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);
223 for (MVT VT : MVT::vector_valuetypes()) {
224 setOperationAction(ISD::ROTL, VT, Expand);
225 setOperationAction(ISD::ROTR, VT, Expand);
228 // AArch64 doesn't have {U|S}MUL_LOHI.
229 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
230 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
233 // Expand the undefined-at-zero variants to cttz/ctlz to their defined-at-zero
234 // counterparts, which AArch64 supports directly.
235 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
236 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
237 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
238 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
240 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
241 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
243 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
244 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
245 setOperationAction(ISD::SREM, MVT::i32, Expand);
246 setOperationAction(ISD::SREM, MVT::i64, Expand);
247 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
248 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
249 setOperationAction(ISD::UREM, MVT::i32, Expand);
250 setOperationAction(ISD::UREM, MVT::i64, Expand);
252 // Custom lower Add/Sub/Mul with overflow.
253 setOperationAction(ISD::SADDO, MVT::i32, Custom);
254 setOperationAction(ISD::SADDO, MVT::i64, Custom);
255 setOperationAction(ISD::UADDO, MVT::i32, Custom);
256 setOperationAction(ISD::UADDO, MVT::i64, Custom);
257 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
258 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
259 setOperationAction(ISD::USUBO, MVT::i32, Custom);
260 setOperationAction(ISD::USUBO, MVT::i64, Custom);
261 setOperationAction(ISD::SMULO, MVT::i32, Custom);
262 setOperationAction(ISD::SMULO, MVT::i64, Custom);
263 setOperationAction(ISD::UMULO, MVT::i32, Custom);
264 setOperationAction(ISD::UMULO, MVT::i64, Custom);
266 setOperationAction(ISD::FSIN, MVT::f32, Expand);
267 setOperationAction(ISD::FSIN, MVT::f64, Expand);
268 setOperationAction(ISD::FCOS, MVT::f32, Expand);
269 setOperationAction(ISD::FCOS, MVT::f64, Expand);
270 setOperationAction(ISD::FPOW, MVT::f32, Expand);
271 setOperationAction(ISD::FPOW, MVT::f64, Expand);
272 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
273 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
275 // f16 is a storage-only type, always promote it to f32.
276 setOperationAction(ISD::SETCC, MVT::f16, Promote);
277 setOperationAction(ISD::BR_CC, MVT::f16, Promote);
278 setOperationAction(ISD::SELECT_CC, MVT::f16, Promote);
279 setOperationAction(ISD::SELECT, MVT::f16, Promote);
280 setOperationAction(ISD::FADD, MVT::f16, Promote);
281 setOperationAction(ISD::FSUB, MVT::f16, Promote);
282 setOperationAction(ISD::FMUL, MVT::f16, Promote);
283 setOperationAction(ISD::FDIV, MVT::f16, Promote);
284 setOperationAction(ISD::FREM, MVT::f16, Promote);
285 setOperationAction(ISD::FMA, MVT::f16, Promote);
286 setOperationAction(ISD::FNEG, MVT::f16, Promote);
287 setOperationAction(ISD::FABS, MVT::f16, Promote);
288 setOperationAction(ISD::FCEIL, MVT::f16, Promote);
289 setOperationAction(ISD::FCOPYSIGN, MVT::f16, Promote);
290 setOperationAction(ISD::FCOS, MVT::f16, Promote);
291 setOperationAction(ISD::FFLOOR, MVT::f16, Promote);
292 setOperationAction(ISD::FNEARBYINT, MVT::f16, Promote);
293 setOperationAction(ISD::FPOW, MVT::f16, Promote);
294 setOperationAction(ISD::FPOWI, MVT::f16, Promote);
295 setOperationAction(ISD::FRINT, MVT::f16, Promote);
296 setOperationAction(ISD::FSIN, MVT::f16, Promote);
297 setOperationAction(ISD::FSINCOS, MVT::f16, Promote);
298 setOperationAction(ISD::FSQRT, MVT::f16, Promote);
299 setOperationAction(ISD::FEXP, MVT::f16, Promote);
300 setOperationAction(ISD::FEXP2, MVT::f16, Promote);
301 setOperationAction(ISD::FLOG, MVT::f16, Promote);
302 setOperationAction(ISD::FLOG2, MVT::f16, Promote);
303 setOperationAction(ISD::FLOG10, MVT::f16, Promote);
304 setOperationAction(ISD::FROUND, MVT::f16, Promote);
305 setOperationAction(ISD::FTRUNC, MVT::f16, Promote);
306 setOperationAction(ISD::FMINNUM, MVT::f16, Promote);
307 setOperationAction(ISD::FMAXNUM, MVT::f16, Promote);
308 setOperationAction(ISD::FMINNAN, MVT::f16, Promote);
309 setOperationAction(ISD::FMAXNAN, MVT::f16, Promote);
311 // v4f16 is also a storage-only type, so promote it to v4f32 when that is
313 setOperationAction(ISD::FADD, MVT::v4f16, Promote);
314 setOperationAction(ISD::FSUB, MVT::v4f16, Promote);
315 setOperationAction(ISD::FMUL, MVT::v4f16, Promote);
316 setOperationAction(ISD::FDIV, MVT::v4f16, Promote);
317 setOperationAction(ISD::FP_EXTEND, MVT::v4f16, Promote);
318 setOperationAction(ISD::FP_ROUND, MVT::v4f16, Promote);
319 AddPromotedToType(ISD::FADD, MVT::v4f16, MVT::v4f32);
320 AddPromotedToType(ISD::FSUB, MVT::v4f16, MVT::v4f32);
321 AddPromotedToType(ISD::FMUL, MVT::v4f16, MVT::v4f32);
322 AddPromotedToType(ISD::FDIV, MVT::v4f16, MVT::v4f32);
323 AddPromotedToType(ISD::FP_EXTEND, MVT::v4f16, MVT::v4f32);
324 AddPromotedToType(ISD::FP_ROUND, MVT::v4f16, MVT::v4f32);
326 // Expand all other v4f16 operations.
327 // FIXME: We could generate better code by promoting some operations to
329 setOperationAction(ISD::FABS, MVT::v4f16, Expand);
330 setOperationAction(ISD::FCEIL, MVT::v4f16, Expand);
331 setOperationAction(ISD::FCOPYSIGN, MVT::v4f16, Expand);
332 setOperationAction(ISD::FCOS, MVT::v4f16, Expand);
333 setOperationAction(ISD::FFLOOR, MVT::v4f16, Expand);
334 setOperationAction(ISD::FMA, MVT::v4f16, Expand);
335 setOperationAction(ISD::FNEARBYINT, MVT::v4f16, Expand);
336 setOperationAction(ISD::FNEG, MVT::v4f16, Expand);
337 setOperationAction(ISD::FPOW, MVT::v4f16, Expand);
338 setOperationAction(ISD::FPOWI, MVT::v4f16, Expand);
339 setOperationAction(ISD::FREM, MVT::v4f16, Expand);
340 setOperationAction(ISD::FROUND, MVT::v4f16, Expand);
341 setOperationAction(ISD::FRINT, MVT::v4f16, Expand);
342 setOperationAction(ISD::FSIN, MVT::v4f16, Expand);
343 setOperationAction(ISD::FSINCOS, MVT::v4f16, Expand);
344 setOperationAction(ISD::FSQRT, MVT::v4f16, Expand);
345 setOperationAction(ISD::FTRUNC, MVT::v4f16, Expand);
346 setOperationAction(ISD::SETCC, MVT::v4f16, Expand);
347 setOperationAction(ISD::BR_CC, MVT::v4f16, Expand);
348 setOperationAction(ISD::SELECT, MVT::v4f16, Expand);
349 setOperationAction(ISD::SELECT_CC, MVT::v4f16, Expand);
350 setOperationAction(ISD::FEXP, MVT::v4f16, Expand);
351 setOperationAction(ISD::FEXP2, MVT::v4f16, Expand);
352 setOperationAction(ISD::FLOG, MVT::v4f16, Expand);
353 setOperationAction(ISD::FLOG2, MVT::v4f16, Expand);
354 setOperationAction(ISD::FLOG10, MVT::v4f16, Expand);
357 // v8f16 is also a storage-only type, so expand it.
358 setOperationAction(ISD::FABS, MVT::v8f16, Expand);
359 setOperationAction(ISD::FADD, MVT::v8f16, Expand);
360 setOperationAction(ISD::FCEIL, MVT::v8f16, Expand);
361 setOperationAction(ISD::FCOPYSIGN, MVT::v8f16, Expand);
362 setOperationAction(ISD::FCOS, MVT::v8f16, Expand);
363 setOperationAction(ISD::FDIV, MVT::v8f16, Expand);
364 setOperationAction(ISD::FFLOOR, MVT::v8f16, Expand);
365 setOperationAction(ISD::FMA, MVT::v8f16, Expand);
366 setOperationAction(ISD::FMUL, MVT::v8f16, Expand);
367 setOperationAction(ISD::FNEARBYINT, MVT::v8f16, Expand);
368 setOperationAction(ISD::FNEG, MVT::v8f16, Expand);
369 setOperationAction(ISD::FPOW, MVT::v8f16, Expand);
370 setOperationAction(ISD::FPOWI, MVT::v8f16, Expand);
371 setOperationAction(ISD::FREM, MVT::v8f16, Expand);
372 setOperationAction(ISD::FROUND, MVT::v8f16, Expand);
373 setOperationAction(ISD::FRINT, MVT::v8f16, Expand);
374 setOperationAction(ISD::FSIN, MVT::v8f16, Expand);
375 setOperationAction(ISD::FSINCOS, MVT::v8f16, Expand);
376 setOperationAction(ISD::FSQRT, MVT::v8f16, Expand);
377 setOperationAction(ISD::FSUB, MVT::v8f16, Expand);
378 setOperationAction(ISD::FTRUNC, MVT::v8f16, Expand);
379 setOperationAction(ISD::SETCC, MVT::v8f16, Expand);
380 setOperationAction(ISD::BR_CC, MVT::v8f16, Expand);
381 setOperationAction(ISD::SELECT, MVT::v8f16, Expand);
382 setOperationAction(ISD::SELECT_CC, MVT::v8f16, Expand);
383 setOperationAction(ISD::FP_EXTEND, MVT::v8f16, Expand);
384 setOperationAction(ISD::FEXP, MVT::v8f16, Expand);
385 setOperationAction(ISD::FEXP2, MVT::v8f16, Expand);
386 setOperationAction(ISD::FLOG, MVT::v8f16, Expand);
387 setOperationAction(ISD::FLOG2, MVT::v8f16, Expand);
388 setOperationAction(ISD::FLOG10, MVT::v8f16, Expand);
390 // AArch64 has implementations of a lot of rounding-like FP operations.
391 for (MVT Ty : {MVT::f32, MVT::f64}) {
392 setOperationAction(ISD::FFLOOR, Ty, Legal);
393 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
394 setOperationAction(ISD::FCEIL, Ty, Legal);
395 setOperationAction(ISD::FRINT, Ty, Legal);
396 setOperationAction(ISD::FTRUNC, Ty, Legal);
397 setOperationAction(ISD::FROUND, Ty, Legal);
398 setOperationAction(ISD::FMINNUM, Ty, Legal);
399 setOperationAction(ISD::FMAXNUM, Ty, Legal);
400 setOperationAction(ISD::FMINNAN, Ty, Legal);
401 setOperationAction(ISD::FMAXNAN, Ty, Legal);
404 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
406 // Lower READCYCLECOUNTER using an mrs from PMCCNTR_EL0.
407 // This requires the Performance Monitors extension.
408 if (Subtarget->hasPerfMon())
409 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Legal);
411 if (Subtarget->isTargetMachO()) {
412 // For iOS, we don't want to the normal expansion of a libcall to
413 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
415 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
416 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
418 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
419 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
422 // Make floating-point constants legal for the large code model, so they don't
423 // become loads from the constant pool.
424 if (Subtarget->isTargetMachO() && TM.getCodeModel() == CodeModel::Large) {
425 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
426 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
429 // AArch64 does not have floating-point extending loads, i1 sign-extending
430 // load, floating-point truncating stores, or v2i32->v2i16 truncating store.
431 for (MVT VT : MVT::fp_valuetypes()) {
432 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
433 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
434 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f64, Expand);
435 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f80, Expand);
437 for (MVT VT : MVT::integer_valuetypes())
438 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Expand);
440 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
441 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
442 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
443 setTruncStoreAction(MVT::f128, MVT::f80, Expand);
444 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
445 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
446 setTruncStoreAction(MVT::f128, MVT::f16, Expand);
448 setOperationAction(ISD::BITCAST, MVT::i16, Custom);
449 setOperationAction(ISD::BITCAST, MVT::f16, Custom);
451 // Indexed loads and stores are supported.
452 for (unsigned im = (unsigned)ISD::PRE_INC;
453 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
454 setIndexedLoadAction(im, MVT::i8, Legal);
455 setIndexedLoadAction(im, MVT::i16, Legal);
456 setIndexedLoadAction(im, MVT::i32, Legal);
457 setIndexedLoadAction(im, MVT::i64, Legal);
458 setIndexedLoadAction(im, MVT::f64, Legal);
459 setIndexedLoadAction(im, MVT::f32, Legal);
460 setIndexedLoadAction(im, MVT::f16, Legal);
461 setIndexedStoreAction(im, MVT::i8, Legal);
462 setIndexedStoreAction(im, MVT::i16, Legal);
463 setIndexedStoreAction(im, MVT::i32, Legal);
464 setIndexedStoreAction(im, MVT::i64, Legal);
465 setIndexedStoreAction(im, MVT::f64, Legal);
466 setIndexedStoreAction(im, MVT::f32, Legal);
467 setIndexedStoreAction(im, MVT::f16, Legal);
471 setOperationAction(ISD::TRAP, MVT::Other, Legal);
473 // We combine OR nodes for bitfield operations.
474 setTargetDAGCombine(ISD::OR);
476 // Vector add and sub nodes may conceal a high-half opportunity.
477 // Also, try to fold ADD into CSINC/CSINV..
478 setTargetDAGCombine(ISD::ADD);
479 setTargetDAGCombine(ISD::SUB);
481 setTargetDAGCombine(ISD::XOR);
482 setTargetDAGCombine(ISD::SINT_TO_FP);
483 setTargetDAGCombine(ISD::UINT_TO_FP);
485 setTargetDAGCombine(ISD::FP_TO_SINT);
486 setTargetDAGCombine(ISD::FP_TO_UINT);
487 setTargetDAGCombine(ISD::FDIV);
489 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
491 setTargetDAGCombine(ISD::ANY_EXTEND);
492 setTargetDAGCombine(ISD::ZERO_EXTEND);
493 setTargetDAGCombine(ISD::SIGN_EXTEND);
494 setTargetDAGCombine(ISD::BITCAST);
495 setTargetDAGCombine(ISD::CONCAT_VECTORS);
496 setTargetDAGCombine(ISD::STORE);
498 setTargetDAGCombine(ISD::MUL);
500 setTargetDAGCombine(ISD::SELECT);
501 setTargetDAGCombine(ISD::VSELECT);
503 setTargetDAGCombine(ISD::INTRINSIC_VOID);
504 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
505 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
506 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
508 MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8;
509 MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4;
510 MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 4;
512 setStackPointerRegisterToSaveRestore(AArch64::SP);
514 setSchedulingPreference(Sched::Hybrid);
517 MaskAndBranchFoldingIsLegal = true;
518 EnableExtLdPromotion = true;
520 setMinFunctionAlignment(2);
522 setHasExtractBitsInsn(true);
524 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
526 if (Subtarget->hasNEON()) {
527 // FIXME: v1f64 shouldn't be legal if we can avoid it, because it leads to
528 // silliness like this:
529 setOperationAction(ISD::FABS, MVT::v1f64, Expand);
530 setOperationAction(ISD::FADD, MVT::v1f64, Expand);
531 setOperationAction(ISD::FCEIL, MVT::v1f64, Expand);
532 setOperationAction(ISD::FCOPYSIGN, MVT::v1f64, Expand);
533 setOperationAction(ISD::FCOS, MVT::v1f64, Expand);
534 setOperationAction(ISD::FDIV, MVT::v1f64, Expand);
535 setOperationAction(ISD::FFLOOR, MVT::v1f64, Expand);
536 setOperationAction(ISD::FMA, MVT::v1f64, Expand);
537 setOperationAction(ISD::FMUL, MVT::v1f64, Expand);
538 setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Expand);
539 setOperationAction(ISD::FNEG, MVT::v1f64, Expand);
540 setOperationAction(ISD::FPOW, MVT::v1f64, Expand);
541 setOperationAction(ISD::FREM, MVT::v1f64, Expand);
542 setOperationAction(ISD::FROUND, MVT::v1f64, Expand);
543 setOperationAction(ISD::FRINT, MVT::v1f64, Expand);
544 setOperationAction(ISD::FSIN, MVT::v1f64, Expand);
545 setOperationAction(ISD::FSINCOS, MVT::v1f64, Expand);
546 setOperationAction(ISD::FSQRT, MVT::v1f64, Expand);
547 setOperationAction(ISD::FSUB, MVT::v1f64, Expand);
548 setOperationAction(ISD::FTRUNC, MVT::v1f64, Expand);
549 setOperationAction(ISD::SETCC, MVT::v1f64, Expand);
550 setOperationAction(ISD::BR_CC, MVT::v1f64, Expand);
551 setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
552 setOperationAction(ISD::SELECT_CC, MVT::v1f64, Expand);
553 setOperationAction(ISD::FP_EXTEND, MVT::v1f64, Expand);
555 setOperationAction(ISD::FP_TO_SINT, MVT::v1i64, Expand);
556 setOperationAction(ISD::FP_TO_UINT, MVT::v1i64, Expand);
557 setOperationAction(ISD::SINT_TO_FP, MVT::v1i64, Expand);
558 setOperationAction(ISD::UINT_TO_FP, MVT::v1i64, Expand);
559 setOperationAction(ISD::FP_ROUND, MVT::v1f64, Expand);
561 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
563 // AArch64 doesn't have a direct vector ->f32 conversion instructions for
564 // elements smaller than i32, so promote the input to i32 first.
565 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Promote);
566 setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Promote);
567 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Promote);
568 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Promote);
569 // i8 and i16 vector elements also need promotion to i32 for v8i8 or v8i16
570 // -> v8f16 conversions.
571 setOperationAction(ISD::SINT_TO_FP, MVT::v8i8, Promote);
572 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Promote);
573 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
574 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Promote);
575 // Similarly, there is no direct i32 -> f64 vector conversion instruction.
576 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
577 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
578 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
579 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
580 // Or, direct i32 -> f16 vector conversion. Set it so custom, so the
581 // conversion happens in two steps: v4i32 -> v4f32 -> v4f16
582 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Custom);
583 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
585 // AArch64 doesn't have MUL.2d:
586 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
587 // Custom handling for some quad-vector types to detect MULL.
588 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
589 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
590 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
592 setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal);
593 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
594 // Likewise, narrowing and extending vector loads/stores aren't handled
596 for (MVT VT : MVT::vector_valuetypes()) {
597 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
599 setOperationAction(ISD::MULHS, VT, Expand);
600 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
601 setOperationAction(ISD::MULHU, VT, Expand);
602 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
604 setOperationAction(ISD::BSWAP, VT, Expand);
606 for (MVT InnerVT : MVT::vector_valuetypes()) {
607 setTruncStoreAction(VT, InnerVT, Expand);
608 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
609 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
610 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
614 // AArch64 has implementations of a lot of rounding-like FP operations.
615 for (MVT Ty : {MVT::v2f32, MVT::v4f32, MVT::v2f64}) {
616 setOperationAction(ISD::FFLOOR, Ty, Legal);
617 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
618 setOperationAction(ISD::FCEIL, Ty, Legal);
619 setOperationAction(ISD::FRINT, Ty, Legal);
620 setOperationAction(ISD::FTRUNC, Ty, Legal);
621 setOperationAction(ISD::FROUND, Ty, Legal);
625 // Prefer likely predicted branches to selects on out-of-order cores.
626 if (Subtarget->isCortexA57())
627 PredictableSelectIsExpensive = true;
630 void AArch64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) {
631 if (VT == MVT::v2f32 || VT == MVT::v4f16) {
632 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
633 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i32);
635 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
636 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i32);
637 } else if (VT == MVT::v2f64 || VT == MVT::v4f32 || VT == MVT::v8f16) {
638 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
639 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i64);
641 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
642 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i64);
645 // Mark vector float intrinsics as expand.
646 if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) {
647 setOperationAction(ISD::FSIN, VT.getSimpleVT(), Expand);
648 setOperationAction(ISD::FCOS, VT.getSimpleVT(), Expand);
649 setOperationAction(ISD::FPOWI, VT.getSimpleVT(), Expand);
650 setOperationAction(ISD::FPOW, VT.getSimpleVT(), Expand);
651 setOperationAction(ISD::FLOG, VT.getSimpleVT(), Expand);
652 setOperationAction(ISD::FLOG2, VT.getSimpleVT(), Expand);
653 setOperationAction(ISD::FLOG10, VT.getSimpleVT(), Expand);
654 setOperationAction(ISD::FEXP, VT.getSimpleVT(), Expand);
655 setOperationAction(ISD::FEXP2, VT.getSimpleVT(), Expand);
657 // But we do support custom-lowering for FCOPYSIGN.
658 setOperationAction(ISD::FCOPYSIGN, VT.getSimpleVT(), Custom);
661 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
662 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
663 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
664 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
665 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Custom);
666 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
667 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
668 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
669 setOperationAction(ISD::AND, VT.getSimpleVT(), Custom);
670 setOperationAction(ISD::OR, VT.getSimpleVT(), Custom);
671 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
672 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
674 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
675 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
676 setOperationAction(ISD::VSELECT, VT.getSimpleVT(), Expand);
677 for (MVT InnerVT : MVT::all_valuetypes())
678 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT.getSimpleVT(), Expand);
680 // CNT supports only B element sizes.
681 if (VT != MVT::v8i8 && VT != MVT::v16i8)
682 setOperationAction(ISD::CTPOP, VT.getSimpleVT(), Expand);
684 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
685 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
686 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
687 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
688 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
690 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
691 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
693 // [SU][MIN|MAX] and [SU]ABSDIFF are available for all NEON types apart from
695 if (!VT.isFloatingPoint() &&
696 VT.getSimpleVT() != MVT::v2i64 && VT.getSimpleVT() != MVT::v1i64)
697 for (unsigned Opcode : {ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX,
698 ISD::SABSDIFF, ISD::UABSDIFF})
699 setOperationAction(Opcode, VT.getSimpleVT(), Legal);
701 // F[MIN|MAX][NUM|NAN] are available for all FP NEON types (not f16 though!).
702 if (VT.isFloatingPoint() && VT.getVectorElementType() != MVT::f16)
703 for (unsigned Opcode : {ISD::FMINNAN, ISD::FMAXNAN,
704 ISD::FMINNUM, ISD::FMAXNUM})
705 setOperationAction(Opcode, VT.getSimpleVT(), Legal);
707 if (Subtarget->isLittleEndian()) {
708 for (unsigned im = (unsigned)ISD::PRE_INC;
709 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
710 setIndexedLoadAction(im, VT.getSimpleVT(), Legal);
711 setIndexedStoreAction(im, VT.getSimpleVT(), Legal);
716 void AArch64TargetLowering::addDRTypeForNEON(MVT VT) {
717 addRegisterClass(VT, &AArch64::FPR64RegClass);
718 addTypeForNEON(VT, MVT::v2i32);
721 void AArch64TargetLowering::addQRTypeForNEON(MVT VT) {
722 addRegisterClass(VT, &AArch64::FPR128RegClass);
723 addTypeForNEON(VT, MVT::v4i32);
726 EVT AArch64TargetLowering::getSetCCResultType(const DataLayout &, LLVMContext &,
730 return VT.changeVectorElementTypeToInteger();
733 /// computeKnownBitsForTargetNode - Determine which of the bits specified in
734 /// Mask are known to be either zero or one and return them in the
735 /// KnownZero/KnownOne bitsets.
736 void AArch64TargetLowering::computeKnownBitsForTargetNode(
737 const SDValue Op, APInt &KnownZero, APInt &KnownOne,
738 const SelectionDAG &DAG, unsigned Depth) const {
739 switch (Op.getOpcode()) {
742 case AArch64ISD::CSEL: {
743 APInt KnownZero2, KnownOne2;
744 DAG.computeKnownBits(Op->getOperand(0), KnownZero, KnownOne, Depth + 1);
745 DAG.computeKnownBits(Op->getOperand(1), KnownZero2, KnownOne2, Depth + 1);
746 KnownZero &= KnownZero2;
747 KnownOne &= KnownOne2;
750 case ISD::INTRINSIC_W_CHAIN: {
751 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
752 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
755 case Intrinsic::aarch64_ldaxr:
756 case Intrinsic::aarch64_ldxr: {
757 unsigned BitWidth = KnownOne.getBitWidth();
758 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
759 unsigned MemBits = VT.getScalarType().getSizeInBits();
760 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
766 case ISD::INTRINSIC_WO_CHAIN:
767 case ISD::INTRINSIC_VOID: {
768 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
772 case Intrinsic::aarch64_neon_umaxv:
773 case Intrinsic::aarch64_neon_uminv: {
774 // Figure out the datatype of the vector operand. The UMINV instruction
775 // will zero extend the result, so we can mark as known zero all the
776 // bits larger than the element datatype. 32-bit or larget doesn't need
777 // this as those are legal types and will be handled by isel directly.
778 MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
779 unsigned BitWidth = KnownZero.getBitWidth();
780 if (VT == MVT::v8i8 || VT == MVT::v16i8) {
781 assert(BitWidth >= 8 && "Unexpected width!");
782 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8);
784 } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
785 assert(BitWidth >= 16 && "Unexpected width!");
786 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
796 MVT AArch64TargetLowering::getScalarShiftAmountTy(const DataLayout &DL,
801 bool AArch64TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
805 if (Subtarget->requiresStrictAlign())
808 // FIXME: This is mostly true for Cyclone, but not necessarily others.
810 // FIXME: Define an attribute for slow unaligned accesses instead of
811 // relying on the CPU type as a proxy.
812 // On Cyclone, unaligned 128-bit stores are slow.
813 *Fast = !Subtarget->isCyclone() || VT.getStoreSize() != 16 ||
814 // See comments in performSTORECombine() for more details about
817 // Code that uses clang vector extensions can mark that it
818 // wants unaligned accesses to be treated as fast by
819 // underspecifying alignment to be 1 or 2.
822 // Disregard v2i64. Memcpy lowering produces those and splitting
823 // them regresses performance on micro-benchmarks and olden/bh.
830 AArch64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
831 const TargetLibraryInfo *libInfo) const {
832 return AArch64::createFastISel(funcInfo, libInfo);
835 const char *AArch64TargetLowering::getTargetNodeName(unsigned Opcode) const {
836 switch ((AArch64ISD::NodeType)Opcode) {
837 case AArch64ISD::FIRST_NUMBER: break;
838 case AArch64ISD::CALL: return "AArch64ISD::CALL";
839 case AArch64ISD::ADRP: return "AArch64ISD::ADRP";
840 case AArch64ISD::ADDlow: return "AArch64ISD::ADDlow";
841 case AArch64ISD::LOADgot: return "AArch64ISD::LOADgot";
842 case AArch64ISD::RET_FLAG: return "AArch64ISD::RET_FLAG";
843 case AArch64ISD::BRCOND: return "AArch64ISD::BRCOND";
844 case AArch64ISD::CSEL: return "AArch64ISD::CSEL";
845 case AArch64ISD::FCSEL: return "AArch64ISD::FCSEL";
846 case AArch64ISD::CSINV: return "AArch64ISD::CSINV";
847 case AArch64ISD::CSNEG: return "AArch64ISD::CSNEG";
848 case AArch64ISD::CSINC: return "AArch64ISD::CSINC";
849 case AArch64ISD::THREAD_POINTER: return "AArch64ISD::THREAD_POINTER";
850 case AArch64ISD::TLSDESC_CALLSEQ: return "AArch64ISD::TLSDESC_CALLSEQ";
851 case AArch64ISD::ADC: return "AArch64ISD::ADC";
852 case AArch64ISD::SBC: return "AArch64ISD::SBC";
853 case AArch64ISD::ADDS: return "AArch64ISD::ADDS";
854 case AArch64ISD::SUBS: return "AArch64ISD::SUBS";
855 case AArch64ISD::ADCS: return "AArch64ISD::ADCS";
856 case AArch64ISD::SBCS: return "AArch64ISD::SBCS";
857 case AArch64ISD::ANDS: return "AArch64ISD::ANDS";
858 case AArch64ISD::CCMP: return "AArch64ISD::CCMP";
859 case AArch64ISD::CCMN: return "AArch64ISD::CCMN";
860 case AArch64ISD::FCCMP: return "AArch64ISD::FCCMP";
861 case AArch64ISD::FCMP: return "AArch64ISD::FCMP";
862 case AArch64ISD::DUP: return "AArch64ISD::DUP";
863 case AArch64ISD::DUPLANE8: return "AArch64ISD::DUPLANE8";
864 case AArch64ISD::DUPLANE16: return "AArch64ISD::DUPLANE16";
865 case AArch64ISD::DUPLANE32: return "AArch64ISD::DUPLANE32";
866 case AArch64ISD::DUPLANE64: return "AArch64ISD::DUPLANE64";
867 case AArch64ISD::MOVI: return "AArch64ISD::MOVI";
868 case AArch64ISD::MOVIshift: return "AArch64ISD::MOVIshift";
869 case AArch64ISD::MOVIedit: return "AArch64ISD::MOVIedit";
870 case AArch64ISD::MOVImsl: return "AArch64ISD::MOVImsl";
871 case AArch64ISD::FMOV: return "AArch64ISD::FMOV";
872 case AArch64ISD::MVNIshift: return "AArch64ISD::MVNIshift";
873 case AArch64ISD::MVNImsl: return "AArch64ISD::MVNImsl";
874 case AArch64ISD::BICi: return "AArch64ISD::BICi";
875 case AArch64ISD::ORRi: return "AArch64ISD::ORRi";
876 case AArch64ISD::BSL: return "AArch64ISD::BSL";
877 case AArch64ISD::NEG: return "AArch64ISD::NEG";
878 case AArch64ISD::EXTR: return "AArch64ISD::EXTR";
879 case AArch64ISD::ZIP1: return "AArch64ISD::ZIP1";
880 case AArch64ISD::ZIP2: return "AArch64ISD::ZIP2";
881 case AArch64ISD::UZP1: return "AArch64ISD::UZP1";
882 case AArch64ISD::UZP2: return "AArch64ISD::UZP2";
883 case AArch64ISD::TRN1: return "AArch64ISD::TRN1";
884 case AArch64ISD::TRN2: return "AArch64ISD::TRN2";
885 case AArch64ISD::REV16: return "AArch64ISD::REV16";
886 case AArch64ISD::REV32: return "AArch64ISD::REV32";
887 case AArch64ISD::REV64: return "AArch64ISD::REV64";
888 case AArch64ISD::EXT: return "AArch64ISD::EXT";
889 case AArch64ISD::VSHL: return "AArch64ISD::VSHL";
890 case AArch64ISD::VLSHR: return "AArch64ISD::VLSHR";
891 case AArch64ISD::VASHR: return "AArch64ISD::VASHR";
892 case AArch64ISD::CMEQ: return "AArch64ISD::CMEQ";
893 case AArch64ISD::CMGE: return "AArch64ISD::CMGE";
894 case AArch64ISD::CMGT: return "AArch64ISD::CMGT";
895 case AArch64ISD::CMHI: return "AArch64ISD::CMHI";
896 case AArch64ISD::CMHS: return "AArch64ISD::CMHS";
897 case AArch64ISD::FCMEQ: return "AArch64ISD::FCMEQ";
898 case AArch64ISD::FCMGE: return "AArch64ISD::FCMGE";
899 case AArch64ISD::FCMGT: return "AArch64ISD::FCMGT";
900 case AArch64ISD::CMEQz: return "AArch64ISD::CMEQz";
901 case AArch64ISD::CMGEz: return "AArch64ISD::CMGEz";
902 case AArch64ISD::CMGTz: return "AArch64ISD::CMGTz";
903 case AArch64ISD::CMLEz: return "AArch64ISD::CMLEz";
904 case AArch64ISD::CMLTz: return "AArch64ISD::CMLTz";
905 case AArch64ISD::FCMEQz: return "AArch64ISD::FCMEQz";
906 case AArch64ISD::FCMGEz: return "AArch64ISD::FCMGEz";
907 case AArch64ISD::FCMGTz: return "AArch64ISD::FCMGTz";
908 case AArch64ISD::FCMLEz: return "AArch64ISD::FCMLEz";
909 case AArch64ISD::FCMLTz: return "AArch64ISD::FCMLTz";
910 case AArch64ISD::SADDV: return "AArch64ISD::SADDV";
911 case AArch64ISD::UADDV: return "AArch64ISD::UADDV";
912 case AArch64ISD::SMINV: return "AArch64ISD::SMINV";
913 case AArch64ISD::UMINV: return "AArch64ISD::UMINV";
914 case AArch64ISD::SMAXV: return "AArch64ISD::SMAXV";
915 case AArch64ISD::UMAXV: return "AArch64ISD::UMAXV";
916 case AArch64ISD::NOT: return "AArch64ISD::NOT";
917 case AArch64ISD::BIT: return "AArch64ISD::BIT";
918 case AArch64ISD::CBZ: return "AArch64ISD::CBZ";
919 case AArch64ISD::CBNZ: return "AArch64ISD::CBNZ";
920 case AArch64ISD::TBZ: return "AArch64ISD::TBZ";
921 case AArch64ISD::TBNZ: return "AArch64ISD::TBNZ";
922 case AArch64ISD::TC_RETURN: return "AArch64ISD::TC_RETURN";
923 case AArch64ISD::PREFETCH: return "AArch64ISD::PREFETCH";
924 case AArch64ISD::SITOF: return "AArch64ISD::SITOF";
925 case AArch64ISD::UITOF: return "AArch64ISD::UITOF";
926 case AArch64ISD::NVCAST: return "AArch64ISD::NVCAST";
927 case AArch64ISD::SQSHL_I: return "AArch64ISD::SQSHL_I";
928 case AArch64ISD::UQSHL_I: return "AArch64ISD::UQSHL_I";
929 case AArch64ISD::SRSHR_I: return "AArch64ISD::SRSHR_I";
930 case AArch64ISD::URSHR_I: return "AArch64ISD::URSHR_I";
931 case AArch64ISD::SQSHLU_I: return "AArch64ISD::SQSHLU_I";
932 case AArch64ISD::WrapperLarge: return "AArch64ISD::WrapperLarge";
933 case AArch64ISD::LD2post: return "AArch64ISD::LD2post";
934 case AArch64ISD::LD3post: return "AArch64ISD::LD3post";
935 case AArch64ISD::LD4post: return "AArch64ISD::LD4post";
936 case AArch64ISD::ST2post: return "AArch64ISD::ST2post";
937 case AArch64ISD::ST3post: return "AArch64ISD::ST3post";
938 case AArch64ISD::ST4post: return "AArch64ISD::ST4post";
939 case AArch64ISD::LD1x2post: return "AArch64ISD::LD1x2post";
940 case AArch64ISD::LD1x3post: return "AArch64ISD::LD1x3post";
941 case AArch64ISD::LD1x4post: return "AArch64ISD::LD1x4post";
942 case AArch64ISD::ST1x2post: return "AArch64ISD::ST1x2post";
943 case AArch64ISD::ST1x3post: return "AArch64ISD::ST1x3post";
944 case AArch64ISD::ST1x4post: return "AArch64ISD::ST1x4post";
945 case AArch64ISD::LD1DUPpost: return "AArch64ISD::LD1DUPpost";
946 case AArch64ISD::LD2DUPpost: return "AArch64ISD::LD2DUPpost";
947 case AArch64ISD::LD3DUPpost: return "AArch64ISD::LD3DUPpost";
948 case AArch64ISD::LD4DUPpost: return "AArch64ISD::LD4DUPpost";
949 case AArch64ISD::LD1LANEpost: return "AArch64ISD::LD1LANEpost";
950 case AArch64ISD::LD2LANEpost: return "AArch64ISD::LD2LANEpost";
951 case AArch64ISD::LD3LANEpost: return "AArch64ISD::LD3LANEpost";
952 case AArch64ISD::LD4LANEpost: return "AArch64ISD::LD4LANEpost";
953 case AArch64ISD::ST2LANEpost: return "AArch64ISD::ST2LANEpost";
954 case AArch64ISD::ST3LANEpost: return "AArch64ISD::ST3LANEpost";
955 case AArch64ISD::ST4LANEpost: return "AArch64ISD::ST4LANEpost";
956 case AArch64ISD::SMULL: return "AArch64ISD::SMULL";
957 case AArch64ISD::UMULL: return "AArch64ISD::UMULL";
963 AArch64TargetLowering::EmitF128CSEL(MachineInstr *MI,
964 MachineBasicBlock *MBB) const {
965 // We materialise the F128CSEL pseudo-instruction as some control flow and a
969 // [... previous instrs leading to comparison ...]
975 // Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB]
977 MachineFunction *MF = MBB->getParent();
978 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
979 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
980 DebugLoc DL = MI->getDebugLoc();
981 MachineFunction::iterator It = ++MBB->getIterator();
983 unsigned DestReg = MI->getOperand(0).getReg();
984 unsigned IfTrueReg = MI->getOperand(1).getReg();
985 unsigned IfFalseReg = MI->getOperand(2).getReg();
986 unsigned CondCode = MI->getOperand(3).getImm();
987 bool NZCVKilled = MI->getOperand(4).isKill();
989 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
990 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
991 MF->insert(It, TrueBB);
992 MF->insert(It, EndBB);
994 // Transfer rest of current basic-block to EndBB
995 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
997 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
999 BuildMI(MBB, DL, TII->get(AArch64::Bcc)).addImm(CondCode).addMBB(TrueBB);
1000 BuildMI(MBB, DL, TII->get(AArch64::B)).addMBB(EndBB);
1001 MBB->addSuccessor(TrueBB);
1002 MBB->addSuccessor(EndBB);
1004 // TrueBB falls through to the end.
1005 TrueBB->addSuccessor(EndBB);
1008 TrueBB->addLiveIn(AArch64::NZCV);
1009 EndBB->addLiveIn(AArch64::NZCV);
1012 BuildMI(*EndBB, EndBB->begin(), DL, TII->get(AArch64::PHI), DestReg)
1018 MI->eraseFromParent();
1023 AArch64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
1024 MachineBasicBlock *BB) const {
1025 switch (MI->getOpcode()) {
1030 llvm_unreachable("Unexpected instruction for custom inserter!");
1032 case AArch64::F128CSEL:
1033 return EmitF128CSEL(MI, BB);
1035 case TargetOpcode::STACKMAP:
1036 case TargetOpcode::PATCHPOINT:
1037 return emitPatchPoint(MI, BB);
1041 //===----------------------------------------------------------------------===//
1042 // AArch64 Lowering private implementation.
1043 //===----------------------------------------------------------------------===//
1045 //===----------------------------------------------------------------------===//
1047 //===----------------------------------------------------------------------===//
1049 /// changeIntCCToAArch64CC - Convert a DAG integer condition code to an AArch64
1051 static AArch64CC::CondCode changeIntCCToAArch64CC(ISD::CondCode CC) {
1054 llvm_unreachable("Unknown condition code!");
1056 return AArch64CC::NE;
1058 return AArch64CC::EQ;
1060 return AArch64CC::GT;
1062 return AArch64CC::GE;
1064 return AArch64CC::LT;
1066 return AArch64CC::LE;
1068 return AArch64CC::HI;
1070 return AArch64CC::HS;
1072 return AArch64CC::LO;
1074 return AArch64CC::LS;
1078 /// changeFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64 CC.
1079 static void changeFPCCToAArch64CC(ISD::CondCode CC,
1080 AArch64CC::CondCode &CondCode,
1081 AArch64CC::CondCode &CondCode2) {
1082 CondCode2 = AArch64CC::AL;
1085 llvm_unreachable("Unknown FP condition!");
1088 CondCode = AArch64CC::EQ;
1092 CondCode = AArch64CC::GT;
1096 CondCode = AArch64CC::GE;
1099 CondCode = AArch64CC::MI;
1102 CondCode = AArch64CC::LS;
1105 CondCode = AArch64CC::MI;
1106 CondCode2 = AArch64CC::GT;
1109 CondCode = AArch64CC::VC;
1112 CondCode = AArch64CC::VS;
1115 CondCode = AArch64CC::EQ;
1116 CondCode2 = AArch64CC::VS;
1119 CondCode = AArch64CC::HI;
1122 CondCode = AArch64CC::PL;
1126 CondCode = AArch64CC::LT;
1130 CondCode = AArch64CC::LE;
1134 CondCode = AArch64CC::NE;
1139 /// changeVectorFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64
1140 /// CC usable with the vector instructions. Fewer operations are available
1141 /// without a real NZCV register, so we have to use less efficient combinations
1142 /// to get the same effect.
1143 static void changeVectorFPCCToAArch64CC(ISD::CondCode CC,
1144 AArch64CC::CondCode &CondCode,
1145 AArch64CC::CondCode &CondCode2,
1150 // Mostly the scalar mappings work fine.
1151 changeFPCCToAArch64CC(CC, CondCode, CondCode2);
1154 Invert = true; // Fallthrough
1156 CondCode = AArch64CC::MI;
1157 CondCode2 = AArch64CC::GE;
1164 // All of the compare-mask comparisons are ordered, but we can switch
1165 // between the two by a double inversion. E.g. ULE == !OGT.
1167 changeFPCCToAArch64CC(getSetCCInverse(CC, false), CondCode, CondCode2);
1172 static bool isLegalArithImmed(uint64_t C) {
1173 // Matches AArch64DAGToDAGISel::SelectArithImmed().
1174 return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
1177 static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1178 SDLoc dl, SelectionDAG &DAG) {
1179 EVT VT = LHS.getValueType();
1181 if (VT.isFloatingPoint())
1182 return DAG.getNode(AArch64ISD::FCMP, dl, VT, LHS, RHS);
1184 // The CMP instruction is just an alias for SUBS, and representing it as
1185 // SUBS means that it's possible to get CSE with subtract operations.
1186 // A later phase can perform the optimization of setting the destination
1187 // register to WZR/XZR if it ends up being unused.
1188 unsigned Opcode = AArch64ISD::SUBS;
1190 if (RHS.getOpcode() == ISD::SUB && isa<ConstantSDNode>(RHS.getOperand(0)) &&
1191 cast<ConstantSDNode>(RHS.getOperand(0))->getZExtValue() == 0 &&
1192 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1193 // We'd like to combine a (CMP op1, (sub 0, op2) into a CMN instruction on
1194 // the grounds that "op1 - (-op2) == op1 + op2". However, the C and V flags
1195 // can be set differently by this operation. It comes down to whether
1196 // "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are then
1197 // everything is fine. If not then the optimization is wrong. Thus general
1198 // comparisons are only valid if op2 != 0.
1200 // So, finally, the only LLVM-native comparisons that don't mention C and V
1201 // are SETEQ and SETNE. They're the only ones we can safely use CMN for in
1202 // the absence of information about op2.
1203 Opcode = AArch64ISD::ADDS;
1204 RHS = RHS.getOperand(1);
1205 } else if (LHS.getOpcode() == ISD::AND && isa<ConstantSDNode>(RHS) &&
1206 cast<ConstantSDNode>(RHS)->getZExtValue() == 0 &&
1207 !isUnsignedIntSetCC(CC)) {
1208 // Similarly, (CMP (and X, Y), 0) can be implemented with a TST
1209 // (a.k.a. ANDS) except that the flags are only guaranteed to work for one
1210 // of the signed comparisons.
1211 Opcode = AArch64ISD::ANDS;
1212 RHS = LHS.getOperand(1);
1213 LHS = LHS.getOperand(0);
1216 return DAG.getNode(Opcode, dl, DAG.getVTList(VT, MVT_CC), LHS, RHS)
1220 /// \defgroup AArch64CCMP CMP;CCMP matching
1222 /// These functions deal with the formation of CMP;CCMP;... sequences.
1223 /// The CCMP/CCMN/FCCMP/FCCMPE instructions allow the conditional execution of
1224 /// a comparison. They set the NZCV flags to a predefined value if their
1225 /// predicate is false. This allows to express arbitrary conjunctions, for
1226 /// example "cmp 0 (and (setCA (cmp A)) (setCB (cmp B))))"
1229 /// ccmp B, inv(CB), CA
1230 /// check for CB flags
1232 /// In general we can create code for arbitrary "... (and (and A B) C)"
1233 /// sequences. We can also implement some "or" expressions, because "(or A B)"
1234 /// is equivalent to "not (and (not A) (not B))" and we can implement some
1235 /// negation operations:
1236 /// We can negate the results of a single comparison by inverting the flags
1237 /// used when the predicate fails and inverting the flags tested in the next
1238 /// instruction; We can also negate the results of the whole previous
1239 /// conditional compare sequence by inverting the flags tested in the next
1240 /// instruction. However there is no way to negate the result of a partial
1243 /// Therefore on encountering an "or" expression we can negate the subtree on
1244 /// one side and have to be able to push the negate to the leafs of the subtree
1245 /// on the other side (see also the comments in code). As complete example:
1246 /// "or (or (setCA (cmp A)) (setCB (cmp B)))
1247 /// (and (setCC (cmp C)) (setCD (cmp D)))"
1248 /// is transformed to
1249 /// "not (and (not (and (setCC (cmp C)) (setCC (cmp D))))
1250 /// (and (not (setCA (cmp A)) (not (setCB (cmp B))))))"
1251 /// and implemented as:
1253 /// ccmp D, inv(CD), CC
1254 /// ccmp A, CA, inv(CD)
1255 /// ccmp B, CB, inv(CA)
1256 /// check for CB flags
1257 /// A counterexample is "or (and A B) (and C D)" which cannot be implemented
1258 /// by conditional compare sequences.
1261 /// Create a conditional comparison; Use CCMP, CCMN or FCCMP as appropriate.
1262 static SDValue emitConditionalComparison(SDValue LHS, SDValue RHS,
1263 ISD::CondCode CC, SDValue CCOp,
1264 SDValue Condition, unsigned NZCV,
1265 SDLoc DL, SelectionDAG &DAG) {
1266 unsigned Opcode = 0;
1267 if (LHS.getValueType().isFloatingPoint())
1268 Opcode = AArch64ISD::FCCMP;
1269 else if (RHS.getOpcode() == ISD::SUB) {
1270 SDValue SubOp0 = RHS.getOperand(0);
1271 if (const ConstantSDNode *SubOp0C = dyn_cast<ConstantSDNode>(SubOp0))
1272 if (SubOp0C->isNullValue() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1273 // See emitComparison() on why we can only do this for SETEQ and SETNE.
1274 Opcode = AArch64ISD::CCMN;
1275 RHS = RHS.getOperand(1);
1279 Opcode = AArch64ISD::CCMP;
1281 SDValue NZCVOp = DAG.getConstant(NZCV, DL, MVT::i32);
1282 return DAG.getNode(Opcode, DL, MVT_CC, LHS, RHS, NZCVOp, Condition, CCOp);
1285 /// Returns true if @p Val is a tree of AND/OR/SETCC operations.
1286 /// CanPushNegate is set to true if we can push a negate operation through
1287 /// the tree in a was that we are left with AND operations and negate operations
1288 /// at the leafs only. i.e. "not (or (or x y) z)" can be changed to
1289 /// "and (and (not x) (not y)) (not z)"; "not (or (and x y) z)" cannot be
1290 /// brought into such a form.
1291 static bool isConjunctionDisjunctionTree(const SDValue Val, bool &CanPushNegate,
1292 unsigned Depth = 0) {
1293 if (!Val.hasOneUse())
1295 unsigned Opcode = Val->getOpcode();
1296 if (Opcode == ISD::SETCC) {
1297 CanPushNegate = true;
1300 // Protect against stack overflow.
1303 if (Opcode == ISD::AND || Opcode == ISD::OR) {
1304 SDValue O0 = Val->getOperand(0);
1305 SDValue O1 = Val->getOperand(1);
1306 bool CanPushNegateL;
1307 if (!isConjunctionDisjunctionTree(O0, CanPushNegateL, Depth+1))
1309 bool CanPushNegateR;
1310 if (!isConjunctionDisjunctionTree(O1, CanPushNegateR, Depth+1))
1312 // We cannot push a negate through an AND operation (it would become an OR),
1313 // we can however change a (not (or x y)) to (and (not x) (not y)) if we can
1314 // push the negate through the x/y subtrees.
1315 CanPushNegate = (Opcode == ISD::OR) && CanPushNegateL && CanPushNegateR;
1321 /// Emit conjunction or disjunction tree with the CMP/FCMP followed by a chain
1322 /// of CCMP/CFCMP ops. See @ref AArch64CCMP.
1323 /// Tries to transform the given i1 producing node @p Val to a series compare
1324 /// and conditional compare operations. @returns an NZCV flags producing node
1325 /// and sets @p OutCC to the flags that should be tested or returns SDValue() if
1326 /// transformation was not possible.
1327 /// On recursive invocations @p PushNegate may be set to true to have negation
1328 /// effects pushed to the tree leafs; @p Predicate is an NZCV flag predicate
1329 /// for the comparisons in the current subtree; @p Depth limits the search
1330 /// depth to avoid stack overflow.
1331 static SDValue emitConjunctionDisjunctionTree(SelectionDAG &DAG, SDValue Val,
1332 AArch64CC::CondCode &OutCC, bool PushNegate = false,
1333 SDValue CCOp = SDValue(), AArch64CC::CondCode Predicate = AArch64CC::AL,
1334 unsigned Depth = 0) {
1335 // We're at a tree leaf, produce a conditional comparison operation.
1336 unsigned Opcode = Val->getOpcode();
1337 if (Opcode == ISD::SETCC) {
1338 SDValue LHS = Val->getOperand(0);
1339 SDValue RHS = Val->getOperand(1);
1340 ISD::CondCode CC = cast<CondCodeSDNode>(Val->getOperand(2))->get();
1341 bool isInteger = LHS.getValueType().isInteger();
1343 CC = getSetCCInverse(CC, isInteger);
1345 // Determine OutCC and handle FP special case.
1347 OutCC = changeIntCCToAArch64CC(CC);
1349 assert(LHS.getValueType().isFloatingPoint());
1350 AArch64CC::CondCode ExtraCC;
1351 changeFPCCToAArch64CC(CC, OutCC, ExtraCC);
1352 // Surpisingly some floating point conditions can't be tested with a
1353 // single condition code. Construct an additional comparison in this case.
1354 // See comment below on how we deal with OR conditions.
1355 if (ExtraCC != AArch64CC::AL) {
1357 if (!CCOp.getNode())
1358 ExtraCmp = emitComparison(LHS, RHS, CC, DL, DAG);
1360 SDValue ConditionOp = DAG.getConstant(Predicate, DL, MVT_CC);
1361 // Note that we want the inverse of ExtraCC, so NZCV is not inversed.
1362 unsigned NZCV = AArch64CC::getNZCVToSatisfyCondCode(ExtraCC);
1363 ExtraCmp = emitConditionalComparison(LHS, RHS, CC, CCOp, ConditionOp,
1367 Predicate = AArch64CC::getInvertedCondCode(ExtraCC);
1368 OutCC = AArch64CC::getInvertedCondCode(OutCC);
1372 // Produce a normal comparison if we are first in the chain
1373 if (!CCOp.getNode())
1374 return emitComparison(LHS, RHS, CC, DL, DAG);
1375 // Otherwise produce a ccmp.
1376 SDValue ConditionOp = DAG.getConstant(Predicate, DL, MVT_CC);
1377 AArch64CC::CondCode InvOutCC = AArch64CC::getInvertedCondCode(OutCC);
1378 unsigned NZCV = AArch64CC::getNZCVToSatisfyCondCode(InvOutCC);
1379 return emitConditionalComparison(LHS, RHS, CC, CCOp, ConditionOp, NZCV, DL,
1381 } else if ((Opcode != ISD::AND && Opcode != ISD::OR) || !Val->hasOneUse())
1384 assert((Opcode == ISD::OR || !PushNegate)
1385 && "Can only push negate through OR operation");
1387 // Check if both sides can be transformed.
1388 SDValue LHS = Val->getOperand(0);
1389 SDValue RHS = Val->getOperand(1);
1390 bool CanPushNegateL;
1391 if (!isConjunctionDisjunctionTree(LHS, CanPushNegateL, Depth+1))
1393 bool CanPushNegateR;
1394 if (!isConjunctionDisjunctionTree(RHS, CanPushNegateR, Depth+1))
1397 // Do we need to negate our operands?
1398 bool NegateOperands = Opcode == ISD::OR;
1399 // We can negate the results of all previous operations by inverting the
1400 // predicate flags giving us a free negation for one side. For the other side
1401 // we need to be able to push the negation to the leafs of the tree.
1402 if (NegateOperands) {
1403 if (!CanPushNegateL && !CanPushNegateR)
1405 // Order the side where we can push the negate through to LHS.
1406 if (!CanPushNegateL && CanPushNegateR)
1407 std::swap(LHS, RHS);
1409 bool NeedsNegOutL = LHS->getOpcode() == ISD::OR;
1410 bool NeedsNegOutR = RHS->getOpcode() == ISD::OR;
1411 if (NeedsNegOutL && NeedsNegOutR)
1413 // Order the side where we need to negate the output flags to RHS so it
1414 // gets emitted first.
1416 std::swap(LHS, RHS);
1419 // Emit RHS. If we want to negate the tree we only need to push a negate
1420 // through if we are already in a PushNegate case, otherwise we can negate
1421 // the "flags to test" afterwards.
1422 AArch64CC::CondCode RHSCC;
1423 SDValue CmpR = emitConjunctionDisjunctionTree(DAG, RHS, RHSCC, PushNegate,
1424 CCOp, Predicate, Depth+1);
1425 if (NegateOperands && !PushNegate)
1426 RHSCC = AArch64CC::getInvertedCondCode(RHSCC);
1427 // Emit LHS. We must push the negate through if we need to negate it.
1428 SDValue CmpL = emitConjunctionDisjunctionTree(DAG, LHS, OutCC, NegateOperands,
1429 CmpR, RHSCC, Depth+1);
1430 // If we transformed an OR to and AND then we have to negate the result
1431 // (or absorb a PushNegate resulting in a double negation).
1432 if (Opcode == ISD::OR && !PushNegate)
1433 OutCC = AArch64CC::getInvertedCondCode(OutCC);
1439 static SDValue getAArch64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1440 SDValue &AArch64cc, SelectionDAG &DAG, SDLoc dl) {
1441 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1442 EVT VT = RHS.getValueType();
1443 uint64_t C = RHSC->getZExtValue();
1444 if (!isLegalArithImmed(C)) {
1445 // Constant does not fit, try adjusting it by one?
1451 if ((VT == MVT::i32 && C != 0x80000000 &&
1452 isLegalArithImmed((uint32_t)(C - 1))) ||
1453 (VT == MVT::i64 && C != 0x80000000ULL &&
1454 isLegalArithImmed(C - 1ULL))) {
1455 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1456 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1457 RHS = DAG.getConstant(C, dl, VT);
1462 if ((VT == MVT::i32 && C != 0 &&
1463 isLegalArithImmed((uint32_t)(C - 1))) ||
1464 (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) {
1465 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1466 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1467 RHS = DAG.getConstant(C, dl, VT);
1472 if ((VT == MVT::i32 && C != INT32_MAX &&
1473 isLegalArithImmed((uint32_t)(C + 1))) ||
1474 (VT == MVT::i64 && C != INT64_MAX &&
1475 isLegalArithImmed(C + 1ULL))) {
1476 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1477 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1478 RHS = DAG.getConstant(C, dl, VT);
1483 if ((VT == MVT::i32 && C != UINT32_MAX &&
1484 isLegalArithImmed((uint32_t)(C + 1))) ||
1485 (VT == MVT::i64 && C != UINT64_MAX &&
1486 isLegalArithImmed(C + 1ULL))) {
1487 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1488 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1489 RHS = DAG.getConstant(C, dl, VT);
1496 AArch64CC::CondCode AArch64CC;
1497 if ((CC == ISD::SETEQ || CC == ISD::SETNE) && isa<ConstantSDNode>(RHS)) {
1498 const ConstantSDNode *RHSC = cast<ConstantSDNode>(RHS);
1500 // The imm operand of ADDS is an unsigned immediate, in the range 0 to 4095.
1501 // For the i8 operand, the largest immediate is 255, so this can be easily
1502 // encoded in the compare instruction. For the i16 operand, however, the
1503 // largest immediate cannot be encoded in the compare.
1504 // Therefore, use a sign extending load and cmn to avoid materializing the
1505 // -1 constant. For example,
1507 // ldrh w0, [x0, #0]
1510 // ldrsh w0, [x0, #0]
1512 // Fundamental, we're relying on the property that (zext LHS) == (zext RHS)
1513 // if and only if (sext LHS) == (sext RHS). The checks are in place to
1514 // ensure both the LHS and RHS are truly zero extended and to make sure the
1515 // transformation is profitable.
1516 if ((RHSC->getZExtValue() >> 16 == 0) && isa<LoadSDNode>(LHS) &&
1517 cast<LoadSDNode>(LHS)->getExtensionType() == ISD::ZEXTLOAD &&
1518 cast<LoadSDNode>(LHS)->getMemoryVT() == MVT::i16 &&
1519 LHS.getNode()->hasNUsesOfValue(1, 0)) {
1520 int16_t ValueofRHS = cast<ConstantSDNode>(RHS)->getZExtValue();
1521 if (ValueofRHS < 0 && isLegalArithImmed(-ValueofRHS)) {
1523 DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, LHS.getValueType(), LHS,
1524 DAG.getValueType(MVT::i16));
1525 Cmp = emitComparison(SExt, DAG.getConstant(ValueofRHS, dl,
1526 RHS.getValueType()),
1528 AArch64CC = changeIntCCToAArch64CC(CC);
1532 if (!Cmp && (RHSC->isNullValue() || RHSC->isOne())) {
1533 if ((Cmp = emitConjunctionDisjunctionTree(DAG, LHS, AArch64CC))) {
1534 if ((CC == ISD::SETNE) ^ RHSC->isNullValue())
1535 AArch64CC = AArch64CC::getInvertedCondCode(AArch64CC);
1541 Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
1542 AArch64CC = changeIntCCToAArch64CC(CC);
1544 AArch64cc = DAG.getConstant(AArch64CC, dl, MVT_CC);
1548 static std::pair<SDValue, SDValue>
1549 getAArch64XALUOOp(AArch64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
1550 assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
1551 "Unsupported value type");
1552 SDValue Value, Overflow;
1554 SDValue LHS = Op.getOperand(0);
1555 SDValue RHS = Op.getOperand(1);
1557 switch (Op.getOpcode()) {
1559 llvm_unreachable("Unknown overflow instruction!");
1561 Opc = AArch64ISD::ADDS;
1565 Opc = AArch64ISD::ADDS;
1569 Opc = AArch64ISD::SUBS;
1573 Opc = AArch64ISD::SUBS;
1576 // Multiply needs a little bit extra work.
1580 bool IsSigned = Op.getOpcode() == ISD::SMULO;
1581 if (Op.getValueType() == MVT::i32) {
1582 unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1583 // For a 32 bit multiply with overflow check we want the instruction
1584 // selector to generate a widening multiply (SMADDL/UMADDL). For that we
1585 // need to generate the following pattern:
1586 // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b))
1587 LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
1588 RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
1589 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1590 SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul,
1591 DAG.getConstant(0, DL, MVT::i64));
1592 // On AArch64 the upper 32 bits are always zero extended for a 32 bit
1593 // operation. We need to clear out the upper 32 bits, because we used a
1594 // widening multiply that wrote all 64 bits. In the end this should be a
1596 Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add);
1598 // The signed overflow check requires more than just a simple check for
1599 // any bit set in the upper 32 bits of the result. These bits could be
1600 // just the sign bits of a negative number. To perform the overflow
1601 // check we have to arithmetic shift right the 32nd bit of the result by
1602 // 31 bits. Then we compare the result to the upper 32 bits.
1603 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add,
1604 DAG.getConstant(32, DL, MVT::i64));
1605 UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits);
1606 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value,
1607 DAG.getConstant(31, DL, MVT::i64));
1608 // It is important that LowerBits is last, otherwise the arithmetic
1609 // shift will not be folded into the compare (SUBS).
1610 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32);
1611 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1614 // The overflow check for unsigned multiply is easy. We only need to
1615 // check if any of the upper 32 bits are set. This can be done with a
1616 // CMP (shifted register). For that we need to generate the following
1618 // (i64 AArch64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32)
1619 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
1620 DAG.getConstant(32, DL, MVT::i64));
1621 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1623 DAG.getNode(AArch64ISD::SUBS, DL, VTs,
1624 DAG.getConstant(0, DL, MVT::i64),
1625 UpperBits).getValue(1);
1629 assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
1630 // For the 64 bit multiply
1631 Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1633 SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
1634 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
1635 DAG.getConstant(63, DL, MVT::i64));
1636 // It is important that LowerBits is last, otherwise the arithmetic
1637 // shift will not be folded into the compare (SUBS).
1638 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1639 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1642 SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
1643 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1645 DAG.getNode(AArch64ISD::SUBS, DL, VTs,
1646 DAG.getConstant(0, DL, MVT::i64),
1647 UpperBits).getValue(1);
1654 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);
1656 // Emit the AArch64 operation with overflow check.
1657 Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
1658 Overflow = Value.getValue(1);
1660 return std::make_pair(Value, Overflow);
1663 SDValue AArch64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG,
1664 RTLIB::Libcall Call) const {
1665 SmallVector<SDValue, 2> Ops(Op->op_begin(), Op->op_end());
1666 return makeLibCall(DAG, Call, MVT::f128, Ops, false, SDLoc(Op)).first;
1669 static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) {
1670 SDValue Sel = Op.getOperand(0);
1671 SDValue Other = Op.getOperand(1);
1673 // If neither operand is a SELECT_CC, give up.
1674 if (Sel.getOpcode() != ISD::SELECT_CC)
1675 std::swap(Sel, Other);
1676 if (Sel.getOpcode() != ISD::SELECT_CC)
1679 // The folding we want to perform is:
1680 // (xor x, (select_cc a, b, cc, 0, -1) )
1682 // (csel x, (xor x, -1), cc ...)
1684 // The latter will get matched to a CSINV instruction.
1686 ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
1687 SDValue LHS = Sel.getOperand(0);
1688 SDValue RHS = Sel.getOperand(1);
1689 SDValue TVal = Sel.getOperand(2);
1690 SDValue FVal = Sel.getOperand(3);
1693 // FIXME: This could be generalized to non-integer comparisons.
1694 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
1697 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
1698 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
1700 // The values aren't constants, this isn't the pattern we're looking for.
1701 if (!CFVal || !CTVal)
1704 // We can commute the SELECT_CC by inverting the condition. This
1705 // might be needed to make this fit into a CSINV pattern.
1706 if (CTVal->isAllOnesValue() && CFVal->isNullValue()) {
1707 std::swap(TVal, FVal);
1708 std::swap(CTVal, CFVal);
1709 CC = ISD::getSetCCInverse(CC, true);
1712 // If the constants line up, perform the transform!
1713 if (CTVal->isNullValue() && CFVal->isAllOnesValue()) {
1715 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
1718 TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other,
1719 DAG.getConstant(-1ULL, dl, Other.getValueType()));
1721 return DAG.getNode(AArch64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal,
1728 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
1729 EVT VT = Op.getValueType();
1731 // Let legalize expand this if it isn't a legal type yet.
1732 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
1735 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
1738 bool ExtraOp = false;
1739 switch (Op.getOpcode()) {
1741 llvm_unreachable("Invalid code");
1743 Opc = AArch64ISD::ADDS;
1746 Opc = AArch64ISD::SUBS;
1749 Opc = AArch64ISD::ADCS;
1753 Opc = AArch64ISD::SBCS;
1759 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1));
1760 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1),
1764 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
1765 // Let legalize expand this if it isn't a legal type yet.
1766 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
1770 AArch64CC::CondCode CC;
1771 // The actual operation that sets the overflow or carry flag.
1772 SDValue Value, Overflow;
1773 std::tie(Value, Overflow) = getAArch64XALUOOp(CC, Op, DAG);
1775 // We use 0 and 1 as false and true values.
1776 SDValue TVal = DAG.getConstant(1, dl, MVT::i32);
1777 SDValue FVal = DAG.getConstant(0, dl, MVT::i32);
1779 // We use an inverted condition, because the conditional select is inverted
1780 // too. This will allow it to be selected to a single instruction:
1781 // CSINC Wd, WZR, WZR, invert(cond).
1782 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), dl, MVT::i32);
1783 Overflow = DAG.getNode(AArch64ISD::CSEL, dl, MVT::i32, FVal, TVal,
1786 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
1787 return DAG.getNode(ISD::MERGE_VALUES, dl, VTs, Value, Overflow);
1790 // Prefetch operands are:
1791 // 1: Address to prefetch
1793 // 3: int locality (0 = no locality ... 3 = extreme locality)
1794 // 4: bool isDataCache
1795 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
1797 unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
1798 unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
1799 unsigned IsData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
1801 bool IsStream = !Locality;
1802 // When the locality number is set
1804 // The front-end should have filtered out the out-of-range values
1805 assert(Locality <= 3 && "Prefetch locality out-of-range");
1806 // The locality degree is the opposite of the cache speed.
1807 // Put the number the other way around.
1808 // The encoding starts at 0 for level 1
1809 Locality = 3 - Locality;
1812 // built the mask value encoding the expected behavior.
1813 unsigned PrfOp = (IsWrite << 4) | // Load/Store bit
1814 (!IsData << 3) | // IsDataCache bit
1815 (Locality << 1) | // Cache level bits
1816 (unsigned)IsStream; // Stream bit
1817 return DAG.getNode(AArch64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
1818 DAG.getConstant(PrfOp, DL, MVT::i32), Op.getOperand(1));
1821 SDValue AArch64TargetLowering::LowerFP_EXTEND(SDValue Op,
1822 SelectionDAG &DAG) const {
1823 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
1826 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
1828 return LowerF128Call(Op, DAG, LC);
1831 SDValue AArch64TargetLowering::LowerFP_ROUND(SDValue Op,
1832 SelectionDAG &DAG) const {
1833 if (Op.getOperand(0).getValueType() != MVT::f128) {
1834 // It's legal except when f128 is involved
1839 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
1841 // FP_ROUND node has a second operand indicating whether it is known to be
1842 // precise. That doesn't take part in the LibCall so we can't directly use
1844 SDValue SrcVal = Op.getOperand(0);
1845 return makeLibCall(DAG, LC, Op.getValueType(), SrcVal, /*isSigned*/ false,
1849 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1850 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1851 // Any additional optimization in this function should be recorded
1852 // in the cost tables.
1853 EVT InVT = Op.getOperand(0).getValueType();
1854 EVT VT = Op.getValueType();
1856 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1859 DAG.getNode(Op.getOpcode(), dl, InVT.changeVectorElementTypeToInteger(),
1861 return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
1864 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1867 MVT::getVectorVT(MVT::getFloatingPointVT(VT.getScalarSizeInBits()),
1868 VT.getVectorNumElements());
1869 SDValue Ext = DAG.getNode(ISD::FP_EXTEND, dl, ExtVT, Op.getOperand(0));
1870 return DAG.getNode(Op.getOpcode(), dl, VT, Ext);
1873 // Type changing conversions are illegal.
1877 SDValue AArch64TargetLowering::LowerFP_TO_INT(SDValue Op,
1878 SelectionDAG &DAG) const {
1879 if (Op.getOperand(0).getValueType().isVector())
1880 return LowerVectorFP_TO_INT(Op, DAG);
1882 // f16 conversions are promoted to f32.
1883 if (Op.getOperand(0).getValueType() == MVT::f16) {
1886 Op.getOpcode(), dl, Op.getValueType(),
1887 DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, Op.getOperand(0)));
1890 if (Op.getOperand(0).getValueType() != MVT::f128) {
1891 // It's legal except when f128 is involved
1896 if (Op.getOpcode() == ISD::FP_TO_SINT)
1897 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
1899 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
1901 SmallVector<SDValue, 2> Ops(Op->op_begin(), Op->op_end());
1902 return makeLibCall(DAG, LC, Op.getValueType(), Ops, false, SDLoc(Op)).first;
1905 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1906 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1907 // Any additional optimization in this function should be recorded
1908 // in the cost tables.
1909 EVT VT = Op.getValueType();
1911 SDValue In = Op.getOperand(0);
1912 EVT InVT = In.getValueType();
1914 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1916 MVT::getVectorVT(MVT::getFloatingPointVT(InVT.getScalarSizeInBits()),
1917 InVT.getVectorNumElements());
1918 In = DAG.getNode(Op.getOpcode(), dl, CastVT, In);
1919 return DAG.getNode(ISD::FP_ROUND, dl, VT, In, DAG.getIntPtrConstant(0, dl));
1922 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1924 Op.getOpcode() == ISD::SINT_TO_FP ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1925 EVT CastVT = VT.changeVectorElementTypeToInteger();
1926 In = DAG.getNode(CastOpc, dl, CastVT, In);
1927 return DAG.getNode(Op.getOpcode(), dl, VT, In);
1933 SDValue AArch64TargetLowering::LowerINT_TO_FP(SDValue Op,
1934 SelectionDAG &DAG) const {
1935 if (Op.getValueType().isVector())
1936 return LowerVectorINT_TO_FP(Op, DAG);
1938 // f16 conversions are promoted to f32.
1939 if (Op.getValueType() == MVT::f16) {
1942 ISD::FP_ROUND, dl, MVT::f16,
1943 DAG.getNode(Op.getOpcode(), dl, MVT::f32, Op.getOperand(0)),
1944 DAG.getIntPtrConstant(0, dl));
1947 // i128 conversions are libcalls.
1948 if (Op.getOperand(0).getValueType() == MVT::i128)
1951 // Other conversions are legal, unless it's to the completely software-based
1953 if (Op.getValueType() != MVT::f128)
1957 if (Op.getOpcode() == ISD::SINT_TO_FP)
1958 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1960 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1962 return LowerF128Call(Op, DAG, LC);
1965 SDValue AArch64TargetLowering::LowerFSINCOS(SDValue Op,
1966 SelectionDAG &DAG) const {
1967 // For iOS, we want to call an alternative entry point: __sincos_stret,
1968 // which returns the values in two S / D registers.
1970 SDValue Arg = Op.getOperand(0);
1971 EVT ArgVT = Arg.getValueType();
1972 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1979 Entry.isSExt = false;
1980 Entry.isZExt = false;
1981 Args.push_back(Entry);
1983 const char *LibcallName =
1984 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
1986 DAG.getExternalSymbol(LibcallName, getPointerTy(DAG.getDataLayout()));
1988 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
1989 TargetLowering::CallLoweringInfo CLI(DAG);
1990 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
1991 .setCallee(CallingConv::Fast, RetTy, Callee, std::move(Args), 0);
1993 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1994 return CallResult.first;
1997 static SDValue LowerBITCAST(SDValue Op, SelectionDAG &DAG) {
1998 if (Op.getValueType() != MVT::f16)
2001 assert(Op.getOperand(0).getValueType() == MVT::i16);
2004 Op = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op.getOperand(0));
2005 Op = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Op);
2007 DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL, MVT::f16, Op,
2008 DAG.getTargetConstant(AArch64::hsub, DL, MVT::i32)),
2012 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
2013 if (OrigVT.getSizeInBits() >= 64)
2016 assert(OrigVT.isSimple() && "Expecting a simple value type");
2018 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
2019 switch (OrigSimpleTy) {
2020 default: llvm_unreachable("Unexpected Vector Type");
2029 static SDValue addRequiredExtensionForVectorMULL(SDValue N, SelectionDAG &DAG,
2032 unsigned ExtOpcode) {
2033 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
2034 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
2035 // 64-bits we need to insert a new extension so that it will be 64-bits.
2036 assert(ExtTy.is128BitVector() && "Unexpected extension size");
2037 if (OrigTy.getSizeInBits() >= 64)
2040 // Must extend size to at least 64 bits to be used as an operand for VMULL.
2041 EVT NewVT = getExtensionTo64Bits(OrigTy);
2043 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
2046 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
2048 EVT VT = N->getValueType(0);
2050 if (N->getOpcode() != ISD::BUILD_VECTOR)
2053 for (const SDValue &Elt : N->op_values()) {
2054 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
2055 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
2056 unsigned HalfSize = EltSize / 2;
2058 if (!isIntN(HalfSize, C->getSExtValue()))
2061 if (!isUIntN(HalfSize, C->getZExtValue()))
2072 static SDValue skipExtensionForVectorMULL(SDNode *N, SelectionDAG &DAG) {
2073 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
2074 return addRequiredExtensionForVectorMULL(N->getOperand(0), DAG,
2075 N->getOperand(0)->getValueType(0),
2079 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
2080 EVT VT = N->getValueType(0);
2082 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
2083 unsigned NumElts = VT.getVectorNumElements();
2084 MVT TruncVT = MVT::getIntegerVT(EltSize);
2085 SmallVector<SDValue, 8> Ops;
2086 for (unsigned i = 0; i != NumElts; ++i) {
2087 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
2088 const APInt &CInt = C->getAPIntValue();
2089 // Element types smaller than 32 bits are not legal, so use i32 elements.
2090 // The values are implicitly truncated so sext vs. zext doesn't matter.
2091 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), dl, MVT::i32));
2093 return DAG.getNode(ISD::BUILD_VECTOR, dl,
2094 MVT::getVectorVT(TruncVT, NumElts), Ops);
2097 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
2098 if (N->getOpcode() == ISD::SIGN_EXTEND)
2100 if (isExtendedBUILD_VECTOR(N, DAG, true))
2105 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
2106 if (N->getOpcode() == ISD::ZERO_EXTEND)
2108 if (isExtendedBUILD_VECTOR(N, DAG, false))
2113 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
2114 unsigned Opcode = N->getOpcode();
2115 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
2116 SDNode *N0 = N->getOperand(0).getNode();
2117 SDNode *N1 = N->getOperand(1).getNode();
2118 return N0->hasOneUse() && N1->hasOneUse() &&
2119 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
2124 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
2125 unsigned Opcode = N->getOpcode();
2126 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
2127 SDNode *N0 = N->getOperand(0).getNode();
2128 SDNode *N1 = N->getOperand(1).getNode();
2129 return N0->hasOneUse() && N1->hasOneUse() &&
2130 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
2135 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
2136 // Multiplications are only custom-lowered for 128-bit vectors so that
2137 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
2138 EVT VT = Op.getValueType();
2139 assert(VT.is128BitVector() && VT.isInteger() &&
2140 "unexpected type for custom-lowering ISD::MUL");
2141 SDNode *N0 = Op.getOperand(0).getNode();
2142 SDNode *N1 = Op.getOperand(1).getNode();
2143 unsigned NewOpc = 0;
2145 bool isN0SExt = isSignExtended(N0, DAG);
2146 bool isN1SExt = isSignExtended(N1, DAG);
2147 if (isN0SExt && isN1SExt)
2148 NewOpc = AArch64ISD::SMULL;
2150 bool isN0ZExt = isZeroExtended(N0, DAG);
2151 bool isN1ZExt = isZeroExtended(N1, DAG);
2152 if (isN0ZExt && isN1ZExt)
2153 NewOpc = AArch64ISD::UMULL;
2154 else if (isN1SExt || isN1ZExt) {
2155 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
2156 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
2157 if (isN1SExt && isAddSubSExt(N0, DAG)) {
2158 NewOpc = AArch64ISD::SMULL;
2160 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
2161 NewOpc = AArch64ISD::UMULL;
2163 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
2165 NewOpc = AArch64ISD::UMULL;
2171 if (VT == MVT::v2i64)
2172 // Fall through to expand this. It is not legal.
2175 // Other vector multiplications are legal.
2180 // Legalize to a S/UMULL instruction
2183 SDValue Op1 = skipExtensionForVectorMULL(N1, DAG);
2185 Op0 = skipExtensionForVectorMULL(N0, DAG);
2186 assert(Op0.getValueType().is64BitVector() &&
2187 Op1.getValueType().is64BitVector() &&
2188 "unexpected types for extended operands to VMULL");
2189 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
2191 // Optimizing (zext A + zext B) * C, to (S/UMULL A, C) + (S/UMULL B, C) during
2192 // isel lowering to take advantage of no-stall back to back s/umul + s/umla.
2193 // This is true for CPUs with accumulate forwarding such as Cortex-A53/A57
2194 SDValue N00 = skipExtensionForVectorMULL(N0->getOperand(0).getNode(), DAG);
2195 SDValue N01 = skipExtensionForVectorMULL(N0->getOperand(1).getNode(), DAG);
2196 EVT Op1VT = Op1.getValueType();
2197 return DAG.getNode(N0->getOpcode(), DL, VT,
2198 DAG.getNode(NewOpc, DL, VT,
2199 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
2200 DAG.getNode(NewOpc, DL, VT,
2201 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
2204 SDValue AArch64TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
2205 SelectionDAG &DAG) const {
2206 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2209 default: return SDValue(); // Don't custom lower most intrinsics.
2210 case Intrinsic::aarch64_thread_pointer: {
2211 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2212 return DAG.getNode(AArch64ISD::THREAD_POINTER, dl, PtrVT);
2214 case Intrinsic::aarch64_neon_smax:
2215 return DAG.getNode(ISD::SMAX, dl, Op.getValueType(),
2216 Op.getOperand(1), Op.getOperand(2));
2217 case Intrinsic::aarch64_neon_umax:
2218 return DAG.getNode(ISD::UMAX, dl, Op.getValueType(),
2219 Op.getOperand(1), Op.getOperand(2));
2220 case Intrinsic::aarch64_neon_smin:
2221 return DAG.getNode(ISD::SMIN, dl, Op.getValueType(),
2222 Op.getOperand(1), Op.getOperand(2));
2223 case Intrinsic::aarch64_neon_umin:
2224 return DAG.getNode(ISD::UMIN, dl, Op.getValueType(),
2225 Op.getOperand(1), Op.getOperand(2));
2229 SDValue AArch64TargetLowering::LowerOperation(SDValue Op,
2230 SelectionDAG &DAG) const {
2231 switch (Op.getOpcode()) {
2233 llvm_unreachable("unimplemented operand");
2236 return LowerBITCAST(Op, DAG);
2237 case ISD::GlobalAddress:
2238 return LowerGlobalAddress(Op, DAG);
2239 case ISD::GlobalTLSAddress:
2240 return LowerGlobalTLSAddress(Op, DAG);
2242 return LowerSETCC(Op, DAG);
2244 return LowerBR_CC(Op, DAG);
2246 return LowerSELECT(Op, DAG);
2247 case ISD::SELECT_CC:
2248 return LowerSELECT_CC(Op, DAG);
2249 case ISD::JumpTable:
2250 return LowerJumpTable(Op, DAG);
2251 case ISD::ConstantPool:
2252 return LowerConstantPool(Op, DAG);
2253 case ISD::BlockAddress:
2254 return LowerBlockAddress(Op, DAG);
2256 return LowerVASTART(Op, DAG);
2258 return LowerVACOPY(Op, DAG);
2260 return LowerVAARG(Op, DAG);
2265 return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
2272 return LowerXALUO(Op, DAG);
2274 return LowerF128Call(Op, DAG, RTLIB::ADD_F128);
2276 return LowerF128Call(Op, DAG, RTLIB::SUB_F128);
2278 return LowerF128Call(Op, DAG, RTLIB::MUL_F128);
2280 return LowerF128Call(Op, DAG, RTLIB::DIV_F128);
2282 return LowerFP_ROUND(Op, DAG);
2283 case ISD::FP_EXTEND:
2284 return LowerFP_EXTEND(Op, DAG);
2285 case ISD::FRAMEADDR:
2286 return LowerFRAMEADDR(Op, DAG);
2287 case ISD::RETURNADDR:
2288 return LowerRETURNADDR(Op, DAG);
2289 case ISD::INSERT_VECTOR_ELT:
2290 return LowerINSERT_VECTOR_ELT(Op, DAG);
2291 case ISD::EXTRACT_VECTOR_ELT:
2292 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2293 case ISD::BUILD_VECTOR:
2294 return LowerBUILD_VECTOR(Op, DAG);
2295 case ISD::VECTOR_SHUFFLE:
2296 return LowerVECTOR_SHUFFLE(Op, DAG);
2297 case ISD::EXTRACT_SUBVECTOR:
2298 return LowerEXTRACT_SUBVECTOR(Op, DAG);
2302 return LowerVectorSRA_SRL_SHL(Op, DAG);
2303 case ISD::SHL_PARTS:
2304 return LowerShiftLeftParts(Op, DAG);
2305 case ISD::SRL_PARTS:
2306 case ISD::SRA_PARTS:
2307 return LowerShiftRightParts(Op, DAG);
2309 return LowerCTPOP(Op, DAG);
2310 case ISD::FCOPYSIGN:
2311 return LowerFCOPYSIGN(Op, DAG);
2313 return LowerVectorAND(Op, DAG);
2315 return LowerVectorOR(Op, DAG);
2317 return LowerXOR(Op, DAG);
2319 return LowerPREFETCH(Op, DAG);
2320 case ISD::SINT_TO_FP:
2321 case ISD::UINT_TO_FP:
2322 return LowerINT_TO_FP(Op, DAG);
2323 case ISD::FP_TO_SINT:
2324 case ISD::FP_TO_UINT:
2325 return LowerFP_TO_INT(Op, DAG);
2327 return LowerFSINCOS(Op, DAG);
2329 return LowerMUL(Op, DAG);
2330 case ISD::INTRINSIC_WO_CHAIN:
2331 return LowerINTRINSIC_WO_CHAIN(Op, DAG);
2335 /// getFunctionAlignment - Return the Log2 alignment of this function.
2336 unsigned AArch64TargetLowering::getFunctionAlignment(const Function *F) const {
2340 //===----------------------------------------------------------------------===//
2341 // Calling Convention Implementation
2342 //===----------------------------------------------------------------------===//
2344 #include "AArch64GenCallingConv.inc"
2346 /// Selects the correct CCAssignFn for a given CallingConvention value.
2347 CCAssignFn *AArch64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
2348 bool IsVarArg) const {
2351 llvm_unreachable("Unsupported calling convention.");
2352 case CallingConv::WebKit_JS:
2353 return CC_AArch64_WebKit_JS;
2354 case CallingConv::GHC:
2355 return CC_AArch64_GHC;
2356 case CallingConv::C:
2357 case CallingConv::Fast:
2358 if (!Subtarget->isTargetDarwin())
2359 return CC_AArch64_AAPCS;
2360 return IsVarArg ? CC_AArch64_DarwinPCS_VarArg : CC_AArch64_DarwinPCS;
2364 SDValue AArch64TargetLowering::LowerFormalArguments(
2365 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2366 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2367 SmallVectorImpl<SDValue> &InVals) const {
2368 MachineFunction &MF = DAG.getMachineFunction();
2369 MachineFrameInfo *MFI = MF.getFrameInfo();
2371 // Assign locations to all of the incoming arguments.
2372 SmallVector<CCValAssign, 16> ArgLocs;
2373 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
2376 // At this point, Ins[].VT may already be promoted to i32. To correctly
2377 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2378 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
2379 // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
2380 // we use a special version of AnalyzeFormalArguments to pass in ValVT and
2382 unsigned NumArgs = Ins.size();
2383 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2384 unsigned CurArgIdx = 0;
2385 for (unsigned i = 0; i != NumArgs; ++i) {
2386 MVT ValVT = Ins[i].VT;
2387 if (Ins[i].isOrigArg()) {
2388 std::advance(CurOrigArg, Ins[i].getOrigArgIndex() - CurArgIdx);
2389 CurArgIdx = Ins[i].getOrigArgIndex();
2391 // Get type of the original argument.
2392 EVT ActualVT = getValueType(DAG.getDataLayout(), CurOrigArg->getType(),
2393 /*AllowUnknown*/ true);
2394 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
2395 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2396 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2398 else if (ActualMVT == MVT::i16)
2401 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2403 AssignFn(i, ValVT, ValVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
2404 assert(!Res && "Call operand has unhandled type");
2407 assert(ArgLocs.size() == Ins.size());
2408 SmallVector<SDValue, 16> ArgValues;
2409 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2410 CCValAssign &VA = ArgLocs[i];
2412 if (Ins[i].Flags.isByVal()) {
2413 // Byval is used for HFAs in the PCS, but the system should work in a
2414 // non-compliant manner for larger structs.
2415 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2416 int Size = Ins[i].Flags.getByValSize();
2417 unsigned NumRegs = (Size + 7) / 8;
2419 // FIXME: This works on big-endian for composite byvals, which are the common
2420 // case. It should also work for fundamental types too.
2422 MFI->CreateFixedObject(8 * NumRegs, VA.getLocMemOffset(), false);
2423 SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrVT);
2424 InVals.push_back(FrameIdxN);
2429 if (VA.isRegLoc()) {
2430 // Arguments stored in registers.
2431 EVT RegVT = VA.getLocVT();
2434 const TargetRegisterClass *RC;
2436 if (RegVT == MVT::i32)
2437 RC = &AArch64::GPR32RegClass;
2438 else if (RegVT == MVT::i64)
2439 RC = &AArch64::GPR64RegClass;
2440 else if (RegVT == MVT::f16)
2441 RC = &AArch64::FPR16RegClass;
2442 else if (RegVT == MVT::f32)
2443 RC = &AArch64::FPR32RegClass;
2444 else if (RegVT == MVT::f64 || RegVT.is64BitVector())
2445 RC = &AArch64::FPR64RegClass;
2446 else if (RegVT == MVT::f128 || RegVT.is128BitVector())
2447 RC = &AArch64::FPR128RegClass;
2449 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2451 // Transform the arguments in physical registers into virtual ones.
2452 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2453 ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2455 // If this is an 8, 16 or 32-bit value, it is really passed promoted
2456 // to 64 bits. Insert an assert[sz]ext to capture this, then
2457 // truncate to the right size.
2458 switch (VA.getLocInfo()) {
2460 llvm_unreachable("Unknown loc info!");
2461 case CCValAssign::Full:
2463 case CCValAssign::BCvt:
2464 ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue);
2466 case CCValAssign::AExt:
2467 case CCValAssign::SExt:
2468 case CCValAssign::ZExt:
2469 // SelectionDAGBuilder will insert appropriate AssertZExt & AssertSExt
2470 // nodes after our lowering.
2471 assert(RegVT == Ins[i].VT && "incorrect register location selected");
2475 InVals.push_back(ArgValue);
2477 } else { // VA.isRegLoc()
2478 assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
2479 unsigned ArgOffset = VA.getLocMemOffset();
2480 unsigned ArgSize = VA.getValVT().getSizeInBits() / 8;
2482 uint32_t BEAlign = 0;
2483 if (!Subtarget->isLittleEndian() && ArgSize < 8 &&
2484 !Ins[i].Flags.isInConsecutiveRegs())
2485 BEAlign = 8 - ArgSize;
2487 int FI = MFI->CreateFixedObject(ArgSize, ArgOffset + BEAlign, true);
2489 // Create load nodes to retrieve arguments from the stack.
2490 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2493 // For NON_EXTLOAD, generic code in getLoad assert(ValVT == MemVT)
2494 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
2495 MVT MemVT = VA.getValVT();
2497 switch (VA.getLocInfo()) {
2500 case CCValAssign::BCvt:
2501 MemVT = VA.getLocVT();
2503 case CCValAssign::SExt:
2504 ExtType = ISD::SEXTLOAD;
2506 case CCValAssign::ZExt:
2507 ExtType = ISD::ZEXTLOAD;
2509 case CCValAssign::AExt:
2510 ExtType = ISD::EXTLOAD;
2514 ArgValue = DAG.getExtLoad(
2515 ExtType, DL, VA.getLocVT(), Chain, FIN,
2516 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
2517 MemVT, false, false, false, 0);
2519 InVals.push_back(ArgValue);
2525 if (!Subtarget->isTargetDarwin()) {
2526 // The AAPCS variadic function ABI is identical to the non-variadic
2527 // one. As a result there may be more arguments in registers and we should
2528 // save them for future reference.
2529 saveVarArgRegisters(CCInfo, DAG, DL, Chain);
2532 AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>();
2533 // This will point to the next argument passed via stack.
2534 unsigned StackOffset = CCInfo.getNextStackOffset();
2535 // We currently pass all varargs at 8-byte alignment.
2536 StackOffset = ((StackOffset + 7) & ~7);
2537 AFI->setVarArgsStackIndex(MFI->CreateFixedObject(4, StackOffset, true));
2540 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2541 unsigned StackArgSize = CCInfo.getNextStackOffset();
2542 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2543 if (DoesCalleeRestoreStack(CallConv, TailCallOpt)) {
2544 // This is a non-standard ABI so by fiat I say we're allowed to make full
2545 // use of the stack area to be popped, which must be aligned to 16 bytes in
2547 StackArgSize = RoundUpToAlignment(StackArgSize, 16);
2549 // If we're expected to restore the stack (e.g. fastcc) then we'll be adding
2550 // a multiple of 16.
2551 FuncInfo->setArgumentStackToRestore(StackArgSize);
2553 // This realignment carries over to the available bytes below. Our own
2554 // callers will guarantee the space is free by giving an aligned value to
2557 // Even if we're not expected to free up the space, it's useful to know how
2558 // much is there while considering tail calls (because we can reuse it).
2559 FuncInfo->setBytesInStackArgArea(StackArgSize);
2564 void AArch64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
2565 SelectionDAG &DAG, SDLoc DL,
2566 SDValue &Chain) const {
2567 MachineFunction &MF = DAG.getMachineFunction();
2568 MachineFrameInfo *MFI = MF.getFrameInfo();
2569 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2570 auto PtrVT = getPointerTy(DAG.getDataLayout());
2572 SmallVector<SDValue, 8> MemOps;
2574 static const MCPhysReg GPRArgRegs[] = { AArch64::X0, AArch64::X1, AArch64::X2,
2575 AArch64::X3, AArch64::X4, AArch64::X5,
2576 AArch64::X6, AArch64::X7 };
2577 static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs);
2578 unsigned FirstVariadicGPR = CCInfo.getFirstUnallocated(GPRArgRegs);
2580 unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
2582 if (GPRSaveSize != 0) {
2583 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
2585 SDValue FIN = DAG.getFrameIndex(GPRIdx, PtrVT);
2587 for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
2588 unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &AArch64::GPR64RegClass);
2589 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
2590 SDValue Store = DAG.getStore(
2591 Val.getValue(1), DL, Val, FIN,
2592 MachinePointerInfo::getStack(DAG.getMachineFunction(), i * 8), false,
2594 MemOps.push_back(Store);
2596 DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getConstant(8, DL, PtrVT));
2599 FuncInfo->setVarArgsGPRIndex(GPRIdx);
2600 FuncInfo->setVarArgsGPRSize(GPRSaveSize);
2602 if (Subtarget->hasFPARMv8()) {
2603 static const MCPhysReg FPRArgRegs[] = {
2604 AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3,
2605 AArch64::Q4, AArch64::Q5, AArch64::Q6, AArch64::Q7};
2606 static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs);
2607 unsigned FirstVariadicFPR = CCInfo.getFirstUnallocated(FPRArgRegs);
2609 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
2611 if (FPRSaveSize != 0) {
2612 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
2614 SDValue FIN = DAG.getFrameIndex(FPRIdx, PtrVT);
2616 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
2617 unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &AArch64::FPR128RegClass);
2618 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f128);
2620 SDValue Store = DAG.getStore(
2621 Val.getValue(1), DL, Val, FIN,
2622 MachinePointerInfo::getStack(DAG.getMachineFunction(), i * 16),
2624 MemOps.push_back(Store);
2625 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN,
2626 DAG.getConstant(16, DL, PtrVT));
2629 FuncInfo->setVarArgsFPRIndex(FPRIdx);
2630 FuncInfo->setVarArgsFPRSize(FPRSaveSize);
2633 if (!MemOps.empty()) {
2634 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
2638 /// LowerCallResult - Lower the result values of a call into the
2639 /// appropriate copies out of appropriate physical registers.
2640 SDValue AArch64TargetLowering::LowerCallResult(
2641 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
2642 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2643 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
2644 SDValue ThisVal) const {
2645 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2646 ? RetCC_AArch64_WebKit_JS
2647 : RetCC_AArch64_AAPCS;
2648 // Assign locations to each value returned by this call.
2649 SmallVector<CCValAssign, 16> RVLocs;
2650 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2652 CCInfo.AnalyzeCallResult(Ins, RetCC);
2654 // Copy all of the result registers out of their specified physreg.
2655 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2656 CCValAssign VA = RVLocs[i];
2658 // Pass 'this' value directly from the argument to return value, to avoid
2659 // reg unit interference
2660 if (i == 0 && isThisReturn) {
2661 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 &&
2662 "unexpected return calling convention register assignment");
2663 InVals.push_back(ThisVal);
2668 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag);
2669 Chain = Val.getValue(1);
2670 InFlag = Val.getValue(2);
2672 switch (VA.getLocInfo()) {
2674 llvm_unreachable("Unknown loc info!");
2675 case CCValAssign::Full:
2677 case CCValAssign::BCvt:
2678 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
2682 InVals.push_back(Val);
2688 bool AArch64TargetLowering::isEligibleForTailCallOptimization(
2689 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
2690 bool isCalleeStructRet, bool isCallerStructRet,
2691 const SmallVectorImpl<ISD::OutputArg> &Outs,
2692 const SmallVectorImpl<SDValue> &OutVals,
2693 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
2694 // For CallingConv::C this function knows whether the ABI needs
2695 // changing. That's not true for other conventions so they will have to opt in
2697 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
2700 const MachineFunction &MF = DAG.getMachineFunction();
2701 const Function *CallerF = MF.getFunction();
2702 CallingConv::ID CallerCC = CallerF->getCallingConv();
2703 bool CCMatch = CallerCC == CalleeCC;
2705 // Byval parameters hand the function a pointer directly into the stack area
2706 // we want to reuse during a tail call. Working around this *is* possible (see
2707 // X86) but less efficient and uglier in LowerCall.
2708 for (Function::const_arg_iterator i = CallerF->arg_begin(),
2709 e = CallerF->arg_end();
2711 if (i->hasByValAttr())
2714 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
2715 if (IsTailCallConvention(CalleeCC) && CCMatch)
2720 // Externally-defined functions with weak linkage should not be
2721 // tail-called on AArch64 when the OS does not support dynamic
2722 // pre-emption of symbols, as the AAELF spec requires normal calls
2723 // to undefined weak functions to be replaced with a NOP or jump to the
2724 // next instruction. The behaviour of branch instructions in this
2725 // situation (as used for tail calls) is implementation-defined, so we
2726 // cannot rely on the linker replacing the tail call with a return.
2727 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2728 const GlobalValue *GV = G->getGlobal();
2729 const Triple &TT = getTargetMachine().getTargetTriple();
2730 if (GV->hasExternalWeakLinkage() &&
2731 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2735 // Now we search for cases where we can use a tail call without changing the
2736 // ABI. Sibcall is used in some places (particularly gcc) to refer to this
2739 // I want anyone implementing a new calling convention to think long and hard
2740 // about this assert.
2741 assert((!isVarArg || CalleeCC == CallingConv::C) &&
2742 "Unexpected variadic calling convention");
2744 if (isVarArg && !Outs.empty()) {
2745 // At least two cases here: if caller is fastcc then we can't have any
2746 // memory arguments (we'd be expected to clean up the stack afterwards). If
2747 // caller is C then we could potentially use its argument area.
2749 // FIXME: for now we take the most conservative of these in both cases:
2750 // disallow all variadic memory operands.
2751 SmallVector<CCValAssign, 16> ArgLocs;
2752 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2755 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, true));
2756 for (const CCValAssign &ArgLoc : ArgLocs)
2757 if (!ArgLoc.isRegLoc())
2761 // If the calling conventions do not match, then we'd better make sure the
2762 // results are returned in the same way as what the caller expects.
2764 SmallVector<CCValAssign, 16> RVLocs1;
2765 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2767 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForCall(CalleeCC, isVarArg));
2769 SmallVector<CCValAssign, 16> RVLocs2;
2770 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2772 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForCall(CallerCC, isVarArg));
2774 if (RVLocs1.size() != RVLocs2.size())
2776 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2777 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2779 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2781 if (RVLocs1[i].isRegLoc()) {
2782 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2785 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2791 // Nothing more to check if the callee is taking no arguments
2795 SmallVector<CCValAssign, 16> ArgLocs;
2796 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2799 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, isVarArg));
2801 const AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2803 // If the stack arguments for this call would fit into our own save area then
2804 // the call can be made tail.
2805 return CCInfo.getNextStackOffset() <= FuncInfo->getBytesInStackArgArea();
2808 SDValue AArch64TargetLowering::addTokenForArgument(SDValue Chain,
2810 MachineFrameInfo *MFI,
2811 int ClobberedFI) const {
2812 SmallVector<SDValue, 8> ArgChains;
2813 int64_t FirstByte = MFI->getObjectOffset(ClobberedFI);
2814 int64_t LastByte = FirstByte + MFI->getObjectSize(ClobberedFI) - 1;
2816 // Include the original chain at the beginning of the list. When this is
2817 // used by target LowerCall hooks, this helps legalize find the
2818 // CALLSEQ_BEGIN node.
2819 ArgChains.push_back(Chain);
2821 // Add a chain value for each stack argument corresponding
2822 for (SDNode::use_iterator U = DAG.getEntryNode().getNode()->use_begin(),
2823 UE = DAG.getEntryNode().getNode()->use_end();
2825 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
2826 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
2827 if (FI->getIndex() < 0) {
2828 int64_t InFirstByte = MFI->getObjectOffset(FI->getIndex());
2829 int64_t InLastByte = InFirstByte;
2830 InLastByte += MFI->getObjectSize(FI->getIndex()) - 1;
2832 if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
2833 (FirstByte <= InFirstByte && InFirstByte <= LastByte))
2834 ArgChains.push_back(SDValue(L, 1));
2837 // Build a tokenfactor for all the chains.
2838 return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
2841 bool AArch64TargetLowering::DoesCalleeRestoreStack(CallingConv::ID CallCC,
2842 bool TailCallOpt) const {
2843 return CallCC == CallingConv::Fast && TailCallOpt;
2846 bool AArch64TargetLowering::IsTailCallConvention(CallingConv::ID CallCC) const {
2847 return CallCC == CallingConv::Fast;
2850 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
2851 /// and add input and output parameter nodes.
2853 AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
2854 SmallVectorImpl<SDValue> &InVals) const {
2855 SelectionDAG &DAG = CLI.DAG;
2857 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2858 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2859 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2860 SDValue Chain = CLI.Chain;
2861 SDValue Callee = CLI.Callee;
2862 bool &IsTailCall = CLI.IsTailCall;
2863 CallingConv::ID CallConv = CLI.CallConv;
2864 bool IsVarArg = CLI.IsVarArg;
2866 MachineFunction &MF = DAG.getMachineFunction();
2867 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
2868 bool IsThisReturn = false;
2870 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2871 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2872 bool IsSibCall = false;
2875 // Check if it's really possible to do a tail call.
2876 IsTailCall = isEligibleForTailCallOptimization(
2877 Callee, CallConv, IsVarArg, IsStructRet,
2878 MF.getFunction()->hasStructRetAttr(), Outs, OutVals, Ins, DAG);
2879 if (!IsTailCall && CLI.CS && CLI.CS->isMustTailCall())
2880 report_fatal_error("failed to perform tail call elimination on a call "
2881 "site marked musttail");
2883 // A sibling call is one where we're under the usual C ABI and not planning
2884 // to change that but can still do a tail call:
2885 if (!TailCallOpt && IsTailCall)
2892 // Analyze operands of the call, assigning locations to each operand.
2893 SmallVector<CCValAssign, 16> ArgLocs;
2894 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
2898 // Handle fixed and variable vector arguments differently.
2899 // Variable vector arguments always go into memory.
2900 unsigned NumArgs = Outs.size();
2902 for (unsigned i = 0; i != NumArgs; ++i) {
2903 MVT ArgVT = Outs[i].VT;
2904 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2905 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
2906 /*IsVarArg=*/ !Outs[i].IsFixed);
2907 bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
2908 assert(!Res && "Call operand has unhandled type");
2912 // At this point, Outs[].VT may already be promoted to i32. To correctly
2913 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2914 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
2915 // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here
2916 // we use a special version of AnalyzeCallOperands to pass in ValVT and
2918 unsigned NumArgs = Outs.size();
2919 for (unsigned i = 0; i != NumArgs; ++i) {
2920 MVT ValVT = Outs[i].VT;
2921 // Get type of the original argument.
2922 EVT ActualVT = getValueType(DAG.getDataLayout(),
2923 CLI.getArgs()[Outs[i].OrigArgIndex].Ty,
2924 /*AllowUnknown*/ true);
2925 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT;
2926 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2927 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2928 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2930 else if (ActualMVT == MVT::i16)
2933 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2934 bool Res = AssignFn(i, ValVT, ValVT, CCValAssign::Full, ArgFlags, CCInfo);
2935 assert(!Res && "Call operand has unhandled type");
2940 // Get a count of how many bytes are to be pushed on the stack.
2941 unsigned NumBytes = CCInfo.getNextStackOffset();
2944 // Since we're not changing the ABI to make this a tail call, the memory
2945 // operands are already available in the caller's incoming argument space.
2949 // FPDiff is the byte offset of the call's argument area from the callee's.
2950 // Stores to callee stack arguments will be placed in FixedStackSlots offset
2951 // by this amount for a tail call. In a sibling call it must be 0 because the
2952 // caller will deallocate the entire stack and the callee still expects its
2953 // arguments to begin at SP+0. Completely unused for non-tail calls.
2956 if (IsTailCall && !IsSibCall) {
2957 unsigned NumReusableBytes = FuncInfo->getBytesInStackArgArea();
2959 // Since callee will pop argument stack as a tail call, we must keep the
2960 // popped size 16-byte aligned.
2961 NumBytes = RoundUpToAlignment(NumBytes, 16);
2963 // FPDiff will be negative if this tail call requires more space than we
2964 // would automatically have in our incoming argument space. Positive if we
2965 // can actually shrink the stack.
2966 FPDiff = NumReusableBytes - NumBytes;
2968 // The stack pointer must be 16-byte aligned at all times it's used for a
2969 // memory operation, which in practice means at *all* times and in
2970 // particular across call boundaries. Therefore our own arguments started at
2971 // a 16-byte aligned SP and the delta applied for the tail call should
2972 // satisfy the same constraint.
2973 assert(FPDiff % 16 == 0 && "unaligned stack on tail call");
2976 // Adjust the stack pointer for the new arguments...
2977 // These operations are automatically eliminated by the prolog/epilog pass
2979 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, DL,
2983 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, AArch64::SP,
2984 getPointerTy(DAG.getDataLayout()));
2986 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2987 SmallVector<SDValue, 8> MemOpChains;
2988 auto PtrVT = getPointerTy(DAG.getDataLayout());
2990 // Walk the register/memloc assignments, inserting copies/loads.
2991 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2992 ++i, ++realArgIdx) {
2993 CCValAssign &VA = ArgLocs[i];
2994 SDValue Arg = OutVals[realArgIdx];
2995 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2997 // Promote the value if needed.
2998 switch (VA.getLocInfo()) {
3000 llvm_unreachable("Unknown loc info!");
3001 case CCValAssign::Full:
3003 case CCValAssign::SExt:
3004 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
3006 case CCValAssign::ZExt:
3007 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
3009 case CCValAssign::AExt:
3010 if (Outs[realArgIdx].ArgVT == MVT::i1) {
3011 // AAPCS requires i1 to be zero-extended to 8-bits by the caller.
3012 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
3013 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i8, Arg);
3015 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
3017 case CCValAssign::BCvt:
3018 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
3020 case CCValAssign::FPExt:
3021 Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
3025 if (VA.isRegLoc()) {
3026 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i64) {
3027 assert(VA.getLocVT() == MVT::i64 &&
3028 "unexpected calling convention register assignment");
3029 assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
3030 "unexpected use of 'returned'");
3031 IsThisReturn = true;
3033 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3035 assert(VA.isMemLoc());
3038 MachinePointerInfo DstInfo;
3040 // FIXME: This works on big-endian for composite byvals, which are the
3041 // common case. It should also work for fundamental types too.
3042 uint32_t BEAlign = 0;
3043 unsigned OpSize = Flags.isByVal() ? Flags.getByValSize() * 8
3044 : VA.getValVT().getSizeInBits();
3045 OpSize = (OpSize + 7) / 8;
3046 if (!Subtarget->isLittleEndian() && !Flags.isByVal() &&
3047 !Flags.isInConsecutiveRegs()) {
3049 BEAlign = 8 - OpSize;
3051 unsigned LocMemOffset = VA.getLocMemOffset();
3052 int32_t Offset = LocMemOffset + BEAlign;
3053 SDValue PtrOff = DAG.getIntPtrConstant(Offset, DL);
3054 PtrOff = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, PtrOff);
3057 Offset = Offset + FPDiff;
3058 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3060 DstAddr = DAG.getFrameIndex(FI, PtrVT);
3062 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI);
3064 // Make sure any stack arguments overlapping with where we're storing
3065 // are loaded before this eventual operation. Otherwise they'll be
3067 Chain = addTokenForArgument(Chain, DAG, MF.getFrameInfo(), FI);
3069 SDValue PtrOff = DAG.getIntPtrConstant(Offset, DL);
3071 DstAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, PtrOff);
3072 DstInfo = MachinePointerInfo::getStack(DAG.getMachineFunction(),
3076 if (Outs[i].Flags.isByVal()) {
3078 DAG.getConstant(Outs[i].Flags.getByValSize(), DL, MVT::i64);
3079 SDValue Cpy = DAG.getMemcpy(
3080 Chain, DL, DstAddr, Arg, SizeNode, Outs[i].Flags.getByValAlign(),
3081 /*isVol = */ false, /*AlwaysInline = */ false,
3082 /*isTailCall = */ false,
3083 DstInfo, MachinePointerInfo());
3085 MemOpChains.push_back(Cpy);
3087 // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
3088 // promoted to a legal register type i32, we should truncate Arg back to
3090 if (VA.getValVT() == MVT::i1 || VA.getValVT() == MVT::i8 ||
3091 VA.getValVT() == MVT::i16)
3092 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Arg);
3095 DAG.getStore(Chain, DL, Arg, DstAddr, DstInfo, false, false, 0);
3096 MemOpChains.push_back(Store);
3101 if (!MemOpChains.empty())
3102 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
3104 // Build a sequence of copy-to-reg nodes chained together with token chain
3105 // and flag operands which copy the outgoing args into the appropriate regs.
3107 for (auto &RegToPass : RegsToPass) {
3108 Chain = DAG.getCopyToReg(Chain, DL, RegToPass.first,
3109 RegToPass.second, InFlag);
3110 InFlag = Chain.getValue(1);
3113 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
3114 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
3115 // node so that legalize doesn't hack it.
3116 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3117 Subtarget->isTargetMachO()) {
3118 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3119 const GlobalValue *GV = G->getGlobal();
3120 bool InternalLinkage = GV->hasInternalLinkage();
3121 if (InternalLinkage)
3122 Callee = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 0);
3125 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_GOT);
3126 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, Callee);
3128 } else if (ExternalSymbolSDNode *S =
3129 dyn_cast<ExternalSymbolSDNode>(Callee)) {
3130 const char *Sym = S->getSymbol();
3131 Callee = DAG.getTargetExternalSymbol(Sym, PtrVT, AArch64II::MO_GOT);
3132 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, Callee);
3134 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3135 const GlobalValue *GV = G->getGlobal();
3136 Callee = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 0);
3137 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3138 const char *Sym = S->getSymbol();
3139 Callee = DAG.getTargetExternalSymbol(Sym, PtrVT, 0);
3142 // We don't usually want to end the call-sequence here because we would tidy
3143 // the frame up *after* the call, however in the ABI-changing tail-call case
3144 // we've carefully laid out the parameters so that when sp is reset they'll be
3145 // in the correct location.
3146 if (IsTailCall && !IsSibCall) {
3147 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, DL, true),
3148 DAG.getIntPtrConstant(0, DL, true), InFlag, DL);
3149 InFlag = Chain.getValue(1);
3152 std::vector<SDValue> Ops;
3153 Ops.push_back(Chain);
3154 Ops.push_back(Callee);
3157 // Each tail call may have to adjust the stack by a different amount, so
3158 // this information must travel along with the operation for eventual
3159 // consumption by emitEpilogue.
3160 Ops.push_back(DAG.getTargetConstant(FPDiff, DL, MVT::i32));
3163 // Add argument registers to the end of the list so that they are known live
3165 for (auto &RegToPass : RegsToPass)
3166 Ops.push_back(DAG.getRegister(RegToPass.first,
3167 RegToPass.second.getValueType()));
3169 // Add a register mask operand representing the call-preserved registers.
3170 const uint32_t *Mask;
3171 const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();
3173 // For 'this' returns, use the X0-preserving mask if applicable
3174 Mask = TRI->getThisReturnPreservedMask(MF, CallConv);
3176 IsThisReturn = false;
3177 Mask = TRI->getCallPreservedMask(MF, CallConv);
3180 Mask = TRI->getCallPreservedMask(MF, CallConv);
3182 assert(Mask && "Missing call preserved mask for calling convention");
3183 Ops.push_back(DAG.getRegisterMask(Mask));
3185 if (InFlag.getNode())
3186 Ops.push_back(InFlag);
3188 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3190 // If we're doing a tall call, use a TC_RETURN here rather than an
3191 // actual call instruction.
3193 MF.getFrameInfo()->setHasTailCall();
3194 return DAG.getNode(AArch64ISD::TC_RETURN, DL, NodeTys, Ops);
3197 // Returns a chain and a flag for retval copy to use.
3198 Chain = DAG.getNode(AArch64ISD::CALL, DL, NodeTys, Ops);
3199 InFlag = Chain.getValue(1);
3201 uint64_t CalleePopBytes = DoesCalleeRestoreStack(CallConv, TailCallOpt)
3202 ? RoundUpToAlignment(NumBytes, 16)
3205 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, DL, true),
3206 DAG.getIntPtrConstant(CalleePopBytes, DL, true),
3209 InFlag = Chain.getValue(1);
3211 // Handle result values, copying them out of physregs into vregs that we
3213 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
3214 InVals, IsThisReturn,
3215 IsThisReturn ? OutVals[0] : SDValue());
3218 bool AArch64TargetLowering::CanLowerReturn(
3219 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
3220 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
3221 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
3222 ? RetCC_AArch64_WebKit_JS
3223 : RetCC_AArch64_AAPCS;
3224 SmallVector<CCValAssign, 16> RVLocs;
3225 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
3226 return CCInfo.CheckReturn(Outs, RetCC);
3230 AArch64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
3232 const SmallVectorImpl<ISD::OutputArg> &Outs,
3233 const SmallVectorImpl<SDValue> &OutVals,
3234 SDLoc DL, SelectionDAG &DAG) const {
3235 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
3236 ? RetCC_AArch64_WebKit_JS
3237 : RetCC_AArch64_AAPCS;
3238 SmallVector<CCValAssign, 16> RVLocs;
3239 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
3241 CCInfo.AnalyzeReturn(Outs, RetCC);
3243 // Copy the result values into the output registers.
3245 SmallVector<SDValue, 4> RetOps(1, Chain);
3246 for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size();
3247 ++i, ++realRVLocIdx) {
3248 CCValAssign &VA = RVLocs[i];
3249 assert(VA.isRegLoc() && "Can only return in registers!");
3250 SDValue Arg = OutVals[realRVLocIdx];
3252 switch (VA.getLocInfo()) {
3254 llvm_unreachable("Unknown loc info!");
3255 case CCValAssign::Full:
3256 if (Outs[i].ArgVT == MVT::i1) {
3257 // AAPCS requires i1 to be zero-extended to i8 by the producer of the
3258 // value. This is strictly redundant on Darwin (which uses "zeroext
3259 // i1"), but will be optimised out before ISel.
3260 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
3261 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
3264 case CCValAssign::BCvt:
3265 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
3269 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
3270 Flag = Chain.getValue(1);
3271 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
3274 RetOps[0] = Chain; // Update chain.
3276 // Add the flag if we have it.
3278 RetOps.push_back(Flag);
3280 return DAG.getNode(AArch64ISD::RET_FLAG, DL, MVT::Other, RetOps);
3283 //===----------------------------------------------------------------------===//
3284 // Other Lowering Code
3285 //===----------------------------------------------------------------------===//
3287 SDValue AArch64TargetLowering::LowerGlobalAddress(SDValue Op,
3288 SelectionDAG &DAG) const {
3289 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3291 const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
3292 const GlobalValue *GV = GN->getGlobal();
3293 unsigned char OpFlags =
3294 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
3296 assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
3297 "unexpected offset in global node");
3299 // This also catched the large code model case for Darwin.
3300 if ((OpFlags & AArch64II::MO_GOT) != 0) {
3301 SDValue GotAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
3302 // FIXME: Once remat is capable of dealing with instructions with register
3303 // operands, expand this into two nodes instead of using a wrapper node.
3304 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
3307 if ((OpFlags & AArch64II::MO_CONSTPOOL) != 0) {
3308 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
3309 "use of MO_CONSTPOOL only supported on small model");
3310 SDValue Hi = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, AArch64II::MO_PAGE);
3311 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3312 unsigned char LoFlags = AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
3313 SDValue Lo = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, LoFlags);
3314 SDValue PoolAddr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3315 SDValue GlobalAddr = DAG.getLoad(
3316 PtrVT, DL, DAG.getEntryNode(), PoolAddr,
3317 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
3318 /*isVolatile=*/false,
3319 /*isNonTemporal=*/true,
3320 /*isInvariant=*/true, 8);
3321 if (GN->getOffset() != 0)
3322 return DAG.getNode(ISD::ADD, DL, PtrVT, GlobalAddr,
3323 DAG.getConstant(GN->getOffset(), DL, PtrVT));
3327 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
3328 const unsigned char MO_NC = AArch64II::MO_NC;
3330 AArch64ISD::WrapperLarge, DL, PtrVT,
3331 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G3),
3332 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
3333 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
3334 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
3336 // Use ADRP/ADD or ADRP/LDR for everything else: the small model on ELF and
3337 // the only correct model on Darwin.
3338 SDValue Hi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
3339 OpFlags | AArch64II::MO_PAGE);
3340 unsigned char LoFlags = OpFlags | AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
3341 SDValue Lo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, LoFlags);
3343 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3344 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3348 /// \brief Convert a TLS address reference into the correct sequence of loads
3349 /// and calls to compute the variable's address (for Darwin, currently) and
3350 /// return an SDValue containing the final node.
3352 /// Darwin only has one TLS scheme which must be capable of dealing with the
3353 /// fully general situation, in the worst case. This means:
3354 /// + "extern __thread" declaration.
3355 /// + Defined in a possibly unknown dynamic library.
3357 /// The general system is that each __thread variable has a [3 x i64] descriptor
3358 /// which contains information used by the runtime to calculate the address. The
3359 /// only part of this the compiler needs to know about is the first xword, which
3360 /// contains a function pointer that must be called with the address of the
3361 /// entire descriptor in "x0".
3363 /// Since this descriptor may be in a different unit, in general even the
3364 /// descriptor must be accessed via an indirect load. The "ideal" code sequence
3366 /// adrp x0, _var@TLVPPAGE
3367 /// ldr x0, [x0, _var@TLVPPAGEOFF] ; x0 now contains address of descriptor
3368 /// ldr x1, [x0] ; x1 contains 1st entry of descriptor,
3369 /// ; the function pointer
3370 /// blr x1 ; Uses descriptor address in x0
3371 /// ; Address of _var is now in x0.
3373 /// If the address of _var's descriptor *is* known to the linker, then it can
3374 /// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
3375 /// a slight efficiency gain.
3377 AArch64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
3378 SelectionDAG &DAG) const {
3379 assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");
3382 MVT PtrVT = getPointerTy(DAG.getDataLayout());
3383 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3386 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3387 SDValue DescAddr = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TLVPAddr);
3389 // The first entry in the descriptor is a function pointer that we must call
3390 // to obtain the address of the variable.
3391 SDValue Chain = DAG.getEntryNode();
3392 SDValue FuncTLVGet =
3393 DAG.getLoad(MVT::i64, DL, Chain, DescAddr,
3394 MachinePointerInfo::getGOT(DAG.getMachineFunction()), false,
3396 Chain = FuncTLVGet.getValue(1);
3398 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3399 MFI->setAdjustsStack(true);
3401 // TLS calls preserve all registers except those that absolutely must be
3402 // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
3404 const uint32_t *Mask =
3405 Subtarget->getRegisterInfo()->getTLSCallPreservedMask();
3407 // Finally, we can make the call. This is just a degenerate version of a
3408 // normal AArch64 call node: x0 takes the address of the descriptor, and
3409 // returns the address of the variable in this thread.
3410 Chain = DAG.getCopyToReg(Chain, DL, AArch64::X0, DescAddr, SDValue());
3412 DAG.getNode(AArch64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
3413 Chain, FuncTLVGet, DAG.getRegister(AArch64::X0, MVT::i64),
3414 DAG.getRegisterMask(Mask), Chain.getValue(1));
3415 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Chain.getValue(1));
3418 /// When accessing thread-local variables under either the general-dynamic or
3419 /// local-dynamic system, we make a "TLS-descriptor" call. The variable will
3420 /// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
3421 /// is a function pointer to carry out the resolution.
3423 /// The sequence is:
3424 /// adrp x0, :tlsdesc:var
3425 /// ldr x1, [x0, #:tlsdesc_lo12:var]
3426 /// add x0, x0, #:tlsdesc_lo12:var
3427 /// .tlsdesccall var
3429 /// (TPIDR_EL0 offset now in x0)
3431 /// The above sequence must be produced unscheduled, to enable the linker to
3432 /// optimize/relax this sequence.
3433 /// Therefore, a pseudo-instruction (TLSDESC_CALLSEQ) is used to represent the
3434 /// above sequence, and expanded really late in the compilation flow, to ensure
3435 /// the sequence is produced as per above.
3436 SDValue AArch64TargetLowering::LowerELFTLSDescCallSeq(SDValue SymAddr, SDLoc DL,
3437 SelectionDAG &DAG) const {
3438 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3440 SDValue Chain = DAG.getEntryNode();
3441 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3443 SmallVector<SDValue, 2> Ops;
3444 Ops.push_back(Chain);
3445 Ops.push_back(SymAddr);
3447 Chain = DAG.getNode(AArch64ISD::TLSDESC_CALLSEQ, DL, NodeTys, Ops);
3448 SDValue Glue = Chain.getValue(1);
3450 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Glue);
3454 AArch64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
3455 SelectionDAG &DAG) const {
3456 assert(Subtarget->isTargetELF() && "This function expects an ELF target");
3457 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
3458 "ELF TLS only supported in small memory model");
3459 // Different choices can be made for the maximum size of the TLS area for a
3460 // module. For the small address model, the default TLS size is 16MiB and the
3461 // maximum TLS size is 4GiB.
3462 // FIXME: add -mtls-size command line option and make it control the 16MiB
3463 // vs. 4GiB code sequence generation.
3464 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
3466 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
3468 if (DAG.getTarget().Options.EmulatedTLS)
3469 return LowerToTLSEmulatedModel(GA, DAG);
3471 if (!EnableAArch64ELFLocalDynamicTLSGeneration) {
3472 if (Model == TLSModel::LocalDynamic)
3473 Model = TLSModel::GeneralDynamic;
3477 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3479 const GlobalValue *GV = GA->getGlobal();
3481 SDValue ThreadBase = DAG.getNode(AArch64ISD::THREAD_POINTER, DL, PtrVT);
3483 if (Model == TLSModel::LocalExec) {
3484 SDValue HiVar = DAG.getTargetGlobalAddress(
3485 GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_HI12);
3486 SDValue LoVar = DAG.getTargetGlobalAddress(
3488 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3490 SDValue TPWithOff_lo =
3491 SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, ThreadBase,
3493 DAG.getTargetConstant(0, DL, MVT::i32)),
3496 SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPWithOff_lo,
3498 DAG.getTargetConstant(0, DL, MVT::i32)),
3501 } else if (Model == TLSModel::InitialExec) {
3502 TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3503 TPOff = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TPOff);
3504 } else if (Model == TLSModel::LocalDynamic) {
3505 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
3506 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
3507 // the beginning of the module's TLS region, followed by a DTPREL offset
3510 // These accesses will need deduplicating if there's more than one.
3511 AArch64FunctionInfo *MFI =
3512 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
3513 MFI->incNumLocalDynamicTLSAccesses();
3515 // The call needs a relocation too for linker relaxation. It doesn't make
3516 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
3518 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
3521 // Now we can calculate the offset from TPIDR_EL0 to this module's
3522 // thread-local area.
3523 TPOff = LowerELFTLSDescCallSeq(SymAddr, DL, DAG);
3525 // Now use :dtprel_whatever: operations to calculate this variable's offset
3526 // in its thread-storage area.
3527 SDValue HiVar = DAG.getTargetGlobalAddress(
3528 GV, DL, MVT::i64, 0, AArch64II::MO_TLS | AArch64II::MO_HI12);
3529 SDValue LoVar = DAG.getTargetGlobalAddress(
3530 GV, DL, MVT::i64, 0,
3531 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3533 TPOff = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPOff, HiVar,
3534 DAG.getTargetConstant(0, DL, MVT::i32)),
3536 TPOff = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPOff, LoVar,
3537 DAG.getTargetConstant(0, DL, MVT::i32)),
3539 } else if (Model == TLSModel::GeneralDynamic) {
3540 // The call needs a relocation too for linker relaxation. It doesn't make
3541 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
3544 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
3546 // Finally we can make a call to calculate the offset from tpidr_el0.
3547 TPOff = LowerELFTLSDescCallSeq(SymAddr, DL, DAG);
3549 llvm_unreachable("Unsupported ELF TLS access model");
3551 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
3554 SDValue AArch64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
3555 SelectionDAG &DAG) const {
3556 if (Subtarget->isTargetDarwin())
3557 return LowerDarwinGlobalTLSAddress(Op, DAG);
3558 else if (Subtarget->isTargetELF())
3559 return LowerELFGlobalTLSAddress(Op, DAG);
3561 llvm_unreachable("Unexpected platform trying to use TLS");
3563 SDValue AArch64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3564 SDValue Chain = Op.getOperand(0);
3565 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3566 SDValue LHS = Op.getOperand(2);
3567 SDValue RHS = Op.getOperand(3);
3568 SDValue Dest = Op.getOperand(4);
3571 // Handle f128 first, since lowering it will result in comparing the return
3572 // value of a libcall against zero, which is just what the rest of LowerBR_CC
3573 // is expecting to deal with.
3574 if (LHS.getValueType() == MVT::f128) {
3575 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3577 // If softenSetCCOperands returned a scalar, we need to compare the result
3578 // against zero to select between true and false values.
3579 if (!RHS.getNode()) {
3580 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3585 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
3587 unsigned Opc = LHS.getOpcode();
3588 if (LHS.getResNo() == 1 && isa<ConstantSDNode>(RHS) &&
3589 cast<ConstantSDNode>(RHS)->isOne() &&
3590 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3591 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3592 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
3593 "Unexpected condition code.");
3594 // Only lower legal XALUO ops.
3595 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
3598 // The actual operation with overflow check.
3599 AArch64CC::CondCode OFCC;
3600 SDValue Value, Overflow;
3601 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, LHS.getValue(0), DAG);
3603 if (CC == ISD::SETNE)
3604 OFCC = getInvertedCondCode(OFCC);
3605 SDValue CCVal = DAG.getConstant(OFCC, dl, MVT::i32);
3607 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3611 if (LHS.getValueType().isInteger()) {
3612 assert((LHS.getValueType() == RHS.getValueType()) &&
3613 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3615 // If the RHS of the comparison is zero, we can potentially fold this
3616 // to a specialized branch.
3617 const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
3618 if (RHSC && RHSC->getZExtValue() == 0) {
3619 if (CC == ISD::SETEQ) {
3620 // See if we can use a TBZ to fold in an AND as well.
3621 // TBZ has a smaller branch displacement than CBZ. If the offset is
3622 // out of bounds, a late MI-layer pass rewrites branches.
3623 // 403.gcc is an example that hits this case.
3624 if (LHS.getOpcode() == ISD::AND &&
3625 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3626 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3627 SDValue Test = LHS.getOperand(0);
3628 uint64_t Mask = LHS.getConstantOperandVal(1);
3629 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, Test,
3630 DAG.getConstant(Log2_64(Mask), dl, MVT::i64),
3634 return DAG.getNode(AArch64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest);
3635 } else if (CC == ISD::SETNE) {
3636 // See if we can use a TBZ to fold in an AND as well.
3637 // TBZ has a smaller branch displacement than CBZ. If the offset is
3638 // out of bounds, a late MI-layer pass rewrites branches.
3639 // 403.gcc is an example that hits this case.
3640 if (LHS.getOpcode() == ISD::AND &&
3641 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3642 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3643 SDValue Test = LHS.getOperand(0);
3644 uint64_t Mask = LHS.getConstantOperandVal(1);
3645 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, Test,
3646 DAG.getConstant(Log2_64(Mask), dl, MVT::i64),
3650 return DAG.getNode(AArch64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest);
3651 } else if (CC == ISD::SETLT && LHS.getOpcode() != ISD::AND) {
3652 // Don't combine AND since emitComparison converts the AND to an ANDS
3653 // (a.k.a. TST) and the test in the test bit and branch instruction
3654 // becomes redundant. This would also increase register pressure.
3655 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
3656 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, LHS,
3657 DAG.getConstant(Mask, dl, MVT::i64), Dest);
3660 if (RHSC && RHSC->getSExtValue() == -1 && CC == ISD::SETGT &&
3661 LHS.getOpcode() != ISD::AND) {
3662 // Don't combine AND since emitComparison converts the AND to an ANDS
3663 // (a.k.a. TST) and the test in the test bit and branch instruction
3664 // becomes redundant. This would also increase register pressure.
3665 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
3666 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, LHS,
3667 DAG.getConstant(Mask, dl, MVT::i64), Dest);
3671 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3672 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3676 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3678 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
3679 // clean. Some of them require two branches to implement.
3680 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3681 AArch64CC::CondCode CC1, CC2;
3682 changeFPCCToAArch64CC(CC, CC1, CC2);
3683 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
3685 DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp);
3686 if (CC2 != AArch64CC::AL) {
3687 SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32);
3688 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val,
3695 SDValue AArch64TargetLowering::LowerFCOPYSIGN(SDValue Op,
3696 SelectionDAG &DAG) const {
3697 EVT VT = Op.getValueType();
3700 SDValue In1 = Op.getOperand(0);
3701 SDValue In2 = Op.getOperand(1);
3702 EVT SrcVT = In2.getValueType();
3704 if (SrcVT.bitsLT(VT))
3705 In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2);
3706 else if (SrcVT.bitsGT(VT))
3707 In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, DAG.getIntPtrConstant(0, DL));
3712 SDValue VecVal1, VecVal2;
3713 if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) {
3715 VecVT = (VT == MVT::v2f32 ? MVT::v2i32 : MVT::v4i32);
3716 EltMask = 0x80000000ULL;
3718 if (!VT.isVector()) {
3719 VecVal1 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3720 DAG.getUNDEF(VecVT), In1);
3721 VecVal2 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3722 DAG.getUNDEF(VecVT), In2);
3724 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3725 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3727 } else if (VT == MVT::f64 || VT == MVT::v2f64) {
3731 // We want to materialize a mask with the high bit set, but the AdvSIMD
3732 // immediate moves cannot materialize that in a single instruction for
3733 // 64-bit elements. Instead, materialize zero and then negate it.
3736 if (!VT.isVector()) {
3737 VecVal1 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3738 DAG.getUNDEF(VecVT), In1);
3739 VecVal2 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3740 DAG.getUNDEF(VecVT), In2);
3742 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3743 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3746 llvm_unreachable("Invalid type for copysign!");
3749 SDValue BuildVec = DAG.getConstant(EltMask, DL, VecVT);
3751 // If we couldn't materialize the mask above, then the mask vector will be
3752 // the zero vector, and we need to negate it here.
3753 if (VT == MVT::f64 || VT == MVT::v2f64) {
3754 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec);
3755 BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec);
3756 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec);
3760 DAG.getNode(AArch64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec);
3763 return DAG.getTargetExtractSubreg(AArch64::ssub, DL, VT, Sel);
3764 else if (VT == MVT::f64)
3765 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, VT, Sel);
3767 return DAG.getNode(ISD::BITCAST, DL, VT, Sel);
3770 SDValue AArch64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
3771 if (DAG.getMachineFunction().getFunction()->hasFnAttribute(
3772 Attribute::NoImplicitFloat))
3775 if (!Subtarget->hasNEON())
3778 // While there is no integer popcount instruction, it can
3779 // be more efficiently lowered to the following sequence that uses
3780 // AdvSIMD registers/instructions as long as the copies to/from
3781 // the AdvSIMD registers are cheap.
3782 // FMOV D0, X0 // copy 64-bit int to vector, high bits zero'd
3783 // CNT V0.8B, V0.8B // 8xbyte pop-counts
3784 // ADDV B0, V0.8B // sum 8xbyte pop-counts
3785 // UMOV X0, V0.B[0] // copy byte result back to integer reg
3786 SDValue Val = Op.getOperand(0);
3788 EVT VT = Op.getValueType();
3791 Val = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Val);
3792 Val = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);
3794 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, Val);
3795 SDValue UaddLV = DAG.getNode(
3796 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
3797 DAG.getConstant(Intrinsic::aarch64_neon_uaddlv, DL, MVT::i32), CtPop);
3800 UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV);
3804 SDValue AArch64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
3806 if (Op.getValueType().isVector())
3807 return LowerVSETCC(Op, DAG);
3809 SDValue LHS = Op.getOperand(0);
3810 SDValue RHS = Op.getOperand(1);
3811 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
3814 // We chose ZeroOrOneBooleanContents, so use zero and one.
3815 EVT VT = Op.getValueType();
3816 SDValue TVal = DAG.getConstant(1, dl, VT);
3817 SDValue FVal = DAG.getConstant(0, dl, VT);
3819 // Handle f128 first, since one possible outcome is a normal integer
3820 // comparison which gets picked up by the next if statement.
3821 if (LHS.getValueType() == MVT::f128) {
3822 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3824 // If softenSetCCOperands returned a scalar, use it.
3825 if (!RHS.getNode()) {
3826 assert(LHS.getValueType() == Op.getValueType() &&
3827 "Unexpected setcc expansion!");
3832 if (LHS.getValueType().isInteger()) {
3835 getAArch64Cmp(LHS, RHS, ISD::getSetCCInverse(CC, true), CCVal, DAG, dl);
3837 // Note that we inverted the condition above, so we reverse the order of
3838 // the true and false operands here. This will allow the setcc to be
3839 // matched to a single CSINC instruction.
3840 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp);
3843 // Now we know we're dealing with FP values.
3844 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3846 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3847 // and do the comparison.
3848 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3850 AArch64CC::CondCode CC1, CC2;
3851 changeFPCCToAArch64CC(CC, CC1, CC2);
3852 if (CC2 == AArch64CC::AL) {
3853 changeFPCCToAArch64CC(ISD::getSetCCInverse(CC, false), CC1, CC2);
3854 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
3856 // Note that we inverted the condition above, so we reverse the order of
3857 // the true and false operands here. This will allow the setcc to be
3858 // matched to a single CSINC instruction.
3859 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp);
3861 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't
3862 // totally clean. Some of them require two CSELs to implement. As is in
3863 // this case, we emit the first CSEL and then emit a second using the output
3864 // of the first as the RHS. We're effectively OR'ing the two CC's together.
3866 // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
3867 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
3869 DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3871 SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32);
3872 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3876 SDValue AArch64TargetLowering::LowerSELECT_CC(ISD::CondCode CC, SDValue LHS,
3877 SDValue RHS, SDValue TVal,
3878 SDValue FVal, SDLoc dl,
3879 SelectionDAG &DAG) const {
3880 // Handle f128 first, because it will result in a comparison of some RTLIB
3881 // call result against zero.
3882 if (LHS.getValueType() == MVT::f128) {
3883 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3885 // If softenSetCCOperands returned a scalar, we need to compare the result
3886 // against zero to select between true and false values.
3887 if (!RHS.getNode()) {
3888 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3893 // Handle integers first.
3894 if (LHS.getValueType().isInteger()) {
3895 assert((LHS.getValueType() == RHS.getValueType()) &&
3896 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3898 unsigned Opcode = AArch64ISD::CSEL;
3900 // If both the TVal and the FVal are constants, see if we can swap them in
3901 // order to for a CSINV or CSINC out of them.
3902 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
3903 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
3905 if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) {
3906 std::swap(TVal, FVal);
3907 std::swap(CTVal, CFVal);
3908 CC = ISD::getSetCCInverse(CC, true);
3909 } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) {
3910 std::swap(TVal, FVal);
3911 std::swap(CTVal, CFVal);
3912 CC = ISD::getSetCCInverse(CC, true);
3913 } else if (TVal.getOpcode() == ISD::XOR) {
3914 // If TVal is a NOT we want to swap TVal and FVal so that we can match
3915 // with a CSINV rather than a CSEL.
3916 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(1));
3918 if (CVal && CVal->isAllOnesValue()) {
3919 std::swap(TVal, FVal);
3920 std::swap(CTVal, CFVal);
3921 CC = ISD::getSetCCInverse(CC, true);
3923 } else if (TVal.getOpcode() == ISD::SUB) {
3924 // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
3925 // that we can match with a CSNEG rather than a CSEL.
3926 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(0));
3928 if (CVal && CVal->isNullValue()) {
3929 std::swap(TVal, FVal);
3930 std::swap(CTVal, CFVal);
3931 CC = ISD::getSetCCInverse(CC, true);
3933 } else if (CTVal && CFVal) {
3934 const int64_t TrueVal = CTVal->getSExtValue();
3935 const int64_t FalseVal = CFVal->getSExtValue();
3938 // If both TVal and FVal are constants, see if FVal is the
3939 // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
3940 // instead of a CSEL in that case.
3941 if (TrueVal == ~FalseVal) {
3942 Opcode = AArch64ISD::CSINV;
3943 } else if (TrueVal == -FalseVal) {
3944 Opcode = AArch64ISD::CSNEG;
3945 } else if (TVal.getValueType() == MVT::i32) {
3946 // If our operands are only 32-bit wide, make sure we use 32-bit
3947 // arithmetic for the check whether we can use CSINC. This ensures that
3948 // the addition in the check will wrap around properly in case there is
3949 // an overflow (which would not be the case if we do the check with
3950 // 64-bit arithmetic).
3951 const uint32_t TrueVal32 = CTVal->getZExtValue();
3952 const uint32_t FalseVal32 = CFVal->getZExtValue();
3954 if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
3955 Opcode = AArch64ISD::CSINC;
3957 if (TrueVal32 > FalseVal32) {
3961 // 64-bit check whether we can use CSINC.
3962 } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) {
3963 Opcode = AArch64ISD::CSINC;
3965 if (TrueVal > FalseVal) {
3970 // Swap TVal and FVal if necessary.
3972 std::swap(TVal, FVal);
3973 std::swap(CTVal, CFVal);
3974 CC = ISD::getSetCCInverse(CC, true);
3977 if (Opcode != AArch64ISD::CSEL) {
3978 // Drop FVal since we can get its value by simply inverting/negating
3985 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3987 EVT VT = TVal.getValueType();
3988 return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp);
3991 // Now we know we're dealing with FP values.
3992 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3993 assert(LHS.getValueType() == RHS.getValueType());
3994 EVT VT = TVal.getValueType();
3995 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3997 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
3998 // clean. Some of them require two CSELs to implement.
3999 AArch64CC::CondCode CC1, CC2;
4000 changeFPCCToAArch64CC(CC, CC1, CC2);
4001 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
4002 SDValue CS1 = DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
4004 // If we need a second CSEL, emit it, using the output of the first as the
4005 // RHS. We're effectively OR'ing the two CC's together.
4006 if (CC2 != AArch64CC::AL) {
4007 SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32);
4008 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
4011 // Otherwise, return the output of the first CSEL.
4015 SDValue AArch64TargetLowering::LowerSELECT_CC(SDValue Op,
4016 SelectionDAG &DAG) const {
4017 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
4018 SDValue LHS = Op.getOperand(0);
4019 SDValue RHS = Op.getOperand(1);
4020 SDValue TVal = Op.getOperand(2);
4021 SDValue FVal = Op.getOperand(3);
4023 return LowerSELECT_CC(CC, LHS, RHS, TVal, FVal, DL, DAG);
4026 SDValue AArch64TargetLowering::LowerSELECT(SDValue Op,
4027 SelectionDAG &DAG) const {
4028 SDValue CCVal = Op->getOperand(0);
4029 SDValue TVal = Op->getOperand(1);
4030 SDValue FVal = Op->getOperand(2);
4033 unsigned Opc = CCVal.getOpcode();
4034 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
4036 if (CCVal.getResNo() == 1 &&
4037 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
4038 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
4039 // Only lower legal XALUO ops.
4040 if (!DAG.getTargetLoweringInfo().isTypeLegal(CCVal->getValueType(0)))
4043 AArch64CC::CondCode OFCC;
4044 SDValue Value, Overflow;
4045 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, CCVal.getValue(0), DAG);
4046 SDValue CCVal = DAG.getConstant(OFCC, DL, MVT::i32);
4048 return DAG.getNode(AArch64ISD::CSEL, DL, Op.getValueType(), TVal, FVal,
4052 // Lower it the same way as we would lower a SELECT_CC node.
4055 if (CCVal.getOpcode() == ISD::SETCC) {
4056 LHS = CCVal.getOperand(0);
4057 RHS = CCVal.getOperand(1);
4058 CC = cast<CondCodeSDNode>(CCVal->getOperand(2))->get();
4061 RHS = DAG.getConstant(0, DL, CCVal.getValueType());
4064 return LowerSELECT_CC(CC, LHS, RHS, TVal, FVal, DL, DAG);
4067 SDValue AArch64TargetLowering::LowerJumpTable(SDValue Op,
4068 SelectionDAG &DAG) const {
4069 // Jump table entries as PC relative offsets. No additional tweaking
4070 // is necessary here. Just get the address of the jump table.
4071 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4072 EVT PtrVT = getPointerTy(DAG.getDataLayout());
4075 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
4076 !Subtarget->isTargetMachO()) {
4077 const unsigned char MO_NC = AArch64II::MO_NC;
4079 AArch64ISD::WrapperLarge, DL, PtrVT,
4080 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G3),
4081 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G2 | MO_NC),
4082 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G1 | MO_NC),
4083 DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
4084 AArch64II::MO_G0 | MO_NC));
4088 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_PAGE);
4089 SDValue Lo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
4090 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
4091 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
4092 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
4095 SDValue AArch64TargetLowering::LowerConstantPool(SDValue Op,
4096 SelectionDAG &DAG) const {
4097 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4098 EVT PtrVT = getPointerTy(DAG.getDataLayout());
4101 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
4102 // Use the GOT for the large code model on iOS.
4103 if (Subtarget->isTargetMachO()) {
4104 SDValue GotAddr = DAG.getTargetConstantPool(
4105 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
4107 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
4110 const unsigned char MO_NC = AArch64II::MO_NC;
4112 AArch64ISD::WrapperLarge, DL, PtrVT,
4113 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4114 CP->getOffset(), AArch64II::MO_G3),
4115 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4116 CP->getOffset(), AArch64II::MO_G2 | MO_NC),
4117 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4118 CP->getOffset(), AArch64II::MO_G1 | MO_NC),
4119 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4120 CP->getOffset(), AArch64II::MO_G0 | MO_NC));
4122 // Use ADRP/ADD or ADRP/LDR for everything else: the small memory model on
4123 // ELF, the only valid one on Darwin.
4125 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
4126 CP->getOffset(), AArch64II::MO_PAGE);
4127 SDValue Lo = DAG.getTargetConstantPool(
4128 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
4129 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
4131 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
4132 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
4136 SDValue AArch64TargetLowering::LowerBlockAddress(SDValue Op,
4137 SelectionDAG &DAG) const {
4138 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
4139 EVT PtrVT = getPointerTy(DAG.getDataLayout());
4141 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
4142 !Subtarget->isTargetMachO()) {
4143 const unsigned char MO_NC = AArch64II::MO_NC;
4145 AArch64ISD::WrapperLarge, DL, PtrVT,
4146 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G3),
4147 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
4148 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
4149 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
4151 SDValue Hi = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGE);
4152 SDValue Lo = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGEOFF |
4154 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
4155 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
4159 SDValue AArch64TargetLowering::LowerDarwin_VASTART(SDValue Op,
4160 SelectionDAG &DAG) const {
4161 AArch64FunctionInfo *FuncInfo =
4162 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
4165 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(),
4166 getPointerTy(DAG.getDataLayout()));
4167 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
4168 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
4169 MachinePointerInfo(SV), false, false, 0);
4172 SDValue AArch64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
4173 SelectionDAG &DAG) const {
4174 // The layout of the va_list struct is specified in the AArch64 Procedure Call
4175 // Standard, section B.3.
4176 MachineFunction &MF = DAG.getMachineFunction();
4177 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
4178 auto PtrVT = getPointerTy(DAG.getDataLayout());
4181 SDValue Chain = Op.getOperand(0);
4182 SDValue VAList = Op.getOperand(1);
4183 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
4184 SmallVector<SDValue, 4> MemOps;
4186 // void *__stack at offset 0
4187 SDValue Stack = DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), PtrVT);
4188 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
4189 MachinePointerInfo(SV), false, false, 8));
4191 // void *__gr_top at offset 8
4192 int GPRSize = FuncInfo->getVarArgsGPRSize();
4194 SDValue GRTop, GRTopAddr;
4197 DAG.getNode(ISD::ADD, DL, PtrVT, VAList, DAG.getConstant(8, DL, PtrVT));
4199 GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), PtrVT);
4200 GRTop = DAG.getNode(ISD::ADD, DL, PtrVT, GRTop,
4201 DAG.getConstant(GPRSize, DL, PtrVT));
4203 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
4204 MachinePointerInfo(SV, 8), false, false, 8));
4207 // void *__vr_top at offset 16
4208 int FPRSize = FuncInfo->getVarArgsFPRSize();
4210 SDValue VRTop, VRTopAddr;
4211 VRTopAddr = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
4212 DAG.getConstant(16, DL, PtrVT));
4214 VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), PtrVT);
4215 VRTop = DAG.getNode(ISD::ADD, DL, PtrVT, VRTop,
4216 DAG.getConstant(FPRSize, DL, PtrVT));
4218 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
4219 MachinePointerInfo(SV, 16), false, false, 8));
4222 // int __gr_offs at offset 24
4223 SDValue GROffsAddr =
4224 DAG.getNode(ISD::ADD, DL, PtrVT, VAList, DAG.getConstant(24, DL, PtrVT));
4225 MemOps.push_back(DAG.getStore(Chain, DL,
4226 DAG.getConstant(-GPRSize, DL, MVT::i32),
4227 GROffsAddr, MachinePointerInfo(SV, 24), false,
4230 // int __vr_offs at offset 28
4231 SDValue VROffsAddr =
4232 DAG.getNode(ISD::ADD, DL, PtrVT, VAList, DAG.getConstant(28, DL, PtrVT));
4233 MemOps.push_back(DAG.getStore(Chain, DL,
4234 DAG.getConstant(-FPRSize, DL, MVT::i32),
4235 VROffsAddr, MachinePointerInfo(SV, 28), false,
4238 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
4241 SDValue AArch64TargetLowering::LowerVASTART(SDValue Op,
4242 SelectionDAG &DAG) const {
4243 return Subtarget->isTargetDarwin() ? LowerDarwin_VASTART(Op, DAG)
4244 : LowerAAPCS_VASTART(Op, DAG);
4247 SDValue AArch64TargetLowering::LowerVACOPY(SDValue Op,
4248 SelectionDAG &DAG) const {
4249 // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
4252 unsigned VaListSize = Subtarget->isTargetDarwin() ? 8 : 32;
4253 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
4254 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
4256 return DAG.getMemcpy(Op.getOperand(0), DL, Op.getOperand(1),
4258 DAG.getConstant(VaListSize, DL, MVT::i32),
4259 8, false, false, false, MachinePointerInfo(DestSV),
4260 MachinePointerInfo(SrcSV));
4263 SDValue AArch64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
4264 assert(Subtarget->isTargetDarwin() &&
4265 "automatic va_arg instruction only works on Darwin");
4267 const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
4268 EVT VT = Op.getValueType();
4270 SDValue Chain = Op.getOperand(0);
4271 SDValue Addr = Op.getOperand(1);
4272 unsigned Align = Op.getConstantOperandVal(3);
4273 auto PtrVT = getPointerTy(DAG.getDataLayout());
4275 SDValue VAList = DAG.getLoad(PtrVT, DL, Chain, Addr, MachinePointerInfo(V),
4276 false, false, false, 0);
4277 Chain = VAList.getValue(1);
4280 assert(((Align & (Align - 1)) == 0) && "Expected Align to be a power of 2");
4281 VAList = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
4282 DAG.getConstant(Align - 1, DL, PtrVT));
4283 VAList = DAG.getNode(ISD::AND, DL, PtrVT, VAList,
4284 DAG.getConstant(-(int64_t)Align, DL, PtrVT));
4287 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
4288 uint64_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
4290 // Scalar integer and FP values smaller than 64 bits are implicitly extended
4291 // up to 64 bits. At the very least, we have to increase the striding of the
4292 // vaargs list to match this, and for FP values we need to introduce
4293 // FP_ROUND nodes as well.
4294 if (VT.isInteger() && !VT.isVector())
4296 bool NeedFPTrunc = false;
4297 if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
4302 // Increment the pointer, VAList, to the next vaarg
4303 SDValue VANext = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
4304 DAG.getConstant(ArgSize, DL, PtrVT));
4305 // Store the incremented VAList to the legalized pointer
4306 SDValue APStore = DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V),
4309 // Load the actual argument out of the pointer VAList
4311 // Load the value as an f64.
4312 SDValue WideFP = DAG.getLoad(MVT::f64, DL, APStore, VAList,
4313 MachinePointerInfo(), false, false, false, 0);
4314 // Round the value down to an f32.
4315 SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
4316 DAG.getIntPtrConstant(1, DL));
4317 SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
4318 // Merge the rounded value with the chain output of the load.
4319 return DAG.getMergeValues(Ops, DL);
4322 return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo(), false,
4326 SDValue AArch64TargetLowering::LowerFRAMEADDR(SDValue Op,
4327 SelectionDAG &DAG) const {
4328 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4329 MFI->setFrameAddressIsTaken(true);
4331 EVT VT = Op.getValueType();
4333 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4335 DAG.getCopyFromReg(DAG.getEntryNode(), DL, AArch64::FP, VT);
4337 FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
4338 MachinePointerInfo(), false, false, false, 0);
4342 // FIXME? Maybe this could be a TableGen attribute on some registers and
4343 // this table could be generated automatically from RegInfo.
4344 unsigned AArch64TargetLowering::getRegisterByName(const char* RegName, EVT VT,
4345 SelectionDAG &DAG) const {
4346 unsigned Reg = StringSwitch<unsigned>(RegName)
4347 .Case("sp", AArch64::SP)
4351 report_fatal_error(Twine("Invalid register name \""
4352 + StringRef(RegName) + "\"."));
4355 SDValue AArch64TargetLowering::LowerRETURNADDR(SDValue Op,
4356 SelectionDAG &DAG) const {
4357 MachineFunction &MF = DAG.getMachineFunction();
4358 MachineFrameInfo *MFI = MF.getFrameInfo();
4359 MFI->setReturnAddressIsTaken(true);
4361 EVT VT = Op.getValueType();
4363 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4365 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4366 SDValue Offset = DAG.getConstant(8, DL, getPointerTy(DAG.getDataLayout()));
4367 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
4368 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
4369 MachinePointerInfo(), false, false, false, 0);
4372 // Return LR, which contains the return address. Mark it an implicit live-in.
4373 unsigned Reg = MF.addLiveIn(AArch64::LR, &AArch64::GPR64RegClass);
4374 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
4377 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4378 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4379 SDValue AArch64TargetLowering::LowerShiftRightParts(SDValue Op,
4380 SelectionDAG &DAG) const {
4381 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4382 EVT VT = Op.getValueType();
4383 unsigned VTBits = VT.getSizeInBits();
4385 SDValue ShOpLo = Op.getOperand(0);
4386 SDValue ShOpHi = Op.getOperand(1);
4387 SDValue ShAmt = Op.getOperand(2);
4389 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4391 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4393 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4394 DAG.getConstant(VTBits, dl, MVT::i64), ShAmt);
4395 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4396 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4397 DAG.getConstant(VTBits, dl, MVT::i64));
4398 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4400 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, dl, MVT::i64),
4401 ISD::SETGE, dl, DAG);
4402 SDValue CCVal = DAG.getConstant(AArch64CC::GE, dl, MVT::i32);
4404 SDValue FalseValLo = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4405 SDValue TrueValLo = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4407 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
4409 // AArch64 shifts larger than the register width are wrapped rather than
4410 // clamped, so we can't just emit "hi >> x".
4411 SDValue FalseValHi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4412 SDValue TrueValHi = Opc == ISD::SRA
4413 ? DAG.getNode(Opc, dl, VT, ShOpHi,
4414 DAG.getConstant(VTBits - 1, dl,
4416 : DAG.getConstant(0, dl, VT);
4418 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValHi, FalseValHi, CCVal, Cmp);
4420 SDValue Ops[2] = { Lo, Hi };
4421 return DAG.getMergeValues(Ops, dl);
4424 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4425 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4426 SDValue AArch64TargetLowering::LowerShiftLeftParts(SDValue Op,
4427 SelectionDAG &DAG) const {
4428 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4429 EVT VT = Op.getValueType();
4430 unsigned VTBits = VT.getSizeInBits();
4432 SDValue ShOpLo = Op.getOperand(0);
4433 SDValue ShOpHi = Op.getOperand(1);
4434 SDValue ShAmt = Op.getOperand(2);
4437 assert(Op.getOpcode() == ISD::SHL_PARTS);
4438 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4439 DAG.getConstant(VTBits, dl, MVT::i64), ShAmt);
4440 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4441 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4442 DAG.getConstant(VTBits, dl, MVT::i64));
4443 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4444 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4446 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4448 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, dl, MVT::i64),
4449 ISD::SETGE, dl, DAG);
4450 SDValue CCVal = DAG.getConstant(AArch64CC::GE, dl, MVT::i32);
4452 DAG.getNode(AArch64ISD::CSEL, dl, VT, Tmp3, FalseVal, CCVal, Cmp);
4454 // AArch64 shifts of larger than register sizes are wrapped rather than
4455 // clamped, so we can't just emit "lo << a" if a is too big.
4456 SDValue TrueValLo = DAG.getConstant(0, dl, VT);
4457 SDValue FalseValLo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4459 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
4461 SDValue Ops[2] = { Lo, Hi };
4462 return DAG.getMergeValues(Ops, dl);
4465 bool AArch64TargetLowering::isOffsetFoldingLegal(
4466 const GlobalAddressSDNode *GA) const {
4467 // The AArch64 target doesn't support folding offsets into global addresses.
4471 bool AArch64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4472 // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
4473 // FIXME: We should be able to handle f128 as well with a clever lowering.
4474 if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32))
4478 return AArch64_AM::getFP64Imm(Imm) != -1;
4479 else if (VT == MVT::f32)
4480 return AArch64_AM::getFP32Imm(Imm) != -1;
4484 //===----------------------------------------------------------------------===//
4485 // AArch64 Optimization Hooks
4486 //===----------------------------------------------------------------------===//
4488 //===----------------------------------------------------------------------===//
4489 // AArch64 Inline Assembly Support
4490 //===----------------------------------------------------------------------===//
4492 // Table of Constraints
4493 // TODO: This is the current set of constraints supported by ARM for the
4494 // compiler, not all of them may make sense, e.g. S may be difficult to support.
4496 // r - A general register
4497 // w - An FP/SIMD register of some size in the range v0-v31
4498 // x - An FP/SIMD register of some size in the range v0-v15
4499 // I - Constant that can be used with an ADD instruction
4500 // J - Constant that can be used with a SUB instruction
4501 // K - Constant that can be used with a 32-bit logical instruction
4502 // L - Constant that can be used with a 64-bit logical instruction
4503 // M - Constant that can be used as a 32-bit MOV immediate
4504 // N - Constant that can be used as a 64-bit MOV immediate
4505 // Q - A memory reference with base register and no offset
4506 // S - A symbolic address
4507 // Y - Floating point constant zero
4508 // Z - Integer constant zero
4510 // Note that general register operands will be output using their 64-bit x
4511 // register name, whatever the size of the variable, unless the asm operand
4512 // is prefixed by the %w modifier. Floating-point and SIMD register operands
4513 // will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
4516 /// getConstraintType - Given a constraint letter, return the type of
4517 /// constraint it is for this target.
4518 AArch64TargetLowering::ConstraintType
4519 AArch64TargetLowering::getConstraintType(StringRef Constraint) const {
4520 if (Constraint.size() == 1) {
4521 switch (Constraint[0]) {
4528 return C_RegisterClass;
4529 // An address with a single base register. Due to the way we
4530 // currently handle addresses it is the same as 'r'.
4535 return TargetLowering::getConstraintType(Constraint);
4538 /// Examine constraint type and operand type and determine a weight value.
4539 /// This object must already have been set up with the operand type
4540 /// and the current alternative constraint selected.
4541 TargetLowering::ConstraintWeight
4542 AArch64TargetLowering::getSingleConstraintMatchWeight(
4543 AsmOperandInfo &info, const char *constraint) const {
4544 ConstraintWeight weight = CW_Invalid;
4545 Value *CallOperandVal = info.CallOperandVal;
4546 // If we don't have a value, we can't do a match,
4547 // but allow it at the lowest weight.
4548 if (!CallOperandVal)
4550 Type *type = CallOperandVal->getType();
4551 // Look at the constraint type.
4552 switch (*constraint) {
4554 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
4558 if (type->isFloatingPointTy() || type->isVectorTy())
4559 weight = CW_Register;
4562 weight = CW_Constant;
4568 std::pair<unsigned, const TargetRegisterClass *>
4569 AArch64TargetLowering::getRegForInlineAsmConstraint(
4570 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
4571 if (Constraint.size() == 1) {
4572 switch (Constraint[0]) {
4574 if (VT.getSizeInBits() == 64)
4575 return std::make_pair(0U, &AArch64::GPR64commonRegClass);
4576 return std::make_pair(0U, &AArch64::GPR32commonRegClass);
4579 return std::make_pair(0U, &AArch64::FPR32RegClass);
4580 if (VT.getSizeInBits() == 64)
4581 return std::make_pair(0U, &AArch64::FPR64RegClass);
4582 if (VT.getSizeInBits() == 128)
4583 return std::make_pair(0U, &AArch64::FPR128RegClass);
4585 // The instructions that this constraint is designed for can
4586 // only take 128-bit registers so just use that regclass.
4588 if (VT.getSizeInBits() == 128)
4589 return std::make_pair(0U, &AArch64::FPR128_loRegClass);
4593 if (StringRef("{cc}").equals_lower(Constraint))
4594 return std::make_pair(unsigned(AArch64::NZCV), &AArch64::CCRRegClass);
4596 // Use the default implementation in TargetLowering to convert the register
4597 // constraint into a member of a register class.
4598 std::pair<unsigned, const TargetRegisterClass *> Res;
4599 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
4601 // Not found as a standard register?
4603 unsigned Size = Constraint.size();
4604 if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
4605 tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
4607 bool Failed = Constraint.slice(2, Size - 1).getAsInteger(10, RegNo);
4608 if (!Failed && RegNo >= 0 && RegNo <= 31) {
4609 // v0 - v31 are aliases of q0 - q31.
4610 // By default we'll emit v0-v31 for this unless there's a modifier where
4611 // we'll emit the correct register as well.
4612 Res.first = AArch64::FPR128RegClass.getRegister(RegNo);
4613 Res.second = &AArch64::FPR128RegClass;
4621 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4622 /// vector. If it is invalid, don't add anything to Ops.
4623 void AArch64TargetLowering::LowerAsmOperandForConstraint(
4624 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
4625 SelectionDAG &DAG) const {
4628 // Currently only support length 1 constraints.
4629 if (Constraint.length() != 1)
4632 char ConstraintLetter = Constraint[0];
4633 switch (ConstraintLetter) {
4637 // This set of constraints deal with valid constants for various instructions.
4638 // Validate and return a target constant for them if we can.
4640 // 'z' maps to xzr or wzr so it needs an input of 0.
4641 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4642 if (!C || C->getZExtValue() != 0)
4645 if (Op.getValueType() == MVT::i64)
4646 Result = DAG.getRegister(AArch64::XZR, MVT::i64);
4648 Result = DAG.getRegister(AArch64::WZR, MVT::i32);
4658 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4662 // Grab the value and do some validation.
4663 uint64_t CVal = C->getZExtValue();
4664 switch (ConstraintLetter) {
4665 // The I constraint applies only to simple ADD or SUB immediate operands:
4666 // i.e. 0 to 4095 with optional shift by 12
4667 // The J constraint applies only to ADD or SUB immediates that would be
4668 // valid when negated, i.e. if [an add pattern] were to be output as a SUB
4669 // instruction [or vice versa], in other words -1 to -4095 with optional
4670 // left shift by 12.
4672 if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
4676 uint64_t NVal = -C->getSExtValue();
4677 if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal)) {
4678 CVal = C->getSExtValue();
4683 // The K and L constraints apply *only* to logical immediates, including
4684 // what used to be the MOVI alias for ORR (though the MOVI alias has now
4685 // been removed and MOV should be used). So these constraints have to
4686 // distinguish between bit patterns that are valid 32-bit or 64-bit
4687 // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
4688 // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
4691 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4695 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4698 // The M and N constraints are a superset of K and L respectively, for use
4699 // with the MOV (immediate) alias. As well as the logical immediates they
4700 // also match 32 or 64-bit immediates that can be loaded either using a
4701 // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
4702 // (M) or 64-bit 0x1234000000000000 (N) etc.
4703 // As a note some of this code is liberally stolen from the asm parser.
4705 if (!isUInt<32>(CVal))
4707 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4709 if ((CVal & 0xFFFF) == CVal)
4711 if ((CVal & 0xFFFF0000ULL) == CVal)
4713 uint64_t NCVal = ~(uint32_t)CVal;
4714 if ((NCVal & 0xFFFFULL) == NCVal)
4716 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4721 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4723 if ((CVal & 0xFFFFULL) == CVal)
4725 if ((CVal & 0xFFFF0000ULL) == CVal)
4727 if ((CVal & 0xFFFF00000000ULL) == CVal)
4729 if ((CVal & 0xFFFF000000000000ULL) == CVal)
4731 uint64_t NCVal = ~CVal;
4732 if ((NCVal & 0xFFFFULL) == NCVal)
4734 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4736 if ((NCVal & 0xFFFF00000000ULL) == NCVal)
4738 if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
4746 // All assembler immediates are 64-bit integers.
4747 Result = DAG.getTargetConstant(CVal, SDLoc(Op), MVT::i64);
4751 if (Result.getNode()) {
4752 Ops.push_back(Result);
4756 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
4759 //===----------------------------------------------------------------------===//
4760 // AArch64 Advanced SIMD Support
4761 //===----------------------------------------------------------------------===//
4763 /// WidenVector - Given a value in the V64 register class, produce the
4764 /// equivalent value in the V128 register class.
4765 static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
4766 EVT VT = V64Reg.getValueType();
4767 unsigned NarrowSize = VT.getVectorNumElements();
4768 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4769 MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
4772 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
4773 V64Reg, DAG.getConstant(0, DL, MVT::i32));
4776 /// getExtFactor - Determine the adjustment factor for the position when
4777 /// generating an "extract from vector registers" instruction.
4778 static unsigned getExtFactor(SDValue &V) {
4779 EVT EltType = V.getValueType().getVectorElementType();
4780 return EltType.getSizeInBits() / 8;
4783 /// NarrowVector - Given a value in the V128 register class, produce the
4784 /// equivalent value in the V64 register class.
4785 static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) {
4786 EVT VT = V128Reg.getValueType();
4787 unsigned WideSize = VT.getVectorNumElements();
4788 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4789 MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2);
4792 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, NarrowTy, V128Reg);
4795 // Gather data to see if the operation can be modelled as a
4796 // shuffle in combination with VEXTs.
4797 SDValue AArch64TargetLowering::ReconstructShuffle(SDValue Op,
4798 SelectionDAG &DAG) const {
4799 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
4801 EVT VT = Op.getValueType();
4802 unsigned NumElts = VT.getVectorNumElements();
4804 struct ShuffleSourceInfo {
4809 // We may insert some combination of BITCASTs and VEXT nodes to force Vec to
4810 // be compatible with the shuffle we intend to construct. As a result
4811 // ShuffleVec will be some sliding window into the original Vec.
4814 // Code should guarantee that element i in Vec starts at element "WindowBase
4815 // + i * WindowScale in ShuffleVec".
4819 bool operator ==(SDValue OtherVec) { return Vec == OtherVec; }
4820 ShuffleSourceInfo(SDValue Vec)
4821 : Vec(Vec), MinElt(UINT_MAX), MaxElt(0), ShuffleVec(Vec), WindowBase(0),
4825 // First gather all vectors used as an immediate source for this BUILD_VECTOR
4827 SmallVector<ShuffleSourceInfo, 2> Sources;
4828 for (unsigned i = 0; i < NumElts; ++i) {
4829 SDValue V = Op.getOperand(i);
4830 if (V.getOpcode() == ISD::UNDEF)
4832 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4833 // A shuffle can only come from building a vector from various
4834 // elements of other vectors.
4838 // Add this element source to the list if it's not already there.
4839 SDValue SourceVec = V.getOperand(0);
4840 auto Source = std::find(Sources.begin(), Sources.end(), SourceVec);
4841 if (Source == Sources.end())
4842 Source = Sources.insert(Sources.end(), ShuffleSourceInfo(SourceVec));
4844 // Update the minimum and maximum lane number seen.
4845 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4846 Source->MinElt = std::min(Source->MinElt, EltNo);
4847 Source->MaxElt = std::max(Source->MaxElt, EltNo);
4850 // Currently only do something sane when at most two source vectors
4852 if (Sources.size() > 2)
4855 // Find out the smallest element size among result and two sources, and use
4856 // it as element size to build the shuffle_vector.
4857 EVT SmallestEltTy = VT.getVectorElementType();
4858 for (auto &Source : Sources) {
4859 EVT SrcEltTy = Source.Vec.getValueType().getVectorElementType();
4860 if (SrcEltTy.bitsLT(SmallestEltTy)) {
4861 SmallestEltTy = SrcEltTy;
4864 unsigned ResMultiplier =
4865 VT.getVectorElementType().getSizeInBits() / SmallestEltTy.getSizeInBits();
4866 NumElts = VT.getSizeInBits() / SmallestEltTy.getSizeInBits();
4867 EVT ShuffleVT = EVT::getVectorVT(*DAG.getContext(), SmallestEltTy, NumElts);
4869 // If the source vector is too wide or too narrow, we may nevertheless be able
4870 // to construct a compatible shuffle either by concatenating it with UNDEF or
4871 // extracting a suitable range of elements.
4872 for (auto &Src : Sources) {
4873 EVT SrcVT = Src.ShuffleVec.getValueType();
4875 if (SrcVT.getSizeInBits() == VT.getSizeInBits())
4878 // This stage of the search produces a source with the same element type as
4879 // the original, but with a total width matching the BUILD_VECTOR output.
4880 EVT EltVT = SrcVT.getVectorElementType();
4881 unsigned NumSrcElts = VT.getSizeInBits() / EltVT.getSizeInBits();
4882 EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumSrcElts);
4884 if (SrcVT.getSizeInBits() < VT.getSizeInBits()) {
4885 assert(2 * SrcVT.getSizeInBits() == VT.getSizeInBits());
4886 // We can pad out the smaller vector for free, so if it's part of a
4889 DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, Src.ShuffleVec,
4890 DAG.getUNDEF(Src.ShuffleVec.getValueType()));
4894 assert(SrcVT.getSizeInBits() == 2 * VT.getSizeInBits());
4896 if (Src.MaxElt - Src.MinElt >= NumSrcElts) {
4897 // Span too large for a VEXT to cope
4901 if (Src.MinElt >= NumSrcElts) {
4902 // The extraction can just take the second half
4904 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4905 DAG.getConstant(NumSrcElts, dl, MVT::i64));
4906 Src.WindowBase = -NumSrcElts;
4907 } else if (Src.MaxElt < NumSrcElts) {
4908 // The extraction can just take the first half
4910 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4911 DAG.getConstant(0, dl, MVT::i64));
4913 // An actual VEXT is needed
4915 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4916 DAG.getConstant(0, dl, MVT::i64));
4918 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4919 DAG.getConstant(NumSrcElts, dl, MVT::i64));
4920 unsigned Imm = Src.MinElt * getExtFactor(VEXTSrc1);
4922 Src.ShuffleVec = DAG.getNode(AArch64ISD::EXT, dl, DestVT, VEXTSrc1,
4924 DAG.getConstant(Imm, dl, MVT::i32));
4925 Src.WindowBase = -Src.MinElt;
4929 // Another possible incompatibility occurs from the vector element types. We
4930 // can fix this by bitcasting the source vectors to the same type we intend
4932 for (auto &Src : Sources) {
4933 EVT SrcEltTy = Src.ShuffleVec.getValueType().getVectorElementType();
4934 if (SrcEltTy == SmallestEltTy)
4936 assert(ShuffleVT.getVectorElementType() == SmallestEltTy);
4937 Src.ShuffleVec = DAG.getNode(ISD::BITCAST, dl, ShuffleVT, Src.ShuffleVec);
4938 Src.WindowScale = SrcEltTy.getSizeInBits() / SmallestEltTy.getSizeInBits();
4939 Src.WindowBase *= Src.WindowScale;
4942 // Final sanity check before we try to actually produce a shuffle.
4944 for (auto Src : Sources)
4945 assert(Src.ShuffleVec.getValueType() == ShuffleVT);
4948 // The stars all align, our next step is to produce the mask for the shuffle.
4949 SmallVector<int, 8> Mask(ShuffleVT.getVectorNumElements(), -1);
4950 int BitsPerShuffleLane = ShuffleVT.getVectorElementType().getSizeInBits();
4951 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
4952 SDValue Entry = Op.getOperand(i);
4953 if (Entry.getOpcode() == ISD::UNDEF)
4956 auto Src = std::find(Sources.begin(), Sources.end(), Entry.getOperand(0));
4957 int EltNo = cast<ConstantSDNode>(Entry.getOperand(1))->getSExtValue();
4959 // EXTRACT_VECTOR_ELT performs an implicit any_ext; BUILD_VECTOR an implicit
4960 // trunc. So only std::min(SrcBits, DestBits) actually get defined in this
4962 EVT OrigEltTy = Entry.getOperand(0).getValueType().getVectorElementType();
4963 int BitsDefined = std::min(OrigEltTy.getSizeInBits(),
4964 VT.getVectorElementType().getSizeInBits());
4965 int LanesDefined = BitsDefined / BitsPerShuffleLane;
4967 // This source is expected to fill ResMultiplier lanes of the final shuffle,
4968 // starting at the appropriate offset.
4969 int *LaneMask = &Mask[i * ResMultiplier];
4971 int ExtractBase = EltNo * Src->WindowScale + Src->WindowBase;
4972 ExtractBase += NumElts * (Src - Sources.begin());
4973 for (int j = 0; j < LanesDefined; ++j)
4974 LaneMask[j] = ExtractBase + j;
4977 // Final check before we try to produce nonsense...
4978 if (!isShuffleMaskLegal(Mask, ShuffleVT))
4981 SDValue ShuffleOps[] = { DAG.getUNDEF(ShuffleVT), DAG.getUNDEF(ShuffleVT) };
4982 for (unsigned i = 0; i < Sources.size(); ++i)
4983 ShuffleOps[i] = Sources[i].ShuffleVec;
4985 SDValue Shuffle = DAG.getVectorShuffle(ShuffleVT, dl, ShuffleOps[0],
4986 ShuffleOps[1], &Mask[0]);
4987 return DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
4990 // check if an EXT instruction can handle the shuffle mask when the
4991 // vector sources of the shuffle are the same.
4992 static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4993 unsigned NumElts = VT.getVectorNumElements();
4995 // Assume that the first shuffle index is not UNDEF. Fail if it is.
5001 // If this is a VEXT shuffle, the immediate value is the index of the first
5002 // element. The other shuffle indices must be the successive elements after
5004 unsigned ExpectedElt = Imm;
5005 for (unsigned i = 1; i < NumElts; ++i) {
5006 // Increment the expected index. If it wraps around, just follow it
5007 // back to index zero and keep going.
5009 if (ExpectedElt == NumElts)
5013 continue; // ignore UNDEF indices
5014 if (ExpectedElt != static_cast<unsigned>(M[i]))
5021 // check if an EXT instruction can handle the shuffle mask when the
5022 // vector sources of the shuffle are different.
5023 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
5025 // Look for the first non-undef element.
5026 const int *FirstRealElt = std::find_if(M.begin(), M.end(),
5027 [](int Elt) {return Elt >= 0;});
5029 // Benefit form APInt to handle overflow when calculating expected element.
5030 unsigned NumElts = VT.getVectorNumElements();
5031 unsigned MaskBits = APInt(32, NumElts * 2).logBase2();
5032 APInt ExpectedElt = APInt(MaskBits, *FirstRealElt + 1);
5033 // The following shuffle indices must be the successive elements after the
5034 // first real element.
5035 const int *FirstWrongElt = std::find_if(FirstRealElt + 1, M.end(),
5036 [&](int Elt) {return Elt != ExpectedElt++ && Elt != -1;});
5037 if (FirstWrongElt != M.end())
5040 // The index of an EXT is the first element if it is not UNDEF.
5041 // Watch out for the beginning UNDEFs. The EXT index should be the expected
5042 // value of the first element. E.g.
5043 // <-1, -1, 3, ...> is treated as <1, 2, 3, ...>.
5044 // <-1, -1, 0, 1, ...> is treated as <2*NumElts-2, 2*NumElts-1, 0, 1, ...>.
5045 // ExpectedElt is the last mask index plus 1.
5046 Imm = ExpectedElt.getZExtValue();
5048 // There are two difference cases requiring to reverse input vectors.
5049 // For example, for vector <4 x i32> we have the following cases,
5050 // Case 1: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, -1, 0>)
5051 // Case 2: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, 7, 0>)
5052 // For both cases, we finally use mask <5, 6, 7, 0>, which requires
5053 // to reverse two input vectors.
5062 /// isREVMask - Check if a vector shuffle corresponds to a REV
5063 /// instruction with the specified blocksize. (The order of the elements
5064 /// within each block of the vector is reversed.)
5065 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
5066 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
5067 "Only possible block sizes for REV are: 16, 32, 64");
5069 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5073 unsigned NumElts = VT.getVectorNumElements();
5074 unsigned BlockElts = M[0] + 1;
5075 // If the first shuffle index is UNDEF, be optimistic.
5077 BlockElts = BlockSize / EltSz;
5079 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
5082 for (unsigned i = 0; i < NumElts; ++i) {
5084 continue; // ignore UNDEF indices
5085 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
5092 static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5093 unsigned NumElts = VT.getVectorNumElements();
5094 WhichResult = (M[0] == 0 ? 0 : 1);
5095 unsigned Idx = WhichResult * NumElts / 2;
5096 for (unsigned i = 0; i != NumElts; i += 2) {
5097 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
5098 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
5106 static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5107 unsigned NumElts = VT.getVectorNumElements();
5108 WhichResult = (M[0] == 0 ? 0 : 1);
5109 for (unsigned i = 0; i != NumElts; ++i) {
5111 continue; // ignore UNDEF indices
5112 if ((unsigned)M[i] != 2 * i + WhichResult)
5119 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5120 unsigned NumElts = VT.getVectorNumElements();
5121 WhichResult = (M[0] == 0 ? 0 : 1);
5122 for (unsigned i = 0; i < NumElts; i += 2) {
5123 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
5124 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + NumElts + WhichResult))
5130 /// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of
5131 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5132 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
5133 static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5134 unsigned NumElts = VT.getVectorNumElements();
5135 WhichResult = (M[0] == 0 ? 0 : 1);
5136 unsigned Idx = WhichResult * NumElts / 2;
5137 for (unsigned i = 0; i != NumElts; i += 2) {
5138 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
5139 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx))
5147 /// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of
5148 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5149 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
5150 static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5151 unsigned Half = VT.getVectorNumElements() / 2;
5152 WhichResult = (M[0] == 0 ? 0 : 1);
5153 for (unsigned j = 0; j != 2; ++j) {
5154 unsigned Idx = WhichResult;
5155 for (unsigned i = 0; i != Half; ++i) {
5156 int MIdx = M[i + j * Half];
5157 if (MIdx >= 0 && (unsigned)MIdx != Idx)
5166 /// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of
5167 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5168 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
5169 static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5170 unsigned NumElts = VT.getVectorNumElements();
5171 WhichResult = (M[0] == 0 ? 0 : 1);
5172 for (unsigned i = 0; i < NumElts; i += 2) {
5173 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
5174 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult))
5180 static bool isINSMask(ArrayRef<int> M, int NumInputElements,
5181 bool &DstIsLeft, int &Anomaly) {
5182 if (M.size() != static_cast<size_t>(NumInputElements))
5185 int NumLHSMatch = 0, NumRHSMatch = 0;
5186 int LastLHSMismatch = -1, LastRHSMismatch = -1;
5188 for (int i = 0; i < NumInputElements; ++i) {
5198 LastLHSMismatch = i;
5200 if (M[i] == i + NumInputElements)
5203 LastRHSMismatch = i;
5206 if (NumLHSMatch == NumInputElements - 1) {
5208 Anomaly = LastLHSMismatch;
5210 } else if (NumRHSMatch == NumInputElements - 1) {
5212 Anomaly = LastRHSMismatch;
5219 static bool isConcatMask(ArrayRef<int> Mask, EVT VT, bool SplitLHS) {
5220 if (VT.getSizeInBits() != 128)
5223 unsigned NumElts = VT.getVectorNumElements();
5225 for (int I = 0, E = NumElts / 2; I != E; I++) {
5230 int Offset = NumElts / 2;
5231 for (int I = NumElts / 2, E = NumElts; I != E; I++) {
5232 if (Mask[I] != I + SplitLHS * Offset)
5239 static SDValue tryFormConcatFromShuffle(SDValue Op, SelectionDAG &DAG) {
5241 EVT VT = Op.getValueType();
5242 SDValue V0 = Op.getOperand(0);
5243 SDValue V1 = Op.getOperand(1);
5244 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op)->getMask();
5246 if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() ||
5247 VT.getVectorElementType() != V1.getValueType().getVectorElementType())
5250 bool SplitV0 = V0.getValueType().getSizeInBits() == 128;
5252 if (!isConcatMask(Mask, VT, SplitV0))
5255 EVT CastVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
5256 VT.getVectorNumElements() / 2);
5258 V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0,
5259 DAG.getConstant(0, DL, MVT::i64));
5261 if (V1.getValueType().getSizeInBits() == 128) {
5262 V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1,
5263 DAG.getConstant(0, DL, MVT::i64));
5265 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1);
5268 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5269 /// the specified operations to build the shuffle.
5270 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5271 SDValue RHS, SelectionDAG &DAG,
5273 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5274 unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1);
5275 unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1);
5278 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5287 OP_VUZPL, // VUZP, left result
5288 OP_VUZPR, // VUZP, right result
5289 OP_VZIPL, // VZIP, left result
5290 OP_VZIPR, // VZIP, right result
5291 OP_VTRNL, // VTRN, left result
5292 OP_VTRNR // VTRN, right result
5295 if (OpNum == OP_COPY) {
5296 if (LHSID == (1 * 9 + 2) * 9 + 3)
5298 assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!");
5302 SDValue OpLHS, OpRHS;
5303 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5304 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5305 EVT VT = OpLHS.getValueType();
5309 llvm_unreachable("Unknown shuffle opcode!");
5311 // VREV divides the vector in half and swaps within the half.
5312 if (VT.getVectorElementType() == MVT::i32 ||
5313 VT.getVectorElementType() == MVT::f32)
5314 return DAG.getNode(AArch64ISD::REV64, dl, VT, OpLHS);
5315 // vrev <4 x i16> -> REV32
5316 if (VT.getVectorElementType() == MVT::i16 ||
5317 VT.getVectorElementType() == MVT::f16)
5318 return DAG.getNode(AArch64ISD::REV32, dl, VT, OpLHS);
5319 // vrev <4 x i8> -> REV16
5320 assert(VT.getVectorElementType() == MVT::i8);
5321 return DAG.getNode(AArch64ISD::REV16, dl, VT, OpLHS);
5326 EVT EltTy = VT.getVectorElementType();
5328 if (EltTy == MVT::i8)
5329 Opcode = AArch64ISD::DUPLANE8;
5330 else if (EltTy == MVT::i16 || EltTy == MVT::f16)
5331 Opcode = AArch64ISD::DUPLANE16;
5332 else if (EltTy == MVT::i32 || EltTy == MVT::f32)
5333 Opcode = AArch64ISD::DUPLANE32;
5334 else if (EltTy == MVT::i64 || EltTy == MVT::f64)
5335 Opcode = AArch64ISD::DUPLANE64;
5337 llvm_unreachable("Invalid vector element type?");
5339 if (VT.getSizeInBits() == 64)
5340 OpLHS = WidenVector(OpLHS, DAG);
5341 SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, dl, MVT::i64);
5342 return DAG.getNode(Opcode, dl, VT, OpLHS, Lane);
5347 unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
5348 return DAG.getNode(AArch64ISD::EXT, dl, VT, OpLHS, OpRHS,
5349 DAG.getConstant(Imm, dl, MVT::i32));
5352 return DAG.getNode(AArch64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS,
5355 return DAG.getNode(AArch64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS,
5358 return DAG.getNode(AArch64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS,
5361 return DAG.getNode(AArch64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS,
5364 return DAG.getNode(AArch64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS,
5367 return DAG.getNode(AArch64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS,
5372 static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask,
5373 SelectionDAG &DAG) {
5374 // Check to see if we can use the TBL instruction.
5375 SDValue V1 = Op.getOperand(0);
5376 SDValue V2 = Op.getOperand(1);
5379 EVT EltVT = Op.getValueType().getVectorElementType();
5380 unsigned BytesPerElt = EltVT.getSizeInBits() / 8;
5382 SmallVector<SDValue, 8> TBLMask;
5383 for (int Val : ShuffleMask) {
5384 for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
5385 unsigned Offset = Byte + Val * BytesPerElt;
5386 TBLMask.push_back(DAG.getConstant(Offset, DL, MVT::i32));
5390 MVT IndexVT = MVT::v8i8;
5391 unsigned IndexLen = 8;
5392 if (Op.getValueType().getSizeInBits() == 128) {
5393 IndexVT = MVT::v16i8;
5397 SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
5398 SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);
5401 if (V2.getNode()->getOpcode() == ISD::UNDEF) {
5403 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
5404 Shuffle = DAG.getNode(
5405 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5406 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, DL, MVT::i32), V1Cst,
5407 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5408 makeArrayRef(TBLMask.data(), IndexLen)));
5410 if (IndexLen == 8) {
5411 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
5412 Shuffle = DAG.getNode(
5413 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5414 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, DL, MVT::i32), V1Cst,
5415 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5416 makeArrayRef(TBLMask.data(), IndexLen)));
5418 // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
5419 // cannot currently represent the register constraints on the input
5421 // Shuffle = DAG.getNode(AArch64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
5422 // DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5423 // &TBLMask[0], IndexLen));
5424 Shuffle = DAG.getNode(
5425 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
5426 DAG.getConstant(Intrinsic::aarch64_neon_tbl2, DL, MVT::i32),
5428 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
5429 makeArrayRef(TBLMask.data(), IndexLen)));
5432 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
5435 static unsigned getDUPLANEOp(EVT EltType) {
5436 if (EltType == MVT::i8)
5437 return AArch64ISD::DUPLANE8;
5438 if (EltType == MVT::i16 || EltType == MVT::f16)
5439 return AArch64ISD::DUPLANE16;
5440 if (EltType == MVT::i32 || EltType == MVT::f32)
5441 return AArch64ISD::DUPLANE32;
5442 if (EltType == MVT::i64 || EltType == MVT::f64)
5443 return AArch64ISD::DUPLANE64;
5445 llvm_unreachable("Invalid vector element type?");
5448 SDValue AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
5449 SelectionDAG &DAG) const {
5451 EVT VT = Op.getValueType();
5453 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5455 // Convert shuffles that are directly supported on NEON to target-specific
5456 // DAG nodes, instead of keeping them as shuffles and matching them again
5457 // during code selection. This is more efficient and avoids the possibility
5458 // of inconsistencies between legalization and selection.
5459 ArrayRef<int> ShuffleMask = SVN->getMask();
5461 SDValue V1 = Op.getOperand(0);
5462 SDValue V2 = Op.getOperand(1);
5464 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0],
5465 V1.getValueType().getSimpleVT())) {
5466 int Lane = SVN->getSplatIndex();
5467 // If this is undef splat, generate it via "just" vdup, if possible.
5471 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
5472 return DAG.getNode(AArch64ISD::DUP, dl, V1.getValueType(),
5474 // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
5475 // constant. If so, we can just reference the lane's definition directly.
5476 if (V1.getOpcode() == ISD::BUILD_VECTOR &&
5477 !isa<ConstantSDNode>(V1.getOperand(Lane)))
5478 return DAG.getNode(AArch64ISD::DUP, dl, VT, V1.getOperand(Lane));
5480 // Otherwise, duplicate from the lane of the input vector.
5481 unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
5483 // SelectionDAGBuilder may have "helpfully" already extracted or conatenated
5484 // to make a vector of the same size as this SHUFFLE. We can ignore the
5485 // extract entirely, and canonicalise the concat using WidenVector.
5486 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
5487 Lane += cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
5488 V1 = V1.getOperand(0);
5489 } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
5490 unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
5491 Lane -= Idx * VT.getVectorNumElements() / 2;
5492 V1 = WidenVector(V1.getOperand(Idx), DAG);
5493 } else if (VT.getSizeInBits() == 64)
5494 V1 = WidenVector(V1, DAG);
5496 return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, dl, MVT::i64));
5499 if (isREVMask(ShuffleMask, VT, 64))
5500 return DAG.getNode(AArch64ISD::REV64, dl, V1.getValueType(), V1, V2);
5501 if (isREVMask(ShuffleMask, VT, 32))
5502 return DAG.getNode(AArch64ISD::REV32, dl, V1.getValueType(), V1, V2);
5503 if (isREVMask(ShuffleMask, VT, 16))
5504 return DAG.getNode(AArch64ISD::REV16, dl, V1.getValueType(), V1, V2);
5506 bool ReverseEXT = false;
5508 if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
5511 Imm *= getExtFactor(V1);
5512 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V2,
5513 DAG.getConstant(Imm, dl, MVT::i32));
5514 } else if (V2->getOpcode() == ISD::UNDEF &&
5515 isSingletonEXTMask(ShuffleMask, VT, Imm)) {
5516 Imm *= getExtFactor(V1);
5517 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V1,
5518 DAG.getConstant(Imm, dl, MVT::i32));
5521 unsigned WhichResult;
5522 if (isZIPMask(ShuffleMask, VT, WhichResult)) {
5523 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
5524 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5526 if (isUZPMask(ShuffleMask, VT, WhichResult)) {
5527 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
5528 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5530 if (isTRNMask(ShuffleMask, VT, WhichResult)) {
5531 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
5532 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
5535 if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5536 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
5537 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5539 if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5540 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
5541 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5543 if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
5544 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
5545 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
5548 SDValue Concat = tryFormConcatFromShuffle(Op, DAG);
5549 if (Concat.getNode())
5554 int NumInputElements = V1.getValueType().getVectorNumElements();
5555 if (isINSMask(ShuffleMask, NumInputElements, DstIsLeft, Anomaly)) {
5556 SDValue DstVec = DstIsLeft ? V1 : V2;
5557 SDValue DstLaneV = DAG.getConstant(Anomaly, dl, MVT::i64);
5559 SDValue SrcVec = V1;
5560 int SrcLane = ShuffleMask[Anomaly];
5561 if (SrcLane >= NumInputElements) {
5563 SrcLane -= VT.getVectorNumElements();
5565 SDValue SrcLaneV = DAG.getConstant(SrcLane, dl, MVT::i64);
5567 EVT ScalarVT = VT.getVectorElementType();
5569 if (ScalarVT.getSizeInBits() < 32 && ScalarVT.isInteger())
5570 ScalarVT = MVT::i32;
5573 ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5574 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ScalarVT, SrcVec, SrcLaneV),
5578 // If the shuffle is not directly supported and it has 4 elements, use
5579 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5580 unsigned NumElts = VT.getVectorNumElements();
5582 unsigned PFIndexes[4];
5583 for (unsigned i = 0; i != 4; ++i) {
5584 if (ShuffleMask[i] < 0)
5587 PFIndexes[i] = ShuffleMask[i];
5590 // Compute the index in the perfect shuffle table.
5591 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
5592 PFIndexes[2] * 9 + PFIndexes[3];
5593 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5594 unsigned Cost = (PFEntry >> 30);
5597 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5600 return GenerateTBL(Op, ShuffleMask, DAG);
5603 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
5605 EVT VT = BVN->getValueType(0);
5606 APInt SplatBits, SplatUndef;
5607 unsigned SplatBitSize;
5609 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5610 unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;
5612 for (unsigned i = 0; i < NumSplats; ++i) {
5613 CnstBits <<= SplatBitSize;
5614 UndefBits <<= SplatBitSize;
5615 CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits());
5616 UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits());
5625 SDValue AArch64TargetLowering::LowerVectorAND(SDValue Op,
5626 SelectionDAG &DAG) const {
5627 BuildVectorSDNode *BVN =
5628 dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5629 SDValue LHS = Op.getOperand(0);
5631 EVT VT = Op.getValueType();
5636 APInt CnstBits(VT.getSizeInBits(), 0);
5637 APInt UndefBits(VT.getSizeInBits(), 0);
5638 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5639 // We only have BIC vector immediate instruction, which is and-not.
5640 CnstBits = ~CnstBits;
5642 // We make use of a little bit of goto ickiness in order to avoid having to
5643 // duplicate the immediate matching logic for the undef toggled case.
5644 bool SecondTry = false;
5647 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5648 CnstBits = CnstBits.zextOrTrunc(64);
5649 uint64_t CnstVal = CnstBits.getZExtValue();
5651 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5652 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5653 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5654 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5655 DAG.getConstant(CnstVal, dl, MVT::i32),
5656 DAG.getConstant(0, dl, MVT::i32));
5657 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5660 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5661 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5662 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5663 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5664 DAG.getConstant(CnstVal, dl, MVT::i32),
5665 DAG.getConstant(8, dl, MVT::i32));
5666 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5669 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5670 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5671 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5672 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5673 DAG.getConstant(CnstVal, dl, MVT::i32),
5674 DAG.getConstant(16, dl, MVT::i32));
5675 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5678 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5679 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5680 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5681 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5682 DAG.getConstant(CnstVal, dl, MVT::i32),
5683 DAG.getConstant(24, dl, MVT::i32));
5684 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5687 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5688 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5689 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5690 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5691 DAG.getConstant(CnstVal, dl, MVT::i32),
5692 DAG.getConstant(0, dl, MVT::i32));
5693 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5696 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5697 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5698 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5699 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5700 DAG.getConstant(CnstVal, dl, MVT::i32),
5701 DAG.getConstant(8, dl, MVT::i32));
5702 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5709 CnstBits = ~UndefBits;
5713 // We can always fall back to a non-immediate AND.
5718 // Specialized code to quickly find if PotentialBVec is a BuildVector that
5719 // consists of only the same constant int value, returned in reference arg
5721 static bool isAllConstantBuildVector(const SDValue &PotentialBVec,
5722 uint64_t &ConstVal) {
5723 BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec);
5726 ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0));
5729 EVT VT = Bvec->getValueType(0);
5730 unsigned NumElts = VT.getVectorNumElements();
5731 for (unsigned i = 1; i < NumElts; ++i)
5732 if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt)
5734 ConstVal = FirstElt->getZExtValue();
5738 static unsigned getIntrinsicID(const SDNode *N) {
5739 unsigned Opcode = N->getOpcode();
5742 return Intrinsic::not_intrinsic;
5743 case ISD::INTRINSIC_WO_CHAIN: {
5744 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
5745 if (IID < Intrinsic::num_intrinsics)
5747 return Intrinsic::not_intrinsic;
5752 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
5753 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
5754 // BUILD_VECTORs with constant element C1, C2 is a constant, and C1 == ~C2.
5755 // Also, logical shift right -> sri, with the same structure.
5756 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
5757 EVT VT = N->getValueType(0);
5764 // Is the first op an AND?
5765 const SDValue And = N->getOperand(0);
5766 if (And.getOpcode() != ISD::AND)
5769 // Is the second op an shl or lshr?
5770 SDValue Shift = N->getOperand(1);
5771 // This will have been turned into: AArch64ISD::VSHL vector, #shift
5772 // or AArch64ISD::VLSHR vector, #shift
5773 unsigned ShiftOpc = Shift.getOpcode();
5774 if ((ShiftOpc != AArch64ISD::VSHL && ShiftOpc != AArch64ISD::VLSHR))
5776 bool IsShiftRight = ShiftOpc == AArch64ISD::VLSHR;
5778 // Is the shift amount constant?
5779 ConstantSDNode *C2node = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
5783 // Is the and mask vector all constant?
5785 if (!isAllConstantBuildVector(And.getOperand(1), C1))
5788 // Is C1 == ~C2, taking into account how much one can shift elements of a
5790 uint64_t C2 = C2node->getZExtValue();
5791 unsigned ElemSizeInBits = VT.getVectorElementType().getSizeInBits();
5792 if (C2 > ElemSizeInBits)
5794 unsigned ElemMask = (1 << ElemSizeInBits) - 1;
5795 if ((C1 & ElemMask) != (~C2 & ElemMask))
5798 SDValue X = And.getOperand(0);
5799 SDValue Y = Shift.getOperand(0);
5802 IsShiftRight ? Intrinsic::aarch64_neon_vsri : Intrinsic::aarch64_neon_vsli;
5804 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5805 DAG.getConstant(Intrin, DL, MVT::i32), X, Y,
5806 Shift.getOperand(1));
5808 DEBUG(dbgs() << "aarch64-lower: transformed: \n");
5809 DEBUG(N->dump(&DAG));
5810 DEBUG(dbgs() << "into: \n");
5811 DEBUG(ResultSLI->dump(&DAG));
5817 SDValue AArch64TargetLowering::LowerVectorOR(SDValue Op,
5818 SelectionDAG &DAG) const {
5819 // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
5820 if (EnableAArch64SlrGeneration) {
5821 SDValue Res = tryLowerToSLI(Op.getNode(), DAG);
5826 BuildVectorSDNode *BVN =
5827 dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode());
5828 SDValue LHS = Op.getOperand(1);
5830 EVT VT = Op.getValueType();
5832 // OR commutes, so try swapping the operands.
5834 LHS = Op.getOperand(0);
5835 BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5840 APInt CnstBits(VT.getSizeInBits(), 0);
5841 APInt UndefBits(VT.getSizeInBits(), 0);
5842 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5843 // We make use of a little bit of goto ickiness in order to avoid having to
5844 // duplicate the immediate matching logic for the undef toggled case.
5845 bool SecondTry = false;
5848 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5849 CnstBits = CnstBits.zextOrTrunc(64);
5850 uint64_t CnstVal = CnstBits.getZExtValue();
5852 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5853 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5854 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5855 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5856 DAG.getConstant(CnstVal, dl, MVT::i32),
5857 DAG.getConstant(0, dl, MVT::i32));
5858 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5861 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5862 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5863 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5864 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5865 DAG.getConstant(CnstVal, dl, MVT::i32),
5866 DAG.getConstant(8, dl, MVT::i32));
5867 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5870 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5871 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5872 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5873 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5874 DAG.getConstant(CnstVal, dl, MVT::i32),
5875 DAG.getConstant(16, dl, MVT::i32));
5876 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5879 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5880 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5881 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5882 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5883 DAG.getConstant(CnstVal, dl, MVT::i32),
5884 DAG.getConstant(24, dl, MVT::i32));
5885 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5888 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5889 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5890 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5891 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5892 DAG.getConstant(CnstVal, dl, MVT::i32),
5893 DAG.getConstant(0, dl, MVT::i32));
5894 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5897 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5898 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5899 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5900 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5901 DAG.getConstant(CnstVal, dl, MVT::i32),
5902 DAG.getConstant(8, dl, MVT::i32));
5903 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5910 CnstBits = UndefBits;
5914 // We can always fall back to a non-immediate OR.
5919 // Normalize the operands of BUILD_VECTOR. The value of constant operands will
5920 // be truncated to fit element width.
5921 static SDValue NormalizeBuildVector(SDValue Op,
5922 SelectionDAG &DAG) {
5923 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
5925 EVT VT = Op.getValueType();
5926 EVT EltTy= VT.getVectorElementType();
5928 if (EltTy.isFloatingPoint() || EltTy.getSizeInBits() > 16)
5931 SmallVector<SDValue, 16> Ops;
5932 for (SDValue Lane : Op->ops()) {
5933 if (auto *CstLane = dyn_cast<ConstantSDNode>(Lane)) {
5934 APInt LowBits(EltTy.getSizeInBits(),
5935 CstLane->getZExtValue());
5936 Lane = DAG.getConstant(LowBits.getZExtValue(), dl, MVT::i32);
5938 Ops.push_back(Lane);
5940 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5943 SDValue AArch64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
5944 SelectionDAG &DAG) const {
5946 EVT VT = Op.getValueType();
5947 Op = NormalizeBuildVector(Op, DAG);
5948 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5950 APInt CnstBits(VT.getSizeInBits(), 0);
5951 APInt UndefBits(VT.getSizeInBits(), 0);
5952 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5953 // We make use of a little bit of goto ickiness in order to avoid having to
5954 // duplicate the immediate matching logic for the undef toggled case.
5955 bool SecondTry = false;
5958 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5959 CnstBits = CnstBits.zextOrTrunc(64);
5960 uint64_t CnstVal = CnstBits.getZExtValue();
5962 // Certain magic vector constants (used to express things like NOT
5963 // and NEG) are passed through unmodified. This allows codegen patterns
5964 // for these operations to match. Special-purpose patterns will lower
5965 // these immediates to MOVIs if it proves necessary.
5966 if (VT.isInteger() && (CnstVal == 0 || CnstVal == ~0ULL))
5969 // The many faces of MOVI...
5970 if (AArch64_AM::isAdvSIMDModImmType10(CnstVal)) {
5971 CnstVal = AArch64_AM::encodeAdvSIMDModImmType10(CnstVal);
5972 if (VT.getSizeInBits() == 128) {
5973 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::v2i64,
5974 DAG.getConstant(CnstVal, dl, MVT::i32));
5975 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5978 // Support the V64 version via subregister insertion.
5979 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::f64,
5980 DAG.getConstant(CnstVal, dl, MVT::i32));
5981 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5984 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5985 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5986 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5987 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5988 DAG.getConstant(CnstVal, dl, MVT::i32),
5989 DAG.getConstant(0, dl, MVT::i32));
5990 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
5993 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5994 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5995 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5996 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5997 DAG.getConstant(CnstVal, dl, MVT::i32),
5998 DAG.getConstant(8, dl, MVT::i32));
5999 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6002 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
6003 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
6004 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6005 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
6006 DAG.getConstant(CnstVal, dl, MVT::i32),
6007 DAG.getConstant(16, dl, MVT::i32));
6008 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6011 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
6012 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
6013 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6014 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
6015 DAG.getConstant(CnstVal, dl, MVT::i32),
6016 DAG.getConstant(24, dl, MVT::i32));
6017 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6020 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
6021 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
6022 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
6023 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
6024 DAG.getConstant(CnstVal, dl, MVT::i32),
6025 DAG.getConstant(0, dl, MVT::i32));
6026 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6029 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
6030 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
6031 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
6032 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
6033 DAG.getConstant(CnstVal, dl, MVT::i32),
6034 DAG.getConstant(8, dl, MVT::i32));
6035 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6038 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
6039 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
6040 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6041 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
6042 DAG.getConstant(CnstVal, dl, MVT::i32),
6043 DAG.getConstant(264, dl, MVT::i32));
6044 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6047 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
6048 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
6049 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6050 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
6051 DAG.getConstant(CnstVal, dl, MVT::i32),
6052 DAG.getConstant(272, dl, MVT::i32));
6053 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6056 if (AArch64_AM::isAdvSIMDModImmType9(CnstVal)) {
6057 CnstVal = AArch64_AM::encodeAdvSIMDModImmType9(CnstVal);
6058 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;
6059 SDValue Mov = DAG.getNode(AArch64ISD::MOVI, dl, MovTy,
6060 DAG.getConstant(CnstVal, dl, MVT::i32));
6061 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6064 // The few faces of FMOV...
6065 if (AArch64_AM::isAdvSIMDModImmType11(CnstVal)) {
6066 CnstVal = AArch64_AM::encodeAdvSIMDModImmType11(CnstVal);
6067 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4f32 : MVT::v2f32;
6068 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MovTy,
6069 DAG.getConstant(CnstVal, dl, MVT::i32));
6070 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6073 if (AArch64_AM::isAdvSIMDModImmType12(CnstVal) &&
6074 VT.getSizeInBits() == 128) {
6075 CnstVal = AArch64_AM::encodeAdvSIMDModImmType12(CnstVal);
6076 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MVT::v2f64,
6077 DAG.getConstant(CnstVal, dl, MVT::i32));
6078 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6081 // The many faces of MVNI...
6083 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
6084 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
6085 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6086 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6087 DAG.getConstant(CnstVal, dl, MVT::i32),
6088 DAG.getConstant(0, dl, MVT::i32));
6089 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6092 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
6093 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
6094 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6095 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6096 DAG.getConstant(CnstVal, dl, MVT::i32),
6097 DAG.getConstant(8, dl, MVT::i32));
6098 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6101 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
6102 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
6103 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6104 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6105 DAG.getConstant(CnstVal, dl, MVT::i32),
6106 DAG.getConstant(16, dl, MVT::i32));
6107 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6110 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
6111 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
6112 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6113 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6114 DAG.getConstant(CnstVal, dl, MVT::i32),
6115 DAG.getConstant(24, dl, MVT::i32));
6116 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6119 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
6120 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
6121 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
6122 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6123 DAG.getConstant(CnstVal, dl, MVT::i32),
6124 DAG.getConstant(0, dl, MVT::i32));
6125 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6128 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
6129 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
6130 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
6131 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
6132 DAG.getConstant(CnstVal, dl, MVT::i32),
6133 DAG.getConstant(8, dl, MVT::i32));
6134 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6137 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
6138 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
6139 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6140 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
6141 DAG.getConstant(CnstVal, dl, MVT::i32),
6142 DAG.getConstant(264, dl, MVT::i32));
6143 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6146 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
6147 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
6148 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
6149 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
6150 DAG.getConstant(CnstVal, dl, MVT::i32),
6151 DAG.getConstant(272, dl, MVT::i32));
6152 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov);
6159 CnstBits = UndefBits;
6164 // Scan through the operands to find some interesting properties we can
6166 // 1) If only one value is used, we can use a DUP, or
6167 // 2) if only the low element is not undef, we can just insert that, or
6168 // 3) if only one constant value is used (w/ some non-constant lanes),
6169 // we can splat the constant value into the whole vector then fill
6170 // in the non-constant lanes.
6171 // 4) FIXME: If different constant values are used, but we can intelligently
6172 // select the values we'll be overwriting for the non-constant
6173 // lanes such that we can directly materialize the vector
6174 // some other way (MOVI, e.g.), we can be sneaky.
6175 unsigned NumElts = VT.getVectorNumElements();
6176 bool isOnlyLowElement = true;
6177 bool usesOnlyOneValue = true;
6178 bool usesOnlyOneConstantValue = true;
6179 bool isConstant = true;
6180 unsigned NumConstantLanes = 0;
6182 SDValue ConstantValue;
6183 for (unsigned i = 0; i < NumElts; ++i) {
6184 SDValue V = Op.getOperand(i);
6185 if (V.getOpcode() == ISD::UNDEF)
6188 isOnlyLowElement = false;
6189 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
6192 if (isa<ConstantSDNode>(V) || isa<ConstantFPSDNode>(V)) {
6194 if (!ConstantValue.getNode())
6196 else if (ConstantValue != V)
6197 usesOnlyOneConstantValue = false;
6200 if (!Value.getNode())
6202 else if (V != Value)
6203 usesOnlyOneValue = false;
6206 if (!Value.getNode())
6207 return DAG.getUNDEF(VT);
6209 if (isOnlyLowElement)
6210 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
6212 // Use DUP for non-constant splats. For f32 constant splats, reduce to
6213 // i32 and try again.
6214 if (usesOnlyOneValue) {
6216 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6217 Value.getValueType() != VT)
6218 return DAG.getNode(AArch64ISD::DUP, dl, VT, Value);
6220 // This is actually a DUPLANExx operation, which keeps everything vectory.
6222 // DUPLANE works on 128-bit vectors, widen it if necessary.
6223 SDValue Lane = Value.getOperand(1);
6224 Value = Value.getOperand(0);
6225 if (Value.getValueType().getSizeInBits() == 64)
6226 Value = WidenVector(Value, DAG);
6228 unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
6229 return DAG.getNode(Opcode, dl, VT, Value, Lane);
6232 if (VT.getVectorElementType().isFloatingPoint()) {
6233 SmallVector<SDValue, 8> Ops;
6234 EVT EltTy = VT.getVectorElementType();
6235 assert ((EltTy == MVT::f16 || EltTy == MVT::f32 || EltTy == MVT::f64) &&
6236 "Unsupported floating-point vector type");
6237 MVT NewType = MVT::getIntegerVT(EltTy.getSizeInBits());
6238 for (unsigned i = 0; i < NumElts; ++i)
6239 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, NewType, Op.getOperand(i)));
6240 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts);
6241 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
6242 Val = LowerBUILD_VECTOR(Val, DAG);
6244 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
6248 // If there was only one constant value used and for more than one lane,
6249 // start by splatting that value, then replace the non-constant lanes. This
6250 // is better than the default, which will perform a separate initialization
6252 if (NumConstantLanes > 0 && usesOnlyOneConstantValue) {
6253 SDValue Val = DAG.getNode(AArch64ISD::DUP, dl, VT, ConstantValue);
6254 // Now insert the non-constant lanes.
6255 for (unsigned i = 0; i < NumElts; ++i) {
6256 SDValue V = Op.getOperand(i);
6257 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i64);
6258 if (!isa<ConstantSDNode>(V) && !isa<ConstantFPSDNode>(V)) {
6259 // Note that type legalization likely mucked about with the VT of the
6260 // source operand, so we may have to convert it here before inserting.
6261 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, V, LaneIdx);
6267 // If all elements are constants and the case above didn't get hit, fall back
6268 // to the default expansion, which will generate a load from the constant
6273 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
6275 if (SDValue shuffle = ReconstructShuffle(Op, DAG))
6279 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
6280 // know the default expansion would otherwise fall back on something even
6281 // worse. For a vector with one or two non-undef values, that's
6282 // scalar_to_vector for the elements followed by a shuffle (provided the
6283 // shuffle is valid for the target) and materialization element by element
6284 // on the stack followed by a load for everything else.
6285 if (!isConstant && !usesOnlyOneValue) {
6286 SDValue Vec = DAG.getUNDEF(VT);
6287 SDValue Op0 = Op.getOperand(0);
6288 unsigned ElemSize = VT.getVectorElementType().getSizeInBits();
6290 // For 32 and 64 bit types, use INSERT_SUBREG for lane zero to
6291 // a) Avoid a RMW dependency on the full vector register, and
6292 // b) Allow the register coalescer to fold away the copy if the
6293 // value is already in an S or D register.
6294 // Do not do this for UNDEF/LOAD nodes because we have better patterns
6295 // for those avoiding the SCALAR_TO_VECTOR/BUILD_VECTOR.
6296 if (Op0.getOpcode() != ISD::UNDEF && Op0.getOpcode() != ISD::LOAD &&
6297 (ElemSize == 32 || ElemSize == 64)) {
6298 unsigned SubIdx = ElemSize == 32 ? AArch64::ssub : AArch64::dsub;
6300 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0,
6301 DAG.getTargetConstant(SubIdx, dl, MVT::i32));
6302 Vec = SDValue(N, 0);
6305 for (; i < NumElts; ++i) {
6306 SDValue V = Op.getOperand(i);
6307 if (V.getOpcode() == ISD::UNDEF)
6309 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i64);
6310 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
6315 // Just use the default expansion. We failed to find a better alternative.
6319 SDValue AArch64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
6320 SelectionDAG &DAG) const {
6321 assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!");
6323 // Check for non-constant or out of range lane.
6324 EVT VT = Op.getOperand(0).getValueType();
6325 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(2));
6326 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
6330 // Insertion/extraction are legal for V128 types.
6331 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
6332 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
6336 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
6337 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
6340 // For V64 types, we perform insertion by expanding the value
6341 // to a V128 type and perform the insertion on that.
6343 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
6344 EVT WideTy = WideVec.getValueType();
6346 SDValue Node = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideTy, WideVec,
6347 Op.getOperand(1), Op.getOperand(2));
6348 // Re-narrow the resultant vector.
6349 return NarrowVector(Node, DAG);
6353 AArch64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
6354 SelectionDAG &DAG) const {
6355 assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!");
6357 // Check for non-constant or out of range lane.
6358 EVT VT = Op.getOperand(0).getValueType();
6359 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6360 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
6364 // Insertion/extraction are legal for V128 types.
6365 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
6366 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
6370 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
6371 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16)
6374 // For V64 types, we perform extraction by expanding the value
6375 // to a V128 type and perform the extraction on that.
6377 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
6378 EVT WideTy = WideVec.getValueType();
6380 EVT ExtrTy = WideTy.getVectorElementType();
6381 if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8)
6384 // For extractions, we just return the result directly.
6385 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec,
6389 SDValue AArch64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
6390 SelectionDAG &DAG) const {
6391 EVT VT = Op.getOperand(0).getValueType();
6397 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6400 unsigned Val = Cst->getZExtValue();
6402 unsigned Size = Op.getValueType().getSizeInBits();
6406 return DAG.getTargetExtractSubreg(AArch64::bsub, dl, Op.getValueType(),
6409 return DAG.getTargetExtractSubreg(AArch64::hsub, dl, Op.getValueType(),
6412 return DAG.getTargetExtractSubreg(AArch64::ssub, dl, Op.getValueType(),
6415 return DAG.getTargetExtractSubreg(AArch64::dsub, dl, Op.getValueType(),
6418 llvm_unreachable("Unexpected vector type in extract_subvector!");
6421 // If this is extracting the upper 64-bits of a 128-bit vector, we match
6423 if (Size == 64 && Val * VT.getVectorElementType().getSizeInBits() == 64)
6429 bool AArch64TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
6431 if (VT.getVectorNumElements() == 4 &&
6432 (VT.is128BitVector() || VT.is64BitVector())) {
6433 unsigned PFIndexes[4];
6434 for (unsigned i = 0; i != 4; ++i) {
6438 PFIndexes[i] = M[i];
6441 // Compute the index in the perfect shuffle table.
6442 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
6443 PFIndexes[2] * 9 + PFIndexes[3];
6444 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
6445 unsigned Cost = (PFEntry >> 30);
6453 unsigned DummyUnsigned;
6455 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || isREVMask(M, VT, 64) ||
6456 isREVMask(M, VT, 32) || isREVMask(M, VT, 16) ||
6457 isEXTMask(M, VT, DummyBool, DummyUnsigned) ||
6458 // isTBLMask(M, VT) || // FIXME: Port TBL support from ARM.
6459 isTRNMask(M, VT, DummyUnsigned) || isUZPMask(M, VT, DummyUnsigned) ||
6460 isZIPMask(M, VT, DummyUnsigned) ||
6461 isTRN_v_undef_Mask(M, VT, DummyUnsigned) ||
6462 isUZP_v_undef_Mask(M, VT, DummyUnsigned) ||
6463 isZIP_v_undef_Mask(M, VT, DummyUnsigned) ||
6464 isINSMask(M, VT.getVectorNumElements(), DummyBool, DummyInt) ||
6465 isConcatMask(M, VT, VT.getSizeInBits() == 128));
6468 /// getVShiftImm - Check if this is a valid build_vector for the immediate
6469 /// operand of a vector shift operation, where all the elements of the
6470 /// build_vector must have the same constant integer value.
6471 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
6472 // Ignore bit_converts.
6473 while (Op.getOpcode() == ISD::BITCAST)
6474 Op = Op.getOperand(0);
6475 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
6476 APInt SplatBits, SplatUndef;
6477 unsigned SplatBitSize;
6479 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
6480 HasAnyUndefs, ElementBits) ||
6481 SplatBitSize > ElementBits)
6483 Cnt = SplatBits.getSExtValue();
6487 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
6488 /// operand of a vector shift left operation. That value must be in the range:
6489 /// 0 <= Value < ElementBits for a left shift; or
6490 /// 0 <= Value <= ElementBits for a long left shift.
6491 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
6492 assert(VT.isVector() && "vector shift count is not a vector type");
6493 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
6494 if (!getVShiftImm(Op, ElementBits, Cnt))
6496 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
6499 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
6500 /// operand of a vector shift right operation. The value must be in the range:
6501 /// 1 <= Value <= ElementBits for a right shift; or
6502 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, int64_t &Cnt) {
6503 assert(VT.isVector() && "vector shift count is not a vector type");
6504 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
6505 if (!getVShiftImm(Op, ElementBits, Cnt))
6507 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
6510 SDValue AArch64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
6511 SelectionDAG &DAG) const {
6512 EVT VT = Op.getValueType();
6516 if (!Op.getOperand(1).getValueType().isVector())
6518 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
6520 switch (Op.getOpcode()) {
6522 llvm_unreachable("unexpected shift opcode");
6525 if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
6526 return DAG.getNode(AArch64ISD::VSHL, DL, VT, Op.getOperand(0),
6527 DAG.getConstant(Cnt, DL, MVT::i32));
6528 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
6529 DAG.getConstant(Intrinsic::aarch64_neon_ushl, DL,
6531 Op.getOperand(0), Op.getOperand(1));
6534 // Right shift immediate
6535 if (isVShiftRImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize) {
6537 (Op.getOpcode() == ISD::SRA) ? AArch64ISD::VASHR : AArch64ISD::VLSHR;
6538 return DAG.getNode(Opc, DL, VT, Op.getOperand(0),
6539 DAG.getConstant(Cnt, DL, MVT::i32));
6542 // Right shift register. Note, there is not a shift right register
6543 // instruction, but the shift left register instruction takes a signed
6544 // value, where negative numbers specify a right shift.
6545 unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::aarch64_neon_sshl
6546 : Intrinsic::aarch64_neon_ushl;
6547 // negate the shift amount
6548 SDValue NegShift = DAG.getNode(AArch64ISD::NEG, DL, VT, Op.getOperand(1));
6549 SDValue NegShiftLeft =
6550 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
6551 DAG.getConstant(Opc, DL, MVT::i32), Op.getOperand(0),
6553 return NegShiftLeft;
6559 static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS,
6560 AArch64CC::CondCode CC, bool NoNans, EVT VT,
6561 SDLoc dl, SelectionDAG &DAG) {
6562 EVT SrcVT = LHS.getValueType();
6563 assert(VT.getSizeInBits() == SrcVT.getSizeInBits() &&
6564 "function only supposed to emit natural comparisons");
6566 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
6567 APInt CnstBits(VT.getSizeInBits(), 0);
6568 APInt UndefBits(VT.getSizeInBits(), 0);
6569 bool IsCnst = BVN && resolveBuildVector(BVN, CnstBits, UndefBits);
6570 bool IsZero = IsCnst && (CnstBits == 0);
6572 if (SrcVT.getVectorElementType().isFloatingPoint()) {
6576 case AArch64CC::NE: {
6579 Fcmeq = DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
6581 Fcmeq = DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
6582 return DAG.getNode(AArch64ISD::NOT, dl, VT, Fcmeq);
6586 return DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
6587 return DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
6590 return DAG.getNode(AArch64ISD::FCMGEz, dl, VT, LHS);
6591 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, LHS, RHS);
6594 return DAG.getNode(AArch64ISD::FCMGTz, dl, VT, LHS);
6595 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, LHS, RHS);
6598 return DAG.getNode(AArch64ISD::FCMLEz, dl, VT, LHS);
6599 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, RHS, LHS);
6603 // If we ignore NaNs then we can use to the MI implementation.
6607 return DAG.getNode(AArch64ISD::FCMLTz, dl, VT, LHS);
6608 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, RHS, LHS);
6615 case AArch64CC::NE: {
6618 Cmeq = DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
6620 Cmeq = DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
6621 return DAG.getNode(AArch64ISD::NOT, dl, VT, Cmeq);
6625 return DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
6626 return DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
6629 return DAG.getNode(AArch64ISD::CMGEz, dl, VT, LHS);
6630 return DAG.getNode(AArch64ISD::CMGE, dl, VT, LHS, RHS);
6633 return DAG.getNode(AArch64ISD::CMGTz, dl, VT, LHS);
6634 return DAG.getNode(AArch64ISD::CMGT, dl, VT, LHS, RHS);
6637 return DAG.getNode(AArch64ISD::CMLEz, dl, VT, LHS);
6638 return DAG.getNode(AArch64ISD::CMGE, dl, VT, RHS, LHS);
6640 return DAG.getNode(AArch64ISD::CMHS, dl, VT, RHS, LHS);
6642 return DAG.getNode(AArch64ISD::CMHI, dl, VT, RHS, LHS);
6645 return DAG.getNode(AArch64ISD::CMLTz, dl, VT, LHS);
6646 return DAG.getNode(AArch64ISD::CMGT, dl, VT, RHS, LHS);
6648 return DAG.getNode(AArch64ISD::CMHI, dl, VT, LHS, RHS);
6650 return DAG.getNode(AArch64ISD::CMHS, dl, VT, LHS, RHS);
6654 SDValue AArch64TargetLowering::LowerVSETCC(SDValue Op,
6655 SelectionDAG &DAG) const {
6656 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6657 SDValue LHS = Op.getOperand(0);
6658 SDValue RHS = Op.getOperand(1);
6659 EVT CmpVT = LHS.getValueType().changeVectorElementTypeToInteger();
6662 if (LHS.getValueType().getVectorElementType().isInteger()) {
6663 assert(LHS.getValueType() == RHS.getValueType());
6664 AArch64CC::CondCode AArch64CC = changeIntCCToAArch64CC(CC);
6666 EmitVectorComparison(LHS, RHS, AArch64CC, false, CmpVT, dl, DAG);
6667 return DAG.getSExtOrTrunc(Cmp, dl, Op.getValueType());
6670 assert(LHS.getValueType().getVectorElementType() == MVT::f32 ||
6671 LHS.getValueType().getVectorElementType() == MVT::f64);
6673 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
6674 // clean. Some of them require two branches to implement.
6675 AArch64CC::CondCode CC1, CC2;
6677 changeVectorFPCCToAArch64CC(CC, CC1, CC2, ShouldInvert);
6679 bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath;
6681 EmitVectorComparison(LHS, RHS, CC1, NoNaNs, CmpVT, dl, DAG);
6685 if (CC2 != AArch64CC::AL) {
6687 EmitVectorComparison(LHS, RHS, CC2, NoNaNs, CmpVT, dl, DAG);
6688 if (!Cmp2.getNode())
6691 Cmp = DAG.getNode(ISD::OR, dl, CmpVT, Cmp, Cmp2);
6694 Cmp = DAG.getSExtOrTrunc(Cmp, dl, Op.getValueType());
6697 return Cmp = DAG.getNOT(dl, Cmp, Cmp.getValueType());
6702 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
6703 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
6704 /// specified in the intrinsic calls.
6705 bool AArch64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
6707 unsigned Intrinsic) const {
6708 auto &DL = I.getModule()->getDataLayout();
6709 switch (Intrinsic) {
6710 case Intrinsic::aarch64_neon_ld2:
6711 case Intrinsic::aarch64_neon_ld3:
6712 case Intrinsic::aarch64_neon_ld4:
6713 case Intrinsic::aarch64_neon_ld1x2:
6714 case Intrinsic::aarch64_neon_ld1x3:
6715 case Intrinsic::aarch64_neon_ld1x4:
6716 case Intrinsic::aarch64_neon_ld2lane:
6717 case Intrinsic::aarch64_neon_ld3lane:
6718 case Intrinsic::aarch64_neon_ld4lane:
6719 case Intrinsic::aarch64_neon_ld2r:
6720 case Intrinsic::aarch64_neon_ld3r:
6721 case Intrinsic::aarch64_neon_ld4r: {
6722 Info.opc = ISD::INTRINSIC_W_CHAIN;
6723 // Conservatively set memVT to the entire set of vectors loaded.
6724 uint64_t NumElts = DL.getTypeAllocSize(I.getType()) / 8;
6725 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6726 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6729 Info.vol = false; // volatile loads with NEON intrinsics not supported
6730 Info.readMem = true;
6731 Info.writeMem = false;
6734 case Intrinsic::aarch64_neon_st2:
6735 case Intrinsic::aarch64_neon_st3:
6736 case Intrinsic::aarch64_neon_st4:
6737 case Intrinsic::aarch64_neon_st1x2:
6738 case Intrinsic::aarch64_neon_st1x3:
6739 case Intrinsic::aarch64_neon_st1x4:
6740 case Intrinsic::aarch64_neon_st2lane:
6741 case Intrinsic::aarch64_neon_st3lane:
6742 case Intrinsic::aarch64_neon_st4lane: {
6743 Info.opc = ISD::INTRINSIC_VOID;
6744 // Conservatively set memVT to the entire set of vectors stored.
6745 unsigned NumElts = 0;
6746 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
6747 Type *ArgTy = I.getArgOperand(ArgI)->getType();
6748 if (!ArgTy->isVectorTy())
6750 NumElts += DL.getTypeAllocSize(ArgTy) / 8;
6752 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6753 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6756 Info.vol = false; // volatile stores with NEON intrinsics not supported
6757 Info.readMem = false;
6758 Info.writeMem = true;
6761 case Intrinsic::aarch64_ldaxr:
6762 case Intrinsic::aarch64_ldxr: {
6763 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
6764 Info.opc = ISD::INTRINSIC_W_CHAIN;
6765 Info.memVT = MVT::getVT(PtrTy->getElementType());
6766 Info.ptrVal = I.getArgOperand(0);
6768 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
6770 Info.readMem = true;
6771 Info.writeMem = false;
6774 case Intrinsic::aarch64_stlxr:
6775 case Intrinsic::aarch64_stxr: {
6776 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
6777 Info.opc = ISD::INTRINSIC_W_CHAIN;
6778 Info.memVT = MVT::getVT(PtrTy->getElementType());
6779 Info.ptrVal = I.getArgOperand(1);
6781 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
6783 Info.readMem = false;
6784 Info.writeMem = true;
6787 case Intrinsic::aarch64_ldaxp:
6788 case Intrinsic::aarch64_ldxp: {
6789 Info.opc = ISD::INTRINSIC_W_CHAIN;
6790 Info.memVT = MVT::i128;
6791 Info.ptrVal = I.getArgOperand(0);
6795 Info.readMem = true;
6796 Info.writeMem = false;
6799 case Intrinsic::aarch64_stlxp:
6800 case Intrinsic::aarch64_stxp: {
6801 Info.opc = ISD::INTRINSIC_W_CHAIN;
6802 Info.memVT = MVT::i128;
6803 Info.ptrVal = I.getArgOperand(2);
6807 Info.readMem = false;
6808 Info.writeMem = true;
6818 // Truncations from 64-bit GPR to 32-bit GPR is free.
6819 bool AArch64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
6820 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6822 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6823 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6824 return NumBits1 > NumBits2;
6826 bool AArch64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
6827 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6829 unsigned NumBits1 = VT1.getSizeInBits();
6830 unsigned NumBits2 = VT2.getSizeInBits();
6831 return NumBits1 > NumBits2;
6834 /// Check if it is profitable to hoist instruction in then/else to if.
6835 /// Not profitable if I and it's user can form a FMA instruction
6836 /// because we prefer FMSUB/FMADD.
6837 bool AArch64TargetLowering::isProfitableToHoist(Instruction *I) const {
6838 if (I->getOpcode() != Instruction::FMul)
6841 if (I->getNumUses() != 1)
6844 Instruction *User = I->user_back();
6847 !(User->getOpcode() == Instruction::FSub ||
6848 User->getOpcode() == Instruction::FAdd))
6851 const TargetOptions &Options = getTargetMachine().Options;
6852 const DataLayout &DL = I->getModule()->getDataLayout();
6853 EVT VT = getValueType(DL, User->getOperand(0)->getType());
6855 if (isFMAFasterThanFMulAndFAdd(VT) &&
6856 isOperationLegalOrCustom(ISD::FMA, VT) &&
6857 (Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath))
6863 // All 32-bit GPR operations implicitly zero the high-half of the corresponding
6865 bool AArch64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
6866 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6868 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6869 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6870 return NumBits1 == 32 && NumBits2 == 64;
6872 bool AArch64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
6873 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6875 unsigned NumBits1 = VT1.getSizeInBits();
6876 unsigned NumBits2 = VT2.getSizeInBits();
6877 return NumBits1 == 32 && NumBits2 == 64;
6880 bool AArch64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
6881 EVT VT1 = Val.getValueType();
6882 if (isZExtFree(VT1, VT2)) {
6886 if (Val.getOpcode() != ISD::LOAD)
6889 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
6890 return (VT1.isSimple() && !VT1.isVector() && VT1.isInteger() &&
6891 VT2.isSimple() && !VT2.isVector() && VT2.isInteger() &&
6892 VT1.getSizeInBits() <= 32);
6895 bool AArch64TargetLowering::isExtFreeImpl(const Instruction *Ext) const {
6896 if (isa<FPExtInst>(Ext))
6899 // Vector types are next free.
6900 if (Ext->getType()->isVectorTy())
6903 for (const Use &U : Ext->uses()) {
6904 // The extension is free if we can fold it with a left shift in an
6905 // addressing mode or an arithmetic operation: add, sub, and cmp.
6907 // Is there a shift?
6908 const Instruction *Instr = cast<Instruction>(U.getUser());
6910 // Is this a constant shift?
6911 switch (Instr->getOpcode()) {
6912 case Instruction::Shl:
6913 if (!isa<ConstantInt>(Instr->getOperand(1)))
6916 case Instruction::GetElementPtr: {
6917 gep_type_iterator GTI = gep_type_begin(Instr);
6918 auto &DL = Ext->getModule()->getDataLayout();
6919 std::advance(GTI, U.getOperandNo());
6921 // This extension will end up with a shift because of the scaling factor.
6922 // 8-bit sized types have a scaling factor of 1, thus a shift amount of 0.
6923 // Get the shift amount based on the scaling factor:
6924 // log2(sizeof(IdxTy)) - log2(8).
6926 countTrailingZeros(DL.getTypeStoreSizeInBits(IdxTy)) - 3;
6927 // Is the constant foldable in the shift of the addressing mode?
6928 // I.e., shift amount is between 1 and 4 inclusive.
6929 if (ShiftAmt == 0 || ShiftAmt > 4)
6933 case Instruction::Trunc:
6934 // Check if this is a noop.
6935 // trunc(sext ty1 to ty2) to ty1.
6936 if (Instr->getType() == Ext->getOperand(0)->getType())
6943 // At this point we can use the bfm family, so this extension is free
6949 bool AArch64TargetLowering::hasPairedLoad(Type *LoadedType,
6950 unsigned &RequiredAligment) const {
6951 if (!LoadedType->isIntegerTy() && !LoadedType->isFloatTy())
6953 // Cyclone supports unaligned accesses.
6954 RequiredAligment = 0;
6955 unsigned NumBits = LoadedType->getPrimitiveSizeInBits();
6956 return NumBits == 32 || NumBits == 64;
6959 bool AArch64TargetLowering::hasPairedLoad(EVT LoadedType,
6960 unsigned &RequiredAligment) const {
6961 if (!LoadedType.isSimple() ||
6962 (!LoadedType.isInteger() && !LoadedType.isFloatingPoint()))
6964 // Cyclone supports unaligned accesses.
6965 RequiredAligment = 0;
6966 unsigned NumBits = LoadedType.getSizeInBits();
6967 return NumBits == 32 || NumBits == 64;
6970 /// \brief Lower an interleaved load into a ldN intrinsic.
6972 /// E.g. Lower an interleaved load (Factor = 2):
6973 /// %wide.vec = load <8 x i32>, <8 x i32>* %ptr
6974 /// %v0 = shuffle %wide.vec, undef, <0, 2, 4, 6> ; Extract even elements
6975 /// %v1 = shuffle %wide.vec, undef, <1, 3, 5, 7> ; Extract odd elements
6978 /// %ld2 = { <4 x i32>, <4 x i32> } call llvm.aarch64.neon.ld2(%ptr)
6979 /// %vec0 = extractelement { <4 x i32>, <4 x i32> } %ld2, i32 0
6980 /// %vec1 = extractelement { <4 x i32>, <4 x i32> } %ld2, i32 1
6981 bool AArch64TargetLowering::lowerInterleavedLoad(
6982 LoadInst *LI, ArrayRef<ShuffleVectorInst *> Shuffles,
6983 ArrayRef<unsigned> Indices, unsigned Factor) const {
6984 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
6985 "Invalid interleave factor");
6986 assert(!Shuffles.empty() && "Empty shufflevector input");
6987 assert(Shuffles.size() == Indices.size() &&
6988 "Unmatched number of shufflevectors and indices");
6990 const DataLayout &DL = LI->getModule()->getDataLayout();
6992 VectorType *VecTy = Shuffles[0]->getType();
6993 unsigned VecSize = DL.getTypeAllocSizeInBits(VecTy);
6995 // Skip if we do not have NEON and skip illegal vector types.
6996 if (!Subtarget->hasNEON() || (VecSize != 64 && VecSize != 128))
6999 // A pointer vector can not be the return type of the ldN intrinsics. Need to
7000 // load integer vectors first and then convert to pointer vectors.
7001 Type *EltTy = VecTy->getVectorElementType();
7002 if (EltTy->isPointerTy())
7004 VectorType::get(DL.getIntPtrType(EltTy), VecTy->getVectorNumElements());
7006 Type *PtrTy = VecTy->getPointerTo(LI->getPointerAddressSpace());
7007 Type *Tys[2] = {VecTy, PtrTy};
7008 static const Intrinsic::ID LoadInts[3] = {Intrinsic::aarch64_neon_ld2,
7009 Intrinsic::aarch64_neon_ld3,
7010 Intrinsic::aarch64_neon_ld4};
7012 Intrinsic::getDeclaration(LI->getModule(), LoadInts[Factor - 2], Tys);
7014 IRBuilder<> Builder(LI);
7015 Value *Ptr = Builder.CreateBitCast(LI->getPointerOperand(), PtrTy);
7017 CallInst *LdN = Builder.CreateCall(LdNFunc, Ptr, "ldN");
7019 // Replace uses of each shufflevector with the corresponding vector loaded
7021 for (unsigned i = 0; i < Shuffles.size(); i++) {
7022 ShuffleVectorInst *SVI = Shuffles[i];
7023 unsigned Index = Indices[i];
7025 Value *SubVec = Builder.CreateExtractValue(LdN, Index);
7027 // Convert the integer vector to pointer vector if the element is pointer.
7028 if (EltTy->isPointerTy())
7029 SubVec = Builder.CreateIntToPtr(SubVec, SVI->getType());
7031 SVI->replaceAllUsesWith(SubVec);
7037 /// \brief Get a mask consisting of sequential integers starting from \p Start.
7039 /// I.e. <Start, Start + 1, ..., Start + NumElts - 1>
7040 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned Start,
7042 SmallVector<Constant *, 16> Mask;
7043 for (unsigned i = 0; i < NumElts; i++)
7044 Mask.push_back(Builder.getInt32(Start + i));
7046 return ConstantVector::get(Mask);
7049 /// \brief Lower an interleaved store into a stN intrinsic.
7051 /// E.g. Lower an interleaved store (Factor = 3):
7052 /// %i.vec = shuffle <8 x i32> %v0, <8 x i32> %v1,
7053 /// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11>
7054 /// store <12 x i32> %i.vec, <12 x i32>* %ptr
7057 /// %sub.v0 = shuffle <8 x i32> %v0, <8 x i32> v1, <0, 1, 2, 3>
7058 /// %sub.v1 = shuffle <8 x i32> %v0, <8 x i32> v1, <4, 5, 6, 7>
7059 /// %sub.v2 = shuffle <8 x i32> %v0, <8 x i32> v1, <8, 9, 10, 11>
7060 /// call void llvm.aarch64.neon.st3(%sub.v0, %sub.v1, %sub.v2, %ptr)
7062 /// Note that the new shufflevectors will be removed and we'll only generate one
7063 /// st3 instruction in CodeGen.
7064 bool AArch64TargetLowering::lowerInterleavedStore(StoreInst *SI,
7065 ShuffleVectorInst *SVI,
7066 unsigned Factor) const {
7067 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
7068 "Invalid interleave factor");
7070 VectorType *VecTy = SVI->getType();
7071 assert(VecTy->getVectorNumElements() % Factor == 0 &&
7072 "Invalid interleaved store");
7074 unsigned NumSubElts = VecTy->getVectorNumElements() / Factor;
7075 Type *EltTy = VecTy->getVectorElementType();
7076 VectorType *SubVecTy = VectorType::get(EltTy, NumSubElts);
7078 const DataLayout &DL = SI->getModule()->getDataLayout();
7079 unsigned SubVecSize = DL.getTypeAllocSizeInBits(SubVecTy);
7081 // Skip if we do not have NEON and skip illegal vector types.
7082 if (!Subtarget->hasNEON() || (SubVecSize != 64 && SubVecSize != 128))
7085 Value *Op0 = SVI->getOperand(0);
7086 Value *Op1 = SVI->getOperand(1);
7087 IRBuilder<> Builder(SI);
7089 // StN intrinsics don't support pointer vectors as arguments. Convert pointer
7090 // vectors to integer vectors.
7091 if (EltTy->isPointerTy()) {
7092 Type *IntTy = DL.getIntPtrType(EltTy);
7093 unsigned NumOpElts =
7094 dyn_cast<VectorType>(Op0->getType())->getVectorNumElements();
7096 // Convert to the corresponding integer vector.
7097 Type *IntVecTy = VectorType::get(IntTy, NumOpElts);
7098 Op0 = Builder.CreatePtrToInt(Op0, IntVecTy);
7099 Op1 = Builder.CreatePtrToInt(Op1, IntVecTy);
7101 SubVecTy = VectorType::get(IntTy, NumSubElts);
7104 Type *PtrTy = SubVecTy->getPointerTo(SI->getPointerAddressSpace());
7105 Type *Tys[2] = {SubVecTy, PtrTy};
7106 static const Intrinsic::ID StoreInts[3] = {Intrinsic::aarch64_neon_st2,
7107 Intrinsic::aarch64_neon_st3,
7108 Intrinsic::aarch64_neon_st4};
7110 Intrinsic::getDeclaration(SI->getModule(), StoreInts[Factor - 2], Tys);
7112 SmallVector<Value *, 5> Ops;
7114 // Split the shufflevector operands into sub vectors for the new stN call.
7115 for (unsigned i = 0; i < Factor; i++)
7116 Ops.push_back(Builder.CreateShuffleVector(
7117 Op0, Op1, getSequentialMask(Builder, NumSubElts * i, NumSubElts)));
7119 Ops.push_back(Builder.CreateBitCast(SI->getPointerOperand(), PtrTy));
7120 Builder.CreateCall(StNFunc, Ops);
7124 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
7125 unsigned AlignCheck) {
7126 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
7127 (DstAlign == 0 || DstAlign % AlignCheck == 0));
7130 EVT AArch64TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
7131 unsigned SrcAlign, bool IsMemset,
7134 MachineFunction &MF) const {
7135 // Don't use AdvSIMD to implement 16-byte memset. It would have taken one
7136 // instruction to materialize the v2i64 zero and one store (with restrictive
7137 // addressing mode). Just do two i64 store of zero-registers.
7139 const Function *F = MF.getFunction();
7140 if (Subtarget->hasFPARMv8() && !IsMemset && Size >= 16 &&
7141 !F->hasFnAttribute(Attribute::NoImplicitFloat) &&
7142 (memOpAlign(SrcAlign, DstAlign, 16) ||
7143 (allowsMisalignedMemoryAccesses(MVT::f128, 0, 1, &Fast) && Fast)))
7147 (memOpAlign(SrcAlign, DstAlign, 8) ||
7148 (allowsMisalignedMemoryAccesses(MVT::i64, 0, 1, &Fast) && Fast)))
7152 (memOpAlign(SrcAlign, DstAlign, 4) ||
7153 (allowsMisalignedMemoryAccesses(MVT::i32, 0, 1, &Fast) && Fast)))
7159 // 12-bit optionally shifted immediates are legal for adds.
7160 bool AArch64TargetLowering::isLegalAddImmediate(int64_t Immed) const {
7161 if ((Immed >> 12) == 0 || ((Immed & 0xfff) == 0 && Immed >> 24 == 0))
7166 // Integer comparisons are implemented with ADDS/SUBS, so the range of valid
7167 // immediates is the same as for an add or a sub.
7168 bool AArch64TargetLowering::isLegalICmpImmediate(int64_t Immed) const {
7171 return isLegalAddImmediate(Immed);
7174 /// isLegalAddressingMode - Return true if the addressing mode represented
7175 /// by AM is legal for this target, for a load/store of the specified type.
7176 bool AArch64TargetLowering::isLegalAddressingMode(const DataLayout &DL,
7177 const AddrMode &AM, Type *Ty,
7178 unsigned AS) const {
7179 // AArch64 has five basic addressing modes:
7181 // reg + 9-bit signed offset
7182 // reg + SIZE_IN_BYTES * 12-bit unsigned offset
7184 // reg + SIZE_IN_BYTES * reg
7186 // No global is ever allowed as a base.
7190 // No reg+reg+imm addressing.
7191 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
7194 // check reg + imm case:
7195 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
7196 uint64_t NumBytes = 0;
7197 if (Ty->isSized()) {
7198 uint64_t NumBits = DL.getTypeSizeInBits(Ty);
7199 NumBytes = NumBits / 8;
7200 if (!isPowerOf2_64(NumBits))
7205 int64_t Offset = AM.BaseOffs;
7207 // 9-bit signed offset
7208 if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
7211 // 12-bit unsigned offset
7212 unsigned shift = Log2_64(NumBytes);
7213 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
7214 // Must be a multiple of NumBytes (NumBytes is a power of 2)
7215 (Offset >> shift) << shift == Offset)
7220 // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2
7222 if (!AM.Scale || AM.Scale == 1 ||
7223 (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
7228 int AArch64TargetLowering::getScalingFactorCost(const DataLayout &DL,
7229 const AddrMode &AM, Type *Ty,
7230 unsigned AS) const {
7231 // Scaling factors are not free at all.
7232 // Operands | Rt Latency
7233 // -------------------------------------------
7235 // -------------------------------------------
7236 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
7237 // Rt, [Xn, Wm, <extend> #imm] |
7238 if (isLegalAddressingMode(DL, AM, Ty, AS))
7239 // Scale represents reg2 * scale, thus account for 1 if
7240 // it is not equal to 0 or 1.
7241 return AM.Scale != 0 && AM.Scale != 1;
7245 bool AArch64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
7246 VT = VT.getScalarType();
7251 switch (VT.getSimpleVT().SimpleTy) {
7263 AArch64TargetLowering::getScratchRegisters(CallingConv::ID) const {
7264 // LR is a callee-save register, but we must treat it as clobbered by any call
7265 // site. Hence we include LR in the scratch registers, which are in turn added
7266 // as implicit-defs for stackmaps and patchpoints.
7267 static const MCPhysReg ScratchRegs[] = {
7268 AArch64::X16, AArch64::X17, AArch64::LR, 0
7274 AArch64TargetLowering::isDesirableToCommuteWithShift(const SDNode *N) const {
7275 EVT VT = N->getValueType(0);
7276 // If N is unsigned bit extraction: ((x >> C) & mask), then do not combine
7277 // it with shift to let it be lowered to UBFX.
7278 if (N->getOpcode() == ISD::AND && (VT == MVT::i32 || VT == MVT::i64) &&
7279 isa<ConstantSDNode>(N->getOperand(1))) {
7280 uint64_t TruncMask = N->getConstantOperandVal(1);
7281 if (isMask_64(TruncMask) &&
7282 N->getOperand(0).getOpcode() == ISD::SRL &&
7283 isa<ConstantSDNode>(N->getOperand(0)->getOperand(1)))
7289 bool AArch64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
7291 assert(Ty->isIntegerTy());
7293 unsigned BitSize = Ty->getPrimitiveSizeInBits();
7297 int64_t Val = Imm.getSExtValue();
7298 if (Val == 0 || AArch64_AM::isLogicalImmediate(Val, BitSize))
7301 if ((int64_t)Val < 0)
7304 Val &= (1LL << 32) - 1;
7306 unsigned LZ = countLeadingZeros((uint64_t)Val);
7307 unsigned Shift = (63 - LZ) / 16;
7308 // MOVZ is free so return true for one or fewer MOVK.
7312 // Generate SUBS and CSEL for integer abs.
7313 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
7314 EVT VT = N->getValueType(0);
7316 SDValue N0 = N->getOperand(0);
7317 SDValue N1 = N->getOperand(1);
7320 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
7321 // and change it to SUB and CSEL.
7322 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
7323 N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
7324 N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0))
7325 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
7326 if (Y1C->getAPIntValue() == VT.getSizeInBits() - 1) {
7327 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT),
7329 // Generate SUBS & CSEL.
7331 DAG.getNode(AArch64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32),
7332 N0.getOperand(0), DAG.getConstant(0, DL, VT));
7333 return DAG.getNode(AArch64ISD::CSEL, DL, VT, N0.getOperand(0), Neg,
7334 DAG.getConstant(AArch64CC::PL, DL, MVT::i32),
7335 SDValue(Cmp.getNode(), 1));
7340 // performXorCombine - Attempts to handle integer ABS.
7341 static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
7342 TargetLowering::DAGCombinerInfo &DCI,
7343 const AArch64Subtarget *Subtarget) {
7344 if (DCI.isBeforeLegalizeOps())
7347 return performIntegerAbsCombine(N, DAG);
7351 AArch64TargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor,
7353 std::vector<SDNode *> *Created) const {
7354 // fold (sdiv X, pow2)
7355 EVT VT = N->getValueType(0);
7356 if ((VT != MVT::i32 && VT != MVT::i64) ||
7357 !(Divisor.isPowerOf2() || (-Divisor).isPowerOf2()))
7361 SDValue N0 = N->getOperand(0);
7362 unsigned Lg2 = Divisor.countTrailingZeros();
7363 SDValue Zero = DAG.getConstant(0, DL, VT);
7364 SDValue Pow2MinusOne = DAG.getConstant((1ULL << Lg2) - 1, DL, VT);
7366 // Add (N0 < 0) ? Pow2 - 1 : 0;
7368 SDValue Cmp = getAArch64Cmp(N0, Zero, ISD::SETLT, CCVal, DAG, DL);
7369 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0, Pow2MinusOne);
7370 SDValue CSel = DAG.getNode(AArch64ISD::CSEL, DL, VT, Add, N0, CCVal, Cmp);
7373 Created->push_back(Cmp.getNode());
7374 Created->push_back(Add.getNode());
7375 Created->push_back(CSel.getNode());
7380 DAG.getNode(ISD::SRA, DL, VT, CSel, DAG.getConstant(Lg2, DL, MVT::i64));
7382 // If we're dividing by a positive value, we're done. Otherwise, we must
7383 // negate the result.
7384 if (Divisor.isNonNegative())
7388 Created->push_back(SRA.getNode());
7389 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), SRA);
7392 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG,
7393 TargetLowering::DAGCombinerInfo &DCI,
7394 const AArch64Subtarget *Subtarget) {
7395 if (DCI.isBeforeLegalizeOps())
7398 // Multiplication of a power of two plus/minus one can be done more
7399 // cheaply as as shift+add/sub. For now, this is true unilaterally. If
7400 // future CPUs have a cheaper MADD instruction, this may need to be
7401 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
7402 // 64-bit is 5 cycles, so this is always a win.
7403 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
7404 APInt Value = C->getAPIntValue();
7405 EVT VT = N->getValueType(0);
7407 if (Value.isNonNegative()) {
7408 // (mul x, 2^N + 1) => (add (shl x, N), x)
7409 APInt VM1 = Value - 1;
7410 if (VM1.isPowerOf2()) {
7411 SDValue ShiftedVal =
7412 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7413 DAG.getConstant(VM1.logBase2(), DL, MVT::i64));
7414 return DAG.getNode(ISD::ADD, DL, VT, ShiftedVal,
7417 // (mul x, 2^N - 1) => (sub (shl x, N), x)
7418 APInt VP1 = Value + 1;
7419 if (VP1.isPowerOf2()) {
7420 SDValue ShiftedVal =
7421 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7422 DAG.getConstant(VP1.logBase2(), DL, MVT::i64));
7423 return DAG.getNode(ISD::SUB, DL, VT, ShiftedVal,
7427 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
7428 APInt VNP1 = -Value + 1;
7429 if (VNP1.isPowerOf2()) {
7430 SDValue ShiftedVal =
7431 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7432 DAG.getConstant(VNP1.logBase2(), DL, MVT::i64));
7433 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0),
7436 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
7437 APInt VNM1 = -Value - 1;
7438 if (VNM1.isPowerOf2()) {
7439 SDValue ShiftedVal =
7440 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
7441 DAG.getConstant(VNM1.logBase2(), DL, MVT::i64));
7443 DAG.getNode(ISD::ADD, DL, VT, ShiftedVal, N->getOperand(0));
7444 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Add);
7451 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
7452 SelectionDAG &DAG) {
7453 // Take advantage of vector comparisons producing 0 or -1 in each lane to
7454 // optimize away operation when it's from a constant.
7456 // The general transformation is:
7457 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
7458 // AND(VECTOR_CMP(x,y), constant2)
7459 // constant2 = UNARYOP(constant)
7461 // Early exit if this isn't a vector operation, the operand of the
7462 // unary operation isn't a bitwise AND, or if the sizes of the operations
7464 EVT VT = N->getValueType(0);
7465 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
7466 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
7467 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
7470 // Now check that the other operand of the AND is a constant. We could
7471 // make the transformation for non-constant splats as well, but it's unclear
7472 // that would be a benefit as it would not eliminate any operations, just
7473 // perform one more step in scalar code before moving to the vector unit.
7474 if (BuildVectorSDNode *BV =
7475 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
7476 // Bail out if the vector isn't a constant.
7477 if (!BV->isConstant())
7480 // Everything checks out. Build up the new and improved node.
7482 EVT IntVT = BV->getValueType(0);
7483 // Create a new constant of the appropriate type for the transformed
7485 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
7486 // The AND node needs bitcasts to/from an integer vector type around it.
7487 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
7488 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
7489 N->getOperand(0)->getOperand(0), MaskConst);
7490 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
7497 static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG,
7498 const AArch64Subtarget *Subtarget) {
7499 // First try to optimize away the conversion when it's conditionally from
7500 // a constant. Vectors only.
7501 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
7504 EVT VT = N->getValueType(0);
7505 if (VT != MVT::f32 && VT != MVT::f64)
7508 // Only optimize when the source and destination types have the same width.
7509 if (VT.getSizeInBits() != N->getOperand(0).getValueType().getSizeInBits())
7512 // If the result of an integer load is only used by an integer-to-float
7513 // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
7514 // This eliminates an "integer-to-vector-move" UOP and improves throughput.
7515 SDValue N0 = N->getOperand(0);
7516 if (Subtarget->hasNEON() && ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7517 // Do not change the width of a volatile load.
7518 !cast<LoadSDNode>(N0)->isVolatile()) {
7519 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7520 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
7521 LN0->getPointerInfo(), LN0->isVolatile(),
7522 LN0->isNonTemporal(), LN0->isInvariant(),
7523 LN0->getAlignment());
7525 // Make sure successors of the original load stay after it by updating them
7526 // to use the new Chain.
7527 DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));
7530 (N->getOpcode() == ISD::SINT_TO_FP) ? AArch64ISD::SITOF : AArch64ISD::UITOF;
7531 return DAG.getNode(Opcode, SDLoc(N), VT, Load);
7537 /// Fold a floating-point multiply by power of two into floating-point to
7538 /// fixed-point conversion.
7539 static SDValue performFpToIntCombine(SDNode *N, SelectionDAG &DAG,
7540 const AArch64Subtarget *Subtarget) {
7541 if (!Subtarget->hasNEON())
7544 SDValue Op = N->getOperand(0);
7545 if (!Op.getValueType().isVector() || Op.getOpcode() != ISD::FMUL)
7548 SDValue ConstVec = Op->getOperand(1);
7549 if (!isa<BuildVectorSDNode>(ConstVec))
7552 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
7553 uint32_t FloatBits = FloatTy.getSizeInBits();
7554 if (FloatBits != 32 && FloatBits != 64)
7557 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
7558 uint32_t IntBits = IntTy.getSizeInBits();
7559 if (IntBits != 16 && IntBits != 32 && IntBits != 64)
7562 // Avoid conversions where iN is larger than the float (e.g., float -> i64).
7563 if (IntBits > FloatBits)
7566 BitVector UndefElements;
7567 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(ConstVec);
7568 int32_t Bits = IntBits == 64 ? 64 : 32;
7569 int32_t C = BV->getConstantFPSplatPow2ToLog2Int(&UndefElements, Bits + 1);
7570 if (C == -1 || C == 0 || C > Bits)
7574 unsigned NumLanes = Op.getValueType().getVectorNumElements();
7579 ResTy = FloatBits == 32 ? MVT::v2i32 : MVT::v2i64;
7587 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
7588 unsigned IntrinsicOpcode = IsSigned ? Intrinsic::aarch64_neon_vcvtfp2fxs
7589 : Intrinsic::aarch64_neon_vcvtfp2fxu;
7591 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, ResTy,
7592 DAG.getConstant(IntrinsicOpcode, DL, MVT::i32),
7593 Op->getOperand(0), DAG.getConstant(C, DL, MVT::i32));
7594 // We can handle smaller integers by generating an extra trunc.
7595 if (IntBits < FloatBits)
7596 FixConv = DAG.getNode(ISD::TRUNCATE, DL, N->getValueType(0), FixConv);
7601 /// Fold a floating-point divide by power of two into fixed-point to
7602 /// floating-point conversion.
7603 static SDValue performFDivCombine(SDNode *N, SelectionDAG &DAG,
7604 const AArch64Subtarget *Subtarget) {
7605 if (!Subtarget->hasNEON())
7608 SDValue Op = N->getOperand(0);
7609 unsigned Opc = Op->getOpcode();
7610 if (!Op.getValueType().isVector() ||
7611 (Opc != ISD::SINT_TO_FP && Opc != ISD::UINT_TO_FP))
7614 SDValue ConstVec = N->getOperand(1);
7615 if (!isa<BuildVectorSDNode>(ConstVec))
7618 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
7619 int32_t IntBits = IntTy.getSizeInBits();
7620 if (IntBits != 16 && IntBits != 32 && IntBits != 64)
7623 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
7624 int32_t FloatBits = FloatTy.getSizeInBits();
7625 if (FloatBits != 32 && FloatBits != 64)
7628 // Avoid conversions where iN is larger than the float (e.g., i64 -> float).
7629 if (IntBits > FloatBits)
7632 BitVector UndefElements;
7633 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(ConstVec);
7634 int32_t C = BV->getConstantFPSplatPow2ToLog2Int(&UndefElements, FloatBits + 1);
7635 if (C == -1 || C == 0 || C > FloatBits)
7639 unsigned NumLanes = Op.getValueType().getVectorNumElements();
7644 ResTy = FloatBits == 32 ? MVT::v2i32 : MVT::v2i64;
7652 SDValue ConvInput = Op.getOperand(0);
7653 bool IsSigned = Opc == ISD::SINT_TO_FP;
7654 if (IntBits < FloatBits)
7655 ConvInput = DAG.getNode(IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND, DL,
7658 unsigned IntrinsicOpcode = IsSigned ? Intrinsic::aarch64_neon_vcvtfxs2fp
7659 : Intrinsic::aarch64_neon_vcvtfxu2fp;
7660 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, Op.getValueType(),
7661 DAG.getConstant(IntrinsicOpcode, DL, MVT::i32), ConvInput,
7662 DAG.getConstant(C, DL, MVT::i32));
7665 /// An EXTR instruction is made up of two shifts, ORed together. This helper
7666 /// searches for and classifies those shifts.
7667 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
7669 if (N.getOpcode() == ISD::SHL)
7671 else if (N.getOpcode() == ISD::SRL)
7676 if (!isa<ConstantSDNode>(N.getOperand(1)))
7679 ShiftAmount = N->getConstantOperandVal(1);
7680 Src = N->getOperand(0);
7684 /// EXTR instruction extracts a contiguous chunk of bits from two existing
7685 /// registers viewed as a high/low pair. This function looks for the pattern:
7686 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
7687 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
7689 static SDValue tryCombineToEXTR(SDNode *N,
7690 TargetLowering::DAGCombinerInfo &DCI) {
7691 SelectionDAG &DAG = DCI.DAG;
7693 EVT VT = N->getValueType(0);
7695 assert(N->getOpcode() == ISD::OR && "Unexpected root");
7697 if (VT != MVT::i32 && VT != MVT::i64)
7701 uint32_t ShiftLHS = 0;
7703 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
7707 uint32_t ShiftRHS = 0;
7709 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
7712 // If they're both trying to come from the high part of the register, they're
7713 // not really an EXTR.
7714 if (LHSFromHi == RHSFromHi)
7717 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
7721 std::swap(LHS, RHS);
7722 std::swap(ShiftLHS, ShiftRHS);
7725 return DAG.getNode(AArch64ISD::EXTR, DL, VT, LHS, RHS,
7726 DAG.getConstant(ShiftRHS, DL, MVT::i64));
7729 static SDValue tryCombineToBSL(SDNode *N,
7730 TargetLowering::DAGCombinerInfo &DCI) {
7731 EVT VT = N->getValueType(0);
7732 SelectionDAG &DAG = DCI.DAG;
7738 SDValue N0 = N->getOperand(0);
7739 if (N0.getOpcode() != ISD::AND)
7742 SDValue N1 = N->getOperand(1);
7743 if (N1.getOpcode() != ISD::AND)
7746 // We only have to look for constant vectors here since the general, variable
7747 // case can be handled in TableGen.
7748 unsigned Bits = VT.getVectorElementType().getSizeInBits();
7749 uint64_t BitMask = Bits == 64 ? -1ULL : ((1ULL << Bits) - 1);
7750 for (int i = 1; i >= 0; --i)
7751 for (int j = 1; j >= 0; --j) {
7752 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(i));
7753 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(j));
7757 bool FoundMatch = true;
7758 for (unsigned k = 0; k < VT.getVectorNumElements(); ++k) {
7759 ConstantSDNode *CN0 = dyn_cast<ConstantSDNode>(BVN0->getOperand(k));
7760 ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(BVN1->getOperand(k));
7762 CN0->getZExtValue() != (BitMask & ~CN1->getZExtValue())) {
7769 return DAG.getNode(AArch64ISD::BSL, DL, VT, SDValue(BVN0, 0),
7770 N0->getOperand(1 - i), N1->getOperand(1 - j));
7776 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
7777 const AArch64Subtarget *Subtarget) {
7778 // Attempt to form an EXTR from (or (shl VAL1, #N), (srl VAL2, #RegWidth-N))
7779 if (!EnableAArch64ExtrGeneration)
7781 SelectionDAG &DAG = DCI.DAG;
7782 EVT VT = N->getValueType(0);
7784 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7787 SDValue Res = tryCombineToEXTR(N, DCI);
7791 Res = tryCombineToBSL(N, DCI);
7798 static SDValue performBitcastCombine(SDNode *N,
7799 TargetLowering::DAGCombinerInfo &DCI,
7800 SelectionDAG &DAG) {
7801 // Wait 'til after everything is legalized to try this. That way we have
7802 // legal vector types and such.
7803 if (DCI.isBeforeLegalizeOps())
7806 // Remove extraneous bitcasts around an extract_subvector.
7808 // (v4i16 (bitconvert
7809 // (extract_subvector (v2i64 (bitconvert (v8i16 ...)), (i64 1)))))
7811 // (extract_subvector ((v8i16 ...), (i64 4)))
7813 // Only interested in 64-bit vectors as the ultimate result.
7814 EVT VT = N->getValueType(0);
7817 if (VT.getSimpleVT().getSizeInBits() != 64)
7819 // Is the operand an extract_subvector starting at the beginning or halfway
7820 // point of the vector? A low half may also come through as an
7821 // EXTRACT_SUBREG, so look for that, too.
7822 SDValue Op0 = N->getOperand(0);
7823 if (Op0->getOpcode() != ISD::EXTRACT_SUBVECTOR &&
7824 !(Op0->isMachineOpcode() &&
7825 Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG))
7827 uint64_t idx = cast<ConstantSDNode>(Op0->getOperand(1))->getZExtValue();
7828 if (Op0->getOpcode() == ISD::EXTRACT_SUBVECTOR) {
7829 if (Op0->getValueType(0).getVectorNumElements() != idx && idx != 0)
7831 } else if (Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG) {
7832 if (idx != AArch64::dsub)
7834 // The dsub reference is equivalent to a lane zero subvector reference.
7837 // Look through the bitcast of the input to the extract.
7838 if (Op0->getOperand(0)->getOpcode() != ISD::BITCAST)
7840 SDValue Source = Op0->getOperand(0)->getOperand(0);
7841 // If the source type has twice the number of elements as our destination
7842 // type, we know this is an extract of the high or low half of the vector.
7843 EVT SVT = Source->getValueType(0);
7844 if (SVT.getVectorNumElements() != VT.getVectorNumElements() * 2)
7847 DEBUG(dbgs() << "aarch64-lower: bitcast extract_subvector simplification\n");
7849 // Create the simplified form to just extract the low or high half of the
7850 // vector directly rather than bothering with the bitcasts.
7852 unsigned NumElements = VT.getVectorNumElements();
7854 SDValue HalfIdx = DAG.getConstant(NumElements, dl, MVT::i64);
7855 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Source, HalfIdx);
7857 SDValue SubReg = DAG.getTargetConstant(AArch64::dsub, dl, MVT::i32);
7858 return SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl, VT,
7864 static SDValue performConcatVectorsCombine(SDNode *N,
7865 TargetLowering::DAGCombinerInfo &DCI,
7866 SelectionDAG &DAG) {
7868 EVT VT = N->getValueType(0);
7869 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
7871 // Optimize concat_vectors of truncated vectors, where the intermediate
7872 // type is illegal, to avoid said illegality, e.g.,
7873 // (v4i16 (concat_vectors (v2i16 (truncate (v2i64))),
7874 // (v2i16 (truncate (v2i64)))))
7876 // (v4i16 (truncate (vector_shuffle (v4i32 (bitcast (v2i64))),
7877 // (v4i32 (bitcast (v2i64))),
7879 // This isn't really target-specific, but ISD::TRUNCATE legality isn't keyed
7880 // on both input and result type, so we might generate worse code.
7881 // On AArch64 we know it's fine for v2i64->v4i16 and v4i32->v8i8.
7882 if (N->getNumOperands() == 2 &&
7883 N0->getOpcode() == ISD::TRUNCATE &&
7884 N1->getOpcode() == ISD::TRUNCATE) {
7885 SDValue N00 = N0->getOperand(0);
7886 SDValue N10 = N1->getOperand(0);
7887 EVT N00VT = N00.getValueType();
7889 if (N00VT == N10.getValueType() &&
7890 (N00VT == MVT::v2i64 || N00VT == MVT::v4i32) &&
7891 N00VT.getScalarSizeInBits() == 4 * VT.getScalarSizeInBits()) {
7892 MVT MidVT = (N00VT == MVT::v2i64 ? MVT::v4i32 : MVT::v8i16);
7893 SmallVector<int, 8> Mask(MidVT.getVectorNumElements());
7894 for (size_t i = 0; i < Mask.size(); ++i)
7896 return DAG.getNode(ISD::TRUNCATE, dl, VT,
7897 DAG.getVectorShuffle(
7899 DAG.getNode(ISD::BITCAST, dl, MidVT, N00),
7900 DAG.getNode(ISD::BITCAST, dl, MidVT, N10), Mask));
7904 // Wait 'til after everything is legalized to try this. That way we have
7905 // legal vector types and such.
7906 if (DCI.isBeforeLegalizeOps())
7909 // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector
7910 // splat. The indexed instructions are going to be expecting a DUPLANE64, so
7911 // canonicalise to that.
7912 if (N0 == N1 && VT.getVectorNumElements() == 2) {
7913 assert(VT.getVectorElementType().getSizeInBits() == 64);
7914 return DAG.getNode(AArch64ISD::DUPLANE64, dl, VT, WidenVector(N0, DAG),
7915 DAG.getConstant(0, dl, MVT::i64));
7918 // Canonicalise concat_vectors so that the right-hand vector has as few
7919 // bit-casts as possible before its real operation. The primary matching
7920 // destination for these operations will be the narrowing "2" instructions,
7921 // which depend on the operation being performed on this right-hand vector.
7923 // (concat_vectors LHS, (v1i64 (bitconvert (v4i16 RHS))))
7925 // (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS))
7927 if (N1->getOpcode() != ISD::BITCAST)
7929 SDValue RHS = N1->getOperand(0);
7930 MVT RHSTy = RHS.getValueType().getSimpleVT();
7931 // If the RHS is not a vector, this is not the pattern we're looking for.
7932 if (!RHSTy.isVector())
7935 DEBUG(dbgs() << "aarch64-lower: concat_vectors bitcast simplification\n");
7937 MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(),
7938 RHSTy.getVectorNumElements() * 2);
7939 return DAG.getNode(ISD::BITCAST, dl, VT,
7940 DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatTy,
7941 DAG.getNode(ISD::BITCAST, dl, RHSTy, N0),
7945 static SDValue tryCombineFixedPointConvert(SDNode *N,
7946 TargetLowering::DAGCombinerInfo &DCI,
7947 SelectionDAG &DAG) {
7948 // Wait 'til after everything is legalized to try this. That way we have
7949 // legal vector types and such.
7950 if (DCI.isBeforeLegalizeOps())
7952 // Transform a scalar conversion of a value from a lane extract into a
7953 // lane extract of a vector conversion. E.g., from foo1 to foo2:
7954 // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); }
7955 // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; }
7957 // The second form interacts better with instruction selection and the
7958 // register allocator to avoid cross-class register copies that aren't
7959 // coalescable due to a lane reference.
7961 // Check the operand and see if it originates from a lane extract.
7962 SDValue Op1 = N->getOperand(1);
7963 if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7964 // Yep, no additional predication needed. Perform the transform.
7965 SDValue IID = N->getOperand(0);
7966 SDValue Shift = N->getOperand(2);
7967 SDValue Vec = Op1.getOperand(0);
7968 SDValue Lane = Op1.getOperand(1);
7969 EVT ResTy = N->getValueType(0);
7973 // The vector width should be 128 bits by the time we get here, even
7974 // if it started as 64 bits (the extract_vector handling will have
7976 assert(Vec.getValueType().getSizeInBits() == 128 &&
7977 "unexpected vector size on extract_vector_elt!");
7978 if (Vec.getValueType() == MVT::v4i32)
7979 VecResTy = MVT::v4f32;
7980 else if (Vec.getValueType() == MVT::v2i64)
7981 VecResTy = MVT::v2f64;
7983 llvm_unreachable("unexpected vector type!");
7986 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift);
7987 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane);
7992 // AArch64 high-vector "long" operations are formed by performing the non-high
7993 // version on an extract_subvector of each operand which gets the high half:
7995 // (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS))
7997 // However, there are cases which don't have an extract_high explicitly, but
7998 // have another operation that can be made compatible with one for free. For
8001 // (dupv64 scalar) --> (extract_high (dup128 scalar))
8003 // This routine does the actual conversion of such DUPs, once outer routines
8004 // have determined that everything else is in order.
8005 // It also supports immediate DUP-like nodes (MOVI/MVNi), which we can fold
8007 static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) {
8008 switch (N.getOpcode()) {
8009 case AArch64ISD::DUP:
8010 case AArch64ISD::DUPLANE8:
8011 case AArch64ISD::DUPLANE16:
8012 case AArch64ISD::DUPLANE32:
8013 case AArch64ISD::DUPLANE64:
8014 case AArch64ISD::MOVI:
8015 case AArch64ISD::MOVIshift:
8016 case AArch64ISD::MOVIedit:
8017 case AArch64ISD::MOVImsl:
8018 case AArch64ISD::MVNIshift:
8019 case AArch64ISD::MVNImsl:
8022 // FMOV could be supported, but isn't very useful, as it would only occur
8023 // if you passed a bitcast' floating point immediate to an eligible long
8024 // integer op (addl, smull, ...).
8028 MVT NarrowTy = N.getSimpleValueType();
8029 if (!NarrowTy.is64BitVector())
8032 MVT ElementTy = NarrowTy.getVectorElementType();
8033 unsigned NumElems = NarrowTy.getVectorNumElements();
8034 MVT NewVT = MVT::getVectorVT(ElementTy, NumElems * 2);
8037 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NarrowTy,
8038 DAG.getNode(N->getOpcode(), dl, NewVT, N->ops()),
8039 DAG.getConstant(NumElems, dl, MVT::i64));
8042 static bool isEssentiallyExtractSubvector(SDValue N) {
8043 if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR)
8046 return N.getOpcode() == ISD::BITCAST &&
8047 N.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR;
8050 /// \brief Helper structure to keep track of ISD::SET_CC operands.
8051 struct GenericSetCCInfo {
8052 const SDValue *Opnd0;
8053 const SDValue *Opnd1;
8057 /// \brief Helper structure to keep track of a SET_CC lowered into AArch64 code.
8058 struct AArch64SetCCInfo {
8060 AArch64CC::CondCode CC;
8063 /// \brief Helper structure to keep track of SetCC information.
8065 GenericSetCCInfo Generic;
8066 AArch64SetCCInfo AArch64;
8069 /// \brief Helper structure to be able to read SetCC information. If set to
8070 /// true, IsAArch64 field, Info is a AArch64SetCCInfo, otherwise Info is a
8071 /// GenericSetCCInfo.
8072 struct SetCCInfoAndKind {
8077 /// \brief Check whether or not \p Op is a SET_CC operation, either a generic or
8079 /// AArch64 lowered one.
8080 /// \p SetCCInfo is filled accordingly.
8081 /// \post SetCCInfo is meanginfull only when this function returns true.
8082 /// \return True when Op is a kind of SET_CC operation.
8083 static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) {
8084 // If this is a setcc, this is straight forward.
8085 if (Op.getOpcode() == ISD::SETCC) {
8086 SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0);
8087 SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1);
8088 SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
8089 SetCCInfo.IsAArch64 = false;
8092 // Otherwise, check if this is a matching csel instruction.
8096 if (Op.getOpcode() != AArch64ISD::CSEL)
8098 // Set the information about the operands.
8099 // TODO: we want the operands of the Cmp not the csel
8100 SetCCInfo.Info.AArch64.Cmp = &Op.getOperand(3);
8101 SetCCInfo.IsAArch64 = true;
8102 SetCCInfo.Info.AArch64.CC = static_cast<AArch64CC::CondCode>(
8103 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
8105 // Check that the operands matches the constraints:
8106 // (1) Both operands must be constants.
8107 // (2) One must be 1 and the other must be 0.
8108 ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0));
8109 ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1));
8112 if (!TValue || !FValue)
8116 if (!TValue->isOne()) {
8117 // Update the comparison when we are interested in !cc.
8118 std::swap(TValue, FValue);
8119 SetCCInfo.Info.AArch64.CC =
8120 AArch64CC::getInvertedCondCode(SetCCInfo.Info.AArch64.CC);
8122 return TValue->isOne() && FValue->isNullValue();
8125 // Returns true if Op is setcc or zext of setcc.
8126 static bool isSetCCOrZExtSetCC(const SDValue& Op, SetCCInfoAndKind &Info) {
8127 if (isSetCC(Op, Info))
8129 return ((Op.getOpcode() == ISD::ZERO_EXTEND) &&
8130 isSetCC(Op->getOperand(0), Info));
8133 // The folding we want to perform is:
8134 // (add x, [zext] (setcc cc ...) )
8136 // (csel x, (add x, 1), !cc ...)
8138 // The latter will get matched to a CSINC instruction.
8139 static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) {
8140 assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!");
8141 SDValue LHS = Op->getOperand(0);
8142 SDValue RHS = Op->getOperand(1);
8143 SetCCInfoAndKind InfoAndKind;
8145 // If neither operand is a SET_CC, give up.
8146 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind)) {
8147 std::swap(LHS, RHS);
8148 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind))
8152 // FIXME: This could be generatized to work for FP comparisons.
8153 EVT CmpVT = InfoAndKind.IsAArch64
8154 ? InfoAndKind.Info.AArch64.Cmp->getOperand(0).getValueType()
8155 : InfoAndKind.Info.Generic.Opnd0->getValueType();
8156 if (CmpVT != MVT::i32 && CmpVT != MVT::i64)
8162 if (InfoAndKind.IsAArch64) {
8163 CCVal = DAG.getConstant(
8164 AArch64CC::getInvertedCondCode(InfoAndKind.Info.AArch64.CC), dl,
8166 Cmp = *InfoAndKind.Info.AArch64.Cmp;
8168 Cmp = getAArch64Cmp(*InfoAndKind.Info.Generic.Opnd0,
8169 *InfoAndKind.Info.Generic.Opnd1,
8170 ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, true),
8173 EVT VT = Op->getValueType(0);
8174 LHS = DAG.getNode(ISD::ADD, dl, VT, RHS, DAG.getConstant(1, dl, VT));
8175 return DAG.getNode(AArch64ISD::CSEL, dl, VT, RHS, LHS, CCVal, Cmp);
8178 // The basic add/sub long vector instructions have variants with "2" on the end
8179 // which act on the high-half of their inputs. They are normally matched by
8182 // (add (zeroext (extract_high LHS)),
8183 // (zeroext (extract_high RHS)))
8184 // -> uaddl2 vD, vN, vM
8186 // However, if one of the extracts is something like a duplicate, this
8187 // instruction can still be used profitably. This function puts the DAG into a
8188 // more appropriate form for those patterns to trigger.
8189 static SDValue performAddSubLongCombine(SDNode *N,
8190 TargetLowering::DAGCombinerInfo &DCI,
8191 SelectionDAG &DAG) {
8192 if (DCI.isBeforeLegalizeOps())
8195 MVT VT = N->getSimpleValueType(0);
8196 if (!VT.is128BitVector()) {
8197 if (N->getOpcode() == ISD::ADD)
8198 return performSetccAddFolding(N, DAG);
8202 // Make sure both branches are extended in the same way.
8203 SDValue LHS = N->getOperand(0);
8204 SDValue RHS = N->getOperand(1);
8205 if ((LHS.getOpcode() != ISD::ZERO_EXTEND &&
8206 LHS.getOpcode() != ISD::SIGN_EXTEND) ||
8207 LHS.getOpcode() != RHS.getOpcode())
8210 unsigned ExtType = LHS.getOpcode();
8212 // It's not worth doing if at least one of the inputs isn't already an
8213 // extract, but we don't know which it'll be so we have to try both.
8214 if (isEssentiallyExtractSubvector(LHS.getOperand(0))) {
8215 RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG);
8219 RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS);
8220 } else if (isEssentiallyExtractSubvector(RHS.getOperand(0))) {
8221 LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG);
8225 LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS);
8228 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS);
8231 // Massage DAGs which we can use the high-half "long" operations on into
8232 // something isel will recognize better. E.g.
8234 // (aarch64_neon_umull (extract_high vec) (dupv64 scalar)) -->
8235 // (aarch64_neon_umull (extract_high (v2i64 vec)))
8236 // (extract_high (v2i64 (dup128 scalar)))))
8238 static SDValue tryCombineLongOpWithDup(SDNode *N,
8239 TargetLowering::DAGCombinerInfo &DCI,
8240 SelectionDAG &DAG) {
8241 if (DCI.isBeforeLegalizeOps())
8244 bool IsIntrinsic = N->getOpcode() == ISD::INTRINSIC_WO_CHAIN;
8245 SDValue LHS = N->getOperand(IsIntrinsic ? 1 : 0);
8246 SDValue RHS = N->getOperand(IsIntrinsic ? 2 : 1);
8247 assert(LHS.getValueType().is64BitVector() &&
8248 RHS.getValueType().is64BitVector() &&
8249 "unexpected shape for long operation");
8251 // Either node could be a DUP, but it's not worth doing both of them (you'd
8252 // just as well use the non-high version) so look for a corresponding extract
8253 // operation on the other "wing".
8254 if (isEssentiallyExtractSubvector(LHS)) {
8255 RHS = tryExtendDUPToExtractHigh(RHS, DAG);
8258 } else if (isEssentiallyExtractSubvector(RHS)) {
8259 LHS = tryExtendDUPToExtractHigh(LHS, DAG);
8264 // N could either be an intrinsic or a sabsdiff/uabsdiff node.
8266 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0),
8267 N->getOperand(0), LHS, RHS);
8269 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
8273 static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
8274 MVT ElemTy = N->getSimpleValueType(0).getScalarType();
8275 unsigned ElemBits = ElemTy.getSizeInBits();
8277 int64_t ShiftAmount;
8278 if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) {
8279 APInt SplatValue, SplatUndef;
8280 unsigned SplatBitSize;
8282 if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
8283 HasAnyUndefs, ElemBits) ||
8284 SplatBitSize != ElemBits)
8287 ShiftAmount = SplatValue.getSExtValue();
8288 } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) {
8289 ShiftAmount = CVN->getSExtValue();
8297 llvm_unreachable("Unknown shift intrinsic");
8298 case Intrinsic::aarch64_neon_sqshl:
8299 Opcode = AArch64ISD::SQSHL_I;
8300 IsRightShift = false;
8302 case Intrinsic::aarch64_neon_uqshl:
8303 Opcode = AArch64ISD::UQSHL_I;
8304 IsRightShift = false;
8306 case Intrinsic::aarch64_neon_srshl:
8307 Opcode = AArch64ISD::SRSHR_I;
8308 IsRightShift = true;
8310 case Intrinsic::aarch64_neon_urshl:
8311 Opcode = AArch64ISD::URSHR_I;
8312 IsRightShift = true;
8314 case Intrinsic::aarch64_neon_sqshlu:
8315 Opcode = AArch64ISD::SQSHLU_I;
8316 IsRightShift = false;
8320 if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits) {
8322 return DAG.getNode(Opcode, dl, N->getValueType(0), N->getOperand(1),
8323 DAG.getConstant(-ShiftAmount, dl, MVT::i32));
8324 } else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount < ElemBits) {
8326 return DAG.getNode(Opcode, dl, N->getValueType(0), N->getOperand(1),
8327 DAG.getConstant(ShiftAmount, dl, MVT::i32));
8333 // The CRC32[BH] instructions ignore the high bits of their data operand. Since
8334 // the intrinsics must be legal and take an i32, this means there's almost
8335 // certainly going to be a zext in the DAG which we can eliminate.
8336 static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) {
8337 SDValue AndN = N->getOperand(2);
8338 if (AndN.getOpcode() != ISD::AND)
8341 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1));
8342 if (!CMask || CMask->getZExtValue() != Mask)
8345 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32,
8346 N->getOperand(0), N->getOperand(1), AndN.getOperand(0));
8349 static SDValue combineAcrossLanesIntrinsic(unsigned Opc, SDNode *N,
8350 SelectionDAG &DAG) {
8352 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0),
8353 DAG.getNode(Opc, dl,
8354 N->getOperand(1).getSimpleValueType(),
8356 DAG.getConstant(0, dl, MVT::i64));
8359 static SDValue performIntrinsicCombine(SDNode *N,
8360 TargetLowering::DAGCombinerInfo &DCI,
8361 const AArch64Subtarget *Subtarget) {
8362 SelectionDAG &DAG = DCI.DAG;
8363 unsigned IID = getIntrinsicID(N);
8367 case Intrinsic::aarch64_neon_vcvtfxs2fp:
8368 case Intrinsic::aarch64_neon_vcvtfxu2fp:
8369 return tryCombineFixedPointConvert(N, DCI, DAG);
8370 case Intrinsic::aarch64_neon_saddv:
8371 return combineAcrossLanesIntrinsic(AArch64ISD::SADDV, N, DAG);
8372 case Intrinsic::aarch64_neon_uaddv:
8373 return combineAcrossLanesIntrinsic(AArch64ISD::UADDV, N, DAG);
8374 case Intrinsic::aarch64_neon_sminv:
8375 return combineAcrossLanesIntrinsic(AArch64ISD::SMINV, N, DAG);
8376 case Intrinsic::aarch64_neon_uminv:
8377 return combineAcrossLanesIntrinsic(AArch64ISD::UMINV, N, DAG);
8378 case Intrinsic::aarch64_neon_smaxv:
8379 return combineAcrossLanesIntrinsic(AArch64ISD::SMAXV, N, DAG);
8380 case Intrinsic::aarch64_neon_umaxv:
8381 return combineAcrossLanesIntrinsic(AArch64ISD::UMAXV, N, DAG);
8382 case Intrinsic::aarch64_neon_fmax:
8383 return DAG.getNode(ISD::FMAXNAN, SDLoc(N), N->getValueType(0),
8384 N->getOperand(1), N->getOperand(2));
8385 case Intrinsic::aarch64_neon_fmin:
8386 return DAG.getNode(ISD::FMINNAN, SDLoc(N), N->getValueType(0),
8387 N->getOperand(1), N->getOperand(2));
8388 case Intrinsic::aarch64_neon_sabd:
8389 return DAG.getNode(ISD::SABSDIFF, SDLoc(N), N->getValueType(0),
8390 N->getOperand(1), N->getOperand(2));
8391 case Intrinsic::aarch64_neon_uabd:
8392 return DAG.getNode(ISD::UABSDIFF, SDLoc(N), N->getValueType(0),
8393 N->getOperand(1), N->getOperand(2));
8394 case Intrinsic::aarch64_neon_fmaxnm:
8395 return DAG.getNode(ISD::FMAXNUM, SDLoc(N), N->getValueType(0),
8396 N->getOperand(1), N->getOperand(2));
8397 case Intrinsic::aarch64_neon_fminnm:
8398 return DAG.getNode(ISD::FMINNUM, SDLoc(N), N->getValueType(0),
8399 N->getOperand(1), N->getOperand(2));
8400 case Intrinsic::aarch64_neon_smull:
8401 case Intrinsic::aarch64_neon_umull:
8402 case Intrinsic::aarch64_neon_pmull:
8403 case Intrinsic::aarch64_neon_sqdmull:
8404 return tryCombineLongOpWithDup(N, DCI, DAG);
8405 case Intrinsic::aarch64_neon_sqshl:
8406 case Intrinsic::aarch64_neon_uqshl:
8407 case Intrinsic::aarch64_neon_sqshlu:
8408 case Intrinsic::aarch64_neon_srshl:
8409 case Intrinsic::aarch64_neon_urshl:
8410 return tryCombineShiftImm(IID, N, DAG);
8411 case Intrinsic::aarch64_crc32b:
8412 case Intrinsic::aarch64_crc32cb:
8413 return tryCombineCRC32(0xff, N, DAG);
8414 case Intrinsic::aarch64_crc32h:
8415 case Intrinsic::aarch64_crc32ch:
8416 return tryCombineCRC32(0xffff, N, DAG);
8421 static SDValue performExtendCombine(SDNode *N,
8422 TargetLowering::DAGCombinerInfo &DCI,
8423 SelectionDAG &DAG) {
8424 // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then
8425 // we can convert that DUP into another extract_high (of a bigger DUP), which
8426 // helps the backend to decide that an sabdl2 would be useful, saving a real
8427 // extract_high operation.
8428 if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND &&
8429 (N->getOperand(0).getOpcode() == ISD::SABSDIFF ||
8430 N->getOperand(0).getOpcode() == ISD::UABSDIFF)) {
8431 SDNode *ABDNode = N->getOperand(0).getNode();
8432 SDValue NewABD = tryCombineLongOpWithDup(ABDNode, DCI, DAG);
8433 if (!NewABD.getNode())
8436 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0),
8440 // This is effectively a custom type legalization for AArch64.
8442 // Type legalization will split an extend of a small, legal, type to a larger
8443 // illegal type by first splitting the destination type, often creating
8444 // illegal source types, which then get legalized in isel-confusing ways,
8445 // leading to really terrible codegen. E.g.,
8446 // %result = v8i32 sext v8i8 %value
8448 // %losrc = extract_subreg %value, ...
8449 // %hisrc = extract_subreg %value, ...
8450 // %lo = v4i32 sext v4i8 %losrc
8451 // %hi = v4i32 sext v4i8 %hisrc
8452 // Things go rapidly downhill from there.
8454 // For AArch64, the [sz]ext vector instructions can only go up one element
8455 // size, so we can, e.g., extend from i8 to i16, but to go from i8 to i32
8456 // take two instructions.
8458 // This implies that the most efficient way to do the extend from v8i8
8459 // to two v4i32 values is to first extend the v8i8 to v8i16, then do
8460 // the normal splitting to happen for the v8i16->v8i32.
8462 // This is pre-legalization to catch some cases where the default
8463 // type legalization will create ill-tempered code.
8464 if (!DCI.isBeforeLegalizeOps())
8467 // We're only interested in cleaning things up for non-legal vector types
8468 // here. If both the source and destination are legal, things will just
8469 // work naturally without any fiddling.
8470 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8471 EVT ResVT = N->getValueType(0);
8472 if (!ResVT.isVector() || TLI.isTypeLegal(ResVT))
8474 // If the vector type isn't a simple VT, it's beyond the scope of what
8475 // we're worried about here. Let legalization do its thing and hope for
8477 SDValue Src = N->getOperand(0);
8478 EVT SrcVT = Src->getValueType(0);
8479 if (!ResVT.isSimple() || !SrcVT.isSimple())
8482 // If the source VT is a 64-bit vector, we can play games and get the
8483 // better results we want.
8484 if (SrcVT.getSizeInBits() != 64)
8487 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
8488 unsigned ElementCount = SrcVT.getVectorNumElements();
8489 SrcVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize * 2), ElementCount);
8491 Src = DAG.getNode(N->getOpcode(), DL, SrcVT, Src);
8493 // Now split the rest of the operation into two halves, each with a 64
8497 unsigned NumElements = ResVT.getVectorNumElements();
8498 assert(!(NumElements & 1) && "Splitting vector, but not in half!");
8499 LoVT = HiVT = EVT::getVectorVT(*DAG.getContext(),
8500 ResVT.getVectorElementType(), NumElements / 2);
8502 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
8503 LoVT.getVectorNumElements());
8504 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
8505 DAG.getConstant(0, DL, MVT::i64));
8506 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
8507 DAG.getConstant(InNVT.getVectorNumElements(), DL, MVT::i64));
8508 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, Lo);
8509 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, Hi);
8511 // Now combine the parts back together so we still have a single result
8512 // like the combiner expects.
8513 return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
8516 /// Replace a splat of a scalar to a vector store by scalar stores of the scalar
8517 /// value. The load store optimizer pass will merge them to store pair stores.
8518 /// This has better performance than a splat of the scalar followed by a split
8519 /// vector store. Even if the stores are not merged it is four stores vs a dup,
8520 /// followed by an ext.b and two stores.
8521 static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) {
8522 SDValue StVal = St->getValue();
8523 EVT VT = StVal.getValueType();
8525 // Don't replace floating point stores, they possibly won't be transformed to
8526 // stp because of the store pair suppress pass.
8527 if (VT.isFloatingPoint())
8530 // Check for insert vector elements.
8531 if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT)
8534 // We can express a splat as store pair(s) for 2 or 4 elements.
8535 unsigned NumVecElts = VT.getVectorNumElements();
8536 if (NumVecElts != 4 && NumVecElts != 2)
8538 SDValue SplatVal = StVal.getOperand(1);
8539 unsigned RemainInsertElts = NumVecElts - 1;
8541 // Check that this is a splat.
8542 while (--RemainInsertElts) {
8543 SDValue NextInsertElt = StVal.getOperand(0);
8544 if (NextInsertElt.getOpcode() != ISD::INSERT_VECTOR_ELT)
8546 if (NextInsertElt.getOperand(1) != SplatVal)
8548 StVal = NextInsertElt;
8550 unsigned OrigAlignment = St->getAlignment();
8551 unsigned EltOffset = NumVecElts == 4 ? 4 : 8;
8552 unsigned Alignment = std::min(OrigAlignment, EltOffset);
8554 // Create scalar stores. This is at least as good as the code sequence for a
8555 // split unaligned store which is a dup.s, ext.b, and two stores.
8556 // Most of the time the three stores should be replaced by store pair
8557 // instructions (stp).
8559 SDValue BasePtr = St->getBasePtr();
8561 DAG.getStore(St->getChain(), DL, SplatVal, BasePtr, St->getPointerInfo(),
8562 St->isVolatile(), St->isNonTemporal(), St->getAlignment());
8564 unsigned Offset = EltOffset;
8565 while (--NumVecElts) {
8566 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
8567 DAG.getConstant(Offset, DL, MVT::i64));
8568 NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr,
8569 St->getPointerInfo(), St->isVolatile(),
8570 St->isNonTemporal(), Alignment);
8571 Offset += EltOffset;
8576 static SDValue performSTORECombine(SDNode *N,
8577 TargetLowering::DAGCombinerInfo &DCI,
8579 const AArch64Subtarget *Subtarget) {
8580 if (!DCI.isBeforeLegalize())
8583 StoreSDNode *S = cast<StoreSDNode>(N);
8584 if (S->isVolatile())
8587 // FIXME: The logic for deciding if an unaligned store should be split should
8588 // be included in TLI.allowsMisalignedMemoryAccesses(), and there should be
8589 // a call to that function here.
8591 // Cyclone has bad performance on unaligned 16B stores when crossing line and
8592 // page boundaries. We want to split such stores.
8593 if (!Subtarget->isCyclone())
8596 // Don't split at -Oz.
8597 if (DAG.getMachineFunction().getFunction()->optForMinSize())
8600 SDValue StVal = S->getValue();
8601 EVT VT = StVal.getValueType();
8603 // Don't split v2i64 vectors. Memcpy lowering produces those and splitting
8604 // those up regresses performance on micro-benchmarks and olden/bh.
8605 if (!VT.isVector() || VT.getVectorNumElements() < 2 || VT == MVT::v2i64)
8608 // Split unaligned 16B stores. They are terrible for performance.
8609 // Don't split stores with alignment of 1 or 2. Code that uses clang vector
8610 // extensions can use this to mark that it does not want splitting to happen
8611 // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of
8612 // eliminating alignment hazards is only 1 in 8 for alignment of 2.
8613 if (VT.getSizeInBits() != 128 || S->getAlignment() >= 16 ||
8614 S->getAlignment() <= 2)
8617 // If we get a splat of a scalar convert this vector store to a store of
8618 // scalars. They will be merged into store pairs thereby removing two
8620 if (SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S))
8621 return ReplacedSplat;
8624 unsigned NumElts = VT.getVectorNumElements() / 2;
8625 // Split VT into two.
8627 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), NumElts);
8628 SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
8629 DAG.getConstant(0, DL, MVT::i64));
8630 SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
8631 DAG.getConstant(NumElts, DL, MVT::i64));
8632 SDValue BasePtr = S->getBasePtr();
8634 DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
8635 S->isVolatile(), S->isNonTemporal(), S->getAlignment());
8636 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
8637 DAG.getConstant(8, DL, MVT::i64));
8638 return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr,
8639 S->getPointerInfo(), S->isVolatile(), S->isNonTemporal(),
8643 /// Target-specific DAG combine function for post-increment LD1 (lane) and
8644 /// post-increment LD1R.
8645 static SDValue performPostLD1Combine(SDNode *N,
8646 TargetLowering::DAGCombinerInfo &DCI,
8648 if (DCI.isBeforeLegalizeOps())
8651 SelectionDAG &DAG = DCI.DAG;
8652 EVT VT = N->getValueType(0);
8654 unsigned LoadIdx = IsLaneOp ? 1 : 0;
8655 SDNode *LD = N->getOperand(LoadIdx).getNode();
8656 // If it is not LOAD, can not do such combine.
8657 if (LD->getOpcode() != ISD::LOAD)
8660 LoadSDNode *LoadSDN = cast<LoadSDNode>(LD);
8661 EVT MemVT = LoadSDN->getMemoryVT();
8662 // Check if memory operand is the same type as the vector element.
8663 if (MemVT != VT.getVectorElementType())
8666 // Check if there are other uses. If so, do not combine as it will introduce
8668 for (SDNode::use_iterator UI = LD->use_begin(), UE = LD->use_end(); UI != UE;
8670 if (UI.getUse().getResNo() == 1) // Ignore uses of the chain result.
8676 SDValue Addr = LD->getOperand(1);
8677 SDValue Vector = N->getOperand(0);
8678 // Search for a use of the address operand that is an increment.
8679 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(), UE =
8680 Addr.getNode()->use_end(); UI != UE; ++UI) {
8682 if (User->getOpcode() != ISD::ADD
8683 || UI.getUse().getResNo() != Addr.getResNo())
8686 // Check that the add is independent of the load. Otherwise, folding it
8687 // would create a cycle.
8688 if (User->isPredecessorOf(LD) || LD->isPredecessorOf(User))
8690 // Also check that add is not used in the vector operand. This would also
8692 if (User->isPredecessorOf(Vector.getNode()))
8695 // If the increment is a constant, it must match the memory ref size.
8696 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8697 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8698 uint32_t IncVal = CInc->getZExtValue();
8699 unsigned NumBytes = VT.getScalarSizeInBits() / 8;
8700 if (IncVal != NumBytes)
8702 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
8705 // Finally, check that the vector doesn't depend on the load.
8706 // Again, this would create a cycle.
8707 // The load depending on the vector is fine, as that's the case for the
8708 // LD1*post we'll eventually generate anyway.
8709 if (LoadSDN->isPredecessorOf(Vector.getNode()))
8712 SmallVector<SDValue, 8> Ops;
8713 Ops.push_back(LD->getOperand(0)); // Chain
8715 Ops.push_back(Vector); // The vector to be inserted
8716 Ops.push_back(N->getOperand(2)); // The lane to be inserted in the vector
8718 Ops.push_back(Addr);
8721 EVT Tys[3] = { VT, MVT::i64, MVT::Other };
8722 SDVTList SDTys = DAG.getVTList(Tys);
8723 unsigned NewOp = IsLaneOp ? AArch64ISD::LD1LANEpost : AArch64ISD::LD1DUPpost;
8724 SDValue UpdN = DAG.getMemIntrinsicNode(NewOp, SDLoc(N), SDTys, Ops,
8726 LoadSDN->getMemOperand());
8729 SmallVector<SDValue, 2> NewResults;
8730 NewResults.push_back(SDValue(LD, 0)); // The result of load
8731 NewResults.push_back(SDValue(UpdN.getNode(), 2)); // Chain
8732 DCI.CombineTo(LD, NewResults);
8733 DCI.CombineTo(N, SDValue(UpdN.getNode(), 0)); // Dup/Inserted Result
8734 DCI.CombineTo(User, SDValue(UpdN.getNode(), 1)); // Write back register
8741 /// This function handles the log2-shuffle pattern produced by the
8742 /// LoopVectorizer for the across vector reduction. It consists of
8743 /// log2(NumVectorElements) steps and, in each step, 2^(s) elements
8744 /// are reduced, where s is an induction variable from 0 to
8745 /// log2(NumVectorElements).
8746 static SDValue tryMatchAcrossLaneShuffleForReduction(SDNode *N, SDValue OpV,
8748 SelectionDAG &DAG) {
8749 EVT VTy = OpV->getOperand(0).getValueType();
8750 if (!VTy.isVector())
8753 int NumVecElts = VTy.getVectorNumElements();
8754 if (Op == ISD::FMAXNUM || Op == ISD::FMINNUM) {
8755 if (NumVecElts != 4)
8758 if (NumVecElts != 4 && NumVecElts != 8 && NumVecElts != 16)
8762 int NumExpectedSteps = APInt(8, NumVecElts).logBase2();
8763 SDValue PreOp = OpV;
8764 // Iterate over each step of the across vector reduction.
8765 for (int CurStep = 0; CurStep != NumExpectedSteps; ++CurStep) {
8766 SDValue CurOp = PreOp.getOperand(0);
8767 SDValue Shuffle = PreOp.getOperand(1);
8768 if (Shuffle.getOpcode() != ISD::VECTOR_SHUFFLE) {
8769 // Try to swap the 1st and 2nd operand as add and min/max instructions
8771 CurOp = PreOp.getOperand(1);
8772 Shuffle = PreOp.getOperand(0);
8773 if (Shuffle.getOpcode() != ISD::VECTOR_SHUFFLE)
8777 // Check if the input vector is fed by the operator we want to handle,
8778 // except the last step; the very first input vector is not necessarily
8779 // the same operator we are handling.
8780 if (CurOp.getOpcode() != Op && (CurStep != (NumExpectedSteps - 1)))
8783 // Check if it forms one step of the across vector reduction.
8785 // %cur = add %1, %0
8786 // %shuffle = vector_shuffle %cur, <2, 3, u, u>
8787 // %pre = add %cur, %shuffle
8788 if (Shuffle.getOperand(0) != CurOp)
8791 int NumMaskElts = 1 << CurStep;
8792 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Shuffle)->getMask();
8793 // Check mask values in each step.
8794 // We expect the shuffle mask in each step follows a specific pattern
8795 // denoted here by the <M, U> form, where M is a sequence of integers
8796 // starting from NumMaskElts, increasing by 1, and the number integers
8797 // in M should be NumMaskElts. U is a sequence of UNDEFs and the number
8798 // of undef in U should be NumVecElts - NumMaskElts.
8799 // E.g., for <8 x i16>, mask values in each step should be :
8800 // step 0 : <1,u,u,u,u,u,u,u>
8801 // step 1 : <2,3,u,u,u,u,u,u>
8802 // step 2 : <4,5,6,7,u,u,u,u>
8803 for (int i = 0; i < NumVecElts; ++i)
8804 if ((i < NumMaskElts && Mask[i] != (NumMaskElts + i)) ||
8805 (i >= NumMaskElts && !(Mask[i] < 0)))
8811 bool IsIntrinsic = false;
8815 llvm_unreachable("Unexpected operator for across vector reduction");
8817 Opcode = AArch64ISD::UADDV;
8820 Opcode = AArch64ISD::SMAXV;
8823 Opcode = AArch64ISD::UMAXV;
8826 Opcode = AArch64ISD::SMINV;
8829 Opcode = AArch64ISD::UMINV;
8832 Opcode = Intrinsic::aarch64_neon_fmaxnmv;
8836 Opcode = Intrinsic::aarch64_neon_fminnmv;
8843 ? DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, N->getValueType(0),
8844 DAG.getConstant(Opcode, DL, MVT::i32), PreOp)
8846 ISD::EXTRACT_VECTOR_ELT, DL, N->getValueType(0),
8847 DAG.getNode(Opcode, DL, PreOp.getSimpleValueType(), PreOp),
8848 DAG.getConstant(0, DL, MVT::i64));
8851 /// Target-specific DAG combine for the across vector min/max reductions.
8852 /// This function specifically handles the final clean-up step of the vector
8853 /// min/max reductions produced by the LoopVectorizer. It is the log2-shuffle
8854 /// pattern, which narrows down and finds the final min/max value from all
8855 /// elements of the vector.
8856 /// For example, for a <16 x i8> vector :
8857 /// svn0 = vector_shuffle %0, undef<8,9,10,11,12,13,14,15,u,u,u,u,u,u,u,u>
8858 /// %smax0 = smax %arr, svn0
8859 /// %svn1 = vector_shuffle %smax0, undef<4,5,6,7,u,u,u,u,u,u,u,u,u,u,u,u>
8860 /// %smax1 = smax %smax0, %svn1
8861 /// %svn2 = vector_shuffle %smax1, undef<2,3,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
8862 /// %smax2 = smax %smax1, svn2
8863 /// %svn3 = vector_shuffle %smax2, undef<1,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
8864 /// %sc = setcc %smax2, %svn3, gt
8865 /// %n0 = extract_vector_elt %sc, #0
8866 /// %n1 = extract_vector_elt %smax2, #0
8867 /// %n2 = extract_vector_elt $smax2, #1
8868 /// %result = select %n0, %n1, n2
8871 /// %result = extract_vector_elt %1, 0
8873 performAcrossLaneMinMaxReductionCombine(SDNode *N, SelectionDAG &DAG,
8874 const AArch64Subtarget *Subtarget) {
8875 if (!Subtarget->hasNEON())
8878 SDValue N0 = N->getOperand(0);
8879 SDValue IfTrue = N->getOperand(1);
8880 SDValue IfFalse = N->getOperand(2);
8882 // Check if the SELECT merges up the final result of the min/max
8884 if (N0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
8885 IfTrue.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
8886 IfFalse.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8889 // Expect N0 is fed by SETCC.
8890 SDValue SetCC = N0.getOperand(0);
8891 EVT SetCCVT = SetCC.getValueType();
8892 if (SetCC.getOpcode() != ISD::SETCC || !SetCCVT.isVector() ||
8893 SetCCVT.getVectorElementType() != MVT::i1)
8896 SDValue VectorOp = SetCC.getOperand(0);
8897 unsigned Op = VectorOp->getOpcode();
8898 // Check if the input vector is fed by the operator we want to handle.
8899 if (Op != ISD::SMAX && Op != ISD::UMAX && Op != ISD::SMIN &&
8900 Op != ISD::UMIN && Op != ISD::FMAXNUM && Op != ISD::FMINNUM)
8903 EVT VTy = VectorOp.getValueType();
8904 if (!VTy.isVector())
8907 if (VTy.getSizeInBits() < 64)
8910 EVT EltTy = VTy.getVectorElementType();
8911 if (Op == ISD::FMAXNUM || Op == ISD::FMINNUM) {
8912 if (EltTy != MVT::f32)
8915 if (EltTy != MVT::i32 && EltTy != MVT::i16 && EltTy != MVT::i8)
8919 // Check if extracting from the same vector.
8921 // %sc = setcc %vector, %svn1, gt
8922 // %n0 = extract_vector_elt %sc, #0
8923 // %n1 = extract_vector_elt %vector, #0
8924 // %n2 = extract_vector_elt $vector, #1
8925 if (!(VectorOp == IfTrue->getOperand(0) &&
8926 VectorOp == IfFalse->getOperand(0)))
8929 // Check if the condition code is matched with the operator type.
8930 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC->getOperand(2))->get();
8931 if ((Op == ISD::SMAX && CC != ISD::SETGT && CC != ISD::SETGE) ||
8932 (Op == ISD::UMAX && CC != ISD::SETUGT && CC != ISD::SETUGE) ||
8933 (Op == ISD::SMIN && CC != ISD::SETLT && CC != ISD::SETLE) ||
8934 (Op == ISD::UMIN && CC != ISD::SETULT && CC != ISD::SETULE) ||
8935 (Op == ISD::FMAXNUM && CC != ISD::SETOGT && CC != ISD::SETOGE &&
8936 CC != ISD::SETUGT && CC != ISD::SETUGE && CC != ISD::SETGT &&
8937 CC != ISD::SETGE) ||
8938 (Op == ISD::FMINNUM && CC != ISD::SETOLT && CC != ISD::SETOLE &&
8939 CC != ISD::SETULT && CC != ISD::SETULE && CC != ISD::SETLT &&
8943 // Expect to check only lane 0 from the vector SETCC.
8944 if (!isa<ConstantSDNode>(N0.getOperand(1)) ||
8945 cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue() != 0)
8948 // Expect to extract the true value from lane 0.
8949 if (!isa<ConstantSDNode>(IfTrue.getOperand(1)) ||
8950 cast<ConstantSDNode>(IfTrue.getOperand(1))->getZExtValue() != 0)
8953 // Expect to extract the false value from lane 1.
8954 if (!isa<ConstantSDNode>(IfFalse.getOperand(1)) ||
8955 cast<ConstantSDNode>(IfFalse.getOperand(1))->getZExtValue() != 1)
8958 return tryMatchAcrossLaneShuffleForReduction(N, SetCC, Op, DAG);
8961 /// Target-specific DAG combine for the across vector add reduction.
8962 /// This function specifically handles the final clean-up step of the vector
8963 /// add reduction produced by the LoopVectorizer. It is the log2-shuffle
8964 /// pattern, which adds all elements of a vector together.
8965 /// For example, for a <4 x i32> vector :
8966 /// %1 = vector_shuffle %0, <2,3,u,u>
8968 /// %3 = vector_shuffle %2, <1,u,u,u>
8970 /// %result = extract_vector_elt %4, 0
8973 /// %result = extract_vector_elt %0, 0
8975 performAcrossLaneAddReductionCombine(SDNode *N, SelectionDAG &DAG,
8976 const AArch64Subtarget *Subtarget) {
8977 if (!Subtarget->hasNEON())
8979 SDValue N0 = N->getOperand(0);
8980 SDValue N1 = N->getOperand(1);
8982 // Check if the input vector is fed by the ADD.
8983 if (N0->getOpcode() != ISD::ADD)
8986 // The vector extract idx must constant zero because we only expect the final
8987 // result of the reduction is placed in lane 0.
8988 if (!isa<ConstantSDNode>(N1) || cast<ConstantSDNode>(N1)->getZExtValue() != 0)
8991 EVT VTy = N0.getValueType();
8992 if (!VTy.isVector())
8995 EVT EltTy = VTy.getVectorElementType();
8996 if (EltTy != MVT::i32 && EltTy != MVT::i16 && EltTy != MVT::i8)
8999 if (VTy.getSizeInBits() < 64)
9002 return tryMatchAcrossLaneShuffleForReduction(N, N0, ISD::ADD, DAG);
9005 /// Target-specific DAG combine function for NEON load/store intrinsics
9006 /// to merge base address updates.
9007 static SDValue performNEONPostLDSTCombine(SDNode *N,
9008 TargetLowering::DAGCombinerInfo &DCI,
9009 SelectionDAG &DAG) {
9010 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9013 unsigned AddrOpIdx = N->getNumOperands() - 1;
9014 SDValue Addr = N->getOperand(AddrOpIdx);
9016 // Search for a use of the address operand that is an increment.
9017 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
9018 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
9020 if (User->getOpcode() != ISD::ADD ||
9021 UI.getUse().getResNo() != Addr.getResNo())
9024 // Check that the add is independent of the load/store. Otherwise, folding
9025 // it would create a cycle.
9026 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
9029 // Find the new opcode for the updating load/store.
9030 bool IsStore = false;
9031 bool IsLaneOp = false;
9032 bool IsDupOp = false;
9033 unsigned NewOpc = 0;
9034 unsigned NumVecs = 0;
9035 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
9037 default: llvm_unreachable("unexpected intrinsic for Neon base update");
9038 case Intrinsic::aarch64_neon_ld2: NewOpc = AArch64ISD::LD2post;
9040 case Intrinsic::aarch64_neon_ld3: NewOpc = AArch64ISD::LD3post;
9042 case Intrinsic::aarch64_neon_ld4: NewOpc = AArch64ISD::LD4post;
9044 case Intrinsic::aarch64_neon_st2: NewOpc = AArch64ISD::ST2post;
9045 NumVecs = 2; IsStore = true; break;
9046 case Intrinsic::aarch64_neon_st3: NewOpc = AArch64ISD::ST3post;
9047 NumVecs = 3; IsStore = true; break;
9048 case Intrinsic::aarch64_neon_st4: NewOpc = AArch64ISD::ST4post;
9049 NumVecs = 4; IsStore = true; break;
9050 case Intrinsic::aarch64_neon_ld1x2: NewOpc = AArch64ISD::LD1x2post;
9052 case Intrinsic::aarch64_neon_ld1x3: NewOpc = AArch64ISD::LD1x3post;
9054 case Intrinsic::aarch64_neon_ld1x4: NewOpc = AArch64ISD::LD1x4post;
9056 case Intrinsic::aarch64_neon_st1x2: NewOpc = AArch64ISD::ST1x2post;
9057 NumVecs = 2; IsStore = true; break;
9058 case Intrinsic::aarch64_neon_st1x3: NewOpc = AArch64ISD::ST1x3post;
9059 NumVecs = 3; IsStore = true; break;
9060 case Intrinsic::aarch64_neon_st1x4: NewOpc = AArch64ISD::ST1x4post;
9061 NumVecs = 4; IsStore = true; break;
9062 case Intrinsic::aarch64_neon_ld2r: NewOpc = AArch64ISD::LD2DUPpost;
9063 NumVecs = 2; IsDupOp = true; break;
9064 case Intrinsic::aarch64_neon_ld3r: NewOpc = AArch64ISD::LD3DUPpost;
9065 NumVecs = 3; IsDupOp = true; break;
9066 case Intrinsic::aarch64_neon_ld4r: NewOpc = AArch64ISD::LD4DUPpost;
9067 NumVecs = 4; IsDupOp = true; break;
9068 case Intrinsic::aarch64_neon_ld2lane: NewOpc = AArch64ISD::LD2LANEpost;
9069 NumVecs = 2; IsLaneOp = true; break;
9070 case Intrinsic::aarch64_neon_ld3lane: NewOpc = AArch64ISD::LD3LANEpost;
9071 NumVecs = 3; IsLaneOp = true; break;
9072 case Intrinsic::aarch64_neon_ld4lane: NewOpc = AArch64ISD::LD4LANEpost;
9073 NumVecs = 4; IsLaneOp = true; break;
9074 case Intrinsic::aarch64_neon_st2lane: NewOpc = AArch64ISD::ST2LANEpost;
9075 NumVecs = 2; IsStore = true; IsLaneOp = true; break;
9076 case Intrinsic::aarch64_neon_st3lane: NewOpc = AArch64ISD::ST3LANEpost;
9077 NumVecs = 3; IsStore = true; IsLaneOp = true; break;
9078 case Intrinsic::aarch64_neon_st4lane: NewOpc = AArch64ISD::ST4LANEpost;
9079 NumVecs = 4; IsStore = true; IsLaneOp = true; break;
9084 VecTy = N->getOperand(2).getValueType();
9086 VecTy = N->getValueType(0);
9088 // If the increment is a constant, it must match the memory ref size.
9089 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
9090 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
9091 uint32_t IncVal = CInc->getZExtValue();
9092 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
9093 if (IsLaneOp || IsDupOp)
9094 NumBytes /= VecTy.getVectorNumElements();
9095 if (IncVal != NumBytes)
9097 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
9099 SmallVector<SDValue, 8> Ops;
9100 Ops.push_back(N->getOperand(0)); // Incoming chain
9101 // Load lane and store have vector list as input.
9102 if (IsLaneOp || IsStore)
9103 for (unsigned i = 2; i < AddrOpIdx; ++i)
9104 Ops.push_back(N->getOperand(i));
9105 Ops.push_back(Addr); // Base register
9110 unsigned NumResultVecs = (IsStore ? 0 : NumVecs);
9112 for (n = 0; n < NumResultVecs; ++n)
9114 Tys[n++] = MVT::i64; // Type of write back register
9115 Tys[n] = MVT::Other; // Type of the chain
9116 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs + 2));
9118 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
9119 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys, Ops,
9120 MemInt->getMemoryVT(),
9121 MemInt->getMemOperand());
9124 std::vector<SDValue> NewResults;
9125 for (unsigned i = 0; i < NumResultVecs; ++i) {
9126 NewResults.push_back(SDValue(UpdN.getNode(), i));
9128 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs + 1));
9129 DCI.CombineTo(N, NewResults);
9130 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9137 // Checks to see if the value is the prescribed width and returns information
9138 // about its extension mode.
9140 bool checkValueWidth(SDValue V, unsigned width, ISD::LoadExtType &ExtType) {
9141 ExtType = ISD::NON_EXTLOAD;
9142 switch(V.getNode()->getOpcode()) {
9146 LoadSDNode *LoadNode = cast<LoadSDNode>(V.getNode());
9147 if ((LoadNode->getMemoryVT() == MVT::i8 && width == 8)
9148 || (LoadNode->getMemoryVT() == MVT::i16 && width == 16)) {
9149 ExtType = LoadNode->getExtensionType();
9154 case ISD::AssertSext: {
9155 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
9156 if ((TypeNode->getVT() == MVT::i8 && width == 8)
9157 || (TypeNode->getVT() == MVT::i16 && width == 16)) {
9158 ExtType = ISD::SEXTLOAD;
9163 case ISD::AssertZext: {
9164 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
9165 if ((TypeNode->getVT() == MVT::i8 && width == 8)
9166 || (TypeNode->getVT() == MVT::i16 && width == 16)) {
9167 ExtType = ISD::ZEXTLOAD;
9173 case ISD::TargetConstant: {
9174 if (std::abs(cast<ConstantSDNode>(V.getNode())->getSExtValue()) <
9184 // This function does a whole lot of voodoo to determine if the tests are
9185 // equivalent without and with a mask. Essentially what happens is that given a
9188 // +-------------+ +-------------+ +-------------+ +-------------+
9189 // | Input | | AddConstant | | CompConstant| | CC |
9190 // +-------------+ +-------------+ +-------------+ +-------------+
9192 // V V | +----------+
9193 // +-------------+ +----+ | |
9194 // | ADD | |0xff| | |
9195 // +-------------+ +----+ | |
9198 // +-------------+ | |
9200 // +-------------+ | |
9209 // The AND node may be safely removed for some combinations of inputs. In
9210 // particular we need to take into account the extension type of the Input,
9211 // the exact values of AddConstant, CompConstant, and CC, along with the nominal
9212 // width of the input (this can work for any width inputs, the above graph is
9213 // specific to 8 bits.
9215 // The specific equations were worked out by generating output tables for each
9216 // AArch64CC value in terms of and AddConstant (w1), CompConstant(w2). The
9217 // problem was simplified by working with 4 bit inputs, which means we only
9218 // needed to reason about 24 distinct bit patterns: 8 patterns unique to zero
9219 // extension (8,15), 8 patterns unique to sign extensions (-8,-1), and 8
9220 // patterns present in both extensions (0,7). For every distinct set of
9221 // AddConstant and CompConstants bit patterns we can consider the masked and
9222 // unmasked versions to be equivalent if the result of this function is true for
9223 // all 16 distinct bit patterns of for the current extension type of Input (w0).
9226 // and w10, w8, #0x0f
9228 // cset w9, AArch64CC
9230 // cset w11, AArch64CC
9235 // Since the above function shows when the outputs are equivalent it defines
9236 // when it is safe to remove the AND. Unfortunately it only runs on AArch64 and
9237 // would be expensive to run during compiles. The equations below were written
9238 // in a test harness that confirmed they gave equivalent outputs to the above
9239 // for all inputs function, so they can be used determine if the removal is
9242 // isEquivalentMaskless() is the code for testing if the AND can be removed
9243 // factored out of the DAG recognition as the DAG can take several forms.
9246 bool isEquivalentMaskless(unsigned CC, unsigned width,
9247 ISD::LoadExtType ExtType, signed AddConstant,
9248 signed CompConstant) {
9249 // By being careful about our equations and only writing the in term
9250 // symbolic values and well known constants (0, 1, -1, MaxUInt) we can
9251 // make them generally applicable to all bit widths.
9252 signed MaxUInt = (1 << width);
9254 // For the purposes of these comparisons sign extending the type is
9255 // equivalent to zero extending the add and displacing it by half the integer
9256 // width. Provided we are careful and make sure our equations are valid over
9257 // the whole range we can just adjust the input and avoid writing equations
9258 // for sign extended inputs.
9259 if (ExtType == ISD::SEXTLOAD)
9260 AddConstant -= (1 << (width-1));
9264 case AArch64CC::GT: {
9265 if ((AddConstant == 0) ||
9266 (CompConstant == MaxUInt - 1 && AddConstant < 0) ||
9267 (AddConstant >= 0 && CompConstant < 0) ||
9268 (AddConstant <= 0 && CompConstant <= 0 && CompConstant < AddConstant))
9272 case AArch64CC::GE: {
9273 if ((AddConstant == 0) ||
9274 (AddConstant >= 0 && CompConstant <= 0) ||
9275 (AddConstant <= 0 && CompConstant <= 0 && CompConstant <= AddConstant))
9279 case AArch64CC::LS: {
9280 if ((AddConstant >= 0 && CompConstant < 0) ||
9281 (AddConstant <= 0 && CompConstant >= -1 &&
9282 CompConstant < AddConstant + MaxUInt))
9286 case AArch64CC::MI: {
9287 if ((AddConstant == 0) ||
9288 (AddConstant > 0 && CompConstant <= 0) ||
9289 (AddConstant < 0 && CompConstant <= AddConstant))
9293 case AArch64CC::HS: {
9294 if ((AddConstant >= 0 && CompConstant <= 0) ||
9295 (AddConstant <= 0 && CompConstant >= 0 &&
9296 CompConstant <= AddConstant + MaxUInt))
9300 case AArch64CC::NE: {
9301 if ((AddConstant > 0 && CompConstant < 0) ||
9302 (AddConstant < 0 && CompConstant >= 0 &&
9303 CompConstant < AddConstant + MaxUInt) ||
9304 (AddConstant >= 0 && CompConstant >= 0 &&
9305 CompConstant >= AddConstant) ||
9306 (AddConstant <= 0 && CompConstant < 0 && CompConstant < AddConstant))
9315 case AArch64CC::Invalid:
9323 SDValue performCONDCombine(SDNode *N,
9324 TargetLowering::DAGCombinerInfo &DCI,
9325 SelectionDAG &DAG, unsigned CCIndex,
9326 unsigned CmpIndex) {
9327 unsigned CC = cast<ConstantSDNode>(N->getOperand(CCIndex))->getSExtValue();
9328 SDNode *SubsNode = N->getOperand(CmpIndex).getNode();
9329 unsigned CondOpcode = SubsNode->getOpcode();
9331 if (CondOpcode != AArch64ISD::SUBS)
9334 // There is a SUBS feeding this condition. Is it fed by a mask we can
9337 SDNode *AndNode = SubsNode->getOperand(0).getNode();
9338 unsigned MaskBits = 0;
9340 if (AndNode->getOpcode() != ISD::AND)
9343 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(AndNode->getOperand(1))) {
9344 uint32_t CNV = CN->getZExtValue();
9347 else if (CNV == 65535)
9354 SDValue AddValue = AndNode->getOperand(0);
9356 if (AddValue.getOpcode() != ISD::ADD)
9359 // The basic dag structure is correct, grab the inputs and validate them.
9361 SDValue AddInputValue1 = AddValue.getNode()->getOperand(0);
9362 SDValue AddInputValue2 = AddValue.getNode()->getOperand(1);
9363 SDValue SubsInputValue = SubsNode->getOperand(1);
9365 // The mask is present and the provenance of all the values is a smaller type,
9366 // lets see if the mask is superfluous.
9368 if (!isa<ConstantSDNode>(AddInputValue2.getNode()) ||
9369 !isa<ConstantSDNode>(SubsInputValue.getNode()))
9372 ISD::LoadExtType ExtType;
9374 if (!checkValueWidth(SubsInputValue, MaskBits, ExtType) ||
9375 !checkValueWidth(AddInputValue2, MaskBits, ExtType) ||
9376 !checkValueWidth(AddInputValue1, MaskBits, ExtType) )
9379 if(!isEquivalentMaskless(CC, MaskBits, ExtType,
9380 cast<ConstantSDNode>(AddInputValue2.getNode())->getSExtValue(),
9381 cast<ConstantSDNode>(SubsInputValue.getNode())->getSExtValue()))
9384 // The AND is not necessary, remove it.
9386 SDVTList VTs = DAG.getVTList(SubsNode->getValueType(0),
9387 SubsNode->getValueType(1));
9388 SDValue Ops[] = { AddValue, SubsNode->getOperand(1) };
9390 SDValue NewValue = DAG.getNode(CondOpcode, SDLoc(SubsNode), VTs, Ops);
9391 DAG.ReplaceAllUsesWith(SubsNode, NewValue.getNode());
9393 return SDValue(N, 0);
9396 // Optimize compare with zero and branch.
9397 static SDValue performBRCONDCombine(SDNode *N,
9398 TargetLowering::DAGCombinerInfo &DCI,
9399 SelectionDAG &DAG) {
9400 SDValue NV = performCONDCombine(N, DCI, DAG, 2, 3);
9403 SDValue Chain = N->getOperand(0);
9404 SDValue Dest = N->getOperand(1);
9405 SDValue CCVal = N->getOperand(2);
9406 SDValue Cmp = N->getOperand(3);
9408 assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!");
9409 unsigned CC = cast<ConstantSDNode>(CCVal)->getZExtValue();
9410 if (CC != AArch64CC::EQ && CC != AArch64CC::NE)
9413 unsigned CmpOpc = Cmp.getOpcode();
9414 if (CmpOpc != AArch64ISD::ADDS && CmpOpc != AArch64ISD::SUBS)
9417 // Only attempt folding if there is only one use of the flag and no use of the
9419 if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1))
9422 SDValue LHS = Cmp.getOperand(0);
9423 SDValue RHS = Cmp.getOperand(1);
9425 assert(LHS.getValueType() == RHS.getValueType() &&
9426 "Expected the value type to be the same for both operands!");
9427 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
9430 if (isa<ConstantSDNode>(LHS) && cast<ConstantSDNode>(LHS)->isNullValue())
9431 std::swap(LHS, RHS);
9433 if (!isa<ConstantSDNode>(RHS) || !cast<ConstantSDNode>(RHS)->isNullValue())
9436 if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA ||
9437 LHS.getOpcode() == ISD::SRL)
9440 // Fold the compare into the branch instruction.
9442 if (CC == AArch64CC::EQ)
9443 BR = DAG.getNode(AArch64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
9445 BR = DAG.getNode(AArch64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
9447 // Do not add new nodes to DAG combiner worklist.
9448 DCI.CombineTo(N, BR, false);
9453 // vselect (v1i1 setcc) ->
9454 // vselect (v1iXX setcc) (XX is the size of the compared operand type)
9455 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
9456 // condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
9458 static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) {
9459 SDValue N0 = N->getOperand(0);
9460 EVT CCVT = N0.getValueType();
9462 if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
9463 CCVT.getVectorElementType() != MVT::i1)
9466 EVT ResVT = N->getValueType(0);
9467 EVT CmpVT = N0.getOperand(0).getValueType();
9468 // Only combine when the result type is of the same size as the compared
9470 if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
9473 SDValue IfTrue = N->getOperand(1);
9474 SDValue IfFalse = N->getOperand(2);
9476 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
9477 N0.getOperand(0), N0.getOperand(1),
9478 cast<CondCodeSDNode>(N0.getOperand(2))->get());
9479 return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
9483 /// A vector select: "(select vL, vR, (setcc LHS, RHS))" is best performed with
9484 /// the compare-mask instructions rather than going via NZCV, even if LHS and
9485 /// RHS are really scalar. This replaces any scalar setcc in the above pattern
9486 /// with a vector one followed by a DUP shuffle on the result.
9487 static SDValue performSelectCombine(SDNode *N,
9488 TargetLowering::DAGCombinerInfo &DCI) {
9489 SelectionDAG &DAG = DCI.DAG;
9490 SDValue N0 = N->getOperand(0);
9491 EVT ResVT = N->getValueType(0);
9493 if (N0.getOpcode() != ISD::SETCC)
9496 // Make sure the SETCC result is either i1 (initial DAG), or i32, the lowered
9497 // scalar SetCCResultType. We also don't expect vectors, because we assume
9498 // that selects fed by vector SETCCs are canonicalized to VSELECT.
9499 assert((N0.getValueType() == MVT::i1 || N0.getValueType() == MVT::i32) &&
9500 "Scalar-SETCC feeding SELECT has unexpected result type!");
9502 // If NumMaskElts == 0, the comparison is larger than select result. The
9503 // largest real NEON comparison is 64-bits per lane, which means the result is
9504 // at most 32-bits and an illegal vector. Just bail out for now.
9505 EVT SrcVT = N0.getOperand(0).getValueType();
9507 // Don't try to do this optimization when the setcc itself has i1 operands.
9508 // There are no legal vectors of i1, so this would be pointless.
9509 if (SrcVT == MVT::i1)
9512 int NumMaskElts = ResVT.getSizeInBits() / SrcVT.getSizeInBits();
9513 if (!ResVT.isVector() || NumMaskElts == 0)
9516 SrcVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumMaskElts);
9517 EVT CCVT = SrcVT.changeVectorElementTypeToInteger();
9519 // Also bail out if the vector CCVT isn't the same size as ResVT.
9520 // This can happen if the SETCC operand size doesn't divide the ResVT size
9521 // (e.g., f64 vs v3f32).
9522 if (CCVT.getSizeInBits() != ResVT.getSizeInBits())
9525 // Make sure we didn't create illegal types, if we're not supposed to.
9526 assert(DCI.isBeforeLegalize() ||
9527 DAG.getTargetLoweringInfo().isTypeLegal(SrcVT));
9529 // First perform a vector comparison, where lane 0 is the one we're interested
9533 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(0));
9535 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(1));
9536 SDValue SetCC = DAG.getNode(ISD::SETCC, DL, CCVT, LHS, RHS, N0.getOperand(2));
9538 // Now duplicate the comparison mask we want across all other lanes.
9539 SmallVector<int, 8> DUPMask(CCVT.getVectorNumElements(), 0);
9540 SDValue Mask = DAG.getVectorShuffle(CCVT, DL, SetCC, SetCC, DUPMask.data());
9541 Mask = DAG.getNode(ISD::BITCAST, DL,
9542 ResVT.changeVectorElementTypeToInteger(), Mask);
9544 return DAG.getSelect(DL, ResVT, Mask, N->getOperand(1), N->getOperand(2));
9547 /// Get rid of unnecessary NVCASTs (that don't change the type).
9548 static SDValue performNVCASTCombine(SDNode *N) {
9549 if (N->getValueType(0) == N->getOperand(0).getValueType())
9550 return N->getOperand(0);
9555 SDValue AArch64TargetLowering::PerformDAGCombine(SDNode *N,
9556 DAGCombinerInfo &DCI) const {
9557 SelectionDAG &DAG = DCI.DAG;
9558 switch (N->getOpcode()) {
9563 return performAddSubLongCombine(N, DCI, DAG);
9565 return performXorCombine(N, DAG, DCI, Subtarget);
9567 return performMulCombine(N, DAG, DCI, Subtarget);
9568 case ISD::SINT_TO_FP:
9569 case ISD::UINT_TO_FP:
9570 return performIntToFpCombine(N, DAG, Subtarget);
9571 case ISD::FP_TO_SINT:
9572 case ISD::FP_TO_UINT:
9573 return performFpToIntCombine(N, DAG, Subtarget);
9575 return performFDivCombine(N, DAG, Subtarget);
9577 return performORCombine(N, DCI, Subtarget);
9578 case ISD::INTRINSIC_WO_CHAIN:
9579 return performIntrinsicCombine(N, DCI, Subtarget);
9580 case ISD::ANY_EXTEND:
9581 case ISD::ZERO_EXTEND:
9582 case ISD::SIGN_EXTEND:
9583 return performExtendCombine(N, DCI, DAG);
9585 return performBitcastCombine(N, DCI, DAG);
9586 case ISD::CONCAT_VECTORS:
9587 return performConcatVectorsCombine(N, DCI, DAG);
9589 SDValue RV = performSelectCombine(N, DCI);
9591 RV = performAcrossLaneMinMaxReductionCombine(N, DAG, Subtarget);
9595 return performVSelectCombine(N, DCI.DAG);
9597 return performSTORECombine(N, DCI, DAG, Subtarget);
9598 case AArch64ISD::BRCOND:
9599 return performBRCONDCombine(N, DCI, DAG);
9600 case AArch64ISD::CSEL:
9601 return performCONDCombine(N, DCI, DAG, 2, 3);
9602 case AArch64ISD::DUP:
9603 return performPostLD1Combine(N, DCI, false);
9604 case AArch64ISD::NVCAST:
9605 return performNVCASTCombine(N);
9606 case ISD::INSERT_VECTOR_ELT:
9607 return performPostLD1Combine(N, DCI, true);
9608 case ISD::EXTRACT_VECTOR_ELT:
9609 return performAcrossLaneAddReductionCombine(N, DAG, Subtarget);
9610 case ISD::INTRINSIC_VOID:
9611 case ISD::INTRINSIC_W_CHAIN:
9612 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9613 case Intrinsic::aarch64_neon_ld2:
9614 case Intrinsic::aarch64_neon_ld3:
9615 case Intrinsic::aarch64_neon_ld4:
9616 case Intrinsic::aarch64_neon_ld1x2:
9617 case Intrinsic::aarch64_neon_ld1x3:
9618 case Intrinsic::aarch64_neon_ld1x4:
9619 case Intrinsic::aarch64_neon_ld2lane:
9620 case Intrinsic::aarch64_neon_ld3lane:
9621 case Intrinsic::aarch64_neon_ld4lane:
9622 case Intrinsic::aarch64_neon_ld2r:
9623 case Intrinsic::aarch64_neon_ld3r:
9624 case Intrinsic::aarch64_neon_ld4r:
9625 case Intrinsic::aarch64_neon_st2:
9626 case Intrinsic::aarch64_neon_st3:
9627 case Intrinsic::aarch64_neon_st4:
9628 case Intrinsic::aarch64_neon_st1x2:
9629 case Intrinsic::aarch64_neon_st1x3:
9630 case Intrinsic::aarch64_neon_st1x4:
9631 case Intrinsic::aarch64_neon_st2lane:
9632 case Intrinsic::aarch64_neon_st3lane:
9633 case Intrinsic::aarch64_neon_st4lane:
9634 return performNEONPostLDSTCombine(N, DCI, DAG);
9642 // Check if the return value is used as only a return value, as otherwise
9643 // we can't perform a tail-call. In particular, we need to check for
9644 // target ISD nodes that are returns and any other "odd" constructs
9645 // that the generic analysis code won't necessarily catch.
9646 bool AArch64TargetLowering::isUsedByReturnOnly(SDNode *N,
9647 SDValue &Chain) const {
9648 if (N->getNumValues() != 1)
9650 if (!N->hasNUsesOfValue(1, 0))
9653 SDValue TCChain = Chain;
9654 SDNode *Copy = *N->use_begin();
9655 if (Copy->getOpcode() == ISD::CopyToReg) {
9656 // If the copy has a glue operand, we conservatively assume it isn't safe to
9657 // perform a tail call.
9658 if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() ==
9661 TCChain = Copy->getOperand(0);
9662 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
9665 bool HasRet = false;
9666 for (SDNode *Node : Copy->uses()) {
9667 if (Node->getOpcode() != AArch64ISD::RET_FLAG)
9679 // Return whether the an instruction can potentially be optimized to a tail
9680 // call. This will cause the optimizers to attempt to move, or duplicate,
9681 // return instructions to help enable tail call optimizations for this
9683 bool AArch64TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
9684 if (!CI->isTailCall())
9690 bool AArch64TargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base,
9692 ISD::MemIndexedMode &AM,
9694 SelectionDAG &DAG) const {
9695 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
9698 Base = Op->getOperand(0);
9699 // All of the indexed addressing mode instructions take a signed
9700 // 9 bit immediate offset.
9701 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
9702 int64_t RHSC = (int64_t)RHS->getZExtValue();
9703 if (RHSC >= 256 || RHSC <= -256)
9705 IsInc = (Op->getOpcode() == ISD::ADD);
9706 Offset = Op->getOperand(1);
9712 bool AArch64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
9714 ISD::MemIndexedMode &AM,
9715 SelectionDAG &DAG) const {
9718 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9719 VT = LD->getMemoryVT();
9720 Ptr = LD->getBasePtr();
9721 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9722 VT = ST->getMemoryVT();
9723 Ptr = ST->getBasePtr();
9728 if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, IsInc, DAG))
9730 AM = IsInc ? ISD::PRE_INC : ISD::PRE_DEC;
9734 bool AArch64TargetLowering::getPostIndexedAddressParts(
9735 SDNode *N, SDNode *Op, SDValue &Base, SDValue &Offset,
9736 ISD::MemIndexedMode &AM, SelectionDAG &DAG) const {
9739 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9740 VT = LD->getMemoryVT();
9741 Ptr = LD->getBasePtr();
9742 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9743 VT = ST->getMemoryVT();
9744 Ptr = ST->getBasePtr();
9749 if (!getIndexedAddressParts(Op, Base, Offset, AM, IsInc, DAG))
9751 // Post-indexing updates the base, so it's not a valid transform
9752 // if that's not the same as the load's pointer.
9755 AM = IsInc ? ISD::POST_INC : ISD::POST_DEC;
9759 static void ReplaceBITCASTResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
9760 SelectionDAG &DAG) {
9762 SDValue Op = N->getOperand(0);
9764 if (N->getValueType(0) != MVT::i16 || Op.getValueType() != MVT::f16)
9768 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, DL, MVT::f32,
9769 DAG.getUNDEF(MVT::i32), Op,
9770 DAG.getTargetConstant(AArch64::hsub, DL, MVT::i32)),
9772 Op = DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op);
9773 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, Op));
9776 static void ReplaceReductionResults(SDNode *N,
9777 SmallVectorImpl<SDValue> &Results,
9778 SelectionDAG &DAG, unsigned InterOp,
9779 unsigned AcrossOp) {
9783 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
9784 std::tie(Lo, Hi) = DAG.SplitVectorOperand(N, 0);
9785 SDValue InterVal = DAG.getNode(InterOp, dl, LoVT, Lo, Hi);
9786 SDValue SplitVal = DAG.getNode(AcrossOp, dl, LoVT, InterVal);
9787 Results.push_back(SplitVal);
9790 void AArch64TargetLowering::ReplaceNodeResults(
9791 SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
9792 switch (N->getOpcode()) {
9794 llvm_unreachable("Don't know how to custom expand this");
9796 ReplaceBITCASTResults(N, Results, DAG);
9798 case AArch64ISD::SADDV:
9799 ReplaceReductionResults(N, Results, DAG, ISD::ADD, AArch64ISD::SADDV);
9801 case AArch64ISD::UADDV:
9802 ReplaceReductionResults(N, Results, DAG, ISD::ADD, AArch64ISD::UADDV);
9804 case AArch64ISD::SMINV:
9805 ReplaceReductionResults(N, Results, DAG, ISD::SMIN, AArch64ISD::SMINV);
9807 case AArch64ISD::UMINV:
9808 ReplaceReductionResults(N, Results, DAG, ISD::UMIN, AArch64ISD::UMINV);
9810 case AArch64ISD::SMAXV:
9811 ReplaceReductionResults(N, Results, DAG, ISD::SMAX, AArch64ISD::SMAXV);
9813 case AArch64ISD::UMAXV:
9814 ReplaceReductionResults(N, Results, DAG, ISD::UMAX, AArch64ISD::UMAXV);
9816 case ISD::FP_TO_UINT:
9817 case ISD::FP_TO_SINT:
9818 assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion");
9819 // Let normal code take care of it by not adding anything to Results.
9824 bool AArch64TargetLowering::useLoadStackGuardNode() const {
9828 unsigned AArch64TargetLowering::combineRepeatedFPDivisors() const {
9829 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
9830 // reciprocal if there are three or more FDIVs.
9834 TargetLoweringBase::LegalizeTypeAction
9835 AArch64TargetLowering::getPreferredVectorAction(EVT VT) const {
9836 MVT SVT = VT.getSimpleVT();
9837 // During type legalization, we prefer to widen v1i8, v1i16, v1i32 to v8i8,
9838 // v4i16, v2i32 instead of to promote.
9839 if (SVT == MVT::v1i8 || SVT == MVT::v1i16 || SVT == MVT::v1i32
9840 || SVT == MVT::v1f32)
9841 return TypeWidenVector;
9843 return TargetLoweringBase::getPreferredVectorAction(VT);
9846 // Loads and stores less than 128-bits are already atomic; ones above that
9847 // are doomed anyway, so defer to the default libcall and blame the OS when
9849 bool AArch64TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
9850 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
9854 // Loads and stores less than 128-bits are already atomic; ones above that
9855 // are doomed anyway, so defer to the default libcall and blame the OS when
9857 TargetLowering::AtomicExpansionKind
9858 AArch64TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
9859 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
9860 return Size == 128 ? AtomicExpansionKind::LLSC : AtomicExpansionKind::None;
9863 // For the real atomic operations, we have ldxr/stxr up to 128 bits,
9864 TargetLowering::AtomicExpansionKind
9865 AArch64TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
9866 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
9867 return Size <= 128 ? AtomicExpansionKind::LLSC : AtomicExpansionKind::None;
9870 bool AArch64TargetLowering::shouldExpandAtomicCmpXchgInIR(
9871 AtomicCmpXchgInst *AI) const {
9875 Value *AArch64TargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
9876 AtomicOrdering Ord) const {
9877 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
9878 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
9879 bool IsAcquire = isAtLeastAcquire(Ord);
9881 // Since i128 isn't legal and intrinsics don't get type-lowered, the ldrexd
9882 // intrinsic must return {i64, i64} and we have to recombine them into a
9883 // single i128 here.
9884 if (ValTy->getPrimitiveSizeInBits() == 128) {
9886 IsAcquire ? Intrinsic::aarch64_ldaxp : Intrinsic::aarch64_ldxp;
9887 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int);
9889 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
9890 Value *LoHi = Builder.CreateCall(Ldxr, Addr, "lohi");
9892 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
9893 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
9894 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
9895 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
9896 return Builder.CreateOr(
9897 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 64)), "val64");
9900 Type *Tys[] = { Addr->getType() };
9902 IsAcquire ? Intrinsic::aarch64_ldaxr : Intrinsic::aarch64_ldxr;
9903 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int, Tys);
9905 return Builder.CreateTruncOrBitCast(
9906 Builder.CreateCall(Ldxr, Addr),
9907 cast<PointerType>(Addr->getType())->getElementType());
9910 void AArch64TargetLowering::emitAtomicCmpXchgNoStoreLLBalance(
9911 IRBuilder<> &Builder) const {
9912 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
9914 llvm::Intrinsic::getDeclaration(M, Intrinsic::aarch64_clrex));
9917 Value *AArch64TargetLowering::emitStoreConditional(IRBuilder<> &Builder,
9918 Value *Val, Value *Addr,
9919 AtomicOrdering Ord) const {
9920 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
9921 bool IsRelease = isAtLeastRelease(Ord);
9923 // Since the intrinsics must have legal type, the i128 intrinsics take two
9924 // parameters: "i64, i64". We must marshal Val into the appropriate form
9926 if (Val->getType()->getPrimitiveSizeInBits() == 128) {
9928 IsRelease ? Intrinsic::aarch64_stlxp : Intrinsic::aarch64_stxp;
9929 Function *Stxr = Intrinsic::getDeclaration(M, Int);
9930 Type *Int64Ty = Type::getInt64Ty(M->getContext());
9932 Value *Lo = Builder.CreateTrunc(Val, Int64Ty, "lo");
9933 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 64), Int64Ty, "hi");
9934 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
9935 return Builder.CreateCall(Stxr, {Lo, Hi, Addr});
9939 IsRelease ? Intrinsic::aarch64_stlxr : Intrinsic::aarch64_stxr;
9940 Type *Tys[] = { Addr->getType() };
9941 Function *Stxr = Intrinsic::getDeclaration(M, Int, Tys);
9943 return Builder.CreateCall(Stxr,
9944 {Builder.CreateZExtOrBitCast(
9945 Val, Stxr->getFunctionType()->getParamType(0)),
9949 bool AArch64TargetLowering::functionArgumentNeedsConsecutiveRegisters(
9950 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
9951 return Ty->isArrayTy();
9954 bool AArch64TargetLowering::shouldNormalizeToSelectSequence(LLVMContext &,
9959 Value *AArch64TargetLowering::getSafeStackPointerLocation(IRBuilder<> &IRB) const {
9960 if (!Subtarget->isTargetAndroid())
9961 return TargetLowering::getSafeStackPointerLocation(IRB);
9963 // Android provides a fixed TLS slot for the SafeStack pointer. See the
9964 // definition of TLS_SLOT_SAFESTACK in
9965 // https://android.googlesource.com/platform/bionic/+/master/libc/private/bionic_tls.h
9966 const unsigned TlsOffset = 0x48;
9967 Module *M = IRB.GetInsertBlock()->getParent()->getParent();
9968 Function *ThreadPointerFunc =
9969 Intrinsic::getDeclaration(M, Intrinsic::aarch64_thread_pointer);
9970 return IRB.CreatePointerCast(
9971 IRB.CreateConstGEP1_32(IRB.CreateCall(ThreadPointerFunc), TlsOffset),
9972 Type::getInt8PtrTy(IRB.getContext())->getPointerTo(0));