1 //===-- AArch64ISelLowering.cpp - AArch64 DAG Lowering Implementation ----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the AArch64TargetLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "AArch64ISelLowering.h"
15 #include "AArch64CallingConvention.h"
16 #include "AArch64MachineFunctionInfo.h"
17 #include "AArch64PerfectShuffle.h"
18 #include "AArch64Subtarget.h"
19 #include "AArch64TargetMachine.h"
20 #include "AArch64TargetObjectFile.h"
21 #include "MCTargetDesc/AArch64AddressingModes.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/CodeGen/CallingConvLower.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/GetElementPtrTypeIterator.h"
29 #include "llvm/IR/Intrinsics.h"
30 #include "llvm/IR/Type.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Target/TargetOptions.h"
38 #define DEBUG_TYPE "aarch64-lower"
40 STATISTIC(NumTailCalls, "Number of tail calls");
41 STATISTIC(NumShiftInserts, "Number of vector shift inserts");
43 // Place holder until extr generation is tested fully.
45 EnableAArch64ExtrGeneration("aarch64-extr-generation", cl::Hidden,
46 cl::desc("Allow AArch64 (or (shift)(shift))->extract"),
50 EnableAArch64SlrGeneration("aarch64-shift-insert-generation", cl::Hidden,
51 cl::desc("Allow AArch64 SLI/SRI formation"),
54 // FIXME: The necessary dtprel relocations don't seem to be supported
55 // well in the GNU bfd and gold linkers at the moment. Therefore, by
56 // default, for now, fall back to GeneralDynamic code generation.
57 cl::opt<bool> EnableAArch64ELFLocalDynamicTLSGeneration(
58 "aarch64-elf-ldtls-generation", cl::Hidden,
59 cl::desc("Allow AArch64 Local Dynamic TLS code generation"),
62 /// Value type used for condition codes.
63 static const MVT MVT_CC = MVT::i32;
65 AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
66 const AArch64Subtarget &STI)
67 : TargetLowering(TM), Subtarget(&STI) {
69 // AArch64 doesn't have comparisons which set GPRs or setcc instructions, so
70 // we have to make something up. Arbitrarily, choose ZeroOrOne.
71 setBooleanContents(ZeroOrOneBooleanContent);
72 // When comparing vectors the result sets the different elements in the
73 // vector to all-one or all-zero.
74 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
76 // Set up the register classes.
77 addRegisterClass(MVT::i32, &AArch64::GPR32allRegClass);
78 addRegisterClass(MVT::i64, &AArch64::GPR64allRegClass);
80 if (Subtarget->hasFPARMv8()) {
81 addRegisterClass(MVT::f16, &AArch64::FPR16RegClass);
82 addRegisterClass(MVT::f32, &AArch64::FPR32RegClass);
83 addRegisterClass(MVT::f64, &AArch64::FPR64RegClass);
84 addRegisterClass(MVT::f128, &AArch64::FPR128RegClass);
87 if (Subtarget->hasNEON()) {
88 addRegisterClass(MVT::v16i8, &AArch64::FPR8RegClass);
89 addRegisterClass(MVT::v8i16, &AArch64::FPR16RegClass);
90 // Someone set us up the NEON.
91 addDRTypeForNEON(MVT::v2f32);
92 addDRTypeForNEON(MVT::v8i8);
93 addDRTypeForNEON(MVT::v4i16);
94 addDRTypeForNEON(MVT::v2i32);
95 addDRTypeForNEON(MVT::v1i64);
96 addDRTypeForNEON(MVT::v1f64);
97 addDRTypeForNEON(MVT::v4f16);
99 addQRTypeForNEON(MVT::v4f32);
100 addQRTypeForNEON(MVT::v2f64);
101 addQRTypeForNEON(MVT::v16i8);
102 addQRTypeForNEON(MVT::v8i16);
103 addQRTypeForNEON(MVT::v4i32);
104 addQRTypeForNEON(MVT::v2i64);
105 addQRTypeForNEON(MVT::v8f16);
108 // Compute derived properties from the register classes
109 computeRegisterProperties(Subtarget->getRegisterInfo());
111 // Provide all sorts of operation actions
112 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
113 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
114 setOperationAction(ISD::SETCC, MVT::i32, Custom);
115 setOperationAction(ISD::SETCC, MVT::i64, Custom);
116 setOperationAction(ISD::SETCC, MVT::f32, Custom);
117 setOperationAction(ISD::SETCC, MVT::f64, Custom);
118 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
119 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
120 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
121 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
122 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
123 setOperationAction(ISD::SELECT, MVT::i32, Custom);
124 setOperationAction(ISD::SELECT, MVT::i64, Custom);
125 setOperationAction(ISD::SELECT, MVT::f32, Custom);
126 setOperationAction(ISD::SELECT, MVT::f64, Custom);
127 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
128 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
129 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
130 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
131 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
132 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
134 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
135 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
136 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
138 setOperationAction(ISD::FREM, MVT::f32, Expand);
139 setOperationAction(ISD::FREM, MVT::f64, Expand);
140 setOperationAction(ISD::FREM, MVT::f80, Expand);
142 // Custom lowering hooks are needed for XOR
143 // to fold it into CSINC/CSINV.
144 setOperationAction(ISD::XOR, MVT::i32, Custom);
145 setOperationAction(ISD::XOR, MVT::i64, Custom);
147 // Virtually no operation on f128 is legal, but LLVM can't expand them when
148 // there's a valid register class, so we need custom operations in most cases.
149 setOperationAction(ISD::FABS, MVT::f128, Expand);
150 setOperationAction(ISD::FADD, MVT::f128, Custom);
151 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
152 setOperationAction(ISD::FCOS, MVT::f128, Expand);
153 setOperationAction(ISD::FDIV, MVT::f128, Custom);
154 setOperationAction(ISD::FMA, MVT::f128, Expand);
155 setOperationAction(ISD::FMUL, MVT::f128, Custom);
156 setOperationAction(ISD::FNEG, MVT::f128, Expand);
157 setOperationAction(ISD::FPOW, MVT::f128, Expand);
158 setOperationAction(ISD::FREM, MVT::f128, Expand);
159 setOperationAction(ISD::FRINT, MVT::f128, Expand);
160 setOperationAction(ISD::FSIN, MVT::f128, Expand);
161 setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
162 setOperationAction(ISD::FSQRT, MVT::f128, Expand);
163 setOperationAction(ISD::FSUB, MVT::f128, Custom);
164 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
165 setOperationAction(ISD::SETCC, MVT::f128, Custom);
166 setOperationAction(ISD::BR_CC, MVT::f128, Custom);
167 setOperationAction(ISD::SELECT, MVT::f128, Custom);
168 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
169 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);
171 // Lowering for many of the conversions is actually specified by the non-f128
172 // type. The LowerXXX function will be trivial when f128 isn't involved.
173 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
174 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
175 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
176 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
177 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
178 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
179 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
180 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
181 setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
182 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
183 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
184 setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
185 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
186 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
188 // Variable arguments.
189 setOperationAction(ISD::VASTART, MVT::Other, Custom);
190 setOperationAction(ISD::VAARG, MVT::Other, Custom);
191 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
192 setOperationAction(ISD::VAEND, MVT::Other, Expand);
194 // Variable-sized objects.
195 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
196 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
197 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
199 // Exception handling.
200 // FIXME: These are guesses. Has this been defined yet?
201 setExceptionPointerRegister(AArch64::X0);
202 setExceptionSelectorRegister(AArch64::X1);
204 // Constant pool entries
205 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
208 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
210 // Add/Sub overflow ops with MVT::Glues are lowered to NZCV dependences.
211 setOperationAction(ISD::ADDC, MVT::i32, Custom);
212 setOperationAction(ISD::ADDE, MVT::i32, Custom);
213 setOperationAction(ISD::SUBC, MVT::i32, Custom);
214 setOperationAction(ISD::SUBE, MVT::i32, Custom);
215 setOperationAction(ISD::ADDC, MVT::i64, Custom);
216 setOperationAction(ISD::ADDE, MVT::i64, Custom);
217 setOperationAction(ISD::SUBC, MVT::i64, Custom);
218 setOperationAction(ISD::SUBE, MVT::i64, Custom);
220 // AArch64 lacks both left-rotate and popcount instructions.
221 setOperationAction(ISD::ROTL, MVT::i32, Expand);
222 setOperationAction(ISD::ROTL, MVT::i64, Expand);
224 // AArch64 doesn't have {U|S}MUL_LOHI.
225 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
226 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
229 // Expand the undefined-at-zero variants to cttz/ctlz to their defined-at-zero
230 // counterparts, which AArch64 supports directly.
231 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
232 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
233 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
234 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
236 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
237 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
239 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
240 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
241 setOperationAction(ISD::SREM, MVT::i32, Expand);
242 setOperationAction(ISD::SREM, MVT::i64, Expand);
243 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
244 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
245 setOperationAction(ISD::UREM, MVT::i32, Expand);
246 setOperationAction(ISD::UREM, MVT::i64, Expand);
248 // Custom lower Add/Sub/Mul with overflow.
249 setOperationAction(ISD::SADDO, MVT::i32, Custom);
250 setOperationAction(ISD::SADDO, MVT::i64, Custom);
251 setOperationAction(ISD::UADDO, MVT::i32, Custom);
252 setOperationAction(ISD::UADDO, MVT::i64, Custom);
253 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
254 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
255 setOperationAction(ISD::USUBO, MVT::i32, Custom);
256 setOperationAction(ISD::USUBO, MVT::i64, Custom);
257 setOperationAction(ISD::SMULO, MVT::i32, Custom);
258 setOperationAction(ISD::SMULO, MVT::i64, Custom);
259 setOperationAction(ISD::UMULO, MVT::i32, Custom);
260 setOperationAction(ISD::UMULO, MVT::i64, Custom);
262 setOperationAction(ISD::FSIN, MVT::f32, Expand);
263 setOperationAction(ISD::FSIN, MVT::f64, Expand);
264 setOperationAction(ISD::FCOS, MVT::f32, Expand);
265 setOperationAction(ISD::FCOS, MVT::f64, Expand);
266 setOperationAction(ISD::FPOW, MVT::f32, Expand);
267 setOperationAction(ISD::FPOW, MVT::f64, Expand);
268 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
269 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
271 // f16 is a storage-only type, always promote it to f32.
272 setOperationAction(ISD::SETCC, MVT::f16, Promote);
273 setOperationAction(ISD::BR_CC, MVT::f16, Promote);
274 setOperationAction(ISD::SELECT_CC, MVT::f16, Promote);
275 setOperationAction(ISD::SELECT, MVT::f16, Promote);
276 setOperationAction(ISD::FADD, MVT::f16, Promote);
277 setOperationAction(ISD::FSUB, MVT::f16, Promote);
278 setOperationAction(ISD::FMUL, MVT::f16, Promote);
279 setOperationAction(ISD::FDIV, MVT::f16, Promote);
280 setOperationAction(ISD::FREM, MVT::f16, Promote);
281 setOperationAction(ISD::FMA, MVT::f16, Promote);
282 setOperationAction(ISD::FNEG, MVT::f16, Promote);
283 setOperationAction(ISD::FABS, MVT::f16, Promote);
284 setOperationAction(ISD::FCEIL, MVT::f16, Promote);
285 setOperationAction(ISD::FCOPYSIGN, MVT::f16, Promote);
286 setOperationAction(ISD::FCOS, MVT::f16, Promote);
287 setOperationAction(ISD::FFLOOR, MVT::f16, Promote);
288 setOperationAction(ISD::FNEARBYINT, MVT::f16, Promote);
289 setOperationAction(ISD::FPOW, MVT::f16, Promote);
290 setOperationAction(ISD::FPOWI, MVT::f16, Promote);
291 setOperationAction(ISD::FRINT, MVT::f16, Promote);
292 setOperationAction(ISD::FSIN, MVT::f16, Promote);
293 setOperationAction(ISD::FSINCOS, MVT::f16, Promote);
294 setOperationAction(ISD::FSQRT, MVT::f16, Promote);
295 setOperationAction(ISD::FEXP, MVT::f16, Promote);
296 setOperationAction(ISD::FEXP2, MVT::f16, Promote);
297 setOperationAction(ISD::FLOG, MVT::f16, Promote);
298 setOperationAction(ISD::FLOG2, MVT::f16, Promote);
299 setOperationAction(ISD::FLOG10, MVT::f16, Promote);
300 setOperationAction(ISD::FROUND, MVT::f16, Promote);
301 setOperationAction(ISD::FTRUNC, MVT::f16, Promote);
302 setOperationAction(ISD::FMINNUM, MVT::f16, Promote);
303 setOperationAction(ISD::FMAXNUM, MVT::f16, Promote);
305 // v4f16 is also a storage-only type, so promote it to v4f32 when that is
307 setOperationAction(ISD::FADD, MVT::v4f16, Promote);
308 setOperationAction(ISD::FSUB, MVT::v4f16, Promote);
309 setOperationAction(ISD::FMUL, MVT::v4f16, Promote);
310 setOperationAction(ISD::FDIV, MVT::v4f16, Promote);
311 setOperationAction(ISD::FP_EXTEND, MVT::v4f16, Promote);
312 setOperationAction(ISD::FP_ROUND, MVT::v4f16, Promote);
313 AddPromotedToType(ISD::FADD, MVT::v4f16, MVT::v4f32);
314 AddPromotedToType(ISD::FSUB, MVT::v4f16, MVT::v4f32);
315 AddPromotedToType(ISD::FMUL, MVT::v4f16, MVT::v4f32);
316 AddPromotedToType(ISD::FDIV, MVT::v4f16, MVT::v4f32);
317 AddPromotedToType(ISD::FP_EXTEND, MVT::v4f16, MVT::v4f32);
318 AddPromotedToType(ISD::FP_ROUND, MVT::v4f16, MVT::v4f32);
320 // Expand all other v4f16 operations.
321 // FIXME: We could generate better code by promoting some operations to
323 setOperationAction(ISD::FABS, MVT::v4f16, Expand);
324 setOperationAction(ISD::FCEIL, MVT::v4f16, Expand);
325 setOperationAction(ISD::FCOPYSIGN, MVT::v4f16, Expand);
326 setOperationAction(ISD::FCOS, MVT::v4f16, Expand);
327 setOperationAction(ISD::FFLOOR, MVT::v4f16, Expand);
328 setOperationAction(ISD::FMA, MVT::v4f16, Expand);
329 setOperationAction(ISD::FNEARBYINT, MVT::v4f16, Expand);
330 setOperationAction(ISD::FNEG, MVT::v4f16, Expand);
331 setOperationAction(ISD::FPOW, MVT::v4f16, Expand);
332 setOperationAction(ISD::FPOWI, MVT::v4f16, Expand);
333 setOperationAction(ISD::FREM, MVT::v4f16, Expand);
334 setOperationAction(ISD::FROUND, MVT::v4f16, Expand);
335 setOperationAction(ISD::FRINT, MVT::v4f16, Expand);
336 setOperationAction(ISD::FSIN, MVT::v4f16, Expand);
337 setOperationAction(ISD::FSINCOS, MVT::v4f16, Expand);
338 setOperationAction(ISD::FSQRT, MVT::v4f16, Expand);
339 setOperationAction(ISD::FTRUNC, MVT::v4f16, Expand);
340 setOperationAction(ISD::SETCC, MVT::v4f16, Expand);
341 setOperationAction(ISD::BR_CC, MVT::v4f16, Expand);
342 setOperationAction(ISD::SELECT, MVT::v4f16, Expand);
343 setOperationAction(ISD::SELECT_CC, MVT::v4f16, Expand);
344 setOperationAction(ISD::FEXP, MVT::v4f16, Expand);
345 setOperationAction(ISD::FEXP2, MVT::v4f16, Expand);
346 setOperationAction(ISD::FLOG, MVT::v4f16, Expand);
347 setOperationAction(ISD::FLOG2, MVT::v4f16, Expand);
348 setOperationAction(ISD::FLOG10, MVT::v4f16, Expand);
351 // v8f16 is also a storage-only type, so expand it.
352 setOperationAction(ISD::FABS, MVT::v8f16, Expand);
353 setOperationAction(ISD::FADD, MVT::v8f16, Expand);
354 setOperationAction(ISD::FCEIL, MVT::v8f16, Expand);
355 setOperationAction(ISD::FCOPYSIGN, MVT::v8f16, Expand);
356 setOperationAction(ISD::FCOS, MVT::v8f16, Expand);
357 setOperationAction(ISD::FDIV, MVT::v8f16, Expand);
358 setOperationAction(ISD::FFLOOR, MVT::v8f16, Expand);
359 setOperationAction(ISD::FMA, MVT::v8f16, Expand);
360 setOperationAction(ISD::FMUL, MVT::v8f16, Expand);
361 setOperationAction(ISD::FNEARBYINT, MVT::v8f16, Expand);
362 setOperationAction(ISD::FNEG, MVT::v8f16, Expand);
363 setOperationAction(ISD::FPOW, MVT::v8f16, Expand);
364 setOperationAction(ISD::FPOWI, MVT::v8f16, Expand);
365 setOperationAction(ISD::FREM, MVT::v8f16, Expand);
366 setOperationAction(ISD::FROUND, MVT::v8f16, Expand);
367 setOperationAction(ISD::FRINT, MVT::v8f16, Expand);
368 setOperationAction(ISD::FSIN, MVT::v8f16, Expand);
369 setOperationAction(ISD::FSINCOS, MVT::v8f16, Expand);
370 setOperationAction(ISD::FSQRT, MVT::v8f16, Expand);
371 setOperationAction(ISD::FSUB, MVT::v8f16, Expand);
372 setOperationAction(ISD::FTRUNC, MVT::v8f16, Expand);
373 setOperationAction(ISD::SETCC, MVT::v8f16, Expand);
374 setOperationAction(ISD::BR_CC, MVT::v8f16, Expand);
375 setOperationAction(ISD::SELECT, MVT::v8f16, Expand);
376 setOperationAction(ISD::SELECT_CC, MVT::v8f16, Expand);
377 setOperationAction(ISD::FP_EXTEND, MVT::v8f16, Expand);
378 setOperationAction(ISD::FEXP, MVT::v8f16, Expand);
379 setOperationAction(ISD::FEXP2, MVT::v8f16, Expand);
380 setOperationAction(ISD::FLOG, MVT::v8f16, Expand);
381 setOperationAction(ISD::FLOG2, MVT::v8f16, Expand);
382 setOperationAction(ISD::FLOG10, MVT::v8f16, Expand);
384 // AArch64 has implementations of a lot of rounding-like FP operations.
385 for (MVT Ty : {MVT::f32, MVT::f64}) {
386 setOperationAction(ISD::FFLOOR, Ty, Legal);
387 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
388 setOperationAction(ISD::FCEIL, Ty, Legal);
389 setOperationAction(ISD::FRINT, Ty, Legal);
390 setOperationAction(ISD::FTRUNC, Ty, Legal);
391 setOperationAction(ISD::FROUND, Ty, Legal);
394 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
396 if (Subtarget->isTargetMachO()) {
397 // For iOS, we don't want to the normal expansion of a libcall to
398 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
400 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
401 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
403 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
404 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
407 // Make floating-point constants legal for the large code model, so they don't
408 // become loads from the constant pool.
409 if (Subtarget->isTargetMachO() && TM.getCodeModel() == CodeModel::Large) {
410 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
411 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
414 // AArch64 does not have floating-point extending loads, i1 sign-extending
415 // load, floating-point truncating stores, or v2i32->v2i16 truncating store.
416 for (MVT VT : MVT::fp_valuetypes()) {
417 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
418 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
419 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f64, Expand);
420 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f80, Expand);
422 for (MVT VT : MVT::integer_valuetypes())
423 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Expand);
425 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
426 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
427 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
428 setTruncStoreAction(MVT::f128, MVT::f80, Expand);
429 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
430 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
431 setTruncStoreAction(MVT::f128, MVT::f16, Expand);
433 setOperationAction(ISD::BITCAST, MVT::i16, Custom);
434 setOperationAction(ISD::BITCAST, MVT::f16, Custom);
436 // Indexed loads and stores are supported.
437 for (unsigned im = (unsigned)ISD::PRE_INC;
438 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
439 setIndexedLoadAction(im, MVT::i8, Legal);
440 setIndexedLoadAction(im, MVT::i16, Legal);
441 setIndexedLoadAction(im, MVT::i32, Legal);
442 setIndexedLoadAction(im, MVT::i64, Legal);
443 setIndexedLoadAction(im, MVT::f64, Legal);
444 setIndexedLoadAction(im, MVT::f32, Legal);
445 setIndexedStoreAction(im, MVT::i8, Legal);
446 setIndexedStoreAction(im, MVT::i16, Legal);
447 setIndexedStoreAction(im, MVT::i32, Legal);
448 setIndexedStoreAction(im, MVT::i64, Legal);
449 setIndexedStoreAction(im, MVT::f64, Legal);
450 setIndexedStoreAction(im, MVT::f32, Legal);
454 setOperationAction(ISD::TRAP, MVT::Other, Legal);
456 // We combine OR nodes for bitfield operations.
457 setTargetDAGCombine(ISD::OR);
459 // Vector add and sub nodes may conceal a high-half opportunity.
460 // Also, try to fold ADD into CSINC/CSINV..
461 setTargetDAGCombine(ISD::ADD);
462 setTargetDAGCombine(ISD::SUB);
464 setTargetDAGCombine(ISD::XOR);
465 setTargetDAGCombine(ISD::SINT_TO_FP);
466 setTargetDAGCombine(ISD::UINT_TO_FP);
468 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
470 setTargetDAGCombine(ISD::ANY_EXTEND);
471 setTargetDAGCombine(ISD::ZERO_EXTEND);
472 setTargetDAGCombine(ISD::SIGN_EXTEND);
473 setTargetDAGCombine(ISD::BITCAST);
474 setTargetDAGCombine(ISD::CONCAT_VECTORS);
475 setTargetDAGCombine(ISD::STORE);
477 setTargetDAGCombine(ISD::MUL);
479 setTargetDAGCombine(ISD::SELECT);
480 setTargetDAGCombine(ISD::VSELECT);
481 setTargetDAGCombine(ISD::SELECT_CC);
483 setTargetDAGCombine(ISD::INTRINSIC_VOID);
484 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
485 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
487 MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8;
488 MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4;
489 MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 4;
491 setStackPointerRegisterToSaveRestore(AArch64::SP);
493 setSchedulingPreference(Sched::Hybrid);
496 MaskAndBranchFoldingIsLegal = true;
497 EnableExtLdPromotion = true;
499 setMinFunctionAlignment(2);
501 setHasExtractBitsInsn(true);
503 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
505 if (Subtarget->hasNEON()) {
506 // FIXME: v1f64 shouldn't be legal if we can avoid it, because it leads to
507 // silliness like this:
508 setOperationAction(ISD::FABS, MVT::v1f64, Expand);
509 setOperationAction(ISD::FADD, MVT::v1f64, Expand);
510 setOperationAction(ISD::FCEIL, MVT::v1f64, Expand);
511 setOperationAction(ISD::FCOPYSIGN, MVT::v1f64, Expand);
512 setOperationAction(ISD::FCOS, MVT::v1f64, Expand);
513 setOperationAction(ISD::FDIV, MVT::v1f64, Expand);
514 setOperationAction(ISD::FFLOOR, MVT::v1f64, Expand);
515 setOperationAction(ISD::FMA, MVT::v1f64, Expand);
516 setOperationAction(ISD::FMUL, MVT::v1f64, Expand);
517 setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Expand);
518 setOperationAction(ISD::FNEG, MVT::v1f64, Expand);
519 setOperationAction(ISD::FPOW, MVT::v1f64, Expand);
520 setOperationAction(ISD::FREM, MVT::v1f64, Expand);
521 setOperationAction(ISD::FROUND, MVT::v1f64, Expand);
522 setOperationAction(ISD::FRINT, MVT::v1f64, Expand);
523 setOperationAction(ISD::FSIN, MVT::v1f64, Expand);
524 setOperationAction(ISD::FSINCOS, MVT::v1f64, Expand);
525 setOperationAction(ISD::FSQRT, MVT::v1f64, Expand);
526 setOperationAction(ISD::FSUB, MVT::v1f64, Expand);
527 setOperationAction(ISD::FTRUNC, MVT::v1f64, Expand);
528 setOperationAction(ISD::SETCC, MVT::v1f64, Expand);
529 setOperationAction(ISD::BR_CC, MVT::v1f64, Expand);
530 setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
531 setOperationAction(ISD::SELECT_CC, MVT::v1f64, Expand);
532 setOperationAction(ISD::FP_EXTEND, MVT::v1f64, Expand);
534 setOperationAction(ISD::FP_TO_SINT, MVT::v1i64, Expand);
535 setOperationAction(ISD::FP_TO_UINT, MVT::v1i64, Expand);
536 setOperationAction(ISD::SINT_TO_FP, MVT::v1i64, Expand);
537 setOperationAction(ISD::UINT_TO_FP, MVT::v1i64, Expand);
538 setOperationAction(ISD::FP_ROUND, MVT::v1f64, Expand);
540 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
542 // AArch64 doesn't have a direct vector ->f32 conversion instructions for
543 // elements smaller than i32, so promote the input to i32 first.
544 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Promote);
545 setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Promote);
546 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Promote);
547 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Promote);
548 // i8 and i16 vector elements also need promotion to i32 for v8i8 or v8i16
549 // -> v8f16 conversions.
550 setOperationAction(ISD::SINT_TO_FP, MVT::v8i8, Promote);
551 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Promote);
552 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
553 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Promote);
554 // Similarly, there is no direct i32 -> f64 vector conversion instruction.
555 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
556 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
557 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
558 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
559 // Or, direct i32 -> f16 vector conversion. Set it so custom, so the
560 // conversion happens in two steps: v4i32 -> v4f32 -> v4f16
561 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Custom);
562 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
564 // AArch64 doesn't have MUL.2d:
565 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
566 // Custom handling for some quad-vector types to detect MULL.
567 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
568 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
569 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
571 setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal);
572 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
573 // Likewise, narrowing and extending vector loads/stores aren't handled
575 for (MVT VT : MVT::vector_valuetypes()) {
576 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
578 setOperationAction(ISD::MULHS, VT, Expand);
579 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
580 setOperationAction(ISD::MULHU, VT, Expand);
581 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
583 setOperationAction(ISD::BSWAP, VT, Expand);
585 for (MVT InnerVT : MVT::vector_valuetypes()) {
586 setTruncStoreAction(VT, InnerVT, Expand);
587 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
588 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
589 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
593 // AArch64 has implementations of a lot of rounding-like FP operations.
594 for (MVT Ty : {MVT::v2f32, MVT::v4f32, MVT::v2f64}) {
595 setOperationAction(ISD::FFLOOR, Ty, Legal);
596 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
597 setOperationAction(ISD::FCEIL, Ty, Legal);
598 setOperationAction(ISD::FRINT, Ty, Legal);
599 setOperationAction(ISD::FTRUNC, Ty, Legal);
600 setOperationAction(ISD::FROUND, Ty, Legal);
604 // Prefer likely predicted branches to selects on out-of-order cores.
605 if (Subtarget->isCortexA57())
606 PredictableSelectIsExpensive = true;
609 void AArch64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) {
610 if (VT == MVT::v2f32 || VT == MVT::v4f16) {
611 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
612 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i32);
614 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
615 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i32);
616 } else if (VT == MVT::v2f64 || VT == MVT::v4f32 || VT == MVT::v8f16) {
617 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
618 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i64);
620 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
621 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i64);
624 // Mark vector float intrinsics as expand.
625 if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) {
626 setOperationAction(ISD::FSIN, VT.getSimpleVT(), Expand);
627 setOperationAction(ISD::FCOS, VT.getSimpleVT(), Expand);
628 setOperationAction(ISD::FPOWI, VT.getSimpleVT(), Expand);
629 setOperationAction(ISD::FPOW, VT.getSimpleVT(), Expand);
630 setOperationAction(ISD::FLOG, VT.getSimpleVT(), Expand);
631 setOperationAction(ISD::FLOG2, VT.getSimpleVT(), Expand);
632 setOperationAction(ISD::FLOG10, VT.getSimpleVT(), Expand);
633 setOperationAction(ISD::FEXP, VT.getSimpleVT(), Expand);
634 setOperationAction(ISD::FEXP2, VT.getSimpleVT(), Expand);
637 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
638 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
639 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
640 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
641 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Custom);
642 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
643 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
644 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
645 setOperationAction(ISD::AND, VT.getSimpleVT(), Custom);
646 setOperationAction(ISD::OR, VT.getSimpleVT(), Custom);
647 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
648 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
650 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
651 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
652 setOperationAction(ISD::VSELECT, VT.getSimpleVT(), Expand);
653 for (MVT InnerVT : MVT::all_valuetypes())
654 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT.getSimpleVT(), Expand);
656 // CNT supports only B element sizes.
657 if (VT != MVT::v8i8 && VT != MVT::v16i8)
658 setOperationAction(ISD::CTPOP, VT.getSimpleVT(), Expand);
660 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
661 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
662 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
663 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
664 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
666 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
667 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
669 // [SU][MIN|MAX] and [SU]ABSDIFF are available for all NEON types apart from
671 if (!VT.isFloatingPoint() &&
672 VT.getSimpleVT() != MVT::v2i64 && VT.getSimpleVT() != MVT::v1i64)
673 for (unsigned Opcode : {ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX,
674 ISD::SABSDIFF, ISD::UABSDIFF})
675 setOperationAction(Opcode, VT.getSimpleVT(), Legal);
677 if (Subtarget->isLittleEndian()) {
678 for (unsigned im = (unsigned)ISD::PRE_INC;
679 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
680 setIndexedLoadAction(im, VT.getSimpleVT(), Legal);
681 setIndexedStoreAction(im, VT.getSimpleVT(), Legal);
686 void AArch64TargetLowering::addDRTypeForNEON(MVT VT) {
687 addRegisterClass(VT, &AArch64::FPR64RegClass);
688 addTypeForNEON(VT, MVT::v2i32);
691 void AArch64TargetLowering::addQRTypeForNEON(MVT VT) {
692 addRegisterClass(VT, &AArch64::FPR128RegClass);
693 addTypeForNEON(VT, MVT::v4i32);
696 EVT AArch64TargetLowering::getSetCCResultType(const DataLayout &, LLVMContext &,
700 return VT.changeVectorElementTypeToInteger();
703 /// computeKnownBitsForTargetNode - Determine which of the bits specified in
704 /// Mask are known to be either zero or one and return them in the
705 /// KnownZero/KnownOne bitsets.
706 void AArch64TargetLowering::computeKnownBitsForTargetNode(
707 const SDValue Op, APInt &KnownZero, APInt &KnownOne,
708 const SelectionDAG &DAG, unsigned Depth) const {
709 switch (Op.getOpcode()) {
712 case AArch64ISD::CSEL: {
713 APInt KnownZero2, KnownOne2;
714 DAG.computeKnownBits(Op->getOperand(0), KnownZero, KnownOne, Depth + 1);
715 DAG.computeKnownBits(Op->getOperand(1), KnownZero2, KnownOne2, Depth + 1);
716 KnownZero &= KnownZero2;
717 KnownOne &= KnownOne2;
720 case ISD::INTRINSIC_W_CHAIN: {
721 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
722 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
725 case Intrinsic::aarch64_ldaxr:
726 case Intrinsic::aarch64_ldxr: {
727 unsigned BitWidth = KnownOne.getBitWidth();
728 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
729 unsigned MemBits = VT.getScalarType().getSizeInBits();
730 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
736 case ISD::INTRINSIC_WO_CHAIN:
737 case ISD::INTRINSIC_VOID: {
738 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
742 case Intrinsic::aarch64_neon_umaxv:
743 case Intrinsic::aarch64_neon_uminv: {
744 // Figure out the datatype of the vector operand. The UMINV instruction
745 // will zero extend the result, so we can mark as known zero all the
746 // bits larger than the element datatype. 32-bit or larget doesn't need
747 // this as those are legal types and will be handled by isel directly.
748 MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
749 unsigned BitWidth = KnownZero.getBitWidth();
750 if (VT == MVT::v8i8 || VT == MVT::v16i8) {
751 assert(BitWidth >= 8 && "Unexpected width!");
752 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8);
754 } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
755 assert(BitWidth >= 16 && "Unexpected width!");
756 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
766 MVT AArch64TargetLowering::getScalarShiftAmountTy(const DataLayout &DL,
771 bool AArch64TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
775 if (Subtarget->requiresStrictAlign())
777 // FIXME: True for Cyclone, but not necessary others.
784 AArch64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
785 const TargetLibraryInfo *libInfo) const {
786 return AArch64::createFastISel(funcInfo, libInfo);
789 const char *AArch64TargetLowering::getTargetNodeName(unsigned Opcode) const {
790 switch ((AArch64ISD::NodeType)Opcode) {
791 case AArch64ISD::FIRST_NUMBER: break;
792 case AArch64ISD::CALL: return "AArch64ISD::CALL";
793 case AArch64ISD::ADRP: return "AArch64ISD::ADRP";
794 case AArch64ISD::ADDlow: return "AArch64ISD::ADDlow";
795 case AArch64ISD::LOADgot: return "AArch64ISD::LOADgot";
796 case AArch64ISD::RET_FLAG: return "AArch64ISD::RET_FLAG";
797 case AArch64ISD::BRCOND: return "AArch64ISD::BRCOND";
798 case AArch64ISD::CSEL: return "AArch64ISD::CSEL";
799 case AArch64ISD::FCSEL: return "AArch64ISD::FCSEL";
800 case AArch64ISD::CSINV: return "AArch64ISD::CSINV";
801 case AArch64ISD::CSNEG: return "AArch64ISD::CSNEG";
802 case AArch64ISD::CSINC: return "AArch64ISD::CSINC";
803 case AArch64ISD::THREAD_POINTER: return "AArch64ISD::THREAD_POINTER";
804 case AArch64ISD::TLSDESC_CALLSEQ: return "AArch64ISD::TLSDESC_CALLSEQ";
805 case AArch64ISD::ADC: return "AArch64ISD::ADC";
806 case AArch64ISD::SBC: return "AArch64ISD::SBC";
807 case AArch64ISD::ADDS: return "AArch64ISD::ADDS";
808 case AArch64ISD::SUBS: return "AArch64ISD::SUBS";
809 case AArch64ISD::ADCS: return "AArch64ISD::ADCS";
810 case AArch64ISD::SBCS: return "AArch64ISD::SBCS";
811 case AArch64ISD::ANDS: return "AArch64ISD::ANDS";
812 case AArch64ISD::CCMP: return "AArch64ISD::CCMP";
813 case AArch64ISD::CCMN: return "AArch64ISD::CCMN";
814 case AArch64ISD::FCCMP: return "AArch64ISD::FCCMP";
815 case AArch64ISD::FCMP: return "AArch64ISD::FCMP";
816 case AArch64ISD::FMIN: return "AArch64ISD::FMIN";
817 case AArch64ISD::FMAX: return "AArch64ISD::FMAX";
818 case AArch64ISD::DUP: return "AArch64ISD::DUP";
819 case AArch64ISD::DUPLANE8: return "AArch64ISD::DUPLANE8";
820 case AArch64ISD::DUPLANE16: return "AArch64ISD::DUPLANE16";
821 case AArch64ISD::DUPLANE32: return "AArch64ISD::DUPLANE32";
822 case AArch64ISD::DUPLANE64: return "AArch64ISD::DUPLANE64";
823 case AArch64ISD::MOVI: return "AArch64ISD::MOVI";
824 case AArch64ISD::MOVIshift: return "AArch64ISD::MOVIshift";
825 case AArch64ISD::MOVIedit: return "AArch64ISD::MOVIedit";
826 case AArch64ISD::MOVImsl: return "AArch64ISD::MOVImsl";
827 case AArch64ISD::FMOV: return "AArch64ISD::FMOV";
828 case AArch64ISD::MVNIshift: return "AArch64ISD::MVNIshift";
829 case AArch64ISD::MVNImsl: return "AArch64ISD::MVNImsl";
830 case AArch64ISD::BICi: return "AArch64ISD::BICi";
831 case AArch64ISD::ORRi: return "AArch64ISD::ORRi";
832 case AArch64ISD::BSL: return "AArch64ISD::BSL";
833 case AArch64ISD::NEG: return "AArch64ISD::NEG";
834 case AArch64ISD::EXTR: return "AArch64ISD::EXTR";
835 case AArch64ISD::ZIP1: return "AArch64ISD::ZIP1";
836 case AArch64ISD::ZIP2: return "AArch64ISD::ZIP2";
837 case AArch64ISD::UZP1: return "AArch64ISD::UZP1";
838 case AArch64ISD::UZP2: return "AArch64ISD::UZP2";
839 case AArch64ISD::TRN1: return "AArch64ISD::TRN1";
840 case AArch64ISD::TRN2: return "AArch64ISD::TRN2";
841 case AArch64ISD::REV16: return "AArch64ISD::REV16";
842 case AArch64ISD::REV32: return "AArch64ISD::REV32";
843 case AArch64ISD::REV64: return "AArch64ISD::REV64";
844 case AArch64ISD::EXT: return "AArch64ISD::EXT";
845 case AArch64ISD::VSHL: return "AArch64ISD::VSHL";
846 case AArch64ISD::VLSHR: return "AArch64ISD::VLSHR";
847 case AArch64ISD::VASHR: return "AArch64ISD::VASHR";
848 case AArch64ISD::CMEQ: return "AArch64ISD::CMEQ";
849 case AArch64ISD::CMGE: return "AArch64ISD::CMGE";
850 case AArch64ISD::CMGT: return "AArch64ISD::CMGT";
851 case AArch64ISD::CMHI: return "AArch64ISD::CMHI";
852 case AArch64ISD::CMHS: return "AArch64ISD::CMHS";
853 case AArch64ISD::FCMEQ: return "AArch64ISD::FCMEQ";
854 case AArch64ISD::FCMGE: return "AArch64ISD::FCMGE";
855 case AArch64ISD::FCMGT: return "AArch64ISD::FCMGT";
856 case AArch64ISD::CMEQz: return "AArch64ISD::CMEQz";
857 case AArch64ISD::CMGEz: return "AArch64ISD::CMGEz";
858 case AArch64ISD::CMGTz: return "AArch64ISD::CMGTz";
859 case AArch64ISD::CMLEz: return "AArch64ISD::CMLEz";
860 case AArch64ISD::CMLTz: return "AArch64ISD::CMLTz";
861 case AArch64ISD::FCMEQz: return "AArch64ISD::FCMEQz";
862 case AArch64ISD::FCMGEz: return "AArch64ISD::FCMGEz";
863 case AArch64ISD::FCMGTz: return "AArch64ISD::FCMGTz";
864 case AArch64ISD::FCMLEz: return "AArch64ISD::FCMLEz";
865 case AArch64ISD::FCMLTz: return "AArch64ISD::FCMLTz";
866 case AArch64ISD::SADDV: return "AArch64ISD::SADDV";
867 case AArch64ISD::UADDV: return "AArch64ISD::UADDV";
868 case AArch64ISD::SMINV: return "AArch64ISD::SMINV";
869 case AArch64ISD::UMINV: return "AArch64ISD::UMINV";
870 case AArch64ISD::SMAXV: return "AArch64ISD::SMAXV";
871 case AArch64ISD::UMAXV: return "AArch64ISD::UMAXV";
872 case AArch64ISD::NOT: return "AArch64ISD::NOT";
873 case AArch64ISD::BIT: return "AArch64ISD::BIT";
874 case AArch64ISD::CBZ: return "AArch64ISD::CBZ";
875 case AArch64ISD::CBNZ: return "AArch64ISD::CBNZ";
876 case AArch64ISD::TBZ: return "AArch64ISD::TBZ";
877 case AArch64ISD::TBNZ: return "AArch64ISD::TBNZ";
878 case AArch64ISD::TC_RETURN: return "AArch64ISD::TC_RETURN";
879 case AArch64ISD::PREFETCH: return "AArch64ISD::PREFETCH";
880 case AArch64ISD::SITOF: return "AArch64ISD::SITOF";
881 case AArch64ISD::UITOF: return "AArch64ISD::UITOF";
882 case AArch64ISD::NVCAST: return "AArch64ISD::NVCAST";
883 case AArch64ISD::SQSHL_I: return "AArch64ISD::SQSHL_I";
884 case AArch64ISD::UQSHL_I: return "AArch64ISD::UQSHL_I";
885 case AArch64ISD::SRSHR_I: return "AArch64ISD::SRSHR_I";
886 case AArch64ISD::URSHR_I: return "AArch64ISD::URSHR_I";
887 case AArch64ISD::SQSHLU_I: return "AArch64ISD::SQSHLU_I";
888 case AArch64ISD::WrapperLarge: return "AArch64ISD::WrapperLarge";
889 case AArch64ISD::LD2post: return "AArch64ISD::LD2post";
890 case AArch64ISD::LD3post: return "AArch64ISD::LD3post";
891 case AArch64ISD::LD4post: return "AArch64ISD::LD4post";
892 case AArch64ISD::ST2post: return "AArch64ISD::ST2post";
893 case AArch64ISD::ST3post: return "AArch64ISD::ST3post";
894 case AArch64ISD::ST4post: return "AArch64ISD::ST4post";
895 case AArch64ISD::LD1x2post: return "AArch64ISD::LD1x2post";
896 case AArch64ISD::LD1x3post: return "AArch64ISD::LD1x3post";
897 case AArch64ISD::LD1x4post: return "AArch64ISD::LD1x4post";
898 case AArch64ISD::ST1x2post: return "AArch64ISD::ST1x2post";
899 case AArch64ISD::ST1x3post: return "AArch64ISD::ST1x3post";
900 case AArch64ISD::ST1x4post: return "AArch64ISD::ST1x4post";
901 case AArch64ISD::LD1DUPpost: return "AArch64ISD::LD1DUPpost";
902 case AArch64ISD::LD2DUPpost: return "AArch64ISD::LD2DUPpost";
903 case AArch64ISD::LD3DUPpost: return "AArch64ISD::LD3DUPpost";
904 case AArch64ISD::LD4DUPpost: return "AArch64ISD::LD4DUPpost";
905 case AArch64ISD::LD1LANEpost: return "AArch64ISD::LD1LANEpost";
906 case AArch64ISD::LD2LANEpost: return "AArch64ISD::LD2LANEpost";
907 case AArch64ISD::LD3LANEpost: return "AArch64ISD::LD3LANEpost";
908 case AArch64ISD::LD4LANEpost: return "AArch64ISD::LD4LANEpost";
909 case AArch64ISD::ST2LANEpost: return "AArch64ISD::ST2LANEpost";
910 case AArch64ISD::ST3LANEpost: return "AArch64ISD::ST3LANEpost";
911 case AArch64ISD::ST4LANEpost: return "AArch64ISD::ST4LANEpost";
912 case AArch64ISD::SMULL: return "AArch64ISD::SMULL";
913 case AArch64ISD::UMULL: return "AArch64ISD::UMULL";
919 AArch64TargetLowering::EmitF128CSEL(MachineInstr *MI,
920 MachineBasicBlock *MBB) const {
921 // We materialise the F128CSEL pseudo-instruction as some control flow and a
925 // [... previous instrs leading to comparison ...]
931 // Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB]
933 MachineFunction *MF = MBB->getParent();
934 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
935 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
936 DebugLoc DL = MI->getDebugLoc();
937 MachineFunction::iterator It = MBB;
940 unsigned DestReg = MI->getOperand(0).getReg();
941 unsigned IfTrueReg = MI->getOperand(1).getReg();
942 unsigned IfFalseReg = MI->getOperand(2).getReg();
943 unsigned CondCode = MI->getOperand(3).getImm();
944 bool NZCVKilled = MI->getOperand(4).isKill();
946 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
947 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
948 MF->insert(It, TrueBB);
949 MF->insert(It, EndBB);
951 // Transfer rest of current basic-block to EndBB
952 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
954 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
956 BuildMI(MBB, DL, TII->get(AArch64::Bcc)).addImm(CondCode).addMBB(TrueBB);
957 BuildMI(MBB, DL, TII->get(AArch64::B)).addMBB(EndBB);
958 MBB->addSuccessor(TrueBB);
959 MBB->addSuccessor(EndBB);
961 // TrueBB falls through to the end.
962 TrueBB->addSuccessor(EndBB);
965 TrueBB->addLiveIn(AArch64::NZCV);
966 EndBB->addLiveIn(AArch64::NZCV);
969 BuildMI(*EndBB, EndBB->begin(), DL, TII->get(AArch64::PHI), DestReg)
975 MI->eraseFromParent();
980 AArch64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
981 MachineBasicBlock *BB) const {
982 switch (MI->getOpcode()) {
987 llvm_unreachable("Unexpected instruction for custom inserter!");
989 case AArch64::F128CSEL:
990 return EmitF128CSEL(MI, BB);
992 case TargetOpcode::STACKMAP:
993 case TargetOpcode::PATCHPOINT:
994 return emitPatchPoint(MI, BB);
998 //===----------------------------------------------------------------------===//
999 // AArch64 Lowering private implementation.
1000 //===----------------------------------------------------------------------===//
1002 //===----------------------------------------------------------------------===//
1004 //===----------------------------------------------------------------------===//
1006 /// changeIntCCToAArch64CC - Convert a DAG integer condition code to an AArch64
1008 static AArch64CC::CondCode changeIntCCToAArch64CC(ISD::CondCode CC) {
1011 llvm_unreachable("Unknown condition code!");
1013 return AArch64CC::NE;
1015 return AArch64CC::EQ;
1017 return AArch64CC::GT;
1019 return AArch64CC::GE;
1021 return AArch64CC::LT;
1023 return AArch64CC::LE;
1025 return AArch64CC::HI;
1027 return AArch64CC::HS;
1029 return AArch64CC::LO;
1031 return AArch64CC::LS;
1035 /// changeFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64 CC.
1036 static void changeFPCCToAArch64CC(ISD::CondCode CC,
1037 AArch64CC::CondCode &CondCode,
1038 AArch64CC::CondCode &CondCode2) {
1039 CondCode2 = AArch64CC::AL;
1042 llvm_unreachable("Unknown FP condition!");
1045 CondCode = AArch64CC::EQ;
1049 CondCode = AArch64CC::GT;
1053 CondCode = AArch64CC::GE;
1056 CondCode = AArch64CC::MI;
1059 CondCode = AArch64CC::LS;
1062 CondCode = AArch64CC::MI;
1063 CondCode2 = AArch64CC::GT;
1066 CondCode = AArch64CC::VC;
1069 CondCode = AArch64CC::VS;
1072 CondCode = AArch64CC::EQ;
1073 CondCode2 = AArch64CC::VS;
1076 CondCode = AArch64CC::HI;
1079 CondCode = AArch64CC::PL;
1083 CondCode = AArch64CC::LT;
1087 CondCode = AArch64CC::LE;
1091 CondCode = AArch64CC::NE;
1096 /// changeVectorFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64
1097 /// CC usable with the vector instructions. Fewer operations are available
1098 /// without a real NZCV register, so we have to use less efficient combinations
1099 /// to get the same effect.
1100 static void changeVectorFPCCToAArch64CC(ISD::CondCode CC,
1101 AArch64CC::CondCode &CondCode,
1102 AArch64CC::CondCode &CondCode2,
1107 // Mostly the scalar mappings work fine.
1108 changeFPCCToAArch64CC(CC, CondCode, CondCode2);
1111 Invert = true; // Fallthrough
1113 CondCode = AArch64CC::MI;
1114 CondCode2 = AArch64CC::GE;
1121 // All of the compare-mask comparisons are ordered, but we can switch
1122 // between the two by a double inversion. E.g. ULE == !OGT.
1124 changeFPCCToAArch64CC(getSetCCInverse(CC, false), CondCode, CondCode2);
1129 static bool isLegalArithImmed(uint64_t C) {
1130 // Matches AArch64DAGToDAGISel::SelectArithImmed().
1131 return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
1134 static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1135 SDLoc dl, SelectionDAG &DAG) {
1136 EVT VT = LHS.getValueType();
1138 if (VT.isFloatingPoint())
1139 return DAG.getNode(AArch64ISD::FCMP, dl, VT, LHS, RHS);
1141 // The CMP instruction is just an alias for SUBS, and representing it as
1142 // SUBS means that it's possible to get CSE with subtract operations.
1143 // A later phase can perform the optimization of setting the destination
1144 // register to WZR/XZR if it ends up being unused.
1145 unsigned Opcode = AArch64ISD::SUBS;
1147 if (RHS.getOpcode() == ISD::SUB && isa<ConstantSDNode>(RHS.getOperand(0)) &&
1148 cast<ConstantSDNode>(RHS.getOperand(0))->getZExtValue() == 0 &&
1149 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1150 // We'd like to combine a (CMP op1, (sub 0, op2) into a CMN instruction on
1151 // the grounds that "op1 - (-op2) == op1 + op2". However, the C and V flags
1152 // can be set differently by this operation. It comes down to whether
1153 // "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are then
1154 // everything is fine. If not then the optimization is wrong. Thus general
1155 // comparisons are only valid if op2 != 0.
1157 // So, finally, the only LLVM-native comparisons that don't mention C and V
1158 // are SETEQ and SETNE. They're the only ones we can safely use CMN for in
1159 // the absence of information about op2.
1160 Opcode = AArch64ISD::ADDS;
1161 RHS = RHS.getOperand(1);
1162 } else if (LHS.getOpcode() == ISD::AND && isa<ConstantSDNode>(RHS) &&
1163 cast<ConstantSDNode>(RHS)->getZExtValue() == 0 &&
1164 !isUnsignedIntSetCC(CC)) {
1165 // Similarly, (CMP (and X, Y), 0) can be implemented with a TST
1166 // (a.k.a. ANDS) except that the flags are only guaranteed to work for one
1167 // of the signed comparisons.
1168 Opcode = AArch64ISD::ANDS;
1169 RHS = LHS.getOperand(1);
1170 LHS = LHS.getOperand(0);
1173 return DAG.getNode(Opcode, dl, DAG.getVTList(VT, MVT_CC), LHS, RHS)
1177 /// \defgroup AArch64CCMP CMP;CCMP matching
1179 /// These functions deal with the formation of CMP;CCMP;... sequences.
1180 /// The CCMP/CCMN/FCCMP/FCCMPE instructions allow the conditional execution of
1181 /// a comparison. They set the NZCV flags to a predefined value if their
1182 /// predicate is false. This allows to express arbitrary conjunctions, for
1183 /// example "cmp 0 (and (setCA (cmp A)) (setCB (cmp B))))"
1186 /// ccmp B, inv(CB), CA
1187 /// check for CB flags
1189 /// In general we can create code for arbitrary "... (and (and A B) C)"
1190 /// sequences. We can also implement some "or" expressions, because "(or A B)"
1191 /// is equivalent to "not (and (not A) (not B))" and we can implement some
1192 /// negation operations:
1193 /// We can negate the results of a single comparison by inverting the flags
1194 /// used when the predicate fails and inverting the flags tested in the next
1195 /// instruction; We can also negate the results of the whole previous
1196 /// conditional compare sequence by inverting the flags tested in the next
1197 /// instruction. However there is no way to negate the result of a partial
1200 /// Therefore on encountering an "or" expression we can negate the subtree on
1201 /// one side and have to be able to push the negate to the leafs of the subtree
1202 /// on the other side (see also the comments in code). As complete example:
1203 /// "or (or (setCA (cmp A)) (setCB (cmp B)))
1204 /// (and (setCC (cmp C)) (setCD (cmp D)))"
1205 /// is transformed to
1206 /// "not (and (not (and (setCC (cmp C)) (setCC (cmp D))))
1207 /// (and (not (setCA (cmp A)) (not (setCB (cmp B))))))"
1208 /// and implemented as:
1210 /// ccmp D, inv(CD), CC
1211 /// ccmp A, CA, inv(CD)
1212 /// ccmp B, CB, inv(CA)
1213 /// check for CB flags
1214 /// A counterexample is "or (and A B) (and C D)" which cannot be implemented
1215 /// by conditional compare sequences.
1218 /// Create a conditional comparison; Use CCMP, CCMN or FCCMP as appropriate.
1219 static SDValue emitConditionalComparison(SDValue LHS, SDValue RHS,
1220 ISD::CondCode CC, SDValue CCOp,
1221 SDValue Condition, unsigned NZCV,
1222 SDLoc DL, SelectionDAG &DAG) {
1223 unsigned Opcode = 0;
1224 if (LHS.getValueType().isFloatingPoint())
1225 Opcode = AArch64ISD::FCCMP;
1226 else if (RHS.getOpcode() == ISD::SUB) {
1227 SDValue SubOp0 = RHS.getOperand(0);
1228 if (const ConstantSDNode *SubOp0C = dyn_cast<ConstantSDNode>(SubOp0))
1229 if (SubOp0C->isNullValue() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1230 // See emitComparison() on why we can only do this for SETEQ and SETNE.
1231 Opcode = AArch64ISD::CCMN;
1232 RHS = RHS.getOperand(1);
1236 Opcode = AArch64ISD::CCMP;
1238 SDValue NZCVOp = DAG.getConstant(NZCV, DL, MVT::i32);
1239 return DAG.getNode(Opcode, DL, MVT_CC, LHS, RHS, NZCVOp, Condition, CCOp);
1242 /// Returns true if @p Val is a tree of AND/OR/SETCC operations.
1243 /// CanPushNegate is set to true if we can push a negate operation through
1244 /// the tree in a was that we are left with AND operations and negate operations
1245 /// at the leafs only. i.e. "not (or (or x y) z)" can be changed to
1246 /// "and (and (not x) (not y)) (not z)"; "not (or (and x y) z)" cannot be
1247 /// brought into such a form.
1248 static bool isConjunctionDisjunctionTree(const SDValue Val, bool &CanPushNegate,
1249 unsigned Depth = 0) {
1250 if (!Val.hasOneUse())
1252 unsigned Opcode = Val->getOpcode();
1253 if (Opcode == ISD::SETCC) {
1254 CanPushNegate = true;
1257 // Protect against stack overflow.
1260 if (Opcode == ISD::AND || Opcode == ISD::OR) {
1261 SDValue O0 = Val->getOperand(0);
1262 SDValue O1 = Val->getOperand(1);
1263 bool CanPushNegateL;
1264 if (!isConjunctionDisjunctionTree(O0, CanPushNegateL, Depth+1))
1266 bool CanPushNegateR;
1267 if (!isConjunctionDisjunctionTree(O1, CanPushNegateR, Depth+1))
1269 // We cannot push a negate through an AND operation (it would become an OR),
1270 // we can however change a (not (or x y)) to (and (not x) (not y)) if we can
1271 // push the negate through the x/y subtrees.
1272 CanPushNegate = (Opcode == ISD::OR) && CanPushNegateL && CanPushNegateR;
1278 /// Emit conjunction or disjunction tree with the CMP/FCMP followed by a chain
1279 /// of CCMP/CFCMP ops. See @ref AArch64CCMP.
1280 /// Tries to transform the given i1 producing node @p Val to a series compare
1281 /// and conditional compare operations. @returns an NZCV flags producing node
1282 /// and sets @p OutCC to the flags that should be tested or returns SDValue() if
1283 /// transformation was not possible.
1284 /// On recursive invocations @p PushNegate may be set to true to have negation
1285 /// effects pushed to the tree leafs; @p Predicate is an NZCV flag predicate
1286 /// for the comparisons in the current subtree; @p Depth limits the search
1287 /// depth to avoid stack overflow.
1288 static SDValue emitConjunctionDisjunctionTree(SelectionDAG &DAG, SDValue Val,
1289 AArch64CC::CondCode &OutCC, bool PushNegate = false,
1290 SDValue CCOp = SDValue(), AArch64CC::CondCode Predicate = AArch64CC::AL,
1291 unsigned Depth = 0) {
1292 // We're at a tree leaf, produce a conditional comparison operation.
1293 unsigned Opcode = Val->getOpcode();
1294 if (Opcode == ISD::SETCC) {
1295 SDValue LHS = Val->getOperand(0);
1296 SDValue RHS = Val->getOperand(1);
1297 ISD::CondCode CC = cast<CondCodeSDNode>(Val->getOperand(2))->get();
1298 bool isInteger = LHS.getValueType().isInteger();
1300 CC = getSetCCInverse(CC, isInteger);
1302 // Determine OutCC and handle FP special case.
1304 OutCC = changeIntCCToAArch64CC(CC);
1306 assert(LHS.getValueType().isFloatingPoint());
1307 AArch64CC::CondCode ExtraCC;
1308 changeFPCCToAArch64CC(CC, OutCC, ExtraCC);
1309 // Surpisingly some floating point conditions can't be tested with a
1310 // single condition code. Construct an additional comparison in this case.
1311 // See comment below on how we deal with OR conditions.
1312 if (ExtraCC != AArch64CC::AL) {
1314 if (!CCOp.getNode())
1315 ExtraCmp = emitComparison(LHS, RHS, CC, DL, DAG);
1317 SDValue ConditionOp = DAG.getConstant(Predicate, DL, MVT_CC);
1318 // Note that we want the inverse of ExtraCC, so NZCV is not inversed.
1319 unsigned NZCV = AArch64CC::getNZCVToSatisfyCondCode(ExtraCC);
1320 ExtraCmp = emitConditionalComparison(LHS, RHS, CC, CCOp, ConditionOp,
1324 Predicate = AArch64CC::getInvertedCondCode(ExtraCC);
1325 OutCC = AArch64CC::getInvertedCondCode(OutCC);
1329 // Produce a normal comparison if we are first in the chain
1330 if (!CCOp.getNode())
1331 return emitComparison(LHS, RHS, CC, DL, DAG);
1332 // Otherwise produce a ccmp.
1333 SDValue ConditionOp = DAG.getConstant(Predicate, DL, MVT_CC);
1334 AArch64CC::CondCode InvOutCC = AArch64CC::getInvertedCondCode(OutCC);
1335 unsigned NZCV = AArch64CC::getNZCVToSatisfyCondCode(InvOutCC);
1336 return emitConditionalComparison(LHS, RHS, CC, CCOp, ConditionOp, NZCV, DL,
1338 } else if (Opcode != ISD::AND && Opcode != ISD::OR)
1341 assert((Opcode == ISD::OR || !PushNegate)
1342 && "Can only push negate through OR operation");
1344 // Check if both sides can be transformed.
1345 SDValue LHS = Val->getOperand(0);
1346 SDValue RHS = Val->getOperand(1);
1347 bool CanPushNegateL;
1348 if (!isConjunctionDisjunctionTree(LHS, CanPushNegateL, Depth+1))
1350 bool CanPushNegateR;
1351 if (!isConjunctionDisjunctionTree(RHS, CanPushNegateR, Depth+1))
1354 // Do we need to negate our operands?
1355 bool NegateOperands = Opcode == ISD::OR;
1356 // We can negate the results of all previous operations by inverting the
1357 // predicate flags giving us a free negation for one side. For the other side
1358 // we need to be able to push the negation to the leafs of the tree.
1359 if (NegateOperands) {
1360 if (!CanPushNegateL && !CanPushNegateR)
1362 // Order the side where we can push the negate through to LHS.
1363 if (!CanPushNegateL && CanPushNegateR) {
1364 std::swap(LHS, RHS);
1365 CanPushNegateL = true;
1369 // Emit RHS. If we want to negate the tree we only need to push a negate
1370 // through if we are already in a PushNegate case, otherwise we can negate
1371 // the "flags to test" afterwards.
1372 AArch64CC::CondCode RHSCC;
1373 SDValue CmpR = emitConjunctionDisjunctionTree(DAG, RHS, RHSCC, PushNegate,
1374 CCOp, Predicate, Depth+1);
1375 if (NegateOperands && !PushNegate)
1376 RHSCC = AArch64CC::getInvertedCondCode(RHSCC);
1377 // Emit LHS. We must push the negate through if we need to negate it.
1378 SDValue CmpL = emitConjunctionDisjunctionTree(DAG, LHS, OutCC, NegateOperands,
1379 CmpR, RHSCC, Depth+1);
1380 // If we transformed an OR to and AND then we have to negate the result
1381 // (or absorb a PushNegate resulting in a double negation).
1382 if (Opcode == ISD::OR && !PushNegate)
1383 OutCC = AArch64CC::getInvertedCondCode(OutCC);
1389 static SDValue getAArch64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1390 SDValue &AArch64cc, SelectionDAG &DAG, SDLoc dl) {
1391 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1392 EVT VT = RHS.getValueType();
1393 uint64_t C = RHSC->getZExtValue();
1394 if (!isLegalArithImmed(C)) {
1395 // Constant does not fit, try adjusting it by one?
1401 if ((VT == MVT::i32 && C != 0x80000000 &&
1402 isLegalArithImmed((uint32_t)(C - 1))) ||
1403 (VT == MVT::i64 && C != 0x80000000ULL &&
1404 isLegalArithImmed(C - 1ULL))) {
1405 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1406 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1407 RHS = DAG.getConstant(C, dl, VT);
1412 if ((VT == MVT::i32 && C != 0 &&
1413 isLegalArithImmed((uint32_t)(C - 1))) ||
1414 (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) {
1415 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1416 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1417 RHS = DAG.getConstant(C, dl, VT);
1422 if ((VT == MVT::i32 && C != INT32_MAX &&
1423 isLegalArithImmed((uint32_t)(C + 1))) ||
1424 (VT == MVT::i64 && C != INT64_MAX &&
1425 isLegalArithImmed(C + 1ULL))) {
1426 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1427 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1428 RHS = DAG.getConstant(C, dl, VT);
1433 if ((VT == MVT::i32 && C != UINT32_MAX &&
1434 isLegalArithImmed((uint32_t)(C + 1))) ||
1435 (VT == MVT::i64 && C != UINT64_MAX &&
1436 isLegalArithImmed(C + 1ULL))) {
1437 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1438 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1439 RHS = DAG.getConstant(C, dl, VT);
1446 AArch64CC::CondCode AArch64CC;
1447 if ((CC == ISD::SETEQ || CC == ISD::SETNE) && isa<ConstantSDNode>(RHS)) {
1448 const ConstantSDNode *RHSC = cast<ConstantSDNode>(RHS);
1450 // The imm operand of ADDS is an unsigned immediate, in the range 0 to 4095.
1451 // For the i8 operand, the largest immediate is 255, so this can be easily
1452 // encoded in the compare instruction. For the i16 operand, however, the
1453 // largest immediate cannot be encoded in the compare.
1454 // Therefore, use a sign extending load and cmn to avoid materializing the
1455 // -1 constant. For example,
1457 // ldrh w0, [x0, #0]
1460 // ldrsh w0, [x0, #0]
1462 // Fundamental, we're relying on the property that (zext LHS) == (zext RHS)
1463 // if and only if (sext LHS) == (sext RHS). The checks are in place to
1464 // ensure both the LHS and RHS are truely zero extended and to make sure the
1465 // transformation is profitable.
1466 if ((RHSC->getZExtValue() >> 16 == 0) && isa<LoadSDNode>(LHS) &&
1467 cast<LoadSDNode>(LHS)->getExtensionType() == ISD::ZEXTLOAD &&
1468 cast<LoadSDNode>(LHS)->getMemoryVT() == MVT::i16 &&
1469 LHS.getNode()->hasNUsesOfValue(1, 0)) {
1470 int16_t ValueofRHS = cast<ConstantSDNode>(RHS)->getZExtValue();
1471 if (ValueofRHS < 0 && isLegalArithImmed(-ValueofRHS)) {
1473 DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, LHS.getValueType(), LHS,
1474 DAG.getValueType(MVT::i16));
1475 Cmp = emitComparison(SExt, DAG.getConstant(ValueofRHS, dl,
1476 RHS.getValueType()),
1478 AArch64CC = changeIntCCToAArch64CC(CC);
1482 if (!Cmp && (RHSC->isNullValue() || RHSC->isOne())) {
1483 if ((Cmp = emitConjunctionDisjunctionTree(DAG, LHS, AArch64CC))) {
1484 if ((CC == ISD::SETNE) ^ RHSC->isNullValue())
1485 AArch64CC = AArch64CC::getInvertedCondCode(AArch64CC);
1491 Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
1492 AArch64CC = changeIntCCToAArch64CC(CC);
1494 AArch64cc = DAG.getConstant(AArch64CC, dl, MVT_CC);
1498 static std::pair<SDValue, SDValue>
1499 getAArch64XALUOOp(AArch64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
1500 assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
1501 "Unsupported value type");
1502 SDValue Value, Overflow;
1504 SDValue LHS = Op.getOperand(0);
1505 SDValue RHS = Op.getOperand(1);
1507 switch (Op.getOpcode()) {
1509 llvm_unreachable("Unknown overflow instruction!");
1511 Opc = AArch64ISD::ADDS;
1515 Opc = AArch64ISD::ADDS;
1519 Opc = AArch64ISD::SUBS;
1523 Opc = AArch64ISD::SUBS;
1526 // Multiply needs a little bit extra work.
1530 bool IsSigned = Op.getOpcode() == ISD::SMULO;
1531 if (Op.getValueType() == MVT::i32) {
1532 unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1533 // For a 32 bit multiply with overflow check we want the instruction
1534 // selector to generate a widening multiply (SMADDL/UMADDL). For that we
1535 // need to generate the following pattern:
1536 // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b))
1537 LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
1538 RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
1539 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1540 SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul,
1541 DAG.getConstant(0, DL, MVT::i64));
1542 // On AArch64 the upper 32 bits are always zero extended for a 32 bit
1543 // operation. We need to clear out the upper 32 bits, because we used a
1544 // widening multiply that wrote all 64 bits. In the end this should be a
1546 Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add);
1548 // The signed overflow check requires more than just a simple check for
1549 // any bit set in the upper 32 bits of the result. These bits could be
1550 // just the sign bits of a negative number. To perform the overflow
1551 // check we have to arithmetic shift right the 32nd bit of the result by
1552 // 31 bits. Then we compare the result to the upper 32 bits.
1553 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add,
1554 DAG.getConstant(32, DL, MVT::i64));
1555 UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits);
1556 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value,
1557 DAG.getConstant(31, DL, MVT::i64));
1558 // It is important that LowerBits is last, otherwise the arithmetic
1559 // shift will not be folded into the compare (SUBS).
1560 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32);
1561 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1564 // The overflow check for unsigned multiply is easy. We only need to
1565 // check if any of the upper 32 bits are set. This can be done with a
1566 // CMP (shifted register). For that we need to generate the following
1568 // (i64 AArch64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32)
1569 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
1570 DAG.getConstant(32, DL, MVT::i64));
1571 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1573 DAG.getNode(AArch64ISD::SUBS, DL, VTs,
1574 DAG.getConstant(0, DL, MVT::i64),
1575 UpperBits).getValue(1);
1579 assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
1580 // For the 64 bit multiply
1581 Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1583 SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
1584 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
1585 DAG.getConstant(63, DL, MVT::i64));
1586 // It is important that LowerBits is last, otherwise the arithmetic
1587 // shift will not be folded into the compare (SUBS).
1588 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1589 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1592 SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
1593 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1595 DAG.getNode(AArch64ISD::SUBS, DL, VTs,
1596 DAG.getConstant(0, DL, MVT::i64),
1597 UpperBits).getValue(1);
1604 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);
1606 // Emit the AArch64 operation with overflow check.
1607 Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
1608 Overflow = Value.getValue(1);
1610 return std::make_pair(Value, Overflow);
1613 SDValue AArch64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG,
1614 RTLIB::Libcall Call) const {
1615 SmallVector<SDValue, 2> Ops(Op->op_begin(), Op->op_end());
1616 return makeLibCall(DAG, Call, MVT::f128, &Ops[0], Ops.size(), false,
1620 static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) {
1621 SDValue Sel = Op.getOperand(0);
1622 SDValue Other = Op.getOperand(1);
1624 // If neither operand is a SELECT_CC, give up.
1625 if (Sel.getOpcode() != ISD::SELECT_CC)
1626 std::swap(Sel, Other);
1627 if (Sel.getOpcode() != ISD::SELECT_CC)
1630 // The folding we want to perform is:
1631 // (xor x, (select_cc a, b, cc, 0, -1) )
1633 // (csel x, (xor x, -1), cc ...)
1635 // The latter will get matched to a CSINV instruction.
1637 ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
1638 SDValue LHS = Sel.getOperand(0);
1639 SDValue RHS = Sel.getOperand(1);
1640 SDValue TVal = Sel.getOperand(2);
1641 SDValue FVal = Sel.getOperand(3);
1644 // FIXME: This could be generalized to non-integer comparisons.
1645 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
1648 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
1649 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
1651 // The values aren't constants, this isn't the pattern we're looking for.
1652 if (!CFVal || !CTVal)
1655 // We can commute the SELECT_CC by inverting the condition. This
1656 // might be needed to make this fit into a CSINV pattern.
1657 if (CTVal->isAllOnesValue() && CFVal->isNullValue()) {
1658 std::swap(TVal, FVal);
1659 std::swap(CTVal, CFVal);
1660 CC = ISD::getSetCCInverse(CC, true);
1663 // If the constants line up, perform the transform!
1664 if (CTVal->isNullValue() && CFVal->isAllOnesValue()) {
1666 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
1669 TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other,
1670 DAG.getConstant(-1ULL, dl, Other.getValueType()));
1672 return DAG.getNode(AArch64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal,
1679 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
1680 EVT VT = Op.getValueType();
1682 // Let legalize expand this if it isn't a legal type yet.
1683 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
1686 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
1689 bool ExtraOp = false;
1690 switch (Op.getOpcode()) {
1692 llvm_unreachable("Invalid code");
1694 Opc = AArch64ISD::ADDS;
1697 Opc = AArch64ISD::SUBS;
1700 Opc = AArch64ISD::ADCS;
1704 Opc = AArch64ISD::SBCS;
1710 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1));
1711 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1),
1715 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
1716 // Let legalize expand this if it isn't a legal type yet.
1717 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
1721 AArch64CC::CondCode CC;
1722 // The actual operation that sets the overflow or carry flag.
1723 SDValue Value, Overflow;
1724 std::tie(Value, Overflow) = getAArch64XALUOOp(CC, Op, DAG);
1726 // We use 0 and 1 as false and true values.
1727 SDValue TVal = DAG.getConstant(1, dl, MVT::i32);
1728 SDValue FVal = DAG.getConstant(0, dl, MVT::i32);
1730 // We use an inverted condition, because the conditional select is inverted
1731 // too. This will allow it to be selected to a single instruction:
1732 // CSINC Wd, WZR, WZR, invert(cond).
1733 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), dl, MVT::i32);
1734 Overflow = DAG.getNode(AArch64ISD::CSEL, dl, MVT::i32, FVal, TVal,
1737 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
1738 return DAG.getNode(ISD::MERGE_VALUES, dl, VTs, Value, Overflow);
1741 // Prefetch operands are:
1742 // 1: Address to prefetch
1744 // 3: int locality (0 = no locality ... 3 = extreme locality)
1745 // 4: bool isDataCache
1746 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
1748 unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
1749 unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
1750 unsigned IsData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
1752 bool IsStream = !Locality;
1753 // When the locality number is set
1755 // The front-end should have filtered out the out-of-range values
1756 assert(Locality <= 3 && "Prefetch locality out-of-range");
1757 // The locality degree is the opposite of the cache speed.
1758 // Put the number the other way around.
1759 // The encoding starts at 0 for level 1
1760 Locality = 3 - Locality;
1763 // built the mask value encoding the expected behavior.
1764 unsigned PrfOp = (IsWrite << 4) | // Load/Store bit
1765 (!IsData << 3) | // IsDataCache bit
1766 (Locality << 1) | // Cache level bits
1767 (unsigned)IsStream; // Stream bit
1768 return DAG.getNode(AArch64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
1769 DAG.getConstant(PrfOp, DL, MVT::i32), Op.getOperand(1));
1772 SDValue AArch64TargetLowering::LowerFP_EXTEND(SDValue Op,
1773 SelectionDAG &DAG) const {
1774 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
1777 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
1779 return LowerF128Call(Op, DAG, LC);
1782 SDValue AArch64TargetLowering::LowerFP_ROUND(SDValue Op,
1783 SelectionDAG &DAG) const {
1784 if (Op.getOperand(0).getValueType() != MVT::f128) {
1785 // It's legal except when f128 is involved
1790 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
1792 // FP_ROUND node has a second operand indicating whether it is known to be
1793 // precise. That doesn't take part in the LibCall so we can't directly use
1795 SDValue SrcVal = Op.getOperand(0);
1796 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
1797 /*isSigned*/ false, SDLoc(Op)).first;
1800 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1801 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1802 // Any additional optimization in this function should be recorded
1803 // in the cost tables.
1804 EVT InVT = Op.getOperand(0).getValueType();
1805 EVT VT = Op.getValueType();
1807 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1810 DAG.getNode(Op.getOpcode(), dl, InVT.changeVectorElementTypeToInteger(),
1812 return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
1815 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1818 MVT::getVectorVT(MVT::getFloatingPointVT(VT.getScalarSizeInBits()),
1819 VT.getVectorNumElements());
1820 SDValue Ext = DAG.getNode(ISD::FP_EXTEND, dl, ExtVT, Op.getOperand(0));
1821 return DAG.getNode(Op.getOpcode(), dl, VT, Ext);
1824 // Type changing conversions are illegal.
1828 SDValue AArch64TargetLowering::LowerFP_TO_INT(SDValue Op,
1829 SelectionDAG &DAG) const {
1830 if (Op.getOperand(0).getValueType().isVector())
1831 return LowerVectorFP_TO_INT(Op, DAG);
1833 // f16 conversions are promoted to f32.
1834 if (Op.getOperand(0).getValueType() == MVT::f16) {
1837 Op.getOpcode(), dl, Op.getValueType(),
1838 DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, Op.getOperand(0)));
1841 if (Op.getOperand(0).getValueType() != MVT::f128) {
1842 // It's legal except when f128 is involved
1847 if (Op.getOpcode() == ISD::FP_TO_SINT)
1848 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
1850 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
1852 SmallVector<SDValue, 2> Ops(Op->op_begin(), Op->op_end());
1853 return makeLibCall(DAG, LC, Op.getValueType(), &Ops[0], Ops.size(), false,
1857 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1858 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1859 // Any additional optimization in this function should be recorded
1860 // in the cost tables.
1861 EVT VT = Op.getValueType();
1863 SDValue In = Op.getOperand(0);
1864 EVT InVT = In.getValueType();
1866 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1868 MVT::getVectorVT(MVT::getFloatingPointVT(InVT.getScalarSizeInBits()),
1869 InVT.getVectorNumElements());
1870 In = DAG.getNode(Op.getOpcode(), dl, CastVT, In);
1871 return DAG.getNode(ISD::FP_ROUND, dl, VT, In, DAG.getIntPtrConstant(0, dl));
1874 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1876 Op.getOpcode() == ISD::SINT_TO_FP ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1877 EVT CastVT = VT.changeVectorElementTypeToInteger();
1878 In = DAG.getNode(CastOpc, dl, CastVT, In);
1879 return DAG.getNode(Op.getOpcode(), dl, VT, In);
1885 SDValue AArch64TargetLowering::LowerINT_TO_FP(SDValue Op,
1886 SelectionDAG &DAG) const {
1887 if (Op.getValueType().isVector())
1888 return LowerVectorINT_TO_FP(Op, DAG);
1890 // f16 conversions are promoted to f32.
1891 if (Op.getValueType() == MVT::f16) {
1894 ISD::FP_ROUND, dl, MVT::f16,
1895 DAG.getNode(Op.getOpcode(), dl, MVT::f32, Op.getOperand(0)),
1896 DAG.getIntPtrConstant(0, dl));
1899 // i128 conversions are libcalls.
1900 if (Op.getOperand(0).getValueType() == MVT::i128)
1903 // Other conversions are legal, unless it's to the completely software-based
1905 if (Op.getValueType() != MVT::f128)
1909 if (Op.getOpcode() == ISD::SINT_TO_FP)
1910 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1912 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1914 return LowerF128Call(Op, DAG, LC);
1917 SDValue AArch64TargetLowering::LowerFSINCOS(SDValue Op,
1918 SelectionDAG &DAG) const {
1919 // For iOS, we want to call an alternative entry point: __sincos_stret,
1920 // which returns the values in two S / D registers.
1922 SDValue Arg = Op.getOperand(0);
1923 EVT ArgVT = Arg.getValueType();
1924 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1931 Entry.isSExt = false;
1932 Entry.isZExt = false;
1933 Args.push_back(Entry);
1935 const char *LibcallName =
1936 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
1938 DAG.getExternalSymbol(LibcallName, getPointerTy(DAG.getDataLayout()));
1940 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
1941 TargetLowering::CallLoweringInfo CLI(DAG);
1942 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
1943 .setCallee(CallingConv::Fast, RetTy, Callee, std::move(Args), 0);
1945 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1946 return CallResult.first;
1949 static SDValue LowerBITCAST(SDValue Op, SelectionDAG &DAG) {
1950 if (Op.getValueType() != MVT::f16)
1953 assert(Op.getOperand(0).getValueType() == MVT::i16);
1956 Op = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op.getOperand(0));
1957 Op = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Op);
1959 DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL, MVT::f16, Op,
1960 DAG.getTargetConstant(AArch64::hsub, DL, MVT::i32)),
1964 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
1965 if (OrigVT.getSizeInBits() >= 64)
1968 assert(OrigVT.isSimple() && "Expecting a simple value type");
1970 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
1971 switch (OrigSimpleTy) {
1972 default: llvm_unreachable("Unexpected Vector Type");
1981 static SDValue addRequiredExtensionForVectorMULL(SDValue N, SelectionDAG &DAG,
1984 unsigned ExtOpcode) {
1985 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
1986 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
1987 // 64-bits we need to insert a new extension so that it will be 64-bits.
1988 assert(ExtTy.is128BitVector() && "Unexpected extension size");
1989 if (OrigTy.getSizeInBits() >= 64)
1992 // Must extend size to at least 64 bits to be used as an operand for VMULL.
1993 EVT NewVT = getExtensionTo64Bits(OrigTy);
1995 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
1998 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
2000 EVT VT = N->getValueType(0);
2002 if (N->getOpcode() != ISD::BUILD_VECTOR)
2005 for (const SDValue &Elt : N->op_values()) {
2006 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
2007 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
2008 unsigned HalfSize = EltSize / 2;
2010 if (!isIntN(HalfSize, C->getSExtValue()))
2013 if (!isUIntN(HalfSize, C->getZExtValue()))
2024 static SDValue skipExtensionForVectorMULL(SDNode *N, SelectionDAG &DAG) {
2025 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
2026 return addRequiredExtensionForVectorMULL(N->getOperand(0), DAG,
2027 N->getOperand(0)->getValueType(0),
2031 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
2032 EVT VT = N->getValueType(0);
2034 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
2035 unsigned NumElts = VT.getVectorNumElements();
2036 MVT TruncVT = MVT::getIntegerVT(EltSize);
2037 SmallVector<SDValue, 8> Ops;
2038 for (unsigned i = 0; i != NumElts; ++i) {
2039 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
2040 const APInt &CInt = C->getAPIntValue();
2041 // Element types smaller than 32 bits are not legal, so use i32 elements.
2042 // The values are implicitly truncated so sext vs. zext doesn't matter.
2043 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), dl, MVT::i32));
2045 return DAG.getNode(ISD::BUILD_VECTOR, dl,
2046 MVT::getVectorVT(TruncVT, NumElts), Ops);
2049 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
2050 if (N->getOpcode() == ISD::SIGN_EXTEND)
2052 if (isExtendedBUILD_VECTOR(N, DAG, true))
2057 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
2058 if (N->getOpcode() == ISD::ZERO_EXTEND)
2060 if (isExtendedBUILD_VECTOR(N, DAG, false))
2065 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
2066 unsigned Opcode = N->getOpcode();
2067 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
2068 SDNode *N0 = N->getOperand(0).getNode();
2069 SDNode *N1 = N->getOperand(1).getNode();
2070 return N0->hasOneUse() && N1->hasOneUse() &&
2071 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
2076 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
2077 unsigned Opcode = N->getOpcode();
2078 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
2079 SDNode *N0 = N->getOperand(0).getNode();
2080 SDNode *N1 = N->getOperand(1).getNode();
2081 return N0->hasOneUse() && N1->hasOneUse() &&
2082 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
2087 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
2088 // Multiplications are only custom-lowered for 128-bit vectors so that
2089 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
2090 EVT VT = Op.getValueType();
2091 assert(VT.is128BitVector() && VT.isInteger() &&
2092 "unexpected type for custom-lowering ISD::MUL");
2093 SDNode *N0 = Op.getOperand(0).getNode();
2094 SDNode *N1 = Op.getOperand(1).getNode();
2095 unsigned NewOpc = 0;
2097 bool isN0SExt = isSignExtended(N0, DAG);
2098 bool isN1SExt = isSignExtended(N1, DAG);
2099 if (isN0SExt && isN1SExt)
2100 NewOpc = AArch64ISD::SMULL;
2102 bool isN0ZExt = isZeroExtended(N0, DAG);
2103 bool isN1ZExt = isZeroExtended(N1, DAG);
2104 if (isN0ZExt && isN1ZExt)
2105 NewOpc = AArch64ISD::UMULL;
2106 else if (isN1SExt || isN1ZExt) {
2107 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
2108 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
2109 if (isN1SExt && isAddSubSExt(N0, DAG)) {
2110 NewOpc = AArch64ISD::SMULL;
2112 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
2113 NewOpc = AArch64ISD::UMULL;
2115 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
2117 NewOpc = AArch64ISD::UMULL;
2123 if (VT == MVT::v2i64)
2124 // Fall through to expand this. It is not legal.
2127 // Other vector multiplications are legal.
2132 // Legalize to a S/UMULL instruction
2135 SDValue Op1 = skipExtensionForVectorMULL(N1, DAG);
2137 Op0 = skipExtensionForVectorMULL(N0, DAG);
2138 assert(Op0.getValueType().is64BitVector() &&
2139 Op1.getValueType().is64BitVector() &&
2140 "unexpected types for extended operands to VMULL");
2141 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
2143 // Optimizing (zext A + zext B) * C, to (S/UMULL A, C) + (S/UMULL B, C) during
2144 // isel lowering to take advantage of no-stall back to back s/umul + s/umla.
2145 // This is true for CPUs with accumulate forwarding such as Cortex-A53/A57
2146 SDValue N00 = skipExtensionForVectorMULL(N0->getOperand(0).getNode(), DAG);
2147 SDValue N01 = skipExtensionForVectorMULL(N0->getOperand(1).getNode(), DAG);
2148 EVT Op1VT = Op1.getValueType();
2149 return DAG.getNode(N0->getOpcode(), DL, VT,
2150 DAG.getNode(NewOpc, DL, VT,
2151 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
2152 DAG.getNode(NewOpc, DL, VT,
2153 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
2156 SDValue AArch64TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
2157 SelectionDAG &DAG) const {
2158 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2161 default: return SDValue(); // Don't custom lower most intrinsics.
2162 case Intrinsic::aarch64_thread_pointer: {
2163 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2164 return DAG.getNode(AArch64ISD::THREAD_POINTER, dl, PtrVT);
2169 SDValue AArch64TargetLowering::LowerOperation(SDValue Op,
2170 SelectionDAG &DAG) const {
2171 switch (Op.getOpcode()) {
2173 llvm_unreachable("unimplemented operand");
2176 return LowerBITCAST(Op, DAG);
2177 case ISD::GlobalAddress:
2178 return LowerGlobalAddress(Op, DAG);
2179 case ISD::GlobalTLSAddress:
2180 return LowerGlobalTLSAddress(Op, DAG);
2182 return LowerSETCC(Op, DAG);
2184 return LowerBR_CC(Op, DAG);
2186 return LowerSELECT(Op, DAG);
2187 case ISD::SELECT_CC:
2188 return LowerSELECT_CC(Op, DAG);
2189 case ISD::JumpTable:
2190 return LowerJumpTable(Op, DAG);
2191 case ISD::ConstantPool:
2192 return LowerConstantPool(Op, DAG);
2193 case ISD::BlockAddress:
2194 return LowerBlockAddress(Op, DAG);
2196 return LowerVASTART(Op, DAG);
2198 return LowerVACOPY(Op, DAG);
2200 return LowerVAARG(Op, DAG);
2205 return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
2212 return LowerXALUO(Op, DAG);
2214 return LowerF128Call(Op, DAG, RTLIB::ADD_F128);
2216 return LowerF128Call(Op, DAG, RTLIB::SUB_F128);
2218 return LowerF128Call(Op, DAG, RTLIB::MUL_F128);
2220 return LowerF128Call(Op, DAG, RTLIB::DIV_F128);
2222 return LowerFP_ROUND(Op, DAG);
2223 case ISD::FP_EXTEND:
2224 return LowerFP_EXTEND(Op, DAG);
2225 case ISD::FRAMEADDR:
2226 return LowerFRAMEADDR(Op, DAG);
2227 case ISD::RETURNADDR:
2228 return LowerRETURNADDR(Op, DAG);
2229 case ISD::INSERT_VECTOR_ELT:
2230 return LowerINSERT_VECTOR_ELT(Op, DAG);
2231 case ISD::EXTRACT_VECTOR_ELT:
2232 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2233 case ISD::BUILD_VECTOR:
2234 return LowerBUILD_VECTOR(Op, DAG);
2235 case ISD::VECTOR_SHUFFLE:
2236 return LowerVECTOR_SHUFFLE(Op, DAG);
2237 case ISD::EXTRACT_SUBVECTOR:
2238 return LowerEXTRACT_SUBVECTOR(Op, DAG);
2242 return LowerVectorSRA_SRL_SHL(Op, DAG);
2243 case ISD::SHL_PARTS:
2244 return LowerShiftLeftParts(Op, DAG);
2245 case ISD::SRL_PARTS:
2246 case ISD::SRA_PARTS:
2247 return LowerShiftRightParts(Op, DAG);
2249 return LowerCTPOP(Op, DAG);
2250 case ISD::FCOPYSIGN:
2251 return LowerFCOPYSIGN(Op, DAG);
2253 return LowerVectorAND(Op, DAG);
2255 return LowerVectorOR(Op, DAG);
2257 return LowerXOR(Op, DAG);
2259 return LowerPREFETCH(Op, DAG);
2260 case ISD::SINT_TO_FP:
2261 case ISD::UINT_TO_FP:
2262 return LowerINT_TO_FP(Op, DAG);
2263 case ISD::FP_TO_SINT:
2264 case ISD::FP_TO_UINT:
2265 return LowerFP_TO_INT(Op, DAG);
2267 return LowerFSINCOS(Op, DAG);
2269 return LowerMUL(Op, DAG);
2270 case ISD::INTRINSIC_WO_CHAIN:
2271 return LowerINTRINSIC_WO_CHAIN(Op, DAG);
2275 /// getFunctionAlignment - Return the Log2 alignment of this function.
2276 unsigned AArch64TargetLowering::getFunctionAlignment(const Function *F) const {
2280 //===----------------------------------------------------------------------===//
2281 // Calling Convention Implementation
2282 //===----------------------------------------------------------------------===//
2284 #include "AArch64GenCallingConv.inc"
2286 /// Selects the correct CCAssignFn for a given CallingConvention value.
2287 CCAssignFn *AArch64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
2288 bool IsVarArg) const {
2291 llvm_unreachable("Unsupported calling convention.");
2292 case CallingConv::WebKit_JS:
2293 return CC_AArch64_WebKit_JS;
2294 case CallingConv::GHC:
2295 return CC_AArch64_GHC;
2296 case CallingConv::C:
2297 case CallingConv::Fast:
2298 if (!Subtarget->isTargetDarwin())
2299 return CC_AArch64_AAPCS;
2300 return IsVarArg ? CC_AArch64_DarwinPCS_VarArg : CC_AArch64_DarwinPCS;
2304 SDValue AArch64TargetLowering::LowerFormalArguments(
2305 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2306 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2307 SmallVectorImpl<SDValue> &InVals) const {
2308 MachineFunction &MF = DAG.getMachineFunction();
2309 MachineFrameInfo *MFI = MF.getFrameInfo();
2311 // Assign locations to all of the incoming arguments.
2312 SmallVector<CCValAssign, 16> ArgLocs;
2313 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
2316 // At this point, Ins[].VT may already be promoted to i32. To correctly
2317 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2318 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
2319 // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
2320 // we use a special version of AnalyzeFormalArguments to pass in ValVT and
2322 unsigned NumArgs = Ins.size();
2323 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2324 unsigned CurArgIdx = 0;
2325 for (unsigned i = 0; i != NumArgs; ++i) {
2326 MVT ValVT = Ins[i].VT;
2327 if (Ins[i].isOrigArg()) {
2328 std::advance(CurOrigArg, Ins[i].getOrigArgIndex() - CurArgIdx);
2329 CurArgIdx = Ins[i].getOrigArgIndex();
2331 // Get type of the original argument.
2332 EVT ActualVT = getValueType(DAG.getDataLayout(), CurOrigArg->getType(),
2333 /*AllowUnknown*/ true);
2334 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
2335 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2336 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2338 else if (ActualMVT == MVT::i16)
2341 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2343 AssignFn(i, ValVT, ValVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
2344 assert(!Res && "Call operand has unhandled type");
2347 assert(ArgLocs.size() == Ins.size());
2348 SmallVector<SDValue, 16> ArgValues;
2349 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2350 CCValAssign &VA = ArgLocs[i];
2352 if (Ins[i].Flags.isByVal()) {
2353 // Byval is used for HFAs in the PCS, but the system should work in a
2354 // non-compliant manner for larger structs.
2355 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2356 int Size = Ins[i].Flags.getByValSize();
2357 unsigned NumRegs = (Size + 7) / 8;
2359 // FIXME: This works on big-endian for composite byvals, which are the common
2360 // case. It should also work for fundamental types too.
2362 MFI->CreateFixedObject(8 * NumRegs, VA.getLocMemOffset(), false);
2363 SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrVT);
2364 InVals.push_back(FrameIdxN);
2369 if (VA.isRegLoc()) {
2370 // Arguments stored in registers.
2371 EVT RegVT = VA.getLocVT();
2374 const TargetRegisterClass *RC;
2376 if (RegVT == MVT::i32)
2377 RC = &AArch64::GPR32RegClass;
2378 else if (RegVT == MVT::i64)
2379 RC = &AArch64::GPR64RegClass;
2380 else if (RegVT == MVT::f16)
2381 RC = &AArch64::FPR16RegClass;
2382 else if (RegVT == MVT::f32)
2383 RC = &AArch64::FPR32RegClass;
2384 else if (RegVT == MVT::f64 || RegVT.is64BitVector())
2385 RC = &AArch64::FPR64RegClass;
2386 else if (RegVT == MVT::f128 || RegVT.is128BitVector())
2387 RC = &AArch64::FPR128RegClass;
2389 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2391 // Transform the arguments in physical registers into virtual ones.
2392 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2393 ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2395 // If this is an 8, 16 or 32-bit value, it is really passed promoted
2396 // to 64 bits. Insert an assert[sz]ext to capture this, then
2397 // truncate to the right size.
2398 switch (VA.getLocInfo()) {
2400 llvm_unreachable("Unknown loc info!");
2401 case CCValAssign::Full:
2403 case CCValAssign::BCvt:
2404 ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue);
2406 case CCValAssign::AExt:
2407 case CCValAssign::SExt:
2408 case CCValAssign::ZExt:
2409 // SelectionDAGBuilder will insert appropriate AssertZExt & AssertSExt
2410 // nodes after our lowering.
2411 assert(RegVT == Ins[i].VT && "incorrect register location selected");
2415 InVals.push_back(ArgValue);
2417 } else { // VA.isRegLoc()
2418 assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
2419 unsigned ArgOffset = VA.getLocMemOffset();
2420 unsigned ArgSize = VA.getValVT().getSizeInBits() / 8;
2422 uint32_t BEAlign = 0;
2423 if (!Subtarget->isLittleEndian() && ArgSize < 8 &&
2424 !Ins[i].Flags.isInConsecutiveRegs())
2425 BEAlign = 8 - ArgSize;
2427 int FI = MFI->CreateFixedObject(ArgSize, ArgOffset + BEAlign, true);
2429 // Create load nodes to retrieve arguments from the stack.
2430 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2433 // For NON_EXTLOAD, generic code in getLoad assert(ValVT == MemVT)
2434 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
2435 MVT MemVT = VA.getValVT();
2437 switch (VA.getLocInfo()) {
2440 case CCValAssign::BCvt:
2441 MemVT = VA.getLocVT();
2443 case CCValAssign::SExt:
2444 ExtType = ISD::SEXTLOAD;
2446 case CCValAssign::ZExt:
2447 ExtType = ISD::ZEXTLOAD;
2449 case CCValAssign::AExt:
2450 ExtType = ISD::EXTLOAD;
2454 ArgValue = DAG.getExtLoad(ExtType, DL, VA.getLocVT(), Chain, FIN,
2455 MachinePointerInfo::getFixedStack(FI),
2456 MemVT, false, false, false, 0);
2458 InVals.push_back(ArgValue);
2464 if (!Subtarget->isTargetDarwin()) {
2465 // The AAPCS variadic function ABI is identical to the non-variadic
2466 // one. As a result there may be more arguments in registers and we should
2467 // save them for future reference.
2468 saveVarArgRegisters(CCInfo, DAG, DL, Chain);
2471 AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>();
2472 // This will point to the next argument passed via stack.
2473 unsigned StackOffset = CCInfo.getNextStackOffset();
2474 // We currently pass all varargs at 8-byte alignment.
2475 StackOffset = ((StackOffset + 7) & ~7);
2476 AFI->setVarArgsStackIndex(MFI->CreateFixedObject(4, StackOffset, true));
2479 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2480 unsigned StackArgSize = CCInfo.getNextStackOffset();
2481 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2482 if (DoesCalleeRestoreStack(CallConv, TailCallOpt)) {
2483 // This is a non-standard ABI so by fiat I say we're allowed to make full
2484 // use of the stack area to be popped, which must be aligned to 16 bytes in
2486 StackArgSize = RoundUpToAlignment(StackArgSize, 16);
2488 // If we're expected to restore the stack (e.g. fastcc) then we'll be adding
2489 // a multiple of 16.
2490 FuncInfo->setArgumentStackToRestore(StackArgSize);
2492 // This realignment carries over to the available bytes below. Our own
2493 // callers will guarantee the space is free by giving an aligned value to
2496 // Even if we're not expected to free up the space, it's useful to know how
2497 // much is there while considering tail calls (because we can reuse it).
2498 FuncInfo->setBytesInStackArgArea(StackArgSize);
2503 void AArch64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
2504 SelectionDAG &DAG, SDLoc DL,
2505 SDValue &Chain) const {
2506 MachineFunction &MF = DAG.getMachineFunction();
2507 MachineFrameInfo *MFI = MF.getFrameInfo();
2508 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2509 auto PtrVT = getPointerTy(DAG.getDataLayout());
2511 SmallVector<SDValue, 8> MemOps;
2513 static const MCPhysReg GPRArgRegs[] = { AArch64::X0, AArch64::X1, AArch64::X2,
2514 AArch64::X3, AArch64::X4, AArch64::X5,
2515 AArch64::X6, AArch64::X7 };
2516 static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs);
2517 unsigned FirstVariadicGPR = CCInfo.getFirstUnallocated(GPRArgRegs);
2519 unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
2521 if (GPRSaveSize != 0) {
2522 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
2524 SDValue FIN = DAG.getFrameIndex(GPRIdx, PtrVT);
2526 for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
2527 unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &AArch64::GPR64RegClass);
2528 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
2530 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2531 MachinePointerInfo::getStack(i * 8), false, false, 0);
2532 MemOps.push_back(Store);
2534 DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getConstant(8, DL, PtrVT));
2537 FuncInfo->setVarArgsGPRIndex(GPRIdx);
2538 FuncInfo->setVarArgsGPRSize(GPRSaveSize);
2540 if (Subtarget->hasFPARMv8()) {
2541 static const MCPhysReg FPRArgRegs[] = {
2542 AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3,
2543 AArch64::Q4, AArch64::Q5, AArch64::Q6, AArch64::Q7};
2544 static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs);
2545 unsigned FirstVariadicFPR = CCInfo.getFirstUnallocated(FPRArgRegs);
2547 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
2549 if (FPRSaveSize != 0) {
2550 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
2552 SDValue FIN = DAG.getFrameIndex(FPRIdx, PtrVT);
2554 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
2555 unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &AArch64::FPR128RegClass);
2556 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f128);
2559 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2560 MachinePointerInfo::getStack(i * 16), false, false, 0);
2561 MemOps.push_back(Store);
2562 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN,
2563 DAG.getConstant(16, DL, PtrVT));
2566 FuncInfo->setVarArgsFPRIndex(FPRIdx);
2567 FuncInfo->setVarArgsFPRSize(FPRSaveSize);
2570 if (!MemOps.empty()) {
2571 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
2575 /// LowerCallResult - Lower the result values of a call into the
2576 /// appropriate copies out of appropriate physical registers.
2577 SDValue AArch64TargetLowering::LowerCallResult(
2578 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
2579 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2580 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
2581 SDValue ThisVal) const {
2582 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2583 ? RetCC_AArch64_WebKit_JS
2584 : RetCC_AArch64_AAPCS;
2585 // Assign locations to each value returned by this call.
2586 SmallVector<CCValAssign, 16> RVLocs;
2587 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2589 CCInfo.AnalyzeCallResult(Ins, RetCC);
2591 // Copy all of the result registers out of their specified physreg.
2592 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2593 CCValAssign VA = RVLocs[i];
2595 // Pass 'this' value directly from the argument to return value, to avoid
2596 // reg unit interference
2597 if (i == 0 && isThisReturn) {
2598 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 &&
2599 "unexpected return calling convention register assignment");
2600 InVals.push_back(ThisVal);
2605 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag);
2606 Chain = Val.getValue(1);
2607 InFlag = Val.getValue(2);
2609 switch (VA.getLocInfo()) {
2611 llvm_unreachable("Unknown loc info!");
2612 case CCValAssign::Full:
2614 case CCValAssign::BCvt:
2615 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
2619 InVals.push_back(Val);
2625 bool AArch64TargetLowering::isEligibleForTailCallOptimization(
2626 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
2627 bool isCalleeStructRet, bool isCallerStructRet,
2628 const SmallVectorImpl<ISD::OutputArg> &Outs,
2629 const SmallVectorImpl<SDValue> &OutVals,
2630 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
2631 // For CallingConv::C this function knows whether the ABI needs
2632 // changing. That's not true for other conventions so they will have to opt in
2634 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
2637 const MachineFunction &MF = DAG.getMachineFunction();
2638 const Function *CallerF = MF.getFunction();
2639 CallingConv::ID CallerCC = CallerF->getCallingConv();
2640 bool CCMatch = CallerCC == CalleeCC;
2642 // Byval parameters hand the function a pointer directly into the stack area
2643 // we want to reuse during a tail call. Working around this *is* possible (see
2644 // X86) but less efficient and uglier in LowerCall.
2645 for (Function::const_arg_iterator i = CallerF->arg_begin(),
2646 e = CallerF->arg_end();
2648 if (i->hasByValAttr())
2651 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
2652 if (IsTailCallConvention(CalleeCC) && CCMatch)
2657 // Externally-defined functions with weak linkage should not be
2658 // tail-called on AArch64 when the OS does not support dynamic
2659 // pre-emption of symbols, as the AAELF spec requires normal calls
2660 // to undefined weak functions to be replaced with a NOP or jump to the
2661 // next instruction. The behaviour of branch instructions in this
2662 // situation (as used for tail calls) is implementation-defined, so we
2663 // cannot rely on the linker replacing the tail call with a return.
2664 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2665 const GlobalValue *GV = G->getGlobal();
2666 const Triple &TT = getTargetMachine().getTargetTriple();
2667 if (GV->hasExternalWeakLinkage() &&
2668 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2672 // Now we search for cases where we can use a tail call without changing the
2673 // ABI. Sibcall is used in some places (particularly gcc) to refer to this
2676 // I want anyone implementing a new calling convention to think long and hard
2677 // about this assert.
2678 assert((!isVarArg || CalleeCC == CallingConv::C) &&
2679 "Unexpected variadic calling convention");
2681 if (isVarArg && !Outs.empty()) {
2682 // At least two cases here: if caller is fastcc then we can't have any
2683 // memory arguments (we'd be expected to clean up the stack afterwards). If
2684 // caller is C then we could potentially use its argument area.
2686 // FIXME: for now we take the most conservative of these in both cases:
2687 // disallow all variadic memory operands.
2688 SmallVector<CCValAssign, 16> ArgLocs;
2689 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2692 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, true));
2693 for (const CCValAssign &ArgLoc : ArgLocs)
2694 if (!ArgLoc.isRegLoc())
2698 // If the calling conventions do not match, then we'd better make sure the
2699 // results are returned in the same way as what the caller expects.
2701 SmallVector<CCValAssign, 16> RVLocs1;
2702 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2704 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForCall(CalleeCC, isVarArg));
2706 SmallVector<CCValAssign, 16> RVLocs2;
2707 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2709 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForCall(CallerCC, isVarArg));
2711 if (RVLocs1.size() != RVLocs2.size())
2713 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2714 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2716 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2718 if (RVLocs1[i].isRegLoc()) {
2719 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2722 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2728 // Nothing more to check if the callee is taking no arguments
2732 SmallVector<CCValAssign, 16> ArgLocs;
2733 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2736 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, isVarArg));
2738 const AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2740 // If the stack arguments for this call would fit into our own save area then
2741 // the call can be made tail.
2742 return CCInfo.getNextStackOffset() <= FuncInfo->getBytesInStackArgArea();
2745 SDValue AArch64TargetLowering::addTokenForArgument(SDValue Chain,
2747 MachineFrameInfo *MFI,
2748 int ClobberedFI) const {
2749 SmallVector<SDValue, 8> ArgChains;
2750 int64_t FirstByte = MFI->getObjectOffset(ClobberedFI);
2751 int64_t LastByte = FirstByte + MFI->getObjectSize(ClobberedFI) - 1;
2753 // Include the original chain at the beginning of the list. When this is
2754 // used by target LowerCall hooks, this helps legalize find the
2755 // CALLSEQ_BEGIN node.
2756 ArgChains.push_back(Chain);
2758 // Add a chain value for each stack argument corresponding
2759 for (SDNode::use_iterator U = DAG.getEntryNode().getNode()->use_begin(),
2760 UE = DAG.getEntryNode().getNode()->use_end();
2762 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
2763 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
2764 if (FI->getIndex() < 0) {
2765 int64_t InFirstByte = MFI->getObjectOffset(FI->getIndex());
2766 int64_t InLastByte = InFirstByte;
2767 InLastByte += MFI->getObjectSize(FI->getIndex()) - 1;
2769 if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
2770 (FirstByte <= InFirstByte && InFirstByte <= LastByte))
2771 ArgChains.push_back(SDValue(L, 1));
2774 // Build a tokenfactor for all the chains.
2775 return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
2778 bool AArch64TargetLowering::DoesCalleeRestoreStack(CallingConv::ID CallCC,
2779 bool TailCallOpt) const {
2780 return CallCC == CallingConv::Fast && TailCallOpt;
2783 bool AArch64TargetLowering::IsTailCallConvention(CallingConv::ID CallCC) const {
2784 return CallCC == CallingConv::Fast;
2787 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
2788 /// and add input and output parameter nodes.
2790 AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
2791 SmallVectorImpl<SDValue> &InVals) const {
2792 SelectionDAG &DAG = CLI.DAG;
2794 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2795 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2796 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2797 SDValue Chain = CLI.Chain;
2798 SDValue Callee = CLI.Callee;
2799 bool &IsTailCall = CLI.IsTailCall;
2800 CallingConv::ID CallConv = CLI.CallConv;
2801 bool IsVarArg = CLI.IsVarArg;
2803 MachineFunction &MF = DAG.getMachineFunction();
2804 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
2805 bool IsThisReturn = false;
2807 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2808 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2809 bool IsSibCall = false;
2812 // Check if it's really possible to do a tail call.
2813 IsTailCall = isEligibleForTailCallOptimization(
2814 Callee, CallConv, IsVarArg, IsStructRet,
2815 MF.getFunction()->hasStructRetAttr(), Outs, OutVals, Ins, DAG);
2816 if (!IsTailCall && CLI.CS && CLI.CS->isMustTailCall())
2817 report_fatal_error("failed to perform tail call elimination on a call "
2818 "site marked musttail");
2820 // A sibling call is one where we're under the usual C ABI and not planning
2821 // to change that but can still do a tail call:
2822 if (!TailCallOpt && IsTailCall)
2829 // Analyze operands of the call, assigning locations to each operand.
2830 SmallVector<CCValAssign, 16> ArgLocs;
2831 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
2835 // Handle fixed and variable vector arguments differently.
2836 // Variable vector arguments always go into memory.
2837 unsigned NumArgs = Outs.size();
2839 for (unsigned i = 0; i != NumArgs; ++i) {
2840 MVT ArgVT = Outs[i].VT;
2841 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2842 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
2843 /*IsVarArg=*/ !Outs[i].IsFixed);
2844 bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
2845 assert(!Res && "Call operand has unhandled type");
2849 // At this point, Outs[].VT may already be promoted to i32. To correctly
2850 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2851 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
2852 // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here
2853 // we use a special version of AnalyzeCallOperands to pass in ValVT and
2855 unsigned NumArgs = Outs.size();
2856 for (unsigned i = 0; i != NumArgs; ++i) {
2857 MVT ValVT = Outs[i].VT;
2858 // Get type of the original argument.
2859 EVT ActualVT = getValueType(DAG.getDataLayout(),
2860 CLI.getArgs()[Outs[i].OrigArgIndex].Ty,
2861 /*AllowUnknown*/ true);
2862 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT;
2863 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2864 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2865 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2867 else if (ActualMVT == MVT::i16)
2870 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2871 bool Res = AssignFn(i, ValVT, ValVT, CCValAssign::Full, ArgFlags, CCInfo);
2872 assert(!Res && "Call operand has unhandled type");
2877 // Get a count of how many bytes are to be pushed on the stack.
2878 unsigned NumBytes = CCInfo.getNextStackOffset();
2881 // Since we're not changing the ABI to make this a tail call, the memory
2882 // operands are already available in the caller's incoming argument space.
2886 // FPDiff is the byte offset of the call's argument area from the callee's.
2887 // Stores to callee stack arguments will be placed in FixedStackSlots offset
2888 // by this amount for a tail call. In a sibling call it must be 0 because the
2889 // caller will deallocate the entire stack and the callee still expects its
2890 // arguments to begin at SP+0. Completely unused for non-tail calls.
2893 if (IsTailCall && !IsSibCall) {
2894 unsigned NumReusableBytes = FuncInfo->getBytesInStackArgArea();
2896 // Since callee will pop argument stack as a tail call, we must keep the
2897 // popped size 16-byte aligned.
2898 NumBytes = RoundUpToAlignment(NumBytes, 16);
2900 // FPDiff will be negative if this tail call requires more space than we
2901 // would automatically have in our incoming argument space. Positive if we
2902 // can actually shrink the stack.
2903 FPDiff = NumReusableBytes - NumBytes;
2905 // The stack pointer must be 16-byte aligned at all times it's used for a
2906 // memory operation, which in practice means at *all* times and in
2907 // particular across call boundaries. Therefore our own arguments started at
2908 // a 16-byte aligned SP and the delta applied for the tail call should
2909 // satisfy the same constraint.
2910 assert(FPDiff % 16 == 0 && "unaligned stack on tail call");
2913 // Adjust the stack pointer for the new arguments...
2914 // These operations are automatically eliminated by the prolog/epilog pass
2916 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, DL,
2920 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, AArch64::SP,
2921 getPointerTy(DAG.getDataLayout()));
2923 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2924 SmallVector<SDValue, 8> MemOpChains;
2925 auto PtrVT = getPointerTy(DAG.getDataLayout());
2927 // Walk the register/memloc assignments, inserting copies/loads.
2928 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2929 ++i, ++realArgIdx) {
2930 CCValAssign &VA = ArgLocs[i];
2931 SDValue Arg = OutVals[realArgIdx];
2932 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2934 // Promote the value if needed.
2935 switch (VA.getLocInfo()) {
2937 llvm_unreachable("Unknown loc info!");
2938 case CCValAssign::Full:
2940 case CCValAssign::SExt:
2941 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
2943 case CCValAssign::ZExt:
2944 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2946 case CCValAssign::AExt:
2947 if (Outs[realArgIdx].ArgVT == MVT::i1) {
2948 // AAPCS requires i1 to be zero-extended to 8-bits by the caller.
2949 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
2950 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i8, Arg);
2952 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
2954 case CCValAssign::BCvt:
2955 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2957 case CCValAssign::FPExt:
2958 Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
2962 if (VA.isRegLoc()) {
2963 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i64) {
2964 assert(VA.getLocVT() == MVT::i64 &&
2965 "unexpected calling convention register assignment");
2966 assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
2967 "unexpected use of 'returned'");
2968 IsThisReturn = true;
2970 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2972 assert(VA.isMemLoc());
2975 MachinePointerInfo DstInfo;
2977 // FIXME: This works on big-endian for composite byvals, which are the
2978 // common case. It should also work for fundamental types too.
2979 uint32_t BEAlign = 0;
2980 unsigned OpSize = Flags.isByVal() ? Flags.getByValSize() * 8
2981 : VA.getValVT().getSizeInBits();
2982 OpSize = (OpSize + 7) / 8;
2983 if (!Subtarget->isLittleEndian() && !Flags.isByVal() &&
2984 !Flags.isInConsecutiveRegs()) {
2986 BEAlign = 8 - OpSize;
2988 unsigned LocMemOffset = VA.getLocMemOffset();
2989 int32_t Offset = LocMemOffset + BEAlign;
2990 SDValue PtrOff = DAG.getIntPtrConstant(Offset, DL);
2991 PtrOff = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, PtrOff);
2994 Offset = Offset + FPDiff;
2995 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2997 DstAddr = DAG.getFrameIndex(FI, PtrVT);
2998 DstInfo = MachinePointerInfo::getFixedStack(FI);
3000 // Make sure any stack arguments overlapping with where we're storing
3001 // are loaded before this eventual operation. Otherwise they'll be
3003 Chain = addTokenForArgument(Chain, DAG, MF.getFrameInfo(), FI);
3005 SDValue PtrOff = DAG.getIntPtrConstant(Offset, DL);
3007 DstAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, PtrOff);
3008 DstInfo = MachinePointerInfo::getStack(LocMemOffset);
3011 if (Outs[i].Flags.isByVal()) {
3013 DAG.getConstant(Outs[i].Flags.getByValSize(), DL, MVT::i64);
3014 SDValue Cpy = DAG.getMemcpy(
3015 Chain, DL, DstAddr, Arg, SizeNode, Outs[i].Flags.getByValAlign(),
3016 /*isVol = */ false, /*AlwaysInline = */ false,
3017 /*isTailCall = */ false,
3018 DstInfo, MachinePointerInfo());
3020 MemOpChains.push_back(Cpy);
3022 // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
3023 // promoted to a legal register type i32, we should truncate Arg back to
3025 if (VA.getValVT() == MVT::i1 || VA.getValVT() == MVT::i8 ||
3026 VA.getValVT() == MVT::i16)
3027 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Arg);
3030 DAG.getStore(Chain, DL, Arg, DstAddr, DstInfo, false, false, 0);
3031 MemOpChains.push_back(Store);
3036 if (!MemOpChains.empty())
3037 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
3039 // Build a sequence of copy-to-reg nodes chained together with token chain
3040 // and flag operands which copy the outgoing args into the appropriate regs.
3042 for (auto &RegToPass : RegsToPass) {
3043 Chain = DAG.getCopyToReg(Chain, DL, RegToPass.first,
3044 RegToPass.second, InFlag);
3045 InFlag = Chain.getValue(1);
3048 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
3049 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
3050 // node so that legalize doesn't hack it.
3051 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3052 Subtarget->isTargetMachO()) {
3053 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3054 const GlobalValue *GV = G->getGlobal();
3055 bool InternalLinkage = GV->hasInternalLinkage();
3056 if (InternalLinkage)
3057 Callee = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 0);
3060 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_GOT);
3061 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, Callee);
3063 } else if (ExternalSymbolSDNode *S =
3064 dyn_cast<ExternalSymbolSDNode>(Callee)) {
3065 const char *Sym = S->getSymbol();
3066 Callee = DAG.getTargetExternalSymbol(Sym, PtrVT, AArch64II::MO_GOT);
3067 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, Callee);
3069 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3070 const GlobalValue *GV = G->getGlobal();
3071 Callee = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 0);
3072 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3073 const char *Sym = S->getSymbol();
3074 Callee = DAG.getTargetExternalSymbol(Sym, PtrVT, 0);
3077 // We don't usually want to end the call-sequence here because we would tidy
3078 // the frame up *after* the call, however in the ABI-changing tail-call case
3079 // we've carefully laid out the parameters so that when sp is reset they'll be
3080 // in the correct location.
3081 if (IsTailCall && !IsSibCall) {
3082 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, DL, true),
3083 DAG.getIntPtrConstant(0, DL, true), InFlag, DL);
3084 InFlag = Chain.getValue(1);
3087 std::vector<SDValue> Ops;
3088 Ops.push_back(Chain);
3089 Ops.push_back(Callee);
3092 // Each tail call may have to adjust the stack by a different amount, so
3093 // this information must travel along with the operation for eventual
3094 // consumption by emitEpilogue.
3095 Ops.push_back(DAG.getTargetConstant(FPDiff, DL, MVT::i32));
3098 // Add argument registers to the end of the list so that they are known live
3100 for (auto &RegToPass : RegsToPass)
3101 Ops.push_back(DAG.getRegister(RegToPass.first,
3102 RegToPass.second.getValueType()));
3104 // Add a register mask operand representing the call-preserved registers.
3105 const uint32_t *Mask;
3106 const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();
3108 // For 'this' returns, use the X0-preserving mask if applicable
3109 Mask = TRI->getThisReturnPreservedMask(MF, CallConv);
3111 IsThisReturn = false;
3112 Mask = TRI->getCallPreservedMask(MF, CallConv);
3115 Mask = TRI->getCallPreservedMask(MF, CallConv);
3117 assert(Mask && "Missing call preserved mask for calling convention");
3118 Ops.push_back(DAG.getRegisterMask(Mask));
3120 if (InFlag.getNode())
3121 Ops.push_back(InFlag);
3123 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3125 // If we're doing a tall call, use a TC_RETURN here rather than an
3126 // actual call instruction.
3128 MF.getFrameInfo()->setHasTailCall();
3129 return DAG.getNode(AArch64ISD::TC_RETURN, DL, NodeTys, Ops);
3132 // Returns a chain and a flag for retval copy to use.
3133 Chain = DAG.getNode(AArch64ISD::CALL, DL, NodeTys, Ops);
3134 InFlag = Chain.getValue(1);
3136 uint64_t CalleePopBytes = DoesCalleeRestoreStack(CallConv, TailCallOpt)
3137 ? RoundUpToAlignment(NumBytes, 16)
3140 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, DL, true),
3141 DAG.getIntPtrConstant(CalleePopBytes, DL, true),
3144 InFlag = Chain.getValue(1);
3146 // Handle result values, copying them out of physregs into vregs that we
3148 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
3149 InVals, IsThisReturn,
3150 IsThisReturn ? OutVals[0] : SDValue());
3153 bool AArch64TargetLowering::CanLowerReturn(
3154 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
3155 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
3156 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
3157 ? RetCC_AArch64_WebKit_JS
3158 : RetCC_AArch64_AAPCS;
3159 SmallVector<CCValAssign, 16> RVLocs;
3160 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
3161 return CCInfo.CheckReturn(Outs, RetCC);
3165 AArch64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
3167 const SmallVectorImpl<ISD::OutputArg> &Outs,
3168 const SmallVectorImpl<SDValue> &OutVals,
3169 SDLoc DL, SelectionDAG &DAG) const {
3170 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
3171 ? RetCC_AArch64_WebKit_JS
3172 : RetCC_AArch64_AAPCS;
3173 SmallVector<CCValAssign, 16> RVLocs;
3174 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
3176 CCInfo.AnalyzeReturn(Outs, RetCC);
3178 // Copy the result values into the output registers.
3180 SmallVector<SDValue, 4> RetOps(1, Chain);
3181 for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size();
3182 ++i, ++realRVLocIdx) {
3183 CCValAssign &VA = RVLocs[i];
3184 assert(VA.isRegLoc() && "Can only return in registers!");
3185 SDValue Arg = OutVals[realRVLocIdx];
3187 switch (VA.getLocInfo()) {
3189 llvm_unreachable("Unknown loc info!");
3190 case CCValAssign::Full:
3191 if (Outs[i].ArgVT == MVT::i1) {
3192 // AAPCS requires i1 to be zero-extended to i8 by the producer of the
3193 // value. This is strictly redundant on Darwin (which uses "zeroext
3194 // i1"), but will be optimised out before ISel.
3195 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
3196 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
3199 case CCValAssign::BCvt:
3200 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
3204 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
3205 Flag = Chain.getValue(1);
3206 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
3209 RetOps[0] = Chain; // Update chain.
3211 // Add the flag if we have it.
3213 RetOps.push_back(Flag);
3215 return DAG.getNode(AArch64ISD::RET_FLAG, DL, MVT::Other, RetOps);
3218 //===----------------------------------------------------------------------===//
3219 // Other Lowering Code
3220 //===----------------------------------------------------------------------===//
3222 SDValue AArch64TargetLowering::LowerGlobalAddress(SDValue Op,
3223 SelectionDAG &DAG) const {
3224 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3226 const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
3227 const GlobalValue *GV = GN->getGlobal();
3228 unsigned char OpFlags =
3229 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
3231 assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
3232 "unexpected offset in global node");
3234 // This also catched the large code model case for Darwin.
3235 if ((OpFlags & AArch64II::MO_GOT) != 0) {
3236 SDValue GotAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
3237 // FIXME: Once remat is capable of dealing with instructions with register
3238 // operands, expand this into two nodes instead of using a wrapper node.
3239 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
3242 if ((OpFlags & AArch64II::MO_CONSTPOOL) != 0) {
3243 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
3244 "use of MO_CONSTPOOL only supported on small model");
3245 SDValue Hi = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, AArch64II::MO_PAGE);
3246 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3247 unsigned char LoFlags = AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
3248 SDValue Lo = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, LoFlags);
3249 SDValue PoolAddr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3250 SDValue GlobalAddr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), PoolAddr,
3251 MachinePointerInfo::getConstantPool(),
3252 /*isVolatile=*/ false,
3253 /*isNonTemporal=*/ true,
3254 /*isInvariant=*/ true, 8);
3255 if (GN->getOffset() != 0)
3256 return DAG.getNode(ISD::ADD, DL, PtrVT, GlobalAddr,
3257 DAG.getConstant(GN->getOffset(), DL, PtrVT));
3261 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
3262 const unsigned char MO_NC = AArch64II::MO_NC;
3264 AArch64ISD::WrapperLarge, DL, PtrVT,
3265 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G3),
3266 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
3267 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
3268 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
3270 // Use ADRP/ADD or ADRP/LDR for everything else: the small model on ELF and
3271 // the only correct model on Darwin.
3272 SDValue Hi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
3273 OpFlags | AArch64II::MO_PAGE);
3274 unsigned char LoFlags = OpFlags | AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
3275 SDValue Lo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, LoFlags);
3277 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3278 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3282 /// \brief Convert a TLS address reference into the correct sequence of loads
3283 /// and calls to compute the variable's address (for Darwin, currently) and
3284 /// return an SDValue containing the final node.
3286 /// Darwin only has one TLS scheme which must be capable of dealing with the
3287 /// fully general situation, in the worst case. This means:
3288 /// + "extern __thread" declaration.
3289 /// + Defined in a possibly unknown dynamic library.
3291 /// The general system is that each __thread variable has a [3 x i64] descriptor
3292 /// which contains information used by the runtime to calculate the address. The
3293 /// only part of this the compiler needs to know about is the first xword, which
3294 /// contains a function pointer that must be called with the address of the
3295 /// entire descriptor in "x0".
3297 /// Since this descriptor may be in a different unit, in general even the
3298 /// descriptor must be accessed via an indirect load. The "ideal" code sequence
3300 /// adrp x0, _var@TLVPPAGE
3301 /// ldr x0, [x0, _var@TLVPPAGEOFF] ; x0 now contains address of descriptor
3302 /// ldr x1, [x0] ; x1 contains 1st entry of descriptor,
3303 /// ; the function pointer
3304 /// blr x1 ; Uses descriptor address in x0
3305 /// ; Address of _var is now in x0.
3307 /// If the address of _var's descriptor *is* known to the linker, then it can
3308 /// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
3309 /// a slight efficiency gain.
3311 AArch64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
3312 SelectionDAG &DAG) const {
3313 assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");
3316 MVT PtrVT = getPointerTy(DAG.getDataLayout());
3317 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3320 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3321 SDValue DescAddr = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TLVPAddr);
3323 // The first entry in the descriptor is a function pointer that we must call
3324 // to obtain the address of the variable.
3325 SDValue Chain = DAG.getEntryNode();
3326 SDValue FuncTLVGet =
3327 DAG.getLoad(MVT::i64, DL, Chain, DescAddr, MachinePointerInfo::getGOT(),
3328 false, true, true, 8);
3329 Chain = FuncTLVGet.getValue(1);
3331 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3332 MFI->setAdjustsStack(true);
3334 // TLS calls preserve all registers except those that absolutely must be
3335 // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
3337 const uint32_t *Mask =
3338 Subtarget->getRegisterInfo()->getTLSCallPreservedMask();
3340 // Finally, we can make the call. This is just a degenerate version of a
3341 // normal AArch64 call node: x0 takes the address of the descriptor, and
3342 // returns the address of the variable in this thread.
3343 Chain = DAG.getCopyToReg(Chain, DL, AArch64::X0, DescAddr, SDValue());
3345 DAG.getNode(AArch64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
3346 Chain, FuncTLVGet, DAG.getRegister(AArch64::X0, MVT::i64),
3347 DAG.getRegisterMask(Mask), Chain.getValue(1));
3348 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Chain.getValue(1));
3351 /// When accessing thread-local variables under either the general-dynamic or
3352 /// local-dynamic system, we make a "TLS-descriptor" call. The variable will
3353 /// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
3354 /// is a function pointer to carry out the resolution.
3356 /// The sequence is:
3357 /// adrp x0, :tlsdesc:var
3358 /// ldr x1, [x0, #:tlsdesc_lo12:var]
3359 /// add x0, x0, #:tlsdesc_lo12:var
3360 /// .tlsdesccall var
3362 /// (TPIDR_EL0 offset now in x0)
3364 /// The above sequence must be produced unscheduled, to enable the linker to
3365 /// optimize/relax this sequence.
3366 /// Therefore, a pseudo-instruction (TLSDESC_CALLSEQ) is used to represent the
3367 /// above sequence, and expanded really late in the compilation flow, to ensure
3368 /// the sequence is produced as per above.
3369 SDValue AArch64TargetLowering::LowerELFTLSDescCallSeq(SDValue SymAddr, SDLoc DL,
3370 SelectionDAG &DAG) const {
3371 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3373 SDValue Chain = DAG.getEntryNode();
3374 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3376 SmallVector<SDValue, 2> Ops;
3377 Ops.push_back(Chain);
3378 Ops.push_back(SymAddr);
3380 Chain = DAG.getNode(AArch64ISD::TLSDESC_CALLSEQ, DL, NodeTys, Ops);
3381 SDValue Glue = Chain.getValue(1);
3383 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Glue);
3387 AArch64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
3388 SelectionDAG &DAG) const {
3389 assert(Subtarget->isTargetELF() && "This function expects an ELF target");
3390 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
3391 "ELF TLS only supported in small memory model");
3392 // Different choices can be made for the maximum size of the TLS area for a
3393 // module. For the small address model, the default TLS size is 16MiB and the
3394 // maximum TLS size is 4GiB.
3395 // FIXME: add -mtls-size command line option and make it control the 16MiB
3396 // vs. 4GiB code sequence generation.
3397 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3399 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
3401 if (DAG.getTarget().Options.EmulatedTLS)
3402 return LowerToTLSEmulatedModel(GA, DAG);
3404 if (!EnableAArch64ELFLocalDynamicTLSGeneration) {
3405 if (Model == TLSModel::LocalDynamic)
3406 Model = TLSModel::GeneralDynamic;
3410 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3412 const GlobalValue *GV = GA->getGlobal();
3414 SDValue ThreadBase = DAG.getNode(AArch64ISD::THREAD_POINTER, DL, PtrVT);
3416 if (Model == TLSModel::LocalExec) {
3417 SDValue HiVar = DAG.getTargetGlobalAddress(
3418 GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_HI12);
3419 SDValue LoVar = DAG.getTargetGlobalAddress(
3421 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3423 SDValue TPWithOff_lo =
3424 SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, ThreadBase,
3426 DAG.getTargetConstant(0, DL, MVT::i32)),
3429 SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPWithOff_lo,
3431 DAG.getTargetConstant(0, DL, MVT::i32)),
3434 } else if (Model == TLSModel::InitialExec) {
3435 TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3436 TPOff = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TPOff);
3437 } else if (Model == TLSModel::LocalDynamic) {
3438 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
3439 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
3440 // the beginning of the module's TLS region, followed by a DTPREL offset
3443 // These accesses will need deduplicating if there's more than one.
3444 AArch64FunctionInfo *MFI =
3445 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
3446 MFI->incNumLocalDynamicTLSAccesses();
3448 // The call needs a relocation too for linker relaxation. It doesn't make
3449 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
3451 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
3454 // Now we can calculate the offset from TPIDR_EL0 to this module's
3455 // thread-local area.
3456 TPOff = LowerELFTLSDescCallSeq(SymAddr, DL, DAG);
3458 // Now use :dtprel_whatever: operations to calculate this variable's offset
3459 // in its thread-storage area.
3460 SDValue HiVar = DAG.getTargetGlobalAddress(
3461 GV, DL, MVT::i64, 0, AArch64II::MO_TLS | AArch64II::MO_HI12);
3462 SDValue LoVar = DAG.getTargetGlobalAddress(
3463 GV, DL, MVT::i64, 0,
3464 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3466 TPOff = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPOff, HiVar,
3467 DAG.getTargetConstant(0, DL, MVT::i32)),
3469 TPOff = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPOff, LoVar,
3470 DAG.getTargetConstant(0, DL, MVT::i32)),
3472 } else if (Model == TLSModel::GeneralDynamic) {
3473 // The call needs a relocation too for linker relaxation. It doesn't make
3474 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
3477 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3479 // Finally we can make a call to calculate the offset from tpidr_el0.
3480 TPOff = LowerELFTLSDescCallSeq(SymAddr, DL, DAG);
3482 llvm_unreachable("Unsupported ELF TLS access model");
3484 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
3487 SDValue AArch64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
3488 SelectionDAG &DAG) const {
3489 if (Subtarget->isTargetDarwin())
3490 return LowerDarwinGlobalTLSAddress(Op, DAG);
3491 else if (Subtarget->isTargetELF())
3492 return LowerELFGlobalTLSAddress(Op, DAG);
3494 llvm_unreachable("Unexpected platform trying to use TLS");
3496 SDValue AArch64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3497 SDValue Chain = Op.getOperand(0);
3498 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3499 SDValue LHS = Op.getOperand(2);
3500 SDValue RHS = Op.getOperand(3);
3501 SDValue Dest = Op.getOperand(4);
3504 // Handle f128 first, since lowering it will result in comparing the return
3505 // value of a libcall against zero, which is just what the rest of LowerBR_CC
3506 // is expecting to deal with.
3507 if (LHS.getValueType() == MVT::f128) {
3508 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3510 // If softenSetCCOperands returned a scalar, we need to compare the result
3511 // against zero to select between true and false values.
3512 if (!RHS.getNode()) {
3513 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3518 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
3520 unsigned Opc = LHS.getOpcode();
3521 if (LHS.getResNo() == 1 && isa<ConstantSDNode>(RHS) &&
3522 cast<ConstantSDNode>(RHS)->isOne() &&
3523 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3524 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3525 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
3526 "Unexpected condition code.");
3527 // Only lower legal XALUO ops.
3528 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
3531 // The actual operation with overflow check.
3532 AArch64CC::CondCode OFCC;
3533 SDValue Value, Overflow;
3534 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, LHS.getValue(0), DAG);
3536 if (CC == ISD::SETNE)
3537 OFCC = getInvertedCondCode(OFCC);
3538 SDValue CCVal = DAG.getConstant(OFCC, dl, MVT::i32);
3540 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3544 if (LHS.getValueType().isInteger()) {
3545 assert((LHS.getValueType() == RHS.getValueType()) &&
3546 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3548 // If the RHS of the comparison is zero, we can potentially fold this
3549 // to a specialized branch.
3550 const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
3551 if (RHSC && RHSC->getZExtValue() == 0) {
3552 if (CC == ISD::SETEQ) {
3553 // See if we can use a TBZ to fold in an AND as well.
3554 // TBZ has a smaller branch displacement than CBZ. If the offset is
3555 // out of bounds, a late MI-layer pass rewrites branches.
3556 // 403.gcc is an example that hits this case.
3557 if (LHS.getOpcode() == ISD::AND &&
3558 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3559 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3560 SDValue Test = LHS.getOperand(0);
3561 uint64_t Mask = LHS.getConstantOperandVal(1);
3562 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, Test,
3563 DAG.getConstant(Log2_64(Mask), dl, MVT::i64),
3567 return DAG.getNode(AArch64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest);
3568 } else if (CC == ISD::SETNE) {
3569 // See if we can use a TBZ to fold in an AND as well.
3570 // TBZ has a smaller branch displacement than CBZ. If the offset is
3571 // out of bounds, a late MI-layer pass rewrites branches.
3572 // 403.gcc is an example that hits this case.
3573 if (LHS.getOpcode() == ISD::AND &&
3574 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3575 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3576 SDValue Test = LHS.getOperand(0);
3577 uint64_t Mask = LHS.getConstantOperandVal(1);
3578 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, Test,
3579 DAG.getConstant(Log2_64(Mask), dl, MVT::i64),
3583 return DAG.getNode(AArch64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest);
3584 } else if (CC == ISD::SETLT && LHS.getOpcode() != ISD::AND) {
3585 // Don't combine AND since emitComparison converts the AND to an ANDS
3586 // (a.k.a. TST) and the test in the test bit and branch instruction
3587 // becomes redundant. This would also increase register pressure.
3588 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
3589 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, LHS,
3590 DAG.getConstant(Mask, dl, MVT::i64), Dest);
3593 if (RHSC && RHSC->getSExtValue() == -1 && CC == ISD::SETGT &&
3594 LHS.getOpcode() != ISD::AND) {
3595 // Don't combine AND since emitComparison converts the AND to an ANDS
3596 // (a.k.a. TST) and the test in the test bit and branch instruction
3597 // becomes redundant. This would also increase register pressure.
3598 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
3599 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, LHS,
3600 DAG.getConstant(Mask, dl, MVT::i64), Dest);
3604 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3605 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3609 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3611 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
3612 // clean. Some of them require two branches to implement.
3613 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3614 AArch64CC::CondCode CC1, CC2;
3615 changeFPCCToAArch64CC(CC, CC1, CC2);
3616 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
3618 DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp);
3619 if (CC2 != AArch64CC::AL) {
3620 SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32);
3621 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val,
3628 SDValue AArch64TargetLowering::LowerFCOPYSIGN(SDValue Op,
3629 SelectionDAG &DAG) const {
3630 EVT VT = Op.getValueType();
3633 SDValue In1 = Op.getOperand(0);
3634 SDValue In2 = Op.getOperand(1);
3635 EVT SrcVT = In2.getValueType();
3637 if (SrcVT == MVT::f32 && VT == MVT::f64)
3638 In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2);
3639 else if (SrcVT == MVT::f64 && VT == MVT::f32)
3640 In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2,
3641 DAG.getIntPtrConstant(0, DL));
3643 // FIXME: Src type is different, bail out for now. Can VT really be a
3651 SDValue VecVal1, VecVal2;
3652 if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) {
3655 EltMask = 0x80000000ULL;
3657 if (!VT.isVector()) {
3658 VecVal1 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3659 DAG.getUNDEF(VecVT), In1);
3660 VecVal2 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3661 DAG.getUNDEF(VecVT), In2);
3663 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3664 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3666 } else if (VT == MVT::f64 || VT == MVT::v2f64) {
3670 // We want to materialize a mask with the high bit set, but the AdvSIMD
3671 // immediate moves cannot materialize that in a single instruction for
3672 // 64-bit elements. Instead, materialize zero and then negate it.
3675 if (!VT.isVector()) {
3676 VecVal1 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3677 DAG.getUNDEF(VecVT), In1);
3678 VecVal2 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3679 DAG.getUNDEF(VecVT), In2);
3681 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3682 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3685 llvm_unreachable("Invalid type for copysign!");
3688 SDValue BuildVec = DAG.getConstant(EltMask, DL, VecVT);
3690 // If we couldn't materialize the mask above, then the mask vector will be
3691 // the zero vector, and we need to negate it here.
3692 if (VT == MVT::f64 || VT == MVT::v2f64) {
3693 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec);
3694 BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec);
3695 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec);
3699 DAG.getNode(AArch64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec);
3702 return DAG.getTargetExtractSubreg(AArch64::ssub, DL, VT, Sel);
3703 else if (VT == MVT::f64)
3704 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, VT, Sel);
3706 return DAG.getNode(ISD::BITCAST, DL, VT, Sel);
3709 SDValue AArch64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
3710 if (DAG.getMachineFunction().getFunction()->hasFnAttribute(
3711 Attribute::NoImplicitFloat))
3714 if (!Subtarget->hasNEON())
3717 // While there is no integer popcount instruction, it can
3718 // be more efficiently lowered to the following sequence that uses
3719 // AdvSIMD registers/instructions as long as the copies to/from
3720 // the AdvSIMD registers are cheap.
3721 // FMOV D0, X0 // copy 64-bit int to vector, high bits zero'd
3722 // CNT V0.8B, V0.8B // 8xbyte pop-counts
3723 // ADDV B0, V0.8B // sum 8xbyte pop-counts
3724 // UMOV X0, V0.B[0] // copy byte result back to integer reg
3725 SDValue Val = Op.getOperand(0);
3727 EVT VT = Op.getValueType();
3730 Val = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Val);
3731 Val = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);
3733 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, Val);
3734 SDValue UaddLV = DAG.getNode(
3735 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
3736 DAG.getConstant(Intrinsic::aarch64_neon_uaddlv, DL, MVT::i32), CtPop);
3739 UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV);
3743 SDValue AArch64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
3745 if (Op.getValueType().isVector())
3746 return LowerVSETCC(Op, DAG);
3748 SDValue LHS = Op.getOperand(0);
3749 SDValue RHS = Op.getOperand(1);
3750 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
3753 // We chose ZeroOrOneBooleanContents, so use zero and one.
3754 EVT VT = Op.getValueType();
3755 SDValue TVal = DAG.getConstant(1, dl, VT);
3756 SDValue FVal = DAG.getConstant(0, dl, VT);
3758 // Handle f128 first, since one possible outcome is a normal integer
3759 // comparison which gets picked up by the next if statement.
3760 if (LHS.getValueType() == MVT::f128) {
3761 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3763 // If softenSetCCOperands returned a scalar, use it.
3764 if (!RHS.getNode()) {
3765 assert(LHS.getValueType() == Op.getValueType() &&
3766 "Unexpected setcc expansion!");
3771 if (LHS.getValueType().isInteger()) {
3774 getAArch64Cmp(LHS, RHS, ISD::getSetCCInverse(CC, true), CCVal, DAG, dl);
3776 // Note that we inverted the condition above, so we reverse the order of
3777 // the true and false operands here. This will allow the setcc to be
3778 // matched to a single CSINC instruction.
3779 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp);
3782 // Now we know we're dealing with FP values.
3783 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3785 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3786 // and do the comparison.
3787 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3789 AArch64CC::CondCode CC1, CC2;
3790 changeFPCCToAArch64CC(CC, CC1, CC2);
3791 if (CC2 == AArch64CC::AL) {
3792 changeFPCCToAArch64CC(ISD::getSetCCInverse(CC, false), CC1, CC2);
3793 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
3795 // Note that we inverted the condition above, so we reverse the order of
3796 // the true and false operands here. This will allow the setcc to be
3797 // matched to a single CSINC instruction.
3798 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp);
3800 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't
3801 // totally clean. Some of them require two CSELs to implement. As is in
3802 // this case, we emit the first CSEL and then emit a second using the output
3803 // of the first as the RHS. We're effectively OR'ing the two CC's together.
3805 // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
3806 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
3808 DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3810 SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32);
3811 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3815 /// A SELECT_CC operation is really some kind of max or min if both values being
3816 /// compared are, in some sense, equal to the results in either case. However,
3817 /// it is permissible to compare f32 values and produce directly extended f64
3820 /// Extending the comparison operands would also be allowed, but is less likely
3821 /// to happen in practice since their use is right here. Note that truncate
3822 /// operations would *not* be semantically equivalent.
3823 static bool selectCCOpsAreFMaxCompatible(SDValue Cmp, SDValue Result) {
3825 return (Cmp.getValueType() == MVT::f32 ||
3826 Cmp.getValueType() == MVT::f64);
3828 ConstantFPSDNode *CCmp = dyn_cast<ConstantFPSDNode>(Cmp);
3829 ConstantFPSDNode *CResult = dyn_cast<ConstantFPSDNode>(Result);
3830 if (CCmp && CResult && Cmp.getValueType() == MVT::f32 &&
3831 Result.getValueType() == MVT::f64) {
3833 APFloat CmpVal = CCmp->getValueAPF();
3834 CmpVal.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &Lossy);
3835 return CResult->getValueAPF().bitwiseIsEqual(CmpVal);
3838 return Result->getOpcode() == ISD::FP_EXTEND && Result->getOperand(0) == Cmp;
3841 SDValue AArch64TargetLowering::LowerSELECT_CC(ISD::CondCode CC, SDValue LHS,
3842 SDValue RHS, SDValue TVal,
3843 SDValue FVal, SDLoc dl,
3844 SelectionDAG &DAG) const {
3845 // Handle f128 first, because it will result in a comparison of some RTLIB
3846 // call result against zero.
3847 if (LHS.getValueType() == MVT::f128) {
3848 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3850 // If softenSetCCOperands returned a scalar, we need to compare the result
3851 // against zero to select between true and false values.
3852 if (!RHS.getNode()) {
3853 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3858 // Handle integers first.
3859 if (LHS.getValueType().isInteger()) {
3860 assert((LHS.getValueType() == RHS.getValueType()) &&
3861 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3863 unsigned Opcode = AArch64ISD::CSEL;
3865 // If both the TVal and the FVal are constants, see if we can swap them in
3866 // order to for a CSINV or CSINC out of them.
3867 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
3868 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
3870 if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) {
3871 std::swap(TVal, FVal);
3872 std::swap(CTVal, CFVal);
3873 CC = ISD::getSetCCInverse(CC, true);
3874 } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) {
3875 std::swap(TVal, FVal);
3876 std::swap(CTVal, CFVal);
3877 CC = ISD::getSetCCInverse(CC, true);
3878 } else if (TVal.getOpcode() == ISD::XOR) {
3879 // If TVal is a NOT we want to swap TVal and FVal so that we can match
3880 // with a CSINV rather than a CSEL.
3881 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(1));
3883 if (CVal && CVal->isAllOnesValue()) {
3884 std::swap(TVal, FVal);
3885 std::swap(CTVal, CFVal);
3886 CC = ISD::getSetCCInverse(CC, true);
3888 } else if (TVal.getOpcode() == ISD::SUB) {
3889 // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
3890 // that we can match with a CSNEG rather than a CSEL.
3891 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(0));
3893 if (CVal && CVal->isNullValue()) {
3894 std::swap(TVal, FVal);
3895 std::swap(CTVal, CFVal);
3896 CC = ISD::getSetCCInverse(CC, true);
3898 } else if (CTVal && CFVal) {
3899 const int64_t TrueVal = CTVal->getSExtValue();
3900 const int64_t FalseVal = CFVal->getSExtValue();
3903 // If both TVal and FVal are constants, see if FVal is the
3904 // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
3905 // instead of a CSEL in that case.
3906 if (TrueVal == ~FalseVal) {
3907 Opcode = AArch64ISD::CSINV;
3908 } else if (TrueVal == -FalseVal) {
3909 Opcode = AArch64ISD::CSNEG;
3910 } else if (TVal.getValueType() == MVT::i32) {
3911 // If our operands are only 32-bit wide, make sure we use 32-bit
3912 // arithmetic for the check whether we can use CSINC. This ensures that
3913 // the addition in the check will wrap around properly in case there is
3914 // an overflow (which would not be the case if we do the check with
3915 // 64-bit arithmetic).
3916 const uint32_t TrueVal32 = CTVal->getZExtValue();
3917 const uint32_t FalseVal32 = CFVal->getZExtValue();
3919 if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
3920 Opcode = AArch64ISD::CSINC;
3922 if (TrueVal32 > FalseVal32) {
3926 // 64-bit check whether we can use CSINC.
3927 } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) {
3928 Opcode = AArch64ISD::CSINC;
3930 if (TrueVal > FalseVal) {
3935 // Swap TVal and FVal if necessary.
3937 std::swap(TVal, FVal);
3938 std::swap(CTVal, CFVal);
3939 CC = ISD::getSetCCInverse(CC, true);
3942 if (Opcode != AArch64ISD::CSEL) {
3943 // Drop FVal since we can get its value by simply inverting/negating
3950 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3952 EVT VT = TVal.getValueType();
3953 return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp);
3956 // Now we know we're dealing with FP values.
3957 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3958 assert(LHS.getValueType() == RHS.getValueType());
3959 EVT VT = TVal.getValueType();
3960 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3962 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
3963 // clean. Some of them require two CSELs to implement.
3964 AArch64CC::CondCode CC1, CC2;
3965 changeFPCCToAArch64CC(CC, CC1, CC2);
3966 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
3967 SDValue CS1 = DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3969 // If we need a second CSEL, emit it, using the output of the first as the
3970 // RHS. We're effectively OR'ing the two CC's together.
3971 if (CC2 != AArch64CC::AL) {
3972 SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32);
3973 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3976 // Otherwise, return the output of the first CSEL.
3980 SDValue AArch64TargetLowering::LowerSELECT_CC(SDValue Op,
3981 SelectionDAG &DAG) const {
3982 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3983 SDValue LHS = Op.getOperand(0);
3984 SDValue RHS = Op.getOperand(1);
3985 SDValue TVal = Op.getOperand(2);
3986 SDValue FVal = Op.getOperand(3);
3988 return LowerSELECT_CC(CC, LHS, RHS, TVal, FVal, DL, DAG);
3991 SDValue AArch64TargetLowering::LowerSELECT(SDValue Op,
3992 SelectionDAG &DAG) const {
3993 SDValue CCVal = Op->getOperand(0);
3994 SDValue TVal = Op->getOperand(1);
3995 SDValue FVal = Op->getOperand(2);
3998 unsigned Opc = CCVal.getOpcode();
3999 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
4001 if (CCVal.getResNo() == 1 &&
4002 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
4003 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
4004 // Only lower legal XALUO ops.
4005 if (!DAG.getTargetLoweringInfo().isTypeLegal(CCVal->getValueType(0)))
4008 AArch64CC::CondCode OFCC;
4009 SDValue Value, Overflow;
4010 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, CCVal.getValue(0), DAG);
4011 SDValue CCVal = DAG.getConstant(OFCC, DL, MVT::i32);
4013 return DAG.getNode(AArch64ISD::CSEL, DL, Op.getValueType(), TVal, FVal,
4017 // Lower it the same way as we would lower a SELECT_CC node.
4020 if (CCVal.getOpcode() == ISD::SETCC) {
4021 LHS = CCVal.getOperand(0);
4022 RHS = CCVal.getOperand(1);
4023 CC = cast<CondCodeSDNode>(CCVal->getOperand(2))->get();
4026 RHS = DAG.getConstant(0, DL, CCVal.getValueType());
4029 return LowerSELECT_CC(CC, LHS, RHS, TVal, FVal, DL, DAG);
4032 SDValue AArch64TargetLowering::LowerJumpTable(SDValue Op,
4033 SelectionDAG &DAG) const {
4034 // Jump table entries as PC relative offsets. No additional tweaking
4035 // is necessary here. Just get the address of the jump table.
4036 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4037 EVT PtrVT = getPointerTy(DAG.getDataLayout());
4040 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
4041 !Subtarget->isTargetMachO()) {
4042 const unsigned char MO_NC = AArch64II::MO_NC;
4044 AArch64ISD::WrapperLarge, DL, PtrVT,
4045 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G3),
4046 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G2 | MO_NC),
4047 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G1 | MO_NC),
4048 DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
4049 AArch64II::MO_G0 | MO_NC));
4053 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_PAGE);
4054 SDValue Lo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
4055 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
4056 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
4057 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
4060 SDValue AArch64TargetLowering::LowerConstantPool(SDValue Op,
4061 SelectionDAG &DAG) const {
4062 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4063 EVT PtrVT = getPointerTy(DAG.getDataLayout());
4066 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
4067 // Use the GOT for the large code model on iOS.
4068 if (Subtarget->isTargetMachO()) {
4069 SDValue GotAddr = DAG.getTargetConstantPool(
4070 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
4072 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
4075 const unsigned char MO_NC = AArch64II::MO_NC;
4077 AArch64ISD::WrapperLarge, DL, PtrVT,
4078 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4079 CP->getOffset(), AArch64II::MO_G3),
4080 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4081 CP->getOffset(), AArch64II::MO_G2 | MO_NC),
4082 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4083 CP->getOffset(), AArch64II::MO_G1 | MO_NC),
4084 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4085 CP->getOffset(), AArch64II::MO_G0 | MO_NC));
4087 // Use ADRP/ADD or ADRP/LDR for everything else: the small memory model on
4088 // ELF, the only valid one on Darwin.
4090 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4091 CP->getOffset(), AArch64II::MO_PAGE);
4092 SDValue Lo = DAG.getTargetConstantPool(
4093 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
4094 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
4096 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
4097 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
4101 SDValue AArch64TargetLowering::LowerBlockAddress(SDValue Op,
4102 SelectionDAG &DAG) const {
4103 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
4104 EVT PtrVT = getPointerTy(DAG.getDataLayout());
4106 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
4107 !Subtarget->isTargetMachO()) {
4108 const unsigned char MO_NC = AArch64II::MO_NC;
4110 AArch64ISD::WrapperLarge, DL, PtrVT,
4111 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G3),
4112 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
4113 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
4114 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
4116 SDValue Hi = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGE);
4117 SDValue Lo = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGEOFF |
4119 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
4120 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
4124 SDValue AArch64TargetLowering::LowerDarwin_VASTART(SDValue Op,
4125 SelectionDAG &DAG) const {
4126 AArch64FunctionInfo *FuncInfo =
4127 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
4130 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(),
4131 getPointerTy(DAG.getDataLayout()));
4132 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
4133 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
4134 MachinePointerInfo(SV), false, false, 0);
4137 SDValue AArch64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
4138 SelectionDAG &DAG) const {
4139 // The layout of the va_list struct is specified in the AArch64 Procedure Call
4140 // Standard, section B.3.
4141 MachineFunction &MF = DAG.getMachineFunction();
4142 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
4143 auto PtrVT = getPointerTy(DAG.getDataLayout());
4146 SDValue Chain = Op.getOperand(0);
4147 SDValue VAList = Op.getOperand(1);
4148 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
4149 SmallVector<SDValue, 4> MemOps;
4151 // void *__stack at offset 0
4152 SDValue Stack = DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), PtrVT);
4153 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
4154 MachinePointerInfo(SV), false, false, 8));
4156 // void *__gr_top at offset 8
4157 int GPRSize = FuncInfo->getVarArgsGPRSize();
4159 SDValue GRTop, GRTopAddr;
4162 DAG.getNode(ISD::ADD, DL, PtrVT, VAList, DAG.getConstant(8, DL, PtrVT));
4164 GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), PtrVT);
4165 GRTop = DAG.getNode(ISD::ADD, DL, PtrVT, GRTop,
4166 DAG.getConstant(GPRSize, DL, PtrVT));
4168 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
4169 MachinePointerInfo(SV, 8), false, false, 8));
4172 // void *__vr_top at offset 16
4173 int FPRSize = FuncInfo->getVarArgsFPRSize();
4175 SDValue VRTop, VRTopAddr;
4176 VRTopAddr = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
4177 DAG.getConstant(16, DL, PtrVT));
4179 VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), PtrVT);
4180 VRTop = DAG.getNode(ISD::ADD, DL, PtrVT, VRTop,
4181 DAG.getConstant(FPRSize, DL, PtrVT));
4183 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
4184 MachinePointerInfo(SV, 16), false, false, 8));
4187 // int __gr_offs at offset 24
4188 SDValue GROffsAddr =
4189 DAG.getNode(ISD::ADD, DL, PtrVT, VAList, DAG.getConstant(24, DL, PtrVT));
4190 MemOps.push_back(DAG.getStore(Chain, DL,
4191 DAG.getConstant(-GPRSize, DL, MVT::i32),
4192 GROffsAddr, MachinePointerInfo(SV, 24), false,
4195 // int __vr_offs at offset 28
4196 SDValue VROffsAddr =
4197 DAG.getNode(ISD::ADD, DL, PtrVT, VAList, DAG.getConstant(28, DL, PtrVT));
4198 MemOps.push_back(DAG.getStore(Chain, DL,
4199 DAG.getConstant(-FPRSize, DL, MVT::i32),
4200 VROffsAddr, MachinePointerInfo(SV, 28), false,
4203 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
4206 SDValue AArch64TargetLowering::LowerVASTART(SDValue Op,
4207 SelectionDAG &DAG) const {
4208 return Subtarget->isTargetDarwin() ? LowerDarwin_VASTART(Op, DAG)
4209 : LowerAAPCS_VASTART(Op, DAG);
4212 SDValue AArch64TargetLowering::LowerVACOPY(SDValue Op,
4213 SelectionDAG &DAG) const {
4214 // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
4217 unsigned VaListSize = Subtarget->isTargetDarwin() ? 8 : 32;
4218 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
4219 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
4221 return DAG.getMemcpy(Op.getOperand(0), DL, Op.getOperand(1),
4223 DAG.getConstant(VaListSize, DL, MVT::i32),
4224 8, false, false, false, MachinePointerInfo(DestSV),
4225 MachinePointerInfo(SrcSV));
4228 SDValue AArch64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
4229 assert(Subtarget->isTargetDarwin() &&
4230 "automatic va_arg instruction only works on Darwin");
4232 const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
4233 EVT VT = Op.getValueType();
4235 SDValue Chain = Op.getOperand(0);
4236 SDValue Addr = Op.getOperand(1);
4237 unsigned Align = Op.getConstantOperandVal(3);
4238 auto PtrVT = getPointerTy(DAG.getDataLayout());
4240 SDValue VAList = DAG.getLoad(PtrVT, DL, Chain, Addr, MachinePointerInfo(V),
4241 false, false, false, 0);
4242 Chain = VAList.getValue(1);
4245 assert(((Align & (Align - 1)) == 0) && "Expected Align to be a power of 2");
4246 VAList = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
4247 DAG.getConstant(Align - 1, DL, PtrVT));
4248 VAList = DAG.getNode(ISD::AND, DL, PtrVT, VAList,
4249 DAG.getConstant(-(int64_t)Align, DL, PtrVT));
4252 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
4253 uint64_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
4255 // Scalar integer and FP values smaller than 64 bits are implicitly extended
4256 // up to 64 bits. At the very least, we have to increase the striding of the
4257 // vaargs list to match this, and for FP values we need to introduce
4258 // FP_ROUND nodes as well.
4259 if (VT.isInteger() && !VT.isVector())
4261 bool NeedFPTrunc = false;
4262 if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
4267 // Increment the pointer, VAList, to the next vaarg
4268 SDValue VANext = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
4269 DAG.getConstant(ArgSize, DL, PtrVT));
4270 // Store the incremented VAList to the legalized pointer
4271 SDValue APStore = DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V),
4274 // Load the actual argument out of the pointer VAList
4276 // Load the value as an f64.
4277 SDValue WideFP = DAG.getLoad(MVT::f64, DL, APStore, VAList,
4278 MachinePointerInfo(), false, false, false, 0);
4279 // Round the value down to an f32.
4280 SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
4281 DAG.getIntPtrConstant(1, DL));
4282 SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
4283 // Merge the rounded value with the chain output of the load.
4284 return DAG.getMergeValues(Ops, DL);
4287 return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo(), false,
4291 SDValue AArch64TargetLowering::LowerFRAMEADDR(SDValue Op,
4292 SelectionDAG &DAG) const {
4293 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4294 MFI->setFrameAddressIsTaken(true);
4296 EVT VT = Op.getValueType();
4298 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4300 DAG.getCopyFromReg(DAG.getEntryNode(), DL, AArch64::FP, VT);
4302 FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
4303 MachinePointerInfo(), false, false, false, 0);
4307 // FIXME? Maybe this could be a TableGen attribute on some registers and
4308 // this table could be generated automatically from RegInfo.
4309 unsigned AArch64TargetLowering::getRegisterByName(const char* RegName, EVT VT,
4310 SelectionDAG &DAG) const {
4311 unsigned Reg = StringSwitch<unsigned>(RegName)
4312 .Case("sp", AArch64::SP)
4316 report_fatal_error(Twine("Invalid register name \""
4317 + StringRef(RegName) + "\"."));
4320 SDValue AArch64TargetLowering::LowerRETURNADDR(SDValue Op,
4321 SelectionDAG &DAG) const {
4322 MachineFunction &MF = DAG.getMachineFunction();
4323 MachineFrameInfo *MFI = MF.getFrameInfo();
4324 MFI->setReturnAddressIsTaken(true);
4326 EVT VT = Op.getValueType();
4328 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4330 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4331 SDValue Offset = DAG.getConstant(8, DL, getPointerTy(DAG.getDataLayout()));
4332 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
4333 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
4334 MachinePointerInfo(), false, false, false, 0);
4337 // Return LR, which contains the return address. Mark it an implicit live-in.
4338 unsigned Reg = MF.addLiveIn(AArch64::LR, &AArch64::GPR64RegClass);
4339 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
4342 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4343 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4344 SDValue AArch64TargetLowering::LowerShiftRightParts(SDValue Op,
4345 SelectionDAG &DAG) const {
4346 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4347 EVT VT = Op.getValueType();
4348 unsigned VTBits = VT.getSizeInBits();
4350 SDValue ShOpLo = Op.getOperand(0);
4351 SDValue ShOpHi = Op.getOperand(1);
4352 SDValue ShAmt = Op.getOperand(2);
4354 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4356 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4358 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4359 DAG.getConstant(VTBits, dl, MVT::i64), ShAmt);
4360 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4361 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4362 DAG.getConstant(VTBits, dl, MVT::i64));
4363 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4365 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, dl, MVT::i64),
4366 ISD::SETGE, dl, DAG);
4367 SDValue CCVal = DAG.getConstant(AArch64CC::GE, dl, MVT::i32);
4369 SDValue FalseValLo = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4370 SDValue TrueValLo = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4372 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
4374 // AArch64 shifts larger than the register width are wrapped rather than
4375 // clamped, so we can't just emit "hi >> x".
4376 SDValue FalseValHi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4377 SDValue TrueValHi = Opc == ISD::SRA
4378 ? DAG.getNode(Opc, dl, VT, ShOpHi,
4379 DAG.getConstant(VTBits - 1, dl,
4381 : DAG.getConstant(0, dl, VT);
4383 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValHi, FalseValHi, CCVal, Cmp);
4385 SDValue Ops[2] = { Lo, Hi };
4386 return DAG.getMergeValues(Ops, dl);
4389 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4390 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4391 SDValue AArch64TargetLowering::LowerShiftLeftParts(SDValue Op,
4392 SelectionDAG &DAG) const {
4393 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4394 EVT VT = Op.getValueType();
4395 unsigned VTBits = VT.getSizeInBits();
4397 SDValue ShOpLo = Op.getOperand(0);
4398 SDValue ShOpHi = Op.getOperand(1);
4399 SDValue ShAmt = Op.getOperand(2);
4402 assert(Op.getOpcode() == ISD::SHL_PARTS);
4403 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4404 DAG.getConstant(VTBits, dl, MVT::i64), ShAmt);
4405 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4406 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4407 DAG.getConstant(VTBits, dl, MVT::i64));
4408 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4409 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4411 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4413 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, dl, MVT::i64),
4414 ISD::SETGE, dl, DAG);
4415 SDValue CCVal = DAG.getConstant(AArch64CC::GE, dl, MVT::i32);
4417 DAG.getNode(AArch64ISD::CSEL, dl, VT, Tmp3, FalseVal, CCVal, Cmp);
4419 // AArch64 shifts of larger than register sizes are wrapped rather than
4420 // clamped, so we can't just emit "lo << a" if a is too big.
4421 SDValue TrueValLo = DAG.getConstant(0, dl, VT);
4422 SDValue FalseValLo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4424 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
4426 SDValue Ops[2] = { Lo, Hi };
4427 return DAG.getMergeValues(Ops, dl);
4430 bool AArch64TargetLowering::isOffsetFoldingLegal(
4431 const GlobalAddressSDNode *GA) const {
4432 // The AArch64 target doesn't support folding offsets into global addresses.
4436 bool AArch64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4437 // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
4438 // FIXME: We should be able to handle f128 as well with a clever lowering.
4439 if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32))
4443 return AArch64_AM::getFP64Imm(Imm) != -1;
4444 else if (VT == MVT::f32)
4445 return AArch64_AM::getFP32Imm(Imm) != -1;
4449 //===----------------------------------------------------------------------===//
4450 // AArch64 Optimization Hooks
4451 //===----------------------------------------------------------------------===//
4453 //===----------------------------------------------------------------------===//
4454 // AArch64 Inline Assembly Support
4455 //===----------------------------------------------------------------------===//
4457 // Table of Constraints
4458 // TODO: This is the current set of constraints supported by ARM for the
4459 // compiler, not all of them may make sense, e.g. S may be difficult to support.
4461 // r - A general register
4462 // w - An FP/SIMD register of some size in the range v0-v31
4463 // x - An FP/SIMD register of some size in the range v0-v15
4464 // I - Constant that can be used with an ADD instruction
4465 // J - Constant that can be used with a SUB instruction
4466 // K - Constant that can be used with a 32-bit logical instruction
4467 // L - Constant that can be used with a 64-bit logical instruction
4468 // M - Constant that can be used as a 32-bit MOV immediate
4469 // N - Constant that can be used as a 64-bit MOV immediate
4470 // Q - A memory reference with base register and no offset
4471 // S - A symbolic address
4472 // Y - Floating point constant zero
4473 // Z - Integer constant zero
4475 // Note that general register operands will be output using their 64-bit x
4476 // register name, whatever the size of the variable, unless the asm operand
4477 // is prefixed by the %w modifier. Floating-point and SIMD register operands
4478 // will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
4481 /// getConstraintType - Given a constraint letter, return the type of
4482 /// constraint it is for this target.
4483 AArch64TargetLowering::ConstraintType
4484 AArch64TargetLowering::getConstraintType(StringRef Constraint) const {
4485 if (Constraint.size() == 1) {
4486 switch (Constraint[0]) {
4493 return C_RegisterClass;
4494 // An address with a single base register. Due to the way we
4495 // currently handle addresses it is the same as 'r'.
4500 return TargetLowering::getConstraintType(Constraint);
4503 /// Examine constraint type and operand type and determine a weight value.
4504 /// This object must already have been set up with the operand type
4505 /// and the current alternative constraint selected.
4506 TargetLowering::ConstraintWeight
4507 AArch64TargetLowering::getSingleConstraintMatchWeight(
4508 AsmOperandInfo &info, const char *constraint) const {
4509 ConstraintWeight weight = CW_Invalid;
4510 Value *CallOperandVal = info.CallOperandVal;
4511 // If we don't have a value, we can't do a match,
4512 // but allow it at the lowest weight.
4513 if (!CallOperandVal)
4515 Type *type = CallOperandVal->getType();
4516 // Look at the constraint type.
4517 switch (*constraint) {
4519 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
4523 if (type->isFloatingPointTy() || type->isVectorTy())
4524 weight = CW_Register;
4527 weight = CW_Constant;
4533 std::pair<unsigned, const TargetRegisterClass *>
4534 AArch64TargetLowering::getRegForInlineAsmConstraint(
4535 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
4536 if (Constraint.size() == 1) {
4537 switch (Constraint[0]) {
4539 if (VT.getSizeInBits() == 64)
4540 return std::make_pair(0U, &AArch64::GPR64commonRegClass);
4541 return std::make_pair(0U, &AArch64::GPR32commonRegClass);
4544 return std::make_pair(0U, &AArch64::FPR32RegClass);
4545 if (VT.getSizeInBits() == 64)
4546 return std::make_pair(0U, &AArch64::FPR64RegClass);
4547 if (VT.getSizeInBits() == 128)
4548 return std::make_pair(0U, &AArch64::FPR128RegClass);
4550 // The instructions that this constraint is designed for can
4551 // only take 128-bit registers so just use that regclass.
4553 if (VT.getSizeInBits() == 128)
4554 return std::make_pair(0U, &AArch64::FPR128_loRegClass);
4558 if (StringRef("{cc}").equals_lower(Constraint))
4559 return std::make_pair(unsigned(AArch64::NZCV), &AArch64::CCRRegClass);
4561 // Use the default implementation in TargetLowering to convert the register
4562 // constraint into a member of a register class.
4563 std::pair<unsigned, const TargetRegisterClass *> Res;
4564 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
4566 // Not found as a standard register?
4568 unsigned Size = Constraint.size();
4569 if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
4570 tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
4572 bool Failed = Constraint.slice(2, Size - 1).getAsInteger(10, RegNo);
4573 if (!Failed && RegNo >= 0 && RegNo <= 31) {
4574 // v0 - v31 are aliases of q0 - q31.
4575 // By default we'll emit v0-v31 for this unless there's a modifier where
4576 // we'll emit the correct register as well.
4577 Res.first = AArch64::FPR128RegClass.getRegister(RegNo);
4578 Res.second = &AArch64::FPR128RegClass;
4586 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4587 /// vector. If it is invalid, don't add anything to Ops.
4588 void AArch64TargetLowering::LowerAsmOperandForConstraint(
4589 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
4590 SelectionDAG &DAG) const {
4593 // Currently only support length 1 constraints.
4594 if (Constraint.length() != 1)
4597 char ConstraintLetter = Constraint[0];
4598 switch (ConstraintLetter) {
4602 // This set of constraints deal with valid constants for various instructions.
4603 // Validate and return a target constant for them if we can.
4605 // 'z' maps to xzr or wzr so it needs an input of 0.
4606 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4607 if (!C || C->getZExtValue() != 0)
4610 if (Op.getValueType() == MVT::i64)
4611 Result = DAG.getRegister(AArch64::XZR, MVT::i64);
4613 Result = DAG.getRegister(AArch64::WZR, MVT::i32);
4623 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4627 // Grab the value and do some validation.
4628 uint64_t CVal = C->getZExtValue();
4629 switch (ConstraintLetter) {
4630 // The I constraint applies only to simple ADD or SUB immediate operands:
4631 // i.e. 0 to 4095 with optional shift by 12
4632 // The J constraint applies only to ADD or SUB immediates that would be
4633 // valid when negated, i.e. if [an add pattern] were to be output as a SUB
4634 // instruction [or vice versa], in other words -1 to -4095 with optional
4635 // left shift by 12.
4637 if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
4641 uint64_t NVal = -C->getSExtValue();
4642 if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal)) {
4643 CVal = C->getSExtValue();
4648 // The K and L constraints apply *only* to logical immediates, including
4649 // what used to be the MOVI alias for ORR (though the MOVI alias has now
4650 // been removed and MOV should be used). So these constraints have to
4651 // distinguish between bit patterns that are valid 32-bit or 64-bit
4652 // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
4653 // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
4656 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4660 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4663 // The M and N constraints are a superset of K and L respectively, for use
4664 // with the MOV (immediate) alias. As well as the logical immediates they
4665 // also match 32 or 64-bit immediates that can be loaded either using a
4666 // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
4667 // (M) or 64-bit 0x1234000000000000 (N) etc.
4668 // As a note some of this code is liberally stolen from the asm parser.
4670 if (!isUInt<32>(CVal))
4672 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4674 if ((CVal & 0xFFFF) == CVal)
4676 if ((CVal & 0xFFFF0000ULL) == CVal)
4678 uint64_t NCVal = ~(uint32_t)CVal;
4679 if ((NCVal & 0xFFFFULL) == NCVal)
4681 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4686 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4688 if ((CVal & 0xFFFFULL) == CVal)
4690 if ((CVal & 0xFFFF0000ULL) == CVal)
4692 if ((CVal & 0xFFFF00000000ULL) == CVal)
4694 if ((CVal & 0xFFFF000000000000ULL) == CVal)
4696 uint64_t NCVal = ~CVal;
4697 if ((NCVal & 0xFFFFULL) == NCVal)
4699 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4701 if ((NCVal & 0xFFFF00000000ULL) == NCVal)
4703 if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
4711 // All assembler immediates are 64-bit integers.
4712 Result = DAG.getTargetConstant(CVal, SDLoc(Op), MVT::i64);
4716 if (Result.getNode()) {
4717 Ops.push_back(Result);
4721 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
4724 //===----------------------------------------------------------------------===//
4725 // AArch64 Advanced SIMD Support
4726 //===----------------------------------------------------------------------===//
4728 /// WidenVector - Given a value in the V64 register class, produce the
4729 /// equivalent value in the V128 register class.
4730 static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
4731 EVT VT = V64Reg.getValueType();
4732 unsigned NarrowSize = VT.getVectorNumElements();
4733 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4734 MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
4737 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
4738 V64Reg, DAG.getConstant(0, DL, MVT::i32));
4741 /// getExtFactor - Determine the adjustment factor for the position when
4742 /// generating an "extract from vector registers" instruction.
4743 static unsigned getExtFactor(SDValue &V) {
4744 EVT EltType = V.getValueType().getVectorElementType();
4745 return EltType.getSizeInBits() / 8;
4748 /// NarrowVector - Given a value in the V128 register class, produce the
4749 /// equivalent value in the V64 register class.
4750 static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) {
4751 EVT VT = V128Reg.getValueType();
4752 unsigned WideSize = VT.getVectorNumElements();
4753 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4754 MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2);
4757 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, NarrowTy, V128Reg);
4760 // Gather data to see if the operation can be modelled as a
4761 // shuffle in combination with VEXTs.
4762 SDValue AArch64TargetLowering::ReconstructShuffle(SDValue Op,
4763 SelectionDAG &DAG) const {
4764 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
4766 EVT VT = Op.getValueType();
4767 unsigned NumElts = VT.getVectorNumElements();
4769 struct ShuffleSourceInfo {
4774 // We may insert some combination of BITCASTs and VEXT nodes to force Vec to
4775 // be compatible with the shuffle we intend to construct. As a result
4776 // ShuffleVec will be some sliding window into the original Vec.
4779 // Code should guarantee that element i in Vec starts at element "WindowBase
4780 // + i * WindowScale in ShuffleVec".
4784 bool operator ==(SDValue OtherVec) { return Vec == OtherVec; }
4785 ShuffleSourceInfo(SDValue Vec)
4786 : Vec(Vec), MinElt(UINT_MAX), MaxElt(0), ShuffleVec(Vec), WindowBase(0),
4790 // First gather all vectors used as an immediate source for this BUILD_VECTOR
4792 SmallVector<ShuffleSourceInfo, 2> Sources;
4793 for (unsigned i = 0; i < NumElts; ++i) {
4794 SDValue V = Op.getOperand(i);
4795 if (V.getOpcode() == ISD::UNDEF)
4797 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4798 // A shuffle can only come from building a vector from various
4799 // elements of other vectors.
4803 // Add this element source to the list if it's not already there.
4804 SDValue SourceVec = V.getOperand(0);
4805 auto Source = std::find(Sources.begin(), Sources.end(), SourceVec);
4806 if (Source == Sources.end())
4807 Source = Sources.insert(Sources.end(), ShuffleSourceInfo(SourceVec));
4809 // Update the minimum and maximum lane number seen.
4810 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4811 Source->MinElt = std::min(Source->MinElt, EltNo);
4812 Source->MaxElt = std::max(Source->MaxElt, EltNo);
4815 // Currently only do something sane when at most two source vectors
4817 if (Sources.size() > 2)
4820 // Find out the smallest element size among result and two sources, and use
4821 // it as element size to build the shuffle_vector.
4822 EVT SmallestEltTy = VT.getVectorElementType();
4823 for (auto &Source : Sources) {
4824 EVT SrcEltTy = Source.Vec.getValueType().getVectorElementType();
4825 if (SrcEltTy.bitsLT(SmallestEltTy)) {
4826 SmallestEltTy = SrcEltTy;
4829 unsigned ResMultiplier =
4830 VT.getVectorElementType().getSizeInBits() / SmallestEltTy.getSizeInBits();
4831 NumElts = VT.getSizeInBits() / SmallestEltTy.getSizeInBits();
4832 EVT ShuffleVT = EVT::getVectorVT(*DAG.getContext(), SmallestEltTy, NumElts);
4834 // If the source vector is too wide or too narrow, we may nevertheless be able
4835 // to construct a compatible shuffle either by concatenating it with UNDEF or
4836 // extracting a suitable range of elements.
4837 for (auto &Src : Sources) {
4838 EVT SrcVT = Src.ShuffleVec.getValueType();
4840 if (SrcVT.getSizeInBits() == VT.getSizeInBits())
4843 // This stage of the search produces a source with the same element type as
4844 // the original, but with a total width matching the BUILD_VECTOR output.
4845 EVT EltVT = SrcVT.getVectorElementType();
4846 unsigned NumSrcElts = VT.getSizeInBits() / EltVT.getSizeInBits();
4847 EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumSrcElts);
4849 if (SrcVT.getSizeInBits() < VT.getSizeInBits()) {
4850 assert(2 * SrcVT.getSizeInBits() == VT.getSizeInBits());
4851 // We can pad out the smaller vector for free, so if it's part of a
4854 DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, Src.ShuffleVec,
4855 DAG.getUNDEF(Src.ShuffleVec.getValueType()));
4859 assert(SrcVT.getSizeInBits() == 2 * VT.getSizeInBits());
4861 if (Src.MaxElt - Src.MinElt >= NumSrcElts) {
4862 // Span too large for a VEXT to cope
4866 if (Src.MinElt >= NumSrcElts) {
4867 // The extraction can just take the second half
4869 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4870 DAG.getConstant(NumSrcElts, dl, MVT::i64));
4871 Src.WindowBase = -NumSrcElts;
4872 } else if (Src.MaxElt < NumSrcElts) {
4873 // The extraction can just take the first half
4875 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4876 DAG.getConstant(0, dl, MVT::i64));
4878 // An actual VEXT is needed
4880 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4881 DAG.getConstant(0, dl, MVT::i64));
4883 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4884 DAG.getConstant(NumSrcElts, dl, MVT::i64));
4885 unsigned Imm = Src.MinElt * getExtFactor(VEXTSrc1);
4887 Src.ShuffleVec = DAG.getNode(AArch64ISD::EXT, dl, DestVT, VEXTSrc1,
4889 DAG.getConstant(Imm, dl, MVT::i32));
4890 Src.WindowBase = -Src.MinElt;
4894 // Another possible incompatibility occurs from the vector element types. We
4895 // can fix this by bitcasting the source vectors to the same type we intend
4897 for (auto &Src : Sources) {
4898 EVT SrcEltTy = Src.ShuffleVec.getValueType().getVectorElementType();
4899 if (SrcEltTy == SmallestEltTy)
4901 assert(ShuffleVT.getVectorElementType() == SmallestEltTy);
4902 Src.ShuffleVec = DAG.getNode(ISD::BITCAST, dl, ShuffleVT, Src.ShuffleVec);
4903 Src.WindowScale = SrcEltTy.getSizeInBits() / SmallestEltTy.getSizeInBits();
4904 Src.WindowBase *= Src.WindowScale;
4907 // Final sanity check before we try to actually produce a shuffle.
4909 for (auto Src : Sources)
4910 assert(Src.ShuffleVec.getValueType() == ShuffleVT);
4913 // The stars all align, our next step is to produce the mask for the shuffle.
4914 SmallVector<int, 8> Mask(ShuffleVT.getVectorNumElements(), -1);
4915 int BitsPerShuffleLane = ShuffleVT.getVectorElementType().getSizeInBits();
4916 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
4917 SDValue Entry = Op.getOperand(i);
4918 if (Entry.getOpcode() == ISD::UNDEF)
4921 auto Src = std::find(Sources.begin(), Sources.end(), Entry.getOperand(0));
4922 int EltNo = cast<ConstantSDNode>(Entry.getOperand(1))->getSExtValue();
4924 // EXTRACT_VECTOR_ELT performs an implicit any_ext; BUILD_VECTOR an implicit
4925 // trunc. So only std::min(SrcBits, DestBits) actually get defined in this
4927 EVT OrigEltTy = Entry.getOperand(0).getValueType().getVectorElementType();
4928 int BitsDefined = std::min(OrigEltTy.getSizeInBits(),
4929 VT.getVectorElementType().getSizeInBits());
4930 int LanesDefined = BitsDefined / BitsPerShuffleLane;
4932 // This source is expected to fill ResMultiplier lanes of the final shuffle,
4933 // starting at the appropriate offset.
4934 int *LaneMask = &Mask[i * ResMultiplier];
4936 int ExtractBase = EltNo * Src->WindowScale + Src->WindowBase;
4937 ExtractBase += NumElts * (Src - Sources.begin());
4938 for (int j = 0; j < LanesDefined; ++j)
4939 LaneMask[j] = ExtractBase + j;
4942 // Final check before we try to produce nonsense...
4943 if (!isShuffleMaskLegal(Mask, ShuffleVT))
4946 SDValue ShuffleOps[] = { DAG.getUNDEF(ShuffleVT), DAG.getUNDEF(ShuffleVT) };
4947 for (unsigned i = 0; i < Sources.size(); ++i)
4948 ShuffleOps[i] = Sources[i].ShuffleVec;
4950 SDValue Shuffle = DAG.getVectorShuffle(ShuffleVT, dl, ShuffleOps[0],
4951 ShuffleOps[1], &Mask[0]);
4952 return DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
4955 // check if an EXT instruction can handle the shuffle mask when the
4956 // vector sources of the shuffle are the same.
4957 static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4958 unsigned NumElts = VT.getVectorNumElements();
4960 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4966 // If this is a VEXT shuffle, the immediate value is the index of the first
4967 // element. The other shuffle indices must be the successive elements after
4969 unsigned ExpectedElt = Imm;
4970 for (unsigned i = 1; i < NumElts; ++i) {
4971 // Increment the expected index. If it wraps around, just follow it
4972 // back to index zero and keep going.
4974 if (ExpectedElt == NumElts)
4978 continue; // ignore UNDEF indices
4979 if (ExpectedElt != static_cast<unsigned>(M[i]))
4986 // check if an EXT instruction can handle the shuffle mask when the
4987 // vector sources of the shuffle are different.
4988 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
4990 // Look for the first non-undef element.
4991 const int *FirstRealElt = std::find_if(M.begin(), M.end(),
4992 [](int Elt) {return Elt >= 0;});
4994 // Benefit form APInt to handle overflow when calculating expected element.
4995 unsigned NumElts = VT.getVectorNumElements();
4996 unsigned MaskBits = APInt(32, NumElts * 2).logBase2();
4997 APInt ExpectedElt = APInt(MaskBits, *FirstRealElt + 1);
4998 // The following shuffle indices must be the successive elements after the
4999 // first real element.
5000 const int *FirstWrongElt = std::find_if(FirstRealElt + 1, M.end(),
5001 [&](int Elt) {return Elt != ExpectedElt++ && Elt != -1;});
5002 if (FirstWrongElt != M.end())
5005 // The index of an EXT is the first element if it is not UNDEF.
5006 // Watch out for the beginning UNDEFs. The EXT index should be the expected
5007 // value of the first element. E.g.
5008 // <-1, -1, 3, ...> is treated as <1, 2, 3, ...>.
5009 // <-1, -1, 0, 1, ...> is treated as <2*NumElts-2, 2*NumElts-1, 0, 1, ...>.
5010 // ExpectedElt is the last mask index plus 1.
5011 Imm = ExpectedElt.getZExtValue();
5013 // There are two difference cases requiring to reverse input vectors.
5014 // For example, for vector <4 x i32> we have the following cases,
5015 // Case 1: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, -1, 0>)
5016 // Case 2: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, 7, 0>)
5017 // For both cases, we finally use mask <5, 6, 7, 0>, which requires
5018 // to reverse two input vectors.
5027 /// isREVMask - Check if a vector shuffle corresponds to a REV
5028 /// instruction with the specified blocksize. (The order of the elements
5029 /// within each block of the vector is reversed.)
5030 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
5031 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
5032 "Only possible block sizes for REV are: 16, 32, 64");
5034 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5038 unsigned NumElts = VT.getVectorNumElements();
5039 unsigned BlockElts = M[0] + 1;
5040 // If the first shuffle index is UNDEF, be optimistic.
5042 BlockElts = BlockSize / EltSz;
5044 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
5047 for (unsigned i = 0; i < NumElts; ++i) {
5049 continue; // ignore UNDEF indices
5050 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
5057 static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5058 unsigned NumElts = VT.getVectorNumElements();
5059 WhichResult = (M[0] == 0 ? 0 : 1);
5060 unsigned Idx = WhichResult * NumElts / 2;
5061 for (unsigned i = 0; i != NumElts; i += 2) {
5062 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
5063 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
5071 static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5072 unsigned NumElts = VT.getVectorNumElements();
5073 WhichResult = (M[0] == 0 ? 0 : 1);
5074 for (unsigned i = 0; i != NumElts; ++i) {
5076 continue; // ignore UNDEF indices
5077 if ((unsigned)M[i] != 2 * i + WhichResult)
5084 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5085 unsigned NumElts = VT.getVectorNumElements();
5086 WhichResult = (M[0] == 0 ? 0 : 1);
5087 for (unsigned i = 0; i < NumElts; i += 2) {
5088 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
5089 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + NumElts + WhichResult))
5095 /// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of
5096 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5097 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
5098 static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5099 unsigned NumElts = VT.getVectorNumElements();
5100 WhichResult = (M[0] == 0 ? 0 : 1);
5101 unsigned Idx = WhichResult * NumElts / 2;
5102 for (unsigned i = 0; i != NumElts; i += 2) {
5103 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
5104 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx))
5112 /// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of
5113 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5114 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
5115 static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5116 unsigned Half = VT.getVectorNumElements() / 2;
5117 WhichResult = (M[0] == 0 ? 0 : 1);
5118 for (unsigned j = 0; j != 2; ++j) {
5119 unsigned Idx = WhichResult;
5120 for (unsigned i = 0; i != Half; ++i) {
5121 int MIdx = M[i + j * Half];
5122 if (MIdx >= 0 && (unsigned)MIdx != Idx)
5131 /// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of
5132 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5133 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
5134 static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5135 unsigned NumElts = VT.getVectorNumElements();
5136 WhichResult = (M[0] == 0 ? 0 : 1);
5137 for (unsigned i = 0; i < NumElts; i += 2) {
5138 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
5139 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult))
5145 static bool isINSMask(ArrayRef<int> M, int NumInputElements,
5146 bool &DstIsLeft, int &Anomaly) {
5147 if (M.size() != static_cast<size_t>(NumInputElements))
5150 int NumLHSMatch = 0, NumRHSMatch = 0;
5151 int LastLHSMismatch = -1, LastRHSMismatch = -1;
5153 for (int i = 0; i < NumInputElements; ++i) {
5163 LastLHSMismatch = i;
5165 if (M[i] == i + NumInputElements)
5168 LastRHSMismatch = i;
5171 if (NumLHSMatch == NumInputElements - 1) {
5173 Anomaly = LastLHSMismatch;
5175 } else if (NumRHSMatch == NumInputElements - 1) {
5177 Anomaly = LastRHSMismatch;
5184 static bool isConcatMask(ArrayRef<int> Mask, EVT VT, bool SplitLHS) {
5185 if (VT.getSizeInBits() != 128)
5188 unsigned NumElts = VT.getVectorNumElements();
5190 for (int I = 0, E = NumElts / 2; I != E; I++) {
5195 int Offset = NumElts / 2;
5196 for (int I = NumElts / 2, E = NumElts; I != E; I++) {
5197 if (Mask[I] != I + SplitLHS * Offset)
5204 static SDValue tryFormConcatFromShuffle(SDValue Op, SelectionDAG &DAG) {
5206 EVT VT = Op.getValueType();
5207 SDValue V0 = Op.getOperand(0);
5208 SDValue V1 = Op.getOperand(1);
5209 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op)->getMask();
5211 if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() ||
5212 VT.getVectorElementType() != V1.getValueType().getVectorElementType())
5215 bool SplitV0 = V0.getValueType().getSizeInBits() == 128;
5217 if (!isConcatMask(Mask, VT, SplitV0))
5220 EVT CastVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
5221 VT.getVectorNumElements() / 2);
5223 V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0,
5224 DAG.getConstant(0, DL, MVT::i64));
5226 if (V1.getValueType().getSizeInBits() == 128) {
5227 V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1,
5228 DAG.getConstant(0, DL, MVT::i64));
5230 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1);
5233 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5234 /// the specified operations to build the shuffle.
5235 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5236 SDValue RHS, SelectionDAG &DAG,
5238 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5239 unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1);
5240 unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1);
5243 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5252 OP_VUZPL, // VUZP, left result
5253 OP_VUZPR, // VUZP, right result
5254 OP_VZIPL, // VZIP, left result
5255 OP_VZIPR, // VZIP, right result
5256 OP_VTRNL, // VTRN, left result
5257 OP_VTRNR // VTRN, right result
5260 if (OpNum == OP_COPY) {
5261 if (LHSID == (1 * 9 + 2) * 9 + 3)
5263 assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!");
5267 SDValue OpLHS, OpRHS;
5268 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5269 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5270 EVT VT = OpLHS.getValueType();
5274 llvm_unreachable("Unknown shuffle opcode!");
5276 // VREV divides the vector in half and swaps within the half.
5277 if (VT.getVectorElementType() == MVT::i32 ||
5278 VT.getVectorElementType() == MVT::f32)
5279 return DAG.getNode(AArch64ISD::REV64, dl, VT, OpLHS);
5280 // vrev <4 x i16> -> REV32
5281 if (VT.getVectorElementType() == MVT::i16 ||
5282 VT.getVectorElementType() == MVT::f16)
5283 return DAG.getNode(AArch64ISD::REV32, dl, VT, OpLHS);
5284 // vrev <4 x i8> -> REV16
5285 assert(VT.getVectorElementType() == MVT::i8);
5286 return DAG.getNode(AArch64ISD::REV16, dl, VT, OpLHS);
5291 EVT EltTy = VT.getVectorElementType();
5293 if (EltTy == MVT::i8)
5294 Opcode = AArch64ISD::DUPLANE8;
5295 else if (EltTy == MVT::i16 || EltTy == MVT::f16)
5296 Opcode = AArch64ISD::DUPLANE16;
5297 else if (EltTy == MVT::i32 || EltTy == MVT::f32)
5298 Opcode = AArch64ISD::DUPLANE32;
5299 else if (EltTy == MVT::i64 || EltTy == MVT::f64)
5300 Opcode = AArch64ISD::DUPLANE64;
5302 llvm_unreachable("Invalid vector element type?");
5304 if (VT.getSizeInBits() == 64)
5305 OpLHS = WidenVector(OpLHS, DAG);
5306 SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, dl, MVT::i64);
5307 return DAG.getNode(Opcode, dl, VT, OpLHS, Lane);
5312 unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
5313 return DAG.getNode(AArch64ISD::EXT, dl, VT, OpLHS, OpRHS,
5314 DAG.getConstant(Imm, dl, MVT::i32));
5317 return DAG.getNode(AArch64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS,
5320 return DAG.getNode(AArch64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS,
5323 return DAG.getNode(AArch64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS,
5326 return DAG.getNode(AArch64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS,
5329 return DAG.getNode(AArch64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS,
5332 return DAG.getNode(AArch64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS,
5337 static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask,
5338 SelectionDAG &DAG) {
5339 // Check to see if we can use the TBL instruction.
5340 SDValue V1 = Op.getOperand(0);
5341 SDValue V2 = Op.getOperand(1);
5344 EVT EltVT = Op.getValueType().getVectorElementType();
5345 unsigned BytesPerElt = EltVT.getSizeInBits() / 8;
5347 SmallVector<SDValue, 8> TBLMask;
5348 for (int Val : ShuffleMask) {
5349 for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
5350 unsigned Offset = Byte + Val * BytesPerElt;
5351 TBLMask.push_back(DAG.getConstant(Offset, DL, MVT::i32));
5355 MVT IndexVT = MVT::v8i8;
5356 unsigned IndexLen = 8;
5357 if (Op.getValueType().getSizeInBits() == 128) {
5358 IndexVT = MVT::v16i8;
5362 SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
5363 SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);
5366 if (V2.getNode()->getOpcode() == ISD::UNDEF) {
5368 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
5369 Shuffle = DAG.getNode(
5370 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5371 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, DL, MVT::i32), V1Cst,
5372 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5373 makeArrayRef(TBLMask.data(), IndexLen)));
5375 if (IndexLen == 8) {
5376 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
5377 Shuffle = DAG.getNode(
5378 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5379 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, DL, MVT::i32), V1Cst,
5380 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5381 makeArrayRef(TBLMask.data(), IndexLen)));
5383 // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
5384 // cannot currently represent the register constraints on the input
5386 // Shuffle = DAG.getNode(AArch64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
5387 // DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5388 // &TBLMask[0], IndexLen));
5389 Shuffle = DAG.getNode(
5390 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5391 DAG.getConstant(Intrinsic::aarch64_neon_tbl2, DL, MVT::i32),
5393 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5394 makeArrayRef(TBLMask.data(), IndexLen)));
5397 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
5400 static unsigned getDUPLANEOp(EVT EltType) {
5401 if (EltType == MVT::i8)
5402 return AArch64ISD::DUPLANE8;
5403 if (EltType == MVT::i16 || EltType == MVT::f16)
5404 return AArch64ISD::DUPLANE16;
5405 if (EltType == MVT::i32 || EltType == MVT::f32)
5406 return AArch64ISD::DUPLANE32;
5407 if (EltType == MVT::i64 || EltType == MVT::f64)
5408 return AArch64ISD::DUPLANE64;
5410 llvm_unreachable("Invalid vector element type?");
5413 SDValue AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
5414 SelectionDAG &DAG) const {
5416 EVT VT = Op.getValueType();
5418 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5420 // Convert shuffles that are directly supported on NEON to target-specific
5421 // DAG nodes, instead of keeping them as shuffles and matching them again
5422 // during code selection. This is more efficient and avoids the possibility
5423 // of inconsistencies between legalization and selection.
5424 ArrayRef<int> ShuffleMask = SVN->getMask();
5426 SDValue V1 = Op.getOperand(0);
5427 SDValue V2 = Op.getOperand(1);
5429 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0],
5430 V1.getValueType().getSimpleVT())) {
5431 int Lane = SVN->getSplatIndex();
5432 // If this is undef splat, generate it via "just" vdup, if possible.
5436 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
5437 return DAG.getNode(AArch64ISD::DUP, dl, V1.getValueType(),
5439 // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
5440 // constant. If so, we can just reference the lane's definition directly.
5441 if (V1.getOpcode() == ISD::BUILD_VECTOR &&
5442 !isa<ConstantSDNode>(V1.getOperand(Lane)))
5443 return DAG.getNode(AArch64ISD::DUP, dl, VT, V1.getOperand(Lane));
5445 // Otherwise, duplicate from the lane of the input vector.
5446 unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
5448 // SelectionDAGBuilder may have "helpfully" already extracted or conatenated
5449 // to make a vector of the same size as this SHUFFLE. We can ignore the
5450 // extract entirely, and canonicalise the concat using WidenVector.
5451 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
5452 Lane += cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
5453 V1 = V1.getOperand(0);
5454 } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
5455 unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
5456 Lane -= Idx * VT.getVectorNumElements() / 2;
5457 V1 = WidenVector(V1.getOperand(Idx), DAG);
5458 } else if (VT.getSizeInBits() == 64)
5459 V1 = WidenVector(V1, DAG);
5461 return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, dl, MVT::i64));
5464 if (isREVMask(ShuffleMask, VT, 64))
5465 return DAG.getNode(AArch64ISD::REV64, dl, V1.getValueType(), V1, V2);
5466 if (isREVMask(ShuffleMask, VT, 32))
5467 return DAG.getNode(AArch64ISD::REV32, dl, V1.getValueType(), V1, V2);
5468 if (isREVMask(ShuffleMask, VT, 16))
5469 return DAG.getNode(AArch64ISD::REV16, dl, V1.getValueType(), V1, V2);
5471 bool ReverseEXT = false;
5473 if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
5476 Imm *= getExtFactor(V1);
5477 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V2,
5478 DAG.getConstant(Imm, dl, MVT::i32));
5479 } else if (V2->getOpcode() == ISD::UNDEF &&
5480 isSingletonEXTMask(ShuffleMask, VT, Imm)) {
5481 Imm *= getExtFactor(V1);
5482 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V1,
5483 DAG.getConstant(Imm, dl, MVT::i32));
5486 unsigned WhichResult;
5487 if (isZIPMask(ShuffleMask, VT, WhichResult)) {
5488 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
5489 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5491 if (isUZPMask(ShuffleMask, VT, WhichResult)) {
5492 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
5493 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5495 if (isTRNMask(ShuffleMask, VT, WhichResult)) {
5496 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
5497 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5500 if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5501 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
5502 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5504 if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5505 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
5506 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5508 if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5509 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
5510 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5513 SDValue Concat = tryFormConcatFromShuffle(Op, DAG);
5514 if (Concat.getNode())
5519 int NumInputElements = V1.getValueType().getVectorNumElements();
5520 if (isINSMask(ShuffleMask, NumInputElements, DstIsLeft, Anomaly)) {
5521 SDValue DstVec = DstIsLeft ? V1 : V2;
5522 SDValue DstLaneV = DAG.getConstant(Anomaly, dl, MVT::i64);
5524 SDValue SrcVec = V1;
5525 int SrcLane = ShuffleMask[Anomaly];
5526 if (SrcLane >= NumInputElements) {
5528 SrcLane -= VT.getVectorNumElements();
5530 SDValue SrcLaneV = DAG.getConstant(SrcLane, dl, MVT::i64);
5532 EVT ScalarVT = VT.getVectorElementType();
5534 if (ScalarVT.getSizeInBits() < 32 && ScalarVT.isInteger())
5535 ScalarVT = MVT::i32;
5538 ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5539 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ScalarVT, SrcVec, SrcLaneV),
5543 // If the shuffle is not directly supported and it has 4 elements, use
5544 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5545 unsigned NumElts = VT.getVectorNumElements();
5547 unsigned PFIndexes[4];
5548 for (unsigned i = 0; i != 4; ++i) {
5549 if (ShuffleMask[i] < 0)
5552 PFIndexes[i] = ShuffleMask[i];
5555 // Compute the index in the perfect shuffle table.
5556 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
5557 PFIndexes[2] * 9 + PFIndexes[3];
5558 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5559 unsigned Cost = (PFEntry >> 30);
5562 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5565 return GenerateTBL(Op, ShuffleMask, DAG);
5568 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
5570 EVT VT = BVN->getValueType(0);
5571 APInt SplatBits, SplatUndef;
5572 unsigned SplatBitSize;
5574 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5575 unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;
5577 for (unsigned i = 0; i < NumSplats; ++i) {
5578 CnstBits <<= SplatBitSize;
5579 UndefBits <<= SplatBitSize;
5580 CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits());
5581 UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits());
5590 SDValue AArch64TargetLowering::LowerVectorAND(SDValue Op,
5591 SelectionDAG &DAG) const {
5592 BuildVectorSDNode *BVN =
5593 dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5594 SDValue LHS = Op.getOperand(0);
5596 EVT VT = Op.getValueType();
5601 APInt CnstBits(VT.getSizeInBits(), 0);
5602 APInt UndefBits(VT.getSizeInBits(), 0);
5603 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5604 // We only have BIC vector immediate instruction, which is and-not.
5605 CnstBits = ~CnstBits;
5607 // We make use of a little bit of goto ickiness in order to avoid having to
5608 // duplicate the immediate matching logic for the undef toggled case.
5609 bool SecondTry = false;
5612 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5613 CnstBits = CnstBits.zextOrTrunc(64);
5614 uint64_t CnstVal = CnstBits.getZExtValue();
5616 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5617 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5618 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5619 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5620 DAG.getConstant(CnstVal, dl, MVT::i32),
5621 DAG.getConstant(0, dl, MVT::i32));
5622 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5625 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5626 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5627 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5628 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5629 DAG.getConstant(CnstVal, dl, MVT::i32),
5630 DAG.getConstant(8, dl, MVT::i32));
5631 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5634 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5635 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5636 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5637 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5638 DAG.getConstant(CnstVal, dl, MVT::i32),
5639 DAG.getConstant(16, dl, MVT::i32));
5640 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5643 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5644 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5645 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5646 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5647 DAG.getConstant(CnstVal, dl, MVT::i32),
5648 DAG.getConstant(24, dl, MVT::i32));
5649 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5652 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5653 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5654 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5655 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5656 DAG.getConstant(CnstVal, dl, MVT::i32),
5657 DAG.getConstant(0, dl, MVT::i32));
5658 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5661 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5662 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5663 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5664 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5665 DAG.getConstant(CnstVal, dl, MVT::i32),
5666 DAG.getConstant(8, dl, MVT::i32));
5667 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5674 CnstBits = ~UndefBits;
5678 // We can always fall back to a non-immediate AND.
5683 // Specialized code to quickly find if PotentialBVec is a BuildVector that
5684 // consists of only the same constant int value, returned in reference arg
5686 static bool isAllConstantBuildVector(const SDValue &PotentialBVec,
5687 uint64_t &ConstVal) {
5688 BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec);
5691 ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0));
5694 EVT VT = Bvec->getValueType(0);
5695 unsigned NumElts = VT.getVectorNumElements();
5696 for (unsigned i = 1; i < NumElts; ++i)
5697 if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt)
5699 ConstVal = FirstElt->getZExtValue();
5703 static unsigned getIntrinsicID(const SDNode *N) {
5704 unsigned Opcode = N->getOpcode();
5707 return Intrinsic::not_intrinsic;
5708 case ISD::INTRINSIC_WO_CHAIN: {
5709 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
5710 if (IID < Intrinsic::num_intrinsics)
5712 return Intrinsic::not_intrinsic;
5717 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
5718 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
5719 // BUILD_VECTORs with constant element C1, C2 is a constant, and C1 == ~C2.
5720 // Also, logical shift right -> sri, with the same structure.
5721 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
5722 EVT VT = N->getValueType(0);
5729 // Is the first op an AND?
5730 const SDValue And = N->getOperand(0);
5731 if (And.getOpcode() != ISD::AND)
5734 // Is the second op an shl or lshr?
5735 SDValue Shift = N->getOperand(1);
5736 // This will have been turned into: AArch64ISD::VSHL vector, #shift
5737 // or AArch64ISD::VLSHR vector, #shift
5738 unsigned ShiftOpc = Shift.getOpcode();
5739 if ((ShiftOpc != AArch64ISD::VSHL && ShiftOpc != AArch64ISD::VLSHR))
5741 bool IsShiftRight = ShiftOpc == AArch64ISD::VLSHR;
5743 // Is the shift amount constant?
5744 ConstantSDNode *C2node = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
5748 // Is the and mask vector all constant?
5750 if (!isAllConstantBuildVector(And.getOperand(1), C1))
5753 // Is C1 == ~C2, taking into account how much one can shift elements of a
5755 uint64_t C2 = C2node->getZExtValue();
5756 unsigned ElemSizeInBits = VT.getVectorElementType().getSizeInBits();
5757 if (C2 > ElemSizeInBits)
5759 unsigned ElemMask = (1 << ElemSizeInBits) - 1;
5760 if ((C1 & ElemMask) != (~C2 & ElemMask))
5763 SDValue X = And.getOperand(0);
5764 SDValue Y = Shift.getOperand(0);
5767 IsShiftRight ? Intrinsic::aarch64_neon_vsri : Intrinsic::aarch64_neon_vsli;
5769 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5770 DAG.getConstant(Intrin, DL, MVT::i32), X, Y,
5771 Shift.getOperand(1));
5773 DEBUG(dbgs() << "aarch64-lower: transformed: \n");
5774 DEBUG(N->dump(&DAG));
5775 DEBUG(dbgs() << "into: \n");
5776 DEBUG(ResultSLI->dump(&DAG));
5782 SDValue AArch64TargetLowering::LowerVectorOR(SDValue Op,
5783 SelectionDAG &DAG) const {
5784 // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
5785 if (EnableAArch64SlrGeneration) {
5786 SDValue Res = tryLowerToSLI(Op.getNode(), DAG);
5791 BuildVectorSDNode *BVN =
5792 dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode());
5793 SDValue LHS = Op.getOperand(1);
5795 EVT VT = Op.getValueType();
5797 // OR commutes, so try swapping the operands.
5799 LHS = Op.getOperand(0);
5800 BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5805 APInt CnstBits(VT.getSizeInBits(), 0);
5806 APInt UndefBits(VT.getSizeInBits(), 0);
5807 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5808 // We make use of a little bit of goto ickiness in order to avoid having to
5809 // duplicate the immediate matching logic for the undef toggled case.
5810 bool SecondTry = false;
5813 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5814 CnstBits = CnstBits.zextOrTrunc(64);
5815 uint64_t CnstVal = CnstBits.getZExtValue();
5817 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5818 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5819 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5820 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5821 DAG.getConstant(CnstVal, dl, MVT::i32),
5822 DAG.getConstant(0, dl, MVT::i32));
5823 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5826 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5827 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5828 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5829 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5830 DAG.getConstant(CnstVal, dl, MVT::i32),
5831 DAG.getConstant(8, dl, MVT::i32));
5832 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5835 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5836 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5837 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5838 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5839 DAG.getConstant(CnstVal, dl, MVT::i32),
5840 DAG.getConstant(16, dl, MVT::i32));
5841 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5844 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5845 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5846 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5847 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5848 DAG.getConstant(CnstVal, dl, MVT::i32),
5849 DAG.getConstant(24, dl, MVT::i32));
5850 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5853 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5854 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5855 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5856 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5857 DAG.getConstant(CnstVal, dl, MVT::i32),
5858 DAG.getConstant(0, dl, MVT::i32));
5859 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5862 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5863 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5864 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5865 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5866 DAG.getConstant(CnstVal, dl, MVT::i32),
5867 DAG.getConstant(8, dl, MVT::i32));
5868 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5875 CnstBits = UndefBits;
5879 // We can always fall back to a non-immediate OR.
5884 // Normalize the operands of BUILD_VECTOR. The value of constant operands will
5885 // be truncated to fit element width.
5886 static SDValue NormalizeBuildVector(SDValue Op,
5887 SelectionDAG &DAG) {
5888 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
5890 EVT VT = Op.getValueType();
5891 EVT EltTy= VT.getVectorElementType();
5893 if (EltTy.isFloatingPoint() || EltTy.getSizeInBits() > 16)
5896 SmallVector<SDValue, 16> Ops;
5897 for (SDValue Lane : Op->ops()) {
5898 if (auto *CstLane = dyn_cast<ConstantSDNode>(Lane)) {
5899 APInt LowBits(EltTy.getSizeInBits(),
5900 CstLane->getZExtValue());
5901 Lane = DAG.getConstant(LowBits.getZExtValue(), dl, MVT::i32);
5903 Ops.push_back(Lane);
5905 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5908 SDValue AArch64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
5909 SelectionDAG &DAG) const {
5911 EVT VT = Op.getValueType();
5912 Op = NormalizeBuildVector(Op, DAG);
5913 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5915 APInt CnstBits(VT.getSizeInBits(), 0);
5916 APInt UndefBits(VT.getSizeInBits(), 0);
5917 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5918 // We make use of a little bit of goto ickiness in order to avoid having to
5919 // duplicate the immediate matching logic for the undef toggled case.
5920 bool SecondTry = false;
5923 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5924 CnstBits = CnstBits.zextOrTrunc(64);
5925 uint64_t CnstVal = CnstBits.getZExtValue();
5927 // Certain magic vector constants (used to express things like NOT
5928 // and NEG) are passed through unmodified. This allows codegen patterns
5929 // for these operations to match. Special-purpose patterns will lower
5930 // these immediates to MOVIs if it proves necessary.
5931 if (VT.isInteger() && (CnstVal == 0 || CnstVal == ~0ULL))
5934 // The many faces of MOVI...
5935 if (AArch64_AM::isAdvSIMDModImmType10(CnstVal)) {
5936 CnstVal = AArch64_AM::encodeAdvSIMDModImmType10(CnstVal);
5937 if (VT.getSizeInBits() == 128) {
5938 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::v2i64,
5939 DAG.getConstant(CnstVal, dl, MVT::i32));
5940 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5943 // Support the V64 version via subregister insertion.
5944 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::f64,
5945 DAG.getConstant(CnstVal, dl, MVT::i32));
5946 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5949 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5950 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5951 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5952 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5953 DAG.getConstant(CnstVal, dl, MVT::i32),
5954 DAG.getConstant(0, dl, MVT::i32));
5955 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5958 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5959 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5960 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5961 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5962 DAG.getConstant(CnstVal, dl, MVT::i32),
5963 DAG.getConstant(8, dl, MVT::i32));
5964 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5967 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5968 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5969 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5970 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5971 DAG.getConstant(CnstVal, dl, MVT::i32),
5972 DAG.getConstant(16, dl, MVT::i32));
5973 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5976 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5977 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5978 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5979 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5980 DAG.getConstant(CnstVal, dl, MVT::i32),
5981 DAG.getConstant(24, dl, MVT::i32));
5982 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5985 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5986 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5987 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5988 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5989 DAG.getConstant(CnstVal, dl, MVT::i32),
5990 DAG.getConstant(0, dl, MVT::i32));
5991 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5994 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5995 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5996 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5997 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5998 DAG.getConstant(CnstVal, dl, MVT::i32),
5999 DAG.getConstant(8, dl, MVT::i32));
6000 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6003 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
6004 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
6005 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6006 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
6007 DAG.getConstant(CnstVal, dl, MVT::i32),
6008 DAG.getConstant(264, dl, MVT::i32));
6009 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6012 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
6013 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
6014 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6015 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
6016 DAG.getConstant(CnstVal, dl, MVT::i32),
6017 DAG.getConstant(272, dl, MVT::i32));
6018 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6021 if (AArch64_AM::isAdvSIMDModImmType9(CnstVal)) {
6022 CnstVal = AArch64_AM::encodeAdvSIMDModImmType9(CnstVal);
6023 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;
6024 SDValue Mov = DAG.getNode(AArch64ISD::MOVI, dl, MovTy,
6025 DAG.getConstant(CnstVal, dl, MVT::i32));
6026 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6029 // The few faces of FMOV...
6030 if (AArch64_AM::isAdvSIMDModImmType11(CnstVal)) {
6031 CnstVal = AArch64_AM::encodeAdvSIMDModImmType11(CnstVal);
6032 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4f32 : MVT::v2f32;
6033 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MovTy,
6034 DAG.getConstant(CnstVal, dl, MVT::i32));
6035 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6038 if (AArch64_AM::isAdvSIMDModImmType12(CnstVal) &&
6039 VT.getSizeInBits() == 128) {
6040 CnstVal = AArch64_AM::encodeAdvSIMDModImmType12(CnstVal);
6041 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MVT::v2f64,
6042 DAG.getConstant(CnstVal, dl, MVT::i32));
6043 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6046 // The many faces of MVNI...
6048 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
6049 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
6050 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6051 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6052 DAG.getConstant(CnstVal, dl, MVT::i32),
6053 DAG.getConstant(0, dl, MVT::i32));
6054 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6057 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
6058 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
6059 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6060 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6061 DAG.getConstant(CnstVal, dl, MVT::i32),
6062 DAG.getConstant(8, dl, MVT::i32));
6063 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6066 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
6067 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
6068 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6069 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6070 DAG.getConstant(CnstVal, dl, MVT::i32),
6071 DAG.getConstant(16, dl, MVT::i32));
6072 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6075 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
6076 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
6077 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6078 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6079 DAG.getConstant(CnstVal, dl, MVT::i32),
6080 DAG.getConstant(24, dl, MVT::i32));
6081 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6084 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
6085 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
6086 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
6087 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6088 DAG.getConstant(CnstVal, dl, MVT::i32),
6089 DAG.getConstant(0, dl, MVT::i32));
6090 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6093 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
6094 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
6095 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
6096 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6097 DAG.getConstant(CnstVal, dl, MVT::i32),
6098 DAG.getConstant(8, dl, MVT::i32));
6099 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6102 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
6103 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
6104 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6105 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
6106 DAG.getConstant(CnstVal, dl, MVT::i32),
6107 DAG.getConstant(264, dl, MVT::i32));
6108 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6111 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
6112 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
6113 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6114 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
6115 DAG.getConstant(CnstVal, dl, MVT::i32),
6116 DAG.getConstant(272, dl, MVT::i32));
6117 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6124 CnstBits = UndefBits;
6129 // Scan through the operands to find some interesting properties we can
6131 // 1) If only one value is used, we can use a DUP, or
6132 // 2) if only the low element is not undef, we can just insert that, or
6133 // 3) if only one constant value is used (w/ some non-constant lanes),
6134 // we can splat the constant value into the whole vector then fill
6135 // in the non-constant lanes.
6136 // 4) FIXME: If different constant values are used, but we can intelligently
6137 // select the values we'll be overwriting for the non-constant
6138 // lanes such that we can directly materialize the vector
6139 // some other way (MOVI, e.g.), we can be sneaky.
6140 unsigned NumElts = VT.getVectorNumElements();
6141 bool isOnlyLowElement = true;
6142 bool usesOnlyOneValue = true;
6143 bool usesOnlyOneConstantValue = true;
6144 bool isConstant = true;
6145 unsigned NumConstantLanes = 0;
6147 SDValue ConstantValue;
6148 for (unsigned i = 0; i < NumElts; ++i) {
6149 SDValue V = Op.getOperand(i);
6150 if (V.getOpcode() == ISD::UNDEF)
6153 isOnlyLowElement = false;
6154 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
6157 if (isa<ConstantSDNode>(V) || isa<ConstantFPSDNode>(V)) {
6159 if (!ConstantValue.getNode())
6161 else if (ConstantValue != V)
6162 usesOnlyOneConstantValue = false;
6165 if (!Value.getNode())
6167 else if (V != Value)
6168 usesOnlyOneValue = false;
6171 if (!Value.getNode())
6172 return DAG.getUNDEF(VT);
6174 if (isOnlyLowElement)
6175 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
6177 // Use DUP for non-constant splats. For f32 constant splats, reduce to
6178 // i32 and try again.
6179 if (usesOnlyOneValue) {
6181 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6182 Value.getValueType() != VT)
6183 return DAG.getNode(AArch64ISD::DUP, dl, VT, Value);
6185 // This is actually a DUPLANExx operation, which keeps everything vectory.
6187 // DUPLANE works on 128-bit vectors, widen it if necessary.
6188 SDValue Lane = Value.getOperand(1);
6189 Value = Value.getOperand(0);
6190 if (Value.getValueType().getSizeInBits() == 64)
6191 Value = WidenVector(Value, DAG);
6193 unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
6194 return DAG.getNode(Opcode, dl, VT, Value, Lane);
6197 if (VT.getVectorElementType().isFloatingPoint()) {
6198 SmallVector<SDValue, 8> Ops;
6199 EVT EltTy = VT.getVectorElementType();
6200 assert ((EltTy == MVT::f16 || EltTy == MVT::f32 || EltTy == MVT::f64) &&
6201 "Unsupported floating-point vector type");
6202 MVT NewType = MVT::getIntegerVT(EltTy.getSizeInBits());
6203 for (unsigned i = 0; i < NumElts; ++i)
6204 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, NewType, Op.getOperand(i)));
6205 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts);
6206 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
6207 Val = LowerBUILD_VECTOR(Val, DAG);
6209 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
6213 // If there was only one constant value used and for more than one lane,
6214 // start by splatting that value, then replace the non-constant lanes. This
6215 // is better than the default, which will perform a separate initialization
6217 if (NumConstantLanes > 0 && usesOnlyOneConstantValue) {
6218 SDValue Val = DAG.getNode(AArch64ISD::DUP, dl, VT, ConstantValue);
6219 // Now insert the non-constant lanes.
6220 for (unsigned i = 0; i < NumElts; ++i) {
6221 SDValue V = Op.getOperand(i);
6222 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i64);
6223 if (!isa<ConstantSDNode>(V) && !isa<ConstantFPSDNode>(V)) {
6224 // Note that type legalization likely mucked about with the VT of the
6225 // source operand, so we may have to convert it here before inserting.
6226 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, V, LaneIdx);
6232 // If all elements are constants and the case above didn't get hit, fall back
6233 // to the default expansion, which will generate a load from the constant
6238 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
6240 SDValue shuffle = ReconstructShuffle(Op, DAG);
6241 if (shuffle != SDValue())
6245 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
6246 // know the default expansion would otherwise fall back on something even
6247 // worse. For a vector with one or two non-undef values, that's
6248 // scalar_to_vector for the elements followed by a shuffle (provided the
6249 // shuffle is valid for the target) and materialization element by element
6250 // on the stack followed by a load for everything else.
6251 if (!isConstant && !usesOnlyOneValue) {
6252 SDValue Vec = DAG.getUNDEF(VT);
6253 SDValue Op0 = Op.getOperand(0);
6254 unsigned ElemSize = VT.getVectorElementType().getSizeInBits();
6256 // For 32 and 64 bit types, use INSERT_SUBREG for lane zero to
6257 // a) Avoid a RMW dependency on the full vector register, and
6258 // b) Allow the register coalescer to fold away the copy if the
6259 // value is already in an S or D register.
6260 if (Op0.getOpcode() != ISD::UNDEF && (ElemSize == 32 || ElemSize == 64)) {
6261 unsigned SubIdx = ElemSize == 32 ? AArch64::ssub : AArch64::dsub;
6263 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0,
6264 DAG.getTargetConstant(SubIdx, dl, MVT::i32));
6265 Vec = SDValue(N, 0);
6268 for (; i < NumElts; ++i) {
6269 SDValue V = Op.getOperand(i);
6270 if (V.getOpcode() == ISD::UNDEF)
6272 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i64);
6273 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
6278 // Just use the default expansion. We failed to find a better alternative.
6282 SDValue AArch64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
6283 SelectionDAG &DAG) const {
6284 assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!");
6286 // Check for non-constant or out of range lane.
6287 EVT VT = Op.getOperand(0).getValueType();
6288 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(2));
6289 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
6293 // Insertion/extraction are legal for V128 types.
6294 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
6295 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
6299 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
6300 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
6303 // For V64 types, we perform insertion by expanding the value
6304 // to a V128 type and perform the insertion on that.
6306 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
6307 EVT WideTy = WideVec.getValueType();
6309 SDValue Node = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideTy, WideVec,
6310 Op.getOperand(1), Op.getOperand(2));
6311 // Re-narrow the resultant vector.
6312 return NarrowVector(Node, DAG);
6316 AArch64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
6317 SelectionDAG &DAG) const {
6318 assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!");
6320 // Check for non-constant or out of range lane.
6321 EVT VT = Op.getOperand(0).getValueType();
6322 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6323 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
6327 // Insertion/extraction are legal for V128 types.
6328 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
6329 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
6333 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
6334 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
6337 // For V64 types, we perform extraction by expanding the value
6338 // to a V128 type and perform the extraction on that.
6340 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
6341 EVT WideTy = WideVec.getValueType();
6343 EVT ExtrTy = WideTy.getVectorElementType();
6344 if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8)
6347 // For extractions, we just return the result directly.
6348 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec,
6352 SDValue AArch64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
6353 SelectionDAG &DAG) const {
6354 EVT VT = Op.getOperand(0).getValueType();
6360 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6363 unsigned Val = Cst->getZExtValue();
6365 unsigned Size = Op.getValueType().getSizeInBits();
6369 return DAG.getTargetExtractSubreg(AArch64::bsub, dl, Op.getValueType(),
6372 return DAG.getTargetExtractSubreg(AArch64::hsub, dl, Op.getValueType(),
6375 return DAG.getTargetExtractSubreg(AArch64::ssub, dl, Op.getValueType(),
6378 return DAG.getTargetExtractSubreg(AArch64::dsub, dl, Op.getValueType(),
6381 llvm_unreachable("Unexpected vector type in extract_subvector!");
6384 // If this is extracting the upper 64-bits of a 128-bit vector, we match
6386 if (Size == 64 && Val * VT.getVectorElementType().getSizeInBits() == 64)
6392 bool AArch64TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
6394 if (VT.getVectorNumElements() == 4 &&
6395 (VT.is128BitVector() || VT.is64BitVector())) {
6396 unsigned PFIndexes[4];
6397 for (unsigned i = 0; i != 4; ++i) {
6401 PFIndexes[i] = M[i];
6404 // Compute the index in the perfect shuffle table.
6405 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
6406 PFIndexes[2] * 9 + PFIndexes[3];
6407 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
6408 unsigned Cost = (PFEntry >> 30);
6416 unsigned DummyUnsigned;
6418 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || isREVMask(M, VT, 64) ||
6419 isREVMask(M, VT, 32) || isREVMask(M, VT, 16) ||
6420 isEXTMask(M, VT, DummyBool, DummyUnsigned) ||
6421 // isTBLMask(M, VT) || // FIXME: Port TBL support from ARM.
6422 isTRNMask(M, VT, DummyUnsigned) || isUZPMask(M, VT, DummyUnsigned) ||
6423 isZIPMask(M, VT, DummyUnsigned) ||
6424 isTRN_v_undef_Mask(M, VT, DummyUnsigned) ||
6425 isUZP_v_undef_Mask(M, VT, DummyUnsigned) ||
6426 isZIP_v_undef_Mask(M, VT, DummyUnsigned) ||
6427 isINSMask(M, VT.getVectorNumElements(), DummyBool, DummyInt) ||
6428 isConcatMask(M, VT, VT.getSizeInBits() == 128));
6431 /// getVShiftImm - Check if this is a valid build_vector for the immediate
6432 /// operand of a vector shift operation, where all the elements of the
6433 /// build_vector must have the same constant integer value.
6434 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
6435 // Ignore bit_converts.
6436 while (Op.getOpcode() == ISD::BITCAST)
6437 Op = Op.getOperand(0);
6438 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
6439 APInt SplatBits, SplatUndef;
6440 unsigned SplatBitSize;
6442 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
6443 HasAnyUndefs, ElementBits) ||
6444 SplatBitSize > ElementBits)
6446 Cnt = SplatBits.getSExtValue();
6450 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
6451 /// operand of a vector shift left operation. That value must be in the range:
6452 /// 0 <= Value < ElementBits for a left shift; or
6453 /// 0 <= Value <= ElementBits for a long left shift.
6454 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
6455 assert(VT.isVector() && "vector shift count is not a vector type");
6456 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
6457 if (!getVShiftImm(Op, ElementBits, Cnt))
6459 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
6462 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
6463 /// operand of a vector shift right operation. The value must be in the range:
6464 /// 1 <= Value <= ElementBits for a right shift; or
6465 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, int64_t &Cnt) {
6466 assert(VT.isVector() && "vector shift count is not a vector type");
6467 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
6468 if (!getVShiftImm(Op, ElementBits, Cnt))
6470 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
6473 SDValue AArch64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
6474 SelectionDAG &DAG) const {
6475 EVT VT = Op.getValueType();
6479 if (!Op.getOperand(1).getValueType().isVector())
6481 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
6483 switch (Op.getOpcode()) {
6485 llvm_unreachable("unexpected shift opcode");
6488 if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
6489 return DAG.getNode(AArch64ISD::VSHL, DL, VT, Op.getOperand(0),
6490 DAG.getConstant(Cnt, DL, MVT::i32));
6491 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
6492 DAG.getConstant(Intrinsic::aarch64_neon_ushl, DL,
6494 Op.getOperand(0), Op.getOperand(1));
6497 // Right shift immediate
6498 if (isVShiftRImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize) {
6500 (Op.getOpcode() == ISD::SRA) ? AArch64ISD::VASHR : AArch64ISD::VLSHR;
6501 return DAG.getNode(Opc, DL, VT, Op.getOperand(0),
6502 DAG.getConstant(Cnt, DL, MVT::i32));
6505 // Right shift register. Note, there is not a shift right register
6506 // instruction, but the shift left register instruction takes a signed
6507 // value, where negative numbers specify a right shift.
6508 unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::aarch64_neon_sshl
6509 : Intrinsic::aarch64_neon_ushl;
6510 // negate the shift amount
6511 SDValue NegShift = DAG.getNode(AArch64ISD::NEG, DL, VT, Op.getOperand(1));
6512 SDValue NegShiftLeft =
6513 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
6514 DAG.getConstant(Opc, DL, MVT::i32), Op.getOperand(0),
6516 return NegShiftLeft;
6522 static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS,
6523 AArch64CC::CondCode CC, bool NoNans, EVT VT,
6524 SDLoc dl, SelectionDAG &DAG) {
6525 EVT SrcVT = LHS.getValueType();
6526 assert(VT.getSizeInBits() == SrcVT.getSizeInBits() &&
6527 "function only supposed to emit natural comparisons");
6529 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
6530 APInt CnstBits(VT.getSizeInBits(), 0);
6531 APInt UndefBits(VT.getSizeInBits(), 0);
6532 bool IsCnst = BVN && resolveBuildVector(BVN, CnstBits, UndefBits);
6533 bool IsZero = IsCnst && (CnstBits == 0);
6535 if (SrcVT.getVectorElementType().isFloatingPoint()) {
6539 case AArch64CC::NE: {
6542 Fcmeq = DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
6544 Fcmeq = DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
6545 return DAG.getNode(AArch64ISD::NOT, dl, VT, Fcmeq);
6549 return DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
6550 return DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
6553 return DAG.getNode(AArch64ISD::FCMGEz, dl, VT, LHS);
6554 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, LHS, RHS);
6557 return DAG.getNode(AArch64ISD::FCMGTz, dl, VT, LHS);
6558 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, LHS, RHS);
6561 return DAG.getNode(AArch64ISD::FCMLEz, dl, VT, LHS);
6562 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, RHS, LHS);
6566 // If we ignore NaNs then we can use to the MI implementation.
6570 return DAG.getNode(AArch64ISD::FCMLTz, dl, VT, LHS);
6571 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, RHS, LHS);
6578 case AArch64CC::NE: {
6581 Cmeq = DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
6583 Cmeq = DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
6584 return DAG.getNode(AArch64ISD::NOT, dl, VT, Cmeq);
6588 return DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
6589 return DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
6592 return DAG.getNode(AArch64ISD::CMGEz, dl, VT, LHS);
6593 return DAG.getNode(AArch64ISD::CMGE, dl, VT, LHS, RHS);
6596 return DAG.getNode(AArch64ISD::CMGTz, dl, VT, LHS);
6597 return DAG.getNode(AArch64ISD::CMGT, dl, VT, LHS, RHS);
6600 return DAG.getNode(AArch64ISD::CMLEz, dl, VT, LHS);
6601 return DAG.getNode(AArch64ISD::CMGE, dl, VT, RHS, LHS);
6603 return DAG.getNode(AArch64ISD::CMHS, dl, VT, RHS, LHS);
6605 return DAG.getNode(AArch64ISD::CMHI, dl, VT, RHS, LHS);
6608 return DAG.getNode(AArch64ISD::CMLTz, dl, VT, LHS);
6609 return DAG.getNode(AArch64ISD::CMGT, dl, VT, RHS, LHS);
6611 return DAG.getNode(AArch64ISD::CMHI, dl, VT, LHS, RHS);
6613 return DAG.getNode(AArch64ISD::CMHS, dl, VT, LHS, RHS);
6617 SDValue AArch64TargetLowering::LowerVSETCC(SDValue Op,
6618 SelectionDAG &DAG) const {
6619 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6620 SDValue LHS = Op.getOperand(0);
6621 SDValue RHS = Op.getOperand(1);
6622 EVT CmpVT = LHS.getValueType().changeVectorElementTypeToInteger();
6625 if (LHS.getValueType().getVectorElementType().isInteger()) {
6626 assert(LHS.getValueType() == RHS.getValueType());
6627 AArch64CC::CondCode AArch64CC = changeIntCCToAArch64CC(CC);
6629 EmitVectorComparison(LHS, RHS, AArch64CC, false, CmpVT, dl, DAG);
6630 return DAG.getSExtOrTrunc(Cmp, dl, Op.getValueType());
6633 assert(LHS.getValueType().getVectorElementType() == MVT::f32 ||
6634 LHS.getValueType().getVectorElementType() == MVT::f64);
6636 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
6637 // clean. Some of them require two branches to implement.
6638 AArch64CC::CondCode CC1, CC2;
6640 changeVectorFPCCToAArch64CC(CC, CC1, CC2, ShouldInvert);
6642 bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath;
6644 EmitVectorComparison(LHS, RHS, CC1, NoNaNs, CmpVT, dl, DAG);
6648 if (CC2 != AArch64CC::AL) {
6650 EmitVectorComparison(LHS, RHS, CC2, NoNaNs, CmpVT, dl, DAG);
6651 if (!Cmp2.getNode())
6654 Cmp = DAG.getNode(ISD::OR, dl, CmpVT, Cmp, Cmp2);
6657 Cmp = DAG.getSExtOrTrunc(Cmp, dl, Op.getValueType());
6660 return Cmp = DAG.getNOT(dl, Cmp, Cmp.getValueType());
6665 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
6666 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
6667 /// specified in the intrinsic calls.
6668 bool AArch64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
6670 unsigned Intrinsic) const {
6671 auto &DL = I.getModule()->getDataLayout();
6672 switch (Intrinsic) {
6673 case Intrinsic::aarch64_neon_ld2:
6674 case Intrinsic::aarch64_neon_ld3:
6675 case Intrinsic::aarch64_neon_ld4:
6676 case Intrinsic::aarch64_neon_ld1x2:
6677 case Intrinsic::aarch64_neon_ld1x3:
6678 case Intrinsic::aarch64_neon_ld1x4:
6679 case Intrinsic::aarch64_neon_ld2lane:
6680 case Intrinsic::aarch64_neon_ld3lane:
6681 case Intrinsic::aarch64_neon_ld4lane:
6682 case Intrinsic::aarch64_neon_ld2r:
6683 case Intrinsic::aarch64_neon_ld3r:
6684 case Intrinsic::aarch64_neon_ld4r: {
6685 Info.opc = ISD::INTRINSIC_W_CHAIN;
6686 // Conservatively set memVT to the entire set of vectors loaded.
6687 uint64_t NumElts = DL.getTypeAllocSize(I.getType()) / 8;
6688 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6689 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6692 Info.vol = false; // volatile loads with NEON intrinsics not supported
6693 Info.readMem = true;
6694 Info.writeMem = false;
6697 case Intrinsic::aarch64_neon_st2:
6698 case Intrinsic::aarch64_neon_st3:
6699 case Intrinsic::aarch64_neon_st4:
6700 case Intrinsic::aarch64_neon_st1x2:
6701 case Intrinsic::aarch64_neon_st1x3:
6702 case Intrinsic::aarch64_neon_st1x4:
6703 case Intrinsic::aarch64_neon_st2lane:
6704 case Intrinsic::aarch64_neon_st3lane:
6705 case Intrinsic::aarch64_neon_st4lane: {
6706 Info.opc = ISD::INTRINSIC_VOID;
6707 // Conservatively set memVT to the entire set of vectors stored.
6708 unsigned NumElts = 0;
6709 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
6710 Type *ArgTy = I.getArgOperand(ArgI)->getType();
6711 if (!ArgTy->isVectorTy())
6713 NumElts += DL.getTypeAllocSize(ArgTy) / 8;
6715 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6716 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6719 Info.vol = false; // volatile stores with NEON intrinsics not supported
6720 Info.readMem = false;
6721 Info.writeMem = true;
6724 case Intrinsic::aarch64_ldaxr:
6725 case Intrinsic::aarch64_ldxr: {
6726 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
6727 Info.opc = ISD::INTRINSIC_W_CHAIN;
6728 Info.memVT = MVT::getVT(PtrTy->getElementType());
6729 Info.ptrVal = I.getArgOperand(0);
6731 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
6733 Info.readMem = true;
6734 Info.writeMem = false;
6737 case Intrinsic::aarch64_stlxr:
6738 case Intrinsic::aarch64_stxr: {
6739 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
6740 Info.opc = ISD::INTRINSIC_W_CHAIN;
6741 Info.memVT = MVT::getVT(PtrTy->getElementType());
6742 Info.ptrVal = I.getArgOperand(1);
6744 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
6746 Info.readMem = false;
6747 Info.writeMem = true;
6750 case Intrinsic::aarch64_ldaxp:
6751 case Intrinsic::aarch64_ldxp: {
6752 Info.opc = ISD::INTRINSIC_W_CHAIN;
6753 Info.memVT = MVT::i128;
6754 Info.ptrVal = I.getArgOperand(0);
6758 Info.readMem = true;
6759 Info.writeMem = false;
6762 case Intrinsic::aarch64_stlxp:
6763 case Intrinsic::aarch64_stxp: {
6764 Info.opc = ISD::INTRINSIC_W_CHAIN;
6765 Info.memVT = MVT::i128;
6766 Info.ptrVal = I.getArgOperand(2);
6770 Info.readMem = false;
6771 Info.writeMem = true;
6781 // Truncations from 64-bit GPR to 32-bit GPR is free.
6782 bool AArch64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
6783 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6785 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6786 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6787 return NumBits1 > NumBits2;
6789 bool AArch64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
6790 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6792 unsigned NumBits1 = VT1.getSizeInBits();
6793 unsigned NumBits2 = VT2.getSizeInBits();
6794 return NumBits1 > NumBits2;
6797 /// Check if it is profitable to hoist instruction in then/else to if.
6798 /// Not profitable if I and it's user can form a FMA instruction
6799 /// because we prefer FMSUB/FMADD.
6800 bool AArch64TargetLowering::isProfitableToHoist(Instruction *I) const {
6801 if (I->getOpcode() != Instruction::FMul)
6804 if (I->getNumUses() != 1)
6807 Instruction *User = I->user_back();
6810 !(User->getOpcode() == Instruction::FSub ||
6811 User->getOpcode() == Instruction::FAdd))
6814 const TargetOptions &Options = getTargetMachine().Options;
6815 const DataLayout &DL = I->getModule()->getDataLayout();
6816 EVT VT = getValueType(DL, User->getOperand(0)->getType());
6818 if (isFMAFasterThanFMulAndFAdd(VT) &&
6819 isOperationLegalOrCustom(ISD::FMA, VT) &&
6820 (Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath))
6826 // All 32-bit GPR operations implicitly zero the high-half of the corresponding
6828 bool AArch64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
6829 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6831 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6832 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6833 return NumBits1 == 32 && NumBits2 == 64;
6835 bool AArch64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
6836 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6838 unsigned NumBits1 = VT1.getSizeInBits();
6839 unsigned NumBits2 = VT2.getSizeInBits();
6840 return NumBits1 == 32 && NumBits2 == 64;
6843 bool AArch64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
6844 EVT VT1 = Val.getValueType();
6845 if (isZExtFree(VT1, VT2)) {
6849 if (Val.getOpcode() != ISD::LOAD)
6852 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
6853 return (VT1.isSimple() && !VT1.isVector() && VT1.isInteger() &&
6854 VT2.isSimple() && !VT2.isVector() && VT2.isInteger() &&
6855 VT1.getSizeInBits() <= 32);
6858 bool AArch64TargetLowering::isExtFreeImpl(const Instruction *Ext) const {
6859 if (isa<FPExtInst>(Ext))
6862 // Vector types are next free.
6863 if (Ext->getType()->isVectorTy())
6866 for (const Use &U : Ext->uses()) {
6867 // The extension is free if we can fold it with a left shift in an
6868 // addressing mode or an arithmetic operation: add, sub, and cmp.
6870 // Is there a shift?
6871 const Instruction *Instr = cast<Instruction>(U.getUser());
6873 // Is this a constant shift?
6874 switch (Instr->getOpcode()) {
6875 case Instruction::Shl:
6876 if (!isa<ConstantInt>(Instr->getOperand(1)))
6879 case Instruction::GetElementPtr: {
6880 gep_type_iterator GTI = gep_type_begin(Instr);
6881 auto &DL = Ext->getModule()->getDataLayout();
6882 std::advance(GTI, U.getOperandNo());
6884 // This extension will end up with a shift because of the scaling factor.
6885 // 8-bit sized types have a scaling factor of 1, thus a shift amount of 0.
6886 // Get the shift amount based on the scaling factor:
6887 // log2(sizeof(IdxTy)) - log2(8).
6889 countTrailingZeros(DL.getTypeStoreSizeInBits(IdxTy)) - 3;
6890 // Is the constant foldable in the shift of the addressing mode?
6891 // I.e., shift amount is between 1 and 4 inclusive.
6892 if (ShiftAmt == 0 || ShiftAmt > 4)
6896 case Instruction::Trunc:
6897 // Check if this is a noop.
6898 // trunc(sext ty1 to ty2) to ty1.
6899 if (Instr->getType() == Ext->getOperand(0)->getType())
6906 // At this point we can use the bfm family, so this extension is free
6912 bool AArch64TargetLowering::hasPairedLoad(Type *LoadedType,
6913 unsigned &RequiredAligment) const {
6914 if (!LoadedType->isIntegerTy() && !LoadedType->isFloatTy())
6916 // Cyclone supports unaligned accesses.
6917 RequiredAligment = 0;
6918 unsigned NumBits = LoadedType->getPrimitiveSizeInBits();
6919 return NumBits == 32 || NumBits == 64;
6922 bool AArch64TargetLowering::hasPairedLoad(EVT LoadedType,
6923 unsigned &RequiredAligment) const {
6924 if (!LoadedType.isSimple() ||
6925 (!LoadedType.isInteger() && !LoadedType.isFloatingPoint()))
6927 // Cyclone supports unaligned accesses.
6928 RequiredAligment = 0;
6929 unsigned NumBits = LoadedType.getSizeInBits();
6930 return NumBits == 32 || NumBits == 64;
6933 /// \brief Lower an interleaved load into a ldN intrinsic.
6935 /// E.g. Lower an interleaved load (Factor = 2):
6936 /// %wide.vec = load <8 x i32>, <8 x i32>* %ptr
6937 /// %v0 = shuffle %wide.vec, undef, <0, 2, 4, 6> ; Extract even elements
6938 /// %v1 = shuffle %wide.vec, undef, <1, 3, 5, 7> ; Extract odd elements
6941 /// %ld2 = { <4 x i32>, <4 x i32> } call llvm.aarch64.neon.ld2(%ptr)
6942 /// %vec0 = extractelement { <4 x i32>, <4 x i32> } %ld2, i32 0
6943 /// %vec1 = extractelement { <4 x i32>, <4 x i32> } %ld2, i32 1
6944 bool AArch64TargetLowering::lowerInterleavedLoad(
6945 LoadInst *LI, ArrayRef<ShuffleVectorInst *> Shuffles,
6946 ArrayRef<unsigned> Indices, unsigned Factor) const {
6947 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
6948 "Invalid interleave factor");
6949 assert(!Shuffles.empty() && "Empty shufflevector input");
6950 assert(Shuffles.size() == Indices.size() &&
6951 "Unmatched number of shufflevectors and indices");
6953 const DataLayout &DL = LI->getModule()->getDataLayout();
6955 VectorType *VecTy = Shuffles[0]->getType();
6956 unsigned VecSize = DL.getTypeAllocSizeInBits(VecTy);
6958 // Skip illegal vector types.
6959 if (VecSize != 64 && VecSize != 128)
6962 // A pointer vector can not be the return type of the ldN intrinsics. Need to
6963 // load integer vectors first and then convert to pointer vectors.
6964 Type *EltTy = VecTy->getVectorElementType();
6965 if (EltTy->isPointerTy())
6967 VectorType::get(DL.getIntPtrType(EltTy), VecTy->getVectorNumElements());
6969 Type *PtrTy = VecTy->getPointerTo(LI->getPointerAddressSpace());
6970 Type *Tys[2] = {VecTy, PtrTy};
6971 static const Intrinsic::ID LoadInts[3] = {Intrinsic::aarch64_neon_ld2,
6972 Intrinsic::aarch64_neon_ld3,
6973 Intrinsic::aarch64_neon_ld4};
6975 Intrinsic::getDeclaration(LI->getModule(), LoadInts[Factor - 2], Tys);
6977 IRBuilder<> Builder(LI);
6978 Value *Ptr = Builder.CreateBitCast(LI->getPointerOperand(), PtrTy);
6980 CallInst *LdN = Builder.CreateCall(LdNFunc, Ptr, "ldN");
6982 // Replace uses of each shufflevector with the corresponding vector loaded
6984 for (unsigned i = 0; i < Shuffles.size(); i++) {
6985 ShuffleVectorInst *SVI = Shuffles[i];
6986 unsigned Index = Indices[i];
6988 Value *SubVec = Builder.CreateExtractValue(LdN, Index);
6990 // Convert the integer vector to pointer vector if the element is pointer.
6991 if (EltTy->isPointerTy())
6992 SubVec = Builder.CreateIntToPtr(SubVec, SVI->getType());
6994 SVI->replaceAllUsesWith(SubVec);
7000 /// \brief Get a mask consisting of sequential integers starting from \p Start.
7002 /// I.e. <Start, Start + 1, ..., Start + NumElts - 1>
7003 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned Start,
7005 SmallVector<Constant *, 16> Mask;
7006 for (unsigned i = 0; i < NumElts; i++)
7007 Mask.push_back(Builder.getInt32(Start + i));
7009 return ConstantVector::get(Mask);
7012 /// \brief Lower an interleaved store into a stN intrinsic.
7014 /// E.g. Lower an interleaved store (Factor = 3):
7015 /// %i.vec = shuffle <8 x i32> %v0, <8 x i32> %v1,
7016 /// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11>
7017 /// store <12 x i32> %i.vec, <12 x i32>* %ptr
7020 /// %sub.v0 = shuffle <8 x i32> %v0, <8 x i32> v1, <0, 1, 2, 3>
7021 /// %sub.v1 = shuffle <8 x i32> %v0, <8 x i32> v1, <4, 5, 6, 7>
7022 /// %sub.v2 = shuffle <8 x i32> %v0, <8 x i32> v1, <8, 9, 10, 11>
7023 /// call void llvm.aarch64.neon.st3(%sub.v0, %sub.v1, %sub.v2, %ptr)
7025 /// Note that the new shufflevectors will be removed and we'll only generate one
7026 /// st3 instruction in CodeGen.
7027 bool AArch64TargetLowering::lowerInterleavedStore(StoreInst *SI,
7028 ShuffleVectorInst *SVI,
7029 unsigned Factor) const {
7030 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
7031 "Invalid interleave factor");
7033 VectorType *VecTy = SVI->getType();
7034 assert(VecTy->getVectorNumElements() % Factor == 0 &&
7035 "Invalid interleaved store");
7037 unsigned NumSubElts = VecTy->getVectorNumElements() / Factor;
7038 Type *EltTy = VecTy->getVectorElementType();
7039 VectorType *SubVecTy = VectorType::get(EltTy, NumSubElts);
7041 const DataLayout &DL = SI->getModule()->getDataLayout();
7042 unsigned SubVecSize = DL.getTypeAllocSizeInBits(SubVecTy);
7044 // Skip illegal vector types.
7045 if (SubVecSize != 64 && SubVecSize != 128)
7048 Value *Op0 = SVI->getOperand(0);
7049 Value *Op1 = SVI->getOperand(1);
7050 IRBuilder<> Builder(SI);
7052 // StN intrinsics don't support pointer vectors as arguments. Convert pointer
7053 // vectors to integer vectors.
7054 if (EltTy->isPointerTy()) {
7055 Type *IntTy = DL.getIntPtrType(EltTy);
7056 unsigned NumOpElts =
7057 dyn_cast<VectorType>(Op0->getType())->getVectorNumElements();
7059 // Convert to the corresponding integer vector.
7060 Type *IntVecTy = VectorType::get(IntTy, NumOpElts);
7061 Op0 = Builder.CreatePtrToInt(Op0, IntVecTy);
7062 Op1 = Builder.CreatePtrToInt(Op1, IntVecTy);
7064 SubVecTy = VectorType::get(IntTy, NumSubElts);
7067 Type *PtrTy = SubVecTy->getPointerTo(SI->getPointerAddressSpace());
7068 Type *Tys[2] = {SubVecTy, PtrTy};
7069 static const Intrinsic::ID StoreInts[3] = {Intrinsic::aarch64_neon_st2,
7070 Intrinsic::aarch64_neon_st3,
7071 Intrinsic::aarch64_neon_st4};
7073 Intrinsic::getDeclaration(SI->getModule(), StoreInts[Factor - 2], Tys);
7075 SmallVector<Value *, 5> Ops;
7077 // Split the shufflevector operands into sub vectors for the new stN call.
7078 for (unsigned i = 0; i < Factor; i++)
7079 Ops.push_back(Builder.CreateShuffleVector(
7080 Op0, Op1, getSequentialMask(Builder, NumSubElts * i, NumSubElts)));
7082 Ops.push_back(Builder.CreateBitCast(SI->getPointerOperand(), PtrTy));
7083 Builder.CreateCall(StNFunc, Ops);
7087 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
7088 unsigned AlignCheck) {
7089 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
7090 (DstAlign == 0 || DstAlign % AlignCheck == 0));
7093 EVT AArch64TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
7094 unsigned SrcAlign, bool IsMemset,
7097 MachineFunction &MF) const {
7098 // Don't use AdvSIMD to implement 16-byte memset. It would have taken one
7099 // instruction to materialize the v2i64 zero and one store (with restrictive
7100 // addressing mode). Just do two i64 store of zero-registers.
7102 const Function *F = MF.getFunction();
7103 if (Subtarget->hasFPARMv8() && !IsMemset && Size >= 16 &&
7104 !F->hasFnAttribute(Attribute::NoImplicitFloat) &&
7105 (memOpAlign(SrcAlign, DstAlign, 16) ||
7106 (allowsMisalignedMemoryAccesses(MVT::f128, 0, 1, &Fast) && Fast)))
7110 (memOpAlign(SrcAlign, DstAlign, 8) ||
7111 (allowsMisalignedMemoryAccesses(MVT::i64, 0, 1, &Fast) && Fast)))
7115 (memOpAlign(SrcAlign, DstAlign, 4) ||
7116 (allowsMisalignedMemoryAccesses(MVT::i32, 0, 1, &Fast) && Fast)))
7122 // 12-bit optionally shifted immediates are legal for adds.
7123 bool AArch64TargetLowering::isLegalAddImmediate(int64_t Immed) const {
7124 if ((Immed >> 12) == 0 || ((Immed & 0xfff) == 0 && Immed >> 24 == 0))
7129 // Integer comparisons are implemented with ADDS/SUBS, so the range of valid
7130 // immediates is the same as for an add or a sub.
7131 bool AArch64TargetLowering::isLegalICmpImmediate(int64_t Immed) const {
7134 return isLegalAddImmediate(Immed);
7137 /// isLegalAddressingMode - Return true if the addressing mode represented
7138 /// by AM is legal for this target, for a load/store of the specified type.
7139 bool AArch64TargetLowering::isLegalAddressingMode(const DataLayout &DL,
7140 const AddrMode &AM, Type *Ty,
7141 unsigned AS) const {
7142 // AArch64 has five basic addressing modes:
7144 // reg + 9-bit signed offset
7145 // reg + SIZE_IN_BYTES * 12-bit unsigned offset
7147 // reg + SIZE_IN_BYTES * reg
7149 // No global is ever allowed as a base.
7153 // No reg+reg+imm addressing.
7154 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
7157 // check reg + imm case:
7158 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
7159 uint64_t NumBytes = 0;
7160 if (Ty->isSized()) {
7161 uint64_t NumBits = DL.getTypeSizeInBits(Ty);
7162 NumBytes = NumBits / 8;
7163 if (!isPowerOf2_64(NumBits))
7168 int64_t Offset = AM.BaseOffs;
7170 // 9-bit signed offset
7171 if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
7174 // 12-bit unsigned offset
7175 unsigned shift = Log2_64(NumBytes);
7176 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
7177 // Must be a multiple of NumBytes (NumBytes is a power of 2)
7178 (Offset >> shift) << shift == Offset)
7183 // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2
7185 if (!AM.Scale || AM.Scale == 1 ||
7186 (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
7191 int AArch64TargetLowering::getScalingFactorCost(const DataLayout &DL,
7192 const AddrMode &AM, Type *Ty,
7193 unsigned AS) const {
7194 // Scaling factors are not free at all.
7195 // Operands | Rt Latency
7196 // -------------------------------------------
7198 // -------------------------------------------
7199 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
7200 // Rt, [Xn, Wm, <extend> #imm] |
7201 if (isLegalAddressingMode(DL, AM, Ty, AS))
7202 // Scale represents reg2 * scale, thus account for 1 if
7203 // it is not equal to 0 or 1.
7204 return AM.Scale != 0 && AM.Scale != 1;
7208 bool AArch64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
7209 VT = VT.getScalarType();
7214 switch (VT.getSimpleVT().SimpleTy) {
7226 AArch64TargetLowering::getScratchRegisters(CallingConv::ID) const {
7227 // LR is a callee-save register, but we must treat it as clobbered by any call
7228 // site. Hence we include LR in the scratch registers, which are in turn added
7229 // as implicit-defs for stackmaps and patchpoints.
7230 static const MCPhysReg ScratchRegs[] = {
7231 AArch64::X16, AArch64::X17, AArch64::LR, 0
7237 AArch64TargetLowering::isDesirableToCommuteWithShift(const SDNode *N) const {
7238 EVT VT = N->getValueType(0);
7239 // If N is unsigned bit extraction: ((x >> C) & mask), then do not combine
7240 // it with shift to let it be lowered to UBFX.
7241 if (N->getOpcode() == ISD::AND && (VT == MVT::i32 || VT == MVT::i64) &&
7242 isa<ConstantSDNode>(N->getOperand(1))) {
7243 uint64_t TruncMask = N->getConstantOperandVal(1);
7244 if (isMask_64(TruncMask) &&
7245 N->getOperand(0).getOpcode() == ISD::SRL &&
7246 isa<ConstantSDNode>(N->getOperand(0)->getOperand(1)))
7252 bool AArch64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
7254 assert(Ty->isIntegerTy());
7256 unsigned BitSize = Ty->getPrimitiveSizeInBits();
7260 int64_t Val = Imm.getSExtValue();
7261 if (Val == 0 || AArch64_AM::isLogicalImmediate(Val, BitSize))
7264 if ((int64_t)Val < 0)
7267 Val &= (1LL << 32) - 1;
7269 unsigned LZ = countLeadingZeros((uint64_t)Val);
7270 unsigned Shift = (63 - LZ) / 16;
7271 // MOVZ is free so return true for one or fewer MOVK.
7275 // Generate SUBS and CSEL for integer abs.
7276 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
7277 EVT VT = N->getValueType(0);
7279 SDValue N0 = N->getOperand(0);
7280 SDValue N1 = N->getOperand(1);
7283 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
7284 // and change it to SUB and CSEL.
7285 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
7286 N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
7287 N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0))
7288 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
7289 if (Y1C->getAPIntValue() == VT.getSizeInBits() - 1) {
7290 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT),
7292 // Generate SUBS & CSEL.
7294 DAG.getNode(AArch64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32),
7295 N0.getOperand(0), DAG.getConstant(0, DL, VT));
7296 return DAG.getNode(AArch64ISD::CSEL, DL, VT, N0.getOperand(0), Neg,
7297 DAG.getConstant(AArch64CC::PL, DL, MVT::i32),
7298 SDValue(Cmp.getNode(), 1));
7303 // performXorCombine - Attempts to handle integer ABS.
7304 static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
7305 TargetLowering::DAGCombinerInfo &DCI,
7306 const AArch64Subtarget *Subtarget) {
7307 if (DCI.isBeforeLegalizeOps())
7310 return performIntegerAbsCombine(N, DAG);
7314 AArch64TargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor,
7316 std::vector<SDNode *> *Created) const {
7317 // fold (sdiv X, pow2)
7318 EVT VT = N->getValueType(0);
7319 if ((VT != MVT::i32 && VT != MVT::i64) ||
7320 !(Divisor.isPowerOf2() || (-Divisor).isPowerOf2()))
7324 SDValue N0 = N->getOperand(0);
7325 unsigned Lg2 = Divisor.countTrailingZeros();
7326 SDValue Zero = DAG.getConstant(0, DL, VT);
7327 SDValue Pow2MinusOne = DAG.getConstant((1ULL << Lg2) - 1, DL, VT);
7329 // Add (N0 < 0) ? Pow2 - 1 : 0;
7331 SDValue Cmp = getAArch64Cmp(N0, Zero, ISD::SETLT, CCVal, DAG, DL);
7332 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0, Pow2MinusOne);
7333 SDValue CSel = DAG.getNode(AArch64ISD::CSEL, DL, VT, Add, N0, CCVal, Cmp);
7336 Created->push_back(Cmp.getNode());
7337 Created->push_back(Add.getNode());
7338 Created->push_back(CSel.getNode());
7343 DAG.getNode(ISD::SRA, DL, VT, CSel, DAG.getConstant(Lg2, DL, MVT::i64));
7345 // If we're dividing by a positive value, we're done. Otherwise, we must
7346 // negate the result.
7347 if (Divisor.isNonNegative())
7351 Created->push_back(SRA.getNode());
7352 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), SRA);
7355 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG,
7356 TargetLowering::DAGCombinerInfo &DCI,
7357 const AArch64Subtarget *Subtarget) {
7358 if (DCI.isBeforeLegalizeOps())
7361 // Multiplication of a power of two plus/minus one can be done more
7362 // cheaply as as shift+add/sub. For now, this is true unilaterally. If
7363 // future CPUs have a cheaper MADD instruction, this may need to be
7364 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
7365 // 64-bit is 5 cycles, so this is always a win.
7366 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
7367 APInt Value = C->getAPIntValue();
7368 EVT VT = N->getValueType(0);
7370 if (Value.isNonNegative()) {
7371 // (mul x, 2^N + 1) => (add (shl x, N), x)
7372 APInt VM1 = Value - 1;
7373 if (VM1.isPowerOf2()) {
7374 SDValue ShiftedVal =
7375 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7376 DAG.getConstant(VM1.logBase2(), DL, MVT::i64));
7377 return DAG.getNode(ISD::ADD, DL, VT, ShiftedVal,
7380 // (mul x, 2^N - 1) => (sub (shl x, N), x)
7381 APInt VP1 = Value + 1;
7382 if (VP1.isPowerOf2()) {
7383 SDValue ShiftedVal =
7384 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7385 DAG.getConstant(VP1.logBase2(), DL, MVT::i64));
7386 return DAG.getNode(ISD::SUB, DL, VT, ShiftedVal,
7390 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
7391 APInt VNP1 = -Value + 1;
7392 if (VNP1.isPowerOf2()) {
7393 SDValue ShiftedVal =
7394 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7395 DAG.getConstant(VNP1.logBase2(), DL, MVT::i64));
7396 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0),
7399 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
7400 APInt VNM1 = -Value - 1;
7401 if (VNM1.isPowerOf2()) {
7402 SDValue ShiftedVal =
7403 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7404 DAG.getConstant(VNM1.logBase2(), DL, MVT::i64));
7406 DAG.getNode(ISD::ADD, DL, VT, ShiftedVal, N->getOperand(0));
7407 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Add);
7414 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
7415 SelectionDAG &DAG) {
7416 // Take advantage of vector comparisons producing 0 or -1 in each lane to
7417 // optimize away operation when it's from a constant.
7419 // The general transformation is:
7420 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
7421 // AND(VECTOR_CMP(x,y), constant2)
7422 // constant2 = UNARYOP(constant)
7424 // Early exit if this isn't a vector operation, the operand of the
7425 // unary operation isn't a bitwise AND, or if the sizes of the operations
7427 EVT VT = N->getValueType(0);
7428 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
7429 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
7430 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
7433 // Now check that the other operand of the AND is a constant. We could
7434 // make the transformation for non-constant splats as well, but it's unclear
7435 // that would be a benefit as it would not eliminate any operations, just
7436 // perform one more step in scalar code before moving to the vector unit.
7437 if (BuildVectorSDNode *BV =
7438 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
7439 // Bail out if the vector isn't a constant.
7440 if (!BV->isConstant())
7443 // Everything checks out. Build up the new and improved node.
7445 EVT IntVT = BV->getValueType(0);
7446 // Create a new constant of the appropriate type for the transformed
7448 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
7449 // The AND node needs bitcasts to/from an integer vector type around it.
7450 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
7451 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
7452 N->getOperand(0)->getOperand(0), MaskConst);
7453 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
7460 static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG,
7461 const AArch64Subtarget *Subtarget) {
7462 // First try to optimize away the conversion when it's conditionally from
7463 // a constant. Vectors only.
7464 SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
7465 if (Res != SDValue())
7468 EVT VT = N->getValueType(0);
7469 if (VT != MVT::f32 && VT != MVT::f64)
7472 // Only optimize when the source and destination types have the same width.
7473 if (VT.getSizeInBits() != N->getOperand(0).getValueType().getSizeInBits())
7476 // If the result of an integer load is only used by an integer-to-float
7477 // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
7478 // This eliminates an "integer-to-vector-move UOP and improve throughput.
7479 SDValue N0 = N->getOperand(0);
7480 if (Subtarget->hasNEON() && ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7481 // Do not change the width of a volatile load.
7482 !cast<LoadSDNode>(N0)->isVolatile()) {
7483 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7484 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
7485 LN0->getPointerInfo(), LN0->isVolatile(),
7486 LN0->isNonTemporal(), LN0->isInvariant(),
7487 LN0->getAlignment());
7489 // Make sure successors of the original load stay after it by updating them
7490 // to use the new Chain.
7491 DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));
7494 (N->getOpcode() == ISD::SINT_TO_FP) ? AArch64ISD::SITOF : AArch64ISD::UITOF;
7495 return DAG.getNode(Opcode, SDLoc(N), VT, Load);
7501 /// An EXTR instruction is made up of two shifts, ORed together. This helper
7502 /// searches for and classifies those shifts.
7503 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
7505 if (N.getOpcode() == ISD::SHL)
7507 else if (N.getOpcode() == ISD::SRL)
7512 if (!isa<ConstantSDNode>(N.getOperand(1)))
7515 ShiftAmount = N->getConstantOperandVal(1);
7516 Src = N->getOperand(0);
7520 /// EXTR instruction extracts a contiguous chunk of bits from two existing
7521 /// registers viewed as a high/low pair. This function looks for the pattern:
7522 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
7523 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
7525 static SDValue tryCombineToEXTR(SDNode *N,
7526 TargetLowering::DAGCombinerInfo &DCI) {
7527 SelectionDAG &DAG = DCI.DAG;
7529 EVT VT = N->getValueType(0);
7531 assert(N->getOpcode() == ISD::OR && "Unexpected root");
7533 if (VT != MVT::i32 && VT != MVT::i64)
7537 uint32_t ShiftLHS = 0;
7539 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
7543 uint32_t ShiftRHS = 0;
7545 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
7548 // If they're both trying to come from the high part of the register, they're
7549 // not really an EXTR.
7550 if (LHSFromHi == RHSFromHi)
7553 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
7557 std::swap(LHS, RHS);
7558 std::swap(ShiftLHS, ShiftRHS);
7561 return DAG.getNode(AArch64ISD::EXTR, DL, VT, LHS, RHS,
7562 DAG.getConstant(ShiftRHS, DL, MVT::i64));
7565 static SDValue tryCombineToBSL(SDNode *N,
7566 TargetLowering::DAGCombinerInfo &DCI) {
7567 EVT VT = N->getValueType(0);
7568 SelectionDAG &DAG = DCI.DAG;
7574 SDValue N0 = N->getOperand(0);
7575 if (N0.getOpcode() != ISD::AND)
7578 SDValue N1 = N->getOperand(1);
7579 if (N1.getOpcode() != ISD::AND)
7582 // We only have to look for constant vectors here since the general, variable
7583 // case can be handled in TableGen.
7584 unsigned Bits = VT.getVectorElementType().getSizeInBits();
7585 uint64_t BitMask = Bits == 64 ? -1ULL : ((1ULL << Bits) - 1);
7586 for (int i = 1; i >= 0; --i)
7587 for (int j = 1; j >= 0; --j) {
7588 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(i));
7589 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(j));
7593 bool FoundMatch = true;
7594 for (unsigned k = 0; k < VT.getVectorNumElements(); ++k) {
7595 ConstantSDNode *CN0 = dyn_cast<ConstantSDNode>(BVN0->getOperand(k));
7596 ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(BVN1->getOperand(k));
7598 CN0->getZExtValue() != (BitMask & ~CN1->getZExtValue())) {
7605 return DAG.getNode(AArch64ISD::BSL, DL, VT, SDValue(BVN0, 0),
7606 N0->getOperand(1 - i), N1->getOperand(1 - j));
7612 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
7613 const AArch64Subtarget *Subtarget) {
7614 // Attempt to form an EXTR from (or (shl VAL1, #N), (srl VAL2, #RegWidth-N))
7615 if (!EnableAArch64ExtrGeneration)
7617 SelectionDAG &DAG = DCI.DAG;
7618 EVT VT = N->getValueType(0);
7620 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7623 SDValue Res = tryCombineToEXTR(N, DCI);
7627 Res = tryCombineToBSL(N, DCI);
7634 static SDValue performBitcastCombine(SDNode *N,
7635 TargetLowering::DAGCombinerInfo &DCI,
7636 SelectionDAG &DAG) {
7637 // Wait 'til after everything is legalized to try this. That way we have
7638 // legal vector types and such.
7639 if (DCI.isBeforeLegalizeOps())
7642 // Remove extraneous bitcasts around an extract_subvector.
7644 // (v4i16 (bitconvert
7645 // (extract_subvector (v2i64 (bitconvert (v8i16 ...)), (i64 1)))))
7647 // (extract_subvector ((v8i16 ...), (i64 4)))
7649 // Only interested in 64-bit vectors as the ultimate result.
7650 EVT VT = N->getValueType(0);
7653 if (VT.getSimpleVT().getSizeInBits() != 64)
7655 // Is the operand an extract_subvector starting at the beginning or halfway
7656 // point of the vector? A low half may also come through as an
7657 // EXTRACT_SUBREG, so look for that, too.
7658 SDValue Op0 = N->getOperand(0);
7659 if (Op0->getOpcode() != ISD::EXTRACT_SUBVECTOR &&
7660 !(Op0->isMachineOpcode() &&
7661 Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG))
7663 uint64_t idx = cast<ConstantSDNode>(Op0->getOperand(1))->getZExtValue();
7664 if (Op0->getOpcode() == ISD::EXTRACT_SUBVECTOR) {
7665 if (Op0->getValueType(0).getVectorNumElements() != idx && idx != 0)
7667 } else if (Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG) {
7668 if (idx != AArch64::dsub)
7670 // The dsub reference is equivalent to a lane zero subvector reference.
7673 // Look through the bitcast of the input to the extract.
7674 if (Op0->getOperand(0)->getOpcode() != ISD::BITCAST)
7676 SDValue Source = Op0->getOperand(0)->getOperand(0);
7677 // If the source type has twice the number of elements as our destination
7678 // type, we know this is an extract of the high or low half of the vector.
7679 EVT SVT = Source->getValueType(0);
7680 if (SVT.getVectorNumElements() != VT.getVectorNumElements() * 2)
7683 DEBUG(dbgs() << "aarch64-lower: bitcast extract_subvector simplification\n");
7685 // Create the simplified form to just extract the low or high half of the
7686 // vector directly rather than bothering with the bitcasts.
7688 unsigned NumElements = VT.getVectorNumElements();
7690 SDValue HalfIdx = DAG.getConstant(NumElements, dl, MVT::i64);
7691 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Source, HalfIdx);
7693 SDValue SubReg = DAG.getTargetConstant(AArch64::dsub, dl, MVT::i32);
7694 return SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl, VT,
7700 static SDValue performConcatVectorsCombine(SDNode *N,
7701 TargetLowering::DAGCombinerInfo &DCI,
7702 SelectionDAG &DAG) {
7704 EVT VT = N->getValueType(0);
7705 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
7707 // Optimize concat_vectors of truncated vectors, where the intermediate
7708 // type is illegal, to avoid said illegality, e.g.,
7709 // (v4i16 (concat_vectors (v2i16 (truncate (v2i64))),
7710 // (v2i16 (truncate (v2i64)))))
7712 // (v4i16 (truncate (vector_shuffle (v4i32 (bitcast (v2i64))),
7713 // (v4i32 (bitcast (v2i64))),
7715 // This isn't really target-specific, but ISD::TRUNCATE legality isn't keyed
7716 // on both input and result type, so we might generate worse code.
7717 // On AArch64 we know it's fine for v2i64->v4i16 and v4i32->v8i8.
7718 if (N->getNumOperands() == 2 &&
7719 N0->getOpcode() == ISD::TRUNCATE &&
7720 N1->getOpcode() == ISD::TRUNCATE) {
7721 SDValue N00 = N0->getOperand(0);
7722 SDValue N10 = N1->getOperand(0);
7723 EVT N00VT = N00.getValueType();
7725 if (N00VT == N10.getValueType() &&
7726 (N00VT == MVT::v2i64 || N00VT == MVT::v4i32) &&
7727 N00VT.getScalarSizeInBits() == 4 * VT.getScalarSizeInBits()) {
7728 MVT MidVT = (N00VT == MVT::v2i64 ? MVT::v4i32 : MVT::v8i16);
7729 SmallVector<int, 8> Mask(MidVT.getVectorNumElements());
7730 for (size_t i = 0; i < Mask.size(); ++i)
7732 return DAG.getNode(ISD::TRUNCATE, dl, VT,
7733 DAG.getVectorShuffle(
7735 DAG.getNode(ISD::BITCAST, dl, MidVT, N00),
7736 DAG.getNode(ISD::BITCAST, dl, MidVT, N10), Mask));
7740 // Wait 'til after everything is legalized to try this. That way we have
7741 // legal vector types and such.
7742 if (DCI.isBeforeLegalizeOps())
7745 // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector
7746 // splat. The indexed instructions are going to be expecting a DUPLANE64, so
7747 // canonicalise to that.
7748 if (N0 == N1 && VT.getVectorNumElements() == 2) {
7749 assert(VT.getVectorElementType().getSizeInBits() == 64);
7750 return DAG.getNode(AArch64ISD::DUPLANE64, dl, VT, WidenVector(N0, DAG),
7751 DAG.getConstant(0, dl, MVT::i64));
7754 // Canonicalise concat_vectors so that the right-hand vector has as few
7755 // bit-casts as possible before its real operation. The primary matching
7756 // destination for these operations will be the narrowing "2" instructions,
7757 // which depend on the operation being performed on this right-hand vector.
7759 // (concat_vectors LHS, (v1i64 (bitconvert (v4i16 RHS))))
7761 // (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS))
7763 if (N1->getOpcode() != ISD::BITCAST)
7765 SDValue RHS = N1->getOperand(0);
7766 MVT RHSTy = RHS.getValueType().getSimpleVT();
7767 // If the RHS is not a vector, this is not the pattern we're looking for.
7768 if (!RHSTy.isVector())
7771 DEBUG(dbgs() << "aarch64-lower: concat_vectors bitcast simplification\n");
7773 MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(),
7774 RHSTy.getVectorNumElements() * 2);
7775 return DAG.getNode(ISD::BITCAST, dl, VT,
7776 DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatTy,
7777 DAG.getNode(ISD::BITCAST, dl, RHSTy, N0),
7781 static SDValue tryCombineFixedPointConvert(SDNode *N,
7782 TargetLowering::DAGCombinerInfo &DCI,
7783 SelectionDAG &DAG) {
7784 // Wait 'til after everything is legalized to try this. That way we have
7785 // legal vector types and such.
7786 if (DCI.isBeforeLegalizeOps())
7788 // Transform a scalar conversion of a value from a lane extract into a
7789 // lane extract of a vector conversion. E.g., from foo1 to foo2:
7790 // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); }
7791 // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; }
7793 // The second form interacts better with instruction selection and the
7794 // register allocator to avoid cross-class register copies that aren't
7795 // coalescable due to a lane reference.
7797 // Check the operand and see if it originates from a lane extract.
7798 SDValue Op1 = N->getOperand(1);
7799 if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7800 // Yep, no additional predication needed. Perform the transform.
7801 SDValue IID = N->getOperand(0);
7802 SDValue Shift = N->getOperand(2);
7803 SDValue Vec = Op1.getOperand(0);
7804 SDValue Lane = Op1.getOperand(1);
7805 EVT ResTy = N->getValueType(0);
7809 // The vector width should be 128 bits by the time we get here, even
7810 // if it started as 64 bits (the extract_vector handling will have
7812 assert(Vec.getValueType().getSizeInBits() == 128 &&
7813 "unexpected vector size on extract_vector_elt!");
7814 if (Vec.getValueType() == MVT::v4i32)
7815 VecResTy = MVT::v4f32;
7816 else if (Vec.getValueType() == MVT::v2i64)
7817 VecResTy = MVT::v2f64;
7819 llvm_unreachable("unexpected vector type!");
7822 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift);
7823 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane);
7828 // AArch64 high-vector "long" operations are formed by performing the non-high
7829 // version on an extract_subvector of each operand which gets the high half:
7831 // (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS))
7833 // However, there are cases which don't have an extract_high explicitly, but
7834 // have another operation that can be made compatible with one for free. For
7837 // (dupv64 scalar) --> (extract_high (dup128 scalar))
7839 // This routine does the actual conversion of such DUPs, once outer routines
7840 // have determined that everything else is in order.
7841 // It also supports immediate DUP-like nodes (MOVI/MVNi), which we can fold
7843 static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) {
7844 switch (N.getOpcode()) {
7845 case AArch64ISD::DUP:
7846 case AArch64ISD::DUPLANE8:
7847 case AArch64ISD::DUPLANE16:
7848 case AArch64ISD::DUPLANE32:
7849 case AArch64ISD::DUPLANE64:
7850 case AArch64ISD::MOVI:
7851 case AArch64ISD::MOVIshift:
7852 case AArch64ISD::MOVIedit:
7853 case AArch64ISD::MOVImsl:
7854 case AArch64ISD::MVNIshift:
7855 case AArch64ISD::MVNImsl:
7858 // FMOV could be supported, but isn't very useful, as it would only occur
7859 // if you passed a bitcast' floating point immediate to an eligible long
7860 // integer op (addl, smull, ...).
7864 MVT NarrowTy = N.getSimpleValueType();
7865 if (!NarrowTy.is64BitVector())
7868 MVT ElementTy = NarrowTy.getVectorElementType();
7869 unsigned NumElems = NarrowTy.getVectorNumElements();
7870 MVT NewVT = MVT::getVectorVT(ElementTy, NumElems * 2);
7873 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NarrowTy,
7874 DAG.getNode(N->getOpcode(), dl, NewVT, N->ops()),
7875 DAG.getConstant(NumElems, dl, MVT::i64));
7878 static bool isEssentiallyExtractSubvector(SDValue N) {
7879 if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR)
7882 return N.getOpcode() == ISD::BITCAST &&
7883 N.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR;
7886 /// \brief Helper structure to keep track of ISD::SET_CC operands.
7887 struct GenericSetCCInfo {
7888 const SDValue *Opnd0;
7889 const SDValue *Opnd1;
7893 /// \brief Helper structure to keep track of a SET_CC lowered into AArch64 code.
7894 struct AArch64SetCCInfo {
7896 AArch64CC::CondCode CC;
7899 /// \brief Helper structure to keep track of SetCC information.
7901 GenericSetCCInfo Generic;
7902 AArch64SetCCInfo AArch64;
7905 /// \brief Helper structure to be able to read SetCC information. If set to
7906 /// true, IsAArch64 field, Info is a AArch64SetCCInfo, otherwise Info is a
7907 /// GenericSetCCInfo.
7908 struct SetCCInfoAndKind {
7913 /// \brief Check whether or not \p Op is a SET_CC operation, either a generic or
7915 /// AArch64 lowered one.
7916 /// \p SetCCInfo is filled accordingly.
7917 /// \post SetCCInfo is meanginfull only when this function returns true.
7918 /// \return True when Op is a kind of SET_CC operation.
7919 static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) {
7920 // If this is a setcc, this is straight forward.
7921 if (Op.getOpcode() == ISD::SETCC) {
7922 SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0);
7923 SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1);
7924 SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
7925 SetCCInfo.IsAArch64 = false;
7928 // Otherwise, check if this is a matching csel instruction.
7932 if (Op.getOpcode() != AArch64ISD::CSEL)
7934 // Set the information about the operands.
7935 // TODO: we want the operands of the Cmp not the csel
7936 SetCCInfo.Info.AArch64.Cmp = &Op.getOperand(3);
7937 SetCCInfo.IsAArch64 = true;
7938 SetCCInfo.Info.AArch64.CC = static_cast<AArch64CC::CondCode>(
7939 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
7941 // Check that the operands matches the constraints:
7942 // (1) Both operands must be constants.
7943 // (2) One must be 1 and the other must be 0.
7944 ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0));
7945 ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7948 if (!TValue || !FValue)
7952 if (!TValue->isOne()) {
7953 // Update the comparison when we are interested in !cc.
7954 std::swap(TValue, FValue);
7955 SetCCInfo.Info.AArch64.CC =
7956 AArch64CC::getInvertedCondCode(SetCCInfo.Info.AArch64.CC);
7958 return TValue->isOne() && FValue->isNullValue();
7961 // Returns true if Op is setcc or zext of setcc.
7962 static bool isSetCCOrZExtSetCC(const SDValue& Op, SetCCInfoAndKind &Info) {
7963 if (isSetCC(Op, Info))
7965 return ((Op.getOpcode() == ISD::ZERO_EXTEND) &&
7966 isSetCC(Op->getOperand(0), Info));
7969 // The folding we want to perform is:
7970 // (add x, [zext] (setcc cc ...) )
7972 // (csel x, (add x, 1), !cc ...)
7974 // The latter will get matched to a CSINC instruction.
7975 static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) {
7976 assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!");
7977 SDValue LHS = Op->getOperand(0);
7978 SDValue RHS = Op->getOperand(1);
7979 SetCCInfoAndKind InfoAndKind;
7981 // If neither operand is a SET_CC, give up.
7982 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind)) {
7983 std::swap(LHS, RHS);
7984 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind))
7988 // FIXME: This could be generatized to work for FP comparisons.
7989 EVT CmpVT = InfoAndKind.IsAArch64
7990 ? InfoAndKind.Info.AArch64.Cmp->getOperand(0).getValueType()
7991 : InfoAndKind.Info.Generic.Opnd0->getValueType();
7992 if (CmpVT != MVT::i32 && CmpVT != MVT::i64)
7998 if (InfoAndKind.IsAArch64) {
7999 CCVal = DAG.getConstant(
8000 AArch64CC::getInvertedCondCode(InfoAndKind.Info.AArch64.CC), dl,
8002 Cmp = *InfoAndKind.Info.AArch64.Cmp;
8004 Cmp = getAArch64Cmp(*InfoAndKind.Info.Generic.Opnd0,
8005 *InfoAndKind.Info.Generic.Opnd1,
8006 ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, true),
8009 EVT VT = Op->getValueType(0);
8010 LHS = DAG.getNode(ISD::ADD, dl, VT, RHS, DAG.getConstant(1, dl, VT));
8011 return DAG.getNode(AArch64ISD::CSEL, dl, VT, RHS, LHS, CCVal, Cmp);
8014 // The basic add/sub long vector instructions have variants with "2" on the end
8015 // which act on the high-half of their inputs. They are normally matched by
8018 // (add (zeroext (extract_high LHS)),
8019 // (zeroext (extract_high RHS)))
8020 // -> uaddl2 vD, vN, vM
8022 // However, if one of the extracts is something like a duplicate, this
8023 // instruction can still be used profitably. This function puts the DAG into a
8024 // more appropriate form for those patterns to trigger.
8025 static SDValue performAddSubLongCombine(SDNode *N,
8026 TargetLowering::DAGCombinerInfo &DCI,
8027 SelectionDAG &DAG) {
8028 if (DCI.isBeforeLegalizeOps())
8031 MVT VT = N->getSimpleValueType(0);
8032 if (!VT.is128BitVector()) {
8033 if (N->getOpcode() == ISD::ADD)
8034 return performSetccAddFolding(N, DAG);
8038 // Make sure both branches are extended in the same way.
8039 SDValue LHS = N->getOperand(0);
8040 SDValue RHS = N->getOperand(1);
8041 if ((LHS.getOpcode() != ISD::ZERO_EXTEND &&
8042 LHS.getOpcode() != ISD::SIGN_EXTEND) ||
8043 LHS.getOpcode() != RHS.getOpcode())
8046 unsigned ExtType = LHS.getOpcode();
8048 // It's not worth doing if at least one of the inputs isn't already an
8049 // extract, but we don't know which it'll be so we have to try both.
8050 if (isEssentiallyExtractSubvector(LHS.getOperand(0))) {
8051 RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG);
8055 RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS);
8056 } else if (isEssentiallyExtractSubvector(RHS.getOperand(0))) {
8057 LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG);
8061 LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS);
8064 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS);
8067 // Massage DAGs which we can use the high-half "long" operations on into
8068 // something isel will recognize better. E.g.
8070 // (aarch64_neon_umull (extract_high vec) (dupv64 scalar)) -->
8071 // (aarch64_neon_umull (extract_high (v2i64 vec)))
8072 // (extract_high (v2i64 (dup128 scalar)))))
8074 static SDValue tryCombineLongOpWithDup(SDNode *N,
8075 TargetLowering::DAGCombinerInfo &DCI,
8076 SelectionDAG &DAG) {
8077 if (DCI.isBeforeLegalizeOps())
8080 bool IsIntrinsic = N->getOpcode() == ISD::INTRINSIC_WO_CHAIN;
8081 SDValue LHS = N->getOperand(IsIntrinsic ? 1 : 0);
8082 SDValue RHS = N->getOperand(IsIntrinsic ? 2 : 1);
8083 assert(LHS.getValueType().is64BitVector() &&
8084 RHS.getValueType().is64BitVector() &&
8085 "unexpected shape for long operation");
8087 // Either node could be a DUP, but it's not worth doing both of them (you'd
8088 // just as well use the non-high version) so look for a corresponding extract
8089 // operation on the other "wing".
8090 if (isEssentiallyExtractSubvector(LHS)) {
8091 RHS = tryExtendDUPToExtractHigh(RHS, DAG);
8094 } else if (isEssentiallyExtractSubvector(RHS)) {
8095 LHS = tryExtendDUPToExtractHigh(LHS, DAG);
8100 // N could either be an intrinsic or a sabsdiff/uabsdiff node.
8102 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0),
8103 N->getOperand(0), LHS, RHS);
8105 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
8109 static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
8110 MVT ElemTy = N->getSimpleValueType(0).getScalarType();
8111 unsigned ElemBits = ElemTy.getSizeInBits();
8113 int64_t ShiftAmount;
8114 if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) {
8115 APInt SplatValue, SplatUndef;
8116 unsigned SplatBitSize;
8118 if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
8119 HasAnyUndefs, ElemBits) ||
8120 SplatBitSize != ElemBits)
8123 ShiftAmount = SplatValue.getSExtValue();
8124 } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) {
8125 ShiftAmount = CVN->getSExtValue();
8133 llvm_unreachable("Unknown shift intrinsic");
8134 case Intrinsic::aarch64_neon_sqshl:
8135 Opcode = AArch64ISD::SQSHL_I;
8136 IsRightShift = false;
8138 case Intrinsic::aarch64_neon_uqshl:
8139 Opcode = AArch64ISD::UQSHL_I;
8140 IsRightShift = false;
8142 case Intrinsic::aarch64_neon_srshl:
8143 Opcode = AArch64ISD::SRSHR_I;
8144 IsRightShift = true;
8146 case Intrinsic::aarch64_neon_urshl:
8147 Opcode = AArch64ISD::URSHR_I;
8148 IsRightShift = true;
8150 case Intrinsic::aarch64_neon_sqshlu:
8151 Opcode = AArch64ISD::SQSHLU_I;
8152 IsRightShift = false;
8156 if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits) {
8158 return DAG.getNode(Opcode, dl, N->getValueType(0), N->getOperand(1),
8159 DAG.getConstant(-ShiftAmount, dl, MVT::i32));
8160 } else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount < ElemBits) {
8162 return DAG.getNode(Opcode, dl, N->getValueType(0), N->getOperand(1),
8163 DAG.getConstant(ShiftAmount, dl, MVT::i32));
8169 // The CRC32[BH] instructions ignore the high bits of their data operand. Since
8170 // the intrinsics must be legal and take an i32, this means there's almost
8171 // certainly going to be a zext in the DAG which we can eliminate.
8172 static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) {
8173 SDValue AndN = N->getOperand(2);
8174 if (AndN.getOpcode() != ISD::AND)
8177 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1));
8178 if (!CMask || CMask->getZExtValue() != Mask)
8181 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32,
8182 N->getOperand(0), N->getOperand(1), AndN.getOperand(0));
8185 static SDValue combineAcrossLanesIntrinsic(unsigned Opc, SDNode *N,
8186 SelectionDAG &DAG) {
8188 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0),
8189 DAG.getNode(Opc, dl,
8190 N->getOperand(1).getSimpleValueType(),
8192 DAG.getConstant(0, dl, MVT::i64));
8195 static SDValue performIntrinsicCombine(SDNode *N,
8196 TargetLowering::DAGCombinerInfo &DCI,
8197 const AArch64Subtarget *Subtarget) {
8198 SelectionDAG &DAG = DCI.DAG;
8199 unsigned IID = getIntrinsicID(N);
8203 case Intrinsic::aarch64_neon_vcvtfxs2fp:
8204 case Intrinsic::aarch64_neon_vcvtfxu2fp:
8205 return tryCombineFixedPointConvert(N, DCI, DAG);
8207 case Intrinsic::aarch64_neon_saddv:
8208 return combineAcrossLanesIntrinsic(AArch64ISD::SADDV, N, DAG);
8209 case Intrinsic::aarch64_neon_uaddv:
8210 return combineAcrossLanesIntrinsic(AArch64ISD::UADDV, N, DAG);
8211 case Intrinsic::aarch64_neon_sminv:
8212 return combineAcrossLanesIntrinsic(AArch64ISD::SMINV, N, DAG);
8213 case Intrinsic::aarch64_neon_uminv:
8214 return combineAcrossLanesIntrinsic(AArch64ISD::UMINV, N, DAG);
8215 case Intrinsic::aarch64_neon_smaxv:
8216 return combineAcrossLanesIntrinsic(AArch64ISD::SMAXV, N, DAG);
8217 case Intrinsic::aarch64_neon_umaxv:
8218 return combineAcrossLanesIntrinsic(AArch64ISD::UMAXV, N, DAG);
8219 case Intrinsic::aarch64_neon_fmax:
8220 return DAG.getNode(AArch64ISD::FMAX, SDLoc(N), N->getValueType(0),
8221 N->getOperand(1), N->getOperand(2));
8222 case Intrinsic::aarch64_neon_fmin:
8223 return DAG.getNode(AArch64ISD::FMIN, SDLoc(N), N->getValueType(0),
8224 N->getOperand(1), N->getOperand(2));
8225 case Intrinsic::aarch64_neon_sabd:
8226 return DAG.getNode(ISD::SABSDIFF, SDLoc(N), N->getValueType(0),
8227 N->getOperand(1), N->getOperand(2));
8228 case Intrinsic::aarch64_neon_uabd:
8229 return DAG.getNode(ISD::UABSDIFF, SDLoc(N), N->getValueType(0),
8230 N->getOperand(1), N->getOperand(2));
8231 case Intrinsic::aarch64_neon_smull:
8232 case Intrinsic::aarch64_neon_umull:
8233 case Intrinsic::aarch64_neon_pmull:
8234 case Intrinsic::aarch64_neon_sqdmull:
8235 return tryCombineLongOpWithDup(N, DCI, DAG);
8236 case Intrinsic::aarch64_neon_sqshl:
8237 case Intrinsic::aarch64_neon_uqshl:
8238 case Intrinsic::aarch64_neon_sqshlu:
8239 case Intrinsic::aarch64_neon_srshl:
8240 case Intrinsic::aarch64_neon_urshl:
8241 return tryCombineShiftImm(IID, N, DAG);
8242 case Intrinsic::aarch64_crc32b:
8243 case Intrinsic::aarch64_crc32cb:
8244 return tryCombineCRC32(0xff, N, DAG);
8245 case Intrinsic::aarch64_crc32h:
8246 case Intrinsic::aarch64_crc32ch:
8247 return tryCombineCRC32(0xffff, N, DAG);
8252 static SDValue performExtendCombine(SDNode *N,
8253 TargetLowering::DAGCombinerInfo &DCI,
8254 SelectionDAG &DAG) {
8255 // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then
8256 // we can convert that DUP into another extract_high (of a bigger DUP), which
8257 // helps the backend to decide that an sabdl2 would be useful, saving a real
8258 // extract_high operation.
8259 if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND &&
8260 (N->getOperand(0).getOpcode() == ISD::SABSDIFF ||
8261 N->getOperand(0).getOpcode() == ISD::UABSDIFF)) {
8262 SDNode *ABDNode = N->getOperand(0).getNode();
8263 SDValue NewABD = tryCombineLongOpWithDup(ABDNode, DCI, DAG);
8264 if (!NewABD.getNode())
8267 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0),
8271 // This is effectively a custom type legalization for AArch64.
8273 // Type legalization will split an extend of a small, legal, type to a larger
8274 // illegal type by first splitting the destination type, often creating
8275 // illegal source types, which then get legalized in isel-confusing ways,
8276 // leading to really terrible codegen. E.g.,
8277 // %result = v8i32 sext v8i8 %value
8279 // %losrc = extract_subreg %value, ...
8280 // %hisrc = extract_subreg %value, ...
8281 // %lo = v4i32 sext v4i8 %losrc
8282 // %hi = v4i32 sext v4i8 %hisrc
8283 // Things go rapidly downhill from there.
8285 // For AArch64, the [sz]ext vector instructions can only go up one element
8286 // size, so we can, e.g., extend from i8 to i16, but to go from i8 to i32
8287 // take two instructions.
8289 // This implies that the most efficient way to do the extend from v8i8
8290 // to two v4i32 values is to first extend the v8i8 to v8i16, then do
8291 // the normal splitting to happen for the v8i16->v8i32.
8293 // This is pre-legalization to catch some cases where the default
8294 // type legalization will create ill-tempered code.
8295 if (!DCI.isBeforeLegalizeOps())
8298 // We're only interested in cleaning things up for non-legal vector types
8299 // here. If both the source and destination are legal, things will just
8300 // work naturally without any fiddling.
8301 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8302 EVT ResVT = N->getValueType(0);
8303 if (!ResVT.isVector() || TLI.isTypeLegal(ResVT))
8305 // If the vector type isn't a simple VT, it's beyond the scope of what
8306 // we're worried about here. Let legalization do its thing and hope for
8308 SDValue Src = N->getOperand(0);
8309 EVT SrcVT = Src->getValueType(0);
8310 if (!ResVT.isSimple() || !SrcVT.isSimple())
8313 // If the source VT is a 64-bit vector, we can play games and get the
8314 // better results we want.
8315 if (SrcVT.getSizeInBits() != 64)
8318 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
8319 unsigned ElementCount = SrcVT.getVectorNumElements();
8320 SrcVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize * 2), ElementCount);
8322 Src = DAG.getNode(N->getOpcode(), DL, SrcVT, Src);
8324 // Now split the rest of the operation into two halves, each with a 64
8328 unsigned NumElements = ResVT.getVectorNumElements();
8329 assert(!(NumElements & 1) && "Splitting vector, but not in half!");
8330 LoVT = HiVT = EVT::getVectorVT(*DAG.getContext(),
8331 ResVT.getVectorElementType(), NumElements / 2);
8333 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
8334 LoVT.getVectorNumElements());
8335 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
8336 DAG.getConstant(0, DL, MVT::i64));
8337 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
8338 DAG.getConstant(InNVT.getVectorNumElements(), DL, MVT::i64));
8339 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, Lo);
8340 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, Hi);
8342 // Now combine the parts back together so we still have a single result
8343 // like the combiner expects.
8344 return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
8347 /// Replace a splat of a scalar to a vector store by scalar stores of the scalar
8348 /// value. The load store optimizer pass will merge them to store pair stores.
8349 /// This has better performance than a splat of the scalar followed by a split
8350 /// vector store. Even if the stores are not merged it is four stores vs a dup,
8351 /// followed by an ext.b and two stores.
8352 static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) {
8353 SDValue StVal = St->getValue();
8354 EVT VT = StVal.getValueType();
8356 // Don't replace floating point stores, they possibly won't be transformed to
8357 // stp because of the store pair suppress pass.
8358 if (VT.isFloatingPoint())
8361 // Check for insert vector elements.
8362 if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT)
8365 // We can express a splat as store pair(s) for 2 or 4 elements.
8366 unsigned NumVecElts = VT.getVectorNumElements();
8367 if (NumVecElts != 4 && NumVecElts != 2)
8369 SDValue SplatVal = StVal.getOperand(1);
8370 unsigned RemainInsertElts = NumVecElts - 1;
8372 // Check that this is a splat.
8373 while (--RemainInsertElts) {
8374 SDValue NextInsertElt = StVal.getOperand(0);
8375 if (NextInsertElt.getOpcode() != ISD::INSERT_VECTOR_ELT)
8377 if (NextInsertElt.getOperand(1) != SplatVal)
8379 StVal = NextInsertElt;
8381 unsigned OrigAlignment = St->getAlignment();
8382 unsigned EltOffset = NumVecElts == 4 ? 4 : 8;
8383 unsigned Alignment = std::min(OrigAlignment, EltOffset);
8385 // Create scalar stores. This is at least as good as the code sequence for a
8386 // split unaligned store wich is a dup.s, ext.b, and two stores.
8387 // Most of the time the three stores should be replaced by store pair
8388 // instructions (stp).
8390 SDValue BasePtr = St->getBasePtr();
8392 DAG.getStore(St->getChain(), DL, SplatVal, BasePtr, St->getPointerInfo(),
8393 St->isVolatile(), St->isNonTemporal(), St->getAlignment());
8395 unsigned Offset = EltOffset;
8396 while (--NumVecElts) {
8397 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
8398 DAG.getConstant(Offset, DL, MVT::i64));
8399 NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr,
8400 St->getPointerInfo(), St->isVolatile(),
8401 St->isNonTemporal(), Alignment);
8402 Offset += EltOffset;
8407 static SDValue performSTORECombine(SDNode *N,
8408 TargetLowering::DAGCombinerInfo &DCI,
8410 const AArch64Subtarget *Subtarget) {
8411 if (!DCI.isBeforeLegalize())
8414 StoreSDNode *S = cast<StoreSDNode>(N);
8415 if (S->isVolatile())
8418 // Cyclone has bad performance on unaligned 16B stores when crossing line and
8419 // page boundaries. We want to split such stores.
8420 if (!Subtarget->isCyclone())
8423 // Don't split at Oz.
8424 MachineFunction &MF = DAG.getMachineFunction();
8425 bool IsMinSize = MF.getFunction()->hasFnAttribute(Attribute::MinSize);
8429 SDValue StVal = S->getValue();
8430 EVT VT = StVal.getValueType();
8432 // Don't split v2i64 vectors. Memcpy lowering produces those and splitting
8433 // those up regresses performance on micro-benchmarks and olden/bh.
8434 if (!VT.isVector() || VT.getVectorNumElements() < 2 || VT == MVT::v2i64)
8437 // Split unaligned 16B stores. They are terrible for performance.
8438 // Don't split stores with alignment of 1 or 2. Code that uses clang vector
8439 // extensions can use this to mark that it does not want splitting to happen
8440 // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of
8441 // eliminating alignment hazards is only 1 in 8 for alignment of 2.
8442 if (VT.getSizeInBits() != 128 || S->getAlignment() >= 16 ||
8443 S->getAlignment() <= 2)
8446 // If we get a splat of a scalar convert this vector store to a store of
8447 // scalars. They will be merged into store pairs thereby removing two
8449 SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S);
8450 if (ReplacedSplat != SDValue())
8451 return ReplacedSplat;
8454 unsigned NumElts = VT.getVectorNumElements() / 2;
8455 // Split VT into two.
8457 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), NumElts);
8458 SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
8459 DAG.getConstant(0, DL, MVT::i64));
8460 SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
8461 DAG.getConstant(NumElts, DL, MVT::i64));
8462 SDValue BasePtr = S->getBasePtr();
8464 DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
8465 S->isVolatile(), S->isNonTemporal(), S->getAlignment());
8466 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
8467 DAG.getConstant(8, DL, MVT::i64));
8468 return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr,
8469 S->getPointerInfo(), S->isVolatile(), S->isNonTemporal(),
8473 /// Target-specific DAG combine function for post-increment LD1 (lane) and
8474 /// post-increment LD1R.
8475 static SDValue performPostLD1Combine(SDNode *N,
8476 TargetLowering::DAGCombinerInfo &DCI,
8478 if (DCI.isBeforeLegalizeOps())
8481 SelectionDAG &DAG = DCI.DAG;
8482 EVT VT = N->getValueType(0);
8484 unsigned LoadIdx = IsLaneOp ? 1 : 0;
8485 SDNode *LD = N->getOperand(LoadIdx).getNode();
8486 // If it is not LOAD, can not do such combine.
8487 if (LD->getOpcode() != ISD::LOAD)
8490 LoadSDNode *LoadSDN = cast<LoadSDNode>(LD);
8491 EVT MemVT = LoadSDN->getMemoryVT();
8492 // Check if memory operand is the same type as the vector element.
8493 if (MemVT != VT.getVectorElementType())
8496 // Check if there are other uses. If so, do not combine as it will introduce
8498 for (SDNode::use_iterator UI = LD->use_begin(), UE = LD->use_end(); UI != UE;
8500 if (UI.getUse().getResNo() == 1) // Ignore uses of the chain result.
8506 SDValue Addr = LD->getOperand(1);
8507 SDValue Vector = N->getOperand(0);
8508 // Search for a use of the address operand that is an increment.
8509 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(), UE =
8510 Addr.getNode()->use_end(); UI != UE; ++UI) {
8512 if (User->getOpcode() != ISD::ADD
8513 || UI.getUse().getResNo() != Addr.getResNo())
8516 // Check that the add is independent of the load. Otherwise, folding it
8517 // would create a cycle.
8518 if (User->isPredecessorOf(LD) || LD->isPredecessorOf(User))
8520 // Also check that add is not used in the vector operand. This would also
8522 if (User->isPredecessorOf(Vector.getNode()))
8525 // If the increment is a constant, it must match the memory ref size.
8526 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8527 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8528 uint32_t IncVal = CInc->getZExtValue();
8529 unsigned NumBytes = VT.getScalarSizeInBits() / 8;
8530 if (IncVal != NumBytes)
8532 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
8535 // Finally, check that the vector doesn't depend on the load.
8536 // Again, this would create a cycle.
8537 // The load depending on the vector is fine, as that's the case for the
8538 // LD1*post we'll eventually generate anyway.
8539 if (LoadSDN->isPredecessorOf(Vector.getNode()))
8542 SmallVector<SDValue, 8> Ops;
8543 Ops.push_back(LD->getOperand(0)); // Chain
8545 Ops.push_back(Vector); // The vector to be inserted
8546 Ops.push_back(N->getOperand(2)); // The lane to be inserted in the vector
8548 Ops.push_back(Addr);
8551 EVT Tys[3] = { VT, MVT::i64, MVT::Other };
8552 SDVTList SDTys = DAG.getVTList(Tys);
8553 unsigned NewOp = IsLaneOp ? AArch64ISD::LD1LANEpost : AArch64ISD::LD1DUPpost;
8554 SDValue UpdN = DAG.getMemIntrinsicNode(NewOp, SDLoc(N), SDTys, Ops,
8556 LoadSDN->getMemOperand());
8559 SmallVector<SDValue, 2> NewResults;
8560 NewResults.push_back(SDValue(LD, 0)); // The result of load
8561 NewResults.push_back(SDValue(UpdN.getNode(), 2)); // Chain
8562 DCI.CombineTo(LD, NewResults);
8563 DCI.CombineTo(N, SDValue(UpdN.getNode(), 0)); // Dup/Inserted Result
8564 DCI.CombineTo(User, SDValue(UpdN.getNode(), 1)); // Write back register
8571 /// Target-specific DAG combine function for NEON load/store intrinsics
8572 /// to merge base address updates.
8573 static SDValue performNEONPostLDSTCombine(SDNode *N,
8574 TargetLowering::DAGCombinerInfo &DCI,
8575 SelectionDAG &DAG) {
8576 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8579 unsigned AddrOpIdx = N->getNumOperands() - 1;
8580 SDValue Addr = N->getOperand(AddrOpIdx);
8582 // Search for a use of the address operand that is an increment.
8583 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8584 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8586 if (User->getOpcode() != ISD::ADD ||
8587 UI.getUse().getResNo() != Addr.getResNo())
8590 // Check that the add is independent of the load/store. Otherwise, folding
8591 // it would create a cycle.
8592 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8595 // Find the new opcode for the updating load/store.
8596 bool IsStore = false;
8597 bool IsLaneOp = false;
8598 bool IsDupOp = false;
8599 unsigned NewOpc = 0;
8600 unsigned NumVecs = 0;
8601 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8603 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8604 case Intrinsic::aarch64_neon_ld2: NewOpc = AArch64ISD::LD2post;
8606 case Intrinsic::aarch64_neon_ld3: NewOpc = AArch64ISD::LD3post;
8608 case Intrinsic::aarch64_neon_ld4: NewOpc = AArch64ISD::LD4post;
8610 case Intrinsic::aarch64_neon_st2: NewOpc = AArch64ISD::ST2post;
8611 NumVecs = 2; IsStore = true; break;
8612 case Intrinsic::aarch64_neon_st3: NewOpc = AArch64ISD::ST3post;
8613 NumVecs = 3; IsStore = true; break;
8614 case Intrinsic::aarch64_neon_st4: NewOpc = AArch64ISD::ST4post;
8615 NumVecs = 4; IsStore = true; break;
8616 case Intrinsic::aarch64_neon_ld1x2: NewOpc = AArch64ISD::LD1x2post;
8618 case Intrinsic::aarch64_neon_ld1x3: NewOpc = AArch64ISD::LD1x3post;
8620 case Intrinsic::aarch64_neon_ld1x4: NewOpc = AArch64ISD::LD1x4post;
8622 case Intrinsic::aarch64_neon_st1x2: NewOpc = AArch64ISD::ST1x2post;
8623 NumVecs = 2; IsStore = true; break;
8624 case Intrinsic::aarch64_neon_st1x3: NewOpc = AArch64ISD::ST1x3post;
8625 NumVecs = 3; IsStore = true; break;
8626 case Intrinsic::aarch64_neon_st1x4: NewOpc = AArch64ISD::ST1x4post;
8627 NumVecs = 4; IsStore = true; break;
8628 case Intrinsic::aarch64_neon_ld2r: NewOpc = AArch64ISD::LD2DUPpost;
8629 NumVecs = 2; IsDupOp = true; break;
8630 case Intrinsic::aarch64_neon_ld3r: NewOpc = AArch64ISD::LD3DUPpost;
8631 NumVecs = 3; IsDupOp = true; break;
8632 case Intrinsic::aarch64_neon_ld4r: NewOpc = AArch64ISD::LD4DUPpost;
8633 NumVecs = 4; IsDupOp = true; break;
8634 case Intrinsic::aarch64_neon_ld2lane: NewOpc = AArch64ISD::LD2LANEpost;
8635 NumVecs = 2; IsLaneOp = true; break;
8636 case Intrinsic::aarch64_neon_ld3lane: NewOpc = AArch64ISD::LD3LANEpost;
8637 NumVecs = 3; IsLaneOp = true; break;
8638 case Intrinsic::aarch64_neon_ld4lane: NewOpc = AArch64ISD::LD4LANEpost;
8639 NumVecs = 4; IsLaneOp = true; break;
8640 case Intrinsic::aarch64_neon_st2lane: NewOpc = AArch64ISD::ST2LANEpost;
8641 NumVecs = 2; IsStore = true; IsLaneOp = true; break;
8642 case Intrinsic::aarch64_neon_st3lane: NewOpc = AArch64ISD::ST3LANEpost;
8643 NumVecs = 3; IsStore = true; IsLaneOp = true; break;
8644 case Intrinsic::aarch64_neon_st4lane: NewOpc = AArch64ISD::ST4LANEpost;
8645 NumVecs = 4; IsStore = true; IsLaneOp = true; break;
8650 VecTy = N->getOperand(2).getValueType();
8652 VecTy = N->getValueType(0);
8654 // If the increment is a constant, it must match the memory ref size.
8655 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8656 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8657 uint32_t IncVal = CInc->getZExtValue();
8658 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8659 if (IsLaneOp || IsDupOp)
8660 NumBytes /= VecTy.getVectorNumElements();
8661 if (IncVal != NumBytes)
8663 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
8665 SmallVector<SDValue, 8> Ops;
8666 Ops.push_back(N->getOperand(0)); // Incoming chain
8667 // Load lane and store have vector list as input.
8668 if (IsLaneOp || IsStore)
8669 for (unsigned i = 2; i < AddrOpIdx; ++i)
8670 Ops.push_back(N->getOperand(i));
8671 Ops.push_back(Addr); // Base register
8676 unsigned NumResultVecs = (IsStore ? 0 : NumVecs);
8678 for (n = 0; n < NumResultVecs; ++n)
8680 Tys[n++] = MVT::i64; // Type of write back register
8681 Tys[n] = MVT::Other; // Type of the chain
8682 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs + 2));
8684 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8685 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys, Ops,
8686 MemInt->getMemoryVT(),
8687 MemInt->getMemOperand());
8690 std::vector<SDValue> NewResults;
8691 for (unsigned i = 0; i < NumResultVecs; ++i) {
8692 NewResults.push_back(SDValue(UpdN.getNode(), i));
8694 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs + 1));
8695 DCI.CombineTo(N, NewResults);
8696 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8703 // Checks to see if the value is the prescribed width and returns information
8704 // about its extension mode.
8706 bool checkValueWidth(SDValue V, unsigned width, ISD::LoadExtType &ExtType) {
8707 ExtType = ISD::NON_EXTLOAD;
8708 switch(V.getNode()->getOpcode()) {
8712 LoadSDNode *LoadNode = cast<LoadSDNode>(V.getNode());
8713 if ((LoadNode->getMemoryVT() == MVT::i8 && width == 8)
8714 || (LoadNode->getMemoryVT() == MVT::i16 && width == 16)) {
8715 ExtType = LoadNode->getExtensionType();
8720 case ISD::AssertSext: {
8721 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
8722 if ((TypeNode->getVT() == MVT::i8 && width == 8)
8723 || (TypeNode->getVT() == MVT::i16 && width == 16)) {
8724 ExtType = ISD::SEXTLOAD;
8729 case ISD::AssertZext: {
8730 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
8731 if ((TypeNode->getVT() == MVT::i8 && width == 8)
8732 || (TypeNode->getVT() == MVT::i16 && width == 16)) {
8733 ExtType = ISD::ZEXTLOAD;
8739 case ISD::TargetConstant: {
8740 if (std::abs(cast<ConstantSDNode>(V.getNode())->getSExtValue()) <
8750 // This function does a whole lot of voodoo to determine if the tests are
8751 // equivalent without and with a mask. Essentially what happens is that given a
8754 // +-------------+ +-------------+ +-------------+ +-------------+
8755 // | Input | | AddConstant | | CompConstant| | CC |
8756 // +-------------+ +-------------+ +-------------+ +-------------+
8758 // V V | +----------+
8759 // +-------------+ +----+ | |
8760 // | ADD | |0xff| | |
8761 // +-------------+ +----+ | |
8764 // +-------------+ | |
8766 // +-------------+ | |
8775 // The AND node may be safely removed for some combinations of inputs. In
8776 // particular we need to take into account the extension type of the Input,
8777 // the exact values of AddConstant, CompConstant, and CC, along with the nominal
8778 // width of the input (this can work for any width inputs, the above graph is
8779 // specific to 8 bits.
8781 // The specific equations were worked out by generating output tables for each
8782 // AArch64CC value in terms of and AddConstant (w1), CompConstant(w2). The
8783 // problem was simplified by working with 4 bit inputs, which means we only
8784 // needed to reason about 24 distinct bit patterns: 8 patterns unique to zero
8785 // extension (8,15), 8 patterns unique to sign extensions (-8,-1), and 8
8786 // patterns present in both extensions (0,7). For every distinct set of
8787 // AddConstant and CompConstants bit patterns we can consider the masked and
8788 // unmasked versions to be equivalent if the result of this function is true for
8789 // all 16 distinct bit patterns of for the current extension type of Input (w0).
8792 // and w10, w8, #0x0f
8794 // cset w9, AArch64CC
8796 // cset w11, AArch64CC
8801 // Since the above function shows when the outputs are equivalent it defines
8802 // when it is safe to remove the AND. Unfortunately it only runs on AArch64 and
8803 // would be expensive to run during compiles. The equations below were written
8804 // in a test harness that confirmed they gave equivalent outputs to the above
8805 // for all inputs function, so they can be used determine if the removal is
8808 // isEquivalentMaskless() is the code for testing if the AND can be removed
8809 // factored out of the DAG recognition as the DAG can take several forms.
8812 bool isEquivalentMaskless(unsigned CC, unsigned width,
8813 ISD::LoadExtType ExtType, signed AddConstant,
8814 signed CompConstant) {
8815 // By being careful about our equations and only writing the in term
8816 // symbolic values and well known constants (0, 1, -1, MaxUInt) we can
8817 // make them generally applicable to all bit widths.
8818 signed MaxUInt = (1 << width);
8820 // For the purposes of these comparisons sign extending the type is
8821 // equivalent to zero extending the add and displacing it by half the integer
8822 // width. Provided we are careful and make sure our equations are valid over
8823 // the whole range we can just adjust the input and avoid writing equations
8824 // for sign extended inputs.
8825 if (ExtType == ISD::SEXTLOAD)
8826 AddConstant -= (1 << (width-1));
8830 case AArch64CC::GT: {
8831 if ((AddConstant == 0) ||
8832 (CompConstant == MaxUInt - 1 && AddConstant < 0) ||
8833 (AddConstant >= 0 && CompConstant < 0) ||
8834 (AddConstant <= 0 && CompConstant <= 0 && CompConstant < AddConstant))
8838 case AArch64CC::GE: {
8839 if ((AddConstant == 0) ||
8840 (AddConstant >= 0 && CompConstant <= 0) ||
8841 (AddConstant <= 0 && CompConstant <= 0 && CompConstant <= AddConstant))
8845 case AArch64CC::LS: {
8846 if ((AddConstant >= 0 && CompConstant < 0) ||
8847 (AddConstant <= 0 && CompConstant >= -1 &&
8848 CompConstant < AddConstant + MaxUInt))
8852 case AArch64CC::MI: {
8853 if ((AddConstant == 0) ||
8854 (AddConstant > 0 && CompConstant <= 0) ||
8855 (AddConstant < 0 && CompConstant <= AddConstant))
8859 case AArch64CC::HS: {
8860 if ((AddConstant >= 0 && CompConstant <= 0) ||
8861 (AddConstant <= 0 && CompConstant >= 0 &&
8862 CompConstant <= AddConstant + MaxUInt))
8866 case AArch64CC::NE: {
8867 if ((AddConstant > 0 && CompConstant < 0) ||
8868 (AddConstant < 0 && CompConstant >= 0 &&
8869 CompConstant < AddConstant + MaxUInt) ||
8870 (AddConstant >= 0 && CompConstant >= 0 &&
8871 CompConstant >= AddConstant) ||
8872 (AddConstant <= 0 && CompConstant < 0 && CompConstant < AddConstant))
8881 case AArch64CC::Invalid:
8889 SDValue performCONDCombine(SDNode *N,
8890 TargetLowering::DAGCombinerInfo &DCI,
8891 SelectionDAG &DAG, unsigned CCIndex,
8892 unsigned CmpIndex) {
8893 unsigned CC = cast<ConstantSDNode>(N->getOperand(CCIndex))->getSExtValue();
8894 SDNode *SubsNode = N->getOperand(CmpIndex).getNode();
8895 unsigned CondOpcode = SubsNode->getOpcode();
8897 if (CondOpcode != AArch64ISD::SUBS)
8900 // There is a SUBS feeding this condition. Is it fed by a mask we can
8903 SDNode *AndNode = SubsNode->getOperand(0).getNode();
8904 unsigned MaskBits = 0;
8906 if (AndNode->getOpcode() != ISD::AND)
8909 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(AndNode->getOperand(1))) {
8910 uint32_t CNV = CN->getZExtValue();
8913 else if (CNV == 65535)
8920 SDValue AddValue = AndNode->getOperand(0);
8922 if (AddValue.getOpcode() != ISD::ADD)
8925 // The basic dag structure is correct, grab the inputs and validate them.
8927 SDValue AddInputValue1 = AddValue.getNode()->getOperand(0);
8928 SDValue AddInputValue2 = AddValue.getNode()->getOperand(1);
8929 SDValue SubsInputValue = SubsNode->getOperand(1);
8931 // The mask is present and the provenance of all the values is a smaller type,
8932 // lets see if the mask is superfluous.
8934 if (!isa<ConstantSDNode>(AddInputValue2.getNode()) ||
8935 !isa<ConstantSDNode>(SubsInputValue.getNode()))
8938 ISD::LoadExtType ExtType;
8940 if (!checkValueWidth(SubsInputValue, MaskBits, ExtType) ||
8941 !checkValueWidth(AddInputValue2, MaskBits, ExtType) ||
8942 !checkValueWidth(AddInputValue1, MaskBits, ExtType) )
8945 if(!isEquivalentMaskless(CC, MaskBits, ExtType,
8946 cast<ConstantSDNode>(AddInputValue2.getNode())->getSExtValue(),
8947 cast<ConstantSDNode>(SubsInputValue.getNode())->getSExtValue()))
8950 // The AND is not necessary, remove it.
8952 SDVTList VTs = DAG.getVTList(SubsNode->getValueType(0),
8953 SubsNode->getValueType(1));
8954 SDValue Ops[] = { AddValue, SubsNode->getOperand(1) };
8956 SDValue NewValue = DAG.getNode(CondOpcode, SDLoc(SubsNode), VTs, Ops);
8957 DAG.ReplaceAllUsesWith(SubsNode, NewValue.getNode());
8959 return SDValue(N, 0);
8962 // Optimize compare with zero and branch.
8963 static SDValue performBRCONDCombine(SDNode *N,
8964 TargetLowering::DAGCombinerInfo &DCI,
8965 SelectionDAG &DAG) {
8966 SDValue NV = performCONDCombine(N, DCI, DAG, 2, 3);
8969 SDValue Chain = N->getOperand(0);
8970 SDValue Dest = N->getOperand(1);
8971 SDValue CCVal = N->getOperand(2);
8972 SDValue Cmp = N->getOperand(3);
8974 assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!");
8975 unsigned CC = cast<ConstantSDNode>(CCVal)->getZExtValue();
8976 if (CC != AArch64CC::EQ && CC != AArch64CC::NE)
8979 unsigned CmpOpc = Cmp.getOpcode();
8980 if (CmpOpc != AArch64ISD::ADDS && CmpOpc != AArch64ISD::SUBS)
8983 // Only attempt folding if there is only one use of the flag and no use of the
8985 if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1))
8988 SDValue LHS = Cmp.getOperand(0);
8989 SDValue RHS = Cmp.getOperand(1);
8991 assert(LHS.getValueType() == RHS.getValueType() &&
8992 "Expected the value type to be the same for both operands!");
8993 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
8996 if (isa<ConstantSDNode>(LHS) && cast<ConstantSDNode>(LHS)->isNullValue())
8997 std::swap(LHS, RHS);
8999 if (!isa<ConstantSDNode>(RHS) || !cast<ConstantSDNode>(RHS)->isNullValue())
9002 if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA ||
9003 LHS.getOpcode() == ISD::SRL)
9006 // Fold the compare into the branch instruction.
9008 if (CC == AArch64CC::EQ)
9009 BR = DAG.getNode(AArch64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
9011 BR = DAG.getNode(AArch64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
9013 // Do not add new nodes to DAG combiner worklist.
9014 DCI.CombineTo(N, BR, false);
9019 // vselect (v1i1 setcc) ->
9020 // vselect (v1iXX setcc) (XX is the size of the compared operand type)
9021 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
9022 // condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
9024 static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) {
9025 SDValue N0 = N->getOperand(0);
9026 EVT CCVT = N0.getValueType();
9028 if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
9029 CCVT.getVectorElementType() != MVT::i1)
9032 EVT ResVT = N->getValueType(0);
9033 EVT CmpVT = N0.getOperand(0).getValueType();
9034 // Only combine when the result type is of the same size as the compared
9036 if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
9039 SDValue IfTrue = N->getOperand(1);
9040 SDValue IfFalse = N->getOperand(2);
9042 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
9043 N0.getOperand(0), N0.getOperand(1),
9044 cast<CondCodeSDNode>(N0.getOperand(2))->get());
9045 return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
9049 /// A vector select: "(select vL, vR, (setcc LHS, RHS))" is best performed with
9050 /// the compare-mask instructions rather than going via NZCV, even if LHS and
9051 /// RHS are really scalar. This replaces any scalar setcc in the above pattern
9052 /// with a vector one followed by a DUP shuffle on the result.
9053 static SDValue performSelectCombine(SDNode *N,
9054 TargetLowering::DAGCombinerInfo &DCI) {
9055 SelectionDAG &DAG = DCI.DAG;
9056 SDValue N0 = N->getOperand(0);
9057 EVT ResVT = N->getValueType(0);
9059 if (N0.getOpcode() != ISD::SETCC)
9062 // Make sure the SETCC result is either i1 (initial DAG), or i32, the lowered
9063 // scalar SetCCResultType. We also don't expect vectors, because we assume
9064 // that selects fed by vector SETCCs are canonicalized to VSELECT.
9065 assert((N0.getValueType() == MVT::i1 || N0.getValueType() == MVT::i32) &&
9066 "Scalar-SETCC feeding SELECT has unexpected result type!");
9068 // If NumMaskElts == 0, the comparison is larger than select result. The
9069 // largest real NEON comparison is 64-bits per lane, which means the result is
9070 // at most 32-bits and an illegal vector. Just bail out for now.
9071 EVT SrcVT = N0.getOperand(0).getValueType();
9073 // Don't try to do this optimization when the setcc itself has i1 operands.
9074 // There are no legal vectors of i1, so this would be pointless.
9075 if (SrcVT == MVT::i1)
9078 int NumMaskElts = ResVT.getSizeInBits() / SrcVT.getSizeInBits();
9079 if (!ResVT.isVector() || NumMaskElts == 0)
9082 SrcVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumMaskElts);
9083 EVT CCVT = SrcVT.changeVectorElementTypeToInteger();
9085 // Also bail out if the vector CCVT isn't the same size as ResVT.
9086 // This can happen if the SETCC operand size doesn't divide the ResVT size
9087 // (e.g., f64 vs v3f32).
9088 if (CCVT.getSizeInBits() != ResVT.getSizeInBits())
9091 // Make sure we didn't create illegal types, if we're not supposed to.
9092 assert(DCI.isBeforeLegalize() ||
9093 DAG.getTargetLoweringInfo().isTypeLegal(SrcVT));
9095 // First perform a vector comparison, where lane 0 is the one we're interested
9099 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(0));
9101 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(1));
9102 SDValue SetCC = DAG.getNode(ISD::SETCC, DL, CCVT, LHS, RHS, N0.getOperand(2));
9104 // Now duplicate the comparison mask we want across all other lanes.
9105 SmallVector<int, 8> DUPMask(CCVT.getVectorNumElements(), 0);
9106 SDValue Mask = DAG.getVectorShuffle(CCVT, DL, SetCC, SetCC, DUPMask.data());
9107 Mask = DAG.getNode(ISD::BITCAST, DL,
9108 ResVT.changeVectorElementTypeToInteger(), Mask);
9110 return DAG.getSelect(DL, ResVT, Mask, N->getOperand(1), N->getOperand(2));
9113 /// performSelectCCCombine - Target-specific DAG combining for ISD::SELECT_CC
9114 /// to match FMIN/FMAX patterns.
9115 static SDValue performSelectCCCombine(SDNode *N, SelectionDAG &DAG) {
9116 // Try to use FMIN/FMAX instructions for FP selects like "x < y ? x : y".
9117 // Unless the NoNaNsFPMath option is set, be careful about NaNs:
9118 // vmax/vmin return NaN if either operand is a NaN;
9119 // only do the transformation when it matches that behavior.
9121 SDValue CondLHS = N->getOperand(0);
9122 SDValue CondRHS = N->getOperand(1);
9123 SDValue LHS = N->getOperand(2);
9124 SDValue RHS = N->getOperand(3);
9125 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9129 if (selectCCOpsAreFMaxCompatible(CondLHS, LHS) &&
9130 selectCCOpsAreFMaxCompatible(CondRHS, RHS)) {
9131 IsReversed = false; // x CC y ? x : y
9132 } else if (selectCCOpsAreFMaxCompatible(CondRHS, LHS) &&
9133 selectCCOpsAreFMaxCompatible(CondLHS, RHS)) {
9134 IsReversed = true ; // x CC y ? y : x
9139 bool IsUnordered = false, IsOrEqual;
9150 IsOrEqual = (CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE);
9151 Opcode = IsReversed ? AArch64ISD::FMAX : AArch64ISD::FMIN;
9161 IsOrEqual = (CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE);
9162 Opcode = IsReversed ? AArch64ISD::FMIN : AArch64ISD::FMAX;
9166 // If LHS is NaN, an ordered comparison will be false and the result will be
9167 // the RHS, but FMIN(NaN, RHS) = FMAX(NaN, RHS) = NaN. Avoid this by checking
9168 // that LHS != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9169 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9172 // For xxx-or-equal comparisons, "+0 <= -0" and "-0 >= +0" will both be true,
9173 // but FMIN will return -0, and FMAX will return +0. So FMIN/FMAX can only be
9174 // used for unsafe math or if one of the operands is known to be nonzero.
9175 if (IsOrEqual && !DAG.getTarget().Options.UnsafeFPMath &&
9176 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9179 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
9182 /// Get rid of unnecessary NVCASTs (that don't change the type).
9183 static SDValue performNVCASTCombine(SDNode *N) {
9184 if (N->getValueType(0) == N->getOperand(0).getValueType())
9185 return N->getOperand(0);
9190 SDValue AArch64TargetLowering::PerformDAGCombine(SDNode *N,
9191 DAGCombinerInfo &DCI) const {
9192 SelectionDAG &DAG = DCI.DAG;
9193 switch (N->getOpcode()) {
9198 return performAddSubLongCombine(N, DCI, DAG);
9200 return performXorCombine(N, DAG, DCI, Subtarget);
9202 return performMulCombine(N, DAG, DCI, Subtarget);
9203 case ISD::SINT_TO_FP:
9204 case ISD::UINT_TO_FP:
9205 return performIntToFpCombine(N, DAG, Subtarget);
9207 return performORCombine(N, DCI, Subtarget);
9208 case ISD::INTRINSIC_WO_CHAIN:
9209 return performIntrinsicCombine(N, DCI, Subtarget);
9210 case ISD::ANY_EXTEND:
9211 case ISD::ZERO_EXTEND:
9212 case ISD::SIGN_EXTEND:
9213 return performExtendCombine(N, DCI, DAG);
9215 return performBitcastCombine(N, DCI, DAG);
9216 case ISD::CONCAT_VECTORS:
9217 return performConcatVectorsCombine(N, DCI, DAG);
9219 return performSelectCombine(N, DCI);
9221 return performVSelectCombine(N, DCI.DAG);
9222 case ISD::SELECT_CC:
9223 return performSelectCCCombine(N, DCI.DAG);
9225 return performSTORECombine(N, DCI, DAG, Subtarget);
9226 case AArch64ISD::BRCOND:
9227 return performBRCONDCombine(N, DCI, DAG);
9228 case AArch64ISD::CSEL:
9229 return performCONDCombine(N, DCI, DAG, 2, 3);
9230 case AArch64ISD::DUP:
9231 return performPostLD1Combine(N, DCI, false);
9232 case AArch64ISD::NVCAST:
9233 return performNVCASTCombine(N);
9234 case ISD::INSERT_VECTOR_ELT:
9235 return performPostLD1Combine(N, DCI, true);
9236 case ISD::INTRINSIC_VOID:
9237 case ISD::INTRINSIC_W_CHAIN:
9238 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9239 case Intrinsic::aarch64_neon_ld2:
9240 case Intrinsic::aarch64_neon_ld3:
9241 case Intrinsic::aarch64_neon_ld4:
9242 case Intrinsic::aarch64_neon_ld1x2:
9243 case Intrinsic::aarch64_neon_ld1x3:
9244 case Intrinsic::aarch64_neon_ld1x4:
9245 case Intrinsic::aarch64_neon_ld2lane:
9246 case Intrinsic::aarch64_neon_ld3lane:
9247 case Intrinsic::aarch64_neon_ld4lane:
9248 case Intrinsic::aarch64_neon_ld2r:
9249 case Intrinsic::aarch64_neon_ld3r:
9250 case Intrinsic::aarch64_neon_ld4r:
9251 case Intrinsic::aarch64_neon_st2:
9252 case Intrinsic::aarch64_neon_st3:
9253 case Intrinsic::aarch64_neon_st4:
9254 case Intrinsic::aarch64_neon_st1x2:
9255 case Intrinsic::aarch64_neon_st1x3:
9256 case Intrinsic::aarch64_neon_st1x4:
9257 case Intrinsic::aarch64_neon_st2lane:
9258 case Intrinsic::aarch64_neon_st3lane:
9259 case Intrinsic::aarch64_neon_st4lane:
9260 return performNEONPostLDSTCombine(N, DCI, DAG);
9268 // Check if the return value is used as only a return value, as otherwise
9269 // we can't perform a tail-call. In particular, we need to check for
9270 // target ISD nodes that are returns and any other "odd" constructs
9271 // that the generic analysis code won't necessarily catch.
9272 bool AArch64TargetLowering::isUsedByReturnOnly(SDNode *N,
9273 SDValue &Chain) const {
9274 if (N->getNumValues() != 1)
9276 if (!N->hasNUsesOfValue(1, 0))
9279 SDValue TCChain = Chain;
9280 SDNode *Copy = *N->use_begin();
9281 if (Copy->getOpcode() == ISD::CopyToReg) {
9282 // If the copy has a glue operand, we conservatively assume it isn't safe to
9283 // perform a tail call.
9284 if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() ==
9287 TCChain = Copy->getOperand(0);
9288 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
9291 bool HasRet = false;
9292 for (SDNode *Node : Copy->uses()) {
9293 if (Node->getOpcode() != AArch64ISD::RET_FLAG)
9305 // Return whether the an instruction can potentially be optimized to a tail
9306 // call. This will cause the optimizers to attempt to move, or duplicate,
9307 // return instructions to help enable tail call optimizations for this
9309 bool AArch64TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
9310 if (!CI->isTailCall())
9316 bool AArch64TargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base,
9318 ISD::MemIndexedMode &AM,
9320 SelectionDAG &DAG) const {
9321 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
9324 Base = Op->getOperand(0);
9325 // All of the indexed addressing mode instructions take a signed
9326 // 9 bit immediate offset.
9327 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
9328 int64_t RHSC = (int64_t)RHS->getZExtValue();
9329 if (RHSC >= 256 || RHSC <= -256)
9331 IsInc = (Op->getOpcode() == ISD::ADD);
9332 Offset = Op->getOperand(1);
9338 bool AArch64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
9340 ISD::MemIndexedMode &AM,
9341 SelectionDAG &DAG) const {
9344 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9345 VT = LD->getMemoryVT();
9346 Ptr = LD->getBasePtr();
9347 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9348 VT = ST->getMemoryVT();
9349 Ptr = ST->getBasePtr();
9354 if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, IsInc, DAG))
9356 AM = IsInc ? ISD::PRE_INC : ISD::PRE_DEC;
9360 bool AArch64TargetLowering::getPostIndexedAddressParts(
9361 SDNode *N, SDNode *Op, SDValue &Base, SDValue &Offset,
9362 ISD::MemIndexedMode &AM, SelectionDAG &DAG) const {
9365 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9366 VT = LD->getMemoryVT();
9367 Ptr = LD->getBasePtr();
9368 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9369 VT = ST->getMemoryVT();
9370 Ptr = ST->getBasePtr();
9375 if (!getIndexedAddressParts(Op, Base, Offset, AM, IsInc, DAG))
9377 // Post-indexing updates the base, so it's not a valid transform
9378 // if that's not the same as the load's pointer.
9381 AM = IsInc ? ISD::POST_INC : ISD::POST_DEC;
9385 static void ReplaceBITCASTResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
9386 SelectionDAG &DAG) {
9388 SDValue Op = N->getOperand(0);
9390 if (N->getValueType(0) != MVT::i16 || Op.getValueType() != MVT::f16)
9394 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, DL, MVT::f32,
9395 DAG.getUNDEF(MVT::i32), Op,
9396 DAG.getTargetConstant(AArch64::hsub, DL, MVT::i32)),
9398 Op = DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op);
9399 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, Op));
9402 void AArch64TargetLowering::ReplaceNodeResults(
9403 SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
9404 switch (N->getOpcode()) {
9406 llvm_unreachable("Don't know how to custom expand this");
9408 ReplaceBITCASTResults(N, Results, DAG);
9410 case ISD::FP_TO_UINT:
9411 case ISD::FP_TO_SINT:
9412 assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion");
9413 // Let normal code take care of it by not adding anything to Results.
9418 bool AArch64TargetLowering::useLoadStackGuardNode() const {
9422 unsigned AArch64TargetLowering::combineRepeatedFPDivisors() const {
9423 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
9424 // reciprocal if there are three or more FDIVs.
9428 TargetLoweringBase::LegalizeTypeAction
9429 AArch64TargetLowering::getPreferredVectorAction(EVT VT) const {
9430 MVT SVT = VT.getSimpleVT();
9431 // During type legalization, we prefer to widen v1i8, v1i16, v1i32 to v8i8,
9432 // v4i16, v2i32 instead of to promote.
9433 if (SVT == MVT::v1i8 || SVT == MVT::v1i16 || SVT == MVT::v1i32
9434 || SVT == MVT::v1f32)
9435 return TypeWidenVector;
9437 return TargetLoweringBase::getPreferredVectorAction(VT);
9440 // Loads and stores less than 128-bits are already atomic; ones above that
9441 // are doomed anyway, so defer to the default libcall and blame the OS when
9443 bool AArch64TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
9444 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
9448 // Loads and stores less than 128-bits are already atomic; ones above that
9449 // are doomed anyway, so defer to the default libcall and blame the OS when
9451 bool AArch64TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
9452 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
9456 // For the real atomic operations, we have ldxr/stxr up to 128 bits,
9457 TargetLoweringBase::AtomicRMWExpansionKind
9458 AArch64TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
9459 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
9460 return Size <= 128 ? AtomicRMWExpansionKind::LLSC
9461 : AtomicRMWExpansionKind::None;
9464 bool AArch64TargetLowering::hasLoadLinkedStoreConditional() const {
9468 Value *AArch64TargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
9469 AtomicOrdering Ord) const {
9470 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
9471 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
9472 bool IsAcquire = isAtLeastAcquire(Ord);
9474 // Since i128 isn't legal and intrinsics don't get type-lowered, the ldrexd
9475 // intrinsic must return {i64, i64} and we have to recombine them into a
9476 // single i128 here.
9477 if (ValTy->getPrimitiveSizeInBits() == 128) {
9479 IsAcquire ? Intrinsic::aarch64_ldaxp : Intrinsic::aarch64_ldxp;
9480 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int);
9482 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
9483 Value *LoHi = Builder.CreateCall(Ldxr, Addr, "lohi");
9485 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
9486 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
9487 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
9488 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
9489 return Builder.CreateOr(
9490 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 64)), "val64");
9493 Type *Tys[] = { Addr->getType() };
9495 IsAcquire ? Intrinsic::aarch64_ldaxr : Intrinsic::aarch64_ldxr;
9496 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int, Tys);
9498 return Builder.CreateTruncOrBitCast(
9499 Builder.CreateCall(Ldxr, Addr),
9500 cast<PointerType>(Addr->getType())->getElementType());
9503 Value *AArch64TargetLowering::emitStoreConditional(IRBuilder<> &Builder,
9504 Value *Val, Value *Addr,
9505 AtomicOrdering Ord) const {
9506 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
9507 bool IsRelease = isAtLeastRelease(Ord);
9509 // Since the intrinsics must have legal type, the i128 intrinsics take two
9510 // parameters: "i64, i64". We must marshal Val into the appropriate form
9512 if (Val->getType()->getPrimitiveSizeInBits() == 128) {
9514 IsRelease ? Intrinsic::aarch64_stlxp : Intrinsic::aarch64_stxp;
9515 Function *Stxr = Intrinsic::getDeclaration(M, Int);
9516 Type *Int64Ty = Type::getInt64Ty(M->getContext());
9518 Value *Lo = Builder.CreateTrunc(Val, Int64Ty, "lo");
9519 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 64), Int64Ty, "hi");
9520 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
9521 return Builder.CreateCall(Stxr, {Lo, Hi, Addr});
9525 IsRelease ? Intrinsic::aarch64_stlxr : Intrinsic::aarch64_stxr;
9526 Type *Tys[] = { Addr->getType() };
9527 Function *Stxr = Intrinsic::getDeclaration(M, Int, Tys);
9529 return Builder.CreateCall(Stxr,
9530 {Builder.CreateZExtOrBitCast(
9531 Val, Stxr->getFunctionType()->getParamType(0)),
9535 bool AArch64TargetLowering::functionArgumentNeedsConsecutiveRegisters(
9536 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
9537 return Ty->isArrayTy();
9540 bool AArch64TargetLowering::shouldNormalizeToSelectSequence(LLVMContext &,