0);
5608 EVT VT = Op.getValueType();
5613 APInt CnstBits(VT.getSizeInBits(), 0);
5614 APInt UndefBits(VT.getSizeInBits(), 0);
5615 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5616 // We only have BIC vector immediate instruction, which is and-not.
5617 CnstBits = ~CnstBits;
5619 // We make use of a little bit of goto ickiness in order to avoid having to
5620 // duplicate the immediate matching logic for the undef toggled case.
5621 bool SecondTry = false;
5624 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5625 CnstBits = CnstBits.zextOrTrunc(64);
5626 uint64_t CnstVal = CnstBits.getZExtValue();
5628 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5629 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5630 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5631 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5632 DAG.getConstant(CnstVal, dl, MVT::i32),
5633 DAG.getConstant(0, dl, MVT::i32));
5634 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5637 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5638 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5639 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5640 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5641 DAG.getConstant(CnstVal, dl, MVT::i32),
5642 DAG.getConstant(8, dl, MVT::i32));
5643 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5646 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5647 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5648 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5649 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5650 DAG.getConstant(CnstVal, dl, MVT::i32),
5651 DAG.getConstant(16, dl, MVT::i32));
5652 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5655 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5656 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5657 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5658 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5659 DAG.getConstant(CnstVal, dl, MVT::i32),
5660 DAG.getConstant(24, dl, MVT::i32));
5661 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5664 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5665 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5666 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5667 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5668 DAG.getConstant(CnstVal, dl, MVT::i32),
5669 DAG.getConstant(0, dl, MVT::i32));
5670 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5673 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5674 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5675 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5676 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5677 DAG.getConstant(CnstVal, dl, MVT::i32),
5678 DAG.getConstant(8, dl, MVT::i32));
5679 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5686 CnstBits = ~UndefBits;
5690 // We can always fall back to a non-immediate AND.
5695 // Specialized code to quickly find if PotentialBVec is a BuildVector that
5696 // consists of only the same constant int value, returned in reference arg
5698 static bool isAllConstantBuildVector(const SDValue &PotentialBVec,
5699 uint64_t &ConstVal) {
5700 BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec);
5703 ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0));
5706 EVT VT = Bvec->getValueType(0);
5707 unsigned NumElts = VT.getVectorNumElements();
5708 for (unsigned i = 1; i < NumElts; ++i)
5709 if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt)
5711 ConstVal = FirstElt->getZExtValue();
5715 static unsigned getIntrinsicID(const SDNode *N) {
5716 unsigned Opcode = N->getOpcode();
5719 return Intrinsic::not_intrinsic;
5720 case ISD::INTRINSIC_WO_CHAIN: {
5721 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
5722 if (IID < Intrinsic::num_intrinsics)
5724 return Intrinsic::not_intrinsic;
5729 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
5730 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
5731 // BUILD_VECTORs with constant element C1, C2 is a constant, and C1 == ~C2.
5732 // Also, logical shift right -> sri, with the same structure.
5733 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
5734 EVT VT = N->getValueType(0);
5741 // Is the first op an AND?
5742 const SDValue And = N->getOperand(0);
5743 if (And.getOpcode() != ISD::AND)
5746 // Is the second op an shl or lshr?
5747 SDValue Shift = N->getOperand(1);
5748 // This will have been turned into: AArch64ISD::VSHL vector, #shift
5749 // or AArch64ISD::VLSHR vector, #shift
5750 unsigned ShiftOpc = Shift.getOpcode();
5751 if ((ShiftOpc != AArch64ISD::VSHL && ShiftOpc != AArch64ISD::VLSHR))
5753 bool IsShiftRight = ShiftOpc == AArch64ISD::VLSHR;
5755 // Is the shift amount constant?
5756 ConstantSDNode *C2node = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
5760 // Is the and mask vector all constant?
5762 if (!isAllConstantBuildVector(And.getOperand(1), C1))
5765 // Is C1 == ~C2, taking into account how much one can shift elements of a
5767 uint64_t C2 = C2node->getZExtValue();
5768 unsigned ElemSizeInBits = VT.getVectorElementType().getSizeInBits();
5769 if (C2 > ElemSizeInBits)
5771 unsigned ElemMask = (1 << ElemSizeInBits) - 1;
5772 if ((C1 & ElemMask) != (~C2 & ElemMask))
5775 SDValue X = And.getOperand(0);
5776 SDValue Y = Shift.getOperand(0);
5779 IsShiftRight ? Intrinsic::aarch64_neon_vsri : Intrinsic::aarch64_neon_vsli;
5781 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5782 DAG.getConstant(Intrin, DL, MVT::i32), X, Y,
5783 Shift.getOperand(1));
5785 DEBUG(dbgs() << "aarch64-lower: transformed: \n");
5786 DEBUG(N->dump(&DAG));
5787 DEBUG(dbgs() << "into: \n");
5788 DEBUG(ResultSLI->dump(&DAG));
5794 SDValue AArch64TargetLowering::LowerVectorOR(SDValue Op,
5795 SelectionDAG &DAG) const {
5796 // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
5797 if (EnableAArch64SlrGeneration) {
5798 SDValue Res = tryLowerToSLI(Op.getNode(), DAG);
5803 BuildVectorSDNode *BVN =
5804 dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode());
5805 SDValue LHS = Op.getOperand(1);
5807 EVT VT = Op.getValueType();
5809 // OR commutes, so try swapping the operands.
5811 LHS = Op.getOperand(0);
5812 BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5817 APInt CnstBits(VT.getSizeInBits(), 0);
5818 APInt UndefBits(VT.getSizeInBits(), 0);
5819 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5820 // We make use of a little bit of goto ickiness in order to avoid having to
5821 // duplicate the immediate matching logic for the undef toggled case.
5822 bool SecondTry = false;
5825 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5826 CnstBits = CnstBits.zextOrTrunc(64);
5827 uint64_t CnstVal = CnstBits.getZExtValue();
5829 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5830 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5831 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5832 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5833 DAG.getConstant(CnstVal, dl, MVT::i32),
5834 DAG.getConstant(0, dl, MVT::i32));
5835 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5838 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5839 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5840 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5841 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5842 DAG.getConstant(CnstVal, dl, MVT::i32),
5843 DAG.getConstant(8, dl, MVT::i32));
5844 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5847 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5848 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5849 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5850 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5851 DAG.getConstant(CnstVal, dl, MVT::i32),
5852 DAG.getConstant(16, dl, MVT::i32));
5853 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5856 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5857 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5858 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5859 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5860 DAG.getConstant(CnstVal, dl, MVT::i32),
5861 DAG.getConstant(24, dl, MVT::i32));
5862 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5865 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5866 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5867 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5868 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5869 DAG.getConstant(CnstVal, dl, MVT::i32),
5870 DAG.getConstant(0, dl, MVT::i32));
5871 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5874 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5875 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5876 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5877 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5878 DAG.getConstant(CnstVal, dl, MVT::i32),
5879 DAG.getConstant(8, dl, MVT::i32));
5880 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5887 CnstBits = UndefBits;
5891 // We can always fall back to a non-immediate OR.
5896 // Normalize the operands of BUILD_VECTOR. The value of constant operands will
5897 // be truncated to fit element width.
5898 static SDValue NormalizeBuildVector(SDValue Op,
5899 SelectionDAG &DAG) {
5900 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
5902 EVT VT = Op.getValueType();
5903 EVT EltTy= VT.getVectorElementType();
5905 if (EltTy.isFloatingPoint() || EltTy.getSizeInBits() > 16)
5908 SmallVector<SDValue, 16> Ops;
5909 for (SDValue Lane : Op->ops()) {
5910 if (auto *CstLane = dyn_cast<ConstantSDNode>(Lane)) {
5911 APInt LowBits(EltTy.getSizeInBits(),
5912 CstLane->getZExtValue());
5913 Lane = DAG.getConstant(LowBits.getZExtValue(), dl, MVT::i32);
5915 Ops.push_back(Lane);
5917 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5920 SDValue AArch64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
5921 SelectionDAG &DAG) const {
5923 EVT VT = Op.getValueType();
5924 Op = NormalizeBuildVector(Op, DAG);
5925 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5927 APInt CnstBits(VT.getSizeInBits(), 0);
5928 APInt UndefBits(VT.getSizeInBits(), 0);
5929 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5930 // We make use of a little bit of goto ickiness in order to avoid having to
5931 // duplicate the immediate matching logic for the undef toggled case.
5932 bool SecondTry = false;
5935 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5936 CnstBits = CnstBits.zextOrTrunc(64);
5937 uint64_t CnstVal = CnstBits.getZExtValue();
5939 // Certain magic vector constants (used to express things like NOT
5940 // and NEG) are passed through unmodified. This allows codegen patterns
5941 // for these operations to match. Special-purpose patterns will lower
5942 // these immediates to MOVIs if it proves necessary.
5943 if (VT.isInteger() && (CnstVal == 0 || CnstVal == ~0ULL))
5946 // The many faces of MOVI...
5947 if (AArch64_AM::isAdvSIMDModImmType10(CnstVal)) {
5948 CnstVal = AArch64_AM::encodeAdvSIMDModImmType10(CnstVal);
5949 if (VT.getSizeInBits() == 128) {
5950 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::v2i64,
5951 DAG.getConstant(CnstVal, dl, MVT::i32));
5952 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5955 // Support the V64 version via subregister insertion.
5956 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::f64,
5957 DAG.getConstant(CnstVal, dl, MVT::i32));
5958 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5961 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5962 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5963 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5964 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5965 DAG.getConstant(CnstVal, dl, MVT::i32),
5966 DAG.getConstant(0, dl, MVT::i32));
5967 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5970 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5971 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5972 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5973 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5974 DAG.getConstant(CnstVal, dl, MVT::i32),
5975 DAG.getConstant(8, dl, MVT::i32));
5976 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5979 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5980 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5981 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5982 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5983 DAG.getConstant(CnstVal, dl, MVT::i32),
5984 DAG.getConstant(16, dl, MVT::i32));
5985 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5988 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5989 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5990 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5991 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5992 DAG.getConstant(CnstVal, dl, MVT::i32),
5993 DAG.getConstant(24, dl, MVT::i32));
5994 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5997 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5998 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5999 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
6000 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
6001 DAG.getConstant(CnstVal, dl, MVT::i32),
6002 DAG.getConstant(0, dl, MVT::i32));
6003 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6006 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
6007 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
6008 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
6009 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
6010 DAG.getConstant(CnstVal, dl, MVT::i32),
6011 DAG.getConstant(8, dl, MVT::i32));
6012 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6015 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
6016 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
6017 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6018 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
6019 DAG.getConstant(CnstVal, dl, MVT::i32),
6020 DAG.getConstant(264, dl, MVT::i32));
6021 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6024 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
6025 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
6026 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6027 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
6028 DAG.getConstant(CnstVal, dl, MVT::i32),
6029 DAG.getConstant(272, dl, MVT::i32));
6030 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6033 if (AArch64_AM::isAdvSIMDModImmType9(CnstVal)) {
6034 CnstVal = AArch64_AM::encodeAdvSIMDModImmType9(CnstVal);
6035 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;
6036 SDValue Mov = DAG.getNode(AArch64ISD::MOVI, dl, MovTy,
6037 DAG.getConstant(CnstVal, dl, MVT::i32));
6038 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6041 // The few faces of FMOV...
6042 if (AArch64_AM::isAdvSIMDModImmType11(CnstVal)) {
6043 CnstVal = AArch64_AM::encodeAdvSIMDModImmType11(CnstVal);
6044 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4f32 : MVT::v2f32;
6045 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MovTy,
6046 DAG.getConstant(CnstVal, dl, MVT::i32));
6047 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6050 if (AArch64_AM::isAdvSIMDModImmType12(CnstVal) &&
6051 VT.getSizeInBits() == 128) {
6052 CnstVal = AArch64_AM::encodeAdvSIMDModImmType12(CnstVal);
6053 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MVT::v2f64,
6054 DAG.getConstant(CnstVal, dl, MVT::i32));
6055 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6058 // The many faces of MVNI...
6060 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
6061 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
6062 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6063 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6064 DAG.getConstant(CnstVal, dl, MVT::i32),
6065 DAG.getConstant(0, dl, MVT::i32));
6066 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6069 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
6070 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
6071 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6072 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6073 DAG.getConstant(CnstVal, dl, MVT::i32),
6074 DAG.getConstant(8, dl, MVT::i32));
6075 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6078 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
6079 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
6080 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6081 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6082 DAG.getConstant(CnstVal, dl, MVT::i32),
6083 DAG.getConstant(16, dl, MVT::i32));
6084 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6087 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
6088 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
6089 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6090 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6091 DAG.getConstant(CnstVal, dl, MVT::i32),
6092 DAG.getConstant(24, dl, MVT::i32));
6093 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6096 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
6097 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
6098 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
6099 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6100 DAG.getConstant(CnstVal, dl, MVT::i32),
6101 DAG.getConstant(0, dl, MVT::i32));
6102 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6105 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
6106 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
6107 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
6108 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6109 DAG.getConstant(CnstVal, dl, MVT::i32),
6110 DAG.getConstant(8, dl, MVT::i32));
6111 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6114 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
6115 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
6116 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6117 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
6118 DAG.getConstant(CnstVal, dl, MVT::i32),
6119 DAG.getConstant(264, dl, MVT::i32));
6120 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6123 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
6124 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
6125 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6126 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
6127 DAG.getConstant(CnstVal, dl, MVT::i32),
6128 DAG.getConstant(272, dl, MVT::i32));
6129 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6136 CnstBits = UndefBits;
6141 // Scan through the operands to find some interesting properties we can
6143 // 1) If only one value is used, we can use a DUP, or
6144 // 2) if only the low element is not undef, we can just insert that, or
6145 // 3) if only one constant value is used (w/ some non-constant lanes),
6146 // we can splat the constant value into the whole vector then fill
6147 // in the non-constant lanes.
6148 // 4) FIXME: If different constant values are used, but we can intelligently
6149 // select the values we'll be overwriting for the non-constant
6150 // lanes such that we can directly materialize the vector
6151 // some other way (MOVI, e.g.), we can be sneaky.
6152 unsigned NumElts = VT.getVectorNumElements();
6153 bool isOnlyLowElement = true;
6154 bool usesOnlyOneValue = true;
6155 bool usesOnlyOneConstantValue = true;
6156 bool isConstant = true;
6157 unsigned NumConstantLanes = 0;
6159 SDValue ConstantValue;
6160 for (unsigned i = 0; i < NumElts; ++i) {
6161 SDValue V = Op.getOperand(i);
6162 if (V.getOpcode() == ISD::UNDEF)
6165 isOnlyLowElement = false;
6166 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
6169 if (isa<ConstantSDNode>(V) || isa<ConstantFPSDNode>(V)) {
6171 if (!ConstantValue.getNode())
6173 else if (ConstantValue != V)
6174 usesOnlyOneConstantValue = false;
6177 if (!Value.getNode())
6179 else if (V != Value)
6180 usesOnlyOneValue = false;
6183 if (!Value.getNode())
6184 return DAG.getUNDEF(VT);
6186 if (isOnlyLowElement)
6187 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
6189 // Use DUP for non-constant splats. For f32 constant splats, reduce to
6190 // i32 and try again.
6191 if (usesOnlyOneValue) {
6193 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6194 Value.getValueType() != VT)
6195 return DAG.getNode(AArch64ISD::DUP, dl, VT, Value);
6197 // This is actually a DUPLANExx operation, which keeps everything vectory.
6199 // DUPLANE works on 128-bit vectors, widen it if necessary.
6200 SDValue Lane = Value.getOperand(1);
6201 Value = Value.getOperand(0);
6202 if (Value.getValueType().getSizeInBits() == 64)
6203 Value = WidenVector(Value, DAG);
6205 unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
6206 return DAG.getNode(Opcode, dl, VT, Value, Lane);
6209 if (VT.getVectorElementType().isFloatingPoint()) {
6210 SmallVector<SDValue, 8> Ops;
6211 EVT EltTy = VT.getVectorElementType();
6212 assert ((EltTy == MVT::f16 || EltTy == MVT::f32 || EltTy == MVT::f64) &&
6213 "Unsupported floating-point vector type");
6214 MVT NewType = MVT::getIntegerVT(EltTy.getSizeInBits());
6215 for (unsigned i = 0; i < NumElts; ++i)
6216 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, NewType, Op.getOperand(i)));
6217 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts);
6218 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
6219 Val = LowerBUILD_VECTOR(Val, DAG);
6221 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
6225 // If there was only one constant value used and for more than one lane,
6226 // start by splatting that value, then replace the non-constant lanes. This
6227 // is better than the default, which will perform a separate initialization
6229 if (NumConstantLanes > 0 && usesOnlyOneConstantValue) {
6230 SDValue Val = DAG.getNode(AArch64ISD::DUP, dl, VT, ConstantValue);
6231 // Now insert the non-constant lanes.
6232 for (unsigned i = 0; i < NumElts; ++i) {
6233 SDValue V = Op.getOperand(i);
6234 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i64);
6235 if (!isa<ConstantSDNode>(V) && !isa<ConstantFPSDNode>(V)) {
6236 // Note that type legalization likely mucked about with the VT of the
6237 // source operand, so we may have to convert it here before inserting.
6238 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, V, LaneIdx);
6244 // If all elements are constants and the case above didn't get hit, fall back
6245 // to the default expansion, which will generate a load from the constant
6250 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
6252 if (SDValue shuffle = ReconstructShuffle(Op, DAG))
6256 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
6257 // know the default expansion would otherwise fall back on something even
6258 // worse. For a vector with one or two non-undef values, that's
6259 // scalar_to_vector for the elements followed by a shuffle (provided the
6260 // shuffle is valid for the target) and materialization element by element
6261 // on the stack followed by a load for everything else.
6262 if (!isConstant && !usesOnlyOneValue) {
6263 SDValue Vec = DAG.getUNDEF(VT);
6264 SDValue Op0 = Op.getOperand(0);
6265 unsigned ElemSize = VT.getVectorElementType().getSizeInBits();
6267 // For 32 and 64 bit types, use INSERT_SUBREG for lane zero to
6268 // a) Avoid a RMW dependency on the full vector register, and
6269 // b) Allow the register coalescer to fold away the copy if the
6270 // value is already in an S or D register.
6271 if (Op0.getOpcode() != ISD::UNDEF && (ElemSize == 32 || ElemSize == 64)) {
6272 unsigned SubIdx = ElemSize == 32 ? AArch64::ssub : AArch64::dsub;
6274 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0,
6275 DAG.getTargetConstant(SubIdx, dl, MVT::i32));
6276 Vec = SDValue(N, 0);
6279 for (; i < NumElts; ++i) {
6280 SDValue V = Op.getOperand(i);
6281 if (V.getOpcode() == ISD::UNDEF)
6283 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i64);
6284 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
6289 // Just use the default expansion. We failed to find a better alternative.
6293 SDValue AArch64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
6294 SelectionDAG &DAG) const {
6295 assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!");
6297 // Check for non-constant or out of range lane.
6298 EVT VT = Op.getOperand(0).getValueType();
6299 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(2));
6300 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
6304 // Insertion/extraction are legal for V128 types.
6305 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
6306 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
6310 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
6311 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
6314 // For V64 types, we perform insertion by expanding the value
6315 // to a V128 type and perform the insertion on that.
6317 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
6318 EVT WideTy = WideVec.getValueType();
6320 SDValue Node = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideTy, WideVec,
6321 Op.getOperand(1), Op.getOperand(2));
6322 // Re-narrow the resultant vector.
6323 return NarrowVector(Node, DAG);
6327 AArch64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
6328 SelectionDAG &DAG) const {
6329 assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!");
6331 // Check for non-constant or out of range lane.
6332 EVT VT = Op.getOperand(0).getValueType();
6333 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6334 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
6338 // Insertion/extraction are legal for V128 types.
6339 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
6340 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
6344 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
6345 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
6348 // For V64 types, we perform extraction by expanding the value
6349 // to a V128 type and perform the extraction on that.
6351 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
6352 EVT WideTy = WideVec.getValueType();
6354 EVT ExtrTy = WideTy.getVectorElementType();
6355 if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8)
6358 // For extractions, we just return the result directly.
6359 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec,
6363 SDValue AArch64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
6364 SelectionDAG &DAG) const {
6365 EVT VT = Op.getOperand(0).getValueType();
6371 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6374 unsigned Val = Cst->getZExtValue();
6376 unsigned Size = Op.getValueType().getSizeInBits();
6380 return DAG.getTargetExtractSubreg(AArch64::bsub, dl, Op.getValueType(),
6383 return DAG.getTargetExtractSubreg(AArch64::hsub, dl, Op.getValueType(),
6386 return DAG.getTargetExtractSubreg(AArch64::ssub, dl, Op.getValueType(),
6389 return DAG.getTargetExtractSubreg(AArch64::dsub, dl, Op.getValueType(),
6392 llvm_unreachable("Unexpected vector type in extract_subvector!");
6395 // If this is extracting the upper 64-bits of a 128-bit vector, we match
6397 if (Size == 64 && Val * VT.getVectorElementType().getSizeInBits() == 64)
6403 bool AArch64TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
6405 if (VT.getVectorNumElements() == 4 &&
6406 (VT.is128BitVector() || VT.is64BitVector())) {
6407 unsigned PFIndexes[4];
6408 for (unsigned i = 0; i != 4; ++i) {
6412 PFIndexes[i] = M[i];
6415 // Compute the index in the perfect shuffle table.
6416 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
6417 PFIndexes[2] * 9 + PFIndexes[3];
6418 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
6419 unsigned Cost = (PFEntry >> 30);
6427 unsigned DummyUnsigned;
6429 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || isREVMask(M, VT, 64) ||
6430 isREVMask(M, VT, 32) || isREVMask(M, VT, 16) ||
6431 isEXTMask(M, VT, DummyBool, DummyUnsigned) ||
6432 // isTBLMask(M, VT) || // FIXME: Port TBL support from ARM.
6433 isTRNMask(M, VT, DummyUnsigned) || isUZPMask(M, VT, DummyUnsigned) ||
6434 isZIPMask(M, VT, DummyUnsigned) ||
6435 isTRN_v_undef_Mask(M, VT, DummyUnsigned) ||
6436 isUZP_v_undef_Mask(M, VT, DummyUnsigned) ||
6437 isZIP_v_undef_Mask(M, VT, DummyUnsigned) ||
6438 isINSMask(M, VT.getVectorNumElements(), DummyBool, DummyInt) ||
6439 isConcatMask(M, VT, VT.getSizeInBits() == 128));
6442 /// getVShiftImm - Check if this is a valid build_vector for the immediate
6443 /// operand of a vector shift operation, where all the elements of the
6444 /// build_vector must have the same constant integer value.
6445 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
6446 // Ignore bit_converts.
6447 while (Op.getOpcode() == ISD::BITCAST)
6448 Op = Op.getOperand(0);
6449 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
6450 APInt SplatBits, SplatUndef;
6451 unsigned SplatBitSize;
6453 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
6454 HasAnyUndefs, ElementBits) ||
6455 SplatBitSize > ElementBits)
6457 Cnt = SplatBits.getSExtValue();
6461 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
6462 /// operand of a vector shift left operation. That value must be in the range:
6463 /// 0 <= Value < ElementBits for a left shift; or
6464 /// 0 <= Value <= ElementBits for a long left shift.
6465 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
6466 assert(VT.isVector() && "vector shift count is not a vector type");
6467 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
6468 if (!getVShiftImm(Op, ElementBits, Cnt))
6470 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
6473 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
6474 /// operand of a vector shift right operation. The value must be in the range:
6475 /// 1 <= Value <= ElementBits for a right shift; or
6476 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, int64_t &Cnt) {
6477 assert(VT.isVector() && "vector shift count is not a vector type");
6478 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
6479 if (!getVShiftImm(Op, ElementBits, Cnt))
6481 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
6484 SDValue AArch64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
6485 SelectionDAG &DAG) const {
6486 EVT VT = Op.getValueType();
6490 if (!Op.getOperand(1).getValueType().isVector())
6492 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
6494 switch (Op.getOpcode()) {
6496 llvm_unreachable("unexpected shift opcode");
6499 if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
6500 return DAG.getNode(AArch64ISD::VSHL, DL, VT, Op.getOperand(0),
6501 DAG.getConstant(Cnt, DL, MVT::i32));
6502 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
6503 DAG.getConstant(Intrinsic::aarch64_neon_ushl, DL,
6505 Op.getOperand(0), Op.getOperand(1));
6508 // Right shift immediate
6509 if (isVShiftRImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize) {
6511 (Op.getOpcode() == ISD::SRA) ? AArch64ISD::VASHR : AArch64ISD::VLSHR;
6512 return DAG.getNode(Opc, DL, VT, Op.getOperand(0),
6513 DAG.getConstant(Cnt, DL, MVT::i32));
6516 // Right shift register. Note, there is not a shift right register
6517 // instruction, but the shift left register instruction takes a signed
6518 // value, where negative numbers specify a right shift.
6519 unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::aarch64_neon_sshl
6520 : Intrinsic::aarch64_neon_ushl;
6521 // negate the shift amount
6522 SDValue NegShift = DAG.getNode(AArch64ISD::NEG, DL, VT, Op.getOperand(1));
6523 SDValue NegShiftLeft =
6524 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
6525 DAG.getConstant(Opc, DL, MVT::i32), Op.getOperand(0),
6527 return NegShiftLeft;
6533 static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS,
6534 AArch64CC::CondCode CC, bool NoNans, EVT VT,
6535 SDLoc dl, SelectionDAG &DAG) {
6536 EVT SrcVT = LHS.getValueType();
6537 assert(VT.getSizeInBits() == SrcVT.getSizeInBits() &&
6538 "function only supposed to emit natural comparisons");
6540 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
6541 APInt CnstBits(VT.getSizeInBits(), 0);
6542 APInt UndefBits(VT.getSizeInBits(), 0);
6543 bool IsCnst = BVN && resolveBuildVector(BVN, CnstBits, UndefBits);
6544 bool IsZero = IsCnst && (CnstBits == 0);
6546 if (SrcVT.getVectorElementType().isFloatingPoint()) {
6550 case AArch64CC::NE: {
6553 Fcmeq = DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
6555 Fcmeq = DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
6556 return DAG.getNode(AArch64ISD::NOT, dl, VT, Fcmeq);
6560 return DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
6561 return DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
6564 return DAG.getNode(AArch64ISD::FCMGEz, dl, VT, LHS);
6565 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, LHS, RHS);
6568 return DAG.getNode(AArch64ISD::FCMGTz, dl, VT, LHS);
6569 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, LHS, RHS);
6572 return DAG.getNode(AArch64ISD::FCMLEz, dl, VT, LHS);
6573 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, RHS, LHS);
6577 // If we ignore NaNs then we can use to the MI implementation.
6581 return DAG.getNode(AArch64ISD::FCMLTz, dl, VT, LHS);
6582 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, RHS, LHS);
6589 case AArch64CC::NE: {
6592 Cmeq = DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
6594 Cmeq = DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
6595 return DAG.getNode(AArch64ISD::NOT, dl, VT, Cmeq);
6599 return DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
6600 return DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
6603 return DAG.getNode(AArch64ISD::CMGEz, dl, VT, LHS);
6604 return DAG.getNode(AArch64ISD::CMGE, dl, VT, LHS, RHS);
6607 return DAG.getNode(AArch64ISD::CMGTz, dl, VT, LHS);
6608 return DAG.getNode(AArch64ISD::CMGT, dl, VT, LHS, RHS);
6611 return DAG.getNode(AArch64ISD::CMLEz, dl, VT, LHS);
6612 return DAG.getNode(AArch64ISD::CMGE, dl, VT, RHS, LHS);
6614 return DAG.getNode(AArch64ISD::CMHS, dl, VT, RHS, LHS);
6616 return DAG.getNode(AArch64ISD::CMHI, dl, VT, RHS, LHS);
6619 return DAG.getNode(AArch64ISD::CMLTz, dl, VT, LHS);
6620 return DAG.getNode(AArch64ISD::CMGT, dl, VT, RHS, LHS);
6622 return DAG.getNode(AArch64ISD::CMHI, dl, VT, LHS, RHS);
6624 return DAG.getNode(AArch64ISD::CMHS, dl, VT, LHS, RHS);
6628 SDValue AArch64TargetLowering::LowerVSETCC(SDValue Op,
6629 SelectionDAG &DAG) const {
6630 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6631 SDValue LHS = Op.getOperand(0);
6632 SDValue RHS = Op.getOperand(1);
6633 EVT CmpVT = LHS.getValueType().changeVectorElementTypeToInteger();
6636 if (LHS.getValueType().getVectorElementType().isInteger()) {
6637 assert(LHS.getValueType() == RHS.getValueType());
6638 AArch64CC::CondCode AArch64CC = changeIntCCToAArch64CC(CC);
6640 EmitVectorComparison(LHS, RHS, AArch64CC, false, CmpVT, dl, DAG);
6641 return DAG.getSExtOrTrunc(Cmp, dl, Op.getValueType());
6644 assert(LHS.getValueType().getVectorElementType() == MVT::f32 ||
6645 LHS.getValueType().getVectorElementType() == MVT::f64);
6647 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
6648 // clean. Some of them require two branches to implement.
6649 AArch64CC::CondCode CC1, CC2;
6651 changeVectorFPCCToAArch64CC(CC, CC1, CC2, ShouldInvert);
6653 bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath;
6655 EmitVectorComparison(LHS, RHS, CC1, NoNaNs, CmpVT, dl, DAG);
6659 if (CC2 != AArch64CC::AL) {
6661 EmitVectorComparison(LHS, RHS, CC2, NoNaNs, CmpVT, dl, DAG);
6662 if (!Cmp2.getNode())
6665 Cmp = DAG.getNode(ISD::OR, dl, CmpVT, Cmp, Cmp2);
6668 Cmp = DAG.getSExtOrTrunc(Cmp, dl, Op.getValueType());
6671 return Cmp = DAG.getNOT(dl, Cmp, Cmp.getValueType());
6676 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
6677 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
6678 /// specified in the intrinsic calls.
6679 bool AArch64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
6681 unsigned Intrinsic) const {
6682 auto &DL = I.getModule()->getDataLayout();
6683 switch (Intrinsic) {
6684 case Intrinsic::aarch64_neon_ld2:
6685 case Intrinsic::aarch64_neon_ld3:
6686 case Intrinsic::aarch64_neon_ld4:
6687 case Intrinsic::aarch64_neon_ld1x2:
6688 case Intrinsic::aarch64_neon_ld1x3:
6689 case Intrinsic::aarch64_neon_ld1x4:
6690 case Intrinsic::aarch64_neon_ld2lane:
6691 case Intrinsic::aarch64_neon_ld3lane:
6692 case Intrinsic::aarch64_neon_ld4lane:
6693 case Intrinsic::aarch64_neon_ld2r:
6694 case Intrinsic::aarch64_neon_ld3r:
6695 case Intrinsic::aarch64_neon_ld4r: {
6696 Info.opc = ISD::INTRINSIC_W_CHAIN;
6697 // Conservatively set memVT to the entire set of vectors loaded.
6698 uint64_t NumElts = DL.getTypeAllocSize(I.getType()) / 8;
6699 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6700 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6703 Info.vol = false; // volatile loads with NEON intrinsics not supported
6704 Info.readMem = true;
6705 Info.writeMem = false;
6708 case Intrinsic::aarch64_neon_st2:
6709 case Intrinsic::aarch64_neon_st3:
6710 case Intrinsic::aarch64_neon_st4:
6711 case Intrinsic::aarch64_neon_st1x2:
6712 case Intrinsic::aarch64_neon_st1x3:
6713 case Intrinsic::aarch64_neon_st1x4:
6714 case Intrinsic::aarch64_neon_st2lane:
6715 case Intrinsic::aarch64_neon_st3lane:
6716 case Intrinsic::aarch64_neon_st4lane: {
6717 Info.opc = ISD::INTRINSIC_VOID;
6718 // Conservatively set memVT to the entire set of vectors stored.
6719 unsigned NumElts = 0;
6720 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
6721 Type *ArgTy = I.getArgOperand(ArgI)->getType();
6722 if (!ArgTy->isVectorTy())
6724 NumElts += DL.getTypeAllocSize(ArgTy) / 8;
6726 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6727 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6730 Info.vol = false; // volatile stores with NEON intrinsics not supported
6731 Info.readMem = false;
6732 Info.writeMem = true;
6735 case Intrinsic::aarch64_ldaxr:
6736 case Intrinsic::aarch64_ldxr: {
6737 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
6738 Info.opc = ISD::INTRINSIC_W_CHAIN;
6739 Info.memVT = MVT::getVT(PtrTy->getElementType());
6740 Info.ptrVal = I.getArgOperand(0);
6742 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
6744 Info.readMem = true;
6745 Info.writeMem = false;
6748 case Intrinsic::aarch64_stlxr:
6749 case Intrinsic::aarch64_stxr: {
6750 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
6751 Info.opc = ISD::INTRINSIC_W_CHAIN;
6752 Info.memVT = MVT::getVT(PtrTy->getElementType());
6753 Info.ptrVal = I.getArgOperand(1);
6755 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
6757 Info.readMem = false;
6758 Info.writeMem = true;
6761 case Intrinsic::aarch64_ldaxp:
6762 case Intrinsic::aarch64_ldxp: {
6763 Info.opc = ISD::INTRINSIC_W_CHAIN;
6764 Info.memVT = MVT::i128;
6765 Info.ptrVal = I.getArgOperand(0);
6769 Info.readMem = true;
6770 Info.writeMem = false;
6773 case Intrinsic::aarch64_stlxp:
6774 case Intrinsic::aarch64_stxp: {
6775 Info.opc = ISD::INTRINSIC_W_CHAIN;
6776 Info.memVT = MVT::i128;
6777 Info.ptrVal = I.getArgOperand(2);
6781 Info.readMem = false;
6782 Info.writeMem = true;
6792 // Truncations from 64-bit GPR to 32-bit GPR is free.
6793 bool AArch64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
6794 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6796 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6797 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6798 return NumBits1 > NumBits2;
6800 bool AArch64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
6801 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6803 unsigned NumBits1 = VT1.getSizeInBits();
6804 unsigned NumBits2 = VT2.getSizeInBits();
6805 return NumBits1 > NumBits2;
6808 /// Check if it is profitable to hoist instruction in then/else to if.
6809 /// Not profitable if I and it's user can form a FMA instruction
6810 /// because we prefer FMSUB/FMADD.
6811 bool AArch64TargetLowering::isProfitableToHoist(Instruction *I) const {
6812 if (I->getOpcode() != Instruction::FMul)
6815 if (I->getNumUses() != 1)
6818 Instruction *User = I->user_back();
6821 !(User->getOpcode() == Instruction::FSub ||
6822 User->getOpcode() == Instruction::FAdd))
6825 const TargetOptions &Options = getTargetMachine().Options;
6826 const DataLayout &DL = I->getModule()->getDataLayout();
6827 EVT VT = getValueType(DL, User->getOperand(0)->getType());
6829 if (isFMAFasterThanFMulAndFAdd(VT) &&
6830 isOperationLegalOrCustom(ISD::FMA, VT) &&
6831 (Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath))
6837 // All 32-bit GPR operations implicitly zero the high-half of the corresponding
6839 bool AArch64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
6840 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6842 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6843 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6844 return NumBits1 == 32 && NumBits2 == 64;
6846 bool AArch64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
6847 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6849 unsigned NumBits1 = VT1.getSizeInBits();
6850 unsigned NumBits2 = VT2.getSizeInBits();
6851 return NumBits1 == 32 && NumBits2 == 64;
6854 bool AArch64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
6855 EVT VT1 = Val.getValueType();
6856 if (isZExtFree(VT1, VT2)) {
6860 if (Val.getOpcode() != ISD::LOAD)
6863 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
6864 return (VT1.isSimple() && !VT1.isVector() && VT1.isInteger() &&
6865 VT2.isSimple() && !VT2.isVector() && VT2.isInteger() &&
6866 VT1.getSizeInBits() <= 32);
6869 bool AArch64TargetLowering::isExtFreeImpl(const Instruction *Ext) const {
6870 if (isa<FPExtInst>(Ext))
6873 // Vector types are next free.
6874 if (Ext->getType()->isVectorTy())
6877 for (const Use &U : Ext->uses()) {
6878 // The extension is free if we can fold it with a left shift in an
6879 // addressing mode or an arithmetic operation: add, sub, and cmp.
6881 // Is there a shift?
6882 const Instruction *Instr = cast<Instruction>(U.getUser());
6884 // Is this a constant shift?
6885 switch (Instr->getOpcode()) {
6886 case Instruction::Shl:
6887 if (!isa<ConstantInt>(Instr->getOperand(1)))
6890 case Instruction::GetElementPtr: {
6891 gep_type_iterator GTI = gep_type_begin(Instr);
6892 auto &DL = Ext->getModule()->getDataLayout();
6893 std::advance(GTI, U.getOperandNo());
6895 // This extension will end up with a shift because of the scaling factor.
6896 // 8-bit sized types have a scaling factor of 1, thus a shift amount of 0.
6897 // Get the shift amount based on the scaling factor:
6898 // log2(sizeof(IdxTy)) - log2(8).
6900 countTrailingZeros(DL.getTypeStoreSizeInBits(IdxTy)) - 3;
6901 // Is the constant foldable in the shift of the addressing mode?
6902 // I.e., shift amount is between 1 and 4 inclusive.
6903 if (ShiftAmt == 0 || ShiftAmt > 4)
6907 case Instruction::Trunc:
6908 // Check if this is a noop.
6909 // trunc(sext ty1 to ty2) to ty1.
6910 if (Instr->getType() == Ext->getOperand(0)->getType())
6917 // At this point we can use the bfm family, so this extension is free
6923 bool AArch64TargetLowering::hasPairedLoad(Type *LoadedType,
6924 unsigned &RequiredAligment) const {
6925 if (!LoadedType->isIntegerTy() && !LoadedType->isFloatTy())
6927 // Cyclone supports unaligned accesses.
6928 RequiredAligment = 0;
6929 unsigned NumBits = LoadedType->getPrimitiveSizeInBits();
6930 return NumBits == 32 || NumBits == 64;
6933 bool AArch64TargetLowering::hasPairedLoad(EVT LoadedType,
6934 unsigned &RequiredAligment) const {
6935 if (!LoadedType.isSimple() ||
6936 (!LoadedType.isInteger() && !LoadedType.isFloatingPoint()))
6938 // Cyclone supports unaligned accesses.
6939 RequiredAligment = 0;
6940 unsigned NumBits = LoadedType.getSizeInBits();
6941 return NumBits == 32 || NumBits == 64;
6944 /// \brief Lower an interleaved load into a ldN intrinsic.
6946 /// E.g. Lower an interleaved load (Factor = 2):
6947 /// %wide.vec = load <8 x i32>, <8 x i32>* %ptr
6948 /// %v0 = shuffle %wide.vec, undef, <0, 2, 4, 6> ; Extract even elements
6949 /// %v1 = shuffle %wide.vec, undef, <1, 3, 5, 7> ; Extract odd elements
6952 /// %ld2 = { <4 x i32>, <4 x i32> } call llvm.aarch64.neon.ld2(%ptr)
6953 /// %vec0 = extractelement { <4 x i32>, <4 x i32> } %ld2, i32 0
6954 /// %vec1 = extractelement { <4 x i32>, <4 x i32> } %ld2, i32 1
6955 bool AArch64TargetLowering::lowerInterleavedLoad(
6956 LoadInst *LI, ArrayRef<ShuffleVectorInst *> Shuffles,
6957 ArrayRef<unsigned> Indices, unsigned Factor) const {
6958 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
6959 "Invalid interleave factor");
6960 assert(!Shuffles.empty() && "Empty shufflevector input");
6961 assert(Shuffles.size() == Indices.size() &&
6962 "Unmatched number of shufflevectors and indices");
6964 const DataLayout &DL = LI->getModule()->getDataLayout();
6966 VectorType *VecTy = Shuffles[0]->getType();
6967 unsigned VecSize = DL.getTypeAllocSizeInBits(VecTy);
6969 // Skip illegal vector types.
6970 if (VecSize != 64 && VecSize != 128)
6973 // A pointer vector can not be the return type of the ldN intrinsics. Need to
6974 // load integer vectors first and then convert to pointer vectors.
6975 Type *EltTy = VecTy->getVectorElementType();
6976 if (EltTy->isPointerTy())
6978 VectorType::get(DL.getIntPtrType(EltTy), VecTy->getVectorNumElements());
6980 Type *PtrTy = VecTy->getPointerTo(LI->getPointerAddressSpace());
6981 Type *Tys[2] = {VecTy, PtrTy};
6982 static const Intrinsic::ID LoadInts[3] = {Intrinsic::aarch64_neon_ld2,
6983 Intrinsic::aarch64_neon_ld3,
6984 Intrinsic::aarch64_neon_ld4};
6986 Intrinsic::getDeclaration(LI->getModule(), LoadInts[Factor - 2], Tys);
6988 IRBuilder<> Builder(LI);
6989 Value *Ptr = Builder.CreateBitCast(LI->getPointerOperand(), PtrTy);
6991 CallInst *LdN = Builder.CreateCall(LdNFunc, Ptr, "ldN");
6993 // Replace uses of each shufflevector with the corresponding vector loaded
6995 for (unsigned i = 0; i < Shuffles.size(); i++) {
6996 ShuffleVectorInst *SVI = Shuffles[i];
6997 unsigned Index = Indices[i];
6999 Value *SubVec = Builder.CreateExtractValue(LdN, Index);
7001 // Convert the integer vector to pointer vector if the element is pointer.
7002 if (EltTy->isPointerTy())
7003 SubVec = Builder.CreateIntToPtr(SubVec, SVI->getType());
7005 SVI->replaceAllUsesWith(SubVec);
7011 /// \brief Get a mask consisting of sequential integers starting from \p Start.
7013 /// I.e. <Start, Start + 1, ..., Start + NumElts - 1>
7014 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned Start,
7016 SmallVector<Constant *, 16> Mask;
7017 for (unsigned i = 0; i < NumElts; i++)
7018 Mask.push_back(Builder.getInt32(Start + i));
7020 return ConstantVector::get(Mask);
7023 /// \brief Lower an interleaved store into a stN intrinsic.
7025 /// E.g. Lower an interleaved store (Factor = 3):
7026 /// %i.vec = shuffle <8 x i32> %v0, <8 x i32> %v1,
7027 /// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11>
7028 /// store <12 x i32> %i.vec, <12 x i32>* %ptr
7031 /// %sub.v0 = shuffle <8 x i32> %v0, <8 x i32> v1, <0, 1, 2, 3>
7032 /// %sub.v1 = shuffle <8 x i32> %v0, <8 x i32> v1, <4, 5, 6, 7>
7033 /// %sub.v2 = shuffle <8 x i32> %v0, <8 x i32> v1, <8, 9, 10, 11>
7034 /// call void llvm.aarch64.neon.st3(%sub.v0, %sub.v1, %sub.v2, %ptr)
7036 /// Note that the new shufflevectors will be removed and we'll only generate one
7037 /// st3 instruction in CodeGen.
7038 bool AArch64TargetLowering::lowerInterleavedStore(StoreInst *SI,
7039 ShuffleVectorInst *SVI,
7040 unsigned Factor) const {
7041 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
7042 "Invalid interleave factor");
7044 VectorType *VecTy = SVI->getType();
7045 assert(VecTy->getVectorNumElements() % Factor == 0 &&
7046 "Invalid interleaved store");
7048 unsigned NumSubElts = VecTy->getVectorNumElements() / Factor;
7049 Type *EltTy = VecTy->getVectorElementType();
7050 VectorType *SubVecTy = VectorType::get(EltTy, NumSubElts);
7052 const DataLayout &DL = SI->getModule()->getDataLayout();
7053 unsigned SubVecSize = DL.getTypeAllocSizeInBits(SubVecTy);
7055 // Skip illegal vector types.
7056 if (SubVecSize != 64 && SubVecSize != 128)
7059 Value *Op0 = SVI->getOperand(0);
7060 Value *Op1 = SVI->getOperand(1);
7061 IRBuilder<> Builder(SI);
7063 // StN intrinsics don't support pointer vectors as arguments. Convert pointer
7064 // vectors to integer vectors.
7065 if (EltTy->isPointerTy()) {
7066 Type *IntTy = DL.getIntPtrType(EltTy);
7067 unsigned NumOpElts =
7068 dyn_cast<VectorType>(Op0->getType())->getVectorNumElements();
7070 // Convert to the corresponding integer vector.
7071 Type *IntVecTy = VectorType::get(IntTy, NumOpElts);
7072 Op0 = Builder.CreatePtrToInt(Op0, IntVecTy);
7073 Op1 = Builder.CreatePtrToInt(Op1, IntVecTy);
7075 SubVecTy = VectorType::get(IntTy, NumSubElts);
7078 Type *PtrTy = SubVecTy->getPointerTo(SI->getPointerAddressSpace());
7079 Type *Tys[2] = {SubVecTy, PtrTy};
7080 static const Intrinsic::ID StoreInts[3] = {Intrinsic::aarch64_neon_st2,
7081 Intrinsic::aarch64_neon_st3,
7082 Intrinsic::aarch64_neon_st4};
7084 Intrinsic::getDeclaration(SI->getModule(), StoreInts[Factor - 2], Tys);
7086 SmallVector<Value *, 5> Ops;
7088 // Split the shufflevector operands into sub vectors for the new stN call.
7089 for (unsigned i = 0; i < Factor; i++)
7090 Ops.push_back(Builder.CreateShuffleVector(
7091 Op0, Op1, getSequentialMask(Builder, NumSubElts * i, NumSubElts)));
7093 Ops.push_back(Builder.CreateBitCast(SI->getPointerOperand(), PtrTy));
7094 Builder.CreateCall(StNFunc, Ops);
7098 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
7099 unsigned AlignCheck) {
7100 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
7101 (DstAlign == 0 || DstAlign % AlignCheck == 0));
7104 EVT AArch64TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
7105 unsigned SrcAlign, bool IsMemset,
7108 MachineFunction &MF) const {
7109 // Don't use AdvSIMD to implement 16-byte memset. It would have taken one
7110 // instruction to materialize the v2i64 zero and one store (with restrictive
7111 // addressing mode). Just do two i64 store of zero-registers.
7113 const Function *F = MF.getFunction();
7114 if (Subtarget->hasFPARMv8() && !IsMemset && Size >= 16 &&
7115 !F->hasFnAttribute(Attribute::NoImplicitFloat) &&
7116 (memOpAlign(SrcAlign, DstAlign, 16) ||
7117 (allowsMisalignedMemoryAccesses(MVT::f128, 0, 1, &Fast) && Fast)))
7121 (memOpAlign(SrcAlign, DstAlign, 8) ||
7122 (allowsMisalignedMemoryAccesses(MVT::i64, 0, 1, &Fast) && Fast)))
7126 (memOpAlign(SrcAlign, DstAlign, 4) ||
7127 (allowsMisalignedMemoryAccesses(MVT::i32, 0, 1, &Fast) && Fast)))
7133 // 12-bit optionally shifted immediates are legal for adds.
7134 bool AArch64TargetLowering::isLegalAddImmediate(int64_t Immed) const {
7135 if ((Immed >> 12) == 0 || ((Immed & 0xfff) == 0 && Immed >> 24 == 0))
7140 // Integer comparisons are implemented with ADDS/SUBS, so the range of valid
7141 // immediates is the same as for an add or a sub.
7142 bool AArch64TargetLowering::isLegalICmpImmediate(int64_t Immed) const {
7145 return isLegalAddImmediate(Immed);
7148 /// isLegalAddressingMode - Return true if the addressing mode represented
7149 /// by AM is legal for this target, for a load/store of the specified type.
7150 bool AArch64TargetLowering::isLegalAddressingMode(const DataLayout &DL,
7151 const AddrMode &AM, Type *Ty,
7152 unsigned AS) const {
7153 // AArch64 has five basic addressing modes:
7155 // reg + 9-bit signed offset
7156 // reg + SIZE_IN_BYTES * 12-bit unsigned offset
7158 // reg + SIZE_IN_BYTES * reg
7160 // No global is ever allowed as a base.
7164 // No reg+reg+imm addressing.
7165 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
7168 // check reg + imm case:
7169 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
7170 uint64_t NumBytes = 0;
7171 if (Ty->isSized()) {
7172 uint64_t NumBits = DL.getTypeSizeInBits(Ty);
7173 NumBytes = NumBits / 8;
7174 if (!isPowerOf2_64(NumBits))
7179 int64_t Offset = AM.BaseOffs;
7181 // 9-bit signed offset
7182 if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
7185 // 12-bit unsigned offset
7186 unsigned shift = Log2_64(NumBytes);
7187 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
7188 // Must be a multiple of NumBytes (NumBytes is a power of 2)
7189 (Offset >> shift) << shift == Offset)
7194 // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2
7196 if (!AM.Scale || AM.Scale == 1 ||
7197 (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
7202 int AArch64TargetLowering::getScalingFactorCost(const DataLayout &DL,
7203 const AddrMode &AM, Type *Ty,
7204 unsigned AS) const {
7205 // Scaling factors are not free at all.
7206 // Operands | Rt Latency
7207 // -------------------------------------------
7209 // -------------------------------------------
7210 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
7211 // Rt, [Xn, Wm, <extend> #imm] |
7212 if (isLegalAddressingMode(DL, AM, Ty, AS))
7213 // Scale represents reg2 * scale, thus account for 1 if
7214 // it is not equal to 0 or 1.
7215 return AM.Scale != 0 && AM.Scale != 1;
7219 bool AArch64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
7220 VT = VT.getScalarType();
7225 switch (VT.getSimpleVT().SimpleTy) {
7237 AArch64TargetLowering::getScratchRegisters(CallingConv::ID) const {
7238 // LR is a callee-save register, but we must treat it as clobbered by any call
7239 // site. Hence we include LR in the scratch registers, which are in turn added
7240 // as implicit-defs for stackmaps and patchpoints.
7241 static const MCPhysReg ScratchRegs[] = {
7242 AArch64::X16, AArch64::X17, AArch64::LR, 0
7248 AArch64TargetLowering::isDesirableToCommuteWithShift(const SDNode *N) const {
7249 EVT VT = N->getValueType(0);
7250 // If N is unsigned bit extraction: ((x >> C) & mask), then do not combine
7251 // it with shift to let it be lowered to UBFX.
7252 if (N->getOpcode() == ISD::AND && (VT == MVT::i32 || VT == MVT::i64) &&
7253 isa<ConstantSDNode>(N->getOperand(1))) {
7254 uint64_t TruncMask = N->getConstantOperandVal(1);
7255 if (isMask_64(TruncMask) &&
7256 N->getOperand(0).getOpcode() == ISD::SRL &&
7257 isa<ConstantSDNode>(N->getOperand(0)->getOperand(1)))
7263 bool AArch64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
7265 assert(Ty->isIntegerTy());
7267 unsigned BitSize = Ty->getPrimitiveSizeInBits();
7271 int64_t Val = Imm.getSExtValue();
7272 if (Val == 0 || AArch64_AM::isLogicalImmediate(Val, BitSize))
7275 if ((int64_t)Val < 0)
7278 Val &= (1LL << 32) - 1;
7280 unsigned LZ = countLeadingZeros((uint64_t)Val);
7281 unsigned Shift = (63 - LZ) / 16;
7282 // MOVZ is free so return true for one or fewer MOVK.
7286 // Generate SUBS and CSEL for integer abs.
7287 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
7288 EVT VT = N->getValueType(0);
7290 SDValue N0 = N->getOperand(0);
7291 SDValue N1 = N->getOperand(1);
7294 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
7295 // and change it to SUB and CSEL.
7296 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
7297 N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
7298 N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0))
7299 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
7300 if (Y1C->getAPIntValue() == VT.getSizeInBits() - 1) {
7301 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT),
7303 // Generate SUBS & CSEL.
7305 DAG.getNode(AArch64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32),
7306 N0.getOperand(0), DAG.getConstant(0, DL, VT));
7307 return DAG.getNode(AArch64ISD::CSEL, DL, VT, N0.getOperand(0), Neg,
7308 DAG.getConstant(AArch64CC::PL, DL, MVT::i32),
7309 SDValue(Cmp.getNode(), 1));
7314 // performXorCombine - Attempts to handle integer ABS.
7315 static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
7316 TargetLowering::DAGCombinerInfo &DCI,
7317 const AArch64Subtarget *Subtarget) {
7318 if (DCI.isBeforeLegalizeOps())
7321 return performIntegerAbsCombine(N, DAG);
7325 AArch64TargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor,
7327 std::vector<SDNode *> *Created) const {
7328 // fold (sdiv X, pow2)
7329 EVT VT = N->getValueType(0);
7330 if ((VT != MVT::i32 && VT != MVT::i64) ||
7331 !(Divisor.isPowerOf2() || (-Divisor).isPowerOf2()))
7335 SDValue N0 = N->getOperand(0);
7336 unsigned Lg2 = Divisor.countTrailingZeros();
7337 SDValue Zero = DAG.getConstant(0, DL, VT);
7338 SDValue Pow2MinusOne = DAG.getConstant((1ULL << Lg2) - 1, DL, VT);
7340 // Add (N0 < 0) ? Pow2 - 1 : 0;
7342 SDValue Cmp = getAArch64Cmp(N0, Zero, ISD::SETLT, CCVal, DAG, DL);
7343 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0, Pow2MinusOne);
7344 SDValue CSel = DAG.getNode(AArch64ISD::CSEL, DL, VT, Add, N0, CCVal, Cmp);
7347 Created->push_back(Cmp.getNode());
7348 Created->push_back(Add.getNode());
7349 Created->push_back(CSel.getNode());
7354 DAG.getNode(ISD::SRA, DL, VT, CSel, DAG.getConstant(Lg2, DL, MVT::i64));
7356 // If we're dividing by a positive value, we're done. Otherwise, we must
7357 // negate the result.
7358 if (Divisor.isNonNegative())
7362 Created->push_back(SRA.getNode());
7363 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), SRA);
7366 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG,
7367 TargetLowering::DAGCombinerInfo &DCI,
7368 const AArch64Subtarget *Subtarget) {
7369 if (DCI.isBeforeLegalizeOps())
7372 // Multiplication of a power of two plus/minus one can be done more
7373 // cheaply as as shift+add/sub. For now, this is true unilaterally. If
7374 // future CPUs have a cheaper MADD instruction, this may need to be
7375 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
7376 // 64-bit is 5 cycles, so this is always a win.
7377 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
7378 APInt Value = C->getAPIntValue();
7379 EVT VT = N->getValueType(0);
7381 if (Value.isNonNegative()) {
7382 // (mul x, 2^N + 1) => (add (shl x, N), x)
7383 APInt VM1 = Value - 1;
7384 if (VM1.isPowerOf2()) {
7385 SDValue ShiftedVal =
7386 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7387 DAG.getConstant(VM1.logBase2(), DL, MVT::i64));
7388 return DAG.getNode(ISD::ADD, DL, VT, ShiftedVal,
7391 // (mul x, 2^N - 1) => (sub (shl x, N), x)
7392 APInt VP1 = Value + 1;
7393 if (VP1.isPowerOf2()) {
7394 SDValue ShiftedVal =
7395 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7396 DAG.getConstant(VP1.logBase2(), DL, MVT::i64));
7397 return DAG.getNode(ISD::SUB, DL, VT, ShiftedVal,
7401 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
7402 APInt VNP1 = -Value + 1;
7403 if (VNP1.isPowerOf2()) {
7404 SDValue ShiftedVal =
7405 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7406 DAG.getConstant(VNP1.logBase2(), DL, MVT::i64));
7407 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0),
7410 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
7411 APInt VNM1 = -Value - 1;
7412 if (VNM1.isPowerOf2()) {
7413 SDValue ShiftedVal =
7414 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7415 DAG.getConstant(VNM1.logBase2(), DL, MVT::i64));
7417 DAG.getNode(ISD::ADD, DL, VT, ShiftedVal, N->getOperand(0));
7418 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Add);
7425 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
7426 SelectionDAG &DAG) {
7427 // Take advantage of vector comparisons producing 0 or -1 in each lane to
7428 // optimize away operation when it's from a constant.
7430 // The general transformation is:
7431 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
7432 // AND(VECTOR_CMP(x,y), constant2)
7433 // constant2 = UNARYOP(constant)
7435 // Early exit if this isn't a vector operation, the operand of the
7436 // unary operation isn't a bitwise AND, or if the sizes of the operations
7438 EVT VT = N->getValueType(0);
7439 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
7440 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
7441 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
7444 // Now check that the other operand of the AND is a constant. We could
7445 // make the transformation for non-constant splats as well, but it's unclear
7446 // that would be a benefit as it would not eliminate any operations, just
7447 // perform one more step in scalar code before moving to the vector unit.
7448 if (BuildVectorSDNode *BV =
7449 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
7450 // Bail out if the vector isn't a constant.
7451 if (!BV->isConstant())
7454 // Everything checks out. Build up the new and improved node.
7456 EVT IntVT = BV->getValueType(0);
7457 // Create a new constant of the appropriate type for the transformed
7459 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
7460 // The AND node needs bitcasts to/from an integer vector type around it.
7461 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
7462 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
7463 N->getOperand(0)->getOperand(0), MaskConst);
7464 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
7471 static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG,
7472 const AArch64Subtarget *Subtarget) {
7473 // First try to optimize away the conversion when it's conditionally from
7474 // a constant. Vectors only.
7475 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
7478 EVT VT = N->getValueType(0);
7479 if (VT != MVT::f32 && VT != MVT::f64)
7482 // Only optimize when the source and destination types have the same width.
7483 if (VT.getSizeInBits() != N->getOperand(0).getValueType().getSizeInBits())
7486 // If the result of an integer load is only used by an integer-to-float
7487 // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
7488 // This eliminates an "integer-to-vector-move UOP and improve throughput.
7489 SDValue N0 = N->getOperand(0);
7490 if (Subtarget->hasNEON() && ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7491 // Do not change the width of a volatile load.
7492 !cast<LoadSDNode>(N0)->isVolatile()) {
7493 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7494 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
7495 LN0->getPointerInfo(), LN0->isVolatile(),
7496 LN0->isNonTemporal(), LN0->isInvariant(),
7497 LN0->getAlignment());
7499 // Make sure successors of the original load stay after it by updating them
7500 // to use the new Chain.
7501 DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));
7504 (N->getOpcode() == ISD::SINT_TO_FP) ? AArch64ISD::SITOF : AArch64ISD::UITOF;
7505 return DAG.getNode(Opcode, SDLoc(N), VT, Load);
7511 /// An EXTR instruction is made up of two shifts, ORed together. This helper
7512 /// searches for and classifies those shifts.
7513 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
7515 if (N.getOpcode() == ISD::SHL)
7517 else if (N.getOpcode() == ISD::SRL)
7522 if (!isa<ConstantSDNode>(N.getOperand(1)))
7525 ShiftAmount = N->getConstantOperandVal(1);
7526 Src = N->getOperand(0);
7530 /// EXTR instruction extracts a contiguous chunk of bits from two existing
7531 /// registers viewed as a high/low pair. This function looks for the pattern:
7532 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
7533 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
7535 static SDValue tryCombineToEXTR(SDNode *N,
7536 TargetLowering::DAGCombinerInfo &DCI) {
7537 SelectionDAG &DAG = DCI.DAG;
7539 EVT VT = N->getValueType(0);
7541 assert(N->getOpcode() == ISD::OR && "Unexpected root");
7543 if (VT != MVT::i32 && VT != MVT::i64)
7547 uint32_t ShiftLHS = 0;
7549 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
7553 uint32_t ShiftRHS = 0;
7555 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
7558 // If they're both trying to come from the high part of the register, they're
7559 // not really an EXTR.
7560 if (LHSFromHi == RHSFromHi)
7563 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
7567 std::swap(LHS, RHS);
7568 std::swap(ShiftLHS, ShiftRHS);
7571 return DAG.getNode(AArch64ISD::EXTR, DL, VT, LHS, RHS,
7572 DAG.getConstant(ShiftRHS, DL, MVT::i64));
7575 static SDValue tryCombineToBSL(SDNode *N,
7576 TargetLowering::DAGCombinerInfo &DCI) {
7577 EVT VT = N->getValueType(0);
7578 SelectionDAG &DAG = DCI.DAG;
7584 SDValue N0 = N->getOperand(0);
7585 if (N0.getOpcode() != ISD::AND)
7588 SDValue N1 = N->getOperand(1);
7589 if (N1.getOpcode() != ISD::AND)
7592 // We only have to look for constant vectors here since the general, variable
7593 // case can be handled in TableGen.
7594 unsigned Bits = VT.getVectorElementType().getSizeInBits();
7595 uint64_t BitMask = Bits == 64 ? -1ULL : ((1ULL << Bits) - 1);
7596 for (int i = 1; i >= 0; --i)
7597 for (int j = 1; j >= 0; --j) {
7598 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(i));
7599 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(j));
7603 bool FoundMatch = true;
7604 for (unsigned k = 0; k < VT.getVectorNumElements(); ++k) {
7605 ConstantSDNode *CN0 = dyn_cast<ConstantSDNode>(BVN0->getOperand(k));
7606 ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(BVN1->getOperand(k));
7608 CN0->getZExtValue() != (BitMask & ~CN1->getZExtValue())) {
7615 return DAG.getNode(AArch64ISD::BSL, DL, VT, SDValue(BVN0, 0),
7616 N0->getOperand(1 - i), N1->getOperand(1 - j));
7622 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
7623 const AArch64Subtarget *Subtarget) {
7624 // Attempt to form an EXTR from (or (shl VAL1, #N), (srl VAL2, #RegWidth-N))
7625 if (!EnableAArch64ExtrGeneration)
7627 SelectionDAG &DAG = DCI.DAG;
7628 EVT VT = N->getValueType(0);
7630 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7633 SDValue Res = tryCombineToEXTR(N, DCI);
7637 Res = tryCombineToBSL(N, DCI);
7644 static SDValue performBitcastCombine(SDNode *N,
7645 TargetLowering::DAGCombinerInfo &DCI,
7646 SelectionDAG &DAG) {
7647 // Wait 'til after everything is legalized to try this. That way we have
7648 // legal vector types and such.
7649 if (DCI.isBeforeLegalizeOps())
7652 // Remove extraneous bitcasts around an extract_subvector.
7654 // (v4i16 (bitconvert
7655 // (extract_subvector (v2i64 (bitconvert (v8i16 ...)), (i64 1)))))
7657 // (extract_subvector ((v8i16 ...), (i64 4)))
7659 // Only interested in 64-bit vectors as the ultimate result.
7660 EVT VT = N->getValueType(0);
7663 if (VT.getSimpleVT().getSizeInBits() != 64)
7665 // Is the operand an extract_subvector starting at the beginning or halfway
7666 // point of the vector? A low half may also come through as an
7667 // EXTRACT_SUBREG, so look for that, too.
7668 SDValue Op0 = N->getOperand(0);
7669 if (Op0->getOpcode() != ISD::EXTRACT_SUBVECTOR &&
7670 !(Op0->isMachineOpcode() &&
7671 Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG))
7673 uint64_t idx = cast<ConstantSDNode>(Op0->getOperand(1))->getZExtValue();
7674 if (Op0->getOpcode() == ISD::EXTRACT_SUBVECTOR) {
7675 if (Op0->getValueType(0).getVectorNumElements() != idx && idx != 0)
7677 } else if (Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG) {
7678 if (idx != AArch64::dsub)
7680 // The dsub reference is equivalent to a lane zero subvector reference.
7683 // Look through the bitcast of the input to the extract.
7684 if (Op0->getOperand(0)->getOpcode() != ISD::BITCAST)
7686 SDValue Source = Op0->getOperand(0)->getOperand(0);
7687 // If the source type has twice the number of elements as our destination
7688 // type, we know this is an extract of the high or low half of the vector.
7689 EVT SVT = Source->getValueType(0);
7690 if (SVT.getVectorNumElements() != VT.getVectorNumElements() * 2)
7693 DEBUG(dbgs() << "aarch64-lower: bitcast extract_subvector simplification\n");
7695 // Create the simplified form to just extract the low or high half of the
7696 // vector directly rather than bothering with the bitcasts.
7698 unsigned NumElements = VT.getVectorNumElements();
7700 SDValue HalfIdx = DAG.getConstant(NumElements, dl, MVT::i64);
7701 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Source, HalfIdx);
7703 SDValue SubReg = DAG.getTargetConstant(AArch64::dsub, dl, MVT::i32);
7704 return SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl, VT,
7710 static SDValue performConcatVectorsCombine(SDNode *N,
7711 TargetLowering::DAGCombinerInfo &DCI,
7712 SelectionDAG &DAG) {
7714 EVT VT = N->getValueType(0);
7715 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
7717 // Optimize concat_vectors of truncated vectors, where the intermediate
7718 // type is illegal, to avoid said illegality, e.g.,
7719 // (v4i16 (concat_vectors (v2i16 (truncate (v2i64))),
7720 // (v2i16 (truncate (v2i64)))))
7722 // (v4i16 (truncate (vector_shuffle (v4i32 (bitcast (v2i64))),
7723 // (v4i32 (bitcast (v2i64))),
7725 // This isn't really target-specific, but ISD::TRUNCATE legality isn't keyed
7726 // on both input and result type, so we might generate worse code.
7727 // On AArch64 we know it's fine for v2i64->v4i16 and v4i32->v8i8.
7728 if (N->getNumOperands() == 2 &&
7729 N0->getOpcode() == ISD::TRUNCATE &&
7730 N1->getOpcode() == ISD::TRUNCATE) {
7731 SDValue N00 = N0->getOperand(0);
7732 SDValue N10 = N1->getOperand(0);
7733 EVT N00VT = N00.getValueType();
7735 if (N00VT == N10.getValueType() &&
7736 (N00VT == MVT::v2i64 || N00VT == MVT::v4i32) &&
7737 N00VT.getScalarSizeInBits() == 4 * VT.getScalarSizeInBits()) {
7738 MVT MidVT = (N00VT == MVT::v2i64 ? MVT::v4i32 : MVT::v8i16);
7739 SmallVector<int, 8> Mask(MidVT.getVectorNumElements());
7740 for (size_t i = 0; i < Mask.size(); ++i)
7742 return DAG.getNode(ISD::TRUNCATE, dl, VT,
7743 DAG.getVectorShuffle(
7745 DAG.getNode(ISD::BITCAST, dl, MidVT, N00),
7746 DAG.getNode(ISD::BITCAST, dl, MidVT, N10), Mask));
7750 // Wait 'til after everything is legalized to try this. That way we have
7751 // legal vector types and such.
7752 if (DCI.isBeforeLegalizeOps())
7755 // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector
7756 // splat. The indexed instructions are going to be expecting a DUPLANE64, so
7757 // canonicalise to that.
7758 if (N0 == N1 && VT.getVectorNumElements() == 2) {
7759 assert(VT.getVectorElementType().getSizeInBits() == 64);
7760 return DAG.getNode(AArch64ISD::DUPLANE64, dl, VT, WidenVector(N0, DAG),
7761 DAG.getConstant(0, dl, MVT::i64));
7764 // Canonicalise concat_vectors so that the right-hand vector has as few
7765 // bit-casts as possible before its real operation. The primary matching
7766 // destination for these operations will be the narrowing "2" instructions,
7767 // which depend on the operation being performed on this right-hand vector.
7769 // (concat_vectors LHS, (v1i64 (bitconvert (v4i16 RHS))))
7771 // (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS))
7773 if (N1->getOpcode() != ISD::BITCAST)
7775 SDValue RHS = N1->getOperand(0);
7776 MVT RHSTy = RHS.getValueType().getSimpleVT();
7777 // If the RHS is not a vector, this is not the pattern we're looking for.
7778 if (!RHSTy.isVector())
7781 DEBUG(dbgs() << "aarch64-lower: concat_vectors bitcast simplification\n");
7783 MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(),
7784 RHSTy.getVectorNumElements() * 2);
7785 return DAG.getNode(ISD::BITCAST, dl, VT,
7786 DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatTy,
7787 DAG.getNode(ISD::BITCAST, dl, RHSTy, N0),
7791 static SDValue tryCombineFixedPointConvert(SDNode *N,
7792 TargetLowering::DAGCombinerInfo &DCI,
7793 SelectionDAG &DAG) {
7794 // Wait 'til after everything is legalized to try this. That way we have
7795 // legal vector types and such.
7796 if (DCI.isBeforeLegalizeOps())
7798 // Transform a scalar conversion of a value from a lane extract into a
7799 // lane extract of a vector conversion. E.g., from foo1 to foo2:
7800 // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); }
7801 // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; }
7803 // The second form interacts better with instruction selection and the
7804 // register allocator to avoid cross-class register copies that aren't
7805 // coalescable due to a lane reference.
7807 // Check the operand and see if it originates from a lane extract.
7808 SDValue Op1 = N->getOperand(1);
7809 if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7810 // Yep, no additional predication needed. Perform the transform.
7811 SDValue IID = N->getOperand(0);
7812 SDValue Shift = N->getOperand(2);
7813 SDValue Vec = Op1.getOperand(0);
7814 SDValue Lane = Op1.getOperand(1);
7815 EVT ResTy = N->getValueType(0);
7819 // The vector width should be 128 bits by the time we get here, even
7820 // if it started as 64 bits (the extract_vector handling will have
7822 assert(Vec.getValueType().getSizeInBits() == 128 &&
7823 "unexpected vector size on extract_vector_elt!");
7824 if (Vec.getValueType() == MVT::v4i32)
7825 VecResTy = MVT::v4f32;
7826 else if (Vec.getValueType() == MVT::v2i64)
7827 VecResTy = MVT::v2f64;
7829 llvm_unreachable("unexpected vector type!");
7832 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift);
7833 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane);
7838 // AArch64 high-vector "long" operations are formed by performing the non-high
7839 // version on an extract_subvector of each operand which gets the high half:
7841 // (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS))
7843 // However, there are cases which don't have an extract_high explicitly, but
7844 // have another operation that can be made compatible with one for free. For
7847 // (dupv64 scalar) --> (extract_high (dup128 scalar))
7849 // This routine does the actual conversion of such DUPs, once outer routines
7850 // have determined that everything else is in order.
7851 // It also supports immediate DUP-like nodes (MOVI/MVNi), which we can fold
7853 static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) {
7854 switch (N.getOpcode()) {
7855 case AArch64ISD::DUP:
7856 case AArch64ISD::DUPLANE8:
7857 case AArch64ISD::DUPLANE16:
7858 case AArch64ISD::DUPLANE32:
7859 case AArch64ISD::DUPLANE64:
7860 case AArch64ISD::MOVI:
7861 case AArch64ISD::MOVIshift:
7862 case AArch64ISD::MOVIedit:
7863 case AArch64ISD::MOVImsl:
7864 case AArch64ISD::MVNIshift:
7865 case AArch64ISD::MVNImsl:
7868 // FMOV could be supported, but isn't very useful, as it would only occur
7869 // if you passed a bitcast' floating point immediate to an eligible long
7870 // integer op (addl, smull, ...).
7874 MVT NarrowTy = N.getSimpleValueType();
7875 if (!NarrowTy.is64BitVector())
7878 MVT ElementTy = NarrowTy.getVectorElementType();
7879 unsigned NumElems = NarrowTy.getVectorNumElements();
7880 MVT NewVT = MVT::getVectorVT(ElementTy, NumElems * 2);
7883 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NarrowTy,
7884 DAG.getNode(N->getOpcode(), dl, NewVT, N->ops()),
7885 DAG.getConstant(NumElems, dl, MVT::i64));
7888 static bool isEssentiallyExtractSubvector(SDValue N) {
7889 if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR)
7892 return N.getOpcode() == ISD::BITCAST &&
7893 N.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR;
7896 /// \brief Helper structure to keep track of ISD::SET_CC operands.
7897 struct GenericSetCCInfo {
7898 const SDValue *Opnd0;
7899 const SDValue *Opnd1;
7903 /// \brief Helper structure to keep track of a SET_CC lowered into AArch64 code.
7904 struct AArch64SetCCInfo {
7906 AArch64CC::CondCode CC;
7909 /// \brief Helper structure to keep track of SetCC information.
7911 GenericSetCCInfo Generic;
7912 AArch64SetCCInfo AArch64;
7915 /// \brief Helper structure to be able to read SetCC information. If set to
7916 /// true, IsAArch64 field, Info is a AArch64SetCCInfo, otherwise Info is a
7917 /// GenericSetCCInfo.
7918 struct SetCCInfoAndKind {
7923 /// \brief Check whether or not \p Op is a SET_CC operation, either a generic or
7925 /// AArch64 lowered one.
7926 /// \p SetCCInfo is filled accordingly.
7927 /// \post SetCCInfo is meanginfull only when this function returns true.
7928 /// \return True when Op is a kind of SET_CC operation.
7929 static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) {
7930 // If this is a setcc, this is straight forward.
7931 if (Op.getOpcode() == ISD::SETCC) {
7932 SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0);
7933 SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1);
7934 SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
7935 SetCCInfo.IsAArch64 = false;
7938 // Otherwise, check if this is a matching csel instruction.
7942 if (Op.getOpcode() != AArch64ISD::CSEL)
7944 // Set the information about the operands.
7945 // TODO: we want the operands of the Cmp not the csel
7946 SetCCInfo.Info.AArch64.Cmp = &Op.getOperand(3);
7947 SetCCInfo.IsAArch64 = true;
7948 SetCCInfo.Info.AArch64.CC = static_cast<AArch64CC::CondCode>(
7949 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
7951 // Check that the operands matches the constraints:
7952 // (1) Both operands must be constants.
7953 // (2) One must be 1 and the other must be 0.
7954 ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0));
7955 ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7958 if (!TValue || !FValue)
7962 if (!TValue->isOne()) {
7963 // Update the comparison when we are interested in !cc.
7964 std::swap(TValue, FValue);
7965 SetCCInfo.Info.AArch64.CC =
7966 AArch64CC::getInvertedCondCode(SetCCInfo.Info.AArch64.CC);
7968 return TValue->isOne() && FValue->isNullValue();
7971 // Returns true if Op is setcc or zext of setcc.
7972 static bool isSetCCOrZExtSetCC(const SDValue& Op, SetCCInfoAndKind &Info) {
7973 if (isSetCC(Op, Info))
7975 return ((Op.getOpcode() == ISD::ZERO_EXTEND) &&
7976 isSetCC(Op->getOperand(0), Info));
7979 // The folding we want to perform is:
7980 // (add x, [zext] (setcc cc ...) )
7982 // (csel x, (add x, 1), !cc ...)
7984 // The latter will get matched to a CSINC instruction.
7985 static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) {
7986 assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!");
7987 SDValue LHS = Op->getOperand(0);
7988 SDValue RHS = Op->getOperand(1);
7989 SetCCInfoAndKind InfoAndKind;
7991 // If neither operand is a SET_CC, give up.
7992 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind)) {
7993 std::swap(LHS, RHS);
7994 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind))
7998 // FIXME: This could be generatized to work for FP comparisons.
7999 EVT CmpVT = InfoAndKind.IsAArch64
8000 ? InfoAndKind.Info.AArch64.Cmp->getOperand(0).getValueType()
8001 : InfoAndKind.Info.Generic.Opnd0->getValueType();
8002 if (CmpVT != MVT::i32 && CmpVT != MVT::i64)
8008 if (InfoAndKind.IsAArch64) {
8009 CCVal = DAG.getConstant(
8010 AArch64CC::getInvertedCondCode(InfoAndKind.Info.AArch64.CC), dl,
8012 Cmp = *InfoAndKind.Info.AArch64.Cmp;
8014 Cmp = getAArch64Cmp(*InfoAndKind.Info.Generic.Opnd0,
8015 *InfoAndKind.Info.Generic.Opnd1,
8016 ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, true),
8019 EVT VT = Op->getValueType(0);
8020 LHS = DAG.getNode(ISD::ADD, dl, VT, RHS, DAG.getConstant(1, dl, VT));
8021 return DAG.getNode(AArch64ISD::CSEL, dl, VT, RHS, LHS, CCVal, Cmp);
8024 // The basic add/sub long vector instructions have variants with "2" on the end
8025 // which act on the high-half of their inputs. They are normally matched by
8028 // (add (zeroext (extract_high LHS)),
8029 // (zeroext (extract_high RHS)))
8030 // -> uaddl2 vD, vN, vM
8032 // However, if one of the extracts is something like a duplicate, this
8033 // instruction can still be used profitably. This function puts the DAG into a
8034 // more appropriate form for those patterns to trigger.
8035 static SDValue performAddSubLongCombine(SDNode *N,
8036 TargetLowering::DAGCombinerInfo &DCI,
8037 SelectionDAG &DAG) {
8038 if (DCI.isBeforeLegalizeOps())
8041 MVT VT = N->getSimpleValueType(0);
8042 if (!VT.is128BitVector()) {
8043 if (N->getOpcode() == ISD::ADD)
8044 return performSetccAddFolding(N, DAG);
8048 // Make sure both branches are extended in the same way.
8049 SDValue LHS = N->getOperand(0);
8050 SDValue RHS = N->getOperand(1);
8051 if ((LHS.getOpcode() != ISD::ZERO_EXTEND &&
8052 LHS.getOpcode() != ISD::SIGN_EXTEND) ||
8053 LHS.getOpcode() != RHS.getOpcode())
8056 unsigned ExtType = LHS.getOpcode();
8058 // It's not worth doing if at least one of the inputs isn't already an
8059 // extract, but we don't know which it'll be so we have to try both.
8060 if (isEssentiallyExtractSubvector(LHS.getOperand(0))) {
8061 RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG);
8065 RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS);
8066 } else if (isEssentiallyExtractSubvector(RHS.getOperand(0))) {
8067 LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG);
8071 LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS);
8074 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS);
8077 // Massage DAGs which we can use the high-half "long" operations on into
8078 // something isel will recognize better. E.g.
8080 // (aarch64_neon_umull (extract_high vec) (dupv64 scalar)) -->
8081 // (aarch64_neon_umull (extract_high (v2i64 vec)))
8082 // (extract_high (v2i64 (dup128 scalar)))))
8084 static SDValue tryCombineLongOpWithDup(SDNode *N,
8085 TargetLowering::DAGCombinerInfo &DCI,
8086 SelectionDAG &DAG) {
8087 if (DCI.isBeforeLegalizeOps())
8090 bool IsIntrinsic = N->getOpcode() == ISD::INTRINSIC_WO_CHAIN;
8091 SDValue LHS = N->getOperand(IsIntrinsic ? 1 : 0);
8092 SDValue RHS = N->getOperand(IsIntrinsic ? 2 : 1);
8093 assert(LHS.getValueType().is64BitVector() &&
8094 RHS.getValueType().is64BitVector() &&
8095 "unexpected shape for long operation");
8097 // Either node could be a DUP, but it's not worth doing both of them (you'd
8098 // just as well use the non-high version) so look for a corresponding extract
8099 // operation on the other "wing".
8100 if (isEssentiallyExtractSubvector(LHS)) {
8101 RHS = tryExtendDUPToExtractHigh(RHS, DAG);
8104 } else if (isEssentiallyExtractSubvector(RHS)) {
8105 LHS = tryExtendDUPToExtractHigh(LHS, DAG);
8110 // N could either be an intrinsic or a sabsdiff/uabsdiff node.
8112 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0),
8113 N->getOperand(0), LHS, RHS);
8115 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
8119 static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
8120 MVT ElemTy = N->getSimpleValueType(0).getScalarType();
8121 unsigned ElemBits = ElemTy.getSizeInBits();
8123 int64_t ShiftAmount;
8124 if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) {
8125 APInt SplatValue, SplatUndef;
8126 unsigned SplatBitSize;
8128 if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
8129 HasAnyUndefs, ElemBits) ||
8130 SplatBitSize != ElemBits)
8133 ShiftAmount = SplatValue.getSExtValue();
8134 } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) {
8135 ShiftAmount = CVN->getSExtValue();
8143 llvm_unreachable("Unknown shift intrinsic");
8144 case Intrinsic::aarch64_neon_sqshl:
8145 Opcode = AArch64ISD::SQSHL_I;
8146 IsRightShift = false;
8148 case Intrinsic::aarch64_neon_uqshl:
8149 Opcode = AArch64ISD::UQSHL_I;
8150 IsRightShift = false;
8152 case Intrinsic::aarch64_neon_srshl:
8153 Opcode = AArch64ISD::SRSHR_I;
8154 IsRightShift = true;
8156 case Intrinsic::aarch64_neon_urshl:
8157 Opcode = AArch64ISD::URSHR_I;
8158 IsRightShift = true;
8160 case Intrinsic::aarch64_neon_sqshlu:
8161 Opcode = AArch64ISD::SQSHLU_I;
8162 IsRightShift = false;
8166 if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits) {
8168 return DAG.getNode(Opcode, dl, N->getValueType(0), N->getOperand(1),
8169 DAG.getConstant(-ShiftAmount, dl, MVT::i32));
8170 } else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount < ElemBits) {
8172 return DAG.getNode(Opcode, dl, N->getValueType(0), N->getOperand(1),
8173 DAG.getConstant(ShiftAmount, dl, MVT::i32));
8179 // The CRC32[BH] instructions ignore the high bits of their data operand. Since
8180 // the intrinsics must be legal and take an i32, this means there's almost
8181 // certainly going to be a zext in the DAG which we can eliminate.
8182 static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) {
8183 SDValue AndN = N->getOperand(2);
8184 if (AndN.getOpcode() != ISD::AND)
8187 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1));
8188 if (!CMask || CMask->getZExtValue() != Mask)
8191 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32,
8192 N->getOperand(0), N->getOperand(1), AndN.getOperand(0));
8195 static SDValue combineAcrossLanesIntrinsic(unsigned Opc, SDNode *N,
8196 SelectionDAG &DAG) {
8198 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0),
8199 DAG.getNode(Opc, dl,
8200 N->getOperand(1).getSimpleValueType(),
8202 DAG.getConstant(0, dl, MVT::i64));
8205 static SDValue performIntrinsicCombine(SDNode *N,
8206 TargetLowering::DAGCombinerInfo &DCI,
8207 const AArch64Subtarget *Subtarget) {
8208 SelectionDAG &DAG = DCI.DAG;
8209 unsigned IID = getIntrinsicID(N);
8213 case Intrinsic::aarch64_neon_vcvtfxs2fp:
8214 case Intrinsic::aarch64_neon_vcvtfxu2fp:
8215 return tryCombineFixedPointConvert(N, DCI, DAG);
8216 case Intrinsic::aarch64_neon_saddv:
8217 return combineAcrossLanesIntrinsic(AArch64ISD::SADDV, N, DAG);
8218 case Intrinsic::aarch64_neon_uaddv:
8219 return combineAcrossLanesIntrinsic(AArch64ISD::UADDV, N, DAG);
8220 case Intrinsic::aarch64_neon_sminv:
8221 return combineAcrossLanesIntrinsic(AArch64ISD::SMINV, N, DAG);
8222 case Intrinsic::aarch64_neon_uminv:
8223 return combineAcrossLanesIntrinsic(AArch64ISD::UMINV, N, DAG);
8224 case Intrinsic::aarch64_neon_smaxv:
8225 return combineAcrossLanesIntrinsic(AArch64ISD::SMAXV, N, DAG);
8226 case Intrinsic::aarch64_neon_umaxv:
8227 return combineAcrossLanesIntrinsic(AArch64ISD::UMAXV, N, DAG);
8228 case Intrinsic::aarch64_neon_fmax:
8229 return DAG.getNode(ISD::FMAXNAN, SDLoc(N), N->getValueType(0),
8230 N->getOperand(1), N->getOperand(2));
8231 case Intrinsic::aarch64_neon_fmin:
8232 return DAG.getNode(ISD::FMINNAN, SDLoc(N), N->getValueType(0),
8233 N->getOperand(1), N->getOperand(2));
8234 case Intrinsic::aarch64_neon_sabd:
8235 return DAG.getNode(ISD::SABSDIFF, SDLoc(N), N->getValueType(0),
8236 N->getOperand(1), N->getOperand(2));
8237 case Intrinsic::aarch64_neon_uabd:
8238 return DAG.getNode(ISD::UABSDIFF, SDLoc(N), N->getValueType(0),
8239 N->getOperand(1), N->getOperand(2));
8240 case Intrinsic::aarch64_neon_fmaxnm:
8241 return DAG.getNode(ISD::FMAXNUM, SDLoc(N), N->getValueType(0),
8242 N->getOperand(1), N->getOperand(2));
8243 case Intrinsic::aarch64_neon_fminnm:
8244 return DAG.getNode(ISD::FMINNUM, SDLoc(N), N->getValueType(0),
8245 N->getOperand(1), N->getOperand(2));
8246 case Intrinsic::aarch64_neon_smull:
8247 case Intrinsic::aarch64_neon_umull:
8248 case Intrinsic::aarch64_neon_pmull:
8249 case Intrinsic::aarch64_neon_sqdmull:
8250 return tryCombineLongOpWithDup(N, DCI, DAG);
8251 case Intrinsic::aarch64_neon_sqshl:
8252 case Intrinsic::aarch64_neon_uqshl:
8253 case Intrinsic::aarch64_neon_sqshlu:
8254 case Intrinsic::aarch64_neon_srshl:
8255 case Intrinsic::aarch64_neon_urshl:
8256 return tryCombineShiftImm(IID, N, DAG);
8257 case Intrinsic::aarch64_crc32b:
8258 case Intrinsic::aarch64_crc32cb:
8259 return tryCombineCRC32(0xff, N, DAG);
8260 case Intrinsic::aarch64_crc32h:
8261 case Intrinsic::aarch64_crc32ch:
8262 return tryCombineCRC32(0xffff, N, DAG);
8267 static SDValue performExtendCombine(SDNode *N,
8268 TargetLowering::DAGCombinerInfo &DCI,
8269 SelectionDAG &DAG) {
8270 // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then
8271 // we can convert that DUP into another extract_high (of a bigger DUP), which
8272 // helps the backend to decide that an sabdl2 would be useful, saving a real
8273 // extract_high operation.
8274 if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND &&
8275 (N->getOperand(0).getOpcode() == ISD::SABSDIFF ||
8276 N->getOperand(0).getOpcode() == ISD::UABSDIFF)) {
8277 SDNode *ABDNode = N->getOperand(0).getNode();
8278 SDValue NewABD = tryCombineLongOpWithDup(ABDNode, DCI, DAG);
8279 if (!NewABD.getNode())
8282 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0),
8286 // This is effectively a custom type legalization for AArch64.
8288 // Type legalization will split an extend of a small, legal, type to a larger
8289 // illegal type by first splitting the destination type, often creating
8290 // illegal source types, which then get legalized in isel-confusing ways,
8291 // leading to really terrible codegen. E.g.,
8292 // %result = v8i32 sext v8i8 %value
8294 // %losrc = extract_subreg %value, ...
8295 // %hisrc = extract_subreg %value, ...
8296 // %lo = v4i32 sext v4i8 %losrc
8297 // %hi = v4i32 sext v4i8 %hisrc
8298 // Things go rapidly downhill from there.
8300 // For AArch64, the [sz]ext vector instructions can only go up one element
8301 // size, so we can, e.g., extend from i8 to i16, but to go from i8 to i32
8302 // take two instructions.
8304 // This implies that the most efficient way to do the extend from v8i8
8305 // to two v4i32 values is to first extend the v8i8 to v8i16, then do
8306 // the normal splitting to happen for the v8i16->v8i32.
8308 // This is pre-legalization to catch some cases where the default
8309 // type legalization will create ill-tempered code.
8310 if (!DCI.isBeforeLegalizeOps())
8313 // We're only interested in cleaning things up for non-legal vector types
8314 // here. If both the source and destination are legal, things will just
8315 // work naturally without any fiddling.
8316 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8317 EVT ResVT = N->getValueType(0);
8318 if (!ResVT.isVector() || TLI.isTypeLegal(ResVT))
8320 // If the vector type isn't a simple VT, it's beyond the scope of what
8321 // we're worried about here. Let legalization do its thing and hope for
8323 SDValue Src = N->getOperand(0);
8324 EVT SrcVT = Src->getValueType(0);
8325 if (!ResVT.isSimple() || !SrcVT.isSimple())
8328 // If the source VT is a 64-bit vector, we can play games and get the
8329 // better results we want.
8330 if (SrcVT.getSizeInBits() != 64)
8333 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
8334 unsigned ElementCount = SrcVT.getVectorNumElements();
8335 SrcVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize * 2), ElementCount);
8337 Src = DAG.getNode(N->getOpcode(), DL, SrcVT, Src);
8339 // Now split the rest of the operation into two halves, each with a 64
8343 unsigned NumElements = ResVT.getVectorNumElements();
8344 assert(!(NumElements & 1) && "Splitting vector, but not in half!");
8345 LoVT = HiVT = EVT::getVectorVT(*DAG.getContext(),
8346 ResVT.getVectorElementType(), NumElements / 2);
8348 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
8349 LoVT.getVectorNumElements());
8350 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
8351 DAG.getConstant(0, DL, MVT::i64));
8352 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
8353 DAG.getConstant(InNVT.getVectorNumElements(), DL, MVT::i64));
8354 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, Lo);
8355 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, Hi);
8357 // Now combine the parts back together so we still have a single result
8358 // like the combiner expects.
8359 return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
8362 /// Replace a splat of a scalar to a vector store by scalar stores of the scalar
8363 /// value. The load store optimizer pass will merge them to store pair stores.
8364 /// This has better performance than a splat of the scalar followed by a split
8365 /// vector store. Even if the stores are not merged it is four stores vs a dup,
8366 /// followed by an ext.b and two stores.
8367 static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) {
8368 SDValue StVal = St->getValue();
8369 EVT VT = StVal.getValueType();
8371 // Don't replace floating point stores, they possibly won't be transformed to
8372 // stp because of the store pair suppress pass.
8373 if (VT.isFloatingPoint())
8376 // Check for insert vector elements.
8377 if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT)
8380 // We can express a splat as store pair(s) for 2 or 4 elements.
8381 unsigned NumVecElts = VT.getVectorNumElements();
8382 if (NumVecElts != 4 && NumVecElts != 2)
8384 SDValue SplatVal = StVal.getOperand(1);
8385 unsigned RemainInsertElts = NumVecElts - 1;
8387 // Check that this is a splat.
8388 while (--RemainInsertElts) {
8389 SDValue NextInsertElt = StVal.getOperand(0);
8390 if (NextInsertElt.getOpcode() != ISD::INSERT_VECTOR_ELT)
8392 if (NextInsertElt.getOperand(1) != SplatVal)
8394 StVal = NextInsertElt;
8396 unsigned OrigAlignment = St->getAlignment();
8397 unsigned EltOffset = NumVecElts == 4 ? 4 : 8;
8398 unsigned Alignment = std::min(OrigAlignment, EltOffset);
8400 // Create scalar stores. This is at least as good as the code sequence for a
8401 // split unaligned store which is a dup.s, ext.b, and two stores.
8402 // Most of the time the three stores should be replaced by store pair
8403 // instructions (stp).
8405 SDValue BasePtr = St->getBasePtr();
8407 DAG.getStore(St->getChain(), DL, SplatVal, BasePtr, St->getPointerInfo(),
8408 St->isVolatile(), St->isNonTemporal(), St->getAlignment());
8410 unsigned Offset = EltOffset;
8411 while (--NumVecElts) {
8412 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
8413 DAG.getConstant(Offset, DL, MVT::i64));
8414 NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr,
8415 St->getPointerInfo(), St->isVolatile(),
8416 St->isNonTemporal(), Alignment);
8417 Offset += EltOffset;
8422 static SDValue performSTORECombine(SDNode *N,
8423 TargetLowering::DAGCombinerInfo &DCI,
8425 const AArch64Subtarget *Subtarget) {
8426 if (!DCI.isBeforeLegalize())
8429 StoreSDNode *S = cast<StoreSDNode>(N);
8430 if (S->isVolatile())
8433 // Cyclone has bad performance on unaligned 16B stores when crossing line and
8434 // page boundaries. We want to split such stores.
8435 if (!Subtarget->isCyclone())
8438 // Don't split at -Oz.
8439 if (DAG.getMachineFunction().getFunction()->optForMinSize())
8442 SDValue StVal = S->getValue();
8443 EVT VT = StVal.getValueType();
8445 // Don't split v2i64 vectors. Memcpy lowering produces those and splitting
8446 // those up regresses performance on micro-benchmarks and olden/bh.
8447 if (!VT.isVector() || VT.getVectorNumElements() < 2 || VT == MVT::v2i64)
8450 // Split unaligned 16B stores. They are terrible for performance.
8451 // Don't split stores with alignment of 1 or 2. Code that uses clang vector
8452 // extensions can use this to mark that it does not want splitting to happen
8453 // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of
8454 // eliminating alignment hazards is only 1 in 8 for alignment of 2.
8455 if (VT.getSizeInBits() != 128 || S->getAlignment() >= 16 ||
8456 S->getAlignment() <= 2)
8459 // If we get a splat of a scalar convert this vector store to a store of
8460 // scalars. They will be merged into store pairs thereby removing two
8462 if (SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S))
8463 return ReplacedSplat;
8466 unsigned NumElts = VT.getVectorNumElements() / 2;
8467 // Split VT into two.
8469 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), NumElts);
8470 SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
8471 DAG.getConstant(0, DL, MVT::i64));
8472 SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
8473 DAG.getConstant(NumElts, DL, MVT::i64));
8474 SDValue BasePtr = S->getBasePtr();
8476 DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
8477 S->isVolatile(), S->isNonTemporal(), S->getAlignment());
8478 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
8479 DAG.getConstant(8, DL, MVT::i64));
8480 return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr,
8481 S->getPointerInfo(), S->isVolatile(), S->isNonTemporal(),
8485 /// Target-specific DAG combine function for post-increment LD1 (lane) and
8486 /// post-increment LD1R.
8487 static SDValue performPostLD1Combine(SDNode *N,
8488 TargetLowering::DAGCombinerInfo &DCI,
8490 if (DCI.isBeforeLegalizeOps())
8493 SelectionDAG &DAG = DCI.DAG;
8494 EVT VT = N->getValueType(0);
8496 unsigned LoadIdx = IsLaneOp ? 1 : 0;
8497 SDNode *LD = N->getOperand(LoadIdx).getNode();
8498 // If it is not LOAD, can not do such combine.
8499 if (LD->getOpcode() != ISD::LOAD)
8502 LoadSDNode *LoadSDN = cast<LoadSDNode>(LD);
8503 EVT MemVT = LoadSDN->getMemoryVT();
8504 // Check if memory operand is the same type as the vector element.
8505 if (MemVT != VT.getVectorElementType())
8508 // Check if there are other uses. If so, do not combine as it will introduce
8510 for (SDNode::use_iterator UI = LD->use_begin(), UE = LD->use_end(); UI != UE;
8512 if (UI.getUse().getResNo() == 1) // Ignore uses of the chain result.
8518 SDValue Addr = LD->getOperand(1);
8519 SDValue Vector = N->getOperand(0);
8520 // Search for a use of the address operand that is an increment.
8521 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(), UE =
8522 Addr.getNode()->use_end(); UI != UE; ++UI) {
8524 if (User->getOpcode() != ISD::ADD
8525 || UI.getUse().getResNo() != Addr.getResNo())
8528 // Check that the add is independent of the load. Otherwise, folding it
8529 // would create a cycle.
8530 if (User->isPredecessorOf(LD) || LD->isPredecessorOf(User))
8532 // Also check that add is not used in the vector operand. This would also
8534 if (User->isPredecessorOf(Vector.getNode()))
8537 // If the increment is a constant, it must match the memory ref size.
8538 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8539 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8540 uint32_t IncVal = CInc->getZExtValue();
8541 unsigned NumBytes = VT.getScalarSizeInBits() / 8;
8542 if (IncVal != NumBytes)
8544 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
8547 // Finally, check that the vector doesn't depend on the load.
8548 // Again, this would create a cycle.
8549 // The load depending on the vector is fine, as that's the case for the
8550 // LD1*post we'll eventually generate anyway.
8551 if (LoadSDN->isPredecessorOf(Vector.getNode()))
8554 SmallVector<SDValue, 8> Ops;
8555 Ops.push_back(LD->getOperand(0)); // Chain
8557 Ops.push_back(Vector); // The vector to be inserted
8558 Ops.push_back(N->getOperand(2)); // The lane to be inserted in the vector
8560 Ops.push_back(Addr);
8563 EVT Tys[3] = { VT, MVT::i64, MVT::Other };
8564 SDVTList SDTys = DAG.getVTList(Tys);
8565 unsigned NewOp = IsLaneOp ? AArch64ISD::LD1LANEpost : AArch64ISD::LD1DUPpost;
8566 SDValue UpdN = DAG.getMemIntrinsicNode(NewOp, SDLoc(N), SDTys, Ops,
8568 LoadSDN->getMemOperand());
8571 SmallVector<SDValue, 2> NewResults;
8572 NewResults.push_back(SDValue(LD, 0)); // The result of load
8573 NewResults.push_back(SDValue(UpdN.getNode(), 2)); // Chain
8574 DCI.CombineTo(LD, NewResults);
8575 DCI.CombineTo(N, SDValue(UpdN.getNode(), 0)); // Dup/Inserted Result
8576 DCI.CombineTo(User, SDValue(UpdN.getNode(), 1)); // Write back register
8583 /// Target-specific DAG combine function for NEON load/store intrinsics
8584 /// to merge base address updates.
8585 static SDValue performNEONPostLDSTCombine(SDNode *N,
8586 TargetLowering::DAGCombinerInfo &DCI,
8587 SelectionDAG &DAG) {
8588 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8591 unsigned AddrOpIdx = N->getNumOperands() - 1;
8592 SDValue Addr = N->getOperand(AddrOpIdx);
8594 // Search for a use of the address operand that is an increment.
8595 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8596 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8598 if (User->getOpcode() != ISD::ADD ||
8599 UI.getUse().getResNo() != Addr.getResNo())
8602 // Check that the add is independent of the load/store. Otherwise, folding
8603 // it would create a cycle.
8604 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8607 // Find the new opcode for the updating load/store.
8608 bool IsStore = false;
8609 bool IsLaneOp = false;
8610 bool IsDupOp = false;
8611 unsigned NewOpc = 0;
8612 unsigned NumVecs = 0;
8613 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8615 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8616 case Intrinsic::aarch64_neon_ld2: NewOpc = AArch64ISD::LD2post;
8618 case Intrinsic::aarch64_neon_ld3: NewOpc = AArch64ISD::LD3post;
8620 case Intrinsic::aarch64_neon_ld4: NewOpc = AArch64ISD::LD4post;
8622 case Intrinsic::aarch64_neon_st2: NewOpc = AArch64ISD::ST2post;
8623 NumVecs = 2; IsStore = true; break;
8624 case Intrinsic::aarch64_neon_st3: NewOpc = AArch64ISD::ST3post;
8625 NumVecs = 3; IsStore = true; break;
8626 case Intrinsic::aarch64_neon_st4: NewOpc = AArch64ISD::ST4post;
8627 NumVecs = 4; IsStore = true; break;
8628 case Intrinsic::aarch64_neon_ld1x2: NewOpc = AArch64ISD::LD1x2post;
8630 case Intrinsic::aarch64_neon_ld1x3: NewOpc = AArch64ISD::LD1x3post;
8632 case Intrinsic::aarch64_neon_ld1x4: NewOpc = AArch64ISD::LD1x4post;
8634 case Intrinsic::aarch64_neon_st1x2: NewOpc = AArch64ISD::ST1x2post;
8635 NumVecs = 2; IsStore = true; break;
8636 case Intrinsic::aarch64_neon_st1x3: NewOpc = AArch64ISD::ST1x3post;
8637 NumVecs = 3; IsStore = true; break;
8638 case Intrinsic::aarch64_neon_st1x4: NewOpc = AArch64ISD::ST1x4post;
8639 NumVecs = 4; IsStore = true; break;
8640 case Intrinsic::aarch64_neon_ld2r: NewOpc = AArch64ISD::LD2DUPpost;
8641 NumVecs = 2; IsDupOp = true; break;
8642 case Intrinsic::aarch64_neon_ld3r: NewOpc = AArch64ISD::LD3DUPpost;
8643 NumVecs = 3; IsDupOp = true; break;
8644 case Intrinsic::aarch64_neon_ld4r: NewOpc = AArch64ISD::LD4DUPpost;
8645 NumVecs = 4; IsDupOp = true; break;
8646 case Intrinsic::aarch64_neon_ld2lane: NewOpc = AArch64ISD::LD2LANEpost;
8647 NumVecs = 2; IsLaneOp = true; break;
8648 case Intrinsic::aarch64_neon_ld3lane: NewOpc = AArch64ISD::LD3LANEpost;
8649 NumVecs = 3; IsLaneOp = true; break;
8650 case Intrinsic::aarch64_neon_ld4lane: NewOpc = AArch64ISD::LD4LANEpost;
8651 NumVecs = 4; IsLaneOp = true; break;
8652 case Intrinsic::aarch64_neon_st2lane: NewOpc = AArch64ISD::ST2LANEpost;
8653 NumVecs = 2; IsStore = true; IsLaneOp = true; break;
8654 case Intrinsic::aarch64_neon_st3lane: NewOpc = AArch64ISD::ST3LANEpost;
8655 NumVecs = 3; IsStore = true; IsLaneOp = true; break;
8656 case Intrinsic::aarch64_neon_st4lane: NewOpc = AArch64ISD::ST4LANEpost;
8657 NumVecs = 4; IsStore = true; IsLaneOp = true; break;
8662 VecTy = N->getOperand(2).getValueType();
8664 VecTy = N->getValueType(0);
8666 // If the increment is a constant, it must match the memory ref size.
8667 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8668 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8669 uint32_t IncVal = CInc->getZExtValue();
8670 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8671 if (IsLaneOp || IsDupOp)
8672 NumBytes /= VecTy.getVectorNumElements();
8673 if (IncVal != NumBytes)
8675 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
8677 SmallVector<SDValue, 8> Ops;
8678 Ops.push_back(N->getOperand(0)); // Incoming chain
8679 // Load lane and store have vector list as input.
8680 if (IsLaneOp || IsStore)
8681 for (unsigned i = 2; i < AddrOpIdx; ++i)
8682 Ops.push_back(N->getOperand(i));
8683 Ops.push_back(Addr); // Base register
8688 unsigned NumResultVecs = (IsStore ? 0 : NumVecs);
8690 for (n = 0; n < NumResultVecs; ++n)
8692 Tys[n++] = MVT::i64; // Type of write back register
8693 Tys[n] = MVT::Other; // Type of the chain
8694 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs + 2));
8696 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8697 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys, Ops,
8698 MemInt->getMemoryVT(),
8699 MemInt->getMemOperand());
8702 std::vector<SDValue> NewResults;
8703 for (unsigned i = 0; i < NumResultVecs; ++i) {
8704 NewResults.push_back(SDValue(UpdN.getNode(), i));
8706 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs + 1));
8707 DCI.CombineTo(N, NewResults);
8708 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8715 // Checks to see if the value is the prescribed width and returns information
8716 // about its extension mode.
8718 bool checkValueWidth(SDValue V, unsigned width, ISD::LoadExtType &ExtType) {
8719 ExtType = ISD::NON_EXTLOAD;
8720 switch(V.getNode()->getOpcode()) {
8724 LoadSDNode *LoadNode = cast<LoadSDNode>(V.getNode());
8725 if ((LoadNode->getMemoryVT() == MVT::i8 && width == 8)
8726 || (LoadNode->getMemoryVT() == MVT::i16 && width == 16)) {
8727 ExtType = LoadNode->getExtensionType();
8732 case ISD::AssertSext: {
8733 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
8734 if ((TypeNode->getVT() == MVT::i8 && width == 8)
8735 || (TypeNode->getVT() == MVT::i16 && width == 16)) {
8736 ExtType = ISD::SEXTLOAD;
8741 case ISD::AssertZext: {
8742 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
8743 if ((TypeNode->getVT() == MVT::i8 && width == 8)
8744 || (TypeNode->getVT() == MVT::i16 && width == 16)) {
8745 ExtType = ISD::ZEXTLOAD;
8751 case ISD::TargetConstant: {
8752 if (std::abs(cast<ConstantSDNode>(V.getNode())->getSExtValue()) <
8762 // This function does a whole lot of voodoo to determine if the tests are
8763 // equivalent without and with a mask. Essentially what happens is that given a
8766 // +-------------+ +-------------+ +-------------+ +-------------+
8767 // | Input | | AddConstant | | CompConstant| | CC |
8768 // +-------------+ +-------------+ +-------------+ +-------------+
8770 // V V | +----------+
8771 // +-------------+ +----+ | |
8772 // | ADD | |0xff| | |
8773 // +-------------+ +----+ | |
8776 // +-------------+ | |
8778 // +-------------+ | |
8787 // The AND node may be safely removed for some combinations of inputs. In
8788 // particular we need to take into account the extension type of the Input,
8789 // the exact values of AddConstant, CompConstant, and CC, along with the nominal
8790 // width of the input (this can work for any width inputs, the above graph is
8791 // specific to 8 bits.
8793 // The specific equations were worked out by generating output tables for each
8794 // AArch64CC value in terms of and AddConstant (w1), CompConstant(w2). The
8795 // problem was simplified by working with 4 bit inputs, which means we only
8796 // needed to reason about 24 distinct bit patterns: 8 patterns unique to zero
8797 // extension (8,15), 8 patterns unique to sign extensions (-8,-1), and 8
8798 // patterns present in both extensions (0,7). For every distinct set of
8799 // AddConstant and CompConstants bit patterns we can consider the masked and
8800 // unmasked versions to be equivalent if the result of this function is true for
8801 // all 16 distinct bit patterns of for the current extension type of Input (w0).
8804 // and w10, w8, #0x0f
8806 // cset w9, AArch64CC
8808 // cset w11, AArch64CC
8813 // Since the above function shows when the outputs are equivalent it defines
8814 // when it is safe to remove the AND. Unfortunately it only runs on AArch64 and
8815 // would be expensive to run during compiles. The equations below were written
8816 // in a test harness that confirmed they gave equivalent outputs to the above
8817 // for all inputs function, so they can be used determine if the removal is
8820 // isEquivalentMaskless() is the code for testing if the AND can be removed
8821 // factored out of the DAG recognition as the DAG can take several forms.
8824 bool isEquivalentMaskless(unsigned CC, unsigned width,
8825 ISD::LoadExtType ExtType, signed AddConstant,
8826 signed CompConstant) {
8827 // By being careful about our equations and only writing the in term
8828 // symbolic values and well known constants (0, 1, -1, MaxUInt) we can
8829 // make them generally applicable to all bit widths.
8830 signed MaxUInt = (1 << width);
8832 // For the purposes of these comparisons sign extending the type is
8833 // equivalent to zero extending the add and displacing it by half the integer
8834 // width. Provided we are careful and make sure our equations are valid over
8835 // the whole range we can just adjust the input and avoid writing equations
8836 // for sign extended inputs.
8837 if (ExtType == ISD::SEXTLOAD)
8838 AddConstant -= (1 << (width-1));
8842 case AArch64CC::GT: {
8843 if ((AddConstant == 0) ||
8844 (CompConstant == MaxUInt - 1 && AddConstant < 0) ||
8845 (AddConstant >= 0 && CompConstant < 0) ||
8846 (AddConstant <= 0 && CompConstant <= 0 && CompConstant < AddConstant))
8850 case AArch64CC::GE: {
8851 if ((AddConstant == 0) ||
8852 (AddConstant >= 0 && CompConstant <= 0) ||
8853 (AddConstant <= 0 && CompConstant <= 0 && CompConstant <= AddConstant))
8857 case AArch64CC::LS: {
8858 if ((AddConstant >= 0 && CompConstant < 0) ||
8859 (AddConstant <= 0 && CompConstant >= -1 &&
8860 CompConstant < AddConstant + MaxUInt))
8864 case AArch64CC::MI: {
8865 if ((AddConstant == 0) ||
8866 (AddConstant > 0 && CompConstant <= 0) ||
8867 (AddConstant < 0 && CompConstant <= AddConstant))
8871 case AArch64CC::HS: {
8872 if ((AddConstant >= 0 && CompConstant <= 0) ||
8873 (AddConstant <= 0 && CompConstant >= 0 &&
8874 CompConstant <= AddConstant + MaxUInt))
8878 case AArch64CC::NE: {
8879 if ((AddConstant > 0 && CompConstant < 0) ||
8880 (AddConstant < 0 && CompConstant >= 0 &&
8881 CompConstant < AddConstant + MaxUInt) ||
8882 (AddConstant >= 0 && CompConstant >= 0 &&
8883 CompConstant >= AddConstant) ||
8884 (AddConstant <= 0 && CompConstant < 0 && CompConstant < AddConstant))
8893 case AArch64CC::Invalid:
8901 SDValue performCONDCombine(SDNode *N,
8902 TargetLowering::DAGCombinerInfo &DCI,
8903 SelectionDAG &DAG, unsigned CCIndex,
8904 unsigned CmpIndex) {
8905 unsigned CC = cast<ConstantSDNode>(N->getOperand(CCIndex))->getSExtValue();
8906 SDNode *SubsNode = N->getOperand(CmpIndex).getNode();
8907 unsigned CondOpcode = SubsNode->getOpcode();
8909 if (CondOpcode != AArch64ISD::SUBS)
8912 // There is a SUBS feeding this condition. Is it fed by a mask we can
8915 SDNode *AndNode = SubsNode->getOperand(0).getNode();
8916 unsigned MaskBits = 0;
8918 if (AndNode->getOpcode() != ISD::AND)
8921 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(AndNode->getOperand(1))) {
8922 uint32_t CNV = CN->getZExtValue();
8925 else if (CNV == 65535)
8932 SDValue AddValue = AndNode->getOperand(0);
8934 if (AddValue.getOpcode() != ISD::ADD)
8937 // The basic dag structure is correct, grab the inputs and validate them.
8939 SDValue AddInputValue1 = AddValue.getNode()->getOperand(0);
8940 SDValue AddInputValue2 = AddValue.getNode()->getOperand(1);
8941 SDValue SubsInputValue = SubsNode->getOperand(1);
8943 // The mask is present and the provenance of all the values is a smaller type,
8944 // lets see if the mask is superfluous.
8946 if (!isa<ConstantSDNode>(AddInputValue2.getNode()) ||
8947 !isa<ConstantSDNode>(SubsInputValue.getNode()))
8950 ISD::LoadExtType ExtType;
8952 if (!checkValueWidth(SubsInputValue, MaskBits, ExtType) ||
8953 !checkValueWidth(AddInputValue2, MaskBits, ExtType) ||
8954 !checkValueWidth(AddInputValue1, MaskBits, ExtType) )
8957 if(!isEquivalentMaskless(CC, MaskBits, ExtType,
8958 cast<ConstantSDNode>(AddInputValue2.getNode())->getSExtValue(),
8959 cast<ConstantSDNode>(SubsInputValue.getNode())->getSExtValue()))
8962 // The AND is not necessary, remove it.
8964 SDVTList VTs = DAG.getVTList(SubsNode->getValueType(0),
8965 SubsNode->getValueType(1));
8966 SDValue Ops[] = { AddValue, SubsNode->getOperand(1) };
8968 SDValue NewValue = DAG.getNode(CondOpcode, SDLoc(SubsNode), VTs, Ops);
8969 DAG.ReplaceAllUsesWith(SubsNode, NewValue.getNode());
8971 return SDValue(N, 0);
8974 // Optimize compare with zero and branch.
8975 static SDValue performBRCONDCombine(SDNode *N,
8976 TargetLowering::DAGCombinerInfo &DCI,
8977 SelectionDAG &DAG) {
8978 SDValue NV = performCONDCombine(N, DCI, DAG, 2, 3);
8981 SDValue Chain = N->getOperand(0);
8982 SDValue Dest = N->getOperand(1);
8983 SDValue CCVal = N->getOperand(2);
8984 SDValue Cmp = N->getOperand(3);
8986 assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!");
8987 unsigned CC = cast<ConstantSDNode>(CCVal)->getZExtValue();
8988 if (CC != AArch64CC::EQ && CC != AArch64CC::NE)
8991 unsigned CmpOpc = Cmp.getOpcode();
8992 if (CmpOpc != AArch64ISD::ADDS && CmpOpc != AArch64ISD::SUBS)
8995 // Only attempt folding if there is only one use of the flag and no use of the
8997 if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1))
9000 SDValue LHS = Cmp.getOperand(0);
9001 SDValue RHS = Cmp.getOperand(1);
9003 assert(LHS.getValueType() == RHS.getValueType() &&
9004 "Expected the value type to be the same for both operands!");
9005 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
9008 if (isa<ConstantSDNode>(LHS) && cast<ConstantSDNode>(LHS)->isNullValue())
9009 std::swap(LHS, RHS);
9011 if (!isa<ConstantSDNode>(RHS) || !cast<ConstantSDNode>(RHS)->isNullValue())
9014 if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA ||
9015 LHS.getOpcode() == ISD::SRL)
9018 // Fold the compare into the branch instruction.
9020 if (CC == AArch64CC::EQ)
9021 BR = DAG.getNode(AArch64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
9023 BR = DAG.getNode(AArch64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
9025 // Do not add new nodes to DAG combiner worklist.
9026 DCI.CombineTo(N, BR, false);
9031 // vselect (v1i1 setcc) ->
9032 // vselect (v1iXX setcc) (XX is the size of the compared operand type)
9033 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
9034 // condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
9036 static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) {
9037 SDValue N0 = N->getOperand(0);
9038 EVT CCVT = N0.getValueType();
9040 if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
9041 CCVT.getVectorElementType() != MVT::i1)
9044 EVT ResVT = N->getValueType(0);
9045 EVT CmpVT = N0.getOperand(0).getValueType();
9046 // Only combine when the result type is of the same size as the compared
9048 if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
9051 SDValue IfTrue = N->getOperand(1);
9052 SDValue IfFalse = N->getOperand(2);
9054 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
9055 N0.getOperand(0), N0.getOperand(1),
9056 cast<CondCodeSDNode>(N0.getOperand(2))->get());
9057 return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
9061 /// A vector select: "(select vL, vR, (setcc LHS, RHS))" is best performed with
9062 /// the compare-mask instructions rather than going via NZCV, even if LHS and
9063 /// RHS are really scalar. This replaces any scalar setcc in the above pattern
9064 /// with a vector one followed by a DUP shuffle on the result.
9065 static SDValue performSelectCombine(SDNode *N,
9066 TargetLowering::DAGCombinerInfo &DCI) {
9067 SelectionDAG &DAG = DCI.DAG;
9068 SDValue N0 = N->getOperand(0);
9069 EVT ResVT = N->getValueType(0);
9071 if (N0.getOpcode() != ISD::SETCC)
9074 // Make sure the SETCC result is either i1 (initial DAG), or i32, the lowered
9075 // scalar SetCCResultType. We also don't expect vectors, because we assume
9076 // that selects fed by vector SETCCs are canonicalized to VSELECT.
9077 assert((N0.getValueType() == MVT::i1 || N0.getValueType() == MVT::i32) &&
9078 "Scalar-SETCC feeding SELECT has unexpected result type!");
9080 // If NumMaskElts == 0, the comparison is larger than select result. The
9081 // largest real NEON comparison is 64-bits per lane, which means the result is
9082 // at most 32-bits and an illegal vector. Just bail out for now.
9083 EVT SrcVT = N0.getOperand(0).getValueType();
9085 // Don't try to do this optimization when the setcc itself has i1 operands.
9086 // There are no legal vectors of i1, so this would be pointless.
9087 if (SrcVT == MVT::i1)
9090 int NumMaskElts = ResVT.getSizeInBits() / SrcVT.getSizeInBits();
9091 if (!ResVT.isVector() || NumMaskElts == 0)
9094 SrcVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumMaskElts);
9095 EVT CCVT = SrcVT.changeVectorElementTypeToInteger();
9097 // Also bail out if the vector CCVT isn't the same size as ResVT.
9098 // This can happen if the SETCC operand size doesn't divide the ResVT size
9099 // (e.g., f64 vs v3f32).
9100 if (CCVT.getSizeInBits() != ResVT.getSizeInBits())
9103 // Make sure we didn't create illegal types, if we're not supposed to.
9104 assert(DCI.isBeforeLegalize() ||
9105 DAG.getTargetLoweringInfo().isTypeLegal(SrcVT));
9107 // First perform a vector comparison, where lane 0 is the one we're interested
9111 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(0));
9113 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(1));
9114 SDValue SetCC = DAG.getNode(ISD::SETCC, DL, CCVT, LHS, RHS, N0.getOperand(2));
9116 // Now duplicate the comparison mask we want across all other lanes.
9117 SmallVector<int, 8> DUPMask(CCVT.getVectorNumElements(), 0);
9118 SDValue Mask = DAG.getVectorShuffle(CCVT, DL, SetCC, SetCC, DUPMask.data());
9119 Mask = DAG.getNode(ISD::BITCAST, DL,
9120 ResVT.changeVectorElementTypeToInteger(), Mask);
9122 return DAG.getSelect(DL, ResVT, Mask, N->getOperand(1), N->getOperand(2));
9125 /// performSelectCCCombine - Target-specific DAG combining for ISD::SELECT_CC
9126 /// to match FMIN/FMAX patterns.
9127 static SDValue performSelectCCCombine(SDNode *N, SelectionDAG &DAG) {
9128 // Try to use FMIN/FMAX instructions for FP selects like "x < y ? x : y".
9129 // Unless the NoNaNsFPMath option is set, be careful about NaNs:
9130 // vmax/vmin return NaN if either operand is a NaN;
9131 // only do the transformation when it matches that behavior.
9133 SDValue CondLHS = N->getOperand(0);
9134 SDValue CondRHS = N->getOperand(1);
9135 SDValue LHS = N->getOperand(2);
9136 SDValue RHS = N->getOperand(3);
9137 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9141 if (selectCCOpsAreFMaxCompatible(CondLHS, LHS) &&
9142 selectCCOpsAreFMaxCompatible(CondRHS, RHS)) {
9143 IsReversed = false; // x CC y ? x : y
9144 } else if (selectCCOpsAreFMaxCompatible(CondRHS, LHS) &&
9145 selectCCOpsAreFMaxCompatible(CondLHS, RHS)) {
9146 IsReversed = true ; // x CC y ? y : x
9151 bool IsUnordered = false, IsOrEqual;
9162 IsOrEqual = (CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE);
9163 Opcode = IsReversed ? ISD::FMAXNAN : ISD::FMINNAN;
9173 IsOrEqual = (CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE);
9174 Opcode = IsReversed ? ISD::FMINNAN : ISD::FMAXNAN;
9178 // If LHS is NaN, an ordered comparison will be false and the result will be
9179 // the RHS, but FMIN(NaN, RHS) = FMAX(NaN, RHS) = NaN. Avoid this by checking
9180 // that LHS != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9181 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9184 // For xxx-or-equal comparisons, "+0 <= -0" and "-0 >= +0" will both be true,
9185 // but FMIN will return -0, and FMAX will return +0. So FMIN/FMAX can only be
9186 // used for unsafe math or if one of the operands is known to be nonzero.
9187 if (IsOrEqual && !DAG.getTarget().Options.UnsafeFPMath &&
9188 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9191 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
9194 /// Get rid of unnecessary NVCASTs (that don't change the type).
9195 static SDValue performNVCASTCombine(SDNode *N) {
9196 if (N->getValueType(0) == N->getOperand(0).getValueType())
9197 return N->getOperand(0);
9202 SDValue AArch64TargetLowering::PerformDAGCombine(SDNode *N,
9203 DAGCombinerInfo &DCI) const {
9204 SelectionDAG &DAG = DCI.DAG;
9205 switch (N->getOpcode()) {
9210 return performAddSubLongCombine(N, DCI, DAG);
9212 return performXorCombine(N, DAG, DCI, Subtarget);
9214 return performMulCombine(N, DAG, DCI, Subtarget);
9215 case ISD::SINT_TO_FP:
9216 case ISD::UINT_TO_FP:
9217 return performIntToFpCombine(N, DAG, Subtarget);
9219 return performORCombine(N, DCI, Subtarget);
9220 case ISD::INTRINSIC_WO_CHAIN:
9221 return performIntrinsicCombine(N, DCI, Subtarget);
9222 case ISD::ANY_EXTEND:
9223 case ISD::ZERO_EXTEND:
9224 case ISD::SIGN_EXTEND:
9225 return performExtendCombine(N, DCI, DAG);
9227 return performBitcastCombine(N, DCI, DAG);
9228 case ISD::CONCAT_VECTORS:
9229 return performConcatVectorsCombine(N, DCI, DAG);
9231 return performSelectCombine(N, DCI);
9233 return performVSelectCombine(N, DCI.DAG);
9234 case ISD::SELECT_CC:
9235 return performSelectCCCombine(N, DCI.DAG);
9237 return performSTORECombine(N, DCI, DAG, Subtarget);
9238 case AArch64ISD::BRCOND:
9239 return performBRCONDCombine(N, DCI, DAG);
9240 case AArch64ISD::CSEL:
9241 return performCONDCombine(N, DCI, DAG, 2, 3);
9242 case AArch64ISD::DUP:
9243 return performPostLD1Combine(N, DCI, false);
9244 case AArch64ISD::NVCAST:
9245 return performNVCASTCombine(N);
9246 case ISD::INSERT_VECTOR_ELT:
9247 return performPostLD1Combine(N, DCI, true);
9248 case ISD::INTRINSIC_VOID:
9249 case ISD::INTRINSIC_W_CHAIN:
9250 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9251 case Intrinsic::aarch64_neon_ld2:
9252 case Intrinsic::aarch64_neon_ld3:
9253 case Intrinsic::aarch64_neon_ld4:
9254 case Intrinsic::aarch64_neon_ld1x2:
9255 case Intrinsic::aarch64_neon_ld1x3:
9256 case Intrinsic::aarch64_neon_ld1x4:
9257 case Intrinsic::aarch64_neon_ld2lane:
9258 case Intrinsic::aarch64_neon_ld3lane:
9259 case Intrinsic::aarch64_neon_ld4lane:
9260 case Intrinsic::aarch64_neon_ld2r:
9261 case Intrinsic::aarch64_neon_ld3r:
9262 case Intrinsic::aarch64_neon_ld4r:
9263 case Intrinsic::aarch64_neon_st2:
9264 case Intrinsic::aarch64_neon_st3:
9265 case Intrinsic::aarch64_neon_st4:
9266 case Intrinsic::aarch64_neon_st1x2:
9267 case Intrinsic::aarch64_neon_st1x3:
9268 case Intrinsic::aarch64_neon_st1x4:
9269 case Intrinsic::aarch64_neon_st2lane:
9270 case Intrinsic::aarch64_neon_st3lane:
9271 case Intrinsic::aarch64_neon_st4lane:
9272 return performNEONPostLDSTCombine(N, DCI, DAG);
9280 // Check if the return value is used as only a return value, as otherwise
9281 // we can't perform a tail-call. In particular, we need to check for
9282 // target ISD nodes that are returns and any other "odd" constructs
9283 // that the generic analysis code won't necessarily catch.
9284 bool AArch64TargetLowering::isUsedByReturnOnly(SDNode *N,
9285 SDValue &Chain) const {
9286 if (N->getNumValues() != 1)
9288 if (!N->hasNUsesOfValue(1, 0))
9291 SDValue TCChain = Chain;
9292 SDNode *Copy = *N->use_begin();
9293 if (Copy->getOpcode() == ISD::CopyToReg) {
9294 // If the copy has a glue operand, we conservatively assume it isn't safe to
9295 // perform a tail call.
9296 if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() ==
9299 TCChain = Copy->getOperand(0);
9300 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
9303 bool HasRet = false;
9304 for (SDNode *Node : Copy->uses()) {
9305 if (Node->getOpcode() != AArch64ISD::RET_FLAG)
9317 // Return whether the an instruction can potentially be optimized to a tail
9318 // call. This will cause the optimizers to attempt to move, or duplicate,
9319 // return instructions to help enable tail call optimizations for this
9321 bool AArch64TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
9322 if (!CI->isTailCall())
9328 bool AArch64TargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base,
9330 ISD::MemIndexedMode &AM,
9332 SelectionDAG &DAG) const {
9333 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
9336 Base = Op->getOperand(0);
9337 // All of the indexed addressing mode instructions take a signed
9338 // 9 bit immediate offset.
9339 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
9340 int64_t RHSC = (int64_t)RHS->getZExtValue();
9341 if (RHSC >= 256 || RHSC <= -256)
9343 IsInc = (Op->getOpcode() == ISD::ADD);
9344 Offset = Op->getOperand(1);
9350 bool AArch64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
9352 ISD::MemIndexedMode &AM,
9353 SelectionDAG &DAG) const {
9356 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9357 VT = LD->getMemoryVT();
9358 Ptr = LD->getBasePtr();
9359 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9360 VT = ST->getMemoryVT();
9361 Ptr = ST->getBasePtr();
9366 if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, IsInc, DAG))
9368 AM = IsInc ? ISD::PRE_INC : ISD::PRE_DEC;
9372 bool AArch64TargetLowering::getPostIndexedAddressParts(
9373 SDNode *N, SDNode *Op, SDValue &Base, SDValue &Offset,
9374 ISD::MemIndexedMode &AM, SelectionDAG &DAG) const {
9377 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9378 VT = LD->getMemoryVT();
9379 Ptr = LD->getBasePtr();
9380 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9381 VT = ST->getMemoryVT();
9382 Ptr = ST->getBasePtr();
9387 if (!getIndexedAddressParts(Op, Base, Offset, AM, IsInc, DAG))
9389 // Post-indexing updates the base, so it's not a valid transform
9390 // if that's not the same as the load's pointer.
9393 AM = IsInc ? ISD::POST_INC : ISD::POST_DEC;
9397 static void ReplaceBITCASTResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
9398 SelectionDAG &DAG) {
9400 SDValue Op = N->getOperand(0);
9402 if (N->getValueType(0) != MVT::i16 || Op.getValueType() != MVT::f16)
9406 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, DL, MVT::f32,
9407 DAG.getUNDEF(MVT::i32), Op,
9408 DAG.getTargetConstant(AArch64::hsub, DL, MVT::i32)),
9410 Op = DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op);
9411 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, Op));
9414 void AArch64TargetLowering::ReplaceNodeResults(
9415 SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
9416 switch (N->getOpcode()) {
9418 llvm_unreachable("Don't know how to custom expand this");
9420 ReplaceBITCASTResults(N, Results, DAG);
9422 case ISD::FP_TO_UINT:
9423 case ISD::FP_TO_SINT:
9424 assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion");
9425 // Let normal code take care of it by not adding anything to Results.
9430 bool AArch64TargetLowering::useLoadStackGuardNode() const {
9434 unsigned AArch64TargetLowering::combineRepeatedFPDivisors() const {
9435 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
9436 // reciprocal if there are three or more FDIVs.
9440 TargetLoweringBase::LegalizeTypeAction
9441 AArch64TargetLowering::getPreferredVectorAction(EVT VT) const {
9442 MVT SVT = VT.getSimpleVT();
9443 // During type legalization, we prefer to widen v1i8, v1i16, v1i32 to v8i8,
9444 // v4i16, v2i32 instead of to promote.
9445 if (SVT == MVT::v1i8 || SVT == MVT::v1i16 || SVT == MVT::v1i32
9446 || SVT == MVT::v1f32)
9447 return TypeWidenVector;
9449 return TargetLoweringBase::getPreferredVectorAction(VT);
9452 // Loads and stores less than 128-bits are already atomic; ones above that
9453 // are doomed anyway, so defer to the default libcall and blame the OS when
9455 bool AArch64TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
9456 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
9460 // Loads and stores less than 128-bits are already atomic; ones above that
9461 // are doomed anyway, so defer to the default libcall and blame the OS when
9463 bool AArch64TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
9464 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
9468 // For the real atomic operations, we have ldxr/stxr up to 128 bits,
9469 TargetLoweringBase::AtomicRMWExpansionKind
9470 AArch64TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
9471 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
9472 return Size <= 128 ? AtomicRMWExpansionKind::LLSC
9473 : AtomicRMWExpansionKind::None;
9476 bool AArch64TargetLowering::hasLoadLinkedStoreConditional() const {
9480 Value *AArch64TargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
9481 AtomicOrdering Ord) const {
9482 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
9483 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
9484 bool IsAcquire = isAtLeastAcquire(Ord);
9486 // Since i128 isn't legal and intrinsics don't get type-lowered, the ldrexd
9487 // intrinsic must return {i64, i64} and we have to recombine them into a
9488 // single i128 here.
9489 if (ValTy->getPrimitiveSizeInBits() == 128) {
9491 IsAcquire ? Intrinsic::aarch64_ldaxp : Intrinsic::aarch64_ldxp;
9492 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int);
9494 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
9495 Value *LoHi = Builder.CreateCall(Ldxr, Addr, "lohi");
9497 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
9498 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
9499 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
9500 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
9501 return Builder.CreateOr(
9502 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 64)), "val64");
9505 Type *Tys[] = { Addr->getType() };
9507 IsAcquire ? Intrinsic::aarch64_ldaxr : Intrinsic::aarch64_ldxr;
9508 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int, Tys);
9510 return Builder.CreateTruncOrBitCast(
9511 Builder.CreateCall(Ldxr, Addr),
9512 cast<PointerType>(Addr->getType())->getElementType());
9515 Value *AArch64TargetLowering::emitStoreConditional(IRBuilder<> &Builder,
9516 Value *Val, Value *Addr,
9517 AtomicOrdering Ord) const {
9518 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
9519 bool IsRelease = isAtLeastRelease(Ord);
9521 // Since the intrinsics must have legal type, the i128 intrinsics take two
9522 // parameters: "i64, i64". We must marshal Val into the appropriate form
9524 if (Val->getType()->getPrimitiveSizeInBits() == 128) {
9526 IsRelease ? Intrinsic::aarch64_stlxp : Intrinsic::aarch64_stxp;
9527 Function *Stxr = Intrinsic::getDeclaration(M, Int);
9528 Type *Int64Ty = Type::getInt64Ty(M->getContext());
9530 Value *Lo = Builder.CreateTrunc(Val, Int64Ty, "lo");
9531 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 64), Int64Ty, "hi");
9532 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
9533 return Builder.CreateCall(Stxr, {Lo, Hi, Addr});
9537 IsRelease ? Intrinsic::aarch64_stlxr : Intrinsic::aarch64_stxr;
9538 Type *Tys[] = { Addr->getType() };
9539 Function *Stxr = Intrinsic::getDeclaration(M, Int, Tys);
9541 return Builder.CreateCall(Stxr,
9542 {Builder.CreateZExtOrBitCast(
9543 Val, Stxr->getFunctionType()->getParamType(0)),
9547 bool AArch64TargetLowering::functionArgumentNeedsConsecutiveRegisters(
9548 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
9549 return Ty->isArrayTy();
9552 bool AArch64TargetLowering::shouldNormalizeToSelectSequence(LLVMContext &,