1 //===-- ARM64ISelLowering.cpp - ARM64 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 ARM64TargetLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "ARM64ISelLowering.h"
15 #include "ARM64PerfectShuffle.h"
16 #include "ARM64Subtarget.h"
17 #include "ARM64CallingConv.h"
18 #include "ARM64MachineFunctionInfo.h"
19 #include "ARM64TargetMachine.h"
20 #include "ARM64TargetObjectFile.h"
21 #include "MCTargetDesc/ARM64AddressingModes.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/CodeGen/CallingConvLower.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/Intrinsics.h"
29 #include "llvm/IR/Type.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetOptions.h"
37 #define DEBUG_TYPE "arm64-lower"
39 STATISTIC(NumTailCalls, "Number of tail calls");
40 STATISTIC(NumShiftInserts, "Number of vector shift inserts");
42 // This option should go away when tail calls fully work.
44 EnableARM64TailCalls("arm64-tail-calls", cl::Hidden,
45 cl::desc("Generate ARM64 tail calls (TEMPORARY OPTION)."),
53 static cl::opt<AlignMode>
54 Align(cl::desc("Load/store alignment support"),
55 cl::Hidden, cl::init(NoStrictAlign),
57 clEnumValN(StrictAlign, "arm64-strict-align",
58 "Disallow all unaligned memory accesses"),
59 clEnumValN(NoStrictAlign, "arm64-no-strict-align",
60 "Allow unaligned memory accesses"),
63 // Place holder until extr generation is tested fully.
65 EnableARM64ExtrGeneration("arm64-extr-generation", cl::Hidden,
66 cl::desc("Allow ARM64 (or (shift)(shift))->extract"),
70 EnableARM64SlrGeneration("arm64-shift-insert-generation", cl::Hidden,
71 cl::desc("Allow ARM64 SLI/SRI formation"),
74 //===----------------------------------------------------------------------===//
75 // ARM64 Lowering public interface.
76 //===----------------------------------------------------------------------===//
77 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
78 if (TM.getSubtarget<ARM64Subtarget>().isTargetDarwin())
79 return new ARM64_MachoTargetObjectFile();
81 return new ARM64_ELFTargetObjectFile();
84 ARM64TargetLowering::ARM64TargetLowering(ARM64TargetMachine &TM)
85 : TargetLowering(TM, createTLOF(TM)) {
86 Subtarget = &TM.getSubtarget<ARM64Subtarget>();
88 // ARM64 doesn't have comparisons which set GPRs or setcc instructions, so
89 // we have to make something up. Arbitrarily, choose ZeroOrOne.
90 setBooleanContents(ZeroOrOneBooleanContent);
91 // When comparing vectors the result sets the different elements in the
92 // vector to all-one or all-zero.
93 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
95 // Set up the register classes.
96 addRegisterClass(MVT::i32, &ARM64::GPR32allRegClass);
97 addRegisterClass(MVT::i64, &ARM64::GPR64allRegClass);
99 if (Subtarget->hasFPARMv8()) {
100 addRegisterClass(MVT::f16, &ARM64::FPR16RegClass);
101 addRegisterClass(MVT::f32, &ARM64::FPR32RegClass);
102 addRegisterClass(MVT::f64, &ARM64::FPR64RegClass);
103 addRegisterClass(MVT::f128, &ARM64::FPR128RegClass);
106 if (Subtarget->hasNEON()) {
107 addRegisterClass(MVT::v16i8, &ARM64::FPR8RegClass);
108 addRegisterClass(MVT::v8i16, &ARM64::FPR16RegClass);
109 // Someone set us up the NEON.
110 addDRTypeForNEON(MVT::v2f32);
111 addDRTypeForNEON(MVT::v8i8);
112 addDRTypeForNEON(MVT::v4i16);
113 addDRTypeForNEON(MVT::v2i32);
114 addDRTypeForNEON(MVT::v1i64);
115 addDRTypeForNEON(MVT::v1f64);
117 addQRTypeForNEON(MVT::v4f32);
118 addQRTypeForNEON(MVT::v2f64);
119 addQRTypeForNEON(MVT::v16i8);
120 addQRTypeForNEON(MVT::v8i16);
121 addQRTypeForNEON(MVT::v4i32);
122 addQRTypeForNEON(MVT::v2i64);
125 // Compute derived properties from the register classes
126 computeRegisterProperties();
128 // Provide all sorts of operation actions
129 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
130 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
131 setOperationAction(ISD::SETCC, MVT::i32, Custom);
132 setOperationAction(ISD::SETCC, MVT::i64, Custom);
133 setOperationAction(ISD::SETCC, MVT::f32, Custom);
134 setOperationAction(ISD::SETCC, MVT::f64, Custom);
135 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
136 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
137 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
138 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
139 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
140 setOperationAction(ISD::SELECT, MVT::i32, Custom);
141 setOperationAction(ISD::SELECT, MVT::i64, Custom);
142 setOperationAction(ISD::SELECT, MVT::f32, Custom);
143 setOperationAction(ISD::SELECT, MVT::f64, Custom);
144 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
145 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
146 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
147 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
148 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
149 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
151 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
152 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
153 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
155 setOperationAction(ISD::FREM, MVT::f32, Expand);
156 setOperationAction(ISD::FREM, MVT::f64, Expand);
157 setOperationAction(ISD::FREM, MVT::f80, Expand);
159 // Custom lowering hooks are needed for XOR
160 // to fold it into CSINC/CSINV.
161 setOperationAction(ISD::XOR, MVT::i32, Custom);
162 setOperationAction(ISD::XOR, MVT::i64, Custom);
164 // Virtually no operation on f128 is legal, but LLVM can't expand them when
165 // there's a valid register class, so we need custom operations in most cases.
166 setOperationAction(ISD::FABS, MVT::f128, Expand);
167 setOperationAction(ISD::FADD, MVT::f128, Custom);
168 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
169 setOperationAction(ISD::FCOS, MVT::f128, Expand);
170 setOperationAction(ISD::FDIV, MVT::f128, Custom);
171 setOperationAction(ISD::FMA, MVT::f128, Expand);
172 setOperationAction(ISD::FMUL, MVT::f128, Custom);
173 setOperationAction(ISD::FNEG, MVT::f128, Expand);
174 setOperationAction(ISD::FPOW, MVT::f128, Expand);
175 setOperationAction(ISD::FREM, MVT::f128, Expand);
176 setOperationAction(ISD::FRINT, MVT::f128, Expand);
177 setOperationAction(ISD::FSIN, MVT::f128, Expand);
178 setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
179 setOperationAction(ISD::FSQRT, MVT::f128, Expand);
180 setOperationAction(ISD::FSUB, MVT::f128, Custom);
181 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
182 setOperationAction(ISD::SETCC, MVT::f128, Custom);
183 setOperationAction(ISD::BR_CC, MVT::f128, Custom);
184 setOperationAction(ISD::SELECT, MVT::f128, Custom);
185 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
186 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);
188 // Lowering for many of the conversions is actually specified by the non-f128
189 // type. The LowerXXX function will be trivial when f128 isn't involved.
190 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
191 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
192 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
193 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
194 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
195 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
196 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
197 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
198 setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
199 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
200 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
201 setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
202 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
203 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
205 // Variable arguments.
206 setOperationAction(ISD::VASTART, MVT::Other, Custom);
207 setOperationAction(ISD::VAARG, MVT::Other, Custom);
208 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
209 setOperationAction(ISD::VAEND, MVT::Other, Expand);
211 // Variable-sized objects.
212 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
213 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
214 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
216 // Exception handling.
217 // FIXME: These are guesses. Has this been defined yet?
218 setExceptionPointerRegister(ARM64::X0);
219 setExceptionSelectorRegister(ARM64::X1);
221 // Constant pool entries
222 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
225 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
227 // Add/Sub overflow ops with MVT::Glues are lowered to NZCV dependences.
228 setOperationAction(ISD::ADDC, MVT::i32, Custom);
229 setOperationAction(ISD::ADDE, MVT::i32, Custom);
230 setOperationAction(ISD::SUBC, MVT::i32, Custom);
231 setOperationAction(ISD::SUBE, MVT::i32, Custom);
232 setOperationAction(ISD::ADDC, MVT::i64, Custom);
233 setOperationAction(ISD::ADDE, MVT::i64, Custom);
234 setOperationAction(ISD::SUBC, MVT::i64, Custom);
235 setOperationAction(ISD::SUBE, MVT::i64, Custom);
237 // ARM64 lacks both left-rotate and popcount instructions.
238 setOperationAction(ISD::ROTL, MVT::i32, Expand);
239 setOperationAction(ISD::ROTL, MVT::i64, Expand);
241 // ARM64 doesn't have {U|S}MUL_LOHI.
242 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
243 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
246 // Expand the undefined-at-zero variants to cttz/ctlz to their defined-at-zero
247 // counterparts, which ARM64 supports directly.
248 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
249 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
250 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
251 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
253 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
254 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
256 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
257 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
258 setOperationAction(ISD::SREM, MVT::i32, Expand);
259 setOperationAction(ISD::SREM, MVT::i64, Expand);
260 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
261 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
262 setOperationAction(ISD::UREM, MVT::i32, Expand);
263 setOperationAction(ISD::UREM, MVT::i64, Expand);
265 // Custom lower Add/Sub/Mul with overflow.
266 setOperationAction(ISD::SADDO, MVT::i32, Custom);
267 setOperationAction(ISD::SADDO, MVT::i64, Custom);
268 setOperationAction(ISD::UADDO, MVT::i32, Custom);
269 setOperationAction(ISD::UADDO, MVT::i64, Custom);
270 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
271 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
272 setOperationAction(ISD::USUBO, MVT::i32, Custom);
273 setOperationAction(ISD::USUBO, MVT::i64, Custom);
274 setOperationAction(ISD::SMULO, MVT::i32, Custom);
275 setOperationAction(ISD::SMULO, MVT::i64, Custom);
276 setOperationAction(ISD::UMULO, MVT::i32, Custom);
277 setOperationAction(ISD::UMULO, MVT::i64, Custom);
279 setOperationAction(ISD::FSIN, MVT::f32, Expand);
280 setOperationAction(ISD::FSIN, MVT::f64, Expand);
281 setOperationAction(ISD::FCOS, MVT::f32, Expand);
282 setOperationAction(ISD::FCOS, MVT::f64, Expand);
283 setOperationAction(ISD::FPOW, MVT::f32, Expand);
284 setOperationAction(ISD::FPOW, MVT::f64, Expand);
285 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
286 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
288 // ARM64 has implementations of a lot of rounding-like FP operations.
289 static MVT RoundingTypes[] = { MVT::f32, MVT::f64};
290 for (unsigned I = 0; I < array_lengthof(RoundingTypes); ++I) {
291 MVT Ty = RoundingTypes[I];
292 setOperationAction(ISD::FFLOOR, Ty, Legal);
293 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
294 setOperationAction(ISD::FCEIL, Ty, Legal);
295 setOperationAction(ISD::FRINT, Ty, Legal);
296 setOperationAction(ISD::FTRUNC, Ty, Legal);
297 setOperationAction(ISD::FROUND, Ty, Legal);
300 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
302 if (Subtarget->isTargetMachO()) {
303 // For iOS, we don't want to the normal expansion of a libcall to
304 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
306 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
307 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
309 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
310 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
313 // ARM64 does not have floating-point extending loads, i1 sign-extending load,
314 // floating-point truncating stores, or v2i32->v2i16 truncating store.
315 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
316 setLoadExtAction(ISD::EXTLOAD, MVT::f64, Expand);
317 setLoadExtAction(ISD::EXTLOAD, MVT::f80, Expand);
318 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Expand);
319 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
320 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
321 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
322 setTruncStoreAction(MVT::f128, MVT::f80, Expand);
323 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
324 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
325 setTruncStoreAction(MVT::f128, MVT::f16, Expand);
326 // Indexed loads and stores are supported.
327 for (unsigned im = (unsigned)ISD::PRE_INC;
328 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
329 setIndexedLoadAction(im, MVT::i8, Legal);
330 setIndexedLoadAction(im, MVT::i16, Legal);
331 setIndexedLoadAction(im, MVT::i32, Legal);
332 setIndexedLoadAction(im, MVT::i64, Legal);
333 setIndexedLoadAction(im, MVT::f64, Legal);
334 setIndexedLoadAction(im, MVT::f32, Legal);
335 setIndexedStoreAction(im, MVT::i8, Legal);
336 setIndexedStoreAction(im, MVT::i16, Legal);
337 setIndexedStoreAction(im, MVT::i32, Legal);
338 setIndexedStoreAction(im, MVT::i64, Legal);
339 setIndexedStoreAction(im, MVT::f64, Legal);
340 setIndexedStoreAction(im, MVT::f32, Legal);
344 setOperationAction(ISD::TRAP, MVT::Other, Legal);
346 // We combine OR nodes for bitfield operations.
347 setTargetDAGCombine(ISD::OR);
349 // Vector add and sub nodes may conceal a high-half opportunity.
350 // Also, try to fold ADD into CSINC/CSINV..
351 setTargetDAGCombine(ISD::ADD);
352 setTargetDAGCombine(ISD::SUB);
354 setTargetDAGCombine(ISD::XOR);
355 setTargetDAGCombine(ISD::SINT_TO_FP);
356 setTargetDAGCombine(ISD::UINT_TO_FP);
358 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
360 setTargetDAGCombine(ISD::ANY_EXTEND);
361 setTargetDAGCombine(ISD::ZERO_EXTEND);
362 setTargetDAGCombine(ISD::SIGN_EXTEND);
363 setTargetDAGCombine(ISD::BITCAST);
364 setTargetDAGCombine(ISD::CONCAT_VECTORS);
365 setTargetDAGCombine(ISD::STORE);
367 setTargetDAGCombine(ISD::MUL);
369 setTargetDAGCombine(ISD::SELECT);
370 setTargetDAGCombine(ISD::VSELECT);
372 setTargetDAGCombine(ISD::INTRINSIC_VOID);
373 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
375 MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8;
376 MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4;
377 MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 4;
379 setStackPointerRegisterToSaveRestore(ARM64::SP);
381 setSchedulingPreference(Sched::Hybrid);
384 MaskAndBranchFoldingIsLegal = true;
386 setMinFunctionAlignment(2);
388 RequireStrictAlign = (Align == StrictAlign);
390 setHasExtractBitsInsn(true);
392 if (Subtarget->hasNEON()) {
393 // FIXME: v1f64 shouldn't be legal if we can avoid it, because it leads to
394 // silliness like this:
395 setOperationAction(ISD::FABS, MVT::v1f64, Expand);
396 setOperationAction(ISD::FADD, MVT::v1f64, Expand);
397 setOperationAction(ISD::FCEIL, MVT::v1f64, Expand);
398 setOperationAction(ISD::FCOPYSIGN, MVT::v1f64, Expand);
399 setOperationAction(ISD::FCOS, MVT::v1f64, Expand);
400 setOperationAction(ISD::FDIV, MVT::v1f64, Expand);
401 setOperationAction(ISD::FFLOOR, MVT::v1f64, Expand);
402 setOperationAction(ISD::FMA, MVT::v1f64, Expand);
403 setOperationAction(ISD::FMUL, MVT::v1f64, Expand);
404 setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Expand);
405 setOperationAction(ISD::FNEG, MVT::v1f64, Expand);
406 setOperationAction(ISD::FPOW, MVT::v1f64, Expand);
407 setOperationAction(ISD::FREM, MVT::v1f64, Expand);
408 setOperationAction(ISD::FROUND, MVT::v1f64, Expand);
409 setOperationAction(ISD::FRINT, MVT::v1f64, Expand);
410 setOperationAction(ISD::FSIN, MVT::v1f64, Expand);
411 setOperationAction(ISD::FSINCOS, MVT::v1f64, Expand);
412 setOperationAction(ISD::FSQRT, MVT::v1f64, Expand);
413 setOperationAction(ISD::FSUB, MVT::v1f64, Expand);
414 setOperationAction(ISD::FTRUNC, MVT::v1f64, Expand);
415 setOperationAction(ISD::SETCC, MVT::v1f64, Expand);
416 setOperationAction(ISD::BR_CC, MVT::v1f64, Expand);
417 setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
418 setOperationAction(ISD::SELECT_CC, MVT::v1f64, Expand);
419 setOperationAction(ISD::FP_EXTEND, MVT::v1f64, Expand);
421 setOperationAction(ISD::FP_TO_SINT, MVT::v1i64, Expand);
422 setOperationAction(ISD::FP_TO_UINT, MVT::v1i64, Expand);
423 setOperationAction(ISD::SINT_TO_FP, MVT::v1i64, Expand);
424 setOperationAction(ISD::UINT_TO_FP, MVT::v1i64, Expand);
425 setOperationAction(ISD::FP_ROUND, MVT::v1f64, Expand);
427 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
429 // ARM64 doesn't have a direct vector ->f32 conversion instructions for
430 // elements smaller than i32, so promote the input to i32 first.
431 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Promote);
432 setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Promote);
433 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Promote);
434 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Promote);
435 // Similarly, there is no direct i32 -> f64 vector conversion instruction.
436 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
437 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
438 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
439 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
441 // ARM64 doesn't have MUL.2d:
442 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
443 setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal);
444 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
445 // Likewise, narrowing and extending vector loads/stores aren't handled
447 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
448 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
450 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,
453 setOperationAction(ISD::MULHS, (MVT::SimpleValueType)VT, Expand);
454 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
455 setOperationAction(ISD::MULHU, (MVT::SimpleValueType)VT, Expand);
456 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
458 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
459 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
460 setTruncStoreAction((MVT::SimpleValueType)VT,
461 (MVT::SimpleValueType)InnerVT, Expand);
462 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
463 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
464 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
467 // ARM64 has implementations of a lot of rounding-like FP operations.
468 static MVT RoundingVecTypes[] = {MVT::v2f32, MVT::v4f32, MVT::v2f64 };
469 for (unsigned I = 0; I < array_lengthof(RoundingVecTypes); ++I) {
470 MVT Ty = RoundingVecTypes[I];
471 setOperationAction(ISD::FFLOOR, Ty, Legal);
472 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
473 setOperationAction(ISD::FCEIL, Ty, Legal);
474 setOperationAction(ISD::FRINT, Ty, Legal);
475 setOperationAction(ISD::FTRUNC, Ty, Legal);
476 setOperationAction(ISD::FROUND, Ty, Legal);
481 void ARM64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) {
482 if (VT == MVT::v2f32) {
483 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
484 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i32);
486 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
487 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i32);
488 } else if (VT == MVT::v2f64 || VT == MVT::v4f32) {
489 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
490 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i64);
492 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
493 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i64);
496 // Mark vector float intrinsics as expand.
497 if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) {
498 setOperationAction(ISD::FSIN, VT.getSimpleVT(), Expand);
499 setOperationAction(ISD::FCOS, VT.getSimpleVT(), Expand);
500 setOperationAction(ISD::FPOWI, VT.getSimpleVT(), Expand);
501 setOperationAction(ISD::FPOW, VT.getSimpleVT(), Expand);
502 setOperationAction(ISD::FLOG, VT.getSimpleVT(), Expand);
503 setOperationAction(ISD::FLOG2, VT.getSimpleVT(), Expand);
504 setOperationAction(ISD::FLOG10, VT.getSimpleVT(), Expand);
505 setOperationAction(ISD::FEXP, VT.getSimpleVT(), Expand);
506 setOperationAction(ISD::FEXP2, VT.getSimpleVT(), Expand);
509 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
510 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
511 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
512 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
513 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Custom);
514 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
515 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
516 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
517 setOperationAction(ISD::AND, VT.getSimpleVT(), Custom);
518 setOperationAction(ISD::OR, VT.getSimpleVT(), Custom);
519 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
520 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
522 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
523 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
524 setOperationAction(ISD::VSELECT, VT.getSimpleVT(), Expand);
525 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
527 // CNT supports only B element sizes.
528 if (VT != MVT::v8i8 && VT != MVT::v16i8)
529 setOperationAction(ISD::CTPOP, VT.getSimpleVT(), Expand);
531 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
532 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
533 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
534 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
535 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
537 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
538 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
540 if (Subtarget->isLittleEndian()) {
541 for (unsigned im = (unsigned)ISD::PRE_INC;
542 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
543 setIndexedLoadAction(im, VT.getSimpleVT(), Legal);
544 setIndexedStoreAction(im, VT.getSimpleVT(), Legal);
549 void ARM64TargetLowering::addDRTypeForNEON(MVT VT) {
550 addRegisterClass(VT, &ARM64::FPR64RegClass);
551 addTypeForNEON(VT, MVT::v2i32);
554 void ARM64TargetLowering::addQRTypeForNEON(MVT VT) {
555 addRegisterClass(VT, &ARM64::FPR128RegClass);
556 addTypeForNEON(VT, MVT::v4i32);
559 EVT ARM64TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
562 return VT.changeVectorElementTypeToInteger();
565 /// computeMaskedBitsForTargetNode - Determine which of the bits specified in
566 /// Mask are known to be either zero or one and return them in the
567 /// KnownZero/KnownOne bitsets.
568 void ARM64TargetLowering::computeMaskedBitsForTargetNode(
569 const SDValue Op, APInt &KnownZero, APInt &KnownOne,
570 const SelectionDAG &DAG, unsigned Depth) const {
571 switch (Op.getOpcode()) {
574 case ARM64ISD::CSEL: {
575 APInt KnownZero2, KnownOne2;
576 DAG.ComputeMaskedBits(Op->getOperand(0), KnownZero, KnownOne, Depth + 1);
577 DAG.ComputeMaskedBits(Op->getOperand(1), KnownZero2, KnownOne2, Depth + 1);
578 KnownZero &= KnownZero2;
579 KnownOne &= KnownOne2;
582 case ISD::INTRINSIC_W_CHAIN: {
583 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
584 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
587 case Intrinsic::arm64_ldaxr:
588 case Intrinsic::arm64_ldxr: {
589 unsigned BitWidth = KnownOne.getBitWidth();
590 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
591 unsigned MemBits = VT.getScalarType().getSizeInBits();
592 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
598 case ISD::INTRINSIC_WO_CHAIN:
599 case ISD::INTRINSIC_VOID: {
600 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
604 case Intrinsic::arm64_neon_umaxv:
605 case Intrinsic::arm64_neon_uminv: {
606 // Figure out the datatype of the vector operand. The UMINV instruction
607 // will zero extend the result, so we can mark as known zero all the
608 // bits larger than the element datatype. 32-bit or larget doesn't need
609 // this as those are legal types and will be handled by isel directly.
610 MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
611 unsigned BitWidth = KnownZero.getBitWidth();
612 if (VT == MVT::v8i8 || VT == MVT::v16i8) {
613 assert(BitWidth >= 8 && "Unexpected width!");
614 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8);
616 } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
617 assert(BitWidth >= 16 && "Unexpected width!");
618 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
628 MVT ARM64TargetLowering::getScalarShiftAmountTy(EVT LHSTy) const {
632 unsigned ARM64TargetLowering::getMaximalGlobalOffset() const {
633 // FIXME: On ARM64, this depends on the type.
634 // Basically, the addressable offsets are o to 4095 * Ty.getSizeInBytes().
635 // and the offset has to be a multiple of the related size in bytes.
640 ARM64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
641 const TargetLibraryInfo *libInfo) const {
642 return ARM64::createFastISel(funcInfo, libInfo);
645 const char *ARM64TargetLowering::getTargetNodeName(unsigned Opcode) const {
649 case ARM64ISD::CALL: return "ARM64ISD::CALL";
650 case ARM64ISD::ADRP: return "ARM64ISD::ADRP";
651 case ARM64ISD::ADDlow: return "ARM64ISD::ADDlow";
652 case ARM64ISD::LOADgot: return "ARM64ISD::LOADgot";
653 case ARM64ISD::RET_FLAG: return "ARM64ISD::RET_FLAG";
654 case ARM64ISD::BRCOND: return "ARM64ISD::BRCOND";
655 case ARM64ISD::CSEL: return "ARM64ISD::CSEL";
656 case ARM64ISD::FCSEL: return "ARM64ISD::FCSEL";
657 case ARM64ISD::CSINV: return "ARM64ISD::CSINV";
658 case ARM64ISD::CSNEG: return "ARM64ISD::CSNEG";
659 case ARM64ISD::CSINC: return "ARM64ISD::CSINC";
660 case ARM64ISD::THREAD_POINTER: return "ARM64ISD::THREAD_POINTER";
661 case ARM64ISD::TLSDESC_CALL: return "ARM64ISD::TLSDESC_CALL";
662 case ARM64ISD::ADC: return "ARM64ISD::ADC";
663 case ARM64ISD::SBC: return "ARM64ISD::SBC";
664 case ARM64ISD::ADDS: return "ARM64ISD::ADDS";
665 case ARM64ISD::SUBS: return "ARM64ISD::SUBS";
666 case ARM64ISD::ADCS: return "ARM64ISD::ADCS";
667 case ARM64ISD::SBCS: return "ARM64ISD::SBCS";
668 case ARM64ISD::ANDS: return "ARM64ISD::ANDS";
669 case ARM64ISD::FCMP: return "ARM64ISD::FCMP";
670 case ARM64ISD::FMIN: return "ARM64ISD::FMIN";
671 case ARM64ISD::FMAX: return "ARM64ISD::FMAX";
672 case ARM64ISD::DUP: return "ARM64ISD::DUP";
673 case ARM64ISD::DUPLANE8: return "ARM64ISD::DUPLANE8";
674 case ARM64ISD::DUPLANE16: return "ARM64ISD::DUPLANE16";
675 case ARM64ISD::DUPLANE32: return "ARM64ISD::DUPLANE32";
676 case ARM64ISD::DUPLANE64: return "ARM64ISD::DUPLANE64";
677 case ARM64ISD::MOVI: return "ARM64ISD::MOVI";
678 case ARM64ISD::MOVIshift: return "ARM64ISD::MOVIshift";
679 case ARM64ISD::MOVIedit: return "ARM64ISD::MOVIedit";
680 case ARM64ISD::MOVImsl: return "ARM64ISD::MOVImsl";
681 case ARM64ISD::FMOV: return "ARM64ISD::FMOV";
682 case ARM64ISD::MVNIshift: return "ARM64ISD::MVNIshift";
683 case ARM64ISD::MVNImsl: return "ARM64ISD::MVNImsl";
684 case ARM64ISD::BICi: return "ARM64ISD::BICi";
685 case ARM64ISD::ORRi: return "ARM64ISD::ORRi";
686 case ARM64ISD::BSL: return "ARM64ISD::BSL";
687 case ARM64ISD::NEG: return "ARM64ISD::NEG";
688 case ARM64ISD::EXTR: return "ARM64ISD::EXTR";
689 case ARM64ISD::ZIP1: return "ARM64ISD::ZIP1";
690 case ARM64ISD::ZIP2: return "ARM64ISD::ZIP2";
691 case ARM64ISD::UZP1: return "ARM64ISD::UZP1";
692 case ARM64ISD::UZP2: return "ARM64ISD::UZP2";
693 case ARM64ISD::TRN1: return "ARM64ISD::TRN1";
694 case ARM64ISD::TRN2: return "ARM64ISD::TRN2";
695 case ARM64ISD::REV16: return "ARM64ISD::REV16";
696 case ARM64ISD::REV32: return "ARM64ISD::REV32";
697 case ARM64ISD::REV64: return "ARM64ISD::REV64";
698 case ARM64ISD::EXT: return "ARM64ISD::EXT";
699 case ARM64ISD::VSHL: return "ARM64ISD::VSHL";
700 case ARM64ISD::VLSHR: return "ARM64ISD::VLSHR";
701 case ARM64ISD::VASHR: return "ARM64ISD::VASHR";
702 case ARM64ISD::CMEQ: return "ARM64ISD::CMEQ";
703 case ARM64ISD::CMGE: return "ARM64ISD::CMGE";
704 case ARM64ISD::CMGT: return "ARM64ISD::CMGT";
705 case ARM64ISD::CMHI: return "ARM64ISD::CMHI";
706 case ARM64ISD::CMHS: return "ARM64ISD::CMHS";
707 case ARM64ISD::FCMEQ: return "ARM64ISD::FCMEQ";
708 case ARM64ISD::FCMGE: return "ARM64ISD::FCMGE";
709 case ARM64ISD::FCMGT: return "ARM64ISD::FCMGT";
710 case ARM64ISD::CMEQz: return "ARM64ISD::CMEQz";
711 case ARM64ISD::CMGEz: return "ARM64ISD::CMGEz";
712 case ARM64ISD::CMGTz: return "ARM64ISD::CMGTz";
713 case ARM64ISD::CMLEz: return "ARM64ISD::CMLEz";
714 case ARM64ISD::CMLTz: return "ARM64ISD::CMLTz";
715 case ARM64ISD::FCMEQz: return "ARM64ISD::FCMEQz";
716 case ARM64ISD::FCMGEz: return "ARM64ISD::FCMGEz";
717 case ARM64ISD::FCMGTz: return "ARM64ISD::FCMGTz";
718 case ARM64ISD::FCMLEz: return "ARM64ISD::FCMLEz";
719 case ARM64ISD::FCMLTz: return "ARM64ISD::FCMLTz";
720 case ARM64ISD::NOT: return "ARM64ISD::NOT";
721 case ARM64ISD::BIT: return "ARM64ISD::BIT";
722 case ARM64ISD::CBZ: return "ARM64ISD::CBZ";
723 case ARM64ISD::CBNZ: return "ARM64ISD::CBNZ";
724 case ARM64ISD::TBZ: return "ARM64ISD::TBZ";
725 case ARM64ISD::TBNZ: return "ARM64ISD::TBNZ";
726 case ARM64ISD::TC_RETURN: return "ARM64ISD::TC_RETURN";
727 case ARM64ISD::SITOF: return "ARM64ISD::SITOF";
728 case ARM64ISD::UITOF: return "ARM64ISD::UITOF";
729 case ARM64ISD::SQSHL_I: return "ARM64ISD::SQSHL_I";
730 case ARM64ISD::UQSHL_I: return "ARM64ISD::UQSHL_I";
731 case ARM64ISD::SRSHR_I: return "ARM64ISD::SRSHR_I";
732 case ARM64ISD::URSHR_I: return "ARM64ISD::URSHR_I";
733 case ARM64ISD::SQSHLU_I: return "ARM64ISD::SQSHLU_I";
734 case ARM64ISD::WrapperLarge: return "ARM64ISD::WrapperLarge";
735 case ARM64ISD::LD2post: return "ARM64ISD::LD2post";
736 case ARM64ISD::LD3post: return "ARM64ISD::LD3post";
737 case ARM64ISD::LD4post: return "ARM64ISD::LD4post";
738 case ARM64ISD::ST2post: return "ARM64ISD::ST2post";
739 case ARM64ISD::ST3post: return "ARM64ISD::ST3post";
740 case ARM64ISD::ST4post: return "ARM64ISD::ST4post";
741 case ARM64ISD::LD1x2post: return "ARM64ISD::LD1x2post";
742 case ARM64ISD::LD1x3post: return "ARM64ISD::LD1x3post";
743 case ARM64ISD::LD1x4post: return "ARM64ISD::LD1x4post";
744 case ARM64ISD::ST1x2post: return "ARM64ISD::ST1x2post";
745 case ARM64ISD::ST1x3post: return "ARM64ISD::ST1x3post";
746 case ARM64ISD::ST1x4post: return "ARM64ISD::ST1x4post";
747 case ARM64ISD::LD2DUPpost: return "ARM64ISD::LD2DUPpost";
748 case ARM64ISD::LD3DUPpost: return "ARM64ISD::LD3DUPpost";
749 case ARM64ISD::LD4DUPpost: return "ARM64ISD::LD4DUPpost";
750 case ARM64ISD::LD2LANEpost: return "ARM64ISD::LD2LANEpost";
751 case ARM64ISD::LD3LANEpost: return "ARM64ISD::LD3LANEpost";
752 case ARM64ISD::LD4LANEpost: return "ARM64ISD::LD4LANEpost";
753 case ARM64ISD::ST2LANEpost: return "ARM64ISD::ST2LANEpost";
754 case ARM64ISD::ST3LANEpost: return "ARM64ISD::ST3LANEpost";
755 case ARM64ISD::ST4LANEpost: return "ARM64ISD::ST4LANEpost";
760 ARM64TargetLowering::EmitF128CSEL(MachineInstr *MI,
761 MachineBasicBlock *MBB) const {
762 // We materialise the F128CSEL pseudo-instruction as some control flow and a
766 // [... previous instrs leading to comparison ...]
772 // Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB]
774 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
775 MachineFunction *MF = MBB->getParent();
776 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
777 DebugLoc DL = MI->getDebugLoc();
778 MachineFunction::iterator It = MBB;
781 unsigned DestReg = MI->getOperand(0).getReg();
782 unsigned IfTrueReg = MI->getOperand(1).getReg();
783 unsigned IfFalseReg = MI->getOperand(2).getReg();
784 unsigned CondCode = MI->getOperand(3).getImm();
785 bool NZCVKilled = MI->getOperand(4).isKill();
787 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
788 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
789 MF->insert(It, TrueBB);
790 MF->insert(It, EndBB);
792 // Transfer rest of current basic-block to EndBB
793 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
795 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
797 BuildMI(MBB, DL, TII->get(ARM64::Bcc)).addImm(CondCode).addMBB(TrueBB);
798 BuildMI(MBB, DL, TII->get(ARM64::B)).addMBB(EndBB);
799 MBB->addSuccessor(TrueBB);
800 MBB->addSuccessor(EndBB);
802 // TrueBB falls through to the end.
803 TrueBB->addSuccessor(EndBB);
806 TrueBB->addLiveIn(ARM64::NZCV);
807 EndBB->addLiveIn(ARM64::NZCV);
810 BuildMI(*EndBB, EndBB->begin(), DL, TII->get(ARM64::PHI), DestReg)
816 MI->eraseFromParent();
821 ARM64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
822 MachineBasicBlock *BB) const {
823 switch (MI->getOpcode()) {
828 assert(0 && "Unexpected instruction for custom inserter!");
831 case ARM64::F128CSEL:
832 return EmitF128CSEL(MI, BB);
834 case TargetOpcode::STACKMAP:
835 case TargetOpcode::PATCHPOINT:
836 return emitPatchPoint(MI, BB);
838 llvm_unreachable("Unexpected instruction for custom inserter!");
841 //===----------------------------------------------------------------------===//
842 // ARM64 Lowering private implementation.
843 //===----------------------------------------------------------------------===//
845 //===----------------------------------------------------------------------===//
847 //===----------------------------------------------------------------------===//
849 /// changeIntCCToARM64CC - Convert a DAG integer condition code to an ARM64 CC
850 static ARM64CC::CondCode changeIntCCToARM64CC(ISD::CondCode CC) {
853 llvm_unreachable("Unknown condition code!");
877 /// changeFPCCToARM64CC - Convert a DAG fp condition code to an ARM64 CC.
878 static void changeFPCCToARM64CC(ISD::CondCode CC, ARM64CC::CondCode &CondCode,
879 ARM64CC::CondCode &CondCode2) {
880 CondCode2 = ARM64CC::AL;
883 llvm_unreachable("Unknown FP condition!");
886 CondCode = ARM64CC::EQ;
890 CondCode = ARM64CC::GT;
894 CondCode = ARM64CC::GE;
897 CondCode = ARM64CC::MI;
900 CondCode = ARM64CC::LS;
903 CondCode = ARM64CC::MI;
904 CondCode2 = ARM64CC::GT;
907 CondCode = ARM64CC::VC;
910 CondCode = ARM64CC::VS;
913 CondCode = ARM64CC::EQ;
914 CondCode2 = ARM64CC::VS;
917 CondCode = ARM64CC::HI;
920 CondCode = ARM64CC::PL;
924 CondCode = ARM64CC::LT;
928 CondCode = ARM64CC::LE;
932 CondCode = ARM64CC::NE;
937 /// changeVectorFPCCToARM64CC - Convert a DAG fp condition code to an ARM64 CC
938 /// usable with the vector instructions. Fewer operations are available without
939 /// a real NZCV register, so we have to use less efficient combinations to get
941 static void changeVectorFPCCToARM64CC(ISD::CondCode CC,
942 ARM64CC::CondCode &CondCode,
943 ARM64CC::CondCode &CondCode2,
948 // Mostly the scalar mappings work fine.
949 changeFPCCToARM64CC(CC, CondCode, CondCode2);
952 Invert = true; // Fallthrough
954 CondCode = ARM64CC::MI;
955 CondCode2 = ARM64CC::GE;
962 // All of the compare-mask comparisons are ordered, but we can switch
963 // between the two by a double inversion. E.g. ULE == !OGT.
965 changeFPCCToARM64CC(getSetCCInverse(CC, false), CondCode, CondCode2);
970 static bool isLegalArithImmed(uint64_t C) {
971 // Matches ARM64DAGToDAGISel::SelectArithImmed().
972 return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
975 static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
976 SDLoc dl, SelectionDAG &DAG) {
977 EVT VT = LHS.getValueType();
979 if (VT.isFloatingPoint())
980 return DAG.getNode(ARM64ISD::FCMP, dl, VT, LHS, RHS);
982 // The CMP instruction is just an alias for SUBS, and representing it as
983 // SUBS means that it's possible to get CSE with subtract operations.
984 // A later phase can perform the optimization of setting the destination
985 // register to WZR/XZR if it ends up being unused.
986 unsigned Opcode = ARM64ISD::SUBS;
988 if (RHS.getOpcode() == ISD::SUB && isa<ConstantSDNode>(RHS.getOperand(0)) &&
989 cast<ConstantSDNode>(RHS.getOperand(0))->getZExtValue() == 0 &&
990 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
991 // We'd like to combine a (CMP op1, (sub 0, op2) into a CMN instruction on
992 // the grounds that "op1 - (-op2) == op1 + op2". However, the C and V flags
993 // can be set differently by this operation. It comes down to whether
994 // "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are then
995 // everything is fine. If not then the optimization is wrong. Thus general
996 // comparisons are only valid if op2 != 0.
998 // So, finally, the only LLVM-native comparisons that don't mention C and V
999 // are SETEQ and SETNE. They're the only ones we can safely use CMN for in
1000 // the absence of information about op2.
1001 Opcode = ARM64ISD::ADDS;
1002 RHS = RHS.getOperand(1);
1003 } else if (LHS.getOpcode() == ISD::AND && isa<ConstantSDNode>(RHS) &&
1004 cast<ConstantSDNode>(RHS)->getZExtValue() == 0 &&
1005 !isUnsignedIntSetCC(CC)) {
1006 // Similarly, (CMP (and X, Y), 0) can be implemented with a TST
1007 // (a.k.a. ANDS) except that the flags are only guaranteed to work for one
1008 // of the signed comparisons.
1009 Opcode = ARM64ISD::ANDS;
1010 RHS = LHS.getOperand(1);
1011 LHS = LHS.getOperand(0);
1014 return DAG.getNode(Opcode, dl, DAG.getVTList(VT, MVT::i32), LHS, RHS)
1018 static SDValue getARM64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1019 SDValue &ARM64cc, SelectionDAG &DAG, SDLoc dl) {
1020 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1021 EVT VT = RHS.getValueType();
1022 uint64_t C = RHSC->getZExtValue();
1023 if (!isLegalArithImmed(C)) {
1024 // Constant does not fit, try adjusting it by one?
1030 if ((VT == MVT::i32 && C != 0x80000000 &&
1031 isLegalArithImmed((uint32_t)(C - 1))) ||
1032 (VT == MVT::i64 && C != 0x80000000ULL &&
1033 isLegalArithImmed(C - 1ULL))) {
1034 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1035 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1036 RHS = DAG.getConstant(C, VT);
1041 if ((VT == MVT::i32 && C != 0 &&
1042 isLegalArithImmed((uint32_t)(C - 1))) ||
1043 (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) {
1044 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1045 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1046 RHS = DAG.getConstant(C, VT);
1051 if ((VT == MVT::i32 && C != 0x7fffffff &&
1052 isLegalArithImmed((uint32_t)(C + 1))) ||
1053 (VT == MVT::i64 && C != 0x7ffffffffffffffULL &&
1054 isLegalArithImmed(C + 1ULL))) {
1055 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1056 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1057 RHS = DAG.getConstant(C, VT);
1062 if ((VT == MVT::i32 && C != 0xffffffff &&
1063 isLegalArithImmed((uint32_t)(C + 1))) ||
1064 (VT == MVT::i64 && C != 0xfffffffffffffffULL &&
1065 isLegalArithImmed(C + 1ULL))) {
1066 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1067 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1068 RHS = DAG.getConstant(C, VT);
1075 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
1076 ARM64CC::CondCode ARM64CC = changeIntCCToARM64CC(CC);
1077 ARM64cc = DAG.getConstant(ARM64CC, MVT::i32);
1081 static std::pair<SDValue, SDValue>
1082 getARM64XALUOOp(ARM64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
1083 assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
1084 "Unsupported value type");
1085 SDValue Value, Overflow;
1087 SDValue LHS = Op.getOperand(0);
1088 SDValue RHS = Op.getOperand(1);
1090 switch (Op.getOpcode()) {
1092 llvm_unreachable("Unknown overflow instruction!");
1094 Opc = ARM64ISD::ADDS;
1098 Opc = ARM64ISD::ADDS;
1102 Opc = ARM64ISD::SUBS;
1106 Opc = ARM64ISD::SUBS;
1109 // Multiply needs a little bit extra work.
1113 bool IsSigned = (Op.getOpcode() == ISD::SMULO) ? true : false;
1114 if (Op.getValueType() == MVT::i32) {
1115 unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1116 // For a 32 bit multiply with overflow check we want the instruction
1117 // selector to generate a widening multiply (SMADDL/UMADDL). For that we
1118 // need to generate the following pattern:
1119 // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b))
1120 LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
1121 RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
1122 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1123 SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul,
1124 DAG.getConstant(0, MVT::i64));
1125 // On ARM64 the upper 32 bits are always zero extended for a 32 bit
1126 // operation. We need to clear out the upper 32 bits, because we used a
1127 // widening multiply that wrote all 64 bits. In the end this should be a
1129 Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add);
1131 // The signed overflow check requires more than just a simple check for
1132 // any bit set in the upper 32 bits of the result. These bits could be
1133 // just the sign bits of a negative number. To perform the overflow
1134 // check we have to arithmetic shift right the 32nd bit of the result by
1135 // 31 bits. Then we compare the result to the upper 32 bits.
1136 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add,
1137 DAG.getConstant(32, MVT::i64));
1138 UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits);
1139 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value,
1140 DAG.getConstant(31, MVT::i64));
1141 // It is important that LowerBits is last, otherwise the arithmetic
1142 // shift will not be folded into the compare (SUBS).
1143 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32);
1144 Overflow = DAG.getNode(ARM64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1147 // The overflow check for unsigned multiply is easy. We only need to
1148 // check if any of the upper 32 bits are set. This can be done with a
1149 // CMP (shifted register). For that we need to generate the following
1151 // (i64 ARM64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32)
1152 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
1153 DAG.getConstant(32, MVT::i64));
1154 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1156 DAG.getNode(ARM64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1157 UpperBits).getValue(1);
1161 assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
1162 // For the 64 bit multiply
1163 Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1165 SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
1166 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
1167 DAG.getConstant(63, MVT::i64));
1168 // It is important that LowerBits is last, otherwise the arithmetic
1169 // shift will not be folded into the compare (SUBS).
1170 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1171 Overflow = DAG.getNode(ARM64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1174 SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
1175 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1177 DAG.getNode(ARM64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1178 UpperBits).getValue(1);
1185 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);
1187 // Emit the ARM64 operation with overflow check.
1188 Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
1189 Overflow = Value.getValue(1);
1191 return std::make_pair(Value, Overflow);
1194 SDValue ARM64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG,
1195 RTLIB::Libcall Call) const {
1196 SmallVector<SDValue, 2> Ops;
1197 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1198 Ops.push_back(Op.getOperand(i));
1200 return makeLibCall(DAG, Call, MVT::f128, &Ops[0], Ops.size(), false,
1204 static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) {
1205 SDValue Sel = Op.getOperand(0);
1206 SDValue Other = Op.getOperand(1);
1208 // If neither operand is a SELECT_CC, give up.
1209 if (Sel.getOpcode() != ISD::SELECT_CC)
1210 std::swap(Sel, Other);
1211 if (Sel.getOpcode() != ISD::SELECT_CC)
1214 // The folding we want to perform is:
1215 // (xor x, (select_cc a, b, cc, 0, -1) )
1217 // (csel x, (xor x, -1), cc ...)
1219 // The latter will get matched to a CSINV instruction.
1221 ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
1222 SDValue LHS = Sel.getOperand(0);
1223 SDValue RHS = Sel.getOperand(1);
1224 SDValue TVal = Sel.getOperand(2);
1225 SDValue FVal = Sel.getOperand(3);
1228 // FIXME: This could be generalized to non-integer comparisons.
1229 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
1232 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
1233 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
1235 // The the values aren't constants, this isn't the pattern we're looking for.
1236 if (!CFVal || !CTVal)
1239 // We can commute the SELECT_CC by inverting the condition. This
1240 // might be needed to make this fit into a CSINV pattern.
1241 if (CTVal->isAllOnesValue() && CFVal->isNullValue()) {
1242 std::swap(TVal, FVal);
1243 std::swap(CTVal, CFVal);
1244 CC = ISD::getSetCCInverse(CC, true);
1247 // If the constants line up, perform the transform!
1248 if (CTVal->isNullValue() && CFVal->isAllOnesValue()) {
1250 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
1253 TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other,
1254 DAG.getConstant(-1ULL, Other.getValueType()));
1256 return DAG.getNode(ARM64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal,
1263 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
1264 EVT VT = Op.getValueType();
1266 // Let legalize expand this if it isn't a legal type yet.
1267 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
1270 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
1273 bool ExtraOp = false;
1274 switch (Op.getOpcode()) {
1276 assert(0 && "Invalid code");
1278 Opc = ARM64ISD::ADDS;
1281 Opc = ARM64ISD::SUBS;
1284 Opc = ARM64ISD::ADCS;
1288 Opc = ARM64ISD::SBCS;
1294 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1));
1295 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1),
1299 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
1300 // Let legalize expand this if it isn't a legal type yet.
1301 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
1304 ARM64CC::CondCode CC;
1305 // The actual operation that sets the overflow or carry flag.
1306 SDValue Value, Overflow;
1307 std::tie(Value, Overflow) = getARM64XALUOOp(CC, Op, DAG);
1309 // We use 0 and 1 as false and true values.
1310 SDValue TVal = DAG.getConstant(1, MVT::i32);
1311 SDValue FVal = DAG.getConstant(0, MVT::i32);
1313 // We use an inverted condition, because the conditional select is inverted
1314 // too. This will allow it to be selected to a single instruction:
1315 // CSINC Wd, WZR, WZR, invert(cond).
1316 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), MVT::i32);
1317 Overflow = DAG.getNode(ARM64ISD::CSEL, SDLoc(Op), MVT::i32, FVal, TVal, CCVal,
1320 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
1321 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
1324 // Prefetch operands are:
1325 // 1: Address to prefetch
1327 // 3: int locality (0 = no locality ... 3 = extreme locality)
1328 // 4: bool isDataCache
1329 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
1331 unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
1332 unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
1333 // The data thing is not used.
1334 // unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
1336 bool IsStream = !Locality;
1337 // When the locality number is set
1339 // The front-end should have filtered out the out-of-range values
1340 assert(Locality <= 3 && "Prefetch locality out-of-range");
1341 // The locality degree is the opposite of the cache speed.
1342 // Put the number the other way around.
1343 // The encoding starts at 0 for level 1
1344 Locality = 3 - Locality;
1347 // built the mask value encoding the expected behavior.
1348 unsigned PrfOp = (IsWrite << 4) | // Load/Store bit
1349 (Locality << 1) | // Cache level bits
1350 (unsigned)IsStream; // Stream bit
1351 return DAG.getNode(ARM64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
1352 DAG.getConstant(PrfOp, MVT::i32), Op.getOperand(1));
1355 SDValue ARM64TargetLowering::LowerFP_EXTEND(SDValue Op,
1356 SelectionDAG &DAG) const {
1357 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
1360 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
1362 return LowerF128Call(Op, DAG, LC);
1365 SDValue ARM64TargetLowering::LowerFP_ROUND(SDValue Op,
1366 SelectionDAG &DAG) const {
1367 if (Op.getOperand(0).getValueType() != MVT::f128) {
1368 // It's legal except when f128 is involved
1373 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
1375 // FP_ROUND node has a second operand indicating whether it is known to be
1376 // precise. That doesn't take part in the LibCall so we can't directly use
1378 SDValue SrcVal = Op.getOperand(0);
1379 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
1380 /*isSigned*/ false, SDLoc(Op)).first;
1383 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1384 // Warning: We maintain cost tables in ARM64TargetTransformInfo.cpp.
1385 // Any additional optimization in this function should be recorded
1386 // in the cost tables.
1387 EVT InVT = Op.getOperand(0).getValueType();
1388 EVT VT = Op.getValueType();
1390 // FP_TO_XINT conversion from the same type are legal.
1391 if (VT.getSizeInBits() == InVT.getSizeInBits())
1394 if (InVT == MVT::v2f64 || InVT == MVT::v4f32) {
1397 DAG.getNode(Op.getOpcode(), dl, InVT.changeVectorElementTypeToInteger(),
1399 return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
1400 } else if (InVT == MVT::v2f32) {
1402 SDValue Ext = DAG.getNode(ISD::FP_EXTEND, dl, MVT::v2f64, Op.getOperand(0));
1403 return DAG.getNode(Op.getOpcode(), dl, VT, Ext);
1406 // Type changing conversions are illegal.
1410 SDValue ARM64TargetLowering::LowerFP_TO_INT(SDValue Op,
1411 SelectionDAG &DAG) const {
1412 if (Op.getOperand(0).getValueType().isVector())
1413 return LowerVectorFP_TO_INT(Op, DAG);
1415 if (Op.getOperand(0).getValueType() != MVT::f128) {
1416 // It's legal except when f128 is involved
1421 if (Op.getOpcode() == ISD::FP_TO_SINT)
1422 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
1424 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
1426 SmallVector<SDValue, 2> Ops;
1427 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1428 Ops.push_back(Op.getOperand(i));
1430 return makeLibCall(DAG, LC, Op.getValueType(), &Ops[0], Ops.size(), false,
1434 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1435 // Warning: We maintain cost tables in ARM64TargetTransformInfo.cpp.
1436 // Any additional optimization in this function should be recorded
1437 // in the cost tables.
1438 EVT VT = Op.getValueType();
1440 SDValue In = Op.getOperand(0);
1441 EVT InVT = In.getValueType();
1443 // v2i32 to v2f32 is legal.
1444 if (VT == MVT::v2f32 && InVT == MVT::v2i32)
1447 // This function only handles v2f64 outputs.
1448 if (VT == MVT::v2f64) {
1449 // Extend the input argument to a v2i64 that we can feed into the
1450 // floating point conversion. Zero or sign extend based on whether
1451 // we're doing a signed or unsigned float conversion.
1453 Op.getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
1454 assert(Op.getNumOperands() == 1 && "FP conversions take one argument");
1455 SDValue Promoted = DAG.getNode(Opc, dl, MVT::v2i64, Op.getOperand(0));
1456 return DAG.getNode(Op.getOpcode(), dl, Op.getValueType(), Promoted);
1459 // Scalarize v2i64 to v2f32 conversions.
1460 std::vector<SDValue> BuildVectorOps;
1461 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
1462 SDValue Sclr = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, In,
1463 DAG.getConstant(i, MVT::i64));
1464 Sclr = DAG.getNode(Op->getOpcode(), dl, MVT::f32, Sclr);
1465 BuildVectorOps.push_back(Sclr);
1468 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, BuildVectorOps);
1471 SDValue ARM64TargetLowering::LowerINT_TO_FP(SDValue Op,
1472 SelectionDAG &DAG) const {
1473 if (Op.getValueType().isVector())
1474 return LowerVectorINT_TO_FP(Op, DAG);
1476 // i128 conversions are libcalls.
1477 if (Op.getOperand(0).getValueType() == MVT::i128)
1480 // Other conversions are legal, unless it's to the completely software-based
1482 if (Op.getValueType() != MVT::f128)
1486 if (Op.getOpcode() == ISD::SINT_TO_FP)
1487 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1489 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1491 return LowerF128Call(Op, DAG, LC);
1494 SDValue ARM64TargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
1495 // For iOS, we want to call an alternative entry point: __sincos_stret,
1496 // which returns the values in two S / D registers.
1498 SDValue Arg = Op.getOperand(0);
1499 EVT ArgVT = Arg.getValueType();
1500 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1507 Entry.isSExt = false;
1508 Entry.isZExt = false;
1509 Args.push_back(Entry);
1511 const char *LibcallName =
1512 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
1513 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
1515 StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
1516 TargetLowering::CallLoweringInfo CLI(
1517 DAG.getEntryNode(), RetTy, false, false, false, false, 0,
1518 CallingConv::Fast, /*isTaillCall=*/false,
1519 /*doesNotRet=*/false, /*isReturnValueUsed*/ true, Callee, Args, DAG, dl);
1520 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1521 return CallResult.first;
1524 SDValue ARM64TargetLowering::LowerOperation(SDValue Op,
1525 SelectionDAG &DAG) const {
1526 switch (Op.getOpcode()) {
1528 llvm_unreachable("unimplemented operand");
1530 case ISD::GlobalAddress:
1531 return LowerGlobalAddress(Op, DAG);
1532 case ISD::GlobalTLSAddress:
1533 return LowerGlobalTLSAddress(Op, DAG);
1535 return LowerSETCC(Op, DAG);
1537 return LowerBR_CC(Op, DAG);
1539 return LowerSELECT(Op, DAG);
1540 case ISD::SELECT_CC:
1541 return LowerSELECT_CC(Op, DAG);
1542 case ISD::JumpTable:
1543 return LowerJumpTable(Op, DAG);
1544 case ISD::ConstantPool:
1545 return LowerConstantPool(Op, DAG);
1546 case ISD::BlockAddress:
1547 return LowerBlockAddress(Op, DAG);
1549 return LowerVASTART(Op, DAG);
1551 return LowerVACOPY(Op, DAG);
1553 return LowerVAARG(Op, DAG);
1558 return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
1565 return LowerXALUO(Op, DAG);
1567 return LowerF128Call(Op, DAG, RTLIB::ADD_F128);
1569 return LowerF128Call(Op, DAG, RTLIB::SUB_F128);
1571 return LowerF128Call(Op, DAG, RTLIB::MUL_F128);
1573 return LowerF128Call(Op, DAG, RTLIB::DIV_F128);
1575 return LowerFP_ROUND(Op, DAG);
1576 case ISD::FP_EXTEND:
1577 return LowerFP_EXTEND(Op, DAG);
1578 case ISD::FRAMEADDR:
1579 return LowerFRAMEADDR(Op, DAG);
1580 case ISD::RETURNADDR:
1581 return LowerRETURNADDR(Op, DAG);
1582 case ISD::INSERT_VECTOR_ELT:
1583 return LowerINSERT_VECTOR_ELT(Op, DAG);
1584 case ISD::EXTRACT_VECTOR_ELT:
1585 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
1586 case ISD::BUILD_VECTOR:
1587 return LowerBUILD_VECTOR(Op, DAG);
1588 case ISD::VECTOR_SHUFFLE:
1589 return LowerVECTOR_SHUFFLE(Op, DAG);
1590 case ISD::EXTRACT_SUBVECTOR:
1591 return LowerEXTRACT_SUBVECTOR(Op, DAG);
1595 return LowerVectorSRA_SRL_SHL(Op, DAG);
1596 case ISD::SHL_PARTS:
1597 return LowerShiftLeftParts(Op, DAG);
1598 case ISD::SRL_PARTS:
1599 case ISD::SRA_PARTS:
1600 return LowerShiftRightParts(Op, DAG);
1602 return LowerCTPOP(Op, DAG);
1603 case ISD::FCOPYSIGN:
1604 return LowerFCOPYSIGN(Op, DAG);
1606 return LowerVectorAND(Op, DAG);
1608 return LowerVectorOR(Op, DAG);
1610 return LowerXOR(Op, DAG);
1612 return LowerPREFETCH(Op, DAG);
1613 case ISD::SINT_TO_FP:
1614 case ISD::UINT_TO_FP:
1615 return LowerINT_TO_FP(Op, DAG);
1616 case ISD::FP_TO_SINT:
1617 case ISD::FP_TO_UINT:
1618 return LowerFP_TO_INT(Op, DAG);
1620 return LowerFSINCOS(Op, DAG);
1624 /// getFunctionAlignment - Return the Log2 alignment of this function.
1625 unsigned ARM64TargetLowering::getFunctionAlignment(const Function *F) const {
1629 //===----------------------------------------------------------------------===//
1630 // Calling Convention Implementation
1631 //===----------------------------------------------------------------------===//
1633 #include "ARM64GenCallingConv.inc"
1635 /// Selects the correct CCAssignFn for a the given CallingConvention
1637 CCAssignFn *ARM64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
1638 bool IsVarArg) const {
1641 llvm_unreachable("Unsupported calling convention.");
1642 case CallingConv::WebKit_JS:
1643 return CC_ARM64_WebKit_JS;
1644 case CallingConv::C:
1645 case CallingConv::Fast:
1646 if (!Subtarget->isTargetDarwin())
1647 return CC_ARM64_AAPCS;
1648 return IsVarArg ? CC_ARM64_DarwinPCS_VarArg : CC_ARM64_DarwinPCS;
1652 SDValue ARM64TargetLowering::LowerFormalArguments(
1653 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1654 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
1655 SmallVectorImpl<SDValue> &InVals) const {
1656 MachineFunction &MF = DAG.getMachineFunction();
1657 MachineFrameInfo *MFI = MF.getFrameInfo();
1659 // Assign locations to all of the incoming arguments.
1660 SmallVector<CCValAssign, 16> ArgLocs;
1661 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1662 getTargetMachine(), ArgLocs, *DAG.getContext());
1664 // At this point, Ins[].VT may already be promoted to i32. To correctly
1665 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
1666 // i8 to CC_ARM64_AAPCS with i32 being ValVT and i8 being LocVT.
1667 // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
1668 // we use a special version of AnalyzeFormalArguments to pass in ValVT and
1670 unsigned NumArgs = Ins.size();
1671 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
1672 unsigned CurArgIdx = 0;
1673 for (unsigned i = 0; i != NumArgs; ++i) {
1674 MVT ValVT = Ins[i].VT;
1675 std::advance(CurOrigArg, Ins[i].OrigArgIndex - CurArgIdx);
1676 CurArgIdx = Ins[i].OrigArgIndex;
1678 // Get type of the original argument.
1679 EVT ActualVT = getValueType(CurOrigArg->getType(), /*AllowUnknown*/ true);
1680 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
1681 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
1683 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
1685 else if (ActualMVT == MVT::i16)
1688 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
1690 AssignFn(i, ValVT, LocVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
1691 assert(!Res && "Call operand has unhandled type");
1694 assert(ArgLocs.size() == Ins.size());
1695 SmallVector<SDValue, 16> ArgValues;
1696 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1697 CCValAssign &VA = ArgLocs[i];
1699 if (Ins[i].Flags.isByVal()) {
1700 // Byval is used for HFAs in the PCS, but the system should work in a
1701 // non-compliant manner for larger structs.
1702 EVT PtrTy = getPointerTy();
1703 int Size = Ins[i].Flags.getByValSize();
1704 unsigned NumRegs = (Size + 7) / 8;
1706 // FIXME: This works on big-endian for composite byvals, which are the common
1707 // case. It should also work for fundamental types too.
1709 MFI->CreateFixedObject(8 * NumRegs, VA.getLocMemOffset(), false);
1710 SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrTy);
1711 InVals.push_back(FrameIdxN);
1714 } if (VA.isRegLoc()) {
1715 // Arguments stored in registers.
1716 EVT RegVT = VA.getLocVT();
1719 const TargetRegisterClass *RC;
1721 if (RegVT == MVT::i32)
1722 RC = &ARM64::GPR32RegClass;
1723 else if (RegVT == MVT::i64)
1724 RC = &ARM64::GPR64RegClass;
1725 else if (RegVT == MVT::f32)
1726 RC = &ARM64::FPR32RegClass;
1727 else if (RegVT == MVT::f64 || RegVT.is64BitVector())
1728 RC = &ARM64::FPR64RegClass;
1729 else if (RegVT == MVT::f128 || RegVT.is128BitVector())
1730 RC = &ARM64::FPR128RegClass;
1732 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1734 // Transform the arguments in physical registers into virtual ones.
1735 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1736 ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
1738 // If this is an 8, 16 or 32-bit value, it is really passed promoted
1739 // to 64 bits. Insert an assert[sz]ext to capture this, then
1740 // truncate to the right size.
1741 switch (VA.getLocInfo()) {
1743 llvm_unreachable("Unknown loc info!");
1744 case CCValAssign::Full:
1746 case CCValAssign::BCvt:
1747 ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue);
1749 case CCValAssign::SExt:
1750 ArgValue = DAG.getNode(ISD::AssertSext, DL, RegVT, ArgValue,
1751 DAG.getValueType(VA.getValVT()));
1752 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), ArgValue);
1754 case CCValAssign::ZExt:
1755 ArgValue = DAG.getNode(ISD::AssertZext, DL, RegVT, ArgValue,
1756 DAG.getValueType(VA.getValVT()));
1757 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), ArgValue);
1761 InVals.push_back(ArgValue);
1763 } else { // VA.isRegLoc()
1764 assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
1765 unsigned ArgOffset = VA.getLocMemOffset();
1766 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1768 uint32_t BEAlign = 0;
1769 if (ArgSize < 8 && !Subtarget->isLittleEndian())
1770 BEAlign = 8 - ArgSize;
1772 int FI = MFI->CreateFixedObject(ArgSize, ArgOffset + BEAlign, true);
1774 // Create load nodes to retrieve arguments from the stack.
1775 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1778 // If the loc type and val type are not the same, create an anyext load.
1779 if (VA.getLocVT().getSizeInBits() != VA.getValVT().getSizeInBits()) {
1780 // We should only get here if this is a pure integer.
1781 assert(!VA.getValVT().isVector() && VA.getValVT().isInteger() &&
1782 "Only integer extension supported!");
1783 ArgValue = DAG.getExtLoad(ISD::EXTLOAD, DL, VA.getValVT(), Chain, FIN,
1784 MachinePointerInfo::getFixedStack(FI),
1786 false, false, false, 0);
1788 ArgValue = DAG.getLoad(VA.getValVT(), DL, Chain, FIN,
1789 MachinePointerInfo::getFixedStack(FI), false,
1793 InVals.push_back(ArgValue);
1799 if (!Subtarget->isTargetDarwin()) {
1800 // The AAPCS variadic function ABI is identical to the non-variadic
1801 // one. As a result there may be more arguments in registers and we should
1802 // save them for future reference.
1803 saveVarArgRegisters(CCInfo, DAG, DL, Chain);
1806 ARM64FunctionInfo *AFI = MF.getInfo<ARM64FunctionInfo>();
1807 // This will point to the next argument passed via stack.
1808 unsigned StackOffset = CCInfo.getNextStackOffset();
1809 // We currently pass all varargs at 8-byte alignment.
1810 StackOffset = ((StackOffset + 7) & ~7);
1811 AFI->setVarArgsStackIndex(MFI->CreateFixedObject(4, StackOffset, true));
1817 void ARM64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
1818 SelectionDAG &DAG, SDLoc DL,
1819 SDValue &Chain) const {
1820 MachineFunction &MF = DAG.getMachineFunction();
1821 MachineFrameInfo *MFI = MF.getFrameInfo();
1822 ARM64FunctionInfo *FuncInfo = MF.getInfo<ARM64FunctionInfo>();
1824 SmallVector<SDValue, 8> MemOps;
1826 static const MCPhysReg GPRArgRegs[] = { ARM64::X0, ARM64::X1, ARM64::X2,
1827 ARM64::X3, ARM64::X4, ARM64::X5,
1828 ARM64::X6, ARM64::X7 };
1829 static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs);
1830 unsigned FirstVariadicGPR =
1831 CCInfo.getFirstUnallocated(GPRArgRegs, NumGPRArgRegs);
1833 unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
1835 if (GPRSaveSize != 0) {
1836 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
1838 SDValue FIN = DAG.getFrameIndex(GPRIdx, getPointerTy());
1840 for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
1841 unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &ARM64::GPR64RegClass);
1842 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
1844 DAG.getStore(Val.getValue(1), DL, Val, FIN,
1845 MachinePointerInfo::getStack(i * 8), false, false, 0);
1846 MemOps.push_back(Store);
1847 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
1848 DAG.getConstant(8, getPointerTy()));
1851 FuncInfo->setVarArgsGPRIndex(GPRIdx);
1852 FuncInfo->setVarArgsGPRSize(GPRSaveSize);
1854 if (Subtarget->hasFPARMv8()) {
1855 static const MCPhysReg FPRArgRegs[] = { ARM64::Q0, ARM64::Q1, ARM64::Q2,
1856 ARM64::Q3, ARM64::Q4, ARM64::Q5,
1857 ARM64::Q6, ARM64::Q7 };
1858 static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs);
1859 unsigned FirstVariadicFPR =
1860 CCInfo.getFirstUnallocated(FPRArgRegs, NumFPRArgRegs);
1862 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
1864 if (FPRSaveSize != 0) {
1865 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
1867 SDValue FIN = DAG.getFrameIndex(FPRIdx, getPointerTy());
1869 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
1870 unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &ARM64::FPR128RegClass);
1871 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f128);
1874 DAG.getStore(Val.getValue(1), DL, Val, FIN,
1875 MachinePointerInfo::getStack(i * 16), false, false, 0);
1876 MemOps.push_back(Store);
1877 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
1878 DAG.getConstant(16, getPointerTy()));
1881 FuncInfo->setVarArgsFPRIndex(FPRIdx);
1882 FuncInfo->setVarArgsFPRSize(FPRSaveSize);
1885 if (!MemOps.empty()) {
1886 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
1890 /// LowerCallResult - Lower the result values of a call into the
1891 /// appropriate copies out of appropriate physical registers.
1892 SDValue ARM64TargetLowering::LowerCallResult(
1893 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
1894 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
1895 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
1896 SDValue ThisVal) const {
1897 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
1898 : RetCC_ARM64_AAPCS;
1899 // Assign locations to each value returned by this call.
1900 SmallVector<CCValAssign, 16> RVLocs;
1901 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1902 getTargetMachine(), RVLocs, *DAG.getContext());
1903 CCInfo.AnalyzeCallResult(Ins, RetCC);
1905 // Copy all of the result registers out of their specified physreg.
1906 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1907 CCValAssign VA = RVLocs[i];
1909 // Pass 'this' value directly from the argument to return value, to avoid
1910 // reg unit interference
1911 if (i == 0 && isThisReturn) {
1912 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 &&
1913 "unexpected return calling convention register assignment");
1914 InVals.push_back(ThisVal);
1919 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag);
1920 Chain = Val.getValue(1);
1921 InFlag = Val.getValue(2);
1923 switch (VA.getLocInfo()) {
1925 llvm_unreachable("Unknown loc info!");
1926 case CCValAssign::Full:
1928 case CCValAssign::BCvt:
1929 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
1933 InVals.push_back(Val);
1939 bool ARM64TargetLowering::isEligibleForTailCallOptimization(
1940 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
1941 bool isCalleeStructRet, bool isCallerStructRet,
1942 const SmallVectorImpl<ISD::OutputArg> &Outs,
1943 const SmallVectorImpl<SDValue> &OutVals,
1944 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
1945 // Look for obvious safe cases to perform tail call optimization that do not
1946 // require ABI changes. This is what gcc calls sibcall.
1948 // Do not sibcall optimize vararg calls unless the call site is not passing
1950 if (isVarArg && !Outs.empty())
1953 // Also avoid sibcall optimization if either caller or callee uses struct
1954 // return semantics.
1955 if (isCalleeStructRet || isCallerStructRet)
1958 // Note that currently ARM64 "C" calling convention and "Fast" calling
1959 // convention are compatible. If/when that ever changes, we'll need to
1960 // add checks here to make sure any interactions are OK.
1962 // If the callee takes no arguments then go on to check the results of the
1964 if (!Outs.empty()) {
1965 // Check if stack adjustment is needed. For now, do not do this if any
1966 // argument is passed on the stack.
1967 SmallVector<CCValAssign, 16> ArgLocs;
1968 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1969 getTargetMachine(), ArgLocs, *DAG.getContext());
1970 CCAssignFn *AssignFn = CCAssignFnForCall(CalleeCC, /*IsVarArg=*/false);
1971 CCInfo.AnalyzeCallOperands(Outs, AssignFn);
1972 if (CCInfo.getNextStackOffset()) {
1973 // Check if the arguments are already laid out in the right way as
1974 // the caller's fixed stack objects.
1975 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
1976 ++i, ++realArgIdx) {
1977 CCValAssign &VA = ArgLocs[i];
1978 if (VA.getLocInfo() == CCValAssign::Indirect)
1980 if (VA.needsCustom()) {
1981 // Just don't handle anything that needs custom adjustments for now.
1982 // If need be, we can revisit later, but we shouldn't ever end up
1985 } else if (!VA.isRegLoc()) {
1986 // Likewise, don't try to handle stack based arguments for the
1996 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
1997 /// and add input and output parameter nodes.
1998 SDValue ARM64TargetLowering::LowerCall(CallLoweringInfo &CLI,
1999 SmallVectorImpl<SDValue> &InVals) const {
2000 SelectionDAG &DAG = CLI.DAG;
2002 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2003 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2004 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2005 SDValue Chain = CLI.Chain;
2006 SDValue Callee = CLI.Callee;
2007 bool &IsTailCall = CLI.IsTailCall;
2008 CallingConv::ID CallConv = CLI.CallConv;
2009 bool IsVarArg = CLI.IsVarArg;
2011 MachineFunction &MF = DAG.getMachineFunction();
2012 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
2013 bool IsThisReturn = false;
2015 // If tail calls are explicitly disabled, make sure not to use them.
2016 if (!EnableARM64TailCalls)
2020 // Check if it's really possible to do a tail call.
2021 IsTailCall = isEligibleForTailCallOptimization(
2022 Callee, CallConv, IsVarArg, IsStructRet,
2023 MF.getFunction()->hasStructRetAttr(), Outs, OutVals, Ins, DAG);
2024 if (!IsTailCall && CLI.CS && CLI.CS->isMustTailCall())
2025 report_fatal_error("failed to perform tail call elimination on a call "
2026 "site marked musttail");
2027 // We don't support GuaranteedTailCallOpt, only automatically
2028 // detected sibcalls.
2029 // FIXME: Re-evaluate. Is this true? Should it be true?
2034 // Analyze operands of the call, assigning locations to each operand.
2035 SmallVector<CCValAssign, 16> ArgLocs;
2036 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2037 getTargetMachine(), ArgLocs, *DAG.getContext());
2040 // Handle fixed and variable vector arguments differently.
2041 // Variable vector arguments always go into memory.
2042 unsigned NumArgs = Outs.size();
2044 for (unsigned i = 0; i != NumArgs; ++i) {
2045 MVT ArgVT = Outs[i].VT;
2046 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2047 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
2048 /*IsVarArg=*/ !Outs[i].IsFixed);
2049 bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
2050 assert(!Res && "Call operand has unhandled type");
2054 // At this point, Outs[].VT may already be promoted to i32. To correctly
2055 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2056 // i8 to CC_ARM64_AAPCS with i32 being ValVT and i8 being LocVT.
2057 // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here
2058 // we use a special version of AnalyzeCallOperands to pass in ValVT and
2060 unsigned NumArgs = Outs.size();
2061 for (unsigned i = 0; i != NumArgs; ++i) {
2062 MVT ValVT = Outs[i].VT;
2063 // Get type of the original argument.
2064 EVT ActualVT = getValueType(CLI.Args[Outs[i].OrigArgIndex].Ty,
2065 /*AllowUnknown*/ true);
2066 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT;
2067 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2068 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2070 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2072 else if (ActualMVT == MVT::i16)
2075 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2076 bool Res = AssignFn(i, ValVT, LocVT, CCValAssign::Full, ArgFlags, CCInfo);
2077 assert(!Res && "Call operand has unhandled type");
2082 // Get a count of how many bytes are to be pushed on the stack.
2083 unsigned NumBytes = CCInfo.getNextStackOffset();
2085 // Adjust the stack pointer for the new arguments...
2086 // These operations are automatically eliminated by the prolog/epilog pass
2089 DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true), DL);
2091 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, ARM64::SP, getPointerTy());
2093 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2094 SmallVector<SDValue, 8> MemOpChains;
2096 // Walk the register/memloc assignments, inserting copies/loads.
2097 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2098 ++i, ++realArgIdx) {
2099 CCValAssign &VA = ArgLocs[i];
2100 SDValue Arg = OutVals[realArgIdx];
2101 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2103 // Promote the value if needed.
2104 switch (VA.getLocInfo()) {
2106 llvm_unreachable("Unknown loc info!");
2107 case CCValAssign::Full:
2109 case CCValAssign::SExt:
2110 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
2112 case CCValAssign::ZExt:
2113 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2115 case CCValAssign::AExt:
2116 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
2118 case CCValAssign::BCvt:
2119 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2121 case CCValAssign::FPExt:
2122 Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
2126 if (VA.isRegLoc()) {
2127 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i64) {
2128 assert(VA.getLocVT() == MVT::i64 &&
2129 "unexpected calling convention register assignment");
2130 assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
2131 "unexpected use of 'returned'");
2132 IsThisReturn = true;
2134 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2136 assert(VA.isMemLoc());
2137 // There's no reason we can't support stack args w/ tailcall, but
2138 // we currently don't, so assert if we see one.
2139 assert(!IsTailCall && "stack argument with tail call!?");
2141 // FIXME: This works on big-endian for composite byvals, which are the common
2142 // case. It should also work for fundamental types too.
2143 uint32_t BEAlign = 0;
2144 if (!Subtarget->isLittleEndian() && !Flags.isByVal()) {
2145 unsigned OpSize = (VA.getLocVT().getSizeInBits() + 7) / 8;
2147 BEAlign = 8 - OpSize;
2150 unsigned LocMemOffset = VA.getLocMemOffset();
2151 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset + BEAlign);
2152 PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff);
2154 if (Outs[i].Flags.isByVal()) {
2156 DAG.getConstant(Outs[i].Flags.getByValSize(), MVT::i64);
2157 SDValue Cpy = DAG.getMemcpy(
2158 Chain, DL, PtrOff, Arg, SizeNode, Outs[i].Flags.getByValAlign(),
2159 /*isVolatile = */ false,
2160 /*alwaysInline = */ false,
2161 MachinePointerInfo::getStack(LocMemOffset), MachinePointerInfo());
2163 MemOpChains.push_back(Cpy);
2165 // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
2166 // promoted to a legal register type i32, we should truncate Arg back to
2168 if (Arg.getValueType().isSimple() &&
2169 Arg.getValueType().getSimpleVT() == MVT::i32 &&
2170 (VA.getLocVT() == MVT::i1 || VA.getLocVT() == MVT::i8 ||
2171 VA.getLocVT() == MVT::i16))
2172 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getLocVT(), Arg);
2174 SDValue Store = DAG.getStore(Chain, DL, Arg, PtrOff,
2175 MachinePointerInfo::getStack(LocMemOffset),
2177 MemOpChains.push_back(Store);
2182 if (!MemOpChains.empty())
2183 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
2185 // Build a sequence of copy-to-reg nodes chained together with token chain
2186 // and flag operands which copy the outgoing args into the appropriate regs.
2188 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2189 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[i].first,
2190 RegsToPass[i].second, InFlag);
2191 InFlag = Chain.getValue(1);
2194 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2195 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2196 // node so that legalize doesn't hack it.
2197 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
2198 Subtarget->isTargetMachO()) {
2199 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2200 const GlobalValue *GV = G->getGlobal();
2201 bool InternalLinkage = GV->hasInternalLinkage();
2202 if (InternalLinkage)
2203 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2205 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0,
2207 Callee = DAG.getNode(ARM64ISD::LOADgot, DL, getPointerTy(), Callee);
2209 } else if (ExternalSymbolSDNode *S =
2210 dyn_cast<ExternalSymbolSDNode>(Callee)) {
2211 const char *Sym = S->getSymbol();
2213 DAG.getTargetExternalSymbol(Sym, getPointerTy(), ARM64II::MO_GOT);
2214 Callee = DAG.getNode(ARM64ISD::LOADgot, DL, getPointerTy(), Callee);
2216 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2217 const GlobalValue *GV = G->getGlobal();
2218 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2219 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2220 const char *Sym = S->getSymbol();
2221 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), 0);
2224 std::vector<SDValue> Ops;
2225 Ops.push_back(Chain);
2226 Ops.push_back(Callee);
2228 // Add argument registers to the end of the list so that they are known live
2230 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2231 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2232 RegsToPass[i].second.getValueType()));
2234 // Add a register mask operand representing the call-preserved registers.
2235 const uint32_t *Mask;
2236 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2237 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2239 // For 'this' returns, use the X0-preserving mask if applicable
2240 Mask = ARI->getThisReturnPreservedMask(CallConv);
2242 IsThisReturn = false;
2243 Mask = ARI->getCallPreservedMask(CallConv);
2246 Mask = ARI->getCallPreservedMask(CallConv);
2248 assert(Mask && "Missing call preserved mask for calling convention");
2249 Ops.push_back(DAG.getRegisterMask(Mask));
2251 if (InFlag.getNode())
2252 Ops.push_back(InFlag);
2254 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2256 // If we're doing a tall call, use a TC_RETURN here rather than an
2257 // actual call instruction.
2259 return DAG.getNode(ARM64ISD::TC_RETURN, DL, NodeTys, Ops);
2261 // Returns a chain and a flag for retval copy to use.
2262 Chain = DAG.getNode(ARM64ISD::CALL, DL, NodeTys, Ops);
2263 InFlag = Chain.getValue(1);
2265 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2266 DAG.getIntPtrConstant(0, true), InFlag, DL);
2268 InFlag = Chain.getValue(1);
2270 // Handle result values, copying them out of physregs into vregs that we
2272 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
2273 InVals, IsThisReturn,
2274 IsThisReturn ? OutVals[0] : SDValue());
2277 bool ARM64TargetLowering::CanLowerReturn(
2278 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2279 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2280 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2281 : RetCC_ARM64_AAPCS;
2282 SmallVector<CCValAssign, 16> RVLocs;
2283 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
2284 return CCInfo.CheckReturn(Outs, RetCC);
2288 ARM64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2290 const SmallVectorImpl<ISD::OutputArg> &Outs,
2291 const SmallVectorImpl<SDValue> &OutVals,
2292 SDLoc DL, SelectionDAG &DAG) const {
2293 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2294 : RetCC_ARM64_AAPCS;
2295 SmallVector<CCValAssign, 16> RVLocs;
2296 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2297 getTargetMachine(), RVLocs, *DAG.getContext());
2298 CCInfo.AnalyzeReturn(Outs, RetCC);
2300 // Copy the result values into the output registers.
2302 SmallVector<SDValue, 4> RetOps(1, Chain);
2303 for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size();
2304 ++i, ++realRVLocIdx) {
2305 CCValAssign &VA = RVLocs[i];
2306 assert(VA.isRegLoc() && "Can only return in registers!");
2307 SDValue Arg = OutVals[realRVLocIdx];
2309 switch (VA.getLocInfo()) {
2311 llvm_unreachable("Unknown loc info!");
2312 case CCValAssign::Full:
2314 case CCValAssign::BCvt:
2315 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2319 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
2320 Flag = Chain.getValue(1);
2321 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2324 RetOps[0] = Chain; // Update chain.
2326 // Add the flag if we have it.
2328 RetOps.push_back(Flag);
2330 return DAG.getNode(ARM64ISD::RET_FLAG, DL, MVT::Other, RetOps);
2333 //===----------------------------------------------------------------------===//
2334 // Other Lowering Code
2335 //===----------------------------------------------------------------------===//
2337 SDValue ARM64TargetLowering::LowerGlobalAddress(SDValue Op,
2338 SelectionDAG &DAG) const {
2339 EVT PtrVT = getPointerTy();
2341 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2342 unsigned char OpFlags =
2343 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
2345 assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
2346 "unexpected offset in global node");
2348 // This also catched the large code model case for Darwin.
2349 if ((OpFlags & ARM64II::MO_GOT) != 0) {
2350 SDValue GotAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
2351 // FIXME: Once remat is capable of dealing with instructions with register
2352 // operands, expand this into two nodes instead of using a wrapper node.
2353 return DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, GotAddr);
2356 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2357 const unsigned char MO_NC = ARM64II::MO_NC;
2359 ARM64ISD::WrapperLarge, DL, PtrVT,
2360 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G3),
2361 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G2 | MO_NC),
2362 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G1 | MO_NC),
2363 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G0 | MO_NC));
2365 // Use ADRP/ADD or ADRP/LDR for everything else: the small model on ELF and
2366 // the only correct model on Darwin.
2367 SDValue Hi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2368 OpFlags | ARM64II::MO_PAGE);
2369 unsigned char LoFlags = OpFlags | ARM64II::MO_PAGEOFF | ARM64II::MO_NC;
2370 SDValue Lo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, LoFlags);
2372 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
2373 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
2377 /// \brief Convert a TLS address reference into the correct sequence of loads
2378 /// and calls to compute the variable's address (for Darwin, currently) and
2379 /// return an SDValue containing the final node.
2381 /// Darwin only has one TLS scheme which must be capable of dealing with the
2382 /// fully general situation, in the worst case. This means:
2383 /// + "extern __thread" declaration.
2384 /// + Defined in a possibly unknown dynamic library.
2386 /// The general system is that each __thread variable has a [3 x i64] descriptor
2387 /// which contains information used by the runtime to calculate the address. The
2388 /// only part of this the compiler needs to know about is the first xword, which
2389 /// contains a function pointer that must be called with the address of the
2390 /// entire descriptor in "x0".
2392 /// Since this descriptor may be in a different unit, in general even the
2393 /// descriptor must be accessed via an indirect load. The "ideal" code sequence
2395 /// adrp x0, _var@TLVPPAGE
2396 /// ldr x0, [x0, _var@TLVPPAGEOFF] ; x0 now contains address of descriptor
2397 /// ldr x1, [x0] ; x1 contains 1st entry of descriptor,
2398 /// ; the function pointer
2399 /// blr x1 ; Uses descriptor address in x0
2400 /// ; Address of _var is now in x0.
2402 /// If the address of _var's descriptor *is* known to the linker, then it can
2403 /// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
2404 /// a slight efficiency gain.
2406 ARM64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
2407 SelectionDAG &DAG) const {
2408 assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");
2411 MVT PtrVT = getPointerTy();
2412 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2415 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2416 SDValue DescAddr = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, TLVPAddr);
2418 // The first entry in the descriptor is a function pointer that we must call
2419 // to obtain the address of the variable.
2420 SDValue Chain = DAG.getEntryNode();
2421 SDValue FuncTLVGet =
2422 DAG.getLoad(MVT::i64, DL, Chain, DescAddr, MachinePointerInfo::getGOT(),
2423 false, true, true, 8);
2424 Chain = FuncTLVGet.getValue(1);
2426 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2427 MFI->setAdjustsStack(true);
2429 // TLS calls preserve all registers except those that absolutely must be
2430 // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
2432 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2433 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2434 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
2436 // Finally, we can make the call. This is just a degenerate version of a
2437 // normal ARM64 call node: x0 takes the address of the descriptor, and returns
2438 // the address of the variable in this thread.
2439 Chain = DAG.getCopyToReg(Chain, DL, ARM64::X0, DescAddr, SDValue());
2440 Chain = DAG.getNode(ARM64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
2441 Chain, FuncTLVGet, DAG.getRegister(ARM64::X0, MVT::i64),
2442 DAG.getRegisterMask(Mask), Chain.getValue(1));
2443 return DAG.getCopyFromReg(Chain, DL, ARM64::X0, PtrVT, Chain.getValue(1));
2446 /// When accessing thread-local variables under either the general-dynamic or
2447 /// local-dynamic system, we make a "TLS-descriptor" call. The variable will
2448 /// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
2449 /// is a function pointer to carry out the resolution. This function takes the
2450 /// address of the descriptor in X0 and returns the TPIDR_EL0 offset in X0. All
2451 /// other registers (except LR, NZCV) are preserved.
2453 /// Thus, the ideal call sequence on AArch64 is:
2455 /// adrp x0, :tlsdesc:thread_var
2456 /// ldr x8, [x0, :tlsdesc_lo12:thread_var]
2457 /// add x0, x0, :tlsdesc_lo12:thread_var
2458 /// .tlsdesccall thread_var
2460 /// (TPIDR_EL0 offset now in x0).
2462 /// The ".tlsdesccall" directive instructs the assembler to insert a particular
2463 /// relocation to help the linker relax this sequence if it turns out to be too
2466 /// FIXME: we currently produce an extra, duplicated, ADRP instruction, but this
2468 SDValue ARM64TargetLowering::LowerELFTLSDescCall(SDValue SymAddr,
2469 SDValue DescAddr, SDLoc DL,
2470 SelectionDAG &DAG) const {
2471 EVT PtrVT = getPointerTy();
2473 // The function we need to call is simply the first entry in the GOT for this
2474 // descriptor, load it in preparation.
2475 SDValue Func = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, SymAddr);
2477 // TLS calls preserve all registers except those that absolutely must be
2478 // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
2480 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2481 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2482 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
2484 // The function takes only one argument: the address of the descriptor itself
2486 SDValue Glue, Chain;
2487 Chain = DAG.getCopyToReg(DAG.getEntryNode(), DL, ARM64::X0, DescAddr, Glue);
2488 Glue = Chain.getValue(1);
2490 // We're now ready to populate the argument list, as with a normal call:
2491 SmallVector<SDValue, 6> Ops;
2492 Ops.push_back(Chain);
2493 Ops.push_back(Func);
2494 Ops.push_back(SymAddr);
2495 Ops.push_back(DAG.getRegister(ARM64::X0, PtrVT));
2496 Ops.push_back(DAG.getRegisterMask(Mask));
2497 Ops.push_back(Glue);
2499 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2500 Chain = DAG.getNode(ARM64ISD::TLSDESC_CALL, DL, NodeTys, Ops);
2501 Glue = Chain.getValue(1);
2503 return DAG.getCopyFromReg(Chain, DL, ARM64::X0, PtrVT, Glue);
2506 SDValue ARM64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
2507 SelectionDAG &DAG) const {
2508 assert(Subtarget->isTargetELF() && "This function expects an ELF target");
2509 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
2510 "ELF TLS only supported in small memory model");
2511 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2513 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
2516 EVT PtrVT = getPointerTy();
2518 const GlobalValue *GV = GA->getGlobal();
2520 SDValue ThreadBase = DAG.getNode(ARM64ISD::THREAD_POINTER, DL, PtrVT);
2522 if (Model == TLSModel::LocalExec) {
2523 SDValue HiVar = DAG.getTargetGlobalAddress(
2524 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_G1);
2525 SDValue LoVar = DAG.getTargetGlobalAddress(
2526 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_G0 | ARM64II::MO_NC);
2528 TPOff = SDValue(DAG.getMachineNode(ARM64::MOVZXi, DL, PtrVT, HiVar,
2529 DAG.getTargetConstant(16, MVT::i32)),
2531 TPOff = SDValue(DAG.getMachineNode(ARM64::MOVKXi, DL, PtrVT, TPOff, LoVar,
2532 DAG.getTargetConstant(0, MVT::i32)),
2534 } else if (Model == TLSModel::InitialExec) {
2535 TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2536 TPOff = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, TPOff);
2537 } else if (Model == TLSModel::LocalDynamic) {
2538 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
2539 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
2540 // the beginning of the module's TLS region, followed by a DTPREL offset
2543 // These accesses will need deduplicating if there's more than one.
2544 ARM64FunctionInfo *MFI =
2545 DAG.getMachineFunction().getInfo<ARM64FunctionInfo>();
2546 MFI->incNumLocalDynamicTLSAccesses();
2548 // Accesses used in this sequence go via the TLS descriptor which lives in
2549 // the GOT. Prepare an address we can use to handle this.
2550 SDValue HiDesc = DAG.getTargetExternalSymbol(
2551 "_TLS_MODULE_BASE_", PtrVT, ARM64II::MO_TLS | ARM64II::MO_PAGE);
2552 SDValue LoDesc = DAG.getTargetExternalSymbol(
2553 "_TLS_MODULE_BASE_", PtrVT,
2554 ARM64II::MO_TLS | ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
2556 // First argument to the descriptor call is the address of the descriptor
2558 SDValue DescAddr = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, HiDesc);
2559 DescAddr = DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
2561 // The call needs a relocation too for linker relaxation. It doesn't make
2562 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
2564 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
2567 // Now we can calculate the offset from TPIDR_EL0 to this module's
2568 // thread-local area.
2569 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
2571 // Now use :dtprel_whatever: operations to calculate this variable's offset
2572 // in its thread-storage area.
2573 SDValue HiVar = DAG.getTargetGlobalAddress(
2574 GV, DL, MVT::i64, 0, ARM64II::MO_TLS | ARM64II::MO_G1);
2575 SDValue LoVar = DAG.getTargetGlobalAddress(
2576 GV, DL, MVT::i64, 0, ARM64II::MO_TLS | ARM64II::MO_G0 | ARM64II::MO_NC);
2579 SDValue(DAG.getMachineNode(ARM64::MOVZXi, DL, PtrVT, HiVar,
2580 DAG.getTargetConstant(16, MVT::i32)),
2582 DTPOff = SDValue(DAG.getMachineNode(ARM64::MOVKXi, DL, PtrVT, DTPOff, LoVar,
2583 DAG.getTargetConstant(0, MVT::i32)),
2586 TPOff = DAG.getNode(ISD::ADD, DL, PtrVT, TPOff, DTPOff);
2587 } else if (Model == TLSModel::GeneralDynamic) {
2588 // Accesses used in this sequence go via the TLS descriptor which lives in
2589 // the GOT. Prepare an address we can use to handle this.
2590 SDValue HiDesc = DAG.getTargetGlobalAddress(
2591 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_PAGE);
2592 SDValue LoDesc = DAG.getTargetGlobalAddress(
2594 ARM64II::MO_TLS | ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
2596 // First argument to the descriptor call is the address of the descriptor
2598 SDValue DescAddr = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, HiDesc);
2599 DescAddr = DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
2601 // The call needs a relocation too for linker relaxation. It doesn't make
2602 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
2605 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2607 // Finally we can make a call to calculate the offset from tpidr_el0.
2608 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
2610 llvm_unreachable("Unsupported ELF TLS access model");
2612 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
2615 SDValue ARM64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
2616 SelectionDAG &DAG) const {
2617 if (Subtarget->isTargetDarwin())
2618 return LowerDarwinGlobalTLSAddress(Op, DAG);
2619 else if (Subtarget->isTargetELF())
2620 return LowerELFGlobalTLSAddress(Op, DAG);
2622 llvm_unreachable("Unexpected platform trying to use TLS");
2624 SDValue ARM64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2625 SDValue Chain = Op.getOperand(0);
2626 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2627 SDValue LHS = Op.getOperand(2);
2628 SDValue RHS = Op.getOperand(3);
2629 SDValue Dest = Op.getOperand(4);
2632 // Handle f128 first, since lowering it will result in comparing the return
2633 // value of a libcall against zero, which is just what the rest of LowerBR_CC
2634 // is expecting to deal with.
2635 if (LHS.getValueType() == MVT::f128) {
2636 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
2638 // If softenSetCCOperands returned a scalar, we need to compare the result
2639 // against zero to select between true and false values.
2640 if (!RHS.getNode()) {
2641 RHS = DAG.getConstant(0, LHS.getValueType());
2646 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
2648 unsigned Opc = LHS.getOpcode();
2649 if (LHS.getResNo() == 1 && isa<ConstantSDNode>(RHS) &&
2650 cast<ConstantSDNode>(RHS)->isOne() &&
2651 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
2652 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
2653 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
2654 "Unexpected condition code.");
2655 // Only lower legal XALUO ops.
2656 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
2659 // The actual operation with overflow check.
2660 ARM64CC::CondCode OFCC;
2661 SDValue Value, Overflow;
2662 std::tie(Value, Overflow) = getARM64XALUOOp(OFCC, LHS.getValue(0), DAG);
2664 if (CC == ISD::SETNE)
2665 OFCC = getInvertedCondCode(OFCC);
2666 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
2668 return DAG.getNode(ARM64ISD::BRCOND, SDLoc(LHS), MVT::Other, Chain, Dest,
2672 if (LHS.getValueType().isInteger()) {
2673 assert((LHS.getValueType() == RHS.getValueType()) &&
2674 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
2676 // If the RHS of the comparison is zero, we can potentially fold this
2677 // to a specialized branch.
2678 const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
2679 if (RHSC && RHSC->getZExtValue() == 0) {
2680 if (CC == ISD::SETEQ) {
2681 // See if we can use a TBZ to fold in an AND as well.
2682 // TBZ has a smaller branch displacement than CBZ. If the offset is
2683 // out of bounds, a late MI-layer pass rewrites branches.
2684 // 403.gcc is an example that hits this case.
2685 if (LHS.getOpcode() == ISD::AND &&
2686 isa<ConstantSDNode>(LHS.getOperand(1)) &&
2687 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
2688 SDValue Test = LHS.getOperand(0);
2689 uint64_t Mask = LHS.getConstantOperandVal(1);
2691 // TBZ only operates on i64's, but the ext should be free.
2692 if (Test.getValueType() == MVT::i32)
2693 Test = DAG.getAnyExtOrTrunc(Test, dl, MVT::i64);
2695 return DAG.getNode(ARM64ISD::TBZ, dl, MVT::Other, Chain, Test,
2696 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
2699 return DAG.getNode(ARM64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest);
2700 } else if (CC == ISD::SETNE) {
2701 // See if we can use a TBZ to fold in an AND as well.
2702 // TBZ has a smaller branch displacement than CBZ. If the offset is
2703 // out of bounds, a late MI-layer pass rewrites branches.
2704 // 403.gcc is an example that hits this case.
2705 if (LHS.getOpcode() == ISD::AND &&
2706 isa<ConstantSDNode>(LHS.getOperand(1)) &&
2707 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
2708 SDValue Test = LHS.getOperand(0);
2709 uint64_t Mask = LHS.getConstantOperandVal(1);
2711 // TBNZ only operates on i64's, but the ext should be free.
2712 if (Test.getValueType() == MVT::i32)
2713 Test = DAG.getAnyExtOrTrunc(Test, dl, MVT::i64);
2715 return DAG.getNode(ARM64ISD::TBNZ, dl, MVT::Other, Chain, Test,
2716 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
2719 return DAG.getNode(ARM64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest);
2724 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
2725 return DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
2729 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2731 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
2732 // clean. Some of them require two branches to implement.
2733 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
2734 ARM64CC::CondCode CC1, CC2;
2735 changeFPCCToARM64CC(CC, CC1, CC2);
2736 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
2738 DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp);
2739 if (CC2 != ARM64CC::AL) {
2740 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
2741 return DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val,
2748 SDValue ARM64TargetLowering::LowerFCOPYSIGN(SDValue Op,
2749 SelectionDAG &DAG) const {
2750 EVT VT = Op.getValueType();
2753 SDValue In1 = Op.getOperand(0);
2754 SDValue In2 = Op.getOperand(1);
2755 EVT SrcVT = In2.getValueType();
2757 if (SrcVT == MVT::f32 && VT == MVT::f64)
2758 In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2);
2759 else if (SrcVT == MVT::f64 && VT == MVT::f32)
2760 In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, DAG.getIntPtrConstant(0));
2762 // FIXME: Src type is different, bail out for now. Can VT really be a
2769 SDValue EltMask, VecVal1, VecVal2;
2770 if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) {
2773 EltMask = DAG.getConstant(0x80000000ULL, EltVT);
2775 if (!VT.isVector()) {
2776 VecVal1 = DAG.getTargetInsertSubreg(ARM64::ssub, DL, VecVT,
2777 DAG.getUNDEF(VecVT), In1);
2778 VecVal2 = DAG.getTargetInsertSubreg(ARM64::ssub, DL, VecVT,
2779 DAG.getUNDEF(VecVT), In2);
2781 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
2782 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
2784 } else if (VT == MVT::f64 || VT == MVT::v2f64) {
2788 // We want to materialize a mask with the the high bit set, but the AdvSIMD
2789 // immediate moves cannot materialize that in a single instruction for
2790 // 64-bit elements. Instead, materialize zero and then negate it.
2791 EltMask = DAG.getConstant(0, EltVT);
2793 if (!VT.isVector()) {
2794 VecVal1 = DAG.getTargetInsertSubreg(ARM64::dsub, DL, VecVT,
2795 DAG.getUNDEF(VecVT), In1);
2796 VecVal2 = DAG.getTargetInsertSubreg(ARM64::dsub, DL, VecVT,
2797 DAG.getUNDEF(VecVT), In2);
2799 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
2800 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
2803 llvm_unreachable("Invalid type for copysign!");
2806 std::vector<SDValue> BuildVectorOps;
2807 for (unsigned i = 0; i < VecVT.getVectorNumElements(); ++i)
2808 BuildVectorOps.push_back(EltMask);
2810 SDValue BuildVec = DAG.getNode(ISD::BUILD_VECTOR, DL, VecVT, BuildVectorOps);
2812 // If we couldn't materialize the mask above, then the mask vector will be
2813 // the zero vector, and we need to negate it here.
2814 if (VT == MVT::f64 || VT == MVT::v2f64) {
2815 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec);
2816 BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec);
2817 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec);
2821 DAG.getNode(ARM64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec);
2824 return DAG.getTargetExtractSubreg(ARM64::ssub, DL, VT, Sel);
2825 else if (VT == MVT::f64)
2826 return DAG.getTargetExtractSubreg(ARM64::dsub, DL, VT, Sel);
2828 return DAG.getNode(ISD::BITCAST, DL, VT, Sel);
2831 SDValue ARM64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
2832 if (DAG.getMachineFunction().getFunction()->getAttributes().hasAttribute(
2833 AttributeSet::FunctionIndex, Attribute::NoImplicitFloat))
2836 // While there is no integer popcount instruction, it can
2837 // be more efficiently lowered to the following sequence that uses
2838 // AdvSIMD registers/instructions as long as the copies to/from
2839 // the AdvSIMD registers are cheap.
2840 // FMOV D0, X0 // copy 64-bit int to vector, high bits zero'd
2841 // CNT V0.8B, V0.8B // 8xbyte pop-counts
2842 // ADDV B0, V0.8B // sum 8xbyte pop-counts
2843 // UMOV X0, V0.B[0] // copy byte result back to integer reg
2844 SDValue Val = Op.getOperand(0);
2846 EVT VT = Op.getValueType();
2847 SDValue ZeroVec = DAG.getUNDEF(MVT::v8i8);
2850 if (VT == MVT::i32) {
2851 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Val);
2853 DAG.getTargetInsertSubreg(ARM64::ssub, DL, MVT::v8i8, ZeroVec, VecVal);
2855 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);
2858 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, VecVal);
2859 SDValue UaddLV = DAG.getNode(
2860 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
2861 DAG.getConstant(Intrinsic::arm64_neon_uaddlv, MVT::i32), CtPop);
2864 UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV);
2868 SDValue ARM64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
2870 if (Op.getValueType().isVector())
2871 return LowerVSETCC(Op, DAG);
2873 SDValue LHS = Op.getOperand(0);
2874 SDValue RHS = Op.getOperand(1);
2875 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
2878 // We chose ZeroOrOneBooleanContents, so use zero and one.
2879 EVT VT = Op.getValueType();
2880 SDValue TVal = DAG.getConstant(1, VT);
2881 SDValue FVal = DAG.getConstant(0, VT);
2883 // Handle f128 first, since one possible outcome is a normal integer
2884 // comparison which gets picked up by the next if statement.
2885 if (LHS.getValueType() == MVT::f128) {
2886 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
2888 // If softenSetCCOperands returned a scalar, use it.
2889 if (!RHS.getNode()) {
2890 assert(LHS.getValueType() == Op.getValueType() &&
2891 "Unexpected setcc expansion!");
2896 if (LHS.getValueType().isInteger()) {
2899 getARM64Cmp(LHS, RHS, ISD::getSetCCInverse(CC, true), CCVal, DAG, dl);
2901 // Note that we inverted the condition above, so we reverse the order of
2902 // the true and false operands here. This will allow the setcc to be
2903 // matched to a single CSINC instruction.
2904 return DAG.getNode(ARM64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp);
2907 // Now we know we're dealing with FP values.
2908 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2910 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
2911 // and do the comparison.
2912 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
2914 ARM64CC::CondCode CC1, CC2;
2915 changeFPCCToARM64CC(CC, CC1, CC2);
2916 if (CC2 == ARM64CC::AL) {
2917 changeFPCCToARM64CC(ISD::getSetCCInverse(CC, false), CC1, CC2);
2918 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
2920 // Note that we inverted the condition above, so we reverse the order of
2921 // the true and false operands here. This will allow the setcc to be
2922 // matched to a single CSINC instruction.
2923 return DAG.getNode(ARM64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp);
2925 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
2926 // clean. Some of them require two CSELs to implement. As is in this case,
2927 // we emit the first CSEL and then emit a second using the output of the
2928 // first as the RHS. We're effectively OR'ing the two CC's together.
2930 // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
2931 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
2932 SDValue CS1 = DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
2934 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
2935 return DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
2939 /// A SELECT_CC operation is really some kind of max or min if both values being
2940 /// compared are, in some sense, equal to the results in either case. However,
2941 /// it is permissible to compare f32 values and produce directly extended f64
2944 /// Extending the comparison operands would also be allowed, but is less likely
2945 /// to happen in practice since their use is right here. Note that truncate
2946 /// operations would *not* be semantically equivalent.
2947 static bool selectCCOpsAreFMaxCompatible(SDValue Cmp, SDValue Result) {
2951 ConstantFPSDNode *CCmp = dyn_cast<ConstantFPSDNode>(Cmp);
2952 ConstantFPSDNode *CResult = dyn_cast<ConstantFPSDNode>(Result);
2953 if (CCmp && CResult && Cmp.getValueType() == MVT::f32 &&
2954 Result.getValueType() == MVT::f64) {
2956 APFloat CmpVal = CCmp->getValueAPF();
2957 CmpVal.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &Lossy);
2958 return CResult->getValueAPF().bitwiseIsEqual(CmpVal);
2961 return Result->getOpcode() == ISD::FP_EXTEND && Result->getOperand(0) == Cmp;
2964 SDValue ARM64TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2965 SDValue CC = Op->getOperand(0);
2966 SDValue TVal = Op->getOperand(1);
2967 SDValue FVal = Op->getOperand(2);
2970 unsigned Opc = CC.getOpcode();
2971 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
2973 if (CC.getResNo() == 1 &&
2974 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
2975 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
2976 // Only lower legal XALUO ops.
2977 if (!DAG.getTargetLoweringInfo().isTypeLegal(CC->getValueType(0)))
2980 ARM64CC::CondCode OFCC;
2981 SDValue Value, Overflow;
2982 std::tie(Value, Overflow) = getARM64XALUOOp(OFCC, CC.getValue(0), DAG);
2983 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
2985 return DAG.getNode(ARM64ISD::CSEL, DL, Op.getValueType(), TVal, FVal, CCVal,
2989 if (CC.getOpcode() == ISD::SETCC)
2990 return DAG.getSelectCC(DL, CC.getOperand(0), CC.getOperand(1), TVal, FVal,
2991 cast<CondCodeSDNode>(CC.getOperand(2))->get());
2993 return DAG.getSelectCC(DL, CC, DAG.getConstant(0, CC.getValueType()), TVal,
2997 SDValue ARM64TargetLowering::LowerSELECT_CC(SDValue Op,
2998 SelectionDAG &DAG) const {
2999 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3000 SDValue LHS = Op.getOperand(0);
3001 SDValue RHS = Op.getOperand(1);
3002 SDValue TVal = Op.getOperand(2);
3003 SDValue FVal = Op.getOperand(3);
3006 // Handle f128 first, because it will result in a comparison of some RTLIB
3007 // call result against zero.
3008 if (LHS.getValueType() == MVT::f128) {
3009 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3011 // If softenSetCCOperands returned a scalar, we need to compare the result
3012 // against zero to select between true and false values.
3013 if (!RHS.getNode()) {
3014 RHS = DAG.getConstant(0, LHS.getValueType());
3019 // Handle integers first.
3020 if (LHS.getValueType().isInteger()) {
3021 assert((LHS.getValueType() == RHS.getValueType()) &&
3022 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3024 unsigned Opcode = ARM64ISD::CSEL;
3026 // If both the TVal and the FVal are constants, see if we can swap them in
3027 // order to for a CSINV or CSINC out of them.
3028 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
3029 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
3031 if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) {
3032 std::swap(TVal, FVal);
3033 std::swap(CTVal, CFVal);
3034 CC = ISD::getSetCCInverse(CC, true);
3035 } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) {
3036 std::swap(TVal, FVal);
3037 std::swap(CTVal, CFVal);
3038 CC = ISD::getSetCCInverse(CC, true);
3039 } else if (TVal.getOpcode() == ISD::XOR) {
3040 // If TVal is a NOT we want to swap TVal and FVal so that we can match
3041 // with a CSINV rather than a CSEL.
3042 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(1));
3044 if (CVal && CVal->isAllOnesValue()) {
3045 std::swap(TVal, FVal);
3046 std::swap(CTVal, CFVal);
3047 CC = ISD::getSetCCInverse(CC, true);
3049 } else if (TVal.getOpcode() == ISD::SUB) {
3050 // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
3051 // that we can match with a CSNEG rather than a CSEL.
3052 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(0));
3054 if (CVal && CVal->isNullValue()) {
3055 std::swap(TVal, FVal);
3056 std::swap(CTVal, CFVal);
3057 CC = ISD::getSetCCInverse(CC, true);
3059 } else if (CTVal && CFVal) {
3060 const int64_t TrueVal = CTVal->getSExtValue();
3061 const int64_t FalseVal = CFVal->getSExtValue();
3064 // If both TVal and FVal are constants, see if FVal is the
3065 // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
3066 // instead of a CSEL in that case.
3067 if (TrueVal == ~FalseVal) {
3068 Opcode = ARM64ISD::CSINV;
3069 } else if (TrueVal == -FalseVal) {
3070 Opcode = ARM64ISD::CSNEG;
3071 } else if (TVal.getValueType() == MVT::i32) {
3072 // If our operands are only 32-bit wide, make sure we use 32-bit
3073 // arithmetic for the check whether we can use CSINC. This ensures that
3074 // the addition in the check will wrap around properly in case there is
3075 // an overflow (which would not be the case if we do the check with
3076 // 64-bit arithmetic).
3077 const uint32_t TrueVal32 = CTVal->getZExtValue();
3078 const uint32_t FalseVal32 = CFVal->getZExtValue();
3080 if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
3081 Opcode = ARM64ISD::CSINC;
3083 if (TrueVal32 > FalseVal32) {
3087 // 64-bit check whether we can use CSINC.
3088 } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) {
3089 Opcode = ARM64ISD::CSINC;
3091 if (TrueVal > FalseVal) {
3096 // Swap TVal and FVal if necessary.
3098 std::swap(TVal, FVal);
3099 std::swap(CTVal, CFVal);
3100 CC = ISD::getSetCCInverse(CC, true);
3103 if (Opcode != ARM64ISD::CSEL) {
3104 // Drop FVal since we can get its value by simply inverting/negating
3111 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3113 EVT VT = Op.getValueType();
3114 return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp);
3117 // Now we know we're dealing with FP values.
3118 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3119 assert(LHS.getValueType() == RHS.getValueType());
3120 EVT VT = Op.getValueType();
3122 // Try to match this select into a max/min operation, which have dedicated
3123 // opcode in the instruction set.
3124 // NOTE: This is not correct in the presence of NaNs, so we only enable this
3126 if (getTargetMachine().Options.NoNaNsFPMath) {
3127 if (selectCCOpsAreFMaxCompatible(LHS, FVal) &&
3128 selectCCOpsAreFMaxCompatible(RHS, TVal)) {
3129 CC = ISD::getSetCCSwappedOperands(CC);
3130 std::swap(TVal, FVal);
3133 if (selectCCOpsAreFMaxCompatible(LHS, TVal) &&
3134 selectCCOpsAreFMaxCompatible(RHS, FVal)) {
3144 return DAG.getNode(ARM64ISD::FMAX, dl, VT, TVal, FVal);
3152 return DAG.getNode(ARM64ISD::FMIN, dl, VT, TVal, FVal);
3158 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3159 // and do the comparison.
3160 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3162 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
3163 // clean. Some of them require two CSELs to implement.
3164 ARM64CC::CondCode CC1, CC2;
3165 changeFPCCToARM64CC(CC, CC1, CC2);
3166 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3167 SDValue CS1 = DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3169 // If we need a second CSEL, emit it, using the output of the first as the
3170 // RHS. We're effectively OR'ing the two CC's together.
3171 if (CC2 != ARM64CC::AL) {
3172 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3173 return DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3176 // Otherwise, return the output of the first CSEL.
3180 SDValue ARM64TargetLowering::LowerJumpTable(SDValue Op,
3181 SelectionDAG &DAG) const {
3182 // Jump table entries as PC relative offsets. No additional tweaking
3183 // is necessary here. Just get the address of the jump table.
3184 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3185 EVT PtrVT = getPointerTy();
3188 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3189 !Subtarget->isTargetMachO()) {
3190 const unsigned char MO_NC = ARM64II::MO_NC;
3192 ARM64ISD::WrapperLarge, DL, PtrVT,
3193 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G3),
3194 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G2 | MO_NC),
3195 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G1 | MO_NC),
3196 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G0 | MO_NC));
3199 SDValue Hi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_PAGE);
3200 SDValue Lo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3201 ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
3202 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3203 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3206 SDValue ARM64TargetLowering::LowerConstantPool(SDValue Op,
3207 SelectionDAG &DAG) const {
3208 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3209 EVT PtrVT = getPointerTy();
3212 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
3213 // Use the GOT for the large code model on iOS.
3214 if (Subtarget->isTargetMachO()) {
3215 SDValue GotAddr = DAG.getTargetConstantPool(
3216 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3218 return DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, GotAddr);
3221 const unsigned char MO_NC = ARM64II::MO_NC;
3223 ARM64ISD::WrapperLarge, DL, PtrVT,
3224 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3225 CP->getOffset(), ARM64II::MO_G3),
3226 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3227 CP->getOffset(), ARM64II::MO_G2 | MO_NC),
3228 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3229 CP->getOffset(), ARM64II::MO_G1 | MO_NC),
3230 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3231 CP->getOffset(), ARM64II::MO_G0 | MO_NC));
3233 // Use ADRP/ADD or ADRP/LDR for everything else: the small memory model on
3234 // ELF, the only valid one on Darwin.
3236 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3237 CP->getOffset(), ARM64II::MO_PAGE);
3238 SDValue Lo = DAG.getTargetConstantPool(
3239 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3240 ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
3242 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3243 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3247 SDValue ARM64TargetLowering::LowerBlockAddress(SDValue Op,
3248 SelectionDAG &DAG) const {
3249 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
3250 EVT PtrVT = getPointerTy();
3252 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3253 !Subtarget->isTargetMachO()) {
3254 const unsigned char MO_NC = ARM64II::MO_NC;
3256 ARM64ISD::WrapperLarge, DL, PtrVT,
3257 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G3),
3258 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G2 | MO_NC),
3259 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G1 | MO_NC),
3260 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G0 | MO_NC));
3262 SDValue Hi = DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_PAGE);
3263 SDValue Lo = DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_PAGEOFF |
3265 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3266 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3270 SDValue ARM64TargetLowering::LowerDarwin_VASTART(SDValue Op,
3271 SelectionDAG &DAG) const {
3272 ARM64FunctionInfo *FuncInfo =
3273 DAG.getMachineFunction().getInfo<ARM64FunctionInfo>();
3277 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3278 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3279 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
3280 MachinePointerInfo(SV), false, false, 0);
3283 SDValue ARM64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
3284 SelectionDAG &DAG) const {
3285 // The layout of the va_list struct is specified in the AArch64 Procedure Call
3286 // Standard, section B.3.
3287 MachineFunction &MF = DAG.getMachineFunction();
3288 ARM64FunctionInfo *FuncInfo = MF.getInfo<ARM64FunctionInfo>();
3291 SDValue Chain = Op.getOperand(0);
3292 SDValue VAList = Op.getOperand(1);
3293 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3294 SmallVector<SDValue, 4> MemOps;
3296 // void *__stack at offset 0
3298 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3299 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
3300 MachinePointerInfo(SV), false, false, 8));
3302 // void *__gr_top at offset 8
3303 int GPRSize = FuncInfo->getVarArgsGPRSize();
3305 SDValue GRTop, GRTopAddr;
3307 GRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3308 DAG.getConstant(8, getPointerTy()));
3310 GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), getPointerTy());
3311 GRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), GRTop,
3312 DAG.getConstant(GPRSize, getPointerTy()));
3314 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
3315 MachinePointerInfo(SV, 8), false, false, 8));
3318 // void *__vr_top at offset 16
3319 int FPRSize = FuncInfo->getVarArgsFPRSize();
3321 SDValue VRTop, VRTopAddr;
3322 VRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3323 DAG.getConstant(16, getPointerTy()));
3325 VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), getPointerTy());
3326 VRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), VRTop,
3327 DAG.getConstant(FPRSize, getPointerTy()));
3329 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
3330 MachinePointerInfo(SV, 16), false, false, 8));
3333 // int __gr_offs at offset 24
3334 SDValue GROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3335 DAG.getConstant(24, getPointerTy()));
3336 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-GPRSize, MVT::i32),
3337 GROffsAddr, MachinePointerInfo(SV, 24), false,
3340 // int __vr_offs at offset 28
3341 SDValue VROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3342 DAG.getConstant(28, getPointerTy()));
3343 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-FPRSize, MVT::i32),
3344 VROffsAddr, MachinePointerInfo(SV, 28), false,
3347 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
3350 SDValue ARM64TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
3351 return Subtarget->isTargetDarwin() ? LowerDarwin_VASTART(Op, DAG)
3352 : LowerAAPCS_VASTART(Op, DAG);
3355 SDValue ARM64TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
3356 // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
3358 unsigned VaListSize = Subtarget->isTargetDarwin() ? 8 : 32;
3359 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
3360 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
3362 return DAG.getMemcpy(Op.getOperand(0), SDLoc(Op), Op.getOperand(1),
3363 Op.getOperand(2), DAG.getConstant(VaListSize, MVT::i32),
3364 8, false, false, MachinePointerInfo(DestSV),
3365 MachinePointerInfo(SrcSV));
3368 SDValue ARM64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
3369 assert(Subtarget->isTargetDarwin() &&
3370 "automatic va_arg instruction only works on Darwin");
3372 const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3373 EVT VT = Op.getValueType();
3375 SDValue Chain = Op.getOperand(0);
3376 SDValue Addr = Op.getOperand(1);
3377 unsigned Align = Op.getConstantOperandVal(3);
3379 SDValue VAList = DAG.getLoad(getPointerTy(), DL, Chain, Addr,
3380 MachinePointerInfo(V), false, false, false, 0);
3381 Chain = VAList.getValue(1);
3384 assert(((Align & (Align - 1)) == 0) && "Expected Align to be a power of 2");
3385 VAList = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3386 DAG.getConstant(Align - 1, getPointerTy()));
3387 VAList = DAG.getNode(ISD::AND, DL, getPointerTy(), VAList,
3388 DAG.getConstant(-(int64_t)Align, getPointerTy()));
3391 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
3392 uint64_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
3394 // Scalar integer and FP values smaller than 64 bits are implicitly extended
3395 // up to 64 bits. At the very least, we have to increase the striding of the
3396 // vaargs list to match this, and for FP values we need to introduce
3397 // FP_ROUND nodes as well.
3398 if (VT.isInteger() && !VT.isVector())
3400 bool NeedFPTrunc = false;
3401 if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
3406 // Increment the pointer, VAList, to the next vaarg
3407 SDValue VANext = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3408 DAG.getConstant(ArgSize, getPointerTy()));
3409 // Store the incremented VAList to the legalized pointer
3410 SDValue APStore = DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V),
3413 // Load the actual argument out of the pointer VAList
3415 // Load the value as an f64.
3416 SDValue WideFP = DAG.getLoad(MVT::f64, DL, APStore, VAList,
3417 MachinePointerInfo(), false, false, false, 0);
3418 // Round the value down to an f32.
3419 SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
3420 DAG.getIntPtrConstant(1));
3421 SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
3422 // Merge the rounded value with the chain output of the load.
3423 return DAG.getMergeValues(Ops, DL);
3426 return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo(), false,
3430 SDValue ARM64TargetLowering::LowerFRAMEADDR(SDValue Op,
3431 SelectionDAG &DAG) const {
3432 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3433 MFI->setFrameAddressIsTaken(true);
3435 EVT VT = Op.getValueType();
3437 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3438 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL, ARM64::FP, VT);
3440 FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
3441 MachinePointerInfo(), false, false, false, 0);
3445 // FIXME? Maybe this could be a TableGen attribute on some registers and
3446 // this table could be generated automatically from RegInfo.
3447 unsigned ARM64TargetLowering::getRegisterByName(const char* RegName) const {
3448 unsigned Reg = StringSwitch<unsigned>(RegName)
3449 .Case("sp", ARM64::SP)
3453 report_fatal_error("Invalid register name global variable");
3456 SDValue ARM64TargetLowering::LowerRETURNADDR(SDValue Op,
3457 SelectionDAG &DAG) const {
3458 MachineFunction &MF = DAG.getMachineFunction();
3459 MachineFrameInfo *MFI = MF.getFrameInfo();
3460 MFI->setReturnAddressIsTaken(true);
3462 EVT VT = Op.getValueType();
3464 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3466 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3467 SDValue Offset = DAG.getConstant(8, getPointerTy());
3468 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
3469 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
3470 MachinePointerInfo(), false, false, false, 0);
3473 // Return LR, which contains the return address. Mark it an implicit live-in.
3474 unsigned Reg = MF.addLiveIn(ARM64::LR, &ARM64::GPR64RegClass);
3475 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
3478 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3479 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
3480 SDValue ARM64TargetLowering::LowerShiftRightParts(SDValue Op,
3481 SelectionDAG &DAG) const {
3482 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3483 EVT VT = Op.getValueType();
3484 unsigned VTBits = VT.getSizeInBits();
3486 SDValue ShOpLo = Op.getOperand(0);
3487 SDValue ShOpHi = Op.getOperand(1);
3488 SDValue ShAmt = Op.getOperand(2);
3490 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3492 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3494 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
3495 DAG.getConstant(VTBits, MVT::i64), ShAmt);
3496 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3497 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
3498 DAG.getConstant(VTBits, MVT::i64));
3499 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3501 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
3502 ISD::SETGE, dl, DAG);
3503 SDValue CCVal = DAG.getConstant(ARM64CC::GE, MVT::i32);
3505 SDValue FalseValLo = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3506 SDValue TrueValLo = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3508 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
3510 // ARM64 shifts larger than the register width are wrapped rather than
3511 // clamped, so we can't just emit "hi >> x".
3512 SDValue FalseValHi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3513 SDValue TrueValHi = Opc == ISD::SRA
3514 ? DAG.getNode(Opc, dl, VT, ShOpHi,
3515 DAG.getConstant(VTBits - 1, MVT::i64))
3516 : DAG.getConstant(0, VT);
3518 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValHi, FalseValHi, CCVal, Cmp);
3520 SDValue Ops[2] = { Lo, Hi };
3521 return DAG.getMergeValues(Ops, dl);
3524 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3525 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
3526 SDValue ARM64TargetLowering::LowerShiftLeftParts(SDValue Op,
3527 SelectionDAG &DAG) const {
3528 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3529 EVT VT = Op.getValueType();
3530 unsigned VTBits = VT.getSizeInBits();
3532 SDValue ShOpLo = Op.getOperand(0);
3533 SDValue ShOpHi = Op.getOperand(1);
3534 SDValue ShAmt = Op.getOperand(2);
3537 assert(Op.getOpcode() == ISD::SHL_PARTS);
3538 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
3539 DAG.getConstant(VTBits, MVT::i64), ShAmt);
3540 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3541 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
3542 DAG.getConstant(VTBits, MVT::i64));
3543 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3544 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3546 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3548 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
3549 ISD::SETGE, dl, DAG);
3550 SDValue CCVal = DAG.getConstant(ARM64CC::GE, MVT::i32);
3551 SDValue Hi = DAG.getNode(ARM64ISD::CSEL, dl, VT, Tmp3, FalseVal, CCVal, Cmp);
3553 // ARM64 shifts of larger than register sizes are wrapped rather than clamped,
3554 // so we can't just emit "lo << a" if a is too big.
3555 SDValue TrueValLo = DAG.getConstant(0, VT);
3556 SDValue FalseValLo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3558 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
3560 SDValue Ops[2] = { Lo, Hi };
3561 return DAG.getMergeValues(Ops, dl);
3565 ARM64TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3566 // The ARM64 target doesn't support folding offsets into global addresses.
3570 bool ARM64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3571 // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
3572 // FIXME: We should be able to handle f128 as well with a clever lowering.
3573 if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32))
3577 return ARM64_AM::getFP64Imm(Imm) != -1;
3578 else if (VT == MVT::f32)
3579 return ARM64_AM::getFP32Imm(Imm) != -1;
3583 //===----------------------------------------------------------------------===//
3584 // ARM64 Optimization Hooks
3585 //===----------------------------------------------------------------------===//
3587 //===----------------------------------------------------------------------===//
3588 // ARM64 Inline Assembly Support
3589 //===----------------------------------------------------------------------===//
3591 // Table of Constraints
3592 // TODO: This is the current set of constraints supported by ARM for the
3593 // compiler, not all of them may make sense, e.g. S may be difficult to support.
3595 // r - A general register
3596 // w - An FP/SIMD register of some size in the range v0-v31
3597 // x - An FP/SIMD register of some size in the range v0-v15
3598 // I - Constant that can be used with an ADD instruction
3599 // J - Constant that can be used with a SUB instruction
3600 // K - Constant that can be used with a 32-bit logical instruction
3601 // L - Constant that can be used with a 64-bit logical instruction
3602 // M - Constant that can be used as a 32-bit MOV immediate
3603 // N - Constant that can be used as a 64-bit MOV immediate
3604 // Q - A memory reference with base register and no offset
3605 // S - A symbolic address
3606 // Y - Floating point constant zero
3607 // Z - Integer constant zero
3609 // Note that general register operands will be output using their 64-bit x
3610 // register name, whatever the size of the variable, unless the asm operand
3611 // is prefixed by the %w modifier. Floating-point and SIMD register operands
3612 // will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
3615 /// getConstraintType - Given a constraint letter, return the type of
3616 /// constraint it is for this target.
3617 ARM64TargetLowering::ConstraintType
3618 ARM64TargetLowering::getConstraintType(const std::string &Constraint) const {
3619 if (Constraint.size() == 1) {
3620 switch (Constraint[0]) {
3627 return C_RegisterClass;
3628 // An address with a single base register. Due to the way we
3629 // currently handle addresses it is the same as 'r'.
3634 return TargetLowering::getConstraintType(Constraint);
3637 /// Examine constraint type and operand type and determine a weight value.
3638 /// This object must already have been set up with the operand type
3639 /// and the current alternative constraint selected.
3640 TargetLowering::ConstraintWeight
3641 ARM64TargetLowering::getSingleConstraintMatchWeight(
3642 AsmOperandInfo &info, const char *constraint) const {
3643 ConstraintWeight weight = CW_Invalid;
3644 Value *CallOperandVal = info.CallOperandVal;
3645 // If we don't have a value, we can't do a match,
3646 // but allow it at the lowest weight.
3647 if (!CallOperandVal)
3649 Type *type = CallOperandVal->getType();
3650 // Look at the constraint type.
3651 switch (*constraint) {
3653 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
3657 if (type->isFloatingPointTy() || type->isVectorTy())
3658 weight = CW_Register;
3661 weight = CW_Constant;
3667 std::pair<unsigned, const TargetRegisterClass *>
3668 ARM64TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3670 if (Constraint.size() == 1) {
3671 switch (Constraint[0]) {
3673 if (VT.getSizeInBits() == 64)
3674 return std::make_pair(0U, &ARM64::GPR64commonRegClass);
3675 return std::make_pair(0U, &ARM64::GPR32commonRegClass);
3678 return std::make_pair(0U, &ARM64::FPR32RegClass);
3679 if (VT.getSizeInBits() == 64)
3680 return std::make_pair(0U, &ARM64::FPR64RegClass);
3681 if (VT.getSizeInBits() == 128)
3682 return std::make_pair(0U, &ARM64::FPR128RegClass);
3684 // The instructions that this constraint is designed for can
3685 // only take 128-bit registers so just use that regclass.
3687 if (VT.getSizeInBits() == 128)
3688 return std::make_pair(0U, &ARM64::FPR128_loRegClass);
3692 if (StringRef("{cc}").equals_lower(Constraint))
3693 return std::make_pair(unsigned(ARM64::NZCV), &ARM64::CCRRegClass);
3695 // Use the default implementation in TargetLowering to convert the register
3696 // constraint into a member of a register class.
3697 std::pair<unsigned, const TargetRegisterClass *> Res;
3698 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3700 // Not found as a standard register?
3702 unsigned Size = Constraint.size();
3703 if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
3704 tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
3705 const std::string Reg =
3706 std::string(&Constraint[2], &Constraint[Size - 1]);
3707 int RegNo = atoi(Reg.c_str());
3708 if (RegNo >= 0 && RegNo <= 31) {
3709 // v0 - v31 are aliases of q0 - q31.
3710 // By default we'll emit v0-v31 for this unless there's a modifier where
3711 // we'll emit the correct register as well.
3712 Res.first = ARM64::FPR128RegClass.getRegister(RegNo);
3713 Res.second = &ARM64::FPR128RegClass;
3721 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3722 /// vector. If it is invalid, don't add anything to Ops.
3723 void ARM64TargetLowering::LowerAsmOperandForConstraint(
3724 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
3725 SelectionDAG &DAG) const {
3728 // Currently only support length 1 constraints.
3729 if (Constraint.length() != 1)
3732 char ConstraintLetter = Constraint[0];
3733 switch (ConstraintLetter) {
3737 // This set of constraints deal with valid constants for various instructions.
3738 // Validate and return a target constant for them if we can.
3740 // 'z' maps to xzr or wzr so it needs an input of 0.
3741 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3742 if (!C || C->getZExtValue() != 0)
3745 if (Op.getValueType() == MVT::i64)
3746 Result = DAG.getRegister(ARM64::XZR, MVT::i64);
3748 Result = DAG.getRegister(ARM64::WZR, MVT::i32);
3758 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3762 // Grab the value and do some validation.
3763 uint64_t CVal = C->getZExtValue();
3764 switch (ConstraintLetter) {
3765 // The I constraint applies only to simple ADD or SUB immediate operands:
3766 // i.e. 0 to 4095 with optional shift by 12
3767 // The J constraint applies only to ADD or SUB immediates that would be
3768 // valid when negated, i.e. if [an add pattern] were to be output as a SUB
3769 // instruction [or vice versa], in other words -1 to -4095 with optional
3770 // left shift by 12.
3772 if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
3776 uint64_t NVal = -C->getSExtValue();
3777 if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal))
3781 // The K and L constraints apply *only* to logical immediates, including
3782 // what used to be the MOVI alias for ORR (though the MOVI alias has now
3783 // been removed and MOV should be used). So these constraints have to
3784 // distinguish between bit patterns that are valid 32-bit or 64-bit
3785 // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
3786 // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
3789 if (ARM64_AM::isLogicalImmediate(CVal, 32))
3793 if (ARM64_AM::isLogicalImmediate(CVal, 64))
3796 // The M and N constraints are a superset of K and L respectively, for use
3797 // with the MOV (immediate) alias. As well as the logical immediates they
3798 // also match 32 or 64-bit immediates that can be loaded either using a
3799 // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
3800 // (M) or 64-bit 0x1234000000000000 (N) etc.
3801 // As a note some of this code is liberally stolen from the asm parser.
3803 if (!isUInt<32>(CVal))
3805 if (ARM64_AM::isLogicalImmediate(CVal, 32))
3807 if ((CVal & 0xFFFF) == CVal)
3809 if ((CVal & 0xFFFF0000ULL) == CVal)
3811 uint64_t NCVal = ~(uint32_t)CVal;
3812 if ((NCVal & 0xFFFFULL) == NCVal)
3814 if ((NCVal & 0xFFFF0000ULL) == NCVal)
3819 if (ARM64_AM::isLogicalImmediate(CVal, 64))
3821 if ((CVal & 0xFFFFULL) == CVal)
3823 if ((CVal & 0xFFFF0000ULL) == CVal)
3825 if ((CVal & 0xFFFF00000000ULL) == CVal)
3827 if ((CVal & 0xFFFF000000000000ULL) == CVal)
3829 uint64_t NCVal = ~CVal;
3830 if ((NCVal & 0xFFFFULL) == NCVal)
3832 if ((NCVal & 0xFFFF0000ULL) == NCVal)
3834 if ((NCVal & 0xFFFF00000000ULL) == NCVal)
3836 if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
3844 // All assembler immediates are 64-bit integers.
3845 Result = DAG.getTargetConstant(CVal, MVT::i64);
3849 if (Result.getNode()) {
3850 Ops.push_back(Result);
3854 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
3857 //===----------------------------------------------------------------------===//
3858 // ARM64 Advanced SIMD Support
3859 //===----------------------------------------------------------------------===//
3861 /// WidenVector - Given a value in the V64 register class, produce the
3862 /// equivalent value in the V128 register class.
3863 static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
3864 EVT VT = V64Reg.getValueType();
3865 unsigned NarrowSize = VT.getVectorNumElements();
3866 MVT EltTy = VT.getVectorElementType().getSimpleVT();
3867 MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
3870 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
3871 V64Reg, DAG.getConstant(0, MVT::i32));
3874 /// getExtFactor - Determine the adjustment factor for the position when
3875 /// generating an "extract from vector registers" instruction.
3876 static unsigned getExtFactor(SDValue &V) {
3877 EVT EltType = V.getValueType().getVectorElementType();
3878 return EltType.getSizeInBits() / 8;
3881 /// NarrowVector - Given a value in the V128 register class, produce the
3882 /// equivalent value in the V64 register class.
3883 static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) {
3884 EVT VT = V128Reg.getValueType();
3885 unsigned WideSize = VT.getVectorNumElements();
3886 MVT EltTy = VT.getVectorElementType().getSimpleVT();
3887 MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2);
3890 return DAG.getTargetExtractSubreg(ARM64::dsub, DL, NarrowTy, V128Reg);
3893 // Gather data to see if the operation can be modelled as a
3894 // shuffle in combination with VEXTs.
3895 SDValue ARM64TargetLowering::ReconstructShuffle(SDValue Op,
3896 SelectionDAG &DAG) const {
3898 EVT VT = Op.getValueType();
3899 unsigned NumElts = VT.getVectorNumElements();
3901 SmallVector<SDValue, 2> SourceVecs;
3902 SmallVector<unsigned, 2> MinElts;
3903 SmallVector<unsigned, 2> MaxElts;
3905 for (unsigned i = 0; i < NumElts; ++i) {
3906 SDValue V = Op.getOperand(i);
3907 if (V.getOpcode() == ISD::UNDEF)
3909 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
3910 // A shuffle can only come from building a vector from various
3911 // elements of other vectors.
3915 // Record this extraction against the appropriate vector if possible...
3916 SDValue SourceVec = V.getOperand(0);
3917 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
3918 bool FoundSource = false;
3919 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
3920 if (SourceVecs[j] == SourceVec) {
3921 if (MinElts[j] > EltNo)
3923 if (MaxElts[j] < EltNo)
3930 // Or record a new source if not...
3932 SourceVecs.push_back(SourceVec);
3933 MinElts.push_back(EltNo);
3934 MaxElts.push_back(EltNo);
3938 // Currently only do something sane when at most two source vectors
3940 if (SourceVecs.size() > 2)
3943 SDValue ShuffleSrcs[2] = { DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
3944 int VEXTOffsets[2] = { 0, 0 };
3946 // This loop extracts the usage patterns of the source vectors
3947 // and prepares appropriate SDValues for a shuffle if possible.
3948 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
3949 if (SourceVecs[i].getValueType() == VT) {
3950 // No VEXT necessary
3951 ShuffleSrcs[i] = SourceVecs[i];
3954 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
3955 // We can pad out the smaller vector for free, so if it's part of a
3957 ShuffleSrcs[i] = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, SourceVecs[i],
3958 DAG.getUNDEF(SourceVecs[i].getValueType()));
3962 // Don't attempt to extract subvectors from BUILD_VECTOR sources
3963 // that expand or trunc the original value.
3964 // TODO: We can try to bitcast and ANY_EXTEND the result but
3965 // we need to consider the cost of vector ANY_EXTEND, and the
3966 // legality of all the types.
3967 if (SourceVecs[i].getValueType().getVectorElementType() !=
3968 VT.getVectorElementType())
3971 // Since only 64-bit and 128-bit vectors are legal on ARM and
3972 // we've eliminated the other cases...
3973 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2 * NumElts &&
3974 "unexpected vector sizes in ReconstructShuffle");
3976 if (MaxElts[i] - MinElts[i] >= NumElts) {
3977 // Span too large for a VEXT to cope
3981 if (MinElts[i] >= NumElts) {
3982 // The extraction can just take the second half
3983 VEXTOffsets[i] = NumElts;
3985 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
3986 DAG.getIntPtrConstant(NumElts));
3987 } else if (MaxElts[i] < NumElts) {
3988 // The extraction can just take the first half
3990 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3991 SourceVecs[i], DAG.getIntPtrConstant(0));
3993 // An actual VEXT is needed
3994 VEXTOffsets[i] = MinElts[i];
3995 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3996 SourceVecs[i], DAG.getIntPtrConstant(0));
3998 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
3999 DAG.getIntPtrConstant(NumElts));
4000 unsigned Imm = VEXTOffsets[i] * getExtFactor(VEXTSrc1);
4001 ShuffleSrcs[i] = DAG.getNode(ARM64ISD::EXT, dl, VT, VEXTSrc1, VEXTSrc2,
4002 DAG.getConstant(Imm, MVT::i32));
4006 SmallVector<int, 8> Mask;
4008 for (unsigned i = 0; i < NumElts; ++i) {
4009 SDValue Entry = Op.getOperand(i);
4010 if (Entry.getOpcode() == ISD::UNDEF) {
4015 SDValue ExtractVec = Entry.getOperand(0);
4017 cast<ConstantSDNode>(Op.getOperand(i).getOperand(1))->getSExtValue();
4018 if (ExtractVec == SourceVecs[0]) {
4019 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4021 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4025 // Final check before we try to produce nonsense...
4026 if (isShuffleMaskLegal(Mask, VT))
4027 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4033 // check if an EXT instruction can handle the shuffle mask when the
4034 // vector sources of the shuffle are the same.
4035 static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4036 unsigned NumElts = VT.getVectorNumElements();
4038 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4044 // If this is a VEXT shuffle, the immediate value is the index of the first
4045 // element. The other shuffle indices must be the successive elements after
4047 unsigned ExpectedElt = Imm;
4048 for (unsigned i = 1; i < NumElts; ++i) {
4049 // Increment the expected index. If it wraps around, just follow it
4050 // back to index zero and keep going.
4052 if (ExpectedElt == NumElts)
4056 continue; // ignore UNDEF indices
4057 if (ExpectedElt != static_cast<unsigned>(M[i]))
4064 // check if an EXT instruction can handle the shuffle mask when the
4065 // vector sources of the shuffle are different.
4066 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
4068 // Look for the first non-undef element.
4069 const int *FirstRealElt = std::find_if(M.begin(), M.end(),
4070 [](int Elt) {return Elt >= 0;});
4072 // Benefit form APInt to handle overflow when calculating expected element.
4073 unsigned NumElts = VT.getVectorNumElements();
4074 unsigned MaskBits = APInt(32, NumElts * 2).logBase2();
4075 APInt ExpectedElt = APInt(MaskBits, *FirstRealElt + 1);
4076 // The following shuffle indices must be the successive elements after the
4077 // first real element.
4078 const int *FirstWrongElt = std::find_if(FirstRealElt + 1, M.end(),
4079 [&](int Elt) {return Elt != ExpectedElt++ && Elt != -1;});
4080 if (FirstWrongElt != M.end())
4083 // The index of an EXT is the first element if it is not UNDEF.
4084 // Watch out for the beginning UNDEFs. The EXT index should be the expected
4085 // value of the first element.
4086 // E.g. <-1, -1, 3, ...> is treated as <1, 2, 3, ...>.
4087 // <-1, -1, 0, 1, ...> is treated as <IDX, IDX+1, 0, 1, ...>. IDX is
4088 // equal to the ExpectedElt.
4089 Imm = (M[0] >= 0) ? static_cast<unsigned>(M[0]) : ExpectedElt.getZExtValue();
4091 // If no beginning UNDEFs, do swap when M[0] >= NumElts.
4092 if (M[0] >= 0 && Imm >= NumElts) {
4095 } else if (M[0] < 0) {
4096 // Only do swap when beginning UNDEFs more than the first real element,
4097 if (*FirstRealElt < FirstRealElt - M.begin())
4106 /// isREVMask - Check if a vector shuffle corresponds to a REV
4107 /// instruction with the specified blocksize. (The order of the elements
4108 /// within each block of the vector is reversed.)
4109 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4110 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
4111 "Only possible block sizes for REV are: 16, 32, 64");
4113 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4117 unsigned NumElts = VT.getVectorNumElements();
4118 unsigned BlockElts = M[0] + 1;
4119 // If the first shuffle index is UNDEF, be optimistic.
4121 BlockElts = BlockSize / EltSz;
4123 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4126 for (unsigned i = 0; i < NumElts; ++i) {
4128 continue; // ignore UNDEF indices
4129 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
4136 static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4137 unsigned NumElts = VT.getVectorNumElements();
4138 WhichResult = (M[0] == 0 ? 0 : 1);
4139 unsigned Idx = WhichResult * NumElts / 2;
4140 for (unsigned i = 0; i != NumElts; i += 2) {
4141 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4142 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
4150 static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4151 unsigned NumElts = VT.getVectorNumElements();
4152 WhichResult = (M[0] == 0 ? 0 : 1);
4153 for (unsigned i = 0; i != NumElts; ++i) {
4155 continue; // ignore UNDEF indices
4156 if ((unsigned)M[i] != 2 * i + WhichResult)
4163 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4164 unsigned NumElts = VT.getVectorNumElements();
4165 WhichResult = (M[0] == 0 ? 0 : 1);
4166 for (unsigned i = 0; i < NumElts; i += 2) {
4167 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4168 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + NumElts + WhichResult))
4174 /// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of
4175 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4176 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4177 static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4178 unsigned NumElts = VT.getVectorNumElements();
4179 WhichResult = (M[0] == 0 ? 0 : 1);
4180 unsigned Idx = WhichResult * NumElts / 2;
4181 for (unsigned i = 0; i != NumElts; i += 2) {
4182 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4183 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx))
4191 /// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of
4192 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4193 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4194 static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4195 unsigned Half = VT.getVectorNumElements() / 2;
4196 WhichResult = (M[0] == 0 ? 0 : 1);
4197 for (unsigned j = 0; j != 2; ++j) {
4198 unsigned Idx = WhichResult;
4199 for (unsigned i = 0; i != Half; ++i) {
4200 int MIdx = M[i + j * Half];
4201 if (MIdx >= 0 && (unsigned)MIdx != Idx)
4210 /// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of
4211 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4212 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4213 static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4214 unsigned NumElts = VT.getVectorNumElements();
4215 WhichResult = (M[0] == 0 ? 0 : 1);
4216 for (unsigned i = 0; i < NumElts; i += 2) {
4217 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4218 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult))
4224 static bool isINSMask(ArrayRef<int> M, int NumInputElements,
4225 bool &DstIsLeft, int &Anomaly) {
4226 if (M.size() != static_cast<size_t>(NumInputElements))
4229 int NumLHSMatch = 0, NumRHSMatch = 0;
4230 int LastLHSMismatch = -1, LastRHSMismatch = -1;
4232 for (int i = 0; i < NumInputElements; ++i) {
4242 LastLHSMismatch = i;
4244 if (M[i] == i + NumInputElements)
4247 LastRHSMismatch = i;
4250 if (NumLHSMatch == NumInputElements - 1) {
4252 Anomaly = LastLHSMismatch;
4254 } else if (NumRHSMatch == NumInputElements - 1) {
4256 Anomaly = LastRHSMismatch;
4263 static bool isConcatMask(ArrayRef<int> Mask, EVT VT, bool SplitLHS) {
4264 if (VT.getSizeInBits() != 128)
4267 unsigned NumElts = VT.getVectorNumElements();
4269 for (int I = 0, E = NumElts / 2; I != E; I++) {
4274 int Offset = NumElts / 2;
4275 for (int I = NumElts / 2, E = NumElts; I != E; I++) {
4276 if (Mask[I] != I + SplitLHS * Offset)
4283 static SDValue tryFormConcatFromShuffle(SDValue Op, SelectionDAG &DAG) {
4285 EVT VT = Op.getValueType();
4286 SDValue V0 = Op.getOperand(0);
4287 SDValue V1 = Op.getOperand(1);
4288 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op)->getMask();
4290 if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() ||
4291 VT.getVectorElementType() != V1.getValueType().getVectorElementType())
4294 bool SplitV0 = V0.getValueType().getSizeInBits() == 128;
4296 if (!isConcatMask(Mask, VT, SplitV0))
4299 EVT CastVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
4300 VT.getVectorNumElements() / 2);
4302 V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0,
4303 DAG.getConstant(0, MVT::i64));
4305 if (V1.getValueType().getSizeInBits() == 128) {
4306 V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1,
4307 DAG.getConstant(0, MVT::i64));
4309 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1);
4312 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4313 /// the specified operations to build the shuffle.
4314 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4315 SDValue RHS, SelectionDAG &DAG,
4317 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4318 unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1);
4319 unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1);
4322 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4331 OP_VUZPL, // VUZP, left result
4332 OP_VUZPR, // VUZP, right result
4333 OP_VZIPL, // VZIP, left result
4334 OP_VZIPR, // VZIP, right result
4335 OP_VTRNL, // VTRN, left result
4336 OP_VTRNR // VTRN, right result
4339 if (OpNum == OP_COPY) {
4340 if (LHSID == (1 * 9 + 2) * 9 + 3)
4342 assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!");
4346 SDValue OpLHS, OpRHS;
4347 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4348 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4349 EVT VT = OpLHS.getValueType();
4353 llvm_unreachable("Unknown shuffle opcode!");
4355 // VREV divides the vector in half and swaps within the half.
4356 if (VT.getVectorElementType() == MVT::i32 ||
4357 VT.getVectorElementType() == MVT::f32)
4358 return DAG.getNode(ARM64ISD::REV64, dl, VT, OpLHS);
4359 // vrev <4 x i16> -> REV32
4360 if (VT.getVectorElementType() == MVT::i16)
4361 return DAG.getNode(ARM64ISD::REV32, dl, VT, OpLHS);
4362 // vrev <4 x i8> -> REV16
4363 assert(VT.getVectorElementType() == MVT::i8);
4364 return DAG.getNode(ARM64ISD::REV16, dl, VT, OpLHS);
4369 EVT EltTy = VT.getVectorElementType();
4371 if (EltTy == MVT::i8)
4372 Opcode = ARM64ISD::DUPLANE8;
4373 else if (EltTy == MVT::i16)
4374 Opcode = ARM64ISD::DUPLANE16;
4375 else if (EltTy == MVT::i32 || EltTy == MVT::f32)
4376 Opcode = ARM64ISD::DUPLANE32;
4377 else if (EltTy == MVT::i64 || EltTy == MVT::f64)
4378 Opcode = ARM64ISD::DUPLANE64;
4380 llvm_unreachable("Invalid vector element type?");
4382 if (VT.getSizeInBits() == 64)
4383 OpLHS = WidenVector(OpLHS, DAG);
4384 SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, MVT::i64);
4385 return DAG.getNode(Opcode, dl, VT, OpLHS, Lane);
4390 unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
4391 return DAG.getNode(ARM64ISD::EXT, dl, VT, OpLHS, OpRHS,
4392 DAG.getConstant(Imm, MVT::i32));
4395 return DAG.getNode(ARM64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4397 return DAG.getNode(ARM64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4399 return DAG.getNode(ARM64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4401 return DAG.getNode(ARM64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4403 return DAG.getNode(ARM64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4405 return DAG.getNode(ARM64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4409 static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask,
4410 SelectionDAG &DAG) {
4411 // Check to see if we can use the TBL instruction.
4412 SDValue V1 = Op.getOperand(0);
4413 SDValue V2 = Op.getOperand(1);
4416 EVT EltVT = Op.getValueType().getVectorElementType();
4417 unsigned BytesPerElt = EltVT.getSizeInBits() / 8;
4419 SmallVector<SDValue, 8> TBLMask;
4420 for (int Val : ShuffleMask) {
4421 for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
4422 unsigned Offset = Byte + Val * BytesPerElt;
4423 TBLMask.push_back(DAG.getConstant(Offset, MVT::i32));
4427 MVT IndexVT = MVT::v8i8;
4428 unsigned IndexLen = 8;
4429 if (Op.getValueType().getSizeInBits() == 128) {
4430 IndexVT = MVT::v16i8;
4434 SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
4435 SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);
4438 if (V2.getNode()->getOpcode() == ISD::UNDEF) {
4440 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
4441 Shuffle = DAG.getNode(
4442 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4443 DAG.getConstant(Intrinsic::arm64_neon_tbl1, MVT::i32), V1Cst,
4444 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4445 makeArrayRef(TBLMask.data(), IndexLen)));
4447 if (IndexLen == 8) {
4448 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
4449 Shuffle = DAG.getNode(
4450 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4451 DAG.getConstant(Intrinsic::arm64_neon_tbl1, MVT::i32), V1Cst,
4452 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4453 makeArrayRef(TBLMask.data(), IndexLen)));
4455 // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
4456 // cannot currently represent the register constraints on the input
4458 // Shuffle = DAG.getNode(ARM64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
4459 // DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4460 // &TBLMask[0], IndexLen));
4461 Shuffle = DAG.getNode(
4462 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4463 DAG.getConstant(Intrinsic::arm64_neon_tbl2, MVT::i32), V1Cst, V2Cst,
4464 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4465 makeArrayRef(TBLMask.data(), IndexLen)));
4468 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
4471 static unsigned getDUPLANEOp(EVT EltType) {
4472 if (EltType == MVT::i8)
4473 return ARM64ISD::DUPLANE8;
4474 if (EltType == MVT::i16)
4475 return ARM64ISD::DUPLANE16;
4476 if (EltType == MVT::i32 || EltType == MVT::f32)
4477 return ARM64ISD::DUPLANE32;
4478 if (EltType == MVT::i64 || EltType == MVT::f64)
4479 return ARM64ISD::DUPLANE64;
4481 llvm_unreachable("Invalid vector element type?");
4484 SDValue ARM64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4485 SelectionDAG &DAG) const {
4487 EVT VT = Op.getValueType();
4489 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4491 // Convert shuffles that are directly supported on NEON to target-specific
4492 // DAG nodes, instead of keeping them as shuffles and matching them again
4493 // during code selection. This is more efficient and avoids the possibility
4494 // of inconsistencies between legalization and selection.
4495 ArrayRef<int> ShuffleMask = SVN->getMask();
4497 SDValue V1 = Op.getOperand(0);
4498 SDValue V2 = Op.getOperand(1);
4500 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0],
4501 V1.getValueType().getSimpleVT())) {
4502 int Lane = SVN->getSplatIndex();
4503 // If this is undef splat, generate it via "just" vdup, if possible.
4507 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
4508 return DAG.getNode(ARM64ISD::DUP, dl, V1.getValueType(),
4510 // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
4511 // constant. If so, we can just reference the lane's definition directly.
4512 if (V1.getOpcode() == ISD::BUILD_VECTOR &&
4513 !isa<ConstantSDNode>(V1.getOperand(Lane)))
4514 return DAG.getNode(ARM64ISD::DUP, dl, VT, V1.getOperand(Lane));
4516 // Otherwise, duplicate from the lane of the input vector.
4517 unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
4519 // SelectionDAGBuilder may have "helpfully" already extracted or conatenated
4520 // to make a vector of the same size as this SHUFFLE. We can ignore the
4521 // extract entirely, and canonicalise the concat using WidenVector.
4522 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
4523 Lane += cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
4524 V1 = V1.getOperand(0);
4525 } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
4526 unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
4527 Lane -= Idx * VT.getVectorNumElements() / 2;
4528 V1 = WidenVector(V1.getOperand(Idx), DAG);
4529 } else if (VT.getSizeInBits() == 64)
4530 V1 = WidenVector(V1, DAG);
4532 return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, MVT::i64));
4535 if (isREVMask(ShuffleMask, VT, 64))
4536 return DAG.getNode(ARM64ISD::REV64, dl, V1.getValueType(), V1, V2);
4537 if (isREVMask(ShuffleMask, VT, 32))
4538 return DAG.getNode(ARM64ISD::REV32, dl, V1.getValueType(), V1, V2);
4539 if (isREVMask(ShuffleMask, VT, 16))
4540 return DAG.getNode(ARM64ISD::REV16, dl, V1.getValueType(), V1, V2);
4542 bool ReverseEXT = false;
4544 if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
4547 Imm *= getExtFactor(V1);
4548 return DAG.getNode(ARM64ISD::EXT, dl, V1.getValueType(), V1, V2,
4549 DAG.getConstant(Imm, MVT::i32));
4550 } else if (V2->getOpcode() == ISD::UNDEF &&
4551 isSingletonEXTMask(ShuffleMask, VT, Imm)) {
4552 Imm *= getExtFactor(V1);
4553 return DAG.getNode(ARM64ISD::EXT, dl, V1.getValueType(), V1, V1,
4554 DAG.getConstant(Imm, MVT::i32));
4557 unsigned WhichResult;
4558 if (isZIPMask(ShuffleMask, VT, WhichResult)) {
4559 unsigned Opc = (WhichResult == 0) ? ARM64ISD::ZIP1 : ARM64ISD::ZIP2;
4560 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4562 if (isUZPMask(ShuffleMask, VT, WhichResult)) {
4563 unsigned Opc = (WhichResult == 0) ? ARM64ISD::UZP1 : ARM64ISD::UZP2;
4564 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4566 if (isTRNMask(ShuffleMask, VT, WhichResult)) {
4567 unsigned Opc = (WhichResult == 0) ? ARM64ISD::TRN1 : ARM64ISD::TRN2;
4568 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4571 if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4572 unsigned Opc = (WhichResult == 0) ? ARM64ISD::ZIP1 : ARM64ISD::ZIP2;
4573 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4575 if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4576 unsigned Opc = (WhichResult == 0) ? ARM64ISD::UZP1 : ARM64ISD::UZP2;
4577 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4579 if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4580 unsigned Opc = (WhichResult == 0) ? ARM64ISD::TRN1 : ARM64ISD::TRN2;
4581 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4584 SDValue Concat = tryFormConcatFromShuffle(Op, DAG);
4585 if (Concat.getNode())
4590 int NumInputElements = V1.getValueType().getVectorNumElements();
4591 if (isINSMask(ShuffleMask, NumInputElements, DstIsLeft, Anomaly)) {
4592 SDValue DstVec = DstIsLeft ? V1 : V2;
4593 SDValue DstLaneV = DAG.getConstant(Anomaly, MVT::i64);
4595 SDValue SrcVec = V1;
4596 int SrcLane = ShuffleMask[Anomaly];
4597 if (SrcLane >= NumInputElements) {
4599 SrcLane -= VT.getVectorNumElements();
4601 SDValue SrcLaneV = DAG.getConstant(SrcLane, MVT::i64);
4603 EVT ScalarVT = VT.getVectorElementType();
4604 if (ScalarVT.getSizeInBits() < 32)
4605 ScalarVT = MVT::i32;
4608 ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
4609 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ScalarVT, SrcVec, SrcLaneV),
4613 // If the shuffle is not directly supported and it has 4 elements, use
4614 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4615 unsigned NumElts = VT.getVectorNumElements();
4617 unsigned PFIndexes[4];
4618 for (unsigned i = 0; i != 4; ++i) {
4619 if (ShuffleMask[i] < 0)
4622 PFIndexes[i] = ShuffleMask[i];
4625 // Compute the index in the perfect shuffle table.
4626 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
4627 PFIndexes[2] * 9 + PFIndexes[3];
4628 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4629 unsigned Cost = (PFEntry >> 30);
4632 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4635 return GenerateTBL(Op, ShuffleMask, DAG);
4638 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
4640 EVT VT = BVN->getValueType(0);
4641 APInt SplatBits, SplatUndef;
4642 unsigned SplatBitSize;
4644 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4645 unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;
4647 for (unsigned i = 0; i < NumSplats; ++i) {
4648 CnstBits <<= SplatBitSize;
4649 UndefBits <<= SplatBitSize;
4650 CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits());
4651 UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits());
4660 SDValue ARM64TargetLowering::LowerVectorAND(SDValue Op,
4661 SelectionDAG &DAG) const {
4662 BuildVectorSDNode *BVN =
4663 dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
4664 SDValue LHS = Op.getOperand(0);
4666 EVT VT = Op.getValueType();
4671 APInt CnstBits(VT.getSizeInBits(), 0);
4672 APInt UndefBits(VT.getSizeInBits(), 0);
4673 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
4674 // We only have BIC vector immediate instruction, which is and-not.
4675 CnstBits = ~CnstBits;
4677 // We make use of a little bit of goto ickiness in order to avoid having to
4678 // duplicate the immediate matching logic for the undef toggled case.
4679 bool SecondTry = false;
4682 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
4683 CnstBits = CnstBits.zextOrTrunc(64);
4684 uint64_t CnstVal = CnstBits.getZExtValue();
4686 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
4687 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
4688 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4689 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4690 DAG.getConstant(CnstVal, MVT::i32),
4691 DAG.getConstant(0, MVT::i32));
4692 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4695 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
4696 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
4697 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4698 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4699 DAG.getConstant(CnstVal, MVT::i32),
4700 DAG.getConstant(8, MVT::i32));
4701 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4704 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
4705 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
4706 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4707 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4708 DAG.getConstant(CnstVal, MVT::i32),
4709 DAG.getConstant(16, MVT::i32));
4710 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4713 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
4714 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
4715 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4716 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4717 DAG.getConstant(CnstVal, MVT::i32),
4718 DAG.getConstant(24, MVT::i32));
4719 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4722 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
4723 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
4724 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4725 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4726 DAG.getConstant(CnstVal, MVT::i32),
4727 DAG.getConstant(0, MVT::i32));
4728 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4731 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
4732 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
4733 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4734 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4735 DAG.getConstant(CnstVal, MVT::i32),
4736 DAG.getConstant(8, MVT::i32));
4737 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4744 CnstBits = ~UndefBits;
4748 // We can always fall back to a non-immediate AND.
4753 // Specialized code to quickly find if PotentialBVec is a BuildVector that
4754 // consists of only the same constant int value, returned in reference arg
4756 static bool isAllConstantBuildVector(const SDValue &PotentialBVec,
4757 uint64_t &ConstVal) {
4758 BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec);
4761 ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0));
4764 EVT VT = Bvec->getValueType(0);
4765 unsigned NumElts = VT.getVectorNumElements();
4766 for (unsigned i = 1; i < NumElts; ++i)
4767 if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt)
4769 ConstVal = FirstElt->getZExtValue();
4773 static unsigned getIntrinsicID(const SDNode *N) {
4774 unsigned Opcode = N->getOpcode();
4777 return Intrinsic::not_intrinsic;
4778 case ISD::INTRINSIC_WO_CHAIN: {
4779 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4780 if (IID < Intrinsic::num_intrinsics)
4782 return Intrinsic::not_intrinsic;
4787 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
4788 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
4789 // BUILD_VECTORs with constant element C1, C2 is a constant, and C1 == ~C2.
4790 // Also, logical shift right -> sri, with the same structure.
4791 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
4792 EVT VT = N->getValueType(0);
4799 // Is the first op an AND?
4800 const SDValue And = N->getOperand(0);
4801 if (And.getOpcode() != ISD::AND)
4804 // Is the second op an shl or lshr?
4805 SDValue Shift = N->getOperand(1);
4806 // This will have been turned into: ARM64ISD::VSHL vector, #shift
4807 // or ARM64ISD::VLSHR vector, #shift
4808 unsigned ShiftOpc = Shift.getOpcode();
4809 if ((ShiftOpc != ARM64ISD::VSHL && ShiftOpc != ARM64ISD::VLSHR))
4811 bool IsShiftRight = ShiftOpc == ARM64ISD::VLSHR;
4813 // Is the shift amount constant?
4814 ConstantSDNode *C2node = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
4818 // Is the and mask vector all constant?
4820 if (!isAllConstantBuildVector(And.getOperand(1), C1))
4823 // Is C1 == ~C2, taking into account how much one can shift elements of a
4825 uint64_t C2 = C2node->getZExtValue();
4826 unsigned ElemSizeInBits = VT.getVectorElementType().getSizeInBits();
4827 if (C2 > ElemSizeInBits)
4829 unsigned ElemMask = (1 << ElemSizeInBits) - 1;
4830 if ((C1 & ElemMask) != (~C2 & ElemMask))
4833 SDValue X = And.getOperand(0);
4834 SDValue Y = Shift.getOperand(0);
4837 IsShiftRight ? Intrinsic::arm64_neon_vsri : Intrinsic::arm64_neon_vsli;
4839 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
4840 DAG.getConstant(Intrin, MVT::i32), X, Y, Shift.getOperand(1));
4842 DEBUG(dbgs() << "arm64-lower: transformed: \n");
4843 DEBUG(N->dump(&DAG));
4844 DEBUG(dbgs() << "into: \n");
4845 DEBUG(ResultSLI->dump(&DAG));
4851 SDValue ARM64TargetLowering::LowerVectorOR(SDValue Op,
4852 SelectionDAG &DAG) const {
4853 // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
4854 if (EnableARM64SlrGeneration) {
4855 SDValue Res = tryLowerToSLI(Op.getNode(), DAG);
4860 BuildVectorSDNode *BVN =
4861 dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode());
4862 SDValue LHS = Op.getOperand(1);
4864 EVT VT = Op.getValueType();
4866 // OR commutes, so try swapping the operands.
4868 LHS = Op.getOperand(0);
4869 BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
4874 APInt CnstBits(VT.getSizeInBits(), 0);
4875 APInt UndefBits(VT.getSizeInBits(), 0);
4876 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
4877 // We make use of a little bit of goto ickiness in order to avoid having to
4878 // duplicate the immediate matching logic for the undef toggled case.
4879 bool SecondTry = false;
4882 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
4883 CnstBits = CnstBits.zextOrTrunc(64);
4884 uint64_t CnstVal = CnstBits.getZExtValue();
4886 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
4887 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
4888 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4889 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
4890 DAG.getConstant(CnstVal, MVT::i32),
4891 DAG.getConstant(0, MVT::i32));
4892 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4895 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
4896 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
4897 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4898 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
4899 DAG.getConstant(CnstVal, MVT::i32),
4900 DAG.getConstant(8, MVT::i32));
4901 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4904 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
4905 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
4906 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4907 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
4908 DAG.getConstant(CnstVal, MVT::i32),
4909 DAG.getConstant(16, MVT::i32));
4910 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4913 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
4914 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
4915 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4916 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
4917 DAG.getConstant(CnstVal, MVT::i32),
4918 DAG.getConstant(24, MVT::i32));
4919 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4922 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
4923 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
4924 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4925 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
4926 DAG.getConstant(CnstVal, MVT::i32),
4927 DAG.getConstant(0, MVT::i32));
4928 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4931 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
4932 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
4933 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4934 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
4935 DAG.getConstant(CnstVal, MVT::i32),
4936 DAG.getConstant(8, MVT::i32));
4937 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4944 CnstBits = UndefBits;
4948 // We can always fall back to a non-immediate OR.
4953 SDValue ARM64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
4954 SelectionDAG &DAG) const {
4955 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4957 EVT VT = Op.getValueType();
4959 APInt CnstBits(VT.getSizeInBits(), 0);
4960 APInt UndefBits(VT.getSizeInBits(), 0);
4961 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
4962 // We make use of a little bit of goto ickiness in order to avoid having to
4963 // duplicate the immediate matching logic for the undef toggled case.
4964 bool SecondTry = false;
4967 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
4968 CnstBits = CnstBits.zextOrTrunc(64);
4969 uint64_t CnstVal = CnstBits.getZExtValue();
4971 // Certain magic vector constants (used to express things like NOT
4972 // and NEG) are passed through unmodified. This allows codegen patterns
4973 // for these operations to match. Special-purpose patterns will lower
4974 // these immediates to MOVIs if it proves necessary.
4975 if (VT.isInteger() && (CnstVal == 0 || CnstVal == ~0ULL))
4978 // The many faces of MOVI...
4979 if (ARM64_AM::isAdvSIMDModImmType10(CnstVal)) {
4980 CnstVal = ARM64_AM::encodeAdvSIMDModImmType10(CnstVal);
4981 if (VT.getSizeInBits() == 128) {
4982 SDValue Mov = DAG.getNode(ARM64ISD::MOVIedit, dl, MVT::v2i64,
4983 DAG.getConstant(CnstVal, MVT::i32));
4984 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4987 // Support the V64 version via subregister insertion.
4988 SDValue Mov = DAG.getNode(ARM64ISD::MOVIedit, dl, MVT::f64,
4989 DAG.getConstant(CnstVal, MVT::i32));
4990 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4993 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
4994 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
4995 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4996 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
4997 DAG.getConstant(CnstVal, MVT::i32),
4998 DAG.getConstant(0, MVT::i32));
4999 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5002 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
5003 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
5004 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5005 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5006 DAG.getConstant(CnstVal, MVT::i32),
5007 DAG.getConstant(8, MVT::i32));
5008 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5011 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
5012 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
5013 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5014 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5015 DAG.getConstant(CnstVal, MVT::i32),
5016 DAG.getConstant(16, MVT::i32));
5017 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5020 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
5021 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
5022 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5023 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5024 DAG.getConstant(CnstVal, MVT::i32),
5025 DAG.getConstant(24, MVT::i32));
5026 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5029 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
5030 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
5031 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5032 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5033 DAG.getConstant(CnstVal, MVT::i32),
5034 DAG.getConstant(0, MVT::i32));
5035 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5038 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
5039 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
5040 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5041 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5042 DAG.getConstant(CnstVal, MVT::i32),
5043 DAG.getConstant(8, MVT::i32));
5044 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5047 if (ARM64_AM::isAdvSIMDModImmType7(CnstVal)) {
5048 CnstVal = ARM64_AM::encodeAdvSIMDModImmType7(CnstVal);
5049 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5050 SDValue Mov = DAG.getNode(ARM64ISD::MOVImsl, dl, MovTy,
5051 DAG.getConstant(CnstVal, MVT::i32),
5052 DAG.getConstant(264, MVT::i32));
5053 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5056 if (ARM64_AM::isAdvSIMDModImmType8(CnstVal)) {
5057 CnstVal = ARM64_AM::encodeAdvSIMDModImmType8(CnstVal);
5058 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5059 SDValue Mov = DAG.getNode(ARM64ISD::MOVImsl, dl, MovTy,
5060 DAG.getConstant(CnstVal, MVT::i32),
5061 DAG.getConstant(272, MVT::i32));
5062 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5065 if (ARM64_AM::isAdvSIMDModImmType9(CnstVal)) {
5066 CnstVal = ARM64_AM::encodeAdvSIMDModImmType9(CnstVal);
5067 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;
5068 SDValue Mov = DAG.getNode(ARM64ISD::MOVI, dl, MovTy,
5069 DAG.getConstant(CnstVal, MVT::i32));
5070 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5073 // The few faces of FMOV...
5074 if (ARM64_AM::isAdvSIMDModImmType11(CnstVal)) {
5075 CnstVal = ARM64_AM::encodeAdvSIMDModImmType11(CnstVal);
5076 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4f32 : MVT::v2f32;
5077 SDValue Mov = DAG.getNode(ARM64ISD::FMOV, dl, MovTy,
5078 DAG.getConstant(CnstVal, MVT::i32));
5079 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5082 if (ARM64_AM::isAdvSIMDModImmType12(CnstVal) &&
5083 VT.getSizeInBits() == 128) {
5084 CnstVal = ARM64_AM::encodeAdvSIMDModImmType12(CnstVal);
5085 SDValue Mov = DAG.getNode(ARM64ISD::FMOV, dl, MVT::v2f64,
5086 DAG.getConstant(CnstVal, MVT::i32));
5087 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5090 // The many faces of MVNI...
5092 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
5093 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
5094 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5095 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5096 DAG.getConstant(CnstVal, MVT::i32),
5097 DAG.getConstant(0, MVT::i32));
5098 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5101 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
5102 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
5103 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5104 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5105 DAG.getConstant(CnstVal, MVT::i32),
5106 DAG.getConstant(8, MVT::i32));
5107 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5110 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
5111 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
5112 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5113 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5114 DAG.getConstant(CnstVal, MVT::i32),
5115 DAG.getConstant(16, MVT::i32));
5116 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5119 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
5120 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
5121 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5122 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5123 DAG.getConstant(CnstVal, MVT::i32),
5124 DAG.getConstant(24, MVT::i32));
5125 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5128 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
5129 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
5130 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5131 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5132 DAG.getConstant(CnstVal, MVT::i32),
5133 DAG.getConstant(0, MVT::i32));
5134 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5137 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
5138 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
5139 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5140 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5141 DAG.getConstant(CnstVal, MVT::i32),
5142 DAG.getConstant(8, MVT::i32));
5143 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5146 if (ARM64_AM::isAdvSIMDModImmType7(CnstVal)) {
5147 CnstVal = ARM64_AM::encodeAdvSIMDModImmType7(CnstVal);
5148 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5149 SDValue Mov = DAG.getNode(ARM64ISD::MVNImsl, dl, MovTy,
5150 DAG.getConstant(CnstVal, MVT::i32),
5151 DAG.getConstant(264, MVT::i32));
5152 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5155 if (ARM64_AM::isAdvSIMDModImmType8(CnstVal)) {
5156 CnstVal = ARM64_AM::encodeAdvSIMDModImmType8(CnstVal);
5157 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5158 SDValue Mov = DAG.getNode(ARM64ISD::MVNImsl, dl, MovTy,
5159 DAG.getConstant(CnstVal, MVT::i32),
5160 DAG.getConstant(272, MVT::i32));
5161 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5168 CnstBits = UndefBits;
5173 // Scan through the operands to find some interesting properties we can
5175 // 1) If only one value is used, we can use a DUP, or
5176 // 2) if only the low element is not undef, we can just insert that, or
5177 // 3) if only one constant value is used (w/ some non-constant lanes),
5178 // we can splat the constant value into the whole vector then fill
5179 // in the non-constant lanes.
5180 // 4) FIXME: If different constant values are used, but we can intelligently
5181 // select the values we'll be overwriting for the non-constant
5182 // lanes such that we can directly materialize the vector
5183 // some other way (MOVI, e.g.), we can be sneaky.
5184 unsigned NumElts = VT.getVectorNumElements();
5185 bool isOnlyLowElement = true;
5186 bool usesOnlyOneValue = true;
5187 bool usesOnlyOneConstantValue = true;
5188 bool isConstant = true;
5189 unsigned NumConstantLanes = 0;
5191 SDValue ConstantValue;
5192 for (unsigned i = 0; i < NumElts; ++i) {
5193 SDValue V = Op.getOperand(i);
5194 if (V.getOpcode() == ISD::UNDEF)
5197 isOnlyLowElement = false;
5198 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5201 if (isa<ConstantSDNode>(V) || isa<ConstantFPSDNode>(V)) {
5203 if (!ConstantValue.getNode())
5205 else if (ConstantValue != V)
5206 usesOnlyOneConstantValue = false;
5209 if (!Value.getNode())
5211 else if (V != Value)
5212 usesOnlyOneValue = false;
5215 if (!Value.getNode())
5216 return DAG.getUNDEF(VT);
5218 if (isOnlyLowElement)
5219 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5221 // Use DUP for non-constant splats. For f32 constant splats, reduce to
5222 // i32 and try again.
5223 if (usesOnlyOneValue) {
5225 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5226 Value.getValueType() != VT)
5227 return DAG.getNode(ARM64ISD::DUP, dl, VT, Value);
5229 // This is actually a DUPLANExx operation, which keeps everything vectory.
5231 // DUPLANE works on 128-bit vectors, widen it if necessary.
5232 SDValue Lane = Value.getOperand(1);
5233 Value = Value.getOperand(0);
5234 if (Value.getValueType().getSizeInBits() == 64)
5235 Value = WidenVector(Value, DAG);
5237 unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
5238 return DAG.getNode(Opcode, dl, VT, Value, Lane);
5241 if (VT.getVectorElementType().isFloatingPoint()) {
5242 SmallVector<SDValue, 8> Ops;
5244 (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
5245 for (unsigned i = 0; i < NumElts; ++i)
5246 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, NewType, Op.getOperand(i)));
5247 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts);
5248 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5249 Val = LowerBUILD_VECTOR(Val, DAG);
5251 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5255 // If there was only one constant value used and for more than one lane,
5256 // start by splatting that value, then replace the non-constant lanes. This
5257 // is better than the default, which will perform a separate initialization
5259 if (NumConstantLanes > 0 && usesOnlyOneConstantValue) {
5260 SDValue Val = DAG.getNode(ARM64ISD::DUP, dl, VT, ConstantValue);
5261 // Now insert the non-constant lanes.
5262 for (unsigned i = 0; i < NumElts; ++i) {
5263 SDValue V = Op.getOperand(i);
5264 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5265 if (!isa<ConstantSDNode>(V) && !isa<ConstantFPSDNode>(V)) {
5266 // Note that type legalization likely mucked about with the VT of the
5267 // source operand, so we may have to convert it here before inserting.
5268 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, V, LaneIdx);
5274 // If all elements are constants and the case above didn't get hit, fall back
5275 // to the default expansion, which will generate a load from the constant
5280 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5282 SDValue shuffle = ReconstructShuffle(Op, DAG);
5283 if (shuffle != SDValue())
5287 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5288 // know the default expansion would otherwise fall back on something even
5289 // worse. For a vector with one or two non-undef values, that's
5290 // scalar_to_vector for the elements followed by a shuffle (provided the
5291 // shuffle is valid for the target) and materialization element by element
5292 // on the stack followed by a load for everything else.
5293 if (!isConstant && !usesOnlyOneValue) {
5294 SDValue Vec = DAG.getUNDEF(VT);
5295 SDValue Op0 = Op.getOperand(0);
5296 unsigned ElemSize = VT.getVectorElementType().getSizeInBits();
5298 // For 32 and 64 bit types, use INSERT_SUBREG for lane zero to
5299 // a) Avoid a RMW dependency on the full vector register, and
5300 // b) Allow the register coalescer to fold away the copy if the
5301 // value is already in an S or D register.
5302 if (Op0.getOpcode() != ISD::UNDEF && (ElemSize == 32 || ElemSize == 64)) {
5303 unsigned SubIdx = ElemSize == 32 ? ARM64::ssub : ARM64::dsub;
5305 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0,
5306 DAG.getTargetConstant(SubIdx, MVT::i32));
5307 Vec = SDValue(N, 0);
5310 for (; i < NumElts; ++i) {
5311 SDValue V = Op.getOperand(i);
5312 if (V.getOpcode() == ISD::UNDEF)
5314 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5315 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5320 // Just use the default expansion. We failed to find a better alternative.
5324 SDValue ARM64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
5325 SelectionDAG &DAG) const {
5326 assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!");
5328 // Check for non-constant lane.
5329 if (!isa<ConstantSDNode>(Op.getOperand(2)))
5332 EVT VT = Op.getOperand(0).getValueType();
5334 // Insertion/extraction are legal for V128 types.
5335 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
5336 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
5339 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
5340 VT != MVT::v1i64 && VT != MVT::v2f32)
5343 // For V64 types, we perform insertion by expanding the value
5344 // to a V128 type and perform the insertion on that.
5346 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
5347 EVT WideTy = WideVec.getValueType();
5349 SDValue Node = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideTy, WideVec,
5350 Op.getOperand(1), Op.getOperand(2));
5351 // Re-narrow the resultant vector.
5352 return NarrowVector(Node, DAG);
5355 SDValue ARM64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5356 SelectionDAG &DAG) const {
5357 assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!");
5359 // Check for non-constant lane.
5360 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5363 EVT VT = Op.getOperand(0).getValueType();
5365 // Insertion/extraction are legal for V128 types.
5366 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
5367 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
5370 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
5371 VT != MVT::v1i64 && VT != MVT::v2f32)
5374 // For V64 types, we perform extraction by expanding the value
5375 // to a V128 type and perform the extraction on that.
5377 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
5378 EVT WideTy = WideVec.getValueType();
5380 EVT ExtrTy = WideTy.getVectorElementType();
5381 if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8)
5384 // For extractions, we just return the result directly.
5385 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec,
5389 SDValue ARM64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
5390 SelectionDAG &DAG) const {
5391 EVT VT = Op.getOperand(0).getValueType();
5397 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5400 unsigned Val = Cst->getZExtValue();
5402 unsigned Size = Op.getValueType().getSizeInBits();
5406 return DAG.getTargetExtractSubreg(ARM64::bsub, dl, Op.getValueType(),
5409 return DAG.getTargetExtractSubreg(ARM64::hsub, dl, Op.getValueType(),
5412 return DAG.getTargetExtractSubreg(ARM64::ssub, dl, Op.getValueType(),
5415 return DAG.getTargetExtractSubreg(ARM64::dsub, dl, Op.getValueType(),
5418 llvm_unreachable("Unexpected vector type in extract_subvector!");
5421 // If this is extracting the upper 64-bits of a 128-bit vector, we match
5423 if (Size == 64 && Val * VT.getVectorElementType().getSizeInBits() == 64)
5429 bool ARM64TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5431 if (VT.getVectorNumElements() == 4 &&
5432 (VT.is128BitVector() || VT.is64BitVector())) {
5433 unsigned PFIndexes[4];
5434 for (unsigned i = 0; i != 4; ++i) {
5438 PFIndexes[i] = M[i];
5441 // Compute the index in the perfect shuffle table.
5442 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
5443 PFIndexes[2] * 9 + PFIndexes[3];
5444 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5445 unsigned Cost = (PFEntry >> 30);
5453 unsigned DummyUnsigned;
5455 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || isREVMask(M, VT, 64) ||
5456 isREVMask(M, VT, 32) || isREVMask(M, VT, 16) ||
5457 isEXTMask(M, VT, DummyBool, DummyUnsigned) ||
5458 // isTBLMask(M, VT) || // FIXME: Port TBL support from ARM.
5459 isTRNMask(M, VT, DummyUnsigned) || isUZPMask(M, VT, DummyUnsigned) ||
5460 isZIPMask(M, VT, DummyUnsigned) ||
5461 isTRN_v_undef_Mask(M, VT, DummyUnsigned) ||
5462 isUZP_v_undef_Mask(M, VT, DummyUnsigned) ||
5463 isZIP_v_undef_Mask(M, VT, DummyUnsigned) ||
5464 isINSMask(M, VT.getVectorNumElements(), DummyBool, DummyInt) ||
5465 isConcatMask(M, VT, VT.getSizeInBits() == 128));
5468 /// getVShiftImm - Check if this is a valid build_vector for the immediate
5469 /// operand of a vector shift operation, where all the elements of the
5470 /// build_vector must have the same constant integer value.
5471 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
5472 // Ignore bit_converts.
5473 while (Op.getOpcode() == ISD::BITCAST)
5474 Op = Op.getOperand(0);
5475 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
5476 APInt SplatBits, SplatUndef;
5477 unsigned SplatBitSize;
5479 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
5480 HasAnyUndefs, ElementBits) ||
5481 SplatBitSize > ElementBits)
5483 Cnt = SplatBits.getSExtValue();
5487 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
5488 /// operand of a vector shift left operation. That value must be in the range:
5489 /// 0 <= Value < ElementBits for a left shift; or
5490 /// 0 <= Value <= ElementBits for a long left shift.
5491 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
5492 assert(VT.isVector() && "vector shift count is not a vector type");
5493 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5494 if (!getVShiftImm(Op, ElementBits, Cnt))
5496 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
5499 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
5500 /// operand of a vector shift right operation. For a shift opcode, the value
5501 /// is positive, but for an intrinsic the value count must be negative. The
5502 /// absolute value must be in the range:
5503 /// 1 <= |Value| <= ElementBits for a right shift; or
5504 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
5505 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
5507 assert(VT.isVector() && "vector shift count is not a vector type");
5508 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5509 if (!getVShiftImm(Op, ElementBits, Cnt))
5513 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
5516 SDValue ARM64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
5517 SelectionDAG &DAG) const {
5518 EVT VT = Op.getValueType();
5522 if (!Op.getOperand(1).getValueType().isVector())
5524 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5526 switch (Op.getOpcode()) {
5528 llvm_unreachable("unexpected shift opcode");
5531 if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
5532 return DAG.getNode(ARM64ISD::VSHL, SDLoc(Op), VT, Op.getOperand(0),
5533 DAG.getConstant(Cnt, MVT::i32));
5534 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5535 DAG.getConstant(Intrinsic::arm64_neon_ushl, MVT::i32),
5536 Op.getOperand(0), Op.getOperand(1));
5539 // Right shift immediate
5540 if (isVShiftRImm(Op.getOperand(1), VT, false, false, Cnt) &&
5543 (Op.getOpcode() == ISD::SRA) ? ARM64ISD::VASHR : ARM64ISD::VLSHR;
5544 return DAG.getNode(Opc, SDLoc(Op), VT, Op.getOperand(0),
5545 DAG.getConstant(Cnt, MVT::i32));
5548 // Right shift register. Note, there is not a shift right register
5549 // instruction, but the shift left register instruction takes a signed
5550 // value, where negative numbers specify a right shift.
5551 unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::arm64_neon_sshl
5552 : Intrinsic::arm64_neon_ushl;
5553 // negate the shift amount
5554 SDValue NegShift = DAG.getNode(ARM64ISD::NEG, DL, VT, Op.getOperand(1));
5555 SDValue NegShiftLeft =
5556 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5557 DAG.getConstant(Opc, MVT::i32), Op.getOperand(0), NegShift);
5558 return NegShiftLeft;
5564 static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS,
5565 ARM64CC::CondCode CC, bool NoNans, EVT VT,
5566 SDLoc dl, SelectionDAG &DAG) {
5567 EVT SrcVT = LHS.getValueType();
5569 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
5570 APInt CnstBits(VT.getSizeInBits(), 0);
5571 APInt UndefBits(VT.getSizeInBits(), 0);
5572 bool IsCnst = BVN && resolveBuildVector(BVN, CnstBits, UndefBits);
5573 bool IsZero = IsCnst && (CnstBits == 0);
5575 if (SrcVT.getVectorElementType().isFloatingPoint()) {
5582 Fcmeq = DAG.getNode(ARM64ISD::FCMEQz, dl, VT, LHS);
5584 Fcmeq = DAG.getNode(ARM64ISD::FCMEQ, dl, VT, LHS, RHS);
5585 return DAG.getNode(ARM64ISD::NOT, dl, VT, Fcmeq);
5589 return DAG.getNode(ARM64ISD::FCMEQz, dl, VT, LHS);
5590 return DAG.getNode(ARM64ISD::FCMEQ, dl, VT, LHS, RHS);
5593 return DAG.getNode(ARM64ISD::FCMGEz, dl, VT, LHS);
5594 return DAG.getNode(ARM64ISD::FCMGE, dl, VT, LHS, RHS);
5597 return DAG.getNode(ARM64ISD::FCMGTz, dl, VT, LHS);
5598 return DAG.getNode(ARM64ISD::FCMGT, dl, VT, LHS, RHS);
5601 return DAG.getNode(ARM64ISD::FCMLEz, dl, VT, LHS);
5602 return DAG.getNode(ARM64ISD::FCMGE, dl, VT, RHS, LHS);
5606 // If we ignore NaNs then we can use to the MI implementation.
5610 return DAG.getNode(ARM64ISD::FCMLTz, dl, VT, LHS);
5611 return DAG.getNode(ARM64ISD::FCMGT, dl, VT, RHS, LHS);
5621 Cmeq = DAG.getNode(ARM64ISD::CMEQz, dl, VT, LHS);
5623 Cmeq = DAG.getNode(ARM64ISD::CMEQ, dl, VT, LHS, RHS);
5624 return DAG.getNode(ARM64ISD::NOT, dl, VT, Cmeq);
5628 return DAG.getNode(ARM64ISD::CMEQz, dl, VT, LHS);
5629 return DAG.getNode(ARM64ISD::CMEQ, dl, VT, LHS, RHS);
5632 return DAG.getNode(ARM64ISD::CMGEz, dl, VT, LHS);
5633 return DAG.getNode(ARM64ISD::CMGE, dl, VT, LHS, RHS);
5636 return DAG.getNode(ARM64ISD::CMGTz, dl, VT, LHS);
5637 return DAG.getNode(ARM64ISD::CMGT, dl, VT, LHS, RHS);
5640 return DAG.getNode(ARM64ISD::CMLEz, dl, VT, LHS);
5641 return DAG.getNode(ARM64ISD::CMGE, dl, VT, RHS, LHS);
5643 return DAG.getNode(ARM64ISD::CMHS, dl, VT, RHS, LHS);
5645 return DAG.getNode(ARM64ISD::CMHI, dl, VT, RHS, LHS);
5648 return DAG.getNode(ARM64ISD::CMLTz, dl, VT, LHS);
5649 return DAG.getNode(ARM64ISD::CMGT, dl, VT, RHS, LHS);
5651 return DAG.getNode(ARM64ISD::CMHI, dl, VT, LHS, RHS);
5653 return DAG.getNode(ARM64ISD::CMHS, dl, VT, LHS, RHS);
5657 SDValue ARM64TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
5658 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5659 SDValue LHS = Op.getOperand(0);
5660 SDValue RHS = Op.getOperand(1);
5663 if (LHS.getValueType().getVectorElementType().isInteger()) {
5664 assert(LHS.getValueType() == RHS.getValueType());
5665 ARM64CC::CondCode ARM64CC = changeIntCCToARM64CC(CC);
5666 return EmitVectorComparison(LHS, RHS, ARM64CC, false, Op.getValueType(), dl,
5670 assert(LHS.getValueType().getVectorElementType() == MVT::f32 ||
5671 LHS.getValueType().getVectorElementType() == MVT::f64);
5673 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
5674 // clean. Some of them require two branches to implement.
5675 ARM64CC::CondCode CC1, CC2;
5677 changeVectorFPCCToARM64CC(CC, CC1, CC2, ShouldInvert);
5679 bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath;
5681 EmitVectorComparison(LHS, RHS, CC1, NoNaNs, Op.getValueType(), dl, DAG);
5685 if (CC2 != ARM64CC::AL) {
5687 EmitVectorComparison(LHS, RHS, CC2, NoNaNs, Op.getValueType(), dl, DAG);
5688 if (!Cmp2.getNode())
5691 Cmp = DAG.getNode(ISD::OR, dl, Cmp.getValueType(), Cmp, Cmp2);
5695 return Cmp = DAG.getNOT(dl, Cmp, Cmp.getValueType());
5700 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
5701 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
5702 /// specified in the intrinsic calls.
5703 bool ARM64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
5705 unsigned Intrinsic) const {
5706 switch (Intrinsic) {
5707 case Intrinsic::arm64_neon_ld2:
5708 case Intrinsic::arm64_neon_ld3:
5709 case Intrinsic::arm64_neon_ld4:
5710 case Intrinsic::arm64_neon_ld1x2:
5711 case Intrinsic::arm64_neon_ld1x3:
5712 case Intrinsic::arm64_neon_ld1x4:
5713 case Intrinsic::arm64_neon_ld2lane:
5714 case Intrinsic::arm64_neon_ld3lane:
5715 case Intrinsic::arm64_neon_ld4lane:
5716 case Intrinsic::arm64_neon_ld2r:
5717 case Intrinsic::arm64_neon_ld3r:
5718 case Intrinsic::arm64_neon_ld4r: {
5719 Info.opc = ISD::INTRINSIC_W_CHAIN;
5720 // Conservatively set memVT to the entire set of vectors loaded.
5721 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
5722 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5723 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
5726 Info.vol = false; // volatile loads with NEON intrinsics not supported
5727 Info.readMem = true;
5728 Info.writeMem = false;
5731 case Intrinsic::arm64_neon_st2:
5732 case Intrinsic::arm64_neon_st3:
5733 case Intrinsic::arm64_neon_st4:
5734 case Intrinsic::arm64_neon_st1x2:
5735 case Intrinsic::arm64_neon_st1x3:
5736 case Intrinsic::arm64_neon_st1x4:
5737 case Intrinsic::arm64_neon_st2lane:
5738 case Intrinsic::arm64_neon_st3lane:
5739 case Intrinsic::arm64_neon_st4lane: {
5740 Info.opc = ISD::INTRINSIC_VOID;
5741 // Conservatively set memVT to the entire set of vectors stored.
5742 unsigned NumElts = 0;
5743 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
5744 Type *ArgTy = I.getArgOperand(ArgI)->getType();
5745 if (!ArgTy->isVectorTy())
5747 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
5749 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5750 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
5753 Info.vol = false; // volatile stores with NEON intrinsics not supported
5754 Info.readMem = false;
5755 Info.writeMem = true;
5758 case Intrinsic::arm64_ldaxr:
5759 case Intrinsic::arm64_ldxr: {
5760 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
5761 Info.opc = ISD::INTRINSIC_W_CHAIN;
5762 Info.memVT = MVT::getVT(PtrTy->getElementType());
5763 Info.ptrVal = I.getArgOperand(0);
5765 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
5767 Info.readMem = true;
5768 Info.writeMem = false;
5771 case Intrinsic::arm64_stlxr:
5772 case Intrinsic::arm64_stxr: {
5773 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
5774 Info.opc = ISD::INTRINSIC_W_CHAIN;
5775 Info.memVT = MVT::getVT(PtrTy->getElementType());
5776 Info.ptrVal = I.getArgOperand(1);
5778 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
5780 Info.readMem = false;
5781 Info.writeMem = true;
5784 case Intrinsic::arm64_ldaxp:
5785 case Intrinsic::arm64_ldxp: {
5786 Info.opc = ISD::INTRINSIC_W_CHAIN;
5787 Info.memVT = MVT::i128;
5788 Info.ptrVal = I.getArgOperand(0);
5792 Info.readMem = true;
5793 Info.writeMem = false;
5796 case Intrinsic::arm64_stlxp:
5797 case Intrinsic::arm64_stxp: {
5798 Info.opc = ISD::INTRINSIC_W_CHAIN;
5799 Info.memVT = MVT::i128;
5800 Info.ptrVal = I.getArgOperand(2);
5804 Info.readMem = false;
5805 Info.writeMem = true;
5815 // Truncations from 64-bit GPR to 32-bit GPR is free.
5816 bool ARM64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
5817 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
5819 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
5820 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
5821 if (NumBits1 <= NumBits2)
5825 bool ARM64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
5826 if (!VT1.isInteger() || !VT2.isInteger())
5828 unsigned NumBits1 = VT1.getSizeInBits();
5829 unsigned NumBits2 = VT2.getSizeInBits();
5830 if (NumBits1 <= NumBits2)
5835 // All 32-bit GPR operations implicitly zero the high-half of the corresponding
5837 bool ARM64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
5838 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
5840 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
5841 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
5842 if (NumBits1 == 32 && NumBits2 == 64)
5846 bool ARM64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
5847 if (!VT1.isInteger() || !VT2.isInteger())
5849 unsigned NumBits1 = VT1.getSizeInBits();
5850 unsigned NumBits2 = VT2.getSizeInBits();
5851 if (NumBits1 == 32 && NumBits2 == 64)
5856 bool ARM64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
5857 EVT VT1 = Val.getValueType();
5858 if (isZExtFree(VT1, VT2)) {
5862 if (Val.getOpcode() != ISD::LOAD)
5865 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
5866 return (VT1.isSimple() && VT1.isInteger() && VT2.isSimple() &&
5867 VT2.isInteger() && VT1.getSizeInBits() <= 32);
5870 bool ARM64TargetLowering::hasPairedLoad(Type *LoadedType,
5871 unsigned &RequiredAligment) const {
5872 if (!LoadedType->isIntegerTy() && !LoadedType->isFloatTy())
5874 // Cyclone supports unaligned accesses.
5875 RequiredAligment = 0;
5876 unsigned NumBits = LoadedType->getPrimitiveSizeInBits();
5877 return NumBits == 32 || NumBits == 64;
5880 bool ARM64TargetLowering::hasPairedLoad(EVT LoadedType,
5881 unsigned &RequiredAligment) const {
5882 if (!LoadedType.isSimple() ||
5883 (!LoadedType.isInteger() && !LoadedType.isFloatingPoint()))
5885 // Cyclone supports unaligned accesses.
5886 RequiredAligment = 0;
5887 unsigned NumBits = LoadedType.getSizeInBits();
5888 return NumBits == 32 || NumBits == 64;
5891 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
5892 unsigned AlignCheck) {
5893 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
5894 (DstAlign == 0 || DstAlign % AlignCheck == 0));
5897 EVT ARM64TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
5898 unsigned SrcAlign, bool IsMemset,
5899 bool ZeroMemset, bool MemcpyStrSrc,
5900 MachineFunction &MF) const {
5901 // Don't use AdvSIMD to implement 16-byte memset. It would have taken one
5902 // instruction to materialize the v2i64 zero and one store (with restrictive
5903 // addressing mode). Just do two i64 store of zero-registers.
5905 const Function *F = MF.getFunction();
5906 if (Subtarget->hasFPARMv8() && !IsMemset && Size >= 16 &&
5907 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
5908 Attribute::NoImplicitFloat) &&
5909 (memOpAlign(SrcAlign, DstAlign, 16) ||
5910 (allowsUnalignedMemoryAccesses(MVT::f128, 0, &Fast) && Fast)))
5913 return Size >= 8 ? MVT::i64 : MVT::i32;
5916 // 12-bit optionally shifted immediates are legal for adds.
5917 bool ARM64TargetLowering::isLegalAddImmediate(int64_t Immed) const {
5918 if ((Immed >> 12) == 0 || ((Immed & 0xfff) == 0 && Immed >> 24 == 0))
5923 // Integer comparisons are implemented with ADDS/SUBS, so the range of valid
5924 // immediates is the same as for an add or a sub.
5925 bool ARM64TargetLowering::isLegalICmpImmediate(int64_t Immed) const {
5928 return isLegalAddImmediate(Immed);
5931 /// isLegalAddressingMode - Return true if the addressing mode represented
5932 /// by AM is legal for this target, for a load/store of the specified type.
5933 bool ARM64TargetLowering::isLegalAddressingMode(const AddrMode &AM,
5935 // ARM64 has five basic addressing modes:
5937 // reg + 9-bit signed offset
5938 // reg + SIZE_IN_BYTES * 12-bit unsigned offset
5940 // reg + SIZE_IN_BYTES * reg
5942 // No global is ever allowed as a base.
5946 // No reg+reg+imm addressing.
5947 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
5950 // check reg + imm case:
5951 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
5952 uint64_t NumBytes = 0;
5953 if (Ty->isSized()) {
5954 uint64_t NumBits = getDataLayout()->getTypeSizeInBits(Ty);
5955 NumBytes = NumBits / 8;
5956 if (!isPowerOf2_64(NumBits))
5961 int64_t Offset = AM.BaseOffs;
5963 // 9-bit signed offset
5964 if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
5967 // 12-bit unsigned offset
5968 unsigned shift = Log2_64(NumBytes);
5969 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
5970 // Must be a multiple of NumBytes (NumBytes is a power of 2)
5971 (Offset >> shift) << shift == Offset)
5976 // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2
5978 if (!AM.Scale || AM.Scale == 1 ||
5979 (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
5984 int ARM64TargetLowering::getScalingFactorCost(const AddrMode &AM,
5986 // Scaling factors are not free at all.
5987 // Operands | Rt Latency
5988 // -------------------------------------------
5990 // -------------------------------------------
5991 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
5992 // Rt, [Xn, Wm, <extend> #imm] |
5993 if (isLegalAddressingMode(AM, Ty))
5994 // Scale represents reg2 * scale, thus account for 1 if
5995 // it is not equal to 0 or 1.
5996 return AM.Scale != 0 && AM.Scale != 1;
6000 bool ARM64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
6001 VT = VT.getScalarType();
6006 switch (VT.getSimpleVT().SimpleTy) {
6018 ARM64TargetLowering::getScratchRegisters(CallingConv::ID) const {
6019 // LR is a callee-save register, but we must treat it as clobbered by any call
6020 // site. Hence we include LR in the scratch registers, which are in turn added
6021 // as implicit-defs for stackmaps and patchpoints.
6022 static const MCPhysReg ScratchRegs[] = {
6023 ARM64::X16, ARM64::X17, ARM64::LR, 0
6028 bool ARM64TargetLowering::isDesirableToCommuteWithShift(const SDNode *N) const {
6029 EVT VT = N->getValueType(0);
6030 // If N is unsigned bit extraction: ((x >> C) & mask), then do not combine
6031 // it with shift to let it be lowered to UBFX.
6032 if (N->getOpcode() == ISD::AND && (VT == MVT::i32 || VT == MVT::i64) &&
6033 isa<ConstantSDNode>(N->getOperand(1))) {
6034 uint64_t TruncMask = N->getConstantOperandVal(1);
6035 if (isMask_64(TruncMask) &&
6036 N->getOperand(0).getOpcode() == ISD::SRL &&
6037 isa<ConstantSDNode>(N->getOperand(0)->getOperand(1)))
6043 bool ARM64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
6045 assert(Ty->isIntegerTy());
6047 unsigned BitSize = Ty->getPrimitiveSizeInBits();
6051 int64_t Val = Imm.getSExtValue();
6052 if (Val == 0 || ARM64_AM::isLogicalImmediate(Val, BitSize))
6055 if ((int64_t)Val < 0)
6058 Val &= (1LL << 32) - 1;
6060 unsigned LZ = countLeadingZeros((uint64_t)Val);
6061 unsigned Shift = (63 - LZ) / 16;
6062 // MOVZ is free so return true for one or fewer MOVK.
6063 return (Shift < 3) ? true : false;
6066 // Generate SUBS and CSEL for integer abs.
6067 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
6068 EVT VT = N->getValueType(0);
6070 SDValue N0 = N->getOperand(0);
6071 SDValue N1 = N->getOperand(1);
6074 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
6075 // and change it to SUB and CSEL.
6076 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
6077 N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
6078 N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0))
6079 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
6080 if (Y1C->getAPIntValue() == VT.getSizeInBits() - 1) {
6081 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
6083 // Generate SUBS & CSEL.
6085 DAG.getNode(ARM64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32),
6086 N0.getOperand(0), DAG.getConstant(0, VT));
6087 return DAG.getNode(ARM64ISD::CSEL, DL, VT, N0.getOperand(0), Neg,
6088 DAG.getConstant(ARM64CC::PL, MVT::i32),
6089 SDValue(Cmp.getNode(), 1));
6094 // performXorCombine - Attempts to handle integer ABS.
6095 static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
6096 TargetLowering::DAGCombinerInfo &DCI,
6097 const ARM64Subtarget *Subtarget) {
6098 if (DCI.isBeforeLegalizeOps())
6101 return performIntegerAbsCombine(N, DAG);
6104 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG,
6105 TargetLowering::DAGCombinerInfo &DCI,
6106 const ARM64Subtarget *Subtarget) {
6107 if (DCI.isBeforeLegalizeOps())
6110 // Multiplication of a power of two plus/minus one can be done more
6111 // cheaply as as shift+add/sub. For now, this is true unilaterally. If
6112 // future CPUs have a cheaper MADD instruction, this may need to be
6113 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
6114 // 64-bit is 5 cycles, so this is always a win.
6115 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
6116 APInt Value = C->getAPIntValue();
6117 EVT VT = N->getValueType(0);
6118 APInt VP1 = Value + 1;
6119 if (VP1.isPowerOf2()) {
6120 // Multiplying by one less than a power of two, replace with a shift
6122 SDValue ShiftedVal =
6123 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6124 DAG.getConstant(VP1.logBase2(), MVT::i64));
6125 return DAG.getNode(ISD::SUB, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
6127 APInt VM1 = Value - 1;
6128 if (VM1.isPowerOf2()) {
6129 // Multiplying by one more than a power of two, replace with a shift
6131 SDValue ShiftedVal =
6132 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6133 DAG.getConstant(VM1.logBase2(), MVT::i64));
6134 return DAG.getNode(ISD::ADD, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
6140 static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG) {
6141 EVT VT = N->getValueType(0);
6142 if (VT != MVT::f32 && VT != MVT::f64)
6144 // Only optimize when the source and destination types have the same width.
6145 if (VT.getSizeInBits() != N->getOperand(0).getValueType().getSizeInBits())
6148 // If the result of an integer load is only used by an integer-to-float
6149 // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
6150 // This eliminates an "integer-to-vector-move UOP and improve throughput.
6151 SDValue N0 = N->getOperand(0);
6152 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6153 // Do not change the width of a volatile load.
6154 !cast<LoadSDNode>(N0)->isVolatile()) {
6155 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6156 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
6157 LN0->getPointerInfo(), LN0->isVolatile(),
6158 LN0->isNonTemporal(), LN0->isInvariant(),
6159 LN0->getAlignment());
6161 // Make sure successors of the original load stay after it by updating them
6162 // to use the new Chain.
6163 DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));
6166 (N->getOpcode() == ISD::SINT_TO_FP) ? ARM64ISD::SITOF : ARM64ISD::UITOF;
6167 return DAG.getNode(Opcode, SDLoc(N), VT, Load);
6173 /// An EXTR instruction is made up of two shifts, ORed together. This helper
6174 /// searches for and classifies those shifts.
6175 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
6177 if (N.getOpcode() == ISD::SHL)
6179 else if (N.getOpcode() == ISD::SRL)
6184 if (!isa<ConstantSDNode>(N.getOperand(1)))
6187 ShiftAmount = N->getConstantOperandVal(1);
6188 Src = N->getOperand(0);
6192 /// EXTR instruction extracts a contiguous chunk of bits from two existing
6193 /// registers viewed as a high/low pair. This function looks for the pattern:
6194 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
6195 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
6197 static SDValue tryCombineToEXTR(SDNode *N,
6198 TargetLowering::DAGCombinerInfo &DCI) {
6199 SelectionDAG &DAG = DCI.DAG;
6201 EVT VT = N->getValueType(0);
6203 assert(N->getOpcode() == ISD::OR && "Unexpected root");
6205 if (VT != MVT::i32 && VT != MVT::i64)
6209 uint32_t ShiftLHS = 0;
6211 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
6215 uint32_t ShiftRHS = 0;
6217 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
6220 // If they're both trying to come from the high part of the register, they're
6221 // not really an EXTR.
6222 if (LHSFromHi == RHSFromHi)
6225 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
6229 std::swap(LHS, RHS);
6230 std::swap(ShiftLHS, ShiftRHS);
6233 return DAG.getNode(ARM64ISD::EXTR, DL, VT, LHS, RHS,
6234 DAG.getConstant(ShiftRHS, MVT::i64));
6237 static SDValue tryCombineToBSL(SDNode *N,
6238 TargetLowering::DAGCombinerInfo &DCI) {
6239 EVT VT = N->getValueType(0);
6240 SelectionDAG &DAG = DCI.DAG;
6246 SDValue N0 = N->getOperand(0);
6247 if (N0.getOpcode() != ISD::AND)
6250 SDValue N1 = N->getOperand(1);
6251 if (N1.getOpcode() != ISD::AND)
6254 // We only have to look for constant vectors here since the general, variable
6255 // case can be handled in TableGen.
6256 unsigned Bits = VT.getVectorElementType().getSizeInBits();
6257 uint64_t BitMask = Bits == 64 ? -1ULL : ((1ULL << Bits) - 1);
6258 for (int i = 1; i >= 0; --i)
6259 for (int j = 1; j >= 0; --j) {
6260 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(i));
6261 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(j));
6265 bool FoundMatch = true;
6266 for (unsigned k = 0; k < VT.getVectorNumElements(); ++k) {
6267 ConstantSDNode *CN0 = dyn_cast<ConstantSDNode>(BVN0->getOperand(k));
6268 ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(BVN1->getOperand(k));
6270 CN0->getZExtValue() != (BitMask & ~CN1->getZExtValue())) {
6277 return DAG.getNode(ARM64ISD::BSL, DL, VT, SDValue(BVN0, 0),
6278 N0->getOperand(1 - i), N1->getOperand(1 - j));
6284 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
6285 const ARM64Subtarget *Subtarget) {
6286 // Attempt to form an EXTR from (or (shl VAL1, #N), (srl VAL2, #RegWidth-N))
6287 if (!EnableARM64ExtrGeneration)
6289 SelectionDAG &DAG = DCI.DAG;
6290 EVT VT = N->getValueType(0);
6292 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6295 SDValue Res = tryCombineToEXTR(N, DCI);
6299 Res = tryCombineToBSL(N, DCI);
6306 static SDValue performBitcastCombine(SDNode *N,
6307 TargetLowering::DAGCombinerInfo &DCI,
6308 SelectionDAG &DAG) {
6309 // Wait 'til after everything is legalized to try this. That way we have
6310 // legal vector types and such.
6311 if (DCI.isBeforeLegalizeOps())
6314 // Remove extraneous bitcasts around an extract_subvector.
6316 // (v4i16 (bitconvert
6317 // (extract_subvector (v2i64 (bitconvert (v8i16 ...)), (i64 1)))))
6319 // (extract_subvector ((v8i16 ...), (i64 4)))
6321 // Only interested in 64-bit vectors as the ultimate result.
6322 EVT VT = N->getValueType(0);
6325 if (VT.getSimpleVT().getSizeInBits() != 64)
6327 // Is the operand an extract_subvector starting at the beginning or halfway
6328 // point of the vector? A low half may also come through as an
6329 // EXTRACT_SUBREG, so look for that, too.
6330 SDValue Op0 = N->getOperand(0);
6331 if (Op0->getOpcode() != ISD::EXTRACT_SUBVECTOR &&
6332 !(Op0->isMachineOpcode() &&
6333 Op0->getMachineOpcode() == ARM64::EXTRACT_SUBREG))
6335 uint64_t idx = cast<ConstantSDNode>(Op0->getOperand(1))->getZExtValue();
6336 if (Op0->getOpcode() == ISD::EXTRACT_SUBVECTOR) {
6337 if (Op0->getValueType(0).getVectorNumElements() != idx && idx != 0)
6339 } else if (Op0->getMachineOpcode() == ARM64::EXTRACT_SUBREG) {
6340 if (idx != ARM64::dsub)
6342 // The dsub reference is equivalent to a lane zero subvector reference.
6345 // Look through the bitcast of the input to the extract.
6346 if (Op0->getOperand(0)->getOpcode() != ISD::BITCAST)
6348 SDValue Source = Op0->getOperand(0)->getOperand(0);
6349 // If the source type has twice the number of elements as our destination
6350 // type, we know this is an extract of the high or low half of the vector.
6351 EVT SVT = Source->getValueType(0);
6352 if (SVT.getVectorNumElements() != VT.getVectorNumElements() * 2)
6355 DEBUG(dbgs() << "arm64-lower: bitcast extract_subvector simplification\n");
6357 // Create the simplified form to just extract the low or high half of the
6358 // vector directly rather than bothering with the bitcasts.
6360 unsigned NumElements = VT.getVectorNumElements();
6362 SDValue HalfIdx = DAG.getConstant(NumElements, MVT::i64);
6363 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Source, HalfIdx);
6365 SDValue SubReg = DAG.getTargetConstant(ARM64::dsub, MVT::i32);
6366 return SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl, VT,
6372 static SDValue performConcatVectorsCombine(SDNode *N,
6373 TargetLowering::DAGCombinerInfo &DCI,
6374 SelectionDAG &DAG) {
6375 // Wait 'til after everything is legalized to try this. That way we have
6376 // legal vector types and such.
6377 if (DCI.isBeforeLegalizeOps())
6381 EVT VT = N->getValueType(0);
6383 // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector
6384 // splat. The indexed instructions are going to be expecting a DUPLANE64, so
6385 // canonicalise to that.
6386 if (N->getOperand(0) == N->getOperand(1) && VT.getVectorNumElements() == 2) {
6387 assert(VT.getVectorElementType().getSizeInBits() == 64);
6388 return DAG.getNode(ARM64ISD::DUPLANE64, dl, VT,
6389 WidenVector(N->getOperand(0), DAG),
6390 DAG.getConstant(0, MVT::i64));
6393 // Canonicalise concat_vectors so that the right-hand vector has as few
6394 // bit-casts as possible before its real operation. The primary matching
6395 // destination for these operations will be the narrowing "2" instructions,
6396 // which depend on the operation being performed on this right-hand vector.
6398 // (concat_vectors LHS, (v1i64 (bitconvert (v4i16 RHS))))
6400 // (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS))
6402 SDValue Op1 = N->getOperand(1);
6403 if (Op1->getOpcode() != ISD::BITCAST)
6405 SDValue RHS = Op1->getOperand(0);
6406 MVT RHSTy = RHS.getValueType().getSimpleVT();
6407 // If the RHS is not a vector, this is not the pattern we're looking for.
6408 if (!RHSTy.isVector())
6411 DEBUG(dbgs() << "arm64-lower: concat_vectors bitcast simplification\n");
6413 MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(),
6414 RHSTy.getVectorNumElements() * 2);
6416 ISD::BITCAST, dl, VT,
6417 DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatTy,
6418 DAG.getNode(ISD::BITCAST, dl, RHSTy, N->getOperand(0)), RHS));
6421 static SDValue tryCombineFixedPointConvert(SDNode *N,
6422 TargetLowering::DAGCombinerInfo &DCI,
6423 SelectionDAG &DAG) {
6424 // Wait 'til after everything is legalized to try this. That way we have
6425 // legal vector types and such.
6426 if (DCI.isBeforeLegalizeOps())
6428 // Transform a scalar conversion of a value from a lane extract into a
6429 // lane extract of a vector conversion. E.g., from foo1 to foo2:
6430 // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); }
6431 // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; }
6433 // The second form interacts better with instruction selection and the
6434 // register allocator to avoid cross-class register copies that aren't
6435 // coalescable due to a lane reference.
6437 // Check the operand and see if it originates from a lane extract.
6438 SDValue Op1 = N->getOperand(1);
6439 if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6440 // Yep, no additional predication needed. Perform the transform.
6441 SDValue IID = N->getOperand(0);
6442 SDValue Shift = N->getOperand(2);
6443 SDValue Vec = Op1.getOperand(0);
6444 SDValue Lane = Op1.getOperand(1);
6445 EVT ResTy = N->getValueType(0);
6449 // The vector width should be 128 bits by the time we get here, even
6450 // if it started as 64 bits (the extract_vector handling will have
6452 assert(Vec.getValueType().getSizeInBits() == 128 &&
6453 "unexpected vector size on extract_vector_elt!");
6454 if (Vec.getValueType() == MVT::v4i32)
6455 VecResTy = MVT::v4f32;
6456 else if (Vec.getValueType() == MVT::v2i64)
6457 VecResTy = MVT::v2f64;
6459 assert(0 && "unexpected vector type!");
6462 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift);
6463 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane);
6468 // AArch64 high-vector "long" operations are formed by performing the non-high
6469 // version on an extract_subvector of each operand which gets the high half:
6471 // (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS))
6473 // However, there are cases which don't have an extract_high explicitly, but
6474 // have another operation that can be made compatible with one for free. For
6477 // (dupv64 scalar) --> (extract_high (dup128 scalar))
6479 // This routine does the actual conversion of such DUPs, once outer routines
6480 // have determined that everything else is in order.
6481 static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) {
6482 // We can handle most types of duplicate, but the lane ones have an extra
6483 // operand saying *which* lane, so we need to know.
6485 switch (N.getOpcode()) {
6489 case ARM64ISD::DUPLANE8:
6490 case ARM64ISD::DUPLANE16:
6491 case ARM64ISD::DUPLANE32:
6492 case ARM64ISD::DUPLANE64:
6499 MVT NarrowTy = N.getSimpleValueType();
6500 if (!NarrowTy.is64BitVector())
6503 MVT ElementTy = NarrowTy.getVectorElementType();
6504 unsigned NumElems = NarrowTy.getVectorNumElements();
6505 MVT NewDUPVT = MVT::getVectorVT(ElementTy, NumElems * 2);
6509 NewDUP = DAG.getNode(N.getOpcode(), SDLoc(N), NewDUPVT, N.getOperand(0),
6512 NewDUP = DAG.getNode(ARM64ISD::DUP, SDLoc(N), NewDUPVT, N.getOperand(0));
6514 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N.getNode()), NarrowTy,
6515 NewDUP, DAG.getConstant(NumElems, MVT::i64));
6518 static bool isEssentiallyExtractSubvector(SDValue N) {
6519 if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR)
6522 return N.getOpcode() == ISD::BITCAST &&
6523 N.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR;
6526 /// \brief Helper structure to keep track of ISD::SET_CC operands.
6527 struct GenericSetCCInfo {
6528 const SDValue *Opnd0;
6529 const SDValue *Opnd1;
6533 /// \brief Helper structure to keep track of a SET_CC lowered into ARM64 code.
6534 struct ARM64SetCCInfo {
6536 ARM64CC::CondCode CC;
6539 /// \brief Helper structure to keep track of SetCC information.
6541 GenericSetCCInfo Generic;
6542 ARM64SetCCInfo ARM64;
6545 /// \brief Helper structure to be able to read SetCC information.
6546 /// If set to true, IsARM64 field, Info is a ARM64SetCCInfo, otherwise Info is
6547 /// a GenericSetCCInfo.
6548 struct SetCCInfoAndKind {
6553 /// \brief Check whether or not \p Op is a SET_CC operation, either a generic or
6555 /// ARM64 lowered one.
6556 /// \p SetCCInfo is filled accordingly.
6557 /// \post SetCCInfo is meanginfull only when this function returns true.
6558 /// \return True when Op is a kind of SET_CC operation.
6559 static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) {
6560 // If this is a setcc, this is straight forward.
6561 if (Op.getOpcode() == ISD::SETCC) {
6562 SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0);
6563 SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1);
6564 SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6565 SetCCInfo.IsARM64 = false;
6568 // Otherwise, check if this is a matching csel instruction.
6572 if (Op.getOpcode() != ARM64ISD::CSEL)
6574 // Set the information about the operands.
6575 // TODO: we want the operands of the Cmp not the csel
6576 SetCCInfo.Info.ARM64.Cmp = &Op.getOperand(3);
6577 SetCCInfo.IsARM64 = true;
6578 SetCCInfo.Info.ARM64.CC = static_cast<ARM64CC::CondCode>(
6579 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
6581 // Check that the operands matches the constraints:
6582 // (1) Both operands must be constants.
6583 // (2) One must be 1 and the other must be 0.
6584 ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0));
6585 ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6588 if (!TValue || !FValue)
6592 if (!TValue->isOne()) {
6593 // Update the comparison when we are interested in !cc.
6594 std::swap(TValue, FValue);
6595 SetCCInfo.Info.ARM64.CC =
6596 ARM64CC::getInvertedCondCode(SetCCInfo.Info.ARM64.CC);
6598 return TValue->isOne() && FValue->isNullValue();
6601 // The folding we want to perform is:
6602 // (add x, (setcc cc ...) )
6604 // (csel x, (add x, 1), !cc ...)
6606 // The latter will get matched to a CSINC instruction.
6607 static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) {
6608 assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!");
6609 SDValue LHS = Op->getOperand(0);
6610 SDValue RHS = Op->getOperand(1);
6611 SetCCInfoAndKind InfoAndKind;
6613 // If neither operand is a SET_CC, give up.
6614 if (!isSetCC(LHS, InfoAndKind)) {
6615 std::swap(LHS, RHS);
6616 if (!isSetCC(LHS, InfoAndKind))
6620 // FIXME: This could be generatized to work for FP comparisons.
6621 EVT CmpVT = InfoAndKind.IsARM64
6622 ? InfoAndKind.Info.ARM64.Cmp->getOperand(0).getValueType()
6623 : InfoAndKind.Info.Generic.Opnd0->getValueType();
6624 if (CmpVT != MVT::i32 && CmpVT != MVT::i64)
6630 if (InfoAndKind.IsARM64) {
6631 CCVal = DAG.getConstant(
6632 ARM64CC::getInvertedCondCode(InfoAndKind.Info.ARM64.CC), MVT::i32);
6633 Cmp = *InfoAndKind.Info.ARM64.Cmp;
6635 Cmp = getARM64Cmp(*InfoAndKind.Info.Generic.Opnd0,
6636 *InfoAndKind.Info.Generic.Opnd1,
6637 ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, true),
6640 EVT VT = Op->getValueType(0);
6641 LHS = DAG.getNode(ISD::ADD, dl, VT, RHS, DAG.getConstant(1, VT));
6642 return DAG.getNode(ARM64ISD::CSEL, dl, VT, RHS, LHS, CCVal, Cmp);
6645 // The basic add/sub long vector instructions have variants with "2" on the end
6646 // which act on the high-half of their inputs. They are normally matched by
6649 // (add (zeroext (extract_high LHS)),
6650 // (zeroext (extract_high RHS)))
6651 // -> uaddl2 vD, vN, vM
6653 // However, if one of the extracts is something like a duplicate, this
6654 // instruction can still be used profitably. This function puts the DAG into a
6655 // more appropriate form for those patterns to trigger.
6656 static SDValue performAddSubLongCombine(SDNode *N,
6657 TargetLowering::DAGCombinerInfo &DCI,
6658 SelectionDAG &DAG) {
6659 if (DCI.isBeforeLegalizeOps())
6662 MVT VT = N->getSimpleValueType(0);
6663 if (!VT.is128BitVector()) {
6664 if (N->getOpcode() == ISD::ADD)
6665 return performSetccAddFolding(N, DAG);
6669 // Make sure both branches are extended in the same way.
6670 SDValue LHS = N->getOperand(0);
6671 SDValue RHS = N->getOperand(1);
6672 if ((LHS.getOpcode() != ISD::ZERO_EXTEND &&
6673 LHS.getOpcode() != ISD::SIGN_EXTEND) ||
6674 LHS.getOpcode() != RHS.getOpcode())
6677 unsigned ExtType = LHS.getOpcode();
6679 // It's not worth doing if at least one of the inputs isn't already an
6680 // extract, but we don't know which it'll be so we have to try both.
6681 if (isEssentiallyExtractSubvector(LHS.getOperand(0))) {
6682 RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG);
6686 RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS);
6687 } else if (isEssentiallyExtractSubvector(RHS.getOperand(0))) {
6688 LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG);
6692 LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS);
6695 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS);
6698 // Massage DAGs which we can use the high-half "long" operations on into
6699 // something isel will recognize better. E.g.
6701 // (arm64_neon_umull (extract_high vec) (dupv64 scalar)) -->
6702 // (arm64_neon_umull (extract_high (v2i64 vec)))
6703 // (extract_high (v2i64 (dup128 scalar)))))
6705 static SDValue tryCombineLongOpWithDup(unsigned IID, SDNode *N,
6706 TargetLowering::DAGCombinerInfo &DCI,
6707 SelectionDAG &DAG) {
6708 if (DCI.isBeforeLegalizeOps())
6711 SDValue LHS = N->getOperand(1);
6712 SDValue RHS = N->getOperand(2);
6713 assert(LHS.getValueType().is64BitVector() &&
6714 RHS.getValueType().is64BitVector() &&
6715 "unexpected shape for long operation");
6717 // Either node could be a DUP, but it's not worth doing both of them (you'd
6718 // just as well use the non-high version) so look for a corresponding extract
6719 // operation on the other "wing".
6720 if (isEssentiallyExtractSubvector(LHS)) {
6721 RHS = tryExtendDUPToExtractHigh(RHS, DAG);
6724 } else if (isEssentiallyExtractSubvector(RHS)) {
6725 LHS = tryExtendDUPToExtractHigh(LHS, DAG);
6730 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0),
6731 N->getOperand(0), LHS, RHS);
6734 static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
6735 MVT ElemTy = N->getSimpleValueType(0).getScalarType();
6736 unsigned ElemBits = ElemTy.getSizeInBits();
6738 int64_t ShiftAmount;
6739 if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) {
6740 APInt SplatValue, SplatUndef;
6741 unsigned SplatBitSize;
6743 if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
6744 HasAnyUndefs, ElemBits) ||
6745 SplatBitSize != ElemBits)
6748 ShiftAmount = SplatValue.getSExtValue();
6749 } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) {
6750 ShiftAmount = CVN->getSExtValue();
6758 llvm_unreachable("Unknown shift intrinsic");
6759 case Intrinsic::arm64_neon_sqshl:
6760 Opcode = ARM64ISD::SQSHL_I;
6761 IsRightShift = false;
6763 case Intrinsic::arm64_neon_uqshl:
6764 Opcode = ARM64ISD::UQSHL_I;
6765 IsRightShift = false;
6767 case Intrinsic::arm64_neon_srshl:
6768 Opcode = ARM64ISD::SRSHR_I;
6769 IsRightShift = true;
6771 case Intrinsic::arm64_neon_urshl:
6772 Opcode = ARM64ISD::URSHR_I;
6773 IsRightShift = true;
6775 case Intrinsic::arm64_neon_sqshlu:
6776 Opcode = ARM64ISD::SQSHLU_I;
6777 IsRightShift = false;
6781 if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits)
6782 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
6783 DAG.getConstant(-ShiftAmount, MVT::i32));
6784 else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount <= ElemBits)
6785 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
6786 DAG.getConstant(ShiftAmount, MVT::i32));
6791 // The CRC32[BH] instructions ignore the high bits of their data operand. Since
6792 // the intrinsics must be legal and take an i32, this means there's almost
6793 // certainly going to be a zext in the DAG which we can eliminate.
6794 static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) {
6795 SDValue AndN = N->getOperand(2);
6796 if (AndN.getOpcode() != ISD::AND)
6799 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1));
6800 if (!CMask || CMask->getZExtValue() != Mask)
6803 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32,
6804 N->getOperand(0), N->getOperand(1), AndN.getOperand(0));
6807 static SDValue performIntrinsicCombine(SDNode *N,
6808 TargetLowering::DAGCombinerInfo &DCI,
6809 const ARM64Subtarget *Subtarget) {
6810 SelectionDAG &DAG = DCI.DAG;
6811 unsigned IID = getIntrinsicID(N);
6815 case Intrinsic::arm64_neon_vcvtfxs2fp:
6816 case Intrinsic::arm64_neon_vcvtfxu2fp:
6817 return tryCombineFixedPointConvert(N, DCI, DAG);
6819 case Intrinsic::arm64_neon_fmax:
6820 return DAG.getNode(ARM64ISD::FMAX, SDLoc(N), N->getValueType(0),
6821 N->getOperand(1), N->getOperand(2));
6822 case Intrinsic::arm64_neon_fmin:
6823 return DAG.getNode(ARM64ISD::FMIN, SDLoc(N), N->getValueType(0),
6824 N->getOperand(1), N->getOperand(2));
6825 case Intrinsic::arm64_neon_smull:
6826 case Intrinsic::arm64_neon_umull:
6827 case Intrinsic::arm64_neon_pmull:
6828 case Intrinsic::arm64_neon_sqdmull:
6829 return tryCombineLongOpWithDup(IID, N, DCI, DAG);
6830 case Intrinsic::arm64_neon_sqshl:
6831 case Intrinsic::arm64_neon_uqshl:
6832 case Intrinsic::arm64_neon_sqshlu:
6833 case Intrinsic::arm64_neon_srshl:
6834 case Intrinsic::arm64_neon_urshl:
6835 return tryCombineShiftImm(IID, N, DAG);
6836 case Intrinsic::arm64_crc32b:
6837 case Intrinsic::arm64_crc32cb:
6838 return tryCombineCRC32(0xff, N, DAG);
6839 case Intrinsic::arm64_crc32h:
6840 case Intrinsic::arm64_crc32ch:
6841 return tryCombineCRC32(0xffff, N, DAG);
6846 static SDValue performExtendCombine(SDNode *N,
6847 TargetLowering::DAGCombinerInfo &DCI,
6848 SelectionDAG &DAG) {
6849 // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then
6850 // we can convert that DUP into another extract_high (of a bigger DUP), which
6851 // helps the backend to decide that an sabdl2 would be useful, saving a real
6852 // extract_high operation.
6853 if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND &&
6854 N->getOperand(0).getOpcode() == ISD::INTRINSIC_WO_CHAIN) {
6855 SDNode *ABDNode = N->getOperand(0).getNode();
6856 unsigned IID = getIntrinsicID(ABDNode);
6857 if (IID == Intrinsic::arm64_neon_sabd ||
6858 IID == Intrinsic::arm64_neon_uabd) {
6859 SDValue NewABD = tryCombineLongOpWithDup(IID, ABDNode, DCI, DAG);
6860 if (!NewABD.getNode())
6863 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0),
6868 // This is effectively a custom type legalization for ARM64.
6870 // Type legalization will split an extend of a small, legal, type to a larger
6871 // illegal type by first splitting the destination type, often creating
6872 // illegal source types, which then get legalized in isel-confusing ways,
6873 // leading to really terrible codegen. E.g.,
6874 // %result = v8i32 sext v8i8 %value
6876 // %losrc = extract_subreg %value, ...
6877 // %hisrc = extract_subreg %value, ...
6878 // %lo = v4i32 sext v4i8 %losrc
6879 // %hi = v4i32 sext v4i8 %hisrc
6880 // Things go rapidly downhill from there.
6882 // For ARM64, the [sz]ext vector instructions can only go up one element
6883 // size, so we can, e.g., extend from i8 to i16, but to go from i8 to i32
6884 // take two instructions.
6886 // This implies that the most efficient way to do the extend from v8i8
6887 // to two v4i32 values is to first extend the v8i8 to v8i16, then do
6888 // the normal splitting to happen for the v8i16->v8i32.
6890 // This is pre-legalization to catch some cases where the default
6891 // type legalization will create ill-tempered code.
6892 if (!DCI.isBeforeLegalizeOps())
6895 // We're only interested in cleaning things up for non-legal vector types
6896 // here. If both the source and destination are legal, things will just
6897 // work naturally without any fiddling.
6898 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6899 EVT ResVT = N->getValueType(0);
6900 if (!ResVT.isVector() || TLI.isTypeLegal(ResVT))
6902 // If the vector type isn't a simple VT, it's beyond the scope of what
6903 // we're worried about here. Let legalization do its thing and hope for
6905 if (!ResVT.isSimple())
6908 SDValue Src = N->getOperand(0);
6909 MVT SrcVT = Src->getValueType(0).getSimpleVT();
6910 // If the source VT is a 64-bit vector, we can play games and get the
6911 // better results we want.
6912 if (SrcVT.getSizeInBits() != 64)
6915 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
6916 unsigned ElementCount = SrcVT.getVectorNumElements();
6917 SrcVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize * 2), ElementCount);
6919 Src = DAG.getNode(N->getOpcode(), DL, SrcVT, Src);
6921 // Now split the rest of the operation into two halves, each with a 64
6925 unsigned NumElements = ResVT.getVectorNumElements();
6926 assert(!(NumElements & 1) && "Splitting vector, but not in half!");
6927 LoVT = HiVT = EVT::getVectorVT(*DAG.getContext(),
6928 ResVT.getVectorElementType(), NumElements / 2);
6930 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
6931 LoVT.getVectorNumElements());
6932 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
6933 DAG.getIntPtrConstant(0));
6934 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
6935 DAG.getIntPtrConstant(InNVT.getVectorNumElements()));
6936 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, Lo);
6937 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, Hi);
6939 // Now combine the parts back together so we still have a single result
6940 // like the combiner expects.
6941 return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
6944 /// Replace a splat of a scalar to a vector store by scalar stores of the scalar
6945 /// value. The load store optimizer pass will merge them to store pair stores.
6946 /// This has better performance than a splat of the scalar followed by a split
6947 /// vector store. Even if the stores are not merged it is four stores vs a dup,
6948 /// followed by an ext.b and two stores.
6949 static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) {
6950 SDValue StVal = St->getValue();
6951 EVT VT = StVal.getValueType();
6953 // Don't replace floating point stores, they possibly won't be transformed to
6954 // stp because of the store pair suppress pass.
6955 if (VT.isFloatingPoint())
6958 // Check for insert vector elements.
6959 if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT)
6962 // We can express a splat as store pair(s) for 2 or 4 elements.
6963 unsigned NumVecElts = VT.getVectorNumElements();
6964 if (NumVecElts != 4 && NumVecElts != 2)
6966 SDValue SplatVal = StVal.getOperand(1);
6967 unsigned RemainInsertElts = NumVecElts - 1;
6969 // Check that this is a splat.
6970 while (--RemainInsertElts) {
6971 SDValue NextInsertElt = StVal.getOperand(0);
6972 if (NextInsertElt.getOpcode() != ISD::INSERT_VECTOR_ELT)
6974 if (NextInsertElt.getOperand(1) != SplatVal)
6976 StVal = NextInsertElt;
6978 unsigned OrigAlignment = St->getAlignment();
6979 unsigned EltOffset = NumVecElts == 4 ? 4 : 8;
6980 unsigned Alignment = std::min(OrigAlignment, EltOffset);
6982 // Create scalar stores. This is at least as good as the code sequence for a
6983 // split unaligned store wich is a dup.s, ext.b, and two stores.
6984 // Most of the time the three stores should be replaced by store pair
6985 // instructions (stp).
6987 SDValue BasePtr = St->getBasePtr();
6989 DAG.getStore(St->getChain(), DL, SplatVal, BasePtr, St->getPointerInfo(),
6990 St->isVolatile(), St->isNonTemporal(), St->getAlignment());
6992 unsigned Offset = EltOffset;
6993 while (--NumVecElts) {
6994 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
6995 DAG.getConstant(Offset, MVT::i64));
6996 NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr,
6997 St->getPointerInfo(), St->isVolatile(),
6998 St->isNonTemporal(), Alignment);
6999 Offset += EltOffset;
7004 static SDValue performSTORECombine(SDNode *N,
7005 TargetLowering::DAGCombinerInfo &DCI,
7007 const ARM64Subtarget *Subtarget) {
7008 if (!DCI.isBeforeLegalize())
7011 StoreSDNode *S = cast<StoreSDNode>(N);
7012 if (S->isVolatile())
7015 // Cyclone has bad performance on unaligned 16B stores when crossing line and
7016 // page boundries. We want to split such stores.
7017 if (!Subtarget->isCyclone())
7020 // Don't split at Oz.
7021 MachineFunction &MF = DAG.getMachineFunction();
7022 bool IsMinSize = MF.getFunction()->getAttributes().hasAttribute(
7023 AttributeSet::FunctionIndex, Attribute::MinSize);
7027 SDValue StVal = S->getValue();
7028 EVT VT = StVal.getValueType();
7030 // Don't split v2i64 vectors. Memcpy lowering produces those and splitting
7031 // those up regresses performance on micro-benchmarks and olden/bh.
7032 if (!VT.isVector() || VT.getVectorNumElements() < 2 || VT == MVT::v2i64)
7035 // Split unaligned 16B stores. They are terrible for performance.
7036 // Don't split stores with alignment of 1 or 2. Code that uses clang vector
7037 // extensions can use this to mark that it does not want splitting to happen
7038 // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of
7039 // eliminating alignment hazards is only 1 in 8 for alignment of 2.
7040 if (VT.getSizeInBits() != 128 || S->getAlignment() >= 16 ||
7041 S->getAlignment() <= 2)
7044 // If we get a splat of a scalar convert this vector store to a store of
7045 // scalars. They will be merged into store pairs thereby removing two
7047 SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S);
7048 if (ReplacedSplat != SDValue())
7049 return ReplacedSplat;
7052 unsigned NumElts = VT.getVectorNumElements() / 2;
7053 // Split VT into two.
7055 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), NumElts);
7056 SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7057 DAG.getIntPtrConstant(0));
7058 SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7059 DAG.getIntPtrConstant(NumElts));
7060 SDValue BasePtr = S->getBasePtr();
7062 DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
7063 S->isVolatile(), S->isNonTemporal(), S->getAlignment());
7064 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7065 DAG.getConstant(8, MVT::i64));
7066 return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr,
7067 S->getPointerInfo(), S->isVolatile(), S->isNonTemporal(),
7071 /// Target-specific DAG combine function for NEON load/store intrinsics
7072 /// to merge base address updates.
7073 static SDValue performNEONPostLDSTCombine(SDNode *N,
7074 TargetLowering::DAGCombinerInfo &DCI,
7075 SelectionDAG &DAG) {
7076 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7079 unsigned AddrOpIdx = N->getNumOperands() - 1;
7080 SDValue Addr = N->getOperand(AddrOpIdx);
7082 // Search for a use of the address operand that is an increment.
7083 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
7084 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
7086 if (User->getOpcode() != ISD::ADD ||
7087 UI.getUse().getResNo() != Addr.getResNo())
7090 // Check that the add is independent of the load/store. Otherwise, folding
7091 // it would create a cycle.
7092 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
7095 // Find the new opcode for the updating load/store.
7096 bool IsStore = false;
7097 bool IsLaneOp = false;
7098 bool IsDupOp = false;
7099 unsigned NewOpc = 0;
7100 unsigned NumVecs = 0;
7101 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
7103 default: llvm_unreachable("unexpected intrinsic for Neon base update");
7104 case Intrinsic::arm64_neon_ld2: NewOpc = ARM64ISD::LD2post;
7106 case Intrinsic::arm64_neon_ld3: NewOpc = ARM64ISD::LD3post;
7108 case Intrinsic::arm64_neon_ld4: NewOpc = ARM64ISD::LD4post;
7110 case Intrinsic::arm64_neon_st2: NewOpc = ARM64ISD::ST2post;
7111 NumVecs = 2; IsStore = true; break;
7112 case Intrinsic::arm64_neon_st3: NewOpc = ARM64ISD::ST3post;
7113 NumVecs = 3; IsStore = true; break;
7114 case Intrinsic::arm64_neon_st4: NewOpc = ARM64ISD::ST4post;
7115 NumVecs = 4; IsStore = true; break;
7116 case Intrinsic::arm64_neon_ld1x2: NewOpc = ARM64ISD::LD1x2post;
7118 case Intrinsic::arm64_neon_ld1x3: NewOpc = ARM64ISD::LD1x3post;
7120 case Intrinsic::arm64_neon_ld1x4: NewOpc = ARM64ISD::LD1x4post;
7122 case Intrinsic::arm64_neon_st1x2: NewOpc = ARM64ISD::ST1x2post;
7123 NumVecs = 2; IsStore = true; break;
7124 case Intrinsic::arm64_neon_st1x3: NewOpc = ARM64ISD::ST1x3post;
7125 NumVecs = 3; IsStore = true; break;
7126 case Intrinsic::arm64_neon_st1x4: NewOpc = ARM64ISD::ST1x4post;
7127 NumVecs = 4; IsStore = true; break;
7128 case Intrinsic::arm64_neon_ld2r: NewOpc = ARM64ISD::LD2DUPpost;
7129 NumVecs = 2; IsDupOp = true; break;
7130 case Intrinsic::arm64_neon_ld3r: NewOpc = ARM64ISD::LD3DUPpost;
7131 NumVecs = 3; IsDupOp = true; break;
7132 case Intrinsic::arm64_neon_ld4r: NewOpc = ARM64ISD::LD4DUPpost;
7133 NumVecs = 4; IsDupOp = true; break;
7134 case Intrinsic::arm64_neon_ld2lane: NewOpc = ARM64ISD::LD2LANEpost;
7135 NumVecs = 2; IsLaneOp = true; break;
7136 case Intrinsic::arm64_neon_ld3lane: NewOpc = ARM64ISD::LD3LANEpost;
7137 NumVecs = 3; IsLaneOp = true; break;
7138 case Intrinsic::arm64_neon_ld4lane: NewOpc = ARM64ISD::LD4LANEpost;
7139 NumVecs = 4; IsLaneOp = true; break;
7140 case Intrinsic::arm64_neon_st2lane: NewOpc = ARM64ISD::ST2LANEpost;
7141 NumVecs = 2; IsStore = true; IsLaneOp = true; break;
7142 case Intrinsic::arm64_neon_st3lane: NewOpc = ARM64ISD::ST3LANEpost;
7143 NumVecs = 3; IsStore = true; IsLaneOp = true; break;
7144 case Intrinsic::arm64_neon_st4lane: NewOpc = ARM64ISD::ST4LANEpost;
7145 NumVecs = 4; IsStore = true; IsLaneOp = true; break;
7150 VecTy = N->getOperand(2).getValueType();
7152 VecTy = N->getValueType(0);
7154 // If the increment is a constant, it must match the memory ref size.
7155 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
7156 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
7157 uint32_t IncVal = CInc->getZExtValue();
7158 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
7159 if (IsLaneOp || IsDupOp)
7160 NumBytes /= VecTy.getVectorNumElements();
7161 if (IncVal != NumBytes)
7163 Inc = DAG.getRegister(ARM64::XZR, MVT::i64);
7165 SmallVector<SDValue, 8> Ops;
7166 Ops.push_back(N->getOperand(0)); // Incoming chain
7167 // Load lane and store have vector list as input.
7168 if (IsLaneOp || IsStore)
7169 for (unsigned i = 2; i < AddrOpIdx; ++i)
7170 Ops.push_back(N->getOperand(i));
7171 Ops.push_back(N->getOperand(AddrOpIdx)); // Base register
7176 unsigned NumResultVecs = (IsStore ? 0 : NumVecs);
7178 for (n = 0; n < NumResultVecs; ++n)
7180 Tys[n++] = MVT::i64; // Type of write back register
7181 Tys[n] = MVT::Other; // Type of the chain
7182 SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, NumResultVecs + 2));
7184 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
7185 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys, Ops,
7186 MemInt->getMemoryVT(),
7187 MemInt->getMemOperand());
7190 std::vector<SDValue> NewResults;
7191 for (unsigned i = 0; i < NumResultVecs; ++i) {
7192 NewResults.push_back(SDValue(UpdN.getNode(), i));
7194 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs + 1));
7195 DCI.CombineTo(N, NewResults);
7196 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
7203 // Optimize compare with zero and branch.
7204 static SDValue performBRCONDCombine(SDNode *N,
7205 TargetLowering::DAGCombinerInfo &DCI,
7206 SelectionDAG &DAG) {
7207 SDValue Chain = N->getOperand(0);
7208 SDValue Dest = N->getOperand(1);
7209 SDValue CCVal = N->getOperand(2);
7210 SDValue Cmp = N->getOperand(3);
7212 assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!");
7213 unsigned CC = cast<ConstantSDNode>(CCVal)->getZExtValue();
7214 if (CC != ARM64CC::EQ && CC != ARM64CC::NE)
7217 unsigned CmpOpc = Cmp.getOpcode();
7218 if (CmpOpc != ARM64ISD::ADDS && CmpOpc != ARM64ISD::SUBS)
7221 // Only attempt folding if there is only one use of the flag and no use of the
7223 if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1))
7226 SDValue LHS = Cmp.getOperand(0);
7227 SDValue RHS = Cmp.getOperand(1);
7229 assert(LHS.getValueType() == RHS.getValueType() &&
7230 "Expected the value type to be the same for both operands!");
7231 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
7234 if (isa<ConstantSDNode>(LHS) && cast<ConstantSDNode>(LHS)->isNullValue())
7235 std::swap(LHS, RHS);
7237 if (!isa<ConstantSDNode>(RHS) || !cast<ConstantSDNode>(RHS)->isNullValue())
7240 if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA ||
7241 LHS.getOpcode() == ISD::SRL)
7244 // Fold the compare into the branch instruction.
7246 if (CC == ARM64CC::EQ)
7247 BR = DAG.getNode(ARM64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
7249 BR = DAG.getNode(ARM64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
7251 // Do not add new nodes to DAG combiner worklist.
7252 DCI.CombineTo(N, BR, false);
7257 // vselect (v1i1 setcc) ->
7258 // vselect (v1iXX setcc) (XX is the size of the compared operand type)
7259 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
7260 // condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
7262 static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) {
7263 SDValue N0 = N->getOperand(0);
7264 EVT CCVT = N0.getValueType();
7266 if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
7267 CCVT.getVectorElementType() != MVT::i1)
7270 EVT ResVT = N->getValueType(0);
7271 EVT CmpVT = N0.getOperand(0).getValueType();
7272 // Only combine when the result type is of the same size as the compared
7274 if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
7277 SDValue IfTrue = N->getOperand(1);
7278 SDValue IfFalse = N->getOperand(2);
7280 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
7281 N0.getOperand(0), N0.getOperand(1),
7282 cast<CondCodeSDNode>(N0.getOperand(2))->get());
7283 return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
7287 /// A vector select: "(select vL, vR, (setcc LHS, RHS))" is best performed with
7288 /// the compare-mask instructions rather than going via NZCV, even if LHS and
7289 /// RHS are really scalar. This replaces any scalar setcc in the above pattern
7290 /// with a vector one followed by a DUP shuffle on the result.
7291 static SDValue performSelectCombine(SDNode *N, SelectionDAG &DAG) {
7292 SDValue N0 = N->getOperand(0);
7293 EVT ResVT = N->getValueType(0);
7295 if (!N->getOperand(1).getValueType().isVector())
7298 if (N0.getOpcode() != ISD::SETCC || N0.getValueType() != MVT::i1)
7303 EVT SrcVT = N0.getOperand(0).getValueType();
7304 SrcVT = EVT::getVectorVT(*DAG.getContext(), SrcVT,
7305 ResVT.getSizeInBits() / SrcVT.getSizeInBits());
7306 EVT CCVT = SrcVT.changeVectorElementTypeToInteger();
7308 // First perform a vector comparison, where lane 0 is the one we're interested
7311 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(0));
7313 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(1));
7314 SDValue SetCC = DAG.getNode(ISD::SETCC, DL, CCVT, LHS, RHS, N0.getOperand(2));
7316 // Now duplicate the comparison mask we want across all other lanes.
7317 SmallVector<int, 8> DUPMask(CCVT.getVectorNumElements(), 0);
7318 SDValue Mask = DAG.getVectorShuffle(CCVT, DL, SetCC, SetCC, DUPMask.data());
7319 Mask = DAG.getNode(ISD::BITCAST, DL, ResVT.changeVectorElementTypeToInteger(),
7322 return DAG.getSelect(DL, ResVT, Mask, N->getOperand(1), N->getOperand(2));
7325 SDValue ARM64TargetLowering::PerformDAGCombine(SDNode *N,
7326 DAGCombinerInfo &DCI) const {
7327 SelectionDAG &DAG = DCI.DAG;
7328 switch (N->getOpcode()) {
7333 return performAddSubLongCombine(N, DCI, DAG);
7335 return performXorCombine(N, DAG, DCI, Subtarget);
7337 return performMulCombine(N, DAG, DCI, Subtarget);
7338 case ISD::SINT_TO_FP:
7339 case ISD::UINT_TO_FP:
7340 return performIntToFpCombine(N, DAG);
7342 return performORCombine(N, DCI, Subtarget);
7343 case ISD::INTRINSIC_WO_CHAIN:
7344 return performIntrinsicCombine(N, DCI, Subtarget);
7345 case ISD::ANY_EXTEND:
7346 case ISD::ZERO_EXTEND:
7347 case ISD::SIGN_EXTEND:
7348 return performExtendCombine(N, DCI, DAG);
7350 return performBitcastCombine(N, DCI, DAG);
7351 case ISD::CONCAT_VECTORS:
7352 return performConcatVectorsCombine(N, DCI, DAG);
7354 return performSelectCombine(N, DAG);
7356 return performVSelectCombine(N, DCI.DAG);
7358 return performSTORECombine(N, DCI, DAG, Subtarget);
7359 case ARM64ISD::BRCOND:
7360 return performBRCONDCombine(N, DCI, DAG);
7361 case ISD::INTRINSIC_VOID:
7362 case ISD::INTRINSIC_W_CHAIN:
7363 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
7364 case Intrinsic::arm64_neon_ld2:
7365 case Intrinsic::arm64_neon_ld3:
7366 case Intrinsic::arm64_neon_ld4:
7367 case Intrinsic::arm64_neon_ld1x2:
7368 case Intrinsic::arm64_neon_ld1x3:
7369 case Intrinsic::arm64_neon_ld1x4:
7370 case Intrinsic::arm64_neon_ld2lane:
7371 case Intrinsic::arm64_neon_ld3lane:
7372 case Intrinsic::arm64_neon_ld4lane:
7373 case Intrinsic::arm64_neon_ld2r:
7374 case Intrinsic::arm64_neon_ld3r:
7375 case Intrinsic::arm64_neon_ld4r:
7376 case Intrinsic::arm64_neon_st2:
7377 case Intrinsic::arm64_neon_st3:
7378 case Intrinsic::arm64_neon_st4:
7379 case Intrinsic::arm64_neon_st1x2:
7380 case Intrinsic::arm64_neon_st1x3:
7381 case Intrinsic::arm64_neon_st1x4:
7382 case Intrinsic::arm64_neon_st2lane:
7383 case Intrinsic::arm64_neon_st3lane:
7384 case Intrinsic::arm64_neon_st4lane:
7385 return performNEONPostLDSTCombine(N, DCI, DAG);
7393 // Check if the return value is used as only a return value, as otherwise
7394 // we can't perform a tail-call. In particular, we need to check for
7395 // target ISD nodes that are returns and any other "odd" constructs
7396 // that the generic analysis code won't necessarily catch.
7397 bool ARM64TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
7398 if (N->getNumValues() != 1)
7400 if (!N->hasNUsesOfValue(1, 0))
7403 SDValue TCChain = Chain;
7404 SDNode *Copy = *N->use_begin();
7405 if (Copy->getOpcode() == ISD::CopyToReg) {
7406 // If the copy has a glue operand, we conservatively assume it isn't safe to
7407 // perform a tail call.
7408 if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() ==
7411 TCChain = Copy->getOperand(0);
7412 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
7415 bool HasRet = false;
7416 for (SDNode *Node : Copy->uses()) {
7417 if (Node->getOpcode() != ARM64ISD::RET_FLAG)
7429 // Return whether the an instruction can potentially be optimized to a tail
7430 // call. This will cause the optimizers to attempt to move, or duplicate,
7431 // return instructions to help enable tail call optimizations for this
7433 bool ARM64TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
7434 if (!EnableARM64TailCalls)
7437 if (!CI->isTailCall())
7443 bool ARM64TargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base,
7445 ISD::MemIndexedMode &AM,
7447 SelectionDAG &DAG) const {
7448 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
7451 Base = Op->getOperand(0);
7452 // All of the indexed addressing mode instructions take a signed
7453 // 9 bit immediate offset.
7454 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
7455 int64_t RHSC = (int64_t)RHS->getZExtValue();
7456 if (RHSC >= 256 || RHSC <= -256)
7458 IsInc = (Op->getOpcode() == ISD::ADD);
7459 Offset = Op->getOperand(1);
7465 bool ARM64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
7467 ISD::MemIndexedMode &AM,
7468 SelectionDAG &DAG) const {
7471 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7472 VT = LD->getMemoryVT();
7473 Ptr = LD->getBasePtr();
7474 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7475 VT = ST->getMemoryVT();
7476 Ptr = ST->getBasePtr();
7481 if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, IsInc, DAG))
7483 AM = IsInc ? ISD::PRE_INC : ISD::PRE_DEC;
7487 bool ARM64TargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
7490 ISD::MemIndexedMode &AM,
7491 SelectionDAG &DAG) const {
7494 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7495 VT = LD->getMemoryVT();
7496 Ptr = LD->getBasePtr();
7497 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7498 VT = ST->getMemoryVT();
7499 Ptr = ST->getBasePtr();
7504 if (!getIndexedAddressParts(Op, Base, Offset, AM, IsInc, DAG))
7506 // Post-indexing updates the base, so it's not a valid transform
7507 // if that's not the same as the load's pointer.
7510 AM = IsInc ? ISD::POST_INC : ISD::POST_DEC;
7514 void ARM64TargetLowering::ReplaceNodeResults(SDNode *N,
7515 SmallVectorImpl<SDValue> &Results,
7516 SelectionDAG &DAG) const {
7517 switch (N->getOpcode()) {
7519 llvm_unreachable("Don't know how to custom expand this");
7520 case ISD::FP_TO_UINT:
7521 case ISD::FP_TO_SINT:
7522 assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion");
7523 // Let normal code take care of it by not adding anything to Results.
7528 bool ARM64TargetLowering::shouldExpandAtomicInIR(Instruction *Inst) const {
7529 // Loads and stores less than 128-bits are already atomic; ones above that
7530 // are doomed anyway, so defer to the default libcall and blame the OS when
7532 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
7533 return SI->getValueOperand()->getType()->getPrimitiveSizeInBits() == 128;
7534 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
7535 return LI->getType()->getPrimitiveSizeInBits() == 128;
7537 // For the real atomic operations, we have ldxr/stxr up to 128 bits.
7538 return Inst->getType()->getPrimitiveSizeInBits() <= 128;
7541 Value *ARM64TargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
7542 AtomicOrdering Ord) const {
7543 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
7544 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
7546 Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent;
7548 // Since i128 isn't legal and intrinsics don't get type-lowered, the ldrexd
7549 // intrinsic must return {i64, i64} and we have to recombine them into a
7550 // single i128 here.
7551 if (ValTy->getPrimitiveSizeInBits() == 128) {
7553 IsAcquire ? Intrinsic::arm64_ldaxp : Intrinsic::arm64_ldxp;
7554 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int);
7556 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
7557 Value *LoHi = Builder.CreateCall(Ldxr, Addr, "lohi");
7559 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
7560 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
7561 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
7562 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
7563 return Builder.CreateOr(
7564 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 64)), "val64");
7567 Type *Tys[] = { Addr->getType() };
7569 IsAcquire ? Intrinsic::arm64_ldaxr : Intrinsic::arm64_ldxr;
7570 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int, Tys);
7572 return Builder.CreateTruncOrBitCast(
7573 Builder.CreateCall(Ldxr, Addr),
7574 cast<PointerType>(Addr->getType())->getElementType());
7577 Value *ARM64TargetLowering::emitStoreConditional(IRBuilder<> &Builder,
7578 Value *Val, Value *Addr,
7579 AtomicOrdering Ord) const {
7580 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
7582 Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent;
7584 // Since the intrinsics must have legal type, the i128 intrinsics take two
7585 // parameters: "i64, i64". We must marshal Val into the appropriate form
7587 if (Val->getType()->getPrimitiveSizeInBits() == 128) {
7589 IsRelease ? Intrinsic::arm64_stlxp : Intrinsic::arm64_stxp;
7590 Function *Stxr = Intrinsic::getDeclaration(M, Int);
7591 Type *Int64Ty = Type::getInt64Ty(M->getContext());
7593 Value *Lo = Builder.CreateTrunc(Val, Int64Ty, "lo");
7594 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 64), Int64Ty, "hi");
7595 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
7596 return Builder.CreateCall3(Stxr, Lo, Hi, Addr);
7600 IsRelease ? Intrinsic::arm64_stlxr : Intrinsic::arm64_stxr;
7601 Type *Tys[] = { Addr->getType() };
7602 Function *Stxr = Intrinsic::getDeclaration(M, Int, Tys);
7604 return Builder.CreateCall2(
7605 Stxr, Builder.CreateZExtOrBitCast(
7606 Val, Stxr->getFunctionType()->getParamType(0)),