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");
47 static cl::opt<AlignMode>
48 Align(cl::desc("Load/store alignment support"),
49 cl::Hidden, cl::init(NoStrictAlign),
51 clEnumValN(StrictAlign, "arm64-strict-align",
52 "Disallow all unaligned memory accesses"),
53 clEnumValN(NoStrictAlign, "arm64-no-strict-align",
54 "Allow unaligned memory accesses"),
57 // Place holder until extr generation is tested fully.
59 EnableARM64ExtrGeneration("arm64-extr-generation", cl::Hidden,
60 cl::desc("Allow ARM64 (or (shift)(shift))->extract"),
64 EnableARM64SlrGeneration("arm64-shift-insert-generation", cl::Hidden,
65 cl::desc("Allow ARM64 SLI/SRI formation"),
68 //===----------------------------------------------------------------------===//
69 // ARM64 Lowering public interface.
70 //===----------------------------------------------------------------------===//
71 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
72 if (TM.getSubtarget<ARM64Subtarget>().isTargetDarwin())
73 return new ARM64_MachoTargetObjectFile();
75 return new ARM64_ELFTargetObjectFile();
78 ARM64TargetLowering::ARM64TargetLowering(ARM64TargetMachine &TM)
79 : TargetLowering(TM, createTLOF(TM)) {
80 Subtarget = &TM.getSubtarget<ARM64Subtarget>();
82 // ARM64 doesn't have comparisons which set GPRs or setcc instructions, so
83 // we have to make something up. Arbitrarily, choose ZeroOrOne.
84 setBooleanContents(ZeroOrOneBooleanContent);
85 // When comparing vectors the result sets the different elements in the
86 // vector to all-one or all-zero.
87 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
89 // Set up the register classes.
90 addRegisterClass(MVT::i32, &ARM64::GPR32allRegClass);
91 addRegisterClass(MVT::i64, &ARM64::GPR64allRegClass);
93 if (Subtarget->hasFPARMv8()) {
94 addRegisterClass(MVT::f16, &ARM64::FPR16RegClass);
95 addRegisterClass(MVT::f32, &ARM64::FPR32RegClass);
96 addRegisterClass(MVT::f64, &ARM64::FPR64RegClass);
97 addRegisterClass(MVT::f128, &ARM64::FPR128RegClass);
100 if (Subtarget->hasNEON()) {
101 addRegisterClass(MVT::v16i8, &ARM64::FPR8RegClass);
102 addRegisterClass(MVT::v8i16, &ARM64::FPR16RegClass);
103 // Someone set us up the NEON.
104 addDRTypeForNEON(MVT::v2f32);
105 addDRTypeForNEON(MVT::v8i8);
106 addDRTypeForNEON(MVT::v4i16);
107 addDRTypeForNEON(MVT::v2i32);
108 addDRTypeForNEON(MVT::v1i64);
109 addDRTypeForNEON(MVT::v1f64);
111 addQRTypeForNEON(MVT::v4f32);
112 addQRTypeForNEON(MVT::v2f64);
113 addQRTypeForNEON(MVT::v16i8);
114 addQRTypeForNEON(MVT::v8i16);
115 addQRTypeForNEON(MVT::v4i32);
116 addQRTypeForNEON(MVT::v2i64);
119 // Compute derived properties from the register classes
120 computeRegisterProperties();
122 // Provide all sorts of operation actions
123 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
124 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
125 setOperationAction(ISD::SETCC, MVT::i32, Custom);
126 setOperationAction(ISD::SETCC, MVT::i64, Custom);
127 setOperationAction(ISD::SETCC, MVT::f32, Custom);
128 setOperationAction(ISD::SETCC, MVT::f64, Custom);
129 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
130 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
131 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
132 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
133 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
134 setOperationAction(ISD::SELECT, MVT::i32, Custom);
135 setOperationAction(ISD::SELECT, MVT::i64, Custom);
136 setOperationAction(ISD::SELECT, MVT::f32, Custom);
137 setOperationAction(ISD::SELECT, MVT::f64, Custom);
138 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
139 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
140 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
141 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
142 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
143 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
145 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
146 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
147 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
149 setOperationAction(ISD::FREM, MVT::f32, Expand);
150 setOperationAction(ISD::FREM, MVT::f64, Expand);
151 setOperationAction(ISD::FREM, MVT::f80, Expand);
153 // Custom lowering hooks are needed for XOR
154 // to fold it into CSINC/CSINV.
155 setOperationAction(ISD::XOR, MVT::i32, Custom);
156 setOperationAction(ISD::XOR, MVT::i64, Custom);
158 // Virtually no operation on f128 is legal, but LLVM can't expand them when
159 // there's a valid register class, so we need custom operations in most cases.
160 setOperationAction(ISD::FABS, MVT::f128, Expand);
161 setOperationAction(ISD::FADD, MVT::f128, Custom);
162 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
163 setOperationAction(ISD::FCOS, MVT::f128, Expand);
164 setOperationAction(ISD::FDIV, MVT::f128, Custom);
165 setOperationAction(ISD::FMA, MVT::f128, Expand);
166 setOperationAction(ISD::FMUL, MVT::f128, Custom);
167 setOperationAction(ISD::FNEG, MVT::f128, Expand);
168 setOperationAction(ISD::FPOW, MVT::f128, Expand);
169 setOperationAction(ISD::FREM, MVT::f128, Expand);
170 setOperationAction(ISD::FRINT, MVT::f128, Expand);
171 setOperationAction(ISD::FSIN, MVT::f128, Expand);
172 setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
173 setOperationAction(ISD::FSQRT, MVT::f128, Expand);
174 setOperationAction(ISD::FSUB, MVT::f128, Custom);
175 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
176 setOperationAction(ISD::SETCC, MVT::f128, Custom);
177 setOperationAction(ISD::BR_CC, MVT::f128, Custom);
178 setOperationAction(ISD::SELECT, MVT::f128, Custom);
179 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
180 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);
182 // Lowering for many of the conversions is actually specified by the non-f128
183 // type. The LowerXXX function will be trivial when f128 isn't involved.
184 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
185 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
186 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
187 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
188 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
189 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
190 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
191 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
192 setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
193 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
194 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
195 setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
196 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
197 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
199 // Variable arguments.
200 setOperationAction(ISD::VASTART, MVT::Other, Custom);
201 setOperationAction(ISD::VAARG, MVT::Other, Custom);
202 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
203 setOperationAction(ISD::VAEND, MVT::Other, Expand);
205 // Variable-sized objects.
206 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
207 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
208 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
210 // Exception handling.
211 // FIXME: These are guesses. Has this been defined yet?
212 setExceptionPointerRegister(ARM64::X0);
213 setExceptionSelectorRegister(ARM64::X1);
215 // Constant pool entries
216 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
219 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
221 // Add/Sub overflow ops with MVT::Glues are lowered to NZCV dependences.
222 setOperationAction(ISD::ADDC, MVT::i32, Custom);
223 setOperationAction(ISD::ADDE, MVT::i32, Custom);
224 setOperationAction(ISD::SUBC, MVT::i32, Custom);
225 setOperationAction(ISD::SUBE, MVT::i32, Custom);
226 setOperationAction(ISD::ADDC, MVT::i64, Custom);
227 setOperationAction(ISD::ADDE, MVT::i64, Custom);
228 setOperationAction(ISD::SUBC, MVT::i64, Custom);
229 setOperationAction(ISD::SUBE, MVT::i64, Custom);
231 // ARM64 lacks both left-rotate and popcount instructions.
232 setOperationAction(ISD::ROTL, MVT::i32, Expand);
233 setOperationAction(ISD::ROTL, MVT::i64, Expand);
235 // ARM64 doesn't have {U|S}MUL_LOHI.
236 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
237 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
240 // Expand the undefined-at-zero variants to cttz/ctlz to their defined-at-zero
241 // counterparts, which ARM64 supports directly.
242 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
243 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
244 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
245 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
247 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
248 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
250 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
251 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
252 setOperationAction(ISD::SREM, MVT::i32, Expand);
253 setOperationAction(ISD::SREM, MVT::i64, Expand);
254 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
255 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
256 setOperationAction(ISD::UREM, MVT::i32, Expand);
257 setOperationAction(ISD::UREM, MVT::i64, Expand);
259 // Custom lower Add/Sub/Mul with overflow.
260 setOperationAction(ISD::SADDO, MVT::i32, Custom);
261 setOperationAction(ISD::SADDO, MVT::i64, Custom);
262 setOperationAction(ISD::UADDO, MVT::i32, Custom);
263 setOperationAction(ISD::UADDO, MVT::i64, Custom);
264 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
265 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
266 setOperationAction(ISD::USUBO, MVT::i32, Custom);
267 setOperationAction(ISD::USUBO, MVT::i64, Custom);
268 setOperationAction(ISD::SMULO, MVT::i32, Custom);
269 setOperationAction(ISD::SMULO, MVT::i64, Custom);
270 setOperationAction(ISD::UMULO, MVT::i32, Custom);
271 setOperationAction(ISD::UMULO, MVT::i64, Custom);
273 setOperationAction(ISD::FSIN, MVT::f32, Expand);
274 setOperationAction(ISD::FSIN, MVT::f64, Expand);
275 setOperationAction(ISD::FCOS, MVT::f32, Expand);
276 setOperationAction(ISD::FCOS, MVT::f64, Expand);
277 setOperationAction(ISD::FPOW, MVT::f32, Expand);
278 setOperationAction(ISD::FPOW, MVT::f64, Expand);
279 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
280 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
282 // ARM64 has implementations of a lot of rounding-like FP operations.
283 static MVT RoundingTypes[] = { MVT::f32, MVT::f64};
284 for (unsigned I = 0; I < array_lengthof(RoundingTypes); ++I) {
285 MVT Ty = RoundingTypes[I];
286 setOperationAction(ISD::FFLOOR, Ty, Legal);
287 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
288 setOperationAction(ISD::FCEIL, Ty, Legal);
289 setOperationAction(ISD::FRINT, Ty, Legal);
290 setOperationAction(ISD::FTRUNC, Ty, Legal);
291 setOperationAction(ISD::FROUND, Ty, Legal);
294 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
296 if (Subtarget->isTargetMachO()) {
297 // For iOS, we don't want to the normal expansion of a libcall to
298 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
300 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
301 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
303 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
304 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
307 // ARM64 does not have floating-point extending loads, i1 sign-extending load,
308 // floating-point truncating stores, or v2i32->v2i16 truncating store.
309 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
310 setLoadExtAction(ISD::EXTLOAD, MVT::f64, Expand);
311 setLoadExtAction(ISD::EXTLOAD, MVT::f80, Expand);
312 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Expand);
313 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
314 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
315 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
316 setTruncStoreAction(MVT::f128, MVT::f80, Expand);
317 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
318 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
319 setTruncStoreAction(MVT::f128, MVT::f16, Expand);
320 // Indexed loads and stores are supported.
321 for (unsigned im = (unsigned)ISD::PRE_INC;
322 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
323 setIndexedLoadAction(im, MVT::i8, Legal);
324 setIndexedLoadAction(im, MVT::i16, Legal);
325 setIndexedLoadAction(im, MVT::i32, Legal);
326 setIndexedLoadAction(im, MVT::i64, Legal);
327 setIndexedLoadAction(im, MVT::f64, Legal);
328 setIndexedLoadAction(im, MVT::f32, Legal);
329 setIndexedStoreAction(im, MVT::i8, Legal);
330 setIndexedStoreAction(im, MVT::i16, Legal);
331 setIndexedStoreAction(im, MVT::i32, Legal);
332 setIndexedStoreAction(im, MVT::i64, Legal);
333 setIndexedStoreAction(im, MVT::f64, Legal);
334 setIndexedStoreAction(im, MVT::f32, Legal);
338 setOperationAction(ISD::TRAP, MVT::Other, Legal);
340 // We combine OR nodes for bitfield operations.
341 setTargetDAGCombine(ISD::OR);
343 // Vector add and sub nodes may conceal a high-half opportunity.
344 // Also, try to fold ADD into CSINC/CSINV..
345 setTargetDAGCombine(ISD::ADD);
346 setTargetDAGCombine(ISD::SUB);
348 setTargetDAGCombine(ISD::XOR);
349 setTargetDAGCombine(ISD::SINT_TO_FP);
350 setTargetDAGCombine(ISD::UINT_TO_FP);
352 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
354 setTargetDAGCombine(ISD::ANY_EXTEND);
355 setTargetDAGCombine(ISD::ZERO_EXTEND);
356 setTargetDAGCombine(ISD::SIGN_EXTEND);
357 setTargetDAGCombine(ISD::BITCAST);
358 setTargetDAGCombine(ISD::CONCAT_VECTORS);
359 setTargetDAGCombine(ISD::STORE);
361 setTargetDAGCombine(ISD::MUL);
363 setTargetDAGCombine(ISD::SELECT);
364 setTargetDAGCombine(ISD::VSELECT);
366 setTargetDAGCombine(ISD::INTRINSIC_VOID);
367 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
368 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
370 MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8;
371 MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4;
372 MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 4;
374 setStackPointerRegisterToSaveRestore(ARM64::SP);
376 setSchedulingPreference(Sched::Hybrid);
379 MaskAndBranchFoldingIsLegal = true;
381 setMinFunctionAlignment(2);
383 RequireStrictAlign = (Align == StrictAlign);
385 setHasExtractBitsInsn(true);
387 if (Subtarget->hasNEON()) {
388 // FIXME: v1f64 shouldn't be legal if we can avoid it, because it leads to
389 // silliness like this:
390 setOperationAction(ISD::FABS, MVT::v1f64, Expand);
391 setOperationAction(ISD::FADD, MVT::v1f64, Expand);
392 setOperationAction(ISD::FCEIL, MVT::v1f64, Expand);
393 setOperationAction(ISD::FCOPYSIGN, MVT::v1f64, Expand);
394 setOperationAction(ISD::FCOS, MVT::v1f64, Expand);
395 setOperationAction(ISD::FDIV, MVT::v1f64, Expand);
396 setOperationAction(ISD::FFLOOR, MVT::v1f64, Expand);
397 setOperationAction(ISD::FMA, MVT::v1f64, Expand);
398 setOperationAction(ISD::FMUL, MVT::v1f64, Expand);
399 setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Expand);
400 setOperationAction(ISD::FNEG, MVT::v1f64, Expand);
401 setOperationAction(ISD::FPOW, MVT::v1f64, Expand);
402 setOperationAction(ISD::FREM, MVT::v1f64, Expand);
403 setOperationAction(ISD::FROUND, MVT::v1f64, Expand);
404 setOperationAction(ISD::FRINT, MVT::v1f64, Expand);
405 setOperationAction(ISD::FSIN, MVT::v1f64, Expand);
406 setOperationAction(ISD::FSINCOS, MVT::v1f64, Expand);
407 setOperationAction(ISD::FSQRT, MVT::v1f64, Expand);
408 setOperationAction(ISD::FSUB, MVT::v1f64, Expand);
409 setOperationAction(ISD::FTRUNC, MVT::v1f64, Expand);
410 setOperationAction(ISD::SETCC, MVT::v1f64, Expand);
411 setOperationAction(ISD::BR_CC, MVT::v1f64, Expand);
412 setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
413 setOperationAction(ISD::SELECT_CC, MVT::v1f64, Expand);
414 setOperationAction(ISD::FP_EXTEND, MVT::v1f64, Expand);
416 setOperationAction(ISD::FP_TO_SINT, MVT::v1i64, Expand);
417 setOperationAction(ISD::FP_TO_UINT, MVT::v1i64, Expand);
418 setOperationAction(ISD::SINT_TO_FP, MVT::v1i64, Expand);
419 setOperationAction(ISD::UINT_TO_FP, MVT::v1i64, Expand);
420 setOperationAction(ISD::FP_ROUND, MVT::v1f64, Expand);
422 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
424 // ARM64 doesn't have a direct vector ->f32 conversion instructions for
425 // elements smaller than i32, so promote the input to i32 first.
426 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Promote);
427 setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Promote);
428 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Promote);
429 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Promote);
430 // Similarly, there is no direct i32 -> f64 vector conversion instruction.
431 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
432 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
433 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
434 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
436 // ARM64 doesn't have MUL.2d:
437 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
438 setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal);
439 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
440 // Likewise, narrowing and extending vector loads/stores aren't handled
442 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
443 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
445 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,
448 setOperationAction(ISD::MULHS, (MVT::SimpleValueType)VT, Expand);
449 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
450 setOperationAction(ISD::MULHU, (MVT::SimpleValueType)VT, Expand);
451 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
453 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
454 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
455 setTruncStoreAction((MVT::SimpleValueType)VT,
456 (MVT::SimpleValueType)InnerVT, Expand);
457 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
458 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
459 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
462 // ARM64 has implementations of a lot of rounding-like FP operations.
463 static MVT RoundingVecTypes[] = {MVT::v2f32, MVT::v4f32, MVT::v2f64 };
464 for (unsigned I = 0; I < array_lengthof(RoundingVecTypes); ++I) {
465 MVT Ty = RoundingVecTypes[I];
466 setOperationAction(ISD::FFLOOR, Ty, Legal);
467 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
468 setOperationAction(ISD::FCEIL, Ty, Legal);
469 setOperationAction(ISD::FRINT, Ty, Legal);
470 setOperationAction(ISD::FTRUNC, Ty, Legal);
471 setOperationAction(ISD::FROUND, Ty, Legal);
476 void ARM64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) {
477 if (VT == MVT::v2f32) {
478 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
479 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i32);
481 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
482 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i32);
483 } else if (VT == MVT::v2f64 || VT == MVT::v4f32) {
484 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
485 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i64);
487 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
488 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i64);
491 // Mark vector float intrinsics as expand.
492 if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) {
493 setOperationAction(ISD::FSIN, VT.getSimpleVT(), Expand);
494 setOperationAction(ISD::FCOS, VT.getSimpleVT(), Expand);
495 setOperationAction(ISD::FPOWI, VT.getSimpleVT(), Expand);
496 setOperationAction(ISD::FPOW, VT.getSimpleVT(), Expand);
497 setOperationAction(ISD::FLOG, VT.getSimpleVT(), Expand);
498 setOperationAction(ISD::FLOG2, VT.getSimpleVT(), Expand);
499 setOperationAction(ISD::FLOG10, VT.getSimpleVT(), Expand);
500 setOperationAction(ISD::FEXP, VT.getSimpleVT(), Expand);
501 setOperationAction(ISD::FEXP2, VT.getSimpleVT(), Expand);
504 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
505 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
506 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
507 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
508 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Custom);
509 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
510 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
511 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
512 setOperationAction(ISD::AND, VT.getSimpleVT(), Custom);
513 setOperationAction(ISD::OR, VT.getSimpleVT(), Custom);
514 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
515 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
517 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
518 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
519 setOperationAction(ISD::VSELECT, VT.getSimpleVT(), Expand);
520 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
522 // CNT supports only B element sizes.
523 if (VT != MVT::v8i8 && VT != MVT::v16i8)
524 setOperationAction(ISD::CTPOP, VT.getSimpleVT(), Expand);
526 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
527 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
528 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
529 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
530 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
532 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
533 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
535 if (Subtarget->isLittleEndian()) {
536 for (unsigned im = (unsigned)ISD::PRE_INC;
537 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
538 setIndexedLoadAction(im, VT.getSimpleVT(), Legal);
539 setIndexedStoreAction(im, VT.getSimpleVT(), Legal);
544 void ARM64TargetLowering::addDRTypeForNEON(MVT VT) {
545 addRegisterClass(VT, &ARM64::FPR64RegClass);
546 addTypeForNEON(VT, MVT::v2i32);
549 void ARM64TargetLowering::addQRTypeForNEON(MVT VT) {
550 addRegisterClass(VT, &ARM64::FPR128RegClass);
551 addTypeForNEON(VT, MVT::v4i32);
554 EVT ARM64TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
557 return VT.changeVectorElementTypeToInteger();
560 /// computeKnownBitsForTargetNode - Determine which of the bits specified in
561 /// Mask are known to be either zero or one and return them in the
562 /// KnownZero/KnownOne bitsets.
563 void ARM64TargetLowering::computeKnownBitsForTargetNode(
564 const SDValue Op, APInt &KnownZero, APInt &KnownOne,
565 const SelectionDAG &DAG, unsigned Depth) const {
566 switch (Op.getOpcode()) {
569 case ARM64ISD::CSEL: {
570 APInt KnownZero2, KnownOne2;
571 DAG.computeKnownBits(Op->getOperand(0), KnownZero, KnownOne, Depth + 1);
572 DAG.computeKnownBits(Op->getOperand(1), KnownZero2, KnownOne2, Depth + 1);
573 KnownZero &= KnownZero2;
574 KnownOne &= KnownOne2;
577 case ISD::INTRINSIC_W_CHAIN: {
578 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
579 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
582 case Intrinsic::arm64_ldaxr:
583 case Intrinsic::arm64_ldxr: {
584 unsigned BitWidth = KnownOne.getBitWidth();
585 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
586 unsigned MemBits = VT.getScalarType().getSizeInBits();
587 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
593 case ISD::INTRINSIC_WO_CHAIN:
594 case ISD::INTRINSIC_VOID: {
595 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
599 case Intrinsic::arm64_neon_umaxv:
600 case Intrinsic::arm64_neon_uminv: {
601 // Figure out the datatype of the vector operand. The UMINV instruction
602 // will zero extend the result, so we can mark as known zero all the
603 // bits larger than the element datatype. 32-bit or larget doesn't need
604 // this as those are legal types and will be handled by isel directly.
605 MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
606 unsigned BitWidth = KnownZero.getBitWidth();
607 if (VT == MVT::v8i8 || VT == MVT::v16i8) {
608 assert(BitWidth >= 8 && "Unexpected width!");
609 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8);
611 } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
612 assert(BitWidth >= 16 && "Unexpected width!");
613 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
623 MVT ARM64TargetLowering::getScalarShiftAmountTy(EVT LHSTy) const {
627 unsigned ARM64TargetLowering::getMaximalGlobalOffset() const {
628 // FIXME: On ARM64, this depends on the type.
629 // Basically, the addressable offsets are o to 4095 * Ty.getSizeInBytes().
630 // and the offset has to be a multiple of the related size in bytes.
635 ARM64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
636 const TargetLibraryInfo *libInfo) const {
637 return ARM64::createFastISel(funcInfo, libInfo);
640 const char *ARM64TargetLowering::getTargetNodeName(unsigned Opcode) const {
644 case ARM64ISD::CALL: return "ARM64ISD::CALL";
645 case ARM64ISD::ADRP: return "ARM64ISD::ADRP";
646 case ARM64ISD::ADDlow: return "ARM64ISD::ADDlow";
647 case ARM64ISD::LOADgot: return "ARM64ISD::LOADgot";
648 case ARM64ISD::RET_FLAG: return "ARM64ISD::RET_FLAG";
649 case ARM64ISD::BRCOND: return "ARM64ISD::BRCOND";
650 case ARM64ISD::CSEL: return "ARM64ISD::CSEL";
651 case ARM64ISD::FCSEL: return "ARM64ISD::FCSEL";
652 case ARM64ISD::CSINV: return "ARM64ISD::CSINV";
653 case ARM64ISD::CSNEG: return "ARM64ISD::CSNEG";
654 case ARM64ISD::CSINC: return "ARM64ISD::CSINC";
655 case ARM64ISD::THREAD_POINTER: return "ARM64ISD::THREAD_POINTER";
656 case ARM64ISD::TLSDESC_CALL: return "ARM64ISD::TLSDESC_CALL";
657 case ARM64ISD::ADC: return "ARM64ISD::ADC";
658 case ARM64ISD::SBC: return "ARM64ISD::SBC";
659 case ARM64ISD::ADDS: return "ARM64ISD::ADDS";
660 case ARM64ISD::SUBS: return "ARM64ISD::SUBS";
661 case ARM64ISD::ADCS: return "ARM64ISD::ADCS";
662 case ARM64ISD::SBCS: return "ARM64ISD::SBCS";
663 case ARM64ISD::ANDS: return "ARM64ISD::ANDS";
664 case ARM64ISD::FCMP: return "ARM64ISD::FCMP";
665 case ARM64ISD::FMIN: return "ARM64ISD::FMIN";
666 case ARM64ISD::FMAX: return "ARM64ISD::FMAX";
667 case ARM64ISD::DUP: return "ARM64ISD::DUP";
668 case ARM64ISD::DUPLANE8: return "ARM64ISD::DUPLANE8";
669 case ARM64ISD::DUPLANE16: return "ARM64ISD::DUPLANE16";
670 case ARM64ISD::DUPLANE32: return "ARM64ISD::DUPLANE32";
671 case ARM64ISD::DUPLANE64: return "ARM64ISD::DUPLANE64";
672 case ARM64ISD::MOVI: return "ARM64ISD::MOVI";
673 case ARM64ISD::MOVIshift: return "ARM64ISD::MOVIshift";
674 case ARM64ISD::MOVIedit: return "ARM64ISD::MOVIedit";
675 case ARM64ISD::MOVImsl: return "ARM64ISD::MOVImsl";
676 case ARM64ISD::FMOV: return "ARM64ISD::FMOV";
677 case ARM64ISD::MVNIshift: return "ARM64ISD::MVNIshift";
678 case ARM64ISD::MVNImsl: return "ARM64ISD::MVNImsl";
679 case ARM64ISD::BICi: return "ARM64ISD::BICi";
680 case ARM64ISD::ORRi: return "ARM64ISD::ORRi";
681 case ARM64ISD::BSL: return "ARM64ISD::BSL";
682 case ARM64ISD::NEG: return "ARM64ISD::NEG";
683 case ARM64ISD::EXTR: return "ARM64ISD::EXTR";
684 case ARM64ISD::ZIP1: return "ARM64ISD::ZIP1";
685 case ARM64ISD::ZIP2: return "ARM64ISD::ZIP2";
686 case ARM64ISD::UZP1: return "ARM64ISD::UZP1";
687 case ARM64ISD::UZP2: return "ARM64ISD::UZP2";
688 case ARM64ISD::TRN1: return "ARM64ISD::TRN1";
689 case ARM64ISD::TRN2: return "ARM64ISD::TRN2";
690 case ARM64ISD::REV16: return "ARM64ISD::REV16";
691 case ARM64ISD::REV32: return "ARM64ISD::REV32";
692 case ARM64ISD::REV64: return "ARM64ISD::REV64";
693 case ARM64ISD::EXT: return "ARM64ISD::EXT";
694 case ARM64ISD::VSHL: return "ARM64ISD::VSHL";
695 case ARM64ISD::VLSHR: return "ARM64ISD::VLSHR";
696 case ARM64ISD::VASHR: return "ARM64ISD::VASHR";
697 case ARM64ISD::CMEQ: return "ARM64ISD::CMEQ";
698 case ARM64ISD::CMGE: return "ARM64ISD::CMGE";
699 case ARM64ISD::CMGT: return "ARM64ISD::CMGT";
700 case ARM64ISD::CMHI: return "ARM64ISD::CMHI";
701 case ARM64ISD::CMHS: return "ARM64ISD::CMHS";
702 case ARM64ISD::FCMEQ: return "ARM64ISD::FCMEQ";
703 case ARM64ISD::FCMGE: return "ARM64ISD::FCMGE";
704 case ARM64ISD::FCMGT: return "ARM64ISD::FCMGT";
705 case ARM64ISD::CMEQz: return "ARM64ISD::CMEQz";
706 case ARM64ISD::CMGEz: return "ARM64ISD::CMGEz";
707 case ARM64ISD::CMGTz: return "ARM64ISD::CMGTz";
708 case ARM64ISD::CMLEz: return "ARM64ISD::CMLEz";
709 case ARM64ISD::CMLTz: return "ARM64ISD::CMLTz";
710 case ARM64ISD::FCMEQz: return "ARM64ISD::FCMEQz";
711 case ARM64ISD::FCMGEz: return "ARM64ISD::FCMGEz";
712 case ARM64ISD::FCMGTz: return "ARM64ISD::FCMGTz";
713 case ARM64ISD::FCMLEz: return "ARM64ISD::FCMLEz";
714 case ARM64ISD::FCMLTz: return "ARM64ISD::FCMLTz";
715 case ARM64ISD::NOT: return "ARM64ISD::NOT";
716 case ARM64ISD::BIT: return "ARM64ISD::BIT";
717 case ARM64ISD::CBZ: return "ARM64ISD::CBZ";
718 case ARM64ISD::CBNZ: return "ARM64ISD::CBNZ";
719 case ARM64ISD::TBZ: return "ARM64ISD::TBZ";
720 case ARM64ISD::TBNZ: return "ARM64ISD::TBNZ";
721 case ARM64ISD::TC_RETURN: return "ARM64ISD::TC_RETURN";
722 case ARM64ISD::SITOF: return "ARM64ISD::SITOF";
723 case ARM64ISD::UITOF: return "ARM64ISD::UITOF";
724 case ARM64ISD::SQSHL_I: return "ARM64ISD::SQSHL_I";
725 case ARM64ISD::UQSHL_I: return "ARM64ISD::UQSHL_I";
726 case ARM64ISD::SRSHR_I: return "ARM64ISD::SRSHR_I";
727 case ARM64ISD::URSHR_I: return "ARM64ISD::URSHR_I";
728 case ARM64ISD::SQSHLU_I: return "ARM64ISD::SQSHLU_I";
729 case ARM64ISD::WrapperLarge: return "ARM64ISD::WrapperLarge";
730 case ARM64ISD::LD2post: return "ARM64ISD::LD2post";
731 case ARM64ISD::LD3post: return "ARM64ISD::LD3post";
732 case ARM64ISD::LD4post: return "ARM64ISD::LD4post";
733 case ARM64ISD::ST2post: return "ARM64ISD::ST2post";
734 case ARM64ISD::ST3post: return "ARM64ISD::ST3post";
735 case ARM64ISD::ST4post: return "ARM64ISD::ST4post";
736 case ARM64ISD::LD1x2post: return "ARM64ISD::LD1x2post";
737 case ARM64ISD::LD1x3post: return "ARM64ISD::LD1x3post";
738 case ARM64ISD::LD1x4post: return "ARM64ISD::LD1x4post";
739 case ARM64ISD::ST1x2post: return "ARM64ISD::ST1x2post";
740 case ARM64ISD::ST1x3post: return "ARM64ISD::ST1x3post";
741 case ARM64ISD::ST1x4post: return "ARM64ISD::ST1x4post";
742 case ARM64ISD::LD1DUPpost: return "ARM64ISD::LD1DUPpost";
743 case ARM64ISD::LD2DUPpost: return "ARM64ISD::LD2DUPpost";
744 case ARM64ISD::LD3DUPpost: return "ARM64ISD::LD3DUPpost";
745 case ARM64ISD::LD4DUPpost: return "ARM64ISD::LD4DUPpost";
746 case ARM64ISD::LD1LANEpost: return "ARM64ISD::LD1LANEpost";
747 case ARM64ISD::LD2LANEpost: return "ARM64ISD::LD2LANEpost";
748 case ARM64ISD::LD3LANEpost: return "ARM64ISD::LD3LANEpost";
749 case ARM64ISD::LD4LANEpost: return "ARM64ISD::LD4LANEpost";
750 case ARM64ISD::ST2LANEpost: return "ARM64ISD::ST2LANEpost";
751 case ARM64ISD::ST3LANEpost: return "ARM64ISD::ST3LANEpost";
752 case ARM64ISD::ST4LANEpost: return "ARM64ISD::ST4LANEpost";
757 ARM64TargetLowering::EmitF128CSEL(MachineInstr *MI,
758 MachineBasicBlock *MBB) const {
759 // We materialise the F128CSEL pseudo-instruction as some control flow and a
763 // [... previous instrs leading to comparison ...]
769 // Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB]
771 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
772 MachineFunction *MF = MBB->getParent();
773 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
774 DebugLoc DL = MI->getDebugLoc();
775 MachineFunction::iterator It = MBB;
778 unsigned DestReg = MI->getOperand(0).getReg();
779 unsigned IfTrueReg = MI->getOperand(1).getReg();
780 unsigned IfFalseReg = MI->getOperand(2).getReg();
781 unsigned CondCode = MI->getOperand(3).getImm();
782 bool NZCVKilled = MI->getOperand(4).isKill();
784 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
785 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
786 MF->insert(It, TrueBB);
787 MF->insert(It, EndBB);
789 // Transfer rest of current basic-block to EndBB
790 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
792 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
794 BuildMI(MBB, DL, TII->get(ARM64::Bcc)).addImm(CondCode).addMBB(TrueBB);
795 BuildMI(MBB, DL, TII->get(ARM64::B)).addMBB(EndBB);
796 MBB->addSuccessor(TrueBB);
797 MBB->addSuccessor(EndBB);
799 // TrueBB falls through to the end.
800 TrueBB->addSuccessor(EndBB);
803 TrueBB->addLiveIn(ARM64::NZCV);
804 EndBB->addLiveIn(ARM64::NZCV);
807 BuildMI(*EndBB, EndBB->begin(), DL, TII->get(ARM64::PHI), DestReg)
813 MI->eraseFromParent();
818 ARM64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
819 MachineBasicBlock *BB) const {
820 switch (MI->getOpcode()) {
825 assert(0 && "Unexpected instruction for custom inserter!");
828 case ARM64::F128CSEL:
829 return EmitF128CSEL(MI, BB);
831 case TargetOpcode::STACKMAP:
832 case TargetOpcode::PATCHPOINT:
833 return emitPatchPoint(MI, BB);
835 llvm_unreachable("Unexpected instruction for custom inserter!");
838 //===----------------------------------------------------------------------===//
839 // ARM64 Lowering private implementation.
840 //===----------------------------------------------------------------------===//
842 //===----------------------------------------------------------------------===//
844 //===----------------------------------------------------------------------===//
846 /// changeIntCCToARM64CC - Convert a DAG integer condition code to an ARM64 CC
847 static ARM64CC::CondCode changeIntCCToARM64CC(ISD::CondCode CC) {
850 llvm_unreachable("Unknown condition code!");
874 /// changeFPCCToARM64CC - Convert a DAG fp condition code to an ARM64 CC.
875 static void changeFPCCToARM64CC(ISD::CondCode CC, ARM64CC::CondCode &CondCode,
876 ARM64CC::CondCode &CondCode2) {
877 CondCode2 = ARM64CC::AL;
880 llvm_unreachable("Unknown FP condition!");
883 CondCode = ARM64CC::EQ;
887 CondCode = ARM64CC::GT;
891 CondCode = ARM64CC::GE;
894 CondCode = ARM64CC::MI;
897 CondCode = ARM64CC::LS;
900 CondCode = ARM64CC::MI;
901 CondCode2 = ARM64CC::GT;
904 CondCode = ARM64CC::VC;
907 CondCode = ARM64CC::VS;
910 CondCode = ARM64CC::EQ;
911 CondCode2 = ARM64CC::VS;
914 CondCode = ARM64CC::HI;
917 CondCode = ARM64CC::PL;
921 CondCode = ARM64CC::LT;
925 CondCode = ARM64CC::LE;
929 CondCode = ARM64CC::NE;
934 /// changeVectorFPCCToARM64CC - Convert a DAG fp condition code to an ARM64 CC
935 /// usable with the vector instructions. Fewer operations are available without
936 /// a real NZCV register, so we have to use less efficient combinations to get
938 static void changeVectorFPCCToARM64CC(ISD::CondCode CC,
939 ARM64CC::CondCode &CondCode,
940 ARM64CC::CondCode &CondCode2,
945 // Mostly the scalar mappings work fine.
946 changeFPCCToARM64CC(CC, CondCode, CondCode2);
949 Invert = true; // Fallthrough
951 CondCode = ARM64CC::MI;
952 CondCode2 = ARM64CC::GE;
959 // All of the compare-mask comparisons are ordered, but we can switch
960 // between the two by a double inversion. E.g. ULE == !OGT.
962 changeFPCCToARM64CC(getSetCCInverse(CC, false), CondCode, CondCode2);
967 static bool isLegalArithImmed(uint64_t C) {
968 // Matches ARM64DAGToDAGISel::SelectArithImmed().
969 return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
972 static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
973 SDLoc dl, SelectionDAG &DAG) {
974 EVT VT = LHS.getValueType();
976 if (VT.isFloatingPoint())
977 return DAG.getNode(ARM64ISD::FCMP, dl, VT, LHS, RHS);
979 // The CMP instruction is just an alias for SUBS, and representing it as
980 // SUBS means that it's possible to get CSE with subtract operations.
981 // A later phase can perform the optimization of setting the destination
982 // register to WZR/XZR if it ends up being unused.
983 unsigned Opcode = ARM64ISD::SUBS;
985 if (RHS.getOpcode() == ISD::SUB && isa<ConstantSDNode>(RHS.getOperand(0)) &&
986 cast<ConstantSDNode>(RHS.getOperand(0))->getZExtValue() == 0 &&
987 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
988 // We'd like to combine a (CMP op1, (sub 0, op2) into a CMN instruction on
989 // the grounds that "op1 - (-op2) == op1 + op2". However, the C and V flags
990 // can be set differently by this operation. It comes down to whether
991 // "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are then
992 // everything is fine. If not then the optimization is wrong. Thus general
993 // comparisons are only valid if op2 != 0.
995 // So, finally, the only LLVM-native comparisons that don't mention C and V
996 // are SETEQ and SETNE. They're the only ones we can safely use CMN for in
997 // the absence of information about op2.
998 Opcode = ARM64ISD::ADDS;
999 RHS = RHS.getOperand(1);
1000 } else if (LHS.getOpcode() == ISD::AND && isa<ConstantSDNode>(RHS) &&
1001 cast<ConstantSDNode>(RHS)->getZExtValue() == 0 &&
1002 !isUnsignedIntSetCC(CC)) {
1003 // Similarly, (CMP (and X, Y), 0) can be implemented with a TST
1004 // (a.k.a. ANDS) except that the flags are only guaranteed to work for one
1005 // of the signed comparisons.
1006 Opcode = ARM64ISD::ANDS;
1007 RHS = LHS.getOperand(1);
1008 LHS = LHS.getOperand(0);
1011 return DAG.getNode(Opcode, dl, DAG.getVTList(VT, MVT::i32), LHS, RHS)
1015 static SDValue getARM64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1016 SDValue &ARM64cc, SelectionDAG &DAG, SDLoc dl) {
1017 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1018 EVT VT = RHS.getValueType();
1019 uint64_t C = RHSC->getZExtValue();
1020 if (!isLegalArithImmed(C)) {
1021 // Constant does not fit, try adjusting it by one?
1027 if ((VT == MVT::i32 && C != 0x80000000 &&
1028 isLegalArithImmed((uint32_t)(C - 1))) ||
1029 (VT == MVT::i64 && C != 0x80000000ULL &&
1030 isLegalArithImmed(C - 1ULL))) {
1031 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1032 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1033 RHS = DAG.getConstant(C, VT);
1038 if ((VT == MVT::i32 && C != 0 &&
1039 isLegalArithImmed((uint32_t)(C - 1))) ||
1040 (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) {
1041 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1042 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1043 RHS = DAG.getConstant(C, VT);
1048 if ((VT == MVT::i32 && C != 0x7fffffff &&
1049 isLegalArithImmed((uint32_t)(C + 1))) ||
1050 (VT == MVT::i64 && C != 0x7ffffffffffffffULL &&
1051 isLegalArithImmed(C + 1ULL))) {
1052 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1053 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1054 RHS = DAG.getConstant(C, VT);
1059 if ((VT == MVT::i32 && C != 0xffffffff &&
1060 isLegalArithImmed((uint32_t)(C + 1))) ||
1061 (VT == MVT::i64 && C != 0xfffffffffffffffULL &&
1062 isLegalArithImmed(C + 1ULL))) {
1063 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1064 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1065 RHS = DAG.getConstant(C, VT);
1072 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
1073 ARM64CC::CondCode ARM64CC = changeIntCCToARM64CC(CC);
1074 ARM64cc = DAG.getConstant(ARM64CC, MVT::i32);
1078 static std::pair<SDValue, SDValue>
1079 getARM64XALUOOp(ARM64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
1080 assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
1081 "Unsupported value type");
1082 SDValue Value, Overflow;
1084 SDValue LHS = Op.getOperand(0);
1085 SDValue RHS = Op.getOperand(1);
1087 switch (Op.getOpcode()) {
1089 llvm_unreachable("Unknown overflow instruction!");
1091 Opc = ARM64ISD::ADDS;
1095 Opc = ARM64ISD::ADDS;
1099 Opc = ARM64ISD::SUBS;
1103 Opc = ARM64ISD::SUBS;
1106 // Multiply needs a little bit extra work.
1110 bool IsSigned = (Op.getOpcode() == ISD::SMULO) ? true : false;
1111 if (Op.getValueType() == MVT::i32) {
1112 unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1113 // For a 32 bit multiply with overflow check we want the instruction
1114 // selector to generate a widening multiply (SMADDL/UMADDL). For that we
1115 // need to generate the following pattern:
1116 // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b))
1117 LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
1118 RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
1119 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1120 SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul,
1121 DAG.getConstant(0, MVT::i64));
1122 // On ARM64 the upper 32 bits are always zero extended for a 32 bit
1123 // operation. We need to clear out the upper 32 bits, because we used a
1124 // widening multiply that wrote all 64 bits. In the end this should be a
1126 Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add);
1128 // The signed overflow check requires more than just a simple check for
1129 // any bit set in the upper 32 bits of the result. These bits could be
1130 // just the sign bits of a negative number. To perform the overflow
1131 // check we have to arithmetic shift right the 32nd bit of the result by
1132 // 31 bits. Then we compare the result to the upper 32 bits.
1133 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add,
1134 DAG.getConstant(32, MVT::i64));
1135 UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits);
1136 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value,
1137 DAG.getConstant(31, MVT::i64));
1138 // It is important that LowerBits is last, otherwise the arithmetic
1139 // shift will not be folded into the compare (SUBS).
1140 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32);
1141 Overflow = DAG.getNode(ARM64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1144 // The overflow check for unsigned multiply is easy. We only need to
1145 // check if any of the upper 32 bits are set. This can be done with a
1146 // CMP (shifted register). For that we need to generate the following
1148 // (i64 ARM64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32)
1149 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
1150 DAG.getConstant(32, MVT::i64));
1151 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1153 DAG.getNode(ARM64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1154 UpperBits).getValue(1);
1158 assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
1159 // For the 64 bit multiply
1160 Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1162 SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
1163 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
1164 DAG.getConstant(63, MVT::i64));
1165 // It is important that LowerBits is last, otherwise the arithmetic
1166 // shift will not be folded into the compare (SUBS).
1167 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1168 Overflow = DAG.getNode(ARM64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1171 SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
1172 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1174 DAG.getNode(ARM64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1175 UpperBits).getValue(1);
1182 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);
1184 // Emit the ARM64 operation with overflow check.
1185 Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
1186 Overflow = Value.getValue(1);
1188 return std::make_pair(Value, Overflow);
1191 SDValue ARM64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG,
1192 RTLIB::Libcall Call) const {
1193 SmallVector<SDValue, 2> Ops;
1194 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1195 Ops.push_back(Op.getOperand(i));
1197 return makeLibCall(DAG, Call, MVT::f128, &Ops[0], Ops.size(), false,
1201 static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) {
1202 SDValue Sel = Op.getOperand(0);
1203 SDValue Other = Op.getOperand(1);
1205 // If neither operand is a SELECT_CC, give up.
1206 if (Sel.getOpcode() != ISD::SELECT_CC)
1207 std::swap(Sel, Other);
1208 if (Sel.getOpcode() != ISD::SELECT_CC)
1211 // The folding we want to perform is:
1212 // (xor x, (select_cc a, b, cc, 0, -1) )
1214 // (csel x, (xor x, -1), cc ...)
1216 // The latter will get matched to a CSINV instruction.
1218 ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
1219 SDValue LHS = Sel.getOperand(0);
1220 SDValue RHS = Sel.getOperand(1);
1221 SDValue TVal = Sel.getOperand(2);
1222 SDValue FVal = Sel.getOperand(3);
1225 // FIXME: This could be generalized to non-integer comparisons.
1226 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
1229 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
1230 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
1232 // The the values aren't constants, this isn't the pattern we're looking for.
1233 if (!CFVal || !CTVal)
1236 // We can commute the SELECT_CC by inverting the condition. This
1237 // might be needed to make this fit into a CSINV pattern.
1238 if (CTVal->isAllOnesValue() && CFVal->isNullValue()) {
1239 std::swap(TVal, FVal);
1240 std::swap(CTVal, CFVal);
1241 CC = ISD::getSetCCInverse(CC, true);
1244 // If the constants line up, perform the transform!
1245 if (CTVal->isNullValue() && CFVal->isAllOnesValue()) {
1247 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
1250 TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other,
1251 DAG.getConstant(-1ULL, Other.getValueType()));
1253 return DAG.getNode(ARM64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal,
1260 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
1261 EVT VT = Op.getValueType();
1263 // Let legalize expand this if it isn't a legal type yet.
1264 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
1267 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
1270 bool ExtraOp = false;
1271 switch (Op.getOpcode()) {
1273 assert(0 && "Invalid code");
1275 Opc = ARM64ISD::ADDS;
1278 Opc = ARM64ISD::SUBS;
1281 Opc = ARM64ISD::ADCS;
1285 Opc = ARM64ISD::SBCS;
1291 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1));
1292 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1),
1296 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
1297 // Let legalize expand this if it isn't a legal type yet.
1298 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
1301 ARM64CC::CondCode CC;
1302 // The actual operation that sets the overflow or carry flag.
1303 SDValue Value, Overflow;
1304 std::tie(Value, Overflow) = getARM64XALUOOp(CC, Op, DAG);
1306 // We use 0 and 1 as false and true values.
1307 SDValue TVal = DAG.getConstant(1, MVT::i32);
1308 SDValue FVal = DAG.getConstant(0, MVT::i32);
1310 // We use an inverted condition, because the conditional select is inverted
1311 // too. This will allow it to be selected to a single instruction:
1312 // CSINC Wd, WZR, WZR, invert(cond).
1313 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), MVT::i32);
1314 Overflow = DAG.getNode(ARM64ISD::CSEL, SDLoc(Op), MVT::i32, FVal, TVal, CCVal,
1317 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
1318 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
1321 // Prefetch operands are:
1322 // 1: Address to prefetch
1324 // 3: int locality (0 = no locality ... 3 = extreme locality)
1325 // 4: bool isDataCache
1326 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
1328 unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
1329 unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
1330 // The data thing is not used.
1331 // unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
1333 bool IsStream = !Locality;
1334 // When the locality number is set
1336 // The front-end should have filtered out the out-of-range values
1337 assert(Locality <= 3 && "Prefetch locality out-of-range");
1338 // The locality degree is the opposite of the cache speed.
1339 // Put the number the other way around.
1340 // The encoding starts at 0 for level 1
1341 Locality = 3 - Locality;
1344 // built the mask value encoding the expected behavior.
1345 unsigned PrfOp = (IsWrite << 4) | // Load/Store bit
1346 (Locality << 1) | // Cache level bits
1347 (unsigned)IsStream; // Stream bit
1348 return DAG.getNode(ARM64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
1349 DAG.getConstant(PrfOp, MVT::i32), Op.getOperand(1));
1352 SDValue ARM64TargetLowering::LowerFP_EXTEND(SDValue Op,
1353 SelectionDAG &DAG) const {
1354 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
1357 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
1359 return LowerF128Call(Op, DAG, LC);
1362 SDValue ARM64TargetLowering::LowerFP_ROUND(SDValue Op,
1363 SelectionDAG &DAG) const {
1364 if (Op.getOperand(0).getValueType() != MVT::f128) {
1365 // It's legal except when f128 is involved
1370 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
1372 // FP_ROUND node has a second operand indicating whether it is known to be
1373 // precise. That doesn't take part in the LibCall so we can't directly use
1375 SDValue SrcVal = Op.getOperand(0);
1376 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
1377 /*isSigned*/ false, SDLoc(Op)).first;
1380 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1381 // Warning: We maintain cost tables in ARM64TargetTransformInfo.cpp.
1382 // Any additional optimization in this function should be recorded
1383 // in the cost tables.
1384 EVT InVT = Op.getOperand(0).getValueType();
1385 EVT VT = Op.getValueType();
1387 // FP_TO_XINT conversion from the same type are legal.
1388 if (VT.getSizeInBits() == InVT.getSizeInBits())
1391 if (InVT == MVT::v2f64 || InVT == MVT::v4f32) {
1394 DAG.getNode(Op.getOpcode(), dl, InVT.changeVectorElementTypeToInteger(),
1396 return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
1397 } else if (InVT == MVT::v2f32) {
1399 SDValue Ext = DAG.getNode(ISD::FP_EXTEND, dl, MVT::v2f64, Op.getOperand(0));
1400 return DAG.getNode(Op.getOpcode(), dl, VT, Ext);
1403 // Type changing conversions are illegal.
1407 SDValue ARM64TargetLowering::LowerFP_TO_INT(SDValue Op,
1408 SelectionDAG &DAG) const {
1409 if (Op.getOperand(0).getValueType().isVector())
1410 return LowerVectorFP_TO_INT(Op, DAG);
1412 if (Op.getOperand(0).getValueType() != MVT::f128) {
1413 // It's legal except when f128 is involved
1418 if (Op.getOpcode() == ISD::FP_TO_SINT)
1419 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
1421 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
1423 SmallVector<SDValue, 2> Ops;
1424 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1425 Ops.push_back(Op.getOperand(i));
1427 return makeLibCall(DAG, LC, Op.getValueType(), &Ops[0], Ops.size(), false,
1431 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1432 // Warning: We maintain cost tables in ARM64TargetTransformInfo.cpp.
1433 // Any additional optimization in this function should be recorded
1434 // in the cost tables.
1435 EVT VT = Op.getValueType();
1437 SDValue In = Op.getOperand(0);
1438 EVT InVT = In.getValueType();
1440 // v2i32 to v2f32 is legal.
1441 if (VT == MVT::v2f32 && InVT == MVT::v2i32)
1444 // This function only handles v2f64 outputs.
1445 if (VT == MVT::v2f64) {
1446 // Extend the input argument to a v2i64 that we can feed into the
1447 // floating point conversion. Zero or sign extend based on whether
1448 // we're doing a signed or unsigned float conversion.
1450 Op.getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
1451 assert(Op.getNumOperands() == 1 && "FP conversions take one argument");
1452 SDValue Promoted = DAG.getNode(Opc, dl, MVT::v2i64, Op.getOperand(0));
1453 return DAG.getNode(Op.getOpcode(), dl, Op.getValueType(), Promoted);
1456 // Scalarize v2i64 to v2f32 conversions.
1457 std::vector<SDValue> BuildVectorOps;
1458 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
1459 SDValue Sclr = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, In,
1460 DAG.getConstant(i, MVT::i64));
1461 Sclr = DAG.getNode(Op->getOpcode(), dl, MVT::f32, Sclr);
1462 BuildVectorOps.push_back(Sclr);
1465 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, BuildVectorOps);
1468 SDValue ARM64TargetLowering::LowerINT_TO_FP(SDValue Op,
1469 SelectionDAG &DAG) const {
1470 if (Op.getValueType().isVector())
1471 return LowerVectorINT_TO_FP(Op, DAG);
1473 // i128 conversions are libcalls.
1474 if (Op.getOperand(0).getValueType() == MVT::i128)
1477 // Other conversions are legal, unless it's to the completely software-based
1479 if (Op.getValueType() != MVT::f128)
1483 if (Op.getOpcode() == ISD::SINT_TO_FP)
1484 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1486 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1488 return LowerF128Call(Op, DAG, LC);
1491 SDValue ARM64TargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
1492 // For iOS, we want to call an alternative entry point: __sincos_stret,
1493 // which returns the values in two S / D registers.
1495 SDValue Arg = Op.getOperand(0);
1496 EVT ArgVT = Arg.getValueType();
1497 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1504 Entry.isSExt = false;
1505 Entry.isZExt = false;
1506 Args.push_back(Entry);
1508 const char *LibcallName =
1509 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
1510 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
1512 StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
1513 TargetLowering::CallLoweringInfo CLI(
1514 DAG.getEntryNode(), RetTy, false, false, false, false, 0,
1515 CallingConv::Fast, /*isTaillCall=*/false,
1516 /*doesNotRet=*/false, /*isReturnValueUsed*/ true, Callee, Args, DAG, dl);
1517 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1518 return CallResult.first;
1521 SDValue ARM64TargetLowering::LowerOperation(SDValue Op,
1522 SelectionDAG &DAG) const {
1523 switch (Op.getOpcode()) {
1525 llvm_unreachable("unimplemented operand");
1527 case ISD::GlobalAddress:
1528 return LowerGlobalAddress(Op, DAG);
1529 case ISD::GlobalTLSAddress:
1530 return LowerGlobalTLSAddress(Op, DAG);
1532 return LowerSETCC(Op, DAG);
1534 return LowerBR_CC(Op, DAG);
1536 return LowerSELECT(Op, DAG);
1537 case ISD::SELECT_CC:
1538 return LowerSELECT_CC(Op, DAG);
1539 case ISD::JumpTable:
1540 return LowerJumpTable(Op, DAG);
1541 case ISD::ConstantPool:
1542 return LowerConstantPool(Op, DAG);
1543 case ISD::BlockAddress:
1544 return LowerBlockAddress(Op, DAG);
1546 return LowerVASTART(Op, DAG);
1548 return LowerVACOPY(Op, DAG);
1550 return LowerVAARG(Op, DAG);
1555 return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
1562 return LowerXALUO(Op, DAG);
1564 return LowerF128Call(Op, DAG, RTLIB::ADD_F128);
1566 return LowerF128Call(Op, DAG, RTLIB::SUB_F128);
1568 return LowerF128Call(Op, DAG, RTLIB::MUL_F128);
1570 return LowerF128Call(Op, DAG, RTLIB::DIV_F128);
1572 return LowerFP_ROUND(Op, DAG);
1573 case ISD::FP_EXTEND:
1574 return LowerFP_EXTEND(Op, DAG);
1575 case ISD::FRAMEADDR:
1576 return LowerFRAMEADDR(Op, DAG);
1577 case ISD::RETURNADDR:
1578 return LowerRETURNADDR(Op, DAG);
1579 case ISD::INSERT_VECTOR_ELT:
1580 return LowerINSERT_VECTOR_ELT(Op, DAG);
1581 case ISD::EXTRACT_VECTOR_ELT:
1582 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
1583 case ISD::BUILD_VECTOR:
1584 return LowerBUILD_VECTOR(Op, DAG);
1585 case ISD::VECTOR_SHUFFLE:
1586 return LowerVECTOR_SHUFFLE(Op, DAG);
1587 case ISD::EXTRACT_SUBVECTOR:
1588 return LowerEXTRACT_SUBVECTOR(Op, DAG);
1592 return LowerVectorSRA_SRL_SHL(Op, DAG);
1593 case ISD::SHL_PARTS:
1594 return LowerShiftLeftParts(Op, DAG);
1595 case ISD::SRL_PARTS:
1596 case ISD::SRA_PARTS:
1597 return LowerShiftRightParts(Op, DAG);
1599 return LowerCTPOP(Op, DAG);
1600 case ISD::FCOPYSIGN:
1601 return LowerFCOPYSIGN(Op, DAG);
1603 return LowerVectorAND(Op, DAG);
1605 return LowerVectorOR(Op, DAG);
1607 return LowerXOR(Op, DAG);
1609 return LowerPREFETCH(Op, DAG);
1610 case ISD::SINT_TO_FP:
1611 case ISD::UINT_TO_FP:
1612 return LowerINT_TO_FP(Op, DAG);
1613 case ISD::FP_TO_SINT:
1614 case ISD::FP_TO_UINT:
1615 return LowerFP_TO_INT(Op, DAG);
1617 return LowerFSINCOS(Op, DAG);
1621 /// getFunctionAlignment - Return the Log2 alignment of this function.
1622 unsigned ARM64TargetLowering::getFunctionAlignment(const Function *F) const {
1626 //===----------------------------------------------------------------------===//
1627 // Calling Convention Implementation
1628 //===----------------------------------------------------------------------===//
1630 #include "ARM64GenCallingConv.inc"
1632 /// Selects the correct CCAssignFn for a the given CallingConvention
1634 CCAssignFn *ARM64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
1635 bool IsVarArg) const {
1638 llvm_unreachable("Unsupported calling convention.");
1639 case CallingConv::WebKit_JS:
1640 return CC_ARM64_WebKit_JS;
1641 case CallingConv::C:
1642 case CallingConv::Fast:
1643 if (!Subtarget->isTargetDarwin())
1644 return CC_ARM64_AAPCS;
1645 return IsVarArg ? CC_ARM64_DarwinPCS_VarArg : CC_ARM64_DarwinPCS;
1649 SDValue ARM64TargetLowering::LowerFormalArguments(
1650 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1651 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
1652 SmallVectorImpl<SDValue> &InVals) const {
1653 MachineFunction &MF = DAG.getMachineFunction();
1654 MachineFrameInfo *MFI = MF.getFrameInfo();
1656 // Assign locations to all of the incoming arguments.
1657 SmallVector<CCValAssign, 16> ArgLocs;
1658 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1659 getTargetMachine(), ArgLocs, *DAG.getContext());
1661 // At this point, Ins[].VT may already be promoted to i32. To correctly
1662 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
1663 // i8 to CC_ARM64_AAPCS with i32 being ValVT and i8 being LocVT.
1664 // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
1665 // we use a special version of AnalyzeFormalArguments to pass in ValVT and
1667 unsigned NumArgs = Ins.size();
1668 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
1669 unsigned CurArgIdx = 0;
1670 for (unsigned i = 0; i != NumArgs; ++i) {
1671 MVT ValVT = Ins[i].VT;
1672 std::advance(CurOrigArg, Ins[i].OrigArgIndex - CurArgIdx);
1673 CurArgIdx = Ins[i].OrigArgIndex;
1675 // Get type of the original argument.
1676 EVT ActualVT = getValueType(CurOrigArg->getType(), /*AllowUnknown*/ true);
1677 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
1678 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
1680 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
1682 else if (ActualMVT == MVT::i16)
1685 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
1687 AssignFn(i, ValVT, LocVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
1688 assert(!Res && "Call operand has unhandled type");
1691 assert(ArgLocs.size() == Ins.size());
1692 SmallVector<SDValue, 16> ArgValues;
1693 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1694 CCValAssign &VA = ArgLocs[i];
1696 if (Ins[i].Flags.isByVal()) {
1697 // Byval is used for HFAs in the PCS, but the system should work in a
1698 // non-compliant manner for larger structs.
1699 EVT PtrTy = getPointerTy();
1700 int Size = Ins[i].Flags.getByValSize();
1701 unsigned NumRegs = (Size + 7) / 8;
1703 // FIXME: This works on big-endian for composite byvals, which are the common
1704 // case. It should also work for fundamental types too.
1706 MFI->CreateFixedObject(8 * NumRegs, VA.getLocMemOffset(), false);
1707 SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrTy);
1708 InVals.push_back(FrameIdxN);
1711 } if (VA.isRegLoc()) {
1712 // Arguments stored in registers.
1713 EVT RegVT = VA.getLocVT();
1716 const TargetRegisterClass *RC;
1718 if (RegVT == MVT::i32)
1719 RC = &ARM64::GPR32RegClass;
1720 else if (RegVT == MVT::i64)
1721 RC = &ARM64::GPR64RegClass;
1722 else if (RegVT == MVT::f32)
1723 RC = &ARM64::FPR32RegClass;
1724 else if (RegVT == MVT::f64 || RegVT.is64BitVector())
1725 RC = &ARM64::FPR64RegClass;
1726 else if (RegVT == MVT::f128 || RegVT.is128BitVector())
1727 RC = &ARM64::FPR128RegClass;
1729 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1731 // Transform the arguments in physical registers into virtual ones.
1732 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1733 ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
1735 // If this is an 8, 16 or 32-bit value, it is really passed promoted
1736 // to 64 bits. Insert an assert[sz]ext to capture this, then
1737 // truncate to the right size.
1738 switch (VA.getLocInfo()) {
1740 llvm_unreachable("Unknown loc info!");
1741 case CCValAssign::Full:
1743 case CCValAssign::BCvt:
1744 ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue);
1746 case CCValAssign::SExt:
1747 ArgValue = DAG.getNode(ISD::AssertSext, DL, RegVT, ArgValue,
1748 DAG.getValueType(VA.getValVT()));
1749 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), ArgValue);
1751 case CCValAssign::ZExt:
1752 ArgValue = DAG.getNode(ISD::AssertZext, DL, RegVT, ArgValue,
1753 DAG.getValueType(VA.getValVT()));
1754 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), ArgValue);
1758 InVals.push_back(ArgValue);
1760 } else { // VA.isRegLoc()
1761 assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
1762 unsigned ArgOffset = VA.getLocMemOffset();
1763 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1765 uint32_t BEAlign = 0;
1766 if (ArgSize < 8 && !Subtarget->isLittleEndian())
1767 BEAlign = 8 - ArgSize;
1769 int FI = MFI->CreateFixedObject(ArgSize, ArgOffset + BEAlign, true);
1771 // Create load nodes to retrieve arguments from the stack.
1772 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1775 // If the loc type and val type are not the same, create an anyext load.
1776 if (VA.getLocVT().getSizeInBits() != VA.getValVT().getSizeInBits()) {
1777 // We should only get here if this is a pure integer.
1778 assert(!VA.getValVT().isVector() && VA.getValVT().isInteger() &&
1779 "Only integer extension supported!");
1780 ArgValue = DAG.getExtLoad(ISD::EXTLOAD, DL, VA.getValVT(), Chain, FIN,
1781 MachinePointerInfo::getFixedStack(FI),
1783 false, false, false, 0);
1785 ArgValue = DAG.getLoad(VA.getValVT(), DL, Chain, FIN,
1786 MachinePointerInfo::getFixedStack(FI), false,
1790 InVals.push_back(ArgValue);
1796 if (!Subtarget->isTargetDarwin()) {
1797 // The AAPCS variadic function ABI is identical to the non-variadic
1798 // one. As a result there may be more arguments in registers and we should
1799 // save them for future reference.
1800 saveVarArgRegisters(CCInfo, DAG, DL, Chain);
1803 ARM64FunctionInfo *AFI = MF.getInfo<ARM64FunctionInfo>();
1804 // This will point to the next argument passed via stack.
1805 unsigned StackOffset = CCInfo.getNextStackOffset();
1806 // We currently pass all varargs at 8-byte alignment.
1807 StackOffset = ((StackOffset + 7) & ~7);
1808 AFI->setVarArgsStackIndex(MFI->CreateFixedObject(4, StackOffset, true));
1811 ARM64FunctionInfo *FuncInfo = MF.getInfo<ARM64FunctionInfo>();
1812 unsigned StackArgSize = CCInfo.getNextStackOffset();
1813 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
1814 if (DoesCalleeRestoreStack(CallConv, TailCallOpt)) {
1815 // This is a non-standard ABI so by fiat I say we're allowed to make full
1816 // use of the stack area to be popped, which must be aligned to 16 bytes in
1818 StackArgSize = RoundUpToAlignment(StackArgSize, 16);
1820 // If we're expected to restore the stack (e.g. fastcc) then we'll be adding
1821 // a multiple of 16.
1822 FuncInfo->setArgumentStackToRestore(StackArgSize);
1824 // This realignment carries over to the available bytes below. Our own
1825 // callers will guarantee the space is free by giving an aligned value to
1828 // Even if we're not expected to free up the space, it's useful to know how
1829 // much is there while considering tail calls (because we can reuse it).
1830 FuncInfo->setBytesInStackArgArea(StackArgSize);
1835 void ARM64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
1836 SelectionDAG &DAG, SDLoc DL,
1837 SDValue &Chain) const {
1838 MachineFunction &MF = DAG.getMachineFunction();
1839 MachineFrameInfo *MFI = MF.getFrameInfo();
1840 ARM64FunctionInfo *FuncInfo = MF.getInfo<ARM64FunctionInfo>();
1842 SmallVector<SDValue, 8> MemOps;
1844 static const MCPhysReg GPRArgRegs[] = { ARM64::X0, ARM64::X1, ARM64::X2,
1845 ARM64::X3, ARM64::X4, ARM64::X5,
1846 ARM64::X6, ARM64::X7 };
1847 static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs);
1848 unsigned FirstVariadicGPR =
1849 CCInfo.getFirstUnallocated(GPRArgRegs, NumGPRArgRegs);
1851 unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
1853 if (GPRSaveSize != 0) {
1854 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
1856 SDValue FIN = DAG.getFrameIndex(GPRIdx, getPointerTy());
1858 for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
1859 unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &ARM64::GPR64RegClass);
1860 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
1862 DAG.getStore(Val.getValue(1), DL, Val, FIN,
1863 MachinePointerInfo::getStack(i * 8), false, false, 0);
1864 MemOps.push_back(Store);
1865 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
1866 DAG.getConstant(8, getPointerTy()));
1869 FuncInfo->setVarArgsGPRIndex(GPRIdx);
1870 FuncInfo->setVarArgsGPRSize(GPRSaveSize);
1872 if (Subtarget->hasFPARMv8()) {
1873 static const MCPhysReg FPRArgRegs[] = { ARM64::Q0, ARM64::Q1, ARM64::Q2,
1874 ARM64::Q3, ARM64::Q4, ARM64::Q5,
1875 ARM64::Q6, ARM64::Q7 };
1876 static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs);
1877 unsigned FirstVariadicFPR =
1878 CCInfo.getFirstUnallocated(FPRArgRegs, NumFPRArgRegs);
1880 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
1882 if (FPRSaveSize != 0) {
1883 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
1885 SDValue FIN = DAG.getFrameIndex(FPRIdx, getPointerTy());
1887 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
1888 unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &ARM64::FPR128RegClass);
1889 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f128);
1892 DAG.getStore(Val.getValue(1), DL, Val, FIN,
1893 MachinePointerInfo::getStack(i * 16), false, false, 0);
1894 MemOps.push_back(Store);
1895 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
1896 DAG.getConstant(16, getPointerTy()));
1899 FuncInfo->setVarArgsFPRIndex(FPRIdx);
1900 FuncInfo->setVarArgsFPRSize(FPRSaveSize);
1903 if (!MemOps.empty()) {
1904 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
1908 /// LowerCallResult - Lower the result values of a call into the
1909 /// appropriate copies out of appropriate physical registers.
1910 SDValue ARM64TargetLowering::LowerCallResult(
1911 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
1912 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
1913 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
1914 SDValue ThisVal) const {
1915 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
1916 : RetCC_ARM64_AAPCS;
1917 // Assign locations to each value returned by this call.
1918 SmallVector<CCValAssign, 16> RVLocs;
1919 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1920 getTargetMachine(), RVLocs, *DAG.getContext());
1921 CCInfo.AnalyzeCallResult(Ins, RetCC);
1923 // Copy all of the result registers out of their specified physreg.
1924 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1925 CCValAssign VA = RVLocs[i];
1927 // Pass 'this' value directly from the argument to return value, to avoid
1928 // reg unit interference
1929 if (i == 0 && isThisReturn) {
1930 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 &&
1931 "unexpected return calling convention register assignment");
1932 InVals.push_back(ThisVal);
1937 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag);
1938 Chain = Val.getValue(1);
1939 InFlag = Val.getValue(2);
1941 switch (VA.getLocInfo()) {
1943 llvm_unreachable("Unknown loc info!");
1944 case CCValAssign::Full:
1946 case CCValAssign::BCvt:
1947 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
1951 InVals.push_back(Val);
1957 bool ARM64TargetLowering::isEligibleForTailCallOptimization(
1958 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
1959 bool isCalleeStructRet, bool isCallerStructRet,
1960 const SmallVectorImpl<ISD::OutputArg> &Outs,
1961 const SmallVectorImpl<SDValue> &OutVals,
1962 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
1963 // For CallingConv::C this function knows whether the ABI needs
1964 // changing. That's not true for other conventions so they will have to opt in
1966 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
1969 const MachineFunction &MF = DAG.getMachineFunction();
1970 const Function *CallerF = MF.getFunction();
1971 CallingConv::ID CallerCC = CallerF->getCallingConv();
1972 bool CCMatch = CallerCC == CalleeCC;
1974 // Byval parameters hand the function a pointer directly into the stack area
1975 // we want to reuse during a tail call. Working around this *is* possible (see
1976 // X86) but less efficient and uglier in LowerCall.
1977 for (Function::const_arg_iterator i = CallerF->arg_begin(),
1978 e = CallerF->arg_end();
1980 if (i->hasByValAttr())
1983 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
1984 if (IsTailCallConvention(CalleeCC) && CCMatch)
1989 // Now we search for cases where we can use a tail call without changing the
1990 // ABI. Sibcall is used in some places (particularly gcc) to refer to this
1993 // I want anyone implementing a new calling convention to think long and hard
1994 // about this assert.
1995 assert((!isVarArg || CalleeCC == CallingConv::C) &&
1996 "Unexpected variadic calling convention");
1998 if (isVarArg && !Outs.empty()) {
1999 // At least two cases here: if caller is fastcc then we can't have any
2000 // memory arguments (we'd be expected to clean up the stack afterwards). If
2001 // caller is C then we could potentially use its argument area.
2003 // FIXME: for now we take the most conservative of these in both cases:
2004 // disallow all variadic memory operands.
2005 SmallVector<CCValAssign, 16> ArgLocs;
2006 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2007 getTargetMachine(), ArgLocs, *DAG.getContext());
2009 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, true));
2010 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
2011 if (!ArgLocs[i].isRegLoc())
2015 // If the calling conventions do not match, then we'd better make sure the
2016 // results are returned in the same way as what the caller expects.
2018 SmallVector<CCValAssign, 16> RVLocs1;
2019 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
2020 getTargetMachine(), RVLocs1, *DAG.getContext());
2021 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForCall(CalleeCC, isVarArg));
2023 SmallVector<CCValAssign, 16> RVLocs2;
2024 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
2025 getTargetMachine(), RVLocs2, *DAG.getContext());
2026 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForCall(CallerCC, isVarArg));
2028 if (RVLocs1.size() != RVLocs2.size())
2030 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2031 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2033 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2035 if (RVLocs1[i].isRegLoc()) {
2036 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2039 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2045 // Nothing more to check if the callee is taking no arguments
2049 SmallVector<CCValAssign, 16> ArgLocs;
2050 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2051 getTargetMachine(), ArgLocs, *DAG.getContext());
2053 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, isVarArg));
2055 const ARM64FunctionInfo *FuncInfo = MF.getInfo<ARM64FunctionInfo>();
2057 // If the stack arguments for this call would fit into our own save area then
2058 // the call can be made tail.
2059 return CCInfo.getNextStackOffset() <= FuncInfo->getBytesInStackArgArea();
2062 SDValue ARM64TargetLowering::addTokenForArgument(SDValue Chain,
2064 MachineFrameInfo *MFI,
2065 int ClobberedFI) const {
2066 SmallVector<SDValue, 8> ArgChains;
2067 int64_t FirstByte = MFI->getObjectOffset(ClobberedFI);
2068 int64_t LastByte = FirstByte + MFI->getObjectSize(ClobberedFI) - 1;
2070 // Include the original chain at the beginning of the list. When this is
2071 // used by target LowerCall hooks, this helps legalize find the
2072 // CALLSEQ_BEGIN node.
2073 ArgChains.push_back(Chain);
2075 // Add a chain value for each stack argument corresponding
2076 for (SDNode::use_iterator U = DAG.getEntryNode().getNode()->use_begin(),
2077 UE = DAG.getEntryNode().getNode()->use_end();
2079 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
2080 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
2081 if (FI->getIndex() < 0) {
2082 int64_t InFirstByte = MFI->getObjectOffset(FI->getIndex());
2083 int64_t InLastByte = InFirstByte;
2084 InLastByte += MFI->getObjectSize(FI->getIndex()) - 1;
2086 if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
2087 (FirstByte <= InFirstByte && InFirstByte <= LastByte))
2088 ArgChains.push_back(SDValue(L, 1));
2091 // Build a tokenfactor for all the chains.
2092 return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
2095 bool ARM64TargetLowering::DoesCalleeRestoreStack(CallingConv::ID CallCC,
2096 bool TailCallOpt) const {
2097 return CallCC == CallingConv::Fast && TailCallOpt;
2100 bool ARM64TargetLowering::IsTailCallConvention(CallingConv::ID CallCC) const {
2101 return CallCC == CallingConv::Fast;
2104 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
2105 /// and add input and output parameter nodes.
2106 SDValue ARM64TargetLowering::LowerCall(CallLoweringInfo &CLI,
2107 SmallVectorImpl<SDValue> &InVals) const {
2108 SelectionDAG &DAG = CLI.DAG;
2110 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2111 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2112 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2113 SDValue Chain = CLI.Chain;
2114 SDValue Callee = CLI.Callee;
2115 bool &IsTailCall = CLI.IsTailCall;
2116 CallingConv::ID CallConv = CLI.CallConv;
2117 bool IsVarArg = CLI.IsVarArg;
2119 MachineFunction &MF = DAG.getMachineFunction();
2120 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
2121 bool IsThisReturn = false;
2123 ARM64FunctionInfo *FuncInfo = MF.getInfo<ARM64FunctionInfo>();
2124 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
2125 bool IsSibCall = false;
2128 // Check if it's really possible to do a tail call.
2129 IsTailCall = isEligibleForTailCallOptimization(
2130 Callee, CallConv, IsVarArg, IsStructRet,
2131 MF.getFunction()->hasStructRetAttr(), Outs, OutVals, Ins, DAG);
2132 if (!IsTailCall && CLI.CS && CLI.CS->isMustTailCall())
2133 report_fatal_error("failed to perform tail call elimination on a call "
2134 "site marked musttail");
2136 // A sibling call is one where we're under the usual C ABI and not planning
2137 // to change that but can still do a tail call:
2138 if (!TailCallOpt && IsTailCall)
2145 // Analyze operands of the call, assigning locations to each operand.
2146 SmallVector<CCValAssign, 16> ArgLocs;
2147 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2148 getTargetMachine(), ArgLocs, *DAG.getContext());
2151 // Handle fixed and variable vector arguments differently.
2152 // Variable vector arguments always go into memory.
2153 unsigned NumArgs = Outs.size();
2155 for (unsigned i = 0; i != NumArgs; ++i) {
2156 MVT ArgVT = Outs[i].VT;
2157 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2158 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
2159 /*IsVarArg=*/ !Outs[i].IsFixed);
2160 bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
2161 assert(!Res && "Call operand has unhandled type");
2165 // At this point, Outs[].VT may already be promoted to i32. To correctly
2166 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2167 // i8 to CC_ARM64_AAPCS with i32 being ValVT and i8 being LocVT.
2168 // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here
2169 // we use a special version of AnalyzeCallOperands to pass in ValVT and
2171 unsigned NumArgs = Outs.size();
2172 for (unsigned i = 0; i != NumArgs; ++i) {
2173 MVT ValVT = Outs[i].VT;
2174 // Get type of the original argument.
2175 EVT ActualVT = getValueType(CLI.Args[Outs[i].OrigArgIndex].Ty,
2176 /*AllowUnknown*/ true);
2177 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT;
2178 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2179 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2181 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2183 else if (ActualMVT == MVT::i16)
2186 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2187 bool Res = AssignFn(i, ValVT, LocVT, CCValAssign::Full, ArgFlags, CCInfo);
2188 assert(!Res && "Call operand has unhandled type");
2193 // Get a count of how many bytes are to be pushed on the stack.
2194 unsigned NumBytes = CCInfo.getNextStackOffset();
2197 // Since we're not changing the ABI to make this a tail call, the memory
2198 // operands are already available in the caller's incoming argument space.
2202 // FPDiff is the byte offset of the call's argument area from the callee's.
2203 // Stores to callee stack arguments will be placed in FixedStackSlots offset
2204 // by this amount for a tail call. In a sibling call it must be 0 because the
2205 // caller will deallocate the entire stack and the callee still expects its
2206 // arguments to begin at SP+0. Completely unused for non-tail calls.
2209 if (IsTailCall && !IsSibCall) {
2210 unsigned NumReusableBytes = FuncInfo->getBytesInStackArgArea();
2212 // Since callee will pop argument stack as a tail call, we must keep the
2213 // popped size 16-byte aligned.
2214 NumBytes = RoundUpToAlignment(NumBytes, 16);
2216 // FPDiff will be negative if this tail call requires more space than we
2217 // would automatically have in our incoming argument space. Positive if we
2218 // can actually shrink the stack.
2219 FPDiff = NumReusableBytes - NumBytes;
2221 // The stack pointer must be 16-byte aligned at all times it's used for a
2222 // memory operation, which in practice means at *all* times and in
2223 // particular across call boundaries. Therefore our own arguments started at
2224 // a 16-byte aligned SP and the delta applied for the tail call should
2225 // satisfy the same constraint.
2226 assert(FPDiff % 16 == 0 && "unaligned stack on tail call");
2229 // Adjust the stack pointer for the new arguments...
2230 // These operations are automatically eliminated by the prolog/epilog pass
2233 DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true), DL);
2235 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, ARM64::SP, getPointerTy());
2237 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2238 SmallVector<SDValue, 8> MemOpChains;
2240 // Walk the register/memloc assignments, inserting copies/loads.
2241 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2242 ++i, ++realArgIdx) {
2243 CCValAssign &VA = ArgLocs[i];
2244 SDValue Arg = OutVals[realArgIdx];
2245 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2247 // Promote the value if needed.
2248 switch (VA.getLocInfo()) {
2250 llvm_unreachable("Unknown loc info!");
2251 case CCValAssign::Full:
2253 case CCValAssign::SExt:
2254 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
2256 case CCValAssign::ZExt:
2257 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2259 case CCValAssign::AExt:
2260 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
2262 case CCValAssign::BCvt:
2263 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2265 case CCValAssign::FPExt:
2266 Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
2270 if (VA.isRegLoc()) {
2271 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i64) {
2272 assert(VA.getLocVT() == MVT::i64 &&
2273 "unexpected calling convention register assignment");
2274 assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
2275 "unexpected use of 'returned'");
2276 IsThisReturn = true;
2278 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2280 assert(VA.isMemLoc());
2283 MachinePointerInfo DstInfo;
2285 // FIXME: This works on big-endian for composite byvals, which are the
2286 // common case. It should also work for fundamental types too.
2287 uint32_t BEAlign = 0;
2288 unsigned OpSize = Flags.isByVal() ? Flags.getByValSize() * 8
2289 : VA.getLocVT().getSizeInBits();
2290 OpSize = (OpSize + 7) / 8;
2291 if (!Subtarget->isLittleEndian() && !Flags.isByVal()) {
2293 BEAlign = 8 - OpSize;
2295 unsigned LocMemOffset = VA.getLocMemOffset();
2296 int32_t Offset = LocMemOffset + BEAlign;
2297 SDValue PtrOff = DAG.getIntPtrConstant(Offset);
2298 PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff);
2301 Offset = Offset + FPDiff;
2302 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2304 DstAddr = DAG.getFrameIndex(FI, getPointerTy());
2305 DstInfo = MachinePointerInfo::getFixedStack(FI);
2307 // Make sure any stack arguments overlapping with where we're storing
2308 // are loaded before this eventual operation. Otherwise they'll be
2310 Chain = addTokenForArgument(Chain, DAG, MF.getFrameInfo(), FI);
2312 SDValue PtrOff = DAG.getIntPtrConstant(Offset);
2314 DstAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff);
2315 DstInfo = MachinePointerInfo::getStack(LocMemOffset);
2318 if (Outs[i].Flags.isByVal()) {
2320 DAG.getConstant(Outs[i].Flags.getByValSize(), MVT::i64);
2321 SDValue Cpy = DAG.getMemcpy(
2322 Chain, DL, DstAddr, Arg, SizeNode, Outs[i].Flags.getByValAlign(),
2323 /*isVolatile = */ false,
2324 /*alwaysInline = */ false, DstInfo, MachinePointerInfo());
2326 MemOpChains.push_back(Cpy);
2328 // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
2329 // promoted to a legal register type i32, we should truncate Arg back to
2331 if (Arg.getValueType().isSimple() &&
2332 Arg.getValueType().getSimpleVT() == MVT::i32 &&
2333 (VA.getLocVT() == MVT::i1 || VA.getLocVT() == MVT::i8 ||
2334 VA.getLocVT() == MVT::i16))
2335 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getLocVT(), Arg);
2338 DAG.getStore(Chain, DL, Arg, DstAddr, DstInfo, false, false, 0);
2339 MemOpChains.push_back(Store);
2344 if (!MemOpChains.empty())
2345 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
2347 // Build a sequence of copy-to-reg nodes chained together with token chain
2348 // and flag operands which copy the outgoing args into the appropriate regs.
2350 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2351 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[i].first,
2352 RegsToPass[i].second, InFlag);
2353 InFlag = Chain.getValue(1);
2356 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2357 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2358 // node so that legalize doesn't hack it.
2359 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
2360 Subtarget->isTargetMachO()) {
2361 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2362 const GlobalValue *GV = G->getGlobal();
2363 bool InternalLinkage = GV->hasInternalLinkage();
2364 if (InternalLinkage)
2365 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2367 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0,
2369 Callee = DAG.getNode(ARM64ISD::LOADgot, DL, getPointerTy(), Callee);
2371 } else if (ExternalSymbolSDNode *S =
2372 dyn_cast<ExternalSymbolSDNode>(Callee)) {
2373 const char *Sym = S->getSymbol();
2375 DAG.getTargetExternalSymbol(Sym, getPointerTy(), ARM64II::MO_GOT);
2376 Callee = DAG.getNode(ARM64ISD::LOADgot, DL, getPointerTy(), Callee);
2378 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2379 const GlobalValue *GV = G->getGlobal();
2380 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2381 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2382 const char *Sym = S->getSymbol();
2383 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), 0);
2386 // We don't usually want to end the call-sequence here because we would tidy
2387 // the frame up *after* the call, however in the ABI-changing tail-call case
2388 // we've carefully laid out the parameters so that when sp is reset they'll be
2389 // in the correct location.
2390 if (IsTailCall && !IsSibCall) {
2391 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2392 DAG.getIntPtrConstant(0, true), InFlag, DL);
2393 InFlag = Chain.getValue(1);
2396 std::vector<SDValue> Ops;
2397 Ops.push_back(Chain);
2398 Ops.push_back(Callee);
2401 // Each tail call may have to adjust the stack by a different amount, so
2402 // this information must travel along with the operation for eventual
2403 // consumption by emitEpilogue.
2404 Ops.push_back(DAG.getTargetConstant(FPDiff, MVT::i32));
2407 // Add argument registers to the end of the list so that they are known live
2409 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2410 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2411 RegsToPass[i].second.getValueType()));
2413 // Add a register mask operand representing the call-preserved registers.
2414 const uint32_t *Mask;
2415 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2416 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2418 // For 'this' returns, use the X0-preserving mask if applicable
2419 Mask = ARI->getThisReturnPreservedMask(CallConv);
2421 IsThisReturn = false;
2422 Mask = ARI->getCallPreservedMask(CallConv);
2425 Mask = ARI->getCallPreservedMask(CallConv);
2427 assert(Mask && "Missing call preserved mask for calling convention");
2428 Ops.push_back(DAG.getRegisterMask(Mask));
2430 if (InFlag.getNode())
2431 Ops.push_back(InFlag);
2433 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2435 // If we're doing a tall call, use a TC_RETURN here rather than an
2436 // actual call instruction.
2438 return DAG.getNode(ARM64ISD::TC_RETURN, DL, NodeTys, Ops);
2440 // Returns a chain and a flag for retval copy to use.
2441 Chain = DAG.getNode(ARM64ISD::CALL, DL, NodeTys, Ops);
2442 InFlag = Chain.getValue(1);
2444 uint64_t CalleePopBytes = DoesCalleeRestoreStack(CallConv, TailCallOpt)
2445 ? RoundUpToAlignment(NumBytes, 16)
2448 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2449 DAG.getIntPtrConstant(CalleePopBytes, true),
2452 InFlag = Chain.getValue(1);
2454 // Handle result values, copying them out of physregs into vregs that we
2456 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
2457 InVals, IsThisReturn,
2458 IsThisReturn ? OutVals[0] : SDValue());
2461 bool ARM64TargetLowering::CanLowerReturn(
2462 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2463 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2464 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2465 : RetCC_ARM64_AAPCS;
2466 SmallVector<CCValAssign, 16> RVLocs;
2467 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
2468 return CCInfo.CheckReturn(Outs, RetCC);
2472 ARM64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2474 const SmallVectorImpl<ISD::OutputArg> &Outs,
2475 const SmallVectorImpl<SDValue> &OutVals,
2476 SDLoc DL, SelectionDAG &DAG) const {
2477 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2478 : RetCC_ARM64_AAPCS;
2479 SmallVector<CCValAssign, 16> RVLocs;
2480 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2481 getTargetMachine(), RVLocs, *DAG.getContext());
2482 CCInfo.AnalyzeReturn(Outs, RetCC);
2484 // Copy the result values into the output registers.
2486 SmallVector<SDValue, 4> RetOps(1, Chain);
2487 for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size();
2488 ++i, ++realRVLocIdx) {
2489 CCValAssign &VA = RVLocs[i];
2490 assert(VA.isRegLoc() && "Can only return in registers!");
2491 SDValue Arg = OutVals[realRVLocIdx];
2493 switch (VA.getLocInfo()) {
2495 llvm_unreachable("Unknown loc info!");
2496 case CCValAssign::Full:
2498 case CCValAssign::BCvt:
2499 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2503 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
2504 Flag = Chain.getValue(1);
2505 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2508 RetOps[0] = Chain; // Update chain.
2510 // Add the flag if we have it.
2512 RetOps.push_back(Flag);
2514 return DAG.getNode(ARM64ISD::RET_FLAG, DL, MVT::Other, RetOps);
2517 //===----------------------------------------------------------------------===//
2518 // Other Lowering Code
2519 //===----------------------------------------------------------------------===//
2521 SDValue ARM64TargetLowering::LowerGlobalAddress(SDValue Op,
2522 SelectionDAG &DAG) const {
2523 EVT PtrVT = getPointerTy();
2525 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2526 unsigned char OpFlags =
2527 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
2529 assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
2530 "unexpected offset in global node");
2532 // This also catched the large code model case for Darwin.
2533 if ((OpFlags & ARM64II::MO_GOT) != 0) {
2534 SDValue GotAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
2535 // FIXME: Once remat is capable of dealing with instructions with register
2536 // operands, expand this into two nodes instead of using a wrapper node.
2537 return DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, GotAddr);
2540 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2541 const unsigned char MO_NC = ARM64II::MO_NC;
2543 ARM64ISD::WrapperLarge, DL, PtrVT,
2544 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G3),
2545 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G2 | MO_NC),
2546 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G1 | MO_NC),
2547 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G0 | MO_NC));
2549 // Use ADRP/ADD or ADRP/LDR for everything else: the small model on ELF and
2550 // the only correct model on Darwin.
2551 SDValue Hi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2552 OpFlags | ARM64II::MO_PAGE);
2553 unsigned char LoFlags = OpFlags | ARM64II::MO_PAGEOFF | ARM64II::MO_NC;
2554 SDValue Lo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, LoFlags);
2556 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
2557 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
2561 /// \brief Convert a TLS address reference into the correct sequence of loads
2562 /// and calls to compute the variable's address (for Darwin, currently) and
2563 /// return an SDValue containing the final node.
2565 /// Darwin only has one TLS scheme which must be capable of dealing with the
2566 /// fully general situation, in the worst case. This means:
2567 /// + "extern __thread" declaration.
2568 /// + Defined in a possibly unknown dynamic library.
2570 /// The general system is that each __thread variable has a [3 x i64] descriptor
2571 /// which contains information used by the runtime to calculate the address. The
2572 /// only part of this the compiler needs to know about is the first xword, which
2573 /// contains a function pointer that must be called with the address of the
2574 /// entire descriptor in "x0".
2576 /// Since this descriptor may be in a different unit, in general even the
2577 /// descriptor must be accessed via an indirect load. The "ideal" code sequence
2579 /// adrp x0, _var@TLVPPAGE
2580 /// ldr x0, [x0, _var@TLVPPAGEOFF] ; x0 now contains address of descriptor
2581 /// ldr x1, [x0] ; x1 contains 1st entry of descriptor,
2582 /// ; the function pointer
2583 /// blr x1 ; Uses descriptor address in x0
2584 /// ; Address of _var is now in x0.
2586 /// If the address of _var's descriptor *is* known to the linker, then it can
2587 /// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
2588 /// a slight efficiency gain.
2590 ARM64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
2591 SelectionDAG &DAG) const {
2592 assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");
2595 MVT PtrVT = getPointerTy();
2596 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2599 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2600 SDValue DescAddr = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, TLVPAddr);
2602 // The first entry in the descriptor is a function pointer that we must call
2603 // to obtain the address of the variable.
2604 SDValue Chain = DAG.getEntryNode();
2605 SDValue FuncTLVGet =
2606 DAG.getLoad(MVT::i64, DL, Chain, DescAddr, MachinePointerInfo::getGOT(),
2607 false, true, true, 8);
2608 Chain = FuncTLVGet.getValue(1);
2610 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2611 MFI->setAdjustsStack(true);
2613 // TLS calls preserve all registers except those that absolutely must be
2614 // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
2616 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2617 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2618 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
2620 // Finally, we can make the call. This is just a degenerate version of a
2621 // normal ARM64 call node: x0 takes the address of the descriptor, and returns
2622 // the address of the variable in this thread.
2623 Chain = DAG.getCopyToReg(Chain, DL, ARM64::X0, DescAddr, SDValue());
2624 Chain = DAG.getNode(ARM64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
2625 Chain, FuncTLVGet, DAG.getRegister(ARM64::X0, MVT::i64),
2626 DAG.getRegisterMask(Mask), Chain.getValue(1));
2627 return DAG.getCopyFromReg(Chain, DL, ARM64::X0, PtrVT, Chain.getValue(1));
2630 /// When accessing thread-local variables under either the general-dynamic or
2631 /// local-dynamic system, we make a "TLS-descriptor" call. The variable will
2632 /// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
2633 /// is a function pointer to carry out the resolution. This function takes the
2634 /// address of the descriptor in X0 and returns the TPIDR_EL0 offset in X0. All
2635 /// other registers (except LR, NZCV) are preserved.
2637 /// Thus, the ideal call sequence on AArch64 is:
2639 /// adrp x0, :tlsdesc:thread_var
2640 /// ldr x8, [x0, :tlsdesc_lo12:thread_var]
2641 /// add x0, x0, :tlsdesc_lo12:thread_var
2642 /// .tlsdesccall thread_var
2644 /// (TPIDR_EL0 offset now in x0).
2646 /// The ".tlsdesccall" directive instructs the assembler to insert a particular
2647 /// relocation to help the linker relax this sequence if it turns out to be too
2650 /// FIXME: we currently produce an extra, duplicated, ADRP instruction, but this
2652 SDValue ARM64TargetLowering::LowerELFTLSDescCall(SDValue SymAddr,
2653 SDValue DescAddr, SDLoc DL,
2654 SelectionDAG &DAG) const {
2655 EVT PtrVT = getPointerTy();
2657 // The function we need to call is simply the first entry in the GOT for this
2658 // descriptor, load it in preparation.
2659 SDValue Func = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, SymAddr);
2661 // TLS calls preserve all registers except those that absolutely must be
2662 // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
2664 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2665 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2666 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
2668 // The function takes only one argument: the address of the descriptor itself
2670 SDValue Glue, Chain;
2671 Chain = DAG.getCopyToReg(DAG.getEntryNode(), DL, ARM64::X0, DescAddr, Glue);
2672 Glue = Chain.getValue(1);
2674 // We're now ready to populate the argument list, as with a normal call:
2675 SmallVector<SDValue, 6> Ops;
2676 Ops.push_back(Chain);
2677 Ops.push_back(Func);
2678 Ops.push_back(SymAddr);
2679 Ops.push_back(DAG.getRegister(ARM64::X0, PtrVT));
2680 Ops.push_back(DAG.getRegisterMask(Mask));
2681 Ops.push_back(Glue);
2683 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2684 Chain = DAG.getNode(ARM64ISD::TLSDESC_CALL, DL, NodeTys, Ops);
2685 Glue = Chain.getValue(1);
2687 return DAG.getCopyFromReg(Chain, DL, ARM64::X0, PtrVT, Glue);
2690 SDValue ARM64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
2691 SelectionDAG &DAG) const {
2692 assert(Subtarget->isTargetELF() && "This function expects an ELF target");
2693 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
2694 "ELF TLS only supported in small memory model");
2695 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2697 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
2700 EVT PtrVT = getPointerTy();
2702 const GlobalValue *GV = GA->getGlobal();
2704 SDValue ThreadBase = DAG.getNode(ARM64ISD::THREAD_POINTER, DL, PtrVT);
2706 if (Model == TLSModel::LocalExec) {
2707 SDValue HiVar = DAG.getTargetGlobalAddress(
2708 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_G1);
2709 SDValue LoVar = DAG.getTargetGlobalAddress(
2710 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_G0 | ARM64II::MO_NC);
2712 TPOff = SDValue(DAG.getMachineNode(ARM64::MOVZXi, DL, PtrVT, HiVar,
2713 DAG.getTargetConstant(16, MVT::i32)),
2715 TPOff = SDValue(DAG.getMachineNode(ARM64::MOVKXi, DL, PtrVT, TPOff, LoVar,
2716 DAG.getTargetConstant(0, MVT::i32)),
2718 } else if (Model == TLSModel::InitialExec) {
2719 TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2720 TPOff = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, TPOff);
2721 } else if (Model == TLSModel::LocalDynamic) {
2722 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
2723 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
2724 // the beginning of the module's TLS region, followed by a DTPREL offset
2727 // These accesses will need deduplicating if there's more than one.
2728 ARM64FunctionInfo *MFI =
2729 DAG.getMachineFunction().getInfo<ARM64FunctionInfo>();
2730 MFI->incNumLocalDynamicTLSAccesses();
2732 // Accesses used in this sequence go via the TLS descriptor which lives in
2733 // the GOT. Prepare an address we can use to handle this.
2734 SDValue HiDesc = DAG.getTargetExternalSymbol(
2735 "_TLS_MODULE_BASE_", PtrVT, ARM64II::MO_TLS | ARM64II::MO_PAGE);
2736 SDValue LoDesc = DAG.getTargetExternalSymbol(
2737 "_TLS_MODULE_BASE_", PtrVT,
2738 ARM64II::MO_TLS | ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
2740 // First argument to the descriptor call is the address of the descriptor
2742 SDValue DescAddr = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, HiDesc);
2743 DescAddr = DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
2745 // The call needs a relocation too for linker relaxation. It doesn't make
2746 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
2748 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
2751 // Now we can calculate the offset from TPIDR_EL0 to this module's
2752 // thread-local area.
2753 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
2755 // Now use :dtprel_whatever: operations to calculate this variable's offset
2756 // in its thread-storage area.
2757 SDValue HiVar = DAG.getTargetGlobalAddress(
2758 GV, DL, MVT::i64, 0, ARM64II::MO_TLS | ARM64II::MO_G1);
2759 SDValue LoVar = DAG.getTargetGlobalAddress(
2760 GV, DL, MVT::i64, 0, ARM64II::MO_TLS | ARM64II::MO_G0 | ARM64II::MO_NC);
2763 SDValue(DAG.getMachineNode(ARM64::MOVZXi, DL, PtrVT, HiVar,
2764 DAG.getTargetConstant(16, MVT::i32)),
2766 DTPOff = SDValue(DAG.getMachineNode(ARM64::MOVKXi, DL, PtrVT, DTPOff, LoVar,
2767 DAG.getTargetConstant(0, MVT::i32)),
2770 TPOff = DAG.getNode(ISD::ADD, DL, PtrVT, TPOff, DTPOff);
2771 } else if (Model == TLSModel::GeneralDynamic) {
2772 // Accesses used in this sequence go via the TLS descriptor which lives in
2773 // the GOT. Prepare an address we can use to handle this.
2774 SDValue HiDesc = DAG.getTargetGlobalAddress(
2775 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_PAGE);
2776 SDValue LoDesc = DAG.getTargetGlobalAddress(
2778 ARM64II::MO_TLS | ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
2780 // First argument to the descriptor call is the address of the descriptor
2782 SDValue DescAddr = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, HiDesc);
2783 DescAddr = DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
2785 // The call needs a relocation too for linker relaxation. It doesn't make
2786 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
2789 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2791 // Finally we can make a call to calculate the offset from tpidr_el0.
2792 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
2794 llvm_unreachable("Unsupported ELF TLS access model");
2796 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
2799 SDValue ARM64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
2800 SelectionDAG &DAG) const {
2801 if (Subtarget->isTargetDarwin())
2802 return LowerDarwinGlobalTLSAddress(Op, DAG);
2803 else if (Subtarget->isTargetELF())
2804 return LowerELFGlobalTLSAddress(Op, DAG);
2806 llvm_unreachable("Unexpected platform trying to use TLS");
2808 SDValue ARM64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2809 SDValue Chain = Op.getOperand(0);
2810 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2811 SDValue LHS = Op.getOperand(2);
2812 SDValue RHS = Op.getOperand(3);
2813 SDValue Dest = Op.getOperand(4);
2816 // Handle f128 first, since lowering it will result in comparing the return
2817 // value of a libcall against zero, which is just what the rest of LowerBR_CC
2818 // is expecting to deal with.
2819 if (LHS.getValueType() == MVT::f128) {
2820 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
2822 // If softenSetCCOperands returned a scalar, we need to compare the result
2823 // against zero to select between true and false values.
2824 if (!RHS.getNode()) {
2825 RHS = DAG.getConstant(0, LHS.getValueType());
2830 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
2832 unsigned Opc = LHS.getOpcode();
2833 if (LHS.getResNo() == 1 && isa<ConstantSDNode>(RHS) &&
2834 cast<ConstantSDNode>(RHS)->isOne() &&
2835 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
2836 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
2837 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
2838 "Unexpected condition code.");
2839 // Only lower legal XALUO ops.
2840 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
2843 // The actual operation with overflow check.
2844 ARM64CC::CondCode OFCC;
2845 SDValue Value, Overflow;
2846 std::tie(Value, Overflow) = getARM64XALUOOp(OFCC, LHS.getValue(0), DAG);
2848 if (CC == ISD::SETNE)
2849 OFCC = getInvertedCondCode(OFCC);
2850 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
2852 return DAG.getNode(ARM64ISD::BRCOND, SDLoc(LHS), MVT::Other, Chain, Dest,
2856 if (LHS.getValueType().isInteger()) {
2857 assert((LHS.getValueType() == RHS.getValueType()) &&
2858 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
2860 // If the RHS of the comparison is zero, we can potentially fold this
2861 // to a specialized branch.
2862 const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
2863 if (RHSC && RHSC->getZExtValue() == 0) {
2864 if (CC == ISD::SETEQ) {
2865 // See if we can use a TBZ to fold in an AND as well.
2866 // TBZ has a smaller branch displacement than CBZ. If the offset is
2867 // out of bounds, a late MI-layer pass rewrites branches.
2868 // 403.gcc is an example that hits this case.
2869 if (LHS.getOpcode() == ISD::AND &&
2870 isa<ConstantSDNode>(LHS.getOperand(1)) &&
2871 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
2872 SDValue Test = LHS.getOperand(0);
2873 uint64_t Mask = LHS.getConstantOperandVal(1);
2875 // TBZ only operates on i64's, but the ext should be free.
2876 if (Test.getValueType() == MVT::i32)
2877 Test = DAG.getAnyExtOrTrunc(Test, dl, MVT::i64);
2879 return DAG.getNode(ARM64ISD::TBZ, dl, MVT::Other, Chain, Test,
2880 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
2883 return DAG.getNode(ARM64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest);
2884 } else if (CC == ISD::SETNE) {
2885 // See if we can use a TBZ to fold in an AND as well.
2886 // TBZ has a smaller branch displacement than CBZ. If the offset is
2887 // out of bounds, a late MI-layer pass rewrites branches.
2888 // 403.gcc is an example that hits this case.
2889 if (LHS.getOpcode() == ISD::AND &&
2890 isa<ConstantSDNode>(LHS.getOperand(1)) &&
2891 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
2892 SDValue Test = LHS.getOperand(0);
2893 uint64_t Mask = LHS.getConstantOperandVal(1);
2895 // TBNZ only operates on i64's, but the ext should be free.
2896 if (Test.getValueType() == MVT::i32)
2897 Test = DAG.getAnyExtOrTrunc(Test, dl, MVT::i64);
2899 return DAG.getNode(ARM64ISD::TBNZ, dl, MVT::Other, Chain, Test,
2900 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
2903 return DAG.getNode(ARM64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest);
2908 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
2909 return DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
2913 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2915 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
2916 // clean. Some of them require two branches to implement.
2917 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
2918 ARM64CC::CondCode CC1, CC2;
2919 changeFPCCToARM64CC(CC, CC1, CC2);
2920 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
2922 DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp);
2923 if (CC2 != ARM64CC::AL) {
2924 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
2925 return DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val,
2932 SDValue ARM64TargetLowering::LowerFCOPYSIGN(SDValue Op,
2933 SelectionDAG &DAG) const {
2934 EVT VT = Op.getValueType();
2937 SDValue In1 = Op.getOperand(0);
2938 SDValue In2 = Op.getOperand(1);
2939 EVT SrcVT = In2.getValueType();
2941 if (SrcVT == MVT::f32 && VT == MVT::f64)
2942 In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2);
2943 else if (SrcVT == MVT::f64 && VT == MVT::f32)
2944 In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, DAG.getIntPtrConstant(0));
2946 // FIXME: Src type is different, bail out for now. Can VT really be a
2953 SDValue EltMask, VecVal1, VecVal2;
2954 if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) {
2957 EltMask = DAG.getConstant(0x80000000ULL, EltVT);
2959 if (!VT.isVector()) {
2960 VecVal1 = DAG.getTargetInsertSubreg(ARM64::ssub, DL, VecVT,
2961 DAG.getUNDEF(VecVT), In1);
2962 VecVal2 = DAG.getTargetInsertSubreg(ARM64::ssub, DL, VecVT,
2963 DAG.getUNDEF(VecVT), In2);
2965 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
2966 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
2968 } else if (VT == MVT::f64 || VT == MVT::v2f64) {
2972 // We want to materialize a mask with the the high bit set, but the AdvSIMD
2973 // immediate moves cannot materialize that in a single instruction for
2974 // 64-bit elements. Instead, materialize zero and then negate it.
2975 EltMask = DAG.getConstant(0, EltVT);
2977 if (!VT.isVector()) {
2978 VecVal1 = DAG.getTargetInsertSubreg(ARM64::dsub, DL, VecVT,
2979 DAG.getUNDEF(VecVT), In1);
2980 VecVal2 = DAG.getTargetInsertSubreg(ARM64::dsub, DL, VecVT,
2981 DAG.getUNDEF(VecVT), In2);
2983 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
2984 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
2987 llvm_unreachable("Invalid type for copysign!");
2990 std::vector<SDValue> BuildVectorOps;
2991 for (unsigned i = 0; i < VecVT.getVectorNumElements(); ++i)
2992 BuildVectorOps.push_back(EltMask);
2994 SDValue BuildVec = DAG.getNode(ISD::BUILD_VECTOR, DL, VecVT, BuildVectorOps);
2996 // If we couldn't materialize the mask above, then the mask vector will be
2997 // the zero vector, and we need to negate it here.
2998 if (VT == MVT::f64 || VT == MVT::v2f64) {
2999 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec);
3000 BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec);
3001 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec);
3005 DAG.getNode(ARM64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec);
3008 return DAG.getTargetExtractSubreg(ARM64::ssub, DL, VT, Sel);
3009 else if (VT == MVT::f64)
3010 return DAG.getTargetExtractSubreg(ARM64::dsub, DL, VT, Sel);
3012 return DAG.getNode(ISD::BITCAST, DL, VT, Sel);
3015 SDValue ARM64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
3016 if (DAG.getMachineFunction().getFunction()->getAttributes().hasAttribute(
3017 AttributeSet::FunctionIndex, Attribute::NoImplicitFloat))
3020 // While there is no integer popcount instruction, it can
3021 // be more efficiently lowered to the following sequence that uses
3022 // AdvSIMD registers/instructions as long as the copies to/from
3023 // the AdvSIMD registers are cheap.
3024 // FMOV D0, X0 // copy 64-bit int to vector, high bits zero'd
3025 // CNT V0.8B, V0.8B // 8xbyte pop-counts
3026 // ADDV B0, V0.8B // sum 8xbyte pop-counts
3027 // UMOV X0, V0.B[0] // copy byte result back to integer reg
3028 SDValue Val = Op.getOperand(0);
3030 EVT VT = Op.getValueType();
3031 SDValue ZeroVec = DAG.getUNDEF(MVT::v8i8);
3034 if (VT == MVT::i32) {
3035 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Val);
3037 DAG.getTargetInsertSubreg(ARM64::ssub, DL, MVT::v8i8, ZeroVec, VecVal);
3039 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);
3042 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, VecVal);
3043 SDValue UaddLV = DAG.getNode(
3044 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
3045 DAG.getConstant(Intrinsic::arm64_neon_uaddlv, MVT::i32), CtPop);
3048 UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV);
3052 SDValue ARM64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
3054 if (Op.getValueType().isVector())
3055 return LowerVSETCC(Op, DAG);
3057 SDValue LHS = Op.getOperand(0);
3058 SDValue RHS = Op.getOperand(1);
3059 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
3062 // We chose ZeroOrOneBooleanContents, so use zero and one.
3063 EVT VT = Op.getValueType();
3064 SDValue TVal = DAG.getConstant(1, VT);
3065 SDValue FVal = DAG.getConstant(0, VT);
3067 // Handle f128 first, since one possible outcome is a normal integer
3068 // comparison which gets picked up by the next if statement.
3069 if (LHS.getValueType() == MVT::f128) {
3070 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3072 // If softenSetCCOperands returned a scalar, use it.
3073 if (!RHS.getNode()) {
3074 assert(LHS.getValueType() == Op.getValueType() &&
3075 "Unexpected setcc expansion!");
3080 if (LHS.getValueType().isInteger()) {
3083 getARM64Cmp(LHS, RHS, ISD::getSetCCInverse(CC, true), CCVal, DAG, dl);
3085 // Note that we inverted the condition above, so we reverse the order of
3086 // the true and false operands here. This will allow the setcc to be
3087 // matched to a single CSINC instruction.
3088 return DAG.getNode(ARM64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp);
3091 // Now we know we're dealing with FP values.
3092 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3094 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3095 // and do the comparison.
3096 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3098 ARM64CC::CondCode CC1, CC2;
3099 changeFPCCToARM64CC(CC, CC1, CC2);
3100 if (CC2 == ARM64CC::AL) {
3101 changeFPCCToARM64CC(ISD::getSetCCInverse(CC, false), CC1, CC2);
3102 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3104 // Note that we inverted the condition above, so we reverse the order of
3105 // the true and false operands here. This will allow the setcc to be
3106 // matched to a single CSINC instruction.
3107 return DAG.getNode(ARM64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp);
3109 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
3110 // clean. Some of them require two CSELs to implement. As is in this case,
3111 // we emit the first CSEL and then emit a second using the output of the
3112 // first as the RHS. We're effectively OR'ing the two CC's together.
3114 // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
3115 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3116 SDValue CS1 = DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3118 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3119 return DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3123 /// A SELECT_CC operation is really some kind of max or min if both values being
3124 /// compared are, in some sense, equal to the results in either case. However,
3125 /// it is permissible to compare f32 values and produce directly extended f64
3128 /// Extending the comparison operands would also be allowed, but is less likely
3129 /// to happen in practice since their use is right here. Note that truncate
3130 /// operations would *not* be semantically equivalent.
3131 static bool selectCCOpsAreFMaxCompatible(SDValue Cmp, SDValue Result) {
3135 ConstantFPSDNode *CCmp = dyn_cast<ConstantFPSDNode>(Cmp);
3136 ConstantFPSDNode *CResult = dyn_cast<ConstantFPSDNode>(Result);
3137 if (CCmp && CResult && Cmp.getValueType() == MVT::f32 &&
3138 Result.getValueType() == MVT::f64) {
3140 APFloat CmpVal = CCmp->getValueAPF();
3141 CmpVal.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &Lossy);
3142 return CResult->getValueAPF().bitwiseIsEqual(CmpVal);
3145 return Result->getOpcode() == ISD::FP_EXTEND && Result->getOperand(0) == Cmp;
3148 SDValue ARM64TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3149 SDValue CC = Op->getOperand(0);
3150 SDValue TVal = Op->getOperand(1);
3151 SDValue FVal = Op->getOperand(2);
3154 unsigned Opc = CC.getOpcode();
3155 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
3157 if (CC.getResNo() == 1 &&
3158 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3159 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3160 // Only lower legal XALUO ops.
3161 if (!DAG.getTargetLoweringInfo().isTypeLegal(CC->getValueType(0)))
3164 ARM64CC::CondCode OFCC;
3165 SDValue Value, Overflow;
3166 std::tie(Value, Overflow) = getARM64XALUOOp(OFCC, CC.getValue(0), DAG);
3167 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
3169 return DAG.getNode(ARM64ISD::CSEL, DL, Op.getValueType(), TVal, FVal, CCVal,
3173 if (CC.getOpcode() == ISD::SETCC)
3174 return DAG.getSelectCC(DL, CC.getOperand(0), CC.getOperand(1), TVal, FVal,
3175 cast<CondCodeSDNode>(CC.getOperand(2))->get());
3177 return DAG.getSelectCC(DL, CC, DAG.getConstant(0, CC.getValueType()), TVal,
3181 SDValue ARM64TargetLowering::LowerSELECT_CC(SDValue Op,
3182 SelectionDAG &DAG) const {
3183 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3184 SDValue LHS = Op.getOperand(0);
3185 SDValue RHS = Op.getOperand(1);
3186 SDValue TVal = Op.getOperand(2);
3187 SDValue FVal = Op.getOperand(3);
3190 // Handle f128 first, because it will result in a comparison of some RTLIB
3191 // call result against zero.
3192 if (LHS.getValueType() == MVT::f128) {
3193 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3195 // If softenSetCCOperands returned a scalar, we need to compare the result
3196 // against zero to select between true and false values.
3197 if (!RHS.getNode()) {
3198 RHS = DAG.getConstant(0, LHS.getValueType());
3203 // Handle integers first.
3204 if (LHS.getValueType().isInteger()) {
3205 assert((LHS.getValueType() == RHS.getValueType()) &&
3206 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3208 unsigned Opcode = ARM64ISD::CSEL;
3210 // If both the TVal and the FVal are constants, see if we can swap them in
3211 // order to for a CSINV or CSINC out of them.
3212 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
3213 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
3215 if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) {
3216 std::swap(TVal, FVal);
3217 std::swap(CTVal, CFVal);
3218 CC = ISD::getSetCCInverse(CC, true);
3219 } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) {
3220 std::swap(TVal, FVal);
3221 std::swap(CTVal, CFVal);
3222 CC = ISD::getSetCCInverse(CC, true);
3223 } else if (TVal.getOpcode() == ISD::XOR) {
3224 // If TVal is a NOT we want to swap TVal and FVal so that we can match
3225 // with a CSINV rather than a CSEL.
3226 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(1));
3228 if (CVal && CVal->isAllOnesValue()) {
3229 std::swap(TVal, FVal);
3230 std::swap(CTVal, CFVal);
3231 CC = ISD::getSetCCInverse(CC, true);
3233 } else if (TVal.getOpcode() == ISD::SUB) {
3234 // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
3235 // that we can match with a CSNEG rather than a CSEL.
3236 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(0));
3238 if (CVal && CVal->isNullValue()) {
3239 std::swap(TVal, FVal);
3240 std::swap(CTVal, CFVal);
3241 CC = ISD::getSetCCInverse(CC, true);
3243 } else if (CTVal && CFVal) {
3244 const int64_t TrueVal = CTVal->getSExtValue();
3245 const int64_t FalseVal = CFVal->getSExtValue();
3248 // If both TVal and FVal are constants, see if FVal is the
3249 // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
3250 // instead of a CSEL in that case.
3251 if (TrueVal == ~FalseVal) {
3252 Opcode = ARM64ISD::CSINV;
3253 } else if (TrueVal == -FalseVal) {
3254 Opcode = ARM64ISD::CSNEG;
3255 } else if (TVal.getValueType() == MVT::i32) {
3256 // If our operands are only 32-bit wide, make sure we use 32-bit
3257 // arithmetic for the check whether we can use CSINC. This ensures that
3258 // the addition in the check will wrap around properly in case there is
3259 // an overflow (which would not be the case if we do the check with
3260 // 64-bit arithmetic).
3261 const uint32_t TrueVal32 = CTVal->getZExtValue();
3262 const uint32_t FalseVal32 = CFVal->getZExtValue();
3264 if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
3265 Opcode = ARM64ISD::CSINC;
3267 if (TrueVal32 > FalseVal32) {
3271 // 64-bit check whether we can use CSINC.
3272 } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) {
3273 Opcode = ARM64ISD::CSINC;
3275 if (TrueVal > FalseVal) {
3280 // Swap TVal and FVal if necessary.
3282 std::swap(TVal, FVal);
3283 std::swap(CTVal, CFVal);
3284 CC = ISD::getSetCCInverse(CC, true);
3287 if (Opcode != ARM64ISD::CSEL) {
3288 // Drop FVal since we can get its value by simply inverting/negating
3295 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3297 EVT VT = Op.getValueType();
3298 return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp);
3301 // Now we know we're dealing with FP values.
3302 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3303 assert(LHS.getValueType() == RHS.getValueType());
3304 EVT VT = Op.getValueType();
3306 // Try to match this select into a max/min operation, which have dedicated
3307 // opcode in the instruction set.
3308 // FIXME: This is not correct in the presence of NaNs, so we only enable this
3310 if (getTargetMachine().Options.NoNaNsFPMath) {
3311 SDValue MinMaxLHS = TVal, MinMaxRHS = FVal;
3312 if (selectCCOpsAreFMaxCompatible(LHS, MinMaxRHS) &&
3313 selectCCOpsAreFMaxCompatible(RHS, MinMaxLHS)) {
3314 CC = ISD::getSetCCSwappedOperands(CC);
3315 std::swap(MinMaxLHS, MinMaxRHS);
3318 if (selectCCOpsAreFMaxCompatible(LHS, MinMaxLHS) &&
3319 selectCCOpsAreFMaxCompatible(RHS, MinMaxRHS)) {
3329 return DAG.getNode(ARM64ISD::FMAX, dl, VT, MinMaxLHS, MinMaxRHS);
3337 return DAG.getNode(ARM64ISD::FMIN, dl, VT, MinMaxLHS, MinMaxRHS);
3343 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3344 // and do the comparison.
3345 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
3347 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
3348 // clean. Some of them require two CSELs to implement.
3349 ARM64CC::CondCode CC1, CC2;
3350 changeFPCCToARM64CC(CC, CC1, CC2);
3351 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3352 SDValue CS1 = DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3354 // If we need a second CSEL, emit it, using the output of the first as the
3355 // RHS. We're effectively OR'ing the two CC's together.
3356 if (CC2 != ARM64CC::AL) {
3357 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3358 return DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3361 // Otherwise, return the output of the first CSEL.
3365 SDValue ARM64TargetLowering::LowerJumpTable(SDValue Op,
3366 SelectionDAG &DAG) const {
3367 // Jump table entries as PC relative offsets. No additional tweaking
3368 // is necessary here. Just get the address of the jump table.
3369 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3370 EVT PtrVT = getPointerTy();
3373 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3374 !Subtarget->isTargetMachO()) {
3375 const unsigned char MO_NC = ARM64II::MO_NC;
3377 ARM64ISD::WrapperLarge, DL, PtrVT,
3378 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G3),
3379 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G2 | MO_NC),
3380 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G1 | MO_NC),
3381 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_G0 | MO_NC));
3384 SDValue Hi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_PAGE);
3385 SDValue Lo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3386 ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
3387 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3388 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3391 SDValue ARM64TargetLowering::LowerConstantPool(SDValue Op,
3392 SelectionDAG &DAG) const {
3393 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3394 EVT PtrVT = getPointerTy();
3397 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
3398 // Use the GOT for the large code model on iOS.
3399 if (Subtarget->isTargetMachO()) {
3400 SDValue GotAddr = DAG.getTargetConstantPool(
3401 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3403 return DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, GotAddr);
3406 const unsigned char MO_NC = ARM64II::MO_NC;
3408 ARM64ISD::WrapperLarge, DL, PtrVT,
3409 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3410 CP->getOffset(), ARM64II::MO_G3),
3411 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3412 CP->getOffset(), ARM64II::MO_G2 | MO_NC),
3413 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3414 CP->getOffset(), ARM64II::MO_G1 | MO_NC),
3415 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3416 CP->getOffset(), ARM64II::MO_G0 | MO_NC));
3418 // Use ADRP/ADD or ADRP/LDR for everything else: the small memory model on
3419 // ELF, the only valid one on Darwin.
3421 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3422 CP->getOffset(), ARM64II::MO_PAGE);
3423 SDValue Lo = DAG.getTargetConstantPool(
3424 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3425 ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
3427 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3428 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3432 SDValue ARM64TargetLowering::LowerBlockAddress(SDValue Op,
3433 SelectionDAG &DAG) const {
3434 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
3435 EVT PtrVT = getPointerTy();
3437 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3438 !Subtarget->isTargetMachO()) {
3439 const unsigned char MO_NC = ARM64II::MO_NC;
3441 ARM64ISD::WrapperLarge, DL, PtrVT,
3442 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G3),
3443 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G2 | MO_NC),
3444 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G1 | MO_NC),
3445 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G0 | MO_NC));
3447 SDValue Hi = DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_PAGE);
3448 SDValue Lo = DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_PAGEOFF |
3450 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3451 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3455 SDValue ARM64TargetLowering::LowerDarwin_VASTART(SDValue Op,
3456 SelectionDAG &DAG) const {
3457 ARM64FunctionInfo *FuncInfo =
3458 DAG.getMachineFunction().getInfo<ARM64FunctionInfo>();
3462 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3463 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3464 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
3465 MachinePointerInfo(SV), false, false, 0);
3468 SDValue ARM64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
3469 SelectionDAG &DAG) const {
3470 // The layout of the va_list struct is specified in the AArch64 Procedure Call
3471 // Standard, section B.3.
3472 MachineFunction &MF = DAG.getMachineFunction();
3473 ARM64FunctionInfo *FuncInfo = MF.getInfo<ARM64FunctionInfo>();
3476 SDValue Chain = Op.getOperand(0);
3477 SDValue VAList = Op.getOperand(1);
3478 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3479 SmallVector<SDValue, 4> MemOps;
3481 // void *__stack at offset 0
3483 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3484 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
3485 MachinePointerInfo(SV), false, false, 8));
3487 // void *__gr_top at offset 8
3488 int GPRSize = FuncInfo->getVarArgsGPRSize();
3490 SDValue GRTop, GRTopAddr;
3492 GRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3493 DAG.getConstant(8, getPointerTy()));
3495 GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), getPointerTy());
3496 GRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), GRTop,
3497 DAG.getConstant(GPRSize, getPointerTy()));
3499 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
3500 MachinePointerInfo(SV, 8), false, false, 8));
3503 // void *__vr_top at offset 16
3504 int FPRSize = FuncInfo->getVarArgsFPRSize();
3506 SDValue VRTop, VRTopAddr;
3507 VRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3508 DAG.getConstant(16, getPointerTy()));
3510 VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), getPointerTy());
3511 VRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), VRTop,
3512 DAG.getConstant(FPRSize, getPointerTy()));
3514 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
3515 MachinePointerInfo(SV, 16), false, false, 8));
3518 // int __gr_offs at offset 24
3519 SDValue GROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3520 DAG.getConstant(24, getPointerTy()));
3521 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-GPRSize, MVT::i32),
3522 GROffsAddr, MachinePointerInfo(SV, 24), false,
3525 // int __vr_offs at offset 28
3526 SDValue VROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3527 DAG.getConstant(28, getPointerTy()));
3528 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-FPRSize, MVT::i32),
3529 VROffsAddr, MachinePointerInfo(SV, 28), false,
3532 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
3535 SDValue ARM64TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
3536 return Subtarget->isTargetDarwin() ? LowerDarwin_VASTART(Op, DAG)
3537 : LowerAAPCS_VASTART(Op, DAG);
3540 SDValue ARM64TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
3541 // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
3543 unsigned VaListSize = Subtarget->isTargetDarwin() ? 8 : 32;
3544 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
3545 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
3547 return DAG.getMemcpy(Op.getOperand(0), SDLoc(Op), Op.getOperand(1),
3548 Op.getOperand(2), DAG.getConstant(VaListSize, MVT::i32),
3549 8, false, false, MachinePointerInfo(DestSV),
3550 MachinePointerInfo(SrcSV));
3553 SDValue ARM64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
3554 assert(Subtarget->isTargetDarwin() &&
3555 "automatic va_arg instruction only works on Darwin");
3557 const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3558 EVT VT = Op.getValueType();
3560 SDValue Chain = Op.getOperand(0);
3561 SDValue Addr = Op.getOperand(1);
3562 unsigned Align = Op.getConstantOperandVal(3);
3564 SDValue VAList = DAG.getLoad(getPointerTy(), DL, Chain, Addr,
3565 MachinePointerInfo(V), false, false, false, 0);
3566 Chain = VAList.getValue(1);
3569 assert(((Align & (Align - 1)) == 0) && "Expected Align to be a power of 2");
3570 VAList = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3571 DAG.getConstant(Align - 1, getPointerTy()));
3572 VAList = DAG.getNode(ISD::AND, DL, getPointerTy(), VAList,
3573 DAG.getConstant(-(int64_t)Align, getPointerTy()));
3576 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
3577 uint64_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
3579 // Scalar integer and FP values smaller than 64 bits are implicitly extended
3580 // up to 64 bits. At the very least, we have to increase the striding of the
3581 // vaargs list to match this, and for FP values we need to introduce
3582 // FP_ROUND nodes as well.
3583 if (VT.isInteger() && !VT.isVector())
3585 bool NeedFPTrunc = false;
3586 if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
3591 // Increment the pointer, VAList, to the next vaarg
3592 SDValue VANext = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3593 DAG.getConstant(ArgSize, getPointerTy()));
3594 // Store the incremented VAList to the legalized pointer
3595 SDValue APStore = DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V),
3598 // Load the actual argument out of the pointer VAList
3600 // Load the value as an f64.
3601 SDValue WideFP = DAG.getLoad(MVT::f64, DL, APStore, VAList,
3602 MachinePointerInfo(), false, false, false, 0);
3603 // Round the value down to an f32.
3604 SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
3605 DAG.getIntPtrConstant(1));
3606 SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
3607 // Merge the rounded value with the chain output of the load.
3608 return DAG.getMergeValues(Ops, DL);
3611 return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo(), false,
3615 SDValue ARM64TargetLowering::LowerFRAMEADDR(SDValue Op,
3616 SelectionDAG &DAG) const {
3617 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3618 MFI->setFrameAddressIsTaken(true);
3620 EVT VT = Op.getValueType();
3622 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3623 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL, ARM64::FP, VT);
3625 FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
3626 MachinePointerInfo(), false, false, false, 0);
3630 // FIXME? Maybe this could be a TableGen attribute on some registers and
3631 // this table could be generated automatically from RegInfo.
3632 unsigned ARM64TargetLowering::getRegisterByName(const char* RegName,
3634 unsigned Reg = StringSwitch<unsigned>(RegName)
3635 .Case("sp", ARM64::SP)
3639 report_fatal_error("Invalid register name global variable");
3642 SDValue ARM64TargetLowering::LowerRETURNADDR(SDValue Op,
3643 SelectionDAG &DAG) const {
3644 MachineFunction &MF = DAG.getMachineFunction();
3645 MachineFrameInfo *MFI = MF.getFrameInfo();
3646 MFI->setReturnAddressIsTaken(true);
3648 EVT VT = Op.getValueType();
3650 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3652 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3653 SDValue Offset = DAG.getConstant(8, getPointerTy());
3654 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
3655 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
3656 MachinePointerInfo(), false, false, false, 0);
3659 // Return LR, which contains the return address. Mark it an implicit live-in.
3660 unsigned Reg = MF.addLiveIn(ARM64::LR, &ARM64::GPR64RegClass);
3661 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
3664 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3665 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
3666 SDValue ARM64TargetLowering::LowerShiftRightParts(SDValue Op,
3667 SelectionDAG &DAG) const {
3668 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3669 EVT VT = Op.getValueType();
3670 unsigned VTBits = VT.getSizeInBits();
3672 SDValue ShOpLo = Op.getOperand(0);
3673 SDValue ShOpHi = Op.getOperand(1);
3674 SDValue ShAmt = Op.getOperand(2);
3676 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3678 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3680 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
3681 DAG.getConstant(VTBits, MVT::i64), ShAmt);
3682 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3683 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
3684 DAG.getConstant(VTBits, MVT::i64));
3685 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3687 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
3688 ISD::SETGE, dl, DAG);
3689 SDValue CCVal = DAG.getConstant(ARM64CC::GE, MVT::i32);
3691 SDValue FalseValLo = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3692 SDValue TrueValLo = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3694 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
3696 // ARM64 shifts larger than the register width are wrapped rather than
3697 // clamped, so we can't just emit "hi >> x".
3698 SDValue FalseValHi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3699 SDValue TrueValHi = Opc == ISD::SRA
3700 ? DAG.getNode(Opc, dl, VT, ShOpHi,
3701 DAG.getConstant(VTBits - 1, MVT::i64))
3702 : DAG.getConstant(0, VT);
3704 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValHi, FalseValHi, CCVal, Cmp);
3706 SDValue Ops[2] = { Lo, Hi };
3707 return DAG.getMergeValues(Ops, dl);
3710 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3711 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
3712 SDValue ARM64TargetLowering::LowerShiftLeftParts(SDValue Op,
3713 SelectionDAG &DAG) const {
3714 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3715 EVT VT = Op.getValueType();
3716 unsigned VTBits = VT.getSizeInBits();
3718 SDValue ShOpLo = Op.getOperand(0);
3719 SDValue ShOpHi = Op.getOperand(1);
3720 SDValue ShAmt = Op.getOperand(2);
3723 assert(Op.getOpcode() == ISD::SHL_PARTS);
3724 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
3725 DAG.getConstant(VTBits, MVT::i64), ShAmt);
3726 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3727 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
3728 DAG.getConstant(VTBits, MVT::i64));
3729 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3730 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3732 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3734 SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64),
3735 ISD::SETGE, dl, DAG);
3736 SDValue CCVal = DAG.getConstant(ARM64CC::GE, MVT::i32);
3737 SDValue Hi = DAG.getNode(ARM64ISD::CSEL, dl, VT, Tmp3, FalseVal, CCVal, Cmp);
3739 // ARM64 shifts of larger than register sizes are wrapped rather than clamped,
3740 // so we can't just emit "lo << a" if a is too big.
3741 SDValue TrueValLo = DAG.getConstant(0, VT);
3742 SDValue FalseValLo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3744 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
3746 SDValue Ops[2] = { Lo, Hi };
3747 return DAG.getMergeValues(Ops, dl);
3751 ARM64TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3752 // The ARM64 target doesn't support folding offsets into global addresses.
3756 bool ARM64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3757 // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
3758 // FIXME: We should be able to handle f128 as well with a clever lowering.
3759 if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32))
3763 return ARM64_AM::getFP64Imm(Imm) != -1;
3764 else if (VT == MVT::f32)
3765 return ARM64_AM::getFP32Imm(Imm) != -1;
3769 //===----------------------------------------------------------------------===//
3770 // ARM64 Optimization Hooks
3771 //===----------------------------------------------------------------------===//
3773 //===----------------------------------------------------------------------===//
3774 // ARM64 Inline Assembly Support
3775 //===----------------------------------------------------------------------===//
3777 // Table of Constraints
3778 // TODO: This is the current set of constraints supported by ARM for the
3779 // compiler, not all of them may make sense, e.g. S may be difficult to support.
3781 // r - A general register
3782 // w - An FP/SIMD register of some size in the range v0-v31
3783 // x - An FP/SIMD register of some size in the range v0-v15
3784 // I - Constant that can be used with an ADD instruction
3785 // J - Constant that can be used with a SUB instruction
3786 // K - Constant that can be used with a 32-bit logical instruction
3787 // L - Constant that can be used with a 64-bit logical instruction
3788 // M - Constant that can be used as a 32-bit MOV immediate
3789 // N - Constant that can be used as a 64-bit MOV immediate
3790 // Q - A memory reference with base register and no offset
3791 // S - A symbolic address
3792 // Y - Floating point constant zero
3793 // Z - Integer constant zero
3795 // Note that general register operands will be output using their 64-bit x
3796 // register name, whatever the size of the variable, unless the asm operand
3797 // is prefixed by the %w modifier. Floating-point and SIMD register operands
3798 // will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
3801 /// getConstraintType - Given a constraint letter, return the type of
3802 /// constraint it is for this target.
3803 ARM64TargetLowering::ConstraintType
3804 ARM64TargetLowering::getConstraintType(const std::string &Constraint) const {
3805 if (Constraint.size() == 1) {
3806 switch (Constraint[0]) {
3813 return C_RegisterClass;
3814 // An address with a single base register. Due to the way we
3815 // currently handle addresses it is the same as 'r'.
3820 return TargetLowering::getConstraintType(Constraint);
3823 /// Examine constraint type and operand type and determine a weight value.
3824 /// This object must already have been set up with the operand type
3825 /// and the current alternative constraint selected.
3826 TargetLowering::ConstraintWeight
3827 ARM64TargetLowering::getSingleConstraintMatchWeight(
3828 AsmOperandInfo &info, const char *constraint) const {
3829 ConstraintWeight weight = CW_Invalid;
3830 Value *CallOperandVal = info.CallOperandVal;
3831 // If we don't have a value, we can't do a match,
3832 // but allow it at the lowest weight.
3833 if (!CallOperandVal)
3835 Type *type = CallOperandVal->getType();
3836 // Look at the constraint type.
3837 switch (*constraint) {
3839 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
3843 if (type->isFloatingPointTy() || type->isVectorTy())
3844 weight = CW_Register;
3847 weight = CW_Constant;
3853 std::pair<unsigned, const TargetRegisterClass *>
3854 ARM64TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3856 if (Constraint.size() == 1) {
3857 switch (Constraint[0]) {
3859 if (VT.getSizeInBits() == 64)
3860 return std::make_pair(0U, &ARM64::GPR64commonRegClass);
3861 return std::make_pair(0U, &ARM64::GPR32commonRegClass);
3864 return std::make_pair(0U, &ARM64::FPR32RegClass);
3865 if (VT.getSizeInBits() == 64)
3866 return std::make_pair(0U, &ARM64::FPR64RegClass);
3867 if (VT.getSizeInBits() == 128)
3868 return std::make_pair(0U, &ARM64::FPR128RegClass);
3870 // The instructions that this constraint is designed for can
3871 // only take 128-bit registers so just use that regclass.
3873 if (VT.getSizeInBits() == 128)
3874 return std::make_pair(0U, &ARM64::FPR128_loRegClass);
3878 if (StringRef("{cc}").equals_lower(Constraint))
3879 return std::make_pair(unsigned(ARM64::NZCV), &ARM64::CCRRegClass);
3881 // Use the default implementation in TargetLowering to convert the register
3882 // constraint into a member of a register class.
3883 std::pair<unsigned, const TargetRegisterClass *> Res;
3884 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3886 // Not found as a standard register?
3888 unsigned Size = Constraint.size();
3889 if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
3890 tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
3891 const std::string Reg =
3892 std::string(&Constraint[2], &Constraint[Size - 1]);
3893 int RegNo = atoi(Reg.c_str());
3894 if (RegNo >= 0 && RegNo <= 31) {
3895 // v0 - v31 are aliases of q0 - q31.
3896 // By default we'll emit v0-v31 for this unless there's a modifier where
3897 // we'll emit the correct register as well.
3898 Res.first = ARM64::FPR128RegClass.getRegister(RegNo);
3899 Res.second = &ARM64::FPR128RegClass;
3907 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3908 /// vector. If it is invalid, don't add anything to Ops.
3909 void ARM64TargetLowering::LowerAsmOperandForConstraint(
3910 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
3911 SelectionDAG &DAG) const {
3914 // Currently only support length 1 constraints.
3915 if (Constraint.length() != 1)
3918 char ConstraintLetter = Constraint[0];
3919 switch (ConstraintLetter) {
3923 // This set of constraints deal with valid constants for various instructions.
3924 // Validate and return a target constant for them if we can.
3926 // 'z' maps to xzr or wzr so it needs an input of 0.
3927 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3928 if (!C || C->getZExtValue() != 0)
3931 if (Op.getValueType() == MVT::i64)
3932 Result = DAG.getRegister(ARM64::XZR, MVT::i64);
3934 Result = DAG.getRegister(ARM64::WZR, MVT::i32);
3944 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3948 // Grab the value and do some validation.
3949 uint64_t CVal = C->getZExtValue();
3950 switch (ConstraintLetter) {
3951 // The I constraint applies only to simple ADD or SUB immediate operands:
3952 // i.e. 0 to 4095 with optional shift by 12
3953 // The J constraint applies only to ADD or SUB immediates that would be
3954 // valid when negated, i.e. if [an add pattern] were to be output as a SUB
3955 // instruction [or vice versa], in other words -1 to -4095 with optional
3956 // left shift by 12.
3958 if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
3962 uint64_t NVal = -C->getSExtValue();
3963 if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal))
3967 // The K and L constraints apply *only* to logical immediates, including
3968 // what used to be the MOVI alias for ORR (though the MOVI alias has now
3969 // been removed and MOV should be used). So these constraints have to
3970 // distinguish between bit patterns that are valid 32-bit or 64-bit
3971 // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
3972 // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
3975 if (ARM64_AM::isLogicalImmediate(CVal, 32))
3979 if (ARM64_AM::isLogicalImmediate(CVal, 64))
3982 // The M and N constraints are a superset of K and L respectively, for use
3983 // with the MOV (immediate) alias. As well as the logical immediates they
3984 // also match 32 or 64-bit immediates that can be loaded either using a
3985 // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
3986 // (M) or 64-bit 0x1234000000000000 (N) etc.
3987 // As a note some of this code is liberally stolen from the asm parser.
3989 if (!isUInt<32>(CVal))
3991 if (ARM64_AM::isLogicalImmediate(CVal, 32))
3993 if ((CVal & 0xFFFF) == CVal)
3995 if ((CVal & 0xFFFF0000ULL) == CVal)
3997 uint64_t NCVal = ~(uint32_t)CVal;
3998 if ((NCVal & 0xFFFFULL) == NCVal)
4000 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4005 if (ARM64_AM::isLogicalImmediate(CVal, 64))
4007 if ((CVal & 0xFFFFULL) == CVal)
4009 if ((CVal & 0xFFFF0000ULL) == CVal)
4011 if ((CVal & 0xFFFF00000000ULL) == CVal)
4013 if ((CVal & 0xFFFF000000000000ULL) == CVal)
4015 uint64_t NCVal = ~CVal;
4016 if ((NCVal & 0xFFFFULL) == NCVal)
4018 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4020 if ((NCVal & 0xFFFF00000000ULL) == NCVal)
4022 if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
4030 // All assembler immediates are 64-bit integers.
4031 Result = DAG.getTargetConstant(CVal, MVT::i64);
4035 if (Result.getNode()) {
4036 Ops.push_back(Result);
4040 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
4043 //===----------------------------------------------------------------------===//
4044 // ARM64 Advanced SIMD Support
4045 //===----------------------------------------------------------------------===//
4047 /// WidenVector - Given a value in the V64 register class, produce the
4048 /// equivalent value in the V128 register class.
4049 static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
4050 EVT VT = V64Reg.getValueType();
4051 unsigned NarrowSize = VT.getVectorNumElements();
4052 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4053 MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
4056 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
4057 V64Reg, DAG.getConstant(0, MVT::i32));
4060 /// getExtFactor - Determine the adjustment factor for the position when
4061 /// generating an "extract from vector registers" instruction.
4062 static unsigned getExtFactor(SDValue &V) {
4063 EVT EltType = V.getValueType().getVectorElementType();
4064 return EltType.getSizeInBits() / 8;
4067 /// NarrowVector - Given a value in the V128 register class, produce the
4068 /// equivalent value in the V64 register class.
4069 static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) {
4070 EVT VT = V128Reg.getValueType();
4071 unsigned WideSize = VT.getVectorNumElements();
4072 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4073 MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2);
4076 return DAG.getTargetExtractSubreg(ARM64::dsub, DL, NarrowTy, V128Reg);
4079 // Gather data to see if the operation can be modelled as a
4080 // shuffle in combination with VEXTs.
4081 SDValue ARM64TargetLowering::ReconstructShuffle(SDValue Op,
4082 SelectionDAG &DAG) const {
4084 EVT VT = Op.getValueType();
4085 unsigned NumElts = VT.getVectorNumElements();
4087 SmallVector<SDValue, 2> SourceVecs;
4088 SmallVector<unsigned, 2> MinElts;
4089 SmallVector<unsigned, 2> MaxElts;
4091 for (unsigned i = 0; i < NumElts; ++i) {
4092 SDValue V = Op.getOperand(i);
4093 if (V.getOpcode() == ISD::UNDEF)
4095 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4096 // A shuffle can only come from building a vector from various
4097 // elements of other vectors.
4101 // Record this extraction against the appropriate vector if possible...
4102 SDValue SourceVec = V.getOperand(0);
4103 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4104 bool FoundSource = false;
4105 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4106 if (SourceVecs[j] == SourceVec) {
4107 if (MinElts[j] > EltNo)
4109 if (MaxElts[j] < EltNo)
4116 // Or record a new source if not...
4118 SourceVecs.push_back(SourceVec);
4119 MinElts.push_back(EltNo);
4120 MaxElts.push_back(EltNo);
4124 // Currently only do something sane when at most two source vectors
4126 if (SourceVecs.size() > 2)
4129 SDValue ShuffleSrcs[2] = { DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4130 int VEXTOffsets[2] = { 0, 0 };
4132 // This loop extracts the usage patterns of the source vectors
4133 // and prepares appropriate SDValues for a shuffle if possible.
4134 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4135 if (SourceVecs[i].getValueType() == VT) {
4136 // No VEXT necessary
4137 ShuffleSrcs[i] = SourceVecs[i];
4140 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4141 // We can pad out the smaller vector for free, so if it's part of a
4143 ShuffleSrcs[i] = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, SourceVecs[i],
4144 DAG.getUNDEF(SourceVecs[i].getValueType()));
4148 // Don't attempt to extract subvectors from BUILD_VECTOR sources
4149 // that expand or trunc the original value.
4150 // TODO: We can try to bitcast and ANY_EXTEND the result but
4151 // we need to consider the cost of vector ANY_EXTEND, and the
4152 // legality of all the types.
4153 if (SourceVecs[i].getValueType().getVectorElementType() !=
4154 VT.getVectorElementType())
4157 // Since only 64-bit and 128-bit vectors are legal on ARM and
4158 // we've eliminated the other cases...
4159 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2 * NumElts &&
4160 "unexpected vector sizes in ReconstructShuffle");
4162 if (MaxElts[i] - MinElts[i] >= NumElts) {
4163 // Span too large for a VEXT to cope
4167 if (MinElts[i] >= NumElts) {
4168 // The extraction can just take the second half
4169 VEXTOffsets[i] = NumElts;
4171 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
4172 DAG.getIntPtrConstant(NumElts));
4173 } else if (MaxElts[i] < NumElts) {
4174 // The extraction can just take the first half
4176 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4177 SourceVecs[i], DAG.getIntPtrConstant(0));
4179 // An actual VEXT is needed
4180 VEXTOffsets[i] = MinElts[i];
4181 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4182 SourceVecs[i], DAG.getIntPtrConstant(0));
4184 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
4185 DAG.getIntPtrConstant(NumElts));
4186 unsigned Imm = VEXTOffsets[i] * getExtFactor(VEXTSrc1);
4187 ShuffleSrcs[i] = DAG.getNode(ARM64ISD::EXT, dl, VT, VEXTSrc1, VEXTSrc2,
4188 DAG.getConstant(Imm, MVT::i32));
4192 SmallVector<int, 8> Mask;
4194 for (unsigned i = 0; i < NumElts; ++i) {
4195 SDValue Entry = Op.getOperand(i);
4196 if (Entry.getOpcode() == ISD::UNDEF) {
4201 SDValue ExtractVec = Entry.getOperand(0);
4203 cast<ConstantSDNode>(Op.getOperand(i).getOperand(1))->getSExtValue();
4204 if (ExtractVec == SourceVecs[0]) {
4205 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4207 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4211 // Final check before we try to produce nonsense...
4212 if (isShuffleMaskLegal(Mask, VT))
4213 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4219 // check if an EXT instruction can handle the shuffle mask when the
4220 // vector sources of the shuffle are the same.
4221 static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4222 unsigned NumElts = VT.getVectorNumElements();
4224 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4230 // If this is a VEXT shuffle, the immediate value is the index of the first
4231 // element. The other shuffle indices must be the successive elements after
4233 unsigned ExpectedElt = Imm;
4234 for (unsigned i = 1; i < NumElts; ++i) {
4235 // Increment the expected index. If it wraps around, just follow it
4236 // back to index zero and keep going.
4238 if (ExpectedElt == NumElts)
4242 continue; // ignore UNDEF indices
4243 if (ExpectedElt != static_cast<unsigned>(M[i]))
4250 // check if an EXT instruction can handle the shuffle mask when the
4251 // vector sources of the shuffle are different.
4252 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
4254 // Look for the first non-undef element.
4255 const int *FirstRealElt = std::find_if(M.begin(), M.end(),
4256 [](int Elt) {return Elt >= 0;});
4258 // Benefit form APInt to handle overflow when calculating expected element.
4259 unsigned NumElts = VT.getVectorNumElements();
4260 unsigned MaskBits = APInt(32, NumElts * 2).logBase2();
4261 APInt ExpectedElt = APInt(MaskBits, *FirstRealElt + 1);
4262 // The following shuffle indices must be the successive elements after the
4263 // first real element.
4264 const int *FirstWrongElt = std::find_if(FirstRealElt + 1, M.end(),
4265 [&](int Elt) {return Elt != ExpectedElt++ && Elt != -1;});
4266 if (FirstWrongElt != M.end())
4269 // The index of an EXT is the first element if it is not UNDEF.
4270 // Watch out for the beginning UNDEFs. The EXT index should be the expected
4271 // value of the first element.
4272 // E.g. <-1, -1, 3, ...> is treated as <1, 2, 3, ...>.
4273 // <-1, -1, 0, 1, ...> is treated as <IDX, IDX+1, 0, 1, ...>. IDX is
4274 // equal to the ExpectedElt.
4275 Imm = (M[0] >= 0) ? static_cast<unsigned>(M[0]) : ExpectedElt.getZExtValue();
4277 // If no beginning UNDEFs, do swap when M[0] >= NumElts.
4278 if (M[0] >= 0 && Imm >= NumElts) {
4281 } else if (M[0] < 0) {
4282 // Only do swap when beginning UNDEFs more than the first real element,
4283 if (*FirstRealElt < FirstRealElt - M.begin())
4292 /// isREVMask - Check if a vector shuffle corresponds to a REV
4293 /// instruction with the specified blocksize. (The order of the elements
4294 /// within each block of the vector is reversed.)
4295 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4296 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
4297 "Only possible block sizes for REV are: 16, 32, 64");
4299 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4303 unsigned NumElts = VT.getVectorNumElements();
4304 unsigned BlockElts = M[0] + 1;
4305 // If the first shuffle index is UNDEF, be optimistic.
4307 BlockElts = BlockSize / EltSz;
4309 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4312 for (unsigned i = 0; i < NumElts; ++i) {
4314 continue; // ignore UNDEF indices
4315 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
4322 static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4323 unsigned NumElts = VT.getVectorNumElements();
4324 WhichResult = (M[0] == 0 ? 0 : 1);
4325 unsigned Idx = WhichResult * NumElts / 2;
4326 for (unsigned i = 0; i != NumElts; i += 2) {
4327 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4328 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
4336 static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4337 unsigned NumElts = VT.getVectorNumElements();
4338 WhichResult = (M[0] == 0 ? 0 : 1);
4339 for (unsigned i = 0; i != NumElts; ++i) {
4341 continue; // ignore UNDEF indices
4342 if ((unsigned)M[i] != 2 * i + WhichResult)
4349 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4350 unsigned NumElts = VT.getVectorNumElements();
4351 WhichResult = (M[0] == 0 ? 0 : 1);
4352 for (unsigned i = 0; i < NumElts; i += 2) {
4353 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4354 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + NumElts + WhichResult))
4360 /// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of
4361 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4362 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4363 static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4364 unsigned NumElts = VT.getVectorNumElements();
4365 WhichResult = (M[0] == 0 ? 0 : 1);
4366 unsigned Idx = WhichResult * NumElts / 2;
4367 for (unsigned i = 0; i != NumElts; i += 2) {
4368 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4369 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx))
4377 /// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of
4378 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4379 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4380 static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4381 unsigned Half = VT.getVectorNumElements() / 2;
4382 WhichResult = (M[0] == 0 ? 0 : 1);
4383 for (unsigned j = 0; j != 2; ++j) {
4384 unsigned Idx = WhichResult;
4385 for (unsigned i = 0; i != Half; ++i) {
4386 int MIdx = M[i + j * Half];
4387 if (MIdx >= 0 && (unsigned)MIdx != Idx)
4396 /// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of
4397 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4398 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4399 static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4400 unsigned NumElts = VT.getVectorNumElements();
4401 WhichResult = (M[0] == 0 ? 0 : 1);
4402 for (unsigned i = 0; i < NumElts; i += 2) {
4403 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4404 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult))
4410 static bool isINSMask(ArrayRef<int> M, int NumInputElements,
4411 bool &DstIsLeft, int &Anomaly) {
4412 if (M.size() != static_cast<size_t>(NumInputElements))
4415 int NumLHSMatch = 0, NumRHSMatch = 0;
4416 int LastLHSMismatch = -1, LastRHSMismatch = -1;
4418 for (int i = 0; i < NumInputElements; ++i) {
4428 LastLHSMismatch = i;
4430 if (M[i] == i + NumInputElements)
4433 LastRHSMismatch = i;
4436 if (NumLHSMatch == NumInputElements - 1) {
4438 Anomaly = LastLHSMismatch;
4440 } else if (NumRHSMatch == NumInputElements - 1) {
4442 Anomaly = LastRHSMismatch;
4449 static bool isConcatMask(ArrayRef<int> Mask, EVT VT, bool SplitLHS) {
4450 if (VT.getSizeInBits() != 128)
4453 unsigned NumElts = VT.getVectorNumElements();
4455 for (int I = 0, E = NumElts / 2; I != E; I++) {
4460 int Offset = NumElts / 2;
4461 for (int I = NumElts / 2, E = NumElts; I != E; I++) {
4462 if (Mask[I] != I + SplitLHS * Offset)
4469 static SDValue tryFormConcatFromShuffle(SDValue Op, SelectionDAG &DAG) {
4471 EVT VT = Op.getValueType();
4472 SDValue V0 = Op.getOperand(0);
4473 SDValue V1 = Op.getOperand(1);
4474 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op)->getMask();
4476 if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() ||
4477 VT.getVectorElementType() != V1.getValueType().getVectorElementType())
4480 bool SplitV0 = V0.getValueType().getSizeInBits() == 128;
4482 if (!isConcatMask(Mask, VT, SplitV0))
4485 EVT CastVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
4486 VT.getVectorNumElements() / 2);
4488 V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0,
4489 DAG.getConstant(0, MVT::i64));
4491 if (V1.getValueType().getSizeInBits() == 128) {
4492 V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1,
4493 DAG.getConstant(0, MVT::i64));
4495 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1);
4498 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4499 /// the specified operations to build the shuffle.
4500 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4501 SDValue RHS, SelectionDAG &DAG,
4503 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4504 unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1);
4505 unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1);
4508 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4517 OP_VUZPL, // VUZP, left result
4518 OP_VUZPR, // VUZP, right result
4519 OP_VZIPL, // VZIP, left result
4520 OP_VZIPR, // VZIP, right result
4521 OP_VTRNL, // VTRN, left result
4522 OP_VTRNR // VTRN, right result
4525 if (OpNum == OP_COPY) {
4526 if (LHSID == (1 * 9 + 2) * 9 + 3)
4528 assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!");
4532 SDValue OpLHS, OpRHS;
4533 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4534 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4535 EVT VT = OpLHS.getValueType();
4539 llvm_unreachable("Unknown shuffle opcode!");
4541 // VREV divides the vector in half and swaps within the half.
4542 if (VT.getVectorElementType() == MVT::i32 ||
4543 VT.getVectorElementType() == MVT::f32)
4544 return DAG.getNode(ARM64ISD::REV64, dl, VT, OpLHS);
4545 // vrev <4 x i16> -> REV32
4546 if (VT.getVectorElementType() == MVT::i16)
4547 return DAG.getNode(ARM64ISD::REV32, dl, VT, OpLHS);
4548 // vrev <4 x i8> -> REV16
4549 assert(VT.getVectorElementType() == MVT::i8);
4550 return DAG.getNode(ARM64ISD::REV16, dl, VT, OpLHS);
4555 EVT EltTy = VT.getVectorElementType();
4557 if (EltTy == MVT::i8)
4558 Opcode = ARM64ISD::DUPLANE8;
4559 else if (EltTy == MVT::i16)
4560 Opcode = ARM64ISD::DUPLANE16;
4561 else if (EltTy == MVT::i32 || EltTy == MVT::f32)
4562 Opcode = ARM64ISD::DUPLANE32;
4563 else if (EltTy == MVT::i64 || EltTy == MVT::f64)
4564 Opcode = ARM64ISD::DUPLANE64;
4566 llvm_unreachable("Invalid vector element type?");
4568 if (VT.getSizeInBits() == 64)
4569 OpLHS = WidenVector(OpLHS, DAG);
4570 SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, MVT::i64);
4571 return DAG.getNode(Opcode, dl, VT, OpLHS, Lane);
4576 unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
4577 return DAG.getNode(ARM64ISD::EXT, dl, VT, OpLHS, OpRHS,
4578 DAG.getConstant(Imm, MVT::i32));
4581 return DAG.getNode(ARM64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4583 return DAG.getNode(ARM64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4585 return DAG.getNode(ARM64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4587 return DAG.getNode(ARM64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4589 return DAG.getNode(ARM64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4591 return DAG.getNode(ARM64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4595 static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask,
4596 SelectionDAG &DAG) {
4597 // Check to see if we can use the TBL instruction.
4598 SDValue V1 = Op.getOperand(0);
4599 SDValue V2 = Op.getOperand(1);
4602 EVT EltVT = Op.getValueType().getVectorElementType();
4603 unsigned BytesPerElt = EltVT.getSizeInBits() / 8;
4605 SmallVector<SDValue, 8> TBLMask;
4606 for (int Val : ShuffleMask) {
4607 for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
4608 unsigned Offset = Byte + Val * BytesPerElt;
4609 TBLMask.push_back(DAG.getConstant(Offset, MVT::i32));
4613 MVT IndexVT = MVT::v8i8;
4614 unsigned IndexLen = 8;
4615 if (Op.getValueType().getSizeInBits() == 128) {
4616 IndexVT = MVT::v16i8;
4620 SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
4621 SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);
4624 if (V2.getNode()->getOpcode() == ISD::UNDEF) {
4626 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
4627 Shuffle = DAG.getNode(
4628 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4629 DAG.getConstant(Intrinsic::arm64_neon_tbl1, MVT::i32), V1Cst,
4630 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4631 makeArrayRef(TBLMask.data(), IndexLen)));
4633 if (IndexLen == 8) {
4634 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
4635 Shuffle = DAG.getNode(
4636 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4637 DAG.getConstant(Intrinsic::arm64_neon_tbl1, MVT::i32), V1Cst,
4638 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4639 makeArrayRef(TBLMask.data(), IndexLen)));
4641 // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
4642 // cannot currently represent the register constraints on the input
4644 // Shuffle = DAG.getNode(ARM64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
4645 // DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4646 // &TBLMask[0], IndexLen));
4647 Shuffle = DAG.getNode(
4648 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4649 DAG.getConstant(Intrinsic::arm64_neon_tbl2, MVT::i32), V1Cst, V2Cst,
4650 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4651 makeArrayRef(TBLMask.data(), IndexLen)));
4654 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
4657 static unsigned getDUPLANEOp(EVT EltType) {
4658 if (EltType == MVT::i8)
4659 return ARM64ISD::DUPLANE8;
4660 if (EltType == MVT::i16)
4661 return ARM64ISD::DUPLANE16;
4662 if (EltType == MVT::i32 || EltType == MVT::f32)
4663 return ARM64ISD::DUPLANE32;
4664 if (EltType == MVT::i64 || EltType == MVT::f64)
4665 return ARM64ISD::DUPLANE64;
4667 llvm_unreachable("Invalid vector element type?");
4670 SDValue ARM64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4671 SelectionDAG &DAG) const {
4673 EVT VT = Op.getValueType();
4675 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4677 // Convert shuffles that are directly supported on NEON to target-specific
4678 // DAG nodes, instead of keeping them as shuffles and matching them again
4679 // during code selection. This is more efficient and avoids the possibility
4680 // of inconsistencies between legalization and selection.
4681 ArrayRef<int> ShuffleMask = SVN->getMask();
4683 SDValue V1 = Op.getOperand(0);
4684 SDValue V2 = Op.getOperand(1);
4686 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0],
4687 V1.getValueType().getSimpleVT())) {
4688 int Lane = SVN->getSplatIndex();
4689 // If this is undef splat, generate it via "just" vdup, if possible.
4693 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
4694 return DAG.getNode(ARM64ISD::DUP, dl, V1.getValueType(),
4696 // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
4697 // constant. If so, we can just reference the lane's definition directly.
4698 if (V1.getOpcode() == ISD::BUILD_VECTOR &&
4699 !isa<ConstantSDNode>(V1.getOperand(Lane)))
4700 return DAG.getNode(ARM64ISD::DUP, dl, VT, V1.getOperand(Lane));
4702 // Otherwise, duplicate from the lane of the input vector.
4703 unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
4705 // SelectionDAGBuilder may have "helpfully" already extracted or conatenated
4706 // to make a vector of the same size as this SHUFFLE. We can ignore the
4707 // extract entirely, and canonicalise the concat using WidenVector.
4708 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
4709 Lane += cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
4710 V1 = V1.getOperand(0);
4711 } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
4712 unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
4713 Lane -= Idx * VT.getVectorNumElements() / 2;
4714 V1 = WidenVector(V1.getOperand(Idx), DAG);
4715 } else if (VT.getSizeInBits() == 64)
4716 V1 = WidenVector(V1, DAG);
4718 return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, MVT::i64));
4721 if (isREVMask(ShuffleMask, VT, 64))
4722 return DAG.getNode(ARM64ISD::REV64, dl, V1.getValueType(), V1, V2);
4723 if (isREVMask(ShuffleMask, VT, 32))
4724 return DAG.getNode(ARM64ISD::REV32, dl, V1.getValueType(), V1, V2);
4725 if (isREVMask(ShuffleMask, VT, 16))
4726 return DAG.getNode(ARM64ISD::REV16, dl, V1.getValueType(), V1, V2);
4728 bool ReverseEXT = false;
4730 if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
4733 Imm *= getExtFactor(V1);
4734 return DAG.getNode(ARM64ISD::EXT, dl, V1.getValueType(), V1, V2,
4735 DAG.getConstant(Imm, MVT::i32));
4736 } else if (V2->getOpcode() == ISD::UNDEF &&
4737 isSingletonEXTMask(ShuffleMask, VT, Imm)) {
4738 Imm *= getExtFactor(V1);
4739 return DAG.getNode(ARM64ISD::EXT, dl, V1.getValueType(), V1, V1,
4740 DAG.getConstant(Imm, MVT::i32));
4743 unsigned WhichResult;
4744 if (isZIPMask(ShuffleMask, VT, WhichResult)) {
4745 unsigned Opc = (WhichResult == 0) ? ARM64ISD::ZIP1 : ARM64ISD::ZIP2;
4746 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4748 if (isUZPMask(ShuffleMask, VT, WhichResult)) {
4749 unsigned Opc = (WhichResult == 0) ? ARM64ISD::UZP1 : ARM64ISD::UZP2;
4750 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4752 if (isTRNMask(ShuffleMask, VT, WhichResult)) {
4753 unsigned Opc = (WhichResult == 0) ? ARM64ISD::TRN1 : ARM64ISD::TRN2;
4754 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4757 if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4758 unsigned Opc = (WhichResult == 0) ? ARM64ISD::ZIP1 : ARM64ISD::ZIP2;
4759 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4761 if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4762 unsigned Opc = (WhichResult == 0) ? ARM64ISD::UZP1 : ARM64ISD::UZP2;
4763 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4765 if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4766 unsigned Opc = (WhichResult == 0) ? ARM64ISD::TRN1 : ARM64ISD::TRN2;
4767 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4770 SDValue Concat = tryFormConcatFromShuffle(Op, DAG);
4771 if (Concat.getNode())
4776 int NumInputElements = V1.getValueType().getVectorNumElements();
4777 if (isINSMask(ShuffleMask, NumInputElements, DstIsLeft, Anomaly)) {
4778 SDValue DstVec = DstIsLeft ? V1 : V2;
4779 SDValue DstLaneV = DAG.getConstant(Anomaly, MVT::i64);
4781 SDValue SrcVec = V1;
4782 int SrcLane = ShuffleMask[Anomaly];
4783 if (SrcLane >= NumInputElements) {
4785 SrcLane -= VT.getVectorNumElements();
4787 SDValue SrcLaneV = DAG.getConstant(SrcLane, MVT::i64);
4789 EVT ScalarVT = VT.getVectorElementType();
4790 if (ScalarVT.getSizeInBits() < 32)
4791 ScalarVT = MVT::i32;
4794 ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
4795 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ScalarVT, SrcVec, SrcLaneV),
4799 // If the shuffle is not directly supported and it has 4 elements, use
4800 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4801 unsigned NumElts = VT.getVectorNumElements();
4803 unsigned PFIndexes[4];
4804 for (unsigned i = 0; i != 4; ++i) {
4805 if (ShuffleMask[i] < 0)
4808 PFIndexes[i] = ShuffleMask[i];
4811 // Compute the index in the perfect shuffle table.
4812 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
4813 PFIndexes[2] * 9 + PFIndexes[3];
4814 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4815 unsigned Cost = (PFEntry >> 30);
4818 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4821 return GenerateTBL(Op, ShuffleMask, DAG);
4824 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
4826 EVT VT = BVN->getValueType(0);
4827 APInt SplatBits, SplatUndef;
4828 unsigned SplatBitSize;
4830 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4831 unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;
4833 for (unsigned i = 0; i < NumSplats; ++i) {
4834 CnstBits <<= SplatBitSize;
4835 UndefBits <<= SplatBitSize;
4836 CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits());
4837 UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits());
4846 SDValue ARM64TargetLowering::LowerVectorAND(SDValue Op,
4847 SelectionDAG &DAG) const {
4848 BuildVectorSDNode *BVN =
4849 dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
4850 SDValue LHS = Op.getOperand(0);
4852 EVT VT = Op.getValueType();
4857 APInt CnstBits(VT.getSizeInBits(), 0);
4858 APInt UndefBits(VT.getSizeInBits(), 0);
4859 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
4860 // We only have BIC vector immediate instruction, which is and-not.
4861 CnstBits = ~CnstBits;
4863 // We make use of a little bit of goto ickiness in order to avoid having to
4864 // duplicate the immediate matching logic for the undef toggled case.
4865 bool SecondTry = false;
4868 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
4869 CnstBits = CnstBits.zextOrTrunc(64);
4870 uint64_t CnstVal = CnstBits.getZExtValue();
4872 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
4873 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
4874 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4875 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4876 DAG.getConstant(CnstVal, MVT::i32),
4877 DAG.getConstant(0, MVT::i32));
4878 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4881 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
4882 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
4883 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4884 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4885 DAG.getConstant(CnstVal, MVT::i32),
4886 DAG.getConstant(8, MVT::i32));
4887 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4890 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
4891 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
4892 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4893 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4894 DAG.getConstant(CnstVal, MVT::i32),
4895 DAG.getConstant(16, MVT::i32));
4896 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4899 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
4900 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
4901 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4902 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4903 DAG.getConstant(CnstVal, MVT::i32),
4904 DAG.getConstant(24, MVT::i32));
4905 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4908 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
4909 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
4910 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4911 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4912 DAG.getConstant(CnstVal, MVT::i32),
4913 DAG.getConstant(0, MVT::i32));
4914 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4917 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
4918 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
4919 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4920 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4921 DAG.getConstant(CnstVal, MVT::i32),
4922 DAG.getConstant(8, MVT::i32));
4923 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4930 CnstBits = ~UndefBits;
4934 // We can always fall back to a non-immediate AND.
4939 // Specialized code to quickly find if PotentialBVec is a BuildVector that
4940 // consists of only the same constant int value, returned in reference arg
4942 static bool isAllConstantBuildVector(const SDValue &PotentialBVec,
4943 uint64_t &ConstVal) {
4944 BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec);
4947 ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0));
4950 EVT VT = Bvec->getValueType(0);
4951 unsigned NumElts = VT.getVectorNumElements();
4952 for (unsigned i = 1; i < NumElts; ++i)
4953 if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt)
4955 ConstVal = FirstElt->getZExtValue();
4959 static unsigned getIntrinsicID(const SDNode *N) {
4960 unsigned Opcode = N->getOpcode();
4963 return Intrinsic::not_intrinsic;
4964 case ISD::INTRINSIC_WO_CHAIN: {
4965 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4966 if (IID < Intrinsic::num_intrinsics)
4968 return Intrinsic::not_intrinsic;
4973 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
4974 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
4975 // BUILD_VECTORs with constant element C1, C2 is a constant, and C1 == ~C2.
4976 // Also, logical shift right -> sri, with the same structure.
4977 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
4978 EVT VT = N->getValueType(0);
4985 // Is the first op an AND?
4986 const SDValue And = N->getOperand(0);
4987 if (And.getOpcode() != ISD::AND)
4990 // Is the second op an shl or lshr?
4991 SDValue Shift = N->getOperand(1);
4992 // This will have been turned into: ARM64ISD::VSHL vector, #shift
4993 // or ARM64ISD::VLSHR vector, #shift
4994 unsigned ShiftOpc = Shift.getOpcode();
4995 if ((ShiftOpc != ARM64ISD::VSHL && ShiftOpc != ARM64ISD::VLSHR))
4997 bool IsShiftRight = ShiftOpc == ARM64ISD::VLSHR;
4999 // Is the shift amount constant?
5000 ConstantSDNode *C2node = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
5004 // Is the and mask vector all constant?
5006 if (!isAllConstantBuildVector(And.getOperand(1), C1))
5009 // Is C1 == ~C2, taking into account how much one can shift elements of a
5011 uint64_t C2 = C2node->getZExtValue();
5012 unsigned ElemSizeInBits = VT.getVectorElementType().getSizeInBits();
5013 if (C2 > ElemSizeInBits)
5015 unsigned ElemMask = (1 << ElemSizeInBits) - 1;
5016 if ((C1 & ElemMask) != (~C2 & ElemMask))
5019 SDValue X = And.getOperand(0);
5020 SDValue Y = Shift.getOperand(0);
5023 IsShiftRight ? Intrinsic::arm64_neon_vsri : Intrinsic::arm64_neon_vsli;
5025 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5026 DAG.getConstant(Intrin, MVT::i32), X, Y, Shift.getOperand(1));
5028 DEBUG(dbgs() << "arm64-lower: transformed: \n");
5029 DEBUG(N->dump(&DAG));
5030 DEBUG(dbgs() << "into: \n");
5031 DEBUG(ResultSLI->dump(&DAG));
5037 SDValue ARM64TargetLowering::LowerVectorOR(SDValue Op,
5038 SelectionDAG &DAG) const {
5039 // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
5040 if (EnableARM64SlrGeneration) {
5041 SDValue Res = tryLowerToSLI(Op.getNode(), DAG);
5046 BuildVectorSDNode *BVN =
5047 dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode());
5048 SDValue LHS = Op.getOperand(1);
5050 EVT VT = Op.getValueType();
5052 // OR commutes, so try swapping the operands.
5054 LHS = Op.getOperand(0);
5055 BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5060 APInt CnstBits(VT.getSizeInBits(), 0);
5061 APInt UndefBits(VT.getSizeInBits(), 0);
5062 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5063 // We make use of a little bit of goto ickiness in order to avoid having to
5064 // duplicate the immediate matching logic for the undef toggled case.
5065 bool SecondTry = false;
5068 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5069 CnstBits = CnstBits.zextOrTrunc(64);
5070 uint64_t CnstVal = CnstBits.getZExtValue();
5072 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
5073 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
5074 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5075 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5076 DAG.getConstant(CnstVal, MVT::i32),
5077 DAG.getConstant(0, MVT::i32));
5078 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5081 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
5082 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
5083 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5084 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5085 DAG.getConstant(CnstVal, MVT::i32),
5086 DAG.getConstant(8, MVT::i32));
5087 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5090 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
5091 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
5092 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5093 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5094 DAG.getConstant(CnstVal, MVT::i32),
5095 DAG.getConstant(16, MVT::i32));
5096 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5099 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
5100 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
5101 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5102 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5103 DAG.getConstant(CnstVal, MVT::i32),
5104 DAG.getConstant(24, MVT::i32));
5105 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5108 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
5109 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
5110 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5111 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5112 DAG.getConstant(CnstVal, MVT::i32),
5113 DAG.getConstant(0, MVT::i32));
5114 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5117 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
5118 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
5119 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5120 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5121 DAG.getConstant(CnstVal, MVT::i32),
5122 DAG.getConstant(8, MVT::i32));
5123 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5130 CnstBits = UndefBits;
5134 // We can always fall back to a non-immediate OR.
5139 SDValue ARM64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
5140 SelectionDAG &DAG) const {
5141 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5143 EVT VT = Op.getValueType();
5145 APInt CnstBits(VT.getSizeInBits(), 0);
5146 APInt UndefBits(VT.getSizeInBits(), 0);
5147 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5148 // We make use of a little bit of goto ickiness in order to avoid having to
5149 // duplicate the immediate matching logic for the undef toggled case.
5150 bool SecondTry = false;
5153 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5154 CnstBits = CnstBits.zextOrTrunc(64);
5155 uint64_t CnstVal = CnstBits.getZExtValue();
5157 // Certain magic vector constants (used to express things like NOT
5158 // and NEG) are passed through unmodified. This allows codegen patterns
5159 // for these operations to match. Special-purpose patterns will lower
5160 // these immediates to MOVIs if it proves necessary.
5161 if (VT.isInteger() && (CnstVal == 0 || CnstVal == ~0ULL))
5164 // The many faces of MOVI...
5165 if (ARM64_AM::isAdvSIMDModImmType10(CnstVal)) {
5166 CnstVal = ARM64_AM::encodeAdvSIMDModImmType10(CnstVal);
5167 if (VT.getSizeInBits() == 128) {
5168 SDValue Mov = DAG.getNode(ARM64ISD::MOVIedit, dl, MVT::v2i64,
5169 DAG.getConstant(CnstVal, MVT::i32));
5170 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5173 // Support the V64 version via subregister insertion.
5174 SDValue Mov = DAG.getNode(ARM64ISD::MOVIedit, dl, MVT::f64,
5175 DAG.getConstant(CnstVal, MVT::i32));
5176 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5179 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
5180 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
5181 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5182 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5183 DAG.getConstant(CnstVal, MVT::i32),
5184 DAG.getConstant(0, MVT::i32));
5185 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5188 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
5189 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
5190 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5191 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5192 DAG.getConstant(CnstVal, MVT::i32),
5193 DAG.getConstant(8, MVT::i32));
5194 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5197 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
5198 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
5199 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5200 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5201 DAG.getConstant(CnstVal, MVT::i32),
5202 DAG.getConstant(16, MVT::i32));
5203 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5206 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
5207 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
5208 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5209 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5210 DAG.getConstant(CnstVal, MVT::i32),
5211 DAG.getConstant(24, MVT::i32));
5212 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5215 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
5216 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
5217 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5218 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5219 DAG.getConstant(CnstVal, MVT::i32),
5220 DAG.getConstant(0, MVT::i32));
5221 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5224 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
5225 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
5226 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5227 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5228 DAG.getConstant(CnstVal, MVT::i32),
5229 DAG.getConstant(8, MVT::i32));
5230 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5233 if (ARM64_AM::isAdvSIMDModImmType7(CnstVal)) {
5234 CnstVal = ARM64_AM::encodeAdvSIMDModImmType7(CnstVal);
5235 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5236 SDValue Mov = DAG.getNode(ARM64ISD::MOVImsl, dl, MovTy,
5237 DAG.getConstant(CnstVal, MVT::i32),
5238 DAG.getConstant(264, MVT::i32));
5239 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5242 if (ARM64_AM::isAdvSIMDModImmType8(CnstVal)) {
5243 CnstVal = ARM64_AM::encodeAdvSIMDModImmType8(CnstVal);
5244 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5245 SDValue Mov = DAG.getNode(ARM64ISD::MOVImsl, dl, MovTy,
5246 DAG.getConstant(CnstVal, MVT::i32),
5247 DAG.getConstant(272, MVT::i32));
5248 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5251 if (ARM64_AM::isAdvSIMDModImmType9(CnstVal)) {
5252 CnstVal = ARM64_AM::encodeAdvSIMDModImmType9(CnstVal);
5253 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;
5254 SDValue Mov = DAG.getNode(ARM64ISD::MOVI, dl, MovTy,
5255 DAG.getConstant(CnstVal, MVT::i32));
5256 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5259 // The few faces of FMOV...
5260 if (ARM64_AM::isAdvSIMDModImmType11(CnstVal)) {
5261 CnstVal = ARM64_AM::encodeAdvSIMDModImmType11(CnstVal);
5262 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4f32 : MVT::v2f32;
5263 SDValue Mov = DAG.getNode(ARM64ISD::FMOV, dl, MovTy,
5264 DAG.getConstant(CnstVal, MVT::i32));
5265 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5268 if (ARM64_AM::isAdvSIMDModImmType12(CnstVal) &&
5269 VT.getSizeInBits() == 128) {
5270 CnstVal = ARM64_AM::encodeAdvSIMDModImmType12(CnstVal);
5271 SDValue Mov = DAG.getNode(ARM64ISD::FMOV, dl, MVT::v2f64,
5272 DAG.getConstant(CnstVal, MVT::i32));
5273 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5276 // The many faces of MVNI...
5278 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
5279 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
5280 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5281 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5282 DAG.getConstant(CnstVal, MVT::i32),
5283 DAG.getConstant(0, MVT::i32));
5284 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5287 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
5288 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
5289 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5290 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5291 DAG.getConstant(CnstVal, MVT::i32),
5292 DAG.getConstant(8, MVT::i32));
5293 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5296 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
5297 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
5298 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5299 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5300 DAG.getConstant(CnstVal, MVT::i32),
5301 DAG.getConstant(16, MVT::i32));
5302 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5305 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
5306 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
5307 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5308 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5309 DAG.getConstant(CnstVal, MVT::i32),
5310 DAG.getConstant(24, MVT::i32));
5311 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5314 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
5315 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
5316 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5317 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5318 DAG.getConstant(CnstVal, MVT::i32),
5319 DAG.getConstant(0, MVT::i32));
5320 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5323 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
5324 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
5325 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5326 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5327 DAG.getConstant(CnstVal, MVT::i32),
5328 DAG.getConstant(8, MVT::i32));
5329 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5332 if (ARM64_AM::isAdvSIMDModImmType7(CnstVal)) {
5333 CnstVal = ARM64_AM::encodeAdvSIMDModImmType7(CnstVal);
5334 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5335 SDValue Mov = DAG.getNode(ARM64ISD::MVNImsl, dl, MovTy,
5336 DAG.getConstant(CnstVal, MVT::i32),
5337 DAG.getConstant(264, MVT::i32));
5338 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5341 if (ARM64_AM::isAdvSIMDModImmType8(CnstVal)) {
5342 CnstVal = ARM64_AM::encodeAdvSIMDModImmType8(CnstVal);
5343 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5344 SDValue Mov = DAG.getNode(ARM64ISD::MVNImsl, dl, MovTy,
5345 DAG.getConstant(CnstVal, MVT::i32),
5346 DAG.getConstant(272, MVT::i32));
5347 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5354 CnstBits = UndefBits;
5359 // Scan through the operands to find some interesting properties we can
5361 // 1) If only one value is used, we can use a DUP, or
5362 // 2) if only the low element is not undef, we can just insert that, or
5363 // 3) if only one constant value is used (w/ some non-constant lanes),
5364 // we can splat the constant value into the whole vector then fill
5365 // in the non-constant lanes.
5366 // 4) FIXME: If different constant values are used, but we can intelligently
5367 // select the values we'll be overwriting for the non-constant
5368 // lanes such that we can directly materialize the vector
5369 // some other way (MOVI, e.g.), we can be sneaky.
5370 unsigned NumElts = VT.getVectorNumElements();
5371 bool isOnlyLowElement = true;
5372 bool usesOnlyOneValue = true;
5373 bool usesOnlyOneConstantValue = true;
5374 bool isConstant = true;
5375 unsigned NumConstantLanes = 0;
5377 SDValue ConstantValue;
5378 for (unsigned i = 0; i < NumElts; ++i) {
5379 SDValue V = Op.getOperand(i);
5380 if (V.getOpcode() == ISD::UNDEF)
5383 isOnlyLowElement = false;
5384 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5387 if (isa<ConstantSDNode>(V) || isa<ConstantFPSDNode>(V)) {
5389 if (!ConstantValue.getNode())
5391 else if (ConstantValue != V)
5392 usesOnlyOneConstantValue = false;
5395 if (!Value.getNode())
5397 else if (V != Value)
5398 usesOnlyOneValue = false;
5401 if (!Value.getNode())
5402 return DAG.getUNDEF(VT);
5404 if (isOnlyLowElement)
5405 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5407 // Use DUP for non-constant splats. For f32 constant splats, reduce to
5408 // i32 and try again.
5409 if (usesOnlyOneValue) {
5411 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5412 Value.getValueType() != VT)
5413 return DAG.getNode(ARM64ISD::DUP, dl, VT, Value);
5415 // This is actually a DUPLANExx operation, which keeps everything vectory.
5417 // DUPLANE works on 128-bit vectors, widen it if necessary.
5418 SDValue Lane = Value.getOperand(1);
5419 Value = Value.getOperand(0);
5420 if (Value.getValueType().getSizeInBits() == 64)
5421 Value = WidenVector(Value, DAG);
5423 unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
5424 return DAG.getNode(Opcode, dl, VT, Value, Lane);
5427 if (VT.getVectorElementType().isFloatingPoint()) {
5428 SmallVector<SDValue, 8> Ops;
5430 (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
5431 for (unsigned i = 0; i < NumElts; ++i)
5432 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, NewType, Op.getOperand(i)));
5433 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts);
5434 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5435 Val = LowerBUILD_VECTOR(Val, DAG);
5437 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5441 // If there was only one constant value used and for more than one lane,
5442 // start by splatting that value, then replace the non-constant lanes. This
5443 // is better than the default, which will perform a separate initialization
5445 if (NumConstantLanes > 0 && usesOnlyOneConstantValue) {
5446 SDValue Val = DAG.getNode(ARM64ISD::DUP, dl, VT, ConstantValue);
5447 // Now insert the non-constant lanes.
5448 for (unsigned i = 0; i < NumElts; ++i) {
5449 SDValue V = Op.getOperand(i);
5450 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5451 if (!isa<ConstantSDNode>(V) && !isa<ConstantFPSDNode>(V)) {
5452 // Note that type legalization likely mucked about with the VT of the
5453 // source operand, so we may have to convert it here before inserting.
5454 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, V, LaneIdx);
5460 // If all elements are constants and the case above didn't get hit, fall back
5461 // to the default expansion, which will generate a load from the constant
5466 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5468 SDValue shuffle = ReconstructShuffle(Op, DAG);
5469 if (shuffle != SDValue())
5473 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5474 // know the default expansion would otherwise fall back on something even
5475 // worse. For a vector with one or two non-undef values, that's
5476 // scalar_to_vector for the elements followed by a shuffle (provided the
5477 // shuffle is valid for the target) and materialization element by element
5478 // on the stack followed by a load for everything else.
5479 if (!isConstant && !usesOnlyOneValue) {
5480 SDValue Vec = DAG.getUNDEF(VT);
5481 SDValue Op0 = Op.getOperand(0);
5482 unsigned ElemSize = VT.getVectorElementType().getSizeInBits();
5484 // For 32 and 64 bit types, use INSERT_SUBREG for lane zero to
5485 // a) Avoid a RMW dependency on the full vector register, and
5486 // b) Allow the register coalescer to fold away the copy if the
5487 // value is already in an S or D register.
5488 if (Op0.getOpcode() != ISD::UNDEF && (ElemSize == 32 || ElemSize == 64)) {
5489 unsigned SubIdx = ElemSize == 32 ? ARM64::ssub : ARM64::dsub;
5491 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0,
5492 DAG.getTargetConstant(SubIdx, MVT::i32));
5493 Vec = SDValue(N, 0);
5496 for (; i < NumElts; ++i) {
5497 SDValue V = Op.getOperand(i);
5498 if (V.getOpcode() == ISD::UNDEF)
5500 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5501 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5506 // Just use the default expansion. We failed to find a better alternative.
5510 SDValue ARM64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
5511 SelectionDAG &DAG) const {
5512 assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!");
5514 // Check for non-constant lane.
5515 if (!isa<ConstantSDNode>(Op.getOperand(2)))
5518 EVT VT = Op.getOperand(0).getValueType();
5520 // Insertion/extraction are legal for V128 types.
5521 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
5522 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
5525 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
5526 VT != MVT::v1i64 && VT != MVT::v2f32)
5529 // For V64 types, we perform insertion by expanding the value
5530 // to a V128 type and perform the insertion on that.
5532 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
5533 EVT WideTy = WideVec.getValueType();
5535 SDValue Node = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideTy, WideVec,
5536 Op.getOperand(1), Op.getOperand(2));
5537 // Re-narrow the resultant vector.
5538 return NarrowVector(Node, DAG);
5541 SDValue ARM64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5542 SelectionDAG &DAG) const {
5543 assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!");
5545 // Check for non-constant lane.
5546 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5549 EVT VT = Op.getOperand(0).getValueType();
5551 // Insertion/extraction are legal for V128 types.
5552 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
5553 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
5556 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
5557 VT != MVT::v1i64 && VT != MVT::v2f32)
5560 // For V64 types, we perform extraction by expanding the value
5561 // to a V128 type and perform the extraction on that.
5563 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
5564 EVT WideTy = WideVec.getValueType();
5566 EVT ExtrTy = WideTy.getVectorElementType();
5567 if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8)
5570 // For extractions, we just return the result directly.
5571 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec,
5575 SDValue ARM64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
5576 SelectionDAG &DAG) const {
5577 EVT VT = Op.getOperand(0).getValueType();
5583 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5586 unsigned Val = Cst->getZExtValue();
5588 unsigned Size = Op.getValueType().getSizeInBits();
5592 return DAG.getTargetExtractSubreg(ARM64::bsub, dl, Op.getValueType(),
5595 return DAG.getTargetExtractSubreg(ARM64::hsub, dl, Op.getValueType(),
5598 return DAG.getTargetExtractSubreg(ARM64::ssub, dl, Op.getValueType(),
5601 return DAG.getTargetExtractSubreg(ARM64::dsub, dl, Op.getValueType(),
5604 llvm_unreachable("Unexpected vector type in extract_subvector!");
5607 // If this is extracting the upper 64-bits of a 128-bit vector, we match
5609 if (Size == 64 && Val * VT.getVectorElementType().getSizeInBits() == 64)
5615 bool ARM64TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5617 if (VT.getVectorNumElements() == 4 &&
5618 (VT.is128BitVector() || VT.is64BitVector())) {
5619 unsigned PFIndexes[4];
5620 for (unsigned i = 0; i != 4; ++i) {
5624 PFIndexes[i] = M[i];
5627 // Compute the index in the perfect shuffle table.
5628 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
5629 PFIndexes[2] * 9 + PFIndexes[3];
5630 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5631 unsigned Cost = (PFEntry >> 30);
5639 unsigned DummyUnsigned;
5641 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || isREVMask(M, VT, 64) ||
5642 isREVMask(M, VT, 32) || isREVMask(M, VT, 16) ||
5643 isEXTMask(M, VT, DummyBool, DummyUnsigned) ||
5644 // isTBLMask(M, VT) || // FIXME: Port TBL support from ARM.
5645 isTRNMask(M, VT, DummyUnsigned) || isUZPMask(M, VT, DummyUnsigned) ||
5646 isZIPMask(M, VT, DummyUnsigned) ||
5647 isTRN_v_undef_Mask(M, VT, DummyUnsigned) ||
5648 isUZP_v_undef_Mask(M, VT, DummyUnsigned) ||
5649 isZIP_v_undef_Mask(M, VT, DummyUnsigned) ||
5650 isINSMask(M, VT.getVectorNumElements(), DummyBool, DummyInt) ||
5651 isConcatMask(M, VT, VT.getSizeInBits() == 128));
5654 /// getVShiftImm - Check if this is a valid build_vector for the immediate
5655 /// operand of a vector shift operation, where all the elements of the
5656 /// build_vector must have the same constant integer value.
5657 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
5658 // Ignore bit_converts.
5659 while (Op.getOpcode() == ISD::BITCAST)
5660 Op = Op.getOperand(0);
5661 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
5662 APInt SplatBits, SplatUndef;
5663 unsigned SplatBitSize;
5665 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
5666 HasAnyUndefs, ElementBits) ||
5667 SplatBitSize > ElementBits)
5669 Cnt = SplatBits.getSExtValue();
5673 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
5674 /// operand of a vector shift left operation. That value must be in the range:
5675 /// 0 <= Value < ElementBits for a left shift; or
5676 /// 0 <= Value <= ElementBits for a long left shift.
5677 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
5678 assert(VT.isVector() && "vector shift count is not a vector type");
5679 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5680 if (!getVShiftImm(Op, ElementBits, Cnt))
5682 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
5685 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
5686 /// operand of a vector shift right operation. For a shift opcode, the value
5687 /// is positive, but for an intrinsic the value count must be negative. The
5688 /// absolute value must be in the range:
5689 /// 1 <= |Value| <= ElementBits for a right shift; or
5690 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
5691 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
5693 assert(VT.isVector() && "vector shift count is not a vector type");
5694 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5695 if (!getVShiftImm(Op, ElementBits, Cnt))
5699 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
5702 SDValue ARM64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
5703 SelectionDAG &DAG) const {
5704 EVT VT = Op.getValueType();
5708 if (!Op.getOperand(1).getValueType().isVector())
5710 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5712 switch (Op.getOpcode()) {
5714 llvm_unreachable("unexpected shift opcode");
5717 if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
5718 return DAG.getNode(ARM64ISD::VSHL, SDLoc(Op), VT, Op.getOperand(0),
5719 DAG.getConstant(Cnt, MVT::i32));
5720 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5721 DAG.getConstant(Intrinsic::arm64_neon_ushl, MVT::i32),
5722 Op.getOperand(0), Op.getOperand(1));
5725 // Right shift immediate
5726 if (isVShiftRImm(Op.getOperand(1), VT, false, false, Cnt) &&
5729 (Op.getOpcode() == ISD::SRA) ? ARM64ISD::VASHR : ARM64ISD::VLSHR;
5730 return DAG.getNode(Opc, SDLoc(Op), VT, Op.getOperand(0),
5731 DAG.getConstant(Cnt, MVT::i32));
5734 // Right shift register. Note, there is not a shift right register
5735 // instruction, but the shift left register instruction takes a signed
5736 // value, where negative numbers specify a right shift.
5737 unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::arm64_neon_sshl
5738 : Intrinsic::arm64_neon_ushl;
5739 // negate the shift amount
5740 SDValue NegShift = DAG.getNode(ARM64ISD::NEG, DL, VT, Op.getOperand(1));
5741 SDValue NegShiftLeft =
5742 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5743 DAG.getConstant(Opc, MVT::i32), Op.getOperand(0), NegShift);
5744 return NegShiftLeft;
5750 static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS,
5751 ARM64CC::CondCode CC, bool NoNans, EVT VT,
5752 SDLoc dl, SelectionDAG &DAG) {
5753 EVT SrcVT = LHS.getValueType();
5755 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
5756 APInt CnstBits(VT.getSizeInBits(), 0);
5757 APInt UndefBits(VT.getSizeInBits(), 0);
5758 bool IsCnst = BVN && resolveBuildVector(BVN, CnstBits, UndefBits);
5759 bool IsZero = IsCnst && (CnstBits == 0);
5761 if (SrcVT.getVectorElementType().isFloatingPoint()) {
5768 Fcmeq = DAG.getNode(ARM64ISD::FCMEQz, dl, VT, LHS);
5770 Fcmeq = DAG.getNode(ARM64ISD::FCMEQ, dl, VT, LHS, RHS);
5771 return DAG.getNode(ARM64ISD::NOT, dl, VT, Fcmeq);
5775 return DAG.getNode(ARM64ISD::FCMEQz, dl, VT, LHS);
5776 return DAG.getNode(ARM64ISD::FCMEQ, dl, VT, LHS, RHS);
5779 return DAG.getNode(ARM64ISD::FCMGEz, dl, VT, LHS);
5780 return DAG.getNode(ARM64ISD::FCMGE, dl, VT, LHS, RHS);
5783 return DAG.getNode(ARM64ISD::FCMGTz, dl, VT, LHS);
5784 return DAG.getNode(ARM64ISD::FCMGT, dl, VT, LHS, RHS);
5787 return DAG.getNode(ARM64ISD::FCMLEz, dl, VT, LHS);
5788 return DAG.getNode(ARM64ISD::FCMGE, dl, VT, RHS, LHS);
5792 // If we ignore NaNs then we can use to the MI implementation.
5796 return DAG.getNode(ARM64ISD::FCMLTz, dl, VT, LHS);
5797 return DAG.getNode(ARM64ISD::FCMGT, dl, VT, RHS, LHS);
5807 Cmeq = DAG.getNode(ARM64ISD::CMEQz, dl, VT, LHS);
5809 Cmeq = DAG.getNode(ARM64ISD::CMEQ, dl, VT, LHS, RHS);
5810 return DAG.getNode(ARM64ISD::NOT, dl, VT, Cmeq);
5814 return DAG.getNode(ARM64ISD::CMEQz, dl, VT, LHS);
5815 return DAG.getNode(ARM64ISD::CMEQ, dl, VT, LHS, RHS);
5818 return DAG.getNode(ARM64ISD::CMGEz, dl, VT, LHS);
5819 return DAG.getNode(ARM64ISD::CMGE, dl, VT, LHS, RHS);
5822 return DAG.getNode(ARM64ISD::CMGTz, dl, VT, LHS);
5823 return DAG.getNode(ARM64ISD::CMGT, dl, VT, LHS, RHS);
5826 return DAG.getNode(ARM64ISD::CMLEz, dl, VT, LHS);
5827 return DAG.getNode(ARM64ISD::CMGE, dl, VT, RHS, LHS);
5829 return DAG.getNode(ARM64ISD::CMHS, dl, VT, RHS, LHS);
5831 return DAG.getNode(ARM64ISD::CMHI, dl, VT, RHS, LHS);
5834 return DAG.getNode(ARM64ISD::CMLTz, dl, VT, LHS);
5835 return DAG.getNode(ARM64ISD::CMGT, dl, VT, RHS, LHS);
5837 return DAG.getNode(ARM64ISD::CMHI, dl, VT, LHS, RHS);
5839 return DAG.getNode(ARM64ISD::CMHS, dl, VT, LHS, RHS);
5843 SDValue ARM64TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
5844 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5845 SDValue LHS = Op.getOperand(0);
5846 SDValue RHS = Op.getOperand(1);
5849 if (LHS.getValueType().getVectorElementType().isInteger()) {
5850 assert(LHS.getValueType() == RHS.getValueType());
5851 ARM64CC::CondCode ARM64CC = changeIntCCToARM64CC(CC);
5852 return EmitVectorComparison(LHS, RHS, ARM64CC, false, Op.getValueType(), dl,
5856 assert(LHS.getValueType().getVectorElementType() == MVT::f32 ||
5857 LHS.getValueType().getVectorElementType() == MVT::f64);
5859 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
5860 // clean. Some of them require two branches to implement.
5861 ARM64CC::CondCode CC1, CC2;
5863 changeVectorFPCCToARM64CC(CC, CC1, CC2, ShouldInvert);
5865 bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath;
5867 EmitVectorComparison(LHS, RHS, CC1, NoNaNs, Op.getValueType(), dl, DAG);
5871 if (CC2 != ARM64CC::AL) {
5873 EmitVectorComparison(LHS, RHS, CC2, NoNaNs, Op.getValueType(), dl, DAG);
5874 if (!Cmp2.getNode())
5877 Cmp = DAG.getNode(ISD::OR, dl, Cmp.getValueType(), Cmp, Cmp2);
5881 return Cmp = DAG.getNOT(dl, Cmp, Cmp.getValueType());
5886 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
5887 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
5888 /// specified in the intrinsic calls.
5889 bool ARM64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
5891 unsigned Intrinsic) const {
5892 switch (Intrinsic) {
5893 case Intrinsic::arm64_neon_ld2:
5894 case Intrinsic::arm64_neon_ld3:
5895 case Intrinsic::arm64_neon_ld4:
5896 case Intrinsic::arm64_neon_ld1x2:
5897 case Intrinsic::arm64_neon_ld1x3:
5898 case Intrinsic::arm64_neon_ld1x4:
5899 case Intrinsic::arm64_neon_ld2lane:
5900 case Intrinsic::arm64_neon_ld3lane:
5901 case Intrinsic::arm64_neon_ld4lane:
5902 case Intrinsic::arm64_neon_ld2r:
5903 case Intrinsic::arm64_neon_ld3r:
5904 case Intrinsic::arm64_neon_ld4r: {
5905 Info.opc = ISD::INTRINSIC_W_CHAIN;
5906 // Conservatively set memVT to the entire set of vectors loaded.
5907 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
5908 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5909 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
5912 Info.vol = false; // volatile loads with NEON intrinsics not supported
5913 Info.readMem = true;
5914 Info.writeMem = false;
5917 case Intrinsic::arm64_neon_st2:
5918 case Intrinsic::arm64_neon_st3:
5919 case Intrinsic::arm64_neon_st4:
5920 case Intrinsic::arm64_neon_st1x2:
5921 case Intrinsic::arm64_neon_st1x3:
5922 case Intrinsic::arm64_neon_st1x4:
5923 case Intrinsic::arm64_neon_st2lane:
5924 case Intrinsic::arm64_neon_st3lane:
5925 case Intrinsic::arm64_neon_st4lane: {
5926 Info.opc = ISD::INTRINSIC_VOID;
5927 // Conservatively set memVT to the entire set of vectors stored.
5928 unsigned NumElts = 0;
5929 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
5930 Type *ArgTy = I.getArgOperand(ArgI)->getType();
5931 if (!ArgTy->isVectorTy())
5933 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
5935 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5936 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
5939 Info.vol = false; // volatile stores with NEON intrinsics not supported
5940 Info.readMem = false;
5941 Info.writeMem = true;
5944 case Intrinsic::arm64_ldaxr:
5945 case Intrinsic::arm64_ldxr: {
5946 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
5947 Info.opc = ISD::INTRINSIC_W_CHAIN;
5948 Info.memVT = MVT::getVT(PtrTy->getElementType());
5949 Info.ptrVal = I.getArgOperand(0);
5951 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
5953 Info.readMem = true;
5954 Info.writeMem = false;
5957 case Intrinsic::arm64_stlxr:
5958 case Intrinsic::arm64_stxr: {
5959 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
5960 Info.opc = ISD::INTRINSIC_W_CHAIN;
5961 Info.memVT = MVT::getVT(PtrTy->getElementType());
5962 Info.ptrVal = I.getArgOperand(1);
5964 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
5966 Info.readMem = false;
5967 Info.writeMem = true;
5970 case Intrinsic::arm64_ldaxp:
5971 case Intrinsic::arm64_ldxp: {
5972 Info.opc = ISD::INTRINSIC_W_CHAIN;
5973 Info.memVT = MVT::i128;
5974 Info.ptrVal = I.getArgOperand(0);
5978 Info.readMem = true;
5979 Info.writeMem = false;
5982 case Intrinsic::arm64_stlxp:
5983 case Intrinsic::arm64_stxp: {
5984 Info.opc = ISD::INTRINSIC_W_CHAIN;
5985 Info.memVT = MVT::i128;
5986 Info.ptrVal = I.getArgOperand(2);
5990 Info.readMem = false;
5991 Info.writeMem = true;
6001 // Truncations from 64-bit GPR to 32-bit GPR is free.
6002 bool ARM64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
6003 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6005 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6006 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6007 if (NumBits1 <= NumBits2)
6011 bool ARM64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
6012 if (!VT1.isInteger() || !VT2.isInteger())
6014 unsigned NumBits1 = VT1.getSizeInBits();
6015 unsigned NumBits2 = VT2.getSizeInBits();
6016 if (NumBits1 <= NumBits2)
6021 // All 32-bit GPR operations implicitly zero the high-half of the corresponding
6023 bool ARM64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
6024 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6026 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6027 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6028 if (NumBits1 == 32 && NumBits2 == 64)
6032 bool ARM64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
6033 if (!VT1.isInteger() || !VT2.isInteger())
6035 unsigned NumBits1 = VT1.getSizeInBits();
6036 unsigned NumBits2 = VT2.getSizeInBits();
6037 if (NumBits1 == 32 && NumBits2 == 64)
6042 bool ARM64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
6043 EVT VT1 = Val.getValueType();
6044 if (isZExtFree(VT1, VT2)) {
6048 if (Val.getOpcode() != ISD::LOAD)
6051 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
6052 return (VT1.isSimple() && VT1.isInteger() && VT2.isSimple() &&
6053 VT2.isInteger() && VT1.getSizeInBits() <= 32);
6056 bool ARM64TargetLowering::hasPairedLoad(Type *LoadedType,
6057 unsigned &RequiredAligment) const {
6058 if (!LoadedType->isIntegerTy() && !LoadedType->isFloatTy())
6060 // Cyclone supports unaligned accesses.
6061 RequiredAligment = 0;
6062 unsigned NumBits = LoadedType->getPrimitiveSizeInBits();
6063 return NumBits == 32 || NumBits == 64;
6066 bool ARM64TargetLowering::hasPairedLoad(EVT LoadedType,
6067 unsigned &RequiredAligment) const {
6068 if (!LoadedType.isSimple() ||
6069 (!LoadedType.isInteger() && !LoadedType.isFloatingPoint()))
6071 // Cyclone supports unaligned accesses.
6072 RequiredAligment = 0;
6073 unsigned NumBits = LoadedType.getSizeInBits();
6074 return NumBits == 32 || NumBits == 64;
6077 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
6078 unsigned AlignCheck) {
6079 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
6080 (DstAlign == 0 || DstAlign % AlignCheck == 0));
6083 EVT ARM64TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
6084 unsigned SrcAlign, bool IsMemset,
6085 bool ZeroMemset, bool MemcpyStrSrc,
6086 MachineFunction &MF) const {
6087 // Don't use AdvSIMD to implement 16-byte memset. It would have taken one
6088 // instruction to materialize the v2i64 zero and one store (with restrictive
6089 // addressing mode). Just do two i64 store of zero-registers.
6091 const Function *F = MF.getFunction();
6092 if (Subtarget->hasFPARMv8() && !IsMemset && Size >= 16 &&
6093 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
6094 Attribute::NoImplicitFloat) &&
6095 (memOpAlign(SrcAlign, DstAlign, 16) ||
6096 (allowsUnalignedMemoryAccesses(MVT::f128, 0, &Fast) && Fast)))
6099 return Size >= 8 ? MVT::i64 : MVT::i32;
6102 // 12-bit optionally shifted immediates are legal for adds.
6103 bool ARM64TargetLowering::isLegalAddImmediate(int64_t Immed) const {
6104 if ((Immed >> 12) == 0 || ((Immed & 0xfff) == 0 && Immed >> 24 == 0))
6109 // Integer comparisons are implemented with ADDS/SUBS, so the range of valid
6110 // immediates is the same as for an add or a sub.
6111 bool ARM64TargetLowering::isLegalICmpImmediate(int64_t Immed) const {
6114 return isLegalAddImmediate(Immed);
6117 /// isLegalAddressingMode - Return true if the addressing mode represented
6118 /// by AM is legal for this target, for a load/store of the specified type.
6119 bool ARM64TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6121 // ARM64 has five basic addressing modes:
6123 // reg + 9-bit signed offset
6124 // reg + SIZE_IN_BYTES * 12-bit unsigned offset
6126 // reg + SIZE_IN_BYTES * reg
6128 // No global is ever allowed as a base.
6132 // No reg+reg+imm addressing.
6133 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
6136 // check reg + imm case:
6137 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
6138 uint64_t NumBytes = 0;
6139 if (Ty->isSized()) {
6140 uint64_t NumBits = getDataLayout()->getTypeSizeInBits(Ty);
6141 NumBytes = NumBits / 8;
6142 if (!isPowerOf2_64(NumBits))
6147 int64_t Offset = AM.BaseOffs;
6149 // 9-bit signed offset
6150 if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
6153 // 12-bit unsigned offset
6154 unsigned shift = Log2_64(NumBytes);
6155 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
6156 // Must be a multiple of NumBytes (NumBytes is a power of 2)
6157 (Offset >> shift) << shift == Offset)
6162 // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2
6164 if (!AM.Scale || AM.Scale == 1 ||
6165 (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
6170 int ARM64TargetLowering::getScalingFactorCost(const AddrMode &AM,
6172 // Scaling factors are not free at all.
6173 // Operands | Rt Latency
6174 // -------------------------------------------
6176 // -------------------------------------------
6177 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
6178 // Rt, [Xn, Wm, <extend> #imm] |
6179 if (isLegalAddressingMode(AM, Ty))
6180 // Scale represents reg2 * scale, thus account for 1 if
6181 // it is not equal to 0 or 1.
6182 return AM.Scale != 0 && AM.Scale != 1;
6186 bool ARM64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
6187 VT = VT.getScalarType();
6192 switch (VT.getSimpleVT().SimpleTy) {
6204 ARM64TargetLowering::getScratchRegisters(CallingConv::ID) const {
6205 // LR is a callee-save register, but we must treat it as clobbered by any call
6206 // site. Hence we include LR in the scratch registers, which are in turn added
6207 // as implicit-defs for stackmaps and patchpoints.
6208 static const MCPhysReg ScratchRegs[] = {
6209 ARM64::X16, ARM64::X17, ARM64::LR, 0
6214 bool ARM64TargetLowering::isDesirableToCommuteWithShift(const SDNode *N) const {
6215 EVT VT = N->getValueType(0);
6216 // If N is unsigned bit extraction: ((x >> C) & mask), then do not combine
6217 // it with shift to let it be lowered to UBFX.
6218 if (N->getOpcode() == ISD::AND && (VT == MVT::i32 || VT == MVT::i64) &&
6219 isa<ConstantSDNode>(N->getOperand(1))) {
6220 uint64_t TruncMask = N->getConstantOperandVal(1);
6221 if (isMask_64(TruncMask) &&
6222 N->getOperand(0).getOpcode() == ISD::SRL &&
6223 isa<ConstantSDNode>(N->getOperand(0)->getOperand(1)))
6229 bool ARM64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
6231 assert(Ty->isIntegerTy());
6233 unsigned BitSize = Ty->getPrimitiveSizeInBits();
6237 int64_t Val = Imm.getSExtValue();
6238 if (Val == 0 || ARM64_AM::isLogicalImmediate(Val, BitSize))
6241 if ((int64_t)Val < 0)
6244 Val &= (1LL << 32) - 1;
6246 unsigned LZ = countLeadingZeros((uint64_t)Val);
6247 unsigned Shift = (63 - LZ) / 16;
6248 // MOVZ is free so return true for one or fewer MOVK.
6249 return (Shift < 3) ? true : false;
6252 // Generate SUBS and CSEL for integer abs.
6253 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
6254 EVT VT = N->getValueType(0);
6256 SDValue N0 = N->getOperand(0);
6257 SDValue N1 = N->getOperand(1);
6260 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
6261 // and change it to SUB and CSEL.
6262 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
6263 N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
6264 N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0))
6265 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
6266 if (Y1C->getAPIntValue() == VT.getSizeInBits() - 1) {
6267 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
6269 // Generate SUBS & CSEL.
6271 DAG.getNode(ARM64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32),
6272 N0.getOperand(0), DAG.getConstant(0, VT));
6273 return DAG.getNode(ARM64ISD::CSEL, DL, VT, N0.getOperand(0), Neg,
6274 DAG.getConstant(ARM64CC::PL, MVT::i32),
6275 SDValue(Cmp.getNode(), 1));
6280 // performXorCombine - Attempts to handle integer ABS.
6281 static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
6282 TargetLowering::DAGCombinerInfo &DCI,
6283 const ARM64Subtarget *Subtarget) {
6284 if (DCI.isBeforeLegalizeOps())
6287 return performIntegerAbsCombine(N, DAG);
6290 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG,
6291 TargetLowering::DAGCombinerInfo &DCI,
6292 const ARM64Subtarget *Subtarget) {
6293 if (DCI.isBeforeLegalizeOps())
6296 // Multiplication of a power of two plus/minus one can be done more
6297 // cheaply as as shift+add/sub. For now, this is true unilaterally. If
6298 // future CPUs have a cheaper MADD instruction, this may need to be
6299 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
6300 // 64-bit is 5 cycles, so this is always a win.
6301 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
6302 APInt Value = C->getAPIntValue();
6303 EVT VT = N->getValueType(0);
6304 APInt VP1 = Value + 1;
6305 if (VP1.isPowerOf2()) {
6306 // Multiplying by one less than a power of two, replace with a shift
6308 SDValue ShiftedVal =
6309 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6310 DAG.getConstant(VP1.logBase2(), MVT::i64));
6311 return DAG.getNode(ISD::SUB, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
6313 APInt VM1 = Value - 1;
6314 if (VM1.isPowerOf2()) {
6315 // Multiplying by one more than a power of two, replace with a shift
6317 SDValue ShiftedVal =
6318 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6319 DAG.getConstant(VM1.logBase2(), MVT::i64));
6320 return DAG.getNode(ISD::ADD, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
6326 static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG) {
6327 EVT VT = N->getValueType(0);
6328 if (VT != MVT::f32 && VT != MVT::f64)
6330 // Only optimize when the source and destination types have the same width.
6331 if (VT.getSizeInBits() != N->getOperand(0).getValueType().getSizeInBits())
6334 // If the result of an integer load is only used by an integer-to-float
6335 // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
6336 // This eliminates an "integer-to-vector-move UOP and improve throughput.
6337 SDValue N0 = N->getOperand(0);
6338 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6339 // Do not change the width of a volatile load.
6340 !cast<LoadSDNode>(N0)->isVolatile()) {
6341 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6342 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
6343 LN0->getPointerInfo(), LN0->isVolatile(),
6344 LN0->isNonTemporal(), LN0->isInvariant(),
6345 LN0->getAlignment());
6347 // Make sure successors of the original load stay after it by updating them
6348 // to use the new Chain.
6349 DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));
6352 (N->getOpcode() == ISD::SINT_TO_FP) ? ARM64ISD::SITOF : ARM64ISD::UITOF;
6353 return DAG.getNode(Opcode, SDLoc(N), VT, Load);
6359 /// An EXTR instruction is made up of two shifts, ORed together. This helper
6360 /// searches for and classifies those shifts.
6361 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
6363 if (N.getOpcode() == ISD::SHL)
6365 else if (N.getOpcode() == ISD::SRL)
6370 if (!isa<ConstantSDNode>(N.getOperand(1)))
6373 ShiftAmount = N->getConstantOperandVal(1);
6374 Src = N->getOperand(0);
6378 /// EXTR instruction extracts a contiguous chunk of bits from two existing
6379 /// registers viewed as a high/low pair. This function looks for the pattern:
6380 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
6381 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
6383 static SDValue tryCombineToEXTR(SDNode *N,
6384 TargetLowering::DAGCombinerInfo &DCI) {
6385 SelectionDAG &DAG = DCI.DAG;
6387 EVT VT = N->getValueType(0);
6389 assert(N->getOpcode() == ISD::OR && "Unexpected root");
6391 if (VT != MVT::i32 && VT != MVT::i64)
6395 uint32_t ShiftLHS = 0;
6397 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
6401 uint32_t ShiftRHS = 0;
6403 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
6406 // If they're both trying to come from the high part of the register, they're
6407 // not really an EXTR.
6408 if (LHSFromHi == RHSFromHi)
6411 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
6415 std::swap(LHS, RHS);
6416 std::swap(ShiftLHS, ShiftRHS);
6419 return DAG.getNode(ARM64ISD::EXTR, DL, VT, LHS, RHS,
6420 DAG.getConstant(ShiftRHS, MVT::i64));
6423 static SDValue tryCombineToBSL(SDNode *N,
6424 TargetLowering::DAGCombinerInfo &DCI) {
6425 EVT VT = N->getValueType(0);
6426 SelectionDAG &DAG = DCI.DAG;
6432 SDValue N0 = N->getOperand(0);
6433 if (N0.getOpcode() != ISD::AND)
6436 SDValue N1 = N->getOperand(1);
6437 if (N1.getOpcode() != ISD::AND)
6440 // We only have to look for constant vectors here since the general, variable
6441 // case can be handled in TableGen.
6442 unsigned Bits = VT.getVectorElementType().getSizeInBits();
6443 uint64_t BitMask = Bits == 64 ? -1ULL : ((1ULL << Bits) - 1);
6444 for (int i = 1; i >= 0; --i)
6445 for (int j = 1; j >= 0; --j) {
6446 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(i));
6447 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(j));
6451 bool FoundMatch = true;
6452 for (unsigned k = 0; k < VT.getVectorNumElements(); ++k) {
6453 ConstantSDNode *CN0 = dyn_cast<ConstantSDNode>(BVN0->getOperand(k));
6454 ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(BVN1->getOperand(k));
6456 CN0->getZExtValue() != (BitMask & ~CN1->getZExtValue())) {
6463 return DAG.getNode(ARM64ISD::BSL, DL, VT, SDValue(BVN0, 0),
6464 N0->getOperand(1 - i), N1->getOperand(1 - j));
6470 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
6471 const ARM64Subtarget *Subtarget) {
6472 // Attempt to form an EXTR from (or (shl VAL1, #N), (srl VAL2, #RegWidth-N))
6473 if (!EnableARM64ExtrGeneration)
6475 SelectionDAG &DAG = DCI.DAG;
6476 EVT VT = N->getValueType(0);
6478 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6481 SDValue Res = tryCombineToEXTR(N, DCI);
6485 Res = tryCombineToBSL(N, DCI);
6492 static SDValue performBitcastCombine(SDNode *N,
6493 TargetLowering::DAGCombinerInfo &DCI,
6494 SelectionDAG &DAG) {
6495 // Wait 'til after everything is legalized to try this. That way we have
6496 // legal vector types and such.
6497 if (DCI.isBeforeLegalizeOps())
6500 // Remove extraneous bitcasts around an extract_subvector.
6502 // (v4i16 (bitconvert
6503 // (extract_subvector (v2i64 (bitconvert (v8i16 ...)), (i64 1)))))
6505 // (extract_subvector ((v8i16 ...), (i64 4)))
6507 // Only interested in 64-bit vectors as the ultimate result.
6508 EVT VT = N->getValueType(0);
6511 if (VT.getSimpleVT().getSizeInBits() != 64)
6513 // Is the operand an extract_subvector starting at the beginning or halfway
6514 // point of the vector? A low half may also come through as an
6515 // EXTRACT_SUBREG, so look for that, too.
6516 SDValue Op0 = N->getOperand(0);
6517 if (Op0->getOpcode() != ISD::EXTRACT_SUBVECTOR &&
6518 !(Op0->isMachineOpcode() &&
6519 Op0->getMachineOpcode() == ARM64::EXTRACT_SUBREG))
6521 uint64_t idx = cast<ConstantSDNode>(Op0->getOperand(1))->getZExtValue();
6522 if (Op0->getOpcode() == ISD::EXTRACT_SUBVECTOR) {
6523 if (Op0->getValueType(0).getVectorNumElements() != idx && idx != 0)
6525 } else if (Op0->getMachineOpcode() == ARM64::EXTRACT_SUBREG) {
6526 if (idx != ARM64::dsub)
6528 // The dsub reference is equivalent to a lane zero subvector reference.
6531 // Look through the bitcast of the input to the extract.
6532 if (Op0->getOperand(0)->getOpcode() != ISD::BITCAST)
6534 SDValue Source = Op0->getOperand(0)->getOperand(0);
6535 // If the source type has twice the number of elements as our destination
6536 // type, we know this is an extract of the high or low half of the vector.
6537 EVT SVT = Source->getValueType(0);
6538 if (SVT.getVectorNumElements() != VT.getVectorNumElements() * 2)
6541 DEBUG(dbgs() << "arm64-lower: bitcast extract_subvector simplification\n");
6543 // Create the simplified form to just extract the low or high half of the
6544 // vector directly rather than bothering with the bitcasts.
6546 unsigned NumElements = VT.getVectorNumElements();
6548 SDValue HalfIdx = DAG.getConstant(NumElements, MVT::i64);
6549 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Source, HalfIdx);
6551 SDValue SubReg = DAG.getTargetConstant(ARM64::dsub, MVT::i32);
6552 return SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl, VT,
6558 static SDValue performConcatVectorsCombine(SDNode *N,
6559 TargetLowering::DAGCombinerInfo &DCI,
6560 SelectionDAG &DAG) {
6561 // Wait 'til after everything is legalized to try this. That way we have
6562 // legal vector types and such.
6563 if (DCI.isBeforeLegalizeOps())
6567 EVT VT = N->getValueType(0);
6569 // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector
6570 // splat. The indexed instructions are going to be expecting a DUPLANE64, so
6571 // canonicalise to that.
6572 if (N->getOperand(0) == N->getOperand(1) && VT.getVectorNumElements() == 2) {
6573 assert(VT.getVectorElementType().getSizeInBits() == 64);
6574 return DAG.getNode(ARM64ISD::DUPLANE64, dl, VT,
6575 WidenVector(N->getOperand(0), DAG),
6576 DAG.getConstant(0, MVT::i64));
6579 // Canonicalise concat_vectors so that the right-hand vector has as few
6580 // bit-casts as possible before its real operation. The primary matching
6581 // destination for these operations will be the narrowing "2" instructions,
6582 // which depend on the operation being performed on this right-hand vector.
6584 // (concat_vectors LHS, (v1i64 (bitconvert (v4i16 RHS))))
6586 // (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS))
6588 SDValue Op1 = N->getOperand(1);
6589 if (Op1->getOpcode() != ISD::BITCAST)
6591 SDValue RHS = Op1->getOperand(0);
6592 MVT RHSTy = RHS.getValueType().getSimpleVT();
6593 // If the RHS is not a vector, this is not the pattern we're looking for.
6594 if (!RHSTy.isVector())
6597 DEBUG(dbgs() << "arm64-lower: concat_vectors bitcast simplification\n");
6599 MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(),
6600 RHSTy.getVectorNumElements() * 2);
6602 ISD::BITCAST, dl, VT,
6603 DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatTy,
6604 DAG.getNode(ISD::BITCAST, dl, RHSTy, N->getOperand(0)), RHS));
6607 static SDValue tryCombineFixedPointConvert(SDNode *N,
6608 TargetLowering::DAGCombinerInfo &DCI,
6609 SelectionDAG &DAG) {
6610 // Wait 'til after everything is legalized to try this. That way we have
6611 // legal vector types and such.
6612 if (DCI.isBeforeLegalizeOps())
6614 // Transform a scalar conversion of a value from a lane extract into a
6615 // lane extract of a vector conversion. E.g., from foo1 to foo2:
6616 // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); }
6617 // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; }
6619 // The second form interacts better with instruction selection and the
6620 // register allocator to avoid cross-class register copies that aren't
6621 // coalescable due to a lane reference.
6623 // Check the operand and see if it originates from a lane extract.
6624 SDValue Op1 = N->getOperand(1);
6625 if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6626 // Yep, no additional predication needed. Perform the transform.
6627 SDValue IID = N->getOperand(0);
6628 SDValue Shift = N->getOperand(2);
6629 SDValue Vec = Op1.getOperand(0);
6630 SDValue Lane = Op1.getOperand(1);
6631 EVT ResTy = N->getValueType(0);
6635 // The vector width should be 128 bits by the time we get here, even
6636 // if it started as 64 bits (the extract_vector handling will have
6638 assert(Vec.getValueType().getSizeInBits() == 128 &&
6639 "unexpected vector size on extract_vector_elt!");
6640 if (Vec.getValueType() == MVT::v4i32)
6641 VecResTy = MVT::v4f32;
6642 else if (Vec.getValueType() == MVT::v2i64)
6643 VecResTy = MVT::v2f64;
6645 assert(0 && "unexpected vector type!");
6648 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift);
6649 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane);
6654 // AArch64 high-vector "long" operations are formed by performing the non-high
6655 // version on an extract_subvector of each operand which gets the high half:
6657 // (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS))
6659 // However, there are cases which don't have an extract_high explicitly, but
6660 // have another operation that can be made compatible with one for free. For
6663 // (dupv64 scalar) --> (extract_high (dup128 scalar))
6665 // This routine does the actual conversion of such DUPs, once outer routines
6666 // have determined that everything else is in order.
6667 static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) {
6668 // We can handle most types of duplicate, but the lane ones have an extra
6669 // operand saying *which* lane, so we need to know.
6671 switch (N.getOpcode()) {
6675 case ARM64ISD::DUPLANE8:
6676 case ARM64ISD::DUPLANE16:
6677 case ARM64ISD::DUPLANE32:
6678 case ARM64ISD::DUPLANE64:
6685 MVT NarrowTy = N.getSimpleValueType();
6686 if (!NarrowTy.is64BitVector())
6689 MVT ElementTy = NarrowTy.getVectorElementType();
6690 unsigned NumElems = NarrowTy.getVectorNumElements();
6691 MVT NewDUPVT = MVT::getVectorVT(ElementTy, NumElems * 2);
6695 NewDUP = DAG.getNode(N.getOpcode(), SDLoc(N), NewDUPVT, N.getOperand(0),
6698 NewDUP = DAG.getNode(ARM64ISD::DUP, SDLoc(N), NewDUPVT, N.getOperand(0));
6700 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N.getNode()), NarrowTy,
6701 NewDUP, DAG.getConstant(NumElems, MVT::i64));
6704 static bool isEssentiallyExtractSubvector(SDValue N) {
6705 if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR)
6708 return N.getOpcode() == ISD::BITCAST &&
6709 N.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR;
6712 /// \brief Helper structure to keep track of ISD::SET_CC operands.
6713 struct GenericSetCCInfo {
6714 const SDValue *Opnd0;
6715 const SDValue *Opnd1;
6719 /// \brief Helper structure to keep track of a SET_CC lowered into ARM64 code.
6720 struct ARM64SetCCInfo {
6722 ARM64CC::CondCode CC;
6725 /// \brief Helper structure to keep track of SetCC information.
6727 GenericSetCCInfo Generic;
6728 ARM64SetCCInfo ARM64;
6731 /// \brief Helper structure to be able to read SetCC information.
6732 /// If set to true, IsARM64 field, Info is a ARM64SetCCInfo, otherwise Info is
6733 /// a GenericSetCCInfo.
6734 struct SetCCInfoAndKind {
6739 /// \brief Check whether or not \p Op is a SET_CC operation, either a generic or
6741 /// ARM64 lowered one.
6742 /// \p SetCCInfo is filled accordingly.
6743 /// \post SetCCInfo is meanginfull only when this function returns true.
6744 /// \return True when Op is a kind of SET_CC operation.
6745 static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) {
6746 // If this is a setcc, this is straight forward.
6747 if (Op.getOpcode() == ISD::SETCC) {
6748 SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0);
6749 SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1);
6750 SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6751 SetCCInfo.IsARM64 = false;
6754 // Otherwise, check if this is a matching csel instruction.
6758 if (Op.getOpcode() != ARM64ISD::CSEL)
6760 // Set the information about the operands.
6761 // TODO: we want the operands of the Cmp not the csel
6762 SetCCInfo.Info.ARM64.Cmp = &Op.getOperand(3);
6763 SetCCInfo.IsARM64 = true;
6764 SetCCInfo.Info.ARM64.CC = static_cast<ARM64CC::CondCode>(
6765 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
6767 // Check that the operands matches the constraints:
6768 // (1) Both operands must be constants.
6769 // (2) One must be 1 and the other must be 0.
6770 ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0));
6771 ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6774 if (!TValue || !FValue)
6778 if (!TValue->isOne()) {
6779 // Update the comparison when we are interested in !cc.
6780 std::swap(TValue, FValue);
6781 SetCCInfo.Info.ARM64.CC =
6782 ARM64CC::getInvertedCondCode(SetCCInfo.Info.ARM64.CC);
6784 return TValue->isOne() && FValue->isNullValue();
6787 // Returns true if Op is setcc or zext of setcc.
6788 static bool isSetCCOrZExtSetCC(const SDValue& Op, SetCCInfoAndKind &Info) {
6789 if (isSetCC(Op, Info))
6791 return ((Op.getOpcode() == ISD::ZERO_EXTEND) &&
6792 isSetCC(Op->getOperand(0), Info));
6795 // The folding we want to perform is:
6796 // (add x, [zext] (setcc cc ...) )
6798 // (csel x, (add x, 1), !cc ...)
6800 // The latter will get matched to a CSINC instruction.
6801 static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) {
6802 assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!");
6803 SDValue LHS = Op->getOperand(0);
6804 SDValue RHS = Op->getOperand(1);
6805 SetCCInfoAndKind InfoAndKind;
6807 // If neither operand is a SET_CC, give up.
6808 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind)) {
6809 std::swap(LHS, RHS);
6810 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind))
6814 // FIXME: This could be generatized to work for FP comparisons.
6815 EVT CmpVT = InfoAndKind.IsARM64
6816 ? InfoAndKind.Info.ARM64.Cmp->getOperand(0).getValueType()
6817 : InfoAndKind.Info.Generic.Opnd0->getValueType();
6818 if (CmpVT != MVT::i32 && CmpVT != MVT::i64)
6824 if (InfoAndKind.IsARM64) {
6825 CCVal = DAG.getConstant(
6826 ARM64CC::getInvertedCondCode(InfoAndKind.Info.ARM64.CC), MVT::i32);
6827 Cmp = *InfoAndKind.Info.ARM64.Cmp;
6829 Cmp = getARM64Cmp(*InfoAndKind.Info.Generic.Opnd0,
6830 *InfoAndKind.Info.Generic.Opnd1,
6831 ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, true),
6834 EVT VT = Op->getValueType(0);
6835 LHS = DAG.getNode(ISD::ADD, dl, VT, RHS, DAG.getConstant(1, VT));
6836 return DAG.getNode(ARM64ISD::CSEL, dl, VT, RHS, LHS, CCVal, Cmp);
6839 // The basic add/sub long vector instructions have variants with "2" on the end
6840 // which act on the high-half of their inputs. They are normally matched by
6843 // (add (zeroext (extract_high LHS)),
6844 // (zeroext (extract_high RHS)))
6845 // -> uaddl2 vD, vN, vM
6847 // However, if one of the extracts is something like a duplicate, this
6848 // instruction can still be used profitably. This function puts the DAG into a
6849 // more appropriate form for those patterns to trigger.
6850 static SDValue performAddSubLongCombine(SDNode *N,
6851 TargetLowering::DAGCombinerInfo &DCI,
6852 SelectionDAG &DAG) {
6853 if (DCI.isBeforeLegalizeOps())
6856 MVT VT = N->getSimpleValueType(0);
6857 if (!VT.is128BitVector()) {
6858 if (N->getOpcode() == ISD::ADD)
6859 return performSetccAddFolding(N, DAG);
6863 // Make sure both branches are extended in the same way.
6864 SDValue LHS = N->getOperand(0);
6865 SDValue RHS = N->getOperand(1);
6866 if ((LHS.getOpcode() != ISD::ZERO_EXTEND &&
6867 LHS.getOpcode() != ISD::SIGN_EXTEND) ||
6868 LHS.getOpcode() != RHS.getOpcode())
6871 unsigned ExtType = LHS.getOpcode();
6873 // It's not worth doing if at least one of the inputs isn't already an
6874 // extract, but we don't know which it'll be so we have to try both.
6875 if (isEssentiallyExtractSubvector(LHS.getOperand(0))) {
6876 RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG);
6880 RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS);
6881 } else if (isEssentiallyExtractSubvector(RHS.getOperand(0))) {
6882 LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG);
6886 LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS);
6889 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS);
6892 // Massage DAGs which we can use the high-half "long" operations on into
6893 // something isel will recognize better. E.g.
6895 // (arm64_neon_umull (extract_high vec) (dupv64 scalar)) -->
6896 // (arm64_neon_umull (extract_high (v2i64 vec)))
6897 // (extract_high (v2i64 (dup128 scalar)))))
6899 static SDValue tryCombineLongOpWithDup(unsigned IID, SDNode *N,
6900 TargetLowering::DAGCombinerInfo &DCI,
6901 SelectionDAG &DAG) {
6902 if (DCI.isBeforeLegalizeOps())
6905 SDValue LHS = N->getOperand(1);
6906 SDValue RHS = N->getOperand(2);
6907 assert(LHS.getValueType().is64BitVector() &&
6908 RHS.getValueType().is64BitVector() &&
6909 "unexpected shape for long operation");
6911 // Either node could be a DUP, but it's not worth doing both of them (you'd
6912 // just as well use the non-high version) so look for a corresponding extract
6913 // operation on the other "wing".
6914 if (isEssentiallyExtractSubvector(LHS)) {
6915 RHS = tryExtendDUPToExtractHigh(RHS, DAG);
6918 } else if (isEssentiallyExtractSubvector(RHS)) {
6919 LHS = tryExtendDUPToExtractHigh(LHS, DAG);
6924 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0),
6925 N->getOperand(0), LHS, RHS);
6928 static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
6929 MVT ElemTy = N->getSimpleValueType(0).getScalarType();
6930 unsigned ElemBits = ElemTy.getSizeInBits();
6932 int64_t ShiftAmount;
6933 if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) {
6934 APInt SplatValue, SplatUndef;
6935 unsigned SplatBitSize;
6937 if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
6938 HasAnyUndefs, ElemBits) ||
6939 SplatBitSize != ElemBits)
6942 ShiftAmount = SplatValue.getSExtValue();
6943 } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) {
6944 ShiftAmount = CVN->getSExtValue();
6952 llvm_unreachable("Unknown shift intrinsic");
6953 case Intrinsic::arm64_neon_sqshl:
6954 Opcode = ARM64ISD::SQSHL_I;
6955 IsRightShift = false;
6957 case Intrinsic::arm64_neon_uqshl:
6958 Opcode = ARM64ISD::UQSHL_I;
6959 IsRightShift = false;
6961 case Intrinsic::arm64_neon_srshl:
6962 Opcode = ARM64ISD::SRSHR_I;
6963 IsRightShift = true;
6965 case Intrinsic::arm64_neon_urshl:
6966 Opcode = ARM64ISD::URSHR_I;
6967 IsRightShift = true;
6969 case Intrinsic::arm64_neon_sqshlu:
6970 Opcode = ARM64ISD::SQSHLU_I;
6971 IsRightShift = false;
6975 if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits)
6976 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
6977 DAG.getConstant(-ShiftAmount, MVT::i32));
6978 else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount <= ElemBits)
6979 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
6980 DAG.getConstant(ShiftAmount, MVT::i32));
6985 // The CRC32[BH] instructions ignore the high bits of their data operand. Since
6986 // the intrinsics must be legal and take an i32, this means there's almost
6987 // certainly going to be a zext in the DAG which we can eliminate.
6988 static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) {
6989 SDValue AndN = N->getOperand(2);
6990 if (AndN.getOpcode() != ISD::AND)
6993 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1));
6994 if (!CMask || CMask->getZExtValue() != Mask)
6997 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32,
6998 N->getOperand(0), N->getOperand(1), AndN.getOperand(0));
7001 static SDValue performIntrinsicCombine(SDNode *N,
7002 TargetLowering::DAGCombinerInfo &DCI,
7003 const ARM64Subtarget *Subtarget) {
7004 SelectionDAG &DAG = DCI.DAG;
7005 unsigned IID = getIntrinsicID(N);
7009 case Intrinsic::arm64_neon_vcvtfxs2fp:
7010 case Intrinsic::arm64_neon_vcvtfxu2fp:
7011 return tryCombineFixedPointConvert(N, DCI, DAG);
7013 case Intrinsic::arm64_neon_fmax:
7014 return DAG.getNode(ARM64ISD::FMAX, SDLoc(N), N->getValueType(0),
7015 N->getOperand(1), N->getOperand(2));
7016 case Intrinsic::arm64_neon_fmin:
7017 return DAG.getNode(ARM64ISD::FMIN, SDLoc(N), N->getValueType(0),
7018 N->getOperand(1), N->getOperand(2));
7019 case Intrinsic::arm64_neon_smull:
7020 case Intrinsic::arm64_neon_umull:
7021 case Intrinsic::arm64_neon_pmull:
7022 case Intrinsic::arm64_neon_sqdmull:
7023 return tryCombineLongOpWithDup(IID, N, DCI, DAG);
7024 case Intrinsic::arm64_neon_sqshl:
7025 case Intrinsic::arm64_neon_uqshl:
7026 case Intrinsic::arm64_neon_sqshlu:
7027 case Intrinsic::arm64_neon_srshl:
7028 case Intrinsic::arm64_neon_urshl:
7029 return tryCombineShiftImm(IID, N, DAG);
7030 case Intrinsic::arm64_crc32b:
7031 case Intrinsic::arm64_crc32cb:
7032 return tryCombineCRC32(0xff, N, DAG);
7033 case Intrinsic::arm64_crc32h:
7034 case Intrinsic::arm64_crc32ch:
7035 return tryCombineCRC32(0xffff, N, DAG);
7040 static SDValue performExtendCombine(SDNode *N,
7041 TargetLowering::DAGCombinerInfo &DCI,
7042 SelectionDAG &DAG) {
7043 // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then
7044 // we can convert that DUP into another extract_high (of a bigger DUP), which
7045 // helps the backend to decide that an sabdl2 would be useful, saving a real
7046 // extract_high operation.
7047 if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND &&
7048 N->getOperand(0).getOpcode() == ISD::INTRINSIC_WO_CHAIN) {
7049 SDNode *ABDNode = N->getOperand(0).getNode();
7050 unsigned IID = getIntrinsicID(ABDNode);
7051 if (IID == Intrinsic::arm64_neon_sabd ||
7052 IID == Intrinsic::arm64_neon_uabd) {
7053 SDValue NewABD = tryCombineLongOpWithDup(IID, ABDNode, DCI, DAG);
7054 if (!NewABD.getNode())
7057 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0),
7062 // This is effectively a custom type legalization for ARM64.
7064 // Type legalization will split an extend of a small, legal, type to a larger
7065 // illegal type by first splitting the destination type, often creating
7066 // illegal source types, which then get legalized in isel-confusing ways,
7067 // leading to really terrible codegen. E.g.,
7068 // %result = v8i32 sext v8i8 %value
7070 // %losrc = extract_subreg %value, ...
7071 // %hisrc = extract_subreg %value, ...
7072 // %lo = v4i32 sext v4i8 %losrc
7073 // %hi = v4i32 sext v4i8 %hisrc
7074 // Things go rapidly downhill from there.
7076 // For ARM64, the [sz]ext vector instructions can only go up one element
7077 // size, so we can, e.g., extend from i8 to i16, but to go from i8 to i32
7078 // take two instructions.
7080 // This implies that the most efficient way to do the extend from v8i8
7081 // to two v4i32 values is to first extend the v8i8 to v8i16, then do
7082 // the normal splitting to happen for the v8i16->v8i32.
7084 // This is pre-legalization to catch some cases where the default
7085 // type legalization will create ill-tempered code.
7086 if (!DCI.isBeforeLegalizeOps())
7089 // We're only interested in cleaning things up for non-legal vector types
7090 // here. If both the source and destination are legal, things will just
7091 // work naturally without any fiddling.
7092 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7093 EVT ResVT = N->getValueType(0);
7094 if (!ResVT.isVector() || TLI.isTypeLegal(ResVT))
7096 // If the vector type isn't a simple VT, it's beyond the scope of what
7097 // we're worried about here. Let legalization do its thing and hope for
7099 if (!ResVT.isSimple())
7102 SDValue Src = N->getOperand(0);
7103 MVT SrcVT = Src->getValueType(0).getSimpleVT();
7104 // If the source VT is a 64-bit vector, we can play games and get the
7105 // better results we want.
7106 if (SrcVT.getSizeInBits() != 64)
7109 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
7110 unsigned ElementCount = SrcVT.getVectorNumElements();
7111 SrcVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize * 2), ElementCount);
7113 Src = DAG.getNode(N->getOpcode(), DL, SrcVT, Src);
7115 // Now split the rest of the operation into two halves, each with a 64
7119 unsigned NumElements = ResVT.getVectorNumElements();
7120 assert(!(NumElements & 1) && "Splitting vector, but not in half!");
7121 LoVT = HiVT = EVT::getVectorVT(*DAG.getContext(),
7122 ResVT.getVectorElementType(), NumElements / 2);
7124 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
7125 LoVT.getVectorNumElements());
7126 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7127 DAG.getIntPtrConstant(0));
7128 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7129 DAG.getIntPtrConstant(InNVT.getVectorNumElements()));
7130 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, Lo);
7131 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, Hi);
7133 // Now combine the parts back together so we still have a single result
7134 // like the combiner expects.
7135 return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
7138 /// Replace a splat of a scalar to a vector store by scalar stores of the scalar
7139 /// value. The load store optimizer pass will merge them to store pair stores.
7140 /// This has better performance than a splat of the scalar followed by a split
7141 /// vector store. Even if the stores are not merged it is four stores vs a dup,
7142 /// followed by an ext.b and two stores.
7143 static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) {
7144 SDValue StVal = St->getValue();
7145 EVT VT = StVal.getValueType();
7147 // Don't replace floating point stores, they possibly won't be transformed to
7148 // stp because of the store pair suppress pass.
7149 if (VT.isFloatingPoint())
7152 // Check for insert vector elements.
7153 if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT)
7156 // We can express a splat as store pair(s) for 2 or 4 elements.
7157 unsigned NumVecElts = VT.getVectorNumElements();
7158 if (NumVecElts != 4 && NumVecElts != 2)
7160 SDValue SplatVal = StVal.getOperand(1);
7161 unsigned RemainInsertElts = NumVecElts - 1;
7163 // Check that this is a splat.
7164 while (--RemainInsertElts) {
7165 SDValue NextInsertElt = StVal.getOperand(0);
7166 if (NextInsertElt.getOpcode() != ISD::INSERT_VECTOR_ELT)
7168 if (NextInsertElt.getOperand(1) != SplatVal)
7170 StVal = NextInsertElt;
7172 unsigned OrigAlignment = St->getAlignment();
7173 unsigned EltOffset = NumVecElts == 4 ? 4 : 8;
7174 unsigned Alignment = std::min(OrigAlignment, EltOffset);
7176 // Create scalar stores. This is at least as good as the code sequence for a
7177 // split unaligned store wich is a dup.s, ext.b, and two stores.
7178 // Most of the time the three stores should be replaced by store pair
7179 // instructions (stp).
7181 SDValue BasePtr = St->getBasePtr();
7183 DAG.getStore(St->getChain(), DL, SplatVal, BasePtr, St->getPointerInfo(),
7184 St->isVolatile(), St->isNonTemporal(), St->getAlignment());
7186 unsigned Offset = EltOffset;
7187 while (--NumVecElts) {
7188 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7189 DAG.getConstant(Offset, MVT::i64));
7190 NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr,
7191 St->getPointerInfo(), St->isVolatile(),
7192 St->isNonTemporal(), Alignment);
7193 Offset += EltOffset;
7198 static SDValue performSTORECombine(SDNode *N,
7199 TargetLowering::DAGCombinerInfo &DCI,
7201 const ARM64Subtarget *Subtarget) {
7202 if (!DCI.isBeforeLegalize())
7205 StoreSDNode *S = cast<StoreSDNode>(N);
7206 if (S->isVolatile())
7209 // Cyclone has bad performance on unaligned 16B stores when crossing line and
7210 // page boundries. We want to split such stores.
7211 if (!Subtarget->isCyclone())
7214 // Don't split at Oz.
7215 MachineFunction &MF = DAG.getMachineFunction();
7216 bool IsMinSize = MF.getFunction()->getAttributes().hasAttribute(
7217 AttributeSet::FunctionIndex, Attribute::MinSize);
7221 SDValue StVal = S->getValue();
7222 EVT VT = StVal.getValueType();
7224 // Don't split v2i64 vectors. Memcpy lowering produces those and splitting
7225 // those up regresses performance on micro-benchmarks and olden/bh.
7226 if (!VT.isVector() || VT.getVectorNumElements() < 2 || VT == MVT::v2i64)
7229 // Split unaligned 16B stores. They are terrible for performance.
7230 // Don't split stores with alignment of 1 or 2. Code that uses clang vector
7231 // extensions can use this to mark that it does not want splitting to happen
7232 // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of
7233 // eliminating alignment hazards is only 1 in 8 for alignment of 2.
7234 if (VT.getSizeInBits() != 128 || S->getAlignment() >= 16 ||
7235 S->getAlignment() <= 2)
7238 // If we get a splat of a scalar convert this vector store to a store of
7239 // scalars. They will be merged into store pairs thereby removing two
7241 SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S);
7242 if (ReplacedSplat != SDValue())
7243 return ReplacedSplat;
7246 unsigned NumElts = VT.getVectorNumElements() / 2;
7247 // Split VT into two.
7249 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), NumElts);
7250 SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7251 DAG.getIntPtrConstant(0));
7252 SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7253 DAG.getIntPtrConstant(NumElts));
7254 SDValue BasePtr = S->getBasePtr();
7256 DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
7257 S->isVolatile(), S->isNonTemporal(), S->getAlignment());
7258 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7259 DAG.getConstant(8, MVT::i64));
7260 return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr,
7261 S->getPointerInfo(), S->isVolatile(), S->isNonTemporal(),
7265 /// Target-specific DAG combine function for post-increment LD1 (lane) and
7266 /// post-increment LD1R.
7267 static SDValue performPostLD1Combine(SDNode *N,
7268 TargetLowering::DAGCombinerInfo &DCI,
7270 if (DCI.isBeforeLegalizeOps())
7273 SelectionDAG &DAG = DCI.DAG;
7274 EVT VT = N->getValueType(0);
7276 unsigned LoadIdx = IsLaneOp ? 1 : 0;
7277 SDNode *LD = N->getOperand(LoadIdx).getNode();
7278 // If it is not LOAD, can not do such combine.
7279 if (LD->getOpcode() != ISD::LOAD)
7282 LoadSDNode *LoadSDN = cast<LoadSDNode>(LD);
7283 EVT MemVT = LoadSDN->getMemoryVT();
7284 // Check if memory operand is the same type as the vector element.
7285 if (MemVT != VT.getVectorElementType())
7288 // Check if there are other uses. If so, do not combine as it will introduce
7290 for (SDNode::use_iterator UI = LD->use_begin(), UE = LD->use_end(); UI != UE;
7292 if (UI.getUse().getResNo() == 1) // Ignore uses of the chain result.
7298 SDValue Addr = LD->getOperand(1);
7299 // Search for a use of the address operand that is an increment.
7300 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(), UE =
7301 Addr.getNode()->use_end(); UI != UE; ++UI) {
7303 if (User->getOpcode() != ISD::ADD
7304 || UI.getUse().getResNo() != Addr.getResNo())
7307 // Check that the add is independent of the load. Otherwise, folding it
7308 // would create a cycle.
7309 if (User->isPredecessorOf(LD) || LD->isPredecessorOf(User))
7312 // If the increment is a constant, it must match the memory ref size.
7313 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
7314 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
7315 uint32_t IncVal = CInc->getZExtValue();
7316 unsigned NumBytes = VT.getScalarSizeInBits() / 8;
7317 if (IncVal != NumBytes)
7319 Inc = DAG.getRegister(ARM64::XZR, MVT::i64);
7322 SmallVector<SDValue, 8> Ops;
7323 Ops.push_back(LD->getOperand(0)); // Chain
7325 Ops.push_back(N->getOperand(0)); // The vector to be inserted
7326 Ops.push_back(N->getOperand(2)); // The lane to be inserted in the vector
7328 Ops.push_back(Addr);
7331 EVT Tys[3] = { VT, MVT::i64, MVT::Other };
7332 SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, 3));
7333 unsigned NewOp = IsLaneOp ? ARM64ISD::LD1LANEpost : ARM64ISD::LD1DUPpost;
7334 SDValue UpdN = DAG.getMemIntrinsicNode(NewOp, SDLoc(N), SDTys, Ops,
7336 LoadSDN->getMemOperand());
7339 std::vector<SDValue> NewResults;
7340 NewResults.push_back(SDValue(LD, 0)); // The result of load
7341 NewResults.push_back(SDValue(UpdN.getNode(), 2)); // Chain
7342 DCI.CombineTo(LD, NewResults);
7343 DCI.CombineTo(N, SDValue(UpdN.getNode(), 0)); // Dup/Inserted Result
7344 DCI.CombineTo(User, SDValue(UpdN.getNode(), 1)); // Write back register
7351 /// Target-specific DAG combine function for NEON load/store intrinsics
7352 /// to merge base address updates.
7353 static SDValue performNEONPostLDSTCombine(SDNode *N,
7354 TargetLowering::DAGCombinerInfo &DCI,
7355 SelectionDAG &DAG) {
7356 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7359 unsigned AddrOpIdx = N->getNumOperands() - 1;
7360 SDValue Addr = N->getOperand(AddrOpIdx);
7362 // Search for a use of the address operand that is an increment.
7363 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
7364 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
7366 if (User->getOpcode() != ISD::ADD ||
7367 UI.getUse().getResNo() != Addr.getResNo())
7370 // Check that the add is independent of the load/store. Otherwise, folding
7371 // it would create a cycle.
7372 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
7375 // Find the new opcode for the updating load/store.
7376 bool IsStore = false;
7377 bool IsLaneOp = false;
7378 bool IsDupOp = false;
7379 unsigned NewOpc = 0;
7380 unsigned NumVecs = 0;
7381 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
7383 default: llvm_unreachable("unexpected intrinsic for Neon base update");
7384 case Intrinsic::arm64_neon_ld2: NewOpc = ARM64ISD::LD2post;
7386 case Intrinsic::arm64_neon_ld3: NewOpc = ARM64ISD::LD3post;
7388 case Intrinsic::arm64_neon_ld4: NewOpc = ARM64ISD::LD4post;
7390 case Intrinsic::arm64_neon_st2: NewOpc = ARM64ISD::ST2post;
7391 NumVecs = 2; IsStore = true; break;
7392 case Intrinsic::arm64_neon_st3: NewOpc = ARM64ISD::ST3post;
7393 NumVecs = 3; IsStore = true; break;
7394 case Intrinsic::arm64_neon_st4: NewOpc = ARM64ISD::ST4post;
7395 NumVecs = 4; IsStore = true; break;
7396 case Intrinsic::arm64_neon_ld1x2: NewOpc = ARM64ISD::LD1x2post;
7398 case Intrinsic::arm64_neon_ld1x3: NewOpc = ARM64ISD::LD1x3post;
7400 case Intrinsic::arm64_neon_ld1x4: NewOpc = ARM64ISD::LD1x4post;
7402 case Intrinsic::arm64_neon_st1x2: NewOpc = ARM64ISD::ST1x2post;
7403 NumVecs = 2; IsStore = true; break;
7404 case Intrinsic::arm64_neon_st1x3: NewOpc = ARM64ISD::ST1x3post;
7405 NumVecs = 3; IsStore = true; break;
7406 case Intrinsic::arm64_neon_st1x4: NewOpc = ARM64ISD::ST1x4post;
7407 NumVecs = 4; IsStore = true; break;
7408 case Intrinsic::arm64_neon_ld2r: NewOpc = ARM64ISD::LD2DUPpost;
7409 NumVecs = 2; IsDupOp = true; break;
7410 case Intrinsic::arm64_neon_ld3r: NewOpc = ARM64ISD::LD3DUPpost;
7411 NumVecs = 3; IsDupOp = true; break;
7412 case Intrinsic::arm64_neon_ld4r: NewOpc = ARM64ISD::LD4DUPpost;
7413 NumVecs = 4; IsDupOp = true; break;
7414 case Intrinsic::arm64_neon_ld2lane: NewOpc = ARM64ISD::LD2LANEpost;
7415 NumVecs = 2; IsLaneOp = true; break;
7416 case Intrinsic::arm64_neon_ld3lane: NewOpc = ARM64ISD::LD3LANEpost;
7417 NumVecs = 3; IsLaneOp = true; break;
7418 case Intrinsic::arm64_neon_ld4lane: NewOpc = ARM64ISD::LD4LANEpost;
7419 NumVecs = 4; IsLaneOp = true; break;
7420 case Intrinsic::arm64_neon_st2lane: NewOpc = ARM64ISD::ST2LANEpost;
7421 NumVecs = 2; IsStore = true; IsLaneOp = true; break;
7422 case Intrinsic::arm64_neon_st3lane: NewOpc = ARM64ISD::ST3LANEpost;
7423 NumVecs = 3; IsStore = true; IsLaneOp = true; break;
7424 case Intrinsic::arm64_neon_st4lane: NewOpc = ARM64ISD::ST4LANEpost;
7425 NumVecs = 4; IsStore = true; IsLaneOp = true; break;
7430 VecTy = N->getOperand(2).getValueType();
7432 VecTy = N->getValueType(0);
7434 // If the increment is a constant, it must match the memory ref size.
7435 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
7436 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
7437 uint32_t IncVal = CInc->getZExtValue();
7438 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
7439 if (IsLaneOp || IsDupOp)
7440 NumBytes /= VecTy.getVectorNumElements();
7441 if (IncVal != NumBytes)
7443 Inc = DAG.getRegister(ARM64::XZR, MVT::i64);
7445 SmallVector<SDValue, 8> Ops;
7446 Ops.push_back(N->getOperand(0)); // Incoming chain
7447 // Load lane and store have vector list as input.
7448 if (IsLaneOp || IsStore)
7449 for (unsigned i = 2; i < AddrOpIdx; ++i)
7450 Ops.push_back(N->getOperand(i));
7451 Ops.push_back(Addr); // Base register
7456 unsigned NumResultVecs = (IsStore ? 0 : NumVecs);
7458 for (n = 0; n < NumResultVecs; ++n)
7460 Tys[n++] = MVT::i64; // Type of write back register
7461 Tys[n] = MVT::Other; // Type of the chain
7462 SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, NumResultVecs + 2));
7464 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
7465 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys, Ops,
7466 MemInt->getMemoryVT(),
7467 MemInt->getMemOperand());
7470 std::vector<SDValue> NewResults;
7471 for (unsigned i = 0; i < NumResultVecs; ++i) {
7472 NewResults.push_back(SDValue(UpdN.getNode(), i));
7474 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs + 1));
7475 DCI.CombineTo(N, NewResults);
7476 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
7483 // Optimize compare with zero and branch.
7484 static SDValue performBRCONDCombine(SDNode *N,
7485 TargetLowering::DAGCombinerInfo &DCI,
7486 SelectionDAG &DAG) {
7487 SDValue Chain = N->getOperand(0);
7488 SDValue Dest = N->getOperand(1);
7489 SDValue CCVal = N->getOperand(2);
7490 SDValue Cmp = N->getOperand(3);
7492 assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!");
7493 unsigned CC = cast<ConstantSDNode>(CCVal)->getZExtValue();
7494 if (CC != ARM64CC::EQ && CC != ARM64CC::NE)
7497 unsigned CmpOpc = Cmp.getOpcode();
7498 if (CmpOpc != ARM64ISD::ADDS && CmpOpc != ARM64ISD::SUBS)
7501 // Only attempt folding if there is only one use of the flag and no use of the
7503 if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1))
7506 SDValue LHS = Cmp.getOperand(0);
7507 SDValue RHS = Cmp.getOperand(1);
7509 assert(LHS.getValueType() == RHS.getValueType() &&
7510 "Expected the value type to be the same for both operands!");
7511 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
7514 if (isa<ConstantSDNode>(LHS) && cast<ConstantSDNode>(LHS)->isNullValue())
7515 std::swap(LHS, RHS);
7517 if (!isa<ConstantSDNode>(RHS) || !cast<ConstantSDNode>(RHS)->isNullValue())
7520 if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA ||
7521 LHS.getOpcode() == ISD::SRL)
7524 // Fold the compare into the branch instruction.
7526 if (CC == ARM64CC::EQ)
7527 BR = DAG.getNode(ARM64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
7529 BR = DAG.getNode(ARM64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
7531 // Do not add new nodes to DAG combiner worklist.
7532 DCI.CombineTo(N, BR, false);
7537 // vselect (v1i1 setcc) ->
7538 // vselect (v1iXX setcc) (XX is the size of the compared operand type)
7539 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
7540 // condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
7542 static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) {
7543 SDValue N0 = N->getOperand(0);
7544 EVT CCVT = N0.getValueType();
7546 if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
7547 CCVT.getVectorElementType() != MVT::i1)
7550 EVT ResVT = N->getValueType(0);
7551 EVT CmpVT = N0.getOperand(0).getValueType();
7552 // Only combine when the result type is of the same size as the compared
7554 if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
7557 SDValue IfTrue = N->getOperand(1);
7558 SDValue IfFalse = N->getOperand(2);
7560 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
7561 N0.getOperand(0), N0.getOperand(1),
7562 cast<CondCodeSDNode>(N0.getOperand(2))->get());
7563 return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
7567 /// A vector select: "(select vL, vR, (setcc LHS, RHS))" is best performed with
7568 /// the compare-mask instructions rather than going via NZCV, even if LHS and
7569 /// RHS are really scalar. This replaces any scalar setcc in the above pattern
7570 /// with a vector one followed by a DUP shuffle on the result.
7571 static SDValue performSelectCombine(SDNode *N, SelectionDAG &DAG) {
7572 SDValue N0 = N->getOperand(0);
7573 EVT ResVT = N->getValueType(0);
7575 if (!N->getOperand(1).getValueType().isVector())
7578 if (N0.getOpcode() != ISD::SETCC || N0.getValueType() != MVT::i1)
7583 EVT SrcVT = N0.getOperand(0).getValueType();
7584 SrcVT = EVT::getVectorVT(*DAG.getContext(), SrcVT,
7585 ResVT.getSizeInBits() / SrcVT.getSizeInBits());
7586 EVT CCVT = SrcVT.changeVectorElementTypeToInteger();
7588 // First perform a vector comparison, where lane 0 is the one we're interested
7591 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(0));
7593 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(1));
7594 SDValue SetCC = DAG.getNode(ISD::SETCC, DL, CCVT, LHS, RHS, N0.getOperand(2));
7596 // Now duplicate the comparison mask we want across all other lanes.
7597 SmallVector<int, 8> DUPMask(CCVT.getVectorNumElements(), 0);
7598 SDValue Mask = DAG.getVectorShuffle(CCVT, DL, SetCC, SetCC, DUPMask.data());
7599 Mask = DAG.getNode(ISD::BITCAST, DL, ResVT.changeVectorElementTypeToInteger(),
7602 return DAG.getSelect(DL, ResVT, Mask, N->getOperand(1), N->getOperand(2));
7605 SDValue ARM64TargetLowering::PerformDAGCombine(SDNode *N,
7606 DAGCombinerInfo &DCI) const {
7607 SelectionDAG &DAG = DCI.DAG;
7608 switch (N->getOpcode()) {
7613 return performAddSubLongCombine(N, DCI, DAG);
7615 return performXorCombine(N, DAG, DCI, Subtarget);
7617 return performMulCombine(N, DAG, DCI, Subtarget);
7618 case ISD::SINT_TO_FP:
7619 case ISD::UINT_TO_FP:
7620 return performIntToFpCombine(N, DAG);
7622 return performORCombine(N, DCI, Subtarget);
7623 case ISD::INTRINSIC_WO_CHAIN:
7624 return performIntrinsicCombine(N, DCI, Subtarget);
7625 case ISD::ANY_EXTEND:
7626 case ISD::ZERO_EXTEND:
7627 case ISD::SIGN_EXTEND:
7628 return performExtendCombine(N, DCI, DAG);
7630 return performBitcastCombine(N, DCI, DAG);
7631 case ISD::CONCAT_VECTORS:
7632 return performConcatVectorsCombine(N, DCI, DAG);
7634 return performSelectCombine(N, DAG);
7636 return performVSelectCombine(N, DCI.DAG);
7638 return performSTORECombine(N, DCI, DAG, Subtarget);
7639 case ARM64ISD::BRCOND:
7640 return performBRCONDCombine(N, DCI, DAG);
7642 return performPostLD1Combine(N, DCI, false);
7643 case ISD::INSERT_VECTOR_ELT:
7644 return performPostLD1Combine(N, DCI, true);
7645 case ISD::INTRINSIC_VOID:
7646 case ISD::INTRINSIC_W_CHAIN:
7647 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
7648 case Intrinsic::arm64_neon_ld2:
7649 case Intrinsic::arm64_neon_ld3:
7650 case Intrinsic::arm64_neon_ld4:
7651 case Intrinsic::arm64_neon_ld1x2:
7652 case Intrinsic::arm64_neon_ld1x3:
7653 case Intrinsic::arm64_neon_ld1x4:
7654 case Intrinsic::arm64_neon_ld2lane:
7655 case Intrinsic::arm64_neon_ld3lane:
7656 case Intrinsic::arm64_neon_ld4lane:
7657 case Intrinsic::arm64_neon_ld2r:
7658 case Intrinsic::arm64_neon_ld3r:
7659 case Intrinsic::arm64_neon_ld4r:
7660 case Intrinsic::arm64_neon_st2:
7661 case Intrinsic::arm64_neon_st3:
7662 case Intrinsic::arm64_neon_st4:
7663 case Intrinsic::arm64_neon_st1x2:
7664 case Intrinsic::arm64_neon_st1x3:
7665 case Intrinsic::arm64_neon_st1x4:
7666 case Intrinsic::arm64_neon_st2lane:
7667 case Intrinsic::arm64_neon_st3lane:
7668 case Intrinsic::arm64_neon_st4lane:
7669 return performNEONPostLDSTCombine(N, DCI, DAG);
7677 // Check if the return value is used as only a return value, as otherwise
7678 // we can't perform a tail-call. In particular, we need to check for
7679 // target ISD nodes that are returns and any other "odd" constructs
7680 // that the generic analysis code won't necessarily catch.
7681 bool ARM64TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
7682 if (N->getNumValues() != 1)
7684 if (!N->hasNUsesOfValue(1, 0))
7687 SDValue TCChain = Chain;
7688 SDNode *Copy = *N->use_begin();
7689 if (Copy->getOpcode() == ISD::CopyToReg) {
7690 // If the copy has a glue operand, we conservatively assume it isn't safe to
7691 // perform a tail call.
7692 if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() ==
7695 TCChain = Copy->getOperand(0);
7696 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
7699 bool HasRet = false;
7700 for (SDNode *Node : Copy->uses()) {
7701 if (Node->getOpcode() != ARM64ISD::RET_FLAG)
7713 // Return whether the an instruction can potentially be optimized to a tail
7714 // call. This will cause the optimizers to attempt to move, or duplicate,
7715 // return instructions to help enable tail call optimizations for this
7717 bool ARM64TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
7718 if (!CI->isTailCall())
7724 bool ARM64TargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base,
7726 ISD::MemIndexedMode &AM,
7728 SelectionDAG &DAG) const {
7729 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
7732 Base = Op->getOperand(0);
7733 // All of the indexed addressing mode instructions take a signed
7734 // 9 bit immediate offset.
7735 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
7736 int64_t RHSC = (int64_t)RHS->getZExtValue();
7737 if (RHSC >= 256 || RHSC <= -256)
7739 IsInc = (Op->getOpcode() == ISD::ADD);
7740 Offset = Op->getOperand(1);
7746 bool ARM64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
7748 ISD::MemIndexedMode &AM,
7749 SelectionDAG &DAG) const {
7752 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7753 VT = LD->getMemoryVT();
7754 Ptr = LD->getBasePtr();
7755 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7756 VT = ST->getMemoryVT();
7757 Ptr = ST->getBasePtr();
7762 if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, IsInc, DAG))
7764 AM = IsInc ? ISD::PRE_INC : ISD::PRE_DEC;
7768 bool ARM64TargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
7771 ISD::MemIndexedMode &AM,
7772 SelectionDAG &DAG) const {
7775 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7776 VT = LD->getMemoryVT();
7777 Ptr = LD->getBasePtr();
7778 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7779 VT = ST->getMemoryVT();
7780 Ptr = ST->getBasePtr();
7785 if (!getIndexedAddressParts(Op, Base, Offset, AM, IsInc, DAG))
7787 // Post-indexing updates the base, so it's not a valid transform
7788 // if that's not the same as the load's pointer.
7791 AM = IsInc ? ISD::POST_INC : ISD::POST_DEC;
7795 void ARM64TargetLowering::ReplaceNodeResults(SDNode *N,
7796 SmallVectorImpl<SDValue> &Results,
7797 SelectionDAG &DAG) const {
7798 switch (N->getOpcode()) {
7800 llvm_unreachable("Don't know how to custom expand this");
7801 case ISD::FP_TO_UINT:
7802 case ISD::FP_TO_SINT:
7803 assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion");
7804 // Let normal code take care of it by not adding anything to Results.
7809 bool ARM64TargetLowering::shouldExpandAtomicInIR(Instruction *Inst) const {
7810 // Loads and stores less than 128-bits are already atomic; ones above that
7811 // are doomed anyway, so defer to the default libcall and blame the OS when
7813 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
7814 return SI->getValueOperand()->getType()->getPrimitiveSizeInBits() == 128;
7815 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
7816 return LI->getType()->getPrimitiveSizeInBits() == 128;
7818 // For the real atomic operations, we have ldxr/stxr up to 128 bits.
7819 return Inst->getType()->getPrimitiveSizeInBits() <= 128;
7822 Value *ARM64TargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
7823 AtomicOrdering Ord) const {
7824 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
7825 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
7827 Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent;
7829 // Since i128 isn't legal and intrinsics don't get type-lowered, the ldrexd
7830 // intrinsic must return {i64, i64} and we have to recombine them into a
7831 // single i128 here.
7832 if (ValTy->getPrimitiveSizeInBits() == 128) {
7834 IsAcquire ? Intrinsic::arm64_ldaxp : Intrinsic::arm64_ldxp;
7835 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int);
7837 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
7838 Value *LoHi = Builder.CreateCall(Ldxr, Addr, "lohi");
7840 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
7841 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
7842 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
7843 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
7844 return Builder.CreateOr(
7845 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 64)), "val64");
7848 Type *Tys[] = { Addr->getType() };
7850 IsAcquire ? Intrinsic::arm64_ldaxr : Intrinsic::arm64_ldxr;
7851 Function *Ldxr = llvm::Intrinsic::getDeclaration(M, Int, Tys);
7853 return Builder.CreateTruncOrBitCast(
7854 Builder.CreateCall(Ldxr, Addr),
7855 cast<PointerType>(Addr->getType())->getElementType());
7858 Value *ARM64TargetLowering::emitStoreConditional(IRBuilder<> &Builder,
7859 Value *Val, Value *Addr,
7860 AtomicOrdering Ord) const {
7861 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
7863 Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent;
7865 // Since the intrinsics must have legal type, the i128 intrinsics take two
7866 // parameters: "i64, i64". We must marshal Val into the appropriate form
7868 if (Val->getType()->getPrimitiveSizeInBits() == 128) {
7870 IsRelease ? Intrinsic::arm64_stlxp : Intrinsic::arm64_stxp;
7871 Function *Stxr = Intrinsic::getDeclaration(M, Int);
7872 Type *Int64Ty = Type::getInt64Ty(M->getContext());
7874 Value *Lo = Builder.CreateTrunc(Val, Int64Ty, "lo");
7875 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 64), Int64Ty, "hi");
7876 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
7877 return Builder.CreateCall3(Stxr, Lo, Hi, Addr);
7881 IsRelease ? Intrinsic::arm64_stlxr : Intrinsic::arm64_stxr;
7882 Type *Tys[] = { Addr->getType() };
7883 Function *Stxr = Intrinsic::getDeclaration(M, Int, Tys);
7885 return Builder.CreateCall2(
7886 Stxr, Builder.CreateZExtOrBitCast(
7887 Val, Stxr->getFunctionType()->getParamType(0)),