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 "AArch64MachineFunctionInfo.h"
16 #include "AArch64PerfectShuffle.h"
17 #include "AArch64Subtarget.h"
18 #include "AArch64TargetMachine.h"
19 #include "AArch64TargetObjectFile.h"
20 #include "MCTargetDesc/AArch64AddressingModes.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/CodeGen/CallingConvLower.h"
23 #include "llvm/CodeGen/MachineFrameInfo.h"
24 #include "llvm/CodeGen/MachineInstrBuilder.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/IR/Function.h"
27 #include "llvm/IR/Intrinsics.h"
28 #include "llvm/IR/Type.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/raw_ostream.h"
33 #include "llvm/Target/TargetOptions.h"
36 #define DEBUG_TYPE "aarch64-lower"
38 STATISTIC(NumTailCalls, "Number of tail calls");
39 STATISTIC(NumShiftInserts, "Number of vector shift inserts");
46 static cl::opt<AlignMode>
47 Align(cl::desc("Load/store alignment support"),
48 cl::Hidden, cl::init(NoStrictAlign),
50 clEnumValN(StrictAlign, "aarch64-strict-align",
51 "Disallow all unaligned memory accesses"),
52 clEnumValN(NoStrictAlign, "aarch64-no-strict-align",
53 "Allow unaligned memory accesses"),
56 // Place holder until extr generation is tested fully.
58 EnableAArch64ExtrGeneration("aarch64-extr-generation", cl::Hidden,
59 cl::desc("Allow AArch64 (or (shift)(shift))->extract"),
63 EnableAArch64SlrGeneration("aarch64-shift-insert-generation", cl::Hidden,
64 cl::desc("Allow AArch64 SLI/SRI formation"),
67 //===----------------------------------------------------------------------===//
68 // AArch64 Lowering public interface.
69 //===----------------------------------------------------------------------===//
70 static TargetLoweringObjectFile *createTLOF(const Triple &TT) {
71 if (TT.isOSBinFormatMachO())
72 return new AArch64_MachoTargetObjectFile();
74 return new AArch64_ELFTargetObjectFile();
77 AArch64TargetLowering::AArch64TargetLowering(TargetMachine &TM)
78 : TargetLowering(TM, createTLOF(Triple(TM.getTargetTriple()))) {
79 Subtarget = &TM.getSubtarget<AArch64Subtarget>();
81 // AArch64 doesn't have comparisons which set GPRs or setcc instructions, so
82 // we have to make something up. Arbitrarily, choose ZeroOrOne.
83 setBooleanContents(ZeroOrOneBooleanContent);
84 // When comparing vectors the result sets the different elements in the
85 // vector to all-one or all-zero.
86 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
88 // Set up the register classes.
89 addRegisterClass(MVT::i32, &AArch64::GPR32allRegClass);
90 addRegisterClass(MVT::i64, &AArch64::GPR64allRegClass);
92 if (Subtarget->hasFPARMv8()) {
93 addRegisterClass(MVT::f16, &AArch64::FPR16RegClass);
94 addRegisterClass(MVT::f32, &AArch64::FPR32RegClass);
95 addRegisterClass(MVT::f64, &AArch64::FPR64RegClass);
96 addRegisterClass(MVT::f128, &AArch64::FPR128RegClass);
99 if (Subtarget->hasNEON()) {
100 addRegisterClass(MVT::v16i8, &AArch64::FPR8RegClass);
101 addRegisterClass(MVT::v8i16, &AArch64::FPR16RegClass);
102 // Someone set us up the NEON.
103 addDRTypeForNEON(MVT::v2f32);
104 addDRTypeForNEON(MVT::v8i8);
105 addDRTypeForNEON(MVT::v4i16);
106 addDRTypeForNEON(MVT::v2i32);
107 addDRTypeForNEON(MVT::v1i64);
108 addDRTypeForNEON(MVT::v1f64);
110 addQRTypeForNEON(MVT::v4f32);
111 addQRTypeForNEON(MVT::v2f64);
112 addQRTypeForNEON(MVT::v16i8);
113 addQRTypeForNEON(MVT::v8i16);
114 addQRTypeForNEON(MVT::v4i32);
115 addQRTypeForNEON(MVT::v2i64);
118 // Compute derived properties from the register classes
119 computeRegisterProperties();
121 // Provide all sorts of operation actions
122 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
123 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
124 setOperationAction(ISD::SETCC, MVT::i32, Custom);
125 setOperationAction(ISD::SETCC, MVT::i64, Custom);
126 setOperationAction(ISD::SETCC, MVT::f32, Custom);
127 setOperationAction(ISD::SETCC, MVT::f64, Custom);
128 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
129 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
130 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
131 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
132 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
133 setOperationAction(ISD::SELECT, MVT::i32, Custom);
134 setOperationAction(ISD::SELECT, MVT::i64, Custom);
135 setOperationAction(ISD::SELECT, MVT::f32, Custom);
136 setOperationAction(ISD::SELECT, MVT::f64, Custom);
137 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
138 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
139 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
140 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
141 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
142 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
144 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
145 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
146 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
148 setOperationAction(ISD::FREM, MVT::f32, Expand);
149 setOperationAction(ISD::FREM, MVT::f64, Expand);
150 setOperationAction(ISD::FREM, MVT::f80, Expand);
152 // Custom lowering hooks are needed for XOR
153 // to fold it into CSINC/CSINV.
154 setOperationAction(ISD::XOR, MVT::i32, Custom);
155 setOperationAction(ISD::XOR, MVT::i64, Custom);
157 // Virtually no operation on f128 is legal, but LLVM can't expand them when
158 // there's a valid register class, so we need custom operations in most cases.
159 setOperationAction(ISD::FABS, MVT::f128, Expand);
160 setOperationAction(ISD::FADD, MVT::f128, Custom);
161 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
162 setOperationAction(ISD::FCOS, MVT::f128, Expand);
163 setOperationAction(ISD::FDIV, MVT::f128, Custom);
164 setOperationAction(ISD::FMA, MVT::f128, Expand);
165 setOperationAction(ISD::FMUL, MVT::f128, Custom);
166 setOperationAction(ISD::FNEG, MVT::f128, Expand);
167 setOperationAction(ISD::FPOW, MVT::f128, Expand);
168 setOperationAction(ISD::FREM, MVT::f128, Expand);
169 setOperationAction(ISD::FRINT, MVT::f128, Expand);
170 setOperationAction(ISD::FSIN, MVT::f128, Expand);
171 setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
172 setOperationAction(ISD::FSQRT, MVT::f128, Expand);
173 setOperationAction(ISD::FSUB, MVT::f128, Custom);
174 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
175 setOperationAction(ISD::SETCC, MVT::f128, Custom);
176 setOperationAction(ISD::BR_CC, MVT::f128, Custom);
177 setOperationAction(ISD::SELECT, MVT::f128, Custom);
178 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
179 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);
181 // Lowering for many of the conversions is actually specified by the non-f128
182 // type. The LowerXXX function will be trivial when f128 isn't involved.
183 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
184 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
185 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
186 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
187 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
188 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
189 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
190 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
191 setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
192 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
193 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
194 setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
195 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
196 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
198 // Variable arguments.
199 setOperationAction(ISD::VASTART, MVT::Other, Custom);
200 setOperationAction(ISD::VAARG, MVT::Other, Custom);
201 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
202 setOperationAction(ISD::VAEND, MVT::Other, Expand);
204 // Variable-sized objects.
205 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
206 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
207 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
209 // Exception handling.
210 // FIXME: These are guesses. Has this been defined yet?
211 setExceptionPointerRegister(AArch64::X0);
212 setExceptionSelectorRegister(AArch64::X1);
214 // Constant pool entries
215 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
218 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
220 // Add/Sub overflow ops with MVT::Glues are lowered to NZCV dependences.
221 setOperationAction(ISD::ADDC, MVT::i32, Custom);
222 setOperationAction(ISD::ADDE, MVT::i32, Custom);
223 setOperationAction(ISD::SUBC, MVT::i32, Custom);
224 setOperationAction(ISD::SUBE, MVT::i32, Custom);
225 setOperationAction(ISD::ADDC, MVT::i64, Custom);
226 setOperationAction(ISD::ADDE, MVT::i64, Custom);
227 setOperationAction(ISD::SUBC, MVT::i64, Custom);
228 setOperationAction(ISD::SUBE, MVT::i64, Custom);
230 // AArch64 lacks both left-rotate and popcount instructions.
231 setOperationAction(ISD::ROTL, MVT::i32, Expand);
232 setOperationAction(ISD::ROTL, MVT::i64, Expand);
234 // AArch64 doesn't have {U|S}MUL_LOHI.
235 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
236 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
239 // Expand the undefined-at-zero variants to cttz/ctlz to their defined-at-zero
240 // counterparts, which AArch64 supports directly.
241 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
242 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
243 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
244 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
246 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
247 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
249 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
250 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
251 setOperationAction(ISD::SREM, MVT::i32, Expand);
252 setOperationAction(ISD::SREM, MVT::i64, Expand);
253 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
254 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
255 setOperationAction(ISD::UREM, MVT::i32, Expand);
256 setOperationAction(ISD::UREM, MVT::i64, Expand);
258 // Custom lower Add/Sub/Mul with overflow.
259 setOperationAction(ISD::SADDO, MVT::i32, Custom);
260 setOperationAction(ISD::SADDO, MVT::i64, Custom);
261 setOperationAction(ISD::UADDO, MVT::i32, Custom);
262 setOperationAction(ISD::UADDO, MVT::i64, Custom);
263 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
264 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
265 setOperationAction(ISD::USUBO, MVT::i32, Custom);
266 setOperationAction(ISD::USUBO, MVT::i64, Custom);
267 setOperationAction(ISD::SMULO, MVT::i32, Custom);
268 setOperationAction(ISD::SMULO, MVT::i64, Custom);
269 setOperationAction(ISD::UMULO, MVT::i32, Custom);
270 setOperationAction(ISD::UMULO, MVT::i64, Custom);
272 setOperationAction(ISD::FSIN, MVT::f32, Expand);
273 setOperationAction(ISD::FSIN, MVT::f64, Expand);
274 setOperationAction(ISD::FCOS, MVT::f32, Expand);
275 setOperationAction(ISD::FCOS, MVT::f64, Expand);
276 setOperationAction(ISD::FPOW, MVT::f32, Expand);
277 setOperationAction(ISD::FPOW, MVT::f64, Expand);
278 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
279 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
281 // AArch64 has implementations of a lot of rounding-like FP operations.
282 static MVT RoundingTypes[] = { MVT::f32, MVT::f64};
283 for (unsigned I = 0; I < array_lengthof(RoundingTypes); ++I) {
284 MVT Ty = RoundingTypes[I];
285 setOperationAction(ISD::FFLOOR, Ty, Legal);
286 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
287 setOperationAction(ISD::FCEIL, Ty, Legal);
288 setOperationAction(ISD::FRINT, Ty, Legal);
289 setOperationAction(ISD::FTRUNC, Ty, Legal);
290 setOperationAction(ISD::FROUND, Ty, Legal);
293 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
295 if (Subtarget->isTargetMachO()) {
296 // For iOS, we don't want to the normal expansion of a libcall to
297 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
299 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
300 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
302 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
303 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
306 // AArch64 does not have floating-point extending loads, i1 sign-extending
307 // load, floating-point truncating stores, or v2i32->v2i16 truncating store.
308 setLoadExtAction(ISD::EXTLOAD, MVT::f16, Expand);
309 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
310 setLoadExtAction(ISD::EXTLOAD, MVT::f64, Expand);
311 setLoadExtAction(ISD::EXTLOAD, MVT::f80, Expand);
312 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Expand);
313 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
314 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
315 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
316 setTruncStoreAction(MVT::f128, MVT::f80, Expand);
317 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
318 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
319 setTruncStoreAction(MVT::f128, MVT::f16, Expand);
321 setOperationAction(ISD::BITCAST, MVT::i16, Custom);
322 setOperationAction(ISD::BITCAST, MVT::f16, Custom);
324 // Indexed loads and stores are supported.
325 for (unsigned im = (unsigned)ISD::PRE_INC;
326 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
327 setIndexedLoadAction(im, MVT::i8, Legal);
328 setIndexedLoadAction(im, MVT::i16, Legal);
329 setIndexedLoadAction(im, MVT::i32, Legal);
330 setIndexedLoadAction(im, MVT::i64, Legal);
331 setIndexedLoadAction(im, MVT::f64, Legal);
332 setIndexedLoadAction(im, MVT::f32, Legal);
333 setIndexedStoreAction(im, MVT::i8, Legal);
334 setIndexedStoreAction(im, MVT::i16, Legal);
335 setIndexedStoreAction(im, MVT::i32, Legal);
336 setIndexedStoreAction(im, MVT::i64, Legal);
337 setIndexedStoreAction(im, MVT::f64, Legal);
338 setIndexedStoreAction(im, MVT::f32, Legal);
342 setOperationAction(ISD::TRAP, MVT::Other, Legal);
344 // We combine OR nodes for bitfield operations.
345 setTargetDAGCombine(ISD::OR);
347 // Vector add and sub nodes may conceal a high-half opportunity.
348 // Also, try to fold ADD into CSINC/CSINV..
349 setTargetDAGCombine(ISD::ADD);
350 setTargetDAGCombine(ISD::SUB);
352 setTargetDAGCombine(ISD::XOR);
353 setTargetDAGCombine(ISD::SINT_TO_FP);
354 setTargetDAGCombine(ISD::UINT_TO_FP);
356 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
358 setTargetDAGCombine(ISD::ANY_EXTEND);
359 setTargetDAGCombine(ISD::ZERO_EXTEND);
360 setTargetDAGCombine(ISD::SIGN_EXTEND);
361 setTargetDAGCombine(ISD::BITCAST);
362 setTargetDAGCombine(ISD::CONCAT_VECTORS);
363 setTargetDAGCombine(ISD::STORE);
365 setTargetDAGCombine(ISD::MUL);
367 setTargetDAGCombine(ISD::SELECT);
368 setTargetDAGCombine(ISD::VSELECT);
370 setTargetDAGCombine(ISD::INTRINSIC_VOID);
371 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
372 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
374 MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8;
375 MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4;
376 MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 4;
378 setStackPointerRegisterToSaveRestore(AArch64::SP);
380 setSchedulingPreference(Sched::Hybrid);
383 MaskAndBranchFoldingIsLegal = true;
385 setMinFunctionAlignment(2);
387 RequireStrictAlign = (Align == StrictAlign);
389 setHasExtractBitsInsn(true);
391 if (Subtarget->hasNEON()) {
392 // FIXME: v1f64 shouldn't be legal if we can avoid it, because it leads to
393 // silliness like this:
394 setOperationAction(ISD::FABS, MVT::v1f64, Expand);
395 setOperationAction(ISD::FADD, MVT::v1f64, Expand);
396 setOperationAction(ISD::FCEIL, MVT::v1f64, Expand);
397 setOperationAction(ISD::FCOPYSIGN, MVT::v1f64, Expand);
398 setOperationAction(ISD::FCOS, MVT::v1f64, Expand);
399 setOperationAction(ISD::FDIV, MVT::v1f64, Expand);
400 setOperationAction(ISD::FFLOOR, MVT::v1f64, Expand);
401 setOperationAction(ISD::FMA, MVT::v1f64, Expand);
402 setOperationAction(ISD::FMUL, MVT::v1f64, Expand);
403 setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Expand);
404 setOperationAction(ISD::FNEG, MVT::v1f64, Expand);
405 setOperationAction(ISD::FPOW, MVT::v1f64, Expand);
406 setOperationAction(ISD::FREM, MVT::v1f64, Expand);
407 setOperationAction(ISD::FROUND, MVT::v1f64, Expand);
408 setOperationAction(ISD::FRINT, MVT::v1f64, Expand);
409 setOperationAction(ISD::FSIN, MVT::v1f64, Expand);
410 setOperationAction(ISD::FSINCOS, MVT::v1f64, Expand);
411 setOperationAction(ISD::FSQRT, MVT::v1f64, Expand);
412 setOperationAction(ISD::FSUB, MVT::v1f64, Expand);
413 setOperationAction(ISD::FTRUNC, MVT::v1f64, Expand);
414 setOperationAction(ISD::SETCC, MVT::v1f64, Expand);
415 setOperationAction(ISD::BR_CC, MVT::v1f64, Expand);
416 setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
417 setOperationAction(ISD::SELECT_CC, MVT::v1f64, Expand);
418 setOperationAction(ISD::FP_EXTEND, MVT::v1f64, Expand);
420 setOperationAction(ISD::FP_TO_SINT, MVT::v1i64, Expand);
421 setOperationAction(ISD::FP_TO_UINT, MVT::v1i64, Expand);
422 setOperationAction(ISD::SINT_TO_FP, MVT::v1i64, Expand);
423 setOperationAction(ISD::UINT_TO_FP, MVT::v1i64, Expand);
424 setOperationAction(ISD::FP_ROUND, MVT::v1f64, Expand);
426 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
428 // AArch64 doesn't have a direct vector ->f32 conversion instructions for
429 // elements smaller than i32, so promote the input to i32 first.
430 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Promote);
431 setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Promote);
432 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Promote);
433 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Promote);
434 // Similarly, there is no direct i32 -> f64 vector conversion instruction.
435 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
436 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
437 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
438 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
440 // AArch64 doesn't have MUL.2d:
441 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
442 setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal);
443 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
444 // Likewise, narrowing and extending vector loads/stores aren't handled
446 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
447 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
449 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,
452 setOperationAction(ISD::MULHS, (MVT::SimpleValueType)VT, Expand);
453 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
454 setOperationAction(ISD::MULHU, (MVT::SimpleValueType)VT, Expand);
455 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
457 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
459 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
460 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
461 setTruncStoreAction((MVT::SimpleValueType)VT,
462 (MVT::SimpleValueType)InnerVT, Expand);
463 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
464 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
465 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
468 // AArch64 has implementations of a lot of rounding-like FP operations.
469 static MVT RoundingVecTypes[] = {MVT::v2f32, MVT::v4f32, MVT::v2f64 };
470 for (unsigned I = 0; I < array_lengthof(RoundingVecTypes); ++I) {
471 MVT Ty = RoundingVecTypes[I];
472 setOperationAction(ISD::FFLOOR, Ty, Legal);
473 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
474 setOperationAction(ISD::FCEIL, Ty, Legal);
475 setOperationAction(ISD::FRINT, Ty, Legal);
476 setOperationAction(ISD::FTRUNC, Ty, Legal);
477 setOperationAction(ISD::FROUND, Ty, Legal);
481 // Prefer likely predicted branches to selects on out-of-order cores.
482 if (Subtarget->isCortexA57())
483 PredictableSelectIsExpensive = true;
486 void AArch64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) {
487 if (VT == MVT::v2f32) {
488 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
489 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i32);
491 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
492 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i32);
493 } else if (VT == MVT::v2f64 || VT == MVT::v4f32) {
494 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
495 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i64);
497 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
498 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i64);
501 // Mark vector float intrinsics as expand.
502 if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) {
503 setOperationAction(ISD::FSIN, VT.getSimpleVT(), Expand);
504 setOperationAction(ISD::FCOS, VT.getSimpleVT(), Expand);
505 setOperationAction(ISD::FPOWI, VT.getSimpleVT(), Expand);
506 setOperationAction(ISD::FPOW, VT.getSimpleVT(), Expand);
507 setOperationAction(ISD::FLOG, VT.getSimpleVT(), Expand);
508 setOperationAction(ISD::FLOG2, VT.getSimpleVT(), Expand);
509 setOperationAction(ISD::FLOG10, VT.getSimpleVT(), Expand);
510 setOperationAction(ISD::FEXP, VT.getSimpleVT(), Expand);
511 setOperationAction(ISD::FEXP2, VT.getSimpleVT(), Expand);
514 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
515 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
516 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
517 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
518 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Custom);
519 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
520 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
521 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
522 setOperationAction(ISD::AND, VT.getSimpleVT(), Custom);
523 setOperationAction(ISD::OR, VT.getSimpleVT(), Custom);
524 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
525 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
527 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
528 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
529 setOperationAction(ISD::VSELECT, VT.getSimpleVT(), Expand);
530 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
532 // CNT supports only B element sizes.
533 if (VT != MVT::v8i8 && VT != MVT::v16i8)
534 setOperationAction(ISD::CTPOP, VT.getSimpleVT(), Expand);
536 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
537 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
538 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
539 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
540 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
542 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
543 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
545 if (Subtarget->isLittleEndian()) {
546 for (unsigned im = (unsigned)ISD::PRE_INC;
547 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
548 setIndexedLoadAction(im, VT.getSimpleVT(), Legal);
549 setIndexedStoreAction(im, VT.getSimpleVT(), Legal);
554 void AArch64TargetLowering::addDRTypeForNEON(MVT VT) {
555 addRegisterClass(VT, &AArch64::FPR64RegClass);
556 addTypeForNEON(VT, MVT::v2i32);
559 void AArch64TargetLowering::addQRTypeForNEON(MVT VT) {
560 addRegisterClass(VT, &AArch64::FPR128RegClass);
561 addTypeForNEON(VT, MVT::v4i32);
564 EVT AArch64TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
567 return VT.changeVectorElementTypeToInteger();
570 /// computeKnownBitsForTargetNode - Determine which of the bits specified in
571 /// Mask are known to be either zero or one and return them in the
572 /// KnownZero/KnownOne bitsets.
573 void AArch64TargetLowering::computeKnownBitsForTargetNode(
574 const SDValue Op, APInt &KnownZero, APInt &KnownOne,
575 const SelectionDAG &DAG, unsigned Depth) const {
576 switch (Op.getOpcode()) {
579 case AArch64ISD::CSEL: {
580 APInt KnownZero2, KnownOne2;
581 DAG.computeKnownBits(Op->getOperand(0), KnownZero, KnownOne, Depth + 1);
582 DAG.computeKnownBits(Op->getOperand(1), KnownZero2, KnownOne2, Depth + 1);
583 KnownZero &= KnownZero2;
584 KnownOne &= KnownOne2;
587 case ISD::INTRINSIC_W_CHAIN: {
588 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
589 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
592 case Intrinsic::aarch64_ldaxr:
593 case Intrinsic::aarch64_ldxr: {
594 unsigned BitWidth = KnownOne.getBitWidth();
595 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
596 unsigned MemBits = VT.getScalarType().getSizeInBits();
597 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
603 case ISD::INTRINSIC_WO_CHAIN:
604 case ISD::INTRINSIC_VOID: {
605 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
609 case Intrinsic::aarch64_neon_umaxv:
610 case Intrinsic::aarch64_neon_uminv: {
611 // Figure out the datatype of the vector operand. The UMINV instruction
612 // will zero extend the result, so we can mark as known zero all the
613 // bits larger than the element datatype. 32-bit or larget doesn't need
614 // this as those are legal types and will be handled by isel directly.
615 MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
616 unsigned BitWidth = KnownZero.getBitWidth();
617 if (VT == MVT::v8i8 || VT == MVT::v16i8) {
618 assert(BitWidth >= 8 && "Unexpected width!");
619 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8);
621 } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
622 assert(BitWidth >= 16 && "Unexpected width!");
623 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
633 MVT AArch64TargetLowering::getScalarShiftAmountTy(EVT LHSTy) const {
637 unsigned AArch64TargetLowering::getMaximalGlobalOffset() const {
638 // FIXME: On AArch64, this depends on the type.
639 // Basically, the addressable offsets are up to 4095 * Ty.getSizeInBytes().
640 // and the offset has to be a multiple of the related size in bytes.
645 AArch64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
646 const TargetLibraryInfo *libInfo) const {
647 return AArch64::createFastISel(funcInfo, libInfo);
650 const char *AArch64TargetLowering::getTargetNodeName(unsigned Opcode) const {
654 case AArch64ISD::CALL: return "AArch64ISD::CALL";
655 case AArch64ISD::ADRP: return "AArch64ISD::ADRP";
656 case AArch64ISD::ADDlow: return "AArch64ISD::ADDlow";
657 case AArch64ISD::LOADgot: return "AArch64ISD::LOADgot";
658 case AArch64ISD::RET_FLAG: return "AArch64ISD::RET_FLAG";
659 case AArch64ISD::BRCOND: return "AArch64ISD::BRCOND";
660 case AArch64ISD::CSEL: return "AArch64ISD::CSEL";
661 case AArch64ISD::FCSEL: return "AArch64ISD::FCSEL";
662 case AArch64ISD::CSINV: return "AArch64ISD::CSINV";
663 case AArch64ISD::CSNEG: return "AArch64ISD::CSNEG";
664 case AArch64ISD::CSINC: return "AArch64ISD::CSINC";
665 case AArch64ISD::THREAD_POINTER: return "AArch64ISD::THREAD_POINTER";
666 case AArch64ISD::TLSDESC_CALL: return "AArch64ISD::TLSDESC_CALL";
667 case AArch64ISD::ADC: return "AArch64ISD::ADC";
668 case AArch64ISD::SBC: return "AArch64ISD::SBC";
669 case AArch64ISD::ADDS: return "AArch64ISD::ADDS";
670 case AArch64ISD::SUBS: return "AArch64ISD::SUBS";
671 case AArch64ISD::ADCS: return "AArch64ISD::ADCS";
672 case AArch64ISD::SBCS: return "AArch64ISD::SBCS";
673 case AArch64ISD::ANDS: return "AArch64ISD::ANDS";
674 case AArch64ISD::FCMP: return "AArch64ISD::FCMP";
675 case AArch64ISD::FMIN: return "AArch64ISD::FMIN";
676 case AArch64ISD::FMAX: return "AArch64ISD::FMAX";
677 case AArch64ISD::DUP: return "AArch64ISD::DUP";
678 case AArch64ISD::DUPLANE8: return "AArch64ISD::DUPLANE8";
679 case AArch64ISD::DUPLANE16: return "AArch64ISD::DUPLANE16";
680 case AArch64ISD::DUPLANE32: return "AArch64ISD::DUPLANE32";
681 case AArch64ISD::DUPLANE64: return "AArch64ISD::DUPLANE64";
682 case AArch64ISD::MOVI: return "AArch64ISD::MOVI";
683 case AArch64ISD::MOVIshift: return "AArch64ISD::MOVIshift";
684 case AArch64ISD::MOVIedit: return "AArch64ISD::MOVIedit";
685 case AArch64ISD::MOVImsl: return "AArch64ISD::MOVImsl";
686 case AArch64ISD::FMOV: return "AArch64ISD::FMOV";
687 case AArch64ISD::MVNIshift: return "AArch64ISD::MVNIshift";
688 case AArch64ISD::MVNImsl: return "AArch64ISD::MVNImsl";
689 case AArch64ISD::BICi: return "AArch64ISD::BICi";
690 case AArch64ISD::ORRi: return "AArch64ISD::ORRi";
691 case AArch64ISD::BSL: return "AArch64ISD::BSL";
692 case AArch64ISD::NEG: return "AArch64ISD::NEG";
693 case AArch64ISD::EXTR: return "AArch64ISD::EXTR";
694 case AArch64ISD::ZIP1: return "AArch64ISD::ZIP1";
695 case AArch64ISD::ZIP2: return "AArch64ISD::ZIP2";
696 case AArch64ISD::UZP1: return "AArch64ISD::UZP1";
697 case AArch64ISD::UZP2: return "AArch64ISD::UZP2";
698 case AArch64ISD::TRN1: return "AArch64ISD::TRN1";
699 case AArch64ISD::TRN2: return "AArch64ISD::TRN2";
700 case AArch64ISD::REV16: return "AArch64ISD::REV16";
701 case AArch64ISD::REV32: return "AArch64ISD::REV32";
702 case AArch64ISD::REV64: return "AArch64ISD::REV64";
703 case AArch64ISD::EXT: return "AArch64ISD::EXT";
704 case AArch64ISD::VSHL: return "AArch64ISD::VSHL";
705 case AArch64ISD::VLSHR: return "AArch64ISD::VLSHR";
706 case AArch64ISD::VASHR: return "AArch64ISD::VASHR";
707 case AArch64ISD::CMEQ: return "AArch64ISD::CMEQ";
708 case AArch64ISD::CMGE: return "AArch64ISD::CMGE";
709 case AArch64ISD::CMGT: return "AArch64ISD::CMGT";
710 case AArch64ISD::CMHI: return "AArch64ISD::CMHI";
711 case AArch64ISD::CMHS: return "AArch64ISD::CMHS";
712 case AArch64ISD::FCMEQ: return "AArch64ISD::FCMEQ";
713 case AArch64ISD::FCMGE: return "AArch64ISD::FCMGE";
714 case AArch64ISD::FCMGT: return "AArch64ISD::FCMGT";
715 case AArch64ISD::CMEQz: return "AArch64ISD::CMEQz";
716 case AArch64ISD::CMGEz: return "AArch64ISD::CMGEz";
717 case AArch64ISD::CMGTz: return "AArch64ISD::CMGTz";
718 case AArch64ISD::CMLEz: return "AArch64ISD::CMLEz";
719 case AArch64ISD::CMLTz: return "AArch64ISD::CMLTz";
720 case AArch64ISD::FCMEQz: return "AArch64ISD::FCMEQz";
721 case AArch64ISD::FCMGEz: return "AArch64ISD::FCMGEz";
722 case AArch64ISD::FCMGTz: return "AArch64ISD::FCMGTz";
723 case AArch64ISD::FCMLEz: return "AArch64ISD::FCMLEz";
724 case AArch64ISD::FCMLTz: return "AArch64ISD::FCMLTz";
725 case AArch64ISD::NOT: return "AArch64ISD::NOT";
726 case AArch64ISD::BIT: return "AArch64ISD::BIT";
727 case AArch64ISD::CBZ: return "AArch64ISD::CBZ";
728 case AArch64ISD::CBNZ: return "AArch64ISD::CBNZ";
729 case AArch64ISD::TBZ: return "AArch64ISD::TBZ";
730 case AArch64ISD::TBNZ: return "AArch64ISD::TBNZ";
731 case AArch64ISD::TC_RETURN: return "AArch64ISD::TC_RETURN";
732 case AArch64ISD::SITOF: return "AArch64ISD::SITOF";
733 case AArch64ISD::UITOF: return "AArch64ISD::UITOF";
734 case AArch64ISD::SQSHL_I: return "AArch64ISD::SQSHL_I";
735 case AArch64ISD::UQSHL_I: return "AArch64ISD::UQSHL_I";
736 case AArch64ISD::SRSHR_I: return "AArch64ISD::SRSHR_I";
737 case AArch64ISD::URSHR_I: return "AArch64ISD::URSHR_I";
738 case AArch64ISD::SQSHLU_I: return "AArch64ISD::SQSHLU_I";
739 case AArch64ISD::WrapperLarge: return "AArch64ISD::WrapperLarge";
740 case AArch64ISD::LD2post: return "AArch64ISD::LD2post";
741 case AArch64ISD::LD3post: return "AArch64ISD::LD3post";
742 case AArch64ISD::LD4post: return "AArch64ISD::LD4post";
743 case AArch64ISD::ST2post: return "AArch64ISD::ST2post";
744 case AArch64ISD::ST3post: return "AArch64ISD::ST3post";
745 case AArch64ISD::ST4post: return "AArch64ISD::ST4post";
746 case AArch64ISD::LD1x2post: return "AArch64ISD::LD1x2post";
747 case AArch64ISD::LD1x3post: return "AArch64ISD::LD1x3post";
748 case AArch64ISD::LD1x4post: return "AArch64ISD::LD1x4post";
749 case AArch64ISD::ST1x2post: return "AArch64ISD::ST1x2post";
750 case AArch64ISD::ST1x3post: return "AArch64ISD::ST1x3post";
751 case AArch64ISD::ST1x4post: return "AArch64ISD::ST1x4post";
752 case AArch64ISD::LD1DUPpost: return "AArch64ISD::LD1DUPpost";
753 case AArch64ISD::LD2DUPpost: return "AArch64ISD::LD2DUPpost";
754 case AArch64ISD::LD3DUPpost: return "AArch64ISD::LD3DUPpost";
755 case AArch64ISD::LD4DUPpost: return "AArch64ISD::LD4DUPpost";
756 case AArch64ISD::LD1LANEpost: return "AArch64ISD::LD1LANEpost";
757 case AArch64ISD::LD2LANEpost: return "AArch64ISD::LD2LANEpost";
758 case AArch64ISD::LD3LANEpost: return "AArch64ISD::LD3LANEpost";
759 case AArch64ISD::LD4LANEpost: return "AArch64ISD::LD4LANEpost";
760 case AArch64ISD::ST2LANEpost: return "AArch64ISD::ST2LANEpost";
761 case AArch64ISD::ST3LANEpost: return "AArch64ISD::ST3LANEpost";
762 case AArch64ISD::ST4LANEpost: return "AArch64ISD::ST4LANEpost";
767 AArch64TargetLowering::EmitF128CSEL(MachineInstr *MI,
768 MachineBasicBlock *MBB) const {
769 // We materialise the F128CSEL pseudo-instruction as some control flow and a
773 // [... previous instrs leading to comparison ...]
779 // Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB]
781 const TargetInstrInfo *TII =
782 getTargetMachine().getSubtargetImpl()->getInstrInfo();
783 MachineFunction *MF = MBB->getParent();
784 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
785 DebugLoc DL = MI->getDebugLoc();
786 MachineFunction::iterator It = MBB;
789 unsigned DestReg = MI->getOperand(0).getReg();
790 unsigned IfTrueReg = MI->getOperand(1).getReg();
791 unsigned IfFalseReg = MI->getOperand(2).getReg();
792 unsigned CondCode = MI->getOperand(3).getImm();
793 bool NZCVKilled = MI->getOperand(4).isKill();
795 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
796 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
797 MF->insert(It, TrueBB);
798 MF->insert(It, EndBB);
800 // Transfer rest of current basic-block to EndBB
801 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
803 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
805 BuildMI(MBB, DL, TII->get(AArch64::Bcc)).addImm(CondCode).addMBB(TrueBB);
806 BuildMI(MBB, DL, TII->get(AArch64::B)).addMBB(EndBB);
807 MBB->addSuccessor(TrueBB);
808 MBB->addSuccessor(EndBB);
810 // TrueBB falls through to the end.
811 TrueBB->addSuccessor(EndBB);
814 TrueBB->addLiveIn(AArch64::NZCV);
815 EndBB->addLiveIn(AArch64::NZCV);
818 BuildMI(*EndBB, EndBB->begin(), DL, TII->get(AArch64::PHI), DestReg)
824 MI->eraseFromParent();
829 AArch64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
830 MachineBasicBlock *BB) const {
831 switch (MI->getOpcode()) {
836 llvm_unreachable("Unexpected instruction for custom inserter!");
838 case AArch64::F128CSEL:
839 return EmitF128CSEL(MI, BB);
841 case TargetOpcode::STACKMAP:
842 case TargetOpcode::PATCHPOINT:
843 return emitPatchPoint(MI, BB);
847 //===----------------------------------------------------------------------===//
848 // AArch64 Lowering private implementation.
849 //===----------------------------------------------------------------------===//
851 //===----------------------------------------------------------------------===//
853 //===----------------------------------------------------------------------===//
855 /// changeIntCCToAArch64CC - Convert a DAG integer condition code to an AArch64
857 static AArch64CC::CondCode changeIntCCToAArch64CC(ISD::CondCode CC) {
860 llvm_unreachable("Unknown condition code!");
862 return AArch64CC::NE;
864 return AArch64CC::EQ;
866 return AArch64CC::GT;
868 return AArch64CC::GE;
870 return AArch64CC::LT;
872 return AArch64CC::LE;
874 return AArch64CC::HI;
876 return AArch64CC::HS;
878 return AArch64CC::LO;
880 return AArch64CC::LS;
884 /// changeFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64 CC.
885 static void changeFPCCToAArch64CC(ISD::CondCode CC,
886 AArch64CC::CondCode &CondCode,
887 AArch64CC::CondCode &CondCode2) {
888 CondCode2 = AArch64CC::AL;
891 llvm_unreachable("Unknown FP condition!");
894 CondCode = AArch64CC::EQ;
898 CondCode = AArch64CC::GT;
902 CondCode = AArch64CC::GE;
905 CondCode = AArch64CC::MI;
908 CondCode = AArch64CC::LS;
911 CondCode = AArch64CC::MI;
912 CondCode2 = AArch64CC::GT;
915 CondCode = AArch64CC::VC;
918 CondCode = AArch64CC::VS;
921 CondCode = AArch64CC::EQ;
922 CondCode2 = AArch64CC::VS;
925 CondCode = AArch64CC::HI;
928 CondCode = AArch64CC::PL;
932 CondCode = AArch64CC::LT;
936 CondCode = AArch64CC::LE;
940 CondCode = AArch64CC::NE;
945 /// changeVectorFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64
946 /// CC usable with the vector instructions. Fewer operations are available
947 /// without a real NZCV register, so we have to use less efficient combinations
948 /// to get the same effect.
949 static void changeVectorFPCCToAArch64CC(ISD::CondCode CC,
950 AArch64CC::CondCode &CondCode,
951 AArch64CC::CondCode &CondCode2,
956 // Mostly the scalar mappings work fine.
957 changeFPCCToAArch64CC(CC, CondCode, CondCode2);
960 Invert = true; // Fallthrough
962 CondCode = AArch64CC::MI;
963 CondCode2 = AArch64CC::GE;
970 // All of the compare-mask comparisons are ordered, but we can switch
971 // between the two by a double inversion. E.g. ULE == !OGT.
973 changeFPCCToAArch64CC(getSetCCInverse(CC, false), CondCode, CondCode2);
978 static bool isLegalArithImmed(uint64_t C) {
979 // Matches AArch64DAGToDAGISel::SelectArithImmed().
980 return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
983 static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
984 SDLoc dl, SelectionDAG &DAG) {
985 EVT VT = LHS.getValueType();
987 if (VT.isFloatingPoint())
988 return DAG.getNode(AArch64ISD::FCMP, dl, VT, LHS, RHS);
990 // The CMP instruction is just an alias for SUBS, and representing it as
991 // SUBS means that it's possible to get CSE with subtract operations.
992 // A later phase can perform the optimization of setting the destination
993 // register to WZR/XZR if it ends up being unused.
994 unsigned Opcode = AArch64ISD::SUBS;
996 if (RHS.getOpcode() == ISD::SUB && isa<ConstantSDNode>(RHS.getOperand(0)) &&
997 cast<ConstantSDNode>(RHS.getOperand(0))->getZExtValue() == 0 &&
998 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
999 // We'd like to combine a (CMP op1, (sub 0, op2) into a CMN instruction on
1000 // the grounds that "op1 - (-op2) == op1 + op2". However, the C and V flags
1001 // can be set differently by this operation. It comes down to whether
1002 // "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are then
1003 // everything is fine. If not then the optimization is wrong. Thus general
1004 // comparisons are only valid if op2 != 0.
1006 // So, finally, the only LLVM-native comparisons that don't mention C and V
1007 // are SETEQ and SETNE. They're the only ones we can safely use CMN for in
1008 // the absence of information about op2.
1009 Opcode = AArch64ISD::ADDS;
1010 RHS = RHS.getOperand(1);
1011 } else if (LHS.getOpcode() == ISD::AND && isa<ConstantSDNode>(RHS) &&
1012 cast<ConstantSDNode>(RHS)->getZExtValue() == 0 &&
1013 !isUnsignedIntSetCC(CC)) {
1014 // Similarly, (CMP (and X, Y), 0) can be implemented with a TST
1015 // (a.k.a. ANDS) except that the flags are only guaranteed to work for one
1016 // of the signed comparisons.
1017 Opcode = AArch64ISD::ANDS;
1018 RHS = LHS.getOperand(1);
1019 LHS = LHS.getOperand(0);
1022 return DAG.getNode(Opcode, dl, DAG.getVTList(VT, MVT::i32), LHS, RHS)
1026 static SDValue getAArch64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1027 SDValue &AArch64cc, SelectionDAG &DAG, SDLoc dl) {
1028 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1029 EVT VT = RHS.getValueType();
1030 uint64_t C = RHSC->getZExtValue();
1031 if (!isLegalArithImmed(C)) {
1032 // Constant does not fit, try adjusting it by one?
1038 if ((VT == MVT::i32 && C != 0x80000000 &&
1039 isLegalArithImmed((uint32_t)(C - 1))) ||
1040 (VT == MVT::i64 && C != 0x80000000ULL &&
1041 isLegalArithImmed(C - 1ULL))) {
1042 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1043 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1044 RHS = DAG.getConstant(C, VT);
1049 if ((VT == MVT::i32 && C != 0 &&
1050 isLegalArithImmed((uint32_t)(C - 1))) ||
1051 (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) {
1052 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1053 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1054 RHS = DAG.getConstant(C, VT);
1059 if ((VT == MVT::i32 && C != 0x7fffffff &&
1060 isLegalArithImmed((uint32_t)(C + 1))) ||
1061 (VT == MVT::i64 && C != 0x7ffffffffffffffULL &&
1062 isLegalArithImmed(C + 1ULL))) {
1063 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1064 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1065 RHS = DAG.getConstant(C, VT);
1070 if ((VT == MVT::i32 && C != 0xffffffff &&
1071 isLegalArithImmed((uint32_t)(C + 1))) ||
1072 (VT == MVT::i64 && C != 0xfffffffffffffffULL &&
1073 isLegalArithImmed(C + 1ULL))) {
1074 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1075 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1076 RHS = DAG.getConstant(C, VT);
1083 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
1084 AArch64CC::CondCode AArch64CC = changeIntCCToAArch64CC(CC);
1085 AArch64cc = DAG.getConstant(AArch64CC, MVT::i32);
1089 static std::pair<SDValue, SDValue>
1090 getAArch64XALUOOp(AArch64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
1091 assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
1092 "Unsupported value type");
1093 SDValue Value, Overflow;
1095 SDValue LHS = Op.getOperand(0);
1096 SDValue RHS = Op.getOperand(1);
1098 switch (Op.getOpcode()) {
1100 llvm_unreachable("Unknown overflow instruction!");
1102 Opc = AArch64ISD::ADDS;
1106 Opc = AArch64ISD::ADDS;
1110 Opc = AArch64ISD::SUBS;
1114 Opc = AArch64ISD::SUBS;
1117 // Multiply needs a little bit extra work.
1121 bool IsSigned = (Op.getOpcode() == ISD::SMULO) ? true : false;
1122 if (Op.getValueType() == MVT::i32) {
1123 unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1124 // For a 32 bit multiply with overflow check we want the instruction
1125 // selector to generate a widening multiply (SMADDL/UMADDL). For that we
1126 // need to generate the following pattern:
1127 // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b))
1128 LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
1129 RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
1130 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1131 SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul,
1132 DAG.getConstant(0, MVT::i64));
1133 // On AArch64 the upper 32 bits are always zero extended for a 32 bit
1134 // operation. We need to clear out the upper 32 bits, because we used a
1135 // widening multiply that wrote all 64 bits. In the end this should be a
1137 Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add);
1139 // The signed overflow check requires more than just a simple check for
1140 // any bit set in the upper 32 bits of the result. These bits could be
1141 // just the sign bits of a negative number. To perform the overflow
1142 // check we have to arithmetic shift right the 32nd bit of the result by
1143 // 31 bits. Then we compare the result to the upper 32 bits.
1144 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add,
1145 DAG.getConstant(32, MVT::i64));
1146 UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits);
1147 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value,
1148 DAG.getConstant(31, MVT::i64));
1149 // It is important that LowerBits is last, otherwise the arithmetic
1150 // shift will not be folded into the compare (SUBS).
1151 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32);
1152 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1155 // The overflow check for unsigned multiply is easy. We only need to
1156 // check if any of the upper 32 bits are set. This can be done with a
1157 // CMP (shifted register). For that we need to generate the following
1159 // (i64 AArch64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32)
1160 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
1161 DAG.getConstant(32, MVT::i64));
1162 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1164 DAG.getNode(AArch64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1165 UpperBits).getValue(1);
1169 assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
1170 // For the 64 bit multiply
1171 Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1173 SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
1174 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
1175 DAG.getConstant(63, MVT::i64));
1176 // It is important that LowerBits is last, otherwise the arithmetic
1177 // shift will not be folded into the compare (SUBS).
1178 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1179 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1182 SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
1183 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1185 DAG.getNode(AArch64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1186 UpperBits).getValue(1);
1193 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);
1195 // Emit the AArch64 operation with overflow check.
1196 Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
1197 Overflow = Value.getValue(1);
1199 return std::make_pair(Value, Overflow);
1202 SDValue AArch64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG,
1203 RTLIB::Libcall Call) const {
1204 SmallVector<SDValue, 2> Ops;
1205 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1206 Ops.push_back(Op.getOperand(i));
1208 return makeLibCall(DAG, Call, MVT::f128, &Ops[0], Ops.size(), false,
1212 static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) {
1213 SDValue Sel = Op.getOperand(0);
1214 SDValue Other = Op.getOperand(1);
1216 // If neither operand is a SELECT_CC, give up.
1217 if (Sel.getOpcode() != ISD::SELECT_CC)
1218 std::swap(Sel, Other);
1219 if (Sel.getOpcode() != ISD::SELECT_CC)
1222 // The folding we want to perform is:
1223 // (xor x, (select_cc a, b, cc, 0, -1) )
1225 // (csel x, (xor x, -1), cc ...)
1227 // The latter will get matched to a CSINV instruction.
1229 ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
1230 SDValue LHS = Sel.getOperand(0);
1231 SDValue RHS = Sel.getOperand(1);
1232 SDValue TVal = Sel.getOperand(2);
1233 SDValue FVal = Sel.getOperand(3);
1236 // FIXME: This could be generalized to non-integer comparisons.
1237 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
1240 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
1241 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
1243 // The the values aren't constants, this isn't the pattern we're looking for.
1244 if (!CFVal || !CTVal)
1247 // We can commute the SELECT_CC by inverting the condition. This
1248 // might be needed to make this fit into a CSINV pattern.
1249 if (CTVal->isAllOnesValue() && CFVal->isNullValue()) {
1250 std::swap(TVal, FVal);
1251 std::swap(CTVal, CFVal);
1252 CC = ISD::getSetCCInverse(CC, true);
1255 // If the constants line up, perform the transform!
1256 if (CTVal->isNullValue() && CFVal->isAllOnesValue()) {
1258 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
1261 TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other,
1262 DAG.getConstant(-1ULL, Other.getValueType()));
1264 return DAG.getNode(AArch64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal,
1271 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
1272 EVT VT = Op.getValueType();
1274 // Let legalize expand this if it isn't a legal type yet.
1275 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
1278 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
1281 bool ExtraOp = false;
1282 switch (Op.getOpcode()) {
1284 llvm_unreachable("Invalid code");
1286 Opc = AArch64ISD::ADDS;
1289 Opc = AArch64ISD::SUBS;
1292 Opc = AArch64ISD::ADCS;
1296 Opc = AArch64ISD::SBCS;
1302 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1));
1303 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1),
1307 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
1308 // Let legalize expand this if it isn't a legal type yet.
1309 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
1312 AArch64CC::CondCode CC;
1313 // The actual operation that sets the overflow or carry flag.
1314 SDValue Value, Overflow;
1315 std::tie(Value, Overflow) = getAArch64XALUOOp(CC, Op, DAG);
1317 // We use 0 and 1 as false and true values.
1318 SDValue TVal = DAG.getConstant(1, MVT::i32);
1319 SDValue FVal = DAG.getConstant(0, MVT::i32);
1321 // We use an inverted condition, because the conditional select is inverted
1322 // too. This will allow it to be selected to a single instruction:
1323 // CSINC Wd, WZR, WZR, invert(cond).
1324 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), MVT::i32);
1325 Overflow = DAG.getNode(AArch64ISD::CSEL, SDLoc(Op), MVT::i32, FVal, TVal,
1328 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
1329 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
1332 // Prefetch operands are:
1333 // 1: Address to prefetch
1335 // 3: int locality (0 = no locality ... 3 = extreme locality)
1336 // 4: bool isDataCache
1337 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
1339 unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
1340 unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
1341 unsigned IsData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
1343 bool IsStream = !Locality;
1344 // When the locality number is set
1346 // The front-end should have filtered out the out-of-range values
1347 assert(Locality <= 3 && "Prefetch locality out-of-range");
1348 // The locality degree is the opposite of the cache speed.
1349 // Put the number the other way around.
1350 // The encoding starts at 0 for level 1
1351 Locality = 3 - Locality;
1354 // built the mask value encoding the expected behavior.
1355 unsigned PrfOp = (IsWrite << 4) | // Load/Store bit
1356 (!IsData << 3) | // IsDataCache bit
1357 (Locality << 1) | // Cache level bits
1358 (unsigned)IsStream; // Stream bit
1359 return DAG.getNode(AArch64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
1360 DAG.getConstant(PrfOp, MVT::i32), Op.getOperand(1));
1363 SDValue AArch64TargetLowering::LowerFP_EXTEND(SDValue Op,
1364 SelectionDAG &DAG) const {
1365 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
1368 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
1370 return LowerF128Call(Op, DAG, LC);
1373 SDValue AArch64TargetLowering::LowerFP_ROUND(SDValue Op,
1374 SelectionDAG &DAG) const {
1375 if (Op.getOperand(0).getValueType() != MVT::f128) {
1376 // It's legal except when f128 is involved
1381 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
1383 // FP_ROUND node has a second operand indicating whether it is known to be
1384 // precise. That doesn't take part in the LibCall so we can't directly use
1386 SDValue SrcVal = Op.getOperand(0);
1387 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
1388 /*isSigned*/ false, SDLoc(Op)).first;
1391 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1392 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1393 // Any additional optimization in this function should be recorded
1394 // in the cost tables.
1395 EVT InVT = Op.getOperand(0).getValueType();
1396 EVT VT = Op.getValueType();
1398 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1401 DAG.getNode(Op.getOpcode(), dl, InVT.changeVectorElementTypeToInteger(),
1403 return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
1406 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1408 SDValue Ext = DAG.getNode(ISD::FP_EXTEND, dl, MVT::v2f64, Op.getOperand(0));
1409 return DAG.getNode(Op.getOpcode(), dl, VT, Ext);
1412 // Type changing conversions are illegal.
1416 SDValue AArch64TargetLowering::LowerFP_TO_INT(SDValue Op,
1417 SelectionDAG &DAG) const {
1418 if (Op.getOperand(0).getValueType().isVector())
1419 return LowerVectorFP_TO_INT(Op, DAG);
1421 if (Op.getOperand(0).getValueType() != MVT::f128) {
1422 // It's legal except when f128 is involved
1427 if (Op.getOpcode() == ISD::FP_TO_SINT)
1428 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
1430 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
1432 SmallVector<SDValue, 2> Ops;
1433 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1434 Ops.push_back(Op.getOperand(i));
1436 return makeLibCall(DAG, LC, Op.getValueType(), &Ops[0], Ops.size(), false,
1440 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1441 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
1442 // Any additional optimization in this function should be recorded
1443 // in the cost tables.
1444 EVT VT = Op.getValueType();
1446 SDValue In = Op.getOperand(0);
1447 EVT InVT = In.getValueType();
1449 if (VT.getSizeInBits() < InVT.getSizeInBits()) {
1451 MVT::getVectorVT(MVT::getFloatingPointVT(InVT.getScalarSizeInBits()),
1452 InVT.getVectorNumElements());
1453 In = DAG.getNode(Op.getOpcode(), dl, CastVT, In);
1454 return DAG.getNode(ISD::FP_ROUND, dl, VT, In, DAG.getIntPtrConstant(0));
1457 if (VT.getSizeInBits() > InVT.getSizeInBits()) {
1459 Op.getOpcode() == ISD::SINT_TO_FP ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1460 EVT CastVT = VT.changeVectorElementTypeToInteger();
1461 In = DAG.getNode(CastOpc, dl, CastVT, In);
1462 return DAG.getNode(Op.getOpcode(), dl, VT, In);
1468 SDValue AArch64TargetLowering::LowerINT_TO_FP(SDValue Op,
1469 SelectionDAG &DAG) const {
1470 if (Op.getValueType().isVector())
1471 return LowerVectorINT_TO_FP(Op, DAG);
1473 // i128 conversions are libcalls.
1474 if (Op.getOperand(0).getValueType() == MVT::i128)
1477 // Other conversions are legal, unless it's to the completely software-based
1479 if (Op.getValueType() != MVT::f128)
1483 if (Op.getOpcode() == ISD::SINT_TO_FP)
1484 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1486 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1488 return LowerF128Call(Op, DAG, LC);
1491 SDValue AArch64TargetLowering::LowerFSINCOS(SDValue Op,
1492 SelectionDAG &DAG) const {
1493 // For iOS, we want to call an alternative entry point: __sincos_stret,
1494 // which returns the values in two S / D registers.
1496 SDValue Arg = Op.getOperand(0);
1497 EVT ArgVT = Arg.getValueType();
1498 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1505 Entry.isSExt = false;
1506 Entry.isZExt = false;
1507 Args.push_back(Entry);
1509 const char *LibcallName =
1510 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
1511 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
1513 StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
1514 TargetLowering::CallLoweringInfo CLI(DAG);
1515 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
1516 .setCallee(CallingConv::Fast, RetTy, Callee, std::move(Args), 0);
1518 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1519 return CallResult.first;
1522 static SDValue LowerBITCAST(SDValue Op, SelectionDAG &DAG) {
1523 if (Op.getValueType() != MVT::f16)
1526 assert(Op.getOperand(0).getValueType() == MVT::i16);
1529 Op = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op.getOperand(0));
1530 Op = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Op);
1532 DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL, MVT::f16, Op,
1533 DAG.getTargetConstant(AArch64::hsub, MVT::i32)),
1538 SDValue AArch64TargetLowering::LowerOperation(SDValue Op,
1539 SelectionDAG &DAG) const {
1540 switch (Op.getOpcode()) {
1542 llvm_unreachable("unimplemented operand");
1545 return LowerBITCAST(Op, DAG);
1546 case ISD::GlobalAddress:
1547 return LowerGlobalAddress(Op, DAG);
1548 case ISD::GlobalTLSAddress:
1549 return LowerGlobalTLSAddress(Op, DAG);
1551 return LowerSETCC(Op, DAG);
1553 return LowerBR_CC(Op, DAG);
1555 return LowerSELECT(Op, DAG);
1556 case ISD::SELECT_CC:
1557 return LowerSELECT_CC(Op, DAG);
1558 case ISD::JumpTable:
1559 return LowerJumpTable(Op, DAG);
1560 case ISD::ConstantPool:
1561 return LowerConstantPool(Op, DAG);
1562 case ISD::BlockAddress:
1563 return LowerBlockAddress(Op, DAG);
1565 return LowerVASTART(Op, DAG);
1567 return LowerVACOPY(Op, DAG);
1569 return LowerVAARG(Op, DAG);
1574 return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
1581 return LowerXALUO(Op, DAG);
1583 return LowerF128Call(Op, DAG, RTLIB::ADD_F128);
1585 return LowerF128Call(Op, DAG, RTLIB::SUB_F128);
1587 return LowerF128Call(Op, DAG, RTLIB::MUL_F128);
1589 return LowerF128Call(Op, DAG, RTLIB::DIV_F128);
1591 return LowerFP_ROUND(Op, DAG);
1592 case ISD::FP_EXTEND:
1593 return LowerFP_EXTEND(Op, DAG);
1594 case ISD::FRAMEADDR:
1595 return LowerFRAMEADDR(Op, DAG);
1596 case ISD::RETURNADDR:
1597 return LowerRETURNADDR(Op, DAG);
1598 case ISD::INSERT_VECTOR_ELT:
1599 return LowerINSERT_VECTOR_ELT(Op, DAG);
1600 case ISD::EXTRACT_VECTOR_ELT:
1601 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
1602 case ISD::BUILD_VECTOR:
1603 return LowerBUILD_VECTOR(Op, DAG);
1604 case ISD::VECTOR_SHUFFLE:
1605 return LowerVECTOR_SHUFFLE(Op, DAG);
1606 case ISD::EXTRACT_SUBVECTOR:
1607 return LowerEXTRACT_SUBVECTOR(Op, DAG);
1611 return LowerVectorSRA_SRL_SHL(Op, DAG);
1612 case ISD::SHL_PARTS:
1613 return LowerShiftLeftParts(Op, DAG);
1614 case ISD::SRL_PARTS:
1615 case ISD::SRA_PARTS:
1616 return LowerShiftRightParts(Op, DAG);
1618 return LowerCTPOP(Op, DAG);
1619 case ISD::FCOPYSIGN:
1620 return LowerFCOPYSIGN(Op, DAG);
1622 return LowerVectorAND(Op, DAG);
1624 return LowerVectorOR(Op, DAG);
1626 return LowerXOR(Op, DAG);
1628 return LowerPREFETCH(Op, DAG);
1629 case ISD::SINT_TO_FP:
1630 case ISD::UINT_TO_FP:
1631 return LowerINT_TO_FP(Op, DAG);
1632 case ISD::FP_TO_SINT:
1633 case ISD::FP_TO_UINT:
1634 return LowerFP_TO_INT(Op, DAG);
1636 return LowerFSINCOS(Op, DAG);
1640 /// getFunctionAlignment - Return the Log2 alignment of this function.
1641 unsigned AArch64TargetLowering::getFunctionAlignment(const Function *F) const {
1645 //===----------------------------------------------------------------------===//
1646 // Calling Convention Implementation
1647 //===----------------------------------------------------------------------===//
1649 #include "AArch64GenCallingConv.inc"
1651 /// Selects the correct CCAssignFn for a the given CallingConvention
1653 CCAssignFn *AArch64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
1654 bool IsVarArg) const {
1657 llvm_unreachable("Unsupported calling convention.");
1658 case CallingConv::WebKit_JS:
1659 return CC_AArch64_WebKit_JS;
1660 case CallingConv::C:
1661 case CallingConv::Fast:
1662 if (!Subtarget->isTargetDarwin())
1663 return CC_AArch64_AAPCS;
1664 return IsVarArg ? CC_AArch64_DarwinPCS_VarArg : CC_AArch64_DarwinPCS;
1668 SDValue AArch64TargetLowering::LowerFormalArguments(
1669 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1670 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
1671 SmallVectorImpl<SDValue> &InVals) const {
1672 MachineFunction &MF = DAG.getMachineFunction();
1673 MachineFrameInfo *MFI = MF.getFrameInfo();
1675 // Assign locations to all of the incoming arguments.
1676 SmallVector<CCValAssign, 16> ArgLocs;
1677 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1680 // At this point, Ins[].VT may already be promoted to i32. To correctly
1681 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
1682 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
1683 // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
1684 // we use a special version of AnalyzeFormalArguments to pass in ValVT and
1686 unsigned NumArgs = Ins.size();
1687 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
1688 unsigned CurArgIdx = 0;
1689 for (unsigned i = 0; i != NumArgs; ++i) {
1690 MVT ValVT = Ins[i].VT;
1691 std::advance(CurOrigArg, Ins[i].OrigArgIndex - CurArgIdx);
1692 CurArgIdx = Ins[i].OrigArgIndex;
1694 // Get type of the original argument.
1695 EVT ActualVT = getValueType(CurOrigArg->getType(), /*AllowUnknown*/ true);
1696 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
1697 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
1698 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
1700 else if (ActualMVT == MVT::i16)
1703 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
1705 AssignFn(i, ValVT, ValVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
1706 assert(!Res && "Call operand has unhandled type");
1709 assert(ArgLocs.size() == Ins.size());
1710 SmallVector<SDValue, 16> ArgValues;
1711 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1712 CCValAssign &VA = ArgLocs[i];
1714 if (Ins[i].Flags.isByVal()) {
1715 // Byval is used for HFAs in the PCS, but the system should work in a
1716 // non-compliant manner for larger structs.
1717 EVT PtrTy = getPointerTy();
1718 int Size = Ins[i].Flags.getByValSize();
1719 unsigned NumRegs = (Size + 7) / 8;
1721 // FIXME: This works on big-endian for composite byvals, which are the common
1722 // case. It should also work for fundamental types too.
1724 MFI->CreateFixedObject(8 * NumRegs, VA.getLocMemOffset(), false);
1725 SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrTy);
1726 InVals.push_back(FrameIdxN);
1731 if (VA.isRegLoc()) {
1732 // Arguments stored in registers.
1733 EVT RegVT = VA.getLocVT();
1736 const TargetRegisterClass *RC;
1738 if (RegVT == MVT::i32)
1739 RC = &AArch64::GPR32RegClass;
1740 else if (RegVT == MVT::i64)
1741 RC = &AArch64::GPR64RegClass;
1742 else if (RegVT == MVT::f16)
1743 RC = &AArch64::FPR16RegClass;
1744 else if (RegVT == MVT::f32)
1745 RC = &AArch64::FPR32RegClass;
1746 else if (RegVT == MVT::f64 || RegVT.is64BitVector())
1747 RC = &AArch64::FPR64RegClass;
1748 else if (RegVT == MVT::f128 || RegVT.is128BitVector())
1749 RC = &AArch64::FPR128RegClass;
1751 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1753 // Transform the arguments in physical registers into virtual ones.
1754 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1755 ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
1757 // If this is an 8, 16 or 32-bit value, it is really passed promoted
1758 // to 64 bits. Insert an assert[sz]ext to capture this, then
1759 // truncate to the right size.
1760 switch (VA.getLocInfo()) {
1762 llvm_unreachable("Unknown loc info!");
1763 case CCValAssign::Full:
1765 case CCValAssign::BCvt:
1766 ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue);
1768 case CCValAssign::AExt:
1769 case CCValAssign::SExt:
1770 case CCValAssign::ZExt:
1771 // SelectionDAGBuilder will insert appropriate AssertZExt & AssertSExt
1772 // nodes after our lowering.
1773 assert(RegVT == Ins[i].VT && "incorrect register location selected");
1777 InVals.push_back(ArgValue);
1779 } else { // VA.isRegLoc()
1780 assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
1781 unsigned ArgOffset = VA.getLocMemOffset();
1782 unsigned ArgSize = VA.getValVT().getSizeInBits() / 8;
1784 uint32_t BEAlign = 0;
1785 if (ArgSize < 8 && !Subtarget->isLittleEndian())
1786 BEAlign = 8 - ArgSize;
1788 int FI = MFI->CreateFixedObject(ArgSize, ArgOffset + BEAlign, true);
1790 // Create load nodes to retrieve arguments from the stack.
1791 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1794 // For NON_EXTLOAD, generic code in getLoad assert(ValVT == MemVT)
1795 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
1796 MVT MemVT = VA.getValVT();
1798 switch (VA.getLocInfo()) {
1801 case CCValAssign::BCvt:
1802 MemVT = VA.getLocVT();
1804 case CCValAssign::SExt:
1805 ExtType = ISD::SEXTLOAD;
1807 case CCValAssign::ZExt:
1808 ExtType = ISD::ZEXTLOAD;
1810 case CCValAssign::AExt:
1811 ExtType = ISD::EXTLOAD;
1815 ArgValue = DAG.getExtLoad(ExtType, DL, VA.getLocVT(), Chain, FIN,
1816 MachinePointerInfo::getFixedStack(FI),
1817 MemVT, false, false, false, 0, nullptr);
1819 InVals.push_back(ArgValue);
1825 if (!Subtarget->isTargetDarwin()) {
1826 // The AAPCS variadic function ABI is identical to the non-variadic
1827 // one. As a result there may be more arguments in registers and we should
1828 // save them for future reference.
1829 saveVarArgRegisters(CCInfo, DAG, DL, Chain);
1832 AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>();
1833 // This will point to the next argument passed via stack.
1834 unsigned StackOffset = CCInfo.getNextStackOffset();
1835 // We currently pass all varargs at 8-byte alignment.
1836 StackOffset = ((StackOffset + 7) & ~7);
1837 AFI->setVarArgsStackIndex(MFI->CreateFixedObject(4, StackOffset, true));
1840 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
1841 unsigned StackArgSize = CCInfo.getNextStackOffset();
1842 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
1843 if (DoesCalleeRestoreStack(CallConv, TailCallOpt)) {
1844 // This is a non-standard ABI so by fiat I say we're allowed to make full
1845 // use of the stack area to be popped, which must be aligned to 16 bytes in
1847 StackArgSize = RoundUpToAlignment(StackArgSize, 16);
1849 // If we're expected to restore the stack (e.g. fastcc) then we'll be adding
1850 // a multiple of 16.
1851 FuncInfo->setArgumentStackToRestore(StackArgSize);
1853 // This realignment carries over to the available bytes below. Our own
1854 // callers will guarantee the space is free by giving an aligned value to
1857 // Even if we're not expected to free up the space, it's useful to know how
1858 // much is there while considering tail calls (because we can reuse it).
1859 FuncInfo->setBytesInStackArgArea(StackArgSize);
1864 void AArch64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
1865 SelectionDAG &DAG, SDLoc DL,
1866 SDValue &Chain) const {
1867 MachineFunction &MF = DAG.getMachineFunction();
1868 MachineFrameInfo *MFI = MF.getFrameInfo();
1869 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
1871 SmallVector<SDValue, 8> MemOps;
1873 static const MCPhysReg GPRArgRegs[] = { AArch64::X0, AArch64::X1, AArch64::X2,
1874 AArch64::X3, AArch64::X4, AArch64::X5,
1875 AArch64::X6, AArch64::X7 };
1876 static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs);
1877 unsigned FirstVariadicGPR =
1878 CCInfo.getFirstUnallocated(GPRArgRegs, NumGPRArgRegs);
1880 unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
1882 if (GPRSaveSize != 0) {
1883 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
1885 SDValue FIN = DAG.getFrameIndex(GPRIdx, getPointerTy());
1887 for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
1888 unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &AArch64::GPR64RegClass);
1889 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
1891 DAG.getStore(Val.getValue(1), DL, Val, FIN,
1892 MachinePointerInfo::getStack(i * 8), false, false, 0);
1893 MemOps.push_back(Store);
1894 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
1895 DAG.getConstant(8, getPointerTy()));
1898 FuncInfo->setVarArgsGPRIndex(GPRIdx);
1899 FuncInfo->setVarArgsGPRSize(GPRSaveSize);
1901 if (Subtarget->hasFPARMv8()) {
1902 static const MCPhysReg FPRArgRegs[] = {
1903 AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3,
1904 AArch64::Q4, AArch64::Q5, AArch64::Q6, AArch64::Q7};
1905 static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs);
1906 unsigned FirstVariadicFPR =
1907 CCInfo.getFirstUnallocated(FPRArgRegs, NumFPRArgRegs);
1909 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
1911 if (FPRSaveSize != 0) {
1912 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
1914 SDValue FIN = DAG.getFrameIndex(FPRIdx, getPointerTy());
1916 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
1917 unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &AArch64::FPR128RegClass);
1918 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f128);
1921 DAG.getStore(Val.getValue(1), DL, Val, FIN,
1922 MachinePointerInfo::getStack(i * 16), false, false, 0);
1923 MemOps.push_back(Store);
1924 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
1925 DAG.getConstant(16, getPointerTy()));
1928 FuncInfo->setVarArgsFPRIndex(FPRIdx);
1929 FuncInfo->setVarArgsFPRSize(FPRSaveSize);
1932 if (!MemOps.empty()) {
1933 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
1937 /// LowerCallResult - Lower the result values of a call into the
1938 /// appropriate copies out of appropriate physical registers.
1939 SDValue AArch64TargetLowering::LowerCallResult(
1940 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
1941 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
1942 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
1943 SDValue ThisVal) const {
1944 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
1945 ? RetCC_AArch64_WebKit_JS
1946 : RetCC_AArch64_AAPCS;
1947 // Assign locations to each value returned by this call.
1948 SmallVector<CCValAssign, 16> RVLocs;
1949 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1951 CCInfo.AnalyzeCallResult(Ins, RetCC);
1953 // Copy all of the result registers out of their specified physreg.
1954 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1955 CCValAssign VA = RVLocs[i];
1957 // Pass 'this' value directly from the argument to return value, to avoid
1958 // reg unit interference
1959 if (i == 0 && isThisReturn) {
1960 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 &&
1961 "unexpected return calling convention register assignment");
1962 InVals.push_back(ThisVal);
1967 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag);
1968 Chain = Val.getValue(1);
1969 InFlag = Val.getValue(2);
1971 switch (VA.getLocInfo()) {
1973 llvm_unreachable("Unknown loc info!");
1974 case CCValAssign::Full:
1976 case CCValAssign::BCvt:
1977 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
1981 InVals.push_back(Val);
1987 bool AArch64TargetLowering::isEligibleForTailCallOptimization(
1988 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
1989 bool isCalleeStructRet, bool isCallerStructRet,
1990 const SmallVectorImpl<ISD::OutputArg> &Outs,
1991 const SmallVectorImpl<SDValue> &OutVals,
1992 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
1993 // For CallingConv::C this function knows whether the ABI needs
1994 // changing. That's not true for other conventions so they will have to opt in
1996 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
1999 const MachineFunction &MF = DAG.getMachineFunction();
2000 const Function *CallerF = MF.getFunction();
2001 CallingConv::ID CallerCC = CallerF->getCallingConv();
2002 bool CCMatch = CallerCC == CalleeCC;
2004 // Byval parameters hand the function a pointer directly into the stack area
2005 // we want to reuse during a tail call. Working around this *is* possible (see
2006 // X86) but less efficient and uglier in LowerCall.
2007 for (Function::const_arg_iterator i = CallerF->arg_begin(),
2008 e = CallerF->arg_end();
2010 if (i->hasByValAttr())
2013 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
2014 if (IsTailCallConvention(CalleeCC) && CCMatch)
2019 // Externally-defined functions with weak linkage should not be
2020 // tail-called on AArch64 when the OS does not support dynamic
2021 // pre-emption of symbols, as the AAELF spec requires normal calls
2022 // to undefined weak functions to be replaced with a NOP or jump to the
2023 // next instruction. The behaviour of branch instructions in this
2024 // situation (as used for tail calls) is implementation-defined, so we
2025 // cannot rely on the linker replacing the tail call with a return.
2026 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2027 const GlobalValue *GV = G->getGlobal();
2028 if (GV->hasExternalWeakLinkage())
2032 // Now we search for cases where we can use a tail call without changing the
2033 // ABI. Sibcall is used in some places (particularly gcc) to refer to this
2036 // I want anyone implementing a new calling convention to think long and hard
2037 // about this assert.
2038 assert((!isVarArg || CalleeCC == CallingConv::C) &&
2039 "Unexpected variadic calling convention");
2041 if (isVarArg && !Outs.empty()) {
2042 // At least two cases here: if caller is fastcc then we can't have any
2043 // memory arguments (we'd be expected to clean up the stack afterwards). If
2044 // caller is C then we could potentially use its argument area.
2046 // FIXME: for now we take the most conservative of these in both cases:
2047 // disallow all variadic memory operands.
2048 SmallVector<CCValAssign, 16> ArgLocs;
2049 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2052 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, true));
2053 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
2054 if (!ArgLocs[i].isRegLoc())
2058 // If the calling conventions do not match, then we'd better make sure the
2059 // results are returned in the same way as what the caller expects.
2061 SmallVector<CCValAssign, 16> RVLocs1;
2062 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2064 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForCall(CalleeCC, isVarArg));
2066 SmallVector<CCValAssign, 16> RVLocs2;
2067 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2069 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForCall(CallerCC, isVarArg));
2071 if (RVLocs1.size() != RVLocs2.size())
2073 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2074 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2076 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2078 if (RVLocs1[i].isRegLoc()) {
2079 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2082 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2088 // Nothing more to check if the callee is taking no arguments
2092 SmallVector<CCValAssign, 16> ArgLocs;
2093 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2096 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, isVarArg));
2098 const AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2100 // If the stack arguments for this call would fit into our own save area then
2101 // the call can be made tail.
2102 return CCInfo.getNextStackOffset() <= FuncInfo->getBytesInStackArgArea();
2105 SDValue AArch64TargetLowering::addTokenForArgument(SDValue Chain,
2107 MachineFrameInfo *MFI,
2108 int ClobberedFI) const {
2109 SmallVector<SDValue, 8> ArgChains;
2110 int64_t FirstByte = MFI->getObjectOffset(ClobberedFI);
2111 int64_t LastByte = FirstByte + MFI->getObjectSize(ClobberedFI) - 1;
2113 // Include the original chain at the beginning of the list. When this is
2114 // used by target LowerCall hooks, this helps legalize find the
2115 // CALLSEQ_BEGIN node.
2116 ArgChains.push_back(Chain);
2118 // Add a chain value for each stack argument corresponding
2119 for (SDNode::use_iterator U = DAG.getEntryNode().getNode()->use_begin(),
2120 UE = DAG.getEntryNode().getNode()->use_end();
2122 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
2123 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
2124 if (FI->getIndex() < 0) {
2125 int64_t InFirstByte = MFI->getObjectOffset(FI->getIndex());
2126 int64_t InLastByte = InFirstByte;
2127 InLastByte += MFI->getObjectSize(FI->getIndex()) - 1;
2129 if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
2130 (FirstByte <= InFirstByte && InFirstByte <= LastByte))
2131 ArgChains.push_back(SDValue(L, 1));
2134 // Build a tokenfactor for all the chains.
2135 return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
2138 bool AArch64TargetLowering::DoesCalleeRestoreStack(CallingConv::ID CallCC,
2139 bool TailCallOpt) const {
2140 return CallCC == CallingConv::Fast && TailCallOpt;
2143 bool AArch64TargetLowering::IsTailCallConvention(CallingConv::ID CallCC) const {
2144 return CallCC == CallingConv::Fast;
2147 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
2148 /// and add input and output parameter nodes.
2150 AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
2151 SmallVectorImpl<SDValue> &InVals) const {
2152 SelectionDAG &DAG = CLI.DAG;
2154 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2155 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2156 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2157 SDValue Chain = CLI.Chain;
2158 SDValue Callee = CLI.Callee;
2159 bool &IsTailCall = CLI.IsTailCall;
2160 CallingConv::ID CallConv = CLI.CallConv;
2161 bool IsVarArg = CLI.IsVarArg;
2163 MachineFunction &MF = DAG.getMachineFunction();
2164 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
2165 bool IsThisReturn = false;
2167 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
2168 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2169 bool IsSibCall = false;
2172 // Check if it's really possible to do a tail call.
2173 IsTailCall = isEligibleForTailCallOptimization(
2174 Callee, CallConv, IsVarArg, IsStructRet,
2175 MF.getFunction()->hasStructRetAttr(), Outs, OutVals, Ins, DAG);
2176 if (!IsTailCall && CLI.CS && CLI.CS->isMustTailCall())
2177 report_fatal_error("failed to perform tail call elimination on a call "
2178 "site marked musttail");
2180 // A sibling call is one where we're under the usual C ABI and not planning
2181 // to change that but can still do a tail call:
2182 if (!TailCallOpt && IsTailCall)
2189 // Analyze operands of the call, assigning locations to each operand.
2190 SmallVector<CCValAssign, 16> ArgLocs;
2191 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
2195 // Handle fixed and variable vector arguments differently.
2196 // Variable vector arguments always go into memory.
2197 unsigned NumArgs = Outs.size();
2199 for (unsigned i = 0; i != NumArgs; ++i) {
2200 MVT ArgVT = Outs[i].VT;
2201 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2202 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
2203 /*IsVarArg=*/ !Outs[i].IsFixed);
2204 bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
2205 assert(!Res && "Call operand has unhandled type");
2209 // At this point, Outs[].VT may already be promoted to i32. To correctly
2210 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2211 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
2212 // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here
2213 // we use a special version of AnalyzeCallOperands to pass in ValVT and
2215 unsigned NumArgs = Outs.size();
2216 for (unsigned i = 0; i != NumArgs; ++i) {
2217 MVT ValVT = Outs[i].VT;
2218 // Get type of the original argument.
2219 EVT ActualVT = getValueType(CLI.getArgs()[Outs[i].OrigArgIndex].Ty,
2220 /*AllowUnknown*/ true);
2221 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT;
2222 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2223 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2224 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2226 else if (ActualMVT == MVT::i16)
2229 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2230 bool Res = AssignFn(i, ValVT, ValVT, CCValAssign::Full, ArgFlags, CCInfo);
2231 assert(!Res && "Call operand has unhandled type");
2236 // Get a count of how many bytes are to be pushed on the stack.
2237 unsigned NumBytes = CCInfo.getNextStackOffset();
2240 // Since we're not changing the ABI to make this a tail call, the memory
2241 // operands are already available in the caller's incoming argument space.
2245 // FPDiff is the byte offset of the call's argument area from the callee's.
2246 // Stores to callee stack arguments will be placed in FixedStackSlots offset
2247 // by this amount for a tail call. In a sibling call it must be 0 because the
2248 // caller will deallocate the entire stack and the callee still expects its
2249 // arguments to begin at SP+0. Completely unused for non-tail calls.
2252 if (IsTailCall && !IsSibCall) {
2253 unsigned NumReusableBytes = FuncInfo->getBytesInStackArgArea();
2255 // Since callee will pop argument stack as a tail call, we must keep the
2256 // popped size 16-byte aligned.
2257 NumBytes = RoundUpToAlignment(NumBytes, 16);
2259 // FPDiff will be negative if this tail call requires more space than we
2260 // would automatically have in our incoming argument space. Positive if we
2261 // can actually shrink the stack.
2262 FPDiff = NumReusableBytes - NumBytes;
2264 // The stack pointer must be 16-byte aligned at all times it's used for a
2265 // memory operation, which in practice means at *all* times and in
2266 // particular across call boundaries. Therefore our own arguments started at
2267 // a 16-byte aligned SP and the delta applied for the tail call should
2268 // satisfy the same constraint.
2269 assert(FPDiff % 16 == 0 && "unaligned stack on tail call");
2272 // Adjust the stack pointer for the new arguments...
2273 // These operations are automatically eliminated by the prolog/epilog pass
2276 DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true), DL);
2278 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, AArch64::SP, getPointerTy());
2280 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2281 SmallVector<SDValue, 8> MemOpChains;
2283 // Walk the register/memloc assignments, inserting copies/loads.
2284 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2285 ++i, ++realArgIdx) {
2286 CCValAssign &VA = ArgLocs[i];
2287 SDValue Arg = OutVals[realArgIdx];
2288 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2290 // Promote the value if needed.
2291 switch (VA.getLocInfo()) {
2293 llvm_unreachable("Unknown loc info!");
2294 case CCValAssign::Full:
2296 case CCValAssign::SExt:
2297 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
2299 case CCValAssign::ZExt:
2300 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2302 case CCValAssign::AExt:
2303 if (Outs[realArgIdx].ArgVT == MVT::i1) {
2304 // AAPCS requires i1 to be zero-extended to 8-bits by the caller.
2305 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
2306 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i8, Arg);
2308 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
2310 case CCValAssign::BCvt:
2311 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2313 case CCValAssign::FPExt:
2314 Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
2318 if (VA.isRegLoc()) {
2319 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i64) {
2320 assert(VA.getLocVT() == MVT::i64 &&
2321 "unexpected calling convention register assignment");
2322 assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
2323 "unexpected use of 'returned'");
2324 IsThisReturn = true;
2326 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2328 assert(VA.isMemLoc());
2331 MachinePointerInfo DstInfo;
2333 // FIXME: This works on big-endian for composite byvals, which are the
2334 // common case. It should also work for fundamental types too.
2335 uint32_t BEAlign = 0;
2336 unsigned OpSize = Flags.isByVal() ? Flags.getByValSize() * 8
2337 : VA.getValVT().getSizeInBits();
2338 OpSize = (OpSize + 7) / 8;
2339 if (!Subtarget->isLittleEndian() && !Flags.isByVal()) {
2341 BEAlign = 8 - OpSize;
2343 unsigned LocMemOffset = VA.getLocMemOffset();
2344 int32_t Offset = LocMemOffset + BEAlign;
2345 SDValue PtrOff = DAG.getIntPtrConstant(Offset);
2346 PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff);
2349 Offset = Offset + FPDiff;
2350 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2352 DstAddr = DAG.getFrameIndex(FI, getPointerTy());
2353 DstInfo = MachinePointerInfo::getFixedStack(FI);
2355 // Make sure any stack arguments overlapping with where we're storing
2356 // are loaded before this eventual operation. Otherwise they'll be
2358 Chain = addTokenForArgument(Chain, DAG, MF.getFrameInfo(), FI);
2360 SDValue PtrOff = DAG.getIntPtrConstant(Offset);
2362 DstAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff);
2363 DstInfo = MachinePointerInfo::getStack(LocMemOffset);
2366 if (Outs[i].Flags.isByVal()) {
2368 DAG.getConstant(Outs[i].Flags.getByValSize(), MVT::i64);
2369 SDValue Cpy = DAG.getMemcpy(
2370 Chain, DL, DstAddr, Arg, SizeNode, Outs[i].Flags.getByValAlign(),
2372 /*AlwaysInline = */ false, DstInfo, MachinePointerInfo());
2374 MemOpChains.push_back(Cpy);
2376 // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
2377 // promoted to a legal register type i32, we should truncate Arg back to
2379 if (VA.getValVT() == MVT::i1 || VA.getValVT() == MVT::i8 ||
2380 VA.getValVT() == MVT::i16)
2381 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Arg);
2384 DAG.getStore(Chain, DL, Arg, DstAddr, DstInfo, false, false, 0);
2385 MemOpChains.push_back(Store);
2390 if (!MemOpChains.empty())
2391 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
2393 // Build a sequence of copy-to-reg nodes chained together with token chain
2394 // and flag operands which copy the outgoing args into the appropriate regs.
2396 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2397 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[i].first,
2398 RegsToPass[i].second, InFlag);
2399 InFlag = Chain.getValue(1);
2402 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2403 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2404 // node so that legalize doesn't hack it.
2405 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
2406 Subtarget->isTargetMachO()) {
2407 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2408 const GlobalValue *GV = G->getGlobal();
2409 bool InternalLinkage = GV->hasInternalLinkage();
2410 if (InternalLinkage)
2411 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2413 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0,
2415 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, getPointerTy(), Callee);
2417 } else if (ExternalSymbolSDNode *S =
2418 dyn_cast<ExternalSymbolSDNode>(Callee)) {
2419 const char *Sym = S->getSymbol();
2421 DAG.getTargetExternalSymbol(Sym, getPointerTy(), AArch64II::MO_GOT);
2422 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, getPointerTy(), Callee);
2424 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2425 const GlobalValue *GV = G->getGlobal();
2426 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2427 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2428 const char *Sym = S->getSymbol();
2429 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), 0);
2432 // We don't usually want to end the call-sequence here because we would tidy
2433 // the frame up *after* the call, however in the ABI-changing tail-call case
2434 // we've carefully laid out the parameters so that when sp is reset they'll be
2435 // in the correct location.
2436 if (IsTailCall && !IsSibCall) {
2437 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2438 DAG.getIntPtrConstant(0, true), InFlag, DL);
2439 InFlag = Chain.getValue(1);
2442 std::vector<SDValue> Ops;
2443 Ops.push_back(Chain);
2444 Ops.push_back(Callee);
2447 // Each tail call may have to adjust the stack by a different amount, so
2448 // this information must travel along with the operation for eventual
2449 // consumption by emitEpilogue.
2450 Ops.push_back(DAG.getTargetConstant(FPDiff, MVT::i32));
2453 // Add argument registers to the end of the list so that they are known live
2455 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2456 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2457 RegsToPass[i].second.getValueType()));
2459 // Add a register mask operand representing the call-preserved registers.
2460 const uint32_t *Mask;
2461 const TargetRegisterInfo *TRI =
2462 getTargetMachine().getSubtargetImpl()->getRegisterInfo();
2463 const AArch64RegisterInfo *ARI =
2464 static_cast<const AArch64RegisterInfo *>(TRI);
2466 // For 'this' returns, use the X0-preserving mask if applicable
2467 Mask = ARI->getThisReturnPreservedMask(CallConv);
2469 IsThisReturn = false;
2470 Mask = ARI->getCallPreservedMask(CallConv);
2473 Mask = ARI->getCallPreservedMask(CallConv);
2475 assert(Mask && "Missing call preserved mask for calling convention");
2476 Ops.push_back(DAG.getRegisterMask(Mask));
2478 if (InFlag.getNode())
2479 Ops.push_back(InFlag);
2481 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2483 // If we're doing a tall call, use a TC_RETURN here rather than an
2484 // actual call instruction.
2486 return DAG.getNode(AArch64ISD::TC_RETURN, DL, NodeTys, Ops);
2488 // Returns a chain and a flag for retval copy to use.
2489 Chain = DAG.getNode(AArch64ISD::CALL, DL, NodeTys, Ops);
2490 InFlag = Chain.getValue(1);
2492 uint64_t CalleePopBytes = DoesCalleeRestoreStack(CallConv, TailCallOpt)
2493 ? RoundUpToAlignment(NumBytes, 16)
2496 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2497 DAG.getIntPtrConstant(CalleePopBytes, true),
2500 InFlag = Chain.getValue(1);
2502 // Handle result values, copying them out of physregs into vregs that we
2504 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
2505 InVals, IsThisReturn,
2506 IsThisReturn ? OutVals[0] : SDValue());
2509 bool AArch64TargetLowering::CanLowerReturn(
2510 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2511 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2512 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2513 ? RetCC_AArch64_WebKit_JS
2514 : RetCC_AArch64_AAPCS;
2515 SmallVector<CCValAssign, 16> RVLocs;
2516 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2517 return CCInfo.CheckReturn(Outs, RetCC);
2521 AArch64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2523 const SmallVectorImpl<ISD::OutputArg> &Outs,
2524 const SmallVectorImpl<SDValue> &OutVals,
2525 SDLoc DL, SelectionDAG &DAG) const {
2526 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
2527 ? RetCC_AArch64_WebKit_JS
2528 : RetCC_AArch64_AAPCS;
2529 SmallVector<CCValAssign, 16> RVLocs;
2530 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2532 CCInfo.AnalyzeReturn(Outs, RetCC);
2534 // Copy the result values into the output registers.
2536 SmallVector<SDValue, 4> RetOps(1, Chain);
2537 for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size();
2538 ++i, ++realRVLocIdx) {
2539 CCValAssign &VA = RVLocs[i];
2540 assert(VA.isRegLoc() && "Can only return in registers!");
2541 SDValue Arg = OutVals[realRVLocIdx];
2543 switch (VA.getLocInfo()) {
2545 llvm_unreachable("Unknown loc info!");
2546 case CCValAssign::Full:
2547 if (Outs[i].ArgVT == MVT::i1) {
2548 // AAPCS requires i1 to be zero-extended to i8 by the producer of the
2549 // value. This is strictly redundant on Darwin (which uses "zeroext
2550 // i1"), but will be optimised out before ISel.
2551 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
2552 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2555 case CCValAssign::BCvt:
2556 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2560 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
2561 Flag = Chain.getValue(1);
2562 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2565 RetOps[0] = Chain; // Update chain.
2567 // Add the flag if we have it.
2569 RetOps.push_back(Flag);
2571 return DAG.getNode(AArch64ISD::RET_FLAG, DL, MVT::Other, RetOps);
2574 //===----------------------------------------------------------------------===//
2575 // Other Lowering Code
2576 //===----------------------------------------------------------------------===//
2578 SDValue AArch64TargetLowering::LowerGlobalAddress(SDValue Op,
2579 SelectionDAG &DAG) const {
2580 EVT PtrVT = getPointerTy();
2582 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2583 unsigned char OpFlags =
2584 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
2586 assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
2587 "unexpected offset in global node");
2589 // This also catched the large code model case for Darwin.
2590 if ((OpFlags & AArch64II::MO_GOT) != 0) {
2591 SDValue GotAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
2592 // FIXME: Once remat is capable of dealing with instructions with register
2593 // operands, expand this into two nodes instead of using a wrapper node.
2594 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
2597 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2598 const unsigned char MO_NC = AArch64II::MO_NC;
2600 AArch64ISD::WrapperLarge, DL, PtrVT,
2601 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G3),
2602 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
2603 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
2604 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
2606 // Use ADRP/ADD or ADRP/LDR for everything else: the small model on ELF and
2607 // the only correct model on Darwin.
2608 SDValue Hi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2609 OpFlags | AArch64II::MO_PAGE);
2610 unsigned char LoFlags = OpFlags | AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
2611 SDValue Lo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, LoFlags);
2613 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
2614 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
2618 /// \brief Convert a TLS address reference into the correct sequence of loads
2619 /// and calls to compute the variable's address (for Darwin, currently) and
2620 /// return an SDValue containing the final node.
2622 /// Darwin only has one TLS scheme which must be capable of dealing with the
2623 /// fully general situation, in the worst case. This means:
2624 /// + "extern __thread" declaration.
2625 /// + Defined in a possibly unknown dynamic library.
2627 /// The general system is that each __thread variable has a [3 x i64] descriptor
2628 /// which contains information used by the runtime to calculate the address. The
2629 /// only part of this the compiler needs to know about is the first xword, which
2630 /// contains a function pointer that must be called with the address of the
2631 /// entire descriptor in "x0".
2633 /// Since this descriptor may be in a different unit, in general even the
2634 /// descriptor must be accessed via an indirect load. The "ideal" code sequence
2636 /// adrp x0, _var@TLVPPAGE
2637 /// ldr x0, [x0, _var@TLVPPAGEOFF] ; x0 now contains address of descriptor
2638 /// ldr x1, [x0] ; x1 contains 1st entry of descriptor,
2639 /// ; the function pointer
2640 /// blr x1 ; Uses descriptor address in x0
2641 /// ; Address of _var is now in x0.
2643 /// If the address of _var's descriptor *is* known to the linker, then it can
2644 /// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
2645 /// a slight efficiency gain.
2647 AArch64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
2648 SelectionDAG &DAG) const {
2649 assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");
2652 MVT PtrVT = getPointerTy();
2653 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2656 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
2657 SDValue DescAddr = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TLVPAddr);
2659 // The first entry in the descriptor is a function pointer that we must call
2660 // to obtain the address of the variable.
2661 SDValue Chain = DAG.getEntryNode();
2662 SDValue FuncTLVGet =
2663 DAG.getLoad(MVT::i64, DL, Chain, DescAddr, MachinePointerInfo::getGOT(),
2664 false, true, true, 8);
2665 Chain = FuncTLVGet.getValue(1);
2667 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2668 MFI->setAdjustsStack(true);
2670 // TLS calls preserve all registers except those that absolutely must be
2671 // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
2673 const TargetRegisterInfo *TRI =
2674 getTargetMachine().getSubtargetImpl()->getRegisterInfo();
2675 const AArch64RegisterInfo *ARI =
2676 static_cast<const AArch64RegisterInfo *>(TRI);
2677 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
2679 // Finally, we can make the call. This is just a degenerate version of a
2680 // normal AArch64 call node: x0 takes the address of the descriptor, and
2681 // returns the address of the variable in this thread.
2682 Chain = DAG.getCopyToReg(Chain, DL, AArch64::X0, DescAddr, SDValue());
2684 DAG.getNode(AArch64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
2685 Chain, FuncTLVGet, DAG.getRegister(AArch64::X0, MVT::i64),
2686 DAG.getRegisterMask(Mask), Chain.getValue(1));
2687 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Chain.getValue(1));
2690 /// When accessing thread-local variables under either the general-dynamic or
2691 /// local-dynamic system, we make a "TLS-descriptor" call. The variable will
2692 /// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
2693 /// is a function pointer to carry out the resolution. This function takes the
2694 /// address of the descriptor in X0 and returns the TPIDR_EL0 offset in X0. All
2695 /// other registers (except LR, NZCV) are preserved.
2697 /// Thus, the ideal call sequence on AArch64 is:
2699 /// adrp x0, :tlsdesc:thread_var
2700 /// ldr x8, [x0, :tlsdesc_lo12:thread_var]
2701 /// add x0, x0, :tlsdesc_lo12:thread_var
2702 /// .tlsdesccall thread_var
2704 /// (TPIDR_EL0 offset now in x0).
2706 /// The ".tlsdesccall" directive instructs the assembler to insert a particular
2707 /// relocation to help the linker relax this sequence if it turns out to be too
2710 /// FIXME: we currently produce an extra, duplicated, ADRP instruction, but this
2712 SDValue AArch64TargetLowering::LowerELFTLSDescCall(SDValue SymAddr,
2713 SDValue DescAddr, SDLoc DL,
2714 SelectionDAG &DAG) const {
2715 EVT PtrVT = getPointerTy();
2717 // The function we need to call is simply the first entry in the GOT for this
2718 // descriptor, load it in preparation.
2719 SDValue Func = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, SymAddr);
2721 // TLS calls preserve all registers except those that absolutely must be
2722 // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
2724 const TargetRegisterInfo *TRI =
2725 getTargetMachine().getSubtargetImpl()->getRegisterInfo();
2726 const AArch64RegisterInfo *ARI =
2727 static_cast<const AArch64RegisterInfo *>(TRI);
2728 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
2730 // The function takes only one argument: the address of the descriptor itself
2732 SDValue Glue, Chain;
2733 Chain = DAG.getCopyToReg(DAG.getEntryNode(), DL, AArch64::X0, DescAddr, Glue);
2734 Glue = Chain.getValue(1);
2736 // We're now ready to populate the argument list, as with a normal call:
2737 SmallVector<SDValue, 6> Ops;
2738 Ops.push_back(Chain);
2739 Ops.push_back(Func);
2740 Ops.push_back(SymAddr);
2741 Ops.push_back(DAG.getRegister(AArch64::X0, PtrVT));
2742 Ops.push_back(DAG.getRegisterMask(Mask));
2743 Ops.push_back(Glue);
2745 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2746 Chain = DAG.getNode(AArch64ISD::TLSDESC_CALL, DL, NodeTys, Ops);
2747 Glue = Chain.getValue(1);
2749 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Glue);
2753 AArch64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
2754 SelectionDAG &DAG) const {
2755 assert(Subtarget->isTargetELF() && "This function expects an ELF target");
2756 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
2757 "ELF TLS only supported in small memory model");
2758 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2760 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
2763 EVT PtrVT = getPointerTy();
2765 const GlobalValue *GV = GA->getGlobal();
2767 SDValue ThreadBase = DAG.getNode(AArch64ISD::THREAD_POINTER, DL, PtrVT);
2769 if (Model == TLSModel::LocalExec) {
2770 SDValue HiVar = DAG.getTargetGlobalAddress(
2771 GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_G1);
2772 SDValue LoVar = DAG.getTargetGlobalAddress(
2774 AArch64II::MO_TLS | AArch64II::MO_G0 | AArch64II::MO_NC);
2776 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVZXi, DL, PtrVT, HiVar,
2777 DAG.getTargetConstant(16, MVT::i32)),
2779 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVKXi, DL, PtrVT, TPOff, LoVar,
2780 DAG.getTargetConstant(0, MVT::i32)),
2782 } else if (Model == TLSModel::InitialExec) {
2783 TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
2784 TPOff = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TPOff);
2785 } else if (Model == TLSModel::LocalDynamic) {
2786 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
2787 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
2788 // the beginning of the module's TLS region, followed by a DTPREL offset
2791 // These accesses will need deduplicating if there's more than one.
2792 AArch64FunctionInfo *MFI =
2793 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
2794 MFI->incNumLocalDynamicTLSAccesses();
2796 // Accesses used in this sequence go via the TLS descriptor which lives in
2797 // the GOT. Prepare an address we can use to handle this.
2798 SDValue HiDesc = DAG.getTargetExternalSymbol(
2799 "_TLS_MODULE_BASE_", PtrVT, AArch64II::MO_TLS | AArch64II::MO_PAGE);
2800 SDValue LoDesc = DAG.getTargetExternalSymbol(
2801 "_TLS_MODULE_BASE_", PtrVT,
2802 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
2804 // First argument to the descriptor call is the address of the descriptor
2806 SDValue DescAddr = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, HiDesc);
2807 DescAddr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
2809 // The call needs a relocation too for linker relaxation. It doesn't make
2810 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
2812 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
2815 // Now we can calculate the offset from TPIDR_EL0 to this module's
2816 // thread-local area.
2817 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
2819 // Now use :dtprel_whatever: operations to calculate this variable's offset
2820 // in its thread-storage area.
2821 SDValue HiVar = DAG.getTargetGlobalAddress(
2822 GV, DL, MVT::i64, 0, AArch64II::MO_TLS | AArch64II::MO_G1);
2823 SDValue LoVar = DAG.getTargetGlobalAddress(
2824 GV, DL, MVT::i64, 0,
2825 AArch64II::MO_TLS | AArch64II::MO_G0 | AArch64II::MO_NC);
2828 SDValue(DAG.getMachineNode(AArch64::MOVZXi, DL, PtrVT, HiVar,
2829 DAG.getTargetConstant(16, MVT::i32)),
2832 SDValue(DAG.getMachineNode(AArch64::MOVKXi, DL, PtrVT, DTPOff, LoVar,
2833 DAG.getTargetConstant(0, MVT::i32)),
2836 TPOff = DAG.getNode(ISD::ADD, DL, PtrVT, TPOff, DTPOff);
2837 } else if (Model == TLSModel::GeneralDynamic) {
2838 // Accesses used in this sequence go via the TLS descriptor which lives in
2839 // the GOT. Prepare an address we can use to handle this.
2840 SDValue HiDesc = DAG.getTargetGlobalAddress(
2841 GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_PAGE);
2842 SDValue LoDesc = DAG.getTargetGlobalAddress(
2844 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
2846 // First argument to the descriptor call is the address of the descriptor
2848 SDValue DescAddr = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, HiDesc);
2849 DescAddr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
2851 // The call needs a relocation too for linker relaxation. It doesn't make
2852 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
2855 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
2857 // Finally we can make a call to calculate the offset from tpidr_el0.
2858 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
2860 llvm_unreachable("Unsupported ELF TLS access model");
2862 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
2865 SDValue AArch64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
2866 SelectionDAG &DAG) const {
2867 if (Subtarget->isTargetDarwin())
2868 return LowerDarwinGlobalTLSAddress(Op, DAG);
2869 else if (Subtarget->isTargetELF())
2870 return LowerELFGlobalTLSAddress(Op, DAG);
2872 llvm_unreachable("Unexpected platform trying to use TLS");
2874 SDValue AArch64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2875 SDValue Chain = Op.getOperand(0);
2876 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2877 SDValue LHS = Op.getOperand(2);
2878 SDValue RHS = Op.getOperand(3);
2879 SDValue Dest = Op.getOperand(4);
2882 // Handle f128 first, since lowering it will result in comparing the return
2883 // value of a libcall against zero, which is just what the rest of LowerBR_CC
2884 // is expecting to deal with.
2885 if (LHS.getValueType() == MVT::f128) {
2886 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
2888 // If softenSetCCOperands returned a scalar, we need to compare the result
2889 // against zero to select between true and false values.
2890 if (!RHS.getNode()) {
2891 RHS = DAG.getConstant(0, LHS.getValueType());
2896 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
2898 unsigned Opc = LHS.getOpcode();
2899 if (LHS.getResNo() == 1 && isa<ConstantSDNode>(RHS) &&
2900 cast<ConstantSDNode>(RHS)->isOne() &&
2901 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
2902 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
2903 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
2904 "Unexpected condition code.");
2905 // Only lower legal XALUO ops.
2906 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
2909 // The actual operation with overflow check.
2910 AArch64CC::CondCode OFCC;
2911 SDValue Value, Overflow;
2912 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, LHS.getValue(0), DAG);
2914 if (CC == ISD::SETNE)
2915 OFCC = getInvertedCondCode(OFCC);
2916 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
2918 return DAG.getNode(AArch64ISD::BRCOND, SDLoc(LHS), MVT::Other, Chain, Dest,
2922 if (LHS.getValueType().isInteger()) {
2923 assert((LHS.getValueType() == RHS.getValueType()) &&
2924 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
2926 // If the RHS of the comparison is zero, we can potentially fold this
2927 // to a specialized branch.
2928 const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
2929 if (RHSC && RHSC->getZExtValue() == 0) {
2930 if (CC == ISD::SETEQ) {
2931 // See if we can use a TBZ to fold in an AND as well.
2932 // TBZ has a smaller branch displacement than CBZ. If the offset is
2933 // out of bounds, a late MI-layer pass rewrites branches.
2934 // 403.gcc is an example that hits this case.
2935 if (LHS.getOpcode() == ISD::AND &&
2936 isa<ConstantSDNode>(LHS.getOperand(1)) &&
2937 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
2938 SDValue Test = LHS.getOperand(0);
2939 uint64_t Mask = LHS.getConstantOperandVal(1);
2940 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, Test,
2941 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
2944 return DAG.getNode(AArch64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest);
2945 } else if (CC == ISD::SETNE) {
2946 // See if we can use a TBZ to fold in an AND as well.
2947 // TBZ has a smaller branch displacement than CBZ. If the offset is
2948 // out of bounds, a late MI-layer pass rewrites branches.
2949 // 403.gcc is an example that hits this case.
2950 if (LHS.getOpcode() == ISD::AND &&
2951 isa<ConstantSDNode>(LHS.getOperand(1)) &&
2952 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
2953 SDValue Test = LHS.getOperand(0);
2954 uint64_t Mask = LHS.getConstantOperandVal(1);
2955 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, Test,
2956 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
2959 return DAG.getNode(AArch64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest);
2960 } else if (CC == ISD::SETLT && LHS.getOpcode() != ISD::AND) {
2961 // Don't combine AND since emitComparison converts the AND to an ANDS
2962 // (a.k.a. TST) and the test in the test bit and branch instruction
2963 // becomes redundant. This would also increase register pressure.
2964 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
2965 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, LHS,
2966 DAG.getConstant(Mask, MVT::i64), Dest);
2969 if (RHSC && RHSC->getSExtValue() == -1 && CC == ISD::SETGT &&
2970 LHS.getOpcode() != ISD::AND) {
2971 // Don't combine AND since emitComparison converts the AND to an ANDS
2972 // (a.k.a. TST) and the test in the test bit and branch instruction
2973 // becomes redundant. This would also increase register pressure.
2974 uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
2975 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, LHS,
2976 DAG.getConstant(Mask, MVT::i64), Dest);
2980 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
2981 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
2985 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2987 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
2988 // clean. Some of them require two branches to implement.
2989 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
2990 AArch64CC::CondCode CC1, CC2;
2991 changeFPCCToAArch64CC(CC, CC1, CC2);
2992 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
2994 DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp);
2995 if (CC2 != AArch64CC::AL) {
2996 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
2997 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val,
3004 SDValue AArch64TargetLowering::LowerFCOPYSIGN(SDValue Op,
3005 SelectionDAG &DAG) const {
3006 EVT VT = Op.getValueType();
3009 SDValue In1 = Op.getOperand(0);
3010 SDValue In2 = Op.getOperand(1);
3011 EVT SrcVT = In2.getValueType();
3013 if (SrcVT == MVT::f32 && VT == MVT::f64)
3014 In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2);
3015 else if (SrcVT == MVT::f64 && VT == MVT::f32)
3016 In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, DAG.getIntPtrConstant(0));
3018 // FIXME: Src type is different, bail out for now. Can VT really be a
3025 SDValue EltMask, VecVal1, VecVal2;
3026 if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) {
3029 EltMask = DAG.getConstant(0x80000000ULL, EltVT);
3031 if (!VT.isVector()) {
3032 VecVal1 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3033 DAG.getUNDEF(VecVT), In1);
3034 VecVal2 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
3035 DAG.getUNDEF(VecVT), In2);
3037 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3038 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3040 } else if (VT == MVT::f64 || VT == MVT::v2f64) {
3044 // We want to materialize a mask with the the high bit set, but the AdvSIMD
3045 // immediate moves cannot materialize that in a single instruction for
3046 // 64-bit elements. Instead, materialize zero and then negate it.
3047 EltMask = DAG.getConstant(0, EltVT);
3049 if (!VT.isVector()) {
3050 VecVal1 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3051 DAG.getUNDEF(VecVT), In1);
3052 VecVal2 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
3053 DAG.getUNDEF(VecVT), In2);
3055 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3056 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3059 llvm_unreachable("Invalid type for copysign!");
3062 std::vector<SDValue> BuildVectorOps;
3063 for (unsigned i = 0; i < VecVT.getVectorNumElements(); ++i)
3064 BuildVectorOps.push_back(EltMask);
3066 SDValue BuildVec = DAG.getNode(ISD::BUILD_VECTOR, DL, VecVT, BuildVectorOps);
3068 // If we couldn't materialize the mask above, then the mask vector will be
3069 // the zero vector, and we need to negate it here.
3070 if (VT == MVT::f64 || VT == MVT::v2f64) {
3071 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec);
3072 BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec);
3073 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec);
3077 DAG.getNode(AArch64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec);
3080 return DAG.getTargetExtractSubreg(AArch64::ssub, DL, VT, Sel);
3081 else if (VT == MVT::f64)
3082 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, VT, Sel);
3084 return DAG.getNode(ISD::BITCAST, DL, VT, Sel);
3087 SDValue AArch64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
3088 if (DAG.getMachineFunction().getFunction()->getAttributes().hasAttribute(
3089 AttributeSet::FunctionIndex, Attribute::NoImplicitFloat))
3092 // While there is no integer popcount instruction, it can
3093 // be more efficiently lowered to the following sequence that uses
3094 // AdvSIMD registers/instructions as long as the copies to/from
3095 // the AdvSIMD registers are cheap.
3096 // FMOV D0, X0 // copy 64-bit int to vector, high bits zero'd
3097 // CNT V0.8B, V0.8B // 8xbyte pop-counts
3098 // ADDV B0, V0.8B // sum 8xbyte pop-counts
3099 // UMOV X0, V0.B[0] // copy byte result back to integer reg
3100 SDValue Val = Op.getOperand(0);
3102 EVT VT = Op.getValueType();
3103 SDValue ZeroVec = DAG.getUNDEF(MVT::v8i8);
3106 if (VT == MVT::i32) {
3107 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Val);
3108 VecVal = DAG.getTargetInsertSubreg(AArch64::ssub, DL, MVT::v8i8, ZeroVec,
3111 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);
3114 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, VecVal);
3115 SDValue UaddLV = DAG.getNode(
3116 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
3117 DAG.getConstant(Intrinsic::aarch64_neon_uaddlv, MVT::i32), CtPop);
3120 UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV);
3124 SDValue AArch64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
3126 if (Op.getValueType().isVector())
3127 return LowerVSETCC(Op, DAG);
3129 SDValue LHS = Op.getOperand(0);
3130 SDValue RHS = Op.getOperand(1);
3131 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
3134 // We chose ZeroOrOneBooleanContents, so use zero and one.
3135 EVT VT = Op.getValueType();
3136 SDValue TVal = DAG.getConstant(1, VT);
3137 SDValue FVal = DAG.getConstant(0, VT);
3139 // Handle f128 first, since one possible outcome is a normal integer
3140 // comparison which gets picked up by the next if statement.
3141 if (LHS.getValueType() == MVT::f128) {
3142 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3144 // If softenSetCCOperands returned a scalar, use it.
3145 if (!RHS.getNode()) {
3146 assert(LHS.getValueType() == Op.getValueType() &&
3147 "Unexpected setcc expansion!");
3152 if (LHS.getValueType().isInteger()) {
3155 getAArch64Cmp(LHS, RHS, ISD::getSetCCInverse(CC, true), CCVal, DAG, dl);
3157 // Note that we inverted the condition above, so we reverse the order of
3158 // the true and false operands here. This will allow the setcc to be
3159 // matched to a single CSINC instruction.
3160 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp);
3163 // Now we know we're dealing with FP values.
3164 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3166 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3167 // and do the comparison.
3168 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3170 AArch64CC::CondCode CC1, CC2;
3171 changeFPCCToAArch64CC(CC, CC1, CC2);
3172 if (CC2 == AArch64CC::AL) {
3173 changeFPCCToAArch64CC(ISD::getSetCCInverse(CC, false), CC1, CC2);
3174 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3176 // Note that we inverted the condition above, so we reverse the order of
3177 // the true and false operands here. This will allow the setcc to be
3178 // matched to a single CSINC instruction.
3179 return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp);
3181 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't
3182 // totally clean. Some of them require two CSELs to implement. As is in
3183 // this case, we emit the first CSEL and then emit a second using the output
3184 // of the first as the RHS. We're effectively OR'ing the two CC's together.
3186 // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
3187 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3189 DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3191 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3192 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3196 /// A SELECT_CC operation is really some kind of max or min if both values being
3197 /// compared are, in some sense, equal to the results in either case. However,
3198 /// it is permissible to compare f32 values and produce directly extended f64
3201 /// Extending the comparison operands would also be allowed, but is less likely
3202 /// to happen in practice since their use is right here. Note that truncate
3203 /// operations would *not* be semantically equivalent.
3204 static bool selectCCOpsAreFMaxCompatible(SDValue Cmp, SDValue Result) {
3208 ConstantFPSDNode *CCmp = dyn_cast<ConstantFPSDNode>(Cmp);
3209 ConstantFPSDNode *CResult = dyn_cast<ConstantFPSDNode>(Result);
3210 if (CCmp && CResult && Cmp.getValueType() == MVT::f32 &&
3211 Result.getValueType() == MVT::f64) {
3213 APFloat CmpVal = CCmp->getValueAPF();
3214 CmpVal.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &Lossy);
3215 return CResult->getValueAPF().bitwiseIsEqual(CmpVal);
3218 return Result->getOpcode() == ISD::FP_EXTEND && Result->getOperand(0) == Cmp;
3221 SDValue AArch64TargetLowering::LowerSELECT(SDValue Op,
3222 SelectionDAG &DAG) const {
3223 SDValue CC = Op->getOperand(0);
3224 SDValue TVal = Op->getOperand(1);
3225 SDValue FVal = Op->getOperand(2);
3228 unsigned Opc = CC.getOpcode();
3229 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
3231 if (CC.getResNo() == 1 &&
3232 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3233 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3234 // Only lower legal XALUO ops.
3235 if (!DAG.getTargetLoweringInfo().isTypeLegal(CC->getValueType(0)))
3238 AArch64CC::CondCode OFCC;
3239 SDValue Value, Overflow;
3240 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, CC.getValue(0), DAG);
3241 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
3243 return DAG.getNode(AArch64ISD::CSEL, DL, Op.getValueType(), TVal, FVal,
3247 if (CC.getOpcode() == ISD::SETCC)
3248 return DAG.getSelectCC(DL, CC.getOperand(0), CC.getOperand(1), TVal, FVal,
3249 cast<CondCodeSDNode>(CC.getOperand(2))->get());
3251 return DAG.getSelectCC(DL, CC, DAG.getConstant(0, CC.getValueType()), TVal,
3255 SDValue AArch64TargetLowering::LowerSELECT_CC(SDValue Op,
3256 SelectionDAG &DAG) const {
3257 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3258 SDValue LHS = Op.getOperand(0);
3259 SDValue RHS = Op.getOperand(1);
3260 SDValue TVal = Op.getOperand(2);
3261 SDValue FVal = Op.getOperand(3);
3264 // Handle f128 first, because it will result in a comparison of some RTLIB
3265 // call result against zero.
3266 if (LHS.getValueType() == MVT::f128) {
3267 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3269 // If softenSetCCOperands returned a scalar, we need to compare the result
3270 // against zero to select between true and false values.
3271 if (!RHS.getNode()) {
3272 RHS = DAG.getConstant(0, LHS.getValueType());
3277 // Handle integers first.
3278 if (LHS.getValueType().isInteger()) {
3279 assert((LHS.getValueType() == RHS.getValueType()) &&
3280 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3282 unsigned Opcode = AArch64ISD::CSEL;
3284 // If both the TVal and the FVal are constants, see if we can swap them in
3285 // order to for a CSINV or CSINC out of them.
3286 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
3287 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
3289 if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) {
3290 std::swap(TVal, FVal);
3291 std::swap(CTVal, CFVal);
3292 CC = ISD::getSetCCInverse(CC, true);
3293 } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) {
3294 std::swap(TVal, FVal);
3295 std::swap(CTVal, CFVal);
3296 CC = ISD::getSetCCInverse(CC, true);
3297 } else if (TVal.getOpcode() == ISD::XOR) {
3298 // If TVal is a NOT we want to swap TVal and FVal so that we can match
3299 // with a CSINV rather than a CSEL.
3300 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(1));
3302 if (CVal && CVal->isAllOnesValue()) {
3303 std::swap(TVal, FVal);
3304 std::swap(CTVal, CFVal);
3305 CC = ISD::getSetCCInverse(CC, true);
3307 } else if (TVal.getOpcode() == ISD::SUB) {
3308 // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
3309 // that we can match with a CSNEG rather than a CSEL.
3310 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(0));
3312 if (CVal && CVal->isNullValue()) {
3313 std::swap(TVal, FVal);
3314 std::swap(CTVal, CFVal);
3315 CC = ISD::getSetCCInverse(CC, true);
3317 } else if (CTVal && CFVal) {
3318 const int64_t TrueVal = CTVal->getSExtValue();
3319 const int64_t FalseVal = CFVal->getSExtValue();
3322 // If both TVal and FVal are constants, see if FVal is the
3323 // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
3324 // instead of a CSEL in that case.
3325 if (TrueVal == ~FalseVal) {
3326 Opcode = AArch64ISD::CSINV;
3327 } else if (TrueVal == -FalseVal) {
3328 Opcode = AArch64ISD::CSNEG;
3329 } else if (TVal.getValueType() == MVT::i32) {
3330 // If our operands are only 32-bit wide, make sure we use 32-bit
3331 // arithmetic for the check whether we can use CSINC. This ensures that
3332 // the addition in the check will wrap around properly in case there is
3333 // an overflow (which would not be the case if we do the check with
3334 // 64-bit arithmetic).
3335 const uint32_t TrueVal32 = CTVal->getZExtValue();
3336 const uint32_t FalseVal32 = CFVal->getZExtValue();
3338 if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
3339 Opcode = AArch64ISD::CSINC;
3341 if (TrueVal32 > FalseVal32) {
3345 // 64-bit check whether we can use CSINC.
3346 } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) {
3347 Opcode = AArch64ISD::CSINC;
3349 if (TrueVal > FalseVal) {
3354 // Swap TVal and FVal if necessary.
3356 std::swap(TVal, FVal);
3357 std::swap(CTVal, CFVal);
3358 CC = ISD::getSetCCInverse(CC, true);
3361 if (Opcode != AArch64ISD::CSEL) {
3362 // Drop FVal since we can get its value by simply inverting/negating
3369 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3371 EVT VT = Op.getValueType();
3372 return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp);
3375 // Now we know we're dealing with FP values.
3376 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3377 assert(LHS.getValueType() == RHS.getValueType());
3378 EVT VT = Op.getValueType();
3380 // Try to match this select into a max/min operation, which have dedicated
3381 // opcode in the instruction set.
3382 // FIXME: This is not correct in the presence of NaNs, so we only enable this
3384 if (getTargetMachine().Options.NoNaNsFPMath) {
3385 SDValue MinMaxLHS = TVal, MinMaxRHS = FVal;
3386 if (selectCCOpsAreFMaxCompatible(LHS, MinMaxRHS) &&
3387 selectCCOpsAreFMaxCompatible(RHS, MinMaxLHS)) {
3388 CC = ISD::getSetCCSwappedOperands(CC);
3389 std::swap(MinMaxLHS, MinMaxRHS);
3392 if (selectCCOpsAreFMaxCompatible(LHS, MinMaxLHS) &&
3393 selectCCOpsAreFMaxCompatible(RHS, MinMaxRHS)) {
3403 return DAG.getNode(AArch64ISD::FMAX, dl, VT, MinMaxLHS, MinMaxRHS);
3411 return DAG.getNode(AArch64ISD::FMIN, dl, VT, MinMaxLHS, MinMaxRHS);
3417 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3418 // and do the comparison.
3419 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3421 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
3422 // clean. Some of them require two CSELs to implement.
3423 AArch64CC::CondCode CC1, CC2;
3424 changeFPCCToAArch64CC(CC, CC1, CC2);
3425 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3426 SDValue CS1 = DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3428 // If we need a second CSEL, emit it, using the output of the first as the
3429 // RHS. We're effectively OR'ing the two CC's together.
3430 if (CC2 != AArch64CC::AL) {
3431 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3432 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3435 // Otherwise, return the output of the first CSEL.
3439 SDValue AArch64TargetLowering::LowerJumpTable(SDValue Op,
3440 SelectionDAG &DAG) const {
3441 // Jump table entries as PC relative offsets. No additional tweaking
3442 // is necessary here. Just get the address of the jump table.
3443 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3444 EVT PtrVT = getPointerTy();
3447 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3448 !Subtarget->isTargetMachO()) {
3449 const unsigned char MO_NC = AArch64II::MO_NC;
3451 AArch64ISD::WrapperLarge, DL, PtrVT,
3452 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G3),
3453 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G2 | MO_NC),
3454 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G1 | MO_NC),
3455 DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3456 AArch64II::MO_G0 | MO_NC));
3460 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_PAGE);
3461 SDValue Lo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3462 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3463 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3464 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3467 SDValue AArch64TargetLowering::LowerConstantPool(SDValue Op,
3468 SelectionDAG &DAG) const {
3469 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3470 EVT PtrVT = getPointerTy();
3473 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
3474 // Use the GOT for the large code model on iOS.
3475 if (Subtarget->isTargetMachO()) {
3476 SDValue GotAddr = DAG.getTargetConstantPool(
3477 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3479 return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
3482 const unsigned char MO_NC = AArch64II::MO_NC;
3484 AArch64ISD::WrapperLarge, DL, PtrVT,
3485 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3486 CP->getOffset(), AArch64II::MO_G3),
3487 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3488 CP->getOffset(), AArch64II::MO_G2 | MO_NC),
3489 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3490 CP->getOffset(), AArch64II::MO_G1 | MO_NC),
3491 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3492 CP->getOffset(), AArch64II::MO_G0 | MO_NC));
3494 // Use ADRP/ADD or ADRP/LDR for everything else: the small memory model on
3495 // ELF, the only valid one on Darwin.
3497 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3498 CP->getOffset(), AArch64II::MO_PAGE);
3499 SDValue Lo = DAG.getTargetConstantPool(
3500 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3501 AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3503 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3504 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3508 SDValue AArch64TargetLowering::LowerBlockAddress(SDValue Op,
3509 SelectionDAG &DAG) const {
3510 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
3511 EVT PtrVT = getPointerTy();
3513 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3514 !Subtarget->isTargetMachO()) {
3515 const unsigned char MO_NC = AArch64II::MO_NC;
3517 AArch64ISD::WrapperLarge, DL, PtrVT,
3518 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G3),
3519 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
3520 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
3521 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
3523 SDValue Hi = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGE);
3524 SDValue Lo = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGEOFF |
3526 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
3527 return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3531 SDValue AArch64TargetLowering::LowerDarwin_VASTART(SDValue Op,
3532 SelectionDAG &DAG) const {
3533 AArch64FunctionInfo *FuncInfo =
3534 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
3538 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3539 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3540 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
3541 MachinePointerInfo(SV), false, false, 0);
3544 SDValue AArch64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
3545 SelectionDAG &DAG) const {
3546 // The layout of the va_list struct is specified in the AArch64 Procedure Call
3547 // Standard, section B.3.
3548 MachineFunction &MF = DAG.getMachineFunction();
3549 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
3552 SDValue Chain = Op.getOperand(0);
3553 SDValue VAList = Op.getOperand(1);
3554 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3555 SmallVector<SDValue, 4> MemOps;
3557 // void *__stack at offset 0
3559 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3560 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
3561 MachinePointerInfo(SV), false, false, 8));
3563 // void *__gr_top at offset 8
3564 int GPRSize = FuncInfo->getVarArgsGPRSize();
3566 SDValue GRTop, GRTopAddr;
3568 GRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3569 DAG.getConstant(8, getPointerTy()));
3571 GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), getPointerTy());
3572 GRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), GRTop,
3573 DAG.getConstant(GPRSize, getPointerTy()));
3575 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
3576 MachinePointerInfo(SV, 8), false, false, 8));
3579 // void *__vr_top at offset 16
3580 int FPRSize = FuncInfo->getVarArgsFPRSize();
3582 SDValue VRTop, VRTopAddr;
3583 VRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3584 DAG.getConstant(16, getPointerTy()));
3586 VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), getPointerTy());
3587 VRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), VRTop,
3588 DAG.getConstant(FPRSize, getPointerTy()));
3590 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
3591 MachinePointerInfo(SV, 16), false, false, 8));
3594 // int __gr_offs at offset 24
3595 SDValue GROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3596 DAG.getConstant(24, getPointerTy()));
3597 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-GPRSize, MVT::i32),
3598 GROffsAddr, MachinePointerInfo(SV, 24), false,
3601 // int __vr_offs at offset 28
3602 SDValue VROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3603 DAG.getConstant(28, getPointerTy()));
3604 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-FPRSize, MVT::i32),
3605 VROffsAddr, MachinePointerInfo(SV, 28), false,
3608 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
3611 SDValue AArch64TargetLowering::LowerVASTART(SDValue Op,
3612 SelectionDAG &DAG) const {
3613 return Subtarget->isTargetDarwin() ? LowerDarwin_VASTART(Op, DAG)
3614 : LowerAAPCS_VASTART(Op, DAG);
3617 SDValue AArch64TargetLowering::LowerVACOPY(SDValue Op,
3618 SelectionDAG &DAG) const {
3619 // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
3621 unsigned VaListSize = Subtarget->isTargetDarwin() ? 8 : 32;
3622 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
3623 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
3625 return DAG.getMemcpy(Op.getOperand(0), SDLoc(Op), Op.getOperand(1),
3626 Op.getOperand(2), DAG.getConstant(VaListSize, MVT::i32),
3627 8, false, false, MachinePointerInfo(DestSV),
3628 MachinePointerInfo(SrcSV));
3631 SDValue AArch64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
3632 assert(Subtarget->isTargetDarwin() &&
3633 "automatic va_arg instruction only works on Darwin");
3635 const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3636 EVT VT = Op.getValueType();
3638 SDValue Chain = Op.getOperand(0);
3639 SDValue Addr = Op.getOperand(1);
3640 unsigned Align = Op.getConstantOperandVal(3);
3642 SDValue VAList = DAG.getLoad(getPointerTy(), DL, Chain, Addr,
3643 MachinePointerInfo(V), false, false, false, 0);
3644 Chain = VAList.getValue(1);
3647 assert(((Align & (Align - 1)) == 0) && "Expected Align to be a power of 2");
3648 VAList = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3649 DAG.getConstant(Align - 1, getPointerTy()));
3650 VAList = DAG.getNode(ISD::AND, DL, getPointerTy(), VAList,
3651 DAG.getConstant(-(int64_t)Align, getPointerTy()));
3654 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
3655 uint64_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
3657 // Scalar integer and FP values smaller than 64 bits are implicitly extended
3658 // up to 64 bits. At the very least, we have to increase the striding of the
3659 // vaargs list to match this, and for FP values we need to introduce
3660 // FP_ROUND nodes as well.
3661 if (VT.isInteger() && !VT.isVector())
3663 bool NeedFPTrunc = false;
3664 if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
3669 // Increment the pointer, VAList, to the next vaarg
3670 SDValue VANext = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3671 DAG.getConstant(ArgSize, getPointerTy()));
3672 // Store the incremented VAList to the legalized pointer
3673 SDValue APStore = DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V),
3676 // Load the actual argument out of the pointer VAList
3678 // Load the value as an f64.
3679 SDValue WideFP = DAG.getLoad(MVT::f64, DL, APStore, VAList,
3680 MachinePointerInfo(), false, false, false, 0);
3681 // Round the value down to an f32.
3682 SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
3683 DAG.getIntPtrConstant(1));
3684 SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
3685 // Merge the rounded value with the chain output of the load.
3686 return DAG.getMergeValues(Ops, DL);
3689 return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo(), false,
3693 SDValue AArch64TargetLowering::LowerFRAMEADDR(SDValue Op,
3694 SelectionDAG &DAG) const {
3695 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3696 MFI->setFrameAddressIsTaken(true);
3698 EVT VT = Op.getValueType();
3700 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3702 DAG.getCopyFromReg(DAG.getEntryNode(), DL, AArch64::FP, VT);
3704 FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
3705 MachinePointerInfo(), false, false, false, 0);
3709 // FIXME? Maybe this could be a TableGen attribute on some registers and
3710 // this table could be generated automatically from RegInfo.
3711 unsigned AArch64TargetLowering::getRegisterByName(const char* RegName,
3713 unsigned Reg = StringSwitch<unsigned>(RegName)
3714 .Case("sp", AArch64::SP)
3718 report_fatal_error("Invalid register name global variable");
3721 SDValue AArch64TargetLowering::LowerRETURNADDR(SDValue Op,
3722 SelectionDAG &DAG) const {
3723 MachineFunction &MF = DAG.getMachineFunction();
3724 MachineFrameInfo *MFI = MF.getFrameInfo();
3725 MFI->setReturnAddressIsTaken(true);
3727 EVT VT = Op.getValueType();
3729 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3731 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3732 SDValue Offset = DAG.getConstant(8, getPointerTy());
3733 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
3734 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
3735 MachinePointerInfo(), false, false, false, 0);
3738 // Return LR, which contains the return address. Mark it an implicit live-in.
3739 unsigned Reg = MF.addLiveIn(AArch64::LR, &AArch64::GPR64RegClass);
3740 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
3743 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3744 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
3745 SDValue AArch64TargetLowering::LowerShiftRightParts(SDValue Op,
3746 SelectionDAG &DAG) const {
3747 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3748 EVT VT = Op.getValueType();
3749 unsigned VTBits = VT.getSizeInBits();
3751 SDValue ShOpLo = Op.getOperand(0);
3752 SDValue ShOpHi = Op.getOperand(1);
3753 SDValue ShAmt = Op.getOperand(2);
3755 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3757 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3759 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
3760 DAG.getConstant(VTBits, MVT::i64), ShAmt);
3761 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3762 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
3763 DAG.getConstant(VTBits, MVT::i64));
3764 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3766 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
3767 ISD::SETGE, dl, DAG);
3768 SDValue CCVal = DAG.getConstant(AArch64CC::GE, MVT::i32);
3770 SDValue FalseValLo = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3771 SDValue TrueValLo = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3773 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
3775 // AArch64 shifts larger than the register width are wrapped rather than
3776 // clamped, so we can't just emit "hi >> x".
3777 SDValue FalseValHi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3778 SDValue TrueValHi = Opc == ISD::SRA
3779 ? DAG.getNode(Opc, dl, VT, ShOpHi,
3780 DAG.getConstant(VTBits - 1, MVT::i64))
3781 : DAG.getConstant(0, VT);
3783 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValHi, FalseValHi, CCVal, Cmp);
3785 SDValue Ops[2] = { Lo, Hi };
3786 return DAG.getMergeValues(Ops, dl);
3789 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3790 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
3791 SDValue AArch64TargetLowering::LowerShiftLeftParts(SDValue Op,
3792 SelectionDAG &DAG) const {
3793 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3794 EVT VT = Op.getValueType();
3795 unsigned VTBits = VT.getSizeInBits();
3797 SDValue ShOpLo = Op.getOperand(0);
3798 SDValue ShOpHi = Op.getOperand(1);
3799 SDValue ShAmt = Op.getOperand(2);
3802 assert(Op.getOpcode() == ISD::SHL_PARTS);
3803 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
3804 DAG.getConstant(VTBits, MVT::i64), ShAmt);
3805 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3806 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
3807 DAG.getConstant(VTBits, MVT::i64));
3808 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3809 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3811 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3813 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
3814 ISD::SETGE, dl, DAG);
3815 SDValue CCVal = DAG.getConstant(AArch64CC::GE, MVT::i32);
3817 DAG.getNode(AArch64ISD::CSEL, dl, VT, Tmp3, FalseVal, CCVal, Cmp);
3819 // AArch64 shifts of larger than register sizes are wrapped rather than
3820 // clamped, so we can't just emit "lo << a" if a is too big.
3821 SDValue TrueValLo = DAG.getConstant(0, VT);
3822 SDValue FalseValLo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3824 DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
3826 SDValue Ops[2] = { Lo, Hi };
3827 return DAG.getMergeValues(Ops, dl);
3830 bool AArch64TargetLowering::isOffsetFoldingLegal(
3831 const GlobalAddressSDNode *GA) const {
3832 // The AArch64 target doesn't support folding offsets into global addresses.
3836 bool AArch64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3837 // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
3838 // FIXME: We should be able to handle f128 as well with a clever lowering.
3839 if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32))
3843 return AArch64_AM::getFP64Imm(Imm) != -1;
3844 else if (VT == MVT::f32)
3845 return AArch64_AM::getFP32Imm(Imm) != -1;
3849 //===----------------------------------------------------------------------===//
3850 // AArch64 Optimization Hooks
3851 //===----------------------------------------------------------------------===//
3853 //===----------------------------------------------------------------------===//
3854 // AArch64 Inline Assembly Support
3855 //===----------------------------------------------------------------------===//
3857 // Table of Constraints
3858 // TODO: This is the current set of constraints supported by ARM for the
3859 // compiler, not all of them may make sense, e.g. S may be difficult to support.
3861 // r - A general register
3862 // w - An FP/SIMD register of some size in the range v0-v31
3863 // x - An FP/SIMD register of some size in the range v0-v15
3864 // I - Constant that can be used with an ADD instruction
3865 // J - Constant that can be used with a SUB instruction
3866 // K - Constant that can be used with a 32-bit logical instruction
3867 // L - Constant that can be used with a 64-bit logical instruction
3868 // M - Constant that can be used as a 32-bit MOV immediate
3869 // N - Constant that can be used as a 64-bit MOV immediate
3870 // Q - A memory reference with base register and no offset
3871 // S - A symbolic address
3872 // Y - Floating point constant zero
3873 // Z - Integer constant zero
3875 // Note that general register operands will be output using their 64-bit x
3876 // register name, whatever the size of the variable, unless the asm operand
3877 // is prefixed by the %w modifier. Floating-point and SIMD register operands
3878 // will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
3881 /// getConstraintType - Given a constraint letter, return the type of
3882 /// constraint it is for this target.
3883 AArch64TargetLowering::ConstraintType
3884 AArch64TargetLowering::getConstraintType(const std::string &Constraint) const {
3885 if (Constraint.size() == 1) {
3886 switch (Constraint[0]) {
3893 return C_RegisterClass;
3894 // An address with a single base register. Due to the way we
3895 // currently handle addresses it is the same as 'r'.
3900 return TargetLowering::getConstraintType(Constraint);
3903 /// Examine constraint type and operand type and determine a weight value.
3904 /// This object must already have been set up with the operand type
3905 /// and the current alternative constraint selected.
3906 TargetLowering::ConstraintWeight
3907 AArch64TargetLowering::getSingleConstraintMatchWeight(
3908 AsmOperandInfo &info, const char *constraint) const {
3909 ConstraintWeight weight = CW_Invalid;
3910 Value *CallOperandVal = info.CallOperandVal;
3911 // If we don't have a value, we can't do a match,
3912 // but allow it at the lowest weight.
3913 if (!CallOperandVal)
3915 Type *type = CallOperandVal->getType();
3916 // Look at the constraint type.
3917 switch (*constraint) {
3919 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
3923 if (type->isFloatingPointTy() || type->isVectorTy())
3924 weight = CW_Register;
3927 weight = CW_Constant;
3933 std::pair<unsigned, const TargetRegisterClass *>
3934 AArch64TargetLowering::getRegForInlineAsmConstraint(
3935 const std::string &Constraint, MVT VT) const {
3936 if (Constraint.size() == 1) {
3937 switch (Constraint[0]) {
3939 if (VT.getSizeInBits() == 64)
3940 return std::make_pair(0U, &AArch64::GPR64commonRegClass);
3941 return std::make_pair(0U, &AArch64::GPR32commonRegClass);
3944 return std::make_pair(0U, &AArch64::FPR32RegClass);
3945 if (VT.getSizeInBits() == 64)
3946 return std::make_pair(0U, &AArch64::FPR64RegClass);
3947 if (VT.getSizeInBits() == 128)
3948 return std::make_pair(0U, &AArch64::FPR128RegClass);
3950 // The instructions that this constraint is designed for can
3951 // only take 128-bit registers so just use that regclass.
3953 if (VT.getSizeInBits() == 128)
3954 return std::make_pair(0U, &AArch64::FPR128_loRegClass);
3958 if (StringRef("{cc}").equals_lower(Constraint))
3959 return std::make_pair(unsigned(AArch64::NZCV), &AArch64::CCRRegClass);
3961 // Use the default implementation in TargetLowering to convert the register
3962 // constraint into a member of a register class.
3963 std::pair<unsigned, const TargetRegisterClass *> Res;
3964 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3966 // Not found as a standard register?
3968 unsigned Size = Constraint.size();
3969 if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
3970 tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
3971 const std::string Reg =
3972 std::string(&Constraint[2], &Constraint[Size - 1]);
3973 int RegNo = atoi(Reg.c_str());
3974 if (RegNo >= 0 && RegNo <= 31) {
3975 // v0 - v31 are aliases of q0 - q31.
3976 // By default we'll emit v0-v31 for this unless there's a modifier where
3977 // we'll emit the correct register as well.
3978 Res.first = AArch64::FPR128RegClass.getRegister(RegNo);
3979 Res.second = &AArch64::FPR128RegClass;
3987 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3988 /// vector. If it is invalid, don't add anything to Ops.
3989 void AArch64TargetLowering::LowerAsmOperandForConstraint(
3990 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
3991 SelectionDAG &DAG) const {
3994 // Currently only support length 1 constraints.
3995 if (Constraint.length() != 1)
3998 char ConstraintLetter = Constraint[0];
3999 switch (ConstraintLetter) {
4003 // This set of constraints deal with valid constants for various instructions.
4004 // Validate and return a target constant for them if we can.
4006 // 'z' maps to xzr or wzr so it needs an input of 0.
4007 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4008 if (!C || C->getZExtValue() != 0)
4011 if (Op.getValueType() == MVT::i64)
4012 Result = DAG.getRegister(AArch64::XZR, MVT::i64);
4014 Result = DAG.getRegister(AArch64::WZR, MVT::i32);
4024 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4028 // Grab the value and do some validation.
4029 uint64_t CVal = C->getZExtValue();
4030 switch (ConstraintLetter) {
4031 // The I constraint applies only to simple ADD or SUB immediate operands:
4032 // i.e. 0 to 4095 with optional shift by 12
4033 // The J constraint applies only to ADD or SUB immediates that would be
4034 // valid when negated, i.e. if [an add pattern] were to be output as a SUB
4035 // instruction [or vice versa], in other words -1 to -4095 with optional
4036 // left shift by 12.
4038 if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
4042 uint64_t NVal = -C->getSExtValue();
4043 if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal)) {
4044 CVal = C->getSExtValue();
4049 // The K and L constraints apply *only* to logical immediates, including
4050 // what used to be the MOVI alias for ORR (though the MOVI alias has now
4051 // been removed and MOV should be used). So these constraints have to
4052 // distinguish between bit patterns that are valid 32-bit or 64-bit
4053 // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
4054 // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
4057 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4061 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4064 // The M and N constraints are a superset of K and L respectively, for use
4065 // with the MOV (immediate) alias. As well as the logical immediates they
4066 // also match 32 or 64-bit immediates that can be loaded either using a
4067 // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
4068 // (M) or 64-bit 0x1234000000000000 (N) etc.
4069 // As a note some of this code is liberally stolen from the asm parser.
4071 if (!isUInt<32>(CVal))
4073 if (AArch64_AM::isLogicalImmediate(CVal, 32))
4075 if ((CVal & 0xFFFF) == CVal)
4077 if ((CVal & 0xFFFF0000ULL) == CVal)
4079 uint64_t NCVal = ~(uint32_t)CVal;
4080 if ((NCVal & 0xFFFFULL) == NCVal)
4082 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4087 if (AArch64_AM::isLogicalImmediate(CVal, 64))
4089 if ((CVal & 0xFFFFULL) == CVal)
4091 if ((CVal & 0xFFFF0000ULL) == CVal)
4093 if ((CVal & 0xFFFF00000000ULL) == CVal)
4095 if ((CVal & 0xFFFF000000000000ULL) == CVal)
4097 uint64_t NCVal = ~CVal;
4098 if ((NCVal & 0xFFFFULL) == NCVal)
4100 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4102 if ((NCVal & 0xFFFF00000000ULL) == NCVal)
4104 if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
4112 // All assembler immediates are 64-bit integers.
4113 Result = DAG.getTargetConstant(CVal, MVT::i64);
4117 if (Result.getNode()) {
4118 Ops.push_back(Result);
4122 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
4125 //===----------------------------------------------------------------------===//
4126 // AArch64 Advanced SIMD Support
4127 //===----------------------------------------------------------------------===//
4129 /// WidenVector - Given a value in the V64 register class, produce the
4130 /// equivalent value in the V128 register class.
4131 static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
4132 EVT VT = V64Reg.getValueType();
4133 unsigned NarrowSize = VT.getVectorNumElements();
4134 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4135 MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
4138 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
4139 V64Reg, DAG.getConstant(0, MVT::i32));
4142 /// getExtFactor - Determine the adjustment factor for the position when
4143 /// generating an "extract from vector registers" instruction.
4144 static unsigned getExtFactor(SDValue &V) {
4145 EVT EltType = V.getValueType().getVectorElementType();
4146 return EltType.getSizeInBits() / 8;
4149 /// NarrowVector - Given a value in the V128 register class, produce the
4150 /// equivalent value in the V64 register class.
4151 static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) {
4152 EVT VT = V128Reg.getValueType();
4153 unsigned WideSize = VT.getVectorNumElements();
4154 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4155 MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2);
4158 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, NarrowTy, V128Reg);
4161 // Gather data to see if the operation can be modelled as a
4162 // shuffle in combination with VEXTs.
4163 SDValue AArch64TargetLowering::ReconstructShuffle(SDValue Op,
4164 SelectionDAG &DAG) const {
4165 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
4167 EVT VT = Op.getValueType();
4168 unsigned NumElts = VT.getVectorNumElements();
4170 struct ShuffleSourceInfo {
4175 // We may insert some combination of BITCASTs and VEXT nodes to force Vec to
4176 // be compatible with the shuffle we intend to construct. As a result
4177 // ShuffleVec will be some sliding window into the original Vec.
4180 // Code should guarantee that element i in Vec starts at element "WindowBase
4181 // + i * WindowScale in ShuffleVec".
4185 bool operator ==(SDValue OtherVec) { return Vec == OtherVec; }
4186 ShuffleSourceInfo(SDValue Vec)
4187 : Vec(Vec), MinElt(UINT_MAX), MaxElt(0), ShuffleVec(Vec), WindowBase(0),
4191 // First gather all vectors used as an immediate source for this BUILD_VECTOR
4193 SmallVector<ShuffleSourceInfo, 2> Sources;
4194 for (unsigned i = 0; i < NumElts; ++i) {
4195 SDValue V = Op.getOperand(i);
4196 if (V.getOpcode() == ISD::UNDEF)
4198 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4199 // A shuffle can only come from building a vector from various
4200 // elements of other vectors.
4204 // Add this element source to the list if it's not already there.
4205 SDValue SourceVec = V.getOperand(0);
4206 auto Source = std::find(Sources.begin(), Sources.end(), SourceVec);
4207 if (Source == Sources.end())
4208 Sources.push_back(ShuffleSourceInfo(SourceVec));
4210 // Update the minimum and maximum lane number seen.
4211 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4212 Source->MinElt = std::min(Source->MinElt, EltNo);
4213 Source->MaxElt = std::max(Source->MaxElt, EltNo);
4216 // Currently only do something sane when at most two source vectors
4218 if (Sources.size() > 2)
4221 // Find out the smallest element size among result and two sources, and use
4222 // it as element size to build the shuffle_vector.
4223 EVT SmallestEltTy = VT.getVectorElementType();
4224 for (auto &Source : Sources) {
4225 EVT SrcEltTy = Source.Vec.getValueType().getVectorElementType();
4226 if (SrcEltTy.bitsLT(SmallestEltTy)) {
4227 SmallestEltTy = SrcEltTy;
4230 unsigned ResMultiplier =
4231 VT.getVectorElementType().getSizeInBits() / SmallestEltTy.getSizeInBits();
4232 NumElts = VT.getSizeInBits() / SmallestEltTy.getSizeInBits();
4233 EVT ShuffleVT = EVT::getVectorVT(*DAG.getContext(), SmallestEltTy, NumElts);
4235 // If the source vector is too wide or too narrow, we may nevertheless be able
4236 // to construct a compatible shuffle either by concatenating it with UNDEF or
4237 // extracting a suitable range of elements.
4238 for (auto &Src : Sources) {
4239 EVT SrcVT = Src.ShuffleVec.getValueType();
4241 if (SrcVT.getSizeInBits() == VT.getSizeInBits())
4244 // This stage of the search produces a source with the same element type as
4245 // the original, but with a total width matching the BUILD_VECTOR output.
4246 EVT EltVT = SrcVT.getVectorElementType();
4247 EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT,
4248 VT.getSizeInBits() / EltVT.getSizeInBits());
4250 if (SrcVT.getSizeInBits() < VT.getSizeInBits()) {
4251 assert(2 * SrcVT.getSizeInBits() == VT.getSizeInBits());
4252 // We can pad out the smaller vector for free, so if it's part of a
4255 DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, Src.ShuffleVec,
4256 DAG.getUNDEF(Src.ShuffleVec.getValueType()));
4260 assert(SrcVT.getSizeInBits() == 2 * VT.getSizeInBits());
4262 if (Src.MaxElt - Src.MinElt >= NumElts) {
4263 // Span too large for a VEXT to cope
4267 if (Src.MinElt >= NumElts) {
4268 // The extraction can just take the second half
4270 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4271 DAG.getIntPtrConstant(NumElts));
4272 Src.WindowBase = -NumElts;
4273 } else if (Src.MaxElt < NumElts) {
4274 // The extraction can just take the first half
4275 Src.ShuffleVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT,
4276 Src.ShuffleVec, DAG.getIntPtrConstant(0));
4278 // An actual VEXT is needed
4279 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT,
4280 Src.ShuffleVec, DAG.getIntPtrConstant(0));
4282 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
4283 DAG.getIntPtrConstant(NumElts));
4284 unsigned Imm = Src.MinElt * getExtFactor(VEXTSrc1);
4286 Src.ShuffleVec = DAG.getNode(AArch64ISD::EXT, dl, DestVT, VEXTSrc1,
4287 VEXTSrc2, DAG.getConstant(Imm, MVT::i32));
4288 Src.WindowBase = -Src.MinElt;
4292 // Another possible incompatibility occurs from the vector element types. We
4293 // can fix this by bitcasting the source vectors to the same type we intend
4295 for (auto &Src : Sources) {
4296 EVT SrcEltTy = Src.ShuffleVec.getValueType().getVectorElementType();
4297 if (SrcEltTy == SmallestEltTy)
4299 assert(ShuffleVT.getVectorElementType() == SmallestEltTy);
4300 Src.ShuffleVec = DAG.getNode(ISD::BITCAST, dl, ShuffleVT, Src.ShuffleVec);
4301 Src.WindowScale = SrcEltTy.getSizeInBits() / SmallestEltTy.getSizeInBits();
4302 Src.WindowBase *= Src.WindowScale;
4305 // Final sanity check before we try to actually produce a shuffle.
4307 for (auto Src : Sources)
4308 assert(Src.ShuffleVec.getValueType() == ShuffleVT);
4311 // The stars all align, our next step is to produce the mask for the shuffle.
4312 SmallVector<int, 8> Mask(ShuffleVT.getVectorNumElements(), -1);
4313 int BitsPerShuffleLane = ShuffleVT.getVectorElementType().getSizeInBits();
4314 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
4315 SDValue Entry = Op.getOperand(i);
4316 if (Entry.getOpcode() == ISD::UNDEF)
4319 auto Src = std::find(Sources.begin(), Sources.end(), Entry.getOperand(0));
4320 int EltNo = cast<ConstantSDNode>(Entry.getOperand(1))->getSExtValue();
4322 // EXTRACT_VECTOR_ELT performs an implicit any_ext; BUILD_VECTOR an implicit
4323 // trunc. So only std::min(SrcBits, DestBits) actually get defined in this
4325 EVT OrigEltTy = Entry.getOperand(0).getValueType().getVectorElementType();
4326 int BitsDefined = std::min(OrigEltTy.getSizeInBits(),
4327 VT.getVectorElementType().getSizeInBits());
4328 int LanesDefined = BitsDefined / BitsPerShuffleLane;
4330 // This source is expected to fill ResMultiplier lanes of the final shuffle,
4331 // starting at the appropriate offset.
4332 int *LaneMask = &Mask[i * ResMultiplier];
4334 int ExtractBase = EltNo * Src->WindowScale + Src->WindowBase;
4335 ExtractBase += NumElts * (Src - Sources.begin());
4336 for (int j = 0; j < LanesDefined; ++j)
4337 LaneMask[j] = ExtractBase + j;
4340 // Final check before we try to produce nonsense...
4341 if (!isShuffleMaskLegal(Mask, ShuffleVT))
4344 SDValue ShuffleOps[] = { DAG.getUNDEF(ShuffleVT), DAG.getUNDEF(ShuffleVT) };
4345 for (unsigned i = 0; i < Sources.size(); ++i)
4346 ShuffleOps[i] = Sources[i].ShuffleVec;
4348 SDValue Shuffle = DAG.getVectorShuffle(ShuffleVT, dl, ShuffleOps[0],
4349 ShuffleOps[1], &Mask[0]);
4350 return DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
4353 // check if an EXT instruction can handle the shuffle mask when the
4354 // vector sources of the shuffle are the same.
4355 static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4356 unsigned NumElts = VT.getVectorNumElements();
4358 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4364 // If this is a VEXT shuffle, the immediate value is the index of the first
4365 // element. The other shuffle indices must be the successive elements after
4367 unsigned ExpectedElt = Imm;
4368 for (unsigned i = 1; i < NumElts; ++i) {
4369 // Increment the expected index. If it wraps around, just follow it
4370 // back to index zero and keep going.
4372 if (ExpectedElt == NumElts)
4376 continue; // ignore UNDEF indices
4377 if (ExpectedElt != static_cast<unsigned>(M[i]))
4384 // check if an EXT instruction can handle the shuffle mask when the
4385 // vector sources of the shuffle are different.
4386 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
4388 // Look for the first non-undef element.
4389 const int *FirstRealElt = std::find_if(M.begin(), M.end(),
4390 [](int Elt) {return Elt >= 0;});
4392 // Benefit form APInt to handle overflow when calculating expected element.
4393 unsigned NumElts = VT.getVectorNumElements();
4394 unsigned MaskBits = APInt(32, NumElts * 2).logBase2();
4395 APInt ExpectedElt = APInt(MaskBits, *FirstRealElt + 1);
4396 // The following shuffle indices must be the successive elements after the
4397 // first real element.
4398 const int *FirstWrongElt = std::find_if(FirstRealElt + 1, M.end(),
4399 [&](int Elt) {return Elt != ExpectedElt++ && Elt != -1;});
4400 if (FirstWrongElt != M.end())
4403 // The index of an EXT is the first element if it is not UNDEF.
4404 // Watch out for the beginning UNDEFs. The EXT index should be the expected
4405 // value of the first element. E.g.
4406 // <-1, -1, 3, ...> is treated as <1, 2, 3, ...>.
4407 // <-1, -1, 0, 1, ...> is treated as <2*NumElts-2, 2*NumElts-1, 0, 1, ...>.
4408 // ExpectedElt is the last mask index plus 1.
4409 Imm = ExpectedElt.getZExtValue();
4411 // There are two difference cases requiring to reverse input vectors.
4412 // For example, for vector <4 x i32> we have the following cases,
4413 // Case 1: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, -1, 0>)
4414 // Case 2: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, 7, 0>)
4415 // For both cases, we finally use mask <5, 6, 7, 0>, which requires
4416 // to reverse two input vectors.
4425 /// isREVMask - Check if a vector shuffle corresponds to a REV
4426 /// instruction with the specified blocksize. (The order of the elements
4427 /// within each block of the vector is reversed.)
4428 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4429 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
4430 "Only possible block sizes for REV are: 16, 32, 64");
4432 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4436 unsigned NumElts = VT.getVectorNumElements();
4437 unsigned BlockElts = M[0] + 1;
4438 // If the first shuffle index is UNDEF, be optimistic.
4440 BlockElts = BlockSize / EltSz;
4442 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4445 for (unsigned i = 0; i < NumElts; ++i) {
4447 continue; // ignore UNDEF indices
4448 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
4455 static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4456 unsigned NumElts = VT.getVectorNumElements();
4457 WhichResult = (M[0] == 0 ? 0 : 1);
4458 unsigned Idx = WhichResult * NumElts / 2;
4459 for (unsigned i = 0; i != NumElts; i += 2) {
4460 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4461 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
4469 static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4470 unsigned NumElts = VT.getVectorNumElements();
4471 WhichResult = (M[0] == 0 ? 0 : 1);
4472 for (unsigned i = 0; i != NumElts; ++i) {
4474 continue; // ignore UNDEF indices
4475 if ((unsigned)M[i] != 2 * i + WhichResult)
4482 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4483 unsigned NumElts = VT.getVectorNumElements();
4484 WhichResult = (M[0] == 0 ? 0 : 1);
4485 for (unsigned i = 0; i < NumElts; i += 2) {
4486 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4487 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + NumElts + WhichResult))
4493 /// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of
4494 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4495 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4496 static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4497 unsigned NumElts = VT.getVectorNumElements();
4498 WhichResult = (M[0] == 0 ? 0 : 1);
4499 unsigned Idx = WhichResult * NumElts / 2;
4500 for (unsigned i = 0; i != NumElts; i += 2) {
4501 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4502 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx))
4510 /// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of
4511 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4512 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4513 static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4514 unsigned Half = VT.getVectorNumElements() / 2;
4515 WhichResult = (M[0] == 0 ? 0 : 1);
4516 for (unsigned j = 0; j != 2; ++j) {
4517 unsigned Idx = WhichResult;
4518 for (unsigned i = 0; i != Half; ++i) {
4519 int MIdx = M[i + j * Half];
4520 if (MIdx >= 0 && (unsigned)MIdx != Idx)
4529 /// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of
4530 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4531 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4532 static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4533 unsigned NumElts = VT.getVectorNumElements();
4534 WhichResult = (M[0] == 0 ? 0 : 1);
4535 for (unsigned i = 0; i < NumElts; i += 2) {
4536 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4537 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult))
4543 static bool isINSMask(ArrayRef<int> M, int NumInputElements,
4544 bool &DstIsLeft, int &Anomaly) {
4545 if (M.size() != static_cast<size_t>(NumInputElements))
4548 int NumLHSMatch = 0, NumRHSMatch = 0;
4549 int LastLHSMismatch = -1, LastRHSMismatch = -1;
4551 for (int i = 0; i < NumInputElements; ++i) {
4561 LastLHSMismatch = i;
4563 if (M[i] == i + NumInputElements)
4566 LastRHSMismatch = i;
4569 if (NumLHSMatch == NumInputElements - 1) {
4571 Anomaly = LastLHSMismatch;
4573 } else if (NumRHSMatch == NumInputElements - 1) {
4575 Anomaly = LastRHSMismatch;
4582 static bool isConcatMask(ArrayRef<int> Mask, EVT VT, bool SplitLHS) {
4583 if (VT.getSizeInBits() != 128)
4586 unsigned NumElts = VT.getVectorNumElements();
4588 for (int I = 0, E = NumElts / 2; I != E; I++) {
4593 int Offset = NumElts / 2;
4594 for (int I = NumElts / 2, E = NumElts; I != E; I++) {
4595 if (Mask[I] != I + SplitLHS * Offset)
4602 static SDValue tryFormConcatFromShuffle(SDValue Op, SelectionDAG &DAG) {
4604 EVT VT = Op.getValueType();
4605 SDValue V0 = Op.getOperand(0);
4606 SDValue V1 = Op.getOperand(1);
4607 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op)->getMask();
4609 if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() ||
4610 VT.getVectorElementType() != V1.getValueType().getVectorElementType())
4613 bool SplitV0 = V0.getValueType().getSizeInBits() == 128;
4615 if (!isConcatMask(Mask, VT, SplitV0))
4618 EVT CastVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
4619 VT.getVectorNumElements() / 2);
4621 V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0,
4622 DAG.getConstant(0, MVT::i64));
4624 if (V1.getValueType().getSizeInBits() == 128) {
4625 V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1,
4626 DAG.getConstant(0, MVT::i64));
4628 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1);
4631 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4632 /// the specified operations to build the shuffle.
4633 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4634 SDValue RHS, SelectionDAG &DAG,
4636 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4637 unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1);
4638 unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1);
4641 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4650 OP_VUZPL, // VUZP, left result
4651 OP_VUZPR, // VUZP, right result
4652 OP_VZIPL, // VZIP, left result
4653 OP_VZIPR, // VZIP, right result
4654 OP_VTRNL, // VTRN, left result
4655 OP_VTRNR // VTRN, right result
4658 if (OpNum == OP_COPY) {
4659 if (LHSID == (1 * 9 + 2) * 9 + 3)
4661 assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!");
4665 SDValue OpLHS, OpRHS;
4666 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4667 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4668 EVT VT = OpLHS.getValueType();
4672 llvm_unreachable("Unknown shuffle opcode!");
4674 // VREV divides the vector in half and swaps within the half.
4675 if (VT.getVectorElementType() == MVT::i32 ||
4676 VT.getVectorElementType() == MVT::f32)
4677 return DAG.getNode(AArch64ISD::REV64, dl, VT, OpLHS);
4678 // vrev <4 x i16> -> REV32
4679 if (VT.getVectorElementType() == MVT::i16)
4680 return DAG.getNode(AArch64ISD::REV32, dl, VT, OpLHS);
4681 // vrev <4 x i8> -> REV16
4682 assert(VT.getVectorElementType() == MVT::i8);
4683 return DAG.getNode(AArch64ISD::REV16, dl, VT, OpLHS);
4688 EVT EltTy = VT.getVectorElementType();
4690 if (EltTy == MVT::i8)
4691 Opcode = AArch64ISD::DUPLANE8;
4692 else if (EltTy == MVT::i16)
4693 Opcode = AArch64ISD::DUPLANE16;
4694 else if (EltTy == MVT::i32 || EltTy == MVT::f32)
4695 Opcode = AArch64ISD::DUPLANE32;
4696 else if (EltTy == MVT::i64 || EltTy == MVT::f64)
4697 Opcode = AArch64ISD::DUPLANE64;
4699 llvm_unreachable("Invalid vector element type?");
4701 if (VT.getSizeInBits() == 64)
4702 OpLHS = WidenVector(OpLHS, DAG);
4703 SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, MVT::i64);
4704 return DAG.getNode(Opcode, dl, VT, OpLHS, Lane);
4709 unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
4710 return DAG.getNode(AArch64ISD::EXT, dl, VT, OpLHS, OpRHS,
4711 DAG.getConstant(Imm, MVT::i32));
4714 return DAG.getNode(AArch64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS,
4717 return DAG.getNode(AArch64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS,
4720 return DAG.getNode(AArch64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS,
4723 return DAG.getNode(AArch64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS,
4726 return DAG.getNode(AArch64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS,
4729 return DAG.getNode(AArch64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS,
4734 static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask,
4735 SelectionDAG &DAG) {
4736 // Check to see if we can use the TBL instruction.
4737 SDValue V1 = Op.getOperand(0);
4738 SDValue V2 = Op.getOperand(1);
4741 EVT EltVT = Op.getValueType().getVectorElementType();
4742 unsigned BytesPerElt = EltVT.getSizeInBits() / 8;
4744 SmallVector<SDValue, 8> TBLMask;
4745 for (int Val : ShuffleMask) {
4746 for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
4747 unsigned Offset = Byte + Val * BytesPerElt;
4748 TBLMask.push_back(DAG.getConstant(Offset, MVT::i32));
4752 MVT IndexVT = MVT::v8i8;
4753 unsigned IndexLen = 8;
4754 if (Op.getValueType().getSizeInBits() == 128) {
4755 IndexVT = MVT::v16i8;
4759 SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
4760 SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);
4763 if (V2.getNode()->getOpcode() == ISD::UNDEF) {
4765 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
4766 Shuffle = DAG.getNode(
4767 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4768 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, MVT::i32), V1Cst,
4769 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4770 makeArrayRef(TBLMask.data(), IndexLen)));
4772 if (IndexLen == 8) {
4773 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
4774 Shuffle = DAG.getNode(
4775 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4776 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, MVT::i32), V1Cst,
4777 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4778 makeArrayRef(TBLMask.data(), IndexLen)));
4780 // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
4781 // cannot currently represent the register constraints on the input
4783 // Shuffle = DAG.getNode(AArch64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
4784 // DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4785 // &TBLMask[0], IndexLen));
4786 Shuffle = DAG.getNode(
4787 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4788 DAG.getConstant(Intrinsic::aarch64_neon_tbl2, MVT::i32), V1Cst, V2Cst,
4789 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4790 makeArrayRef(TBLMask.data(), IndexLen)));
4793 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
4796 static unsigned getDUPLANEOp(EVT EltType) {
4797 if (EltType == MVT::i8)
4798 return AArch64ISD::DUPLANE8;
4799 if (EltType == MVT::i16)
4800 return AArch64ISD::DUPLANE16;
4801 if (EltType == MVT::i32 || EltType == MVT::f32)
4802 return AArch64ISD::DUPLANE32;
4803 if (EltType == MVT::i64 || EltType == MVT::f64)
4804 return AArch64ISD::DUPLANE64;
4806 llvm_unreachable("Invalid vector element type?");
4809 SDValue AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4810 SelectionDAG &DAG) const {
4812 EVT VT = Op.getValueType();
4814 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4816 // Convert shuffles that are directly supported on NEON to target-specific
4817 // DAG nodes, instead of keeping them as shuffles and matching them again
4818 // during code selection. This is more efficient and avoids the possibility
4819 // of inconsistencies between legalization and selection.
4820 ArrayRef<int> ShuffleMask = SVN->getMask();
4822 SDValue V1 = Op.getOperand(0);
4823 SDValue V2 = Op.getOperand(1);
4825 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0],
4826 V1.getValueType().getSimpleVT())) {
4827 int Lane = SVN->getSplatIndex();
4828 // If this is undef splat, generate it via "just" vdup, if possible.
4832 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
4833 return DAG.getNode(AArch64ISD::DUP, dl, V1.getValueType(),
4835 // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
4836 // constant. If so, we can just reference the lane's definition directly.
4837 if (V1.getOpcode() == ISD::BUILD_VECTOR &&
4838 !isa<ConstantSDNode>(V1.getOperand(Lane)))
4839 return DAG.getNode(AArch64ISD::DUP, dl, VT, V1.getOperand(Lane));
4841 // Otherwise, duplicate from the lane of the input vector.
4842 unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
4844 // SelectionDAGBuilder may have "helpfully" already extracted or conatenated
4845 // to make a vector of the same size as this SHUFFLE. We can ignore the
4846 // extract entirely, and canonicalise the concat using WidenVector.
4847 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
4848 Lane += cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
4849 V1 = V1.getOperand(0);
4850 } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
4851 unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
4852 Lane -= Idx * VT.getVectorNumElements() / 2;
4853 V1 = WidenVector(V1.getOperand(Idx), DAG);
4854 } else if (VT.getSizeInBits() == 64)
4855 V1 = WidenVector(V1, DAG);
4857 return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, MVT::i64));
4860 if (isREVMask(ShuffleMask, VT, 64))
4861 return DAG.getNode(AArch64ISD::REV64, dl, V1.getValueType(), V1, V2);
4862 if (isREVMask(ShuffleMask, VT, 32))
4863 return DAG.getNode(AArch64ISD::REV32, dl, V1.getValueType(), V1, V2);
4864 if (isREVMask(ShuffleMask, VT, 16))
4865 return DAG.getNode(AArch64ISD::REV16, dl, V1.getValueType(), V1, V2);
4867 bool ReverseEXT = false;
4869 if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
4872 Imm *= getExtFactor(V1);
4873 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V2,
4874 DAG.getConstant(Imm, MVT::i32));
4875 } else if (V2->getOpcode() == ISD::UNDEF &&
4876 isSingletonEXTMask(ShuffleMask, VT, Imm)) {
4877 Imm *= getExtFactor(V1);
4878 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V1,
4879 DAG.getConstant(Imm, MVT::i32));
4882 unsigned WhichResult;
4883 if (isZIPMask(ShuffleMask, VT, WhichResult)) {
4884 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
4885 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4887 if (isUZPMask(ShuffleMask, VT, WhichResult)) {
4888 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
4889 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4891 if (isTRNMask(ShuffleMask, VT, WhichResult)) {
4892 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
4893 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4896 if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4897 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
4898 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4900 if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4901 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
4902 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4904 if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4905 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
4906 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4909 SDValue Concat = tryFormConcatFromShuffle(Op, DAG);
4910 if (Concat.getNode())
4915 int NumInputElements = V1.getValueType().getVectorNumElements();
4916 if (isINSMask(ShuffleMask, NumInputElements, DstIsLeft, Anomaly)) {
4917 SDValue DstVec = DstIsLeft ? V1 : V2;
4918 SDValue DstLaneV = DAG.getConstant(Anomaly, MVT::i64);
4920 SDValue SrcVec = V1;
4921 int SrcLane = ShuffleMask[Anomaly];
4922 if (SrcLane >= NumInputElements) {
4924 SrcLane -= VT.getVectorNumElements();
4926 SDValue SrcLaneV = DAG.getConstant(SrcLane, MVT::i64);
4928 EVT ScalarVT = VT.getVectorElementType();
4929 if (ScalarVT.getSizeInBits() < 32)
4930 ScalarVT = MVT::i32;
4933 ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
4934 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ScalarVT, SrcVec, SrcLaneV),
4938 // If the shuffle is not directly supported and it has 4 elements, use
4939 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4940 unsigned NumElts = VT.getVectorNumElements();
4942 unsigned PFIndexes[4];
4943 for (unsigned i = 0; i != 4; ++i) {
4944 if (ShuffleMask[i] < 0)
4947 PFIndexes[i] = ShuffleMask[i];
4950 // Compute the index in the perfect shuffle table.
4951 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
4952 PFIndexes[2] * 9 + PFIndexes[3];
4953 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4954 unsigned Cost = (PFEntry >> 30);
4957 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4960 return GenerateTBL(Op, ShuffleMask, DAG);
4963 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
4965 EVT VT = BVN->getValueType(0);
4966 APInt SplatBits, SplatUndef;
4967 unsigned SplatBitSize;
4969 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4970 unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;
4972 for (unsigned i = 0; i < NumSplats; ++i) {
4973 CnstBits <<= SplatBitSize;
4974 UndefBits <<= SplatBitSize;
4975 CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits());
4976 UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits());
4985 SDValue AArch64TargetLowering::LowerVectorAND(SDValue Op,
4986 SelectionDAG &DAG) const {
4987 BuildVectorSDNode *BVN =
4988 dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
4989 SDValue LHS = Op.getOperand(0);
4991 EVT VT = Op.getValueType();
4996 APInt CnstBits(VT.getSizeInBits(), 0);
4997 APInt UndefBits(VT.getSizeInBits(), 0);
4998 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
4999 // We only have BIC vector immediate instruction, which is and-not.
5000 CnstBits = ~CnstBits;
5002 // We make use of a little bit of goto ickiness in order to avoid having to
5003 // duplicate the immediate matching logic for the undef toggled case.
5004 bool SecondTry = false;
5007 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5008 CnstBits = CnstBits.zextOrTrunc(64);
5009 uint64_t CnstVal = CnstBits.getZExtValue();
5011 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5012 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5013 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5014 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5015 DAG.getConstant(CnstVal, MVT::i32),
5016 DAG.getConstant(0, MVT::i32));
5017 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5020 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5021 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5022 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5023 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5024 DAG.getConstant(CnstVal, MVT::i32),
5025 DAG.getConstant(8, MVT::i32));
5026 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5029 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5030 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5031 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5032 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5033 DAG.getConstant(CnstVal, MVT::i32),
5034 DAG.getConstant(16, MVT::i32));
5035 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5038 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5039 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5040 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5041 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5042 DAG.getConstant(CnstVal, MVT::i32),
5043 DAG.getConstant(24, MVT::i32));
5044 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5047 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5048 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5049 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5050 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5051 DAG.getConstant(CnstVal, MVT::i32),
5052 DAG.getConstant(0, MVT::i32));
5053 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5056 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5057 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5058 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5059 SDValue Mov = DAG.getNode(AArch64ISD::BICi, dl, MovTy, LHS,
5060 DAG.getConstant(CnstVal, MVT::i32),
5061 DAG.getConstant(8, MVT::i32));
5062 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5069 CnstBits = ~UndefBits;
5073 // We can always fall back to a non-immediate AND.
5078 // Specialized code to quickly find if PotentialBVec is a BuildVector that
5079 // consists of only the same constant int value, returned in reference arg
5081 static bool isAllConstantBuildVector(const SDValue &PotentialBVec,
5082 uint64_t &ConstVal) {
5083 BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec);
5086 ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0));
5089 EVT VT = Bvec->getValueType(0);
5090 unsigned NumElts = VT.getVectorNumElements();
5091 for (unsigned i = 1; i < NumElts; ++i)
5092 if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt)
5094 ConstVal = FirstElt->getZExtValue();
5098 static unsigned getIntrinsicID(const SDNode *N) {
5099 unsigned Opcode = N->getOpcode();
5102 return Intrinsic::not_intrinsic;
5103 case ISD::INTRINSIC_WO_CHAIN: {
5104 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
5105 if (IID < Intrinsic::num_intrinsics)
5107 return Intrinsic::not_intrinsic;
5112 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
5113 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
5114 // BUILD_VECTORs with constant element C1, C2 is a constant, and C1 == ~C2.
5115 // Also, logical shift right -> sri, with the same structure.
5116 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
5117 EVT VT = N->getValueType(0);
5124 // Is the first op an AND?
5125 const SDValue And = N->getOperand(0);
5126 if (And.getOpcode() != ISD::AND)
5129 // Is the second op an shl or lshr?
5130 SDValue Shift = N->getOperand(1);
5131 // This will have been turned into: AArch64ISD::VSHL vector, #shift
5132 // or AArch64ISD::VLSHR vector, #shift
5133 unsigned ShiftOpc = Shift.getOpcode();
5134 if ((ShiftOpc != AArch64ISD::VSHL && ShiftOpc != AArch64ISD::VLSHR))
5136 bool IsShiftRight = ShiftOpc == AArch64ISD::VLSHR;
5138 // Is the shift amount constant?
5139 ConstantSDNode *C2node = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
5143 // Is the and mask vector all constant?
5145 if (!isAllConstantBuildVector(And.getOperand(1), C1))
5148 // Is C1 == ~C2, taking into account how much one can shift elements of a
5150 uint64_t C2 = C2node->getZExtValue();
5151 unsigned ElemSizeInBits = VT.getVectorElementType().getSizeInBits();
5152 if (C2 > ElemSizeInBits)
5154 unsigned ElemMask = (1 << ElemSizeInBits) - 1;
5155 if ((C1 & ElemMask) != (~C2 & ElemMask))
5158 SDValue X = And.getOperand(0);
5159 SDValue Y = Shift.getOperand(0);
5162 IsShiftRight ? Intrinsic::aarch64_neon_vsri : Intrinsic::aarch64_neon_vsli;
5164 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5165 DAG.getConstant(Intrin, MVT::i32), X, Y, Shift.getOperand(1));
5167 DEBUG(dbgs() << "aarch64-lower: transformed: \n");
5168 DEBUG(N->dump(&DAG));
5169 DEBUG(dbgs() << "into: \n");
5170 DEBUG(ResultSLI->dump(&DAG));
5176 SDValue AArch64TargetLowering::LowerVectorOR(SDValue Op,
5177 SelectionDAG &DAG) const {
5178 // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
5179 if (EnableAArch64SlrGeneration) {
5180 SDValue Res = tryLowerToSLI(Op.getNode(), DAG);
5185 BuildVectorSDNode *BVN =
5186 dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode());
5187 SDValue LHS = Op.getOperand(1);
5189 EVT VT = Op.getValueType();
5191 // OR commutes, so try swapping the operands.
5193 LHS = Op.getOperand(0);
5194 BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5199 APInt CnstBits(VT.getSizeInBits(), 0);
5200 APInt UndefBits(VT.getSizeInBits(), 0);
5201 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5202 // We make use of a little bit of goto ickiness in order to avoid having to
5203 // duplicate the immediate matching logic for the undef toggled case.
5204 bool SecondTry = false;
5207 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5208 CnstBits = CnstBits.zextOrTrunc(64);
5209 uint64_t CnstVal = CnstBits.getZExtValue();
5211 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5212 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5213 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5214 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5215 DAG.getConstant(CnstVal, MVT::i32),
5216 DAG.getConstant(0, MVT::i32));
5217 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5220 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5221 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5222 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5223 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5224 DAG.getConstant(CnstVal, MVT::i32),
5225 DAG.getConstant(8, MVT::i32));
5226 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5229 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5230 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5231 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5232 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5233 DAG.getConstant(CnstVal, MVT::i32),
5234 DAG.getConstant(16, MVT::i32));
5235 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5238 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5239 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5240 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5241 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5242 DAG.getConstant(CnstVal, MVT::i32),
5243 DAG.getConstant(24, MVT::i32));
5244 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5247 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5248 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5249 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5250 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5251 DAG.getConstant(CnstVal, MVT::i32),
5252 DAG.getConstant(0, MVT::i32));
5253 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5256 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5257 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5258 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5259 SDValue Mov = DAG.getNode(AArch64ISD::ORRi, dl, MovTy, LHS,
5260 DAG.getConstant(CnstVal, MVT::i32),
5261 DAG.getConstant(8, MVT::i32));
5262 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5269 CnstBits = UndefBits;
5273 // We can always fall back to a non-immediate OR.
5278 // Normalize the operands of BUILD_VECTOR. The value of constant operands will
5279 // be truncated to fit element width.
5280 static SDValue NormalizeBuildVector(SDValue Op,
5281 SelectionDAG &DAG) {
5282 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
5284 EVT VT = Op.getValueType();
5285 EVT EltTy= VT.getVectorElementType();
5287 if (EltTy.isFloatingPoint() || EltTy.getSizeInBits() > 16)
5290 SmallVector<SDValue, 16> Ops;
5291 for (unsigned I = 0, E = VT.getVectorNumElements(); I != E; ++I) {
5292 SDValue Lane = Op.getOperand(I);
5293 if (Lane.getOpcode() == ISD::Constant) {
5294 APInt LowBits(EltTy.getSizeInBits(),
5295 cast<ConstantSDNode>(Lane)->getZExtValue());
5296 Lane = DAG.getConstant(LowBits.getZExtValue(), MVT::i32);
5298 Ops.push_back(Lane);
5300 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5303 SDValue AArch64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
5304 SelectionDAG &DAG) const {
5306 EVT VT = Op.getValueType();
5307 Op = NormalizeBuildVector(Op, DAG);
5308 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5310 APInt CnstBits(VT.getSizeInBits(), 0);
5311 APInt UndefBits(VT.getSizeInBits(), 0);
5312 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5313 // We make use of a little bit of goto ickiness in order to avoid having to
5314 // duplicate the immediate matching logic for the undef toggled case.
5315 bool SecondTry = false;
5318 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5319 CnstBits = CnstBits.zextOrTrunc(64);
5320 uint64_t CnstVal = CnstBits.getZExtValue();
5322 // Certain magic vector constants (used to express things like NOT
5323 // and NEG) are passed through unmodified. This allows codegen patterns
5324 // for these operations to match. Special-purpose patterns will lower
5325 // these immediates to MOVIs if it proves necessary.
5326 if (VT.isInteger() && (CnstVal == 0 || CnstVal == ~0ULL))
5329 // The many faces of MOVI...
5330 if (AArch64_AM::isAdvSIMDModImmType10(CnstVal)) {
5331 CnstVal = AArch64_AM::encodeAdvSIMDModImmType10(CnstVal);
5332 if (VT.getSizeInBits() == 128) {
5333 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::v2i64,
5334 DAG.getConstant(CnstVal, MVT::i32));
5335 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5338 // Support the V64 version via subregister insertion.
5339 SDValue Mov = DAG.getNode(AArch64ISD::MOVIedit, dl, MVT::f64,
5340 DAG.getConstant(CnstVal, MVT::i32));
5341 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5344 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5345 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5346 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5347 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5348 DAG.getConstant(CnstVal, MVT::i32),
5349 DAG.getConstant(0, MVT::i32));
5350 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5353 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5354 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5355 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5356 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5357 DAG.getConstant(CnstVal, MVT::i32),
5358 DAG.getConstant(8, MVT::i32));
5359 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5362 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5363 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5364 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5365 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5366 DAG.getConstant(CnstVal, MVT::i32),
5367 DAG.getConstant(16, MVT::i32));
5368 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5371 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5372 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5373 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5374 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5375 DAG.getConstant(CnstVal, MVT::i32),
5376 DAG.getConstant(24, MVT::i32));
5377 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5380 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5381 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5382 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5383 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5384 DAG.getConstant(CnstVal, MVT::i32),
5385 DAG.getConstant(0, MVT::i32));
5386 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5389 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5390 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5391 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5392 SDValue Mov = DAG.getNode(AArch64ISD::MOVIshift, dl, MovTy,
5393 DAG.getConstant(CnstVal, MVT::i32),
5394 DAG.getConstant(8, MVT::i32));
5395 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5398 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
5399 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
5400 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5401 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
5402 DAG.getConstant(CnstVal, MVT::i32),
5403 DAG.getConstant(264, MVT::i32));
5404 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5407 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
5408 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
5409 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5410 SDValue Mov = DAG.getNode(AArch64ISD::MOVImsl, dl, MovTy,
5411 DAG.getConstant(CnstVal, MVT::i32),
5412 DAG.getConstant(272, MVT::i32));
5413 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5416 if (AArch64_AM::isAdvSIMDModImmType9(CnstVal)) {
5417 CnstVal = AArch64_AM::encodeAdvSIMDModImmType9(CnstVal);
5418 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;
5419 SDValue Mov = DAG.getNode(AArch64ISD::MOVI, dl, MovTy,
5420 DAG.getConstant(CnstVal, MVT::i32));
5421 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5424 // The few faces of FMOV...
5425 if (AArch64_AM::isAdvSIMDModImmType11(CnstVal)) {
5426 CnstVal = AArch64_AM::encodeAdvSIMDModImmType11(CnstVal);
5427 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4f32 : MVT::v2f32;
5428 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MovTy,
5429 DAG.getConstant(CnstVal, MVT::i32));
5430 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5433 if (AArch64_AM::isAdvSIMDModImmType12(CnstVal) &&
5434 VT.getSizeInBits() == 128) {
5435 CnstVal = AArch64_AM::encodeAdvSIMDModImmType12(CnstVal);
5436 SDValue Mov = DAG.getNode(AArch64ISD::FMOV, dl, MVT::v2f64,
5437 DAG.getConstant(CnstVal, MVT::i32));
5438 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5441 // The many faces of MVNI...
5443 if (AArch64_AM::isAdvSIMDModImmType1(CnstVal)) {
5444 CnstVal = AArch64_AM::encodeAdvSIMDModImmType1(CnstVal);
5445 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5446 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5447 DAG.getConstant(CnstVal, MVT::i32),
5448 DAG.getConstant(0, MVT::i32));
5449 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5452 if (AArch64_AM::isAdvSIMDModImmType2(CnstVal)) {
5453 CnstVal = AArch64_AM::encodeAdvSIMDModImmType2(CnstVal);
5454 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5455 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5456 DAG.getConstant(CnstVal, MVT::i32),
5457 DAG.getConstant(8, MVT::i32));
5458 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5461 if (AArch64_AM::isAdvSIMDModImmType3(CnstVal)) {
5462 CnstVal = AArch64_AM::encodeAdvSIMDModImmType3(CnstVal);
5463 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5464 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5465 DAG.getConstant(CnstVal, MVT::i32),
5466 DAG.getConstant(16, MVT::i32));
5467 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5470 if (AArch64_AM::isAdvSIMDModImmType4(CnstVal)) {
5471 CnstVal = AArch64_AM::encodeAdvSIMDModImmType4(CnstVal);
5472 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5473 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5474 DAG.getConstant(CnstVal, MVT::i32),
5475 DAG.getConstant(24, MVT::i32));
5476 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5479 if (AArch64_AM::isAdvSIMDModImmType5(CnstVal)) {
5480 CnstVal = AArch64_AM::encodeAdvSIMDModImmType5(CnstVal);
5481 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5482 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5483 DAG.getConstant(CnstVal, MVT::i32),
5484 DAG.getConstant(0, MVT::i32));
5485 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5488 if (AArch64_AM::isAdvSIMDModImmType6(CnstVal)) {
5489 CnstVal = AArch64_AM::encodeAdvSIMDModImmType6(CnstVal);
5490 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5491 SDValue Mov = DAG.getNode(AArch64ISD::MVNIshift, dl, MovTy,
5492 DAG.getConstant(CnstVal, MVT::i32),
5493 DAG.getConstant(8, MVT::i32));
5494 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5497 if (AArch64_AM::isAdvSIMDModImmType7(CnstVal)) {
5498 CnstVal = AArch64_AM::encodeAdvSIMDModImmType7(CnstVal);
5499 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5500 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
5501 DAG.getConstant(CnstVal, MVT::i32),
5502 DAG.getConstant(264, MVT::i32));
5503 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5506 if (AArch64_AM::isAdvSIMDModImmType8(CnstVal)) {
5507 CnstVal = AArch64_AM::encodeAdvSIMDModImmType8(CnstVal);
5508 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5509 SDValue Mov = DAG.getNode(AArch64ISD::MVNImsl, dl, MovTy,
5510 DAG.getConstant(CnstVal, MVT::i32),
5511 DAG.getConstant(272, MVT::i32));
5512 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5519 CnstBits = UndefBits;
5524 // Scan through the operands to find some interesting properties we can
5526 // 1) If only one value is used, we can use a DUP, or
5527 // 2) if only the low element is not undef, we can just insert that, or
5528 // 3) if only one constant value is used (w/ some non-constant lanes),
5529 // we can splat the constant value into the whole vector then fill
5530 // in the non-constant lanes.
5531 // 4) FIXME: If different constant values are used, but we can intelligently
5532 // select the values we'll be overwriting for the non-constant
5533 // lanes such that we can directly materialize the vector
5534 // some other way (MOVI, e.g.), we can be sneaky.
5535 unsigned NumElts = VT.getVectorNumElements();
5536 bool isOnlyLowElement = true;
5537 bool usesOnlyOneValue = true;
5538 bool usesOnlyOneConstantValue = true;
5539 bool isConstant = true;
5540 unsigned NumConstantLanes = 0;
5542 SDValue ConstantValue;
5543 for (unsigned i = 0; i < NumElts; ++i) {
5544 SDValue V = Op.getOperand(i);
5545 if (V.getOpcode() == ISD::UNDEF)
5548 isOnlyLowElement = false;
5549 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5552 if (isa<ConstantSDNode>(V) || isa<ConstantFPSDNode>(V)) {
5554 if (!ConstantValue.getNode())
5556 else if (ConstantValue != V)
5557 usesOnlyOneConstantValue = false;
5560 if (!Value.getNode())
5562 else if (V != Value)
5563 usesOnlyOneValue = false;
5566 if (!Value.getNode())
5567 return DAG.getUNDEF(VT);
5569 if (isOnlyLowElement)
5570 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5572 // Use DUP for non-constant splats. For f32 constant splats, reduce to
5573 // i32 and try again.
5574 if (usesOnlyOneValue) {
5576 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5577 Value.getValueType() != VT)
5578 return DAG.getNode(AArch64ISD::DUP, dl, VT, Value);
5580 // This is actually a DUPLANExx operation, which keeps everything vectory.
5582 // DUPLANE works on 128-bit vectors, widen it if necessary.
5583 SDValue Lane = Value.getOperand(1);
5584 Value = Value.getOperand(0);
5585 if (Value.getValueType().getSizeInBits() == 64)
5586 Value = WidenVector(Value, DAG);
5588 unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
5589 return DAG.getNode(Opcode, dl, VT, Value, Lane);
5592 if (VT.getVectorElementType().isFloatingPoint()) {
5593 SmallVector<SDValue, 8> Ops;
5595 (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
5596 for (unsigned i = 0; i < NumElts; ++i)
5597 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, NewType, Op.getOperand(i)));
5598 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts);
5599 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5600 Val = LowerBUILD_VECTOR(Val, DAG);
5602 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5606 // If there was only one constant value used and for more than one lane,
5607 // start by splatting that value, then replace the non-constant lanes. This
5608 // is better than the default, which will perform a separate initialization
5610 if (NumConstantLanes > 0 && usesOnlyOneConstantValue) {
5611 SDValue Val = DAG.getNode(AArch64ISD::DUP, dl, VT, ConstantValue);
5612 // Now insert the non-constant lanes.
5613 for (unsigned i = 0; i < NumElts; ++i) {
5614 SDValue V = Op.getOperand(i);
5615 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5616 if (!isa<ConstantSDNode>(V) && !isa<ConstantFPSDNode>(V)) {
5617 // Note that type legalization likely mucked about with the VT of the
5618 // source operand, so we may have to convert it here before inserting.
5619 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, V, LaneIdx);
5625 // If all elements are constants and the case above didn't get hit, fall back
5626 // to the default expansion, which will generate a load from the constant
5631 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5633 SDValue shuffle = ReconstructShuffle(Op, DAG);
5634 if (shuffle != SDValue())
5638 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5639 // know the default expansion would otherwise fall back on something even
5640 // worse. For a vector with one or two non-undef values, that's
5641 // scalar_to_vector for the elements followed by a shuffle (provided the
5642 // shuffle is valid for the target) and materialization element by element
5643 // on the stack followed by a load for everything else.
5644 if (!isConstant && !usesOnlyOneValue) {
5645 SDValue Vec = DAG.getUNDEF(VT);
5646 SDValue Op0 = Op.getOperand(0);
5647 unsigned ElemSize = VT.getVectorElementType().getSizeInBits();
5649 // For 32 and 64 bit types, use INSERT_SUBREG for lane zero to
5650 // a) Avoid a RMW dependency on the full vector register, and
5651 // b) Allow the register coalescer to fold away the copy if the
5652 // value is already in an S or D register.
5653 if (Op0.getOpcode() != ISD::UNDEF && (ElemSize == 32 || ElemSize == 64)) {
5654 unsigned SubIdx = ElemSize == 32 ? AArch64::ssub : AArch64::dsub;
5656 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0,
5657 DAG.getTargetConstant(SubIdx, MVT::i32));
5658 Vec = SDValue(N, 0);
5661 for (; i < NumElts; ++i) {
5662 SDValue V = Op.getOperand(i);
5663 if (V.getOpcode() == ISD::UNDEF)
5665 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5666 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5671 // Just use the default expansion. We failed to find a better alternative.
5675 SDValue AArch64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
5676 SelectionDAG &DAG) const {
5677 assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!");
5679 // Check for non-constant or out of range lane.
5680 EVT VT = Op.getOperand(0).getValueType();
5681 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(2));
5682 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
5686 // Insertion/extraction are legal for V128 types.
5687 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
5688 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
5691 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
5692 VT != MVT::v1i64 && VT != MVT::v2f32)
5695 // For V64 types, we perform insertion by expanding the value
5696 // to a V128 type and perform the insertion on that.
5698 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
5699 EVT WideTy = WideVec.getValueType();
5701 SDValue Node = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideTy, WideVec,
5702 Op.getOperand(1), Op.getOperand(2));
5703 // Re-narrow the resultant vector.
5704 return NarrowVector(Node, DAG);
5708 AArch64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5709 SelectionDAG &DAG) const {
5710 assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!");
5712 // Check for non-constant or out of range lane.
5713 EVT VT = Op.getOperand(0).getValueType();
5714 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5715 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
5719 // Insertion/extraction are legal for V128 types.
5720 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
5721 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
5724 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
5725 VT != MVT::v1i64 && VT != MVT::v2f32)
5728 // For V64 types, we perform extraction by expanding the value
5729 // to a V128 type and perform the extraction on that.
5731 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
5732 EVT WideTy = WideVec.getValueType();
5734 EVT ExtrTy = WideTy.getVectorElementType();
5735 if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8)
5738 // For extractions, we just return the result directly.
5739 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec,
5743 SDValue AArch64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
5744 SelectionDAG &DAG) const {
5745 EVT VT = Op.getOperand(0).getValueType();
5751 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5754 unsigned Val = Cst->getZExtValue();
5756 unsigned Size = Op.getValueType().getSizeInBits();
5760 return DAG.getTargetExtractSubreg(AArch64::bsub, dl, Op.getValueType(),
5763 return DAG.getTargetExtractSubreg(AArch64::hsub, dl, Op.getValueType(),
5766 return DAG.getTargetExtractSubreg(AArch64::ssub, dl, Op.getValueType(),
5769 return DAG.getTargetExtractSubreg(AArch64::dsub, dl, Op.getValueType(),
5772 llvm_unreachable("Unexpected vector type in extract_subvector!");
5775 // If this is extracting the upper 64-bits of a 128-bit vector, we match
5777 if (Size == 64 && Val * VT.getVectorElementType().getSizeInBits() == 64)
5783 bool AArch64TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5785 if (VT.getVectorNumElements() == 4 &&
5786 (VT.is128BitVector() || VT.is64BitVector())) {
5787 unsigned PFIndexes[4];
5788 for (unsigned i = 0; i != 4; ++i) {
5792 PFIndexes[i] = M[i];
5795 // Compute the index in the perfect shuffle table.
5796 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
5797 PFIndexes[2] * 9 + PFIndexes[3];
5798 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5799 unsigned Cost = (PFEntry >> 30);
5807 unsigned DummyUnsigned;
5809 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || isREVMask(M, VT, 64) ||
5810 isREVMask(M, VT, 32) || isREVMask(M, VT, 16) ||
5811 isEXTMask(M, VT, DummyBool, DummyUnsigned) ||
5812 // isTBLMask(M, VT) || // FIXME: Port TBL support from ARM.
5813 isTRNMask(M, VT, DummyUnsigned) || isUZPMask(M, VT, DummyUnsigned) ||
5814 isZIPMask(M, VT, DummyUnsigned) ||
5815 isTRN_v_undef_Mask(M, VT, DummyUnsigned) ||
5816 isUZP_v_undef_Mask(M, VT, DummyUnsigned) ||
5817 isZIP_v_undef_Mask(M, VT, DummyUnsigned) ||
5818 isINSMask(M, VT.getVectorNumElements(), DummyBool, DummyInt) ||
5819 isConcatMask(M, VT, VT.getSizeInBits() == 128));
5822 /// getVShiftImm - Check if this is a valid build_vector for the immediate
5823 /// operand of a vector shift operation, where all the elements of the
5824 /// build_vector must have the same constant integer value.
5825 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
5826 // Ignore bit_converts.
5827 while (Op.getOpcode() == ISD::BITCAST)
5828 Op = Op.getOperand(0);
5829 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
5830 APInt SplatBits, SplatUndef;
5831 unsigned SplatBitSize;
5833 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
5834 HasAnyUndefs, ElementBits) ||
5835 SplatBitSize > ElementBits)
5837 Cnt = SplatBits.getSExtValue();
5841 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
5842 /// operand of a vector shift left operation. That value must be in the range:
5843 /// 0 <= Value < ElementBits for a left shift; or
5844 /// 0 <= Value <= ElementBits for a long left shift.
5845 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
5846 assert(VT.isVector() && "vector shift count is not a vector type");
5847 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5848 if (!getVShiftImm(Op, ElementBits, Cnt))
5850 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
5853 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
5854 /// operand of a vector shift right operation. For a shift opcode, the value
5855 /// is positive, but for an intrinsic the value count must be negative. The
5856 /// absolute value must be in the range:
5857 /// 1 <= |Value| <= ElementBits for a right shift; or
5858 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
5859 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
5861 assert(VT.isVector() && "vector shift count is not a vector type");
5862 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5863 if (!getVShiftImm(Op, ElementBits, Cnt))
5867 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
5870 SDValue AArch64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
5871 SelectionDAG &DAG) const {
5872 EVT VT = Op.getValueType();
5876 if (!Op.getOperand(1).getValueType().isVector())
5878 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5880 switch (Op.getOpcode()) {
5882 llvm_unreachable("unexpected shift opcode");
5885 if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
5886 return DAG.getNode(AArch64ISD::VSHL, SDLoc(Op), VT, Op.getOperand(0),
5887 DAG.getConstant(Cnt, MVT::i32));
5888 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5889 DAG.getConstant(Intrinsic::aarch64_neon_ushl, MVT::i32),
5890 Op.getOperand(0), Op.getOperand(1));
5893 // Right shift immediate
5894 if (isVShiftRImm(Op.getOperand(1), VT, false, false, Cnt) &&
5897 (Op.getOpcode() == ISD::SRA) ? AArch64ISD::VASHR : AArch64ISD::VLSHR;
5898 return DAG.getNode(Opc, SDLoc(Op), VT, Op.getOperand(0),
5899 DAG.getConstant(Cnt, MVT::i32));
5902 // Right shift register. Note, there is not a shift right register
5903 // instruction, but the shift left register instruction takes a signed
5904 // value, where negative numbers specify a right shift.
5905 unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::aarch64_neon_sshl
5906 : Intrinsic::aarch64_neon_ushl;
5907 // negate the shift amount
5908 SDValue NegShift = DAG.getNode(AArch64ISD::NEG, DL, VT, Op.getOperand(1));
5909 SDValue NegShiftLeft =
5910 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5911 DAG.getConstant(Opc, MVT::i32), Op.getOperand(0), NegShift);
5912 return NegShiftLeft;
5918 static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS,
5919 AArch64CC::CondCode CC, bool NoNans, EVT VT,
5920 SDLoc dl, SelectionDAG &DAG) {
5921 EVT SrcVT = LHS.getValueType();
5923 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
5924 APInt CnstBits(VT.getSizeInBits(), 0);
5925 APInt UndefBits(VT.getSizeInBits(), 0);
5926 bool IsCnst = BVN && resolveBuildVector(BVN, CnstBits, UndefBits);
5927 bool IsZero = IsCnst && (CnstBits == 0);
5929 if (SrcVT.getVectorElementType().isFloatingPoint()) {
5933 case AArch64CC::NE: {
5936 Fcmeq = DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
5938 Fcmeq = DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
5939 return DAG.getNode(AArch64ISD::NOT, dl, VT, Fcmeq);
5943 return DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS);
5944 return DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS);
5947 return DAG.getNode(AArch64ISD::FCMGEz, dl, VT, LHS);
5948 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, LHS, RHS);
5951 return DAG.getNode(AArch64ISD::FCMGTz, dl, VT, LHS);
5952 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, LHS, RHS);
5955 return DAG.getNode(AArch64ISD::FCMLEz, dl, VT, LHS);
5956 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, RHS, LHS);
5960 // If we ignore NaNs then we can use to the MI implementation.
5964 return DAG.getNode(AArch64ISD::FCMLTz, dl, VT, LHS);
5965 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, RHS, LHS);
5972 case AArch64CC::NE: {
5975 Cmeq = DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
5977 Cmeq = DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
5978 return DAG.getNode(AArch64ISD::NOT, dl, VT, Cmeq);
5982 return DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS);
5983 return DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS);
5986 return DAG.getNode(AArch64ISD::CMGEz, dl, VT, LHS);
5987 return DAG.getNode(AArch64ISD::CMGE, dl, VT, LHS, RHS);
5990 return DAG.getNode(AArch64ISD::CMGTz, dl, VT, LHS);
5991 return DAG.getNode(AArch64ISD::CMGT, dl, VT, LHS, RHS);
5994 return DAG.getNode(AArch64ISD::CMLEz, dl, VT, LHS);
5995 return DAG.getNode(AArch64ISD::CMGE, dl, VT, RHS, LHS);
5997 return DAG.getNode(AArch64ISD::CMHS, dl, VT, RHS, LHS);
5999 return DAG.getNode(AArch64ISD::CMHI, dl, VT, RHS, LHS);
6002 return DAG.getNode(AArch64ISD::CMLTz, dl, VT, LHS);
6003 return DAG.getNode(AArch64ISD::CMGT, dl, VT, RHS, LHS);
6005 return DAG.getNode(AArch64ISD::CMHI, dl, VT, LHS, RHS);
6007 return DAG.getNode(AArch64ISD::CMHS, dl, VT, LHS, RHS);
6011 SDValue AArch64TargetLowering::LowerVSETCC(SDValue Op,
6012 SelectionDAG &DAG) const {
6013 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6014 SDValue LHS = Op.getOperand(0);
6015 SDValue RHS = Op.getOperand(1);
6018 if (LHS.getValueType().getVectorElementType().isInteger()) {
6019 assert(LHS.getValueType() == RHS.getValueType());
6020 AArch64CC::CondCode AArch64CC = changeIntCCToAArch64CC(CC);
6021 return EmitVectorComparison(LHS, RHS, AArch64CC, false, Op.getValueType(),
6025 assert(LHS.getValueType().getVectorElementType() == MVT::f32 ||
6026 LHS.getValueType().getVectorElementType() == MVT::f64);
6028 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
6029 // clean. Some of them require two branches to implement.
6030 AArch64CC::CondCode CC1, CC2;
6032 changeVectorFPCCToAArch64CC(CC, CC1, CC2, ShouldInvert);
6034 bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath;
6036 EmitVectorComparison(LHS, RHS, CC1, NoNaNs, Op.getValueType(), dl, DAG);
6040 if (CC2 != AArch64CC::AL) {
6042 EmitVectorComparison(LHS, RHS, CC2, NoNaNs, Op.getValueType(), dl, DAG);
6043 if (!Cmp2.getNode())
6046 Cmp = DAG.getNode(ISD::OR, dl, Cmp.getValueType(), Cmp, Cmp2);
6050 return Cmp = DAG.getNOT(dl, Cmp, Cmp.getValueType());
6055 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
6056 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
6057 /// specified in the intrinsic calls.
6058 bool AArch64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
6060 unsigned Intrinsic) const {
6061 switch (Intrinsic) {
6062 case Intrinsic::aarch64_neon_ld2:
6063 case Intrinsic::aarch64_neon_ld3:
6064 case Intrinsic::aarch64_neon_ld4:
6065 case Intrinsic::aarch64_neon_ld1x2:
6066 case Intrinsic::aarch64_neon_ld1x3:
6067 case Intrinsic::aarch64_neon_ld1x4:
6068 case Intrinsic::aarch64_neon_ld2lane:
6069 case Intrinsic::aarch64_neon_ld3lane:
6070 case Intrinsic::aarch64_neon_ld4lane:
6071 case Intrinsic::aarch64_neon_ld2r:
6072 case Intrinsic::aarch64_neon_ld3r:
6073 case Intrinsic::aarch64_neon_ld4r: {
6074 Info.opc = ISD::INTRINSIC_W_CHAIN;
6075 // Conservatively set memVT to the entire set of vectors loaded.
6076 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
6077 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6078 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6081 Info.vol = false; // volatile loads with NEON intrinsics not supported
6082 Info.readMem = true;
6083 Info.writeMem = false;
6086 case Intrinsic::aarch64_neon_st2:
6087 case Intrinsic::aarch64_neon_st3:
6088 case Intrinsic::aarch64_neon_st4:
6089 case Intrinsic::aarch64_neon_st1x2:
6090 case Intrinsic::aarch64_neon_st1x3:
6091 case Intrinsic::aarch64_neon_st1x4:
6092 case Intrinsic::aarch64_neon_st2lane:
6093 case Intrinsic::aarch64_neon_st3lane:
6094 case Intrinsic::aarch64_neon_st4lane: {
6095 Info.opc = ISD::INTRINSIC_VOID;
6096 // Conservatively set memVT to the entire set of vectors stored.
6097 unsigned NumElts = 0;
6098 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
6099 Type *ArgTy = I.getArgOperand(ArgI)->getType();
6100 if (!ArgTy->isVectorTy())
6102 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
6104 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6105 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
6108 Info.vol = false; // volatile stores with NEON intrinsics not supported
6109 Info.readMem = false;
6110 Info.writeMem = true;
6113 case Intrinsic::aarch64_ldaxr:
6114 case Intrinsic::aarch64_ldxr: {
6115 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
6116 Info.opc = ISD::INTRINSIC_W_CHAIN;
6117 Info.memVT = MVT::getVT(PtrTy->getElementType());
6118 Info.ptrVal = I.getArgOperand(0);
6120 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
6122 Info.readMem = true;
6123 Info.writeMem = false;
6126 case Intrinsic::aarch64_stlxr:
6127 case Intrinsic::aarch64_stxr: {
6128 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
6129 Info.opc = ISD::INTRINSIC_W_CHAIN;
6130 Info.memVT = MVT::getVT(PtrTy->getElementType());
6131 Info.ptrVal = I.getArgOperand(1);
6133 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
6135 Info.readMem = false;
6136 Info.writeMem = true;
6139 case Intrinsic::aarch64_ldaxp:
6140 case Intrinsic::aarch64_ldxp: {
6141 Info.opc = ISD::INTRINSIC_W_CHAIN;
6142 Info.memVT = MVT::i128;
6143 Info.ptrVal = I.getArgOperand(0);
6147 Info.readMem = true;
6148 Info.writeMem = false;
6151 case Intrinsic::aarch64_stlxp:
6152 case Intrinsic::aarch64_stxp: {
6153 Info.opc = ISD::INTRINSIC_W_CHAIN;
6154 Info.memVT = MVT::i128;
6155 Info.ptrVal = I.getArgOperand(2);
6159 Info.readMem = false;
6160 Info.writeMem = true;
6170 // Truncations from 64-bit GPR to 32-bit GPR is free.
6171 bool AArch64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
6172 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6174 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6175 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6176 return NumBits1 > NumBits2;
6178 bool AArch64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
6179 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6181 unsigned NumBits1 = VT1.getSizeInBits();
6182 unsigned NumBits2 = VT2.getSizeInBits();
6183 return NumBits1 > NumBits2;
6186 // All 32-bit GPR operations implicitly zero the high-half of the corresponding
6188 bool AArch64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
6189 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6191 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6192 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6193 return NumBits1 == 32 && NumBits2 == 64;
6195 bool AArch64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
6196 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
6198 unsigned NumBits1 = VT1.getSizeInBits();
6199 unsigned NumBits2 = VT2.getSizeInBits();
6200 return NumBits1 == 32 && NumBits2 == 64;
6203 bool AArch64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
6204 EVT VT1 = Val.getValueType();
6205 if (isZExtFree(VT1, VT2)) {
6209 if (Val.getOpcode() != ISD::LOAD)
6212 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
6213 return (VT1.isSimple() && !VT1.isVector() && VT1.isInteger() &&
6214 VT2.isSimple() && !VT2.isVector() && VT2.isInteger() &&
6215 VT1.getSizeInBits() <= 32);
6218 bool AArch64TargetLowering::hasPairedLoad(Type *LoadedType,
6219 unsigned &RequiredAligment) const {
6220 if (!LoadedType->isIntegerTy() && !LoadedType->isFloatTy())
6222 // Cyclone supports unaligned accesses.
6223 RequiredAligment = 0;
6224 unsigned NumBits = LoadedType->getPrimitiveSizeInBits();
6225 return NumBits == 32 || NumBits == 64;
6228 bool AArch64TargetLowering::hasPairedLoad(EVT LoadedType,
6229 unsigned &RequiredAligment) const {
6230 if (!LoadedType.isSimple() ||
6231 (!LoadedType.isInteger() && !LoadedType.isFloatingPoint()))
6233 // Cyclone supports unaligned accesses.
6234 RequiredAligment = 0;
6235 unsigned NumBits = LoadedType.getSizeInBits();
6236 return NumBits == 32 || NumBits == 64;
6239 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
6240 unsigned AlignCheck) {
6241 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
6242 (DstAlign == 0 || DstAlign % AlignCheck == 0));
6245 EVT AArch64TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
6246 unsigned SrcAlign, bool IsMemset,
6249 MachineFunction &MF) const {
6250 // Don't use AdvSIMD to implement 16-byte memset. It would have taken one
6251 // instruction to materialize the v2i64 zero and one store (with restrictive
6252 // addressing mode). Just do two i64 store of zero-registers.
6254 const Function *F = MF.getFunction();
6255 if (Subtarget->hasFPARMv8() && !IsMemset && Size >= 16 &&
6256 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
6257 Attribute::NoImplicitFloat) &&
6258 (memOpAlign(SrcAlign, DstAlign, 16) ||
6259 (allowsMisalignedMemoryAccesses(MVT::f128, 0, 1, &Fast) && Fast)))
6262 return Size >= 8 ? MVT::i64 : MVT::i32;
6265 // 12-bit optionally shifted immediates are legal for adds.
6266 bool AArch64TargetLowering::isLegalAddImmediate(int64_t Immed) const {
6267 if ((Immed >> 12) == 0 || ((Immed & 0xfff) == 0 && Immed >> 24 == 0))
6272 // Integer comparisons are implemented with ADDS/SUBS, so the range of valid
6273 // immediates is the same as for an add or a sub.
6274 bool AArch64TargetLowering::isLegalICmpImmediate(int64_t Immed) const {
6277 return isLegalAddImmediate(Immed);
6280 /// isLegalAddressingMode - Return true if the addressing mode represented
6281 /// by AM is legal for this target, for a load/store of the specified type.
6282 bool AArch64TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6284 // AArch64 has five basic addressing modes:
6286 // reg + 9-bit signed offset
6287 // reg + SIZE_IN_BYTES * 12-bit unsigned offset
6289 // reg + SIZE_IN_BYTES * reg
6291 // No global is ever allowed as a base.
6295 // No reg+reg+imm addressing.
6296 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
6299 // check reg + imm case:
6300 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
6301 uint64_t NumBytes = 0;
6302 if (Ty->isSized()) {
6303 uint64_t NumBits = getDataLayout()->getTypeSizeInBits(Ty);
6304 NumBytes = NumBits / 8;
6305 if (!isPowerOf2_64(NumBits))
6310 int64_t Offset = AM.BaseOffs;
6312 // 9-bit signed offset
6313 if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
6316 // 12-bit unsigned offset
6317 unsigned shift = Log2_64(NumBytes);
6318 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
6319 // Must be a multiple of NumBytes (NumBytes is a power of 2)
6320 (Offset >> shift) << shift == Offset)
6325 // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2
6327 if (!AM.Scale || AM.Scale == 1 ||
6328 (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
6333 int AArch64TargetLowering::getScalingFactorCost(const AddrMode &AM,
6335 // Scaling factors are not free at all.
6336 // Operands | Rt Latency
6337 // -------------------------------------------
6339 // -------------------------------------------
6340 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
6341 // Rt, [Xn, Wm, <extend> #imm] |
6342 if (isLegalAddressingMode(AM, Ty))
6343 // Scale represents reg2 * scale, thus account for 1 if
6344 // it is not equal to 0 or 1.
6345 return AM.Scale != 0 && AM.Scale != 1;
6349 bool AArch64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
6350 VT = VT.getScalarType();
6355 switch (VT.getSimpleVT().SimpleTy) {
6367 AArch64TargetLowering::getScratchRegisters(CallingConv::ID) const {
6368 // LR is a callee-save register, but we must treat it as clobbered by any call
6369 // site. Hence we include LR in the scratch registers, which are in turn added
6370 // as implicit-defs for stackmaps and patchpoints.
6371 static const MCPhysReg ScratchRegs[] = {
6372 AArch64::X16, AArch64::X17, AArch64::LR, 0
6378 AArch64TargetLowering::isDesirableToCommuteWithShift(const SDNode *N) const {
6379 EVT VT = N->getValueType(0);
6380 // If N is unsigned bit extraction: ((x >> C) & mask), then do not combine
6381 // it with shift to let it be lowered to UBFX.
6382 if (N->getOpcode() == ISD::AND && (VT == MVT::i32 || VT == MVT::i64) &&
6383 isa<ConstantSDNode>(N->getOperand(1))) {
6384 uint64_t TruncMask = N->getConstantOperandVal(1);
6385 if (isMask_64(TruncMask) &&
6386 N->getOperand(0).getOpcode() == ISD::SRL &&
6387 isa<ConstantSDNode>(N->getOperand(0)->getOperand(1)))
6393 bool AArch64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
6395 assert(Ty->isIntegerTy());
6397 unsigned BitSize = Ty->getPrimitiveSizeInBits();
6401 int64_t Val = Imm.getSExtValue();
6402 if (Val == 0 || AArch64_AM::isLogicalImmediate(Val, BitSize))
6405 if ((int64_t)Val < 0)
6408 Val &= (1LL << 32) - 1;
6410 unsigned LZ = countLeadingZeros((uint64_t)Val);
6411 unsigned Shift = (63 - LZ) / 16;
6412 // MOVZ is free so return true for one or fewer MOVK.
6413 return (Shift < 3) ? true : false;
6416 // Generate SUBS and CSEL for integer abs.
6417 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
6418 EVT VT = N->getValueType(0);
6420 SDValue N0 = N->getOperand(0);
6421 SDValue N1 = N->getOperand(1);
6424 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
6425 // and change it to SUB and CSEL.
6426 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
6427 N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
6428 N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0))
6429 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
6430 if (Y1C->getAPIntValue() == VT.getSizeInBits() - 1) {
6431 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
6433 // Generate SUBS & CSEL.
6435 DAG.getNode(AArch64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32),
6436 N0.getOperand(0), DAG.getConstant(0, VT));
6437 return DAG.getNode(AArch64ISD::CSEL, DL, VT, N0.getOperand(0), Neg,
6438 DAG.getConstant(AArch64CC::PL, MVT::i32),
6439 SDValue(Cmp.getNode(), 1));
6444 // performXorCombine - Attempts to handle integer ABS.
6445 static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
6446 TargetLowering::DAGCombinerInfo &DCI,
6447 const AArch64Subtarget *Subtarget) {
6448 if (DCI.isBeforeLegalizeOps())
6451 return performIntegerAbsCombine(N, DAG);
6455 AArch64TargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor,
6457 std::vector<SDNode *> *Created) const {
6458 // fold (sdiv X, pow2)
6459 EVT VT = N->getValueType(0);
6460 if ((VT != MVT::i32 && VT != MVT::i64) ||
6461 !(Divisor.isPowerOf2() || (-Divisor).isPowerOf2()))
6465 SDValue N0 = N->getOperand(0);
6466 unsigned Lg2 = Divisor.countTrailingZeros();
6467 SDValue Zero = DAG.getConstant(0, VT);
6468 SDValue Pow2MinusOne = DAG.getConstant((1 << Lg2) - 1, VT);
6470 // Add (N0 < 0) ? Pow2 - 1 : 0;
6472 SDValue Cmp = getAArch64Cmp(N0, Zero, ISD::SETLT, CCVal, DAG, DL);
6473 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0, Pow2MinusOne);
6474 SDValue CSel = DAG.getNode(AArch64ISD::CSEL, DL, VT, Add, N0, CCVal, Cmp);
6477 Created->push_back(Cmp.getNode());
6478 Created->push_back(Add.getNode());
6479 Created->push_back(CSel.getNode());
6484 DAG.getNode(ISD::SRA, DL, VT, CSel, DAG.getConstant(Lg2, MVT::i64));
6486 // If we're dividing by a positive value, we're done. Otherwise, we must
6487 // negate the result.
6488 if (Divisor.isNonNegative())
6492 Created->push_back(SRA.getNode());
6493 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT), SRA);
6496 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG,
6497 TargetLowering::DAGCombinerInfo &DCI,
6498 const AArch64Subtarget *Subtarget) {
6499 if (DCI.isBeforeLegalizeOps())
6502 // Multiplication of a power of two plus/minus one can be done more
6503 // cheaply as as shift+add/sub. For now, this is true unilaterally. If
6504 // future CPUs have a cheaper MADD instruction, this may need to be
6505 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
6506 // 64-bit is 5 cycles, so this is always a win.
6507 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
6508 APInt Value = C->getAPIntValue();
6509 EVT VT = N->getValueType(0);
6510 if (Value.isNonNegative()) {
6511 // (mul x, 2^N + 1) => (add (shl x, N), x)
6512 APInt VM1 = Value - 1;
6513 if (VM1.isPowerOf2()) {
6514 SDValue ShiftedVal =
6515 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6516 DAG.getConstant(VM1.logBase2(), MVT::i64));
6517 return DAG.getNode(ISD::ADD, SDLoc(N), VT, ShiftedVal,
6520 // (mul x, 2^N - 1) => (sub (shl x, N), x)
6521 APInt VP1 = Value + 1;
6522 if (VP1.isPowerOf2()) {
6523 SDValue ShiftedVal =
6524 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6525 DAG.getConstant(VP1.logBase2(), MVT::i64));
6526 return DAG.getNode(ISD::SUB, SDLoc(N), VT, ShiftedVal,
6530 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
6531 APInt VNM1 = -Value - 1;
6532 if (VNM1.isPowerOf2()) {
6533 SDValue ShiftedVal =
6534 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6535 DAG.getConstant(VNM1.logBase2(), MVT::i64));
6537 DAG.getNode(ISD::ADD, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
6538 return DAG.getNode(ISD::SUB, SDLoc(N), VT, DAG.getConstant(0, VT), Add);
6540 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
6541 APInt VNP1 = -Value + 1;
6542 if (VNP1.isPowerOf2()) {
6543 SDValue ShiftedVal =
6544 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6545 DAG.getConstant(VNP1.logBase2(), MVT::i64));
6546 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N->getOperand(0),
6554 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
6555 SelectionDAG &DAG) {
6556 // Take advantage of vector comparisons producing 0 or -1 in each lane to
6557 // optimize away operation when it's from a constant.
6559 // The general transformation is:
6560 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
6561 // AND(VECTOR_CMP(x,y), constant2)
6562 // constant2 = UNARYOP(constant)
6564 // Early exit if this isn't a vector operation, the operand of the
6565 // unary operation isn't a bitwise AND, or if the sizes of the operations
6567 EVT VT = N->getValueType(0);
6568 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
6569 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
6570 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
6573 // Now check that the other operand of the AND is a constant. We could
6574 // make the transformation for non-constant splats as well, but it's unclear
6575 // that would be a benefit as it would not eliminate any operations, just
6576 // perform one more step in scalar code before moving to the vector unit.
6577 if (BuildVectorSDNode *BV =
6578 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
6579 // Bail out if the vector isn't a constant.
6580 if (!BV->isConstant())
6583 // Everything checks out. Build up the new and improved node.
6585 EVT IntVT = BV->getValueType(0);
6586 // Create a new constant of the appropriate type for the transformed
6588 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
6589 // The AND node needs bitcasts to/from an integer vector type around it.
6590 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
6591 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
6592 N->getOperand(0)->getOperand(0), MaskConst);
6593 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
6600 static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG) {
6601 // First try to optimize away the conversion when it's conditionally from
6602 // a constant. Vectors only.
6603 SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
6604 if (Res != SDValue())
6607 EVT VT = N->getValueType(0);
6608 if (VT != MVT::f32 && VT != MVT::f64)
6611 // Only optimize when the source and destination types have the same width.
6612 if (VT.getSizeInBits() != N->getOperand(0).getValueType().getSizeInBits())
6615 // If the result of an integer load is only used by an integer-to-float
6616 // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
6617 // This eliminates an "integer-to-vector-move UOP and improve throughput.
6618 SDValue N0 = N->getOperand(0);
6619 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6620 // Do not change the width of a volatile load.
6621 !cast<LoadSDNode>(N0)->isVolatile()) {
6622 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6623 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
6624 LN0->getPointerInfo(), LN0->isVolatile(),
6625 LN0->isNonTemporal(), LN0->isInvariant(),
6626 LN0->getAlignment());
6628 // Make sure successors of the original load stay after it by updating them
6629 // to use the new Chain.
6630 DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));
6633 (N->getOpcode() == ISD::SINT_TO_FP) ? AArch64ISD::SITOF : AArch64ISD::UITOF;
6634 return DAG.getNode(Opcode, SDLoc(N), VT, Load);
6640 /// An EXTR instruction is made up of two shifts, ORed together. This helper
6641 /// searches for and classifies those shifts.
6642 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
6644 if (N.getOpcode() == ISD::SHL)
6646 else if (N.getOpcode() == ISD::SRL)
6651 if (!isa<ConstantSDNode>(N.getOperand(1)))
6654 ShiftAmount = N->getConstantOperandVal(1);
6655 Src = N->getOperand(0);
6659 /// EXTR instruction extracts a contiguous chunk of bits from two existing
6660 /// registers viewed as a high/low pair. This function looks for the pattern:
6661 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
6662 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
6664 static SDValue tryCombineToEXTR(SDNode *N,
6665 TargetLowering::DAGCombinerInfo &DCI) {
6666 SelectionDAG &DAG = DCI.DAG;
6668 EVT VT = N->getValueType(0);
6670 assert(N->getOpcode() == ISD::OR && "Unexpected root");
6672 if (VT != MVT::i32 && VT != MVT::i64)
6676 uint32_t ShiftLHS = 0;
6678 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
6682 uint32_t ShiftRHS = 0;
6684 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
6687 // If they're both trying to come from the high part of the register, they're
6688 // not really an EXTR.
6689 if (LHSFromHi == RHSFromHi)
6692 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
6696 std::swap(LHS, RHS);
6697 std::swap(ShiftLHS, ShiftRHS);
6700 return DAG.getNode(AArch64ISD::EXTR, DL, VT, LHS, RHS,
6701 DAG.getConstant(ShiftRHS, MVT::i64));
6704 static SDValue tryCombineToBSL(SDNode *N,
6705 TargetLowering::DAGCombinerInfo &DCI) {
6706 EVT VT = N->getValueType(0);
6707 SelectionDAG &DAG = DCI.DAG;
6713 SDValue N0 = N->getOperand(0);
6714 if (N0.getOpcode() != ISD::AND)
6717 SDValue N1 = N->getOperand(1);
6718 if (N1.getOpcode() != ISD::AND)
6721 // We only have to look for constant vectors here since the general, variable
6722 // case can be handled in TableGen.
6723 unsigned Bits = VT.getVectorElementType().getSizeInBits();
6724 uint64_t BitMask = Bits == 64 ? -1ULL : ((1ULL << Bits) - 1);
6725 for (int i = 1; i >= 0; --i)
6726 for (int j = 1; j >= 0; --j) {
6727 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(i));
6728 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(j));
6732 bool FoundMatch = true;
6733 for (unsigned k = 0; k < VT.getVectorNumElements(); ++k) {
6734 ConstantSDNode *CN0 = dyn_cast<ConstantSDNode>(BVN0->getOperand(k));
6735 ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(BVN1->getOperand(k));
6737 CN0->getZExtValue() != (BitMask & ~CN1->getZExtValue())) {
6744 return DAG.getNode(AArch64ISD::BSL, DL, VT, SDValue(BVN0, 0),
6745 N0->getOperand(1 - i), N1->getOperand(1 - j));
6751 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
6752 const AArch64Subtarget *Subtarget) {
6753 // Attempt to form an EXTR from (or (shl VAL1, #N), (srl VAL2, #RegWidth-N))
6754 if (!EnableAArch64ExtrGeneration)
6756 SelectionDAG &DAG = DCI.DAG;
6757 EVT VT = N->getValueType(0);
6759 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6762 SDValue Res = tryCombineToEXTR(N, DCI);
6766 Res = tryCombineToBSL(N, DCI);
6773 static SDValue performBitcastCombine(SDNode *N,
6774 TargetLowering::DAGCombinerInfo &DCI,
6775 SelectionDAG &DAG) {
6776 // Wait 'til after everything is legalized to try this. That way we have
6777 // legal vector types and such.
6778 if (DCI.isBeforeLegalizeOps())
6781 // Remove extraneous bitcasts around an extract_subvector.
6783 // (v4i16 (bitconvert
6784 // (extract_subvector (v2i64 (bitconvert (v8i16 ...)), (i64 1)))))
6786 // (extract_subvector ((v8i16 ...), (i64 4)))
6788 // Only interested in 64-bit vectors as the ultimate result.
6789 EVT VT = N->getValueType(0);
6792 if (VT.getSimpleVT().getSizeInBits() != 64)
6794 // Is the operand an extract_subvector starting at the beginning or halfway
6795 // point of the vector? A low half may also come through as an
6796 // EXTRACT_SUBREG, so look for that, too.
6797 SDValue Op0 = N->getOperand(0);
6798 if (Op0->getOpcode() != ISD::EXTRACT_SUBVECTOR &&
6799 !(Op0->isMachineOpcode() &&
6800 Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG))
6802 uint64_t idx = cast<ConstantSDNode>(Op0->getOperand(1))->getZExtValue();
6803 if (Op0->getOpcode() == ISD::EXTRACT_SUBVECTOR) {
6804 if (Op0->getValueType(0).getVectorNumElements() != idx && idx != 0)
6806 } else if (Op0->getMachineOpcode() == AArch64::EXTRACT_SUBREG) {
6807 if (idx != AArch64::dsub)
6809 // The dsub reference is equivalent to a lane zero subvector reference.
6812 // Look through the bitcast of the input to the extract.
6813 if (Op0->getOperand(0)->getOpcode() != ISD::BITCAST)
6815 SDValue Source = Op0->getOperand(0)->getOperand(0);
6816 // If the source type has twice the number of elements as our destination
6817 // type, we know this is an extract of the high or low half of the vector.
6818 EVT SVT = Source->getValueType(0);
6819 if (SVT.getVectorNumElements() != VT.getVectorNumElements() * 2)
6822 DEBUG(dbgs() << "aarch64-lower: bitcast extract_subvector simplification\n");
6824 // Create the simplified form to just extract the low or high half of the
6825 // vector directly rather than bothering with the bitcasts.
6827 unsigned NumElements = VT.getVectorNumElements();
6829 SDValue HalfIdx = DAG.getConstant(NumElements, MVT::i64);
6830 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Source, HalfIdx);
6832 SDValue SubReg = DAG.getTargetConstant(AArch64::dsub, MVT::i32);
6833 return SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl, VT,
6839 static SDValue performConcatVectorsCombine(SDNode *N,
6840 TargetLowering::DAGCombinerInfo &DCI,
6841 SelectionDAG &DAG) {
6842 // Wait 'til after everything is legalized to try this. That way we have
6843 // legal vector types and such.
6844 if (DCI.isBeforeLegalizeOps())
6848 EVT VT = N->getValueType(0);
6850 // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector
6851 // splat. The indexed instructions are going to be expecting a DUPLANE64, so
6852 // canonicalise to that.
6853 if (N->getOperand(0) == N->getOperand(1) && VT.getVectorNumElements() == 2) {
6854 assert(VT.getVectorElementType().getSizeInBits() == 64);
6855 return DAG.getNode(AArch64ISD::DUPLANE64, dl, VT,
6856 WidenVector(N->getOperand(0), DAG),
6857 DAG.getConstant(0, MVT::i64));
6860 // Canonicalise concat_vectors so that the right-hand vector has as few
6861 // bit-casts as possible before its real operation. The primary matching
6862 // destination for these operations will be the narrowing "2" instructions,
6863 // which depend on the operation being performed on this right-hand vector.
6865 // (concat_vectors LHS, (v1i64 (bitconvert (v4i16 RHS))))
6867 // (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS))
6869 SDValue Op1 = N->getOperand(1);
6870 if (Op1->getOpcode() != ISD::BITCAST)
6872 SDValue RHS = Op1->getOperand(0);
6873 MVT RHSTy = RHS.getValueType().getSimpleVT();
6874 // If the RHS is not a vector, this is not the pattern we're looking for.
6875 if (!RHSTy.isVector())
6878 DEBUG(dbgs() << "aarch64-lower: concat_vectors bitcast simplification\n");
6880 MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(),
6881 RHSTy.getVectorNumElements() * 2);
6883 ISD::BITCAST, dl, VT,
6884 DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatTy,
6885 DAG.getNode(ISD::BITCAST, dl, RHSTy, N->getOperand(0)), RHS));
6888 static SDValue tryCombineFixedPointConvert(SDNode *N,
6889 TargetLowering::DAGCombinerInfo &DCI,
6890 SelectionDAG &DAG) {
6891 // Wait 'til after everything is legalized to try this. That way we have
6892 // legal vector types and such.
6893 if (DCI.isBeforeLegalizeOps())
6895 // Transform a scalar conversion of a value from a lane extract into a
6896 // lane extract of a vector conversion. E.g., from foo1 to foo2:
6897 // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); }
6898 // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; }
6900 // The second form interacts better with instruction selection and the
6901 // register allocator to avoid cross-class register copies that aren't
6902 // coalescable due to a lane reference.
6904 // Check the operand and see if it originates from a lane extract.
6905 SDValue Op1 = N->getOperand(1);
6906 if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6907 // Yep, no additional predication needed. Perform the transform.
6908 SDValue IID = N->getOperand(0);
6909 SDValue Shift = N->getOperand(2);
6910 SDValue Vec = Op1.getOperand(0);
6911 SDValue Lane = Op1.getOperand(1);
6912 EVT ResTy = N->getValueType(0);
6916 // The vector width should be 128 bits by the time we get here, even
6917 // if it started as 64 bits (the extract_vector handling will have
6919 assert(Vec.getValueType().getSizeInBits() == 128 &&
6920 "unexpected vector size on extract_vector_elt!");
6921 if (Vec.getValueType() == MVT::v4i32)
6922 VecResTy = MVT::v4f32;
6923 else if (Vec.getValueType() == MVT::v2i64)
6924 VecResTy = MVT::v2f64;
6926 llvm_unreachable("unexpected vector type!");
6929 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift);
6930 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane);
6935 // AArch64 high-vector "long" operations are formed by performing the non-high
6936 // version on an extract_subvector of each operand which gets the high half:
6938 // (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS))
6940 // However, there are cases which don't have an extract_high explicitly, but
6941 // have another operation that can be made compatible with one for free. For
6944 // (dupv64 scalar) --> (extract_high (dup128 scalar))
6946 // This routine does the actual conversion of such DUPs, once outer routines
6947 // have determined that everything else is in order.
6948 static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) {
6949 // We can handle most types of duplicate, but the lane ones have an extra
6950 // operand saying *which* lane, so we need to know.
6952 switch (N.getOpcode()) {
6953 case AArch64ISD::DUP:
6956 case AArch64ISD::DUPLANE8:
6957 case AArch64ISD::DUPLANE16:
6958 case AArch64ISD::DUPLANE32:
6959 case AArch64ISD::DUPLANE64:
6966 MVT NarrowTy = N.getSimpleValueType();
6967 if (!NarrowTy.is64BitVector())
6970 MVT ElementTy = NarrowTy.getVectorElementType();
6971 unsigned NumElems = NarrowTy.getVectorNumElements();
6972 MVT NewDUPVT = MVT::getVectorVT(ElementTy, NumElems * 2);
6976 NewDUP = DAG.getNode(N.getOpcode(), SDLoc(N), NewDUPVT, N.getOperand(0),
6979 NewDUP = DAG.getNode(AArch64ISD::DUP, SDLoc(N), NewDUPVT, N.getOperand(0));
6981 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N.getNode()), NarrowTy,
6982 NewDUP, DAG.getConstant(NumElems, MVT::i64));
6985 static bool isEssentiallyExtractSubvector(SDValue N) {
6986 if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR)
6989 return N.getOpcode() == ISD::BITCAST &&
6990 N.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR;
6993 /// \brief Helper structure to keep track of ISD::SET_CC operands.
6994 struct GenericSetCCInfo {
6995 const SDValue *Opnd0;
6996 const SDValue *Opnd1;
7000 /// \brief Helper structure to keep track of a SET_CC lowered into AArch64 code.
7001 struct AArch64SetCCInfo {
7003 AArch64CC::CondCode CC;
7006 /// \brief Helper structure to keep track of SetCC information.
7008 GenericSetCCInfo Generic;
7009 AArch64SetCCInfo AArch64;
7012 /// \brief Helper structure to be able to read SetCC information. If set to
7013 /// true, IsAArch64 field, Info is a AArch64SetCCInfo, otherwise Info is a
7014 /// GenericSetCCInfo.
7015 struct SetCCInfoAndKind {
7020 /// \brief Check whether or not \p Op is a SET_CC operation, either a generic or
7022 /// AArch64 lowered one.
7023 /// \p SetCCInfo is filled accordingly.
7024 /// \post SetCCInfo is meanginfull only when this function returns true.
7025 /// \return True when Op is a kind of SET_CC operation.
7026 static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) {
7027 // If this is a setcc, this is straight forward.
7028 if (Op.getOpcode() == ISD::SETCC) {
7029 SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0);
7030 SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1);
7031 SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
7032 SetCCInfo.IsAArch64 = false;
7035 // Otherwise, check if this is a matching csel instruction.
7039 if (Op.getOpcode() != AArch64ISD::CSEL)
7041 // Set the information about the operands.
7042 // TODO: we want the operands of the Cmp not the csel
7043 SetCCInfo.Info.AArch64.Cmp = &Op.getOperand(3);
7044 SetCCInfo.IsAArch64 = true;
7045 SetCCInfo.Info.AArch64.CC = static_cast<AArch64CC::CondCode>(
7046 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
7048 // Check that the operands matches the constraints:
7049 // (1) Both operands must be constants.
7050 // (2) One must be 1 and the other must be 0.
7051 ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0));
7052 ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7055 if (!TValue || !FValue)
7059 if (!TValue->isOne()) {
7060 // Update the comparison when we are interested in !cc.
7061 std::swap(TValue, FValue);
7062 SetCCInfo.Info.AArch64.CC =
7063 AArch64CC::getInvertedCondCode(SetCCInfo.Info.AArch64.CC);
7065 return TValue->isOne() && FValue->isNullValue();
7068 // Returns true if Op is setcc or zext of setcc.
7069 static bool isSetCCOrZExtSetCC(const SDValue& Op, SetCCInfoAndKind &Info) {
7070 if (isSetCC(Op, Info))
7072 return ((Op.getOpcode() == ISD::ZERO_EXTEND) &&
7073 isSetCC(Op->getOperand(0), Info));
7076 // The folding we want to perform is:
7077 // (add x, [zext] (setcc cc ...) )
7079 // (csel x, (add x, 1), !cc ...)
7081 // The latter will get matched to a CSINC instruction.
7082 static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) {
7083 assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!");
7084 SDValue LHS = Op->getOperand(0);
7085 SDValue RHS = Op->getOperand(1);
7086 SetCCInfoAndKind InfoAndKind;
7088 // If neither operand is a SET_CC, give up.
7089 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind)) {
7090 std::swap(LHS, RHS);
7091 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind))
7095 // FIXME: This could be generatized to work for FP comparisons.
7096 EVT CmpVT = InfoAndKind.IsAArch64
7097 ? InfoAndKind.Info.AArch64.Cmp->getOperand(0).getValueType()
7098 : InfoAndKind.Info.Generic.Opnd0->getValueType();
7099 if (CmpVT != MVT::i32 && CmpVT != MVT::i64)
7105 if (InfoAndKind.IsAArch64) {
7106 CCVal = DAG.getConstant(
7107 AArch64CC::getInvertedCondCode(InfoAndKind.Info.AArch64.CC), MVT::i32);
7108 Cmp = *InfoAndKind.Info.AArch64.Cmp;
7110 Cmp = getAArch64Cmp(*InfoAndKind.Info.Generic.Opnd0,
7111 *InfoAndKind.Info.Generic.Opnd1,
7112 ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, true),
7115 EVT VT = Op->getValueType(0);
7116 LHS = DAG.getNode(ISD::ADD, dl, VT, RHS, DAG.getConstant(1, VT));
7117 return DAG.getNode(AArch64ISD::CSEL, dl, VT, RHS, LHS, CCVal, Cmp);
7120 // The basic add/sub long vector instructions have variants with "2" on the end
7121 // which act on the high-half of their inputs. They are normally matched by
7124 // (add (zeroext (extract_high LHS)),
7125 // (zeroext (extract_high RHS)))
7126 // -> uaddl2 vD, vN, vM
7128 // However, if one of the extracts is something like a duplicate, this
7129 // instruction can still be used profitably. This function puts the DAG into a
7130 // more appropriate form for those patterns to trigger.
7131 static SDValue performAddSubLongCombine(SDNode *N,
7132 TargetLowering::DAGCombinerInfo &DCI,
7133 SelectionDAG &DAG) {
7134 if (DCI.isBeforeLegalizeOps())
7137 MVT VT = N->getSimpleValueType(0);
7138 if (!VT.is128BitVector()) {
7139 if (N->getOpcode() == ISD::ADD)
7140 return performSetccAddFolding(N, DAG);
7144 // Make sure both branches are extended in the same way.
7145 SDValue LHS = N->getOperand(0);
7146 SDValue RHS = N->getOperand(1);
7147 if ((LHS.getOpcode() != ISD::ZERO_EXTEND &&
7148 LHS.getOpcode() != ISD::SIGN_EXTEND) ||
7149 LHS.getOpcode() != RHS.getOpcode())
7152 unsigned ExtType = LHS.getOpcode();
7154 // It's not worth doing if at least one of the inputs isn't already an
7155 // extract, but we don't know which it'll be so we have to try both.
7156 if (isEssentiallyExtractSubvector(LHS.getOperand(0))) {
7157 RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG);
7161 RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS);
7162 } else if (isEssentiallyExtractSubvector(RHS.getOperand(0))) {
7163 LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG);
7167 LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS);
7170 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS);
7173 // Massage DAGs which we can use the high-half "long" operations on into
7174 // something isel will recognize better. E.g.
7176 // (aarch64_neon_umull (extract_high vec) (dupv64 scalar)) -->
7177 // (aarch64_neon_umull (extract_high (v2i64 vec)))
7178 // (extract_high (v2i64 (dup128 scalar)))))
7180 static SDValue tryCombineLongOpWithDup(unsigned IID, SDNode *N,
7181 TargetLowering::DAGCombinerInfo &DCI,
7182 SelectionDAG &DAG) {
7183 if (DCI.isBeforeLegalizeOps())
7186 SDValue LHS = N->getOperand(1);
7187 SDValue RHS = N->getOperand(2);
7188 assert(LHS.getValueType().is64BitVector() &&
7189 RHS.getValueType().is64BitVector() &&
7190 "unexpected shape for long operation");
7192 // Either node could be a DUP, but it's not worth doing both of them (you'd
7193 // just as well use the non-high version) so look for a corresponding extract
7194 // operation on the other "wing".
7195 if (isEssentiallyExtractSubvector(LHS)) {
7196 RHS = tryExtendDUPToExtractHigh(RHS, DAG);
7199 } else if (isEssentiallyExtractSubvector(RHS)) {
7200 LHS = tryExtendDUPToExtractHigh(LHS, DAG);
7205 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0),
7206 N->getOperand(0), LHS, RHS);
7209 static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
7210 MVT ElemTy = N->getSimpleValueType(0).getScalarType();
7211 unsigned ElemBits = ElemTy.getSizeInBits();
7213 int64_t ShiftAmount;
7214 if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) {
7215 APInt SplatValue, SplatUndef;
7216 unsigned SplatBitSize;
7218 if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
7219 HasAnyUndefs, ElemBits) ||
7220 SplatBitSize != ElemBits)
7223 ShiftAmount = SplatValue.getSExtValue();
7224 } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) {
7225 ShiftAmount = CVN->getSExtValue();
7233 llvm_unreachable("Unknown shift intrinsic");
7234 case Intrinsic::aarch64_neon_sqshl:
7235 Opcode = AArch64ISD::SQSHL_I;
7236 IsRightShift = false;
7238 case Intrinsic::aarch64_neon_uqshl:
7239 Opcode = AArch64ISD::UQSHL_I;
7240 IsRightShift = false;
7242 case Intrinsic::aarch64_neon_srshl:
7243 Opcode = AArch64ISD::SRSHR_I;
7244 IsRightShift = true;
7246 case Intrinsic::aarch64_neon_urshl:
7247 Opcode = AArch64ISD::URSHR_I;
7248 IsRightShift = true;
7250 case Intrinsic::aarch64_neon_sqshlu:
7251 Opcode = AArch64ISD::SQSHLU_I;
7252 IsRightShift = false;
7256 if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits)
7257 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
7258 DAG.getConstant(-ShiftAmount, MVT::i32));
7259 else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount < ElemBits)
7260 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
7261 DAG.getConstant(ShiftAmount, MVT::i32));
7266 // The CRC32[BH] instructions ignore the high bits of their data operand. Since
7267 // the intrinsics must be legal and take an i32, this means there's almost
7268 // certainly going to be a zext in the DAG which we can eliminate.
7269 static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) {
7270 SDValue AndN = N->getOperand(2);
7271 if (AndN.getOpcode() != ISD::AND)
7274 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1));
7275 if (!CMask || CMask->getZExtValue() != Mask)
7278 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32,
7279 N->getOperand(0), N->getOperand(1), AndN.getOperand(0));
7282 static SDValue performIntrinsicCombine(SDNode *N,
7283 TargetLowering::DAGCombinerInfo &DCI,
7284 const AArch64Subtarget *Subtarget) {
7285 SelectionDAG &DAG = DCI.DAG;
7286 unsigned IID = getIntrinsicID(N);
7290 case Intrinsic::aarch64_neon_vcvtfxs2fp:
7291 case Intrinsic::aarch64_neon_vcvtfxu2fp:
7292 return tryCombineFixedPointConvert(N, DCI, DAG);
7294 case Intrinsic::aarch64_neon_fmax:
7295 return DAG.getNode(AArch64ISD::FMAX, SDLoc(N), N->getValueType(0),
7296 N->getOperand(1), N->getOperand(2));
7297 case Intrinsic::aarch64_neon_fmin:
7298 return DAG.getNode(AArch64ISD::FMIN, SDLoc(N), N->getValueType(0),
7299 N->getOperand(1), N->getOperand(2));
7300 case Intrinsic::aarch64_neon_smull:
7301 case Intrinsic::aarch64_neon_umull:
7302 case Intrinsic::aarch64_neon_pmull:
7303 case Intrinsic::aarch64_neon_sqdmull:
7304 return tryCombineLongOpWithDup(IID, N, DCI, DAG);
7305 case Intrinsic::aarch64_neon_sqshl:
7306 case Intrinsic::aarch64_neon_uqshl:
7307 case Intrinsic::aarch64_neon_sqshlu:
7308 case Intrinsic::aarch64_neon_srshl:
7309 case Intrinsic::aarch64_neon_urshl:
7310 return tryCombineShiftImm(IID, N, DAG);
7311 case Intrinsic::aarch64_crc32b:
7312 case Intrinsic::aarch64_crc32cb:
7313 return tryCombineCRC32(0xff, N, DAG);
7314 case Intrinsic::aarch64_crc32h:
7315 case Intrinsic::aarch64_crc32ch:
7316 return tryCombineCRC32(0xffff, N, DAG);
7321 static SDValue performExtendCombine(SDNode *N,
7322 TargetLowering::DAGCombinerInfo &DCI,
7323 SelectionDAG &DAG) {
7324 // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then
7325 // we can convert that DUP into another extract_high (of a bigger DUP), which
7326 // helps the backend to decide that an sabdl2 would be useful, saving a real
7327 // extract_high operation.
7328 if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND &&
7329 N->getOperand(0).getOpcode() == ISD::INTRINSIC_WO_CHAIN) {
7330 SDNode *ABDNode = N->getOperand(0).getNode();
7331 unsigned IID = getIntrinsicID(ABDNode);
7332 if (IID == Intrinsic::aarch64_neon_sabd ||
7333 IID == Intrinsic::aarch64_neon_uabd) {
7334 SDValue NewABD = tryCombineLongOpWithDup(IID, ABDNode, DCI, DAG);
7335 if (!NewABD.getNode())
7338 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0),
7343 // This is effectively a custom type legalization for AArch64.
7345 // Type legalization will split an extend of a small, legal, type to a larger
7346 // illegal type by first splitting the destination type, often creating
7347 // illegal source types, which then get legalized in isel-confusing ways,
7348 // leading to really terrible codegen. E.g.,
7349 // %result = v8i32 sext v8i8 %value
7351 // %losrc = extract_subreg %value, ...
7352 // %hisrc = extract_subreg %value, ...
7353 // %lo = v4i32 sext v4i8 %losrc
7354 // %hi = v4i32 sext v4i8 %hisrc
7355 // Things go rapidly downhill from there.
7357 // For AArch64, the [sz]ext vector instructions can only go up one element
7358 // size, so we can, e.g., extend from i8 to i16, but to go from i8 to i32
7359 // take two instructions.
7361 // This implies that the most efficient way to do the extend from v8i8
7362 // to two v4i32 values is to first extend the v8i8 to v8i16, then do
7363 // the normal splitting to happen for the v8i16->v8i32.
7365 // This is pre-legalization to catch some cases where the default
7366 // type legalization will create ill-tempered code.
7367 if (!DCI.isBeforeLegalizeOps())
7370 // We're only interested in cleaning things up for non-legal vector types
7371 // here. If both the source and destination are legal, things will just
7372 // work naturally without any fiddling.
7373 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7374 EVT ResVT = N->getValueType(0);
7375 if (!ResVT.isVector() || TLI.isTypeLegal(ResVT))
7377 // If the vector type isn't a simple VT, it's beyond the scope of what
7378 // we're worried about here. Let legalization do its thing and hope for
7380 if (!ResVT.isSimple())
7383 SDValue Src = N->getOperand(0);
7384 MVT SrcVT = Src->getValueType(0).getSimpleVT();
7385 // If the source VT is a 64-bit vector, we can play games and get the
7386 // better results we want.
7387 if (SrcVT.getSizeInBits() != 64)
7390 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
7391 unsigned ElementCount = SrcVT.getVectorNumElements();
7392 SrcVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize * 2), ElementCount);
7394 Src = DAG.getNode(N->getOpcode(), DL, SrcVT, Src);
7396 // Now split the rest of the operation into two halves, each with a 64
7400 unsigned NumElements = ResVT.getVectorNumElements();
7401 assert(!(NumElements & 1) && "Splitting vector, but not in half!");
7402 LoVT = HiVT = EVT::getVectorVT(*DAG.getContext(),
7403 ResVT.getVectorElementType(), NumElements / 2);
7405 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
7406 LoVT.getVectorNumElements());
7407 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7408 DAG.getIntPtrConstant(0));
7409 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7410 DAG.getIntPtrConstant(InNVT.getVectorNumElements()));
7411 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, Lo);
7412 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, Hi);
7414 // Now combine the parts back together so we still have a single result
7415 // like the combiner expects.
7416 return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
7419 /// Replace a splat of a scalar to a vector store by scalar stores of the scalar
7420 /// value. The load store optimizer pass will merge them to store pair stores.
7421 /// This has better performance than a splat of the scalar followed by a split
7422 /// vector store. Even if the stores are not merged it is four stores vs a dup,
7423 /// followed by an ext.b and two stores.
7424 static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) {
7425 SDValue StVal = St->getValue();
7426 EVT VT = StVal.getValueType();
7428 // Don't replace floating point stores, they possibly won't be transformed to
7429 // stp because of the store pair suppress pass.
7430 if (VT.isFloatingPoint())
7433 // Check for insert vector elements.
7434 if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT)
7437 // We can express a splat as store pair(s) for 2 or 4 elements.
7438 unsigned NumVecElts = VT.getVectorNumElements();
7439 if (NumVecElts != 4 && NumVecElts != 2)
7441 SDValue SplatVal = StVal.getOperand(1);
7442 unsigned RemainInsertElts = NumVecElts - 1;
7444 // Check that this is a splat.
7445 while (--RemainInsertElts) {
7446 SDValue NextInsertElt = StVal.getOperand(0);
7447 if (NextInsertElt.getOpcode() != ISD::INSERT_VECTOR_ELT)
7449 if (NextInsertElt.getOperand(1) != SplatVal)
7451 StVal = NextInsertElt;
7453 unsigned OrigAlignment = St->getAlignment();
7454 unsigned EltOffset = NumVecElts == 4 ? 4 : 8;
7455 unsigned Alignment = std::min(OrigAlignment, EltOffset);
7457 // Create scalar stores. This is at least as good as the code sequence for a
7458 // split unaligned store wich is a dup.s, ext.b, and two stores.
7459 // Most of the time the three stores should be replaced by store pair
7460 // instructions (stp).
7462 SDValue BasePtr = St->getBasePtr();
7464 DAG.getStore(St->getChain(), DL, SplatVal, BasePtr, St->getPointerInfo(),
7465 St->isVolatile(), St->isNonTemporal(), St->getAlignment());
7467 unsigned Offset = EltOffset;
7468 while (--NumVecElts) {
7469 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7470 DAG.getConstant(Offset, MVT::i64));
7471 NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr,
7472 St->getPointerInfo(), St->isVolatile(),
7473 St->isNonTemporal(), Alignment);
7474 Offset += EltOffset;
7479 static SDValue performSTORECombine(SDNode *N,
7480 TargetLowering::DAGCombinerInfo &DCI,
7482 const AArch64Subtarget *Subtarget) {
7483 if (!DCI.isBeforeLegalize())
7486 StoreSDNode *S = cast<StoreSDNode>(N);
7487 if (S->isVolatile())
7490 // Cyclone has bad performance on unaligned 16B stores when crossing line and
7491 // page boundries. We want to split such stores.
7492 if (!Subtarget->isCyclone())
7495 // Don't split at Oz.
7496 MachineFunction &MF = DAG.getMachineFunction();
7497 bool IsMinSize = MF.getFunction()->getAttributes().hasAttribute(
7498 AttributeSet::FunctionIndex, Attribute::MinSize);
7502 SDValue StVal = S->getValue();
7503 EVT VT = StVal.getValueType();
7505 // Don't split v2i64 vectors. Memcpy lowering produces those and splitting
7506 // those up regresses performance on micro-benchmarks and olden/bh.
7507 if (!VT.isVector() || VT.getVectorNumElements() < 2 || VT == MVT::v2i64)
7510 // Split unaligned 16B stores. They are terrible for performance.
7511 // Don't split stores with alignment of 1 or 2. Code that uses clang vector
7512 // extensions can use this to mark that it does not want splitting to happen
7513 // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of
7514 // eliminating alignment hazards is only 1 in 8 for alignment of 2.
7515 if (VT.getSizeInBits() != 128 || S->getAlignment() >= 16 ||
7516 S->getAlignment() <= 2)
7519 // If we get a splat of a scalar convert this vector store to a store of
7520 // scalars. They will be merged into store pairs thereby removing two
7522 SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S);
7523 if (ReplacedSplat != SDValue())
7524 return ReplacedSplat;
7527 unsigned NumElts = VT.getVectorNumElements() / 2;
7528 // Split VT into two.
7530 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), NumElts);
7531 SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7532 DAG.getIntPtrConstant(0));
7533 SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7534 DAG.getIntPtrConstant(NumElts));
7535 SDValue BasePtr = S->getBasePtr();
7537 DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
7538 S->isVolatile(), S->isNonTemporal(), S->getAlignment());
7539 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7540 DAG.getConstant(8, MVT::i64));
7541 return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr,
7542 S->getPointerInfo(), S->isVolatile(), S->isNonTemporal(),
7546 /// Target-specific DAG combine function for post-increment LD1 (lane) and
7547 /// post-increment LD1R.
7548 static SDValue performPostLD1Combine(SDNode *N,
7549 TargetLowering::DAGCombinerInfo &DCI,
7551 if (DCI.isBeforeLegalizeOps())
7554 SelectionDAG &DAG = DCI.DAG;
7555 EVT VT = N->getValueType(0);
7557 unsigned LoadIdx = IsLaneOp ? 1 : 0;
7558 SDNode *LD = N->getOperand(LoadIdx).getNode();
7559 // If it is not LOAD, can not do such combine.
7560 if (LD->getOpcode() != ISD::LOAD)
7563 LoadSDNode *LoadSDN = cast<LoadSDNode>(LD);
7564 EVT MemVT = LoadSDN->getMemoryVT();
7565 // Check if memory operand is the same type as the vector element.
7566 if (MemVT != VT.getVectorElementType())
7569 // Check if there are other uses. If so, do not combine as it will introduce
7571 for (SDNode::use_iterator UI = LD->use_begin(), UE = LD->use_end(); UI != UE;
7573 if (UI.getUse().getResNo() == 1) // Ignore uses of the chain result.
7579 SDValue Addr = LD->getOperand(1);
7580 SDValue Vector = N->getOperand(0);
7581 // Search for a use of the address operand that is an increment.
7582 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(), UE =
7583 Addr.getNode()->use_end(); UI != UE; ++UI) {
7585 if (User->getOpcode() != ISD::ADD
7586 || UI.getUse().getResNo() != Addr.getResNo())
7589 // Check that the add is independent of the load. Otherwise, folding it
7590 // would create a cycle.
7591 if (User->isPredecessorOf(LD) || LD->isPredecessorOf(User))
7593 // Also check that add is not used in the vector operand. This would also
7595 if (User->isPredecessorOf(Vector.getNode()))
7598 // If the increment is a constant, it must match the memory ref size.
7599 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
7600 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
7601 uint32_t IncVal = CInc->getZExtValue();
7602 unsigned NumBytes = VT.getScalarSizeInBits() / 8;
7603 if (IncVal != NumBytes)
7605 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
7608 SmallVector<SDValue, 8> Ops;
7609 Ops.push_back(LD->getOperand(0)); // Chain
7611 Ops.push_back(Vector); // The vector to be inserted
7612 Ops.push_back(N->getOperand(2)); // The lane to be inserted in the vector
7614 Ops.push_back(Addr);
7617 EVT Tys[3] = { VT, MVT::i64, MVT::Other };
7618 SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, 3));
7619 unsigned NewOp = IsLaneOp ? AArch64ISD::LD1LANEpost : AArch64ISD::LD1DUPpost;
7620 SDValue UpdN = DAG.getMemIntrinsicNode(NewOp, SDLoc(N), SDTys, Ops,
7622 LoadSDN->getMemOperand());
7625 std::vector<SDValue> NewResults;
7626 NewResults.push_back(SDValue(LD, 0)); // The result of load
7627 NewResults.push_back(SDValue(UpdN.getNode(), 2)); // Chain
7628 DCI.CombineTo(LD, NewResults);
7629 DCI.CombineTo(N, SDValue(UpdN.getNode(), 0)); // Dup/Inserted Result
7630 DCI.CombineTo(User, SDValue(UpdN.getNode(), 1)); // Write back register
7637 /// Target-specific DAG combine function for NEON load/store intrinsics
7638 /// to merge base address updates.
7639 static SDValue performNEONPostLDSTCombine(SDNode *N,
7640 TargetLowering::DAGCombinerInfo &DCI,
7641 SelectionDAG &DAG) {
7642 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7645 unsigned AddrOpIdx = N->getNumOperands() - 1;
7646 SDValue Addr = N->getOperand(AddrOpIdx);
7648 // Search for a use of the address operand that is an increment.
7649 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
7650 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
7652 if (User->getOpcode() != ISD::ADD ||
7653 UI.getUse().getResNo() != Addr.getResNo())
7656 // Check that the add is independent of the load/store. Otherwise, folding
7657 // it would create a cycle.
7658 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
7661 // Find the new opcode for the updating load/store.
7662 bool IsStore = false;
7663 bool IsLaneOp = false;
7664 bool IsDupOp = false;
7665 unsigned NewOpc = 0;
7666 unsigned NumVecs = 0;
7667 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
7669 default: llvm_unreachable("unexpected intrinsic for Neon base update");
7670 case Intrinsic::aarch64_neon_ld2: NewOpc = AArch64ISD::LD2post;
7672 case Intrinsic::aarch64_neon_ld3: NewOpc = AArch64ISD::LD3post;
7674 case Intrinsic::aarch64_neon_ld4: NewOpc = AArch64ISD::LD4post;
7676 case Intrinsic::aarch64_neon_st2: NewOpc = AArch64ISD::ST2post;
7677 NumVecs = 2; IsStore = true; break;
7678 case Intrinsic::aarch64_neon_st3: NewOpc = AArch64ISD::ST3post;
7679 NumVecs = 3; IsStore = true; break;
7680 case Intrinsic::aarch64_neon_st4: NewOpc = AArch64ISD::ST4post;
7681 NumVecs = 4; IsStore = true; break;
7682 case Intrinsic::aarch64_neon_ld1x2: NewOpc = AArch64ISD::LD1x2post;
7684 case Intrinsic::aarch64_neon_ld1x3: NewOpc = AArch64ISD::LD1x3post;
7686 case Intrinsic::aarch64_neon_ld1x4: NewOpc = AArch64ISD::LD1x4post;
7688 case Intrinsic::aarch64_neon_st1x2: NewOpc = AArch64ISD::ST1x2post;
7689 NumVecs = 2; IsStore = true; break;
7690 case Intrinsic::aarch64_neon_st1x3: NewOpc = AArch64ISD::ST1x3post;
7691 NumVecs = 3; IsStore = true; break;
7692 case Intrinsic::aarch64_neon_st1x4: NewOpc = AArch64ISD::ST1x4post;
7693 NumVecs = 4; IsStore = true; break;
7694 case Intrinsic::aarch64_neon_ld2r: NewOpc = AArch64ISD::LD2DUPpost;
7695 NumVecs = 2; IsDupOp = true; break;
7696 case Intrinsic::aarch64_neon_ld3r: NewOpc = AArch64ISD::LD3DUPpost;
7697 NumVecs = 3; IsDupOp = true; break;
7698 case Intrinsic::aarch64_neon_ld4r: NewOpc = AArch64ISD::LD4DUPpost;
7699 NumVecs = 4; IsDupOp = true; break;
7700 case Intrinsic::aarch64_neon_ld2lane: NewOpc = AArch64ISD::LD2LANEpost;
7701 NumVecs = 2; IsLaneOp = true; break;
7702 case Intrinsic::aarch64_neon_ld3lane: NewOpc = AArch64ISD::LD3LANEpost;
7703 NumVecs = 3; IsLaneOp = true; break;
7704 case Intrinsic::aarch64_neon_ld4lane: NewOpc = AArch64ISD::LD4LANEpost;
7705 NumVecs = 4; IsLaneOp = true; break;
7706 case Intrinsic::aarch64_neon_st2lane: NewOpc = AArch64ISD::ST2LANEpost;
7707 NumVecs = 2; IsStore = true; IsLaneOp = true; break;
7708 case Intrinsic::aarch64_neon_st3lane: NewOpc = AArch64ISD::ST3LANEpost;
7709 NumVecs = 3; IsStore = true; IsLaneOp = true; break;
7710 case Intrinsic::aarch64_neon_st4lane: NewOpc = AArch64ISD::ST4LANEpost;
7711 NumVecs = 4; IsStore = true; IsLaneOp = true; break;
7716 VecTy = N->getOperand(2).getValueType();
7718 VecTy = N->getValueType(0);
7720 // If the increment is a constant, it must match the memory ref size.
7721 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
7722 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
7723 uint32_t IncVal = CInc->getZExtValue();
7724 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
7725 if (IsLaneOp || IsDupOp)
7726 NumBytes /= VecTy.getVectorNumElements();
7727 if (IncVal != NumBytes)
7729 Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
7731 SmallVector<SDValue, 8> Ops;
7732 Ops.push_back(N->getOperand(0)); // Incoming chain
7733 // Load lane and store have vector list as input.
7734 if (IsLaneOp || IsStore)
7735 for (unsigned i = 2; i < AddrOpIdx; ++i)
7736 Ops.push_back(N->getOperand(i));
7737 Ops.push_back(Addr); // Base register
7742 unsigned NumResultVecs = (IsStore ? 0 : NumVecs);
7744 for (n = 0; n < NumResultVecs; ++n)
7746 Tys[n++] = MVT::i64; // Type of write back register
7747 Tys[n] = MVT::Other; // Type of the chain
7748 SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, NumResultVecs + 2));
7750 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
7751 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys, Ops,
7752 MemInt->getMemoryVT(),
7753 MemInt->getMemOperand());
7756 std::vector<SDValue> NewResults;
7757 for (unsigned i = 0; i < NumResultVecs; ++i) {
7758 NewResults.push_back(SDValue(UpdN.getNode(), i));
7760 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs + 1));
7761 DCI.CombineTo(N, NewResults);
7762 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
7769 // Optimize compare with zero and branch.
7770 static SDValue performBRCONDCombine(SDNode *N,
7771 TargetLowering::DAGCombinerInfo &DCI,
7772 SelectionDAG &DAG) {
7773 SDValue Chain = N->getOperand(0);
7774 SDValue Dest = N->getOperand(1);
7775 SDValue CCVal = N->getOperand(2);
7776 SDValue Cmp = N->getOperand(3);
7778 assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!");
7779 unsigned CC = cast<ConstantSDNode>(CCVal)->getZExtValue();
7780 if (CC != AArch64CC::EQ && CC != AArch64CC::NE)
7783 unsigned CmpOpc = Cmp.getOpcode();
7784 if (CmpOpc != AArch64ISD::ADDS && CmpOpc != AArch64ISD::SUBS)
7787 // Only attempt folding if there is only one use of the flag and no use of the
7789 if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1))
7792 SDValue LHS = Cmp.getOperand(0);
7793 SDValue RHS = Cmp.getOperand(1);
7795 assert(LHS.getValueType() == RHS.getValueType() &&
7796 "Expected the value type to be the same for both operands!");
7797 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
7800 if (isa<ConstantSDNode>(LHS) && cast<ConstantSDNode>(LHS)->isNullValue())
7801 std::swap(LHS, RHS);
7803 if (!isa<ConstantSDNode>(RHS) || !cast<ConstantSDNode>(RHS)->isNullValue())
7806 if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA ||
7807 LHS.getOpcode() == ISD::SRL)
7810 // Fold the compare into the branch instruction.
7812 if (CC == AArch64CC::EQ)
7813 BR = DAG.getNode(AArch64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
7815 BR = DAG.getNode(AArch64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
7817 // Do not add new nodes to DAG combiner worklist.
7818 DCI.CombineTo(N, BR, false);
7823 // vselect (v1i1 setcc) ->
7824 // vselect (v1iXX setcc) (XX is the size of the compared operand type)
7825 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
7826 // condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
7828 static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) {
7829 SDValue N0 = N->getOperand(0);
7830 EVT CCVT = N0.getValueType();
7832 if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
7833 CCVT.getVectorElementType() != MVT::i1)
7836 EVT ResVT = N->getValueType(0);
7837 EVT CmpVT = N0.getOperand(0).getValueType();
7838 // Only combine when the result type is of the same size as the compared
7840 if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
7843 SDValue IfTrue = N->getOperand(1);
7844 SDValue IfFalse = N->getOperand(2);
7846 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
7847 N0.getOperand(0), N0.getOperand(1),
7848 cast<CondCodeSDNode>(N0.getOperand(2))->get());
7849 return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
7853 /// A vector select: "(select vL, vR, (setcc LHS, RHS))" is best performed with
7854 /// the compare-mask instructions rather than going via NZCV, even if LHS and
7855 /// RHS are really scalar. This replaces any scalar setcc in the above pattern
7856 /// with a vector one followed by a DUP shuffle on the result.
7857 static SDValue performSelectCombine(SDNode *N, SelectionDAG &DAG) {
7858 SDValue N0 = N->getOperand(0);
7859 EVT ResVT = N->getValueType(0);
7861 if (!N->getOperand(1).getValueType().isVector())
7864 if (N0.getOpcode() != ISD::SETCC || N0.getValueType() != MVT::i1)
7869 EVT SrcVT = N0.getOperand(0).getValueType();
7870 SrcVT = EVT::getVectorVT(*DAG.getContext(), SrcVT,
7871 ResVT.getSizeInBits() / SrcVT.getSizeInBits());
7872 EVT CCVT = SrcVT.changeVectorElementTypeToInteger();
7874 // First perform a vector comparison, where lane 0 is the one we're interested
7877 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(0));
7879 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(1));
7880 SDValue SetCC = DAG.getNode(ISD::SETCC, DL, CCVT, LHS, RHS, N0.getOperand(2));
7882 // Now duplicate the comparison mask we want across all other lanes.
7883 SmallVector<int, 8> DUPMask(CCVT.getVectorNumElements(), 0);
7884 SDValue Mask = DAG.getVectorShuffle(CCVT, DL, SetCC, SetCC, DUPMask.data());
7885 Mask = DAG.getNode(ISD::BITCAST, DL, ResVT.changeVectorElementTypeToInteger(),
7888 return DAG.getSelect(DL, ResVT, Mask, N->getOperand(1), N->getOperand(2));
7891 SDValue AArch64TargetLowering::PerformDAGCombine(SDNode *N,
7892 DAGCombinerInfo &DCI) const {
7893 SelectionDAG &DAG = DCI.DAG;
7894 switch (N->getOpcode()) {
7899 return performAddSubLongCombine(N, DCI, DAG);
7901 return performXorCombine(N, DAG, DCI, Subtarget);
7903 return performMulCombine(N, DAG, DCI, Subtarget);
7904 case ISD::SINT_TO_FP:
7905 case ISD::UINT_TO_FP:
7906 return performIntToFpCombine(N, DAG);
7908 return performORCombine(N, DCI, Subtarget);
7909 case ISD::INTRINSIC_WO_CHAIN:
7910 return performIntrinsicCombine(N, DCI, Subtarget);
7911 case ISD::ANY_EXTEND:
7912 case ISD::ZERO_EXTEND:
7913 case ISD::SIGN_EXTEND:
7914 return performExtendCombine(N, DCI, DAG);
7916 return performBitcastCombine(N, DCI, DAG);
7917 case ISD::CONCAT_VECTORS:
7918 return performConcatVectorsCombine(N, DCI, DAG);
7920 return performSelectCombine(N, DAG);
7922 return performVSelectCombine(N, DCI.DAG);
7924 return performSTORECombine(N, DCI, DAG, Subtarget);
7925 case AArch64ISD::BRCOND:
7926 return performBRCONDCombine(N, DCI, DAG);
7927 case AArch64ISD::DUP:
7928 return performPostLD1Combine(N, DCI, false);
7929 case ISD::INSERT_VECTOR_ELT:
7930 return performPostLD1Combine(N, DCI, true);
7931 case ISD::INTRINSIC_VOID:
7932 case ISD::INTRINSIC_W_CHAIN:
7933 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
7934 case Intrinsic::aarch64_neon_ld2:
7935 case Intrinsic::aarch64_neon_ld3:
7936 case Intrinsic::aarch64_neon_ld4:
7937 case Intrinsic::aarch64_neon_ld1x2:
7938 case Intrinsic::aarch64_neon_ld1x3:
7939 case Intrinsic::aarch64_neon_ld1x4:
7940 case Intrinsic::aarch64_neon_ld2lane:
7941 case Intrinsic::aarch64_neon_ld3lane:
7942 case Intrinsic::aarch64_neon_ld4lane:
7943 case Intrinsic::aarch64_neon_ld2r:
7944 case Intrinsic::aarch64_neon_ld3r:
7945 case Intrinsic::aarch64_neon_ld4r:
7946 case Intrinsic::aarch64_neon_st2:
7947 case Intrinsic::aarch64_neon_st3:
7948 case Intrinsic::aarch64_neon_st4:
7949 case Intrinsic::aarch64_neon_st1x2:
7950 case Intrinsic::aarch64_neon_st1x3:
7951 case Intrinsic::aarch64_neon_st1x4:
7952 case Intrinsic::aarch64_neon_st2lane:
7953 case Intrinsic::aarch64_neon_st3lane:
7954 case Intrinsic::aarch64_neon_st4lane:
7955 return performNEONPostLDSTCombine(N, DCI, DAG);
7963 // Check if the return value is used as only a return value, as otherwise
7964 // we can't perform a tail-call. In particular, we need to check for
7965 // target ISD nodes that are returns and any other "odd" constructs
7966 // that the generic analysis code won't necessarily catch.
7967 bool AArch64TargetLowering::isUsedByReturnOnly(SDNode *N,
7968 SDValue &Chain) const {
7969 if (N->getNumValues() != 1)
7971 if (!N->hasNUsesOfValue(1, 0))
7974 SDValue TCChain = Chain;
7975 SDNode *Copy = *N->use_begin();
7976 if (Copy->getOpcode() == ISD::CopyToReg) {
7977 // If the copy has a glue operand, we conservatively assume it isn't safe to
7978 // perform a tail call.
7979 if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() ==
7982 TCChain = Copy->getOperand(0);
7983 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
7986 bool HasRet = false;
7987 for (SDNode *Node : Copy->uses()) {
7988 if (Node->getOpcode() != AArch64ISD::RET_FLAG)
8000 // Return whether the an instruction can potentially be optimized to a tail
8001 // call. This will cause the optimizers to attempt to move, or duplicate,
8002 // return instructions to help enable tail call optimizations for this
8004 bool AArch64TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
8005 if (!CI->isTailCall())
8011 bool AArch64TargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base,
8013 ISD::MemIndexedMode &AM,
8015 SelectionDAG &DAG) const {
8016 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
8019 Base = Op->getOperand(0);
8020 // All of the indexed addressing mode instructions take a signed
8021 // 9 bit immediate offset.
8022 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
8023 int64_t RHSC = (int64_t)RHS->getZExtValue();
8024 if (RHSC >= 256 || RHSC <= -256)
8026 IsInc = (Op->getOpcode() == ISD::ADD);
8027 Offset = Op->getOperand(1);
8033 bool AArch64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
8035 ISD::MemIndexedMode &AM,
8036 SelectionDAG &DAG) const {
8039 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8040 VT = LD->getMemoryVT();
8041 Ptr = LD->getBasePtr();
8042 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8043 VT = ST->getMemoryVT();
8044 Ptr = ST->getBasePtr();
8049 if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, IsInc, DAG))
8051 AM = IsInc ? ISD::PRE_INC : ISD::PRE_DEC;
8055 bool AArch64TargetLowering::getPostIndexedAddressParts(
8056 SDNode *N, SDNode *Op, SDValue &Base, SDValue &Offset,
8057 ISD::MemIndexedMode &AM, SelectionDAG &DAG) const {
8060 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8061 VT = LD->getMemoryVT();
8062 Ptr = LD->getBasePtr();
8063 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8064 VT = ST->getMemoryVT();
8065 Ptr = ST->getBasePtr();
8070 if (!getIndexedAddressParts(Op, Base, Offset, AM, IsInc, DAG))
8072 // Post-indexing updates the base, so it's not a valid transform
8073 // if that's not the same as the load's pointer.
8076 AM = IsInc ? ISD::POST_INC : ISD::POST_DEC;
8080 static void ReplaceBITCASTResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
8081 SelectionDAG &DAG) {
8082 if (N->getValueType(0) != MVT::i16)
8086 SDValue Op = N->getOperand(0);
8087 assert(Op.getValueType() == MVT::f16 &&
8088 "Inconsistent bitcast? Only 16-bit types should be i16 or f16");
8090 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, DL, MVT::f32,
8091 DAG.getUNDEF(MVT::i32), Op,
8092 DAG.getTargetConstant(AArch64::hsub, MVT::i32)),
8094 Op = DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op);
8095 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, Op));
8098 void AArch64TargetLowering::ReplaceNodeResults(
8099 SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
8100 switch (N->getOpcode()) {
8102 llvm_unreachable("Don't know how to custom expand this");
8104 ReplaceBITCASTResults(N, Results, DAG);
8106 case ISD::FP_TO_UINT:
8107 case ISD::FP_TO_SINT:
8108 assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion");
8109 // Let normal code take care of it by not adding anything to Results.
8114 bool AArch64TargetLowering::shouldExpandAtomicInIR(Instruction *Inst) const {
8115 // Loads and stores less than 128-bits are already atomic; ones above that
8116 // are doomed anyway, so defer to the default libcall and blame the OS when
8118 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
8119 return SI->getValueOperand()->getType()->getPrimitiveSizeInBits() == 128;
8120 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
8121 return LI->getType()->getPrimitiveSizeInBits() == 128;
8123 // For the real atomic operations, we have ldxr/stxr up to 128 bits.
8124 return Inst->getType()->getPrimitiveSizeInBits() <= 128;
8127 bool AArch64TargetLowering::useLoadStackGuardNode() const {
8131 TargetLoweringBase::LegalizeTypeAction
8132 AArch64TargetLowering::getPreferredVectorAction(EVT VT) const {
8133 MVT SVT = VT.getSimpleVT();
8134 // During type legalization, we prefer to widen v1i8, v1i16, v1i32 to v8i8,
8135 // v4i16, v2i32 instead of to promote.
8136 if (SVT == MVT::v1i8 || SVT == MVT::v1i16 || SVT == MVT::v1i32
8137 || SVT == MVT::v1f32)
8138 return TypeWidenVector;
8140 return TargetLoweringBase::getPreferredVectorAction(VT);
8143 Value *AArch64TargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
8144 AtomicOrdering Ord) const {
8145 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
8146 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
8148 Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent;
8150 // Since i128 isn't legal and intrinsics don't get type-lowered, the ldrexd
8151 // intrinsic must return {i64, i64} and we have to recombine them into a
8152 // single i128 here.
8153 if (ValTy->getPrimitiveSizeInBits() == 128) {
8155 IsAcquire ? Intrinsic::aarch64_ldaxp : Intrinsic::aarch64_ldxp;
8156 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int);
8158 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
8159 Value *LoHi = Builder.CreateCall(Ldxr, Addr, "lohi");
8161 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
8162 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
8163 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
8164 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
8165 return Builder.CreateOr(
8166 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 64)), "val64");
8169 Type *Tys[] = { Addr->getType() };
8171 IsAcquire ? Intrinsic::aarch64_ldaxr : Intrinsic::aarch64_ldxr;
8172 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int, Tys);
8174 return Builder.CreateTruncOrBitCast(
8175 Builder.CreateCall(Ldxr, Addr),
8176 cast<PointerType>(Addr->getType())->getElementType());
8179 Value *AArch64TargetLowering::emitStoreConditional(IRBuilder<> &Builder,
8180 Value *Val, Value *Addr,
8181 AtomicOrdering Ord) const {
8182 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
8184 Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent;
8186 // Since the intrinsics must have legal type, the i128 intrinsics take two
8187 // parameters: "i64, i64". We must marshal Val into the appropriate form
8189 if (Val->getType()->getPrimitiveSizeInBits() == 128) {
8191 IsRelease ? Intrinsic::aarch64_stlxp : Intrinsic::aarch64_stxp;
8192 Function *Stxr = Intrinsic::getDeclaration(M, Int);
8193 Type *Int64Ty = Type::getInt64Ty(M->getContext());
8195 Value *Lo = Builder.CreateTrunc(Val, Int64Ty, "lo");
8196 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 64), Int64Ty, "hi");
8197 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
8198 return Builder.CreateCall3(Stxr, Lo, Hi, Addr);
8202 IsRelease ? Intrinsic::aarch64_stlxr : Intrinsic::aarch64_stxr;
8203 Type *Tys[] = { Addr->getType() };
8204 Function *Stxr = Intrinsic::getDeclaration(M, Int, Tys);
8206 return Builder.CreateCall2(
8207 Stxr, Builder.CreateZExtOrBitCast(
8208 Val, Stxr->getFunctionType()->getParamType(0)),