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);
304 setOperationAction(ISD::FMINNAN, MVT::f16, Promote);
305 setOperationAction(ISD::FMAXNAN, MVT::f16, Promote);
307 // v4f16 is also a storage-only type, so promote it to v4f32 when that is
309 setOperationAction(ISD::FADD, MVT::v4f16, Promote);
310 setOperationAction(ISD::FSUB, MVT::v4f16, Promote);
311 setOperationAction(ISD::FMUL, MVT::v4f16, Promote);
312 setOperationAction(ISD::FDIV, MVT::v4f16, Promote);
313 setOperationAction(ISD::FP_EXTEND, MVT::v4f16, Promote);
314 setOperationAction(ISD::FP_ROUND, MVT::v4f16, Promote);
315 AddPromotedToType(ISD::FADD, MVT::v4f16, MVT::v4f32);
316 AddPromotedToType(ISD::FSUB, MVT::v4f16, MVT::v4f32);
317 AddPromotedToType(ISD::FMUL, MVT::v4f16, MVT::v4f32);
318 AddPromotedToType(ISD::FDIV, MVT::v4f16, MVT::v4f32);
319 AddPromotedToType(ISD::FP_EXTEND, MVT::v4f16, MVT::v4f32);
320 AddPromotedToType(ISD::FP_ROUND, MVT::v4f16, MVT::v4f32);
322 // Expand all other v4f16 operations.
323 // FIXME: We could generate better code by promoting some operations to
325 setOperationAction(ISD::FABS, MVT::v4f16, Expand);
326 setOperationAction(ISD::FCEIL, MVT::v4f16, Expand);
327 setOperationAction(ISD::FCOPYSIGN, MVT::v4f16, Expand);
328 setOperationAction(ISD::FCOS, MVT::v4f16, Expand);
329 setOperationAction(ISD::FFLOOR, MVT::v4f16, Expand);
330 setOperationAction(ISD::FMA, MVT::v4f16, Expand);
331 setOperationAction(ISD::FNEARBYINT, MVT::v4f16, Expand);
332 setOperationAction(ISD::FNEG, MVT::v4f16, Expand);
333 setOperationAction(ISD::FPOW, MVT::v4f16, Expand);
334 setOperationAction(ISD::FPOWI, MVT::v4f16, Expand);
335 setOperationAction(ISD::FREM, MVT::v4f16, Expand);
336 setOperationAction(ISD::FROUND, MVT::v4f16, Expand);
337 setOperationAction(ISD::FRINT, MVT::v4f16, Expand);
338 setOperationAction(ISD::FSIN, MVT::v4f16, Expand);
339 setOperationAction(ISD::FSINCOS, MVT::v4f16, Expand);
340 setOperationAction(ISD::FSQRT, MVT::v4f16, Expand);
341 setOperationAction(ISD::FTRUNC, MVT::v4f16, Expand);
342 setOperationAction(ISD::SETCC, MVT::v4f16, Expand);
343 setOperationAction(ISD::BR_CC, MVT::v4f16, Expand);
344 setOperationAction(ISD::SELECT, MVT::v4f16, Expand);
345 setOperationAction(ISD::SELECT_CC, MVT::v4f16, Expand);
346 setOperationAction(ISD::FEXP, MVT::v4f16, Expand);
347 setOperationAction(ISD::FEXP2, MVT::v4f16, Expand);
348 setOperationAction(ISD::FLOG, MVT::v4f16, Expand);
349 setOperationAction(ISD::FLOG2, MVT::v4f16, Expand);
350 setOperationAction(ISD::FLOG10, MVT::v4f16, Expand);
353 // v8f16 is also a storage-only type, so expand it.
354 setOperationAction(ISD::FABS, MVT::v8f16, Expand);
355 setOperationAction(ISD::FADD, MVT::v8f16, Expand);
356 setOperationAction(ISD::FCEIL, MVT::v8f16, Expand);
357 setOperationAction(ISD::FCOPYSIGN, MVT::v8f16, Expand);
358 setOperationAction(ISD::FCOS, MVT::v8f16, Expand);
359 setOperationAction(ISD::FDIV, MVT::v8f16, Expand);
360 setOperationAction(ISD::FFLOOR, MVT::v8f16, Expand);
361 setOperationAction(ISD::FMA, MVT::v8f16, Expand);
362 setOperationAction(ISD::FMUL, MVT::v8f16, Expand);
363 setOperationAction(ISD::FNEARBYINT, MVT::v8f16, Expand);
364 setOperationAction(ISD::FNEG, MVT::v8f16, Expand);
365 setOperationAction(ISD::FPOW, MVT::v8f16, Expand);
366 setOperationAction(ISD::FPOWI, MVT::v8f16, Expand);
367 setOperationAction(ISD::FREM, MVT::v8f16, Expand);
368 setOperationAction(ISD::FROUND, MVT::v8f16, Expand);
369 setOperationAction(ISD::FRINT, MVT::v8f16, Expand);
370 setOperationAction(ISD::FSIN, MVT::v8f16, Expand);
371 setOperationAction(ISD::FSINCOS, MVT::v8f16, Expand);
372 setOperationAction(ISD::FSQRT, MVT::v8f16, Expand);
373 setOperationAction(ISD::FSUB, MVT::v8f16, Expand);
374 setOperationAction(ISD::FTRUNC, MVT::v8f16, Expand);
375 setOperationAction(ISD::SETCC, MVT::v8f16, Expand);
376 setOperationAction(ISD::BR_CC, MVT::v8f16, Expand);
377 setOperationAction(ISD::SELECT, MVT::v8f16, Expand);
378 setOperationAction(ISD::SELECT_CC, MVT::v8f16, Expand);
379 setOperationAction(ISD::FP_EXTEND, MVT::v8f16, Expand);
380 setOperationAction(ISD::FEXP, MVT::v8f16, Expand);
381 setOperationAction(ISD::FEXP2, MVT::v8f16, Expand);
382 setOperationAction(ISD::FLOG, MVT::v8f16, Expand);
383 setOperationAction(ISD::FLOG2, MVT::v8f16, Expand);
384 setOperationAction(ISD::FLOG10, MVT::v8f16, Expand);
386 // AArch64 has implementations of a lot of rounding-like FP operations.
387 for (MVT Ty : {MVT::f32, MVT::f64}) {
388 setOperationAction(ISD::FFLOOR, Ty, Legal);
389 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
390 setOperationAction(ISD::FCEIL, Ty, Legal);
391 setOperationAction(ISD::FRINT, Ty, Legal);
392 setOperationAction(ISD::FTRUNC, Ty, Legal);
393 setOperationAction(ISD::FROUND, Ty, Legal);
394 setOperationAction(ISD::FMINNUM, Ty, Legal);
395 setOperationAction(ISD::FMAXNUM, Ty, Legal);
396 setOperationAction(ISD::FMINNAN, Ty, Legal);
397 setOperationAction(ISD::FMAXNAN, Ty, Legal);
400 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
402 // Lower READCYCLECOUNTER using an mrs from PMCCNTR_EL0.
403 // This requires the Performance Monitors extension.
404 if (Subtarget->hasPerfMon())
405 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Legal);
407 if (Subtarget->isTargetMachO()) {
408 // For iOS, we don't want to the normal expansion of a libcall to
409 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
411 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
412 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
414 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
415 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
418 // Make floating-point constants legal for the large code model, so they don't
419 // become loads from the constant pool.
420 if (Subtarget->isTargetMachO() && TM.getCodeModel() == CodeModel::Large) {
421 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
422 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
425 // AArch64 does not have floating-point extending loads, i1 sign-extending
426 // load, floating-point truncating stores, or v2i32->v2i16 truncating store.
427 for (MVT VT : MVT::fp_valuetypes()) {
428 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
429 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
430 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f64, Expand);
431 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f80, Expand);
433 for (MVT VT : MVT::integer_valuetypes())
434 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Expand);
436 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
437 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
438 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
439 setTruncStoreAction(MVT::f128, MVT::f80, Expand);
440 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
441 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
442 setTruncStoreAction(MVT::f128, MVT::f16, Expand);
444 setOperationAction(ISD::BITCAST, MVT::i16, Custom);
445 setOperationAction(ISD::BITCAST, MVT::f16, Custom);
447 // Indexed loads and stores are supported.
448 for (unsigned im = (unsigned)ISD::PRE_INC;
449 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
450 setIndexedLoadAction(im, MVT::i8, Legal);
451 setIndexedLoadAction(im, MVT::i16, Legal);
452 setIndexedLoadAction(im, MVT::i32, Legal);
453 setIndexedLoadAction(im, MVT::i64, Legal);
454 setIndexedLoadAction(im, MVT::f64, Legal);
455 setIndexedLoadAction(im, MVT::f32, Legal);
456 setIndexedLoadAction(im, MVT::f16, Legal);
457 setIndexedStoreAction(im, MVT::i8, Legal);
458 setIndexedStoreAction(im, MVT::i16, Legal);
459 setIndexedStoreAction(im, MVT::i32, Legal);
460 setIndexedStoreAction(im, MVT::i64, Legal);
461 setIndexedStoreAction(im, MVT::f64, Legal);
462 setIndexedStoreAction(im, MVT::f32, Legal);
463 setIndexedStoreAction(im, MVT::f16, Legal);
467 setOperationAction(ISD::TRAP, MVT::Other, Legal);
469 // We combine OR nodes for bitfield operations.
470 setTargetDAGCombine(ISD::OR);
472 // Vector add and sub nodes may conceal a high-half opportunity.
473 // Also, try to fold ADD into CSINC/CSINV..
474 setTargetDAGCombine(ISD::ADD);
475 setTargetDAGCombine(ISD::SUB);
477 setTargetDAGCombine(ISD::XOR);
478 setTargetDAGCombine(ISD::SINT_TO_FP);
479 setTargetDAGCombine(ISD::UINT_TO_FP);
481 setTargetDAGCombine(ISD::FP_TO_SINT);
482 setTargetDAGCombine(ISD::FP_TO_UINT);
483 setTargetDAGCombine(ISD::FDIV);
485 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
487 setTargetDAGCombine(ISD::ANY_EXTEND);
488 setTargetDAGCombine(ISD::ZERO_EXTEND);
489 setTargetDAGCombine(ISD::SIGN_EXTEND);
490 setTargetDAGCombine(ISD::BITCAST);
491 setTargetDAGCombine(ISD::CONCAT_VECTORS);
492 setTargetDAGCombine(ISD::STORE);
494 setTargetDAGCombine(ISD::MUL);
496 setTargetDAGCombine(ISD::SELECT);
497 setTargetDAGCombine(ISD::VSELECT);
499 setTargetDAGCombine(ISD::INTRINSIC_VOID);
500 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
501 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
502 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
504 MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8;
505 MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4;
506 MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 4;
508 setStackPointerRegisterToSaveRestore(AArch64::SP);
510 setSchedulingPreference(Sched::Hybrid);
513 MaskAndBranchFoldingIsLegal = true;
514 EnableExtLdPromotion = true;
516 setMinFunctionAlignment(2);
518 setHasExtractBitsInsn(true);
520 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
522 if (Subtarget->hasNEON()) {
523 // FIXME: v1f64 shouldn't be legal if we can avoid it, because it leads to
524 // silliness like this:
525 setOperationAction(ISD::FABS, MVT::v1f64, Expand);
526 setOperationAction(ISD::FADD, MVT::v1f64, Expand);
527 setOperationAction(ISD::FCEIL, MVT::v1f64, Expand);
528 setOperationAction(ISD::FCOPYSIGN, MVT::v1f64, Expand);
529 setOperationAction(ISD::FCOS, MVT::v1f64, Expand);
530 setOperationAction(ISD::FDIV, MVT::v1f64, Expand);
531 setOperationAction(ISD::FFLOOR, MVT::v1f64, Expand);
532 setOperationAction(ISD::FMA, MVT::v1f64, Expand);
533 setOperationAction(ISD::FMUL, MVT::v1f64, Expand);
534 setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Expand);
535 setOperationAction(ISD::FNEG, MVT::v1f64, Expand);
536 setOperationAction(ISD::FPOW, MVT::v1f64, Expand);
537 setOperationAction(ISD::FREM, MVT::v1f64, Expand);
538 setOperationAction(ISD::FROUND, MVT::v1f64, Expand);
539 setOperationAction(ISD::FRINT, MVT::v1f64, Expand);
540 setOperationAction(ISD::FSIN, MVT::v1f64, Expand);
541 setOperationAction(ISD::FSINCOS, MVT::v1f64, Expand);
542 setOperationAction(ISD::FSQRT, MVT::v1f64, Expand);
543 setOperationAction(ISD::FSUB, MVT::v1f64, Expand);
544 setOperationAction(ISD::FTRUNC, MVT::v1f64, Expand);
545 setOperationAction(ISD::SETCC, MVT::v1f64, Expand);
546 setOperationAction(ISD::BR_CC, MVT::v1f64, Expand);
547 setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
548 setOperationAction(ISD::SELECT_CC, MVT::v1f64, Expand);
549 setOperationAction(ISD::FP_EXTEND, MVT::v1f64, Expand);
551 setOperationAction(ISD::FP_TO_SINT, MVT::v1i64, Expand);
552 setOperationAction(ISD::FP_TO_UINT, MVT::v1i64, Expand);
553 setOperationAction(ISD::SINT_TO_FP, MVT::v1i64, Expand);
554 setOperationAction(ISD::UINT_TO_FP, MVT::v1i64, Expand);
555 setOperationAction(ISD::FP_ROUND, MVT::v1f64, Expand);
557 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
559 // AArch64 doesn't have a direct vector ->f32 conversion instructions for
560 // elements smaller than i32, so promote the input to i32 first.
561 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Promote);
562 setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Promote);
563 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Promote);
564 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Promote);
565 // i8 and i16 vector elements also need promotion to i32 for v8i8 or v8i16
566 // -> v8f16 conversions.
567 setOperationAction(ISD::SINT_TO_FP, MVT::v8i8, Promote);
568 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Promote);
569 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
570 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Promote);
571 // Similarly, there is no direct i32 -> f64 vector conversion instruction.
572 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
573 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
574 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
575 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
576 // Or, direct i32 -> f16 vector conversion. Set it so custom, so the
577 // conversion happens in two steps: v4i32 -> v4f32 -> v4f16
578 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Custom);
579 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
581 // AArch64 doesn't have MUL.2d:
582 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
583 // Custom handling for some quad-vector types to detect MULL.
584 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
585 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
586 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
588 setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal);
589 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
590 // Likewise, narrowing and extending vector loads/stores aren't handled
592 for (MVT VT : MVT::vector_valuetypes()) {
593 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
595 setOperationAction(ISD::MULHS, VT, Expand);
596 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
597 setOperationAction(ISD::MULHU, VT, Expand);
598 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
600 setOperationAction(ISD::BSWAP, VT, Expand);
602 for (MVT InnerVT : MVT::vector_valuetypes()) {
603 setTruncStoreAction(VT, InnerVT, Expand);
604 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
605 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
606 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
610 // AArch64 has implementations of a lot of rounding-like FP operations.
611 for (MVT Ty : {MVT::v2f32, MVT::v4f32, MVT::v2f64}) {
612 setOperationAction(ISD::FFLOOR, Ty, Legal);
613 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
614 setOperationAction(ISD::FCEIL, Ty, Legal);
615 setOperationAction(ISD::FRINT, Ty, Legal);
616 setOperationAction(ISD::FTRUNC, Ty, Legal);
617 setOperationAction(ISD::FROUND, Ty, Legal);
621 // Prefer likely predicted branches to selects on out-of-order cores.
622 if (Subtarget->isCortexA57())
623 PredictableSelectIsExpensive = true;
626 void AArch64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) {
627 if (VT == MVT::v2f32 || VT == MVT::v4f16) {
628 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
629 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i32);
631 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
632 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i32);
633 } else if (VT == MVT::v2f64 || VT == MVT::v4f32 || VT == MVT::v8f16) {
634 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
635 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i64);
637 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
638 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i64);
641 // Mark vector float intrinsics as expand.
642 if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) {
643 setOperationAction(ISD::FSIN, VT.getSimpleVT(), Expand);
644 setOperationAction(ISD::FCOS, VT.getSimpleVT(), Expand);
645 setOperationAction(ISD::FPOWI, VT.getSimpleVT(), Expand);
646 setOperationAction(ISD::FPOW, VT.getSimpleVT(), Expand);
647 setOperationAction(ISD::FLOG, VT.getSimpleVT(), Expand);
648 setOperationAction(ISD::FLOG2, VT.getSimpleVT(), Expand);
649 setOperationAction(ISD::FLOG10, VT.getSimpleVT(), Expand);
650 setOperationAction(ISD::FEXP, VT.getSimpleVT(), Expand);
651 setOperationAction(ISD::FEXP2, VT.getSimpleVT(), Expand);
653 // But we do support custom-lowering for FCOPYSIGN.
654 setOperationAction(ISD::FCOPYSIGN, VT.getSimpleVT(), Custom);
657 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
658 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
659 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
660 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
661 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Custom);
662 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
663 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
664 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
665 setOperationAction(ISD::AND, VT.getSimpleVT(), Custom);
666 setOperationAction(ISD::OR, VT.getSimpleVT(), Custom);
667 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
668 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
670 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
671 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
672 setOperationAction(ISD::VSELECT, VT.getSimpleVT(), Expand);
673 for (MVT InnerVT : MVT::all_valuetypes())
674 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT.getSimpleVT(), Expand);
676 // CNT supports only B element sizes.
677 if (VT != MVT::v8i8 && VT != MVT::v16i8)
678 setOperationAction(ISD::CTPOP, VT.getSimpleVT(), Expand);
680 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
681 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
682 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
683 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
684 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
686 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
687 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
689 // [SU][MIN|MAX] and [SU]ABSDIFF are available for all NEON types apart from
691 if (!VT.isFloatingPoint() &&
692 VT.getSimpleVT() != MVT::v2i64 && VT.getSimpleVT() != MVT::v1i64)
693 for (unsigned Opcode : {ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX,
694 ISD::SABSDIFF, ISD::UABSDIFF})
695 setOperationAction(Opcode, VT.getSimpleVT(), Legal);
697 // F[MIN|MAX][NUM|NAN] are available for all FP NEON types (not f16 though!).
698 if (VT.isFloatingPoint() && VT.getVectorElementType() != MVT::f16)
699 for (unsigned Opcode : {ISD::FMINNAN, ISD::FMAXNAN,
700 ISD::FMINNUM, ISD::FMAXNUM})
701 setOperationAction(Opcode, VT.getSimpleVT(), Legal);
703 if (Subtarget->isLittleEndian()) {
704 for (unsigned im = (unsigned)ISD::PRE_INC;
705 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
706 setIndexedLoadAction(im, VT.getSimpleVT(), Legal);
707 setIndexedStoreAction(im, VT.getSimpleVT(), Legal);
712 void AArch64TargetLowering::addDRTypeForNEON(MVT VT) {
713 addRegisterClass(VT, &AArch64::FPR64RegClass);
714 addTypeForNEON(VT, MVT::v2i32);
717 void AArch64TargetLowering::addQRTypeForNEON(MVT VT) {
718 addRegisterClass(VT, &AArch64::FPR128RegClass);
719 addTypeForNEON(VT, MVT::v4i32);
722 EVT AArch64TargetLowering::getSetCCResultType(const DataLayout &, LLVMContext &,
726 return VT.changeVectorElementTypeToInteger();
729 /// computeKnownBitsForTargetNode - Determine which of the bits specified in
730 /// Mask are known to be either zero or one and return them in the
731 /// KnownZero/KnownOne bitsets.
732 void AArch64TargetLowering::computeKnownBitsForTargetNode(
733 const SDValue Op, APInt &KnownZero, APInt &KnownOne,
734 const SelectionDAG &DAG, unsigned Depth) const {
735 switch (Op.getOpcode()) {
738 case AArch64ISD::CSEL: {
739 APInt KnownZero2, KnownOne2;
740 DAG.computeKnownBits(Op->getOperand(0), KnownZero, KnownOne, Depth + 1);
741 DAG.computeKnownBits(Op->getOperand(1), KnownZero2, KnownOne2, Depth + 1);
742 KnownZero &= KnownZero2;
743 KnownOne &= KnownOne2;
746 case ISD::INTRINSIC_W_CHAIN: {
747 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
748 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
751 case Intrinsic::aarch64_ldaxr:
752 case Intrinsic::aarch64_ldxr: {
753 unsigned BitWidth = KnownOne.getBitWidth();
754 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
755 unsigned MemBits = VT.getScalarType().getSizeInBits();
756 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
762 case ISD::INTRINSIC_WO_CHAIN:
763 case ISD::INTRINSIC_VOID: {
764 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
768 case Intrinsic::aarch64_neon_umaxv:
769 case Intrinsic::aarch64_neon_uminv: {
770 // Figure out the datatype of the vector operand. The UMINV instruction
771 // will zero extend the result, so we can mark as known zero all the
772 // bits larger than the element datatype. 32-bit or larget doesn't need
773 // this as those are legal types and will be handled by isel directly.
774 MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
775 unsigned BitWidth = KnownZero.getBitWidth();
776 if (VT == MVT::v8i8 || VT == MVT::v16i8) {
777 assert(BitWidth >= 8 && "Unexpected width!");
778 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8);
780 } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
781 assert(BitWidth >= 16 && "Unexpected width!");
782 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
792 MVT AArch64TargetLowering::getScalarShiftAmountTy(const DataLayout &DL,
797 bool AArch64TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
801 if (Subtarget->requiresStrictAlign())
804 // FIXME: This is mostly true for Cyclone, but not necessarily others.
806 // FIXME: Define an attribute for slow unaligned accesses instead of
807 // relying on the CPU type as a proxy.
808 // On Cyclone, unaligned 128-bit stores are slow.
809 *Fast = !Subtarget->isCyclone() || VT.getStoreSize() != 16 ||
810 // See comments in performSTORECombine() for more details about
813 // Code that uses clang vector extensions can mark that it
814 // wants unaligned accesses to be treated as fast by
815 // underspecifying alignment to be 1 or 2.
818 // Disregard v2i64. Memcpy lowering produces those and splitting
819 // them regresses performance on micro-benchmarks and olden/bh.
826 AArch64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
827 const TargetLibraryInfo *libInfo) const {
828 return AArch64::createFastISel(funcInfo, libInfo);
831 const char *AArch64TargetLowering::getTargetNodeName(unsigned Opcode) const {
832 switch ((AArch64ISD::NodeType)Opcode) {
833 case AArch64ISD::FIRST_NUMBER: break;
834 case AArch64ISD::CALL: return "AArch64ISD::CALL";
835 case AArch64ISD::ADRP: return "AArch64ISD::ADRP";
836 case AArch64ISD::ADDlow: return "AArch64ISD::ADDlow";
837 case AArch64ISD::LOADgot: return "AArch64ISD::LOADgot";
838 case AArch64ISD::RET_FLAG: return "AArch64ISD::RET_FLAG";
839 case AArch64ISD::BRCOND: return "AArch64ISD::BRCOND";
840 case AArch64ISD::CSEL: return "AArch64ISD::CSEL";
841 case AArch64ISD::FCSEL: return "AArch64ISD::FCSEL";
842 case AArch64ISD::CSINV: return "AArch64ISD::CSINV";
843 case AArch64ISD::CSNEG: return "AArch64ISD::CSNEG";
844 case AArch64ISD::CSINC: return "AArch64ISD::CSINC";
845 case AArch64ISD::THREAD_POINTER: return "AArch64ISD::THREAD_POINTER";
846 case AArch64ISD::TLSDESC_CALLSEQ: return "AArch64ISD::TLSDESC_CALLSEQ";
847 case AArch64ISD::ADC: return "AArch64ISD::ADC";
848 case AArch64ISD::SBC: return "AArch64ISD::SBC";
849 case AArch64ISD::ADDS: return "AArch64ISD::ADDS";
850 case AArch64ISD::SUBS: return "AArch64ISD::SUBS";
851 case AArch64ISD::ADCS: return "AArch64ISD::ADCS";
852 case AArch64ISD::SBCS: return "AArch64ISD::SBCS";
853 case AArch64ISD::ANDS: return "AArch64ISD::ANDS";
854 case AArch64ISD::CCMP: return "AArch64ISD::CCMP";
855 case AArch64ISD::CCMN: return "AArch64ISD::CCMN";
856 case AArch64ISD::FCCMP: return "AArch64ISD::FCCMP";
857 case AArch64ISD::FCMP: return "AArch64ISD::FCMP";
858 case AArch64ISD::DUP: return "AArch64ISD::DUP";
859 case AArch64ISD::DUPLANE8: return "AArch64ISD::DUPLANE8";
860 case AArch64ISD::DUPLANE16: return "AArch64ISD::DUPLANE16";
861 case AArch64ISD::DUPLANE32: return "AArch64ISD::DUPLANE32";
862 case AArch64ISD::DUPLANE64: return "AArch64ISD::DUPLANE64";
863 case AArch64ISD::MOVI: return "AArch64ISD::MOVI";
864 case AArch64ISD::MOVIshift: return "AArch64ISD::MOVIshift";
865 case AArch64ISD::MOVIedit: return "AArch64ISD::MOVIedit";
866 case AArch64ISD::MOVImsl: return "AArch64ISD::MOVImsl";
867 case AArch64ISD::FMOV: return "AArch64ISD::FMOV";
868 case AArch64ISD::MVNIshift: return "AArch64ISD::MVNIshift";
869 case AArch64ISD::MVNImsl: return "AArch64ISD::MVNImsl";
870 case AArch64ISD::BICi: return "AArch64ISD::BICi";
871 case AArch64ISD::ORRi: return "AArch64ISD::ORRi";
872 case AArch64ISD::BSL: return "AArch64ISD::BSL";
873 case AArch64ISD::NEG: return "AArch64ISD::NEG";
874 case AArch64ISD::EXTR: return "AArch64ISD::EXTR";
875 case AArch64ISD::ZIP1: return "AArch64ISD::ZIP1";
876 case AArch64ISD::ZIP2: return "AArch64ISD::ZIP2";
877 case AArch64ISD::UZP1: return "AArch64ISD::UZP1";
878 case AArch64ISD::UZP2: return "AArch64ISD::UZP2";
879 case AArch64ISD::TRN1: return "AArch64ISD::TRN1";
880 case AArch64ISD::TRN2: return "AArch64ISD::TRN2";
881 case AArch64ISD::REV16: return "AArch64ISD::REV16";
882 case AArch64ISD::REV32: return "AArch64ISD::REV32";
883 case AArch64ISD::REV64: return "AArch64ISD::REV64";
884 case AArch64ISD::EXT: return "AArch64ISD::EXT";
885 case AArch64ISD::VSHL: return "AArch64ISD::VSHL";
886 case AArch64ISD::VLSHR: return "AArch64ISD::VLSHR";
887 case AArch64ISD::VASHR: return "AArch64ISD::VASHR";
888 case AArch64ISD::CMEQ: return "AArch64ISD::CMEQ";
889 case AArch64ISD::CMGE: return "AArch64ISD::CMGE";
890 case AArch64ISD::CMGT: return "AArch64ISD::CMGT";
891 case AArch64ISD::CMHI: return "AArch64ISD::CMHI";
892 case AArch64ISD::CMHS: return "AArch64ISD::CMHS";
893 case AArch64ISD::FCMEQ: return "AArch64ISD::FCMEQ";
894 case AArch64ISD::FCMGE: return "AArch64ISD::FCMGE";
895 case AArch64ISD::FCMGT: return "AArch64ISD::FCMGT";
896 case AArch64ISD::CMEQz: return "AArch64ISD::CMEQz";
897 case AArch64ISD::CMGEz: return "AArch64ISD::CMGEz";
898 case AArch64ISD::CMGTz: return "AArch64ISD::CMGTz";
899 case AArch64ISD::CMLEz: return "AArch64ISD::CMLEz";
900 case AArch64ISD::CMLTz: return "AArch64ISD::CMLTz";
901 case AArch64ISD::FCMEQz: return "AArch64ISD::FCMEQz";
902 case AArch64ISD::FCMGEz: return "AArch64ISD::FCMGEz";
903 case AArch64ISD::FCMGTz: return "AArch64ISD::FCMGTz";
904 case AArch64ISD::FCMLEz: return "AArch64ISD::FCMLEz";
905 case AArch64ISD::FCMLTz: return "AArch64ISD::FCMLTz";
906 case AArch64ISD::SADDV: return "AArch64ISD::SADDV";
907 case AArch64ISD::UADDV: return "AArch64ISD::UADDV";
908 case AArch64ISD::SMINV: return "AArch64ISD::SMINV";
909 case AArch64ISD::UMINV: return "AArch64ISD::UMINV";
910 case AArch64ISD::SMAXV: return "AArch64ISD::SMAXV";
911 case AArch64ISD::UMAXV: return "AArch64ISD::UMAXV";
912 case AArch64ISD::NOT: return "AArch64ISD::NOT";
913 case AArch64ISD::BIT: return "AArch64ISD::BIT";
914 case AArch64ISD::CBZ: return "AArch64ISD::CBZ";
915 case AArch64ISD::CBNZ: return "AArch64ISD::CBNZ";
916 case AArch64ISD::TBZ: return "AArch64ISD::TBZ";
917 case AArch64ISD::TBNZ: return "AArch64ISD::TBNZ";
918 case AArch64ISD::TC_RETURN: return "AArch64ISD::TC_RETURN";
919 case AArch64ISD::PREFETCH: return "AArch64ISD::PREFETCH";
920 case AArch64ISD::SITOF: return "AArch64ISD::SITOF";
921 case AArch64ISD::UITOF: return "AArch64ISD::UITOF";
922 case AArch64ISD::NVCAST: return "AArch64ISD::NVCAST";
923 case AArch64ISD::SQSHL_I: return "AArch64ISD::SQSHL_I";
924 case AArch64ISD::UQSHL_I: return "AArch64ISD::UQSHL_I";
925 case AArch64ISD::SRSHR_I: return "AArch64ISD::SRSHR_I";
926 case AArch64ISD::URSHR_I: return "AArch64ISD::URSHR_I";
927 case AArch64ISD::SQSHLU_I: return "AArch64ISD::SQSHLU_I";
928 case AArch64ISD::WrapperLarge: return "AArch64ISD::WrapperLarge";
929 case AArch64ISD::LD2post: return "AArch64ISD::LD2post";
930 case AArch64ISD::LD3post: return "AArch64ISD::LD3post";
931 case AArch64ISD::LD4post: return "AArch64ISD::LD4post";
932 case AArch64ISD::ST2post: return "AArch64ISD::ST2post";
933 case AArch64ISD::ST3post: return "AArch64ISD::ST3post";
934 case AArch64ISD::ST4post: return "AArch64ISD::ST4post";
935 case AArch64ISD::LD1x2post: return "AArch64ISD::LD1x2post";
936 case AArch64ISD::LD1x3post: return "AArch64ISD::LD1x3post";
937 case AArch64ISD::LD1x4post: return "AArch64ISD::LD1x4post";
938 case AArch64ISD::ST1x2post: return "AArch64ISD::ST1x2post";
939 case AArch64ISD::ST1x3post: return "AArch64ISD::ST1x3post";
940 case AArch64ISD::ST1x4post: return "AArch64ISD::ST1x4post";
941 case AArch64ISD::LD1DUPpost: return "AArch64ISD::LD1DUPpost";
942 case AArch64ISD::LD2DUPpost: return "AArch64ISD::LD2DUPpost";
943 case AArch64ISD::LD3DUPpost: return "AArch64ISD::LD3DUPpost";
944 case AArch64ISD::LD4DUPpost: return "AArch64ISD::LD4DUPpost";
945 case AArch64ISD::LD1LANEpost: return "AArch64ISD::LD1LANEpost";
946 case AArch64ISD::LD2LANEpost: return "AArch64ISD::LD2LANEpost";
947 case AArch64ISD::LD3LANEpost: return "AArch64ISD::LD3LANEpost";
948 case AArch64ISD::LD4LANEpost: return "AArch64ISD::LD4LANEpost";
949 case AArch64ISD::ST2LANEpost: return "AArch64ISD::ST2LANEpost";
950 case AArch64ISD::ST3LANEpost: return "AArch64ISD::ST3LANEpost";
951 case AArch64ISD::ST4LANEpost: return "AArch64ISD::ST4LANEpost";
952 case AArch64ISD::SMULL: return "AArch64ISD::SMULL";
953 case AArch64ISD::UMULL: return "AArch64ISD::UMULL";
959 AArch64TargetLowering::EmitF128CSEL(MachineInstr *MI,
960 MachineBasicBlock *MBB) const {
961 // We materialise the F128CSEL pseudo-instruction as some control flow and a
965 // [... previous instrs leading to comparison ...]
971 // Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB]
973 MachineFunction *MF = MBB->getParent();
974 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
975 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
976 DebugLoc DL = MI->getDebugLoc();
977 MachineFunction::iterator It = MBB;
980 unsigned DestReg = MI->getOperand(0).getReg();
981 unsigned IfTrueReg = MI->getOperand(1).getReg();
982 unsigned IfFalseReg = MI->getOperand(2).getReg();
983 unsigned CondCode = MI->getOperand(3).getImm();
984 bool NZCVKilled = MI->getOperand(4).isKill();
986 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
987 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
988 MF->insert(It, TrueBB);
989 MF->insert(It, EndBB);
991 // Transfer rest of current basic-block to EndBB
992 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
994 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
996 BuildMI(MBB, DL, TII->get(AArch64::Bcc)).addImm(CondCode).addMBB(TrueBB);
997 BuildMI(MBB, DL, TII->get(AArch64::B)).addMBB(EndBB);
998 MBB->addSuccessor(TrueBB);
999 MBB->addSuccessor(EndBB);
1001 // TrueBB falls through to the end.
1002 TrueBB->addSuccessor(EndBB);
1005 TrueBB->addLiveIn(AArch64::NZCV);
1006 EndBB->addLiveIn(AArch64::NZCV);
1009 BuildMI(*EndBB, EndBB->begin(), DL, TII->get(AArch64::PHI), DestReg)
1015 MI->eraseFromParent();
1020 AArch64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
1021 MachineBasicBlock *BB) const {
1022 switch (MI->getOpcode()) {
1027 llvm_unreachable("Unexpected instruction for custom inserter!");
1029 case AArch64::F128CSEL:
1030 return EmitF128CSEL(MI, BB);
1032 case TargetOpcode::STACKMAP:
1033 case TargetOpcode::PATCHPOINT:
1034 return emitPatchPoint(MI, BB);
1038 //===----------------------------------------------------------------------===//
1039 // AArch64 Lowering private implementation.
1040 //===----------------------------------------------------------------------===//
1042 //===----------------------------------------------------------------------===//
1044 //===----------------------------------------------------------------------===//
1046 /// changeIntCCToAArch64CC - Convert a DAG integer condition code to an AArch64
1048 static AArch64CC::CondCode changeIntCCToAArch64CC(ISD::CondCode CC) {
1051 llvm_unreachable("Unknown condition code!");
1053 return AArch64CC::NE;
1055 return AArch64CC::EQ;
1057 return AArch64CC::GT;
1059 return AArch64CC::GE;
1061 return AArch64CC::LT;
1063 return AArch64CC::LE;
1065 return AArch64CC::HI;
1067 return AArch64CC::HS;
1069 return AArch64CC::LO;
1071 return AArch64CC::LS;
1075 /// changeFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64 CC.
1076 static void changeFPCCToAArch64CC(ISD::CondCode CC,
1077 AArch64CC::CondCode &CondCode,
1078 AArch64CC::CondCode &CondCode2) {
1079 CondCode2 = AArch64CC::AL;
1082 llvm_unreachable("Unknown FP condition!");
1085 CondCode = AArch64CC::EQ;
1089 CondCode = AArch64CC::GT;
1093 CondCode = AArch64CC::GE;
1096 CondCode = AArch64CC::MI;
1099 CondCode = AArch64CC::LS;
1102 CondCode = AArch64CC::MI;
1103 CondCode2 = AArch64CC::GT;
1106 CondCode = AArch64CC::VC;
1109 CondCode = AArch64CC::VS;
1112 CondCode = AArch64CC::EQ;
1113 CondCode2 = AArch64CC::VS;
1116 CondCode = AArch64CC::HI;
1119 CondCode = AArch64CC::PL;
1123 CondCode = AArch64CC::LT;
1127 CondCode = AArch64CC::LE;
1131 CondCode = AArch64CC::NE;
1136 /// changeVectorFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64
1137 /// CC usable with the vector instructions. Fewer operations are available
1138 /// without a real NZCV register, so we have to use less efficient combinations
1139 /// to get the same effect.
1140 static void changeVectorFPCCToAArch64CC(ISD::CondCode CC,
1141 AArch64CC::CondCode &CondCode,
1142 AArch64CC::CondCode &CondCode2,
1147 // Mostly the scalar mappings work fine.
1148 changeFPCCToAArch64CC(CC, CondCode, CondCode2);
1151 Invert = true; // Fallthrough
1153 CondCode = AArch64CC::MI;
1154 CondCode2 = AArch64CC::GE;
1161 // All of the compare-mask comparisons are ordered, but we can switch
1162 // between the two by a double inversion. E.g. ULE == !OGT.
1164 changeFPCCToAArch64CC(getSetCCInverse(CC, false), CondCode, CondCode2);
1169 static bool isLegalArithImmed(uint64_t C) {
1170 // Matches AArch64DAGToDAGISel::SelectArithImmed().
1171 return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
1174 static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1175 SDLoc dl, SelectionDAG &DAG) {
1176 EVT VT = LHS.getValueType();
1178 if (VT.isFloatingPoint())
1179 return DAG.getNode(AArch64ISD::FCMP, dl, VT, LHS, RHS);
1181 // The CMP instruction is just an alias for SUBS, and representing it as
1182 // SUBS means that it's possible to get CSE with subtract operations.
1183 // A later phase can perform the optimization of setting the destination
1184 // register to WZR/XZR if it ends up being unused.
1185 unsigned Opcode = AArch64ISD::SUBS;
1187 if (RHS.getOpcode() == ISD::SUB && isa<ConstantSDNode>(RHS.getOperand(0)) &&
1188 cast<ConstantSDNode>(RHS.getOperand(0))->getZExtValue() == 0 &&
1189 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1190 // We'd like to combine a (CMP op1, (sub 0, op2) into a CMN instruction on
1191 // the grounds that "op1 - (-op2) == op1 + op2". However, the C and V flags
1192 // can be set differently by this operation. It comes down to whether
1193 // "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are then
1194 // everything is fine. If not then the optimization is wrong. Thus general
1195 // comparisons are only valid if op2 != 0.
1197 // So, finally, the only LLVM-native comparisons that don't mention C and V
1198 // are SETEQ and SETNE. They're the only ones we can safely use CMN for in
1199 // the absence of information about op2.
1200 Opcode = AArch64ISD::ADDS;
1201 RHS = RHS.getOperand(1);
1202 } else if (LHS.getOpcode() == ISD::AND && isa<ConstantSDNode>(RHS) &&
1203 cast<ConstantSDNode>(RHS)->getZExtValue() == 0 &&
1204 !isUnsignedIntSetCC(CC)) {
1205 // Similarly, (CMP (and X, Y), 0) can be implemented with a TST
1206 // (a.k.a. ANDS) except that the flags are only guaranteed to work for one
1207 // of the signed comparisons.
1208 Opcode = AArch64ISD::ANDS;
1209 RHS = LHS.getOperand(1);
1210 LHS = LHS.getOperand(0);
1213 return DAG.getNode(Opcode, dl, DAG.getVTList(VT, MVT_CC), LHS, RHS)
1217 /// \defgroup AArch64CCMP CMP;CCMP matching
1219 /// These functions deal with the formation of CMP;CCMP;... sequences.
1220 /// The CCMP/CCMN/FCCMP/FCCMPE instructions allow the conditional execution of
1221 /// a comparison. They set the NZCV flags to a predefined value if their
1222 /// predicate is false. This allows to express arbitrary conjunctions, for
1223 /// example "cmp 0 (and (setCA (cmp A)) (setCB (cmp B))))"
1226 /// ccmp B, inv(CB), CA
1227 /// check for CB flags
1229 /// In general we can create code for arbitrary "... (and (and A B) C)"
1230 /// sequences. We can also implement some "or" expressions, because "(or A B)"
1231 /// is equivalent to "not (and (not A) (not B))" and we can implement some
1232 /// negation operations:
1233 /// We can negate the results of a single comparison by inverting the flags
1234 /// used when the predicate fails and inverting the flags tested in the next
1235 /// instruction; We can also negate the results of the whole previous
1236 /// conditional compare sequence by inverting the flags tested in the next
1237 /// instruction. However there is no way to negate the result of a partial
1240 /// Therefore on encountering an "or" expression we can negate the subtree on
1241 /// one side and have to be able to push the negate to the leafs of the subtree
1242 /// on the other side (see also the comments in code). As complete example:
1243 /// "or (or (setCA (cmp A)) (setCB (cmp B)))
1244 /// (and (setCC (cmp C)) (setCD (cmp D)))"
1245 /// is transformed to
1246 /// "not (and (not (and (setCC (cmp C)) (setCC (cmp D))))
1247 /// (and (not (setCA (cmp A)) (not (setCB (cmp B))))))"
1248 /// and implemented as:
1250 /// ccmp D, inv(CD), CC
1251 /// ccmp A, CA, inv(CD)
1252 /// ccmp B, CB, inv(CA)
1253 /// check for CB flags
1254 /// A counterexample is "or (and A B) (and C D)" which cannot be implemented
1255 /// by conditional compare sequences.
1258 /// Create a conditional comparison; Use CCMP, CCMN or FCCMP as appropriate.
1259 static SDValue emitConditionalComparison(SDValue LHS, SDValue RHS,
1260 ISD::CondCode CC, SDValue CCOp,
1261 SDValue Condition, unsigned NZCV,
1262 SDLoc DL, SelectionDAG &DAG) {
1263 unsigned Opcode = 0;
1264 if (LHS.getValueType().isFloatingPoint())
1265 Opcode = AArch64ISD::FCCMP;
1266 else if (RHS.getOpcode() == ISD::SUB) {
1267 SDValue SubOp0 = RHS.getOperand(0);
1268 if (const ConstantSDNode *SubOp0C = dyn_cast<ConstantSDNode>(SubOp0))
1269 if (SubOp0C->isNullValue() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1270 // See emitComparison() on why we can only do this for SETEQ and SETNE.
1271 Opcode = AArch64ISD::CCMN;
1272 RHS = RHS.getOperand(1);
1276 Opcode = AArch64ISD::CCMP;
1278 SDValue NZCVOp = DAG.getConstant(NZCV, DL, MVT::i32);
1279 return DAG.getNode(Opcode, DL, MVT_CC, LHS, RHS, NZCVOp, Condition, CCOp);
1282 /// Returns true if @p Val is a tree of AND/OR/SETCC operations.
1283 /// CanPushNegate is set to true if we can push a negate operation through
1284 /// the tree in a was that we are left with AND operations and negate operations
1285 /// at the leafs only. i.e. "not (or (or x y) z)" can be changed to
1286 /// "and (and (not x) (not y)) (not z)"; "not (or (and x y) z)" cannot be
1287 /// brought into such a form.
1288 static bool isConjunctionDisjunctionTree(const SDValue Val, bool &CanPushNegate,
1289 unsigned Depth = 0) {
1290 if (!Val.hasOneUse())
1292 unsigned Opcode = Val->getOpcode();
1293 if (Opcode == ISD::SETCC) {
1294 CanPushNegate = true;
1297 // Protect against stack overflow.
1300 if (Opcode == ISD::AND || Opcode == ISD::OR) {
1301 SDValue O0 = Val->getOperand(0);
1302 SDValue O1 = Val->getOperand(1);
1303 bool CanPushNegateL;
1304 if (!isConjunctionDisjunctionTree(O0, CanPushNegateL, Depth+1))
1306 bool CanPushNegateR;
1307 if (!isConjunctionDisjunctionTree(O1, CanPushNegateR, Depth+1))
1309 // We cannot push a negate through an AND operation (it would become an OR),
1310 // we can however change a (not (or x y)) to (and (not x) (not y)) if we can
1311 // push the negate through the x/y subtrees.
1312 CanPushNegate = (Opcode == ISD::OR) && CanPushNegateL && CanPushNegateR;
1318 /// Emit conjunction or disjunction tree with the CMP/FCMP followed by a chain
1319 /// of CCMP/CFCMP ops. See @ref AArch64CCMP.
1320 /// Tries to transform the given i1 producing node @p Val to a series compare
1321 /// and conditional compare operations. @returns an NZCV flags producing node
1322 /// and sets @p OutCC to the flags that should be tested or returns SDValue() if
1323 /// transformation was not possible.
1324 /// On recursive invocations @p PushNegate may be set to true to have negation
1325 /// effects pushed to the tree leafs; @p Predicate is an NZCV flag predicate
1326 /// for the comparisons in the current subtree; @p Depth limits the search
1327 /// depth to avoid stack overflow.
1328 static SDValue emitConjunctionDisjunctionTree(SelectionDAG &DAG, SDValue Val,
1329 AArch64CC::CondCode &OutCC, bool PushNegate = false,
1330 SDValue CCOp = SDValue(), AArch64CC::CondCode Predicate = AArch64CC::AL,
1331 unsigned Depth = 0) {
1332 // We're at a tree leaf, produce a conditional comparison operation.
1333 unsigned Opcode = Val->getOpcode();
1334 if (Opcode == ISD::SETCC) {
1335 SDValue LHS = Val->getOperand(0);
1336 SDValue RHS = Val->getOperand(1);
1337 ISD::CondCode CC = cast<CondCodeSDNode>(Val->getOperand(2))->get();
1338 bool isInteger = LHS.getValueType().isInteger();
1340 CC = getSetCCInverse(CC, isInteger);
1342 // Determine OutCC and handle FP special case.
1344 OutCC = changeIntCCToAArch64CC(CC);
1346 assert(LHS.getValueType().isFloatingPoint());
1347 AArch64CC::CondCode ExtraCC;
1348 changeFPCCToAArch64CC(CC, OutCC, ExtraCC);
1349 // Surpisingly some floating point conditions can't be tested with a
1350 // single condition code. Construct an additional comparison in this case.
1351 // See comment below on how we deal with OR conditions.
1352 if (ExtraCC != AArch64CC::AL) {
1354 if (!CCOp.getNode())
1355 ExtraCmp = emitComparison(LHS, RHS, CC, DL, DAG);
1357 SDValue ConditionOp = DAG.getConstant(Predicate, DL, MVT_CC);
1358 // Note that we want the inverse of ExtraCC, so NZCV is not inversed.
1359 unsigned NZCV = AArch64CC::getNZCVToSatisfyCondCode(ExtraCC);
1360 ExtraCmp = emitConditionalComparison(LHS, RHS, CC, CCOp, ConditionOp,
1364 Predicate = AArch64CC::getInvertedCondCode(ExtraCC);
1365 OutCC = AArch64CC::getInvertedCondCode(OutCC);
1369 // Produce a normal comparison if we are first in the chain
1370 if (!CCOp.getNode())
1371 return emitComparison(LHS, RHS, CC, DL, DAG);
1372 // Otherwise produce a ccmp.
1373 SDValue ConditionOp = DAG.getConstant(Predicate, DL, MVT_CC);
1374 AArch64CC::CondCode InvOutCC = AArch64CC::getInvertedCondCode(OutCC);
1375 unsigned NZCV = AArch64CC::getNZCVToSatisfyCondCode(InvOutCC);
1376 return emitConditionalComparison(LHS, RHS, CC, CCOp, ConditionOp, NZCV, DL,
1378 } else if ((Opcode != ISD::AND && Opcode != ISD::OR) || !Val->hasOneUse())
1381 assert((Opcode == ISD::OR || !PushNegate)
1382 && "Can only push negate through OR operation");
1384 // Check if both sides can be transformed.
1385 SDValue LHS = Val->getOperand(0);
1386 SDValue RHS = Val->getOperand(1);
1387 bool CanPushNegateL;
1388 if (!isConjunctionDisjunctionTree(LHS, CanPushNegateL, Depth+1))
1390 bool CanPushNegateR;
1391 if (!isConjunctionDisjunctionTree(RHS, CanPushNegateR, Depth+1))
1394 // Do we need to negate our operands?
1395 bool NegateOperands = Opcode == ISD::OR;
1396 // We can negate the results of all previous operations by inverting the
1397 // predicate flags giving us a free negation for one side. For the other side
1398 // we need to be able to push the negation to the leafs of the tree.
1399 if (NegateOperands) {
1400 if (!CanPushNegateL && !CanPushNegateR)
1402 // Order the side where we can push the negate through to LHS.
1403 if (!CanPushNegateL && CanPushNegateR)
1404 std::swap(LHS, RHS);
1406 bool NeedsNegOutL = LHS->getOpcode() == ISD::OR;
1407 bool NeedsNegOutR = RHS->getOpcode() == ISD::OR;
1408 if (NeedsNegOutL && NeedsNegOutR)
1410 // Order the side where we need to negate the output flags to RHS so it
1411 // gets emitted first.
1413 std::swap(LHS, RHS);
1416 // Emit RHS. If we want to negate the tree we only need to push a negate
1417 // through if we are already in a PushNegate case, otherwise we can negate
1418 // the "flags to test" afterwards.
1419 AArch64CC::CondCode RHSCC;
1420 SDValue CmpR = emitConjunctionDisjunctionTree(DAG, RHS, RHSCC, PushNegate,
1421 CCOp, Predicate, Depth+1);
1422 if (NegateOperands && !PushNegate)
1423 RHSCC = AArch64CC::getInvertedCondCode(RHSCC);
1424 // Emit LHS. We must push the negate through if we need to negate it.
1425 SDValue CmpL = emitConjunctionDisjunctionTree(DAG, LHS, OutCC, NegateOperands,
1426 CmpR, RHSCC, Depth+1);
1427 // If we transformed an OR to and AND then we have to negate the result
1428 // (or absorb a PushNegate resulting in a double negation).
1429 if (Opcode == ISD::OR && !PushNegate)
1430 OutCC = AArch64CC::getInvertedCondCode(OutCC);
1436 static SDValue getAArch64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1437 SDValue &AArch64cc, SelectionDAG &DAG, SDLoc dl) {
1438 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1439 EVT VT = RHS.getValueType();
1440 uint64_t C = RHSC->getZExtValue();
1441 if (!isLegalArithImmed(C)) {
1442 // Constant does not fit, try adjusting it by one?
1448 if ((VT == MVT::i32 && C != 0x80000000 &&
1449 isLegalArithImmed((uint32_t)(C - 1))) ||
1450 (VT == MVT::i64 && C != 0x80000000ULL &&
1451 isLegalArithImmed(C - 1ULL))) {
1452 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1453 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1454 RHS = DAG.getConstant(C, dl, VT);
1459 if ((VT == MVT::i32 && C != 0 &&
1460 isLegalArithImmed((uint32_t)(C - 1))) ||
1461 (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) {
1462 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1463 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1464 RHS = DAG.getConstant(C, dl, VT);
1469 if ((VT == MVT::i32 && C != INT32_MAX &&
1470 isLegalArithImmed((uint32_t)(C + 1))) ||
1471 (VT == MVT::i64 && C != INT64_MAX &&
1472 isLegalArithImmed(C + 1ULL))) {
1473 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1474 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1475 RHS = DAG.getConstant(C, dl, VT);
1480 if ((VT == MVT::i32 && C != UINT32_MAX &&
1481 isLegalArithImmed((uint32_t)(C + 1))) ||
1482 (VT == MVT::i64 && C != UINT64_MAX &&
1483 isLegalArithImmed(C + 1ULL))) {
1484 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1485 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1486 RHS = DAG.getConstant(C, dl, VT);
1493 AArch64CC::CondCode AArch64CC;
1494 if ((CC == ISD::SETEQ || CC == ISD::SETNE) && isa<ConstantSDNode>(RHS)) {
1495 const ConstantSDNode *RHSC = cast<ConstantSDNode>(RHS);
1497 // The imm operand of ADDS is an unsigned immediate, in the range 0 to 4095.
1498 // For the i8 operand, the largest immediate is 255, so this can be easily
1499 // encoded in the compare instruction. For the i16 operand, however, the
1500 // largest immediate cannot be encoded in the compare.
1501 // Therefore, use a sign extending load and cmn to avoid materializing the
1502 // -1 constant. For example,
1504 // ldrh w0, [x0, #0]
1507 // ldrsh w0, [x0, #0]
1509 // Fundamental, we're relying on the property that (zext LHS) == (zext RHS)
1510 // if and only if (sext LHS) == (sext RHS). The checks are in place to
1511 // ensure both the LHS and RHS are truly zero extended and to make sure the
1512 // transformation is profitable.
1513 if ((RHSC->getZExtValue() >> 16 == 0) && isa<LoadSDNode>(LHS) &&
1514 cast<LoadSDNode>(LHS)->getExtensionType() == ISD::ZEXTLOAD &&
1515 cast<LoadSDNode>(LHS)->getMemoryVT() == MVT::i16 &&
1516 LHS.getNode()->hasNUsesOfValue(1, 0)) {
1517 int16_t ValueofRHS = cast<ConstantSDNode>(RHS)->getZExtValue();
1518 if (ValueofRHS < 0 && isLegalArithImmed(-ValueofRHS)) {
1520 DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, LHS.getValueType(), LHS,
1521 DAG.getValueType(MVT::i16));
1522 Cmp = emitComparison(SExt, DAG.getConstant(ValueofRHS, dl,
1523 RHS.getValueType()),
1525 AArch64CC = changeIntCCToAArch64CC(CC);
1529 if (!Cmp && (RHSC->isNullValue() || RHSC->isOne())) {
1530 if ((Cmp = emitConjunctionDisjunctionTree(DAG, LHS, AArch64CC))) {
1531 if ((CC == ISD::SETNE) ^ RHSC->isNullValue())
1532 AArch64CC = AArch64CC::getInvertedCondCode(AArch64CC);
1538 Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
1539 AArch64CC = changeIntCCToAArch64CC(CC);
1541 AArch64cc = DAG.getConstant(AArch64CC, dl, MVT_CC);
1545 static std::pair<SDValue, SDValue>
1546 getAArch64XALUOOp(AArch64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
1547 assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
1548 "Unsupported value type");
1549 SDValue Value, Overflow;
1551 SDValue LHS = Op.getOperand(0);
1552 SDValue RHS = Op.getOperand(1);
1554 switch (Op.getOpcode()) {
1556 llvm_unreachable("Unknown overflow instruction!");
1558 Opc = AArch64ISD::ADDS;
1562 Opc = AArch64ISD::ADDS;
1566 Opc = AArch64ISD::SUBS;
1570 Opc = AArch64ISD::SUBS;
1573 // Multiply needs a little bit extra work.
1577 bool IsSigned = Op.getOpcode() == ISD::SMULO;
1578 if (Op.getValueType() == MVT::i32) {
1579 unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1580 // For a 32 bit multiply with overflow check we want the instruction
1581 // selector to generate a widening multiply (SMADDL/UMADDL). For that we
1582 // need to generate the following pattern:
1583 // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b))
1584 LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
1585 RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
1586 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1587 SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul,
1588 DAG.getConstant(0, DL, MVT::i64));
1589 // On AArch64 the upper 32 bits are always zero extended for a 32 bit
1590 // operation. We need to clear out the upper 32 bits, because we used a
1591 // widening multiply that wrote all 64 bits. In the end this should be a
1593 Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add);
1595 // The signed overflow check requires more than just a simple check for
1596 // any bit set in the upper 32 bits of the result. These bits could be
1597 // just the sign bits of a negative number. To perform the overflow
1598 // check we have to arithmetic shift right the 32nd bit of the result by
1599 // 31 bits. Then we compare the result to the upper 32 bits.
1600 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add,
1601 DAG.getConstant(32, DL, MVT::i64));
1602 UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits);
1603 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value,
1604 DAG.getConstant(31, DL, MVT::i64));
1605 // It is important that LowerBits is last, otherwise the arithmetic
1606 // shift will not be folded into the compare (SUBS).
1607 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32);
1608 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1611 // The overflow check for unsigned multiply is easy. We only need to
1612 // check if any of the upper 32 bits are set. This can be done with a
1613 // CMP (shifted register). For that we need to generate the following
1615 // (i64 AArch64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32)
1616 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
1617 DAG.getConstant(32, DL, MVT::i64));
1618 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1620 DAG.getNode(AArch64ISD::SUBS, DL, VTs,
1621 DAG.getConstant(0, DL, MVT::i64),
1622 UpperBits).getValue(1);
1626 assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
1627 // For the 64 bit multiply
1628 Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1630 SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
1631 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
1632 DAG.getConstant(63, DL, MVT::i64));
1633 // It is important that LowerBits is last, otherwise the arithmetic
1634 // shift will not be folded into the compare (SUBS).
1635 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1636 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1639 SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
1640 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1642 DAG.getNode(AArch64ISD::SUBS, DL, VTs,
1643 DAG.getConstant(0, DL, MVT::i64),
1644 UpperBits).getValue(1);
1651 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);
1653 // Emit the AArch64 operation with overflow check.
1654 Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
1655 Overflow = Value.getValue(1);
1657 return std::make_pair(Value, Overflow);
1660 SDValue AArch64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG,
1661 RTLIB::Libcall Call) const {
1662 SmallVector<SDValue, 2> Ops(Op->op_begin(), Op->op_end());
1663 return makeLibCall(DAG, Call, MVT::f128, &Ops[0], Ops.size(), false,
1667 static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) {
1668 SDValue Sel = Op.getOperand(0);
1669 SDValue Other = Op.getOperand(1);
1671 // If neither operand is a SELECT_CC, give up.
1672 if (Sel.getOpcode() != ISD::SELECT_CC)
1673 std::swap(Sel, Other);
1674 if (Sel.getOpcode() != ISD::SELECT_CC)
1677 // The folding we want to perform is:
1678 // (xor x, (select_cc a, b, cc, 0, -1) )
1680 // (csel x, (xor x, -1), cc ...)
1682 // The latter will get matched to a CSINV instruction.
1684 ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
1685 SDValue LHS = Sel.getOperand(0);
1686 SDValue RHS = Sel.getOperand(1);
1687 SDValue TVal = Sel.getOperand(2);
1688 SDValue FVal = Sel.getOperand(3);
1691 // FIXME: This could be generalized to non-integer comparisons.
1692 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
1695 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
1696 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
1698 // The values aren't constants, this isn't the pattern we're looking for.
1699 if (!CFVal || !CTVal)
1702 // We can commute the SELECT_CC by inverting the condition. This
1703 // might be needed to make this fit into a CSINV pattern.
1704 if (CTVal->isAllOnesValue() && CFVal->isNullValue()) {
1705 std::swap(TVal, FVal);
1706 std::swap(CTVal, CFVal);
1707 CC = ISD::getSetCCInverse(CC, true);
1710 // If the constants line up, perform the transform!
1711 if (CTVal->isNullValue() && CFVal->isAllOnesValue()) {
1713 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
1716 TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other,
1717 DAG.getConstant(-1ULL, dl, Other.getValueType()));
1719 return DAG.getNode(AArch64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal,
1726 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
1727 EVT VT = Op.getValueType();
1729 // Let legalize expand this if it isn't a legal type yet.
1730 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
1733 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
1736 bool ExtraOp = false;
1737 switch (Op.getOpcode()) {
1739 llvm_unreachable("Invalid code");
1741 Opc = AArch64ISD::ADDS;
1744 Opc = AArch64ISD::SUBS;
1747 Opc = AArch64ISD::ADCS;
1751 Opc = AArch64ISD::SBCS;
1757 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1));
1758 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1),
1762 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
1763 // Let legalize expand this if it isn't a legal type yet.
1764 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
1768 AArch64CC::CondCode CC;
1769 // The actual operation that sets the overflow or carry flag.
1770 SDValue Value, Overflow;
1771 std::tie(Value, Overflow) = getAArch64XALUOOp(CC, Op, DAG);
1773 // We use 0 and 1 as false and true values.
1774 SDValue TVal = DAG.getConstant(1, dl, MVT::i32);
1775 SDValue FVal = DAG.getConstant(0, dl, MVT::i32);
1777 // We use an inverted condition, because the conditional select is inverted
1778 // too. This will allow it to be selected to a single instruction:
1779 // CSINC Wd, WZR, WZR, invert(cond).
1780 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), dl, MVT::i32);
1781 Overflow = DAG.getNode(AArch64ISD::CSEL, dl, MVT::i32, FVal, TVal,
1784 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
1785 return DAG.getNode(ISD::MERGE_VALUES, dl, VTs, Value, Overflow);
1788 // Prefetch operands are:
1789 // 1: Address to prefetch
1791 // 3: int locality (0 = no locality ... 3 = extreme locality)
1792 // 4: bool isDataCache
1793 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
1795 unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
1796 unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
1797 unsigned IsData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
1799 bool IsStream = !Locality;
1800 // When the locality number is set
1802 // The front-end should have filtered out the out-of-range values
1803 assert(Locality <= 3 && "Prefetch locality out-of-range");
1804 // The locality degree is the opposite of the cache speed.
1805 // Put the number the other way around.
1806 // The encoding starts at 0 for level 1
1807 Locality = 3 - Locality;
1810 // built the mask value encoding the expected behavior.
1811 unsigned PrfOp = (IsWrite << 4) | // Load/Store bit
1812 (!IsData << 3) | // IsDataCache bit
1813 (Locality << 1) | // Cache level bits
1814 (unsigned)IsStream; // Stream bit
1815 return DAG.getNode(AArch64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
1816 DAG.getConstant(PrfOp, DL, MVT::i32), Op.getOperand(1));
1819 SDValue AArch64TargetLowering::LowerFP_EXTEND(SDValue Op,
1820 SelectionDAG &DAG) const {
1821 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
1824 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
1826 return LowerF128Call(Op, DAG, LC);
1829 SDValue AArch64TargetLowering::LowerFP_ROUND(SDValue Op,
1830 SelectionDAG &DAG) const {
1831 if (Op.getOperand(0).getValueType() != MVT::f128) {
1832 // It's legal except when f128 is involved
1837 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
1839 // FP_ROUND node has a second operand indicating whether it is known to be
1840 // precise. That doesn't take part in the LibCall so we can't directly use
1842 SDValue SrcVal = Op.getOperand(0);
1843 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
1844 /*isSigned*/ false, SDLoc(Op)).first;
1847 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1848 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1849 // Any additional optimization in this function should be recorded
1850 // in the cost tables.
1851 EVT InVT = Op.getOperand(0).getValueType();
1852 EVT VT = Op.getValueType();
1854 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1857 DAG.getNode(Op.getOpcode(), dl, InVT.changeVectorElementTypeToInteger(),
1859 return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
1862 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1865 MVT::getVectorVT(MVT::getFloatingPointVT(VT.getScalarSizeInBits()),
1866 VT.getVectorNumElements());
1867 SDValue Ext = DAG.getNode(ISD::FP_EXTEND, dl, ExtVT, Op.getOperand(0));
1868 return DAG.getNode(Op.getOpcode(), dl, VT, Ext);
1871 // Type changing conversions are illegal.
1875 SDValue AArch64TargetLowering::LowerFP_TO_INT(SDValue Op,
1876 SelectionDAG &DAG) const {
1877 if (Op.getOperand(0).getValueType().isVector())
1878 return LowerVectorFP_TO_INT(Op, DAG);
1880 // f16 conversions are promoted to f32.
1881 if (Op.getOperand(0).getValueType() == MVT::f16) {
1884 Op.getOpcode(), dl, Op.getValueType(),
1885 DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, Op.getOperand(0)));
1888 if (Op.getOperand(0).getValueType() != MVT::f128) {
1889 // It's legal except when f128 is involved
1894 if (Op.getOpcode() == ISD::FP_TO_SINT)
1895 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
1897 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
1899 SmallVector<SDValue, 2> Ops(Op->op_begin(), Op->op_end());
1900 return makeLibCall(DAG, LC, Op.getValueType(), &Ops[0], Ops.size(), false,
1904 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1905 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1906 // Any additional optimization in this function should be recorded
1907 // in the cost tables.
1908 EVT VT = Op.getValueType();
1910 SDValue In = Op.getOperand(0);
1911 EVT InVT = In.getValueType();
1913 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1915 MVT::getVectorVT(MVT::getFloatingPointVT(InVT.getScalarSizeInBits()),
1916 InVT.getVectorNumElements());
1917 In = DAG.getNode(Op.getOpcode(), dl, CastVT, In);
1918 return DAG.getNode(ISD::FP_ROUND, dl, VT, In, DAG.getIntPtrConstant(0, dl));
1921 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1923 Op.getOpcode() == ISD::SINT_TO_FP ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1924 EVT CastVT = VT.changeVectorElementTypeToInteger();
1925 In = DAG.getNode(CastOpc, dl, CastVT, In);
1926 return DAG.getNode(Op.getOpcode(), dl, VT, In);
1932 SDValue AArch64TargetLowering::LowerINT_TO_FP(SDValue Op,
1933 SelectionDAG &DAG) const {
1934 if (Op.getValueType().isVector())
1935 return LowerVectorINT_TO_FP(Op, DAG);
1937 // f16 conversions are promoted to f32.
1938 if (Op.getValueType() == MVT::f16) {
1941 ISD::FP_ROUND, dl, MVT::f16,
1942 DAG.getNode(Op.getOpcode(), dl, MVT::f32, Op.getOperand(0)),
1943 DAG.getIntPtrConstant(0, dl));
1946 // i128 conversions are libcalls.
1947 if (Op.getOperand(0).getValueType() == MVT::i128)
1950 // Other conversions are legal, unless it's to the completely software-based
1952 if (Op.getValueType() != MVT::f128)
1956 if (Op.getOpcode() == ISD::SINT_TO_FP)
1957 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1959 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1961 return LowerF128Call(Op, DAG, LC);
1964 SDValue AArch64TargetLowering::LowerFSINCOS(SDValue Op,
1965 SelectionDAG &DAG) const {
1966 // For iOS, we want to call an alternative entry point: __sincos_stret,
1967 // which returns the values in two S / D registers.
1969 SDValue Arg = Op.getOperand(0);
1970 EVT ArgVT = Arg.getValueType();
1971 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1978 Entry.isSExt = false;
1979 Entry.isZExt = false;
1980 Args.push_back(Entry);
1982 const char *LibcallName =
1983 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
1985 DAG.getExternalSymbol(LibcallName, getPointerTy(DAG.getDataLayout()));
1987 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
1988 TargetLowering::CallLoweringInfo CLI(DAG);
1989 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
1990 .setCallee(CallingConv::Fast, RetTy, Callee, std::move(Args), 0);
1992 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1993 return CallResult.first;
1996 static SDValue LowerBITCAST(SDValue Op, SelectionDAG &DAG) {
1997 if (Op.getValueType() != MVT::f16)
2000 assert(Op.getOperand(0).getValueType() == MVT::i16);
2003 Op = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op.getOperand(0));
2004 Op = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Op);
2006 DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL, MVT::f16, Op,
2007 DAG.getTargetConstant(AArch64::hsub, DL, MVT::i32)),
2011 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
2012 if (OrigVT.getSizeInBits() >= 64)
2015 assert(OrigVT.isSimple() && "Expecting a simple value type");
2017 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
2018 switch (OrigSimpleTy) {
2019 default: llvm_unreachable("Unexpected Vector Type");
2028 static SDValue addRequiredExtensionForVectorMULL(SDValue N, SelectionDAG &DAG,
2031 unsigned ExtOpcode) {
2032 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
2033 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
2034 // 64-bits we need to insert a new extension so that it will be 64-bits.
2035 assert(ExtTy.is128BitVector() && "Unexpected extension size");
2036 if (OrigTy.getSizeInBits() >= 64)
2039 // Must extend size to at least 64 bits to be used as an operand for VMULL.
2040 EVT NewVT = getExtensionTo64Bits(OrigTy);
2042 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
2045 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
2047 EVT VT = N->getValueType(0);
2049 if (N->getOpcode() != ISD::BUILD_VECTOR)
2052 for (const SDValue &Elt : N->op_values()) {
2053 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
2054 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
2055 unsigned HalfSize = EltSize / 2;
2057 if (!isIntN(HalfSize, C->getSExtValue()))
2060 if (!isUIntN(HalfSize, C->getZExtValue()))
2071 static SDValue skipExtensionForVectorMULL(SDNode *N, SelectionDAG &DAG) {
2072 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
2073 return addRequiredExtensionForVectorMULL(N->getOperand(0), DAG,
2074 N->getOperand(0)->getValueType(0),
2078 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
2079 EVT VT = N->getValueType(0);
2081 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
2082 unsigned NumElts = VT.getVectorNumElements();
2083 MVT TruncVT = MVT::getIntegerVT(EltSize);
2084 SmallVector<SDValue, 8> Ops;
2085 for (unsigned i = 0; i != NumElts; ++i) {
2086 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
2087 const APInt &CInt = C->getAPIntValue();
2088 // Element types smaller than 32 bits are not legal, so use i32 elements.
2089 // The values are implicitly truncated so sext vs. zext doesn't matter.
2090 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), dl, MVT::i32));
2092 return DAG.getNode(ISD::BUILD_VECTOR, dl,
2093 MVT::getVectorVT(TruncVT, NumElts), Ops);
2096 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
2097 if (N->getOpcode() == ISD::SIGN_EXTEND)
2099 if (isExtendedBUILD_VECTOR(N, DAG, true))
2104 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
2105 if (N->getOpcode() == ISD::ZERO_EXTEND)
2107 if (isExtendedBUILD_VECTOR(N, DAG, false))
2112 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
2113 unsigned Opcode = N->getOpcode();
2114 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
2115 SDNode *N0 = N->getOperand(0).getNode();
2116 SDNode *N1 = N->getOperand(1).getNode();
2117 return N0->hasOneUse() && N1->hasOneUse() &&
2118 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
2123 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
2124 unsigned Opcode = N->getOpcode();
2125 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
2126 SDNode *N0 = N->getOperand(0).getNode();
2127 SDNode *N1 = N->getOperand(1).getNode();
2128 return N0->hasOneUse() && N1->hasOneUse() &&
2129 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
2134 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
2135 // Multiplications are only custom-lowered for 128-bit vectors so that
2136 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
2137 EVT VT = Op.getValueType();
2138 assert(VT.is128BitVector() && VT.isInteger() &&
2139 "unexpected type for custom-lowering ISD::MUL");
2140 SDNode *N0 = Op.getOperand(0).getNode();
2141 SDNode *N1 = Op.getOperand(1).getNode();
2142 unsigned NewOpc = 0;
2144 bool isN0SExt = isSignExtended(N0, DAG);
2145 bool isN1SExt = isSignExtended(N1, DAG);
2146 if (isN0SExt && isN1SExt)
2147 NewOpc = AArch64ISD::SMULL;
2149 bool isN0ZExt = isZeroExtended(N0, DAG);
2150 bool isN1ZExt = isZeroExtended(N1, DAG);
2151 if (isN0ZExt && isN1ZExt)
2152 NewOpc = AArch64ISD::UMULL;
2153 else if (isN1SExt || isN1ZExt) {
2154 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
2155 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
2156 if (isN1SExt && isAddSubSExt(N0, DAG)) {
2157 NewOpc = AArch64ISD::SMULL;
2159 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
2160 NewOpc = AArch64ISD::UMULL;
2162 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
2164 NewOpc = AArch64ISD::UMULL;
2170 if (VT == MVT::v2i64)
2171 // Fall through to expand this. It is not legal.
2174 // Other vector multiplications are legal.
2179 // Legalize to a S/UMULL instruction
2182 SDValue Op1 = skipExtensionForVectorMULL(N1, DAG);
2184 Op0 = skipExtensionForVectorMULL(N0, DAG);
2185 assert(Op0.getValueType().is64BitVector() &&
2186 Op1.getValueType().is64BitVector() &&
2187 "unexpected types for extended operands to VMULL");
2188 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
2190 // Optimizing (zext A + zext B) * C, to (S/UMULL A, C) + (S/UMULL B, C) during
2191 // isel lowering to take advantage of no-stall back to back s/umul + s/umla.
2192 // This is true for CPUs with accumulate forwarding such as Cortex-A53/A57
2193 SDValue N00 = skipExtensionForVectorMULL(N0->getOperand(0).getNode(), DAG);
2194 SDValue N01 = skipExtensionForVectorMULL(N0->getOperand(1).getNode(), DAG);
2195 EVT Op1VT = Op1.getValueType();
2196 return DAG.getNode(N0->getOpcode(), DL, VT,
2197 DAG.getNode(NewOpc, DL, VT,
2198 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
2199 DAG.getNode(NewOpc, DL, VT,
2200 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
2203 SDValue AArch64TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
2204 SelectionDAG &DAG) const {
2205 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2208 default: return SDValue(); // Don't custom lower most intrinsics.
2209 case Intrinsic::aarch64_thread_pointer: {
2210 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2211 return DAG.getNode(AArch64ISD::THREAD_POINTER, dl, PtrVT);
2213 case Intrinsic::aarch64_neon_smax:
2214 return DAG.getNode(ISD::SMAX, dl, Op.getValueType(),
2215 Op.getOperand(1), Op.getOperand(2));
2216 case Intrinsic::aarch64_neon_umax:
2217 return DAG.getNode(ISD::UMAX, dl, Op.getValueType(),
2218 Op.getOperand(1), Op.getOperand(2));
2219 case Intrinsic::aarch64_neon_smin:
2220 return DAG.getNode(ISD::SMIN, dl, Op.getValueType(),
2221 Op.getOperand(1), Op.getOperand(2));
2222 case Intrinsic::aarch64_neon_umin:
2223 return DAG.getNode(ISD::UMIN, dl, Op.getValueType(),
2224 Op.getOperand(1), Op.getOperand(2));
2228 SDValue AArch64TargetLowering::LowerOperation(SDValue Op,
2229 SelectionDAG &DAG) const {
2230 switch (Op.getOpcode()) {
2232 llvm_unreachable("unimplemented operand");
2235 return LowerBITCAST(Op, DAG);
2236 case ISD::GlobalAddress:
2237 return LowerGlobalAddress(Op, DAG);
2238 case ISD::GlobalTLSAddress:
2239 return LowerGlobalTLSAddress(Op, DAG);
2241 return LowerSETCC(Op, DAG);
2243 return LowerBR_CC(Op, DAG);
2245 return LowerSELECT(Op, DAG);
2246 case ISD::SELECT_CC:
2247 return LowerSELECT_CC(Op, DAG);
2248 case ISD::JumpTable:
2249 return LowerJumpTable(Op, DAG);
2250 case ISD::ConstantPool:
2251 return LowerConstantPool(Op, DAG);
2252 case ISD::BlockAddress:
2253 return LowerBlockAddress(Op, DAG);
2255 return LowerVASTART(Op, DAG);
2257 return LowerVACOPY(Op, DAG);
2259 return LowerVAARG(Op, DAG);
2264 return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
2271 return LowerXALUO(Op, DAG);
2273 return LowerF128Call(Op, DAG, RTLIB::ADD_F128);
2275 return LowerF128Call(Op, DAG, RTLIB::SUB_F128);
2277 return LowerF128Call(Op, DAG, RTLIB::MUL_F128);
2279 return LowerF128Call(Op, DAG, RTLIB::DIV_F128);
2281 return LowerFP_ROUND(Op, DAG);
2282 case ISD::FP_EXTEND:
2283 return LowerFP_EXTEND(Op, DAG);
2284 case ISD::FRAMEADDR:
2285 return LowerFRAMEADDR(Op, DAG);
2286 case ISD::RETURNADDR:
2287 return LowerRETURNADDR(Op, DAG);
2288 case ISD::INSERT_VECTOR_ELT:
2289 return LowerINSERT_VECTOR_ELT(Op, DAG);
2290 case ISD::EXTRACT_VECTOR_ELT:
2291 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2292 case ISD::BUILD_VECTOR:
2293 return LowerBUILD_VECTOR(Op, DAG);
2294 case ISD::VECTOR_SHUFFLE:
2295 return LowerVECTOR_SHUFFLE(Op, DAG);
2296 case ISD::EXTRACT_SUBVECTOR:
2297 return LowerEXTRACT_SUBVECTOR(Op, DAG);
2301 return LowerVectorSRA_SRL_SHL(Op, DAG);
2302 case ISD::SHL_PARTS:
2303 return LowerShiftLeftParts(Op, DAG);
2304 case ISD::SRL_PARTS:
2305 case ISD::SRA_PARTS:
2306 return LowerShiftRightParts(Op, DAG);
2308 return LowerCTPOP(Op, DAG);
2309 case ISD::FCOPYSIGN:
2310 return LowerFCOPYSIGN(Op, DAG);
2312 return LowerVectorAND(Op, DAG);
2314 return LowerVectorOR(Op, DAG);
2316 return LowerXOR(Op, DAG);
2318 return LowerPREFETCH(Op, DAG);
2319 case ISD::SINT_TO_FP:
2320 case ISD::UINT_TO_FP:
2321 return LowerINT_TO_FP(Op, DAG);
2322 case ISD::FP_TO_SINT:
2323 case ISD::FP_TO_UINT:
2324 return LowerFP_TO_INT(Op, DAG);
2326 return LowerFSINCOS(Op, DAG);
2328 return LowerMUL(Op, DAG);
2329 case ISD::INTRINSIC_WO_CHAIN:
2330 return LowerINTRINSIC_WO_CHAIN(Op, DAG);
2334 /// getFunctionAlignment - Return the Log2 alignment of this function.
2335 unsigned AArch64TargetLowering::getFunctionAlignment(const Function *F) const {
2339 //===----------------------------------------------------------------------===//
2340 // Calling Convention Implementation
2341 //===----------------------------------------------------------------------===//
2343 #include "AArch64GenCallingConv.inc"
2345 /// Selects the correct CCAssignFn for a given CallingConvention value.
2346 CCAssignFn *AArch64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
2347 bool IsVarArg) const {
2350 llvm_unreachable("Unsupported calling convention.");
2351 case CallingConv::WebKit_JS:
2352 return CC_AArch64_WebKit_JS;
2353 case CallingConv::GHC:
2354 return CC_AArch64_GHC;
2355 case CallingConv::C:
2356 case CallingConv::Fast:
2357 if (!Subtarget->isTargetDarwin())
2358 return CC_AArch64_AAPCS;
2359 return IsVarArg ? CC_AArch64_DarwinPCS_VarArg : CC_AArch64_DarwinPCS;
2363 SDValue AArch64TargetLowering::LowerFormalArguments(
2364 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2365 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2366 SmallVectorImpl<SDValue> &InVals) const {
2367 MachineFunction &MF = DAG.getMachineFunction();
2368 MachineFrameInfo *MFI = MF.getFrameInfo();
2370 // Assign locations to all of the incoming arguments.
2371 SmallVector<CCValAssign, 16> ArgLocs;
2372 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
2375 // At this point, Ins[].VT may already be promoted to i32. To correctly
2376 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2377 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
2378 // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
2379 // we use a special version of AnalyzeFormalArguments to pass in ValVT and
2381 unsigned NumArgs = Ins.size();
2382 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2383 unsigned CurArgIdx = 0;
2384 for (unsigned i = 0; i != NumArgs; ++i) {
2385 MVT ValVT = Ins[i].VT;
2386 if (Ins[i].isOrigArg()) {
2387 std::advance(CurOrigArg, Ins[i].getOrigArgIndex() - CurArgIdx);
2388 CurArgIdx = Ins[i].getOrigArgIndex();
2390 // Get type of the original argument.
2391 EVT ActualVT = getValueType(DAG.getDataLayout(), CurOrigArg->getType(),
2392 /*AllowUnknown*/ true);
2393 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
2394 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2395 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2397 else if (ActualMVT == MVT::i16)
2400 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2402 AssignFn(i, ValVT, ValVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
2403 assert(!Res && "Call operand has unhandled type");
2406 assert(ArgLocs.size() == Ins.size());
2407 SmallVector<SDValue, 16> ArgValues;
2408 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2409 CCValAssign &VA = ArgLocs[i];
2411 if (Ins[i].Flags.isByVal()) {
2412 // Byval is used for HFAs in the PCS, but the system should work in a
2413 // non-compliant manner for larger structs.
2414 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2415 int Size = Ins[i].Flags.getByValSize();
2416 unsigned NumRegs = (Size + 7) / 8;
2418 // FIXME: This works on big-endian for composite byvals, which are the common
2419 // case. It should also work for fundamental types too.
2421 MFI->CreateFixedObject(8 * NumRegs, VA.getLocMemOffset(), false);
2422 SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrVT);
2423 InVals.push_back(FrameIdxN);
2428 if (VA.isRegLoc()) {
2429 // Arguments stored in registers.
2430 EVT RegVT = VA.getLocVT();
2433 const TargetRegisterClass *RC;
2435 if (RegVT == MVT::i32)
2436 RC = &AArch64::GPR32RegClass;
2437 else if (RegVT == MVT::i64)
2438 RC = &AArch64::GPR64RegClass;
2439 else if (RegVT == MVT::f16)
2440 RC = &AArch64::FPR16RegClass;
2441 else if (RegVT == MVT::f32)
2442 RC = &AArch64::FPR32RegClass;
2443 else if (RegVT == MVT::f64 || RegVT.is64BitVector())
2444 RC = &AArch64::FPR64RegClass;
2445 else if (RegVT == MVT::f128 || RegVT.is128BitVector())
2446 RC = &AArch64::FPR128RegClass;
2448 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2450 // Transform the arguments in physical registers into virtual ones.
2451 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2452 ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2454 // If this is an 8, 16 or 32-bit value, it is really passed promoted
2455 // to 64 bits. Insert an assert[sz]ext to capture this, then
2456 // truncate to the right size.
2457 switch (VA.getLocInfo()) {
2459 llvm_unreachable("Unknown loc info!");
2460 case CCValAssign::Full:
2462 case CCValAssign::BCvt:
2463 ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue);
2465 case CCValAssign::AExt:
2466 case CCValAssign::SExt:
2467 case CCValAssign::ZExt:
2468 // SelectionDAGBuilder will insert appropriate AssertZExt & AssertSExt
2469 // nodes after our lowering.
2470 assert(RegVT == Ins[i].VT && "incorrect register location selected");
2474 InVals.push_back(ArgValue);
2476 } else { // VA.isRegLoc()
2477 assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
2478 unsigned ArgOffset = VA.getLocMemOffset();
2479 unsigned ArgSize = VA.getValVT().getSizeInBits() / 8;
2481 uint32_t BEAlign = 0;
2482 if (!Subtarget->isLittleEndian() && ArgSize < 8 &&
2483 !Ins[i].Flags.isInConsecutiveRegs())
2484 BEAlign = 8 - ArgSize;
2486 int FI = MFI->CreateFixedObject(ArgSize, ArgOffset + BEAlign, true);
2488 // Create load nodes to retrieve arguments from the stack.
2489 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2492 // For NON_EXTLOAD, generic code in getLoad assert(ValVT == MemVT)
2493 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
2494 MVT MemVT = VA.getValVT();
2496 switch (VA.getLocInfo()) {
2499 case CCValAssign::BCvt:
2500 MemVT = VA.getLocVT();
2502 case CCValAssign::SExt:
2503 ExtType = ISD::SEXTLOAD;
2505 case CCValAssign::ZExt:
2506 ExtType = ISD::ZEXTLOAD;
2508 case CCValAssign::AExt:
2509 ExtType = ISD::EXTLOAD;
2513 ArgValue = DAG.getExtLoad(
2514 ExtType, DL, VA.getLocVT(), Chain, FIN,
2515 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
2516 MemVT, false, false, false, 0);
2518 InVals.push_back(ArgValue);
2524 if (!Subtarget->isTargetDarwin()) {
2525 // The AAPCS variadic function ABI is identical to the non-variadic
2526 // one. As a result there may be more arguments in registers and we should
2527 // save them for future reference.
2528 saveVarArgRegisters(CCInfo, DAG, DL, Chain);
2531 AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>();
2532 // This will point to the next argument passed via stack.
2533 unsigned StackOffset = CCInfo.getNextStackOffset();
2534 // We currently pass all varargs at 8-byte alignment.
2535 StackOffset = ((StackOffset + 7) & ~7);
2536 AFI->setVarArgsStackIndex(MFI->CreateFixedObject(4, StackOffset, true));
2539 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2540 unsigned StackArgSize = CCInfo.getNextStackOffset();
2541 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2542 if (DoesCalleeRestoreStack(CallConv, TailCallOpt)) {
2543 // This is a non-standard ABI so by fiat I say we're allowed to make full
2544 // use of the stack area to be popped, which must be aligned to 16 bytes in
2546 StackArgSize = RoundUpToAlignment(StackArgSize, 16);
2548 // If we're expected to restore the stack (e.g. fastcc) then we'll be adding
2549 // a multiple of 16.
2550 FuncInfo->setArgumentStackToRestore(StackArgSize);
2552 // This realignment carries over to the available bytes below. Our own
2553 // callers will guarantee the space is free by giving an aligned value to
2556 // Even if we're not expected to free up the space, it's useful to know how
2557 // much is there while considering tail calls (because we can reuse it).
2558 FuncInfo->setBytesInStackArgArea(StackArgSize);
2563 void AArch64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
2564 SelectionDAG &DAG, SDLoc DL,
2565 SDValue &Chain) const {
2566 MachineFunction &MF = DAG.getMachineFunction();
2567 MachineFrameInfo *MFI = MF.getFrameInfo();
2568 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2569 auto PtrVT = getPointerTy(DAG.getDataLayout());
2571 SmallVector<SDValue, 8> MemOps;
2573 static const MCPhysReg GPRArgRegs[] = { AArch64::X0, AArch64::X1, AArch64::X2,
2574 AArch64::X3, AArch64::X4, AArch64::X5,
2575 AArch64::X6, AArch64::X7 };
2576 static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs);
2577 unsigned FirstVariadicGPR = CCInfo.getFirstUnallocated(GPRArgRegs);
2579 unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
2581 if (GPRSaveSize != 0) {
2582 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
2584 SDValue FIN = DAG.getFrameIndex(GPRIdx, PtrVT);
2586 for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
2587 unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &AArch64::GPR64RegClass);
2588 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
2589 SDValue Store = DAG.getStore(
2590 Val.getValue(1), DL, Val, FIN,
2591 MachinePointerInfo::getStack(DAG.getMachineFunction(), i * 8), false,
2593 MemOps.push_back(Store);
2595 DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getConstant(8, DL, PtrVT));
2598 FuncInfo->setVarArgsGPRIndex(GPRIdx);
2599 FuncInfo->setVarArgsGPRSize(GPRSaveSize);
2601 if (Subtarget->hasFPARMv8()) {
2602 static const MCPhysReg FPRArgRegs[] = {
2603 AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3,
2604 AArch64::Q4, AArch64::Q5, AArch64::Q6, AArch64::Q7};
2605 static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs);
2606 unsigned FirstVariadicFPR = CCInfo.getFirstUnallocated(FPRArgRegs);
2608 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
2610 if (FPRSaveSize != 0) {
2611 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
2613 SDValue FIN = DAG.getFrameIndex(FPRIdx, PtrVT);
2615 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
2616 unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &AArch64::FPR128RegClass);
2617 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f128);
2619 SDValue Store = DAG.getStore(
2620 Val.getValue(1), DL, Val, FIN,
2621 MachinePointerInfo::getStack(DAG.getMachineFunction(), i * 16),
2623 MemOps.push_back(Store);
2624 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN,
2625 DAG.getConstant(16, DL, PtrVT));
2628 FuncInfo->setVarArgsFPRIndex(FPRIdx);
2629 FuncInfo->setVarArgsFPRSize(FPRSaveSize);
2632 if (!MemOps.empty()) {
2633 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
2637 /// LowerCallResult - Lower the result values of a call into the
2638 /// appropriate copies out of appropriate physical registers.
2639 SDValue AArch64TargetLowering::LowerCallResult(
2640 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
2641 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2642 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
2643 SDValue ThisVal) const {
2644 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2645 ? RetCC_AArch64_WebKit_JS
2646 : RetCC_AArch64_AAPCS;
2647 // Assign locations to each value returned by this call.
2648 SmallVector<CCValAssign, 16> RVLocs;
2649 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2651 CCInfo.AnalyzeCallResult(Ins, RetCC);
2653 // Copy all of the result registers out of their specified physreg.
2654 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2655 CCValAssign VA = RVLocs[i];
2657 // Pass 'this' value directly from the argument to return value, to avoid
2658 // reg unit interference
2659 if (i == 0 && isThisReturn) {
2660 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 &&
2661 "unexpected return calling convention register assignment");
2662 InVals.push_back(ThisVal);
2667 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag);
2668 Chain = Val.getValue(1);
2669 InFlag = Val.getValue(2);
2671 switch (VA.getLocInfo()) {
2673 llvm_unreachable("Unknown loc info!");
2674 case CCValAssign::Full:
2676 case CCValAssign::BCvt:
2677 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
2681 InVals.push_back(Val);
2687 bool AArch64TargetLowering::isEligibleForTailCallOptimization(
2688 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
2689 bool isCalleeStructRet, bool isCallerStructRet,
2690 const SmallVectorImpl<ISD::OutputArg> &Outs,
2691 const SmallVectorImpl<SDValue> &OutVals,
2692 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
2693 // For CallingConv::C this function knows whether the ABI needs
2694 // changing. That's not true for other conventions so they will have to opt in
2696 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
2699 const MachineFunction &MF = DAG.getMachineFunction();
2700 const Function *CallerF = MF.getFunction();
2701 CallingConv::ID CallerCC = CallerF->getCallingConv();
2702 bool CCMatch = CallerCC == CalleeCC;
2704 // Byval parameters hand the function a pointer directly into the stack area
2705 // we want to reuse during a tail call. Working around this *is* possible (see
2706 // X86) but less efficient and uglier in LowerCall.
2707 for (Function::const_arg_iterator i = CallerF->arg_begin(),
2708 e = CallerF->arg_end();
2710 if (i->hasByValAttr())
2713 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
2714 if (IsTailCallConvention(CalleeCC) && CCMatch)
2719 // Externally-defined functions with weak linkage should not be
2720 // tail-called on AArch64 when the OS does not support dynamic
2721 // pre-emption of symbols, as the AAELF spec requires normal calls
2722 // to undefined weak functions to be replaced with a NOP or jump to the
2723 // next instruction. The behaviour of branch instructions in this
2724 // situation (as used for tail calls) is implementation-defined, so we
2725 // cannot rely on the linker replacing the tail call with a return.
2726 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2727 const GlobalValue *GV = G->getGlobal();
2728 const Triple &TT = getTargetMachine().getTargetTriple();
2729 if (GV->hasExternalWeakLinkage() &&
2730 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2734 // Now we search for cases where we can use a tail call without changing the
2735 // ABI. Sibcall is used in some places (particularly gcc) to refer to this
2738 // I want anyone implementing a new calling convention to think long and hard
2739 // about this assert.
2740 assert((!isVarArg || CalleeCC == CallingConv::C) &&
2741 "Unexpected variadic calling convention");
2743 if (isVarArg && !Outs.empty()) {
2744 // At least two cases here: if caller is fastcc then we can't have any
2745 // memory arguments (we'd be expected to clean up the stack afterwards). If
2746 // caller is C then we could potentially use its argument area.
2748 // FIXME: for now we take the most conservative of these in both cases:
2749 // disallow all variadic memory operands.
2750 SmallVector<CCValAssign, 16> ArgLocs;
2751 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2754 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, true));
2755 for (const CCValAssign &ArgLoc : ArgLocs)
2756 if (!ArgLoc.isRegLoc())
2760 // If the calling conventions do not match, then we'd better make sure the
2761 // results are returned in the same way as what the caller expects.
2763 SmallVector<CCValAssign, 16> RVLocs1;
2764 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2766 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForCall(CalleeCC, isVarArg));
2768 SmallVector<CCValAssign, 16> RVLocs2;
2769 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2771 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForCall(CallerCC, isVarArg));
2773 if (RVLocs1.size() != RVLocs2.size())
2775 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2776 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2778 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2780 if (RVLocs1[i].isRegLoc()) {
2781 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2784 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2790 // Nothing more to check if the callee is taking no arguments
2794 SmallVector<CCValAssign, 16> ArgLocs;
2795 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2798 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, isVarArg));
2800 const AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2802 // If the stack arguments for this call would fit into our own save area then
2803 // the call can be made tail.
2804 return CCInfo.getNextStackOffset() <= FuncInfo->getBytesInStackArgArea();
2807 SDValue AArch64TargetLowering::addTokenForArgument(SDValue Chain,
2809 MachineFrameInfo *MFI,
2810 int ClobberedFI) const {
2811 SmallVector<SDValue, 8> ArgChains;
2812 int64_t FirstByte = MFI->getObjectOffset(ClobberedFI);
2813 int64_t LastByte = FirstByte + MFI->getObjectSize(ClobberedFI) - 1;
2815 // Include the original chain at the beginning of the list. When this is
2816 // used by target LowerCall hooks, this helps legalize find the
2817 // CALLSEQ_BEGIN node.
2818 ArgChains.push_back(Chain);
2820 // Add a chain value for each stack argument corresponding
2821 for (SDNode::use_iterator U = DAG.getEntryNode().getNode()->use_begin(),
2822 UE = DAG.getEntryNode().getNode()->use_end();
2824 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
2825 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
2826 if (FI->getIndex() < 0) {
2827 int64_t InFirstByte = MFI->getObjectOffset(FI->getIndex());
2828 int64_t InLastByte = InFirstByte;
2829 InLastByte += MFI->getObjectSize(FI->getIndex()) - 1;
2831 if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
2832 (FirstByte <= InFirstByte && InFirstByte <= LastByte))
2833 ArgChains.push_back(SDValue(L, 1));
2836 // Build a tokenfactor for all the chains.
2837 return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
2840 bool AArch64TargetLowering::DoesCalleeRestoreStack(CallingConv::ID CallCC,
2841 bool TailCallOpt) const {
2842 return CallCC == CallingConv::Fast && TailCallOpt;
2845 bool AArch64TargetLowering::IsTailCallConvention(CallingConv::ID CallCC) const {
2846 return CallCC == CallingConv::Fast;
2849 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
2850 /// and add input and output parameter nodes.
2852 AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
2853 SmallVectorImpl<SDValue> &InVals) const {
2854 SelectionDAG &DAG = CLI.DAG;
2856 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2857 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2858 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2859 SDValue Chain = CLI.Chain;
2860 SDValue Callee = CLI.Callee;
2861 bool &IsTailCall = CLI.IsTailCall;
2862 CallingConv::ID CallConv = CLI.CallConv;
2863 bool IsVarArg = CLI.IsVarArg;
2865 MachineFunction &MF = DAG.getMachineFunction();
2866 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
2867 bool IsThisReturn = false;
2869 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2870 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2871 bool IsSibCall = false;
2874 // Check if it's really possible to do a tail call.
2875 IsTailCall = isEligibleForTailCallOptimization(
2876 Callee, CallConv, IsVarArg, IsStructRet,
2877 MF.getFunction()->hasStructRetAttr(), Outs, OutVals, Ins, DAG);
2878 if (!IsTailCall && CLI.CS && CLI.CS->isMustTailCall())
2879 report_fatal_error("failed to perform tail call elimination on a call "
2880 "site marked musttail");
2882 // A sibling call is one where we're under the usual C ABI and not planning
2883 // to change that but can still do a tail call:
2884 if (!TailCallOpt && IsTailCall)
2891 // Analyze operands of the call, assigning locations to each operand.
2892 SmallVector<CCValAssign, 16> ArgLocs;
2893 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
2897 // Handle fixed and variable vector arguments differently.
2898 // Variable vector arguments always go into memory.
2899 unsigned NumArgs = Outs.size();
2901 for (unsigned i = 0; i != NumArgs; ++i) {
2902 MVT ArgVT = Outs[i].VT;
2903 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2904 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
2905 /*IsVarArg=*/ !Outs[i].IsFixed);
2906 bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
2907 assert(!Res && "Call operand has unhandled type");
2911 // At this point, Outs[].VT may already be promoted to i32. To correctly
2912 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2913 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
2914 // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here
2915 // we use a special version of AnalyzeCallOperands to pass in ValVT and
2917 unsigned NumArgs = Outs.size();
2918 for (unsigned i = 0; i != NumArgs; ++i) {
2919 MVT ValVT = Outs[i].VT;
2920 // Get type of the original argument.
2921 EVT ActualVT = getValueType(DAG.getDataLayout(),
2922 CLI.getArgs()[Outs[i].OrigArgIndex].Ty,
2923 /*AllowUnknown*/ true);
2924 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT;
2925 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2926 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2927 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2929 else if (ActualMVT == MVT::i16)
2932 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2933 bool Res = AssignFn(i, ValVT, ValVT, CCValAssign::Full, ArgFlags, CCInfo);
2934 assert(!Res && "Call operand has unhandled type");
2939 // Get a count of how many bytes are to be pushed on the stack.
2940 unsigned NumBytes = CCInfo.getNextStackOffset();
2943 // Since we're not changing the ABI to make this a tail call, the memory
2944 // operands are already available in the caller's incoming argument space.
2948 // FPDiff is the byte offset of the call's argument area from the callee's.
2949 // Stores to callee stack arguments will be placed in FixedStackSlots offset
2950 // by this amount for a tail call. In a sibling call it must be 0 because the
2951 // caller will deallocate the entire stack and the callee still expects its
2952 // arguments to begin at SP+0. Completely unused for non-tail calls.
2955 if (IsTailCall && !IsSibCall) {
2956 unsigned NumReusableBytes = FuncInfo->getBytesInStackArgArea();
2958 // Since callee will pop argument stack as a tail call, we must keep the
2959 // popped size 16-byte aligned.
2960 NumBytes = RoundUpToAlignment(NumBytes, 16);
2962 // FPDiff will be negative if this tail call requires more space than we
2963 // would automatically have in our incoming argument space. Positive if we
2964 // can actually shrink the stack.
2965 FPDiff = NumReusableBytes - NumBytes;
2967 // The stack pointer must be 16-byte aligned at all times it's used for a
2968 // memory operation, which in practice means at *all* times and in
2969 // particular across call boundaries. Therefore our own arguments started at
2970 // a 16-byte aligned SP and the delta applied for the tail call should
2971 // satisfy the same constraint.
2972 assert(FPDiff % 16 == 0 && "unaligned stack on tail call");
2975 // Adjust the stack pointer for the new arguments...
2976 // These operations are automatically eliminated by the prolog/epilog pass
2978 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, DL,
2982 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, AArch64::SP,
2983 getPointerTy(DAG.getDataLayout()));
2985 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2986 SmallVector<SDValue, 8> MemOpChains;
2987 auto PtrVT = getPointerTy(DAG.getDataLayout());
2989 // Walk the register/memloc assignments, inserting copies/loads.
2990 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2991 ++i, ++realArgIdx) {
2992 CCValAssign &VA = ArgLocs[i];
2993 SDValue Arg = OutVals[realArgIdx];
2994 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2996 // Promote the value if needed.
2997 switch (VA.getLocInfo()) {
2999 llvm_unreachable("Unknown loc info!");
3000 case CCValAssign::Full:
3002 case CCValAssign::SExt:
3003 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
3005 case CCValAssign::ZExt:
3006 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
3008 case CCValAssign::AExt:
3009 if (Outs[realArgIdx].ArgVT == MVT::i1) {
3010 // AAPCS requires i1 to be zero-extended to 8-bits by the caller.
3011 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
3012 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i8, Arg);
3014 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
3016 case CCValAssign::BCvt:
3017 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
3019 case CCValAssign::FPExt:
3020 Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
3024 if (VA.isRegLoc()) {
3025 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i64) {
3026 assert(VA.getLocVT() == MVT::i64 &&
3027 "unexpected calling convention register assignment");
3028 assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
3029 "unexpected use of 'returned'");
3030 IsThisReturn = true;
3032 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3034 assert(VA.isMemLoc());
3037 MachinePointerInfo DstInfo;
3039 // FIXME: This works on big-endian for composite byvals, which are the
3040 // common case. It should also work for fundamental types too.
3041 uint32_t BEAlign = 0;
3042 unsigned OpSize = Flags.isByVal() ? Flags.getByValSize() * 8
3043 : VA.getValVT().getSizeInBits();
3044 OpSize = (OpSize + 7) / 8;
3045 if (!Subtarget->isLittleEndian() && !Flags.isByVal() &&
3046 !Flags.isInConsecutiveRegs()) {
3048 BEAlign = 8 - OpSize;
3050 unsigned LocMemOffset = VA.getLocMemOffset();
3051 int32_t Offset = LocMemOffset + BEAlign;
3052 SDValue PtrOff = DAG.getIntPtrConstant(Offset, DL);
3053 PtrOff = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, PtrOff);
3056 Offset = Offset + FPDiff;
3057 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3059 DstAddr = DAG.getFrameIndex(FI, PtrVT);
3061 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI);
3063 // Make sure any stack arguments overlapping with where we're storing
3064 // are loaded before this eventual operation. Otherwise they'll be
3066 Chain = addTokenForArgument(Chain, DAG, MF.getFrameInfo(), FI);
3068 SDValue PtrOff = DAG.getIntPtrConstant(Offset, DL);
3070 DstAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, PtrOff);
3071 DstInfo = MachinePointerInfo::getStack(DAG.getMachineFunction(),
3075 if (Outs[i].Flags.isByVal()) {
3077 DAG.getConstant(Outs[i].Flags.getByValSize(), DL, MVT::i64);
3078 SDValue Cpy = DAG.getMemcpy(
3079 Chain, DL, DstAddr, Arg, SizeNode, Outs[i].Flags.getByValAlign(),
3080 /*isVol = */ false, /*AlwaysInline = */ false,
3081 /*isTailCall = */ false,
3082 DstInfo, MachinePointerInfo());
3084 MemOpChains.push_back(Cpy);
3086 // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
3087 // promoted to a legal register type i32, we should truncate Arg back to
3089 if (VA.getValVT() == MVT::i1 || VA.getValVT() == MVT::i8 ||
3090 VA.getValVT() == MVT::i16)
3091 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Arg);
3094 DAG.getStore(Chain, DL, Arg, DstAddr, DstInfo, false, false, 0);
3095 MemOpChains.push_back(Store);
3100 if (!MemOpChains.empty())
3101 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
3103 // Build a sequence of copy-to-reg nodes chained together with token chain
3104 // and flag operands which copy the outgoing args into the appropriate regs.
3106 for (auto &RegToPass : RegsToPass) {
3107 Chain = DAG.getCopyToReg(Chain, DL, RegToPass.first,
3108 RegToPass.second, InFlag);
3109 InFlag = Chain.getValue(1);
3112 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
3113 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
3114 // node so that legalize doesn't hack it.
3115 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3116 Subtarget->isTargetMachO()) {
3117 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3118 const GlobalValue *GV = G->getGlobal();
3119 bool InternalLinkage = GV->hasInternalLinkage();
3120 if (InternalLinkage)
3121 Callee = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 0);
3124 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_GOT);
3125 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, Callee);
3127 } else if (ExternalSymbolSDNode *S =
3128 dyn_cast<ExternalSymbolSDNode>(Callee)) {
3129 const char *Sym = S->getSymbol();
3130 Callee = DAG.getTargetExternalSymbol(Sym, PtrVT, AArch64II::MO_GOT);
3131 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, Callee);
3133 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3134 const GlobalValue *GV = G->getGlobal();
3135 Callee = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 0);
3136 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3137 const char *Sym = S->getSymbol();
3138 Callee = DAG.getTargetExternalSymbol(Sym, PtrVT, 0);
3141 // We don't usually want to end the call-sequence here because we would tidy
3142 // the frame up *after* the call, however in the ABI-changing tail-call case
3143 // we've carefully laid out the parameters so that when sp is reset they'll be
3144 // in the correct location.
3145 if (IsTailCall && !IsSibCall) {
3146 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, DL, true),
3147 DAG.getIntPtrConstant(0, DL, true), InFlag, DL);
3148 InFlag = Chain.getValue(1);
3151 std::vector<SDValue> Ops;
3152 Ops.push_back(Chain);
3153 Ops.push_back(Callee);
3156 // Each tail call may have to adjust the stack by a different amount, so
3157 // this information must travel along with the operation for eventual
3158 // consumption by emitEpilogue.
3159 Ops.push_back(DAG.getTargetConstant(FPDiff, DL, MVT::i32));
3162 // Add argument registers to the end of the list so that they are known live
3164 for (auto &RegToPass : RegsToPass)
3165 Ops.push_back(DAG.getRegister(RegToPass.first,
3166 RegToPass.second.getValueType()));
3168 // Add a register mask operand representing the call-preserved registers.
3169 const uint32_t *Mask;
3170 const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();
3172 // For 'this' returns, use the X0-preserving mask if applicable
3173 Mask = TRI->getThisReturnPreservedMask(MF, CallConv);
3175 IsThisReturn = false;
3176 Mask = TRI->getCallPreservedMask(MF, CallConv);
3179 Mask = TRI->getCallPreservedMask(MF, CallConv);
3181 assert(Mask && "Missing call preserved mask for calling convention");
3182 Ops.push_back(DAG.getRegisterMask(Mask));
3184 if (InFlag.getNode())
3185 Ops.push_back(InFlag);
3187 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3189 // If we're doing a tall call, use a TC_RETURN here rather than an
3190 // actual call instruction.
3192 MF.getFrameInfo()->setHasTailCall();
3193 return DAG.getNode(AArch64ISD::TC_RETURN, DL, NodeTys, Ops);
3196 // Returns a chain and a flag for retval copy to use.
3197 Chain = DAG.getNode(AArch64ISD::CALL, DL, NodeTys, Ops);
3198 InFlag = Chain.getValue(1);
3200 uint64_t CalleePopBytes = DoesCalleeRestoreStack(CallConv, TailCallOpt)
3201 ? RoundUpToAlignment(NumBytes, 16)
3204 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, DL, true),
3205 DAG.getIntPtrConstant(CalleePopBytes, DL, true),
3208 InFlag = Chain.getValue(1);
3210 // Handle result values, copying them out of physregs into vregs that we
3212 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
3213 InVals, IsThisReturn,
3214 IsThisReturn ? OutVals[0] : SDValue());
3217 bool AArch64TargetLowering::CanLowerReturn(
3218 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
3219 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
3220 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
3221 ? RetCC_AArch64_WebKit_JS
3222 : RetCC_AArch64_AAPCS;
3223 SmallVector<CCValAssign, 16> RVLocs;
3224 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
3225 return CCInfo.CheckReturn(Outs, RetCC);
3229 AArch64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
3231 const SmallVectorImpl<ISD::OutputArg> &Outs,
3232 const SmallVectorImpl<SDValue> &OutVals,
3233 SDLoc DL, SelectionDAG &DAG) const {
3234 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
3235 ? RetCC_AArch64_WebKit_JS
3236 : RetCC_AArch64_AAPCS;
3237 SmallVector<CCValAssign, 16> RVLocs;
3238 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
3240 CCInfo.AnalyzeReturn(Outs, RetCC);
3242 // Copy the result values into the output registers.
3244 SmallVector<SDValue, 4> RetOps(1, Chain);
3245 for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size();
3246 ++i, ++realRVLocIdx) {
3247 CCValAssign &VA = RVLocs[i];
3248 assert(VA.isRegLoc() && "Can only return in registers!");
3249 SDValue Arg = OutVals[realRVLocIdx];
3251 switch (VA.getLocInfo()) {
3253 llvm_unreachable("Unknown loc info!");
3254 case CCValAssign::Full:
3255 if (Outs[i].ArgVT == MVT::i1) {
3256 // AAPCS requires i1 to be zero-extended to i8 by the producer of the
3257 // value. This is strictly redundant on Darwin (which uses "zeroext
3258 // i1"), but will be optimised out before ISel.
3259 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
3260 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
3263 case CCValAssign::BCvt:
3264 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
3268 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
3269 Flag = Chain.getValue(1);
3270 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
3273 RetOps[0] = Chain; // Update chain.
3275 // Add the flag if we have it.
3277 RetOps.push_back(Flag);
3279 return DAG.getNode(AArch64ISD::RET_FLAG, DL, MVT::Other, RetOps);
3282 //===----------------------------------------------------------------------===//
3283 // Other Lowering Code
3284 //===----------------------------------------------------------------------===//
3286 SDValue AArch64TargetLowering::LowerGlobalAddress(SDValue Op,
3287 SelectionDAG &DAG) const {
3288 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3290 const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
3291 const GlobalValue *GV = GN->getGlobal();
3292 unsigned char OpFlags =
3293 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
3295 assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
3296 "unexpected offset in global node");
3298 // This also catched the large code model case for Darwin.
3299 if ((OpFlags & AArch64II::MO_GOT) != 0) {
3300 SDValue GotAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
3301 // FIXME: Once remat is capable of dealing with instructions with register
3302 // operands, expand this into two nodes instead of using a wrapper node.
3303 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
3306 if ((OpFlags & AArch64II::MO_CONSTPOOL) != 0) {
3307 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
3308 "use of MO_CONSTPOOL only supported on small model");
3309 SDValue Hi = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, AArch64II::MO_PAGE);
3310 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3311 unsigned char LoFlags = AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
3312 SDValue Lo = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, LoFlags);
3313 SDValue PoolAddr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3314 SDValue GlobalAddr = DAG.getLoad(
3315 PtrVT, DL, DAG.getEntryNode(), PoolAddr,
3316 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
3317 /*isVolatile=*/false,
3318 /*isNonTemporal=*/true,
3319 /*isInvariant=*/true, 8);
3320 if (GN->getOffset() != 0)
3321 return DAG.getNode(ISD::ADD, DL, PtrVT, GlobalAddr,
3322 DAG.getConstant(GN->getOffset(), DL, PtrVT));
3326 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
3327 const unsigned char MO_NC = AArch64II::MO_NC;
3329 AArch64ISD::WrapperLarge, DL, PtrVT,
3330 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G3),
3331 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
3332 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
3333 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
3335 // Use ADRP/ADD or ADRP/LDR for everything else: the small model on ELF and
3336 // the only correct model on Darwin.
3337 SDValue Hi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
3338 OpFlags | AArch64II::MO_PAGE);
3339 unsigned char LoFlags = OpFlags | AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
3340 SDValue Lo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, LoFlags);
3342 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3343 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3347 /// \brief Convert a TLS address reference into the correct sequence of loads
3348 /// and calls to compute the variable's address (for Darwin, currently) and
3349 /// return an SDValue containing the final node.
3351 /// Darwin only has one TLS scheme which must be capable of dealing with the
3352 /// fully general situation, in the worst case. This means:
3353 /// + "extern __thread" declaration.
3354 /// + Defined in a possibly unknown dynamic library.
3356 /// The general system is that each __thread variable has a [3 x i64] descriptor
3357 /// which contains information used by the runtime to calculate the address. The
3358 /// only part of this the compiler needs to know about is the first xword, which
3359 /// contains a function pointer that must be called with the address of the
3360 /// entire descriptor in "x0".
3362 /// Since this descriptor may be in a different unit, in general even the
3363 /// descriptor must be accessed via an indirect load. The "ideal" code sequence
3365 /// adrp x0, _var@TLVPPAGE
3366 /// ldr x0, [x0, _var@TLVPPAGEOFF] ; x0 now contains address of descriptor
3367 /// ldr x1, [x0] ; x1 contains 1st entry of descriptor,
3368 /// ; the function pointer
3369 /// blr x1 ; Uses descriptor address in x0
3370 /// ; Address of _var is now in x0.
3372 /// If the address of _var's descriptor *is* known to the linker, then it can
3373 /// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
3374 /// a slight efficiency gain.
3376 AArch64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
3377 SelectionDAG &DAG) const {
3378 assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");
3381 MVT PtrVT = getPointerTy(DAG.getDataLayout());
3382 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3385 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3386 SDValue DescAddr = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TLVPAddr);
3388 // The first entry in the descriptor is a function pointer that we must call
3389 // to obtain the address of the variable.
3390 SDValue Chain = DAG.getEntryNode();
3391 SDValue FuncTLVGet =
3392 DAG.getLoad(MVT::i64, DL, Chain, DescAddr,
3393 MachinePointerInfo::getGOT(DAG.getMachineFunction()), false,
3395 Chain = FuncTLVGet.getValue(1);
3397 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3398 MFI->setAdjustsStack(true);
3400 // TLS calls preserve all registers except those that absolutely must be
3401 // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
3403 const uint32_t *Mask =
3404 Subtarget->getRegisterInfo()->getTLSCallPreservedMask();
3406 // Finally, we can make the call. This is just a degenerate version of a
3407 // normal AArch64 call node: x0 takes the address of the descriptor, and
3408 // returns the address of the variable in this thread.
3409 Chain = DAG.getCopyToReg(Chain, DL, AArch64::X0, DescAddr, SDValue());
3411 DAG.getNode(AArch64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
3412 Chain, FuncTLVGet, DAG.getRegister(AArch64::X0, MVT::i64),
3413 DAG.getRegisterMask(Mask), Chain.getValue(1));
3414 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Chain.getValue(1));
3417 /// When accessing thread-local variables under either the general-dynamic or
3418 /// local-dynamic system, we make a "TLS-descriptor" call. The variable will
3419 /// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
3420 /// is a function pointer to carry out the resolution.
3422 /// The sequence is:
3423 /// adrp x0, :tlsdesc:var
3424 /// ldr x1, [x0, #:tlsdesc_lo12:var]
3425 /// add x0, x0, #:tlsdesc_lo12:var
3426 /// .tlsdesccall var
3428 /// (TPIDR_EL0 offset now in x0)
3430 /// The above sequence must be produced unscheduled, to enable the linker to
3431 /// optimize/relax this sequence.
3432 /// Therefore, a pseudo-instruction (TLSDESC_CALLSEQ) is used to represent the
3433 /// above sequence, and expanded really late in the compilation flow, to ensure
3434 /// the sequence is produced as per above.
3435 SDValue AArch64TargetLowering::LowerELFTLSDescCallSeq(SDValue SymAddr, SDLoc DL,
3436 SelectionDAG &DAG) const {
3437 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3439 SDValue Chain = DAG.getEntryNode();
3440 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3442 SmallVector<SDValue, 2> Ops;
3443 Ops.push_back(Chain);
3444 Ops.push_back(SymAddr);
3446 Chain = DAG.getNode(AArch64ISD::TLSDESC_CALLSEQ, DL, NodeTys, Ops);
3447 SDValue Glue = Chain.getValue(1);
3449 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Glue);
3453 AArch64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
3454 SelectionDAG &DAG) const {
3455 assert(Subtarget->isTargetELF() && "This function expects an ELF target");
3456 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
3457 "ELF TLS only supported in small memory model");
3458 // Different choices can be made for the maximum size of the TLS area for a
3459 // module. For the small address model, the default TLS size is 16MiB and the
3460 // maximum TLS size is 4GiB.
3461 // FIXME: add -mtls-size command line option and make it control the 16MiB
3462 // vs. 4GiB code sequence generation.
3463 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3465 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
3467 if (DAG.getTarget().Options.EmulatedTLS)
3468 return LowerToTLSEmulatedModel(GA, DAG);
3470 if (!EnableAArch64ELFLocalDynamicTLSGeneration) {
3471 if (Model == TLSModel::LocalDynamic)
3472 Model = TLSModel::GeneralDynamic;
3476 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3478 const GlobalValue *GV = GA->getGlobal();
3480 SDValue ThreadBase = DAG.getNode(AArch64ISD::THREAD_POINTER, DL, PtrVT);
3482 if (Model == TLSModel::LocalExec) {
3483 SDValue HiVar = DAG.getTargetGlobalAddress(
3484 GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_HI12);
3485 SDValue LoVar = DAG.getTargetGlobalAddress(
3487 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3489 SDValue TPWithOff_lo =
3490 SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, ThreadBase,
3492 DAG.getTargetConstant(0, DL, MVT::i32)),
3495 SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPWithOff_lo,
3497 DAG.getTargetConstant(0, DL, MVT::i32)),
3500 } else if (Model == TLSModel::InitialExec) {
3501 TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3502 TPOff = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TPOff);
3503 } else if (Model == TLSModel::LocalDynamic) {
3504 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
3505 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
3506 // the beginning of the module's TLS region, followed by a DTPREL offset
3509 // These accesses will need deduplicating if there's more than one.
3510 AArch64FunctionInfo *MFI =
3511 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
3512 MFI->incNumLocalDynamicTLSAccesses();
3514 // The call needs a relocation too for linker relaxation. It doesn't make
3515 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
3517 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
3520 // Now we can calculate the offset from TPIDR_EL0 to this module's
3521 // thread-local area.
3522 TPOff = LowerELFTLSDescCallSeq(SymAddr, DL, DAG);
3524 // Now use :dtprel_whatever: operations to calculate this variable's offset
3525 // in its thread-storage area.
3526 SDValue HiVar = DAG.getTargetGlobalAddress(
3527 GV, DL, MVT::i64, 0, AArch64II::MO_TLS | AArch64II::MO_HI12);
3528 SDValue LoVar = DAG.getTargetGlobalAddress(
3529 GV, DL, MVT::i64, 0,
3530 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3532 TPOff = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPOff, HiVar,
3533 DAG.getTargetConstant(0, DL, MVT::i32)),
3535 TPOff = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPOff, LoVar,
3536 DAG.getTargetConstant(0, DL, MVT::i32)),
3538 } else if (Model == TLSModel::GeneralDynamic) {
3539 // The call needs a relocation too for linker relaxation. It doesn't make
3540 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
3543 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3545 // Finally we can make a call to calculate the offset from tpidr_el0.
3546 TPOff = LowerELFTLSDescCallSeq(SymAddr, DL, DAG);
3548 llvm_unreachable("Unsupported ELF TLS access model");
3550 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
3553 SDValue AArch64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
3554 SelectionDAG &DAG) const {
3555 if (Subtarget->isTargetDarwin())
3556 return LowerDarwinGlobalTLSAddress(Op, DAG);
3557 else if (Subtarget->isTargetELF())
3558 return LowerELFGlobalTLSAddress(Op, DAG);
3560 llvm_unreachable("Unexpected platform trying to use TLS");
3562 SDValue AArch64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3563 SDValue Chain = Op.getOperand(0);
3564 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3565 SDValue LHS = Op.getOperand(2);
3566 SDValue RHS = Op.getOperand(3);
3567 SDValue Dest = Op.getOperand(4);
3570 // Handle f128 first, since lowering it will result in comparing the return
3571 // value of a libcall against zero, which is just what the rest of LowerBR_CC
3572 // is expecting to deal with.
3573 if (LHS.getValueType() == MVT::f128) {
3574 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3576 // If softenSetCCOperands returned a scalar, we need to compare the result
3577 // against zero to select between true and false values.
3578 if (!RHS.getNode()) {
3579 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3584 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
3586 unsigned Opc = LHS.getOpcode();
3587 if (LHS.getResNo() == 1 && isa<ConstantSDNode>(RHS) &&
3588 cast<ConstantSDNode>(RHS)->isOne() &&
3589 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3590 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3591 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
3592 "Unexpected condition code.");
3593 // Only lower legal XALUO ops.
3594 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
3597 // The actual operation with overflow check.
3598 AArch64CC::CondCode OFCC;
3599 SDValue Value, Overflow;
3600 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, LHS.getValue(0), DAG);
3602 if (CC == ISD::SETNE)
3603 OFCC = getInvertedCondCode(OFCC);
3604 SDValue CCVal = DAG.getConstant(OFCC, dl, MVT::i32);
3606 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3610 if (LHS.getValueType().isInteger()) {
3611 assert((LHS.getValueType() == RHS.getValueType()) &&
3612 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3614 // If the RHS of the comparison is zero, we can potentially fold this
3615 // to a specialized branch.
3616 const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
3617 if (RHSC && RHSC->getZExtValue() == 0) {
3618 if (CC == ISD::SETEQ) {
3619 // See if we can use a TBZ to fold in an AND as well.
3620 // TBZ has a smaller branch displacement than CBZ. If the offset is
3621 // out of bounds, a late MI-layer pass rewrites branches.
3622 // 403.gcc is an example that hits this case.
3623 if (LHS.getOpcode() == ISD::AND &&
3624 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3625 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3626 SDValue Test = LHS.getOperand(0);
3627 uint64_t Mask = LHS.getConstantOperandVal(1);
3628 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, Test,
3629 DAG.getConstant(Log2_64(Mask), dl, MVT::i64),
3633 return DAG.getNode(AArch64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest);
3634 } else if (CC == ISD::SETNE) {
3635 // See if we can use a TBZ to fold in an AND as well.
3636 // TBZ has a smaller branch displacement than CBZ. If the offset is
3637 // out of bounds, a late MI-layer pass rewrites branches.
3638 // 403.gcc is an example that hits this case.
3639 if (LHS.getOpcode() == ISD::AND &&
3640 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3641 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3642 SDValue Test = LHS.getOperand(0);
3643 uint64_t Mask = LHS.getConstantOperandVal(1);
3644 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, Test,
3645 DAG.getConstant(Log2_64(Mask), dl, MVT::i64),
3649 return DAG.getNode(AArch64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest);
3650 } else if (CC == ISD::SETLT && LHS.getOpcode() != ISD::AND) {
3651 // Don't combine AND since emitComparison converts the AND to an ANDS
3652 // (a.k.a. TST) and the test in the test bit and branch instruction
3653 // becomes redundant. This would also increase register pressure.
3654 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
3655 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, LHS,
3656 DAG.getConstant(Mask, dl, MVT::i64), Dest);
3659 if (RHSC && RHSC->getSExtValue() == -1 && CC == ISD::SETGT &&
3660 LHS.getOpcode() != ISD::AND) {
3661 // Don't combine AND since emitComparison converts the AND to an ANDS
3662 // (a.k.a. TST) and the test in the test bit and branch instruction
3663 // becomes redundant. This would also increase register pressure.
3664 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
3665 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, LHS,
3666 DAG.getConstant(Mask, dl, MVT::i64), Dest);
3670 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3671 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3675 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3677 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
3678 // clean. Some of them require two branches to implement.
3679 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3680 AArch64CC::CondCode CC1, CC2;
3681 changeFPCCToAArch64CC(CC, CC1, CC2);
3682 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
3684 DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp);
3685 if (CC2 != AArch64CC::AL) {
3686 SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32);
3687 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val,
3694 SDValue AArch64TargetLowering::LowerFCOPYSIGN(SDValue Op,
3695 SelectionDAG &DAG) const {
3696 EVT VT = Op.getValueType();
3699 SDValue In1 = Op.getOperand(0);
3700 SDValue In2 = Op.getOperand(1);
3701 EVT SrcVT = In2.getValueType();
3703 if (SrcVT.bitsLT(VT))
3704 In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2);
3705 else if (SrcVT.bitsGT(VT))
3706 In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, DAG.getIntPtrConstant(0, DL));
3711 SDValue VecVal1, VecVal2;
3712 if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) {
3714 VecVT = (VT == MVT::v2f32 ? MVT::v2i32 : MVT::v4i32);
3715 EltMask = 0x80000000ULL;
3717 if (!VT.isVector()) {
3718 VecVal1 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3719 DAG.getUNDEF(VecVT), In1);
3720 VecVal2 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3721 DAG.getUNDEF(VecVT), In2);
3723 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3724 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3726 } else if (VT == MVT::f64 || VT == MVT::v2f64) {
3730 // We want to materialize a mask with the high bit set, but the AdvSIMD
3731 // immediate moves cannot materialize that in a single instruction for
3732 // 64-bit elements. Instead, materialize zero and then negate it.
3735 if (!VT.isVector()) {
3736 VecVal1 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3737 DAG.getUNDEF(VecVT), In1);
3738 VecVal2 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3739 DAG.getUNDEF(VecVT), In2);
3741 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3742 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3745 llvm_unreachable("Invalid type for copysign!");
3748 SDValue BuildVec = DAG.getConstant(EltMask, DL, VecVT);
3750 // If we couldn't materialize the mask above, then the mask vector will be
3751 // the zero vector, and we need to negate it here.
3752 if (VT == MVT::f64 || VT == MVT::v2f64) {
3753 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec);
3754 BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec);
3755 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec);
3759 DAG.getNode(AArch64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec);
3762 return DAG.getTargetExtractSubreg(AArch64::ssub, DL, VT, Sel);
3763 else if (VT == MVT::f64)
3764 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, VT, Sel);
3766 return DAG.getNode(ISD::BITCAST, DL, VT, Sel);
3769 SDValue AArch64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
3770 if (DAG.getMachineFunction().getFunction()->hasFnAttribute(
3771 Attribute::NoImplicitFloat))
3774 if (!Subtarget->hasNEON())
3777 // While there is no integer popcount instruction, it can
3778 // be more efficiently lowered to the following sequence that uses
3779 // AdvSIMD registers/instructions as long as the copies to/from
3780 // the AdvSIMD registers are cheap.
3781 // FMOV D0, X0 // copy 64-bit int to vector, high bits zero'd
3782 // CNT V0.8B, V0.8B // 8xbyte pop-counts
3783 // ADDV B0, V0.8B // sum 8xbyte pop-counts
3784 // UMOV X0, V0.B[0] // copy byte result back to integer reg
3785 SDValue Val = Op.getOperand(0);
3787 EVT VT = Op.getValueType();
3790 Val = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Val);
3791 Val = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);
3793 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, Val);
3794 SDValue UaddLV = DAG.getNode(
3795 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
3796 DAG.getConstant(Intrinsic::aarch64_neon_uaddlv, DL, MVT::i32), CtPop);
3799 UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV);
3803 SDValue AArch64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
3805 if (Op.getValueType().isVector())
3806 return LowerVSETCC(Op, DAG);
3808 SDValue LHS = Op.getOperand(0);
3809 SDValue RHS = Op.getOperand(1);
3810 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
3813 // We chose ZeroOrOneBooleanContents, so use zero and one.
3814 EVT VT = Op.getValueType();
3815 SDValue TVal = DAG.getConstant(1, dl, VT);
3816 SDValue FVal = DAG.getConstant(0, dl, VT);
3818 // Handle f128 first, since one possible outcome is a normal integer
3819 // comparison which gets picked up by the next if statement.
3820 if (LHS.getValueType() == MVT::f128) {
3821 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3823 // If softenSetCCOperands returned a scalar, use it.
3824 if (!RHS.getNode()) {
3825 assert(LHS.getValueType() == Op.getValueType() &&
3826 "Unexpected setcc expansion!");
3831 if (LHS.getValueType().isInteger()) {
3834 getAArch64Cmp(LHS, RHS, ISD::getSetCCInverse(CC, true), CCVal, DAG, dl);
3836 // Note that we inverted the condition above, so we reverse the order of
3837 // the true and false operands here. This will allow the setcc to be
3838 // matched to a single CSINC instruction.
3839 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp);
3842 // Now we know we're dealing with FP values.
3843 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3845 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3846 // and do the comparison.
3847 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3849 AArch64CC::CondCode CC1, CC2;
3850 changeFPCCToAArch64CC(CC, CC1, CC2);
3851 if (CC2 == AArch64CC::AL) {
3852 changeFPCCToAArch64CC(ISD::getSetCCInverse(CC, false), CC1, CC2);
3853 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
3855 // Note that we inverted the condition above, so we reverse the order of
3856 // the true and false operands here. This will allow the setcc to be
3857 // matched to a single CSINC instruction.
3858 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp);
3860 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't
3861 // totally clean. Some of them require two CSELs to implement. As is in
3862 // this case, we emit the first CSEL and then emit a second using the output
3863 // of the first as the RHS. We're effectively OR'ing the two CC's together.
3865 // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
3866 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
3868 DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3870 SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32);
3871 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3875 SDValue AArch64TargetLowering::LowerSELECT_CC(ISD::CondCode CC, SDValue LHS,
3876 SDValue RHS, SDValue TVal,
3877 SDValue FVal, SDLoc dl,
3878 SelectionDAG &DAG) const {
3879 // Handle f128 first, because it will result in a comparison of some RTLIB
3880 // call result against zero.
3881 if (LHS.getValueType() == MVT::f128) {
3882 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3884 // If softenSetCCOperands returned a scalar, we need to compare the result
3885 // against zero to select between true and false values.
3886 if (!RHS.getNode()) {
3887 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3892 // Handle integers first.
3893 if (LHS.getValueType().isInteger()) {
3894 assert((LHS.getValueType() == RHS.getValueType()) &&
3895 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3897 unsigned Opcode = AArch64ISD::CSEL;
3899 // If both the TVal and the FVal are constants, see if we can swap them in
3900 // order to for a CSINV or CSINC out of them.
3901 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
3902 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
3904 if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) {
3905 std::swap(TVal, FVal);
3906 std::swap(CTVal, CFVal);
3907 CC = ISD::getSetCCInverse(CC, true);
3908 } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) {
3909 std::swap(TVal, FVal);
3910 std::swap(CTVal, CFVal);
3911 CC = ISD::getSetCCInverse(CC, true);
3912 } else if (TVal.getOpcode() == ISD::XOR) {
3913 // If TVal is a NOT we want to swap TVal and FVal so that we can match
3914 // with a CSINV rather than a CSEL.
3915 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(1));
3917 if (CVal && CVal->isAllOnesValue()) {
3918 std::swap(TVal, FVal);
3919 std::swap(CTVal, CFVal);
3920 CC = ISD::getSetCCInverse(CC, true);
3922 } else if (TVal.getOpcode() == ISD::SUB) {
3923 // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
3924 // that we can match with a CSNEG rather than a CSEL.
3925 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(0));
3927 if (CVal && CVal->isNullValue()) {
3928 std::swap(TVal, FVal);
3929 std::swap(CTVal, CFVal);
3930 CC = ISD::getSetCCInverse(CC, true);
3932 } else if (CTVal && CFVal) {
3933 const int64_t TrueVal = CTVal->getSExtValue();
3934 const int64_t FalseVal = CFVal->getSExtValue();
3937 // If both TVal and FVal are constants, see if FVal is the
3938 // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
3939 // instead of a CSEL in that case.
3940 if (TrueVal == ~FalseVal) {
3941 Opcode = AArch64ISD::CSINV;
3942 } else if (TrueVal == -FalseVal) {
3943 Opcode = AArch64ISD::CSNEG;
3944 } else if (TVal.getValueType() == MVT::i32) {
3945 // If our operands are only 32-bit wide, make sure we use 32-bit
3946 // arithmetic for the check whether we can use CSINC. This ensures that
3947 // the addition in the check will wrap around properly in case there is
3948 // an overflow (which would not be the case if we do the check with
3949 // 64-bit arithmetic).
3950 const uint32_t TrueVal32 = CTVal->getZExtValue();
3951 const uint32_t FalseVal32 = CFVal->getZExtValue();
3953 if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
3954 Opcode = AArch64ISD::CSINC;
3956 if (TrueVal32 > FalseVal32) {
3960 // 64-bit check whether we can use CSINC.
3961 } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) {
3962 Opcode = AArch64ISD::CSINC;
3964 if (TrueVal > FalseVal) {
3969 // Swap TVal and FVal if necessary.
3971 std::swap(TVal, FVal);
3972 std::swap(CTVal, CFVal);
3973 CC = ISD::getSetCCInverse(CC, true);
3976 if (Opcode != AArch64ISD::CSEL) {
3977 // Drop FVal since we can get its value by simply inverting/negating
3984 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3986 EVT VT = TVal.getValueType();
3987 return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp);
3990 // Now we know we're dealing with FP values.
3991 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3992 assert(LHS.getValueType() == RHS.getValueType());
3993 EVT VT = TVal.getValueType();
3994 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3996 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
3997 // clean. Some of them require two CSELs to implement.
3998 AArch64CC::CondCode CC1, CC2;
3999 changeFPCCToAArch64CC(CC, CC1, CC2);
4000 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
4001 SDValue CS1 = DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
4003 // If we need a second CSEL, emit it, using the output of the first as the
4004 // RHS. We're effectively OR'ing the two CC's together.
4005 if (CC2 != AArch64CC::AL) {
4006 SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32);
4007 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
4010 // Otherwise, return the output of the first CSEL.
4014 SDValue AArch64TargetLowering::LowerSELECT_CC(SDValue Op,
4015 SelectionDAG &DAG) const {
4016 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
4017 SDValue LHS = Op.getOperand(0);
4018 SDValue RHS = Op.getOperand(1);
4019 SDValue TVal = Op.getOperand(2);
4020 SDValue FVal = Op.getOperand(3);
4022 return LowerSELECT_CC(CC, LHS, RHS, TVal, FVal, DL, DAG);
4025 SDValue AArch64TargetLowering::LowerSELECT(SDValue Op,
4026 SelectionDAG &DAG) const {
4027 SDValue CCVal = Op->getOperand(0);
4028 SDValue TVal = Op->getOperand(1);
4029 SDValue FVal = Op->getOperand(2);
4032 unsigned Opc = CCVal.getOpcode();
4033 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
4035 if (CCVal.getResNo() == 1 &&
4036 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
4037 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
4038 // Only lower legal XALUO ops.
4039 if (!DAG.getTargetLoweringInfo().isTypeLegal(CCVal->getValueType(0)))
4042 AArch64CC::CondCode OFCC;
4043 SDValue Value, Overflow;
4044 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, CCVal.getValue(0), DAG);
4045 SDValue CCVal = DAG.getConstant(OFCC, DL, MVT::i32);
4047 return DAG.getNode(AArch64ISD::CSEL, DL, Op.getValueType(), TVal, FVal,
4051 // Lower it the same way as we would lower a SELECT_CC node.
4054 if (CCVal.getOpcode() == ISD::SETCC) {
4055 LHS = CCVal.getOperand(0);
4056 RHS = CCVal.getOperand(1);
4057 CC = cast<CondCodeSDNode>(CCVal->getOperand(2))->get();
4060 RHS = DAG.getConstant(0, DL, CCVal.getValueType());
4063 return LowerSELECT_CC(CC, LHS, RHS, TVal, FVal, DL, DAG);
4066 SDValue AArch64TargetLowering::LowerJumpTable(SDValue Op,
4067 SelectionDAG &DAG) const {
4068 // Jump table entries as PC relative offsets. No additional tweaking
4069 // is necessary here. Just get the address of the jump table.
4070 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4071 EVT PtrVT = getPointerTy(DAG.getDataLayout());
4074 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
4075 !Subtarget->isTargetMachO()) {
4076 const unsigned char MO_NC = AArch64II::MO_NC;
4078 AArch64ISD::WrapperLarge, DL, PtrVT,
4079 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G3),
4080 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G2 | MO_NC),
4081 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G1 | MO_NC),
4082 DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
4083 AArch64II::MO_G0 | MO_NC));
4087 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_PAGE);
4088 SDValue Lo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
4089 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
4090 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
4091 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
4094 SDValue AArch64TargetLowering::LowerConstantPool(SDValue Op,
4095 SelectionDAG &DAG) const {
4096 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4097 EVT PtrVT = getPointerTy(DAG.getDataLayout());
4100 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
4101 // Use the GOT for the large code model on iOS.
4102 if (Subtarget->isTargetMachO()) {
4103 SDValue GotAddr = DAG.getTargetConstantPool(
4104 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
4106 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
4109 const unsigned char MO_NC = AArch64II::MO_NC;
4111 AArch64ISD::WrapperLarge, DL, PtrVT,
4112 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4113 CP->getOffset(), AArch64II::MO_G3),
4114 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4115 CP->getOffset(), AArch64II::MO_G2 | MO_NC),
4116 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4117 CP->getOffset(), AArch64II::MO_G1 | MO_NC),
4118 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4119 CP->getOffset(), AArch64II::MO_G0 | MO_NC));
4121 // Use ADRP/ADD or ADRP/LDR for everything else: the small memory model on
4122 // ELF, the only valid one on Darwin.
4124 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4125 CP->getOffset(), AArch64II::MO_PAGE);
4126 SDValue Lo = DAG.getTargetConstantPool(
4127 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
4128 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
4130 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
4131 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
4135 SDValue AArch64TargetLowering::LowerBlockAddress(SDValue Op,
4136 SelectionDAG &DAG) const {
4137 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
4138 EVT PtrVT = getPointerTy(DAG.getDataLayout());
4140 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
4141 !Subtarget->isTargetMachO()) {
4142 const unsigned char MO_NC = AArch64II::MO_NC;
4144 AArch64ISD::WrapperLarge, DL, PtrVT,
4145 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G3),
4146 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
4147 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
4148 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
4150 SDValue Hi = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGE);
4151 SDValue Lo = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGEOFF |
4153 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
4154 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
4158 SDValue AArch64TargetLowering::LowerDarwin_VASTART(SDValue Op,
4159 SelectionDAG &DAG) const {
4160 AArch64FunctionInfo *FuncInfo =
4161 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
4164 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(),
4165 getPointerTy(DAG.getDataLayout()));
4166 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
4167 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
4168 MachinePointerInfo(SV), false, false, 0);
4171 SDValue AArch64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
4172 SelectionDAG &DAG) const {
4173 // The layout of the va_list struct is specified in the AArch64 Procedure Call
4174 // Standard, section B.3.
4175 MachineFunction &MF = DAG.getMachineFunction();
4176 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
4177 auto PtrVT = getPointerTy(DAG.getDataLayout());
4180 SDValue Chain = Op.getOperand(0);
4181 SDValue VAList = Op.getOperand(1);
4182 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
4183 SmallVector<SDValue, 4> MemOps;
4185 // void *__stack at offset 0
4186 SDValue Stack = DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), PtrVT);
4187 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
4188 MachinePointerInfo(SV), false, false, 8));
4190 // void *__gr_top at offset 8
4191 int GPRSize = FuncInfo->getVarArgsGPRSize();
4193 SDValue GRTop, GRTopAddr;
4196 DAG.getNode(ISD::ADD, DL, PtrVT, VAList, DAG.getConstant(8, DL, PtrVT));
4198 GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), PtrVT);
4199 GRTop = DAG.getNode(ISD::ADD, DL, PtrVT, GRTop,
4200 DAG.getConstant(GPRSize, DL, PtrVT));
4202 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
4203 MachinePointerInfo(SV, 8), false, false, 8));
4206 // void *__vr_top at offset 16
4207 int FPRSize = FuncInfo->getVarArgsFPRSize();
4209 SDValue VRTop, VRTopAddr;
4210 VRTopAddr = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
4211 DAG.getConstant(16, DL, PtrVT));
4213 VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), PtrVT);
4214 VRTop = DAG.getNode(ISD::ADD, DL, PtrVT, VRTop,
4215 DAG.getConstant(FPRSize, DL, PtrVT));
4217 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
4218 MachinePointerInfo(SV, 16), false, false, 8));
4221 // int __gr_offs at offset 24
4222 SDValue GROffsAddr =
4223 DAG.getNode(ISD::ADD, DL, PtrVT, VAList, DAG.getConstant(24, DL, PtrVT));
4224 MemOps.push_back(DAG.getStore(Chain, DL,
4225 DAG.getConstant(-GPRSize, DL, MVT::i32),
4226 GROffsAddr, MachinePointerInfo(SV, 24), false,
4229 // int __vr_offs at offset 28
4230 SDValue VROffsAddr =
4231 DAG.getNode(ISD::ADD, DL, PtrVT, VAList, DAG.getConstant(28, DL, PtrVT));
4232 MemOps.push_back(DAG.getStore(Chain, DL,
4233 DAG.getConstant(-FPRSize, DL, MVT::i32),
4234 VROffsAddr, MachinePointerInfo(SV, 28), false,
4237 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
4240 SDValue AArch64TargetLowering::LowerVASTART(SDValue Op,
4241 SelectionDAG &DAG) const {
4242 return Subtarget->isTargetDarwin() ? LowerDarwin_VASTART(Op, DAG)
4243 : LowerAAPCS_VASTART(Op, DAG);
4246 SDValue AArch64TargetLowering::LowerVACOPY(SDValue Op,
4247 SelectionDAG &DAG) const {
4248 // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
4251 unsigned VaListSize = Subtarget->isTargetDarwin() ? 8 : 32;
4252 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
4253 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
4255 return DAG.getMemcpy(Op.getOperand(0), DL, Op.getOperand(1),
4257 DAG.getConstant(VaListSize, DL, MVT::i32),
4258 8, false, false, false, MachinePointerInfo(DestSV),
4259 MachinePointerInfo(SrcSV));
4262 SDValue AArch64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
4263 assert(Subtarget->isTargetDarwin() &&
4264 "automatic va_arg instruction only works on Darwin");
4266 const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
4267 EVT VT = Op.getValueType();
4269 SDValue Chain = Op.getOperand(0);
4270 SDValue Addr = Op.getOperand(1);
4271 unsigned Align = Op.getConstantOperandVal(3);
4272 auto PtrVT = getPointerTy(DAG.getDataLayout());
4274 SDValue VAList = DAG.getLoad(PtrVT, DL, Chain, Addr, MachinePointerInfo(V),
4275 false, false, false, 0);
4276 Chain = VAList.getValue(1);
4279 assert(((Align & (Align - 1)) == 0) && "Expected Align to be a power of 2");
4280 VAList = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
4281 DAG.getConstant(Align - 1, DL, PtrVT));
4282 VAList = DAG.getNode(ISD::AND, DL, PtrVT, VAList,
4283 DAG.getConstant(-(int64_t)Align, DL, PtrVT));
4286 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
4287 uint64_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
4289 // Scalar integer and FP values smaller than 64 bits are implicitly extended
4290 // up to 64 bits. At the very least, we have to increase the striding of the
4291 // vaargs list to match this, and for FP values we need to introduce
4292 // FP_ROUND nodes as well.
4293 if (VT.isInteger() && !VT.isVector())
4295 bool NeedFPTrunc = false;
4296 if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
4301 // Increment the pointer, VAList, to the next vaarg
4302 SDValue VANext = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
4303 DAG.getConstant(ArgSize, DL, PtrVT));
4304 // Store the incremented VAList to the legalized pointer
4305 SDValue APStore = DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V),
4308 // Load the actual argument out of the pointer VAList
4310 // Load the value as an f64.
4311 SDValue WideFP = DAG.getLoad(MVT::f64, DL, APStore, VAList,
4312 MachinePointerInfo(), false, false, false, 0);
4313 // Round the value down to an f32.
4314 SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
4315 DAG.getIntPtrConstant(1, DL));
4316 SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
4317 // Merge the rounded value with the chain output of the load.
4318 return DAG.getMergeValues(Ops, DL);
4321 return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo(), false,
4325 SDValue AArch64TargetLowering::LowerFRAMEADDR(SDValue Op,
4326 SelectionDAG &DAG) const {
4327 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4328 MFI->setFrameAddressIsTaken(true);
4330 EVT VT = Op.getValueType();
4332 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4334 DAG.getCopyFromReg(DAG.getEntryNode(), DL, AArch64::FP, VT);
4336 FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
4337 MachinePointerInfo(), false, false, false, 0);
4341 // FIXME? Maybe this could be a TableGen attribute on some registers and
4342 // this table could be generated automatically from RegInfo.
4343 unsigned AArch64TargetLowering::getRegisterByName(const char* RegName, EVT VT,
4344 SelectionDAG &DAG) const {
4345 unsigned Reg = StringSwitch<unsigned>(RegName)
4346 .Case("sp", AArch64::SP)
4350 report_fatal_error(Twine("Invalid register name \""
4351 + StringRef(RegName) + "\"."));
4354 SDValue AArch64TargetLowering::LowerRETURNADDR(SDValue Op,
4355 SelectionDAG &DAG) const {
4356 MachineFunction &MF = DAG.getMachineFunction();
4357 MachineFrameInfo *MFI = MF.getFrameInfo();
4358 MFI->setReturnAddressIsTaken(true);
4360 EVT VT = Op.getValueType();
4362 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4364 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4365 SDValue Offset = DAG.getConstant(8, DL, getPointerTy(DAG.getDataLayout()));
4366 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
4367 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
4368 MachinePointerInfo(), false, false, false, 0);
4371 // Return LR, which contains the return address. Mark it an implicit live-in.
4372 unsigned Reg = MF.addLiveIn(AArch64::LR, &AArch64::GPR64RegClass);
4373 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
4376 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4377 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4378 SDValue AArch64TargetLowering::LowerShiftRightParts(SDValue Op,
4379 SelectionDAG &DAG) const {
4380 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4381 EVT VT = Op.getValueType();
4382 unsigned VTBits = VT.getSizeInBits();
4384 SDValue ShOpLo = Op.getOperand(0);
4385 SDValue ShOpHi = Op.getOperand(1);
4386 SDValue ShAmt = Op.getOperand(2);
4388 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4390 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4392 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4393 DAG.getConstant(VTBits, dl, MVT::i64), ShAmt);
4394 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4395 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4396 DAG.getConstant(VTBits, dl, MVT::i64));
4397 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4399 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, dl, MVT::i64),
4400 ISD::SETGE, dl, DAG);
4401 SDValue CCVal = DAG.getConstant(AArch64CC::GE, dl, MVT::i32);
4403 SDValue FalseValLo = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4404 SDValue TrueValLo = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4406 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
4408 // AArch64 shifts larger than the register width are wrapped rather than
4409 // clamped, so we can't just emit "hi >> x".
4410 SDValue FalseValHi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4411 SDValue TrueValHi = Opc == ISD::SRA
4412 ? DAG.getNode(Opc, dl, VT, ShOpHi,
4413 DAG.getConstant(VTBits - 1, dl,
4415 : DAG.getConstant(0, dl, VT);
4417 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValHi, FalseValHi, CCVal, Cmp);
4419 SDValue Ops[2] = { Lo, Hi };
4420 return DAG.getMergeValues(Ops, dl);
4423 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4424 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4425 SDValue AArch64TargetLowering::LowerShiftLeftParts(SDValue Op,
4426 SelectionDAG &DAG) const {
4427 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4428 EVT VT = Op.getValueType();
4429 unsigned VTBits = VT.getSizeInBits();
4431 SDValue ShOpLo = Op.getOperand(0);
4432 SDValue ShOpHi = Op.getOperand(1);
4433 SDValue ShAmt = Op.getOperand(2);
4436 assert(Op.getOpcode() == ISD::SHL_PARTS);
4437 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4438 DAG.getConstant(VTBits, dl, MVT::i64), ShAmt);
4439 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4440 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4441 DAG.getConstant(VTBits, dl, MVT::i64));
4442 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4443 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4445 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4447 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, dl, MVT::i64),
4448 ISD::SETGE, dl, DAG);
4449 SDValue CCVal = DAG.getConstant(AArch64CC::GE, dl, MVT::i32);
4451 DAG.getNode(AArch64ISD::CSEL, dl, VT, Tmp3, FalseVal, CCVal, Cmp);
4453 // AArch64 shifts of larger than register sizes are wrapped rather than
4454 // clamped, so we can't just emit "lo << a" if a is too big.
4455 SDValue TrueValLo = DAG.getConstant(0, dl, VT);
4456 SDValue FalseValLo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4458 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
4460 SDValue Ops[2] = { Lo, Hi };
4461 return DAG.getMergeValues(Ops, dl);
4464 bool AArch64TargetLowering::isOffsetFoldingLegal(
4465 const GlobalAddressSDNode *GA) const {
4466 // The AArch64 target doesn't support folding offsets into global addresses.
4470 bool AArch64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4471 // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
4472 // FIXME: We should be able to handle f128 as well with a clever lowering.
4473 if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32))
4477 return AArch64_AM::getFP64Imm(Imm) != -1;
4478 else if (VT == MVT::f32)
4479 return AArch64_AM::getFP32Imm(Imm) != -1;
4483 //===----------------------------------------------------------------------===//
4484 // AArch64 Optimization Hooks
4485 //===----------------------------------------------------------------------===//
4487 //===----------------------------------------------------------------------===//
4488 // AArch64 Inline Assembly Support
4489 //===----------------------------------------------------------------------===//
4491 // Table of Constraints
4492 // TODO: This is the current set of constraints supported by ARM for the
4493 // compiler, not all of them may make sense, e.g. S may be difficult to support.
4495 // r - A general register
4496 // w - An FP/SIMD register of some size in the range v0-v31
4497 // x - An FP/SIMD register of some size in the range v0-v15
4498 // I - Constant that can be used with an ADD instruction
4499 // J - Constant that can be used with a SUB instruction
4500 // K - Constant that can be used with a 32-bit logical instruction
4501 // L - Constant that can be used with a 64-bit logical instruction
4502 // M - Constant that can be used as a 32-bit MOV immediate
4503 // N - Constant that can be used as a 64-bit MOV immediate
4504 // Q - A memory reference with base register and no offset
4505 // S - A symbolic address
4506 // Y - Floating point constant zero
4507 // Z - Integer constant zero
4509 // Note that general register operands will be output using their 64-bit x
4510 // register name, whatever the size of the variable, unless the asm operand
4511 // is prefixed by the %w modifier. Floating-point and SIMD register operands
4512 // will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
4515 /// getConstraintType - Given a constraint letter, return the type of
4516 /// constraint it is for this target.
4517 AArch64TargetLowering::ConstraintType
4518 AArch64TargetLowering::getConstraintType(StringRef Constraint) const {
4519 if (Constraint.size() == 1) {
4520 switch (Constraint[0]) {
4527 return C_RegisterClass;
4528 // An address with a single base register. Due to the way we
4529 // currently handle addresses it is the same as 'r'.
4534 return TargetLowering::getConstraintType(Constraint);
4537 /// Examine constraint type and operand type and determine a weight value.
4538 /// This object must already have been set up with the operand type
4539 /// and the current alternative constraint selected.
4540 TargetLowering::ConstraintWeight
4541 AArch64TargetLowering::getSingleConstraintMatchWeight(
4542 AsmOperandInfo &info, const char *constraint) const {
4543 ConstraintWeight weight = CW_Invalid;
4544 Value *CallOperandVal = info.CallOperandVal;
4545 // If we don't have a value, we can't do a match,
4546 // but allow it at the lowest weight.
4547 if (!CallOperandVal)
4549 Type *type = CallOperandVal->getType();
4550 // Look at the constraint type.
4551 switch (*constraint) {
4553 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
4557 if (type->isFloatingPointTy() || type->isVectorTy())
4558 weight = CW_Register;
4561 weight = CW_Constant;
4567 std::pair<unsigned, const TargetRegisterClass *>
4568 AArch64TargetLowering::getRegForInlineAsmConstraint(
4569 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
4570 if (Constraint.size() == 1) {
4571 switch (Constraint[0]) {
4573 if (VT.getSizeInBits() == 64)
4574 return std::make_pair(0U, &AArch64::GPR64commonRegClass);
4575 return std::make_pair(0U, &AArch64::GPR32commonRegClass);
4578 return std::make_pair(0U, &AArch64::FPR32RegClass);
4579 if (VT.getSizeInBits() == 64)
4580 return std::make_pair(0U, &AArch64::FPR64RegClass);
4581 if (VT.getSizeInBits() == 128)
4582 return std::make_pair(0U, &AArch64::FPR128RegClass);
4584 // The instructions that this constraint is designed for can
4585 // only take 128-bit registers so just use that regclass.
4587 if (VT.getSizeInBits() == 128)
4588 return std::make_pair(0U, &AArch64::FPR128_loRegClass);
4592 if (StringRef("{cc}").equals_lower(Constraint))
4593 return std::make_pair(unsigned(AArch64::NZCV), &AArch64::CCRRegClass);
4595 // Use the default implementation in TargetLowering to convert the register
4596 // constraint into a member of a register class.
4597 std::pair<unsigned, const TargetRegisterClass *> Res;
4598 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
4600 // Not found as a standard register?
4602 unsigned Size = Constraint.size();
4603 if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
4604 tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
4606 bool Failed = Constraint.slice(2, Size - 1).getAsInteger(10, RegNo);
4607 if (!Failed && RegNo >= 0 && RegNo <= 31) {
4608 // v0 - v31 are aliases of q0 - q31.
4609 // By default we'll emit v0-v31 for this unless there's a modifier where
4610 // we'll emit the correct register as well.
4611 Res.first = AArch64::FPR128RegClass.getRegister(RegNo);
4612 Res.second = &AArch64::FPR128RegClass;
4620 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4621 /// vector. If it is invalid, don't add anything to Ops.
4622 void AArch64TargetLowering::LowerAsmOperandForConstraint(
4623 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
4624 SelectionDAG &DAG) const {
4627 // Currently only support length 1 constraints.
4628 if (Constraint.length() != 1)
4631 char ConstraintLetter = Constraint[0];
4632 switch (ConstraintLetter) {
4636 // This set of constraints deal with valid constants for various instructions.
4637 // Validate and return a target constant for them if we can.
4639 // 'z' maps to xzr or wzr so it needs an input of 0.
4640 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4641 if (!C || C->getZExtValue() != 0)
4644 if (Op.getValueType() == MVT::i64)
4645 Result = DAG.getRegister(AArch64::XZR, MVT::i64);
4647 Result = DAG.getRegister(AArch64::WZR, MVT::i32);
4657 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4661 // Grab the value and do some validation.
4662 uint64_t CVal = C->getZExtValue();
4663 switch (ConstraintLetter) {
4664 // The I constraint applies only to simple ADD or SUB immediate operands:
4665 // i.e. 0 to 4095 with optional shift by 12
4666 // The J constraint applies only to ADD or SUB immediates that would be
4667 // valid when negated, i.e. if [an add pattern] were to be output as a SUB
4668 // instruction [or vice versa], in other words -1 to -4095 with optional
4669 // left shift by 12.
4671 if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
4675 uint64_t NVal = -C->getSExtValue();
4676 if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal)) {
4677 CVal = C->getSExtValue();
4682 // The K and L constraints apply *only* to logical immediates, including
4683 // what used to be the MOVI alias for ORR (though the MOVI alias has now
4684 // been removed and MOV should be used). So these constraints have to
4685 // distinguish between bit patterns that are valid 32-bit or 64-bit
4686 // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
4687 // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
4690 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4694 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4697 // The M and N constraints are a superset of K and L respectively, for use
4698 // with the MOV (immediate) alias. As well as the logical immediates they
4699 // also match 32 or 64-bit immediates that can be loaded either using a
4700 // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
4701 // (M) or 64-bit 0x1234000000000000 (N) etc.
4702 // As a note some of this code is liberally stolen from the asm parser.
4704 if (!isUInt<32>(CVal))
4706 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4708 if ((CVal & 0xFFFF) == CVal)
4710 if ((CVal & 0xFFFF0000ULL) == CVal)
4712 uint64_t NCVal = ~(uint32_t)CVal;
4713 if ((NCVal & 0xFFFFULL) == NCVal)
4715 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4720 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4722 if ((CVal & 0xFFFFULL) == CVal)
4724 if ((CVal & 0xFFFF0000ULL) == CVal)
4726 if ((CVal & 0xFFFF00000000ULL) == CVal)
4728 if ((CVal & 0xFFFF000000000000ULL) == CVal)
4730 uint64_t NCVal = ~CVal;
4731 if ((NCVal & 0xFFFFULL) == NCVal)
4733 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4735 if ((NCVal & 0xFFFF00000000ULL) == NCVal)
4737 if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
4745 // All assembler immediates are 64-bit integers.
4746 Result = DAG.getTargetConstant(CVal, SDLoc(Op), MVT::i64);
4750 if (Result.getNode()) {
4751 Ops.push_back(Result);
4755 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
4758 //===----------------------------------------------------------------------===//
4759 // AArch64 Advanced SIMD Support
4760 //===----------------------------------------------------------------------===//
4762 /// WidenVector - Given a value in the V64 register class, produce the
4763 /// equivalent value in the V128 register class.
4764 static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
4765 EVT VT = V64Reg.getValueType();
4766 unsigned NarrowSize = VT.getVectorNumElements();
4767 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4768 MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
4771 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
4772 V64Reg, DAG.getConstant(0, DL, MVT::i32));
4775 /// getExtFactor - Determine the adjustment factor for the position when
4776 /// generating an "extract from vector registers" instruction.
4777 static unsigned getExtFactor(SDValue &V) {
4778 EVT EltType = V.getValueType().getVectorElementType();
4779 return EltType.getSizeInBits() / 8;
4782 /// NarrowVector - Given a value in the V128 register class, produce the
4783 /// equivalent value in the V64 register class.
4784 static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) {
4785 EVT VT = V128Reg.getValueType();
4786 unsigned WideSize = VT.getVectorNumElements();
4787 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4788 MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2);
4791 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, NarrowTy, V128Reg);
4794 // Gather data to see if the operation can be modelled as a
4795 // shuffle in combination with VEXTs.
4796 SDValue AArch64TargetLowering::ReconstructShuffle(SDValue Op,
4797 SelectionDAG &DAG) const {
4798 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
4800 EVT VT = Op.getValueType();
4801 unsigned NumElts = VT.getVectorNumElements();
4803 struct ShuffleSourceInfo {
4808 // We may insert some combination of BITCASTs and VEXT nodes to force Vec to
4809 // be compatible with the shuffle we intend to construct. As a result
4810 // ShuffleVec will be some sliding window into the original Vec.
4813 // Code should guarantee that element i in Vec starts at element "WindowBase
4814 // + i * WindowScale in ShuffleVec".
4818 bool operator ==(SDValue OtherVec) { return Vec == OtherVec; }
4819 ShuffleSourceInfo(SDValue Vec)
4820 : Vec(Vec), MinElt(UINT_MAX), MaxElt(0), ShuffleVec(Vec), WindowBase(0),
4824 // First gather all vectors used as an immediate source for this BUILD_VECTOR
4826 SmallVector<ShuffleSourceInfo, 2> Sources;
4827 for (unsigned i = 0; i < NumElts; ++i) {
4828 SDValue V = Op.getOperand(i);
4829 if (V.getOpcode() == ISD::UNDEF)
4831 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4832 // A shuffle can only come from building a vector from various
4833 // elements of other vectors.
4837 // Add this element source to the list if it's not already there.
4838 SDValue SourceVec = V.getOperand(0);
4839 auto Source = std::find(Sources.begin(), Sources.end(), SourceVec);
4840 if (Source == Sources.end())
4841 Source = Sources.insert(Sources.end(), ShuffleSourceInfo(SourceVec));
4843 // Update the minimum and maximum lane number seen.
4844 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4845 Source->MinElt = std::min(Source->MinElt, EltNo);
4846 Source->MaxElt = std::max(Source->MaxElt, EltNo);
4849 // Currently only do something sane when at most two source vectors
4851 if (Sources.size() > 2)
4854 // Find out the smallest element size among result and two sources, and use
4855 // it as element size to build the shuffle_vector.
4856 EVT SmallestEltTy = VT.getVectorElementType();
4857 for (auto &Source : Sources) {
4858 EVT SrcEltTy = Source.Vec.getValueType().getVectorElementType();
4859 if (SrcEltTy.bitsLT(SmallestEltTy)) {
4860 SmallestEltTy = SrcEltTy;
4863 unsigned ResMultiplier =
4864 VT.getVectorElementType().getSizeInBits() / SmallestEltTy.getSizeInBits();
4865 NumElts = VT.getSizeInBits() / SmallestEltTy.getSizeInBits();
4866 EVT ShuffleVT = EVT::getVectorVT(*DAG.getContext(), SmallestEltTy, NumElts);
4868 // If the source vector is too wide or too narrow, we may nevertheless be able
4869 // to construct a compatible shuffle either by concatenating it with UNDEF or
4870 // extracting a suitable range of elements.
4871 for (auto &Src : Sources) {
4872 EVT SrcVT = Src.ShuffleVec.getValueType();
4874 if (SrcVT.getSizeInBits() == VT.getSizeInBits())
4877 // This stage of the search produces a source with the same element type as
4878 // the original, but with a total width matching the BUILD_VECTOR output.
4879 EVT EltVT = SrcVT.getVectorElementType();
4880 unsigned NumSrcElts = VT.getSizeInBits() / EltVT.getSizeInBits();
4881 EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumSrcElts);
4883 if (SrcVT.getSizeInBits() < VT.getSizeInBits()) {
4884 assert(2 * SrcVT.getSizeInBits() == VT.getSizeInBits());
4885 // We can pad out the smaller vector for free, so if it's part of a
4888 DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, Src.ShuffleVec,
4889 DAG.getUNDEF(Src.ShuffleVec.getValueType()));
4893 assert(SrcVT.getSizeInBits() == 2 * VT.getSizeInBits());
4895 if (Src.MaxElt - Src.MinElt >= NumSrcElts) {
4896 // Span too large for a VEXT to cope
4900 if (Src.MinElt >= NumSrcElts) {
4901 // The extraction can just take the second half
4903 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4904 DAG.getConstant(NumSrcElts, dl, MVT::i64));
4905 Src.WindowBase = -NumSrcElts;
4906 } else if (Src.MaxElt < NumSrcElts) {
4907 // The extraction can just take the first half
4909 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4910 DAG.getConstant(0, dl, MVT::i64));
4912 // An actual VEXT is needed
4914 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4915 DAG.getConstant(0, dl, MVT::i64));
4917 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4918 DAG.getConstant(NumSrcElts, dl, MVT::i64));
4919 unsigned Imm = Src.MinElt * getExtFactor(VEXTSrc1);
4921 Src.ShuffleVec = DAG.getNode(AArch64ISD::EXT, dl, DestVT, VEXTSrc1,
4923 DAG.getConstant(Imm, dl, MVT::i32));
4924 Src.WindowBase = -Src.MinElt;
4928 // Another possible incompatibility occurs from the vector element types. We
4929 // can fix this by bitcasting the source vectors to the same type we intend
4931 for (auto &Src : Sources) {
4932 EVT SrcEltTy = Src.ShuffleVec.getValueType().getVectorElementType();
4933 if (SrcEltTy == SmallestEltTy)
4935 assert(ShuffleVT.getVectorElementType() == SmallestEltTy);
4936 Src.ShuffleVec = DAG.getNode(ISD::BITCAST, dl, ShuffleVT, Src.ShuffleVec);
4937 Src.WindowScale = SrcEltTy.getSizeInBits() / SmallestEltTy.getSizeInBits();
4938 Src.WindowBase *= Src.WindowScale;
4941 // Final sanity check before we try to actually produce a shuffle.
4943 for (auto Src : Sources)
4944 assert(Src.ShuffleVec.getValueType() == ShuffleVT);
4947 // The stars all align, our next step is to produce the mask for the shuffle.
4948 SmallVector<int, 8> Mask(ShuffleVT.getVectorNumElements(), -1);
4949 int BitsPerShuffleLane = ShuffleVT.getVectorElementType().getSizeInBits();
4950 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
4951 SDValue Entry = Op.getOperand(i);
4952 if (Entry.getOpcode() == ISD::UNDEF)
4955 auto Src = std::find(Sources.begin(), Sources.end(), Entry.getOperand(0));
4956 int EltNo = cast<ConstantSDNode>(Entry.getOperand(1))->getSExtValue();
4958 // EXTRACT_VECTOR_ELT performs an implicit any_ext; BUILD_VECTOR an implicit
4959 // trunc. So only std::min(SrcBits, DestBits) actually get defined in this
4961 EVT OrigEltTy = Entry.getOperand(0).getValueType().getVectorElementType();
4962 int BitsDefined = std::min(OrigEltTy.getSizeInBits(),
4963 VT.getVectorElementType().getSizeInBits());
4964 int LanesDefined = BitsDefined / BitsPerShuffleLane;
4966 // This source is expected to fill ResMultiplier lanes of the final shuffle,
4967 // starting at the appropriate offset.
4968 int *LaneMask = &Mask[i * ResMultiplier];
4970 int ExtractBase = EltNo * Src->WindowScale + Src->WindowBase;
4971 ExtractBase += NumElts * (Src - Sources.begin());
4972 for (int j = 0; j < LanesDefined; ++j)
4973 LaneMask[j] = ExtractBase + j;
4976 // Final check before we try to produce nonsense...
4977 if (!isShuffleMaskLegal(Mask, ShuffleVT))
4980 SDValue ShuffleOps[] = { DAG.getUNDEF(ShuffleVT), DAG.getUNDEF(ShuffleVT) };
4981 for (unsigned i = 0; i < Sources.size(); ++i)
4982 ShuffleOps[i] = Sources[i].ShuffleVec;
4984 SDValue Shuffle = DAG.getVectorShuffle(ShuffleVT, dl, ShuffleOps[0],
4985 ShuffleOps[1], &Mask[0]);
4986 return DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
4989 // check if an EXT instruction can handle the shuffle mask when the
4990 // vector sources of the shuffle are the same.
4991 static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4992 unsigned NumElts = VT.getVectorNumElements();
4994 // Assume that the first shuffle index is not UNDEF. Fail if it is.
5000 // If this is a VEXT shuffle, the immediate value is the index of the first
5001 // element. The other shuffle indices must be the successive elements after
5003 unsigned ExpectedElt = Imm;
5004 for (unsigned i = 1; i < NumElts; ++i) {
5005 // Increment the expected index. If it wraps around, just follow it
5006 // back to index zero and keep going.
5008 if (ExpectedElt == NumElts)
5012 continue; // ignore UNDEF indices
5013 if (ExpectedElt != static_cast<unsigned>(M[i]))
5020 // check if an EXT instruction can handle the shuffle mask when the
5021 // vector sources of the shuffle are different.
5022 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
5024 // Look for the first non-undef element.
5025 const int *FirstRealElt = std::find_if(M.begin(), M.end(),
5026 [](int Elt) {return Elt >= 0;});
5028 // Benefit form APInt to handle overflow when calculating expected element.
5029 unsigned NumElts = VT.getVectorNumElements();
5030 unsigned MaskBits = APInt(32, NumElts * 2).logBase2();
5031 APInt ExpectedElt = APInt(MaskBits, *FirstRealElt + 1);
5032 // The following shuffle indices must be the successive elements after the
5033 // first real element.
5034 const int *FirstWrongElt = std::find_if(FirstRealElt + 1, M.end(),
5035 [&](int Elt) {return Elt != ExpectedElt++ && Elt != -1;});
5036 if (FirstWrongElt != M.end())
5039 // The index of an EXT is the first element if it is not UNDEF.
5040 // Watch out for the beginning UNDEFs. The EXT index should be the expected
5041 // value of the first element. E.g.
5042 // <-1, -1, 3, ...> is treated as <1, 2, 3, ...>.
5043 // <-1, -1, 0, 1, ...> is treated as <2*NumElts-2, 2*NumElts-1, 0, 1, ...>.
5044 // ExpectedElt is the last mask index plus 1.
5045 Imm = ExpectedElt.getZExtValue();
5047 // There are two difference cases requiring to reverse input vectors.
5048 // For example, for vector <4 x i32> we have the following cases,
5049 // Case 1: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, -1, 0>)
5050 // Case 2: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, 7, 0>)
5051 // For both cases, we finally use mask <5, 6, 7, 0>, which requires
5052 // to reverse two input vectors.
5061 /// isREVMask - Check if a vector shuffle corresponds to a REV
5062 /// instruction with the specified blocksize. (The order of the elements
5063 /// within each block of the vector is reversed.)
5064 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
5065 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
5066 "Only possible block sizes for REV are: 16, 32, 64");
5068 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5072 unsigned NumElts = VT.getVectorNumElements();
5073 unsigned BlockElts = M[0] + 1;
5074 // If the first shuffle index is UNDEF, be optimistic.
5076 BlockElts = BlockSize / EltSz;
5078 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
5081 for (unsigned i = 0; i < NumElts; ++i) {
5083 continue; // ignore UNDEF indices
5084 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
5091 static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5092 unsigned NumElts = VT.getVectorNumElements();
5093 WhichResult = (M[0] == 0 ? 0 : 1);
5094 unsigned Idx = WhichResult * NumElts / 2;
5095 for (unsigned i = 0; i != NumElts; i += 2) {
5096 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
5097 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
5105 static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5106 unsigned NumElts = VT.getVectorNumElements();
5107 WhichResult = (M[0] == 0 ? 0 : 1);
5108 for (unsigned i = 0; i != NumElts; ++i) {
5110 continue; // ignore UNDEF indices
5111 if ((unsigned)M[i] != 2 * i + WhichResult)
5118 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5119 unsigned NumElts = VT.getVectorNumElements();
5120 WhichResult = (M[0] == 0 ? 0 : 1);
5121 for (unsigned i = 0; i < NumElts; i += 2) {
5122 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
5123 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + NumElts + WhichResult))
5129 /// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of
5130 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5131 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
5132 static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5133 unsigned NumElts = VT.getVectorNumElements();
5134 WhichResult = (M[0] == 0 ? 0 : 1);
5135 unsigned Idx = WhichResult * NumElts / 2;
5136 for (unsigned i = 0; i != NumElts; i += 2) {
5137 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
5138 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx))
5146 /// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of
5147 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5148 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
5149 static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5150 unsigned Half = VT.getVectorNumElements() / 2;
5151 WhichResult = (M[0] == 0 ? 0 : 1);
5152 for (unsigned j = 0; j != 2; ++j) {
5153 unsigned Idx = WhichResult;
5154 for (unsigned i = 0; i != Half; ++i) {
5155 int MIdx = M[i + j * Half];
5156 if (MIdx >= 0 && (unsigned)MIdx != Idx)
5165 /// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of
5166 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5167 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
5168 static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5169 unsigned NumElts = VT.getVectorNumElements();
5170 WhichResult = (M[0] == 0 ? 0 : 1);
5171 for (unsigned i = 0; i < NumElts; i += 2) {
5172 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
5173 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult))
5179 static bool isINSMask(ArrayRef<int> M, int NumInputElements,
5180 bool &DstIsLeft, int &Anomaly) {
5181 if (M.size() != static_cast<size_t>(NumInputElements))
5184 int NumLHSMatch = 0, NumRHSMatch = 0;
5185 int LastLHSMismatch = -1, LastRHSMismatch = -1;
5187 for (int i = 0; i < NumInputElements; ++i) {
5197 LastLHSMismatch = i;
5199 if (M[i] == i + NumInputElements)
5202 LastRHSMismatch = i;
5205 if (NumLHSMatch == NumInputElements - 1) {
5207 Anomaly = LastLHSMismatch;
5209 } else if (NumRHSMatch == NumInputElements - 1) {
5211 Anomaly = LastRHSMismatch;
5218 static bool isConcatMask(ArrayRef<int> Mask, EVT VT, bool SplitLHS) {
5219 if (VT.getSizeInBits() != 128)
5222 unsigned NumElts = VT.getVectorNumElements();
5224 for (int I = 0, E = NumElts / 2; I != E; I++) {
5229 int Offset = NumElts / 2;
5230 for (int I = NumElts / 2, E = NumElts; I != E; I++) {
5231 if (Mask[I] != I + SplitLHS * Offset)
5238 static SDValue tryFormConcatFromShuffle(SDValue Op, SelectionDAG &DAG) {
5240 EVT VT = Op.getValueType();
5241 SDValue V0 = Op.getOperand(0);
5242 SDValue V1 = Op.getOperand(1);
5243 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op)->getMask();
5245 if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() ||
5246 VT.getVectorElementType() != V1.getValueType().getVectorElementType())
5249 bool SplitV0 = V0.getValueType().getSizeInBits() == 128;
5251 if (!isConcatMask(Mask, VT, SplitV0))
5254 EVT CastVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
5255 VT.getVectorNumElements() / 2);
5257 V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0,
5258 DAG.getConstant(0, DL, MVT::i64));
5260 if (V1.getValueType().getSizeInBits() == 128) {
5261 V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1,
5262 DAG.getConstant(0, DL, MVT::i64));
5264 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1);
5267 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5268 /// the specified operations to build the shuffle.
5269 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5270 SDValue RHS, SelectionDAG &DAG,
5272 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5273 unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1);
5274 unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1);
5277 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5286 OP_VUZPL, // VUZP, left result
5287 OP_VUZPR, // VUZP, right result
5288 OP_VZIPL, // VZIP, left result
5289 OP_VZIPR, // VZIP, right result
5290 OP_VTRNL, // VTRN, left result
5291 OP_VTRNR // VTRN, right result
5294 if (OpNum == OP_COPY) {
5295 if (LHSID == (1 * 9 + 2) * 9 + 3)
5297 assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!");
5301 SDValue OpLHS, OpRHS;
5302 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5303 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5304 EVT VT = OpLHS.getValueType();
5308 llvm_unreachable("Unknown shuffle opcode!");
5310 // VREV divides the vector in half and swaps within the half.
5311 if (VT.getVectorElementType() == MVT::i32 ||
5312 VT.getVectorElementType() == MVT::f32)
5313 return DAG.getNode(AArch64ISD::REV64, dl, VT, OpLHS);
5314 // vrev <4 x i16> -> REV32
5315 if (VT.getVectorElementType() == MVT::i16 ||
5316 VT.getVectorElementType() == MVT::f16)
5317 return DAG.getNode(AArch64ISD::REV32, dl, VT, OpLHS);
5318 // vrev <4 x i8> -> REV16
5319 assert(VT.getVectorElementType() == MVT::i8);
5320 return DAG.getNode(AArch64ISD::REV16, dl, VT, OpLHS);
5325 EVT EltTy = VT.getVectorElementType();
5327 if (EltTy == MVT::i8)
5328 Opcode = AArch64ISD::DUPLANE8;
5329 else if (EltTy == MVT::i16 || EltTy == MVT::f16)
5330 Opcode = AArch64ISD::DUPLANE16;
5331 else if (EltTy == MVT::i32 || EltTy == MVT::f32)
5332 Opcode = AArch64ISD::DUPLANE32;
5333 else if (EltTy == MVT::i64 || EltTy == MVT::f64)
5334 Opcode = AArch64ISD::DUPLANE64;
5336 llvm_unreachable("Invalid vector element type?");
5338 if (VT.getSizeInBits() == 64)
5339 OpLHS = WidenVector(OpLHS, DAG);
5340 SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, dl, MVT::i64);
5341 return DAG.getNode(Opcode, dl, VT, OpLHS, Lane);
5346 unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
5347 return DAG.getNode(AArch64ISD::EXT, dl, VT, OpLHS, OpRHS,
5348 DAG.getConstant(Imm, dl, MVT::i32));
5351 return DAG.getNode(AArch64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS,
5354 return DAG.getNode(AArch64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS,
5357 return DAG.getNode(AArch64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS,
5360 return DAG.getNode(AArch64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS,
5363 return DAG.getNode(AArch64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS,
5366 return DAG.getNode(AArch64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS,
5371 static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask,
5372 SelectionDAG &DAG) {
5373 // Check to see if we can use the TBL instruction.
5374 SDValue V1 = Op.getOperand(0);
5375 SDValue V2 = Op.getOperand(1);
5378 EVT EltVT = Op.getValueType().getVectorElementType();
5379 unsigned BytesPerElt = EltVT.getSizeInBits() / 8;
5381 SmallVector<SDValue, 8> TBLMask;
5382 for (int Val : ShuffleMask) {
5383 for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
5384 unsigned Offset = Byte + Val * BytesPerElt;
5385 TBLMask.push_back(DAG.getConstant(Offset, DL, MVT::i32));
5389 MVT IndexVT = MVT::v8i8;
5390 unsigned IndexLen = 8;
5391 if (Op.getValueType().getSizeInBits() == 128) {
5392 IndexVT = MVT::v16i8;
5396 SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
5397 SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);
5400 if (V2.getNode()->getOpcode() == ISD::UNDEF) {
5402 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
5403 Shuffle = DAG.getNode(
5404 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5405 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, DL, MVT::i32), V1Cst,
5406 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5407 makeArrayRef(TBLMask.data(), IndexLen)));
5409 if (IndexLen == 8) {
5410 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
5411 Shuffle = DAG.getNode(
5412 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5413 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, DL, MVT::i32), V1Cst,
5414 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5415 makeArrayRef(TBLMask.data(), IndexLen)));
5417 // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
5418 // cannot currently represent the register constraints on the input
5420 // Shuffle = DAG.getNode(AArch64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
5421 // DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5422 // &TBLMask[0], IndexLen));
5423 Shuffle = DAG.getNode(
5424 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5425 DAG.getConstant(Intrinsic::aarch64_neon_tbl2, DL, MVT::i32),
5427 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5428 makeArrayRef(TBLMask.data(), IndexLen)));
5431 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
5434 static unsigned getDUPLANEOp(EVT EltType) {
5435 if (EltType == MVT::i8)
5436 return AArch64ISD::DUPLANE8;
5437 if (EltType == MVT::i16 || EltType == MVT::f16)
5438 return AArch64ISD::DUPLANE16;
5439 if (EltType == MVT::i32 || EltType == MVT::f32)
5440 return AArch64ISD::DUPLANE32;
5441 if (EltType == MVT::i64 || EltType == MVT::f64)
5442 return AArch64ISD::DUPLANE64;
5444 llvm_unreachable("Invalid vector element type?");
5447 SDValue AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
5448 SelectionDAG &DAG) const {
5450 EVT VT = Op.getValueType();
5452 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5454 // Convert shuffles that are directly supported on NEON to target-specific
5455 // DAG nodes, instead of keeping them as shuffles and matching them again
5456 // during code selection. This is more efficient and avoids the possibility
5457 // of inconsistencies between legalization and selection.
5458 ArrayRef<int> ShuffleMask = SVN->getMask();
5460 SDValue V1 = Op.getOperand(0);
5461 SDValue V2 = Op.getOperand(1);
5463 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0],
5464 V1.getValueType().getSimpleVT())) {
5465 int Lane = SVN->getSplatIndex();
5466 // If this is undef splat, generate it via "just" vdup, if possible.
5470 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
5471 return DAG.getNode(AArch64ISD::DUP, dl, V1.getValueType(),
5473 // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
5474 // constant. If so, we can just reference the lane's definition directly.
5475 if (V1.getOpcode() == ISD::BUILD_VECTOR &&
5476 !isa<ConstantSDNode>(V1.getOperand(Lane)))
5477 return DAG.getNode(AArch64ISD::DUP, dl, VT, V1.getOperand(Lane));
5479 // Otherwise, duplicate from the lane of the input vector.
5480 unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
5482 // SelectionDAGBuilder may have "helpfully" already extracted or conatenated
5483 // to make a vector of the same size as this SHUFFLE. We can ignore the
5484 // extract entirely, and canonicalise the concat using WidenVector.
5485 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
5486 Lane += cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
5487 V1 = V1.getOperand(0);
5488 } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
5489 unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
5490 Lane -= Idx * VT.getVectorNumElements() / 2;
5491 V1 = WidenVector(V1.getOperand(Idx), DAG);
5492 } else if (VT.getSizeInBits() == 64)
5493 V1 = WidenVector(V1, DAG);
5495 return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, dl, MVT::i64));
5498 if (isREVMask(ShuffleMask, VT, 64))
5499 return DAG.getNode(AArch64ISD::REV64, dl, V1.getValueType(), V1, V2);
5500 if (isREVMask(ShuffleMask, VT, 32))
5501 return DAG.getNode(AArch64ISD::REV32, dl, V1.getValueType(), V1, V2);
5502 if (isREVMask(ShuffleMask, VT, 16))
5503 return DAG.getNode(AArch64ISD::REV16, dl, V1.getValueType(), V1, V2);
5505 bool ReverseEXT = false;
5507 if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
5510 Imm *= getExtFactor(V1);
5511 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V2,
5512 DAG.getConstant(Imm, dl, MVT::i32));
5513 } else if (V2->getOpcode() == ISD::UNDEF &&
5514 isSingletonEXTMask(ShuffleMask, VT, Imm)) {
5515 Imm *= getExtFactor(V1);
5516 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V1,
5517 DAG.getConstant(Imm, dl, MVT::i32));
5520 unsigned WhichResult;
5521 if (isZIPMask(ShuffleMask, VT, WhichResult)) {
5522 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
5523 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5525 if (isUZPMask(ShuffleMask, VT, WhichResult)) {
5526 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
5527 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5529 if (isTRNMask(ShuffleMask, VT, WhichResult)) {
5530 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
5531 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5534 if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5535 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
5536 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5538 if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5539 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
5540 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5542 if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5543 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
5544 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5547 SDValue Concat = tryFormConcatFromShuffle(Op, DAG);
5548 if (Concat.getNode())
5553 int NumInputElements = V1.getValueType().getVectorNumElements();
5554 if (isINSMask(ShuffleMask, NumInputElements, DstIsLeft, Anomaly)) {
5555 SDValue DstVec = DstIsLeft ? V1 : V2;
5556 SDValue DstLaneV = DAG.getConstant(Anomaly, dl, MVT::i64);
5558 SDValue SrcVec = V1;
5559 int SrcLane = ShuffleMask[Anomaly];
5560 if (SrcLane >= NumInputElements) {
5562 SrcLane -= VT.getVectorNumElements();
5564 SDValue SrcLaneV = DAG.getConstant(SrcLane, dl, MVT::i64);
5566 EVT ScalarVT = VT.getVectorElementType();
5568 if (ScalarVT.getSizeInBits() < 32 && ScalarVT.isInteger())
5569 ScalarVT = MVT::i32;
5572 ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5573 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ScalarVT, SrcVec, SrcLaneV),
5577 // If the shuffle is not directly supported and it has 4 elements, use
5578 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5579 unsigned NumElts = VT.getVectorNumElements();
5581 unsigned PFIndexes[4];
5582 for (unsigned i = 0; i != 4; ++i) {
5583 if (ShuffleMask[i] < 0)
5586 PFIndexes[i] = ShuffleMask[i];
5589 // Compute the index in the perfect shuffle table.
5590 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
5591 PFIndexes[2] * 9 + PFIndexes[3];
5592 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5593 unsigned Cost = (PFEntry >> 30);
5596 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5599 return GenerateTBL(Op, ShuffleMask, DAG);
5602 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
5604 EVT VT = BVN->getValueType(0);
5605 APInt SplatBits, SplatUndef;
5606 unsigned SplatBitSize;
5608 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5609 unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;
5611 for (unsigned i = 0; i < NumSplats; ++i) {
5612 CnstBits <<= SplatBitSize;
5613 UndefBits <<= SplatBitSize;
5614 CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits());
5615 UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits());
5624 SDValue AArch64TargetLowering::LowerVectorAND(SDValue Op,
5625 SelectionDAG &DAG) const {
5626 BuildVectorSDNode *BVN =
5627 dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5628 SDValue LHS = Op.getOperand(0);
5630 EVT VT = Op.getValueType();
5635 APInt CnstBits(VT.getSizeInBits(), 0);
5636 APInt UndefBits(VT.getSizeInBits(), 0);
5637 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5638 // We only have BIC vector immediate instruction, which is and-not.
5639 CnstBits = ~CnstBits;
5641 // We make use of a little bit of goto ickiness in order to avoid having to
5642 // duplicate the immediate matching logic for the undef toggled case.
5643 bool SecondTry = false;
5646 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5647 CnstBits = CnstBits.zextOrTrunc(64);
5648 uint64_t CnstVal = CnstBits.getZExtValue();
5650 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5651 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5652 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5653 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5654 DAG.getConstant(CnstVal, dl, MVT::i32),
5655 DAG.getConstant(0, dl, MVT::i32));
5656 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5659 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5660 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5661 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5662 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5663 DAG.getConstant(CnstVal, dl, MVT::i32),
5664 DAG.getConstant(8, dl, MVT::i32));
5665 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5668 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5669 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5670 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5671 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5672 DAG.getConstant(CnstVal, dl, MVT::i32),
5673 DAG.getConstant(16, dl, MVT::i32));
5674 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5677 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5678 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5679 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5680 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5681 DAG.getConstant(CnstVal, dl, MVT::i32),
5682 DAG.getConstant(24, dl, MVT::i32));
5683 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5686 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5687 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5688 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5689 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5690 DAG.getConstant(CnstVal, dl, MVT::i32),
5691 DAG.getConstant(0, dl, MVT::i32));
5692 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5695 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5696 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5697 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5698 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5699 DAG.getConstant(CnstVal, dl, MVT::i32),
5700 DAG.getConstant(8, dl, MVT::i32));
5701 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5708 CnstBits = ~UndefBits;
5712 // We can always fall back to a non-immediate AND.
5717 // Specialized code to quickly find if PotentialBVec is a BuildVector that
5718 // consists of only the same constant int value, returned in reference arg
5720 static bool isAllConstantBuildVector(const SDValue &PotentialBVec,
5721 uint64_t &ConstVal) {
5722 BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec);
5725 ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0));
5728 EVT VT = Bvec->getValueType(0);
5729 unsigned NumElts = VT.getVectorNumElements();
5730 for (unsigned i = 1; i < NumElts; ++i)
5731 if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt)
5733 ConstVal = FirstElt->getZExtValue();
5737 static unsigned getIntrinsicID(const SDNode *N) {
5738 unsigned Opcode = N->getOpcode();
5741 return Intrinsic::not_intrinsic;
5742 case ISD::INTRINSIC_WO_CHAIN: {
5743 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
5744 if (IID < Intrinsic::num_intrinsics)
5746 return Intrinsic::not_intrinsic;
5751 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
5752 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
5753 // BUILD_VECTORs with constant element C1, C2 is a constant, and C1 == ~C2.
5754 // Also, logical shift right -> sri, with the same structure.
5755 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
5756 EVT VT = N->getValueType(0);
5763 // Is the first op an AND?
5764 const SDValue And = N->getOperand(0);
5765 if (And.getOpcode() != ISD::AND)
5768 // Is the second op an shl or lshr?
5769 SDValue Shift = N->getOperand(1);
5770 // This will have been turned into: AArch64ISD::VSHL vector, #shift
5771 // or AArch64ISD::VLSHR vector, #shift
5772 unsigned ShiftOpc = Shift.getOpcode();
5773 if ((ShiftOpc != AArch64ISD::VSHL && ShiftOpc != AArch64ISD::VLSHR))
5775 bool IsShiftRight = ShiftOpc == AArch64ISD::VLSHR;
5777 // Is the shift amount constant?
5778 ConstantSDNode *C2node = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
5782 // Is the and mask vector all constant?
5784 if (!isAllConstantBuildVector(And.getOperand(1), C1))
5787 // Is C1 == ~C2, taking into account how much one can shift elements of a
5789 uint64_t C2 = C2node->getZExtValue();
5790 unsigned ElemSizeInBits = VT.getVectorElementType().getSizeInBits();
5791 if (C2 > ElemSizeInBits)
5793 unsigned ElemMask = (1 << ElemSizeInBits) - 1;
5794 if ((C1 & ElemMask) != (~C2 & ElemMask))
5797 SDValue X = And.getOperand(0);
5798 SDValue Y = Shift.getOperand(0);
5801 IsShiftRight ? Intrinsic::aarch64_neon_vsri : Intrinsic::aarch64_neon_vsli;
5803 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5804 DAG.getConstant(Intrin, DL, MVT::i32), X, Y,
5805 Shift.getOperand(1));
5807 DEBUG(dbgs() << "aarch64-lower: transformed: \n");
5808 DEBUG(N->dump(&DAG));
5809 DEBUG(dbgs() << "into: \n");
5810 DEBUG(ResultSLI->dump(&DAG));
5816 SDValue AArch64TargetLowering::LowerVectorOR(SDValue Op,
5817 SelectionDAG &DAG) const {
5818 // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
5819 if (EnableAArch64SlrGeneration) {
5820 SDValue Res = tryLowerToSLI(Op.getNode(), DAG);
5825 BuildVectorSDNode *BVN =
5826 dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode());
5827 SDValue LHS = Op.getOperand(1);
5829 EVT VT = Op.getValueType();
5831 // OR commutes, so try swapping the operands.
5833 LHS = Op.getOperand(0);
5834 BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5839 APInt CnstBits(VT.getSizeInBits(), 0);
5840 APInt UndefBits(VT.getSizeInBits(), 0);
5841 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5842 // We make use of a little bit of goto ickiness in order to avoid having to
5843 // duplicate the immediate matching logic for the undef toggled case.
5844 bool SecondTry = false;
5847 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5848 CnstBits = CnstBits.zextOrTrunc(64);
5849 uint64_t CnstVal = CnstBits.getZExtValue();
5851 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5852 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5853 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5854 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5855 DAG.getConstant(CnstVal, dl, MVT::i32),
5856 DAG.getConstant(0, dl, MVT::i32));
5857 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5860 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5861 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5862 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5863 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5864 DAG.getConstant(CnstVal, dl, MVT::i32),
5865 DAG.getConstant(8, dl, MVT::i32));
5866 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5869 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5870 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5871 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5872 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5873 DAG.getConstant(CnstVal, dl, MVT::i32),
5874 DAG.getConstant(16, dl, MVT::i32));
5875 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5878 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5879 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5880 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5881 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5882 DAG.getConstant(CnstVal, dl, MVT::i32),
5883 DAG.getConstant(24, dl, MVT::i32));
5884 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5887 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5888 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5889 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5890 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5891 DAG.getConstant(CnstVal, dl, MVT::i32),
5892 DAG.getConstant(0, dl, MVT::i32));
5893 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5896 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5897 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5898 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5899 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5900 DAG.getConstant(CnstVal, dl, MVT::i32),
5901 DAG.getConstant(8, dl, MVT::i32));
5902 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5909 CnstBits = UndefBits;
5913 // We can always fall back to a non-immediate OR.
5918 // Normalize the operands of BUILD_VECTOR. The value of constant operands will
5919 // be truncated to fit element width.
5920 static SDValue NormalizeBuildVector(SDValue Op,
5921 SelectionDAG &DAG) {
5922 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
5924 EVT VT = Op.getValueType();
5925 EVT EltTy= VT.getVectorElementType();
5927 if (EltTy.isFloatingPoint() || EltTy.getSizeInBits() > 16)
5930 SmallVector<SDValue, 16> Ops;
5931 for (SDValue Lane : Op->ops()) {
5932 if (auto *CstLane = dyn_cast<ConstantSDNode>(Lane)) {
5933 APInt LowBits(EltTy.getSizeInBits(),
5934 CstLane->getZExtValue());
5935 Lane = DAG.getConstant(LowBits.getZExtValue(), dl, MVT::i32);
5937 Ops.push_back(Lane);
5939 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5942 SDValue AArch64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
5943 SelectionDAG &DAG) const {
5945 EVT VT = Op.getValueType();
5946 Op = NormalizeBuildVector(Op, DAG);
5947 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5949 APInt CnstBits(VT.getSizeInBits(), 0);
5950 APInt UndefBits(VT.getSizeInBits(), 0);
5951 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5952 // We make use of a little bit of goto ickiness in order to avoid having to
5953 // duplicate the immediate matching logic for the undef toggled case.
5954 bool SecondTry = false;
5957 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5958 CnstBits = CnstBits.zextOrTrunc(64);
5959 uint64_t CnstVal = CnstBits.getZExtValue();
5961 // Certain magic vector constants (used to express things like NOT
5962 // and NEG) are passed through unmodified. This allows codegen patterns
5963 // for these operations to match. Special-purpose patterns will lower
5964 // these immediates to MOVIs if it proves necessary.
5965 if (VT.isInteger() && (CnstVal == 0 || CnstVal == ~0ULL))
5968 // The many faces of MOVI...
5969 if (AArch64_AM::isAdvSIMDModImmType10(CnstVal)) {
5970 CnstVal = AArch64_AM::encodeAdvSIMDModImmType10(CnstVal);
5971 if (VT.getSizeInBits() == 128) {
5972 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::v2i64,
5973 DAG.getConstant(CnstVal, dl, MVT::i32));
5974 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5977 // Support the V64 version via subregister insertion.
5978 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::f64,
5979 DAG.getConstant(CnstVal, dl, MVT::i32));
5980 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5983 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5984 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5985 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5986 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5987 DAG.getConstant(CnstVal, dl, MVT::i32),
5988 DAG.getConstant(0, dl, MVT::i32));
5989 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5992 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5993 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5994 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5995 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5996 DAG.getConstant(CnstVal, dl, MVT::i32),
5997 DAG.getConstant(8, dl, MVT::i32));
5998 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6001 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
6002 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
6003 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6004 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
6005 DAG.getConstant(CnstVal, dl, MVT::i32),
6006 DAG.getConstant(16, dl, MVT::i32));
6007 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6010 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
6011 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
6012 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6013 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
6014 DAG.getConstant(CnstVal, dl, MVT::i32),
6015 DAG.getConstant(24, dl, MVT::i32));
6016 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6019 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
6020 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
6021 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
6022 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
6023 DAG.getConstant(CnstVal, dl, MVT::i32),
6024 DAG.getConstant(0, dl, MVT::i32));
6025 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6028 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
6029 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
6030 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
6031 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
6032 DAG.getConstant(CnstVal, dl, MVT::i32),
6033 DAG.getConstant(8, dl, MVT::i32));
6034 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6037 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
6038 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
6039 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6040 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
6041 DAG.getConstant(CnstVal, dl, MVT::i32),
6042 DAG.getConstant(264, dl, MVT::i32));
6043 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6046 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
6047 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
6048 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6049 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
6050 DAG.getConstant(CnstVal, dl, MVT::i32),
6051 DAG.getConstant(272, dl, MVT::i32));
6052 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6055 if (AArch64_AM::isAdvSIMDModImmType9(CnstVal)) {
6056 CnstVal = AArch64_AM::encodeAdvSIMDModImmType9(CnstVal);
6057 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;
6058 SDValue Mov = DAG.getNode(AArch64ISD::MOVI, dl, MovTy,
6059 DAG.getConstant(CnstVal, dl, MVT::i32));
6060 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6063 // The few faces of FMOV...
6064 if (AArch64_AM::isAdvSIMDModImmType11(CnstVal)) {
6065 CnstVal = AArch64_AM::encodeAdvSIMDModImmType11(CnstVal);
6066 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4f32 : MVT::v2f32;
6067 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MovTy,
6068 DAG.getConstant(CnstVal, dl, MVT::i32));
6069 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6072 if (AArch64_AM::isAdvSIMDModImmType12(CnstVal) &&
6073 VT.getSizeInBits() == 128) {
6074 CnstVal = AArch64_AM::encodeAdvSIMDModImmType12(CnstVal);
6075 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MVT::v2f64,
6076 DAG.getConstant(CnstVal, dl, MVT::i32));
6077 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6080 // The many faces of MVNI...
6082 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
6083 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
6084 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6085 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6086 DAG.getConstant(CnstVal, dl, MVT::i32),
6087 DAG.getConstant(0, dl, MVT::i32));
6088 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6091 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
6092 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
6093 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6094 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6095 DAG.getConstant(CnstVal, dl, MVT::i32),
6096 DAG.getConstant(8, dl, MVT::i32));
6097 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6100 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
6101 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
6102 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6103 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6104 DAG.getConstant(CnstVal, dl, MVT::i32),
6105 DAG.getConstant(16, dl, MVT::i32));
6106 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6109 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
6110 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
6111 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6112 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6113 DAG.getConstant(CnstVal, dl, MVT::i32),
6114 DAG.getConstant(24, dl, MVT::i32));
6115 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6118 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
6119 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
6120 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
6121 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6122 DAG.getConstant(CnstVal, dl, MVT::i32),
6123 DAG.getConstant(0, dl, MVT::i32));
6124 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6127 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
6128 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
6129 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
6130 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6131 DAG.getConstant(CnstVal, dl, MVT::i32),
6132 DAG.getConstant(8, dl, MVT::i32));
6133 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6136 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
6137 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
6138 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6139 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
6140 DAG.getConstant(CnstVal, dl, MVT::i32),
6141 DAG.getConstant(264, dl, MVT::i32));
6142 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6145 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
6146 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
6147 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6148 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
6149 DAG.getConstant(CnstVal, dl, MVT::i32),
6150 DAG.getConstant(272, dl, MVT::i32));
6151 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6158 CnstBits = UndefBits;
6163 // Scan through the operands to find some interesting properties we can
6165 // 1) If only one value is used, we can use a DUP, or
6166 // 2) if only the low element is not undef, we can just insert that, or
6167 // 3) if only one constant value is used (w/ some non-constant lanes),
6168 // we can splat the constant value into the whole vector then fill
6169 // in the non-constant lanes.
6170 // 4) FIXME: If different constant values are used, but we can intelligently
6171 // select the values we'll be overwriting for the non-constant
6172 // lanes such that we can directly materialize the vector
6173 // some other way (MOVI, e.g.), we can be sneaky.
6174 unsigned NumElts = VT.getVectorNumElements();
6175 bool isOnlyLowElement = true;
6176 bool usesOnlyOneValue = true;
6177 bool usesOnlyOneConstantValue = true;
6178 bool isConstant = true;
6179 unsigned NumConstantLanes = 0;
6181 SDValue ConstantValue;
6182 for (unsigned i = 0; i < NumElts; ++i) {
6183 SDValue V = Op.getOperand(i);
6184 if (V.getOpcode() == ISD::UNDEF)
6187 isOnlyLowElement = false;
6188 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
6191 if (isa<ConstantSDNode>(V) || isa<ConstantFPSDNode>(V)) {
6193 if (!ConstantValue.getNode())
6195 else if (ConstantValue != V)
6196 usesOnlyOneConstantValue = false;
6199 if (!Value.getNode())
6201 else if (V != Value)
6202 usesOnlyOneValue = false;
6205 if (!Value.getNode())
6206 return DAG.getUNDEF(VT);
6208 if (isOnlyLowElement)
6209 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
6211 // Use DUP for non-constant splats. For f32 constant splats, reduce to
6212 // i32 and try again.
6213 if (usesOnlyOneValue) {
6215 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6216 Value.getValueType() != VT)
6217 return DAG.getNode(AArch64ISD::DUP, dl, VT, Value);
6219 // This is actually a DUPLANExx operation, which keeps everything vectory.
6221 // DUPLANE works on 128-bit vectors, widen it if necessary.
6222 SDValue Lane = Value.getOperand(1);
6223 Value = Value.getOperand(0);
6224 if (Value.getValueType().getSizeInBits() == 64)
6225 Value = WidenVector(Value, DAG);
6227 unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
6228 return DAG.getNode(Opcode, dl, VT, Value, Lane);
6231 if (VT.getVectorElementType().isFloatingPoint()) {
6232 SmallVector<SDValue, 8> Ops;
6233 EVT EltTy = VT.getVectorElementType();
6234 assert ((EltTy == MVT::f16 || EltTy == MVT::f32 || EltTy == MVT::f64) &&
6235 "Unsupported floating-point vector type");
6236 MVT NewType = MVT::getIntegerVT(EltTy.getSizeInBits());
6237 for (unsigned i = 0; i < NumElts; ++i)
6238 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, NewType, Op.getOperand(i)));
6239 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts);
6240 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
6241 Val = LowerBUILD_VECTOR(Val, DAG);
6243 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
6247 // If there was only one constant value used and for more than one lane,
6248 // start by splatting that value, then replace the non-constant lanes. This
6249 // is better than the default, which will perform a separate initialization
6251 if (NumConstantLanes > 0 && usesOnlyOneConstantValue) {
6252 SDValue Val = DAG.getNode(AArch64ISD::DUP, dl, VT, ConstantValue);
6253 // Now insert the non-constant lanes.
6254 for (unsigned i = 0; i < NumElts; ++i) {
6255 SDValue V = Op.getOperand(i);
6256 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i64);
6257 if (!isa<ConstantSDNode>(V) && !isa<ConstantFPSDNode>(V)) {
6258 // Note that type legalization likely mucked about with the VT of the
6259 // source operand, so we may have to convert it here before inserting.
6260 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, V, LaneIdx);
6266 // If all elements are constants and the case above didn't get hit, fall back
6267 // to the default expansion, which will generate a load from the constant
6272 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
6274 if (SDValue shuffle = ReconstructShuffle(Op, DAG))
6278 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
6279 // know the default expansion would otherwise fall back on something even
6280 // worse. For a vector with one or two non-undef values, that's
6281 // scalar_to_vector for the elements followed by a shuffle (provided the
6282 // shuffle is valid for the target) and materialization element by element
6283 // on the stack followed by a load for everything else.
6284 if (!isConstant && !usesOnlyOneValue) {
6285 SDValue Vec = DAG.getUNDEF(VT);
6286 SDValue Op0 = Op.getOperand(0);
6287 unsigned ElemSize = VT.getVectorElementType().getSizeInBits();
6289 // For 32 and 64 bit types, use INSERT_SUBREG for lane zero to
6290 // a) Avoid a RMW dependency on the full vector register, and
6291 // b) Allow the register coalescer to fold away the copy if the
6292 // value is already in an S or D register.
6293 // Do not do this for UNDEF/LOAD nodes because we have better patterns
6294 // for those avoiding the SCALAR_TO_VECTOR/BUILD_VECTOR.
6295 if (Op0.getOpcode() != ISD::UNDEF && Op0.getOpcode() != ISD::LOAD &&
6296 (ElemSize == 32 || ElemSize == 64)) {
6297 unsigned SubIdx = ElemSize == 32 ? AArch64::ssub : AArch64::dsub;
6299 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0,
6300 DAG.getTargetConstant(SubIdx, dl, MVT::i32));
6301 Vec = SDValue(N, 0);
6304 for (; i < NumElts; ++i) {
6305 SDValue V = Op.getOperand(i);
6306 if (V.getOpcode() == ISD::UNDEF)
6308 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i64);
6309 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
6314 // Just use the default expansion. We failed to find a better alternative.
6318 SDValue AArch64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
6319 SelectionDAG &DAG) const {
6320 assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!");
6322 // Check for non-constant or out of range lane.
6323 EVT VT = Op.getOperand(0).getValueType();
6324 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(2));
6325 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
6329 // Insertion/extraction are legal for V128 types.
6330 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
6331 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
6335 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
6336 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
6339 // For V64 types, we perform insertion by expanding the value
6340 // to a V128 type and perform the insertion on that.
6342 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
6343 EVT WideTy = WideVec.getValueType();
6345 SDValue Node = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideTy, WideVec,
6346 Op.getOperand(1), Op.getOperand(2));
6347 // Re-narrow the resultant vector.
6348 return NarrowVector(Node, DAG);
6352 AArch64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
6353 SelectionDAG &DAG) const {
6354 assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!");
6356 // Check for non-constant or out of range lane.
6357 EVT VT = Op.getOperand(0).getValueType();
6358 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6359 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
6363 // Insertion/extraction are legal for V128 types.
6364 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
6365 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
6369 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
6370 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
6373 // For V64 types, we perform extraction by expanding the value
6374 // to a V128 type and perform the extraction on that.
6376 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
6377 EVT WideTy = WideVec.getValueType();
6379 EVT ExtrTy = WideTy.getVectorElementType();
6380 if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8)
6383 // For extractions, we just return the result directly.
6384 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec,
6388 SDValue AArch64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
6389 SelectionDAG &DAG) const {
6390 EVT VT = Op.getOperand(0).getValueType();
6396 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6399 unsigned Val = Cst->getZExtValue();
6401 unsigned Size = Op.getValueType().getSizeInBits();
6405 return DAG.getTargetExtractSubreg(AArch64::bsub, dl, Op.getValueType(),
6408 return DAG.getTargetExtractSubreg(AArch64::hsub, dl, Op.getValueType(),
6411 return DAG.getTargetExtractSubreg(AArch64::ssub, dl, Op.getValueType(),
6414 return DAG.getTargetExtractSubreg(AArch64::dsub, dl, Op.getValueType(),
6417 llvm_unreachable("Unexpected vector type in extract_subvector!");
6420 // If this is extracting the upper 64-bits of a 128-bit vector, we match
6422 if (Size == 64 && Val * VT.getVectorElementType().getSizeInBits() == 64)
6428 bool AArch64TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
6430 if (VT.getVectorNumElements() == 4 &&
6431 (VT.is128BitVector() || VT.is64BitVector())) {
6432 unsigned PFIndexes[4];
6433 for (unsigned i = 0; i != 4; ++i) {
6437 PFIndexes[i] = M[i];
6440 // Compute the index in the perfect shuffle table.
6441 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
6442 PFIndexes[2] * 9 + PFIndexes[3];
6443 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
6444 unsigned Cost = (PFEntry >> 30);
6452 unsigned DummyUnsigned;
6454 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || isREVMask(M, VT, 64) ||
6455 isREVMask(M, VT, 32) || isREVMask(M, VT, 16) ||
6456 isEXTMask(M, VT, DummyBool, DummyUnsigned) ||
6457 // isTBLMask(M, VT) || // FIXME: Port TBL support from ARM.
6458 isTRNMask(M, VT, DummyUnsigned) || isUZPMask(M, VT, DummyUnsigned) ||
6459 isZIPMask(M, VT, DummyUnsigned) ||
6460 isTRN_v_undef_Mask(M, VT, DummyUnsigned) ||
6461 isUZP_v_undef_Mask(M, VT, DummyUnsigned) ||
6462 isZIP_v_undef_Mask(M, VT, DummyUnsigned) ||
6463 isINSMask(M, VT.getVectorNumElements(), DummyBool, DummyInt) ||
6464 isConcatMask(M, VT, VT.getSizeInBits() == 128));
6467 /// getVShiftImm - Check if this is a valid build_vector for the immediate
6468 /// operand of a vector shift operation, where all the elements of the
6469 /// build_vector must have the same constant integer value.
6470 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
6471 // Ignore bit_converts.
6472 while (Op.getOpcode() == ISD::BITCAST)
6473 Op = Op.getOperand(0);
6474 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
6475 APInt SplatBits, SplatUndef;
6476 unsigned SplatBitSize;
6478 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
6479 HasAnyUndefs, ElementBits) ||
6480 SplatBitSize > ElementBits)
6482 Cnt = SplatBits.getSExtValue();
6486 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
6487 /// operand of a vector shift left operation. That value must be in the range:
6488 /// 0 <= Value < ElementBits for a left shift; or
6489 /// 0 <= Value <= ElementBits for a long left shift.
6490 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
6491 assert(VT.isVector() && "vector shift count is not a vector type");
6492 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
6493 if (!getVShiftImm(Op, ElementBits, Cnt))
6495 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
6498 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
6499 /// operand of a vector shift right operation. The value must be in the range:
6500 /// 1 <= Value <= ElementBits for a right shift; or
6501 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, int64_t &Cnt) {
6502 assert(VT.isVector() && "vector shift count is not a vector type");
6503 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
6504 if (!getVShiftImm(Op, ElementBits, Cnt))
6506 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
6509 SDValue AArch64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
6510 SelectionDAG &DAG) const {
6511 EVT VT = Op.getValueType();
6515 if (!Op.getOperand(1).getValueType().isVector())
6517 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
6519 switch (Op.getOpcode()) {
6521 llvm_unreachable("unexpected shift opcode");
6524 if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
6525 return DAG.getNode(AArch64ISD::VSHL, DL, VT, Op.getOperand(0),
6526 DAG.getConstant(Cnt, DL, MVT::i32));
6527 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
6528 DAG.getConstant(Intrinsic::aarch64_neon_ushl, DL,
6530 Op.getOperand(0), Op.getOperand(1));
6533 // Right shift immediate
6534 if (isVShiftRImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize) {
6536 (Op.getOpcode() == ISD::SRA) ? AArch64ISD::VASHR : AArch64ISD::VLSHR;
6537 return DAG.getNode(Opc, DL, VT, Op.getOperand(0),
6538 DAG.getConstant(Cnt, DL, MVT::i32));
6541 // Right shift register. Note, there is not a shift right register
6542 // instruction, but the shift left register instruction takes a signed
6543 // value, where negative numbers specify a right shift.
6544 unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::aarch64_neon_sshl
6545 : Intrinsic::aarch64_neon_ushl;
6546 // negate the shift amount
6547 SDValue NegShift = DAG.getNode(AArch64ISD::NEG, DL, VT, Op.getOperand(1));
6548 SDValue NegShiftLeft =
6549 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
6550 DAG.getConstant(Opc, DL, MVT::i32), Op.getOperand(0),
6552 return NegShiftLeft;
6558 static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS,
6559 AArch64CC::CondCode CC, bool NoNans, EVT VT,
6560 SDLoc dl, SelectionDAG &DAG) {
6561 EVT SrcVT = LHS.getValueType();
6562 assert(VT.getSizeInBits() == SrcVT.getSizeInBits() &&
6563 "function only supposed to emit natural comparisons");
6565 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
6566 APInt CnstBits(VT.getSizeInBits(), 0);
6567 APInt UndefBits(VT.getSizeInBits(), 0);
6568 bool IsCnst = BVN && resolveBuildVector(BVN, CnstBits, UndefBits);
6569 bool IsZero = IsCnst && (CnstBits == 0);
6571 if (SrcVT.getVectorElementType().isFloatingPoint()) {
6575 case AArch64CC::NE: {
6578 Fcmeq = DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
6580 Fcmeq = DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
6581 return DAG.getNode(AArch64ISD::NOT, dl, VT, Fcmeq);
6585 return DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
6586 return DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
6589 return DAG.getNode(AArch64ISD::FCMGEz, dl, VT, LHS);
6590 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, LHS, RHS);
6593 return DAG.getNode(AArch64ISD::FCMGTz, dl, VT, LHS);
6594 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, LHS, RHS);
6597 return DAG.getNode(AArch64ISD::FCMLEz, dl, VT, LHS);
6598 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, RHS, LHS);
6602 // If we ignore NaNs then we can use to the MI implementation.
6606 return DAG.getNode(AArch64ISD::FCMLTz, dl, VT, LHS);
6607 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, RHS, LHS);
6614 case AArch64CC::NE: {
6617 Cmeq = DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
6619 Cmeq = DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
6620 return DAG.getNode(AArch64ISD::NOT, dl, VT, Cmeq);
6624 return DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
6625 return DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
6628 return DAG.getNode(AArch64ISD::CMGEz, dl, VT, LHS);
6629 return DAG.getNode(AArch64ISD::CMGE, dl, VT, LHS, RHS);
6632 return DAG.getNode(AArch64ISD::CMGTz, dl, VT, LHS);
6633 return DAG.getNode(AArch64ISD::CMGT, dl, VT, LHS, RHS);
6636 return DAG.getNode(AArch64ISD::CMLEz, dl, VT, LHS);
6637 return DAG.getNode(AArch64ISD::CMGE, dl, VT, RHS, LHS);
6639 return DAG.getNode(AArch64ISD::CMHS, dl, VT, RHS, LHS);
6641 return DAG.getNode(AArch64ISD::CMHI, dl, VT, RHS, LHS);
6644 return DAG.getNode(AArch64ISD::CMLTz, dl, VT, LHS);
6645 return DAG.getNode(AArch64ISD::CMGT, dl, VT, RHS, LHS);
6647 return DAG.getNode(AArch64ISD::CMHI, dl, VT, LHS, RHS);
6649 return DAG.getNode(AArch64ISD::CMHS, dl, VT, LHS, RHS);
6653 SDValue AArch64TargetLowering::LowerVSETCC(SDValue Op,
6654 SelectionDAG &DAG) const {
6655 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6656 SDValue LHS = Op.getOperand(0);
6657 SDValue RHS = Op.getOperand(1);
6658 EVT CmpVT = LHS.getValueType().changeVectorElementTypeToInteger();
6661 if (LHS.getValueType().getVectorElementType().isInteger()) {
6662 assert(LHS.getValueType() == RHS.getValueType());
6663 AArch64CC::CondCode AArch64CC = changeIntCCToAArch64CC(CC);
6665 EmitVectorComparison(LHS, RHS, AArch64CC, false, CmpVT, dl, DAG);
6666 return DAG.getSExtOrTrunc(Cmp, dl, Op.getValueType());
6669 assert(LHS.getValueType().getVectorElementType() == MVT::f32 ||
6670 LHS.getValueType().getVectorElementType() == MVT::f64);
6672 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
6673 // clean. Some of them require two branches to implement.
6674 AArch64CC::CondCode CC1, CC2;
6676 changeVectorFPCCToAArch64CC(CC, CC1, CC2, ShouldInvert);
6678 bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath;
6680 EmitVectorComparison(LHS, RHS, CC1, NoNaNs, CmpVT, dl, DAG);
6684 if (CC2 != AArch64CC::AL) {
6686 EmitVectorComparison(LHS, RHS, CC2, NoNaNs, CmpVT, dl, DAG);
6687 if (!Cmp2.getNode())
6690 Cmp = DAG.getNode(ISD::OR, dl, CmpVT, Cmp, Cmp2);
6693 Cmp = DAG.getSExtOrTrunc(Cmp, dl, Op.getValueType());
6696 return Cmp = DAG.getNOT(dl, Cmp, Cmp.getValueType());
6701 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
6702 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
6703 /// specified in the intrinsic calls.
6704 bool AArch64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
6706 unsigned Intrinsic) const {
6707 auto &DL = I.getModule()->getDataLayout();
6708 switch (Intrinsic) {
6709 case Intrinsic::aarch64_neon_ld2:
6710 case Intrinsic::aarch64_neon_ld3:
6711 case Intrinsic::aarch64_neon_ld4:
6712 case Intrinsic::aarch64_neon_ld1x2:
6713 case Intrinsic::aarch64_neon_ld1x3:
6714 case Intrinsic::aarch64_neon_ld1x4:
6715 case Intrinsic::aarch64_neon_ld2lane:
6716 case Intrinsic::aarch64_neon_ld3lane:
6717 case Intrinsic::aarch64_neon_ld4lane:
6718 case Intrinsic::aarch64_neon_ld2r:
6719 case Intrinsic::aarch64_neon_ld3r:
6720 case Intrinsic::aarch64_neon_ld4r: {
6721 Info.opc = ISD::INTRINSIC_W_CHAIN;
6722 // Conservatively set memVT to the entire set of vectors loaded.
6723 uint64_t NumElts = DL.getTypeAllocSize(I.getType()) / 8;
6724 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6725 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6728 Info.vol = false; // volatile loads with NEON intrinsics not supported
6729 Info.readMem = true;
6730 Info.writeMem = false;
6733 case Intrinsic::aarch64_neon_st2:
6734 case Intrinsic::aarch64_neon_st3:
6735 case Intrinsic::aarch64_neon_st4:
6736 case Intrinsic::aarch64_neon_st1x2:
6737 case Intrinsic::aarch64_neon_st1x3:
6738 case Intrinsic::aarch64_neon_st1x4:
6739 case Intrinsic::aarch64_neon_st2lane:
6740 case Intrinsic::aarch64_neon_st3lane:
6741 case Intrinsic::aarch64_neon_st4lane: {
6742 Info.opc = ISD::INTRINSIC_VOID;
6743 // Conservatively set memVT to the entire set of vectors stored.
6744 unsigned NumElts = 0;
6745 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
6746 Type *ArgTy = I.getArgOperand(ArgI)->getType();
6747 if (!ArgTy->isVectorTy())
6749 NumElts += DL.getTypeAllocSize(ArgTy) / 8;
6751 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6752 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6755 Info.vol = false; // volatile stores with NEON intrinsics not supported
6756 Info.readMem = false;
6757 Info.writeMem = true;
6760 case Intrinsic::aarch64_ldaxr:
6761 case Intrinsic::aarch64_ldxr: {
6762 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
6763 Info.opc = ISD::INTRINSIC_W_CHAIN;
6764 Info.memVT = MVT::getVT(PtrTy->getElementType());
6765 Info.ptrVal = I.getArgOperand(0);
6767 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
6769 Info.readMem = true;
6770 Info.writeMem = false;
6773 case Intrinsic::aarch64_stlxr:
6774 case Intrinsic::aarch64_stxr: {
6775 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
6776 Info.opc = ISD::INTRINSIC_W_CHAIN;
6777 Info.memVT = MVT::getVT(PtrTy->getElementType());
6778 Info.ptrVal = I.getArgOperand(1);
6780 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
6782 Info.readMem = false;
6783 Info.writeMem = true;
6786 case Intrinsic::aarch64_ldaxp:
6787 case Intrinsic::aarch64_ldxp: {
6788 Info.opc = ISD::INTRINSIC_W_CHAIN;
6789 Info.memVT = MVT::i128;
6790 Info.ptrVal = I.getArgOperand(0);
6794 Info.readMem = true;
6795 Info.writeMem = false;
6798 case Intrinsic::aarch64_stlxp:
6799 case Intrinsic::aarch64_stxp: {
6800 Info.opc = ISD::INTRINSIC_W_CHAIN;
6801 Info.memVT = MVT::i128;
6802 Info.ptrVal = I.getArgOperand(2);
6806 Info.readMem = false;
6807 Info.writeMem = true;
6817 // Truncations from 64-bit GPR to 32-bit GPR is free.
6818 bool AArch64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
6819 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6821 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6822 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6823 return NumBits1 > NumBits2;
6825 bool AArch64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
6826 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6828 unsigned NumBits1 = VT1.getSizeInBits();
6829 unsigned NumBits2 = VT2.getSizeInBits();
6830 return NumBits1 > NumBits2;
6833 /// Check if it is profitable to hoist instruction in then/else to if.
6834 /// Not profitable if I and it's user can form a FMA instruction
6835 /// because we prefer FMSUB/FMADD.
6836 bool AArch64TargetLowering::isProfitableToHoist(Instruction *I) const {
6837 if (I->getOpcode() != Instruction::FMul)
6840 if (I->getNumUses() != 1)
6843 Instruction *User = I->user_back();
6846 !(User->getOpcode() == Instruction::FSub ||
6847 User->getOpcode() == Instruction::FAdd))
6850 const TargetOptions &Options = getTargetMachine().Options;
6851 const DataLayout &DL = I->getModule()->getDataLayout();
6852 EVT VT = getValueType(DL, User->getOperand(0)->getType());
6854 if (isFMAFasterThanFMulAndFAdd(VT) &&
6855 isOperationLegalOrCustom(ISD::FMA, VT) &&
6856 (Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath))
6862 // All 32-bit GPR operations implicitly zero the high-half of the corresponding
6864 bool AArch64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
6865 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6867 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6868 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6869 return NumBits1 == 32 && NumBits2 == 64;
6871 bool AArch64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
6872 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6874 unsigned NumBits1 = VT1.getSizeInBits();
6875 unsigned NumBits2 = VT2.getSizeInBits();
6876 return NumBits1 == 32 && NumBits2 == 64;
6879 bool AArch64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
6880 EVT VT1 = Val.getValueType();
6881 if (isZExtFree(VT1, VT2)) {
6885 if (Val.getOpcode() != ISD::LOAD)
6888 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
6889 return (VT1.isSimple() && !VT1.isVector() && VT1.isInteger() &&
6890 VT2.isSimple() && !VT2.isVector() && VT2.isInteger() &&
6891 VT1.getSizeInBits() <= 32);
6894 bool AArch64TargetLowering::isExtFreeImpl(const Instruction *Ext) const {
6895 if (isa<FPExtInst>(Ext))
6898 // Vector types are next free.
6899 if (Ext->getType()->isVectorTy())
6902 for (const Use &U : Ext->uses()) {
6903 // The extension is free if we can fold it with a left shift in an
6904 // addressing mode or an arithmetic operation: add, sub, and cmp.
6906 // Is there a shift?
6907 const Instruction *Instr = cast<Instruction>(U.getUser());
6909 // Is this a constant shift?
6910 switch (Instr->getOpcode()) {
6911 case Instruction::Shl:
6912 if (!isa<ConstantInt>(Instr->getOperand(1)))
6915 case Instruction::GetElementPtr: {
6916 gep_type_iterator GTI = gep_type_begin(Instr);
6917 auto &DL = Ext->getModule()->getDataLayout();
6918 std::advance(GTI, U.getOperandNo());
6920 // This extension will end up with a shift because of the scaling factor.
6921 // 8-bit sized types have a scaling factor of 1, thus a shift amount of 0.
6922 // Get the shift amount based on the scaling factor:
6923 // log2(sizeof(IdxTy)) - log2(8).
6925 countTrailingZeros(DL.getTypeStoreSizeInBits(IdxTy)) - 3;
6926 // Is the constant foldable in the shift of the addressing mode?
6927 // I.e., shift amount is between 1 and 4 inclusive.
6928 if (ShiftAmt == 0 || ShiftAmt > 4)
6932 case Instruction::Trunc:
6933 // Check if this is a noop.
6934 // trunc(sext ty1 to ty2) to ty1.
6935 if (Instr->getType() == Ext->getOperand(0)->getType())
6942 // At this point we can use the bfm family, so this extension is free
6948 bool AArch64TargetLowering::hasPairedLoad(Type *LoadedType,
6949 unsigned &RequiredAligment) const {
6950 if (!LoadedType->isIntegerTy() && !LoadedType->isFloatTy())
6952 // Cyclone supports unaligned accesses.
6953 RequiredAligment = 0;
6954 unsigned NumBits = LoadedType->getPrimitiveSizeInBits();
6955 return NumBits == 32 || NumBits == 64;
6958 bool AArch64TargetLowering::hasPairedLoad(EVT LoadedType,
6959 unsigned &RequiredAligment) const {
6960 if (!LoadedType.isSimple() ||
6961 (!LoadedType.isInteger() && !LoadedType.isFloatingPoint()))
6963 // Cyclone supports unaligned accesses.
6964 RequiredAligment = 0;
6965 unsigned NumBits = LoadedType.getSizeInBits();
6966 return NumBits == 32 || NumBits == 64;
6969 /// \brief Lower an interleaved load into a ldN intrinsic.
6971 /// E.g. Lower an interleaved load (Factor = 2):
6972 /// %wide.vec = load <8 x i32>, <8 x i32>* %ptr
6973 /// %v0 = shuffle %wide.vec, undef, <0, 2, 4, 6> ; Extract even elements
6974 /// %v1 = shuffle %wide.vec, undef, <1, 3, 5, 7> ; Extract odd elements
6977 /// %ld2 = { <4 x i32>, <4 x i32> } call llvm.aarch64.neon.ld2(%ptr)
6978 /// %vec0 = extractelement { <4 x i32>, <4 x i32> } %ld2, i32 0
6979 /// %vec1 = extractelement { <4 x i32>, <4 x i32> } %ld2, i32 1
6980 bool AArch64TargetLowering::lowerInterleavedLoad(
6981 LoadInst *LI, ArrayRef<ShuffleVectorInst *> Shuffles,
6982 ArrayRef<unsigned> Indices, unsigned Factor) const {
6983 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
6984 "Invalid interleave factor");
6985 assert(!Shuffles.empty() && "Empty shufflevector input");
6986 assert(Shuffles.size() == Indices.size() &&
6987 "Unmatched number of shufflevectors and indices");
6989 const DataLayout &DL = LI->getModule()->getDataLayout();
6991 VectorType *VecTy = Shuffles[0]->getType();
6992 unsigned VecSize = DL.getTypeAllocSizeInBits(VecTy);
6994 // Skip if we do not have NEON and skip illegal vector types.
6995 if (!Subtarget->hasNEON() || (VecSize != 64 && VecSize != 128))
6998 // A pointer vector can not be the return type of the ldN intrinsics. Need to
6999 // load integer vectors first and then convert to pointer vectors.
7000 Type *EltTy = VecTy->getVectorElementType();
7001 if (EltTy->isPointerTy())
7003 VectorType::get(DL.getIntPtrType(EltTy), VecTy->getVectorNumElements());
7005 Type *PtrTy = VecTy->getPointerTo(LI->getPointerAddressSpace());
7006 Type *Tys[2] = {VecTy, PtrTy};
7007 static const Intrinsic::ID LoadInts[3] = {Intrinsic::aarch64_neon_ld2,
7008 Intrinsic::aarch64_neon_ld3,
7009 Intrinsic::aarch64_neon_ld4};
7011 Intrinsic::getDeclaration(LI->getModule(), LoadInts[Factor - 2], Tys);
7013 IRBuilder<> Builder(LI);
7014 Value *Ptr = Builder.CreateBitCast(LI->getPointerOperand(), PtrTy);
7016 CallInst *LdN = Builder.CreateCall(LdNFunc, Ptr, "ldN");
7018 // Replace uses of each shufflevector with the corresponding vector loaded
7020 for (unsigned i = 0; i < Shuffles.size(); i++) {
7021 ShuffleVectorInst *SVI = Shuffles[i];
7022 unsigned Index = Indices[i];
7024 Value *SubVec = Builder.CreateExtractValue(LdN, Index);
7026 // Convert the integer vector to pointer vector if the element is pointer.
7027 if (EltTy->isPointerTy())
7028 SubVec = Builder.CreateIntToPtr(SubVec, SVI->getType());
7030 SVI->replaceAllUsesWith(SubVec);
7036 /// \brief Get a mask consisting of sequential integers starting from \p Start.
7038 /// I.e. <Start, Start + 1, ..., Start + NumElts - 1>
7039 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned Start,
7041 SmallVector<Constant *, 16> Mask;
7042 for (unsigned i = 0; i < NumElts; i++)
7043 Mask.push_back(Builder.getInt32(Start + i));
7045 return ConstantVector::get(Mask);
7048 /// \brief Lower an interleaved store into a stN intrinsic.
7050 /// E.g. Lower an interleaved store (Factor = 3):
7051 /// %i.vec = shuffle <8 x i32> %v0, <8 x i32> %v1,
7052 /// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11>
7053 /// store <12 x i32> %i.vec, <12 x i32>* %ptr
7056 /// %sub.v0 = shuffle <8 x i32> %v0, <8 x i32> v1, <0, 1, 2, 3>
7057 /// %sub.v1 = shuffle <8 x i32> %v0, <8 x i32> v1, <4, 5, 6, 7>
7058 /// %sub.v2 = shuffle <8 x i32> %v0, <8 x i32> v1, <8, 9, 10, 11>
7059 /// call void llvm.aarch64.neon.st3(%sub.v0, %sub.v1, %sub.v2, %ptr)
7061 /// Note that the new shufflevectors will be removed and we'll only generate one
7062 /// st3 instruction in CodeGen.
7063 bool AArch64TargetLowering::lowerInterleavedStore(StoreInst *SI,
7064 ShuffleVectorInst *SVI,
7065 unsigned Factor) const {
7066 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
7067 "Invalid interleave factor");
7069 VectorType *VecTy = SVI->getType();
7070 assert(VecTy->getVectorNumElements() % Factor == 0 &&
7071 "Invalid interleaved store");
7073 unsigned NumSubElts = VecTy->getVectorNumElements() / Factor;
7074 Type *EltTy = VecTy->getVectorElementType();
7075 VectorType *SubVecTy = VectorType::get(EltTy, NumSubElts);
7077 const DataLayout &DL = SI->getModule()->getDataLayout();
7078 unsigned SubVecSize = DL.getTypeAllocSizeInBits(SubVecTy);
7080 // Skip if we do not have NEON and skip illegal vector types.
7081 if (!Subtarget->hasNEON() || (SubVecSize != 64 && SubVecSize != 128))
7084 Value *Op0 = SVI->getOperand(0);
7085 Value *Op1 = SVI->getOperand(1);
7086 IRBuilder<> Builder(SI);
7088 // StN intrinsics don't support pointer vectors as arguments. Convert pointer
7089 // vectors to integer vectors.
7090 if (EltTy->isPointerTy()) {
7091 Type *IntTy = DL.getIntPtrType(EltTy);
7092 unsigned NumOpElts =
7093 dyn_cast<VectorType>(Op0->getType())->getVectorNumElements();
7095 // Convert to the corresponding integer vector.
7096 Type *IntVecTy = VectorType::get(IntTy, NumOpElts);
7097 Op0 = Builder.CreatePtrToInt(Op0, IntVecTy);
7098 Op1 = Builder.CreatePtrToInt(Op1, IntVecTy);
7100 SubVecTy = VectorType::get(IntTy, NumSubElts);
7103 Type *PtrTy = SubVecTy->getPointerTo(SI->getPointerAddressSpace());
7104 Type *Tys[2] = {SubVecTy, PtrTy};
7105 static const Intrinsic::ID StoreInts[3] = {Intrinsic::aarch64_neon_st2,
7106 Intrinsic::aarch64_neon_st3,
7107 Intrinsic::aarch64_neon_st4};
7109 Intrinsic::getDeclaration(SI->getModule(), StoreInts[Factor - 2], Tys);
7111 SmallVector<Value *, 5> Ops;
7113 // Split the shufflevector operands into sub vectors for the new stN call.
7114 for (unsigned i = 0; i < Factor; i++)
7115 Ops.push_back(Builder.CreateShuffleVector(
7116 Op0, Op1, getSequentialMask(Builder, NumSubElts * i, NumSubElts)));
7118 Ops.push_back(Builder.CreateBitCast(SI->getPointerOperand(), PtrTy));
7119 Builder.CreateCall(StNFunc, Ops);
7123 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
7124 unsigned AlignCheck) {
7125 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
7126 (DstAlign == 0 || DstAlign % AlignCheck == 0));
7129 EVT AArch64TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
7130 unsigned SrcAlign, bool IsMemset,
7133 MachineFunction &MF) const {
7134 // Don't use AdvSIMD to implement 16-byte memset. It would have taken one
7135 // instruction to materialize the v2i64 zero and one store (with restrictive
7136 // addressing mode). Just do two i64 store of zero-registers.
7138 const Function *F = MF.getFunction();
7139 if (Subtarget->hasFPARMv8() && !IsMemset && Size >= 16 &&
7140 !F->hasFnAttribute(Attribute::NoImplicitFloat) &&
7141 (memOpAlign(SrcAlign, DstAlign, 16) ||
7142 (allowsMisalignedMemoryAccesses(MVT::f128, 0, 1, &Fast) && Fast)))
7146 (memOpAlign(SrcAlign, DstAlign, 8) ||
7147 (allowsMisalignedMemoryAccesses(MVT::i64, 0, 1, &Fast) && Fast)))
7151 (memOpAlign(SrcAlign, DstAlign, 4) ||
7152 (allowsMisalignedMemoryAccesses(MVT::i32, 0, 1, &Fast) && Fast)))
7158 // 12-bit optionally shifted immediates are legal for adds.
7159 bool AArch64TargetLowering::isLegalAddImmediate(int64_t Immed) const {
7160 if ((Immed >> 12) == 0 || ((Immed & 0xfff) == 0 && Immed >> 24 == 0))
7165 // Integer comparisons are implemented with ADDS/SUBS, so the range of valid
7166 // immediates is the same as for an add or a sub.
7167 bool AArch64TargetLowering::isLegalICmpImmediate(int64_t Immed) const {
7170 return isLegalAddImmediate(Immed);
7173 /// isLegalAddressingMode - Return true if the addressing mode represented
7174 /// by AM is legal for this target, for a load/store of the specified type.
7175 bool AArch64TargetLowering::isLegalAddressingMode(const DataLayout &DL,
7176 const AddrMode &AM, Type *Ty,
7177 unsigned AS) const {
7178 // AArch64 has five basic addressing modes:
7180 // reg + 9-bit signed offset
7181 // reg + SIZE_IN_BYTES * 12-bit unsigned offset
7183 // reg + SIZE_IN_BYTES * reg
7185 // No global is ever allowed as a base.
7189 // No reg+reg+imm addressing.
7190 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
7193 // check reg + imm case:
7194 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
7195 uint64_t NumBytes = 0;
7196 if (Ty->isSized()) {
7197 uint64_t NumBits = DL.getTypeSizeInBits(Ty);
7198 NumBytes = NumBits / 8;
7199 if (!isPowerOf2_64(NumBits))
7204 int64_t Offset = AM.BaseOffs;
7206 // 9-bit signed offset
7207 if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
7210 // 12-bit unsigned offset
7211 unsigned shift = Log2_64(NumBytes);
7212 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
7213 // Must be a multiple of NumBytes (NumBytes is a power of 2)
7214 (Offset >> shift) << shift == Offset)
7219 // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2
7221 if (!AM.Scale || AM.Scale == 1 ||
7222 (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
7227 int AArch64TargetLowering::getScalingFactorCost(const DataLayout &DL,
7228 const AddrMode &AM, Type *Ty,
7229 unsigned AS) const {
7230 // Scaling factors are not free at all.
7231 // Operands | Rt Latency
7232 // -------------------------------------------
7234 // -------------------------------------------
7235 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
7236 // Rt, [Xn, Wm, <extend> #imm] |
7237 if (isLegalAddressingMode(DL, AM, Ty, AS))
7238 // Scale represents reg2 * scale, thus account for 1 if
7239 // it is not equal to 0 or 1.
7240 return AM.Scale != 0 && AM.Scale != 1;
7244 bool AArch64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
7245 VT = VT.getScalarType();
7250 switch (VT.getSimpleVT().SimpleTy) {
7262 AArch64TargetLowering::getScratchRegisters(CallingConv::ID) const {
7263 // LR is a callee-save register, but we must treat it as clobbered by any call
7264 // site. Hence we include LR in the scratch registers, which are in turn added
7265 // as implicit-defs for stackmaps and patchpoints.
7266 static const MCPhysReg ScratchRegs[] = {
7267 AArch64::X16, AArch64::X17, AArch64::LR, 0
7273 AArch64TargetLowering::isDesirableToCommuteWithShift(const SDNode *N) const {
7274 EVT VT = N->getValueType(0);
7275 // If N is unsigned bit extraction: ((x >> C) & mask), then do not combine
7276 // it with shift to let it be lowered to UBFX.
7277 if (N->getOpcode() == ISD::AND && (VT == MVT::i32 || VT == MVT::i64) &&
7278 isa<ConstantSDNode>(N->getOperand(1))) {
7279 uint64_t TruncMask = N->getConstantOperandVal(1);
7280 if (isMask_64(TruncMask) &&
7281 N->getOperand(0).getOpcode() == ISD::SRL &&
7282 isa<ConstantSDNode>(N->getOperand(0)->getOperand(1)))
7288 bool AArch64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
7290 assert(Ty->isIntegerTy());
7292 unsigned BitSize = Ty->getPrimitiveSizeInBits();
7296 int64_t Val = Imm.getSExtValue();
7297 if (Val == 0 || AArch64_AM::isLogicalImmediate(Val, BitSize))
7300 if ((int64_t)Val < 0)
7303 Val &= (1LL << 32) - 1;
7305 unsigned LZ = countLeadingZeros((uint64_t)Val);
7306 unsigned Shift = (63 - LZ) / 16;
7307 // MOVZ is free so return true for one or fewer MOVK.
7311 // Generate SUBS and CSEL for integer abs.
7312 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
7313 EVT VT = N->getValueType(0);
7315 SDValue N0 = N->getOperand(0);
7316 SDValue N1 = N->getOperand(1);
7319 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
7320 // and change it to SUB and CSEL.
7321 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
7322 N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
7323 N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0))
7324 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
7325 if (Y1C->getAPIntValue() == VT.getSizeInBits() - 1) {
7326 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT),
7328 // Generate SUBS & CSEL.
7330 DAG.getNode(AArch64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32),
7331 N0.getOperand(0), DAG.getConstant(0, DL, VT));
7332 return DAG.getNode(AArch64ISD::CSEL, DL, VT, N0.getOperand(0), Neg,
7333 DAG.getConstant(AArch64CC::PL, DL, MVT::i32),
7334 SDValue(Cmp.getNode(), 1));
7339 // performXorCombine - Attempts to handle integer ABS.
7340 static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
7341 TargetLowering::DAGCombinerInfo &DCI,
7342 const AArch64Subtarget *Subtarget) {
7343 if (DCI.isBeforeLegalizeOps())
7346 return performIntegerAbsCombine(N, DAG);
7350 AArch64TargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor,
7352 std::vector<SDNode *> *Created) const {
7353 // fold (sdiv X, pow2)
7354 EVT VT = N->getValueType(0);
7355 if ((VT != MVT::i32 && VT != MVT::i64) ||
7356 !(Divisor.isPowerOf2() || (-Divisor).isPowerOf2()))
7360 SDValue N0 = N->getOperand(0);
7361 unsigned Lg2 = Divisor.countTrailingZeros();
7362 SDValue Zero = DAG.getConstant(0, DL, VT);
7363 SDValue Pow2MinusOne = DAG.getConstant((1ULL << Lg2) - 1, DL, VT);
7365 // Add (N0 < 0) ? Pow2 - 1 : 0;
7367 SDValue Cmp = getAArch64Cmp(N0, Zero, ISD::SETLT, CCVal, DAG, DL);
7368 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0, Pow2MinusOne);
7369 SDValue CSel = DAG.getNode(AArch64ISD::CSEL, DL, VT, Add, N0, CCVal, Cmp);
7372 Created->push_back(Cmp.getNode());
7373 Created->push_back(Add.getNode());
7374 Created->push_back(CSel.getNode());
7379 DAG.getNode(ISD::SRA, DL, VT, CSel, DAG.getConstant(Lg2, DL, MVT::i64));
7381 // If we're dividing by a positive value, we're done. Otherwise, we must
7382 // negate the result.
7383 if (Divisor.isNonNegative())
7387 Created->push_back(SRA.getNode());
7388 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), SRA);
7391 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG,
7392 TargetLowering::DAGCombinerInfo &DCI,
7393 const AArch64Subtarget *Subtarget) {
7394 if (DCI.isBeforeLegalizeOps())
7397 // Multiplication of a power of two plus/minus one can be done more
7398 // cheaply as as shift+add/sub. For now, this is true unilaterally. If
7399 // future CPUs have a cheaper MADD instruction, this may need to be
7400 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
7401 // 64-bit is 5 cycles, so this is always a win.
7402 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
7403 APInt Value = C->getAPIntValue();
7404 EVT VT = N->getValueType(0);
7406 if (Value.isNonNegative()) {
7407 // (mul x, 2^N + 1) => (add (shl x, N), x)
7408 APInt VM1 = Value - 1;
7409 if (VM1.isPowerOf2()) {
7410 SDValue ShiftedVal =
7411 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7412 DAG.getConstant(VM1.logBase2(), DL, MVT::i64));
7413 return DAG.getNode(ISD::ADD, DL, VT, ShiftedVal,
7416 // (mul x, 2^N - 1) => (sub (shl x, N), x)
7417 APInt VP1 = Value + 1;
7418 if (VP1.isPowerOf2()) {
7419 SDValue ShiftedVal =
7420 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7421 DAG.getConstant(VP1.logBase2(), DL, MVT::i64));
7422 return DAG.getNode(ISD::SUB, DL, VT, ShiftedVal,
7426 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
7427 APInt VNP1 = -Value + 1;
7428 if (VNP1.isPowerOf2()) {
7429 SDValue ShiftedVal =
7430 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7431 DAG.getConstant(VNP1.logBase2(), DL, MVT::i64));
7432 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0),
7435 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
7436 APInt VNM1 = -Value - 1;
7437 if (VNM1.isPowerOf2()) {
7438 SDValue ShiftedVal =
7439 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7440 DAG.getConstant(VNM1.logBase2(), DL, MVT::i64));
7442 DAG.getNode(ISD::ADD, DL, VT, ShiftedVal, N->getOperand(0));
7443 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Add);
7450 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
7451 SelectionDAG &DAG) {
7452 // Take advantage of vector comparisons producing 0 or -1 in each lane to
7453 // optimize away operation when it's from a constant.
7455 // The general transformation is:
7456 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
7457 // AND(VECTOR_CMP(x,y), constant2)
7458 // constant2 = UNARYOP(constant)
7460 // Early exit if this isn't a vector operation, the operand of the
7461 // unary operation isn't a bitwise AND, or if the sizes of the operations
7463 EVT VT = N->getValueType(0);
7464 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
7465 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
7466 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
7469 // Now check that the other operand of the AND is a constant. We could
7470 // make the transformation for non-constant splats as well, but it's unclear
7471 // that would be a benefit as it would not eliminate any operations, just
7472 // perform one more step in scalar code before moving to the vector unit.
7473 if (BuildVectorSDNode *BV =
7474 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
7475 // Bail out if the vector isn't a constant.
7476 if (!BV->isConstant())
7479 // Everything checks out. Build up the new and improved node.
7481 EVT IntVT = BV->getValueType(0);
7482 // Create a new constant of the appropriate type for the transformed
7484 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
7485 // The AND node needs bitcasts to/from an integer vector type around it.
7486 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
7487 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
7488 N->getOperand(0)->getOperand(0), MaskConst);
7489 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
7496 static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG,
7497 const AArch64Subtarget *Subtarget) {
7498 // First try to optimize away the conversion when it's conditionally from
7499 // a constant. Vectors only.
7500 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
7503 EVT VT = N->getValueType(0);
7504 if (VT != MVT::f32 && VT != MVT::f64)
7507 // Only optimize when the source and destination types have the same width.
7508 if (VT.getSizeInBits() != N->getOperand(0).getValueType().getSizeInBits())
7511 // If the result of an integer load is only used by an integer-to-float
7512 // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
7513 // This eliminates an "integer-to-vector-move" UOP and improves throughput.
7514 SDValue N0 = N->getOperand(0);
7515 if (Subtarget->hasNEON() && ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7516 // Do not change the width of a volatile load.
7517 !cast<LoadSDNode>(N0)->isVolatile()) {
7518 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7519 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
7520 LN0->getPointerInfo(), LN0->isVolatile(),
7521 LN0->isNonTemporal(), LN0->isInvariant(),
7522 LN0->getAlignment());
7524 // Make sure successors of the original load stay after it by updating them
7525 // to use the new Chain.
7526 DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));
7529 (N->getOpcode() == ISD::SINT_TO_FP) ? AArch64ISD::SITOF : AArch64ISD::UITOF;
7530 return DAG.getNode(Opcode, SDLoc(N), VT, Load);
7536 /// Fold a floating-point multiply by power of two into floating-point to
7537 /// fixed-point conversion.
7538 static SDValue performFpToIntCombine(SDNode *N, SelectionDAG &DAG,
7539 const AArch64Subtarget *Subtarget) {
7540 if (!Subtarget->hasNEON())
7543 SDValue Op = N->getOperand(0);
7544 if (!Op.getValueType().isVector() || Op.getOpcode() != ISD::FMUL)
7547 SDValue ConstVec = Op->getOperand(1);
7548 if (!isa<BuildVectorSDNode>(ConstVec))
7551 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
7552 uint32_t FloatBits = FloatTy.getSizeInBits();
7553 if (FloatBits != 32 && FloatBits != 64)
7556 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
7557 uint32_t IntBits = IntTy.getSizeInBits();
7558 if (IntBits != 16 && IntBits != 32 && IntBits != 64)
7561 // Avoid conversions where iN is larger than the float (e.g., float -> i64).
7562 if (IntBits > FloatBits)
7565 BitVector UndefElements;
7566 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(ConstVec);
7567 int32_t Bits = IntBits == 64 ? 64 : 32;
7568 int32_t C = BV->getConstantFPSplatPow2ToLog2Int(&UndefElements, Bits + 1);
7569 if (C == -1 || C == 0 || C > Bits)
7573 unsigned NumLanes = Op.getValueType().getVectorNumElements();
7578 ResTy = FloatBits == 32 ? MVT::v2i32 : MVT::v2i64;
7586 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
7587 unsigned IntrinsicOpcode = IsSigned ? Intrinsic::aarch64_neon_vcvtfp2fxs
7588 : Intrinsic::aarch64_neon_vcvtfp2fxu;
7590 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, ResTy,
7591 DAG.getConstant(IntrinsicOpcode, DL, MVT::i32),
7592 Op->getOperand(0), DAG.getConstant(C, DL, MVT::i32));
7593 // We can handle smaller integers by generating an extra trunc.
7594 if (IntBits < FloatBits)
7595 FixConv = DAG.getNode(ISD::TRUNCATE, DL, N->getValueType(0), FixConv);
7600 /// Fold a floating-point divide by power of two into fixed-point to
7601 /// floating-point conversion.
7602 static SDValue performFDivCombine(SDNode *N, SelectionDAG &DAG,
7603 const AArch64Subtarget *Subtarget) {
7604 if (!Subtarget->hasNEON())
7607 SDValue Op = N->getOperand(0);
7608 unsigned Opc = Op->getOpcode();
7609 if (!Op.getValueType().isVector() ||
7610 (Opc != ISD::SINT_TO_FP && Opc != ISD::UINT_TO_FP))
7613 SDValue ConstVec = N->getOperand(1);
7614 if (!isa<BuildVectorSDNode>(ConstVec))
7617 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
7618 int32_t IntBits = IntTy.getSizeInBits();
7619 if (IntBits != 16 && IntBits != 32 && IntBits != 64)
7622 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
7623 int32_t FloatBits = FloatTy.getSizeInBits();
7624 if (FloatBits != 32 && FloatBits != 64)
7627 // Avoid conversions where iN is larger than the float (e.g., i64 -> float).
7628 if (IntBits > FloatBits)
7631 BitVector UndefElements;
7632 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(ConstVec);
7633 int32_t C = BV->getConstantFPSplatPow2ToLog2Int(&UndefElements, FloatBits + 1);
7634 if (C == -1 || C == 0 || C > FloatBits)
7638 unsigned NumLanes = Op.getValueType().getVectorNumElements();
7643 ResTy = FloatBits == 32 ? MVT::v2i32 : MVT::v2i64;
7651 SDValue ConvInput = Op.getOperand(0);
7652 bool IsSigned = Opc == ISD::SINT_TO_FP;
7653 if (IntBits < FloatBits)
7654 ConvInput = DAG.getNode(IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND, DL,
7657 unsigned IntrinsicOpcode = IsSigned ? Intrinsic::aarch64_neon_vcvtfxs2fp
7658 : Intrinsic::aarch64_neon_vcvtfxu2fp;
7659 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, Op.getValueType(),
7660 DAG.getConstant(IntrinsicOpcode, DL, MVT::i32), ConvInput,
7661 DAG.getConstant(C, DL, MVT::i32));
7664 /// An EXTR instruction is made up of two shifts, ORed together. This helper
7665 /// searches for and classifies those shifts.
7666 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
7668 if (N.getOpcode() == ISD::SHL)
7670 else if (N.getOpcode() == ISD::SRL)
7675 if (!isa<ConstantSDNode>(N.getOperand(1)))
7678 ShiftAmount = N->getConstantOperandVal(1);
7679 Src = N->getOperand(0);
7683 /// EXTR instruction extracts a contiguous chunk of bits from two existing
7684 /// registers viewed as a high/low pair. This function looks for the pattern:
7685 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
7686 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
7688 static SDValue tryCombineToEXTR(SDNode *N,
7689 TargetLowering::DAGCombinerInfo &DCI) {
7690 SelectionDAG &DAG = DCI.DAG;
7692 EVT VT = N->getValueType(0);
7694 assert(N->getOpcode() == ISD::OR && "Unexpected root");
7696 if (VT != MVT::i32 && VT != MVT::i64)
7700 uint32_t ShiftLHS = 0;
7702 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
7706 uint32_t ShiftRHS = 0;
7708 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
7711 // If they're both trying to come from the high part of the register, they're
7712 // not really an EXTR.
7713 if (LHSFromHi == RHSFromHi)
7716 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
7720 std::swap(LHS, RHS);
7721 std::swap(ShiftLHS, ShiftRHS);
7724 return DAG.getNode(AArch64ISD::EXTR, DL, VT, LHS, RHS,
7725 DAG.getConstant(ShiftRHS, DL, MVT::i64));
7728 static SDValue tryCombineToBSL(SDNode *N,
7729 TargetLowering::DAGCombinerInfo &DCI) {
7730 EVT VT = N->getValueType(0);
7731 SelectionDAG &DAG = DCI.DAG;
7737 SDValue N0 = N->getOperand(0);
7738 if (N0.getOpcode() != ISD::AND)
7741 SDValue N1 = N->getOperand(1);
7742 if (N1.getOpcode() != ISD::AND)
7745 // We only have to look for constant vectors here since the general, variable
7746 // case can be handled in TableGen.
7747 unsigned Bits = VT.getVectorElementType().getSizeInBits();
7748 uint64_t BitMask = Bits == 64 ? -1ULL : ((1ULL << Bits) - 1);
7749 for (int i = 1; i >= 0; --i)
7750 for (int j = 1; j >= 0; --j) {
7751 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(i));
7752 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(j));
7756 bool FoundMatch = true;
7757 for (unsigned k = 0; k < VT.getVectorNumElements(); ++k) {
7758 ConstantSDNode *CN0 = dyn_cast<ConstantSDNode>(BVN0->getOperand(k));
7759 ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(BVN1->getOperand(k));
7761 CN0->getZExtValue() != (BitMask & ~CN1->getZExtValue())) {
7768 return DAG.getNode(AArch64ISD::BSL, DL, VT, SDValue(BVN0, 0),
7769 N0->getOperand(1 - i), N1->getOperand(1 - j));
7775 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
7776 const AArch64Subtarget *Subtarget) {
7777 // Attempt to form an EXTR from (or (shl VAL1, #N), (srl VAL2, #RegWidth-N))
7778 if (!EnableAArch64ExtrGeneration)
7780 SelectionDAG &DAG = DCI.DAG;
7781 EVT VT = N->getValueType(0);
7783 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7786 SDValue Res = tryCombineToEXTR(N, DCI);
7790 Res = tryCombineToBSL(N, DCI);
7797 static SDValue performBitcastCombine(SDNode *N,
7798 TargetLowering::DAGCombinerInfo &DCI,
7799 SelectionDAG &DAG) {
7800 // Wait 'til after everything is legalized to try this. That way we have
7801 // legal vector types and such.
7802 if (DCI.isBeforeLegalizeOps())
7805 // Remove extraneous bitcasts around an extract_subvector.
7807 // (v4i16 (bitconvert
7808 // (extract_subvector (v2i64 (bitconvert (v8i16 ...)), (i64 1)))))
7810 // (extract_subvector ((v8i16 ...), (i64 4)))
7812 // Only interested in 64-bit vectors as the ultimate result.
7813 EVT VT = N->getValueType(0);
7816 if (VT.getSimpleVT().getSizeInBits() != 64)
7818 // Is the operand an extract_subvector starting at the beginning or halfway
7819 // point of the vector? A low half may also come through as an
7820 // EXTRACT_SUBREG, so look for that, too.
7821 SDValue Op0 = N->getOperand(0);
7822 if (Op0->getOpcode() != ISD::EXTRACT_SUBVECTOR &&
7823 !(Op0->isMachineOpcode() &&
7824 Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG))
7826 uint64_t idx = cast<ConstantSDNode>(Op0->getOperand(1))->getZExtValue();
7827 if (Op0->getOpcode() == ISD::EXTRACT_SUBVECTOR) {
7828 if (Op0->getValueType(0).getVectorNumElements() != idx && idx != 0)
7830 } else if (Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG) {
7831 if (idx != AArch64::dsub)
7833 // The dsub reference is equivalent to a lane zero subvector reference.
7836 // Look through the bitcast of the input to the extract.
7837 if (Op0->getOperand(0)->getOpcode() != ISD::BITCAST)
7839 SDValue Source = Op0->getOperand(0)->getOperand(0);
7840 // If the source type has twice the number of elements as our destination
7841 // type, we know this is an extract of the high or low half of the vector.
7842 EVT SVT = Source->getValueType(0);
7843 if (SVT.getVectorNumElements() != VT.getVectorNumElements() * 2)
7846 DEBUG(dbgs() << "aarch64-lower: bitcast extract_subvector simplification\n");
7848 // Create the simplified form to just extract the low or high half of the
7849 // vector directly rather than bothering with the bitcasts.
7851 unsigned NumElements = VT.getVectorNumElements();
7853 SDValue HalfIdx = DAG.getConstant(NumElements, dl, MVT::i64);
7854 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Source, HalfIdx);
7856 SDValue SubReg = DAG.getTargetConstant(AArch64::dsub, dl, MVT::i32);
7857 return SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl, VT,
7863 static SDValue performConcatVectorsCombine(SDNode *N,
7864 TargetLowering::DAGCombinerInfo &DCI,
7865 SelectionDAG &DAG) {
7867 EVT VT = N->getValueType(0);
7868 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
7870 // Optimize concat_vectors of truncated vectors, where the intermediate
7871 // type is illegal, to avoid said illegality, e.g.,
7872 // (v4i16 (concat_vectors (v2i16 (truncate (v2i64))),
7873 // (v2i16 (truncate (v2i64)))))
7875 // (v4i16 (truncate (vector_shuffle (v4i32 (bitcast (v2i64))),
7876 // (v4i32 (bitcast (v2i64))),
7878 // This isn't really target-specific, but ISD::TRUNCATE legality isn't keyed
7879 // on both input and result type, so we might generate worse code.
7880 // On AArch64 we know it's fine for v2i64->v4i16 and v4i32->v8i8.
7881 if (N->getNumOperands() == 2 &&
7882 N0->getOpcode() == ISD::TRUNCATE &&
7883 N1->getOpcode() == ISD::TRUNCATE) {
7884 SDValue N00 = N0->getOperand(0);
7885 SDValue N10 = N1->getOperand(0);
7886 EVT N00VT = N00.getValueType();
7888 if (N00VT == N10.getValueType() &&
7889 (N00VT == MVT::v2i64 || N00VT == MVT::v4i32) &&
7890 N00VT.getScalarSizeInBits() == 4 * VT.getScalarSizeInBits()) {
7891 MVT MidVT = (N00VT == MVT::v2i64 ? MVT::v4i32 : MVT::v8i16);
7892 SmallVector<int, 8> Mask(MidVT.getVectorNumElements());
7893 for (size_t i = 0; i < Mask.size(); ++i)
7895 return DAG.getNode(ISD::TRUNCATE, dl, VT,
7896 DAG.getVectorShuffle(
7898 DAG.getNode(ISD::BITCAST, dl, MidVT, N00),
7899 DAG.getNode(ISD::BITCAST, dl, MidVT, N10), Mask));
7903 // Wait 'til after everything is legalized to try this. That way we have
7904 // legal vector types and such.
7905 if (DCI.isBeforeLegalizeOps())
7908 // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector
7909 // splat. The indexed instructions are going to be expecting a DUPLANE64, so
7910 // canonicalise to that.
7911 if (N0 == N1 && VT.getVectorNumElements() == 2) {
7912 assert(VT.getVectorElementType().getSizeInBits() == 64);
7913 return DAG.getNode(AArch64ISD::DUPLANE64, dl, VT, WidenVector(N0, DAG),
7914 DAG.getConstant(0, dl, MVT::i64));
7917 // Canonicalise concat_vectors so that the right-hand vector has as few
7918 // bit-casts as possible before its real operation. The primary matching
7919 // destination for these operations will be the narrowing "2" instructions,
7920 // which depend on the operation being performed on this right-hand vector.
7922 // (concat_vectors LHS, (v1i64 (bitconvert (v4i16 RHS))))
7924 // (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS))
7926 if (N1->getOpcode() != ISD::BITCAST)
7928 SDValue RHS = N1->getOperand(0);
7929 MVT RHSTy = RHS.getValueType().getSimpleVT();
7930 // If the RHS is not a vector, this is not the pattern we're looking for.
7931 if (!RHSTy.isVector())
7934 DEBUG(dbgs() << "aarch64-lower: concat_vectors bitcast simplification\n");
7936 MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(),
7937 RHSTy.getVectorNumElements() * 2);
7938 return DAG.getNode(ISD::BITCAST, dl, VT,
7939 DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatTy,
7940 DAG.getNode(ISD::BITCAST, dl, RHSTy, N0),
7944 static SDValue tryCombineFixedPointConvert(SDNode *N,
7945 TargetLowering::DAGCombinerInfo &DCI,
7946 SelectionDAG &DAG) {
7947 // Wait 'til after everything is legalized to try this. That way we have
7948 // legal vector types and such.
7949 if (DCI.isBeforeLegalizeOps())
7951 // Transform a scalar conversion of a value from a lane extract into a
7952 // lane extract of a vector conversion. E.g., from foo1 to foo2:
7953 // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); }
7954 // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; }
7956 // The second form interacts better with instruction selection and the
7957 // register allocator to avoid cross-class register copies that aren't
7958 // coalescable due to a lane reference.
7960 // Check the operand and see if it originates from a lane extract.
7961 SDValue Op1 = N->getOperand(1);
7962 if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7963 // Yep, no additional predication needed. Perform the transform.
7964 SDValue IID = N->getOperand(0);
7965 SDValue Shift = N->getOperand(2);
7966 SDValue Vec = Op1.getOperand(0);
7967 SDValue Lane = Op1.getOperand(1);
7968 EVT ResTy = N->getValueType(0);
7972 // The vector width should be 128 bits by the time we get here, even
7973 // if it started as 64 bits (the extract_vector handling will have
7975 assert(Vec.getValueType().getSizeInBits() == 128 &&
7976 "unexpected vector size on extract_vector_elt!");
7977 if (Vec.getValueType() == MVT::v4i32)
7978 VecResTy = MVT::v4f32;
7979 else if (Vec.getValueType() == MVT::v2i64)
7980 VecResTy = MVT::v2f64;
7982 llvm_unreachable("unexpected vector type!");
7985 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift);
7986 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane);
7991 // AArch64 high-vector "long" operations are formed by performing the non-high
7992 // version on an extract_subvector of each operand which gets the high half:
7994 // (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS))
7996 // However, there are cases which don't have an extract_high explicitly, but
7997 // have another operation that can be made compatible with one for free. For
8000 // (dupv64 scalar) --> (extract_high (dup128 scalar))
8002 // This routine does the actual conversion of such DUPs, once outer routines
8003 // have determined that everything else is in order.
8004 // It also supports immediate DUP-like nodes (MOVI/MVNi), which we can fold
8006 static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) {
8007 switch (N.getOpcode()) {
8008 case AArch64ISD::DUP:
8009 case AArch64ISD::DUPLANE8:
8010 case AArch64ISD::DUPLANE16:
8011 case AArch64ISD::DUPLANE32:
8012 case AArch64ISD::DUPLANE64:
8013 case AArch64ISD::MOVI:
8014 case AArch64ISD::MOVIshift:
8015 case AArch64ISD::MOVIedit:
8016 case AArch64ISD::MOVImsl:
8017 case AArch64ISD::MVNIshift:
8018 case AArch64ISD::MVNImsl:
8021 // FMOV could be supported, but isn't very useful, as it would only occur
8022 // if you passed a bitcast' floating point immediate to an eligible long
8023 // integer op (addl, smull, ...).
8027 MVT NarrowTy = N.getSimpleValueType();
8028 if (!NarrowTy.is64BitVector())
8031 MVT ElementTy = NarrowTy.getVectorElementType();
8032 unsigned NumElems = NarrowTy.getVectorNumElements();
8033 MVT NewVT = MVT::getVectorVT(ElementTy, NumElems * 2);
8036 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NarrowTy,
8037 DAG.getNode(N->getOpcode(), dl, NewVT, N->ops()),
8038 DAG.getConstant(NumElems, dl, MVT::i64));
8041 static bool isEssentiallyExtractSubvector(SDValue N) {
8042 if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR)
8045 return N.getOpcode() == ISD::BITCAST &&
8046 N.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR;
8049 /// \brief Helper structure to keep track of ISD::SET_CC operands.
8050 struct GenericSetCCInfo {
8051 const SDValue *Opnd0;
8052 const SDValue *Opnd1;
8056 /// \brief Helper structure to keep track of a SET_CC lowered into AArch64 code.
8057 struct AArch64SetCCInfo {
8059 AArch64CC::CondCode CC;
8062 /// \brief Helper structure to keep track of SetCC information.
8064 GenericSetCCInfo Generic;
8065 AArch64SetCCInfo AArch64;
8068 /// \brief Helper structure to be able to read SetCC information. If set to
8069 /// true, IsAArch64 field, Info is a AArch64SetCCInfo, otherwise Info is a
8070 /// GenericSetCCInfo.
8071 struct SetCCInfoAndKind {
8076 /// \brief Check whether or not \p Op is a SET_CC operation, either a generic or
8078 /// AArch64 lowered one.
8079 /// \p SetCCInfo is filled accordingly.
8080 /// \post SetCCInfo is meanginfull only when this function returns true.
8081 /// \return True when Op is a kind of SET_CC operation.
8082 static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) {
8083 // If this is a setcc, this is straight forward.
8084 if (Op.getOpcode() == ISD::SETCC) {
8085 SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0);
8086 SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1);
8087 SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
8088 SetCCInfo.IsAArch64 = false;
8091 // Otherwise, check if this is a matching csel instruction.
8095 if (Op.getOpcode() != AArch64ISD::CSEL)
8097 // Set the information about the operands.
8098 // TODO: we want the operands of the Cmp not the csel
8099 SetCCInfo.Info.AArch64.Cmp = &Op.getOperand(3);
8100 SetCCInfo.IsAArch64 = true;
8101 SetCCInfo.Info.AArch64.CC = static_cast<AArch64CC::CondCode>(
8102 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
8104 // Check that the operands matches the constraints:
8105 // (1) Both operands must be constants.
8106 // (2) One must be 1 and the other must be 0.
8107 ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0));
8108 ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1));
8111 if (!TValue || !FValue)
8115 if (!TValue->isOne()) {
8116 // Update the comparison when we are interested in !cc.
8117 std::swap(TValue, FValue);
8118 SetCCInfo.Info.AArch64.CC =
8119 AArch64CC::getInvertedCondCode(SetCCInfo.Info.AArch64.CC);
8121 return TValue->isOne() && FValue->isNullValue();
8124 // Returns true if Op is setcc or zext of setcc.
8125 static bool isSetCCOrZExtSetCC(const SDValue& Op, SetCCInfoAndKind &Info) {
8126 if (isSetCC(Op, Info))
8128 return ((Op.getOpcode() == ISD::ZERO_EXTEND) &&
8129 isSetCC(Op->getOperand(0), Info));
8132 // The folding we want to perform is:
8133 // (add x, [zext] (setcc cc ...) )
8135 // (csel x, (add x, 1), !cc ...)
8137 // The latter will get matched to a CSINC instruction.
8138 static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) {
8139 assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!");
8140 SDValue LHS = Op->getOperand(0);
8141 SDValue RHS = Op->getOperand(1);
8142 SetCCInfoAndKind InfoAndKind;
8144 // If neither operand is a SET_CC, give up.
8145 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind)) {
8146 std::swap(LHS, RHS);
8147 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind))
8151 // FIXME: This could be generatized to work for FP comparisons.
8152 EVT CmpVT = InfoAndKind.IsAArch64
8153 ? InfoAndKind.Info.AArch64.Cmp->getOperand(0).getValueType()
8154 : InfoAndKind.Info.Generic.Opnd0->getValueType();
8155 if (CmpVT != MVT::i32 && CmpVT != MVT::i64)
8161 if (InfoAndKind.IsAArch64) {
8162 CCVal = DAG.getConstant(
8163 AArch64CC::getInvertedCondCode(InfoAndKind.Info.AArch64.CC), dl,
8165 Cmp = *InfoAndKind.Info.AArch64.Cmp;
8167 Cmp = getAArch64Cmp(*InfoAndKind.Info.Generic.Opnd0,
8168 *InfoAndKind.Info.Generic.Opnd1,
8169 ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, true),
8172 EVT VT = Op->getValueType(0);
8173 LHS = DAG.getNode(ISD::ADD, dl, VT, RHS, DAG.getConstant(1, dl, VT));
8174 return DAG.getNode(AArch64ISD::CSEL, dl, VT, RHS, LHS, CCVal, Cmp);
8177 // The basic add/sub long vector instructions have variants with "2" on the end
8178 // which act on the high-half of their inputs. They are normally matched by
8181 // (add (zeroext (extract_high LHS)),
8182 // (zeroext (extract_high RHS)))
8183 // -> uaddl2 vD, vN, vM
8185 // However, if one of the extracts is something like a duplicate, this
8186 // instruction can still be used profitably. This function puts the DAG into a
8187 // more appropriate form for those patterns to trigger.
8188 static SDValue performAddSubLongCombine(SDNode *N,
8189 TargetLowering::DAGCombinerInfo &DCI,
8190 SelectionDAG &DAG) {
8191 if (DCI.isBeforeLegalizeOps())
8194 MVT VT = N->getSimpleValueType(0);
8195 if (!VT.is128BitVector()) {
8196 if (N->getOpcode() == ISD::ADD)
8197 return performSetccAddFolding(N, DAG);
8201 // Make sure both branches are extended in the same way.
8202 SDValue LHS = N->getOperand(0);
8203 SDValue RHS = N->getOperand(1);
8204 if ((LHS.getOpcode() != ISD::ZERO_EXTEND &&
8205 LHS.getOpcode() != ISD::SIGN_EXTEND) ||
8206 LHS.getOpcode() != RHS.getOpcode())
8209 unsigned ExtType = LHS.getOpcode();
8211 // It's not worth doing if at least one of the inputs isn't already an
8212 // extract, but we don't know which it'll be so we have to try both.
8213 if (isEssentiallyExtractSubvector(LHS.getOperand(0))) {
8214 RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG);
8218 RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS);
8219 } else if (isEssentiallyExtractSubvector(RHS.getOperand(0))) {
8220 LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG);
8224 LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS);
8227 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS);
8230 // Massage DAGs which we can use the high-half "long" operations on into
8231 // something isel will recognize better. E.g.
8233 // (aarch64_neon_umull (extract_high vec) (dupv64 scalar)) -->
8234 // (aarch64_neon_umull (extract_high (v2i64 vec)))
8235 // (extract_high (v2i64 (dup128 scalar)))))
8237 static SDValue tryCombineLongOpWithDup(SDNode *N,
8238 TargetLowering::DAGCombinerInfo &DCI,
8239 SelectionDAG &DAG) {
8240 if (DCI.isBeforeLegalizeOps())
8243 bool IsIntrinsic = N->getOpcode() == ISD::INTRINSIC_WO_CHAIN;
8244 SDValue LHS = N->getOperand(IsIntrinsic ? 1 : 0);
8245 SDValue RHS = N->getOperand(IsIntrinsic ? 2 : 1);
8246 assert(LHS.getValueType().is64BitVector() &&
8247 RHS.getValueType().is64BitVector() &&
8248 "unexpected shape for long operation");
8250 // Either node could be a DUP, but it's not worth doing both of them (you'd
8251 // just as well use the non-high version) so look for a corresponding extract
8252 // operation on the other "wing".
8253 if (isEssentiallyExtractSubvector(LHS)) {
8254 RHS = tryExtendDUPToExtractHigh(RHS, DAG);
8257 } else if (isEssentiallyExtractSubvector(RHS)) {
8258 LHS = tryExtendDUPToExtractHigh(LHS, DAG);
8263 // N could either be an intrinsic or a sabsdiff/uabsdiff node.
8265 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0),
8266 N->getOperand(0), LHS, RHS);
8268 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
8272 static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
8273 MVT ElemTy = N->getSimpleValueType(0).getScalarType();
8274 unsigned ElemBits = ElemTy.getSizeInBits();
8276 int64_t ShiftAmount;
8277 if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) {
8278 APInt SplatValue, SplatUndef;
8279 unsigned SplatBitSize;
8281 if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
8282 HasAnyUndefs, ElemBits) ||
8283 SplatBitSize != ElemBits)
8286 ShiftAmount = SplatValue.getSExtValue();
8287 } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) {
8288 ShiftAmount = CVN->getSExtValue();
8296 llvm_unreachable("Unknown shift intrinsic");
8297 case Intrinsic::aarch64_neon_sqshl:
8298 Opcode = AArch64ISD::SQSHL_I;
8299 IsRightShift = false;
8301 case Intrinsic::aarch64_neon_uqshl:
8302 Opcode = AArch64ISD::UQSHL_I;
8303 IsRightShift = false;
8305 case Intrinsic::aarch64_neon_srshl:
8306 Opcode = AArch64ISD::SRSHR_I;
8307 IsRightShift = true;
8309 case Intrinsic::aarch64_neon_urshl:
8310 Opcode = AArch64ISD::URSHR_I;
8311 IsRightShift = true;
8313 case Intrinsic::aarch64_neon_sqshlu:
8314 Opcode = AArch64ISD::SQSHLU_I;
8315 IsRightShift = false;
8319 if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits) {
8321 return DAG.getNode(Opcode, dl, N->getValueType(0), N->getOperand(1),
8322 DAG.getConstant(-ShiftAmount, dl, MVT::i32));
8323 } else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount < ElemBits) {
8325 return DAG.getNode(Opcode, dl, N->getValueType(0), N->getOperand(1),
8326 DAG.getConstant(ShiftAmount, dl, MVT::i32));
8332 // The CRC32[BH] instructions ignore the high bits of their data operand. Since
8333 // the intrinsics must be legal and take an i32, this means there's almost
8334 // certainly going to be a zext in the DAG which we can eliminate.
8335 static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) {
8336 SDValue AndN = N->getOperand(2);
8337 if (AndN.getOpcode() != ISD::AND)
8340 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1));
8341 if (!CMask || CMask->getZExtValue() != Mask)
8344 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32,
8345 N->getOperand(0), N->getOperand(1), AndN.getOperand(0));
8348 static SDValue combineAcrossLanesIntrinsic(unsigned Opc, SDNode *N,
8349 SelectionDAG &DAG) {
8351 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0),
8352 DAG.getNode(Opc, dl,
8353 N->getOperand(1).getSimpleValueType(),
8355 DAG.getConstant(0, dl, MVT::i64));
8358 static SDValue performIntrinsicCombine(SDNode *N,
8359 TargetLowering::DAGCombinerInfo &DCI,
8360 const AArch64Subtarget *Subtarget) {
8361 SelectionDAG &DAG = DCI.DAG;
8362 unsigned IID = getIntrinsicID(N);
8366 case Intrinsic::aarch64_neon_vcvtfxs2fp:
8367 case Intrinsic::aarch64_neon_vcvtfxu2fp:
8368 return tryCombineFixedPointConvert(N, DCI, DAG);
8369 case Intrinsic::aarch64_neon_saddv:
8370 return combineAcrossLanesIntrinsic(AArch64ISD::SADDV, N, DAG);
8371 case Intrinsic::aarch64_neon_uaddv:
8372 return combineAcrossLanesIntrinsic(AArch64ISD::UADDV, N, DAG);
8373 case Intrinsic::aarch64_neon_sminv:
8374 return combineAcrossLanesIntrinsic(AArch64ISD::SMINV, N, DAG);
8375 case Intrinsic::aarch64_neon_uminv:
8376 return combineAcrossLanesIntrinsic(AArch64ISD::UMINV, N, DAG);
8377 case Intrinsic::aarch64_neon_smaxv:
8378 return combineAcrossLanesIntrinsic(AArch64ISD::SMAXV, N, DAG);
8379 case Intrinsic::aarch64_neon_umaxv:
8380 return combineAcrossLanesIntrinsic(AArch64ISD::UMAXV, N, DAG);
8381 case Intrinsic::aarch64_neon_fmax:
8382 return DAG.getNode(ISD::FMAXNAN, SDLoc(N), N->getValueType(0),
8383 N->getOperand(1), N->getOperand(2));
8384 case Intrinsic::aarch64_neon_fmin:
8385 return DAG.getNode(ISD::FMINNAN, SDLoc(N), N->getValueType(0),
8386 N->getOperand(1), N->getOperand(2));
8387 case Intrinsic::aarch64_neon_sabd:
8388 return DAG.getNode(ISD::SABSDIFF, SDLoc(N), N->getValueType(0),
8389 N->getOperand(1), N->getOperand(2));
8390 case Intrinsic::aarch64_neon_uabd:
8391 return DAG.getNode(ISD::UABSDIFF, SDLoc(N), N->getValueType(0),
8392 N->getOperand(1), N->getOperand(2));
8393 case Intrinsic::aarch64_neon_fmaxnm:
8394 return DAG.getNode(ISD::FMAXNUM, SDLoc(N), N->getValueType(0),
8395 N->getOperand(1), N->getOperand(2));
8396 case Intrinsic::aarch64_neon_fminnm:
8397 return DAG.getNode(ISD::FMINNUM, SDLoc(N), N->getValueType(0),
8398 N->getOperand(1), N->getOperand(2));
8399 case Intrinsic::aarch64_neon_smull:
8400 case Intrinsic::aarch64_neon_umull:
8401 case Intrinsic::aarch64_neon_pmull:
8402 case Intrinsic::aarch64_neon_sqdmull:
8403 return tryCombineLongOpWithDup(N, DCI, DAG);
8404 case Intrinsic::aarch64_neon_sqshl:
8405 case Intrinsic::aarch64_neon_uqshl:
8406 case Intrinsic::aarch64_neon_sqshlu:
8407 case Intrinsic::aarch64_neon_srshl:
8408 case Intrinsic::aarch64_neon_urshl:
8409 return tryCombineShiftImm(IID, N, DAG);
8410 case Intrinsic::aarch64_crc32b:
8411 case Intrinsic::aarch64_crc32cb:
8412 return tryCombineCRC32(0xff, N, DAG);
8413 case Intrinsic::aarch64_crc32h:
8414 case Intrinsic::aarch64_crc32ch:
8415 return tryCombineCRC32(0xffff, N, DAG);
8420 static SDValue performExtendCombine(SDNode *N,
8421 TargetLowering::DAGCombinerInfo &DCI,
8422 SelectionDAG &DAG) {
8423 // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then
8424 // we can convert that DUP into another extract_high (of a bigger DUP), which
8425 // helps the backend to decide that an sabdl2 would be useful, saving a real
8426 // extract_high operation.
8427 if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND &&
8428 (N->getOperand(0).getOpcode() == ISD::SABSDIFF ||
8429 N->getOperand(0).getOpcode() == ISD::UABSDIFF)) {
8430 SDNode *ABDNode = N->getOperand(0).getNode();
8431 SDValue NewABD = tryCombineLongOpWithDup(ABDNode, DCI, DAG);
8432 if (!NewABD.getNode())
8435 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0),
8439 // This is effectively a custom type legalization for AArch64.
8441 // Type legalization will split an extend of a small, legal, type to a larger
8442 // illegal type by first splitting the destination type, often creating
8443 // illegal source types, which then get legalized in isel-confusing ways,
8444 // leading to really terrible codegen. E.g.,
8445 // %result = v8i32 sext v8i8 %value
8447 // %losrc = extract_subreg %value, ...
8448 // %hisrc = extract_subreg %value, ...
8449 // %lo = v4i32 sext v4i8 %losrc
8450 // %hi = v4i32 sext v4i8 %hisrc
8451 // Things go rapidly downhill from there.
8453 // For AArch64, the [sz]ext vector instructions can only go up one element
8454 // size, so we can, e.g., extend from i8 to i16, but to go from i8 to i32
8455 // take two instructions.
8457 // This implies that the most efficient way to do the extend from v8i8
8458 // to two v4i32 values is to first extend the v8i8 to v8i16, then do
8459 // the normal splitting to happen for the v8i16->v8i32.
8461 // This is pre-legalization to catch some cases where the default
8462 // type legalization will create ill-tempered code.
8463 if (!DCI.isBeforeLegalizeOps())
8466 // We're only interested in cleaning things up for non-legal vector types
8467 // here. If both the source and destination are legal, things will just
8468 // work naturally without any fiddling.
8469 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8470 EVT ResVT = N->getValueType(0);
8471 if (!ResVT.isVector() || TLI.isTypeLegal(ResVT))
8473 // If the vector type isn't a simple VT, it's beyond the scope of what
8474 // we're worried about here. Let legalization do its thing and hope for
8476 SDValue Src = N->getOperand(0);
8477 EVT SrcVT = Src->getValueType(0);
8478 if (!ResVT.isSimple() || !SrcVT.isSimple())
8481 // If the source VT is a 64-bit vector, we can play games and get the
8482 // better results we want.
8483 if (SrcVT.getSizeInBits() != 64)
8486 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
8487 unsigned ElementCount = SrcVT.getVectorNumElements();
8488 SrcVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize * 2), ElementCount);
8490 Src = DAG.getNode(N->getOpcode(), DL, SrcVT, Src);
8492 // Now split the rest of the operation into two halves, each with a 64
8496 unsigned NumElements = ResVT.getVectorNumElements();
8497 assert(!(NumElements & 1) && "Splitting vector, but not in half!");
8498 LoVT = HiVT = EVT::getVectorVT(*DAG.getContext(),
8499 ResVT.getVectorElementType(), NumElements / 2);
8501 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
8502 LoVT.getVectorNumElements());
8503 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
8504 DAG.getConstant(0, DL, MVT::i64));
8505 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
8506 DAG.getConstant(InNVT.getVectorNumElements(), DL, MVT::i64));
8507 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, Lo);
8508 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, Hi);
8510 // Now combine the parts back together so we still have a single result
8511 // like the combiner expects.
8512 return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
8515 /// Replace a splat of a scalar to a vector store by scalar stores of the scalar
8516 /// value. The load store optimizer pass will merge them to store pair stores.
8517 /// This has better performance than a splat of the scalar followed by a split
8518 /// vector store. Even if the stores are not merged it is four stores vs a dup,
8519 /// followed by an ext.b and two stores.
8520 static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) {
8521 SDValue StVal = St->getValue();
8522 EVT VT = StVal.getValueType();
8524 // Don't replace floating point stores, they possibly won't be transformed to
8525 // stp because of the store pair suppress pass.
8526 if (VT.isFloatingPoint())
8529 // Check for insert vector elements.
8530 if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT)
8533 // We can express a splat as store pair(s) for 2 or 4 elements.
8534 unsigned NumVecElts = VT.getVectorNumElements();
8535 if (NumVecElts != 4 && NumVecElts != 2)
8537 SDValue SplatVal = StVal.getOperand(1);
8538 unsigned RemainInsertElts = NumVecElts - 1;
8540 // Check that this is a splat.
8541 while (--RemainInsertElts) {
8542 SDValue NextInsertElt = StVal.getOperand(0);
8543 if (NextInsertElt.getOpcode() != ISD::INSERT_VECTOR_ELT)
8545 if (NextInsertElt.getOperand(1) != SplatVal)
8547 StVal = NextInsertElt;
8549 unsigned OrigAlignment = St->getAlignment();
8550 unsigned EltOffset = NumVecElts == 4 ? 4 : 8;
8551 unsigned Alignment = std::min(OrigAlignment, EltOffset);
8553 // Create scalar stores. This is at least as good as the code sequence for a
8554 // split unaligned store which is a dup.s, ext.b, and two stores.
8555 // Most of the time the three stores should be replaced by store pair
8556 // instructions (stp).
8558 SDValue BasePtr = St->getBasePtr();
8560 DAG.getStore(St->getChain(), DL, SplatVal, BasePtr, St->getPointerInfo(),
8561 St->isVolatile(), St->isNonTemporal(), St->getAlignment());
8563 unsigned Offset = EltOffset;
8564 while (--NumVecElts) {
8565 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
8566 DAG.getConstant(Offset, DL, MVT::i64));
8567 NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr,
8568 St->getPointerInfo(), St->isVolatile(),
8569 St->isNonTemporal(), Alignment);
8570 Offset += EltOffset;
8575 static SDValue performSTORECombine(SDNode *N,
8576 TargetLowering::DAGCombinerInfo &DCI,
8578 const AArch64Subtarget *Subtarget) {
8579 if (!DCI.isBeforeLegalize())
8582 StoreSDNode *S = cast<StoreSDNode>(N);
8583 if (S->isVolatile())
8586 // FIXME: The logic for deciding if an unaligned store should be split should
8587 // be included in TLI.allowsMisalignedMemoryAccesses(), and there should be
8588 // a call to that function here.
8590 // Cyclone has bad performance on unaligned 16B stores when crossing line and
8591 // page boundaries. We want to split such stores.
8592 if (!Subtarget->isCyclone())
8595 // Don't split at -Oz.
8596 if (DAG.getMachineFunction().getFunction()->optForMinSize())
8599 SDValue StVal = S->getValue();
8600 EVT VT = StVal.getValueType();
8602 // Don't split v2i64 vectors. Memcpy lowering produces those and splitting
8603 // those up regresses performance on micro-benchmarks and olden/bh.
8604 if (!VT.isVector() || VT.getVectorNumElements() < 2 || VT == MVT::v2i64)
8607 // Split unaligned 16B stores. They are terrible for performance.
8608 // Don't split stores with alignment of 1 or 2. Code that uses clang vector
8609 // extensions can use this to mark that it does not want splitting to happen
8610 // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of
8611 // eliminating alignment hazards is only 1 in 8 for alignment of 2.
8612 if (VT.getSizeInBits() != 128 || S->getAlignment() >= 16 ||
8613 S->getAlignment() <= 2)
8616 // If we get a splat of a scalar convert this vector store to a store of
8617 // scalars. They will be merged into store pairs thereby removing two
8619 if (SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S))
8620 return ReplacedSplat;
8623 unsigned NumElts = VT.getVectorNumElements() / 2;
8624 // Split VT into two.
8626 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), NumElts);
8627 SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
8628 DAG.getConstant(0, DL, MVT::i64));
8629 SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
8630 DAG.getConstant(NumElts, DL, MVT::i64));
8631 SDValue BasePtr = S->getBasePtr();
8633 DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
8634 S->isVolatile(), S->isNonTemporal(), S->getAlignment());
8635 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
8636 DAG.getConstant(8, DL, MVT::i64));
8637 return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr,
8638 S->getPointerInfo(), S->isVolatile(), S->isNonTemporal(),
8642 /// Target-specific DAG combine function for post-increment LD1 (lane) and
8643 /// post-increment LD1R.
8644 static SDValue performPostLD1Combine(SDNode *N,
8645 TargetLowering::DAGCombinerInfo &DCI,
8647 if (DCI.isBeforeLegalizeOps())
8650 SelectionDAG &DAG = DCI.DAG;
8651 EVT VT = N->getValueType(0);
8653 unsigned LoadIdx = IsLaneOp ? 1 : 0;
8654 SDNode *LD = N->getOperand(LoadIdx).getNode();
8655 // If it is not LOAD, can not do such combine.
8656 if (LD->getOpcode() != ISD::LOAD)
8659 LoadSDNode *LoadSDN = cast<LoadSDNode>(LD);
8660 EVT MemVT = LoadSDN->getMemoryVT();
8661 // Check if memory operand is the same type as the vector element.
8662 if (MemVT != VT.getVectorElementType())
8665 // Check if there are other uses. If so, do not combine as it will introduce
8667 for (SDNode::use_iterator UI = LD->use_begin(), UE = LD->use_end(); UI != UE;
8669 if (UI.getUse().getResNo() == 1) // Ignore uses of the chain result.
8675 SDValue Addr = LD->getOperand(1);
8676 SDValue Vector = N->getOperand(0);
8677 // Search for a use of the address operand that is an increment.
8678 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(), UE =
8679 Addr.getNode()->use_end(); UI != UE; ++UI) {
8681 if (User->getOpcode() != ISD::ADD
8682 || UI.getUse().getResNo() != Addr.getResNo())
8685 // Check that the add is independent of the load. Otherwise, folding it
8686 // would create a cycle.
8687 if (User->isPredecessorOf(LD) || LD->isPredecessorOf(User))
8689 // Also check that add is not used in the vector operand. This would also
8691 if (User->isPredecessorOf(Vector.getNode()))
8694 // If the increment is a constant, it must match the memory ref size.
8695 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8696 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8697 uint32_t IncVal = CInc->getZExtValue();
8698 unsigned NumBytes = VT.getScalarSizeInBits() / 8;
8699 if (IncVal != NumBytes)
8701 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
8704 // Finally, check that the vector doesn't depend on the load.
8705 // Again, this would create a cycle.
8706 // The load depending on the vector is fine, as that's the case for the
8707 // LD1*post we'll eventually generate anyway.
8708 if (LoadSDN->isPredecessorOf(Vector.getNode()))
8711 SmallVector<SDValue, 8> Ops;
8712 Ops.push_back(LD->getOperand(0)); // Chain
8714 Ops.push_back(Vector); // The vector to be inserted
8715 Ops.push_back(N->getOperand(2)); // The lane to be inserted in the vector
8717 Ops.push_back(Addr);
8720 EVT Tys[3] = { VT, MVT::i64, MVT::Other };
8721 SDVTList SDTys = DAG.getVTList(Tys);
8722 unsigned NewOp = IsLaneOp ? AArch64ISD::LD1LANEpost : AArch64ISD::LD1DUPpost;
8723 SDValue UpdN = DAG.getMemIntrinsicNode(NewOp, SDLoc(N), SDTys, Ops,
8725 LoadSDN->getMemOperand());
8728 SmallVector<SDValue, 2> NewResults;
8729 NewResults.push_back(SDValue(LD, 0)); // The result of load
8730 NewResults.push_back(SDValue(UpdN.getNode(), 2)); // Chain
8731 DCI.CombineTo(LD, NewResults);
8732 DCI.CombineTo(N, SDValue(UpdN.getNode(), 0)); // Dup/Inserted Result
8733 DCI.CombineTo(User, SDValue(UpdN.getNode(), 1)); // Write back register
8740 /// This function handles the log2-shuffle pattern produced by the
8741 /// LoopVectorizer for the across vector reduction. It consists of
8742 /// log2(NumVectorElements) steps and, in each step, 2^(s) elements
8743 /// are reduced, where s is an induction variable from 0 to
8744 /// log2(NumVectorElements).
8745 static SDValue tryMatchAcrossLaneShuffleForReduction(SDNode *N, SDValue OpV,
8747 SelectionDAG &DAG) {
8748 EVT VTy = OpV->getOperand(0).getValueType();
8749 if (!VTy.isVector())
8752 int NumVecElts = VTy.getVectorNumElements();
8753 if (NumVecElts != 4 && NumVecElts != 8 && NumVecElts != 16)
8756 int NumExpectedSteps = APInt(8, NumVecElts).logBase2();
8757 SDValue PreOp = OpV;
8758 // Iterate over each step of the across vector reduction.
8759 for (int CurStep = 0; CurStep != NumExpectedSteps; ++CurStep) {
8760 SDValue CurOp = PreOp.getOperand(0);
8761 SDValue Shuffle = PreOp.getOperand(1);
8762 if (Shuffle.getOpcode() != ISD::VECTOR_SHUFFLE) {
8763 // Try to swap the 1st and 2nd operand as add and min/max instructions
8765 CurOp = PreOp.getOperand(1);
8766 Shuffle = PreOp.getOperand(0);
8767 if (Shuffle.getOpcode() != ISD::VECTOR_SHUFFLE)
8771 // Check if the input vector is fed by the operator we want to handle,
8772 // except the last step; the very first input vector is not necessarily
8773 // the same operator we are handling.
8774 if (CurOp.getOpcode() != Op && (CurStep != (NumExpectedSteps - 1)))
8777 // Check if it forms one step of the across vector reduction.
8779 // %cur = add %1, %0
8780 // %shuffle = vector_shuffle %cur, <2, 3, u, u>
8781 // %pre = add %cur, %shuffle
8782 if (Shuffle.getOperand(0) != CurOp)
8785 int NumMaskElts = 1 << CurStep;
8786 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Shuffle)->getMask();
8787 // Check mask values in each step.
8788 // We expect the shuffle mask in each step follows a specific pattern
8789 // denoted here by the <M, U> form, where M is a sequence of integers
8790 // starting from NumMaskElts, increasing by 1, and the number integers
8791 // in M should be NumMaskElts. U is a sequence of UNDEFs and the number
8792 // of undef in U should be NumVecElts - NumMaskElts.
8793 // E.g., for <8 x i16>, mask values in each step should be :
8794 // step 0 : <1,u,u,u,u,u,u,u>
8795 // step 1 : <2,3,u,u,u,u,u,u>
8796 // step 2 : <4,5,6,7,u,u,u,u>
8797 for (int i = 0; i < NumVecElts; ++i)
8798 if ((i < NumMaskElts && Mask[i] != (NumMaskElts + i)) ||
8799 (i >= NumMaskElts && !(Mask[i] < 0)))
8807 llvm_unreachable("Unexpected operator for across vector reduction");
8809 Opcode = AArch64ISD::UADDV;
8812 Opcode = AArch64ISD::SMAXV;
8815 Opcode = AArch64ISD::UMAXV;
8818 Opcode = AArch64ISD::SMINV;
8821 Opcode = AArch64ISD::UMINV;
8825 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, N->getValueType(0),
8826 DAG.getNode(Opcode, DL, PreOp.getSimpleValueType(), PreOp),
8827 DAG.getConstant(0, DL, MVT::i64));
8830 /// Target-specific DAG combine for the across vector min/max reductions.
8831 /// This function specifically handles the final clean-up step of the vector
8832 /// min/max reductions produced by the LoopVectorizer. It is the log2-shuffle
8833 /// pattern, which narrows down and finds the final min/max value from all
8834 /// elements of the vector.
8835 /// For example, for a <16 x i8> vector :
8836 /// svn0 = vector_shuffle %0, undef<8,9,10,11,12,13,14,15,u,u,u,u,u,u,u,u>
8837 /// %smax0 = smax %arr, svn0
8838 /// %svn1 = vector_shuffle %smax0, undef<4,5,6,7,u,u,u,u,u,u,u,u,u,u,u,u>
8839 /// %smax1 = smax %smax0, %svn1
8840 /// %svn2 = vector_shuffle %smax1, undef<2,3,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
8841 /// %smax2 = smax %smax1, svn2
8842 /// %svn3 = vector_shuffle %smax2, undef<1,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
8843 /// %sc = setcc %smax2, %svn3, gt
8844 /// %n0 = extract_vector_elt %sc, #0
8845 /// %n1 = extract_vector_elt %smax2, #0
8846 /// %n2 = extract_vector_elt $smax2, #1
8847 /// %result = select %n0, %n1, n2
8850 /// %result = extract_vector_elt %1, 0
8851 /// FIXME: Currently this function matches only SMAXV, UMAXV, SMINV, and UMINV.
8852 /// We could also support other types of across lane reduction available
8853 /// in AArch64, including FMAXNMV, FMAXV, FMINNMV, and FMINV.
8855 performAcrossLaneMinMaxReductionCombine(SDNode *N, SelectionDAG &DAG,
8856 const AArch64Subtarget *Subtarget) {
8857 if (!Subtarget->hasNEON())
8860 SDValue N0 = N->getOperand(0);
8861 SDValue IfTrue = N->getOperand(1);
8862 SDValue IfFalse = N->getOperand(2);
8864 // Check if the SELECT merges up the final result of the min/max
8866 if (N0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
8867 IfTrue.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
8868 IfFalse.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8871 // Expect N0 is fed by SETCC.
8872 SDValue SetCC = N0.getOperand(0);
8873 EVT SetCCVT = SetCC.getValueType();
8874 if (SetCC.getOpcode() != ISD::SETCC || !SetCCVT.isVector() ||
8875 SetCCVT.getVectorElementType() != MVT::i1)
8878 SDValue VectorOp = SetCC.getOperand(0);
8879 unsigned Op = VectorOp->getOpcode();
8880 // Check if the input vector is fed by the operator we want to handle.
8881 if (Op != ISD::SMAX && Op != ISD::UMAX && Op != ISD::SMIN && Op != ISD::UMIN)
8884 EVT VTy = VectorOp.getValueType();
8885 if (!VTy.isVector())
8888 EVT EltTy = VTy.getVectorElementType();
8889 if (EltTy != MVT::i32 && EltTy != MVT::i16 && EltTy != MVT::i8)
8892 // Check if extracting from the same vector.
8894 // %sc = setcc %vector, %svn1, gt
8895 // %n0 = extract_vector_elt %sc, #0
8896 // %n1 = extract_vector_elt %vector, #0
8897 // %n2 = extract_vector_elt $vector, #1
8898 if (!(VectorOp == IfTrue->getOperand(0) &&
8899 VectorOp == IfFalse->getOperand(0)))
8902 // Check if the condition code is matched with the operator type.
8903 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC->getOperand(2))->get();
8904 if ((Op == ISD::SMAX && CC != ISD::SETGT && CC != ISD::SETGE) ||
8905 (Op == ISD::UMAX && CC != ISD::SETUGT && CC != ISD::SETUGE) ||
8906 (Op == ISD::SMIN && CC != ISD::SETLT && CC != ISD::SETLE) ||
8907 (Op == ISD::UMIN && CC != ISD::SETULT && CC != ISD::SETULE))
8910 // Expect to check only lane 0 from the vector SETCC.
8911 if (!isa<ConstantSDNode>(N0.getOperand(1)) ||
8912 cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue() != 0)
8915 // Expect to extract the true value from lane 0.
8916 if (!isa<ConstantSDNode>(IfTrue.getOperand(1)) ||
8917 cast<ConstantSDNode>(IfTrue.getOperand(1))->getZExtValue() != 0)
8920 // Expect to extract the false value from lane 1.
8921 if (!isa<ConstantSDNode>(IfFalse.getOperand(1)) ||
8922 cast<ConstantSDNode>(IfFalse.getOperand(1))->getZExtValue() != 1)
8925 return tryMatchAcrossLaneShuffleForReduction(N, SetCC, Op, DAG);
8928 /// Target-specific DAG combine for the across vector add reduction.
8929 /// This function specifically handles the final clean-up step of the vector
8930 /// add reduction produced by the LoopVectorizer. It is the log2-shuffle
8931 /// pattern, which adds all elements of a vector together.
8932 /// For example, for a <4 x i32> vector :
8933 /// %1 = vector_shuffle %0, <2,3,u,u>
8935 /// %3 = vector_shuffle %2, <1,u,u,u>
8937 /// %result = extract_vector_elt %4, 0
8940 /// %result = extract_vector_elt %0, 0
8942 performAcrossLaneAddReductionCombine(SDNode *N, SelectionDAG &DAG,
8943 const AArch64Subtarget *Subtarget) {
8944 if (!Subtarget->hasNEON())
8946 SDValue N0 = N->getOperand(0);
8947 SDValue N1 = N->getOperand(1);
8949 // Check if the input vector is fed by the ADD.
8950 if (N0->getOpcode() != ISD::ADD)
8953 // The vector extract idx must constant zero because we only expect the final
8954 // result of the reduction is placed in lane 0.
8955 if (!isa<ConstantSDNode>(N1) || cast<ConstantSDNode>(N1)->getZExtValue() != 0)
8958 EVT VTy = N0.getValueType();
8959 if (!VTy.isVector())
8962 EVT EltTy = VTy.getVectorElementType();
8963 if (EltTy != MVT::i32 && EltTy != MVT::i16 && EltTy != MVT::i8)
8966 return tryMatchAcrossLaneShuffleForReduction(N, N0, ISD::ADD, DAG);
8969 /// Target-specific DAG combine function for NEON load/store intrinsics
8970 /// to merge base address updates.
8971 static SDValue performNEONPostLDSTCombine(SDNode *N,
8972 TargetLowering::DAGCombinerInfo &DCI,
8973 SelectionDAG &DAG) {
8974 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8977 unsigned AddrOpIdx = N->getNumOperands() - 1;
8978 SDValue Addr = N->getOperand(AddrOpIdx);
8980 // Search for a use of the address operand that is an increment.
8981 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8982 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8984 if (User->getOpcode() != ISD::ADD ||
8985 UI.getUse().getResNo() != Addr.getResNo())
8988 // Check that the add is independent of the load/store. Otherwise, folding
8989 // it would create a cycle.
8990 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8993 // Find the new opcode for the updating load/store.
8994 bool IsStore = false;
8995 bool IsLaneOp = false;
8996 bool IsDupOp = false;
8997 unsigned NewOpc = 0;
8998 unsigned NumVecs = 0;
8999 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
9001 default: llvm_unreachable("unexpected intrinsic for Neon base update");
9002 case Intrinsic::aarch64_neon_ld2: NewOpc = AArch64ISD::LD2post;
9004 case Intrinsic::aarch64_neon_ld3: NewOpc = AArch64ISD::LD3post;
9006 case Intrinsic::aarch64_neon_ld4: NewOpc = AArch64ISD::LD4post;
9008 case Intrinsic::aarch64_neon_st2: NewOpc = AArch64ISD::ST2post;
9009 NumVecs = 2; IsStore = true; break;
9010 case Intrinsic::aarch64_neon_st3: NewOpc = AArch64ISD::ST3post;
9011 NumVecs = 3; IsStore = true; break;
9012 case Intrinsic::aarch64_neon_st4: NewOpc = AArch64ISD::ST4post;
9013 NumVecs = 4; IsStore = true; break;
9014 case Intrinsic::aarch64_neon_ld1x2: NewOpc = AArch64ISD::LD1x2post;
9016 case Intrinsic::aarch64_neon_ld1x3: NewOpc = AArch64ISD::LD1x3post;
9018 case Intrinsic::aarch64_neon_ld1x4: NewOpc = AArch64ISD::LD1x4post;
9020 case Intrinsic::aarch64_neon_st1x2: NewOpc = AArch64ISD::ST1x2post;
9021 NumVecs = 2; IsStore = true; break;
9022 case Intrinsic::aarch64_neon_st1x3: NewOpc = AArch64ISD::ST1x3post;
9023 NumVecs = 3; IsStore = true; break;
9024 case Intrinsic::aarch64_neon_st1x4: NewOpc = AArch64ISD::ST1x4post;
9025 NumVecs = 4; IsStore = true; break;
9026 case Intrinsic::aarch64_neon_ld2r: NewOpc = AArch64ISD::LD2DUPpost;
9027 NumVecs = 2; IsDupOp = true; break;
9028 case Intrinsic::aarch64_neon_ld3r: NewOpc = AArch64ISD::LD3DUPpost;
9029 NumVecs = 3; IsDupOp = true; break;
9030 case Intrinsic::aarch64_neon_ld4r: NewOpc = AArch64ISD::LD4DUPpost;
9031 NumVecs = 4; IsDupOp = true; break;
9032 case Intrinsic::aarch64_neon_ld2lane: NewOpc = AArch64ISD::LD2LANEpost;
9033 NumVecs = 2; IsLaneOp = true; break;
9034 case Intrinsic::aarch64_neon_ld3lane: NewOpc = AArch64ISD::LD3LANEpost;
9035 NumVecs = 3; IsLaneOp = true; break;
9036 case Intrinsic::aarch64_neon_ld4lane: NewOpc = AArch64ISD::LD4LANEpost;
9037 NumVecs = 4; IsLaneOp = true; break;
9038 case Intrinsic::aarch64_neon_st2lane: NewOpc = AArch64ISD::ST2LANEpost;
9039 NumVecs = 2; IsStore = true; IsLaneOp = true; break;
9040 case Intrinsic::aarch64_neon_st3lane: NewOpc = AArch64ISD::ST3LANEpost;
9041 NumVecs = 3; IsStore = true; IsLaneOp = true; break;
9042 case Intrinsic::aarch64_neon_st4lane: NewOpc = AArch64ISD::ST4LANEpost;
9043 NumVecs = 4; IsStore = true; IsLaneOp = true; break;
9048 VecTy = N->getOperand(2).getValueType();
9050 VecTy = N->getValueType(0);
9052 // If the increment is a constant, it must match the memory ref size.
9053 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
9054 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
9055 uint32_t IncVal = CInc->getZExtValue();
9056 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
9057 if (IsLaneOp || IsDupOp)
9058 NumBytes /= VecTy.getVectorNumElements();
9059 if (IncVal != NumBytes)
9061 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
9063 SmallVector<SDValue, 8> Ops;
9064 Ops.push_back(N->getOperand(0)); // Incoming chain
9065 // Load lane and store have vector list as input.
9066 if (IsLaneOp || IsStore)
9067 for (unsigned i = 2; i < AddrOpIdx; ++i)
9068 Ops.push_back(N->getOperand(i));
9069 Ops.push_back(Addr); // Base register
9074 unsigned NumResultVecs = (IsStore ? 0 : NumVecs);
9076 for (n = 0; n < NumResultVecs; ++n)
9078 Tys[n++] = MVT::i64; // Type of write back register
9079 Tys[n] = MVT::Other; // Type of the chain
9080 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs + 2));
9082 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
9083 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys, Ops,
9084 MemInt->getMemoryVT(),
9085 MemInt->getMemOperand());
9088 std::vector<SDValue> NewResults;
9089 for (unsigned i = 0; i < NumResultVecs; ++i) {
9090 NewResults.push_back(SDValue(UpdN.getNode(), i));
9092 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs + 1));
9093 DCI.CombineTo(N, NewResults);
9094 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9101 // Checks to see if the value is the prescribed width and returns information
9102 // about its extension mode.
9104 bool checkValueWidth(SDValue V, unsigned width, ISD::LoadExtType &ExtType) {
9105 ExtType = ISD::NON_EXTLOAD;
9106 switch(V.getNode()->getOpcode()) {
9110 LoadSDNode *LoadNode = cast<LoadSDNode>(V.getNode());
9111 if ((LoadNode->getMemoryVT() == MVT::i8 && width == 8)
9112 || (LoadNode->getMemoryVT() == MVT::i16 && width == 16)) {
9113 ExtType = LoadNode->getExtensionType();
9118 case ISD::AssertSext: {
9119 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
9120 if ((TypeNode->getVT() == MVT::i8 && width == 8)
9121 || (TypeNode->getVT() == MVT::i16 && width == 16)) {
9122 ExtType = ISD::SEXTLOAD;
9127 case ISD::AssertZext: {
9128 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
9129 if ((TypeNode->getVT() == MVT::i8 && width == 8)
9130 || (TypeNode->getVT() == MVT::i16 && width == 16)) {
9131 ExtType = ISD::ZEXTLOAD;
9137 case ISD::TargetConstant: {
9138 if (std::abs(cast<ConstantSDNode>(V.getNode())->getSExtValue()) <
9148 // This function does a whole lot of voodoo to determine if the tests are
9149 // equivalent without and with a mask. Essentially what happens is that given a
9152 // +-------------+ +-------------+ +-------------+ +-------------+
9153 // | Input | | AddConstant | | CompConstant| | CC |
9154 // +-------------+ +-------------+ +-------------+ +-------------+
9156 // V V | +----------+
9157 // +-------------+ +----+ | |
9158 // | ADD | |0xff| | |
9159 // +-------------+ +----+ | |
9162 // +-------------+ | |
9164 // +-------------+ | |
9173 // The AND node may be safely removed for some combinations of inputs. In
9174 // particular we need to take into account the extension type of the Input,
9175 // the exact values of AddConstant, CompConstant, and CC, along with the nominal
9176 // width of the input (this can work for any width inputs, the above graph is
9177 // specific to 8 bits.
9179 // The specific equations were worked out by generating output tables for each
9180 // AArch64CC value in terms of and AddConstant (w1), CompConstant(w2). The
9181 // problem was simplified by working with 4 bit inputs, which means we only
9182 // needed to reason about 24 distinct bit patterns: 8 patterns unique to zero
9183 // extension (8,15), 8 patterns unique to sign extensions (-8,-1), and 8
9184 // patterns present in both extensions (0,7). For every distinct set of
9185 // AddConstant and CompConstants bit patterns we can consider the masked and
9186 // unmasked versions to be equivalent if the result of this function is true for
9187 // all 16 distinct bit patterns of for the current extension type of Input (w0).
9190 // and w10, w8, #0x0f
9192 // cset w9, AArch64CC
9194 // cset w11, AArch64CC
9199 // Since the above function shows when the outputs are equivalent it defines
9200 // when it is safe to remove the AND. Unfortunately it only runs on AArch64 and
9201 // would be expensive to run during compiles. The equations below were written
9202 // in a test harness that confirmed they gave equivalent outputs to the above
9203 // for all inputs function, so they can be used determine if the removal is
9206 // isEquivalentMaskless() is the code for testing if the AND can be removed
9207 // factored out of the DAG recognition as the DAG can take several forms.
9210 bool isEquivalentMaskless(unsigned CC, unsigned width,
9211 ISD::LoadExtType ExtType, signed AddConstant,
9212 signed CompConstant) {
9213 // By being careful about our equations and only writing the in term
9214 // symbolic values and well known constants (0, 1, -1, MaxUInt) we can
9215 // make them generally applicable to all bit widths.
9216 signed MaxUInt = (1 << width);
9218 // For the purposes of these comparisons sign extending the type is
9219 // equivalent to zero extending the add and displacing it by half the integer
9220 // width. Provided we are careful and make sure our equations are valid over
9221 // the whole range we can just adjust the input and avoid writing equations
9222 // for sign extended inputs.
9223 if (ExtType == ISD::SEXTLOAD)
9224 AddConstant -= (1 << (width-1));
9228 case AArch64CC::GT: {
9229 if ((AddConstant == 0) ||
9230 (CompConstant == MaxUInt - 1 && AddConstant < 0) ||
9231 (AddConstant >= 0 && CompConstant < 0) ||
9232 (AddConstant <= 0 && CompConstant <= 0 && CompConstant < AddConstant))
9236 case AArch64CC::GE: {
9237 if ((AddConstant == 0) ||
9238 (AddConstant >= 0 && CompConstant <= 0) ||
9239 (AddConstant <= 0 && CompConstant <= 0 && CompConstant <= AddConstant))
9243 case AArch64CC::LS: {
9244 if ((AddConstant >= 0 && CompConstant < 0) ||
9245 (AddConstant <= 0 && CompConstant >= -1 &&
9246 CompConstant < AddConstant + MaxUInt))
9250 case AArch64CC::MI: {
9251 if ((AddConstant == 0) ||
9252 (AddConstant > 0 && CompConstant <= 0) ||
9253 (AddConstant < 0 && CompConstant <= AddConstant))
9257 case AArch64CC::HS: {
9258 if ((AddConstant >= 0 && CompConstant <= 0) ||
9259 (AddConstant <= 0 && CompConstant >= 0 &&
9260 CompConstant <= AddConstant + MaxUInt))
9264 case AArch64CC::NE: {
9265 if ((AddConstant > 0 && CompConstant < 0) ||
9266 (AddConstant < 0 && CompConstant >= 0 &&
9267 CompConstant < AddConstant + MaxUInt) ||
9268 (AddConstant >= 0 && CompConstant >= 0 &&
9269 CompConstant >= AddConstant) ||
9270 (AddConstant <= 0 && CompConstant < 0 && CompConstant < AddConstant))
9279 case AArch64CC::Invalid:
9287 SDValue performCONDCombine(SDNode *N,
9288 TargetLowering::DAGCombinerInfo &DCI,
9289 SelectionDAG &DAG, unsigned CCIndex,
9290 unsigned CmpIndex) {
9291 unsigned CC = cast<ConstantSDNode>(N->getOperand(CCIndex))->getSExtValue();
9292 SDNode *SubsNode = N->getOperand(CmpIndex).getNode();
9293 unsigned CondOpcode = SubsNode->getOpcode();
9295 if (CondOpcode != AArch64ISD::SUBS)
9298 // There is a SUBS feeding this condition. Is it fed by a mask we can
9301 SDNode *AndNode = SubsNode->getOperand(0).getNode();
9302 unsigned MaskBits = 0;
9304 if (AndNode->getOpcode() != ISD::AND)
9307 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(AndNode->getOperand(1))) {
9308 uint32_t CNV = CN->getZExtValue();
9311 else if (CNV == 65535)
9318 SDValue AddValue = AndNode->getOperand(0);
9320 if (AddValue.getOpcode() != ISD::ADD)
9323 // The basic dag structure is correct, grab the inputs and validate them.
9325 SDValue AddInputValue1 = AddValue.getNode()->getOperand(0);
9326 SDValue AddInputValue2 = AddValue.getNode()->getOperand(1);
9327 SDValue SubsInputValue = SubsNode->getOperand(1);
9329 // The mask is present and the provenance of all the values is a smaller type,
9330 // lets see if the mask is superfluous.
9332 if (!isa<ConstantSDNode>(AddInputValue2.getNode()) ||
9333 !isa<ConstantSDNode>(SubsInputValue.getNode()))
9336 ISD::LoadExtType ExtType;
9338 if (!checkValueWidth(SubsInputValue, MaskBits, ExtType) ||
9339 !checkValueWidth(AddInputValue2, MaskBits, ExtType) ||
9340 !checkValueWidth(AddInputValue1, MaskBits, ExtType) )
9343 if(!isEquivalentMaskless(CC, MaskBits, ExtType,
9344 cast<ConstantSDNode>(AddInputValue2.getNode())->getSExtValue(),
9345 cast<ConstantSDNode>(SubsInputValue.getNode())->getSExtValue()))
9348 // The AND is not necessary, remove it.
9350 SDVTList VTs = DAG.getVTList(SubsNode->getValueType(0),
9351 SubsNode->getValueType(1));
9352 SDValue Ops[] = { AddValue, SubsNode->getOperand(1) };
9354 SDValue NewValue = DAG.getNode(CondOpcode, SDLoc(SubsNode), VTs, Ops);
9355 DAG.ReplaceAllUsesWith(SubsNode, NewValue.getNode());
9357 return SDValue(N, 0);
9360 // Optimize compare with zero and branch.
9361 static SDValue performBRCONDCombine(SDNode *N,
9362 TargetLowering::DAGCombinerInfo &DCI,
9363 SelectionDAG &DAG) {
9364 SDValue NV = performCONDCombine(N, DCI, DAG, 2, 3);
9367 SDValue Chain = N->getOperand(0);
9368 SDValue Dest = N->getOperand(1);
9369 SDValue CCVal = N->getOperand(2);
9370 SDValue Cmp = N->getOperand(3);
9372 assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!");
9373 unsigned CC = cast<ConstantSDNode>(CCVal)->getZExtValue();
9374 if (CC != AArch64CC::EQ && CC != AArch64CC::NE)
9377 unsigned CmpOpc = Cmp.getOpcode();
9378 if (CmpOpc != AArch64ISD::ADDS && CmpOpc != AArch64ISD::SUBS)
9381 // Only attempt folding if there is only one use of the flag and no use of the
9383 if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1))
9386 SDValue LHS = Cmp.getOperand(0);
9387 SDValue RHS = Cmp.getOperand(1);
9389 assert(LHS.getValueType() == RHS.getValueType() &&
9390 "Expected the value type to be the same for both operands!");
9391 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
9394 if (isa<ConstantSDNode>(LHS) && cast<ConstantSDNode>(LHS)->isNullValue())
9395 std::swap(LHS, RHS);
9397 if (!isa<ConstantSDNode>(RHS) || !cast<ConstantSDNode>(RHS)->isNullValue())
9400 if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA ||
9401 LHS.getOpcode() == ISD::SRL)
9404 // Fold the compare into the branch instruction.
9406 if (CC == AArch64CC::EQ)
9407 BR = DAG.getNode(AArch64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
9409 BR = DAG.getNode(AArch64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
9411 // Do not add new nodes to DAG combiner worklist.
9412 DCI.CombineTo(N, BR, false);
9417 // vselect (v1i1 setcc) ->
9418 // vselect (v1iXX setcc) (XX is the size of the compared operand type)
9419 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
9420 // condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
9422 static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) {
9423 SDValue N0 = N->getOperand(0);
9424 EVT CCVT = N0.getValueType();
9426 if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
9427 CCVT.getVectorElementType() != MVT::i1)
9430 EVT ResVT = N->getValueType(0);
9431 EVT CmpVT = N0.getOperand(0).getValueType();
9432 // Only combine when the result type is of the same size as the compared
9434 if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
9437 SDValue IfTrue = N->getOperand(1);
9438 SDValue IfFalse = N->getOperand(2);
9440 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
9441 N0.getOperand(0), N0.getOperand(1),
9442 cast<CondCodeSDNode>(N0.getOperand(2))->get());
9443 return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
9447 /// A vector select: "(select vL, vR, (setcc LHS, RHS))" is best performed with
9448 /// the compare-mask instructions rather than going via NZCV, even if LHS and
9449 /// RHS are really scalar. This replaces any scalar setcc in the above pattern
9450 /// with a vector one followed by a DUP shuffle on the result.
9451 static SDValue performSelectCombine(SDNode *N,
9452 TargetLowering::DAGCombinerInfo &DCI) {
9453 SelectionDAG &DAG = DCI.DAG;
9454 SDValue N0 = N->getOperand(0);
9455 EVT ResVT = N->getValueType(0);
9457 if (N0.getOpcode() != ISD::SETCC)
9460 // Make sure the SETCC result is either i1 (initial DAG), or i32, the lowered
9461 // scalar SetCCResultType. We also don't expect vectors, because we assume
9462 // that selects fed by vector SETCCs are canonicalized to VSELECT.
9463 assert((N0.getValueType() == MVT::i1 || N0.getValueType() == MVT::i32) &&
9464 "Scalar-SETCC feeding SELECT has unexpected result type!");
9466 // If NumMaskElts == 0, the comparison is larger than select result. The
9467 // largest real NEON comparison is 64-bits per lane, which means the result is
9468 // at most 32-bits and an illegal vector. Just bail out for now.
9469 EVT SrcVT = N0.getOperand(0).getValueType();
9471 // Don't try to do this optimization when the setcc itself has i1 operands.
9472 // There are no legal vectors of i1, so this would be pointless.
9473 if (SrcVT == MVT::i1)
9476 int NumMaskElts = ResVT.getSizeInBits() / SrcVT.getSizeInBits();
9477 if (!ResVT.isVector() || NumMaskElts == 0)
9480 SrcVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumMaskElts);
9481 EVT CCVT = SrcVT.changeVectorElementTypeToInteger();
9483 // Also bail out if the vector CCVT isn't the same size as ResVT.
9484 // This can happen if the SETCC operand size doesn't divide the ResVT size
9485 // (e.g., f64 vs v3f32).
9486 if (CCVT.getSizeInBits() != ResVT.getSizeInBits())
9489 // Make sure we didn't create illegal types, if we're not supposed to.
9490 assert(DCI.isBeforeLegalize() ||
9491 DAG.getTargetLoweringInfo().isTypeLegal(SrcVT));
9493 // First perform a vector comparison, where lane 0 is the one we're interested
9497 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(0));
9499 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(1));
9500 SDValue SetCC = DAG.getNode(ISD::SETCC, DL, CCVT, LHS, RHS, N0.getOperand(2));
9502 // Now duplicate the comparison mask we want across all other lanes.
9503 SmallVector<int, 8> DUPMask(CCVT.getVectorNumElements(), 0);
9504 SDValue Mask = DAG.getVectorShuffle(CCVT, DL, SetCC, SetCC, DUPMask.data());
9505 Mask = DAG.getNode(ISD::BITCAST, DL,
9506 ResVT.changeVectorElementTypeToInteger(), Mask);
9508 return DAG.getSelect(DL, ResVT, Mask, N->getOperand(1), N->getOperand(2));
9511 /// Get rid of unnecessary NVCASTs (that don't change the type).
9512 static SDValue performNVCASTCombine(SDNode *N) {
9513 if (N->getValueType(0) == N->getOperand(0).getValueType())
9514 return N->getOperand(0);
9519 SDValue AArch64TargetLowering::PerformDAGCombine(SDNode *N,
9520 DAGCombinerInfo &DCI) const {
9521 SelectionDAG &DAG = DCI.DAG;
9522 switch (N->getOpcode()) {
9527 return performAddSubLongCombine(N, DCI, DAG);
9529 return performXorCombine(N, DAG, DCI, Subtarget);
9531 return performMulCombine(N, DAG, DCI, Subtarget);
9532 case ISD::SINT_TO_FP:
9533 case ISD::UINT_TO_FP:
9534 return performIntToFpCombine(N, DAG, Subtarget);
9535 case ISD::FP_TO_SINT:
9536 case ISD::FP_TO_UINT:
9537 return performFpToIntCombine(N, DAG, Subtarget);
9539 return performFDivCombine(N, DAG, Subtarget);
9541 return performORCombine(N, DCI, Subtarget);
9542 case ISD::INTRINSIC_WO_CHAIN:
9543 return performIntrinsicCombine(N, DCI, Subtarget);
9544 case ISD::ANY_EXTEND:
9545 case ISD::ZERO_EXTEND:
9546 case ISD::SIGN_EXTEND:
9547 return performExtendCombine(N, DCI, DAG);
9549 return performBitcastCombine(N, DCI, DAG);
9550 case ISD::CONCAT_VECTORS:
9551 return performConcatVectorsCombine(N, DCI, DAG);
9553 SDValue RV = performSelectCombine(N, DCI);
9555 RV = performAcrossLaneMinMaxReductionCombine(N, DAG, Subtarget);
9559 return performVSelectCombine(N, DCI.DAG);
9561 return performSTORECombine(N, DCI, DAG, Subtarget);
9562 case AArch64ISD::BRCOND:
9563 return performBRCONDCombine(N, DCI, DAG);
9564 case AArch64ISD::CSEL:
9565 return performCONDCombine(N, DCI, DAG, 2, 3);
9566 case AArch64ISD::DUP:
9567 return performPostLD1Combine(N, DCI, false);
9568 case AArch64ISD::NVCAST:
9569 return performNVCASTCombine(N);
9570 case ISD::INSERT_VECTOR_ELT:
9571 return performPostLD1Combine(N, DCI, true);
9572 case ISD::EXTRACT_VECTOR_ELT:
9573 return performAcrossLaneAddReductionCombine(N, DAG, Subtarget);
9574 case ISD::INTRINSIC_VOID:
9575 case ISD::INTRINSIC_W_CHAIN:
9576 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9577 case Intrinsic::aarch64_neon_ld2:
9578 case Intrinsic::aarch64_neon_ld3:
9579 case Intrinsic::aarch64_neon_ld4:
9580 case Intrinsic::aarch64_neon_ld1x2:
9581 case Intrinsic::aarch64_neon_ld1x3:
9582 case Intrinsic::aarch64_neon_ld1x4:
9583 case Intrinsic::aarch64_neon_ld2lane:
9584 case Intrinsic::aarch64_neon_ld3lane:
9585 case Intrinsic::aarch64_neon_ld4lane:
9586 case Intrinsic::aarch64_neon_ld2r:
9587 case Intrinsic::aarch64_neon_ld3r:
9588 case Intrinsic::aarch64_neon_ld4r:
9589 case Intrinsic::aarch64_neon_st2:
9590 case Intrinsic::aarch64_neon_st3:
9591 case Intrinsic::aarch64_neon_st4:
9592 case Intrinsic::aarch64_neon_st1x2:
9593 case Intrinsic::aarch64_neon_st1x3:
9594 case Intrinsic::aarch64_neon_st1x4:
9595 case Intrinsic::aarch64_neon_st2lane:
9596 case Intrinsic::aarch64_neon_st3lane:
9597 case Intrinsic::aarch64_neon_st4lane:
9598 return performNEONPostLDSTCombine(N, DCI, DAG);
9606 // Check if the return value is used as only a return value, as otherwise
9607 // we can't perform a tail-call. In particular, we need to check for
9608 // target ISD nodes that are returns and any other "odd" constructs
9609 // that the generic analysis code won't necessarily catch.
9610 bool AArch64TargetLowering::isUsedByReturnOnly(SDNode *N,
9611 SDValue &Chain) const {
9612 if (N->getNumValues() != 1)
9614 if (!N->hasNUsesOfValue(1, 0))
9617 SDValue TCChain = Chain;
9618 SDNode *Copy = *N->use_begin();
9619 if (Copy->getOpcode() == ISD::CopyToReg) {
9620 // If the copy has a glue operand, we conservatively assume it isn't safe to
9621 // perform a tail call.
9622 if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() ==
9625 TCChain = Copy->getOperand(0);
9626 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
9629 bool HasRet = false;
9630 for (SDNode *Node : Copy->uses()) {
9631 if (Node->getOpcode() != AArch64ISD::RET_FLAG)
9643 // Return whether the an instruction can potentially be optimized to a tail
9644 // call. This will cause the optimizers to attempt to move, or duplicate,
9645 // return instructions to help enable tail call optimizations for this
9647 bool AArch64TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
9648 if (!CI->isTailCall())
9654 bool AArch64TargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base,
9656 ISD::MemIndexedMode &AM,
9658 SelectionDAG &DAG) const {
9659 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
9662 Base = Op->getOperand(0);
9663 // All of the indexed addressing mode instructions take a signed
9664 // 9 bit immediate offset.
9665 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
9666 int64_t RHSC = (int64_t)RHS->getZExtValue();
9667 if (RHSC >= 256 || RHSC <= -256)
9669 IsInc = (Op->getOpcode() == ISD::ADD);
9670 Offset = Op->getOperand(1);
9676 bool AArch64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
9678 ISD::MemIndexedMode &AM,
9679 SelectionDAG &DAG) const {
9682 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9683 VT = LD->getMemoryVT();
9684 Ptr = LD->getBasePtr();
9685 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9686 VT = ST->getMemoryVT();
9687 Ptr = ST->getBasePtr();
9692 if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, IsInc, DAG))
9694 AM = IsInc ? ISD::PRE_INC : ISD::PRE_DEC;
9698 bool AArch64TargetLowering::getPostIndexedAddressParts(
9699 SDNode *N, SDNode *Op, SDValue &Base, SDValue &Offset,
9700 ISD::MemIndexedMode &AM, SelectionDAG &DAG) const {
9703 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9704 VT = LD->getMemoryVT();
9705 Ptr = LD->getBasePtr();
9706 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9707 VT = ST->getMemoryVT();
9708 Ptr = ST->getBasePtr();
9713 if (!getIndexedAddressParts(Op, Base, Offset, AM, IsInc, DAG))
9715 // Post-indexing updates the base, so it's not a valid transform
9716 // if that's not the same as the load's pointer.
9719 AM = IsInc ? ISD::POST_INC : ISD::POST_DEC;
9723 static void ReplaceBITCASTResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
9724 SelectionDAG &DAG) {
9726 SDValue Op = N->getOperand(0);
9728 if (N->getValueType(0) != MVT::i16 || Op.getValueType() != MVT::f16)
9732 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, DL, MVT::f32,
9733 DAG.getUNDEF(MVT::i32), Op,
9734 DAG.getTargetConstant(AArch64::hsub, DL, MVT::i32)),
9736 Op = DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op);
9737 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, Op));
9740 void AArch64TargetLowering::ReplaceNodeResults(
9741 SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
9742 switch (N->getOpcode()) {
9744 llvm_unreachable("Don't know how to custom expand this");
9746 ReplaceBITCASTResults(N, Results, DAG);
9748 case ISD::FP_TO_UINT:
9749 case ISD::FP_TO_SINT:
9750 assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion");
9751 // Let normal code take care of it by not adding anything to Results.
9756 bool AArch64TargetLowering::useLoadStackGuardNode() const {
9760 unsigned AArch64TargetLowering::combineRepeatedFPDivisors() const {
9761 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
9762 // reciprocal if there are three or more FDIVs.
9766 TargetLoweringBase::LegalizeTypeAction
9767 AArch64TargetLowering::getPreferredVectorAction(EVT VT) const {
9768 MVT SVT = VT.getSimpleVT();
9769 // During type legalization, we prefer to widen v1i8, v1i16, v1i32 to v8i8,
9770 // v4i16, v2i32 instead of to promote.
9771 if (SVT == MVT::v1i8 || SVT == MVT::v1i16 || SVT == MVT::v1i32
9772 || SVT == MVT::v1f32)
9773 return TypeWidenVector;
9775 return TargetLoweringBase::getPreferredVectorAction(VT);
9778 // Loads and stores less than 128-bits are already atomic; ones above that
9779 // are doomed anyway, so defer to the default libcall and blame the OS when
9781 bool AArch64TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
9782 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
9786 // Loads and stores less than 128-bits are already atomic; ones above that
9787 // are doomed anyway, so defer to the default libcall and blame the OS when
9789 TargetLowering::AtomicExpansionKind
9790 AArch64TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
9791 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
9792 return Size == 128 ? AtomicExpansionKind::LLSC : AtomicExpansionKind::None;
9795 // For the real atomic operations, we have ldxr/stxr up to 128 bits,
9796 TargetLowering::AtomicExpansionKind
9797 AArch64TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
9798 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
9799 return Size <= 128 ? AtomicExpansionKind::LLSC : AtomicExpansionKind::None;
9802 bool AArch64TargetLowering::shouldExpandAtomicCmpXchgInIR(
9803 AtomicCmpXchgInst *AI) const {
9807 Value *AArch64TargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
9808 AtomicOrdering Ord) const {
9809 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
9810 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
9811 bool IsAcquire = isAtLeastAcquire(Ord);
9813 // Since i128 isn't legal and intrinsics don't get type-lowered, the ldrexd
9814 // intrinsic must return {i64, i64} and we have to recombine them into a
9815 // single i128 here.
9816 if (ValTy->getPrimitiveSizeInBits() == 128) {
9818 IsAcquire ? Intrinsic::aarch64_ldaxp : Intrinsic::aarch64_ldxp;
9819 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int);
9821 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
9822 Value *LoHi = Builder.CreateCall(Ldxr, Addr, "lohi");
9824 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
9825 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
9826 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
9827 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
9828 return Builder.CreateOr(
9829 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 64)), "val64");
9832 Type *Tys[] = { Addr->getType() };
9834 IsAcquire ? Intrinsic::aarch64_ldaxr : Intrinsic::aarch64_ldxr;
9835 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int, Tys);
9837 return Builder.CreateTruncOrBitCast(
9838 Builder.CreateCall(Ldxr, Addr),
9839 cast<PointerType>(Addr->getType())->getElementType());
9842 void AArch64TargetLowering::emitAtomicCmpXchgNoStoreLLBalance(
9843 IRBuilder<> &Builder) const {
9844 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
9846 llvm::Intrinsic::getDeclaration(M, Intrinsic::aarch64_clrex));
9849 Value *AArch64TargetLowering::emitStoreConditional(IRBuilder<> &Builder,
9850 Value *Val, Value *Addr,
9851 AtomicOrdering Ord) const {
9852 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
9853 bool IsRelease = isAtLeastRelease(Ord);
9855 // Since the intrinsics must have legal type, the i128 intrinsics take two
9856 // parameters: "i64, i64". We must marshal Val into the appropriate form
9858 if (Val->getType()->getPrimitiveSizeInBits() == 128) {
9860 IsRelease ? Intrinsic::aarch64_stlxp : Intrinsic::aarch64_stxp;
9861 Function *Stxr = Intrinsic::getDeclaration(M, Int);
9862 Type *Int64Ty = Type::getInt64Ty(M->getContext());
9864 Value *Lo = Builder.CreateTrunc(Val, Int64Ty, "lo");
9865 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 64), Int64Ty, "hi");
9866 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
9867 return Builder.CreateCall(Stxr, {Lo, Hi, Addr});
9871 IsRelease ? Intrinsic::aarch64_stlxr : Intrinsic::aarch64_stxr;
9872 Type *Tys[] = { Addr->getType() };
9873 Function *Stxr = Intrinsic::getDeclaration(M, Int, Tys);
9875 return Builder.CreateCall(Stxr,
9876 {Builder.CreateZExtOrBitCast(
9877 Val, Stxr->getFunctionType()->getParamType(0)),
9881 bool AArch64TargetLowering::functionArgumentNeedsConsecutiveRegisters(
9882 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
9883 return Ty->isArrayTy();
9886 bool AArch64TargetLowering::shouldNormalizeToSelectSequence(LLVMContext &,