1 //===-- ARMInstrNEON.td - NEON support for ARM -------------*- tablegen -*-===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file describes the ARM NEON instruction set.
11 //===----------------------------------------------------------------------===//
14 //===----------------------------------------------------------------------===//
15 // NEON-specific Operands.
16 //===----------------------------------------------------------------------===//
17 def nModImm : Operand<i32> {
18 let PrintMethod = "printVMOVModImmOperand";
21 def nImmSplatI8AsmOperand : AsmOperandClass { let Name = "NEONi8splat"; }
22 def nImmSplatI8 : Operand<i32> {
23 let PrintMethod = "printVMOVModImmOperand";
24 let ParserMatchClass = nImmSplatI8AsmOperand;
26 def nImmSplatI16AsmOperand : AsmOperandClass { let Name = "NEONi16splat"; }
27 def nImmSplatI16 : Operand<i32> {
28 let PrintMethod = "printVMOVModImmOperand";
29 let ParserMatchClass = nImmSplatI16AsmOperand;
31 def nImmSplatI32AsmOperand : AsmOperandClass { let Name = "NEONi32splat"; }
32 def nImmSplatI32 : Operand<i32> {
33 let PrintMethod = "printVMOVModImmOperand";
34 let ParserMatchClass = nImmSplatI32AsmOperand;
36 def nImmSplatNotI16AsmOperand : AsmOperandClass { let Name = "NEONi16splatNot"; }
37 def nImmSplatNotI16 : Operand<i32> {
38 let ParserMatchClass = nImmSplatNotI16AsmOperand;
40 def nImmSplatNotI32AsmOperand : AsmOperandClass { let Name = "NEONi32splatNot"; }
41 def nImmSplatNotI32 : Operand<i32> {
42 let ParserMatchClass = nImmSplatNotI32AsmOperand;
44 def nImmVMOVI32AsmOperand : AsmOperandClass { let Name = "NEONi32vmov"; }
45 def nImmVMOVI32 : Operand<i32> {
46 let PrintMethod = "printVMOVModImmOperand";
47 let ParserMatchClass = nImmVMOVI32AsmOperand;
50 class nImmVMOVIAsmOperandReplicate<ValueType From, ValueType To>
52 let Name = "NEONi" # To.Size # "vmovi" # From.Size # "Replicate";
53 let PredicateMethod = "isNEONmovReplicate<" # From.Size # ", " # To.Size # ">";
54 let RenderMethod = "addNEONvmovi" # From.Size # "ReplicateOperands";
57 class nImmVINVIAsmOperandReplicate<ValueType From, ValueType To>
59 let Name = "NEONi" # To.Size # "invi" # From.Size # "Replicate";
60 let PredicateMethod = "isNEONinvReplicate<" # From.Size # ", " # To.Size # ">";
61 let RenderMethod = "addNEONinvi" # From.Size # "ReplicateOperands";
64 class nImmVMOVIReplicate<ValueType From, ValueType To> : Operand<i32> {
65 let PrintMethod = "printVMOVModImmOperand";
66 let ParserMatchClass = nImmVMOVIAsmOperandReplicate<From, To>;
69 class nImmVINVIReplicate<ValueType From, ValueType To> : Operand<i32> {
70 let PrintMethod = "printVMOVModImmOperand";
71 let ParserMatchClass = nImmVINVIAsmOperandReplicate<From, To>;
74 def nImmVMOVI32NegAsmOperand : AsmOperandClass { let Name = "NEONi32vmovNeg"; }
75 def nImmVMOVI32Neg : Operand<i32> {
76 let PrintMethod = "printVMOVModImmOperand";
77 let ParserMatchClass = nImmVMOVI32NegAsmOperand;
79 def nImmVMOVF32 : Operand<i32> {
80 let PrintMethod = "printFPImmOperand";
81 let ParserMatchClass = FPImmOperand;
83 def nImmSplatI64AsmOperand : AsmOperandClass { let Name = "NEONi64splat"; }
84 def nImmSplatI64 : Operand<i32> {
85 let PrintMethod = "printVMOVModImmOperand";
86 let ParserMatchClass = nImmSplatI64AsmOperand;
89 def VectorIndex8Operand : AsmOperandClass { let Name = "VectorIndex8"; }
90 def VectorIndex16Operand : AsmOperandClass { let Name = "VectorIndex16"; }
91 def VectorIndex32Operand : AsmOperandClass { let Name = "VectorIndex32"; }
92 def VectorIndex64Operand : AsmOperandClass { let Name = "VectorIndex64"; }
93 def VectorIndex8 : Operand<i32>, ImmLeaf<i32, [{
94 return ((uint64_t)Imm) < 8;
96 let ParserMatchClass = VectorIndex8Operand;
97 let PrintMethod = "printVectorIndex";
98 let MIOperandInfo = (ops i32imm);
100 def VectorIndex16 : Operand<i32>, ImmLeaf<i32, [{
101 return ((uint64_t)Imm) < 4;
103 let ParserMatchClass = VectorIndex16Operand;
104 let PrintMethod = "printVectorIndex";
105 let MIOperandInfo = (ops i32imm);
107 def VectorIndex32 : Operand<i32>, ImmLeaf<i32, [{
108 return ((uint64_t)Imm) < 2;
110 let ParserMatchClass = VectorIndex32Operand;
111 let PrintMethod = "printVectorIndex";
112 let MIOperandInfo = (ops i32imm);
114 def VectorIndex64 : Operand<i32>, ImmLeaf<i32, [{
115 return ((uint64_t)Imm) < 1;
117 let ParserMatchClass = VectorIndex64Operand;
118 let PrintMethod = "printVectorIndex";
119 let MIOperandInfo = (ops i32imm);
122 // Register list of one D register.
123 def VecListOneDAsmOperand : AsmOperandClass {
124 let Name = "VecListOneD";
125 let ParserMethod = "parseVectorList";
126 let RenderMethod = "addVecListOperands";
128 def VecListOneD : RegisterOperand<DPR, "printVectorListOne"> {
129 let ParserMatchClass = VecListOneDAsmOperand;
131 // Register list of two sequential D registers.
132 def VecListDPairAsmOperand : AsmOperandClass {
133 let Name = "VecListDPair";
134 let ParserMethod = "parseVectorList";
135 let RenderMethod = "addVecListOperands";
137 def VecListDPair : RegisterOperand<DPair, "printVectorListTwo"> {
138 let ParserMatchClass = VecListDPairAsmOperand;
140 // Register list of three sequential D registers.
141 def VecListThreeDAsmOperand : AsmOperandClass {
142 let Name = "VecListThreeD";
143 let ParserMethod = "parseVectorList";
144 let RenderMethod = "addVecListOperands";
146 def VecListThreeD : RegisterOperand<DPR, "printVectorListThree"> {
147 let ParserMatchClass = VecListThreeDAsmOperand;
149 // Register list of four sequential D registers.
150 def VecListFourDAsmOperand : AsmOperandClass {
151 let Name = "VecListFourD";
152 let ParserMethod = "parseVectorList";
153 let RenderMethod = "addVecListOperands";
155 def VecListFourD : RegisterOperand<DPR, "printVectorListFour"> {
156 let ParserMatchClass = VecListFourDAsmOperand;
158 // Register list of two D registers spaced by 2 (two sequential Q registers).
159 def VecListDPairSpacedAsmOperand : AsmOperandClass {
160 let Name = "VecListDPairSpaced";
161 let ParserMethod = "parseVectorList";
162 let RenderMethod = "addVecListOperands";
164 def VecListDPairSpaced : RegisterOperand<DPair, "printVectorListTwoSpaced"> {
165 let ParserMatchClass = VecListDPairSpacedAsmOperand;
167 // Register list of three D registers spaced by 2 (three Q registers).
168 def VecListThreeQAsmOperand : AsmOperandClass {
169 let Name = "VecListThreeQ";
170 let ParserMethod = "parseVectorList";
171 let RenderMethod = "addVecListOperands";
173 def VecListThreeQ : RegisterOperand<DPR, "printVectorListThreeSpaced"> {
174 let ParserMatchClass = VecListThreeQAsmOperand;
176 // Register list of three D registers spaced by 2 (three Q registers).
177 def VecListFourQAsmOperand : AsmOperandClass {
178 let Name = "VecListFourQ";
179 let ParserMethod = "parseVectorList";
180 let RenderMethod = "addVecListOperands";
182 def VecListFourQ : RegisterOperand<DPR, "printVectorListFourSpaced"> {
183 let ParserMatchClass = VecListFourQAsmOperand;
186 // Register list of one D register, with "all lanes" subscripting.
187 def VecListOneDAllLanesAsmOperand : AsmOperandClass {
188 let Name = "VecListOneDAllLanes";
189 let ParserMethod = "parseVectorList";
190 let RenderMethod = "addVecListOperands";
192 def VecListOneDAllLanes : RegisterOperand<DPR, "printVectorListOneAllLanes"> {
193 let ParserMatchClass = VecListOneDAllLanesAsmOperand;
195 // Register list of two D registers, with "all lanes" subscripting.
196 def VecListDPairAllLanesAsmOperand : AsmOperandClass {
197 let Name = "VecListDPairAllLanes";
198 let ParserMethod = "parseVectorList";
199 let RenderMethod = "addVecListOperands";
201 def VecListDPairAllLanes : RegisterOperand<DPair,
202 "printVectorListTwoAllLanes"> {
203 let ParserMatchClass = VecListDPairAllLanesAsmOperand;
205 // Register list of two D registers spaced by 2 (two sequential Q registers).
206 def VecListDPairSpacedAllLanesAsmOperand : AsmOperandClass {
207 let Name = "VecListDPairSpacedAllLanes";
208 let ParserMethod = "parseVectorList";
209 let RenderMethod = "addVecListOperands";
211 def VecListDPairSpacedAllLanes : RegisterOperand<DPairSpc,
212 "printVectorListTwoSpacedAllLanes"> {
213 let ParserMatchClass = VecListDPairSpacedAllLanesAsmOperand;
215 // Register list of three D registers, with "all lanes" subscripting.
216 def VecListThreeDAllLanesAsmOperand : AsmOperandClass {
217 let Name = "VecListThreeDAllLanes";
218 let ParserMethod = "parseVectorList";
219 let RenderMethod = "addVecListOperands";
221 def VecListThreeDAllLanes : RegisterOperand<DPR,
222 "printVectorListThreeAllLanes"> {
223 let ParserMatchClass = VecListThreeDAllLanesAsmOperand;
225 // Register list of three D registers spaced by 2 (three sequential Q regs).
226 def VecListThreeQAllLanesAsmOperand : AsmOperandClass {
227 let Name = "VecListThreeQAllLanes";
228 let ParserMethod = "parseVectorList";
229 let RenderMethod = "addVecListOperands";
231 def VecListThreeQAllLanes : RegisterOperand<DPR,
232 "printVectorListThreeSpacedAllLanes"> {
233 let ParserMatchClass = VecListThreeQAllLanesAsmOperand;
235 // Register list of four D registers, with "all lanes" subscripting.
236 def VecListFourDAllLanesAsmOperand : AsmOperandClass {
237 let Name = "VecListFourDAllLanes";
238 let ParserMethod = "parseVectorList";
239 let RenderMethod = "addVecListOperands";
241 def VecListFourDAllLanes : RegisterOperand<DPR, "printVectorListFourAllLanes"> {
242 let ParserMatchClass = VecListFourDAllLanesAsmOperand;
244 // Register list of four D registers spaced by 2 (four sequential Q regs).
245 def VecListFourQAllLanesAsmOperand : AsmOperandClass {
246 let Name = "VecListFourQAllLanes";
247 let ParserMethod = "parseVectorList";
248 let RenderMethod = "addVecListOperands";
250 def VecListFourQAllLanes : RegisterOperand<DPR,
251 "printVectorListFourSpacedAllLanes"> {
252 let ParserMatchClass = VecListFourQAllLanesAsmOperand;
256 // Register list of one D register, with byte lane subscripting.
257 def VecListOneDByteIndexAsmOperand : AsmOperandClass {
258 let Name = "VecListOneDByteIndexed";
259 let ParserMethod = "parseVectorList";
260 let RenderMethod = "addVecListIndexedOperands";
262 def VecListOneDByteIndexed : Operand<i32> {
263 let ParserMatchClass = VecListOneDByteIndexAsmOperand;
264 let MIOperandInfo = (ops DPR:$Vd, i32imm:$idx);
266 // ...with half-word lane subscripting.
267 def VecListOneDHWordIndexAsmOperand : AsmOperandClass {
268 let Name = "VecListOneDHWordIndexed";
269 let ParserMethod = "parseVectorList";
270 let RenderMethod = "addVecListIndexedOperands";
272 def VecListOneDHWordIndexed : Operand<i32> {
273 let ParserMatchClass = VecListOneDHWordIndexAsmOperand;
274 let MIOperandInfo = (ops DPR:$Vd, i32imm:$idx);
276 // ...with word lane subscripting.
277 def VecListOneDWordIndexAsmOperand : AsmOperandClass {
278 let Name = "VecListOneDWordIndexed";
279 let ParserMethod = "parseVectorList";
280 let RenderMethod = "addVecListIndexedOperands";
282 def VecListOneDWordIndexed : Operand<i32> {
283 let ParserMatchClass = VecListOneDWordIndexAsmOperand;
284 let MIOperandInfo = (ops DPR:$Vd, i32imm:$idx);
287 // Register list of two D registers with byte lane subscripting.
288 def VecListTwoDByteIndexAsmOperand : AsmOperandClass {
289 let Name = "VecListTwoDByteIndexed";
290 let ParserMethod = "parseVectorList";
291 let RenderMethod = "addVecListIndexedOperands";
293 def VecListTwoDByteIndexed : Operand<i32> {
294 let ParserMatchClass = VecListTwoDByteIndexAsmOperand;
295 let MIOperandInfo = (ops DPR:$Vd, i32imm:$idx);
297 // ...with half-word lane subscripting.
298 def VecListTwoDHWordIndexAsmOperand : AsmOperandClass {
299 let Name = "VecListTwoDHWordIndexed";
300 let ParserMethod = "parseVectorList";
301 let RenderMethod = "addVecListIndexedOperands";
303 def VecListTwoDHWordIndexed : Operand<i32> {
304 let ParserMatchClass = VecListTwoDHWordIndexAsmOperand;
305 let MIOperandInfo = (ops DPR:$Vd, i32imm:$idx);
307 // ...with word lane subscripting.
308 def VecListTwoDWordIndexAsmOperand : AsmOperandClass {
309 let Name = "VecListTwoDWordIndexed";
310 let ParserMethod = "parseVectorList";
311 let RenderMethod = "addVecListIndexedOperands";
313 def VecListTwoDWordIndexed : Operand<i32> {
314 let ParserMatchClass = VecListTwoDWordIndexAsmOperand;
315 let MIOperandInfo = (ops DPR:$Vd, i32imm:$idx);
317 // Register list of two Q registers with half-word lane subscripting.
318 def VecListTwoQHWordIndexAsmOperand : AsmOperandClass {
319 let Name = "VecListTwoQHWordIndexed";
320 let ParserMethod = "parseVectorList";
321 let RenderMethod = "addVecListIndexedOperands";
323 def VecListTwoQHWordIndexed : Operand<i32> {
324 let ParserMatchClass = VecListTwoQHWordIndexAsmOperand;
325 let MIOperandInfo = (ops DPR:$Vd, i32imm:$idx);
327 // ...with word lane subscripting.
328 def VecListTwoQWordIndexAsmOperand : AsmOperandClass {
329 let Name = "VecListTwoQWordIndexed";
330 let ParserMethod = "parseVectorList";
331 let RenderMethod = "addVecListIndexedOperands";
333 def VecListTwoQWordIndexed : Operand<i32> {
334 let ParserMatchClass = VecListTwoQWordIndexAsmOperand;
335 let MIOperandInfo = (ops DPR:$Vd, i32imm:$idx);
339 // Register list of three D registers with byte lane subscripting.
340 def VecListThreeDByteIndexAsmOperand : AsmOperandClass {
341 let Name = "VecListThreeDByteIndexed";
342 let ParserMethod = "parseVectorList";
343 let RenderMethod = "addVecListIndexedOperands";
345 def VecListThreeDByteIndexed : Operand<i32> {
346 let ParserMatchClass = VecListThreeDByteIndexAsmOperand;
347 let MIOperandInfo = (ops DPR:$Vd, i32imm:$idx);
349 // ...with half-word lane subscripting.
350 def VecListThreeDHWordIndexAsmOperand : AsmOperandClass {
351 let Name = "VecListThreeDHWordIndexed";
352 let ParserMethod = "parseVectorList";
353 let RenderMethod = "addVecListIndexedOperands";
355 def VecListThreeDHWordIndexed : Operand<i32> {
356 let ParserMatchClass = VecListThreeDHWordIndexAsmOperand;
357 let MIOperandInfo = (ops DPR:$Vd, i32imm:$idx);
359 // ...with word lane subscripting.
360 def VecListThreeDWordIndexAsmOperand : AsmOperandClass {
361 let Name = "VecListThreeDWordIndexed";
362 let ParserMethod = "parseVectorList";
363 let RenderMethod = "addVecListIndexedOperands";
365 def VecListThreeDWordIndexed : Operand<i32> {
366 let ParserMatchClass = VecListThreeDWordIndexAsmOperand;
367 let MIOperandInfo = (ops DPR:$Vd, i32imm:$idx);
369 // Register list of three Q registers with half-word lane subscripting.
370 def VecListThreeQHWordIndexAsmOperand : AsmOperandClass {
371 let Name = "VecListThreeQHWordIndexed";
372 let ParserMethod = "parseVectorList";
373 let RenderMethod = "addVecListIndexedOperands";
375 def VecListThreeQHWordIndexed : Operand<i32> {
376 let ParserMatchClass = VecListThreeQHWordIndexAsmOperand;
377 let MIOperandInfo = (ops DPR:$Vd, i32imm:$idx);
379 // ...with word lane subscripting.
380 def VecListThreeQWordIndexAsmOperand : AsmOperandClass {
381 let Name = "VecListThreeQWordIndexed";
382 let ParserMethod = "parseVectorList";
383 let RenderMethod = "addVecListIndexedOperands";
385 def VecListThreeQWordIndexed : Operand<i32> {
386 let ParserMatchClass = VecListThreeQWordIndexAsmOperand;
387 let MIOperandInfo = (ops DPR:$Vd, i32imm:$idx);
390 // Register list of four D registers with byte lane subscripting.
391 def VecListFourDByteIndexAsmOperand : AsmOperandClass {
392 let Name = "VecListFourDByteIndexed";
393 let ParserMethod = "parseVectorList";
394 let RenderMethod = "addVecListIndexedOperands";
396 def VecListFourDByteIndexed : Operand<i32> {
397 let ParserMatchClass = VecListFourDByteIndexAsmOperand;
398 let MIOperandInfo = (ops DPR:$Vd, i32imm:$idx);
400 // ...with half-word lane subscripting.
401 def VecListFourDHWordIndexAsmOperand : AsmOperandClass {
402 let Name = "VecListFourDHWordIndexed";
403 let ParserMethod = "parseVectorList";
404 let RenderMethod = "addVecListIndexedOperands";
406 def VecListFourDHWordIndexed : Operand<i32> {
407 let ParserMatchClass = VecListFourDHWordIndexAsmOperand;
408 let MIOperandInfo = (ops DPR:$Vd, i32imm:$idx);
410 // ...with word lane subscripting.
411 def VecListFourDWordIndexAsmOperand : AsmOperandClass {
412 let Name = "VecListFourDWordIndexed";
413 let ParserMethod = "parseVectorList";
414 let RenderMethod = "addVecListIndexedOperands";
416 def VecListFourDWordIndexed : Operand<i32> {
417 let ParserMatchClass = VecListFourDWordIndexAsmOperand;
418 let MIOperandInfo = (ops DPR:$Vd, i32imm:$idx);
420 // Register list of four Q registers with half-word lane subscripting.
421 def VecListFourQHWordIndexAsmOperand : AsmOperandClass {
422 let Name = "VecListFourQHWordIndexed";
423 let ParserMethod = "parseVectorList";
424 let RenderMethod = "addVecListIndexedOperands";
426 def VecListFourQHWordIndexed : Operand<i32> {
427 let ParserMatchClass = VecListFourQHWordIndexAsmOperand;
428 let MIOperandInfo = (ops DPR:$Vd, i32imm:$idx);
430 // ...with word lane subscripting.
431 def VecListFourQWordIndexAsmOperand : AsmOperandClass {
432 let Name = "VecListFourQWordIndexed";
433 let ParserMethod = "parseVectorList";
434 let RenderMethod = "addVecListIndexedOperands";
436 def VecListFourQWordIndexed : Operand<i32> {
437 let ParserMatchClass = VecListFourQWordIndexAsmOperand;
438 let MIOperandInfo = (ops DPR:$Vd, i32imm:$idx);
441 def dword_alignedload : PatFrag<(ops node:$ptr), (load node:$ptr), [{
442 return cast<LoadSDNode>(N)->getAlignment() >= 8;
444 def dword_alignedstore : PatFrag<(ops node:$val, node:$ptr),
445 (store node:$val, node:$ptr), [{
446 return cast<StoreSDNode>(N)->getAlignment() >= 8;
448 def word_alignedload : PatFrag<(ops node:$ptr), (load node:$ptr), [{
449 return cast<LoadSDNode>(N)->getAlignment() == 4;
451 def word_alignedstore : PatFrag<(ops node:$val, node:$ptr),
452 (store node:$val, node:$ptr), [{
453 return cast<StoreSDNode>(N)->getAlignment() == 4;
455 def hword_alignedload : PatFrag<(ops node:$ptr), (load node:$ptr), [{
456 return cast<LoadSDNode>(N)->getAlignment() == 2;
458 def hword_alignedstore : PatFrag<(ops node:$val, node:$ptr),
459 (store node:$val, node:$ptr), [{
460 return cast<StoreSDNode>(N)->getAlignment() == 2;
462 def byte_alignedload : PatFrag<(ops node:$ptr), (load node:$ptr), [{
463 return cast<LoadSDNode>(N)->getAlignment() == 1;
465 def byte_alignedstore : PatFrag<(ops node:$val, node:$ptr),
466 (store node:$val, node:$ptr), [{
467 return cast<StoreSDNode>(N)->getAlignment() == 1;
469 def non_word_alignedload : PatFrag<(ops node:$ptr), (load node:$ptr), [{
470 return cast<LoadSDNode>(N)->getAlignment() < 4;
472 def non_word_alignedstore : PatFrag<(ops node:$val, node:$ptr),
473 (store node:$val, node:$ptr), [{
474 return cast<StoreSDNode>(N)->getAlignment() < 4;
477 //===----------------------------------------------------------------------===//
478 // NEON-specific DAG Nodes.
479 //===----------------------------------------------------------------------===//
481 def SDTARMVTST : SDTypeProfile<1, 2, [SDTCisInt<0>, SDTCisSameAs<1, 2>]>;
482 def NEONvtst : SDNode<"ARMISD::VTST", SDTARMVTST>;
484 // Types for vector shift by immediates. The "SHX" version is for long and
485 // narrow operations where the source and destination vectors have different
486 // types. The "SHINS" version is for shift and insert operations.
487 def SDTARMVSHXIMM : SDTypeProfile<1, 2, [SDTCisInt<0>, SDTCisInt<1>,
489 def SDTARMVSHINSIMM : SDTypeProfile<1, 3, [SDTCisInt<0>, SDTCisSameAs<0, 1>,
490 SDTCisSameAs<0, 2>, SDTCisVT<3, i32>]>;
492 def NEONvshrnImm : SDNode<"ARMISD::VSHRNIMM", SDTARMVSHXIMM>;
494 def NEONvrshrsImm : SDNode<"ARMISD::VRSHRsIMM", SDTARMVSHIMM>;
495 def NEONvrshruImm : SDNode<"ARMISD::VRSHRuIMM", SDTARMVSHIMM>;
496 def NEONvrshrnImm : SDNode<"ARMISD::VRSHRNIMM", SDTARMVSHXIMM>;
498 def NEONvqshlsImm : SDNode<"ARMISD::VQSHLsIMM", SDTARMVSHIMM>;
499 def NEONvqshluImm : SDNode<"ARMISD::VQSHLuIMM", SDTARMVSHIMM>;
500 def NEONvqshlsuImm : SDNode<"ARMISD::VQSHLsuIMM", SDTARMVSHIMM>;
501 def NEONvqshrnsImm : SDNode<"ARMISD::VQSHRNsIMM", SDTARMVSHXIMM>;
502 def NEONvqshrnuImm : SDNode<"ARMISD::VQSHRNuIMM", SDTARMVSHXIMM>;
503 def NEONvqshrnsuImm : SDNode<"ARMISD::VQSHRNsuIMM", SDTARMVSHXIMM>;
505 def NEONvqrshrnsImm : SDNode<"ARMISD::VQRSHRNsIMM", SDTARMVSHXIMM>;
506 def NEONvqrshrnuImm : SDNode<"ARMISD::VQRSHRNuIMM", SDTARMVSHXIMM>;
507 def NEONvqrshrnsuImm : SDNode<"ARMISD::VQRSHRNsuIMM", SDTARMVSHXIMM>;
509 def NEONvsliImm : SDNode<"ARMISD::VSLIIMM", SDTARMVSHINSIMM>;
510 def NEONvsriImm : SDNode<"ARMISD::VSRIIMM", SDTARMVSHINSIMM>;
512 def NEONvbsl : SDNode<"ARMISD::VBSL",
513 SDTypeProfile<1, 3, [SDTCisVec<0>,
516 SDTCisSameAs<0, 3>]>>;
518 def SDTARMVEXT : SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0, 1>,
519 SDTCisSameAs<0, 2>, SDTCisVT<3, i32>]>;
520 def NEONvext : SDNode<"ARMISD::VEXT", SDTARMVEXT>;
522 def SDTARMVSHUF2 : SDTypeProfile<2, 2, [SDTCisVec<0>, SDTCisSameAs<0, 1>,
524 SDTCisSameAs<0, 3>]>;
525 def NEONzip : SDNode<"ARMISD::VZIP", SDTARMVSHUF2>;
526 def NEONuzp : SDNode<"ARMISD::VUZP", SDTARMVSHUF2>;
527 def NEONtrn : SDNode<"ARMISD::VTRN", SDTARMVSHUF2>;
529 def SDTARMVTBL1 : SDTypeProfile<1, 2, [SDTCisVT<0, v8i8>, SDTCisVT<1, v8i8>,
531 def SDTARMVTBL2 : SDTypeProfile<1, 3, [SDTCisVT<0, v8i8>, SDTCisVT<1, v8i8>,
532 SDTCisVT<2, v8i8>, SDTCisVT<3, v8i8>]>;
533 def NEONvtbl1 : SDNode<"ARMISD::VTBL1", SDTARMVTBL1>;
534 def NEONvtbl2 : SDNode<"ARMISD::VTBL2", SDTARMVTBL2>;
537 def NEONimmAllZerosV: PatLeaf<(ARMvmovImm (i32 timm)), [{
538 ConstantSDNode *ConstVal = cast<ConstantSDNode>(N->getOperand(0));
539 unsigned EltBits = 0;
540 uint64_t EltVal = ARM_AM::decodeVMOVModImm(ConstVal->getZExtValue(), EltBits);
541 return (EltBits == 32 && EltVal == 0);
544 def NEONimmAllOnesV: PatLeaf<(ARMvmovImm (i32 timm)), [{
545 ConstantSDNode *ConstVal = cast<ConstantSDNode>(N->getOperand(0));
546 unsigned EltBits = 0;
547 uint64_t EltVal = ARM_AM::decodeVMOVModImm(ConstVal->getZExtValue(), EltBits);
548 return (EltBits == 8 && EltVal == 0xff);
551 //===----------------------------------------------------------------------===//
552 // NEON load / store instructions
553 //===----------------------------------------------------------------------===//
555 // Use VLDM to load a Q register as a D register pair.
556 // This is a pseudo instruction that is expanded to VLDMD after reg alloc.
558 : PseudoVFPLdStM<(outs DPair:$dst), (ins GPR:$Rn),
560 [(set DPair:$dst, (v2f64 (word_alignedload GPR:$Rn)))]>;
562 // Use VSTM to store a Q register as a D register pair.
563 // This is a pseudo instruction that is expanded to VSTMD after reg alloc.
565 : PseudoVFPLdStM<(outs), (ins DPair:$src, GPR:$Rn),
567 [(word_alignedstore (v2f64 DPair:$src), GPR:$Rn)]>;
569 // Classes for VLD* pseudo-instructions with multi-register operands.
570 // These are expanded to real instructions after register allocation.
571 class VLDQPseudo<InstrItinClass itin>
572 : PseudoNLdSt<(outs QPR:$dst), (ins addrmode6:$addr), itin, "">;
573 class VLDQWBPseudo<InstrItinClass itin>
574 : PseudoNLdSt<(outs QPR:$dst, GPR:$wb),
575 (ins addrmode6:$addr, am6offset:$offset), itin,
577 class VLDQWBfixedPseudo<InstrItinClass itin>
578 : PseudoNLdSt<(outs QPR:$dst, GPR:$wb),
579 (ins addrmode6:$addr), itin,
581 class VLDQWBregisterPseudo<InstrItinClass itin>
582 : PseudoNLdSt<(outs QPR:$dst, GPR:$wb),
583 (ins addrmode6:$addr, rGPR:$offset), itin,
586 class VLDQQPseudo<InstrItinClass itin>
587 : PseudoNLdSt<(outs QQPR:$dst), (ins addrmode6:$addr), itin, "">;
588 class VLDQQWBPseudo<InstrItinClass itin>
589 : PseudoNLdSt<(outs QQPR:$dst, GPR:$wb),
590 (ins addrmode6:$addr, am6offset:$offset), itin,
592 class VLDQQWBfixedPseudo<InstrItinClass itin>
593 : PseudoNLdSt<(outs QQPR:$dst, GPR:$wb),
594 (ins addrmode6:$addr), itin,
596 class VLDQQWBregisterPseudo<InstrItinClass itin>
597 : PseudoNLdSt<(outs QQPR:$dst, GPR:$wb),
598 (ins addrmode6:$addr, rGPR:$offset), itin,
602 class VLDQQQQPseudo<InstrItinClass itin>
603 : PseudoNLdSt<(outs QQQQPR:$dst), (ins addrmode6:$addr, QQQQPR:$src),itin,
605 class VLDQQQQWBPseudo<InstrItinClass itin>
606 : PseudoNLdSt<(outs QQQQPR:$dst, GPR:$wb),
607 (ins addrmode6:$addr, am6offset:$offset, QQQQPR:$src), itin,
608 "$addr.addr = $wb, $src = $dst">;
610 let mayLoad = 1, hasSideEffects = 0, hasExtraDefRegAllocReq = 1 in {
612 // VLD1 : Vector Load (multiple single elements)
613 class VLD1D<bits<4> op7_4, string Dt, Operand AddrMode>
614 : NLdSt<0,0b10,0b0111,op7_4, (outs VecListOneD:$Vd),
615 (ins AddrMode:$Rn), IIC_VLD1,
616 "vld1", Dt, "$Vd, $Rn", "", []>, Sched<[WriteVLD1]> {
619 let DecoderMethod = "DecodeVLDST1Instruction";
621 class VLD1Q<bits<4> op7_4, string Dt, Operand AddrMode>
622 : NLdSt<0,0b10,0b1010,op7_4, (outs VecListDPair:$Vd),
623 (ins AddrMode:$Rn), IIC_VLD1x2,
624 "vld1", Dt, "$Vd, $Rn", "", []>, Sched<[WriteVLD2]> {
626 let Inst{5-4} = Rn{5-4};
627 let DecoderMethod = "DecodeVLDST1Instruction";
630 def VLD1d8 : VLD1D<{0,0,0,?}, "8", addrmode6align64>;
631 def VLD1d16 : VLD1D<{0,1,0,?}, "16", addrmode6align64>;
632 def VLD1d32 : VLD1D<{1,0,0,?}, "32", addrmode6align64>;
633 def VLD1d64 : VLD1D<{1,1,0,?}, "64", addrmode6align64>;
635 def VLD1q8 : VLD1Q<{0,0,?,?}, "8", addrmode6align64or128>;
636 def VLD1q16 : VLD1Q<{0,1,?,?}, "16", addrmode6align64or128>;
637 def VLD1q32 : VLD1Q<{1,0,?,?}, "32", addrmode6align64or128>;
638 def VLD1q64 : VLD1Q<{1,1,?,?}, "64", addrmode6align64or128>;
640 // ...with address register writeback:
641 multiclass VLD1DWB<bits<4> op7_4, string Dt, Operand AddrMode> {
642 def _fixed : NLdSt<0,0b10, 0b0111,op7_4, (outs VecListOneD:$Vd, GPR:$wb),
643 (ins AddrMode:$Rn), IIC_VLD1u,
644 "vld1", Dt, "$Vd, $Rn!",
645 "$Rn.addr = $wb", []>, Sched<[WriteVLD1]> {
646 let Rm = 0b1101; // NLdSt will assign to the right encoding bits.
648 let DecoderMethod = "DecodeVLDST1Instruction";
650 def _register : NLdSt<0,0b10,0b0111,op7_4, (outs VecListOneD:$Vd, GPR:$wb),
651 (ins AddrMode:$Rn, rGPR:$Rm), IIC_VLD1u,
652 "vld1", Dt, "$Vd, $Rn, $Rm",
653 "$Rn.addr = $wb", []>, Sched<[WriteVLD1]> {
655 let DecoderMethod = "DecodeVLDST1Instruction";
658 multiclass VLD1QWB<bits<4> op7_4, string Dt, Operand AddrMode> {
659 def _fixed : NLdSt<0,0b10,0b1010,op7_4, (outs VecListDPair:$Vd, GPR:$wb),
660 (ins AddrMode:$Rn), IIC_VLD1x2u,
661 "vld1", Dt, "$Vd, $Rn!",
662 "$Rn.addr = $wb", []>, Sched<[WriteVLD2]> {
663 let Rm = 0b1101; // NLdSt will assign to the right encoding bits.
664 let Inst{5-4} = Rn{5-4};
665 let DecoderMethod = "DecodeVLDST1Instruction";
667 def _register : NLdSt<0,0b10,0b1010,op7_4, (outs VecListDPair:$Vd, GPR:$wb),
668 (ins AddrMode:$Rn, rGPR:$Rm), IIC_VLD1x2u,
669 "vld1", Dt, "$Vd, $Rn, $Rm",
670 "$Rn.addr = $wb", []>, Sched<[WriteVLD2]> {
671 let Inst{5-4} = Rn{5-4};
672 let DecoderMethod = "DecodeVLDST1Instruction";
676 defm VLD1d8wb : VLD1DWB<{0,0,0,?}, "8", addrmode6align64>;
677 defm VLD1d16wb : VLD1DWB<{0,1,0,?}, "16", addrmode6align64>;
678 defm VLD1d32wb : VLD1DWB<{1,0,0,?}, "32", addrmode6align64>;
679 defm VLD1d64wb : VLD1DWB<{1,1,0,?}, "64", addrmode6align64>;
680 defm VLD1q8wb : VLD1QWB<{0,0,?,?}, "8", addrmode6align64or128>;
681 defm VLD1q16wb : VLD1QWB<{0,1,?,?}, "16", addrmode6align64or128>;
682 defm VLD1q32wb : VLD1QWB<{1,0,?,?}, "32", addrmode6align64or128>;
683 defm VLD1q64wb : VLD1QWB<{1,1,?,?}, "64", addrmode6align64or128>;
685 // ...with 3 registers
686 class VLD1D3<bits<4> op7_4, string Dt, Operand AddrMode>
687 : NLdSt<0,0b10,0b0110,op7_4, (outs VecListThreeD:$Vd),
688 (ins AddrMode:$Rn), IIC_VLD1x3, "vld1", Dt,
689 "$Vd, $Rn", "", []>, Sched<[WriteVLD3]> {
692 let DecoderMethod = "DecodeVLDST1Instruction";
694 multiclass VLD1D3WB<bits<4> op7_4, string Dt, Operand AddrMode> {
695 def _fixed : NLdSt<0,0b10,0b0110, op7_4, (outs VecListThreeD:$Vd, GPR:$wb),
696 (ins AddrMode:$Rn), IIC_VLD1x2u,
697 "vld1", Dt, "$Vd, $Rn!",
698 "$Rn.addr = $wb", []>, Sched<[WriteVLD3]> {
699 let Rm = 0b1101; // NLdSt will assign to the right encoding bits.
701 let DecoderMethod = "DecodeVLDST1Instruction";
703 def _register : NLdSt<0,0b10,0b0110,op7_4, (outs VecListThreeD:$Vd, GPR:$wb),
704 (ins AddrMode:$Rn, rGPR:$Rm), IIC_VLD1x2u,
705 "vld1", Dt, "$Vd, $Rn, $Rm",
706 "$Rn.addr = $wb", []>, Sched<[WriteVLD3]> {
708 let DecoderMethod = "DecodeVLDST1Instruction";
712 def VLD1d8T : VLD1D3<{0,0,0,?}, "8", addrmode6align64>;
713 def VLD1d16T : VLD1D3<{0,1,0,?}, "16", addrmode6align64>;
714 def VLD1d32T : VLD1D3<{1,0,0,?}, "32", addrmode6align64>;
715 def VLD1d64T : VLD1D3<{1,1,0,?}, "64", addrmode6align64>;
717 defm VLD1d8Twb : VLD1D3WB<{0,0,0,?}, "8", addrmode6align64>;
718 defm VLD1d16Twb : VLD1D3WB<{0,1,0,?}, "16", addrmode6align64>;
719 defm VLD1d32Twb : VLD1D3WB<{1,0,0,?}, "32", addrmode6align64>;
720 defm VLD1d64Twb : VLD1D3WB<{1,1,0,?}, "64", addrmode6align64>;
722 def VLD1d8TPseudo : VLDQQPseudo<IIC_VLD1x3>, Sched<[WriteVLD3]>;
723 def VLD1d16TPseudo : VLDQQPseudo<IIC_VLD1x3>, Sched<[WriteVLD3]>;
724 def VLD1d32TPseudo : VLDQQPseudo<IIC_VLD1x3>, Sched<[WriteVLD3]>;
725 def VLD1d64TPseudo : VLDQQPseudo<IIC_VLD1x3>, Sched<[WriteVLD3]>;
726 def VLD1d64TPseudoWB_fixed : VLDQQWBfixedPseudo<IIC_VLD1x3>, Sched<[WriteVLD3]>;
727 def VLD1d64TPseudoWB_register : VLDQQWBregisterPseudo<IIC_VLD1x3>, Sched<[WriteVLD3]>;
729 def VLD1q8HighTPseudo : VLDQQQQPseudo<IIC_VLD1x3>, Sched<[WriteVLD3]>;
730 def VLD1q8LowTPseudo_UPD : VLDQQQQWBPseudo<IIC_VLD1x3>, Sched<[WriteVLD3]>;
731 def VLD1q16HighTPseudo : VLDQQQQPseudo<IIC_VLD1x3>, Sched<[WriteVLD3]>;
732 def VLD1q16LowTPseudo_UPD : VLDQQQQWBPseudo<IIC_VLD1x3>, Sched<[WriteVLD3]>;
733 def VLD1q32HighTPseudo : VLDQQQQPseudo<IIC_VLD1x3>, Sched<[WriteVLD3]>;
734 def VLD1q32LowTPseudo_UPD : VLDQQQQWBPseudo<IIC_VLD1x3>, Sched<[WriteVLD3]>;
735 def VLD1q64HighTPseudo : VLDQQQQPseudo<IIC_VLD1x3>, Sched<[WriteVLD3]>;
736 def VLD1q64LowTPseudo_UPD : VLDQQQQWBPseudo<IIC_VLD1x3>, Sched<[WriteVLD3]>;
738 // ...with 4 registers
739 class VLD1D4<bits<4> op7_4, string Dt, Operand AddrMode>
740 : NLdSt<0, 0b10, 0b0010, op7_4, (outs VecListFourD:$Vd),
741 (ins AddrMode:$Rn), IIC_VLD1x4, "vld1", Dt,
742 "$Vd, $Rn", "", []>, Sched<[WriteVLD4]> {
744 let Inst{5-4} = Rn{5-4};
745 let DecoderMethod = "DecodeVLDST1Instruction";
747 multiclass VLD1D4WB<bits<4> op7_4, string Dt, Operand AddrMode> {
748 def _fixed : NLdSt<0,0b10,0b0010, op7_4, (outs VecListFourD:$Vd, GPR:$wb),
749 (ins AddrMode:$Rn), IIC_VLD1x2u,
750 "vld1", Dt, "$Vd, $Rn!",
751 "$Rn.addr = $wb", []>, Sched<[WriteVLD4]> {
752 let Rm = 0b1101; // NLdSt will assign to the right encoding bits.
753 let Inst{5-4} = Rn{5-4};
754 let DecoderMethod = "DecodeVLDST1Instruction";
756 def _register : NLdSt<0,0b10,0b0010,op7_4, (outs VecListFourD:$Vd, GPR:$wb),
757 (ins AddrMode:$Rn, rGPR:$Rm), IIC_VLD1x2u,
758 "vld1", Dt, "$Vd, $Rn, $Rm",
759 "$Rn.addr = $wb", []>, Sched<[WriteVLD4]> {
760 let Inst{5-4} = Rn{5-4};
761 let DecoderMethod = "DecodeVLDST1Instruction";
765 def VLD1d8Q : VLD1D4<{0,0,?,?}, "8", addrmode6align64or128or256>;
766 def VLD1d16Q : VLD1D4<{0,1,?,?}, "16", addrmode6align64or128or256>;
767 def VLD1d32Q : VLD1D4<{1,0,?,?}, "32", addrmode6align64or128or256>;
768 def VLD1d64Q : VLD1D4<{1,1,?,?}, "64", addrmode6align64or128or256>;
770 defm VLD1d8Qwb : VLD1D4WB<{0,0,?,?}, "8", addrmode6align64or128or256>;
771 defm VLD1d16Qwb : VLD1D4WB<{0,1,?,?}, "16", addrmode6align64or128or256>;
772 defm VLD1d32Qwb : VLD1D4WB<{1,0,?,?}, "32", addrmode6align64or128or256>;
773 defm VLD1d64Qwb : VLD1D4WB<{1,1,?,?}, "64", addrmode6align64or128or256>;
775 def VLD1d8QPseudo : VLDQQPseudo<IIC_VLD1x4>, Sched<[WriteVLD4]>;
776 def VLD1d16QPseudo : VLDQQPseudo<IIC_VLD1x4>, Sched<[WriteVLD4]>;
777 def VLD1d32QPseudo : VLDQQPseudo<IIC_VLD1x4>, Sched<[WriteVLD4]>;
778 def VLD1d64QPseudo : VLDQQPseudo<IIC_VLD1x4>, Sched<[WriteVLD4]>;
779 def VLD1d64QPseudoWB_fixed : VLDQQWBfixedPseudo<IIC_VLD1x4>, Sched<[WriteVLD4]>;
780 def VLD1d64QPseudoWB_register : VLDQQWBregisterPseudo<IIC_VLD1x4>, Sched<[WriteVLD4]>;
782 def VLD1q8LowQPseudo_UPD : VLDQQQQWBPseudo<IIC_VLD1x4>, Sched<[WriteVLD4]>;
783 def VLD1q8HighQPseudo : VLDQQQQPseudo<IIC_VLD1x4>, Sched<[WriteVLD4]>;
784 def VLD1q16LowQPseudo_UPD : VLDQQQQWBPseudo<IIC_VLD1x4>, Sched<[WriteVLD4]>;
785 def VLD1q16HighQPseudo : VLDQQQQPseudo<IIC_VLD1x4>, Sched<[WriteVLD4]>;
786 def VLD1q32LowQPseudo_UPD : VLDQQQQWBPseudo<IIC_VLD1x4>, Sched<[WriteVLD4]>;
787 def VLD1q32HighQPseudo : VLDQQQQPseudo<IIC_VLD1x4>, Sched<[WriteVLD4]>;
788 def VLD1q64LowQPseudo_UPD : VLDQQQQWBPseudo<IIC_VLD1x4>, Sched<[WriteVLD4]>;
789 def VLD1q64HighQPseudo : VLDQQQQPseudo<IIC_VLD1x4>, Sched<[WriteVLD4]>;
791 // VLD2 : Vector Load (multiple 2-element structures)
792 class VLD2<bits<4> op11_8, bits<4> op7_4, string Dt, RegisterOperand VdTy,
793 InstrItinClass itin, Operand AddrMode>
794 : NLdSt<0, 0b10, op11_8, op7_4, (outs VdTy:$Vd),
795 (ins AddrMode:$Rn), itin,
796 "vld2", Dt, "$Vd, $Rn", "", []> {
798 let Inst{5-4} = Rn{5-4};
799 let DecoderMethod = "DecodeVLDST2Instruction";
802 def VLD2d8 : VLD2<0b1000, {0,0,?,?}, "8", VecListDPair, IIC_VLD2,
803 addrmode6align64or128>, Sched<[WriteVLD2]>;
804 def VLD2d16 : VLD2<0b1000, {0,1,?,?}, "16", VecListDPair, IIC_VLD2,
805 addrmode6align64or128>, Sched<[WriteVLD2]>;
806 def VLD2d32 : VLD2<0b1000, {1,0,?,?}, "32", VecListDPair, IIC_VLD2,
807 addrmode6align64or128>, Sched<[WriteVLD2]>;
809 def VLD2q8 : VLD2<0b0011, {0,0,?,?}, "8", VecListFourD, IIC_VLD2x2,
810 addrmode6align64or128or256>, Sched<[WriteVLD4]>;
811 def VLD2q16 : VLD2<0b0011, {0,1,?,?}, "16", VecListFourD, IIC_VLD2x2,
812 addrmode6align64or128or256>, Sched<[WriteVLD4]>;
813 def VLD2q32 : VLD2<0b0011, {1,0,?,?}, "32", VecListFourD, IIC_VLD2x2,
814 addrmode6align64or128or256>, Sched<[WriteVLD4]>;
816 def VLD2q8Pseudo : VLDQQPseudo<IIC_VLD2x2>, Sched<[WriteVLD4]>;
817 def VLD2q16Pseudo : VLDQQPseudo<IIC_VLD2x2>, Sched<[WriteVLD4]>;
818 def VLD2q32Pseudo : VLDQQPseudo<IIC_VLD2x2>, Sched<[WriteVLD4]>;
820 // ...with address register writeback:
821 multiclass VLD2WB<bits<4> op11_8, bits<4> op7_4, string Dt,
822 RegisterOperand VdTy, InstrItinClass itin, Operand AddrMode> {
823 def _fixed : NLdSt<0, 0b10, op11_8, op7_4, (outs VdTy:$Vd, GPR:$wb),
824 (ins AddrMode:$Rn), itin,
825 "vld2", Dt, "$Vd, $Rn!",
826 "$Rn.addr = $wb", []> {
827 let Rm = 0b1101; // NLdSt will assign to the right encoding bits.
828 let Inst{5-4} = Rn{5-4};
829 let DecoderMethod = "DecodeVLDST2Instruction";
831 def _register : NLdSt<0, 0b10, op11_8, op7_4, (outs VdTy:$Vd, GPR:$wb),
832 (ins AddrMode:$Rn, rGPR:$Rm), itin,
833 "vld2", Dt, "$Vd, $Rn, $Rm",
834 "$Rn.addr = $wb", []> {
835 let Inst{5-4} = Rn{5-4};
836 let DecoderMethod = "DecodeVLDST2Instruction";
840 defm VLD2d8wb : VLD2WB<0b1000, {0,0,?,?}, "8", VecListDPair, IIC_VLD2u,
841 addrmode6align64or128>, Sched<[WriteVLD2]>;
842 defm VLD2d16wb : VLD2WB<0b1000, {0,1,?,?}, "16", VecListDPair, IIC_VLD2u,
843 addrmode6align64or128>, Sched<[WriteVLD2]>;
844 defm VLD2d32wb : VLD2WB<0b1000, {1,0,?,?}, "32", VecListDPair, IIC_VLD2u,
845 addrmode6align64or128>, Sched<[WriteVLD2]>;
847 defm VLD2q8wb : VLD2WB<0b0011, {0,0,?,?}, "8", VecListFourD, IIC_VLD2x2u,
848 addrmode6align64or128or256>, Sched<[WriteVLD4]>;
849 defm VLD2q16wb : VLD2WB<0b0011, {0,1,?,?}, "16", VecListFourD, IIC_VLD2x2u,
850 addrmode6align64or128or256>, Sched<[WriteVLD4]>;
851 defm VLD2q32wb : VLD2WB<0b0011, {1,0,?,?}, "32", VecListFourD, IIC_VLD2x2u,
852 addrmode6align64or128or256>, Sched<[WriteVLD4]>;
854 def VLD2q8PseudoWB_fixed : VLDQQWBfixedPseudo<IIC_VLD2x2u>, Sched<[WriteVLD4]>;
855 def VLD2q16PseudoWB_fixed : VLDQQWBfixedPseudo<IIC_VLD2x2u>, Sched<[WriteVLD4]>;
856 def VLD2q32PseudoWB_fixed : VLDQQWBfixedPseudo<IIC_VLD2x2u>, Sched<[WriteVLD4]>;
857 def VLD2q8PseudoWB_register : VLDQQWBregisterPseudo<IIC_VLD2x2u>, Sched<[WriteVLD4]>;
858 def VLD2q16PseudoWB_register : VLDQQWBregisterPseudo<IIC_VLD2x2u>, Sched<[WriteVLD4]>;
859 def VLD2q32PseudoWB_register : VLDQQWBregisterPseudo<IIC_VLD2x2u>, Sched<[WriteVLD4]>;
861 // ...with double-spaced registers
862 def VLD2b8 : VLD2<0b1001, {0,0,?,?}, "8", VecListDPairSpaced, IIC_VLD2,
863 addrmode6align64or128>, Sched<[WriteVLD2]>;
864 def VLD2b16 : VLD2<0b1001, {0,1,?,?}, "16", VecListDPairSpaced, IIC_VLD2,
865 addrmode6align64or128>, Sched<[WriteVLD2]>;
866 def VLD2b32 : VLD2<0b1001, {1,0,?,?}, "32", VecListDPairSpaced, IIC_VLD2,
867 addrmode6align64or128>, Sched<[WriteVLD2]>;
868 defm VLD2b8wb : VLD2WB<0b1001, {0,0,?,?}, "8", VecListDPairSpaced, IIC_VLD2u,
869 addrmode6align64or128>, Sched<[WriteVLD2]>;
870 defm VLD2b16wb : VLD2WB<0b1001, {0,1,?,?}, "16", VecListDPairSpaced, IIC_VLD2u,
871 addrmode6align64or128>, Sched<[WriteVLD2]>;
872 defm VLD2b32wb : VLD2WB<0b1001, {1,0,?,?}, "32", VecListDPairSpaced, IIC_VLD2u,
873 addrmode6align64or128>, Sched<[WriteVLD2]>;
875 // VLD3 : Vector Load (multiple 3-element structures)
876 class VLD3D<bits<4> op11_8, bits<4> op7_4, string Dt>
877 : NLdSt<0, 0b10, op11_8, op7_4, (outs DPR:$Vd, DPR:$dst2, DPR:$dst3),
878 (ins addrmode6:$Rn), IIC_VLD3,
879 "vld3", Dt, "\\{$Vd, $dst2, $dst3\\}, $Rn", "", []>, Sched<[WriteVLD3]> {
882 let DecoderMethod = "DecodeVLDST3Instruction";
885 def VLD3d8 : VLD3D<0b0100, {0,0,0,?}, "8">;
886 def VLD3d16 : VLD3D<0b0100, {0,1,0,?}, "16">;
887 def VLD3d32 : VLD3D<0b0100, {1,0,0,?}, "32">;
889 def VLD3d8Pseudo : VLDQQPseudo<IIC_VLD3>, Sched<[WriteVLD3]>;
890 def VLD3d16Pseudo : VLDQQPseudo<IIC_VLD3>, Sched<[WriteVLD3]>;
891 def VLD3d32Pseudo : VLDQQPseudo<IIC_VLD3>, Sched<[WriteVLD3]>;
893 // ...with address register writeback:
894 class VLD3DWB<bits<4> op11_8, bits<4> op7_4, string Dt>
895 : NLdSt<0, 0b10, op11_8, op7_4,
896 (outs DPR:$Vd, DPR:$dst2, DPR:$dst3, GPR:$wb),
897 (ins addrmode6:$Rn, am6offset:$Rm), IIC_VLD3u,
898 "vld3", Dt, "\\{$Vd, $dst2, $dst3\\}, $Rn$Rm",
899 "$Rn.addr = $wb", []>, Sched<[WriteVLD3]> {
901 let DecoderMethod = "DecodeVLDST3Instruction";
904 def VLD3d8_UPD : VLD3DWB<0b0100, {0,0,0,?}, "8">;
905 def VLD3d16_UPD : VLD3DWB<0b0100, {0,1,0,?}, "16">;
906 def VLD3d32_UPD : VLD3DWB<0b0100, {1,0,0,?}, "32">;
908 def VLD3d8Pseudo_UPD : VLDQQWBPseudo<IIC_VLD3u>, Sched<[WriteVLD3]>;
909 def VLD3d16Pseudo_UPD : VLDQQWBPseudo<IIC_VLD3u>, Sched<[WriteVLD3]>;
910 def VLD3d32Pseudo_UPD : VLDQQWBPseudo<IIC_VLD3u>, Sched<[WriteVLD3]>;
912 // ...with double-spaced registers:
913 def VLD3q8 : VLD3D<0b0101, {0,0,0,?}, "8">;
914 def VLD3q16 : VLD3D<0b0101, {0,1,0,?}, "16">;
915 def VLD3q32 : VLD3D<0b0101, {1,0,0,?}, "32">;
916 def VLD3q8_UPD : VLD3DWB<0b0101, {0,0,0,?}, "8">;
917 def VLD3q16_UPD : VLD3DWB<0b0101, {0,1,0,?}, "16">;
918 def VLD3q32_UPD : VLD3DWB<0b0101, {1,0,0,?}, "32">;
920 def VLD3q8Pseudo_UPD : VLDQQQQWBPseudo<IIC_VLD3u>, Sched<[WriteVLD3]>;
921 def VLD3q16Pseudo_UPD : VLDQQQQWBPseudo<IIC_VLD3u>, Sched<[WriteVLD3]>;
922 def VLD3q32Pseudo_UPD : VLDQQQQWBPseudo<IIC_VLD3u>, Sched<[WriteVLD3]>;
924 // ...alternate versions to be allocated odd register numbers:
925 def VLD3q8oddPseudo : VLDQQQQPseudo<IIC_VLD3>, Sched<[WriteVLD3]>;
926 def VLD3q16oddPseudo : VLDQQQQPseudo<IIC_VLD3>, Sched<[WriteVLD3]>;
927 def VLD3q32oddPseudo : VLDQQQQPseudo<IIC_VLD3>, Sched<[WriteVLD3]>;
929 def VLD3q8oddPseudo_UPD : VLDQQQQWBPseudo<IIC_VLD3u>, Sched<[WriteVLD3]>;
930 def VLD3q16oddPseudo_UPD : VLDQQQQWBPseudo<IIC_VLD3u>, Sched<[WriteVLD3]>;
931 def VLD3q32oddPseudo_UPD : VLDQQQQWBPseudo<IIC_VLD3u>, Sched<[WriteVLD3]>;
933 // VLD4 : Vector Load (multiple 4-element structures)
934 class VLD4D<bits<4> op11_8, bits<4> op7_4, string Dt>
935 : NLdSt<0, 0b10, op11_8, op7_4,
936 (outs DPR:$Vd, DPR:$dst2, DPR:$dst3, DPR:$dst4),
937 (ins addrmode6:$Rn), IIC_VLD4,
938 "vld4", Dt, "\\{$Vd, $dst2, $dst3, $dst4\\}, $Rn", "", []>,
941 let Inst{5-4} = Rn{5-4};
942 let DecoderMethod = "DecodeVLDST4Instruction";
945 def VLD4d8 : VLD4D<0b0000, {0,0,?,?}, "8">;
946 def VLD4d16 : VLD4D<0b0000, {0,1,?,?}, "16">;
947 def VLD4d32 : VLD4D<0b0000, {1,0,?,?}, "32">;
949 def VLD4d8Pseudo : VLDQQPseudo<IIC_VLD4>, Sched<[WriteVLD4]>;
950 def VLD4d16Pseudo : VLDQQPseudo<IIC_VLD4>, Sched<[WriteVLD4]>;
951 def VLD4d32Pseudo : VLDQQPseudo<IIC_VLD4>, Sched<[WriteVLD4]>;
953 // ...with address register writeback:
954 class VLD4DWB<bits<4> op11_8, bits<4> op7_4, string Dt>
955 : NLdSt<0, 0b10, op11_8, op7_4,
956 (outs DPR:$Vd, DPR:$dst2, DPR:$dst3, DPR:$dst4, GPR:$wb),
957 (ins addrmode6:$Rn, am6offset:$Rm), IIC_VLD4u,
958 "vld4", Dt, "\\{$Vd, $dst2, $dst3, $dst4\\}, $Rn$Rm",
959 "$Rn.addr = $wb", []>, Sched<[WriteVLD4]> {
960 let Inst{5-4} = Rn{5-4};
961 let DecoderMethod = "DecodeVLDST4Instruction";
964 def VLD4d8_UPD : VLD4DWB<0b0000, {0,0,?,?}, "8">;
965 def VLD4d16_UPD : VLD4DWB<0b0000, {0,1,?,?}, "16">;
966 def VLD4d32_UPD : VLD4DWB<0b0000, {1,0,?,?}, "32">;
968 def VLD4d8Pseudo_UPD : VLDQQWBPseudo<IIC_VLD4u>, Sched<[WriteVLD4]>;
969 def VLD4d16Pseudo_UPD : VLDQQWBPseudo<IIC_VLD4u>, Sched<[WriteVLD4]>;
970 def VLD4d32Pseudo_UPD : VLDQQWBPseudo<IIC_VLD4u>, Sched<[WriteVLD4]>;
972 // ...with double-spaced registers:
973 def VLD4q8 : VLD4D<0b0001, {0,0,?,?}, "8">;
974 def VLD4q16 : VLD4D<0b0001, {0,1,?,?}, "16">;
975 def VLD4q32 : VLD4D<0b0001, {1,0,?,?}, "32">;
976 def VLD4q8_UPD : VLD4DWB<0b0001, {0,0,?,?}, "8">;
977 def VLD4q16_UPD : VLD4DWB<0b0001, {0,1,?,?}, "16">;
978 def VLD4q32_UPD : VLD4DWB<0b0001, {1,0,?,?}, "32">;
980 def VLD4q8Pseudo_UPD : VLDQQQQWBPseudo<IIC_VLD4u>, Sched<[WriteVLD4]>;
981 def VLD4q16Pseudo_UPD : VLDQQQQWBPseudo<IIC_VLD4u>, Sched<[WriteVLD4]>;
982 def VLD4q32Pseudo_UPD : VLDQQQQWBPseudo<IIC_VLD4u>, Sched<[WriteVLD4]>;
984 // ...alternate versions to be allocated odd register numbers:
985 def VLD4q8oddPseudo : VLDQQQQPseudo<IIC_VLD4>, Sched<[WriteVLD4]>;
986 def VLD4q16oddPseudo : VLDQQQQPseudo<IIC_VLD4>, Sched<[WriteVLD4]>;
987 def VLD4q32oddPseudo : VLDQQQQPseudo<IIC_VLD4>, Sched<[WriteVLD4]>;
989 def VLD4q8oddPseudo_UPD : VLDQQQQWBPseudo<IIC_VLD4u>, Sched<[WriteVLD4]>;
990 def VLD4q16oddPseudo_UPD : VLDQQQQWBPseudo<IIC_VLD4u>, Sched<[WriteVLD4]>;
991 def VLD4q32oddPseudo_UPD : VLDQQQQWBPseudo<IIC_VLD4u>, Sched<[WriteVLD4]>;
993 } // mayLoad = 1, hasSideEffects = 0, hasExtraDefRegAllocReq = 1
995 // Classes for VLD*LN pseudo-instructions with multi-register operands.
996 // These are expanded to real instructions after register allocation.
997 class VLDQLNPseudo<InstrItinClass itin>
998 : PseudoNLdSt<(outs QPR:$dst),
999 (ins addrmode6:$addr, QPR:$src, nohash_imm:$lane),
1000 itin, "$src = $dst">;
1001 class VLDQLNWBPseudo<InstrItinClass itin>
1002 : PseudoNLdSt<(outs QPR:$dst, GPR:$wb),
1003 (ins addrmode6:$addr, am6offset:$offset, QPR:$src,
1004 nohash_imm:$lane), itin, "$addr.addr = $wb, $src = $dst">;
1005 class VLDQQLNPseudo<InstrItinClass itin>
1006 : PseudoNLdSt<(outs QQPR:$dst),
1007 (ins addrmode6:$addr, QQPR:$src, nohash_imm:$lane),
1008 itin, "$src = $dst">;
1009 class VLDQQLNWBPseudo<InstrItinClass itin>
1010 : PseudoNLdSt<(outs QQPR:$dst, GPR:$wb),
1011 (ins addrmode6:$addr, am6offset:$offset, QQPR:$src,
1012 nohash_imm:$lane), itin, "$addr.addr = $wb, $src = $dst">;
1013 class VLDQQQQLNPseudo<InstrItinClass itin>
1014 : PseudoNLdSt<(outs QQQQPR:$dst),
1015 (ins addrmode6:$addr, QQQQPR:$src, nohash_imm:$lane),
1016 itin, "$src = $dst">;
1017 class VLDQQQQLNWBPseudo<InstrItinClass itin>
1018 : PseudoNLdSt<(outs QQQQPR:$dst, GPR:$wb),
1019 (ins addrmode6:$addr, am6offset:$offset, QQQQPR:$src,
1020 nohash_imm:$lane), itin, "$addr.addr = $wb, $src = $dst">;
1022 // VLD1LN : Vector Load (single element to one lane)
1023 class VLD1LN<bits<4> op11_8, bits<4> op7_4, string Dt, ValueType Ty,
1025 : NLdStLn<1, 0b10, op11_8, op7_4, (outs DPR:$Vd),
1026 (ins addrmode6:$Rn, DPR:$src, nohash_imm:$lane),
1027 IIC_VLD1ln, "vld1", Dt, "\\{$Vd[$lane]\\}, $Rn",
1029 [(set DPR:$Vd, (vector_insert (Ty DPR:$src),
1030 (i32 (LoadOp addrmode6:$Rn)),
1033 let DecoderMethod = "DecodeVLD1LN";
1035 class VLD1LN32<bits<4> op11_8, bits<4> op7_4, string Dt, ValueType Ty,
1037 : NLdStLn<1, 0b10, op11_8, op7_4, (outs DPR:$Vd),
1038 (ins addrmode6oneL32:$Rn, DPR:$src, nohash_imm:$lane),
1039 IIC_VLD1ln, "vld1", Dt, "\\{$Vd[$lane]\\}, $Rn",
1041 [(set DPR:$Vd, (vector_insert (Ty DPR:$src),
1042 (i32 (LoadOp addrmode6oneL32:$Rn)),
1043 imm:$lane))]>, Sched<[WriteVLD1]> {
1045 let DecoderMethod = "DecodeVLD1LN";
1047 class VLD1QLNPseudo<ValueType Ty, PatFrag LoadOp> : VLDQLNPseudo<IIC_VLD1ln>,
1048 Sched<[WriteVLD1]> {
1049 let Pattern = [(set QPR:$dst, (vector_insert (Ty QPR:$src),
1050 (i32 (LoadOp addrmode6:$addr)),
1054 def VLD1LNd8 : VLD1LN<0b0000, {?,?,?,0}, "8", v8i8, extloadi8> {
1055 let Inst{7-5} = lane{2-0};
1057 def VLD1LNd16 : VLD1LN<0b0100, {?,?,0,?}, "16", v4i16, extloadi16> {
1058 let Inst{7-6} = lane{1-0};
1059 let Inst{5-4} = Rn{5-4};
1061 def VLD1LNd32 : VLD1LN32<0b1000, {?,0,?,?}, "32", v2i32, load> {
1062 let Inst{7} = lane{0};
1063 let Inst{5-4} = Rn{5-4};
1066 def VLD1LNq8Pseudo : VLD1QLNPseudo<v16i8, extloadi8>;
1067 def VLD1LNq16Pseudo : VLD1QLNPseudo<v8i16, extloadi16>;
1068 def VLD1LNq32Pseudo : VLD1QLNPseudo<v4i32, load>;
1070 let Predicates = [HasNEON] in {
1071 def : Pat<(vector_insert (v4f16 DPR:$src),
1072 (f16 (load addrmode6:$addr)), imm:$lane),
1073 (VLD1LNd16 addrmode6:$addr, DPR:$src, imm:$lane)>;
1074 def : Pat<(vector_insert (v8f16 QPR:$src),
1075 (f16 (load addrmode6:$addr)), imm:$lane),
1076 (VLD1LNq16Pseudo addrmode6:$addr, QPR:$src, imm:$lane)>;
1077 def : Pat<(vector_insert (v4bf16 DPR:$src),
1078 (bf16 (load addrmode6:$addr)), imm:$lane),
1079 (VLD1LNd16 addrmode6:$addr, DPR:$src, imm:$lane)>;
1080 def : Pat<(vector_insert (v8bf16 QPR:$src),
1081 (bf16 (load addrmode6:$addr)), imm:$lane),
1082 (VLD1LNq16Pseudo addrmode6:$addr, QPR:$src, imm:$lane)>;
1083 def : Pat<(vector_insert (v2f32 DPR:$src),
1084 (f32 (load addrmode6:$addr)), imm:$lane),
1085 (VLD1LNd32 addrmode6:$addr, DPR:$src, imm:$lane)>;
1086 def : Pat<(vector_insert (v4f32 QPR:$src),
1087 (f32 (load addrmode6:$addr)), imm:$lane),
1088 (VLD1LNq32Pseudo addrmode6:$addr, QPR:$src, imm:$lane)>;
1090 // A 64-bit subvector insert to the first 128-bit vector position
1091 // is a subregister copy that needs no instruction.
1092 def : Pat<(insert_subvector undef, (v1i64 DPR:$src), (i32 0)),
1093 (INSERT_SUBREG (v2i64 (IMPLICIT_DEF)), DPR:$src, dsub_0)>;
1094 def : Pat<(insert_subvector undef, (v2i32 DPR:$src), (i32 0)),
1095 (INSERT_SUBREG (v4i32 (IMPLICIT_DEF)), DPR:$src, dsub_0)>;
1096 def : Pat<(insert_subvector undef, (v2f32 DPR:$src), (i32 0)),
1097 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), DPR:$src, dsub_0)>;
1098 def : Pat<(insert_subvector undef, (v4i16 DPR:$src), (i32 0)),
1099 (INSERT_SUBREG (v8i16 (IMPLICIT_DEF)), DPR:$src, dsub_0)>;
1100 def : Pat<(insert_subvector undef, (v4f16 DPR:$src), (i32 0)),
1101 (INSERT_SUBREG (v8f16 (IMPLICIT_DEF)), DPR:$src, dsub_0)>;
1102 def : Pat<(insert_subvector (v16i8 undef), (v8i8 DPR:$src), (i32 0)),
1103 (INSERT_SUBREG (v16i8 (IMPLICIT_DEF)), DPR:$src, dsub_0)>;
1107 let mayLoad = 1, hasSideEffects = 0, hasExtraDefRegAllocReq = 1 in {
1109 // ...with address register writeback:
1110 class VLD1LNWB<bits<4> op11_8, bits<4> op7_4, string Dt>
1111 : NLdStLn<1, 0b10, op11_8, op7_4, (outs DPR:$Vd, GPR:$wb),
1112 (ins addrmode6:$Rn, am6offset:$Rm,
1113 DPR:$src, nohash_imm:$lane), IIC_VLD1lnu, "vld1", Dt,
1114 "\\{$Vd[$lane]\\}, $Rn$Rm",
1115 "$src = $Vd, $Rn.addr = $wb", []>, Sched<[WriteVLD1]> {
1116 let DecoderMethod = "DecodeVLD1LN";
1119 def VLD1LNd8_UPD : VLD1LNWB<0b0000, {?,?,?,0}, "8"> {
1120 let Inst{7-5} = lane{2-0};
1122 def VLD1LNd16_UPD : VLD1LNWB<0b0100, {?,?,0,?}, "16"> {
1123 let Inst{7-6} = lane{1-0};
1124 let Inst{4} = Rn{4};
1126 def VLD1LNd32_UPD : VLD1LNWB<0b1000, {?,0,?,?}, "32"> {
1127 let Inst{7} = lane{0};
1128 let Inst{5} = Rn{4};
1129 let Inst{4} = Rn{4};
1132 def VLD1LNq8Pseudo_UPD : VLDQLNWBPseudo<IIC_VLD1lnu>, Sched<[WriteVLD1]>;
1133 def VLD1LNq16Pseudo_UPD : VLDQLNWBPseudo<IIC_VLD1lnu>, Sched<[WriteVLD1]>;
1134 def VLD1LNq32Pseudo_UPD : VLDQLNWBPseudo<IIC_VLD1lnu>, Sched<[WriteVLD1]>;
1136 // VLD2LN : Vector Load (single 2-element structure to one lane)
1137 class VLD2LN<bits<4> op11_8, bits<4> op7_4, string Dt>
1138 : NLdStLn<1, 0b10, op11_8, op7_4, (outs DPR:$Vd, DPR:$dst2),
1139 (ins addrmode6:$Rn, DPR:$src1, DPR:$src2, nohash_imm:$lane),
1140 IIC_VLD2ln, "vld2", Dt, "\\{$Vd[$lane], $dst2[$lane]\\}, $Rn",
1141 "$src1 = $Vd, $src2 = $dst2", []>, Sched<[WriteVLD1]> {
1143 let Inst{4} = Rn{4};
1144 let DecoderMethod = "DecodeVLD2LN";
1147 def VLD2LNd8 : VLD2LN<0b0001, {?,?,?,?}, "8"> {
1148 let Inst{7-5} = lane{2-0};
1150 def VLD2LNd16 : VLD2LN<0b0101, {?,?,0,?}, "16"> {
1151 let Inst{7-6} = lane{1-0};
1153 def VLD2LNd32 : VLD2LN<0b1001, {?,0,0,?}, "32"> {
1154 let Inst{7} = lane{0};
1157 def VLD2LNd8Pseudo : VLDQLNPseudo<IIC_VLD2ln>, Sched<[WriteVLD1]>;
1158 def VLD2LNd16Pseudo : VLDQLNPseudo<IIC_VLD2ln>, Sched<[WriteVLD1]>;
1159 def VLD2LNd32Pseudo : VLDQLNPseudo<IIC_VLD2ln>, Sched<[WriteVLD1]>;
1161 // ...with double-spaced registers:
1162 def VLD2LNq16 : VLD2LN<0b0101, {?,?,1,?}, "16"> {
1163 let Inst{7-6} = lane{1-0};
1165 def VLD2LNq32 : VLD2LN<0b1001, {?,1,0,?}, "32"> {
1166 let Inst{7} = lane{0};
1169 def VLD2LNq16Pseudo : VLDQQLNPseudo<IIC_VLD2ln>, Sched<[WriteVLD1]>;
1170 def VLD2LNq32Pseudo : VLDQQLNPseudo<IIC_VLD2ln>, Sched<[WriteVLD1]>;
1172 // ...with address register writeback:
1173 class VLD2LNWB<bits<4> op11_8, bits<4> op7_4, string Dt>
1174 : NLdStLn<1, 0b10, op11_8, op7_4, (outs DPR:$Vd, DPR:$dst2, GPR:$wb),
1175 (ins addrmode6:$Rn, am6offset:$Rm,
1176 DPR:$src1, DPR:$src2, nohash_imm:$lane), IIC_VLD2lnu, "vld2", Dt,
1177 "\\{$Vd[$lane], $dst2[$lane]\\}, $Rn$Rm",
1178 "$src1 = $Vd, $src2 = $dst2, $Rn.addr = $wb", []> {
1179 let Inst{4} = Rn{4};
1180 let DecoderMethod = "DecodeVLD2LN";
1183 def VLD2LNd8_UPD : VLD2LNWB<0b0001, {?,?,?,?}, "8"> {
1184 let Inst{7-5} = lane{2-0};
1186 def VLD2LNd16_UPD : VLD2LNWB<0b0101, {?,?,0,?}, "16"> {
1187 let Inst{7-6} = lane{1-0};
1189 def VLD2LNd32_UPD : VLD2LNWB<0b1001, {?,0,0,?}, "32"> {
1190 let Inst{7} = lane{0};
1193 def VLD2LNd8Pseudo_UPD : VLDQLNWBPseudo<IIC_VLD2lnu>, Sched<[WriteVLD1]>;
1194 def VLD2LNd16Pseudo_UPD : VLDQLNWBPseudo<IIC_VLD2lnu>, Sched<[WriteVLD1]>;
1195 def VLD2LNd32Pseudo_UPD : VLDQLNWBPseudo<IIC_VLD2lnu>, Sched<[WriteVLD1]>;
1197 def VLD2LNq16_UPD : VLD2LNWB<0b0101, {?,?,1,?}, "16"> {
1198 let Inst{7-6} = lane{1-0};
1200 def VLD2LNq32_UPD : VLD2LNWB<0b1001, {?,1,0,?}, "32"> {
1201 let Inst{7} = lane{0};
1204 def VLD2LNq16Pseudo_UPD : VLDQQLNWBPseudo<IIC_VLD2lnu>, Sched<[WriteVLD1]>;
1205 def VLD2LNq32Pseudo_UPD : VLDQQLNWBPseudo<IIC_VLD2lnu>, Sched<[WriteVLD1]>;
1207 // VLD3LN : Vector Load (single 3-element structure to one lane)
1208 class VLD3LN<bits<4> op11_8, bits<4> op7_4, string Dt>
1209 : NLdStLn<1, 0b10, op11_8, op7_4, (outs DPR:$Vd, DPR:$dst2, DPR:$dst3),
1210 (ins addrmode6:$Rn, DPR:$src1, DPR:$src2, DPR:$src3,
1211 nohash_imm:$lane), IIC_VLD3ln, "vld3", Dt,
1212 "\\{$Vd[$lane], $dst2[$lane], $dst3[$lane]\\}, $Rn",
1213 "$src1 = $Vd, $src2 = $dst2, $src3 = $dst3", []>, Sched<[WriteVLD2]> {
1215 let DecoderMethod = "DecodeVLD3LN";
1218 def VLD3LNd8 : VLD3LN<0b0010, {?,?,?,0}, "8"> {
1219 let Inst{7-5} = lane{2-0};
1221 def VLD3LNd16 : VLD3LN<0b0110, {?,?,0,0}, "16"> {
1222 let Inst{7-6} = lane{1-0};
1224 def VLD3LNd32 : VLD3LN<0b1010, {?,0,0,0}, "32"> {
1225 let Inst{7} = lane{0};
1228 def VLD3LNd8Pseudo : VLDQQLNPseudo<IIC_VLD3ln>, Sched<[WriteVLD2]>;
1229 def VLD3LNd16Pseudo : VLDQQLNPseudo<IIC_VLD3ln>, Sched<[WriteVLD2]>;
1230 def VLD3LNd32Pseudo : VLDQQLNPseudo<IIC_VLD3ln>, Sched<[WriteVLD2]>;
1232 // ...with double-spaced registers:
1233 def VLD3LNq16 : VLD3LN<0b0110, {?,?,1,0}, "16"> {
1234 let Inst{7-6} = lane{1-0};
1236 def VLD3LNq32 : VLD3LN<0b1010, {?,1,0,0}, "32"> {
1237 let Inst{7} = lane{0};
1240 def VLD3LNq16Pseudo : VLDQQQQLNPseudo<IIC_VLD3ln>, Sched<[WriteVLD2]>;
1241 def VLD3LNq32Pseudo : VLDQQQQLNPseudo<IIC_VLD3ln>, Sched<[WriteVLD2]>;
1243 // ...with address register writeback:
1244 class VLD3LNWB<bits<4> op11_8, bits<4> op7_4, string Dt>
1245 : NLdStLn<1, 0b10, op11_8, op7_4,
1246 (outs DPR:$Vd, DPR:$dst2, DPR:$dst3, GPR:$wb),
1247 (ins addrmode6:$Rn, am6offset:$Rm,
1248 DPR:$src1, DPR:$src2, DPR:$src3, nohash_imm:$lane),
1249 IIC_VLD3lnu, "vld3", Dt,
1250 "\\{$Vd[$lane], $dst2[$lane], $dst3[$lane]\\}, $Rn$Rm",
1251 "$src1 = $Vd, $src2 = $dst2, $src3 = $dst3, $Rn.addr = $wb",
1252 []>, Sched<[WriteVLD2]> {
1253 let DecoderMethod = "DecodeVLD3LN";
1256 def VLD3LNd8_UPD : VLD3LNWB<0b0010, {?,?,?,0}, "8"> {
1257 let Inst{7-5} = lane{2-0};
1259 def VLD3LNd16_UPD : VLD3LNWB<0b0110, {?,?,0,0}, "16"> {
1260 let Inst{7-6} = lane{1-0};
1262 def VLD3LNd32_UPD : VLD3LNWB<0b1010, {?,0,0,0}, "32"> {
1263 let Inst{7} = lane{0};
1266 def VLD3LNd8Pseudo_UPD : VLDQQLNWBPseudo<IIC_VLD3lnu>, Sched<[WriteVLD2]>;
1267 def VLD3LNd16Pseudo_UPD : VLDQQLNWBPseudo<IIC_VLD3lnu>, Sched<[WriteVLD2]>;
1268 def VLD3LNd32Pseudo_UPD : VLDQQLNWBPseudo<IIC_VLD3lnu>, Sched<[WriteVLD2]>;
1270 def VLD3LNq16_UPD : VLD3LNWB<0b0110, {?,?,1,0}, "16"> {
1271 let Inst{7-6} = lane{1-0};
1273 def VLD3LNq32_UPD : VLD3LNWB<0b1010, {?,1,0,0}, "32"> {
1274 let Inst{7} = lane{0};
1277 def VLD3LNq16Pseudo_UPD : VLDQQQQLNWBPseudo<IIC_VLD3lnu>, Sched<[WriteVLD2]>;
1278 def VLD3LNq32Pseudo_UPD : VLDQQQQLNWBPseudo<IIC_VLD3lnu>, Sched<[WriteVLD2]>;
1280 // VLD4LN : Vector Load (single 4-element structure to one lane)
1281 class VLD4LN<bits<4> op11_8, bits<4> op7_4, string Dt>
1282 : NLdStLn<1, 0b10, op11_8, op7_4,
1283 (outs DPR:$Vd, DPR:$dst2, DPR:$dst3, DPR:$dst4),
1284 (ins addrmode6:$Rn, DPR:$src1, DPR:$src2, DPR:$src3, DPR:$src4,
1285 nohash_imm:$lane), IIC_VLD4ln, "vld4", Dt,
1286 "\\{$Vd[$lane], $dst2[$lane], $dst3[$lane], $dst4[$lane]\\}, $Rn",
1287 "$src1 = $Vd, $src2 = $dst2, $src3 = $dst3, $src4 = $dst4", []>,
1288 Sched<[WriteVLD2]> {
1290 let Inst{4} = Rn{4};
1291 let DecoderMethod = "DecodeVLD4LN";
1294 def VLD4LNd8 : VLD4LN<0b0011, {?,?,?,?}, "8"> {
1295 let Inst{7-5} = lane{2-0};
1297 def VLD4LNd16 : VLD4LN<0b0111, {?,?,0,?}, "16"> {
1298 let Inst{7-6} = lane{1-0};
1300 def VLD4LNd32 : VLD4LN<0b1011, {?,0,?,?}, "32"> {
1301 let Inst{7} = lane{0};
1302 let Inst{5} = Rn{5};
1305 def VLD4LNd8Pseudo : VLDQQLNPseudo<IIC_VLD4ln>, Sched<[WriteVLD2]>;
1306 def VLD4LNd16Pseudo : VLDQQLNPseudo<IIC_VLD4ln>, Sched<[WriteVLD2]>;
1307 def VLD4LNd32Pseudo : VLDQQLNPseudo<IIC_VLD4ln>, Sched<[WriteVLD2]>;
1309 // ...with double-spaced registers:
1310 def VLD4LNq16 : VLD4LN<0b0111, {?,?,1,?}, "16"> {
1311 let Inst{7-6} = lane{1-0};
1313 def VLD4LNq32 : VLD4LN<0b1011, {?,1,?,?}, "32"> {
1314 let Inst{7} = lane{0};
1315 let Inst{5} = Rn{5};
1318 def VLD4LNq16Pseudo : VLDQQQQLNPseudo<IIC_VLD4ln>, Sched<[WriteVLD2]>;
1319 def VLD4LNq32Pseudo : VLDQQQQLNPseudo<IIC_VLD4ln>, Sched<[WriteVLD2]>;
1321 // ...with address register writeback:
1322 class VLD4LNWB<bits<4> op11_8, bits<4> op7_4, string Dt>
1323 : NLdStLn<1, 0b10, op11_8, op7_4,
1324 (outs DPR:$Vd, DPR:$dst2, DPR:$dst3, DPR:$dst4, GPR:$wb),
1325 (ins addrmode6:$Rn, am6offset:$Rm,
1326 DPR:$src1, DPR:$src2, DPR:$src3, DPR:$src4, nohash_imm:$lane),
1327 IIC_VLD4lnu, "vld4", Dt,
1328 "\\{$Vd[$lane], $dst2[$lane], $dst3[$lane], $dst4[$lane]\\}, $Rn$Rm",
1329 "$src1 = $Vd, $src2 = $dst2, $src3 = $dst3, $src4 = $dst4, $Rn.addr = $wb",
1331 let Inst{4} = Rn{4};
1332 let DecoderMethod = "DecodeVLD4LN" ;
1335 def VLD4LNd8_UPD : VLD4LNWB<0b0011, {?,?,?,?}, "8"> {
1336 let Inst{7-5} = lane{2-0};
1338 def VLD4LNd16_UPD : VLD4LNWB<0b0111, {?,?,0,?}, "16"> {
1339 let Inst{7-6} = lane{1-0};
1341 def VLD4LNd32_UPD : VLD4LNWB<0b1011, {?,0,?,?}, "32"> {
1342 let Inst{7} = lane{0};
1343 let Inst{5} = Rn{5};
1346 def VLD4LNd8Pseudo_UPD : VLDQQLNWBPseudo<IIC_VLD4lnu>, Sched<[WriteVLD2]>;
1347 def VLD4LNd16Pseudo_UPD : VLDQQLNWBPseudo<IIC_VLD4lnu>, Sched<[WriteVLD2]>;
1348 def VLD4LNd32Pseudo_UPD : VLDQQLNWBPseudo<IIC_VLD4lnu>, Sched<[WriteVLD2]>;
1350 def VLD4LNq16_UPD : VLD4LNWB<0b0111, {?,?,1,?}, "16"> {
1351 let Inst{7-6} = lane{1-0};
1353 def VLD4LNq32_UPD : VLD4LNWB<0b1011, {?,1,?,?}, "32"> {
1354 let Inst{7} = lane{0};
1355 let Inst{5} = Rn{5};
1358 def VLD4LNq16Pseudo_UPD : VLDQQQQLNWBPseudo<IIC_VLD4lnu>, Sched<[WriteVLD2]>;
1359 def VLD4LNq32Pseudo_UPD : VLDQQQQLNWBPseudo<IIC_VLD4lnu>, Sched<[WriteVLD2]>;
1361 } // mayLoad = 1, hasSideEffects = 0, hasExtraDefRegAllocReq = 1
1363 // VLD1DUP : Vector Load (single element to all lanes)
1364 class VLD1DUP<bits<4> op7_4, string Dt, ValueType Ty, PatFrag LoadOp,
1366 : NLdSt<1, 0b10, 0b1100, op7_4, (outs VecListOneDAllLanes:$Vd),
1368 IIC_VLD1dup, "vld1", Dt, "$Vd, $Rn", "",
1369 [(set VecListOneDAllLanes:$Vd,
1370 (Ty (ARMvdup (i32 (LoadOp AddrMode:$Rn)))))]>,
1371 Sched<[WriteVLD2]> {
1373 let Inst{4} = Rn{4};
1374 let DecoderMethod = "DecodeVLD1DupInstruction";
1376 def VLD1DUPd8 : VLD1DUP<{0,0,0,?}, "8", v8i8, extloadi8,
1377 addrmode6dupalignNone>;
1378 def VLD1DUPd16 : VLD1DUP<{0,1,0,?}, "16", v4i16, extloadi16,
1379 addrmode6dupalign16>;
1380 def VLD1DUPd32 : VLD1DUP<{1,0,0,?}, "32", v2i32, load,
1381 addrmode6dupalign32>;
1383 let Predicates = [HasNEON] in {
1384 def : Pat<(v2f32 (ARMvdup (f32 (load addrmode6dup:$addr)))),
1385 (VLD1DUPd32 addrmode6:$addr)>;
1388 class VLD1QDUP<bits<4> op7_4, string Dt, ValueType Ty, PatFrag LoadOp,
1390 : NLdSt<1, 0b10, 0b1100, op7_4, (outs VecListDPairAllLanes:$Vd),
1391 (ins AddrMode:$Rn), IIC_VLD1dup,
1392 "vld1", Dt, "$Vd, $Rn", "",
1393 [(set VecListDPairAllLanes:$Vd,
1394 (Ty (ARMvdup (i32 (LoadOp AddrMode:$Rn)))))]> {
1396 let Inst{4} = Rn{4};
1397 let DecoderMethod = "DecodeVLD1DupInstruction";
1400 def VLD1DUPq8 : VLD1QDUP<{0,0,1,0}, "8", v16i8, extloadi8,
1401 addrmode6dupalignNone>;
1402 def VLD1DUPq16 : VLD1QDUP<{0,1,1,?}, "16", v8i16, extloadi16,
1403 addrmode6dupalign16>;
1404 def VLD1DUPq32 : VLD1QDUP<{1,0,1,?}, "32", v4i32, load,
1405 addrmode6dupalign32>;
1407 let Predicates = [HasNEON] in {
1408 def : Pat<(v4f32 (ARMvdup (f32 (load addrmode6dup:$addr)))),
1409 (VLD1DUPq32 addrmode6:$addr)>;
1412 let mayLoad = 1, hasSideEffects = 0, hasExtraDefRegAllocReq = 1 in {
1413 // ...with address register writeback:
1414 multiclass VLD1DUPWB<bits<4> op7_4, string Dt, Operand AddrMode> {
1415 def _fixed : NLdSt<1, 0b10, 0b1100, op7_4,
1416 (outs VecListOneDAllLanes:$Vd, GPR:$wb),
1417 (ins AddrMode:$Rn), IIC_VLD1dupu,
1418 "vld1", Dt, "$Vd, $Rn!",
1419 "$Rn.addr = $wb", []> {
1420 let Rm = 0b1101; // NLdSt will assign to the right encoding bits.
1421 let Inst{4} = Rn{4};
1422 let DecoderMethod = "DecodeVLD1DupInstruction";
1424 def _register : NLdSt<1, 0b10, 0b1100, op7_4,
1425 (outs VecListOneDAllLanes:$Vd, GPR:$wb),
1426 (ins AddrMode:$Rn, rGPR:$Rm), IIC_VLD1dupu,
1427 "vld1", Dt, "$Vd, $Rn, $Rm",
1428 "$Rn.addr = $wb", []> {
1429 let Inst{4} = Rn{4};
1430 let DecoderMethod = "DecodeVLD1DupInstruction";
1433 multiclass VLD1QDUPWB<bits<4> op7_4, string Dt, Operand AddrMode> {
1434 def _fixed : NLdSt<1, 0b10, 0b1100, op7_4,
1435 (outs VecListDPairAllLanes:$Vd, GPR:$wb),
1436 (ins AddrMode:$Rn), IIC_VLD1dupu,
1437 "vld1", Dt, "$Vd, $Rn!",
1438 "$Rn.addr = $wb", []>, Sched<[WriteVLD1]> {
1439 let Rm = 0b1101; // NLdSt will assign to the right encoding bits.
1440 let Inst{4} = Rn{4};
1441 let DecoderMethod = "DecodeVLD1DupInstruction";
1443 def _register : NLdSt<1, 0b10, 0b1100, op7_4,
1444 (outs VecListDPairAllLanes:$Vd, GPR:$wb),
1445 (ins AddrMode:$Rn, rGPR:$Rm), IIC_VLD1dupu,
1446 "vld1", Dt, "$Vd, $Rn, $Rm",
1447 "$Rn.addr = $wb", []> {
1448 let Inst{4} = Rn{4};
1449 let DecoderMethod = "DecodeVLD1DupInstruction";
1453 defm VLD1DUPd8wb : VLD1DUPWB<{0,0,0,0}, "8", addrmode6dupalignNone>;
1454 defm VLD1DUPd16wb : VLD1DUPWB<{0,1,0,?}, "16", addrmode6dupalign16>;
1455 defm VLD1DUPd32wb : VLD1DUPWB<{1,0,0,?}, "32", addrmode6dupalign32>;
1457 defm VLD1DUPq8wb : VLD1QDUPWB<{0,0,1,0}, "8", addrmode6dupalignNone>;
1458 defm VLD1DUPq16wb : VLD1QDUPWB<{0,1,1,?}, "16", addrmode6dupalign16>;
1459 defm VLD1DUPq32wb : VLD1QDUPWB<{1,0,1,?}, "32", addrmode6dupalign32>;
1461 // VLD2DUP : Vector Load (single 2-element structure to all lanes)
1462 class VLD2DUP<bits<4> op7_4, string Dt, RegisterOperand VdTy, Operand AddrMode>
1463 : NLdSt<1, 0b10, 0b1101, op7_4, (outs VdTy:$Vd),
1464 (ins AddrMode:$Rn), IIC_VLD2dup,
1465 "vld2", Dt, "$Vd, $Rn", "", []> {
1467 let Inst{4} = Rn{4};
1468 let DecoderMethod = "DecodeVLD2DupInstruction";
1471 def VLD2DUPd8 : VLD2DUP<{0,0,0,?}, "8", VecListDPairAllLanes,
1472 addrmode6dupalign16>;
1473 def VLD2DUPd16 : VLD2DUP<{0,1,0,?}, "16", VecListDPairAllLanes,
1474 addrmode6dupalign32>;
1475 def VLD2DUPd32 : VLD2DUP<{1,0,0,?}, "32", VecListDPairAllLanes,
1476 addrmode6dupalign64>;
1478 // HACK this one, VLD2DUPd8x2 must be changed at the same time with VLD2b8 or
1479 // "vld2.8 {d0[], d2[]}, [r4:32]" will become "vld2.8 {d0, d2}, [r4:32]".
1480 // ...with double-spaced registers
1481 def VLD2DUPd8x2 : VLD2DUP<{0,0,1,?}, "8", VecListDPairSpacedAllLanes,
1482 addrmode6dupalign16>;
1483 def VLD2DUPd16x2 : VLD2DUP<{0,1,1,?}, "16", VecListDPairSpacedAllLanes,
1484 addrmode6dupalign32>;
1485 def VLD2DUPd32x2 : VLD2DUP<{1,0,1,?}, "32", VecListDPairSpacedAllLanes,
1486 addrmode6dupalign64>;
1488 def VLD2DUPq8EvenPseudo : VLDQQPseudo<IIC_VLD2dup>, Sched<[WriteVLD2]>;
1489 def VLD2DUPq8OddPseudo : VLDQQPseudo<IIC_VLD2dup>, Sched<[WriteVLD2]>;
1490 def VLD2DUPq16EvenPseudo : VLDQQPseudo<IIC_VLD2dup>, Sched<[WriteVLD2]>;
1491 def VLD2DUPq16OddPseudo : VLDQQPseudo<IIC_VLD2dup>, Sched<[WriteVLD2]>;
1492 def VLD2DUPq32EvenPseudo : VLDQQPseudo<IIC_VLD2dup>, Sched<[WriteVLD2]>;
1493 def VLD2DUPq32OddPseudo : VLDQQPseudo<IIC_VLD2dup>, Sched<[WriteVLD2]>;
1495 // ...with address register writeback:
1496 multiclass VLD2DUPWB<bits<4> op7_4, string Dt, RegisterOperand VdTy,
1498 def _fixed : NLdSt<1, 0b10, 0b1101, op7_4,
1499 (outs VdTy:$Vd, GPR:$wb),
1500 (ins AddrMode:$Rn), IIC_VLD2dupu,
1501 "vld2", Dt, "$Vd, $Rn!",
1502 "$Rn.addr = $wb", []>, Sched<[WriteVLD1]> {
1503 let Rm = 0b1101; // NLdSt will assign to the right encoding bits.
1504 let Inst{4} = Rn{4};
1505 let DecoderMethod = "DecodeVLD2DupInstruction";
1507 def _register : NLdSt<1, 0b10, 0b1101, op7_4,
1508 (outs VdTy:$Vd, GPR:$wb),
1509 (ins AddrMode:$Rn, rGPR:$Rm), IIC_VLD2dupu,
1510 "vld2", Dt, "$Vd, $Rn, $Rm",
1511 "$Rn.addr = $wb", []>, Sched<[WriteVLD1]> {
1512 let Inst{4} = Rn{4};
1513 let DecoderMethod = "DecodeVLD2DupInstruction";
1517 defm VLD2DUPd8wb : VLD2DUPWB<{0,0,0,0}, "8", VecListDPairAllLanes,
1518 addrmode6dupalign16>;
1519 defm VLD2DUPd16wb : VLD2DUPWB<{0,1,0,?}, "16", VecListDPairAllLanes,
1520 addrmode6dupalign32>;
1521 defm VLD2DUPd32wb : VLD2DUPWB<{1,0,0,?}, "32", VecListDPairAllLanes,
1522 addrmode6dupalign64>;
1524 defm VLD2DUPd8x2wb : VLD2DUPWB<{0,0,1,0}, "8", VecListDPairSpacedAllLanes,
1525 addrmode6dupalign16>;
1526 defm VLD2DUPd16x2wb : VLD2DUPWB<{0,1,1,?}, "16", VecListDPairSpacedAllLanes,
1527 addrmode6dupalign32>;
1528 defm VLD2DUPd32x2wb : VLD2DUPWB<{1,0,1,?}, "32", VecListDPairSpacedAllLanes,
1529 addrmode6dupalign64>;
1531 // VLD3DUP : Vector Load (single 3-element structure to all lanes)
1532 class VLD3DUP<bits<4> op7_4, string Dt>
1533 : NLdSt<1, 0b10, 0b1110, op7_4, (outs DPR:$Vd, DPR:$dst2, DPR:$dst3),
1534 (ins addrmode6dup:$Rn), IIC_VLD3dup,
1535 "vld3", Dt, "\\{$Vd[], $dst2[], $dst3[]\\}, $Rn", "", []>,
1536 Sched<[WriteVLD2]> {
1539 let DecoderMethod = "DecodeVLD3DupInstruction";
1542 def VLD3DUPd8 : VLD3DUP<{0,0,0,?}, "8">;
1543 def VLD3DUPd16 : VLD3DUP<{0,1,0,?}, "16">;
1544 def VLD3DUPd32 : VLD3DUP<{1,0,0,?}, "32">;
1546 def VLD3DUPd8Pseudo : VLDQQPseudo<IIC_VLD3dup>, Sched<[WriteVLD2]>;
1547 def VLD3DUPd16Pseudo : VLDQQPseudo<IIC_VLD3dup>, Sched<[WriteVLD2]>;
1548 def VLD3DUPd32Pseudo : VLDQQPseudo<IIC_VLD3dup>, Sched<[WriteVLD2]>;
1550 // ...with double-spaced registers (not used for codegen):
1551 def VLD3DUPq8 : VLD3DUP<{0,0,1,?}, "8">;
1552 def VLD3DUPq16 : VLD3DUP<{0,1,1,?}, "16">;
1553 def VLD3DUPq32 : VLD3DUP<{1,0,1,?}, "32">;
1555 def VLD3DUPq8EvenPseudo : VLDQQQQPseudo<IIC_VLD3dup>, Sched<[WriteVLD2]>;
1556 def VLD3DUPq8OddPseudo : VLDQQQQPseudo<IIC_VLD3dup>, Sched<[WriteVLD2]>;
1557 def VLD3DUPq16EvenPseudo : VLDQQQQPseudo<IIC_VLD3dup>, Sched<[WriteVLD2]>;
1558 def VLD3DUPq16OddPseudo : VLDQQQQPseudo<IIC_VLD3dup>, Sched<[WriteVLD2]>;
1559 def VLD3DUPq32EvenPseudo : VLDQQQQPseudo<IIC_VLD3dup>, Sched<[WriteVLD2]>;
1560 def VLD3DUPq32OddPseudo : VLDQQQQPseudo<IIC_VLD3dup>, Sched<[WriteVLD2]>;
1562 // ...with address register writeback:
1563 class VLD3DUPWB<bits<4> op7_4, string Dt, Operand AddrMode>
1564 : NLdSt<1, 0b10, 0b1110, op7_4, (outs DPR:$Vd, DPR:$dst2, DPR:$dst3, GPR:$wb),
1565 (ins AddrMode:$Rn, am6offset:$Rm), IIC_VLD3dupu,
1566 "vld3", Dt, "\\{$Vd[], $dst2[], $dst3[]\\}, $Rn$Rm",
1567 "$Rn.addr = $wb", []>, Sched<[WriteVLD2]> {
1569 let DecoderMethod = "DecodeVLD3DupInstruction";
1572 def VLD3DUPd8_UPD : VLD3DUPWB<{0,0,0,0}, "8", addrmode6dupalign64>;
1573 def VLD3DUPd16_UPD : VLD3DUPWB<{0,1,0,?}, "16", addrmode6dupalign64>;
1574 def VLD3DUPd32_UPD : VLD3DUPWB<{1,0,0,?}, "32", addrmode6dupalign64>;
1576 def VLD3DUPq8_UPD : VLD3DUPWB<{0,0,1,0}, "8", addrmode6dupalign64>;
1577 def VLD3DUPq16_UPD : VLD3DUPWB<{0,1,1,?}, "16", addrmode6dupalign64>;
1578 def VLD3DUPq32_UPD : VLD3DUPWB<{1,0,1,?}, "32", addrmode6dupalign64>;
1580 def VLD3DUPd8Pseudo_UPD : VLDQQWBPseudo<IIC_VLD3dupu>, Sched<[WriteVLD2]>;
1581 def VLD3DUPd16Pseudo_UPD : VLDQQWBPseudo<IIC_VLD3dupu>, Sched<[WriteVLD2]>;
1582 def VLD3DUPd32Pseudo_UPD : VLDQQWBPseudo<IIC_VLD3dupu>, Sched<[WriteVLD2]>;
1584 // VLD4DUP : Vector Load (single 4-element structure to all lanes)
1585 class VLD4DUP<bits<4> op7_4, string Dt>
1586 : NLdSt<1, 0b10, 0b1111, op7_4,
1587 (outs DPR:$Vd, DPR:$dst2, DPR:$dst3, DPR:$dst4),
1588 (ins addrmode6dup:$Rn), IIC_VLD4dup,
1589 "vld4", Dt, "\\{$Vd[], $dst2[], $dst3[], $dst4[]\\}, $Rn", "", []> {
1591 let Inst{4} = Rn{4};
1592 let DecoderMethod = "DecodeVLD4DupInstruction";
1595 def VLD4DUPd8 : VLD4DUP<{0,0,0,?}, "8">;
1596 def VLD4DUPd16 : VLD4DUP<{0,1,0,?}, "16">;
1597 def VLD4DUPd32 : VLD4DUP<{1,?,0,?}, "32"> { let Inst{6} = Rn{5}; }
1599 def VLD4DUPd8Pseudo : VLDQQPseudo<IIC_VLD4dup>, Sched<[WriteVLD2]>;
1600 def VLD4DUPd16Pseudo : VLDQQPseudo<IIC_VLD4dup>, Sched<[WriteVLD2]>;
1601 def VLD4DUPd32Pseudo : VLDQQPseudo<IIC_VLD4dup>, Sched<[WriteVLD2]>;
1603 // ...with double-spaced registers (not used for codegen):
1604 def VLD4DUPq8 : VLD4DUP<{0,0,1,?}, "8">;
1605 def VLD4DUPq16 : VLD4DUP<{0,1,1,?}, "16">;
1606 def VLD4DUPq32 : VLD4DUP<{1,?,1,?}, "32"> { let Inst{6} = Rn{5}; }
1608 def VLD4DUPq8EvenPseudo : VLDQQQQPseudo<IIC_VLD4dup>, Sched<[WriteVLD2]>;
1609 def VLD4DUPq8OddPseudo : VLDQQQQPseudo<IIC_VLD4dup>, Sched<[WriteVLD2]>;
1610 def VLD4DUPq16EvenPseudo : VLDQQQQPseudo<IIC_VLD4dup>, Sched<[WriteVLD2]>;
1611 def VLD4DUPq16OddPseudo : VLDQQQQPseudo<IIC_VLD4dup>, Sched<[WriteVLD2]>;
1612 def VLD4DUPq32EvenPseudo : VLDQQQQPseudo<IIC_VLD4dup>, Sched<[WriteVLD2]>;
1613 def VLD4DUPq32OddPseudo : VLDQQQQPseudo<IIC_VLD4dup>, Sched<[WriteVLD2]>;
1615 // ...with address register writeback:
1616 class VLD4DUPWB<bits<4> op7_4, string Dt>
1617 : NLdSt<1, 0b10, 0b1111, op7_4,
1618 (outs DPR:$Vd, DPR:$dst2, DPR:$dst3, DPR:$dst4, GPR:$wb),
1619 (ins addrmode6dup:$Rn, am6offset:$Rm), IIC_VLD4dupu,
1620 "vld4", Dt, "\\{$Vd[], $dst2[], $dst3[], $dst4[]\\}, $Rn$Rm",
1621 "$Rn.addr = $wb", []>, Sched<[WriteVLD2]> {
1622 let Inst{4} = Rn{4};
1623 let DecoderMethod = "DecodeVLD4DupInstruction";
1626 def VLD4DUPd8_UPD : VLD4DUPWB<{0,0,0,0}, "8">;
1627 def VLD4DUPd16_UPD : VLD4DUPWB<{0,1,0,?}, "16">;
1628 def VLD4DUPd32_UPD : VLD4DUPWB<{1,?,0,?}, "32"> { let Inst{6} = Rn{5}; }
1630 def VLD4DUPq8_UPD : VLD4DUPWB<{0,0,1,0}, "8">;
1631 def VLD4DUPq16_UPD : VLD4DUPWB<{0,1,1,?}, "16">;
1632 def VLD4DUPq32_UPD : VLD4DUPWB<{1,?,1,?}, "32"> { let Inst{6} = Rn{5}; }
1634 def VLD4DUPd8Pseudo_UPD : VLDQQWBPseudo<IIC_VLD4dupu>, Sched<[WriteVLD2]>;
1635 def VLD4DUPd16Pseudo_UPD : VLDQQWBPseudo<IIC_VLD4dupu>, Sched<[WriteVLD2]>;
1636 def VLD4DUPd32Pseudo_UPD : VLDQQWBPseudo<IIC_VLD4dupu>, Sched<[WriteVLD2]>;
1638 } // mayLoad = 1, hasSideEffects = 0, hasExtraDefRegAllocReq = 1
1640 let mayStore = 1, hasSideEffects = 0, hasExtraSrcRegAllocReq = 1 in {
1642 // Classes for VST* pseudo-instructions with multi-register operands.
1643 // These are expanded to real instructions after register allocation.
1644 class VSTQPseudo<InstrItinClass itin>
1645 : PseudoNLdSt<(outs), (ins addrmode6:$addr, QPR:$src), itin, "">;
1646 class VSTQWBPseudo<InstrItinClass itin>
1647 : PseudoNLdSt<(outs GPR:$wb),
1648 (ins addrmode6:$addr, am6offset:$offset, QPR:$src), itin,
1649 "$addr.addr = $wb">;
1650 class VSTQWBfixedPseudo<InstrItinClass itin>
1651 : PseudoNLdSt<(outs GPR:$wb),
1652 (ins addrmode6:$addr, QPR:$src), itin,
1653 "$addr.addr = $wb">;
1654 class VSTQWBregisterPseudo<InstrItinClass itin>
1655 : PseudoNLdSt<(outs GPR:$wb),
1656 (ins addrmode6:$addr, rGPR:$offset, QPR:$src), itin,
1657 "$addr.addr = $wb">;
1658 class VSTQQPseudo<InstrItinClass itin>
1659 : PseudoNLdSt<(outs), (ins addrmode6:$addr, QQPR:$src), itin, "">;
1660 class VSTQQWBPseudo<InstrItinClass itin>
1661 : PseudoNLdSt<(outs GPR:$wb),
1662 (ins addrmode6:$addr, am6offset:$offset, QQPR:$src), itin,
1663 "$addr.addr = $wb">;
1664 class VSTQQWBfixedPseudo<InstrItinClass itin>
1665 : PseudoNLdSt<(outs GPR:$wb),
1666 (ins addrmode6:$addr, QQPR:$src), itin,
1667 "$addr.addr = $wb">;
1668 class VSTQQWBregisterPseudo<InstrItinClass itin>
1669 : PseudoNLdSt<(outs GPR:$wb),
1670 (ins addrmode6:$addr, rGPR:$offset, QQPR:$src), itin,
1671 "$addr.addr = $wb">;
1673 class VSTQQQQPseudo<InstrItinClass itin>
1674 : PseudoNLdSt<(outs), (ins addrmode6:$addr, QQQQPR:$src), itin, "">;
1675 class VSTQQQQWBPseudo<InstrItinClass itin>
1676 : PseudoNLdSt<(outs GPR:$wb),
1677 (ins addrmode6:$addr, am6offset:$offset, QQQQPR:$src), itin,
1678 "$addr.addr = $wb">;
1680 // VST1 : Vector Store (multiple single elements)
1681 class VST1D<bits<4> op7_4, string Dt, Operand AddrMode>
1682 : NLdSt<0,0b00,0b0111,op7_4, (outs), (ins AddrMode:$Rn, VecListOneD:$Vd),
1683 IIC_VST1, "vst1", Dt, "$Vd, $Rn", "", []>, Sched<[WriteVST1]> {
1685 let Inst{4} = Rn{4};
1686 let DecoderMethod = "DecodeVLDST1Instruction";
1688 class VST1Q<bits<4> op7_4, string Dt, Operand AddrMode>
1689 : NLdSt<0,0b00,0b1010,op7_4, (outs), (ins AddrMode:$Rn, VecListDPair:$Vd),
1690 IIC_VST1x2, "vst1", Dt, "$Vd, $Rn", "", []>, Sched<[WriteVST2]> {
1692 let Inst{5-4} = Rn{5-4};
1693 let DecoderMethod = "DecodeVLDST1Instruction";
1696 def VST1d8 : VST1D<{0,0,0,?}, "8", addrmode6align64>;
1697 def VST1d16 : VST1D<{0,1,0,?}, "16", addrmode6align64>;
1698 def VST1d32 : VST1D<{1,0,0,?}, "32", addrmode6align64>;
1699 def VST1d64 : VST1D<{1,1,0,?}, "64", addrmode6align64>;
1701 def VST1q8 : VST1Q<{0,0,?,?}, "8", addrmode6align64or128>;
1702 def VST1q16 : VST1Q<{0,1,?,?}, "16", addrmode6align64or128>;
1703 def VST1q32 : VST1Q<{1,0,?,?}, "32", addrmode6align64or128>;
1704 def VST1q64 : VST1Q<{1,1,?,?}, "64", addrmode6align64or128>;
1706 // ...with address register writeback:
1707 multiclass VST1DWB<bits<4> op7_4, string Dt, Operand AddrMode> {
1708 def _fixed : NLdSt<0,0b00, 0b0111,op7_4, (outs GPR:$wb),
1709 (ins AddrMode:$Rn, VecListOneD:$Vd), IIC_VLD1u,
1710 "vst1", Dt, "$Vd, $Rn!",
1711 "$Rn.addr = $wb", []>, Sched<[WriteVST1]> {
1712 let Rm = 0b1101; // NLdSt will assign to the right encoding bits.
1713 let Inst{4} = Rn{4};
1714 let DecoderMethod = "DecodeVLDST1Instruction";
1716 def _register : NLdSt<0,0b00,0b0111,op7_4, (outs GPR:$wb),
1717 (ins AddrMode:$Rn, rGPR:$Rm, VecListOneD:$Vd),
1719 "vst1", Dt, "$Vd, $Rn, $Rm",
1720 "$Rn.addr = $wb", []>, Sched<[WriteVST1]> {
1721 let Inst{4} = Rn{4};
1722 let DecoderMethod = "DecodeVLDST1Instruction";
1725 multiclass VST1QWB<bits<4> op7_4, string Dt, Operand AddrMode> {
1726 def _fixed : NLdSt<0,0b00,0b1010,op7_4, (outs GPR:$wb),
1727 (ins AddrMode:$Rn, VecListDPair:$Vd), IIC_VLD1x2u,
1728 "vst1", Dt, "$Vd, $Rn!",
1729 "$Rn.addr = $wb", []>, Sched<[WriteVST2]> {
1730 let Rm = 0b1101; // NLdSt will assign to the right encoding bits.
1731 let Inst{5-4} = Rn{5-4};
1732 let DecoderMethod = "DecodeVLDST1Instruction";
1734 def _register : NLdSt<0,0b00,0b1010,op7_4, (outs GPR:$wb),
1735 (ins AddrMode:$Rn, rGPR:$Rm, VecListDPair:$Vd),
1737 "vst1", Dt, "$Vd, $Rn, $Rm",
1738 "$Rn.addr = $wb", []>, Sched<[WriteVST2]> {
1739 let Inst{5-4} = Rn{5-4};
1740 let DecoderMethod = "DecodeVLDST1Instruction";
1744 defm VST1d8wb : VST1DWB<{0,0,0,?}, "8", addrmode6align64>;
1745 defm VST1d16wb : VST1DWB<{0,1,0,?}, "16", addrmode6align64>;
1746 defm VST1d32wb : VST1DWB<{1,0,0,?}, "32", addrmode6align64>;
1747 defm VST1d64wb : VST1DWB<{1,1,0,?}, "64", addrmode6align64>;
1749 defm VST1q8wb : VST1QWB<{0,0,?,?}, "8", addrmode6align64or128>;
1750 defm VST1q16wb : VST1QWB<{0,1,?,?}, "16", addrmode6align64or128>;
1751 defm VST1q32wb : VST1QWB<{1,0,?,?}, "32", addrmode6align64or128>;
1752 defm VST1q64wb : VST1QWB<{1,1,?,?}, "64", addrmode6align64or128>;
1754 // ...with 3 registers
1755 class VST1D3<bits<4> op7_4, string Dt, Operand AddrMode>
1756 : NLdSt<0, 0b00, 0b0110, op7_4, (outs),
1757 (ins AddrMode:$Rn, VecListThreeD:$Vd),
1758 IIC_VST1x3, "vst1", Dt, "$Vd, $Rn", "", []>, Sched<[WriteVST3]> {
1760 let Inst{4} = Rn{4};
1761 let DecoderMethod = "DecodeVLDST1Instruction";
1763 multiclass VST1D3WB<bits<4> op7_4, string Dt, Operand AddrMode> {
1764 def _fixed : NLdSt<0,0b00,0b0110,op7_4, (outs GPR:$wb),
1765 (ins AddrMode:$Rn, VecListThreeD:$Vd), IIC_VLD1x3u,
1766 "vst1", Dt, "$Vd, $Rn!",
1767 "$Rn.addr = $wb", []>, Sched<[WriteVST3]> {
1768 let Rm = 0b1101; // NLdSt will assign to the right encoding bits.
1769 let Inst{5-4} = Rn{5-4};
1770 let DecoderMethod = "DecodeVLDST1Instruction";
1772 def _register : NLdSt<0,0b00,0b0110,op7_4, (outs GPR:$wb),
1773 (ins AddrMode:$Rn, rGPR:$Rm, VecListThreeD:$Vd),
1775 "vst1", Dt, "$Vd, $Rn, $Rm",
1776 "$Rn.addr = $wb", []>, Sched<[WriteVST3]> {
1777 let Inst{5-4} = Rn{5-4};
1778 let DecoderMethod = "DecodeVLDST1Instruction";
1782 def VST1d8T : VST1D3<{0,0,0,?}, "8", addrmode6align64>;
1783 def VST1d16T : VST1D3<{0,1,0,?}, "16", addrmode6align64>;
1784 def VST1d32T : VST1D3<{1,0,0,?}, "32", addrmode6align64>;
1785 def VST1d64T : VST1D3<{1,1,0,?}, "64", addrmode6align64>;
1787 defm VST1d8Twb : VST1D3WB<{0,0,0,?}, "8", addrmode6align64>;
1788 defm VST1d16Twb : VST1D3WB<{0,1,0,?}, "16", addrmode6align64>;
1789 defm VST1d32Twb : VST1D3WB<{1,0,0,?}, "32", addrmode6align64>;
1790 defm VST1d64Twb : VST1D3WB<{1,1,0,?}, "64", addrmode6align64>;
1792 def VST1d8TPseudo : VSTQQPseudo<IIC_VST1x3>, Sched<[WriteVST3]>;
1793 def VST1d16TPseudo : VSTQQPseudo<IIC_VST1x3>, Sched<[WriteVST3]>;
1794 def VST1d32TPseudo : VSTQQPseudo<IIC_VST1x3>, Sched<[WriteVST3]>;
1795 def VST1d64TPseudo : VSTQQPseudo<IIC_VST1x3>, Sched<[WriteVST3]>;
1796 def VST1d64TPseudoWB_fixed : VSTQQWBfixedPseudo<IIC_VST1x3u>, Sched<[WriteVST3]>;
1797 def VST1d64TPseudoWB_register : VSTQQWBPseudo<IIC_VST1x3u>, Sched<[WriteVST3]>;
1799 def VST1q8HighTPseudo : VSTQQQQPseudo<IIC_VST1x3>, Sched<[WriteVST3]>;
1800 def VST1q8LowTPseudo_UPD : VSTQQQQWBPseudo<IIC_VST1x3>, Sched<[WriteVST3]>;
1801 def VST1q16HighTPseudo : VSTQQQQPseudo<IIC_VST1x3>, Sched<[WriteVST3]>;
1802 def VST1q16LowTPseudo_UPD : VSTQQQQWBPseudo<IIC_VST1x3>, Sched<[WriteVST3]>;
1803 def VST1q32HighTPseudo : VSTQQQQPseudo<IIC_VST1x3>, Sched<[WriteVST3]>;
1804 def VST1q32LowTPseudo_UPD : VSTQQQQWBPseudo<IIC_VST1x3>, Sched<[WriteVST3]>;
1805 def VST1q64HighTPseudo : VSTQQQQPseudo<IIC_VST1x3>, Sched<[WriteVST3]>;
1806 def VST1q64LowTPseudo_UPD : VSTQQQQWBPseudo<IIC_VST1x3>, Sched<[WriteVST3]>;
1808 // ...with 4 registers
1809 class VST1D4<bits<4> op7_4, string Dt, Operand AddrMode>
1810 : NLdSt<0, 0b00, 0b0010, op7_4, (outs),
1811 (ins AddrMode:$Rn, VecListFourD:$Vd),
1812 IIC_VST1x4, "vst1", Dt, "$Vd, $Rn", "",
1813 []>, Sched<[WriteVST4]> {
1815 let Inst{5-4} = Rn{5-4};
1816 let DecoderMethod = "DecodeVLDST1Instruction";
1818 multiclass VST1D4WB<bits<4> op7_4, string Dt, Operand AddrMode> {
1819 def _fixed : NLdSt<0,0b00,0b0010,op7_4, (outs GPR:$wb),
1820 (ins AddrMode:$Rn, VecListFourD:$Vd), IIC_VLD1x4u,
1821 "vst1", Dt, "$Vd, $Rn!",
1822 "$Rn.addr = $wb", []>, Sched<[WriteVST4]> {
1823 let Rm = 0b1101; // NLdSt will assign to the right encoding bits.
1824 let Inst{5-4} = Rn{5-4};
1825 let DecoderMethod = "DecodeVLDST1Instruction";
1827 def _register : NLdSt<0,0b00,0b0010,op7_4, (outs GPR:$wb),
1828 (ins AddrMode:$Rn, rGPR:$Rm, VecListFourD:$Vd),
1830 "vst1", Dt, "$Vd, $Rn, $Rm",
1831 "$Rn.addr = $wb", []>, Sched<[WriteVST4]> {
1832 let Inst{5-4} = Rn{5-4};
1833 let DecoderMethod = "DecodeVLDST1Instruction";
1837 def VST1d8Q : VST1D4<{0,0,?,?}, "8", addrmode6align64or128or256>;
1838 def VST1d16Q : VST1D4<{0,1,?,?}, "16", addrmode6align64or128or256>;
1839 def VST1d32Q : VST1D4<{1,0,?,?}, "32", addrmode6align64or128or256>;
1840 def VST1d64Q : VST1D4<{1,1,?,?}, "64", addrmode6align64or128or256>;
1842 defm VST1d8Qwb : VST1D4WB<{0,0,?,?}, "8", addrmode6align64or128or256>;
1843 defm VST1d16Qwb : VST1D4WB<{0,1,?,?}, "16", addrmode6align64or128or256>;
1844 defm VST1d32Qwb : VST1D4WB<{1,0,?,?}, "32", addrmode6align64or128or256>;
1845 defm VST1d64Qwb : VST1D4WB<{1,1,?,?}, "64", addrmode6align64or128or256>;
1847 def VST1d8QPseudo : VSTQQPseudo<IIC_VST1x4>, Sched<[WriteVST4]>;
1848 def VST1d16QPseudo : VSTQQPseudo<IIC_VST1x4>, Sched<[WriteVST4]>;
1849 def VST1d32QPseudo : VSTQQPseudo<IIC_VST1x4>, Sched<[WriteVST4]>;
1850 def VST1d64QPseudo : VSTQQPseudo<IIC_VST1x4>, Sched<[WriteVST4]>;
1851 def VST1d64QPseudoWB_fixed : VSTQQWBfixedPseudo<IIC_VST1x4u>, Sched<[WriteVST4]>;
1852 def VST1d64QPseudoWB_register : VSTQQWBPseudo<IIC_VST1x4u>, Sched<[WriteVST4]>;
1854 def VST1q8HighQPseudo : VSTQQQQPseudo<IIC_VST1x4>, Sched<[WriteVST4]>;
1855 def VST1q8LowQPseudo_UPD : VSTQQQQWBPseudo<IIC_VST1x4>, Sched<[WriteVST4]>;
1856 def VST1q16HighQPseudo : VSTQQQQPseudo<IIC_VST1x4>, Sched<[WriteVST4]>;
1857 def VST1q16LowQPseudo_UPD : VSTQQQQWBPseudo<IIC_VST1x4>, Sched<[WriteVST4]>;
1858 def VST1q32HighQPseudo : VSTQQQQPseudo<IIC_VST1x4>, Sched<[WriteVST4]>;
1859 def VST1q32LowQPseudo_UPD : VSTQQQQWBPseudo<IIC_VST1x4>, Sched<[WriteVST4]>;
1860 def VST1q64HighQPseudo : VSTQQQQPseudo<IIC_VST1x4>, Sched<[WriteVST4]>;
1861 def VST1q64LowQPseudo_UPD : VSTQQQQWBPseudo<IIC_VST1x4>, Sched<[WriteVST4]>;
1863 // VST2 : Vector Store (multiple 2-element structures)
1864 class VST2<bits<4> op11_8, bits<4> op7_4, string Dt, RegisterOperand VdTy,
1865 InstrItinClass itin, Operand AddrMode>
1866 : NLdSt<0, 0b00, op11_8, op7_4, (outs), (ins AddrMode:$Rn, VdTy:$Vd),
1867 itin, "vst2", Dt, "$Vd, $Rn", "", []> {
1869 let Inst{5-4} = Rn{5-4};
1870 let DecoderMethod = "DecodeVLDST2Instruction";
1873 def VST2d8 : VST2<0b1000, {0,0,?,?}, "8", VecListDPair, IIC_VST2,
1874 addrmode6align64or128>, Sched<[WriteVST2]>;
1875 def VST2d16 : VST2<0b1000, {0,1,?,?}, "16", VecListDPair, IIC_VST2,
1876 addrmode6align64or128>, Sched<[WriteVST2]>;
1877 def VST2d32 : VST2<0b1000, {1,0,?,?}, "32", VecListDPair, IIC_VST2,
1878 addrmode6align64or128>, Sched<[WriteVST2]>;
1880 def VST2q8 : VST2<0b0011, {0,0,?,?}, "8", VecListFourD, IIC_VST2x2,
1881 addrmode6align64or128or256>, Sched<[WriteVST4]>;
1882 def VST2q16 : VST2<0b0011, {0,1,?,?}, "16", VecListFourD, IIC_VST2x2,
1883 addrmode6align64or128or256>, Sched<[WriteVST4]>;
1884 def VST2q32 : VST2<0b0011, {1,0,?,?}, "32", VecListFourD, IIC_VST2x2,
1885 addrmode6align64or128or256>, Sched<[WriteVST4]>;
1887 def VST2q8Pseudo : VSTQQPseudo<IIC_VST2x2>, Sched<[WriteVST4]>;
1888 def VST2q16Pseudo : VSTQQPseudo<IIC_VST2x2>, Sched<[WriteVST4]>;
1889 def VST2q32Pseudo : VSTQQPseudo<IIC_VST2x2>, Sched<[WriteVST4]>;
1891 // ...with address register writeback:
1892 multiclass VST2DWB<bits<4> op11_8, bits<4> op7_4, string Dt,
1893 RegisterOperand VdTy, Operand AddrMode> {
1894 def _fixed : NLdSt<0, 0b00, op11_8, op7_4, (outs GPR:$wb),
1895 (ins AddrMode:$Rn, VdTy:$Vd), IIC_VLD1u,
1896 "vst2", Dt, "$Vd, $Rn!",
1897 "$Rn.addr = $wb", []>, Sched<[WriteVST2]> {
1898 let Rm = 0b1101; // NLdSt will assign to the right encoding bits.
1899 let Inst{5-4} = Rn{5-4};
1900 let DecoderMethod = "DecodeVLDST2Instruction";
1902 def _register : NLdSt<0, 0b00, op11_8, op7_4, (outs GPR:$wb),
1903 (ins AddrMode:$Rn, rGPR:$Rm, VdTy:$Vd), IIC_VLD1u,
1904 "vst2", Dt, "$Vd, $Rn, $Rm",
1905 "$Rn.addr = $wb", []>, Sched<[WriteVST2]> {
1906 let Inst{5-4} = Rn{5-4};
1907 let DecoderMethod = "DecodeVLDST2Instruction";
1910 multiclass VST2QWB<bits<4> op7_4, string Dt, Operand AddrMode> {
1911 def _fixed : NLdSt<0, 0b00, 0b0011, op7_4, (outs GPR:$wb),
1912 (ins AddrMode:$Rn, VecListFourD:$Vd), IIC_VLD1u,
1913 "vst2", Dt, "$Vd, $Rn!",
1914 "$Rn.addr = $wb", []>, Sched<[WriteVST4]> {
1915 let Rm = 0b1101; // NLdSt will assign to the right encoding bits.
1916 let Inst{5-4} = Rn{5-4};
1917 let DecoderMethod = "DecodeVLDST2Instruction";
1919 def _register : NLdSt<0, 0b00, 0b0011, op7_4, (outs GPR:$wb),
1920 (ins AddrMode:$Rn, rGPR:$Rm, VecListFourD:$Vd),
1922 "vst2", Dt, "$Vd, $Rn, $Rm",
1923 "$Rn.addr = $wb", []>, Sched<[WriteVST4]> {
1924 let Inst{5-4} = Rn{5-4};
1925 let DecoderMethod = "DecodeVLDST2Instruction";
1929 defm VST2d8wb : VST2DWB<0b1000, {0,0,?,?}, "8", VecListDPair,
1930 addrmode6align64or128>;
1931 defm VST2d16wb : VST2DWB<0b1000, {0,1,?,?}, "16", VecListDPair,
1932 addrmode6align64or128>;
1933 defm VST2d32wb : VST2DWB<0b1000, {1,0,?,?}, "32", VecListDPair,
1934 addrmode6align64or128>;
1936 defm VST2q8wb : VST2QWB<{0,0,?,?}, "8", addrmode6align64or128or256>;
1937 defm VST2q16wb : VST2QWB<{0,1,?,?}, "16", addrmode6align64or128or256>;
1938 defm VST2q32wb : VST2QWB<{1,0,?,?}, "32", addrmode6align64or128or256>;
1940 def VST2q8PseudoWB_fixed : VSTQQWBfixedPseudo<IIC_VST2x2u>, Sched<[WriteVST4]>;
1941 def VST2q16PseudoWB_fixed : VSTQQWBfixedPseudo<IIC_VST2x2u>, Sched<[WriteVST4]>;
1942 def VST2q32PseudoWB_fixed : VSTQQWBfixedPseudo<IIC_VST2x2u>, Sched<[WriteVST4]>;
1943 def VST2q8PseudoWB_register : VSTQQWBregisterPseudo<IIC_VST2x2u>, Sched<[WriteVST4]>;
1944 def VST2q16PseudoWB_register : VSTQQWBregisterPseudo<IIC_VST2x2u>, Sched<[WriteVST4]>;
1945 def VST2q32PseudoWB_register : VSTQQWBregisterPseudo<IIC_VST2x2u>, Sched<[WriteVST4]>;
1947 // ...with double-spaced registers
1948 def VST2b8 : VST2<0b1001, {0,0,?,?}, "8", VecListDPairSpaced, IIC_VST2,
1949 addrmode6align64or128>;
1950 def VST2b16 : VST2<0b1001, {0,1,?,?}, "16", VecListDPairSpaced, IIC_VST2,
1951 addrmode6align64or128>;
1952 def VST2b32 : VST2<0b1001, {1,0,?,?}, "32", VecListDPairSpaced, IIC_VST2,
1953 addrmode6align64or128>;
1954 defm VST2b8wb : VST2DWB<0b1001, {0,0,?,?}, "8", VecListDPairSpaced,
1955 addrmode6align64or128>;
1956 defm VST2b16wb : VST2DWB<0b1001, {0,1,?,?}, "16", VecListDPairSpaced,
1957 addrmode6align64or128>;
1958 defm VST2b32wb : VST2DWB<0b1001, {1,0,?,?}, "32", VecListDPairSpaced,
1959 addrmode6align64or128>;
1961 // VST3 : Vector Store (multiple 3-element structures)
1962 class VST3D<bits<4> op11_8, bits<4> op7_4, string Dt>
1963 : NLdSt<0, 0b00, op11_8, op7_4, (outs),
1964 (ins addrmode6:$Rn, DPR:$Vd, DPR:$src2, DPR:$src3), IIC_VST3,
1965 "vst3", Dt, "\\{$Vd, $src2, $src3\\}, $Rn", "", []>, Sched<[WriteVST3]> {
1967 let Inst{4} = Rn{4};
1968 let DecoderMethod = "DecodeVLDST3Instruction";
1971 def VST3d8 : VST3D<0b0100, {0,0,0,?}, "8">;
1972 def VST3d16 : VST3D<0b0100, {0,1,0,?}, "16">;
1973 def VST3d32 : VST3D<0b0100, {1,0,0,?}, "32">;
1975 def VST3d8Pseudo : VSTQQPseudo<IIC_VST3>, Sched<[WriteVST3]>;
1976 def VST3d16Pseudo : VSTQQPseudo<IIC_VST3>, Sched<[WriteVST3]>;
1977 def VST3d32Pseudo : VSTQQPseudo<IIC_VST3>, Sched<[WriteVST3]>;
1979 // ...with address register writeback:
1980 class VST3DWB<bits<4> op11_8, bits<4> op7_4, string Dt>
1981 : NLdSt<0, 0b00, op11_8, op7_4, (outs GPR:$wb),
1982 (ins addrmode6:$Rn, am6offset:$Rm,
1983 DPR:$Vd, DPR:$src2, DPR:$src3), IIC_VST3u,
1984 "vst3", Dt, "\\{$Vd, $src2, $src3\\}, $Rn$Rm",
1985 "$Rn.addr = $wb", []>, Sched<[WriteVST3]> {
1986 let Inst{4} = Rn{4};
1987 let DecoderMethod = "DecodeVLDST3Instruction";
1990 def VST3d8_UPD : VST3DWB<0b0100, {0,0,0,?}, "8">;
1991 def VST3d16_UPD : VST3DWB<0b0100, {0,1,0,?}, "16">;
1992 def VST3d32_UPD : VST3DWB<0b0100, {1,0,0,?}, "32">;
1994 def VST3d8Pseudo_UPD : VSTQQWBPseudo<IIC_VST3u>, Sched<[WriteVST3]>;
1995 def VST3d16Pseudo_UPD : VSTQQWBPseudo<IIC_VST3u>, Sched<[WriteVST3]>;
1996 def VST3d32Pseudo_UPD : VSTQQWBPseudo<IIC_VST3u>, Sched<[WriteVST3]>;
1998 // ...with double-spaced registers:
1999 def VST3q8 : VST3D<0b0101, {0,0,0,?}, "8">;
2000 def VST3q16 : VST3D<0b0101, {0,1,0,?}, "16">;
2001 def VST3q32 : VST3D<0b0101, {1,0,0,?}, "32">;
2002 def VST3q8_UPD : VST3DWB<0b0101, {0,0,0,?}, "8">;
2003 def VST3q16_UPD : VST3DWB<0b0101, {0,1,0,?}, "16">;
2004 def VST3q32_UPD : VST3DWB<0b0101, {1,0,0,?}, "32">;
2006 def VST3q8Pseudo_UPD : VSTQQQQWBPseudo<IIC_VST3u>, Sched<[WriteVST3]>;
2007 def VST3q16Pseudo_UPD : VSTQQQQWBPseudo<IIC_VST3u>, Sched<[WriteVST3]>;
2008 def VST3q32Pseudo_UPD : VSTQQQQWBPseudo<IIC_VST3u>, Sched<[WriteVST3]>;
2010 // ...alternate versions to be allocated odd register numbers:
2011 def VST3q8oddPseudo : VSTQQQQPseudo<IIC_VST3>, Sched<[WriteVST3]>;
2012 def VST3q16oddPseudo : VSTQQQQPseudo<IIC_VST3>, Sched<[WriteVST3]>;
2013 def VST3q32oddPseudo : VSTQQQQPseudo<IIC_VST3>, Sched<[WriteVST3]>;
2015 def VST3q8oddPseudo_UPD : VSTQQQQWBPseudo<IIC_VST3u>, Sched<[WriteVST3]>;
2016 def VST3q16oddPseudo_UPD : VSTQQQQWBPseudo<IIC_VST3u>, Sched<[WriteVST3]>;
2017 def VST3q32oddPseudo_UPD : VSTQQQQWBPseudo<IIC_VST3u>, Sched<[WriteVST3]>;
2019 // VST4 : Vector Store (multiple 4-element structures)
2020 class VST4D<bits<4> op11_8, bits<4> op7_4, string Dt>
2021 : NLdSt<0, 0b00, op11_8, op7_4, (outs),
2022 (ins addrmode6:$Rn, DPR:$Vd, DPR:$src2, DPR:$src3, DPR:$src4),
2023 IIC_VST4, "vst4", Dt, "\\{$Vd, $src2, $src3, $src4\\}, $Rn",
2024 "", []>, Sched<[WriteVST4]> {
2026 let Inst{5-4} = Rn{5-4};
2027 let DecoderMethod = "DecodeVLDST4Instruction";
2030 def VST4d8 : VST4D<0b0000, {0,0,?,?}, "8">;
2031 def VST4d16 : VST4D<0b0000, {0,1,?,?}, "16">;
2032 def VST4d32 : VST4D<0b0000, {1,0,?,?}, "32">;
2034 def VST4d8Pseudo : VSTQQPseudo<IIC_VST4>, Sched<[WriteVST4]>;
2035 def VST4d16Pseudo : VSTQQPseudo<IIC_VST4>, Sched<[WriteVST4]>;
2036 def VST4d32Pseudo : VSTQQPseudo<IIC_VST4>, Sched<[WriteVST4]>;
2038 // ...with address register writeback:
2039 class VST4DWB<bits<4> op11_8, bits<4> op7_4, string Dt>
2040 : NLdSt<0, 0b00, op11_8, op7_4, (outs GPR:$wb),
2041 (ins addrmode6:$Rn, am6offset:$Rm,
2042 DPR:$Vd, DPR:$src2, DPR:$src3, DPR:$src4), IIC_VST4u,
2043 "vst4", Dt, "\\{$Vd, $src2, $src3, $src4\\}, $Rn$Rm",
2044 "$Rn.addr = $wb", []>, Sched<[WriteVST4]> {
2045 let Inst{5-4} = Rn{5-4};
2046 let DecoderMethod = "DecodeVLDST4Instruction";
2049 def VST4d8_UPD : VST4DWB<0b0000, {0,0,?,?}, "8">;
2050 def VST4d16_UPD : VST4DWB<0b0000, {0,1,?,?}, "16">;
2051 def VST4d32_UPD : VST4DWB<0b0000, {1,0,?,?}, "32">;
2053 def VST4d8Pseudo_UPD : VSTQQWBPseudo<IIC_VST4u>, Sched<[WriteVST4]>;
2054 def VST4d16Pseudo_UPD : VSTQQWBPseudo<IIC_VST4u>, Sched<[WriteVST4]>;
2055 def VST4d32Pseudo_UPD : VSTQQWBPseudo<IIC_VST4u>, Sched<[WriteVST4]>;
2057 // ...with double-spaced registers:
2058 def VST4q8 : VST4D<0b0001, {0,0,?,?}, "8">;
2059 def VST4q16 : VST4D<0b0001, {0,1,?,?}, "16">;
2060 def VST4q32 : VST4D<0b0001, {1,0,?,?}, "32">;
2061 def VST4q8_UPD : VST4DWB<0b0001, {0,0,?,?}, "8">;
2062 def VST4q16_UPD : VST4DWB<0b0001, {0,1,?,?}, "16">;
2063 def VST4q32_UPD : VST4DWB<0b0001, {1,0,?,?}, "32">;
2065 def VST4q8Pseudo_UPD : VSTQQQQWBPseudo<IIC_VST4u>, Sched<[WriteVST4]>;
2066 def VST4q16Pseudo_UPD : VSTQQQQWBPseudo<IIC_VST4u>, Sched<[WriteVST4]>;
2067 def VST4q32Pseudo_UPD : VSTQQQQWBPseudo<IIC_VST4u>, Sched<[WriteVST4]>;
2069 // ...alternate versions to be allocated odd register numbers:
2070 def VST4q8oddPseudo : VSTQQQQPseudo<IIC_VST4>, Sched<[WriteVST4]>;
2071 def VST4q16oddPseudo : VSTQQQQPseudo<IIC_VST4>, Sched<[WriteVST4]>;
2072 def VST4q32oddPseudo : VSTQQQQPseudo<IIC_VST4>, Sched<[WriteVST4]>;
2074 def VST4q8oddPseudo_UPD : VSTQQQQWBPseudo<IIC_VST4u>, Sched<[WriteVST4]>;
2075 def VST4q16oddPseudo_UPD : VSTQQQQWBPseudo<IIC_VST4u>, Sched<[WriteVST4]>;
2076 def VST4q32oddPseudo_UPD : VSTQQQQWBPseudo<IIC_VST4u>, Sched<[WriteVST4]>;
2078 } // mayStore = 1, hasSideEffects = 0, hasExtraSrcRegAllocReq = 1
2080 // Classes for VST*LN pseudo-instructions with multi-register operands.
2081 // These are expanded to real instructions after register allocation.
2082 class VSTQLNPseudo<InstrItinClass itin>
2083 : PseudoNLdSt<(outs), (ins addrmode6:$addr, QPR:$src, nohash_imm:$lane),
2085 class VSTQLNWBPseudo<InstrItinClass itin>
2086 : PseudoNLdSt<(outs GPR:$wb),
2087 (ins addrmode6:$addr, am6offset:$offset, QPR:$src,
2088 nohash_imm:$lane), itin, "$addr.addr = $wb">;
2089 class VSTQQLNPseudo<InstrItinClass itin>
2090 : PseudoNLdSt<(outs), (ins addrmode6:$addr, QQPR:$src, nohash_imm:$lane),
2092 class VSTQQLNWBPseudo<InstrItinClass itin>
2093 : PseudoNLdSt<(outs GPR:$wb),
2094 (ins addrmode6:$addr, am6offset:$offset, QQPR:$src,
2095 nohash_imm:$lane), itin, "$addr.addr = $wb">;
2096 class VSTQQQQLNPseudo<InstrItinClass itin>
2097 : PseudoNLdSt<(outs), (ins addrmode6:$addr, QQQQPR:$src, nohash_imm:$lane),
2099 class VSTQQQQLNWBPseudo<InstrItinClass itin>
2100 : PseudoNLdSt<(outs GPR:$wb),
2101 (ins addrmode6:$addr, am6offset:$offset, QQQQPR:$src,
2102 nohash_imm:$lane), itin, "$addr.addr = $wb">;
2104 // VST1LN : Vector Store (single element from one lane)
2105 class VST1LN<bits<4> op11_8, bits<4> op7_4, string Dt, ValueType Ty,
2106 PatFrag StoreOp, SDNode ExtractOp, Operand AddrMode>
2107 : NLdStLn<1, 0b00, op11_8, op7_4, (outs),
2108 (ins AddrMode:$Rn, DPR:$Vd, nohash_imm:$lane),
2109 IIC_VST1ln, "vst1", Dt, "\\{$Vd[$lane]\\}, $Rn", "",
2110 [(StoreOp (ExtractOp (Ty DPR:$Vd), imm:$lane), AddrMode:$Rn)]>,
2111 Sched<[WriteVST1]> {
2113 let DecoderMethod = "DecodeVST1LN";
2115 class VST1QLNPseudo<ValueType Ty, PatFrag StoreOp, SDNode ExtractOp>
2116 : VSTQLNPseudo<IIC_VST1ln>, Sched<[WriteVST1]> {
2117 let Pattern = [(StoreOp (ExtractOp (Ty QPR:$src), imm:$lane),
2121 def VST1LNd8 : VST1LN<0b0000, {?,?,?,0}, "8", v8i8, truncstorei8,
2122 ARMvgetlaneu, addrmode6> {
2123 let Inst{7-5} = lane{2-0};
2125 def VST1LNd16 : VST1LN<0b0100, {?,?,0,?}, "16", v4i16, truncstorei16,
2126 ARMvgetlaneu, addrmode6> {
2127 let Inst{7-6} = lane{1-0};
2128 let Inst{4} = Rn{4};
2131 def VST1LNd32 : VST1LN<0b1000, {?,0,?,?}, "32", v2i32, store, extractelt,
2133 let Inst{7} = lane{0};
2134 let Inst{5-4} = Rn{5-4};
2137 def VST1LNq8Pseudo : VST1QLNPseudo<v16i8, truncstorei8, ARMvgetlaneu>;
2138 def VST1LNq16Pseudo : VST1QLNPseudo<v8i16, truncstorei16, ARMvgetlaneu>;
2139 def VST1LNq32Pseudo : VST1QLNPseudo<v4i32, store, extractelt>;
2141 let Predicates = [HasNEON] in {
2142 def : Pat<(store (extractelt (v2f32 DPR:$src), imm:$lane), addrmode6:$addr),
2143 (VST1LNd32 addrmode6:$addr, DPR:$src, imm:$lane)>;
2144 def : Pat<(store (extractelt (v4f32 QPR:$src), imm:$lane), addrmode6:$addr),
2145 (VST1LNq32Pseudo addrmode6:$addr, QPR:$src, imm:$lane)>;
2147 def : Pat<(store (extractelt (v4f16 DPR:$src), imm:$lane), addrmode6:$addr),
2148 (VST1LNd16 addrmode6:$addr, DPR:$src, imm:$lane)>;
2149 def : Pat<(store (extractelt (v8f16 QPR:$src), imm:$lane), addrmode6:$addr),
2150 (VST1LNq16Pseudo addrmode6:$addr, QPR:$src, imm:$lane)>;
2153 // ...with address register writeback:
2154 class VST1LNWB<bits<4> op11_8, bits<4> op7_4, string Dt, ValueType Ty,
2155 PatFrag StoreOp, SDNode ExtractOp, Operand AdrMode>
2156 : NLdStLn<1, 0b00, op11_8, op7_4, (outs GPR:$wb),
2157 (ins AdrMode:$Rn, am6offset:$Rm,
2158 DPR:$Vd, nohash_imm:$lane), IIC_VST1lnu, "vst1", Dt,
2159 "\\{$Vd[$lane]\\}, $Rn$Rm",
2161 [(set GPR:$wb, (StoreOp (ExtractOp (Ty DPR:$Vd), imm:$lane),
2162 AdrMode:$Rn, am6offset:$Rm))]>,
2163 Sched<[WriteVST1]> {
2164 let DecoderMethod = "DecodeVST1LN";
2166 class VST1QLNWBPseudo<ValueType Ty, PatFrag StoreOp, SDNode ExtractOp>
2167 : VSTQLNWBPseudo<IIC_VST1lnu>, Sched<[WriteVST1]> {
2168 let Pattern = [(set GPR:$wb, (StoreOp (ExtractOp (Ty QPR:$src), imm:$lane),
2169 addrmode6:$addr, am6offset:$offset))];
2172 def VST1LNd8_UPD : VST1LNWB<0b0000, {?,?,?,0}, "8", v8i8, post_truncsti8,
2173 ARMvgetlaneu, addrmode6> {
2174 let Inst{7-5} = lane{2-0};
2176 def VST1LNd16_UPD : VST1LNWB<0b0100, {?,?,0,?}, "16", v4i16, post_truncsti16,
2177 ARMvgetlaneu, addrmode6> {
2178 let Inst{7-6} = lane{1-0};
2179 let Inst{4} = Rn{4};
2181 def VST1LNd32_UPD : VST1LNWB<0b1000, {?,0,?,?}, "32", v2i32, post_store,
2182 extractelt, addrmode6oneL32> {
2183 let Inst{7} = lane{0};
2184 let Inst{5-4} = Rn{5-4};
2187 def VST1LNq8Pseudo_UPD : VST1QLNWBPseudo<v16i8, post_truncsti8, ARMvgetlaneu>;
2188 def VST1LNq16Pseudo_UPD : VST1QLNWBPseudo<v8i16, post_truncsti16,ARMvgetlaneu>;
2189 def VST1LNq32Pseudo_UPD : VST1QLNWBPseudo<v4i32, post_store, extractelt>;
2191 let mayStore = 1, hasSideEffects = 0, hasExtraSrcRegAllocReq = 1 in {
2193 // VST2LN : Vector Store (single 2-element structure from one lane)
2194 class VST2LN<bits<4> op11_8, bits<4> op7_4, string Dt>
2195 : NLdStLn<1, 0b00, op11_8, op7_4, (outs),
2196 (ins addrmode6:$Rn, DPR:$Vd, DPR:$src2, nohash_imm:$lane),
2197 IIC_VST2ln, "vst2", Dt, "\\{$Vd[$lane], $src2[$lane]\\}, $Rn",
2198 "", []>, Sched<[WriteVST1]> {
2200 let Inst{4} = Rn{4};
2201 let DecoderMethod = "DecodeVST2LN";
2204 def VST2LNd8 : VST2LN<0b0001, {?,?,?,?}, "8"> {
2205 let Inst{7-5} = lane{2-0};
2207 def VST2LNd16 : VST2LN<0b0101, {?,?,0,?}, "16"> {
2208 let Inst{7-6} = lane{1-0};
2210 def VST2LNd32 : VST2LN<0b1001, {?,0,0,?}, "32"> {
2211 let Inst{7} = lane{0};
2214 def VST2LNd8Pseudo : VSTQLNPseudo<IIC_VST2ln>, Sched<[WriteVST1]>;
2215 def VST2LNd16Pseudo : VSTQLNPseudo<IIC_VST2ln>, Sched<[WriteVST1]>;
2216 def VST2LNd32Pseudo : VSTQLNPseudo<IIC_VST2ln>, Sched<[WriteVST1]>;
2218 // ...with double-spaced registers:
2219 def VST2LNq16 : VST2LN<0b0101, {?,?,1,?}, "16"> {
2220 let Inst{7-6} = lane{1-0};
2221 let Inst{4} = Rn{4};
2223 def VST2LNq32 : VST2LN<0b1001, {?,1,0,?}, "32"> {
2224 let Inst{7} = lane{0};
2225 let Inst{4} = Rn{4};
2228 def VST2LNq16Pseudo : VSTQQLNPseudo<IIC_VST2ln>, Sched<[WriteVST1]>;
2229 def VST2LNq32Pseudo : VSTQQLNPseudo<IIC_VST2ln>, Sched<[WriteVST1]>;
2231 // ...with address register writeback:
2232 class VST2LNWB<bits<4> op11_8, bits<4> op7_4, string Dt>
2233 : NLdStLn<1, 0b00, op11_8, op7_4, (outs GPR:$wb),
2234 (ins addrmode6:$Rn, am6offset:$Rm,
2235 DPR:$Vd, DPR:$src2, nohash_imm:$lane), IIC_VST2lnu, "vst2", Dt,
2236 "\\{$Vd[$lane], $src2[$lane]\\}, $Rn$Rm",
2237 "$Rn.addr = $wb", []> {
2238 let Inst{4} = Rn{4};
2239 let DecoderMethod = "DecodeVST2LN";
2242 def VST2LNd8_UPD : VST2LNWB<0b0001, {?,?,?,?}, "8"> {
2243 let Inst{7-5} = lane{2-0};
2245 def VST2LNd16_UPD : VST2LNWB<0b0101, {?,?,0,?}, "16"> {
2246 let Inst{7-6} = lane{1-0};
2248 def VST2LNd32_UPD : VST2LNWB<0b1001, {?,0,0,?}, "32"> {
2249 let Inst{7} = lane{0};
2252 def VST2LNd8Pseudo_UPD : VSTQLNWBPseudo<IIC_VST2lnu>, Sched<[WriteVST1]>;
2253 def VST2LNd16Pseudo_UPD : VSTQLNWBPseudo<IIC_VST2lnu>, Sched<[WriteVST1]>;
2254 def VST2LNd32Pseudo_UPD : VSTQLNWBPseudo<IIC_VST2lnu>, Sched<[WriteVST1]>;
2256 def VST2LNq16_UPD : VST2LNWB<0b0101, {?,?,1,?}, "16"> {
2257 let Inst{7-6} = lane{1-0};
2259 def VST2LNq32_UPD : VST2LNWB<0b1001, {?,1,0,?}, "32"> {
2260 let Inst{7} = lane{0};
2263 def VST2LNq16Pseudo_UPD : VSTQQLNWBPseudo<IIC_VST2lnu>, Sched<[WriteVST1]>;
2264 def VST2LNq32Pseudo_UPD : VSTQQLNWBPseudo<IIC_VST2lnu>, Sched<[WriteVST1]>;
2266 // VST3LN : Vector Store (single 3-element structure from one lane)
2267 class VST3LN<bits<4> op11_8, bits<4> op7_4, string Dt>
2268 : NLdStLn<1, 0b00, op11_8, op7_4, (outs),
2269 (ins addrmode6:$Rn, DPR:$Vd, DPR:$src2, DPR:$src3,
2270 nohash_imm:$lane), IIC_VST3ln, "vst3", Dt,
2271 "\\{$Vd[$lane], $src2[$lane], $src3[$lane]\\}, $Rn", "", []>,
2272 Sched<[WriteVST2]> {
2274 let DecoderMethod = "DecodeVST3LN";
2277 def VST3LNd8 : VST3LN<0b0010, {?,?,?,0}, "8"> {
2278 let Inst{7-5} = lane{2-0};
2280 def VST3LNd16 : VST3LN<0b0110, {?,?,0,0}, "16"> {
2281 let Inst{7-6} = lane{1-0};
2283 def VST3LNd32 : VST3LN<0b1010, {?,0,0,0}, "32"> {
2284 let Inst{7} = lane{0};
2287 def VST3LNd8Pseudo : VSTQQLNPseudo<IIC_VST3ln>, Sched<[WriteVST2]>;
2288 def VST3LNd16Pseudo : VSTQQLNPseudo<IIC_VST3ln>, Sched<[WriteVST2]>;
2289 def VST3LNd32Pseudo : VSTQQLNPseudo<IIC_VST3ln>, Sched<[WriteVST2]>;
2291 // ...with double-spaced registers:
2292 def VST3LNq16 : VST3LN<0b0110, {?,?,1,0}, "16"> {
2293 let Inst{7-6} = lane{1-0};
2295 def VST3LNq32 : VST3LN<0b1010, {?,1,0,0}, "32"> {
2296 let Inst{7} = lane{0};
2299 def VST3LNq16Pseudo : VSTQQQQLNPseudo<IIC_VST3ln>;
2300 def VST3LNq32Pseudo : VSTQQQQLNPseudo<IIC_VST3ln>;
2302 // ...with address register writeback:
2303 class VST3LNWB<bits<4> op11_8, bits<4> op7_4, string Dt>
2304 : NLdStLn<1, 0b00, op11_8, op7_4, (outs GPR:$wb),
2305 (ins addrmode6:$Rn, am6offset:$Rm,
2306 DPR:$Vd, DPR:$src2, DPR:$src3, nohash_imm:$lane),
2307 IIC_VST3lnu, "vst3", Dt,
2308 "\\{$Vd[$lane], $src2[$lane], $src3[$lane]\\}, $Rn$Rm",
2309 "$Rn.addr = $wb", []> {
2310 let DecoderMethod = "DecodeVST3LN";
2313 def VST3LNd8_UPD : VST3LNWB<0b0010, {?,?,?,0}, "8"> {
2314 let Inst{7-5} = lane{2-0};
2316 def VST3LNd16_UPD : VST3LNWB<0b0110, {?,?,0,0}, "16"> {
2317 let Inst{7-6} = lane{1-0};
2319 def VST3LNd32_UPD : VST3LNWB<0b1010, {?,0,0,0}, "32"> {
2320 let Inst{7} = lane{0};
2323 def VST3LNd8Pseudo_UPD : VSTQQLNWBPseudo<IIC_VST3lnu>, Sched<[WriteVST2]>;
2324 def VST3LNd16Pseudo_UPD : VSTQQLNWBPseudo<IIC_VST3lnu>, Sched<[WriteVST2]>;
2325 def VST3LNd32Pseudo_UPD : VSTQQLNWBPseudo<IIC_VST3lnu>, Sched<[WriteVST2]>;
2327 def VST3LNq16_UPD : VST3LNWB<0b0110, {?,?,1,0}, "16"> {
2328 let Inst{7-6} = lane{1-0};
2330 def VST3LNq32_UPD : VST3LNWB<0b1010, {?,1,0,0}, "32"> {
2331 let Inst{7} = lane{0};
2334 def VST3LNq16Pseudo_UPD : VSTQQQQLNWBPseudo<IIC_VST3lnu>, Sched<[WriteVST2]>;
2335 def VST3LNq32Pseudo_UPD : VSTQQQQLNWBPseudo<IIC_VST3lnu>, Sched<[WriteVST2]>;
2337 // VST4LN : Vector Store (single 4-element structure from one lane)
2338 class VST4LN<bits<4> op11_8, bits<4> op7_4, string Dt>
2339 : NLdStLn<1, 0b00, op11_8, op7_4, (outs),
2340 (ins addrmode6:$Rn, DPR:$Vd, DPR:$src2, DPR:$src3, DPR:$src4,
2341 nohash_imm:$lane), IIC_VST4ln, "vst4", Dt,
2342 "\\{$Vd[$lane], $src2[$lane], $src3[$lane], $src4[$lane]\\}, $Rn",
2343 "", []>, Sched<[WriteVST2]> {
2345 let Inst{4} = Rn{4};
2346 let DecoderMethod = "DecodeVST4LN";
2349 def VST4LNd8 : VST4LN<0b0011, {?,?,?,?}, "8"> {
2350 let Inst{7-5} = lane{2-0};
2352 def VST4LNd16 : VST4LN<0b0111, {?,?,0,?}, "16"> {
2353 let Inst{7-6} = lane{1-0};
2355 def VST4LNd32 : VST4LN<0b1011, {?,0,?,?}, "32"> {
2356 let Inst{7} = lane{0};
2357 let Inst{5} = Rn{5};
2360 def VST4LNd8Pseudo : VSTQQLNPseudo<IIC_VST4ln>, Sched<[WriteVST2]>;
2361 def VST4LNd16Pseudo : VSTQQLNPseudo<IIC_VST4ln>, Sched<[WriteVST2]>;
2362 def VST4LNd32Pseudo : VSTQQLNPseudo<IIC_VST4ln>, Sched<[WriteVST2]>;
2364 // ...with double-spaced registers:
2365 def VST4LNq16 : VST4LN<0b0111, {?,?,1,?}, "16"> {
2366 let Inst{7-6} = lane{1-0};
2368 def VST4LNq32 : VST4LN<0b1011, {?,1,?,?}, "32"> {
2369 let Inst{7} = lane{0};
2370 let Inst{5} = Rn{5};
2373 def VST4LNq16Pseudo : VSTQQQQLNPseudo<IIC_VST4ln>, Sched<[WriteVST2]>;
2374 def VST4LNq32Pseudo : VSTQQQQLNPseudo<IIC_VST4ln>, Sched<[WriteVST2]>;
2376 // ...with address register writeback:
2377 class VST4LNWB<bits<4> op11_8, bits<4> op7_4, string Dt>
2378 : NLdStLn<1, 0b00, op11_8, op7_4, (outs GPR:$wb),
2379 (ins addrmode6:$Rn, am6offset:$Rm,
2380 DPR:$Vd, DPR:$src2, DPR:$src3, DPR:$src4, nohash_imm:$lane),
2381 IIC_VST4lnu, "vst4", Dt,
2382 "\\{$Vd[$lane], $src2[$lane], $src3[$lane], $src4[$lane]\\}, $Rn$Rm",
2383 "$Rn.addr = $wb", []> {
2384 let Inst{4} = Rn{4};
2385 let DecoderMethod = "DecodeVST4LN";
2388 def VST4LNd8_UPD : VST4LNWB<0b0011, {?,?,?,?}, "8"> {
2389 let Inst{7-5} = lane{2-0};
2391 def VST4LNd16_UPD : VST4LNWB<0b0111, {?,?,0,?}, "16"> {
2392 let Inst{7-6} = lane{1-0};
2394 def VST4LNd32_UPD : VST4LNWB<0b1011, {?,0,?,?}, "32"> {
2395 let Inst{7} = lane{0};
2396 let Inst{5} = Rn{5};
2399 def VST4LNd8Pseudo_UPD : VSTQQLNWBPseudo<IIC_VST4lnu>, Sched<[WriteVST2]>;
2400 def VST4LNd16Pseudo_UPD : VSTQQLNWBPseudo<IIC_VST4lnu>, Sched<[WriteVST2]>;
2401 def VST4LNd32Pseudo_UPD : VSTQQLNWBPseudo<IIC_VST4lnu>, Sched<[WriteVST2]>;
2403 def VST4LNq16_UPD : VST4LNWB<0b0111, {?,?,1,?}, "16"> {
2404 let Inst{7-6} = lane{1-0};
2406 def VST4LNq32_UPD : VST4LNWB<0b1011, {?,1,?,?}, "32"> {
2407 let Inst{7} = lane{0};
2408 let Inst{5} = Rn{5};
2411 def VST4LNq16Pseudo_UPD : VSTQQQQLNWBPseudo<IIC_VST4lnu>, Sched<[WriteVST2]>;
2412 def VST4LNq32Pseudo_UPD : VSTQQQQLNWBPseudo<IIC_VST4lnu>, Sched<[WriteVST2]>;
2414 } // mayStore = 1, hasSideEffects = 0, hasExtraSrcRegAllocReq = 1
2416 // Use vld1/vst1 for unaligned f64 load / store
2417 let Predicates = [IsLE,HasNEON] in {
2418 def : Pat<(f64 (hword_alignedload addrmode6:$addr)),
2419 (VLD1d16 addrmode6:$addr)>;
2420 def : Pat<(hword_alignedstore (f64 DPR:$value), addrmode6:$addr),
2421 (VST1d16 addrmode6:$addr, DPR:$value)>;
2422 def : Pat<(f64 (byte_alignedload addrmode6:$addr)),
2423 (VLD1d8 addrmode6:$addr)>;
2424 def : Pat<(byte_alignedstore (f64 DPR:$value), addrmode6:$addr),
2425 (VST1d8 addrmode6:$addr, DPR:$value)>;
2427 let Predicates = [IsBE,HasNEON] in {
2428 def : Pat<(f64 (non_word_alignedload addrmode6:$addr)),
2429 (VLD1d64 addrmode6:$addr)>;
2430 def : Pat<(non_word_alignedstore (f64 DPR:$value), addrmode6:$addr),
2431 (VST1d64 addrmode6:$addr, DPR:$value)>;
2434 // Use vld1/vst1 for Q and QQ. Also use them for unaligned v2f64
2435 // load / store if it's legal.
2436 let Predicates = [HasNEON] in {
2437 def : Pat<(v2f64 (dword_alignedload addrmode6:$addr)),
2438 (VLD1q64 addrmode6:$addr)>;
2439 def : Pat<(dword_alignedstore (v2f64 QPR:$value), addrmode6:$addr),
2440 (VST1q64 addrmode6:$addr, QPR:$value)>;
2442 let Predicates = [IsLE,HasNEON] in {
2443 def : Pat<(v2f64 (word_alignedload addrmode6:$addr)),
2444 (VLD1q32 addrmode6:$addr)>;
2445 def : Pat<(word_alignedstore (v2f64 QPR:$value), addrmode6:$addr),
2446 (VST1q32 addrmode6:$addr, QPR:$value)>;
2447 def : Pat<(v2f64 (hword_alignedload addrmode6:$addr)),
2448 (VLD1q16 addrmode6:$addr)>;
2449 def : Pat<(hword_alignedstore (v2f64 QPR:$value), addrmode6:$addr),
2450 (VST1q16 addrmode6:$addr, QPR:$value)>;
2451 def : Pat<(v2f64 (byte_alignedload addrmode6:$addr)),
2452 (VLD1q8 addrmode6:$addr)>;
2453 def : Pat<(byte_alignedstore (v2f64 QPR:$value), addrmode6:$addr),
2454 (VST1q8 addrmode6:$addr, QPR:$value)>;
2457 //===----------------------------------------------------------------------===//
2458 // Instruction Classes
2459 //===----------------------------------------------------------------------===//
2461 // Basic 2-register operations: double- and quad-register.
2462 class N2VD<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
2463 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
2464 string Dt, ValueType ResTy, ValueType OpTy, SDNode OpNode>
2465 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4, (outs DPR:$Vd),
2466 (ins DPR:$Vm), IIC_VUNAD, OpcodeStr, Dt,"$Vd, $Vm", "",
2467 [(set DPR:$Vd, (ResTy (OpNode (OpTy DPR:$Vm))))]>;
2468 class N2VQ<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
2469 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
2470 string Dt, ValueType ResTy, ValueType OpTy, SDNode OpNode>
2471 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4, (outs QPR:$Vd),
2472 (ins QPR:$Vm), IIC_VUNAQ, OpcodeStr, Dt,"$Vd, $Vm", "",
2473 [(set QPR:$Vd, (ResTy (OpNode (OpTy QPR:$Vm))))]>;
2475 // Basic 2-register intrinsics, both double- and quad-register.
2476 class N2VDInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
2477 bits<2> op17_16, bits<5> op11_7, bit op4,
2478 InstrItinClass itin, string OpcodeStr, string Dt,
2479 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
2480 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4, (outs DPR:$Vd),
2481 (ins DPR:$Vm), itin, OpcodeStr, Dt, "$Vd, $Vm", "",
2482 [(set DPR:$Vd, (ResTy (IntOp (OpTy DPR:$Vm))))]>;
2483 class N2VQInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
2484 bits<2> op17_16, bits<5> op11_7, bit op4,
2485 InstrItinClass itin, string OpcodeStr, string Dt,
2486 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
2487 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4, (outs QPR:$Vd),
2488 (ins QPR:$Vm), itin, OpcodeStr, Dt, "$Vd, $Vm", "",
2489 [(set QPR:$Vd, (ResTy (IntOp (OpTy QPR:$Vm))))]>;
2491 // Same as above, but not predicated.
2492 class N2VDIntnp<bits<2> op19_18, bits<2> op17_16, bits<3> op10_8, bit op7,
2493 InstrItinClass itin, string OpcodeStr, string Dt,
2494 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
2495 : N2Vnp<op19_18, op17_16, op10_8, op7, 0, (outs DPR:$Vd), (ins DPR:$Vm),
2496 itin, OpcodeStr, Dt,
2497 [(set DPR:$Vd, (ResTy (IntOp (OpTy DPR:$Vm))))]>;
2499 class N2VQIntnp<bits<2> op19_18, bits<2> op17_16, bits<3> op10_8, bit op7,
2500 InstrItinClass itin, string OpcodeStr, string Dt,
2501 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
2502 : N2Vnp<op19_18, op17_16, op10_8, op7, 1, (outs QPR:$Vd), (ins QPR:$Vm),
2503 itin, OpcodeStr, Dt,
2504 [(set QPR:$Vd, (ResTy (IntOp (OpTy QPR:$Vm))))]>;
2506 // Similar to NV2VQIntnp with some more encoding bits exposed (crypto).
2507 class N2VQIntXnp<bits<2> op19_18, bits<2> op17_16, bits<3> op10_8, bit op6,
2508 bit op7, InstrItinClass itin, string OpcodeStr, string Dt,
2509 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
2510 : N2Vnp<op19_18, op17_16, op10_8, op7, op6, (outs QPR:$Vd), (ins QPR:$Vm),
2511 itin, OpcodeStr, Dt,
2512 [(set QPR:$Vd, (ResTy (IntOp (OpTy QPR:$Vm))))]>;
2514 // Same as N2VQIntXnp but with Vd as a src register.
2515 class N2VQIntX2np<bits<2> op19_18, bits<2> op17_16, bits<3> op10_8, bit op6,
2516 bit op7, InstrItinClass itin, string OpcodeStr, string Dt,
2517 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
2518 : N2Vnp<op19_18, op17_16, op10_8, op7, op6,
2519 (outs QPR:$Vd), (ins QPR:$src, QPR:$Vm),
2520 itin, OpcodeStr, Dt,
2521 [(set QPR:$Vd, (ResTy (IntOp (OpTy QPR:$src), (OpTy QPR:$Vm))))]> {
2522 let Constraints = "$src = $Vd";
2525 // Narrow 2-register operations.
2526 class N2VN<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
2527 bits<2> op17_16, bits<5> op11_7, bit op6, bit op4,
2528 InstrItinClass itin, string OpcodeStr, string Dt,
2529 ValueType TyD, ValueType TyQ, SDNode OpNode>
2530 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, op6, op4, (outs DPR:$Vd),
2531 (ins QPR:$Vm), itin, OpcodeStr, Dt, "$Vd, $Vm", "",
2532 [(set DPR:$Vd, (TyD (OpNode (TyQ QPR:$Vm))))]>;
2534 // Narrow 2-register intrinsics.
2535 class N2VNInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
2536 bits<2> op17_16, bits<5> op11_7, bit op6, bit op4,
2537 InstrItinClass itin, string OpcodeStr, string Dt,
2538 ValueType TyD, ValueType TyQ, SDPatternOperator IntOp>
2539 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, op6, op4, (outs DPR:$Vd),
2540 (ins QPR:$Vm), itin, OpcodeStr, Dt, "$Vd, $Vm", "",
2541 [(set DPR:$Vd, (TyD (IntOp (TyQ QPR:$Vm))))]>;
2543 // Long 2-register operations (currently only used for VMOVL).
2544 class N2VL<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
2545 bits<2> op17_16, bits<5> op11_7, bit op6, bit op4,
2546 InstrItinClass itin, string OpcodeStr, string Dt,
2547 ValueType TyQ, ValueType TyD, SDNode OpNode>
2548 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, op6, op4, (outs QPR:$Vd),
2549 (ins DPR:$Vm), itin, OpcodeStr, Dt, "$Vd, $Vm", "",
2550 [(set QPR:$Vd, (TyQ (OpNode (TyD DPR:$Vm))))]>;
2552 // Long 2-register intrinsics.
2553 class N2VLInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
2554 bits<2> op17_16, bits<5> op11_7, bit op6, bit op4,
2555 InstrItinClass itin, string OpcodeStr, string Dt,
2556 ValueType TyQ, ValueType TyD, SDPatternOperator IntOp>
2557 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, op6, op4, (outs QPR:$Vd),
2558 (ins DPR:$Vm), itin, OpcodeStr, Dt, "$Vd, $Vm", "",
2559 [(set QPR:$Vd, (TyQ (IntOp (TyD DPR:$Vm))))]>;
2561 // 2-register shuffles (VTRN/VZIP/VUZP), both double- and quad-register.
2562 class N2VDShuffle<bits<2> op19_18, bits<5> op11_7, string OpcodeStr, string Dt>
2563 : N2V<0b11, 0b11, op19_18, 0b10, op11_7, 0, 0, (outs DPR:$Vd, DPR:$Vm),
2564 (ins DPR:$src1, DPR:$src2), IIC_VPERMD,
2565 OpcodeStr, Dt, "$Vd, $Vm",
2566 "$src1 = $Vd, $src2 = $Vm", []>;
2567 class N2VQShuffle<bits<2> op19_18, bits<5> op11_7,
2568 InstrItinClass itin, string OpcodeStr, string Dt>
2569 : N2V<0b11, 0b11, op19_18, 0b10, op11_7, 1, 0, (outs QPR:$Vd, QPR:$Vm),
2570 (ins QPR:$src1, QPR:$src2), itin, OpcodeStr, Dt, "$Vd, $Vm",
2571 "$src1 = $Vd, $src2 = $Vm", []>;
2573 // Basic 3-register operations: double- and quad-register.
2574 class N3VD<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
2575 InstrItinClass itin, string OpcodeStr, string Dt,
2576 ValueType ResTy, ValueType OpTy, SDNode OpNode, bit Commutable>
2577 : N3V<op24, op23, op21_20, op11_8, 0, op4,
2578 (outs DPR:$Vd), (ins DPR:$Vn, DPR:$Vm), N3RegFrm, itin,
2579 OpcodeStr, Dt, "$Vd, $Vn, $Vm", "",
2580 [(set DPR:$Vd, (ResTy (OpNode (OpTy DPR:$Vn), (OpTy DPR:$Vm))))]> {
2581 // All of these have a two-operand InstAlias.
2582 let TwoOperandAliasConstraint = "$Vn = $Vd";
2583 let isCommutable = Commutable;
2585 // Same as N3VD but no data type.
2586 class N3VDX<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
2587 InstrItinClass itin, string OpcodeStr,
2588 ValueType ResTy, ValueType OpTy,
2589 SDNode OpNode, bit Commutable>
2590 : N3VX<op24, op23, op21_20, op11_8, 0, op4,
2591 (outs DPR:$Vd), (ins DPR:$Vn, DPR:$Vm), N3RegFrm, itin,
2592 OpcodeStr, "$Vd, $Vn, $Vm", "",
2593 [(set DPR:$Vd, (ResTy (OpNode (OpTy DPR:$Vn), (OpTy DPR:$Vm))))]>{
2594 // All of these have a two-operand InstAlias.
2595 let TwoOperandAliasConstraint = "$Vn = $Vd";
2596 let isCommutable = Commutable;
2599 class N3VDSL<bits<2> op21_20, bits<4> op11_8,
2600 InstrItinClass itin, string OpcodeStr, string Dt,
2601 ValueType Ty, SDNode ShOp>
2602 : N3VLane32<0, 1, op21_20, op11_8, 1, 0,
2603 (outs DPR:$Vd), (ins DPR:$Vn, DPR_VFP2:$Vm, VectorIndex32:$lane),
2604 NVMulSLFrm, itin, OpcodeStr, Dt, "$Vd, $Vn, $Vm$lane", "",
2606 (Ty (ShOp (Ty DPR:$Vn),
2607 (Ty (ARMvduplane (Ty DPR_VFP2:$Vm),imm:$lane)))))]> {
2608 // All of these have a two-operand InstAlias.
2609 let TwoOperandAliasConstraint = "$Vn = $Vd";
2610 let isCommutable = 0;
2612 class N3VDSL16<bits<2> op21_20, bits<4> op11_8,
2613 string OpcodeStr, string Dt, ValueType Ty, SDNode ShOp>
2614 : N3VLane16<0, 1, op21_20, op11_8, 1, 0,
2615 (outs DPR:$Vd), (ins DPR:$Vn, DPR_8:$Vm, VectorIndex16:$lane),
2616 NVMulSLFrm, IIC_VMULi16D, OpcodeStr, Dt,"$Vd, $Vn, $Vm$lane","",
2618 (Ty (ShOp (Ty DPR:$Vn),
2619 (Ty (ARMvduplane (Ty DPR_8:$Vm), imm:$lane)))))]> {
2620 // All of these have a two-operand InstAlias.
2621 let TwoOperandAliasConstraint = "$Vn = $Vd";
2622 let isCommutable = 0;
2625 class N3VQ<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
2626 InstrItinClass itin, string OpcodeStr, string Dt,
2627 ValueType ResTy, ValueType OpTy, SDNode OpNode, bit Commutable>
2628 : N3V<op24, op23, op21_20, op11_8, 1, op4,
2629 (outs QPR:$Vd), (ins QPR:$Vn, QPR:$Vm), N3RegFrm, itin,
2630 OpcodeStr, Dt, "$Vd, $Vn, $Vm", "",
2631 [(set QPR:$Vd, (ResTy (OpNode (OpTy QPR:$Vn), (OpTy QPR:$Vm))))]> {
2632 // All of these have a two-operand InstAlias.
2633 let TwoOperandAliasConstraint = "$Vn = $Vd";
2634 let isCommutable = Commutable;
2636 class N3VQX<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
2637 InstrItinClass itin, string OpcodeStr,
2638 ValueType ResTy, ValueType OpTy, SDNode OpNode, bit Commutable>
2639 : N3VX<op24, op23, op21_20, op11_8, 1, op4,
2640 (outs QPR:$Vd), (ins QPR:$Vn, QPR:$Vm), N3RegFrm, itin,
2641 OpcodeStr, "$Vd, $Vn, $Vm", "",
2642 [(set QPR:$Vd, (ResTy (OpNode (OpTy QPR:$Vn), (OpTy QPR:$Vm))))]>{
2643 // All of these have a two-operand InstAlias.
2644 let TwoOperandAliasConstraint = "$Vn = $Vd";
2645 let isCommutable = Commutable;
2647 class N3VQSL<bits<2> op21_20, bits<4> op11_8,
2648 InstrItinClass itin, string OpcodeStr, string Dt,
2649 ValueType ResTy, ValueType OpTy, SDNode ShOp>
2650 : N3VLane32<1, 1, op21_20, op11_8, 1, 0,
2651 (outs QPR:$Vd), (ins QPR:$Vn, DPR_VFP2:$Vm, VectorIndex32:$lane),
2652 NVMulSLFrm, itin, OpcodeStr, Dt, "$Vd, $Vn, $Vm$lane", "",
2653 [(set (ResTy QPR:$Vd),
2654 (ResTy (ShOp (ResTy QPR:$Vn),
2655 (ResTy (ARMvduplane (OpTy DPR_VFP2:$Vm),
2657 // All of these have a two-operand InstAlias.
2658 let TwoOperandAliasConstraint = "$Vn = $Vd";
2659 let isCommutable = 0;
2661 class N3VQSL16<bits<2> op21_20, bits<4> op11_8, string OpcodeStr, string Dt,
2662 ValueType ResTy, ValueType OpTy, SDNode ShOp>
2663 : N3VLane16<1, 1, op21_20, op11_8, 1, 0,
2664 (outs QPR:$Vd), (ins QPR:$Vn, DPR_8:$Vm, VectorIndex16:$lane),
2665 NVMulSLFrm, IIC_VMULi16Q, OpcodeStr, Dt,"$Vd, $Vn, $Vm$lane", "",
2666 [(set (ResTy QPR:$Vd),
2667 (ResTy (ShOp (ResTy QPR:$Vn),
2668 (ResTy (ARMvduplane (OpTy DPR_8:$Vm),
2670 // All of these have a two-operand InstAlias.
2671 let TwoOperandAliasConstraint = "$Vn = $Vd";
2672 let isCommutable = 0;
2675 // Basic 3-register intrinsics, both double- and quad-register.
2676 class N3VDInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
2677 Format f, InstrItinClass itin, string OpcodeStr, string Dt,
2678 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp, bit Commutable>
2679 : N3V<op24, op23, op21_20, op11_8, 0, op4,
2680 (outs DPR:$Vd), (ins DPR:$Vn, DPR:$Vm), f, itin,
2681 OpcodeStr, Dt, "$Vd, $Vn, $Vm", "",
2682 [(set DPR:$Vd, (ResTy (IntOp (OpTy DPR:$Vn), (OpTy DPR:$Vm))))]> {
2683 // All of these have a two-operand InstAlias.
2684 let TwoOperandAliasConstraint = "$Vn = $Vd";
2685 let isCommutable = Commutable;
2688 class N3VDIntnp<bits<5> op27_23, bits<2> op21_20, bits<4> op11_8, bit op6,
2689 bit op4, Format f, InstrItinClass itin, string OpcodeStr,
2690 string Dt, ValueType ResTy, ValueType OpTy,
2691 SDPatternOperator IntOp, bit Commutable>
2692 : N3Vnp<op27_23, op21_20, op11_8, op6, op4,
2693 (outs DPR:$Vd), (ins DPR:$Vn, DPR:$Vm), N3RegFrm, itin, OpcodeStr, Dt,
2694 [(set DPR:$Vd, (ResTy (IntOp (OpTy DPR:$Vn), (OpTy DPR:$Vm))))]>;
2696 class N3VDIntSL<bits<2> op21_20, bits<4> op11_8, InstrItinClass itin,
2697 string OpcodeStr, string Dt, ValueType Ty, SDPatternOperator IntOp>
2698 : N3VLane32<0, 1, op21_20, op11_8, 1, 0,
2699 (outs DPR:$Vd), (ins DPR:$Vn, DPR_VFP2:$Vm, VectorIndex32:$lane),
2700 NVMulSLFrm, itin, OpcodeStr, Dt, "$Vd, $Vn, $Vm$lane", "",
2702 (Ty (IntOp (Ty DPR:$Vn),
2703 (Ty (ARMvduplane (Ty DPR_VFP2:$Vm),
2705 let isCommutable = 0;
2708 class N3VDIntSL16<bits<2> op21_20, bits<4> op11_8, InstrItinClass itin,
2709 string OpcodeStr, string Dt, ValueType Ty, SDPatternOperator IntOp>
2710 : N3VLane16<0, 1, op21_20, op11_8, 1, 0,
2711 (outs DPR:$Vd), (ins DPR:$Vn, DPR_8:$Vm, VectorIndex16:$lane),
2712 NVMulSLFrm, itin, OpcodeStr, Dt, "$Vd, $Vn, $Vm$lane", "",
2714 (Ty (IntOp (Ty DPR:$Vn),
2715 (Ty (ARMvduplane (Ty DPR_8:$Vm), imm:$lane)))))]> {
2716 let isCommutable = 0;
2718 class N3VDIntSh<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
2719 Format f, InstrItinClass itin, string OpcodeStr, string Dt,
2720 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
2721 : N3V<op24, op23, op21_20, op11_8, 0, op4,
2722 (outs DPR:$Vd), (ins DPR:$Vm, DPR:$Vn), f, itin,
2723 OpcodeStr, Dt, "$Vd, $Vm, $Vn", "",
2724 [(set DPR:$Vd, (ResTy (IntOp (OpTy DPR:$Vm), (OpTy DPR:$Vn))))]> {
2725 let TwoOperandAliasConstraint = "$Vm = $Vd";
2726 let isCommutable = 0;
2729 class N3VQInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
2730 Format f, InstrItinClass itin, string OpcodeStr, string Dt,
2731 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp, bit Commutable>
2732 : N3V<op24, op23, op21_20, op11_8, 1, op4,
2733 (outs QPR:$Vd), (ins QPR:$Vn, QPR:$Vm), f, itin,
2734 OpcodeStr, Dt, "$Vd, $Vn, $Vm", "",
2735 [(set QPR:$Vd, (ResTy (IntOp (OpTy QPR:$Vn), (OpTy QPR:$Vm))))]> {
2736 // All of these have a two-operand InstAlias.
2737 let TwoOperandAliasConstraint = "$Vn = $Vd";
2738 let isCommutable = Commutable;
2741 class N3VQIntnp<bits<5> op27_23, bits<2> op21_20, bits<4> op11_8, bit op6,
2742 bit op4, Format f, InstrItinClass itin, string OpcodeStr,
2743 string Dt, ValueType ResTy, ValueType OpTy,
2744 SDPatternOperator IntOp, bit Commutable>
2745 : N3Vnp<op27_23, op21_20, op11_8, op6, op4,
2746 (outs QPR:$Vd), (ins QPR:$Vn, QPR:$Vm), f, itin, OpcodeStr, Dt,
2747 [(set QPR:$Vd, (ResTy (IntOp (OpTy QPR:$Vn), (OpTy QPR:$Vm))))]>;
2749 // Same as N3VQIntnp but with Vd as a src register.
2750 class N3VQInt3np<bits<5> op27_23, bits<2> op21_20, bits<4> op11_8, bit op6,
2751 bit op4, Format f, InstrItinClass itin, string OpcodeStr,
2752 string Dt, ValueType ResTy, ValueType OpTy,
2753 SDPatternOperator IntOp, bit Commutable>
2754 : N3Vnp<op27_23, op21_20, op11_8, op6, op4,
2755 (outs QPR:$Vd), (ins QPR:$src, QPR:$Vn, QPR:$Vm),
2756 f, itin, OpcodeStr, Dt,
2757 [(set QPR:$Vd, (ResTy (IntOp (OpTy QPR:$src), (OpTy QPR:$Vn),
2758 (OpTy QPR:$Vm))))]> {
2759 let Constraints = "$src = $Vd";
2762 class N3VQIntSL<bits<2> op21_20, bits<4> op11_8, InstrItinClass itin,
2763 string OpcodeStr, string Dt,
2764 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
2765 : N3VLane32<1, 1, op21_20, op11_8, 1, 0,
2766 (outs QPR:$Vd), (ins QPR:$Vn, DPR_VFP2:$Vm, VectorIndex32:$lane),
2767 NVMulSLFrm, itin, OpcodeStr, Dt, "$Vd, $Vn, $Vm$lane", "",
2768 [(set (ResTy QPR:$Vd),
2769 (ResTy (IntOp (ResTy QPR:$Vn),
2770 (ResTy (ARMvduplane (OpTy DPR_VFP2:$Vm),
2772 let isCommutable = 0;
2774 class N3VQIntSL16<bits<2> op21_20, bits<4> op11_8, InstrItinClass itin,
2775 string OpcodeStr, string Dt,
2776 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
2777 : N3VLane16<1, 1, op21_20, op11_8, 1, 0,
2778 (outs QPR:$Vd), (ins QPR:$Vn, DPR_8:$Vm, VectorIndex16:$lane),
2779 NVMulSLFrm, itin, OpcodeStr, Dt, "$Vd, $Vn, $Vm$lane", "",
2780 [(set (ResTy QPR:$Vd),
2781 (ResTy (IntOp (ResTy QPR:$Vn),
2782 (ResTy (ARMvduplane (OpTy DPR_8:$Vm),
2784 let isCommutable = 0;
2786 class N3VQIntSh<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
2787 Format f, InstrItinClass itin, string OpcodeStr, string Dt,
2788 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
2789 : N3V<op24, op23, op21_20, op11_8, 1, op4,
2790 (outs QPR:$Vd), (ins QPR:$Vm, QPR:$Vn), f, itin,
2791 OpcodeStr, Dt, "$Vd, $Vm, $Vn", "",
2792 [(set QPR:$Vd, (ResTy (IntOp (OpTy QPR:$Vm), (OpTy QPR:$Vn))))]> {
2793 let TwoOperandAliasConstraint = "$Vm = $Vd";
2794 let isCommutable = 0;
2797 // Multiply-Add/Sub operations: double- and quad-register.
2798 class N3VDMulOp<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
2799 InstrItinClass itin, string OpcodeStr, string Dt,
2800 ValueType Ty, SDPatternOperator MulOp, SDPatternOperator OpNode>
2801 : N3V<op24, op23, op21_20, op11_8, 0, op4,
2802 (outs DPR:$Vd), (ins DPR:$src1, DPR:$Vn, DPR:$Vm), N3RegFrm, itin,
2803 OpcodeStr, Dt, "$Vd, $Vn, $Vm", "$src1 = $Vd",
2804 [(set DPR:$Vd, (Ty (OpNode DPR:$src1,
2805 (Ty (MulOp DPR:$Vn, DPR:$Vm)))))]>;
2807 class N3VDMulOpSL<bits<2> op21_20, bits<4> op11_8, InstrItinClass itin,
2808 string OpcodeStr, string Dt,
2809 ValueType Ty, SDPatternOperator MulOp, SDPatternOperator ShOp>
2810 : N3VLane32<0, 1, op21_20, op11_8, 1, 0,
2812 (ins DPR:$src1, DPR:$Vn, DPR_VFP2:$Vm, VectorIndex32:$lane),
2814 OpcodeStr, Dt, "$Vd, $Vn, $Vm$lane", "$src1 = $Vd",
2816 (Ty (ShOp (Ty DPR:$src1),
2818 (Ty (ARMvduplane (Ty DPR_VFP2:$Vm),
2820 class N3VDMulOpSL16<bits<2> op21_20, bits<4> op11_8, InstrItinClass itin,
2821 string OpcodeStr, string Dt,
2822 ValueType Ty, SDPatternOperator MulOp, SDPatternOperator ShOp>
2823 : N3VLane16<0, 1, op21_20, op11_8, 1, 0,
2825 (ins DPR:$src1, DPR:$Vn, DPR_8:$Vm, VectorIndex16:$lane),
2827 OpcodeStr, Dt, "$Vd, $Vn, $Vm$lane", "$src1 = $Vd",
2829 (Ty (ShOp (Ty DPR:$src1),
2831 (Ty (ARMvduplane (Ty DPR_8:$Vm),
2834 class N3VQMulOp<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
2835 InstrItinClass itin, string OpcodeStr, string Dt, ValueType Ty,
2836 SDPatternOperator MulOp, SDPatternOperator OpNode>
2837 : N3V<op24, op23, op21_20, op11_8, 1, op4,
2838 (outs QPR:$Vd), (ins QPR:$src1, QPR:$Vn, QPR:$Vm), N3RegFrm, itin,
2839 OpcodeStr, Dt, "$Vd, $Vn, $Vm", "$src1 = $Vd",
2840 [(set QPR:$Vd, (Ty (OpNode QPR:$src1,
2841 (Ty (MulOp QPR:$Vn, QPR:$Vm)))))]>;
2842 class N3VQMulOpSL<bits<2> op21_20, bits<4> op11_8, InstrItinClass itin,
2843 string OpcodeStr, string Dt, ValueType ResTy, ValueType OpTy,
2844 SDPatternOperator MulOp, SDPatternOperator ShOp>
2845 : N3VLane32<1, 1, op21_20, op11_8, 1, 0,
2847 (ins QPR:$src1, QPR:$Vn, DPR_VFP2:$Vm, VectorIndex32:$lane),
2849 OpcodeStr, Dt, "$Vd, $Vn, $Vm$lane", "$src1 = $Vd",
2850 [(set (ResTy QPR:$Vd),
2851 (ResTy (ShOp (ResTy QPR:$src1),
2852 (ResTy (MulOp QPR:$Vn,
2853 (ResTy (ARMvduplane (OpTy DPR_VFP2:$Vm),
2855 class N3VQMulOpSL16<bits<2> op21_20, bits<4> op11_8, InstrItinClass itin,
2856 string OpcodeStr, string Dt,
2857 ValueType ResTy, ValueType OpTy,
2858 SDPatternOperator MulOp, SDPatternOperator ShOp>
2859 : N3VLane16<1, 1, op21_20, op11_8, 1, 0,
2861 (ins QPR:$src1, QPR:$Vn, DPR_8:$Vm, VectorIndex16:$lane),
2863 OpcodeStr, Dt, "$Vd, $Vn, $Vm$lane", "$src1 = $Vd",
2864 [(set (ResTy QPR:$Vd),
2865 (ResTy (ShOp (ResTy QPR:$src1),
2866 (ResTy (MulOp QPR:$Vn,
2867 (ResTy (ARMvduplane (OpTy DPR_8:$Vm),
2870 // Neon Intrinsic-Op instructions (VABA): double- and quad-register.
2871 class N3VDIntOp<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
2872 InstrItinClass itin, string OpcodeStr, string Dt,
2873 ValueType Ty, SDPatternOperator IntOp, SDNode OpNode>
2874 : N3V<op24, op23, op21_20, op11_8, 0, op4,
2875 (outs DPR:$Vd), (ins DPR:$src1, DPR:$Vn, DPR:$Vm), N3RegFrm, itin,
2876 OpcodeStr, Dt, "$Vd, $Vn, $Vm", "$src1 = $Vd",
2877 [(set DPR:$Vd, (Ty (OpNode DPR:$src1,
2878 (Ty (IntOp (Ty DPR:$Vn), (Ty DPR:$Vm))))))]>;
2879 class N3VQIntOp<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
2880 InstrItinClass itin, string OpcodeStr, string Dt,
2881 ValueType Ty, SDPatternOperator IntOp, SDNode OpNode>
2882 : N3V<op24, op23, op21_20, op11_8, 1, op4,
2883 (outs QPR:$Vd), (ins QPR:$src1, QPR:$Vn, QPR:$Vm), N3RegFrm, itin,
2884 OpcodeStr, Dt, "$Vd, $Vn, $Vm", "$src1 = $Vd",
2885 [(set QPR:$Vd, (Ty (OpNode QPR:$src1,
2886 (Ty (IntOp (Ty QPR:$Vn), (Ty QPR:$Vm))))))]>;
2888 // Neon 3-argument intrinsics, both double- and quad-register.
2889 // The destination register is also used as the first source operand register.
2890 class N3VDInt3<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
2891 InstrItinClass itin, string OpcodeStr, string Dt,
2892 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
2893 : N3V<op24, op23, op21_20, op11_8, 0, op4,
2894 (outs DPR:$Vd), (ins DPR:$src1, DPR:$Vn, DPR:$Vm), N3RegFrm, itin,
2895 OpcodeStr, Dt, "$Vd, $Vn, $Vm", "$src1 = $Vd",
2896 [(set DPR:$Vd, (ResTy (IntOp (OpTy DPR:$src1),
2897 (OpTy DPR:$Vn), (OpTy DPR:$Vm))))]>;
2898 class N3VQInt3<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
2899 InstrItinClass itin, string OpcodeStr, string Dt,
2900 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
2901 : N3V<op24, op23, op21_20, op11_8, 1, op4,
2902 (outs QPR:$Vd), (ins QPR:$src1, QPR:$Vn, QPR:$Vm), N3RegFrm, itin,
2903 OpcodeStr, Dt, "$Vd, $Vn, $Vm", "$src1 = $Vd",
2904 [(set QPR:$Vd, (ResTy (IntOp (OpTy QPR:$src1),
2905 (OpTy QPR:$Vn), (OpTy QPR:$Vm))))]>;
2907 // Long Multiply-Add/Sub operations.
2908 class N3VLMulOp<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
2909 InstrItinClass itin, string OpcodeStr, string Dt,
2910 ValueType TyQ, ValueType TyD, SDNode MulOp, SDNode OpNode>
2911 : N3V<op24, op23, op21_20, op11_8, 0, op4,
2912 (outs QPR:$Vd), (ins QPR:$src1, DPR:$Vn, DPR:$Vm), N3RegFrm, itin,
2913 OpcodeStr, Dt, "$Vd, $Vn, $Vm", "$src1 = $Vd",
2914 [(set QPR:$Vd, (OpNode (TyQ QPR:$src1),
2915 (TyQ (MulOp (TyD DPR:$Vn),
2916 (TyD DPR:$Vm)))))]>;
2917 class N3VLMulOpSL<bit op24, bits<2> op21_20, bits<4> op11_8,
2918 InstrItinClass itin, string OpcodeStr, string Dt,
2919 ValueType TyQ, ValueType TyD, SDNode MulOp, SDNode OpNode>
2920 : N3VLane32<op24, 1, op21_20, op11_8, 1, 0, (outs QPR:$Vd),
2921 (ins QPR:$src1, DPR:$Vn, DPR_VFP2:$Vm, VectorIndex32:$lane),
2923 OpcodeStr, Dt, "$Vd, $Vn, $Vm$lane", "$src1 = $Vd",
2925 (OpNode (TyQ QPR:$src1),
2926 (TyQ (MulOp (TyD DPR:$Vn),
2927 (TyD (ARMvduplane (TyD DPR_VFP2:$Vm),
2929 class N3VLMulOpSL16<bit op24, bits<2> op21_20, bits<4> op11_8,
2930 InstrItinClass itin, string OpcodeStr, string Dt,
2931 ValueType TyQ, ValueType TyD, SDNode MulOp, SDNode OpNode>
2932 : N3VLane16<op24, 1, op21_20, op11_8, 1, 0, (outs QPR:$Vd),
2933 (ins QPR:$src1, DPR:$Vn, DPR_8:$Vm, VectorIndex16:$lane),
2935 OpcodeStr, Dt, "$Vd, $Vn, $Vm$lane", "$src1 = $Vd",
2937 (OpNode (TyQ QPR:$src1),
2938 (TyQ (MulOp (TyD DPR:$Vn),
2939 (TyD (ARMvduplane (TyD DPR_8:$Vm),
2942 // Long Intrinsic-Op vector operations with explicit extend (VABAL).
2943 class N3VLIntExtOp<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
2944 InstrItinClass itin, string OpcodeStr, string Dt,
2945 ValueType TyQ, ValueType TyD, SDPatternOperator IntOp, SDNode ExtOp,
2947 : N3V<op24, op23, op21_20, op11_8, 0, op4,
2948 (outs QPR:$Vd), (ins QPR:$src1, DPR:$Vn, DPR:$Vm), N3RegFrm, itin,
2949 OpcodeStr, Dt, "$Vd, $Vn, $Vm", "$src1 = $Vd",
2950 [(set QPR:$Vd, (OpNode (TyQ QPR:$src1),
2951 (TyQ (ExtOp (TyD (IntOp (TyD DPR:$Vn),
2952 (TyD DPR:$Vm)))))))]>;
2954 // Neon Long 3-argument intrinsic. The destination register is
2955 // a quad-register and is also used as the first source operand register.
2956 class N3VLInt3<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
2957 InstrItinClass itin, string OpcodeStr, string Dt,
2958 ValueType TyQ, ValueType TyD, SDPatternOperator IntOp>
2959 : N3V<op24, op23, op21_20, op11_8, 0, op4,
2960 (outs QPR:$Vd), (ins QPR:$src1, DPR:$Vn, DPR:$Vm), N3RegFrm, itin,
2961 OpcodeStr, Dt, "$Vd, $Vn, $Vm", "$src1 = $Vd",
2963 (TyQ (IntOp (TyQ QPR:$src1), (TyD DPR:$Vn), (TyD DPR:$Vm))))]>;
2964 class N3VLInt3SL<bit op24, bits<2> op21_20, bits<4> op11_8, InstrItinClass itin,
2965 string OpcodeStr, string Dt,
2966 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
2967 : N3VLane32<op24, 1, op21_20, op11_8, 1, 0,
2969 (ins QPR:$src1, DPR:$Vn, DPR_VFP2:$Vm, VectorIndex32:$lane),
2971 OpcodeStr, Dt, "$Vd, $Vn, $Vm$lane", "$src1 = $Vd",
2972 [(set (ResTy QPR:$Vd),
2973 (ResTy (IntOp (ResTy QPR:$src1),
2975 (OpTy (ARMvduplane (OpTy DPR_VFP2:$Vm),
2977 class N3VLInt3SL16<bit op24, bits<2> op21_20, bits<4> op11_8,
2978 InstrItinClass itin, string OpcodeStr, string Dt,
2979 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
2980 : N3VLane16<op24, 1, op21_20, op11_8, 1, 0,
2982 (ins QPR:$src1, DPR:$Vn, DPR_8:$Vm, VectorIndex16:$lane),
2984 OpcodeStr, Dt, "$Vd, $Vn, $Vm$lane", "$src1 = $Vd",
2985 [(set (ResTy QPR:$Vd),
2986 (ResTy (IntOp (ResTy QPR:$src1),
2988 (OpTy (ARMvduplane (OpTy DPR_8:$Vm),
2991 // Narrowing 3-register intrinsics.
2992 class N3VNInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
2993 string OpcodeStr, string Dt, ValueType TyD, ValueType TyQ,
2994 SDPatternOperator IntOp, bit Commutable>
2995 : N3V<op24, op23, op21_20, op11_8, 0, op4,
2996 (outs DPR:$Vd), (ins QPR:$Vn, QPR:$Vm), N3RegFrm, IIC_VBINi4D,
2997 OpcodeStr, Dt, "$Vd, $Vn, $Vm", "",
2998 [(set DPR:$Vd, (TyD (IntOp (TyQ QPR:$Vn), (TyQ QPR:$Vm))))]> {
2999 let isCommutable = Commutable;
3002 // Long 3-register operations.
3003 class N3VL<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
3004 InstrItinClass itin, string OpcodeStr, string Dt,
3005 ValueType TyQ, ValueType TyD, SDNode OpNode, bit Commutable>
3006 : N3V<op24, op23, op21_20, op11_8, 0, op4,
3007 (outs QPR:$Vd), (ins DPR:$Vn, DPR:$Vm), N3RegFrm, itin,
3008 OpcodeStr, Dt, "$Vd, $Vn, $Vm", "",
3009 [(set QPR:$Vd, (TyQ (OpNode (TyD DPR:$Vn), (TyD DPR:$Vm))))]> {
3010 let isCommutable = Commutable;
3013 class N3VLSL<bit op24, bits<2> op21_20, bits<4> op11_8,
3014 InstrItinClass itin, string OpcodeStr, string Dt,
3015 ValueType TyQ, ValueType TyD, SDNode OpNode>
3016 : N3VLane32<op24, 1, op21_20, op11_8, 1, 0,
3017 (outs QPR:$Vd), (ins DPR:$Vn, DPR_VFP2:$Vm, VectorIndex32:$lane),
3018 NVMulSLFrm, itin, OpcodeStr, Dt, "$Vd, $Vn, $Vm$lane", "",
3020 (TyQ (OpNode (TyD DPR:$Vn),
3021 (TyD (ARMvduplane (TyD DPR_VFP2:$Vm),imm:$lane)))))]>;
3022 class N3VLSL16<bit op24, bits<2> op21_20, bits<4> op11_8,
3023 InstrItinClass itin, string OpcodeStr, string Dt,
3024 ValueType TyQ, ValueType TyD, SDNode OpNode>
3025 : N3VLane16<op24, 1, op21_20, op11_8, 1, 0,
3026 (outs QPR:$Vd), (ins DPR:$Vn, DPR_8:$Vm, VectorIndex16:$lane),
3027 NVMulSLFrm, itin, OpcodeStr, Dt, "$Vd, $Vn, $Vm$lane", "",
3029 (TyQ (OpNode (TyD DPR:$Vn),
3030 (TyD (ARMvduplane (TyD DPR_8:$Vm), imm:$lane)))))]>;
3032 // Long 3-register operations with explicitly extended operands.
3033 class N3VLExt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
3034 InstrItinClass itin, string OpcodeStr, string Dt,
3035 ValueType TyQ, ValueType TyD, SDNode OpNode, SDNode ExtOp,
3037 : N3V<op24, op23, op21_20, op11_8, 0, op4,
3038 (outs QPR:$Vd), (ins DPR:$Vn, DPR:$Vm), N3RegFrm, itin,
3039 OpcodeStr, Dt, "$Vd, $Vn, $Vm", "",
3040 [(set QPR:$Vd, (OpNode (TyQ (ExtOp (TyD DPR:$Vn))),
3041 (TyQ (ExtOp (TyD DPR:$Vm)))))]> {
3042 let isCommutable = Commutable;
3045 // Long 3-register intrinsics with explicit extend (VABDL).
3046 class N3VLIntExt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
3047 InstrItinClass itin, string OpcodeStr, string Dt,
3048 ValueType TyQ, ValueType TyD, SDPatternOperator IntOp, SDNode ExtOp,
3050 : N3V<op24, op23, op21_20, op11_8, 0, op4,
3051 (outs QPR:$Vd), (ins DPR:$Vn, DPR:$Vm), N3RegFrm, itin,
3052 OpcodeStr, Dt, "$Vd, $Vn, $Vm", "",
3053 [(set QPR:$Vd, (TyQ (ExtOp (TyD (IntOp (TyD DPR:$Vn),
3054 (TyD DPR:$Vm))))))]> {
3055 let isCommutable = Commutable;
3058 // Long 3-register intrinsics.
3059 class N3VLInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
3060 InstrItinClass itin, string OpcodeStr, string Dt,
3061 ValueType TyQ, ValueType TyD, SDPatternOperator IntOp, bit Commutable>
3062 : N3V<op24, op23, op21_20, op11_8, 0, op4,
3063 (outs QPR:$Vd), (ins DPR:$Vn, DPR:$Vm), N3RegFrm, itin,
3064 OpcodeStr, Dt, "$Vd, $Vn, $Vm", "",
3065 [(set QPR:$Vd, (TyQ (IntOp (TyD DPR:$Vn), (TyD DPR:$Vm))))]> {
3066 let isCommutable = Commutable;
3069 // Same as above, but not predicated.
3070 class N3VLIntnp<bits<5> op27_23, bits<2> op21_20, bits<4> op11_8, bit op6,
3071 bit op4, InstrItinClass itin, string OpcodeStr,
3072 string Dt, ValueType ResTy, ValueType OpTy,
3073 SDPatternOperator IntOp, bit Commutable>
3074 : N3Vnp<op27_23, op21_20, op11_8, op6, op4,
3075 (outs QPR:$Vd), (ins DPR:$Vn, DPR:$Vm), N3RegFrm, itin, OpcodeStr, Dt,
3076 [(set QPR:$Vd, (ResTy (IntOp (OpTy DPR:$Vn), (OpTy DPR:$Vm))))]>;
3078 class N3VLIntSL<bit op24, bits<2> op21_20, bits<4> op11_8, InstrItinClass itin,
3079 string OpcodeStr, string Dt,
3080 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
3081 : N3VLane32<op24, 1, op21_20, op11_8, 1, 0,
3082 (outs QPR:$Vd), (ins DPR:$Vn, DPR_VFP2:$Vm, VectorIndex32:$lane),
3083 NVMulSLFrm, itin, OpcodeStr, Dt, "$Vd, $Vn, $Vm$lane", "",
3084 [(set (ResTy QPR:$Vd),
3085 (ResTy (IntOp (OpTy DPR:$Vn),
3086 (OpTy (ARMvduplane (OpTy DPR_VFP2:$Vm),
3088 class N3VLIntSL16<bit op24, bits<2> op21_20, bits<4> op11_8,
3089 InstrItinClass itin, string OpcodeStr, string Dt,
3090 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
3091 : N3VLane16<op24, 1, op21_20, op11_8, 1, 0,
3092 (outs QPR:$Vd), (ins DPR:$Vn, DPR_8:$Vm, VectorIndex16:$lane),
3093 NVMulSLFrm, itin, OpcodeStr, Dt, "$Vd, $Vn, $Vm$lane", "",
3094 [(set (ResTy QPR:$Vd),
3095 (ResTy (IntOp (OpTy DPR:$Vn),
3096 (OpTy (ARMvduplane (OpTy DPR_8:$Vm),
3099 // Wide 3-register operations.
3100 class N3VW<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
3101 string OpcodeStr, string Dt, ValueType TyQ, ValueType TyD,
3102 SDNode OpNode, SDNode ExtOp, bit Commutable>
3103 : N3V<op24, op23, op21_20, op11_8, 0, op4,
3104 (outs QPR:$Vd), (ins QPR:$Vn, DPR:$Vm), N3RegFrm, IIC_VSUBiD,
3105 OpcodeStr, Dt, "$Vd, $Vn, $Vm", "",
3106 [(set QPR:$Vd, (OpNode (TyQ QPR:$Vn),
3107 (TyQ (ExtOp (TyD DPR:$Vm)))))]> {
3108 // All of these have a two-operand InstAlias.
3109 let TwoOperandAliasConstraint = "$Vn = $Vd";
3110 let isCommutable = Commutable;
3113 // Pairwise long 2-register intrinsics, both double- and quad-register.
3114 class N2VDPLInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
3115 bits<2> op17_16, bits<5> op11_7, bit op4,
3116 string OpcodeStr, string Dt,
3117 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
3118 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4, (outs DPR:$Vd),
3119 (ins DPR:$Vm), IIC_VSHLiD, OpcodeStr, Dt, "$Vd, $Vm", "",
3120 [(set DPR:$Vd, (ResTy (IntOp (OpTy DPR:$Vm))))]>;
3121 class N2VQPLInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
3122 bits<2> op17_16, bits<5> op11_7, bit op4,
3123 string OpcodeStr, string Dt,
3124 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
3125 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4, (outs QPR:$Vd),
3126 (ins QPR:$Vm), IIC_VSHLiD, OpcodeStr, Dt, "$Vd, $Vm", "",
3127 [(set QPR:$Vd, (ResTy (IntOp (OpTy QPR:$Vm))))]>;
3129 // Pairwise long 2-register accumulate intrinsics,
3130 // both double- and quad-register.
3131 // The destination register is also used as the first source operand register.
3132 class N2VDPLInt2<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
3133 bits<2> op17_16, bits<5> op11_7, bit op4,
3134 string OpcodeStr, string Dt,
3135 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
3136 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4,
3137 (outs DPR:$Vd), (ins DPR:$src1, DPR:$Vm), IIC_VPALiD,
3138 OpcodeStr, Dt, "$Vd, $Vm", "$src1 = $Vd",
3139 [(set DPR:$Vd, (ResTy (IntOp (ResTy DPR:$src1), (OpTy DPR:$Vm))))]>;
3140 class N2VQPLInt2<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
3141 bits<2> op17_16, bits<5> op11_7, bit op4,
3142 string OpcodeStr, string Dt,
3143 ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp>
3144 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4,
3145 (outs QPR:$Vd), (ins QPR:$src1, QPR:$Vm), IIC_VPALiQ,
3146 OpcodeStr, Dt, "$Vd, $Vm", "$src1 = $Vd",
3147 [(set QPR:$Vd, (ResTy (IntOp (ResTy QPR:$src1), (OpTy QPR:$Vm))))]>;
3149 // Shift by immediate,
3150 // both double- and quad-register.
3151 let TwoOperandAliasConstraint = "$Vm = $Vd" in {
3152 class N2VDSh<bit op24, bit op23, bits<4> op11_8, bit op7, bit op4,
3153 Format f, InstrItinClass itin, Operand ImmTy,
3154 string OpcodeStr, string Dt, ValueType Ty, SDNode OpNode>
3155 : N2VImm<op24, op23, op11_8, op7, 0, op4,
3156 (outs DPR:$Vd), (ins DPR:$Vm, ImmTy:$SIMM), f, itin,
3157 OpcodeStr, Dt, "$Vd, $Vm, $SIMM", "",
3158 [(set DPR:$Vd, (Ty (OpNode (Ty DPR:$Vm), (i32 imm:$SIMM))))]>;
3159 class N2VQSh<bit op24, bit op23, bits<4> op11_8, bit op7, bit op4,
3160 Format f, InstrItinClass itin, Operand ImmTy,
3161 string OpcodeStr, string Dt, ValueType Ty, SDNode OpNode>
3162 : N2VImm<op24, op23, op11_8, op7, 1, op4,
3163 (outs QPR:$Vd), (ins QPR:$Vm, ImmTy:$SIMM), f, itin,
3164 OpcodeStr, Dt, "$Vd, $Vm, $SIMM", "",
3165 [(set QPR:$Vd, (Ty (OpNode (Ty QPR:$Vm), (i32 imm:$SIMM))))]>;
3168 // Long shift by immediate.
3169 class N2VLSh<bit op24, bit op23, bits<4> op11_8, bit op7, bit op6, bit op4,
3170 string OpcodeStr, string Dt,
3171 ValueType ResTy, ValueType OpTy, Operand ImmTy,
3172 SDPatternOperator OpNode>
3173 : N2VImm<op24, op23, op11_8, op7, op6, op4,
3174 (outs QPR:$Vd), (ins DPR:$Vm, ImmTy:$SIMM), N2RegVShLFrm,
3175 IIC_VSHLiD, OpcodeStr, Dt, "$Vd, $Vm, $SIMM", "",
3176 [(set QPR:$Vd, (ResTy (OpNode (OpTy DPR:$Vm), ImmTy:$SIMM)))]>;
3178 // Narrow shift by immediate.
3179 class N2VNSh<bit op24, bit op23, bits<4> op11_8, bit op7, bit op6, bit op4,
3180 InstrItinClass itin, string OpcodeStr, string Dt,
3181 ValueType ResTy, ValueType OpTy, Operand ImmTy,
3182 SDPatternOperator OpNode>
3183 : N2VImm<op24, op23, op11_8, op7, op6, op4,
3184 (outs DPR:$Vd), (ins QPR:$Vm, ImmTy:$SIMM), N2RegVShRFrm, itin,
3185 OpcodeStr, Dt, "$Vd, $Vm, $SIMM", "",
3186 [(set DPR:$Vd, (ResTy (OpNode (OpTy QPR:$Vm),
3187 (i32 ImmTy:$SIMM))))]>;
3189 // Shift right by immediate and accumulate,
3190 // both double- and quad-register.
3191 let TwoOperandAliasConstraint = "$Vm = $Vd" in {
3192 class N2VDShAdd<bit op24, bit op23, bits<4> op11_8, bit op7, bit op4,
3193 Operand ImmTy, string OpcodeStr, string Dt,
3194 ValueType Ty, SDNode ShOp>
3195 : N2VImm<op24, op23, op11_8, op7, 0, op4, (outs DPR:$Vd),
3196 (ins DPR:$src1, DPR:$Vm, ImmTy:$SIMM), N2RegVShRFrm, IIC_VPALiD,
3197 OpcodeStr, Dt, "$Vd, $Vm, $SIMM", "$src1 = $Vd",
3198 [(set DPR:$Vd, (Ty (add DPR:$src1,
3199 (Ty (ShOp DPR:$Vm, (i32 imm:$SIMM))))))]>;
3200 class N2VQShAdd<bit op24, bit op23, bits<4> op11_8, bit op7, bit op4,
3201 Operand ImmTy, string OpcodeStr, string Dt,
3202 ValueType Ty, SDNode ShOp>
3203 : N2VImm<op24, op23, op11_8, op7, 1, op4, (outs QPR:$Vd),
3204 (ins QPR:$src1, QPR:$Vm, ImmTy:$SIMM), N2RegVShRFrm, IIC_VPALiD,
3205 OpcodeStr, Dt, "$Vd, $Vm, $SIMM", "$src1 = $Vd",
3206 [(set QPR:$Vd, (Ty (add QPR:$src1,
3207 (Ty (ShOp QPR:$Vm, (i32 imm:$SIMM))))))]>;
3210 // Shift by immediate and insert,
3211 // both double- and quad-register.
3212 let TwoOperandAliasConstraint = "$Vm = $Vd" in {
3213 class N2VDShIns<bit op24, bit op23, bits<4> op11_8, bit op7, bit op4,
3214 Operand ImmTy, Format f, string OpcodeStr, string Dt,
3215 ValueType Ty,SDNode ShOp>
3216 : N2VImm<op24, op23, op11_8, op7, 0, op4, (outs DPR:$Vd),
3217 (ins DPR:$src1, DPR:$Vm, ImmTy:$SIMM), f, IIC_VSHLiD,
3218 OpcodeStr, Dt, "$Vd, $Vm, $SIMM", "$src1 = $Vd",
3219 [(set DPR:$Vd, (Ty (ShOp DPR:$src1, DPR:$Vm, (i32 imm:$SIMM))))]>;
3220 class N2VQShIns<bit op24, bit op23, bits<4> op11_8, bit op7, bit op4,
3221 Operand ImmTy, Format f, string OpcodeStr, string Dt,
3222 ValueType Ty,SDNode ShOp>
3223 : N2VImm<op24, op23, op11_8, op7, 1, op4, (outs QPR:$Vd),
3224 (ins QPR:$src1, QPR:$Vm, ImmTy:$SIMM), f, IIC_VSHLiQ,
3225 OpcodeStr, Dt, "$Vd, $Vm, $SIMM", "$src1 = $Vd",
3226 [(set QPR:$Vd, (Ty (ShOp QPR:$src1, QPR:$Vm, (i32 imm:$SIMM))))]>;
3229 // Convert, with fractional bits immediate,
3230 // both double- and quad-register.
3231 class N2VCvtD<bit op24, bit op23, bits<4> op11_8, bit op7, bit op4,
3232 string OpcodeStr, string Dt, ValueType ResTy, ValueType OpTy,
3233 SDPatternOperator IntOp>
3234 : N2VImm<op24, op23, op11_8, op7, 0, op4,
3235 (outs DPR:$Vd), (ins DPR:$Vm, neon_vcvt_imm32:$SIMM), NVCVTFrm,
3236 IIC_VUNAD, OpcodeStr, Dt, "$Vd, $Vm, $SIMM", "",
3237 [(set DPR:$Vd, (ResTy (IntOp (OpTy DPR:$Vm), (i32 imm:$SIMM))))]>;
3238 class N2VCvtQ<bit op24, bit op23, bits<4> op11_8, bit op7, bit op4,
3239 string OpcodeStr, string Dt, ValueType ResTy, ValueType OpTy,
3240 SDPatternOperator IntOp>
3241 : N2VImm<op24, op23, op11_8, op7, 1, op4,
3242 (outs QPR:$Vd), (ins QPR:$Vm, neon_vcvt_imm32:$SIMM), NVCVTFrm,
3243 IIC_VUNAQ, OpcodeStr, Dt, "$Vd, $Vm, $SIMM", "",
3244 [(set QPR:$Vd, (ResTy (IntOp (OpTy QPR:$Vm), (i32 imm:$SIMM))))]>;
3246 //===----------------------------------------------------------------------===//
3248 //===----------------------------------------------------------------------===//
3250 // Abbreviations used in multiclass suffixes:
3251 // Q = quarter int (8 bit) elements
3252 // H = half int (16 bit) elements
3253 // S = single int (32 bit) elements
3254 // D = double int (64 bit) elements
3256 // Neon 2-register vector operations and intrinsics.
3258 // Neon 2-register comparisons.
3259 // source operand element sizes of 8, 16 and 32 bits:
3260 multiclass N2V_QHS_cmp<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
3261 bits<5> op11_7, bit op4, string opc, string Dt,
3262 string asm, PatFrag fc> {
3263 // 64-bit vector types.
3264 def v8i8 : N2V<op24_23, op21_20, 0b00, op17_16, op11_7, 0, op4,
3265 (outs DPR:$Vd), (ins DPR:$Vm), NoItinerary,
3266 opc, !strconcat(Dt, "8"), asm, "",
3267 [(set DPR:$Vd, (v8i8 (ARMvcmpz (v8i8 DPR:$Vm), fc)))]>;
3268 def v4i16 : N2V<op24_23, op21_20, 0b01, op17_16, op11_7, 0, op4,
3269 (outs DPR:$Vd), (ins DPR:$Vm), NoItinerary,
3270 opc, !strconcat(Dt, "16"), asm, "",
3271 [(set DPR:$Vd, (v4i16 (ARMvcmpz (v4i16 DPR:$Vm), fc)))]>;
3272 def v2i32 : N2V<op24_23, op21_20, 0b10, op17_16, op11_7, 0, op4,
3273 (outs DPR:$Vd), (ins DPR:$Vm), NoItinerary,
3274 opc, !strconcat(Dt, "32"), asm, "",
3275 [(set DPR:$Vd, (v2i32 (ARMvcmpz (v2i32 DPR:$Vm), fc)))]>;
3276 def v2f32 : N2V<op24_23, op21_20, 0b10, op17_16, op11_7, 0, op4,
3277 (outs DPR:$Vd), (ins DPR:$Vm), NoItinerary,
3278 opc, "f32", asm, "",
3279 [(set DPR:$Vd, (v2i32 (ARMvcmpz (v2f32 DPR:$Vm), fc)))]> {
3280 let Inst{10} = 1; // overwrite F = 1
3282 def v4f16 : N2V<op24_23, op21_20, 0b01, op17_16, op11_7, 0, op4,
3283 (outs DPR:$Vd), (ins DPR:$Vm), NoItinerary,
3284 opc, "f16", asm, "",
3285 [(set DPR:$Vd, (v4i16 (ARMvcmpz (v4f16 DPR:$Vm), fc)))]>,
3286 Requires<[HasNEON,HasFullFP16]> {
3287 let Inst{10} = 1; // overwrite F = 1
3290 // 128-bit vector types.
3291 def v16i8 : N2V<op24_23, op21_20, 0b00, op17_16, op11_7, 1, op4,
3292 (outs QPR:$Vd), (ins QPR:$Vm), NoItinerary,
3293 opc, !strconcat(Dt, "8"), asm, "",
3294 [(set QPR:$Vd, (v16i8 (ARMvcmpz (v16i8 QPR:$Vm), fc)))]>;
3295 def v8i16 : N2V<op24_23, op21_20, 0b01, op17_16, op11_7, 1, op4,
3296 (outs QPR:$Vd), (ins QPR:$Vm), NoItinerary,
3297 opc, !strconcat(Dt, "16"), asm, "",
3298 [(set QPR:$Vd, (v8i16 (ARMvcmpz (v8i16 QPR:$Vm), fc)))]>;
3299 def v4i32 : N2V<op24_23, op21_20, 0b10, op17_16, op11_7, 1, op4,
3300 (outs QPR:$Vd), (ins QPR:$Vm), NoItinerary,
3301 opc, !strconcat(Dt, "32"), asm, "",
3302 [(set QPR:$Vd, (v4i32 (ARMvcmpz (v4i32 QPR:$Vm), fc)))]>;
3303 def v4f32 : N2V<op24_23, op21_20, 0b10, op17_16, op11_7, 1, op4,
3304 (outs QPR:$Vd), (ins QPR:$Vm), NoItinerary,
3305 opc, "f32", asm, "",
3306 [(set QPR:$Vd, (v4i32 (ARMvcmpz (v4f32 QPR:$Vm), fc)))]> {
3307 let Inst{10} = 1; // overwrite F = 1
3309 def v8f16 : N2V<op24_23, op21_20, 0b01, op17_16, op11_7, 1, op4,
3310 (outs QPR:$Vd), (ins QPR:$Vm), NoItinerary,
3311 opc, "f16", asm, "",
3312 [(set QPR:$Vd, (v8i16 (ARMvcmpz (v8f16 QPR:$Vm), fc)))]>,
3313 Requires<[HasNEON,HasFullFP16]> {
3314 let Inst{10} = 1; // overwrite F = 1
3318 // Neon 3-register comparisons.
3319 class N3VQ_cmp<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
3320 InstrItinClass itin, string OpcodeStr, string Dt,
3321 ValueType ResTy, ValueType OpTy, PatFrag fc, bit Commutable>
3322 : N3V<op24, op23, op21_20, op11_8, 1, op4,
3323 (outs QPR:$Vd), (ins QPR:$Vn, QPR:$Vm), N3RegFrm, itin,
3324 OpcodeStr, Dt, "$Vd, $Vn, $Vm", "",
3325 [(set QPR:$Vd, (ResTy (ARMvcmp (OpTy QPR:$Vn), (OpTy QPR:$Vm), fc)))]> {
3326 // All of these have a two-operand InstAlias.
3327 let TwoOperandAliasConstraint = "$Vn = $Vd";
3328 let isCommutable = Commutable;
3331 class N3VD_cmp<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
3332 InstrItinClass itin, string OpcodeStr, string Dt,
3333 ValueType ResTy, ValueType OpTy, PatFrag fc, bit Commutable>
3334 : N3V<op24, op23, op21_20, op11_8, 0, op4,
3335 (outs DPR:$Vd), (ins DPR:$Vn, DPR:$Vm), N3RegFrm, itin,
3336 OpcodeStr, Dt, "$Vd, $Vn, $Vm", "",
3337 [(set DPR:$Vd, (ResTy (ARMvcmp (OpTy DPR:$Vn), (OpTy DPR:$Vm), fc)))]> {
3338 // All of these have a two-operand InstAlias.
3339 let TwoOperandAliasConstraint = "$Vn = $Vd";
3340 let isCommutable = Commutable;
3343 multiclass N3V_QHS_cmp<bit op24, bit op23, bits<4> op11_8, bit op4,
3344 InstrItinClass itinD16, InstrItinClass itinD32,
3345 InstrItinClass itinQ16, InstrItinClass itinQ32,
3346 string OpcodeStr, string Dt,
3347 PatFrag fc, bit Commutable = 0> {
3348 // 64-bit vector types.
3349 def v8i8 : N3VD_cmp<op24, op23, 0b00, op11_8, op4, itinD16,
3350 OpcodeStr, !strconcat(Dt, "8"),
3351 v8i8, v8i8, fc, Commutable>;
3352 def v4i16 : N3VD_cmp<op24, op23, 0b01, op11_8, op4, itinD16,
3353 OpcodeStr, !strconcat(Dt, "16"),
3354 v4i16, v4i16, fc, Commutable>;
3355 def v2i32 : N3VD_cmp<op24, op23, 0b10, op11_8, op4, itinD32,
3356 OpcodeStr, !strconcat(Dt, "32"),
3357 v2i32, v2i32, fc, Commutable>;
3359 // 128-bit vector types.
3360 def v16i8 : N3VQ_cmp<op24, op23, 0b00, op11_8, op4, itinQ16,
3361 OpcodeStr, !strconcat(Dt, "8"),
3362 v16i8, v16i8, fc, Commutable>;
3363 def v8i16 : N3VQ_cmp<op24, op23, 0b01, op11_8, op4, itinQ16,
3364 OpcodeStr, !strconcat(Dt, "16"),
3365 v8i16, v8i16, fc, Commutable>;
3366 def v4i32 : N3VQ_cmp<op24, op23, 0b10, op11_8, op4, itinQ32,
3367 OpcodeStr, !strconcat(Dt, "32"),
3368 v4i32, v4i32, fc, Commutable>;
3372 // Neon 2-register vector intrinsics,
3373 // element sizes of 8, 16 and 32 bits:
3374 multiclass N2VInt_QHS<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
3375 bits<5> op11_7, bit op4,
3376 InstrItinClass itinD, InstrItinClass itinQ,
3377 string OpcodeStr, string Dt, SDPatternOperator IntOp> {
3378 // 64-bit vector types.
3379 def v8i8 : N2VDInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
3380 itinD, OpcodeStr, !strconcat(Dt, "8"), v8i8, v8i8, IntOp>;
3381 def v4i16 : N2VDInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
3382 itinD, OpcodeStr, !strconcat(Dt, "16"),v4i16,v4i16,IntOp>;
3383 def v2i32 : N2VDInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
3384 itinD, OpcodeStr, !strconcat(Dt, "32"),v2i32,v2i32,IntOp>;
3386 // 128-bit vector types.
3387 def v16i8 : N2VQInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
3388 itinQ, OpcodeStr, !strconcat(Dt, "8"), v16i8,v16i8,IntOp>;
3389 def v8i16 : N2VQInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
3390 itinQ, OpcodeStr, !strconcat(Dt, "16"),v8i16,v8i16,IntOp>;
3391 def v4i32 : N2VQInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
3392 itinQ, OpcodeStr, !strconcat(Dt, "32"),v4i32,v4i32,IntOp>;
3396 // Neon Narrowing 2-register vector operations,
3397 // source operand element sizes of 16, 32 and 64 bits:
3398 multiclass N2VN_HSD<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
3399 bits<5> op11_7, bit op6, bit op4,
3400 InstrItinClass itin, string OpcodeStr, string Dt,
3402 def v8i8 : N2VN<op24_23, op21_20, 0b00, op17_16, op11_7, op6, op4,
3403 itin, OpcodeStr, !strconcat(Dt, "16"),
3404 v8i8, v8i16, OpNode>;
3405 def v4i16 : N2VN<op24_23, op21_20, 0b01, op17_16, op11_7, op6, op4,
3406 itin, OpcodeStr, !strconcat(Dt, "32"),
3407 v4i16, v4i32, OpNode>;
3408 def v2i32 : N2VN<op24_23, op21_20, 0b10, op17_16, op11_7, op6, op4,
3409 itin, OpcodeStr, !strconcat(Dt, "64"),
3410 v2i32, v2i64, OpNode>;
3413 // Neon Narrowing 2-register vector intrinsics,
3414 // source operand element sizes of 16, 32 and 64 bits:
3415 multiclass N2VNInt_HSD<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
3416 bits<5> op11_7, bit op6, bit op4,
3417 InstrItinClass itin, string OpcodeStr, string Dt,
3418 SDPatternOperator IntOp> {
3419 def v8i8 : N2VNInt<op24_23, op21_20, 0b00, op17_16, op11_7, op6, op4,
3420 itin, OpcodeStr, !strconcat(Dt, "16"),
3421 v8i8, v8i16, IntOp>;
3422 def v4i16 : N2VNInt<op24_23, op21_20, 0b01, op17_16, op11_7, op6, op4,
3423 itin, OpcodeStr, !strconcat(Dt, "32"),
3424 v4i16, v4i32, IntOp>;
3425 def v2i32 : N2VNInt<op24_23, op21_20, 0b10, op17_16, op11_7, op6, op4,
3426 itin, OpcodeStr, !strconcat(Dt, "64"),
3427 v2i32, v2i64, IntOp>;
3431 // Neon Lengthening 2-register vector intrinsic (currently specific to VMOVL).
3432 // source operand element sizes of 16, 32 and 64 bits:
3433 multiclass N2VL_QHS<bits<2> op24_23, bits<5> op11_7, bit op6, bit op4,
3434 string OpcodeStr, string Dt, SDNode OpNode> {
3435 def v8i16 : N2VL<op24_23, 0b00, 0b10, 0b00, op11_7, op6, op4, IIC_VQUNAiD,
3436 OpcodeStr, !strconcat(Dt, "8"), v8i16, v8i8, OpNode>;
3437 def v4i32 : N2VL<op24_23, 0b01, 0b00, 0b00, op11_7, op6, op4, IIC_VQUNAiD,
3438 OpcodeStr, !strconcat(Dt, "16"), v4i32, v4i16, OpNode>;
3439 def v2i64 : N2VL<op24_23, 0b10, 0b00, 0b00, op11_7, op6, op4, IIC_VQUNAiD,
3440 OpcodeStr, !strconcat(Dt, "32"), v2i64, v2i32, OpNode>;
3444 // Neon 3-register vector operations.
3446 // First with only element sizes of 8, 16 and 32 bits:
3447 multiclass N3V_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
3448 InstrItinClass itinD16, InstrItinClass itinD32,
3449 InstrItinClass itinQ16, InstrItinClass itinQ32,
3450 string OpcodeStr, string Dt,
3451 SDNode OpNode, bit Commutable = 0> {
3452 // 64-bit vector types.
3453 def v8i8 : N3VD<op24, op23, 0b00, op11_8, op4, itinD16,
3454 OpcodeStr, !strconcat(Dt, "8"),
3455 v8i8, v8i8, OpNode, Commutable>;
3456 def v4i16 : N3VD<op24, op23, 0b01, op11_8, op4, itinD16,
3457 OpcodeStr, !strconcat(Dt, "16"),
3458 v4i16, v4i16, OpNode, Commutable>;
3459 def v2i32 : N3VD<op24, op23, 0b10, op11_8, op4, itinD32,
3460 OpcodeStr, !strconcat(Dt, "32"),
3461 v2i32, v2i32, OpNode, Commutable>;
3463 // 128-bit vector types.
3464 def v16i8 : N3VQ<op24, op23, 0b00, op11_8, op4, itinQ16,
3465 OpcodeStr, !strconcat(Dt, "8"),
3466 v16i8, v16i8, OpNode, Commutable>;
3467 def v8i16 : N3VQ<op24, op23, 0b01, op11_8, op4, itinQ16,
3468 OpcodeStr, !strconcat(Dt, "16"),
3469 v8i16, v8i16, OpNode, Commutable>;
3470 def v4i32 : N3VQ<op24, op23, 0b10, op11_8, op4, itinQ32,
3471 OpcodeStr, !strconcat(Dt, "32"),
3472 v4i32, v4i32, OpNode, Commutable>;
3475 multiclass N3VSL_HS<bits<4> op11_8, string OpcodeStr, SDNode ShOp> {
3476 def v4i16 : N3VDSL16<0b01, op11_8, OpcodeStr, "i16", v4i16, ShOp>;
3477 def v2i32 : N3VDSL<0b10, op11_8, IIC_VMULi32D, OpcodeStr, "i32", v2i32, ShOp>;
3478 def v8i16 : N3VQSL16<0b01, op11_8, OpcodeStr, "i16", v8i16, v4i16, ShOp>;
3479 def v4i32 : N3VQSL<0b10, op11_8, IIC_VMULi32Q, OpcodeStr, "i32",
3480 v4i32, v2i32, ShOp>;
3483 // ....then also with element size 64 bits:
3484 multiclass N3V_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
3485 InstrItinClass itinD, InstrItinClass itinQ,
3486 string OpcodeStr, string Dt,
3487 SDNode OpNode, bit Commutable = 0>
3488 : N3V_QHS<op24, op23, op11_8, op4, itinD, itinD, itinQ, itinQ,
3489 OpcodeStr, Dt, OpNode, Commutable> {
3490 def v1i64 : N3VD<op24, op23, 0b11, op11_8, op4, itinD,
3491 OpcodeStr, !strconcat(Dt, "64"),
3492 v1i64, v1i64, OpNode, Commutable>;
3493 def v2i64 : N3VQ<op24, op23, 0b11, op11_8, op4, itinQ,
3494 OpcodeStr, !strconcat(Dt, "64"),
3495 v2i64, v2i64, OpNode, Commutable>;
3499 // Neon 3-register vector intrinsics.
3501 // First with only element sizes of 16 and 32 bits:
3502 multiclass N3VInt_HS<bit op24, bit op23, bits<4> op11_8, bit op4, Format f,
3503 InstrItinClass itinD16, InstrItinClass itinD32,
3504 InstrItinClass itinQ16, InstrItinClass itinQ32,
3505 string OpcodeStr, string Dt,
3506 SDPatternOperator IntOp, bit Commutable = 0> {
3507 // 64-bit vector types.
3508 def v4i16 : N3VDInt<op24, op23, 0b01, op11_8, op4, f, itinD16,
3509 OpcodeStr, !strconcat(Dt, "16"),
3510 v4i16, v4i16, IntOp, Commutable>;
3511 def v2i32 : N3VDInt<op24, op23, 0b10, op11_8, op4, f, itinD32,
3512 OpcodeStr, !strconcat(Dt, "32"),
3513 v2i32, v2i32, IntOp, Commutable>;
3515 // 128-bit vector types.
3516 def v8i16 : N3VQInt<op24, op23, 0b01, op11_8, op4, f, itinQ16,
3517 OpcodeStr, !strconcat(Dt, "16"),
3518 v8i16, v8i16, IntOp, Commutable>;
3519 def v4i32 : N3VQInt<op24, op23, 0b10, op11_8, op4, f, itinQ32,
3520 OpcodeStr, !strconcat(Dt, "32"),
3521 v4i32, v4i32, IntOp, Commutable>;
3523 multiclass N3VInt_HSSh<bit op24, bit op23, bits<4> op11_8, bit op4, Format f,
3524 InstrItinClass itinD16, InstrItinClass itinD32,
3525 InstrItinClass itinQ16, InstrItinClass itinQ32,
3526 string OpcodeStr, string Dt,
3527 SDPatternOperator IntOp> {
3528 // 64-bit vector types.
3529 def v4i16 : N3VDIntSh<op24, op23, 0b01, op11_8, op4, f, itinD16,
3530 OpcodeStr, !strconcat(Dt, "16"),
3531 v4i16, v4i16, IntOp>;
3532 def v2i32 : N3VDIntSh<op24, op23, 0b10, op11_8, op4, f, itinD32,
3533 OpcodeStr, !strconcat(Dt, "32"),
3534 v2i32, v2i32, IntOp>;
3536 // 128-bit vector types.
3537 def v8i16 : N3VQIntSh<op24, op23, 0b01, op11_8, op4, f, itinQ16,
3538 OpcodeStr, !strconcat(Dt, "16"),
3539 v8i16, v8i16, IntOp>;
3540 def v4i32 : N3VQIntSh<op24, op23, 0b10, op11_8, op4, f, itinQ32,
3541 OpcodeStr, !strconcat(Dt, "32"),
3542 v4i32, v4i32, IntOp>;
3545 multiclass N3VIntSL_HS<bits<4> op11_8,
3546 InstrItinClass itinD16, InstrItinClass itinD32,
3547 InstrItinClass itinQ16, InstrItinClass itinQ32,
3548 string OpcodeStr, string Dt, SDPatternOperator IntOp> {
3549 def v4i16 : N3VDIntSL16<0b01, op11_8, itinD16,
3550 OpcodeStr, !strconcat(Dt, "16"), v4i16, IntOp>;
3551 def v2i32 : N3VDIntSL<0b10, op11_8, itinD32,
3552 OpcodeStr, !strconcat(Dt, "32"), v2i32, IntOp>;
3553 def v8i16 : N3VQIntSL16<0b01, op11_8, itinQ16,
3554 OpcodeStr, !strconcat(Dt, "16"), v8i16, v4i16, IntOp>;
3555 def v4i32 : N3VQIntSL<0b10, op11_8, itinQ32,
3556 OpcodeStr, !strconcat(Dt, "32"), v4i32, v2i32, IntOp>;
3559 // ....then also with element size of 8 bits:
3560 multiclass N3VInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op4, Format f,
3561 InstrItinClass itinD16, InstrItinClass itinD32,
3562 InstrItinClass itinQ16, InstrItinClass itinQ32,
3563 string OpcodeStr, string Dt,
3564 SDPatternOperator IntOp, bit Commutable = 0>
3565 : N3VInt_HS<op24, op23, op11_8, op4, f, itinD16, itinD32, itinQ16, itinQ32,
3566 OpcodeStr, Dt, IntOp, Commutable> {
3567 def v8i8 : N3VDInt<op24, op23, 0b00, op11_8, op4, f, itinD16,
3568 OpcodeStr, !strconcat(Dt, "8"),
3569 v8i8, v8i8, IntOp, Commutable>;
3570 def v16i8 : N3VQInt<op24, op23, 0b00, op11_8, op4, f, itinQ16,
3571 OpcodeStr, !strconcat(Dt, "8"),
3572 v16i8, v16i8, IntOp, Commutable>;
3574 multiclass N3VInt_QHSSh<bit op24, bit op23, bits<4> op11_8, bit op4, Format f,
3575 InstrItinClass itinD16, InstrItinClass itinD32,
3576 InstrItinClass itinQ16, InstrItinClass itinQ32,
3577 string OpcodeStr, string Dt,
3578 SDPatternOperator IntOp>
3579 : N3VInt_HSSh<op24, op23, op11_8, op4, f, itinD16, itinD32, itinQ16, itinQ32,
3580 OpcodeStr, Dt, IntOp> {
3581 def v8i8 : N3VDIntSh<op24, op23, 0b00, op11_8, op4, f, itinD16,
3582 OpcodeStr, !strconcat(Dt, "8"),
3584 def v16i8 : N3VQIntSh<op24, op23, 0b00, op11_8, op4, f, itinQ16,
3585 OpcodeStr, !strconcat(Dt, "8"),
3586 v16i8, v16i8, IntOp>;
3590 // ....then also with element size of 64 bits:
3591 multiclass N3VInt_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4, Format f,
3592 InstrItinClass itinD16, InstrItinClass itinD32,
3593 InstrItinClass itinQ16, InstrItinClass itinQ32,
3594 string OpcodeStr, string Dt,
3595 SDPatternOperator IntOp, bit Commutable = 0>
3596 : N3VInt_QHS<op24, op23, op11_8, op4, f, itinD16, itinD32, itinQ16, itinQ32,
3597 OpcodeStr, Dt, IntOp, Commutable> {
3598 def v1i64 : N3VDInt<op24, op23, 0b11, op11_8, op4, f, itinD32,
3599 OpcodeStr, !strconcat(Dt, "64"),
3600 v1i64, v1i64, IntOp, Commutable>;
3601 def v2i64 : N3VQInt<op24, op23, 0b11, op11_8, op4, f, itinQ32,
3602 OpcodeStr, !strconcat(Dt, "64"),
3603 v2i64, v2i64, IntOp, Commutable>;
3605 multiclass N3VInt_QHSDSh<bit op24, bit op23, bits<4> op11_8, bit op4, Format f,
3606 InstrItinClass itinD16, InstrItinClass itinD32,
3607 InstrItinClass itinQ16, InstrItinClass itinQ32,
3608 string OpcodeStr, string Dt,
3609 SDPatternOperator IntOp>
3610 : N3VInt_QHSSh<op24, op23, op11_8, op4, f, itinD16, itinD32, itinQ16, itinQ32,
3611 OpcodeStr, Dt, IntOp> {
3612 def v1i64 : N3VDIntSh<op24, op23, 0b11, op11_8, op4, f, itinD32,
3613 OpcodeStr, !strconcat(Dt, "64"),
3614 v1i64, v1i64, IntOp>;
3615 def v2i64 : N3VQIntSh<op24, op23, 0b11, op11_8, op4, f, itinQ32,
3616 OpcodeStr, !strconcat(Dt, "64"),
3617 v2i64, v2i64, IntOp>;
3620 // Neon Narrowing 3-register vector intrinsics,
3621 // source operand element sizes of 16, 32 and 64 bits:
3622 multiclass N3VNInt_HSD<bit op24, bit op23, bits<4> op11_8, bit op4,
3623 string OpcodeStr, string Dt,
3624 SDPatternOperator IntOp, bit Commutable = 0> {
3625 def v8i8 : N3VNInt<op24, op23, 0b00, op11_8, op4,
3626 OpcodeStr, !strconcat(Dt, "16"),
3627 v8i8, v8i16, IntOp, Commutable>;
3628 def v4i16 : N3VNInt<op24, op23, 0b01, op11_8, op4,
3629 OpcodeStr, !strconcat(Dt, "32"),
3630 v4i16, v4i32, IntOp, Commutable>;
3631 def v2i32 : N3VNInt<op24, op23, 0b10, op11_8, op4,
3632 OpcodeStr, !strconcat(Dt, "64"),
3633 v2i32, v2i64, IntOp, Commutable>;
3637 // Neon Long 3-register vector operations.
3639 multiclass N3VL_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
3640 InstrItinClass itin16, InstrItinClass itin32,
3641 string OpcodeStr, string Dt,
3642 SDNode OpNode, bit Commutable = 0> {
3643 def v8i16 : N3VL<op24, op23, 0b00, op11_8, op4, itin16,
3644 OpcodeStr, !strconcat(Dt, "8"),
3645 v8i16, v8i8, OpNode, Commutable>;
3646 def v4i32 : N3VL<op24, op23, 0b01, op11_8, op4, itin16,
3647 OpcodeStr, !strconcat(Dt, "16"),
3648 v4i32, v4i16, OpNode, Commutable>;
3649 def v2i64 : N3VL<op24, op23, 0b10, op11_8, op4, itin32,
3650 OpcodeStr, !strconcat(Dt, "32"),
3651 v2i64, v2i32, OpNode, Commutable>;
3654 multiclass N3VLSL_HS<bit op24, bits<4> op11_8,
3655 InstrItinClass itin, string OpcodeStr, string Dt,
3657 def v4i16 : N3VLSL16<op24, 0b01, op11_8, itin, OpcodeStr,
3658 !strconcat(Dt, "16"), v4i32, v4i16, OpNode>;
3659 def v2i32 : N3VLSL<op24, 0b10, op11_8, itin, OpcodeStr,
3660 !strconcat(Dt, "32"), v2i64, v2i32, OpNode>;
3663 multiclass N3VLExt_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
3664 InstrItinClass itin16, InstrItinClass itin32,
3665 string OpcodeStr, string Dt,
3666 SDNode OpNode, SDNode ExtOp, bit Commutable = 0> {
3667 def v8i16 : N3VLExt<op24, op23, 0b00, op11_8, op4, itin16,
3668 OpcodeStr, !strconcat(Dt, "8"),
3669 v8i16, v8i8, OpNode, ExtOp, Commutable>;
3670 def v4i32 : N3VLExt<op24, op23, 0b01, op11_8, op4, itin16,
3671 OpcodeStr, !strconcat(Dt, "16"),
3672 v4i32, v4i16, OpNode, ExtOp, Commutable>;
3673 def v2i64 : N3VLExt<op24, op23, 0b10, op11_8, op4, itin32,
3674 OpcodeStr, !strconcat(Dt, "32"),
3675 v2i64, v2i32, OpNode, ExtOp, Commutable>;
3678 // Neon Long 3-register vector intrinsics.
3680 // First with only element sizes of 16 and 32 bits:
3681 multiclass N3VLInt_HS<bit op24, bit op23, bits<4> op11_8, bit op4,
3682 InstrItinClass itin16, InstrItinClass itin32,
3683 string OpcodeStr, string Dt,
3684 SDPatternOperator IntOp, bit Commutable = 0> {
3685 def v4i32 : N3VLInt<op24, op23, 0b01, op11_8, op4, itin16,
3686 OpcodeStr, !strconcat(Dt, "16"),
3687 v4i32, v4i16, IntOp, Commutable>;
3688 def v2i64 : N3VLInt<op24, op23, 0b10, op11_8, op4, itin32,
3689 OpcodeStr, !strconcat(Dt, "32"),
3690 v2i64, v2i32, IntOp, Commutable>;
3693 multiclass N3VLIntSL_HS<bit op24, bits<4> op11_8,
3694 InstrItinClass itin, string OpcodeStr, string Dt,
3695 SDPatternOperator IntOp> {
3696 def v4i16 : N3VLIntSL16<op24, 0b01, op11_8, itin,
3697 OpcodeStr, !strconcat(Dt, "16"), v4i32, v4i16, IntOp>;
3698 def v2i32 : N3VLIntSL<op24, 0b10, op11_8, itin,
3699 OpcodeStr, !strconcat(Dt, "32"), v2i64, v2i32, IntOp>;
3702 // ....then also with element size of 8 bits:
3703 multiclass N3VLInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
3704 InstrItinClass itin16, InstrItinClass itin32,
3705 string OpcodeStr, string Dt,
3706 SDPatternOperator IntOp, bit Commutable = 0>
3707 : N3VLInt_HS<op24, op23, op11_8, op4, itin16, itin32, OpcodeStr, Dt,
3708 IntOp, Commutable> {
3709 def v8i16 : N3VLInt<op24, op23, 0b00, op11_8, op4, itin16,
3710 OpcodeStr, !strconcat(Dt, "8"),
3711 v8i16, v8i8, IntOp, Commutable>;
3714 // ....with explicit extend (VABDL).
3715 multiclass N3VLIntExt_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
3716 InstrItinClass itin, string OpcodeStr, string Dt,
3717 SDPatternOperator IntOp, SDNode ExtOp, bit Commutable = 0> {
3718 def v8i16 : N3VLIntExt<op24, op23, 0b00, op11_8, op4, itin,
3719 OpcodeStr, !strconcat(Dt, "8"),
3720 v8i16, v8i8, IntOp, ExtOp, Commutable>;
3721 def v4i32 : N3VLIntExt<op24, op23, 0b01, op11_8, op4, itin,
3722 OpcodeStr, !strconcat(Dt, "16"),
3723 v4i32, v4i16, IntOp, ExtOp, Commutable>;
3724 def v2i64 : N3VLIntExt<op24, op23, 0b10, op11_8, op4, itin,
3725 OpcodeStr, !strconcat(Dt, "32"),
3726 v2i64, v2i32, IntOp, ExtOp, Commutable>;
3730 // Neon Wide 3-register vector intrinsics,
3731 // source operand element sizes of 8, 16 and 32 bits:
3732 multiclass N3VW_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
3733 string OpcodeStr, string Dt,
3734 SDNode OpNode, SDNode ExtOp, bit Commutable = 0> {
3735 def v8i16 : N3VW<op24, op23, 0b00, op11_8, op4,
3736 OpcodeStr, !strconcat(Dt, "8"),
3737 v8i16, v8i8, OpNode, ExtOp, Commutable>;
3738 def v4i32 : N3VW<op24, op23, 0b01, op11_8, op4,
3739 OpcodeStr, !strconcat(Dt, "16"),
3740 v4i32, v4i16, OpNode, ExtOp, Commutable>;
3741 def v2i64 : N3VW<op24, op23, 0b10, op11_8, op4,
3742 OpcodeStr, !strconcat(Dt, "32"),
3743 v2i64, v2i32, OpNode, ExtOp, Commutable>;
3747 // Neon Multiply-Op vector operations,
3748 // element sizes of 8, 16 and 32 bits:
3749 multiclass N3VMulOp_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
3750 InstrItinClass itinD16, InstrItinClass itinD32,
3751 InstrItinClass itinQ16, InstrItinClass itinQ32,
3752 string OpcodeStr, string Dt, SDNode OpNode> {
3753 // 64-bit vector types.
3754 def v8i8 : N3VDMulOp<op24, op23, 0b00, op11_8, op4, itinD16,
3755 OpcodeStr, !strconcat(Dt, "8"), v8i8, mul, OpNode>;
3756 def v4i16 : N3VDMulOp<op24, op23, 0b01, op11_8, op4, itinD16,
3757 OpcodeStr, !strconcat(Dt, "16"), v4i16, mul, OpNode>;
3758 def v2i32 : N3VDMulOp<op24, op23, 0b10, op11_8, op4, itinD32,
3759 OpcodeStr, !strconcat(Dt, "32"), v2i32, mul, OpNode>;
3761 // 128-bit vector types.
3762 def v16i8 : N3VQMulOp<op24, op23, 0b00, op11_8, op4, itinQ16,
3763 OpcodeStr, !strconcat(Dt, "8"), v16i8, mul, OpNode>;
3764 def v8i16 : N3VQMulOp<op24, op23, 0b01, op11_8, op4, itinQ16,
3765 OpcodeStr, !strconcat(Dt, "16"), v8i16, mul, OpNode>;
3766 def v4i32 : N3VQMulOp<op24, op23, 0b10, op11_8, op4, itinQ32,
3767 OpcodeStr, !strconcat(Dt, "32"), v4i32, mul, OpNode>;
3770 multiclass N3VMulOpSL_HS<bits<4> op11_8,
3771 InstrItinClass itinD16, InstrItinClass itinD32,
3772 InstrItinClass itinQ16, InstrItinClass itinQ32,
3773 string OpcodeStr, string Dt, SDPatternOperator ShOp> {
3774 def v4i16 : N3VDMulOpSL16<0b01, op11_8, itinD16,
3775 OpcodeStr, !strconcat(Dt, "16"), v4i16, mul, ShOp>;
3776 def v2i32 : N3VDMulOpSL<0b10, op11_8, itinD32,
3777 OpcodeStr, !strconcat(Dt, "32"), v2i32, mul, ShOp>;
3778 def v8i16 : N3VQMulOpSL16<0b01, op11_8, itinQ16,
3779 OpcodeStr, !strconcat(Dt, "16"), v8i16, v4i16,
3781 def v4i32 : N3VQMulOpSL<0b10, op11_8, itinQ32,
3782 OpcodeStr, !strconcat(Dt, "32"), v4i32, v2i32,
3786 // Neon Intrinsic-Op vector operations,
3787 // element sizes of 8, 16 and 32 bits:
3788 multiclass N3VIntOp_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
3789 InstrItinClass itinD, InstrItinClass itinQ,
3790 string OpcodeStr, string Dt, SDPatternOperator IntOp,
3792 // 64-bit vector types.
3793 def v8i8 : N3VDIntOp<op24, op23, 0b00, op11_8, op4, itinD,
3794 OpcodeStr, !strconcat(Dt, "8"), v8i8, IntOp, OpNode>;
3795 def v4i16 : N3VDIntOp<op24, op23, 0b01, op11_8, op4, itinD,
3796 OpcodeStr, !strconcat(Dt, "16"), v4i16, IntOp, OpNode>;
3797 def v2i32 : N3VDIntOp<op24, op23, 0b10, op11_8, op4, itinD,
3798 OpcodeStr, !strconcat(Dt, "32"), v2i32, IntOp, OpNode>;
3800 // 128-bit vector types.
3801 def v16i8 : N3VQIntOp<op24, op23, 0b00, op11_8, op4, itinQ,
3802 OpcodeStr, !strconcat(Dt, "8"), v16i8, IntOp, OpNode>;
3803 def v8i16 : N3VQIntOp<op24, op23, 0b01, op11_8, op4, itinQ,
3804 OpcodeStr, !strconcat(Dt, "16"), v8i16, IntOp, OpNode>;
3805 def v4i32 : N3VQIntOp<op24, op23, 0b10, op11_8, op4, itinQ,
3806 OpcodeStr, !strconcat(Dt, "32"), v4i32, IntOp, OpNode>;
3809 // Neon 3-argument intrinsics,
3810 // element sizes of 16 and 32 bits:
3811 multiclass N3VInt3_HS<bit op24, bit op23, bits<4> op11_8, bit op4,
3812 InstrItinClass itinD16, InstrItinClass itinD32,
3813 InstrItinClass itinQ16, InstrItinClass itinQ32,
3814 string OpcodeStr, string Dt, SDPatternOperator IntOp> {
3815 // 64-bit vector types.
3816 def v4i16 : N3VDInt3<op24, op23, 0b01, op11_8, op4, itinD16,
3817 OpcodeStr, !strconcat(Dt, "16"), v4i16, v4i16, IntOp>;
3818 def v2i32 : N3VDInt3<op24, op23, 0b10, op11_8, op4, itinD32,
3819 OpcodeStr, !strconcat(Dt, "32"), v2i32, v2i32, IntOp>;
3821 // 128-bit vector types.
3822 def v8i16 : N3VQInt3<op24, op23, 0b01, op11_8, op4, itinQ16,
3823 OpcodeStr, !strconcat(Dt, "16"), v8i16, v8i16, IntOp>;
3824 def v4i32 : N3VQInt3<op24, op23, 0b10, op11_8, op4, itinQ32,
3825 OpcodeStr, !strconcat(Dt, "32"), v4i32, v4i32, IntOp>;
3828 // element sizes of 8, 16 and 32 bits:
3829 multiclass N3VInt3_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
3830 InstrItinClass itinD16, InstrItinClass itinD32,
3831 InstrItinClass itinQ16, InstrItinClass itinQ32,
3832 string OpcodeStr, string Dt, SDPatternOperator IntOp>
3833 :N3VInt3_HS <op24, op23, op11_8, op4, itinD16, itinD32,
3834 itinQ16, itinQ32, OpcodeStr, Dt, IntOp>{
3835 // 64-bit vector types.
3836 def v8i8 : N3VDInt3<op24, op23, 0b00, op11_8, op4, itinD16,
3837 OpcodeStr, !strconcat(Dt, "8"), v8i8, v8i8, IntOp>;
3838 // 128-bit vector types.
3839 def v16i8 : N3VQInt3<op24, op23, 0b00, op11_8, op4, itinQ16,
3840 OpcodeStr, !strconcat(Dt, "8"), v16i8, v16i8, IntOp>;
3843 // Neon Long Multiply-Op vector operations,
3844 // element sizes of 8, 16 and 32 bits:
3845 multiclass N3VLMulOp_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
3846 InstrItinClass itin16, InstrItinClass itin32,
3847 string OpcodeStr, string Dt, SDNode MulOp,
3849 def v8i16 : N3VLMulOp<op24, op23, 0b00, op11_8, op4, itin16, OpcodeStr,
3850 !strconcat(Dt, "8"), v8i16, v8i8, MulOp, OpNode>;
3851 def v4i32 : N3VLMulOp<op24, op23, 0b01, op11_8, op4, itin16, OpcodeStr,
3852 !strconcat(Dt, "16"), v4i32, v4i16, MulOp, OpNode>;
3853 def v2i64 : N3VLMulOp<op24, op23, 0b10, op11_8, op4, itin32, OpcodeStr,
3854 !strconcat(Dt, "32"), v2i64, v2i32, MulOp, OpNode>;
3857 multiclass N3VLMulOpSL_HS<bit op24, bits<4> op11_8, string OpcodeStr,
3858 string Dt, SDNode MulOp, SDNode OpNode> {
3859 def v4i16 : N3VLMulOpSL16<op24, 0b01, op11_8, IIC_VMACi16D, OpcodeStr,
3860 !strconcat(Dt,"16"), v4i32, v4i16, MulOp, OpNode>;
3861 def v2i32 : N3VLMulOpSL<op24, 0b10, op11_8, IIC_VMACi32D, OpcodeStr,
3862 !strconcat(Dt, "32"), v2i64, v2i32, MulOp, OpNode>;
3866 // Neon Long 3-argument intrinsics.
3868 // First with only element sizes of 16 and 32 bits:
3869 multiclass N3VLInt3_HS<bit op24, bit op23, bits<4> op11_8, bit op4,
3870 InstrItinClass itin16, InstrItinClass itin32,
3871 string OpcodeStr, string Dt, SDPatternOperator IntOp> {
3872 def v4i32 : N3VLInt3<op24, op23, 0b01, op11_8, op4, itin16,
3873 OpcodeStr, !strconcat(Dt, "16"), v4i32, v4i16, IntOp>;
3874 def v2i64 : N3VLInt3<op24, op23, 0b10, op11_8, op4, itin32,
3875 OpcodeStr, !strconcat(Dt, "32"), v2i64, v2i32, IntOp>;
3878 multiclass N3VLInt3SL_HS<bit op24, bits<4> op11_8,
3879 string OpcodeStr, string Dt, SDPatternOperator IntOp> {
3880 def v4i16 : N3VLInt3SL16<op24, 0b01, op11_8, IIC_VMACi16D,
3881 OpcodeStr, !strconcat(Dt,"16"), v4i32, v4i16, IntOp>;
3882 def v2i32 : N3VLInt3SL<op24, 0b10, op11_8, IIC_VMACi32D,
3883 OpcodeStr, !strconcat(Dt, "32"), v2i64, v2i32, IntOp>;
3886 // ....then also with element size of 8 bits:
3887 multiclass N3VLInt3_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
3888 InstrItinClass itin16, InstrItinClass itin32,
3889 string OpcodeStr, string Dt, SDPatternOperator IntOp>
3890 : N3VLInt3_HS<op24, op23, op11_8, op4, itin16, itin32, OpcodeStr, Dt, IntOp> {
3891 def v8i16 : N3VLInt3<op24, op23, 0b00, op11_8, op4, itin16,
3892 OpcodeStr, !strconcat(Dt, "8"), v8i16, v8i8, IntOp>;
3895 // ....with explicit extend (VABAL).
3896 multiclass N3VLIntExtOp_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
3897 InstrItinClass itin, string OpcodeStr, string Dt,
3898 SDPatternOperator IntOp, SDNode ExtOp, SDNode OpNode> {
3899 def v8i16 : N3VLIntExtOp<op24, op23, 0b00, op11_8, op4, itin,
3900 OpcodeStr, !strconcat(Dt, "8"), v8i16, v8i8,
3901 IntOp, ExtOp, OpNode>;
3902 def v4i32 : N3VLIntExtOp<op24, op23, 0b01, op11_8, op4, itin,
3903 OpcodeStr, !strconcat(Dt, "16"), v4i32, v4i16,
3904 IntOp, ExtOp, OpNode>;
3905 def v2i64 : N3VLIntExtOp<op24, op23, 0b10, op11_8, op4, itin,
3906 OpcodeStr, !strconcat(Dt, "32"), v2i64, v2i32,
3907 IntOp, ExtOp, OpNode>;
3911 // Neon Pairwise long 2-register intrinsics,
3912 // element sizes of 8, 16 and 32 bits:
3913 multiclass N2VPLInt_QHS<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
3914 bits<5> op11_7, bit op4,
3915 string OpcodeStr, string Dt, SDPatternOperator IntOp> {
3916 // 64-bit vector types.
3917 def v8i8 : N2VDPLInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
3918 OpcodeStr, !strconcat(Dt, "8"), v4i16, v8i8, IntOp>;
3919 def v4i16 : N2VDPLInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
3920 OpcodeStr, !strconcat(Dt, "16"), v2i32, v4i16, IntOp>;
3921 def v2i32 : N2VDPLInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
3922 OpcodeStr, !strconcat(Dt, "32"), v1i64, v2i32, IntOp>;
3924 // 128-bit vector types.
3925 def v16i8 : N2VQPLInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
3926 OpcodeStr, !strconcat(Dt, "8"), v8i16, v16i8, IntOp>;
3927 def v8i16 : N2VQPLInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
3928 OpcodeStr, !strconcat(Dt, "16"), v4i32, v8i16, IntOp>;
3929 def v4i32 : N2VQPLInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
3930 OpcodeStr, !strconcat(Dt, "32"), v2i64, v4i32, IntOp>;
3934 // Neon Pairwise long 2-register accumulate intrinsics,
3935 // element sizes of 8, 16 and 32 bits:
3936 multiclass N2VPLInt2_QHS<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
3937 bits<5> op11_7, bit op4,
3938 string OpcodeStr, string Dt, SDPatternOperator IntOp> {
3939 // 64-bit vector types.
3940 def v8i8 : N2VDPLInt2<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
3941 OpcodeStr, !strconcat(Dt, "8"), v4i16, v8i8, IntOp>;
3942 def v4i16 : N2VDPLInt2<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
3943 OpcodeStr, !strconcat(Dt, "16"), v2i32, v4i16, IntOp>;
3944 def v2i32 : N2VDPLInt2<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
3945 OpcodeStr, !strconcat(Dt, "32"), v1i64, v2i32, IntOp>;
3947 // 128-bit vector types.
3948 def v16i8 : N2VQPLInt2<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
3949 OpcodeStr, !strconcat(Dt, "8"), v8i16, v16i8, IntOp>;
3950 def v8i16 : N2VQPLInt2<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
3951 OpcodeStr, !strconcat(Dt, "16"), v4i32, v8i16, IntOp>;
3952 def v4i32 : N2VQPLInt2<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
3953 OpcodeStr, !strconcat(Dt, "32"), v2i64, v4i32, IntOp>;
3957 // Neon 2-register vector shift by immediate,
3958 // with f of either N2RegVShLFrm or N2RegVShRFrm
3959 // element sizes of 8, 16, 32 and 64 bits:
3960 multiclass N2VShL_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
3961 InstrItinClass itin, string OpcodeStr, string Dt,
3963 // 64-bit vector types.
3964 def v8i8 : N2VDSh<op24, op23, op11_8, 0, op4, N2RegVShLFrm, itin, i32imm,
3965 OpcodeStr, !strconcat(Dt, "8"), v8i8, OpNode> {
3966 let Inst{21-19} = 0b001; // imm6 = 001xxx
3968 def v4i16 : N2VDSh<op24, op23, op11_8, 0, op4, N2RegVShLFrm, itin, i32imm,
3969 OpcodeStr, !strconcat(Dt, "16"), v4i16, OpNode> {
3970 let Inst{21-20} = 0b01; // imm6 = 01xxxx
3972 def v2i32 : N2VDSh<op24, op23, op11_8, 0, op4, N2RegVShLFrm, itin, i32imm,
3973 OpcodeStr, !strconcat(Dt, "32"), v2i32, OpNode> {
3974 let Inst{21} = 0b1; // imm6 = 1xxxxx
3976 def v1i64 : N2VDSh<op24, op23, op11_8, 1, op4, N2RegVShLFrm, itin, i32imm,
3977 OpcodeStr, !strconcat(Dt, "64"), v1i64, OpNode>;
3980 // 128-bit vector types.
3981 def v16i8 : N2VQSh<op24, op23, op11_8, 0, op4, N2RegVShLFrm, itin, i32imm,
3982 OpcodeStr, !strconcat(Dt, "8"), v16i8, OpNode> {
3983 let Inst{21-19} = 0b001; // imm6 = 001xxx
3985 def v8i16 : N2VQSh<op24, op23, op11_8, 0, op4, N2RegVShLFrm, itin, i32imm,
3986 OpcodeStr, !strconcat(Dt, "16"), v8i16, OpNode> {
3987 let Inst{21-20} = 0b01; // imm6 = 01xxxx
3989 def v4i32 : N2VQSh<op24, op23, op11_8, 0, op4, N2RegVShLFrm, itin, i32imm,
3990 OpcodeStr, !strconcat(Dt, "32"), v4i32, OpNode> {
3991 let Inst{21} = 0b1; // imm6 = 1xxxxx
3993 def v2i64 : N2VQSh<op24, op23, op11_8, 1, op4, N2RegVShLFrm, itin, i32imm,
3994 OpcodeStr, !strconcat(Dt, "64"), v2i64, OpNode>;
3997 multiclass N2VShR_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
3998 InstrItinClass itin, string OpcodeStr, string Dt,
3999 string baseOpc, SDNode OpNode> {
4000 // 64-bit vector types.
4001 def v8i8 : N2VDSh<op24, op23, op11_8, 0, op4, N2RegVShRFrm, itin, shr_imm8,
4002 OpcodeStr, !strconcat(Dt, "8"), v8i8, OpNode> {
4003 let Inst{21-19} = 0b001; // imm6 = 001xxx
4005 def v4i16 : N2VDSh<op24, op23, op11_8, 0, op4, N2RegVShRFrm, itin, shr_imm16,
4006 OpcodeStr, !strconcat(Dt, "16"), v4i16, OpNode> {
4007 let Inst{21-20} = 0b01; // imm6 = 01xxxx
4009 def v2i32 : N2VDSh<op24, op23, op11_8, 0, op4, N2RegVShRFrm, itin, shr_imm32,
4010 OpcodeStr, !strconcat(Dt, "32"), v2i32, OpNode> {
4011 let Inst{21} = 0b1; // imm6 = 1xxxxx
4013 def v1i64 : N2VDSh<op24, op23, op11_8, 1, op4, N2RegVShRFrm, itin, shr_imm64,
4014 OpcodeStr, !strconcat(Dt, "64"), v1i64, OpNode>;
4017 // 128-bit vector types.
4018 def v16i8 : N2VQSh<op24, op23, op11_8, 0, op4, N2RegVShRFrm, itin, shr_imm8,
4019 OpcodeStr, !strconcat(Dt, "8"), v16i8, OpNode> {
4020 let Inst{21-19} = 0b001; // imm6 = 001xxx
4022 def v8i16 : N2VQSh<op24, op23, op11_8, 0, op4, N2RegVShRFrm, itin, shr_imm16,
4023 OpcodeStr, !strconcat(Dt, "16"), v8i16, OpNode> {
4024 let Inst{21-20} = 0b01; // imm6 = 01xxxx
4026 def v4i32 : N2VQSh<op24, op23, op11_8, 0, op4, N2RegVShRFrm, itin, shr_imm32,
4027 OpcodeStr, !strconcat(Dt, "32"), v4i32, OpNode> {
4028 let Inst{21} = 0b1; // imm6 = 1xxxxx
4030 def v2i64 : N2VQSh<op24, op23, op11_8, 1, op4, N2RegVShRFrm, itin, shr_imm64,
4031 OpcodeStr, !strconcat(Dt, "64"), v2i64, OpNode>;
4035 // Neon Shift-Accumulate vector operations,
4036 // element sizes of 8, 16, 32 and 64 bits:
4037 multiclass N2VShAdd_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
4038 string OpcodeStr, string Dt, SDNode ShOp> {
4039 // 64-bit vector types.
4040 def v8i8 : N2VDShAdd<op24, op23, op11_8, 0, op4, shr_imm8,
4041 OpcodeStr, !strconcat(Dt, "8"), v8i8, ShOp> {
4042 let Inst{21-19} = 0b001; // imm6 = 001xxx
4044 def v4i16 : N2VDShAdd<op24, op23, op11_8, 0, op4, shr_imm16,
4045 OpcodeStr, !strconcat(Dt, "16"), v4i16, ShOp> {
4046 let Inst{21-20} = 0b01; // imm6 = 01xxxx
4048 def v2i32 : N2VDShAdd<op24, op23, op11_8, 0, op4, shr_imm32,
4049 OpcodeStr, !strconcat(Dt, "32"), v2i32, ShOp> {
4050 let Inst{21} = 0b1; // imm6 = 1xxxxx
4052 def v1i64 : N2VDShAdd<op24, op23, op11_8, 1, op4, shr_imm64,
4053 OpcodeStr, !strconcat(Dt, "64"), v1i64, ShOp>;
4056 // 128-bit vector types.
4057 def v16i8 : N2VQShAdd<op24, op23, op11_8, 0, op4, shr_imm8,
4058 OpcodeStr, !strconcat(Dt, "8"), v16i8, ShOp> {
4059 let Inst{21-19} = 0b001; // imm6 = 001xxx
4061 def v8i16 : N2VQShAdd<op24, op23, op11_8, 0, op4, shr_imm16,
4062 OpcodeStr, !strconcat(Dt, "16"), v8i16, ShOp> {
4063 let Inst{21-20} = 0b01; // imm6 = 01xxxx
4065 def v4i32 : N2VQShAdd<op24, op23, op11_8, 0, op4, shr_imm32,
4066 OpcodeStr, !strconcat(Dt, "32"), v4i32, ShOp> {
4067 let Inst{21} = 0b1; // imm6 = 1xxxxx
4069 def v2i64 : N2VQShAdd<op24, op23, op11_8, 1, op4, shr_imm64,
4070 OpcodeStr, !strconcat(Dt, "64"), v2i64, ShOp>;
4074 // Neon Shift-Insert vector operations,
4075 // with f of either N2RegVShLFrm or N2RegVShRFrm
4076 // element sizes of 8, 16, 32 and 64 bits:
4077 multiclass N2VShInsL_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
4079 // 64-bit vector types.
4080 def v8i8 : N2VDShIns<op24, op23, op11_8, 0, op4, i32imm,
4081 N2RegVShLFrm, OpcodeStr, "8", v8i8, NEONvsliImm> {
4082 let Inst{21-19} = 0b001; // imm6 = 001xxx
4084 def v4i16 : N2VDShIns<op24, op23, op11_8, 0, op4, i32imm,
4085 N2RegVShLFrm, OpcodeStr, "16", v4i16, NEONvsliImm> {
4086 let Inst{21-20} = 0b01; // imm6 = 01xxxx
4088 def v2i32 : N2VDShIns<op24, op23, op11_8, 0, op4, i32imm,
4089 N2RegVShLFrm, OpcodeStr, "32", v2i32, NEONvsliImm> {
4090 let Inst{21} = 0b1; // imm6 = 1xxxxx
4092 def v1i64 : N2VDShIns<op24, op23, op11_8, 1, op4, i32imm,
4093 N2RegVShLFrm, OpcodeStr, "64", v1i64, NEONvsliImm>;
4096 // 128-bit vector types.
4097 def v16i8 : N2VQShIns<op24, op23, op11_8, 0, op4, i32imm,
4098 N2RegVShLFrm, OpcodeStr, "8", v16i8, NEONvsliImm> {
4099 let Inst{21-19} = 0b001; // imm6 = 001xxx
4101 def v8i16 : N2VQShIns<op24, op23, op11_8, 0, op4, i32imm,
4102 N2RegVShLFrm, OpcodeStr, "16", v8i16, NEONvsliImm> {
4103 let Inst{21-20} = 0b01; // imm6 = 01xxxx
4105 def v4i32 : N2VQShIns<op24, op23, op11_8, 0, op4, i32imm,
4106 N2RegVShLFrm, OpcodeStr, "32", v4i32, NEONvsliImm> {
4107 let Inst{21} = 0b1; // imm6 = 1xxxxx
4109 def v2i64 : N2VQShIns<op24, op23, op11_8, 1, op4, i32imm,
4110 N2RegVShLFrm, OpcodeStr, "64", v2i64, NEONvsliImm>;
4113 multiclass N2VShInsR_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
4115 // 64-bit vector types.
4116 def v8i8 : N2VDShIns<op24, op23, op11_8, 0, op4, shr_imm8,
4117 N2RegVShRFrm, OpcodeStr, "8", v8i8, NEONvsriImm> {
4118 let Inst{21-19} = 0b001; // imm6 = 001xxx
4120 def v4i16 : N2VDShIns<op24, op23, op11_8, 0, op4, shr_imm16,
4121 N2RegVShRFrm, OpcodeStr, "16", v4i16, NEONvsriImm> {
4122 let Inst{21-20} = 0b01; // imm6 = 01xxxx
4124 def v2i32 : N2VDShIns<op24, op23, op11_8, 0, op4, shr_imm32,
4125 N2RegVShRFrm, OpcodeStr, "32", v2i32, NEONvsriImm> {
4126 let Inst{21} = 0b1; // imm6 = 1xxxxx
4128 def v1i64 : N2VDShIns<op24, op23, op11_8, 1, op4, shr_imm64,
4129 N2RegVShRFrm, OpcodeStr, "64", v1i64, NEONvsriImm>;
4132 // 128-bit vector types.
4133 def v16i8 : N2VQShIns<op24, op23, op11_8, 0, op4, shr_imm8,
4134 N2RegVShRFrm, OpcodeStr, "8", v16i8, NEONvsriImm> {
4135 let Inst{21-19} = 0b001; // imm6 = 001xxx
4137 def v8i16 : N2VQShIns<op24, op23, op11_8, 0, op4, shr_imm16,
4138 N2RegVShRFrm, OpcodeStr, "16", v8i16, NEONvsriImm> {
4139 let Inst{21-20} = 0b01; // imm6 = 01xxxx
4141 def v4i32 : N2VQShIns<op24, op23, op11_8, 0, op4, shr_imm32,
4142 N2RegVShRFrm, OpcodeStr, "32", v4i32, NEONvsriImm> {
4143 let Inst{21} = 0b1; // imm6 = 1xxxxx
4145 def v2i64 : N2VQShIns<op24, op23, op11_8, 1, op4, shr_imm64,
4146 N2RegVShRFrm, OpcodeStr, "64", v2i64, NEONvsriImm>;
4150 // Neon Shift Long operations,
4151 // element sizes of 8, 16, 32 bits:
4152 multiclass N2VLSh_QHS<bit op24, bit op23, bits<4> op11_8, bit op7, bit op6,
4153 bit op4, string OpcodeStr, string Dt,
4154 SDPatternOperator OpNode> {
4155 def v8i16 : N2VLSh<op24, op23, op11_8, op7, op6, op4,
4156 OpcodeStr, !strconcat(Dt, "8"), v8i16, v8i8, imm1_7, OpNode> {
4157 let Inst{21-19} = 0b001; // imm6 = 001xxx
4159 def v4i32 : N2VLSh<op24, op23, op11_8, op7, op6, op4,
4160 OpcodeStr, !strconcat(Dt, "16"), v4i32, v4i16, imm1_15, OpNode> {
4161 let Inst{21-20} = 0b01; // imm6 = 01xxxx
4163 def v2i64 : N2VLSh<op24, op23, op11_8, op7, op6, op4,
4164 OpcodeStr, !strconcat(Dt, "32"), v2i64, v2i32, imm1_31, OpNode> {
4165 let Inst{21} = 0b1; // imm6 = 1xxxxx
4169 // Neon Shift Narrow operations,
4170 // element sizes of 16, 32, 64 bits:
4171 multiclass N2VNSh_HSD<bit op24, bit op23, bits<4> op11_8, bit op7, bit op6,
4172 bit op4, InstrItinClass itin, string OpcodeStr, string Dt,
4173 SDPatternOperator OpNode> {
4174 def v8i8 : N2VNSh<op24, op23, op11_8, op7, op6, op4, itin,
4175 OpcodeStr, !strconcat(Dt, "16"),
4176 v8i8, v8i16, shr_imm8, OpNode> {
4177 let Inst{21-19} = 0b001; // imm6 = 001xxx
4179 def v4i16 : N2VNSh<op24, op23, op11_8, op7, op6, op4, itin,
4180 OpcodeStr, !strconcat(Dt, "32"),
4181 v4i16, v4i32, shr_imm16, OpNode> {
4182 let Inst{21-20} = 0b01; // imm6 = 01xxxx
4184 def v2i32 : N2VNSh<op24, op23, op11_8, op7, op6, op4, itin,
4185 OpcodeStr, !strconcat(Dt, "64"),
4186 v2i32, v2i64, shr_imm32, OpNode> {
4187 let Inst{21} = 0b1; // imm6 = 1xxxxx
4191 //===----------------------------------------------------------------------===//
4192 // Instruction Definitions.
4193 //===----------------------------------------------------------------------===//
4195 // Vector Add Operations.
4197 // VADD : Vector Add (integer and floating-point)
4198 defm VADD : N3V_QHSD<0, 0, 0b1000, 0, IIC_VBINiD, IIC_VBINiQ, "vadd", "i",
4200 def VADDfd : N3VD<0, 0, 0b00, 0b1101, 0, IIC_VBIND, "vadd", "f32",
4201 v2f32, v2f32, fadd, 1>;
4202 def VADDfq : N3VQ<0, 0, 0b00, 0b1101, 0, IIC_VBINQ, "vadd", "f32",
4203 v4f32, v4f32, fadd, 1>;
4204 def VADDhd : N3VD<0, 0, 0b01, 0b1101, 0, IIC_VBIND, "vadd", "f16",
4205 v4f16, v4f16, fadd, 1>,
4206 Requires<[HasNEON,HasFullFP16]>;
4207 def VADDhq : N3VQ<0, 0, 0b01, 0b1101, 0, IIC_VBINQ, "vadd", "f16",
4208 v8f16, v8f16, fadd, 1>,
4209 Requires<[HasNEON,HasFullFP16]>;
4210 // VADDL : Vector Add Long (Q = D + D)
4211 defm VADDLs : N3VLExt_QHS<0,1,0b0000,0, IIC_VSHLiD, IIC_VSHLiD,
4212 "vaddl", "s", add, sext, 1>;
4213 defm VADDLu : N3VLExt_QHS<1,1,0b0000,0, IIC_VSHLiD, IIC_VSHLiD,
4214 "vaddl", "u", add, zext, 1>;
4215 // VADDW : Vector Add Wide (Q = Q + D)
4216 defm VADDWs : N3VW_QHS<0,1,0b0001,0, "vaddw", "s", add, sext, 0>;
4217 defm VADDWu : N3VW_QHS<1,1,0b0001,0, "vaddw", "u", add, zext, 0>;
4218 // VHADD : Vector Halving Add
4219 defm VHADDs : N3VInt_QHS<0, 0, 0b0000, 0, N3RegFrm,
4220 IIC_VBINi4D, IIC_VBINi4D, IIC_VBINi4Q, IIC_VBINi4Q,
4221 "vhadd", "s", int_arm_neon_vhadds, 1>;
4222 defm VHADDu : N3VInt_QHS<1, 0, 0b0000, 0, N3RegFrm,
4223 IIC_VBINi4D, IIC_VBINi4D, IIC_VBINi4Q, IIC_VBINi4Q,
4224 "vhadd", "u", int_arm_neon_vhaddu, 1>;
4225 // VRHADD : Vector Rounding Halving Add
4226 defm VRHADDs : N3VInt_QHS<0, 0, 0b0001, 0, N3RegFrm,
4227 IIC_VBINi4D, IIC_VBINi4D, IIC_VBINi4Q, IIC_VBINi4Q,
4228 "vrhadd", "s", int_arm_neon_vrhadds, 1>;
4229 defm VRHADDu : N3VInt_QHS<1, 0, 0b0001, 0, N3RegFrm,
4230 IIC_VBINi4D, IIC_VBINi4D, IIC_VBINi4Q, IIC_VBINi4Q,
4231 "vrhadd", "u", int_arm_neon_vrhaddu, 1>;
4232 // VQADD : Vector Saturating Add
4233 defm VQADDs : N3VInt_QHSD<0, 0, 0b0000, 1, N3RegFrm,
4234 IIC_VBINi4D, IIC_VBINi4D, IIC_VBINi4Q, IIC_VBINi4Q,
4235 "vqadd", "s", saddsat, 1>;
4236 defm VQADDu : N3VInt_QHSD<1, 0, 0b0000, 1, N3RegFrm,
4237 IIC_VBINi4D, IIC_VBINi4D, IIC_VBINi4Q, IIC_VBINi4Q,
4238 "vqadd", "u", uaddsat, 1>;
4239 // VADDHN : Vector Add and Narrow Returning High Half (D = Q + Q)
4240 defm VADDHN : N3VNInt_HSD<0,1,0b0100,0, "vaddhn", "i", null_frag, 1>;
4241 // VRADDHN : Vector Rounding Add and Narrow Returning High Half (D = Q + Q)
4242 defm VRADDHN : N3VNInt_HSD<1,1,0b0100,0, "vraddhn", "i",
4243 int_arm_neon_vraddhn, 1>;
4245 let Predicates = [HasNEON] in {
4246 def : Pat<(v8i8 (trunc (ARMvshruImm (add (v8i16 QPR:$Vn), QPR:$Vm), 8))),
4247 (VADDHNv8i8 QPR:$Vn, QPR:$Vm)>;
4248 def : Pat<(v4i16 (trunc (ARMvshruImm (add (v4i32 QPR:$Vn), QPR:$Vm), 16))),
4249 (VADDHNv4i16 QPR:$Vn, QPR:$Vm)>;
4250 def : Pat<(v2i32 (trunc (ARMvshruImm (add (v2i64 QPR:$Vn), QPR:$Vm), 32))),
4251 (VADDHNv2i32 QPR:$Vn, QPR:$Vm)>;
4254 // Vector Multiply Operations.
4256 // VMUL : Vector Multiply (integer, polynomial and floating-point)
4257 defm VMUL : N3V_QHS<0, 0, 0b1001, 1, IIC_VMULi16D, IIC_VMULi32D,
4258 IIC_VMULi16Q, IIC_VMULi32Q, "vmul", "i", mul, 1>;
4259 def VMULpd : N3VDInt<1, 0, 0b00, 0b1001, 1, N3RegFrm, IIC_VMULi16D, "vmul",
4260 "p8", v8i8, v8i8, int_arm_neon_vmulp, 1>;
4261 def VMULpq : N3VQInt<1, 0, 0b00, 0b1001, 1, N3RegFrm, IIC_VMULi16Q, "vmul",
4262 "p8", v16i8, v16i8, int_arm_neon_vmulp, 1>;
4263 def VMULfd : N3VD<1, 0, 0b00, 0b1101, 1, IIC_VFMULD, "vmul", "f32",
4264 v2f32, v2f32, fmul, 1>;
4265 def VMULfq : N3VQ<1, 0, 0b00, 0b1101, 1, IIC_VFMULQ, "vmul", "f32",
4266 v4f32, v4f32, fmul, 1>;
4267 def VMULhd : N3VD<1, 0, 0b01, 0b1101, 1, IIC_VFMULD, "vmul", "f16",
4268 v4f16, v4f16, fmul, 1>,
4269 Requires<[HasNEON,HasFullFP16]>;
4270 def VMULhq : N3VQ<1, 0, 0b01, 0b1101, 1, IIC_VFMULQ, "vmul", "f16",
4271 v8f16, v8f16, fmul, 1>,
4272 Requires<[HasNEON,HasFullFP16]>;
4273 defm VMULsl : N3VSL_HS<0b1000, "vmul", mul>;
4274 def VMULslfd : N3VDSL<0b10, 0b1001, IIC_VBIND, "vmul", "f32", v2f32, fmul>;
4275 def VMULslfq : N3VQSL<0b10, 0b1001, IIC_VBINQ, "vmul", "f32", v4f32,
4277 def VMULslhd : N3VDSL16<0b01, 0b1001, "vmul", "f16", v4f16, fmul>,
4278 Requires<[HasNEON,HasFullFP16]>;
4279 def VMULslhq : N3VQSL16<0b01, 0b1001, "vmul", "f16", v8f16,
4281 Requires<[HasNEON,HasFullFP16]>;
4283 let Predicates = [HasNEON] in {
4284 def : Pat<(v8i16 (mul (v8i16 QPR:$src1),
4285 (v8i16 (ARMvduplane (v8i16 QPR:$src2), imm:$lane)))),
4286 (v8i16 (VMULslv8i16 (v8i16 QPR:$src1),
4287 (v4i16 (EXTRACT_SUBREG QPR:$src2,
4288 (DSubReg_i16_reg imm:$lane))),
4289 (SubReg_i16_lane imm:$lane)))>;
4290 def : Pat<(v4i32 (mul (v4i32 QPR:$src1),
4291 (v4i32 (ARMvduplane (v4i32 QPR:$src2), imm:$lane)))),
4292 (v4i32 (VMULslv4i32 (v4i32 QPR:$src1),
4293 (v2i32 (EXTRACT_SUBREG QPR:$src2,
4294 (DSubReg_i32_reg imm:$lane))),
4295 (SubReg_i32_lane imm:$lane)))>;
4296 def : Pat<(v4f32 (fmul (v4f32 QPR:$src1),
4297 (v4f32 (ARMvduplane (v4f32 QPR:$src2), imm:$lane)))),
4298 (v4f32 (VMULslfq (v4f32 QPR:$src1),
4299 (v2f32 (EXTRACT_SUBREG QPR:$src2,
4300 (DSubReg_i32_reg imm:$lane))),
4301 (SubReg_i32_lane imm:$lane)))>;
4302 def : Pat<(v8f16 (fmul (v8f16 QPR:$src1),
4303 (v8f16 (ARMvduplane (v8f16 QPR:$src2), imm:$lane)))),
4304 (v8f16 (VMULslhq(v8f16 QPR:$src1),
4305 (v4f16 (EXTRACT_SUBREG QPR:$src2,
4306 (DSubReg_i16_reg imm:$lane))),
4307 (SubReg_i16_lane imm:$lane)))>;
4309 def : Pat<(v2f32 (fmul DPR:$Rn, (ARMvdup (f32 SPR:$Rm)))),
4311 (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$Rm, ssub_0),
4313 def : Pat<(v4f16 (fmul DPR:$Rn, (ARMvdup (f16 HPR:$Rm)))),
4315 (INSERT_SUBREG (v4f16 (IMPLICIT_DEF)), (f16 HPR:$Rm), ssub_0),
4317 def : Pat<(v4f32 (fmul QPR:$Rn, (ARMvdup (f32 SPR:$Rm)))),
4319 (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$Rm, ssub_0),
4321 def : Pat<(v8f16 (fmul QPR:$Rn, (ARMvdup (f16 HPR:$Rm)))),
4323 (INSERT_SUBREG (v4f16 (IMPLICIT_DEF)), (f16 HPR:$Rm), ssub_0),
4327 // VQDMULH : Vector Saturating Doubling Multiply Returning High Half
4328 defm VQDMULH : N3VInt_HS<0, 0, 0b1011, 0, N3RegFrm, IIC_VMULi16D, IIC_VMULi32D,
4329 IIC_VMULi16Q, IIC_VMULi32Q,
4330 "vqdmulh", "s", int_arm_neon_vqdmulh, 1>;
4331 defm VQDMULHsl: N3VIntSL_HS<0b1100, IIC_VMULi16D, IIC_VMULi32D,
4332 IIC_VMULi16Q, IIC_VMULi32Q,
4333 "vqdmulh", "s", int_arm_neon_vqdmulh>;
4335 let Predicates = [HasNEON] in {
4336 def : Pat<(v8i16 (int_arm_neon_vqdmulh (v8i16 QPR:$src1),
4337 (v8i16 (ARMvduplane (v8i16 QPR:$src2),
4339 (v8i16 (VQDMULHslv8i16 (v8i16 QPR:$src1),
4340 (v4i16 (EXTRACT_SUBREG QPR:$src2,
4341 (DSubReg_i16_reg imm:$lane))),
4342 (SubReg_i16_lane imm:$lane)))>;
4343 def : Pat<(v4i32 (int_arm_neon_vqdmulh (v4i32 QPR:$src1),
4344 (v4i32 (ARMvduplane (v4i32 QPR:$src2),
4346 (v4i32 (VQDMULHslv4i32 (v4i32 QPR:$src1),
4347 (v2i32 (EXTRACT_SUBREG QPR:$src2,
4348 (DSubReg_i32_reg imm:$lane))),
4349 (SubReg_i32_lane imm:$lane)))>;
4352 // VQRDMULH : Vector Rounding Saturating Doubling Multiply Returning High Half
4353 defm VQRDMULH : N3VInt_HS<1, 0, 0b1011, 0, N3RegFrm,
4354 IIC_VMULi16D,IIC_VMULi32D,IIC_VMULi16Q,IIC_VMULi32Q,
4355 "vqrdmulh", "s", int_arm_neon_vqrdmulh, 1>;
4356 defm VQRDMULHsl : N3VIntSL_HS<0b1101, IIC_VMULi16D, IIC_VMULi32D,
4357 IIC_VMULi16Q, IIC_VMULi32Q,
4358 "vqrdmulh", "s", int_arm_neon_vqrdmulh>;
4360 let Predicates = [HasNEON] in {
4361 def : Pat<(v8i16 (int_arm_neon_vqrdmulh (v8i16 QPR:$src1),
4362 (v8i16 (ARMvduplane (v8i16 QPR:$src2),
4364 (v8i16 (VQRDMULHslv8i16 (v8i16 QPR:$src1),
4365 (v4i16 (EXTRACT_SUBREG QPR:$src2,
4366 (DSubReg_i16_reg imm:$lane))),
4367 (SubReg_i16_lane imm:$lane)))>;
4368 def : Pat<(v4i32 (int_arm_neon_vqrdmulh (v4i32 QPR:$src1),
4369 (v4i32 (ARMvduplane (v4i32 QPR:$src2),
4371 (v4i32 (VQRDMULHslv4i32 (v4i32 QPR:$src1),
4372 (v2i32 (EXTRACT_SUBREG QPR:$src2,
4373 (DSubReg_i32_reg imm:$lane))),
4374 (SubReg_i32_lane imm:$lane)))>;
4377 // VMULL : Vector Multiply Long (integer and polynomial) (Q = D * D)
4378 let PostEncoderMethod = "NEONThumb2DataIPostEncoder",
4379 DecoderNamespace = "NEONData" in {
4380 defm VMULLs : N3VL_QHS<0,1,0b1100,0, IIC_VMULi16D, IIC_VMULi32D,
4381 "vmull", "s", ARMvmulls, 1>;
4382 defm VMULLu : N3VL_QHS<1,1,0b1100,0, IIC_VMULi16D, IIC_VMULi32D,
4383 "vmull", "u", ARMvmullu, 1>;
4384 def VMULLp8 : N3VLInt<0, 1, 0b00, 0b1110, 0, IIC_VMULi16D, "vmull", "p8",
4385 v8i16, v8i8, int_arm_neon_vmullp, 1>;
4386 def VMULLp64 : N3VLIntnp<0b00101, 0b10, 0b1110, 0, 0, NoItinerary,
4387 "vmull", "p64", v2i64, v1i64, int_arm_neon_vmullp, 1>,
4388 Requires<[HasV8, HasCrypto]>;
4390 defm VMULLsls : N3VLSL_HS<0, 0b1010, IIC_VMULi16D, "vmull", "s", ARMvmulls>;
4391 defm VMULLslu : N3VLSL_HS<1, 0b1010, IIC_VMULi16D, "vmull", "u", ARMvmullu>;
4393 // VQDMULL : Vector Saturating Doubling Multiply Long (Q = D * D)
4394 defm VQDMULL : N3VLInt_HS<0,1,0b1101,0, IIC_VMULi16D, IIC_VMULi32D,
4395 "vqdmull", "s", int_arm_neon_vqdmull, 1>;
4396 defm VQDMULLsl: N3VLIntSL_HS<0, 0b1011, IIC_VMULi16D,
4397 "vqdmull", "s", int_arm_neon_vqdmull>;
4399 // Vector Multiply-Accumulate and Multiply-Subtract Operations.
4401 // VMLA : Vector Multiply Accumulate (integer and floating-point)
4402 defm VMLA : N3VMulOp_QHS<0, 0, 0b1001, 0, IIC_VMACi16D, IIC_VMACi32D,
4403 IIC_VMACi16Q, IIC_VMACi32Q, "vmla", "i", add>;
4404 def VMLAfd : N3VDMulOp<0, 0, 0b00, 0b1101, 1, IIC_VMACD, "vmla", "f32",
4405 v2f32, fmul_su, fadd_mlx>,
4406 Requires<[HasNEON, UseFPVMLx]>;
4407 def VMLAfq : N3VQMulOp<0, 0, 0b00, 0b1101, 1, IIC_VMACQ, "vmla", "f32",
4408 v4f32, fmul_su, fadd_mlx>,
4409 Requires<[HasNEON, UseFPVMLx]>;
4410 def VMLAhd : N3VDMulOp<0, 0, 0b01, 0b1101, 1, IIC_VMACD, "vmla", "f16",
4411 v4f16, fmul_su, fadd_mlx>,
4412 Requires<[HasNEON, HasFullFP16, UseFPVMLx]>;
4413 def VMLAhq : N3VQMulOp<0, 0, 0b01, 0b1101, 1, IIC_VMACQ, "vmla", "f16",
4414 v8f16, fmul_su, fadd_mlx>,
4415 Requires<[HasNEON, HasFullFP16, UseFPVMLx]>;
4416 defm VMLAsl : N3VMulOpSL_HS<0b0000, IIC_VMACi16D, IIC_VMACi32D,
4417 IIC_VMACi16Q, IIC_VMACi32Q, "vmla", "i", add>;
4418 def VMLAslfd : N3VDMulOpSL<0b10, 0b0001, IIC_VMACD, "vmla", "f32",
4419 v2f32, fmul_su, fadd_mlx>,
4420 Requires<[HasNEON, UseFPVMLx]>;
4421 def VMLAslfq : N3VQMulOpSL<0b10, 0b0001, IIC_VMACQ, "vmla", "f32",
4422 v4f32, v2f32, fmul_su, fadd_mlx>,
4423 Requires<[HasNEON, UseFPVMLx]>;
4424 def VMLAslhd : N3VDMulOpSL16<0b01, 0b0001, IIC_VMACD, "vmla", "f16",
4426 Requires<[HasNEON, HasFullFP16, UseFPVMLx]>;
4427 def VMLAslhq : N3VQMulOpSL16<0b01, 0b0001, IIC_VMACQ, "vmla", "f16",
4428 v8f16, v4f16, fmul, fadd>,
4429 Requires<[HasNEON, HasFullFP16, UseFPVMLx]>;
4431 let Predicates = [HasNEON] in {
4432 def : Pat<(v8i16 (add (v8i16 QPR:$src1),
4433 (mul (v8i16 QPR:$src2),
4434 (v8i16 (ARMvduplane (v8i16 QPR:$src3), imm:$lane))))),
4435 (v8i16 (VMLAslv8i16 (v8i16 QPR:$src1), (v8i16 QPR:$src2),
4436 (v4i16 (EXTRACT_SUBREG QPR:$src3,
4437 (DSubReg_i16_reg imm:$lane))),
4438 (SubReg_i16_lane imm:$lane)))>;
4440 def : Pat<(v4i32 (add (v4i32 QPR:$src1),
4441 (mul (v4i32 QPR:$src2),
4442 (v4i32 (ARMvduplane (v4i32 QPR:$src3), imm:$lane))))),
4443 (v4i32 (VMLAslv4i32 (v4i32 QPR:$src1), (v4i32 QPR:$src2),
4444 (v2i32 (EXTRACT_SUBREG QPR:$src3,
4445 (DSubReg_i32_reg imm:$lane))),
4446 (SubReg_i32_lane imm:$lane)))>;
4449 def : Pat<(v4f32 (fadd_mlx (v4f32 QPR:$src1),
4450 (fmul_su (v4f32 QPR:$src2),
4451 (v4f32 (ARMvduplane (v4f32 QPR:$src3), imm:$lane))))),
4452 (v4f32 (VMLAslfq (v4f32 QPR:$src1),
4454 (v2f32 (EXTRACT_SUBREG QPR:$src3,
4455 (DSubReg_i32_reg imm:$lane))),
4456 (SubReg_i32_lane imm:$lane)))>,
4457 Requires<[HasNEON, UseFPVMLx]>;
4459 // VMLAL : Vector Multiply Accumulate Long (Q += D * D)
4460 defm VMLALs : N3VLMulOp_QHS<0,1,0b1000,0, IIC_VMACi16D, IIC_VMACi32D,
4461 "vmlal", "s", ARMvmulls, add>;
4462 defm VMLALu : N3VLMulOp_QHS<1,1,0b1000,0, IIC_VMACi16D, IIC_VMACi32D,
4463 "vmlal", "u", ARMvmullu, add>;
4465 defm VMLALsls : N3VLMulOpSL_HS<0, 0b0010, "vmlal", "s", ARMvmulls, add>;
4466 defm VMLALslu : N3VLMulOpSL_HS<1, 0b0010, "vmlal", "u", ARMvmullu, add>;
4468 let Predicates = [HasNEON, HasV8_1a] in {
4469 // v8.1a Neon Rounding Double Multiply-Op vector operations,
4470 // VQRDMLAH : Vector Saturating Rounding Doubling Multiply Accumulate Long
4472 defm VQRDMLAH : N3VInt3_HS<1, 0, 0b1011, 1, IIC_VMACi16D, IIC_VMACi32D,
4473 IIC_VMACi16Q, IIC_VMACi32Q, "vqrdmlah", "s",
4475 def : Pat<(v4i16 (saddsat
4477 (v4i16 (int_arm_neon_vqrdmulh (v4i16 DPR:$Vn),
4478 (v4i16 DPR:$Vm))))),
4479 (v4i16 (VQRDMLAHv4i16 DPR:$src1, DPR:$Vn, DPR:$Vm))>;
4480 def : Pat<(v2i32 (saddsat
4482 (v2i32 (int_arm_neon_vqrdmulh (v2i32 DPR:$Vn),
4483 (v2i32 DPR:$Vm))))),
4484 (v2i32 (VQRDMLAHv2i32 DPR:$src1, DPR:$Vn, DPR:$Vm))>;
4485 def : Pat<(v8i16 (saddsat
4487 (v8i16 (int_arm_neon_vqrdmulh (v8i16 QPR:$Vn),
4488 (v8i16 QPR:$Vm))))),
4489 (v8i16 (VQRDMLAHv8i16 QPR:$src1, QPR:$Vn, QPR:$Vm))>;
4490 def : Pat<(v4i32 (saddsat
4492 (v4i32 (int_arm_neon_vqrdmulh (v4i32 QPR:$Vn),
4493 (v4i32 QPR:$Vm))))),
4494 (v4i32 (VQRDMLAHv4i32 QPR:$src1, QPR:$Vn, QPR:$Vm))>;
4496 defm VQRDMLAHsl : N3VMulOpSL_HS<0b1110, IIC_VMACi16D, IIC_VMACi32D,
4497 IIC_VMACi16Q, IIC_VMACi32Q, "vqrdmlah", "s",
4499 def : Pat<(v4i16 (saddsat
4501 (v4i16 (int_arm_neon_vqrdmulh
4503 (v4i16 (ARMvduplane (v4i16 DPR_8:$Vm),
4505 (v4i16 (VQRDMLAHslv4i16 DPR:$src1, DPR:$Vn, DPR_8:$Vm,
4507 def : Pat<(v2i32 (saddsat
4509 (v2i32 (int_arm_neon_vqrdmulh
4511 (v2i32 (ARMvduplane (v2i32 DPR_VFP2:$Vm),
4513 (v2i32 (VQRDMLAHslv2i32 DPR:$src1, DPR:$Vn, DPR_VFP2:$Vm,
4515 def : Pat<(v8i16 (saddsat
4517 (v8i16 (int_arm_neon_vqrdmulh
4519 (v8i16 (ARMvduplane (v8i16 QPR:$src3),
4521 (v8i16 (VQRDMLAHslv8i16 (v8i16 QPR:$src1),
4523 (v4i16 (EXTRACT_SUBREG
4525 (DSubReg_i16_reg imm:$lane))),
4526 (SubReg_i16_lane imm:$lane)))>;
4527 def : Pat<(v4i32 (saddsat
4529 (v4i32 (int_arm_neon_vqrdmulh
4531 (v4i32 (ARMvduplane (v4i32 QPR:$src3),
4533 (v4i32 (VQRDMLAHslv4i32 (v4i32 QPR:$src1),
4535 (v2i32 (EXTRACT_SUBREG
4537 (DSubReg_i32_reg imm:$lane))),
4538 (SubReg_i32_lane imm:$lane)))>;
4540 // VQRDMLSH : Vector Saturating Rounding Doubling Multiply Subtract Long
4542 defm VQRDMLSH : N3VInt3_HS<1, 0, 0b1100, 1, IIC_VMACi16D, IIC_VMACi32D,
4543 IIC_VMACi16Q, IIC_VMACi32Q, "vqrdmlsh", "s",
4545 def : Pat<(v4i16 (ssubsat
4547 (v4i16 (int_arm_neon_vqrdmulh (v4i16 DPR:$Vn),
4548 (v4i16 DPR:$Vm))))),
4549 (v4i16 (VQRDMLSHv4i16 DPR:$src1, DPR:$Vn, DPR:$Vm))>;
4550 def : Pat<(v2i32 (ssubsat
4552 (v2i32 (int_arm_neon_vqrdmulh (v2i32 DPR:$Vn),
4553 (v2i32 DPR:$Vm))))),
4554 (v2i32 (VQRDMLSHv2i32 DPR:$src1, DPR:$Vn, DPR:$Vm))>;
4555 def : Pat<(v8i16 (ssubsat
4557 (v8i16 (int_arm_neon_vqrdmulh (v8i16 QPR:$Vn),
4558 (v8i16 QPR:$Vm))))),
4559 (v8i16 (VQRDMLSHv8i16 QPR:$src1, QPR:$Vn, QPR:$Vm))>;
4560 def : Pat<(v4i32 (ssubsat
4562 (v4i32 (int_arm_neon_vqrdmulh (v4i32 QPR:$Vn),
4563 (v4i32 QPR:$Vm))))),
4564 (v4i32 (VQRDMLSHv4i32 QPR:$src1, QPR:$Vn, QPR:$Vm))>;
4566 defm VQRDMLSHsl : N3VMulOpSL_HS<0b1111, IIC_VMACi16D, IIC_VMACi32D,
4567 IIC_VMACi16Q, IIC_VMACi32Q, "vqrdmlsh", "s",
4569 def : Pat<(v4i16 (ssubsat
4571 (v4i16 (int_arm_neon_vqrdmulh
4573 (v4i16 (ARMvduplane (v4i16 DPR_8:$Vm),
4575 (v4i16 (VQRDMLSHslv4i16 DPR:$src1, DPR:$Vn, DPR_8:$Vm, imm:$lane))>;
4576 def : Pat<(v2i32 (ssubsat
4578 (v2i32 (int_arm_neon_vqrdmulh
4580 (v2i32 (ARMvduplane (v2i32 DPR_VFP2:$Vm),
4582 (v2i32 (VQRDMLSHslv2i32 DPR:$src1, DPR:$Vn, DPR_VFP2:$Vm,
4584 def : Pat<(v8i16 (ssubsat
4586 (v8i16 (int_arm_neon_vqrdmulh
4588 (v8i16 (ARMvduplane (v8i16 QPR:$src3),
4590 (v8i16 (VQRDMLSHslv8i16 (v8i16 QPR:$src1),
4592 (v4i16 (EXTRACT_SUBREG
4594 (DSubReg_i16_reg imm:$lane))),
4595 (SubReg_i16_lane imm:$lane)))>;
4596 def : Pat<(v4i32 (ssubsat
4598 (v4i32 (int_arm_neon_vqrdmulh
4600 (v4i32 (ARMvduplane (v4i32 QPR:$src3),
4602 (v4i32 (VQRDMLSHslv4i32 (v4i32 QPR:$src1),
4604 (v2i32 (EXTRACT_SUBREG
4606 (DSubReg_i32_reg imm:$lane))),
4607 (SubReg_i32_lane imm:$lane)))>;
4609 // VQDMLAL : Vector Saturating Doubling Multiply Accumulate Long (Q += D * D)
4610 defm VQDMLAL : N3VLInt3_HS<0, 1, 0b1001, 0, IIC_VMACi16D, IIC_VMACi32D,
4611 "vqdmlal", "s", null_frag>;
4612 defm VQDMLALsl: N3VLInt3SL_HS<0, 0b0011, "vqdmlal", "s", null_frag>;
4614 let Predicates = [HasNEON] in {
4615 def : Pat<(v4i32 (saddsat (v4i32 QPR:$src1),
4616 (v4i32 (int_arm_neon_vqdmull (v4i16 DPR:$Vn),
4617 (v4i16 DPR:$Vm))))),
4618 (VQDMLALv4i32 QPR:$src1, DPR:$Vn, DPR:$Vm)>;
4619 def : Pat<(v2i64 (saddsat (v2i64 QPR:$src1),
4620 (v2i64 (int_arm_neon_vqdmull (v2i32 DPR:$Vn),
4621 (v2i32 DPR:$Vm))))),
4622 (VQDMLALv2i64 QPR:$src1, DPR:$Vn, DPR:$Vm)>;
4623 def : Pat<(v4i32 (saddsat (v4i32 QPR:$src1),
4624 (v4i32 (int_arm_neon_vqdmull (v4i16 DPR:$Vn),
4625 (v4i16 (ARMvduplane (v4i16 DPR_8:$Vm),
4627 (VQDMLALslv4i16 QPR:$src1, DPR:$Vn, DPR_8:$Vm, imm:$lane)>;
4628 def : Pat<(v2i64 (saddsat (v2i64 QPR:$src1),
4629 (v2i64 (int_arm_neon_vqdmull (v2i32 DPR:$Vn),
4630 (v2i32 (ARMvduplane (v2i32 DPR_VFP2:$Vm),
4632 (VQDMLALslv2i32 QPR:$src1, DPR:$Vn, DPR_VFP2:$Vm, imm:$lane)>;
4635 // VMLS : Vector Multiply Subtract (integer and floating-point)
4636 defm VMLS : N3VMulOp_QHS<1, 0, 0b1001, 0, IIC_VMACi16D, IIC_VMACi32D,
4637 IIC_VMACi16Q, IIC_VMACi32Q, "vmls", "i", sub>;
4638 def VMLSfd : N3VDMulOp<0, 0, 0b10, 0b1101, 1, IIC_VMACD, "vmls", "f32",
4639 v2f32, fmul_su, fsub_mlx>,
4640 Requires<[HasNEON, UseFPVMLx]>;
4641 def VMLSfq : N3VQMulOp<0, 0, 0b10, 0b1101, 1, IIC_VMACQ, "vmls", "f32",
4642 v4f32, fmul_su, fsub_mlx>,
4643 Requires<[HasNEON, UseFPVMLx]>;
4644 def VMLShd : N3VDMulOp<0, 0, 0b11, 0b1101, 1, IIC_VMACD, "vmls", "f16",
4646 Requires<[HasNEON, HasFullFP16, UseFPVMLx]>;
4647 def VMLShq : N3VQMulOp<0, 0, 0b11, 0b1101, 1, IIC_VMACQ, "vmls", "f16",
4649 Requires<[HasNEON, HasFullFP16, UseFPVMLx]>;
4650 defm VMLSsl : N3VMulOpSL_HS<0b0100, IIC_VMACi16D, IIC_VMACi32D,
4651 IIC_VMACi16Q, IIC_VMACi32Q, "vmls", "i", sub>;
4652 def VMLSslfd : N3VDMulOpSL<0b10, 0b0101, IIC_VMACD, "vmls", "f32",
4653 v2f32, fmul_su, fsub_mlx>,
4654 Requires<[HasNEON, UseFPVMLx]>;
4655 def VMLSslfq : N3VQMulOpSL<0b10, 0b0101, IIC_VMACQ, "vmls", "f32",
4656 v4f32, v2f32, fmul_su, fsub_mlx>,
4657 Requires<[HasNEON, UseFPVMLx]>;
4658 def VMLSslhd : N3VDMulOpSL16<0b01, 0b0101, IIC_VMACD, "vmls", "f16",
4660 Requires<[HasNEON, HasFullFP16, UseFPVMLx]>;
4661 def VMLSslhq : N3VQMulOpSL16<0b01, 0b0101, IIC_VMACQ, "vmls", "f16",
4662 v8f16, v4f16, fmul, fsub>,
4663 Requires<[HasNEON, HasFullFP16, UseFPVMLx]>;
4665 let Predicates = [HasNEON] in {
4666 def : Pat<(v8i16 (sub (v8i16 QPR:$src1),
4667 (mul (v8i16 QPR:$src2),
4668 (v8i16 (ARMvduplane (v8i16 QPR:$src3), imm:$lane))))),
4669 (v8i16 (VMLSslv8i16 (v8i16 QPR:$src1), (v8i16 QPR:$src2),
4670 (v4i16 (EXTRACT_SUBREG QPR:$src3,
4671 (DSubReg_i16_reg imm:$lane))),
4672 (SubReg_i16_lane imm:$lane)))>;
4674 def : Pat<(v4i32 (sub (v4i32 QPR:$src1),
4675 (mul (v4i32 QPR:$src2),
4676 (v4i32 (ARMvduplane (v4i32 QPR:$src3), imm:$lane))))),
4677 (v4i32 (VMLSslv4i32 (v4i32 QPR:$src1), (v4i32 QPR:$src2),
4678 (v2i32 (EXTRACT_SUBREG QPR:$src3,
4679 (DSubReg_i32_reg imm:$lane))),
4680 (SubReg_i32_lane imm:$lane)))>;
4683 def : Pat<(v4f32 (fsub_mlx (v4f32 QPR:$src1),
4684 (fmul_su (v4f32 QPR:$src2),
4685 (v4f32 (ARMvduplane (v4f32 QPR:$src3), imm:$lane))))),
4686 (v4f32 (VMLSslfq (v4f32 QPR:$src1), (v4f32 QPR:$src2),
4687 (v2f32 (EXTRACT_SUBREG QPR:$src3,
4688 (DSubReg_i32_reg imm:$lane))),
4689 (SubReg_i32_lane imm:$lane)))>,
4690 Requires<[HasNEON, UseFPVMLx]>;
4692 // VMLSL : Vector Multiply Subtract Long (Q -= D * D)
4693 defm VMLSLs : N3VLMulOp_QHS<0,1,0b1010,0, IIC_VMACi16D, IIC_VMACi32D,
4694 "vmlsl", "s", ARMvmulls, sub>;
4695 defm VMLSLu : N3VLMulOp_QHS<1,1,0b1010,0, IIC_VMACi16D, IIC_VMACi32D,
4696 "vmlsl", "u", ARMvmullu, sub>;
4698 defm VMLSLsls : N3VLMulOpSL_HS<0, 0b0110, "vmlsl", "s", ARMvmulls, sub>;
4699 defm VMLSLslu : N3VLMulOpSL_HS<1, 0b0110, "vmlsl", "u", ARMvmullu, sub>;
4701 // VQDMLSL : Vector Saturating Doubling Multiply Subtract Long (Q -= D * D)
4702 defm VQDMLSL : N3VLInt3_HS<0, 1, 0b1011, 0, IIC_VMACi16D, IIC_VMACi32D,
4703 "vqdmlsl", "s", null_frag>;
4704 defm VQDMLSLsl: N3VLInt3SL_HS<0, 0b0111, "vqdmlsl", "s", null_frag>;
4706 let Predicates = [HasNEON] in {
4707 def : Pat<(v4i32 (ssubsat (v4i32 QPR:$src1),
4708 (v4i32 (int_arm_neon_vqdmull (v4i16 DPR:$Vn),
4709 (v4i16 DPR:$Vm))))),
4710 (VQDMLSLv4i32 QPR:$src1, DPR:$Vn, DPR:$Vm)>;
4711 def : Pat<(v2i64 (ssubsat (v2i64 QPR:$src1),
4712 (v2i64 (int_arm_neon_vqdmull (v2i32 DPR:$Vn),
4713 (v2i32 DPR:$Vm))))),
4714 (VQDMLSLv2i64 QPR:$src1, DPR:$Vn, DPR:$Vm)>;
4715 def : Pat<(v4i32 (ssubsat (v4i32 QPR:$src1),
4716 (v4i32 (int_arm_neon_vqdmull (v4i16 DPR:$Vn),
4717 (v4i16 (ARMvduplane (v4i16 DPR_8:$Vm),
4719 (VQDMLSLslv4i16 QPR:$src1, DPR:$Vn, DPR_8:$Vm, imm:$lane)>;
4720 def : Pat<(v2i64 (ssubsat (v2i64 QPR:$src1),
4721 (v2i64 (int_arm_neon_vqdmull (v2i32 DPR:$Vn),
4722 (v2i32 (ARMvduplane (v2i32 DPR_VFP2:$Vm),
4724 (VQDMLSLslv2i32 QPR:$src1, DPR:$Vn, DPR_VFP2:$Vm, imm:$lane)>;
4727 // Fused Vector Multiply-Accumulate and Fused Multiply-Subtract Operations.
4728 def VFMAfd : N3VDMulOp<0, 0, 0b00, 0b1100, 1, IIC_VFMACD, "vfma", "f32",
4729 v2f32, fmul_su, fadd_mlx>,
4730 Requires<[HasNEON,HasVFP4,UseFusedMAC]>;
4732 def VFMAfq : N3VQMulOp<0, 0, 0b00, 0b1100, 1, IIC_VFMACQ, "vfma", "f32",
4733 v4f32, fmul_su, fadd_mlx>,
4734 Requires<[HasNEON,HasVFP4,UseFusedMAC]>;
4735 def VFMAhd : N3VDMulOp<0, 0, 0b01, 0b1100, 1, IIC_VFMACD, "vfma", "f16",
4737 Requires<[HasNEON,HasFullFP16,UseFusedMAC]>;
4739 def VFMAhq : N3VQMulOp<0, 0, 0b01, 0b1100, 1, IIC_VFMACQ, "vfma", "f16",
4741 Requires<[HasNEON,HasFullFP16,UseFusedMAC]>;
4743 // Fused Vector Multiply Subtract (floating-point)
4744 def VFMSfd : N3VDMulOp<0, 0, 0b10, 0b1100, 1, IIC_VFMACD, "vfms", "f32",
4745 v2f32, fmul_su, fsub_mlx>,
4746 Requires<[HasNEON,HasVFP4,UseFusedMAC]>;
4747 def VFMSfq : N3VQMulOp<0, 0, 0b10, 0b1100, 1, IIC_VFMACQ, "vfms", "f32",
4748 v4f32, fmul_su, fsub_mlx>,
4749 Requires<[HasNEON,HasVFP4,UseFusedMAC]>;
4750 def VFMShd : N3VDMulOp<0, 0, 0b11, 0b1100, 1, IIC_VFMACD, "vfms", "f16",
4752 Requires<[HasNEON,HasFullFP16,UseFusedMAC]>;
4753 def VFMShq : N3VQMulOp<0, 0, 0b11, 0b1100, 1, IIC_VFMACQ, "vfms", "f16",
4755 Requires<[HasNEON,HasFullFP16,UseFusedMAC]>;
4757 // Match @llvm.fma.* intrinsics
4758 def : Pat<(v4f16 (fma DPR:$Vn, DPR:$Vm, DPR:$src1)),
4759 (VFMAhd DPR:$src1, DPR:$Vn, DPR:$Vm)>,
4760 Requires<[HasNEON,HasFullFP16]>;
4761 def : Pat<(v8f16 (fma QPR:$Vn, QPR:$Vm, QPR:$src1)),
4762 (VFMAhq QPR:$src1, QPR:$Vn, QPR:$Vm)>,
4763 Requires<[HasNEON,HasFullFP16]>;
4764 def : Pat<(v2f32 (fma DPR:$Vn, DPR:$Vm, DPR:$src1)),
4765 (VFMAfd DPR:$src1, DPR:$Vn, DPR:$Vm)>,
4766 Requires<[HasNEON,HasVFP4]>;
4767 def : Pat<(v4f32 (fma QPR:$Vn, QPR:$Vm, QPR:$src1)),
4768 (VFMAfq QPR:$src1, QPR:$Vn, QPR:$Vm)>,
4769 Requires<[HasNEON,HasVFP4]>;
4770 def : Pat<(v2f32 (fma (fneg DPR:$Vn), DPR:$Vm, DPR:$src1)),
4771 (VFMSfd DPR:$src1, DPR:$Vn, DPR:$Vm)>,
4772 Requires<[HasNEON,HasVFP4]>;
4773 def : Pat<(v4f32 (fma (fneg QPR:$Vn), QPR:$Vm, QPR:$src1)),
4774 (VFMSfq QPR:$src1, QPR:$Vn, QPR:$Vm)>,
4775 Requires<[HasNEON,HasVFP4]>;
4777 // ARMv8.2a dot product instructions.
4778 // We put them in the VFPV8 decoder namespace because the ARM and Thumb
4779 // encodings are the same and thus no further bit twiddling is necessary
4780 // in the disassembler.
4781 class VDOT<bit op6, bit op4, bit op23, RegisterClass RegTy, string Asm,
4782 string AsmTy, ValueType AccumTy, ValueType InputTy,
4783 SDPatternOperator OpNode> :
4784 N3Vnp<{0b1100, op23}, 0b10, 0b1101, op6, op4, (outs RegTy:$dst),
4785 (ins RegTy:$Vd, RegTy:$Vn, RegTy:$Vm), N3RegFrm, IIC_VDOTPROD,
4787 [(set (AccumTy RegTy:$dst),
4788 (OpNode (AccumTy RegTy:$Vd),
4789 (InputTy RegTy:$Vn),
4790 (InputTy RegTy:$Vm)))]> {
4791 let Predicates = [HasDotProd];
4792 let DecoderNamespace = "VFPV8";
4793 let Constraints = "$dst = $Vd";
4796 def VUDOTD : VDOT<0, 1, 0, DPR, "vudot", "u8", v2i32, v8i8, int_arm_neon_udot>;
4797 def VSDOTD : VDOT<0, 0, 0, DPR, "vsdot", "s8", v2i32, v8i8, int_arm_neon_sdot>;
4798 def VUDOTQ : VDOT<1, 1, 0, QPR, "vudot", "u8", v4i32, v16i8, int_arm_neon_udot>;
4799 def VSDOTQ : VDOT<1, 0, 0, QPR, "vsdot", "s8", v4i32, v16i8, int_arm_neon_sdot>;
4801 // Indexed dot product instructions:
4802 multiclass DOTI<string opc, string dt, bit Q, bit U, RegisterClass Ty,
4803 ValueType AccumType, ValueType InputType, SDPatternOperator OpNode,
4805 def "" : N3Vnp<0b11100, 0b10, 0b1101, Q, U, (outs Ty:$dst),
4806 (ins Ty:$Vd, Ty:$Vn, DPR_VFP2:$Vm, VectorIndex32:$lane),
4807 N3RegFrm, IIC_VDOTPROD, opc, dt, []> {
4810 let AsmString = !strconcat(opc, ".", dt, "\t$Vd, $Vn, $Vm$lane");
4811 let Constraints = "$dst = $Vd";
4812 let Predicates = [HasDotProd];
4813 let DecoderNamespace = "VFPV8";
4817 (AccumType (OpNode (AccumType Ty:$Vd),
4819 (InputType (bitconvert (AccumType
4820 (ARMvduplane (AccumType Ty:$Vm),
4821 VectorIndex32:$lane)))))),
4822 (!cast<Instruction>(NAME) Ty:$Vd, Ty:$Vn, RHS, VectorIndex32:$lane)>;
4825 defm VUDOTDI : DOTI<"vudot", "u8", 0b0, 0b1, DPR, v2i32, v8i8,
4826 int_arm_neon_udot, (v2i32 DPR_VFP2:$Vm)>;
4827 defm VSDOTDI : DOTI<"vsdot", "s8", 0b0, 0b0, DPR, v2i32, v8i8,
4828 int_arm_neon_sdot, (v2i32 DPR_VFP2:$Vm)>;
4829 defm VUDOTQI : DOTI<"vudot", "u8", 0b1, 0b1, QPR, v4i32, v16i8,
4830 int_arm_neon_udot, (EXTRACT_SUBREG QPR:$Vm, dsub_0)>;
4831 defm VSDOTQI : DOTI<"vsdot", "s8", 0b1, 0b0, QPR, v4i32, v16i8,
4832 int_arm_neon_sdot, (EXTRACT_SUBREG QPR:$Vm, dsub_0)>;
4834 // v8.6A matrix multiplication extension
4835 let Predicates = [HasMatMulInt8] in {
4836 class N3VMatMul<bit B, bit U, string Asm, string AsmTy,
4837 SDPatternOperator OpNode>
4838 : N3Vnp<{0b1100, B}, 0b10, 0b1100, 1, U, (outs QPR:$dst),
4839 (ins QPR:$Vd, QPR:$Vn, QPR:$Vm), N3RegFrm, NoItinerary,
4841 [(set (v4i32 QPR:$dst), (OpNode (v4i32 QPR:$Vd),
4843 (v16i8 QPR:$Vm)))]> {
4844 let DecoderNamespace = "VFPV8";
4845 let Constraints = "$dst = $Vd";
4848 multiclass N3VMixedDotLane<bit Q, bit U, string Asm, string AsmTy, RegisterClass RegTy,
4849 ValueType AccumTy, ValueType InputTy, SDPatternOperator OpNode,
4852 def "" : N3Vnp<0b11101, 0b00, 0b1101, Q, U, (outs RegTy:$dst),
4853 (ins RegTy:$Vd, RegTy:$Vn, DPR_VFP2:$Vm, VectorIndex32:$lane), N3RegFrm,
4854 NoItinerary, Asm, AsmTy, []> {
4857 let AsmString = !strconcat(Asm, ".", AsmTy, "\t$Vd, $Vn, $Vm$lane");
4858 let DecoderNamespace = "VFPV8";
4859 let Constraints = "$dst = $Vd";
4863 (AccumTy (OpNode (AccumTy RegTy:$Vd),
4864 (InputTy RegTy:$Vn),
4865 (InputTy (bitconvert (AccumTy
4866 (ARMvduplane (AccumTy RegTy:$Vm),
4867 VectorIndex32:$lane)))))),
4868 (!cast<Instruction>(NAME) RegTy:$Vd, RegTy:$Vn, RHS, VectorIndex32:$lane)>;
4872 multiclass SUDOTLane<bit Q, RegisterClass RegTy, ValueType AccumTy, ValueType InputTy, dag RHS>
4873 : N3VMixedDotLane<Q, 1, "vsudot", "u8", RegTy, AccumTy, InputTy, null_frag, null_frag> {
4875 (AccumTy (int_arm_neon_usdot (AccumTy RegTy:$Vd),
4876 (InputTy (bitconvert (AccumTy
4877 (ARMvduplane (AccumTy RegTy:$Vm),
4878 VectorIndex32:$lane)))),
4879 (InputTy RegTy:$Vn))),
4880 (!cast<Instruction>(NAME) RegTy:$Vd, RegTy:$Vn, RHS, VectorIndex32:$lane)>;
4883 def VSMMLA : N3VMatMul<0, 0, "vsmmla", "s8", int_arm_neon_smmla>;
4884 def VUMMLA : N3VMatMul<0, 1, "vummla", "u8", int_arm_neon_ummla>;
4885 def VUSMMLA : N3VMatMul<1, 0, "vusmmla", "s8", int_arm_neon_usmmla>;
4886 def VUSDOTD : VDOT<0, 0, 1, DPR, "vusdot", "s8", v2i32, v8i8, int_arm_neon_usdot>;
4887 def VUSDOTQ : VDOT<1, 0, 1, QPR, "vusdot", "s8", v4i32, v16i8, int_arm_neon_usdot>;
4889 defm VUSDOTDI : N3VMixedDotLane<0, 0, "vusdot", "s8", DPR, v2i32, v8i8,
4890 int_arm_neon_usdot, (v2i32 DPR_VFP2:$Vm)>;
4891 defm VUSDOTQI : N3VMixedDotLane<1, 0, "vusdot", "s8", QPR, v4i32, v16i8,
4892 int_arm_neon_usdot, (EXTRACT_SUBREG QPR:$Vm, dsub_0)>;
4893 defm VSUDOTDI : SUDOTLane<0, DPR, v2i32, v8i8, (v2i32 DPR_VFP2:$Vm)>;
4894 defm VSUDOTQI : SUDOTLane<1, QPR, v4i32, v16i8, (EXTRACT_SUBREG QPR:$Vm, dsub_0)>;
4897 // ARMv8.3 complex operations
4898 class BaseN3VCP8ComplexTied<bit op21, bit op4, bit s, bit q,
4899 InstrItinClass itin, dag oops, dag iops,
4900 string opc, string dt, list<dag> pattern>
4901 : N3VCP8<{?,?}, {op21,s}, q, op4, oops,
4902 iops, itin, opc, dt, "$Vd, $Vn, $Vm, $rot", "$src1 = $Vd", pattern>{
4904 let Inst{24-23} = rot;
4907 class BaseN3VCP8ComplexOdd<bit op23, bit op21, bit op4, bit s, bit q,
4908 InstrItinClass itin, dag oops, dag iops, string opc,
4909 string dt, list<dag> pattern>
4910 : N3VCP8<{?,op23}, {op21,s}, q, op4, oops,
4911 iops, itin, opc, dt, "$Vd, $Vn, $Vm, $rot", "", pattern> {
4916 class BaseN3VCP8ComplexTiedLane32<bit op4, bit s, bit q, InstrItinClass itin,
4917 dag oops, dag iops, string opc, string dt,
4919 : N3VLaneCP8<s, {?,?}, q, op4, oops, iops, itin, opc, dt,
4920 "$Vd, $Vn, $Vm$lane, $rot", "$src1 = $Vd", pattern> {
4924 let Inst{21-20} = rot;
4928 class BaseN3VCP8ComplexTiedLane64<bit op4, bit s, bit q, InstrItinClass itin,
4929 dag oops, dag iops, string opc, string dt,
4931 : N3VLaneCP8<s, {?,?}, q, op4, oops, iops, itin, opc, dt,
4932 "$Vd, $Vn, $Vm$lane, $rot", "$src1 = $Vd", pattern> {
4936 let Inst{21-20} = rot;
4937 let Inst{5} = Vm{4};
4938 // This is needed because the lane operand does not have any bits in the
4939 // encoding (it only has one possible value), so we need to manually set it
4940 // to it's default value.
4941 let DecoderMethod = "DecodeNEONComplexLane64Instruction";
4944 multiclass N3VCP8ComplexTied<bit op21, bit op4,
4945 string OpcodeStr, SDPatternOperator Op> {
4946 let Predicates = [HasNEON,HasV8_3a,HasFullFP16] in {
4947 def v4f16 : BaseN3VCP8ComplexTied<op21, op4, 0, 0, IIC_VMACD, (outs DPR:$Vd),
4948 (ins DPR:$src1, DPR:$Vn, DPR:$Vm, complexrotateop:$rot),
4949 OpcodeStr, "f16", []>;
4950 def v8f16 : BaseN3VCP8ComplexTied<op21, op4, 0, 1, IIC_VMACQ, (outs QPR:$Vd),
4951 (ins QPR:$src1, QPR:$Vn, QPR:$Vm, complexrotateop:$rot),
4952 OpcodeStr, "f16", []>;
4954 let Predicates = [HasNEON,HasV8_3a] in {
4955 def v2f32 : BaseN3VCP8ComplexTied<op21, op4, 1, 0, IIC_VMACD, (outs DPR:$Vd),
4956 (ins DPR:$src1, DPR:$Vn, DPR:$Vm, complexrotateop:$rot),
4957 OpcodeStr, "f32", []>;
4958 def v4f32 : BaseN3VCP8ComplexTied<op21, op4, 1, 1, IIC_VMACQ, (outs QPR:$Vd),
4959 (ins QPR:$src1, QPR:$Vn, QPR:$Vm, complexrotateop:$rot),
4960 OpcodeStr, "f32", []>;
4964 multiclass N3VCP8ComplexOdd<bit op23, bit op21, bit op4,
4965 string OpcodeStr, SDPatternOperator Op> {
4966 let Predicates = [HasNEON,HasV8_3a,HasFullFP16] in {
4967 def v4f16 : BaseN3VCP8ComplexOdd<op23, op21, op4, 0, 0, IIC_VMACD,
4969 (ins DPR:$Vn, DPR:$Vm, complexrotateopodd:$rot),
4970 OpcodeStr, "f16", []>;
4971 def v8f16 : BaseN3VCP8ComplexOdd<op23, op21, op4, 0, 1, IIC_VMACQ,
4973 (ins QPR:$Vn, QPR:$Vm, complexrotateopodd:$rot),
4974 OpcodeStr, "f16", []>;
4976 let Predicates = [HasNEON,HasV8_3a] in {
4977 def v2f32 : BaseN3VCP8ComplexOdd<op23, op21, op4, 1, 0, IIC_VMACD,
4979 (ins DPR:$Vn, DPR:$Vm, complexrotateopodd:$rot),
4980 OpcodeStr, "f32", []>;
4981 def v4f32 : BaseN3VCP8ComplexOdd<op23, op21, op4, 1, 1, IIC_VMACQ,
4983 (ins QPR:$Vn, QPR:$Vm, complexrotateopodd:$rot),
4984 OpcodeStr, "f32", []>;
4988 // These instructions index by pairs of lanes, so the VectorIndexes are twice
4989 // as wide as the data types.
4990 multiclass N3VCP8ComplexTiedLane<bit op4, string OpcodeStr,
4991 SDPatternOperator Op> {
4992 let Predicates = [HasNEON,HasV8_3a,HasFullFP16] in {
4993 def v4f16_indexed : BaseN3VCP8ComplexTiedLane32<op4, 0, 0, IIC_VMACD,
4995 (ins DPR:$src1, DPR:$Vn, DPR_VFP2:$Vm,
4996 VectorIndex32:$lane, complexrotateop:$rot),
4997 OpcodeStr, "f16", []>;
4998 def v8f16_indexed : BaseN3VCP8ComplexTiedLane32<op4, 0, 1, IIC_VMACQ,
5000 (ins QPR:$src1, QPR:$Vn, DPR_VFP2:$Vm,
5001 VectorIndex32:$lane, complexrotateop:$rot),
5002 OpcodeStr, "f16", []>;
5004 let Predicates = [HasNEON,HasV8_3a] in {
5005 def v2f32_indexed : BaseN3VCP8ComplexTiedLane64<op4, 1, 0, IIC_VMACD,
5007 (ins DPR:$src1, DPR:$Vn, DPR:$Vm, VectorIndex64:$lane,
5008 complexrotateop:$rot),
5009 OpcodeStr, "f32", []>;
5010 def v4f32_indexed : BaseN3VCP8ComplexTiedLane64<op4, 1, 1, IIC_VMACQ,
5012 (ins QPR:$src1, QPR:$Vn, DPR:$Vm, VectorIndex64:$lane,
5013 complexrotateop:$rot),
5014 OpcodeStr, "f32", []>;
5018 defm VCMLA : N3VCP8ComplexTied<1, 0, "vcmla", null_frag>;
5019 defm VCADD : N3VCP8ComplexOdd<1, 0, 0, "vcadd", null_frag>;
5020 defm VCMLA : N3VCP8ComplexTiedLane<0, "vcmla", null_frag>;
5022 let Predicates = [HasNEON,HasV8_3a,HasFullFP16] in {
5023 def : Pat<(v4f16 (int_arm_neon_vcadd_rot90 (v4f16 DPR:$Rn), (v4f16 DPR:$Rm))),
5024 (VCADDv4f16 (v4f16 DPR:$Rn), (v4f16 DPR:$Rm), (i32 0))>;
5025 def : Pat<(v4f16 (int_arm_neon_vcadd_rot270 (v4f16 DPR:$Rn), (v4f16 DPR:$Rm))),
5026 (VCADDv4f16 (v4f16 DPR:$Rn), (v4f16 DPR:$Rm), (i32 1))>;
5027 def : Pat<(v8f16 (int_arm_neon_vcadd_rot90 (v8f16 QPR:$Rn), (v8f16 QPR:$Rm))),
5028 (VCADDv8f16 (v8f16 QPR:$Rn), (v8f16 QPR:$Rm), (i32 0))>;
5029 def : Pat<(v8f16 (int_arm_neon_vcadd_rot270 (v8f16 QPR:$Rn), (v8f16 QPR:$Rm))),
5030 (VCADDv8f16 (v8f16 QPR:$Rn), (v8f16 QPR:$Rm), (i32 1))>;
5032 let Predicates = [HasNEON,HasV8_3a] in {
5033 def : Pat<(v2f32 (int_arm_neon_vcadd_rot90 (v2f32 DPR:$Rn), (v2f32 DPR:$Rm))),
5034 (VCADDv2f32 (v2f32 DPR:$Rn), (v2f32 DPR:$Rm), (i32 0))>;
5035 def : Pat<(v2f32 (int_arm_neon_vcadd_rot270 (v2f32 DPR:$Rn), (v2f32 DPR:$Rm))),
5036 (VCADDv2f32 (v2f32 DPR:$Rn), (v2f32 DPR:$Rm), (i32 1))>;
5037 def : Pat<(v4f32 (int_arm_neon_vcadd_rot90 (v4f32 QPR:$Rn), (v4f32 QPR:$Rm))),
5038 (VCADDv4f32 (v4f32 QPR:$Rn), (v4f32 QPR:$Rm), (i32 0))>;
5039 def : Pat<(v4f32 (int_arm_neon_vcadd_rot270 (v4f32 QPR:$Rn), (v4f32 QPR:$Rm))),
5040 (VCADDv4f32 (v4f32 QPR:$Rn), (v4f32 QPR:$Rm), (i32 1))>;
5043 // Vector Subtract Operations.
5045 // VSUB : Vector Subtract (integer and floating-point)
5046 defm VSUB : N3V_QHSD<1, 0, 0b1000, 0, IIC_VSUBiD, IIC_VSUBiQ,
5047 "vsub", "i", sub, 0>;
5048 def VSUBfd : N3VD<0, 0, 0b10, 0b1101, 0, IIC_VBIND, "vsub", "f32",
5049 v2f32, v2f32, fsub, 0>;
5050 def VSUBfq : N3VQ<0, 0, 0b10, 0b1101, 0, IIC_VBINQ, "vsub", "f32",
5051 v4f32, v4f32, fsub, 0>;
5052 def VSUBhd : N3VD<0, 0, 0b11, 0b1101, 0, IIC_VBIND, "vsub", "f16",
5053 v4f16, v4f16, fsub, 0>,
5054 Requires<[HasNEON,HasFullFP16]>;
5055 def VSUBhq : N3VQ<0, 0, 0b11, 0b1101, 0, IIC_VBINQ, "vsub", "f16",
5056 v8f16, v8f16, fsub, 0>,
5057 Requires<[HasNEON,HasFullFP16]>;
5058 // VSUBL : Vector Subtract Long (Q = D - D)
5059 defm VSUBLs : N3VLExt_QHS<0,1,0b0010,0, IIC_VSHLiD, IIC_VSHLiD,
5060 "vsubl", "s", sub, sext, 0>;
5061 defm VSUBLu : N3VLExt_QHS<1,1,0b0010,0, IIC_VSHLiD, IIC_VSHLiD,
5062 "vsubl", "u", sub, zext, 0>;
5063 // VSUBW : Vector Subtract Wide (Q = Q - D)
5064 defm VSUBWs : N3VW_QHS<0,1,0b0011,0, "vsubw", "s", sub, sext, 0>;
5065 defm VSUBWu : N3VW_QHS<1,1,0b0011,0, "vsubw", "u", sub, zext, 0>;
5066 // VHSUB : Vector Halving Subtract
5067 defm VHSUBs : N3VInt_QHS<0, 0, 0b0010, 0, N3RegFrm,
5068 IIC_VSUBi4D, IIC_VSUBi4D, IIC_VSUBi4Q, IIC_VSUBi4Q,
5069 "vhsub", "s", int_arm_neon_vhsubs, 0>;
5070 defm VHSUBu : N3VInt_QHS<1, 0, 0b0010, 0, N3RegFrm,
5071 IIC_VSUBi4D, IIC_VSUBi4D, IIC_VSUBi4Q, IIC_VSUBi4Q,
5072 "vhsub", "u", int_arm_neon_vhsubu, 0>;
5073 // VQSUB : Vector Saturing Subtract
5074 defm VQSUBs : N3VInt_QHSD<0, 0, 0b0010, 1, N3RegFrm,
5075 IIC_VSUBi4D, IIC_VSUBi4D, IIC_VSUBi4Q, IIC_VSUBi4Q,
5076 "vqsub", "s", ssubsat, 0>;
5077 defm VQSUBu : N3VInt_QHSD<1, 0, 0b0010, 1, N3RegFrm,
5078 IIC_VSUBi4D, IIC_VSUBi4D, IIC_VSUBi4Q, IIC_VSUBi4Q,
5079 "vqsub", "u", usubsat, 0>;
5080 // VSUBHN : Vector Subtract and Narrow Returning High Half (D = Q - Q)
5081 defm VSUBHN : N3VNInt_HSD<0,1,0b0110,0, "vsubhn", "i", null_frag, 0>;
5082 // VRSUBHN : Vector Rounding Subtract and Narrow Returning High Half (D=Q-Q)
5083 defm VRSUBHN : N3VNInt_HSD<1,1,0b0110,0, "vrsubhn", "i",
5084 int_arm_neon_vrsubhn, 0>;
5086 let Predicates = [HasNEON] in {
5087 def : Pat<(v8i8 (trunc (ARMvshruImm (sub (v8i16 QPR:$Vn), QPR:$Vm), 8))),
5088 (VSUBHNv8i8 QPR:$Vn, QPR:$Vm)>;
5089 def : Pat<(v4i16 (trunc (ARMvshruImm (sub (v4i32 QPR:$Vn), QPR:$Vm), 16))),
5090 (VSUBHNv4i16 QPR:$Vn, QPR:$Vm)>;
5091 def : Pat<(v2i32 (trunc (ARMvshruImm (sub (v2i64 QPR:$Vn), QPR:$Vm), 32))),
5092 (VSUBHNv2i32 QPR:$Vn, QPR:$Vm)>;
5095 // Vector Comparisons.
5097 // VCEQ : Vector Compare Equal
5098 defm VCEQ : N3V_QHS_cmp<1, 0, 0b1000, 1, IIC_VSUBi4D, IIC_VSUBi4D, IIC_VSUBi4Q,
5099 IIC_VSUBi4Q, "vceq", "i", ARMCCeq, 1>;
5100 def VCEQfd : N3VD_cmp<0,0,0b00,0b1110,0, IIC_VBIND, "vceq", "f32", v2i32, v2f32,
5102 def VCEQfq : N3VQ_cmp<0,0,0b00,0b1110,0, IIC_VBINQ, "vceq", "f32", v4i32, v4f32,
5104 def VCEQhd : N3VD_cmp<0,0,0b01,0b1110,0, IIC_VBIND, "vceq", "f16", v4i16, v4f16,
5106 Requires<[HasNEON, HasFullFP16]>;
5107 def VCEQhq : N3VQ_cmp<0,0,0b01,0b1110,0, IIC_VBINQ, "vceq", "f16", v8i16, v8f16,
5109 Requires<[HasNEON, HasFullFP16]>;
5111 let TwoOperandAliasConstraint = "$Vm = $Vd" in
5112 defm VCEQz : N2V_QHS_cmp<0b11, 0b11, 0b01, 0b00010, 0, "vceq", "i",
5113 "$Vd, $Vm, #0", ARMCCeq>;
5115 // VCGE : Vector Compare Greater Than or Equal
5116 defm VCGEs : N3V_QHS_cmp<0, 0, 0b0011, 1, IIC_VSUBi4D, IIC_VSUBi4D, IIC_VSUBi4Q,
5117 IIC_VSUBi4Q, "vcge", "s", ARMCCge, 0>;
5118 defm VCGEu : N3V_QHS_cmp<1, 0, 0b0011, 1, IIC_VSUBi4D, IIC_VSUBi4D, IIC_VSUBi4Q,
5119 IIC_VSUBi4Q, "vcge", "u", ARMCChs, 0>;
5120 def VCGEfd : N3VD_cmp<1,0,0b00,0b1110,0, IIC_VBIND, "vcge", "f32", v2i32, v2f32,
5122 def VCGEfq : N3VQ_cmp<1,0,0b00,0b1110,0, IIC_VBINQ, "vcge", "f32", v4i32, v4f32,
5124 def VCGEhd : N3VD_cmp<1,0,0b01,0b1110,0, IIC_VBIND, "vcge", "f16", v4i16, v4f16,
5126 Requires<[HasNEON, HasFullFP16]>;
5127 def VCGEhq : N3VQ_cmp<1,0,0b01,0b1110,0, IIC_VBINQ, "vcge", "f16", v8i16, v8f16,
5129 Requires<[HasNEON, HasFullFP16]>;
5131 let TwoOperandAliasConstraint = "$Vm = $Vd" in {
5132 defm VCGEz : N2V_QHS_cmp<0b11, 0b11, 0b01, 0b00001, 0, "vcge", "s",
5133 "$Vd, $Vm, #0", ARMCCge>;
5134 defm VCLEz : N2V_QHS_cmp<0b11, 0b11, 0b01, 0b00011, 0, "vcle", "s",
5135 "$Vd, $Vm, #0", ARMCCle>;
5138 // VCGT : Vector Compare Greater Than
5139 defm VCGTs : N3V_QHS_cmp<0, 0, 0b0011, 0, IIC_VSUBi4D, IIC_VSUBi4D, IIC_VSUBi4Q,
5140 IIC_VSUBi4Q, "vcgt", "s", ARMCCgt, 0>;
5141 defm VCGTu : N3V_QHS_cmp<1, 0, 0b0011, 0, IIC_VSUBi4D, IIC_VSUBi4D, IIC_VSUBi4Q,
5142 IIC_VSUBi4Q, "vcgt", "u", ARMCChi, 0>;
5143 def VCGTfd : N3VD_cmp<1,0,0b10,0b1110,0, IIC_VBIND, "vcgt", "f32", v2i32, v2f32,
5145 def VCGTfq : N3VQ_cmp<1,0,0b10,0b1110,0, IIC_VBINQ, "vcgt", "f32", v4i32, v4f32,
5147 def VCGThd : N3VD_cmp<1,0,0b11,0b1110,0, IIC_VBIND, "vcgt", "f16", v4i16, v4f16,
5149 Requires<[HasNEON, HasFullFP16]>;
5150 def VCGThq : N3VQ_cmp<1,0,0b11,0b1110,0, IIC_VBINQ, "vcgt", "f16", v8i16, v8f16,
5152 Requires<[HasNEON, HasFullFP16]>;
5154 let TwoOperandAliasConstraint = "$Vm = $Vd" in {
5155 defm VCGTz : N2V_QHS_cmp<0b11, 0b11, 0b01, 0b00000, 0, "vcgt", "s",
5156 "$Vd, $Vm, #0", ARMCCgt>;
5157 defm VCLTz : N2V_QHS_cmp<0b11, 0b11, 0b01, 0b00100, 0, "vclt", "s",
5158 "$Vd, $Vm, #0", ARMCClt>;
5161 // VACGE : Vector Absolute Compare Greater Than or Equal (aka VCAGE)
5162 def VACGEfd : N3VDInt<1, 0, 0b00, 0b1110, 1, N3RegFrm, IIC_VBIND, "vacge",
5163 "f32", v2i32, v2f32, int_arm_neon_vacge, 0>;
5164 def VACGEfq : N3VQInt<1, 0, 0b00, 0b1110, 1, N3RegFrm, IIC_VBINQ, "vacge",
5165 "f32", v4i32, v4f32, int_arm_neon_vacge, 0>;
5166 def VACGEhd : N3VDInt<1, 0, 0b01, 0b1110, 1, N3RegFrm, IIC_VBIND, "vacge",
5167 "f16", v4i16, v4f16, int_arm_neon_vacge, 0>,
5168 Requires<[HasNEON, HasFullFP16]>;
5169 def VACGEhq : N3VQInt<1, 0, 0b01, 0b1110, 1, N3RegFrm, IIC_VBINQ, "vacge",
5170 "f16", v8i16, v8f16, int_arm_neon_vacge, 0>,
5171 Requires<[HasNEON, HasFullFP16]>;
5172 // VACGT : Vector Absolute Compare Greater Than (aka VCAGT)
5173 def VACGTfd : N3VDInt<1, 0, 0b10, 0b1110, 1, N3RegFrm, IIC_VBIND, "vacgt",
5174 "f32", v2i32, v2f32, int_arm_neon_vacgt, 0>;
5175 def VACGTfq : N3VQInt<1, 0, 0b10, 0b1110, 1, N3RegFrm, IIC_VBINQ, "vacgt",
5176 "f32", v4i32, v4f32, int_arm_neon_vacgt, 0>;
5177 def VACGThd : N3VDInt<1, 0, 0b11, 0b1110, 1, N3RegFrm, IIC_VBIND, "vacgt",
5178 "f16", v4i16, v4f16, int_arm_neon_vacgt, 0>,
5179 Requires<[HasNEON, HasFullFP16]>;
5180 def VACGThq : N3VQInt<1, 0, 0b11, 0b1110, 1, N3RegFrm, IIC_VBINQ, "vacgt",
5181 "f16", v8i16, v8f16, int_arm_neon_vacgt, 0>,
5182 Requires<[HasNEON, HasFullFP16]>;
5183 // VTST : Vector Test Bits
5184 defm VTST : N3V_QHS<0, 0, 0b1000, 1, IIC_VBINi4D, IIC_VBINi4D, IIC_VBINi4Q,
5185 IIC_VBINi4Q, "vtst", "", NEONvtst, 1>;
5187 def: NEONInstAlias<"vaclt${p}.f32 $Vd, $Vn, $Vm",
5188 (VACGTfd DPR:$Vd, DPR:$Vm, DPR:$Vn, pred:$p)>;
5189 def: NEONInstAlias<"vaclt${p}.f32 $Vd, $Vn, $Vm",
5190 (VACGTfq QPR:$Vd, QPR:$Vm, QPR:$Vn, pred:$p)>;
5191 def: NEONInstAlias<"vacle${p}.f32 $Vd, $Vn, $Vm",
5192 (VACGEfd DPR:$Vd, DPR:$Vm, DPR:$Vn, pred:$p)>;
5193 def: NEONInstAlias<"vacle${p}.f32 $Vd, $Vn, $Vm",
5194 (VACGEfq QPR:$Vd, QPR:$Vm, QPR:$Vn, pred:$p)>;
5195 let Predicates = [HasNEON, HasFullFP16] in {
5196 def: NEONInstAlias<"vaclt${p}.f16 $Vd, $Vn, $Vm",
5197 (VACGThd DPR:$Vd, DPR:$Vm, DPR:$Vn, pred:$p)>;
5198 def: NEONInstAlias<"vaclt${p}.f16 $Vd, $Vn, $Vm",
5199 (VACGThq QPR:$Vd, QPR:$Vm, QPR:$Vn, pred:$p)>;
5200 def: NEONInstAlias<"vacle${p}.f16 $Vd, $Vn, $Vm",
5201 (VACGEhd DPR:$Vd, DPR:$Vm, DPR:$Vn, pred:$p)>;
5202 def: NEONInstAlias<"vacle${p}.f16 $Vd, $Vn, $Vm",
5203 (VACGEhq QPR:$Vd, QPR:$Vm, QPR:$Vn, pred:$p)>;
5206 // +fp16fml Floating Point Multiplication Variants
5207 let Predicates = [HasNEON, HasFP16FML], DecoderNamespace= "VFPV8" in {
5209 class N3VCP8F16Q1<string asm, RegisterClass Td, RegisterClass Tn,
5210 RegisterClass Tm, bits<2> op1, bits<2> op2, bit op3>
5211 : N3VCP8<op1, op2, 1, op3, (outs Td:$Vd), (ins Tn:$Vn, Tm:$Vm), NoItinerary,
5212 asm, "f16", "$Vd, $Vn, $Vm", "", []>;
5214 class N3VCP8F16Q0<string asm, RegisterClass Td, RegisterClass Tn,
5215 RegisterClass Tm, bits<2> op1, bits<2> op2, bit op3>
5216 : N3VCP8Q0<op1, op2, 0, op3, (outs Td:$Vd), (ins Tn:$Vn, Tm:$Vm), NoItinerary,
5217 asm, "f16", "$Vd, $Vn, $Vm", "", []>;
5219 // Vd, Vs, Vs[0-15], Idx[0-1]
5220 class VFMD<string opc, string type, bits<2> S>
5221 : N3VLaneCP8<0, S, 0, 1, (outs DPR:$Vd),
5222 (ins SPR:$Vn, SPR_8:$Vm, VectorIndex32:$idx),
5223 IIC_VMACD, opc, type, "$Vd, $Vn, $Vm$idx", "", []> {
5226 let Inst{19-16} = Vn{4-1};
5227 let Inst{7} = Vn{0};
5228 let Inst{5} = Vm{0};
5229 let Inst{2-0} = Vm{3-1};
5232 // Vq, Vd, Vd[0-7], Idx[0-3]
5233 class VFMQ<string opc, string type, bits<2> S>
5234 : N3VLaneCP8<0, S, 1, 1, (outs QPR:$Vd),
5235 (ins DPR:$Vn, DPR_8:$Vm, VectorIndex16:$idx),
5236 IIC_VMACD, opc, type, "$Vd, $Vn, $Vm$idx", "", []> {
5238 let Inst{5} = idx{1};
5239 let Inst{3} = idx{0};
5243 def VFMALD : N3VCP8F16Q0<"vfmal", DPR, SPR, SPR, 0b00, 0b10, 1>;
5244 def VFMSLD : N3VCP8F16Q0<"vfmsl", DPR, SPR, SPR, 0b01, 0b10, 1>;
5245 def VFMALQ : N3VCP8F16Q1<"vfmal", QPR, DPR, DPR, 0b00, 0b10, 1>;
5246 def VFMSLQ : N3VCP8F16Q1<"vfmsl", QPR, DPR, DPR, 0b01, 0b10, 1>;
5247 def VFMALDI : VFMD<"vfmal", "f16", 0b00>;
5248 def VFMSLDI : VFMD<"vfmsl", "f16", 0b01>;
5249 def VFMALQI : VFMQ<"vfmal", "f16", 0b00>;
5250 def VFMSLQI : VFMQ<"vfmsl", "f16", 0b01>;
5251 } // HasNEON, HasFP16FML
5254 def: NEONInstAlias<"vaclt${p}.f32 $Vd, $Vm",
5255 (VACGTfd DPR:$Vd, DPR:$Vm, DPR:$Vd, pred:$p)>;
5256 def: NEONInstAlias<"vaclt${p}.f32 $Vd, $Vm",
5257 (VACGTfq QPR:$Vd, QPR:$Vm, QPR:$Vd, pred:$p)>;
5258 def: NEONInstAlias<"vacle${p}.f32 $Vd, $Vm",
5259 (VACGEfd DPR:$Vd, DPR:$Vm, DPR:$Vd, pred:$p)>;
5260 def: NEONInstAlias<"vacle${p}.f32 $Vd, $Vm",
5261 (VACGEfq QPR:$Vd, QPR:$Vm, QPR:$Vd, pred:$p)>;
5262 let Predicates = [HasNEON, HasFullFP16] in {
5263 def: NEONInstAlias<"vaclt${p}.f16 $Vd, $Vm",
5264 (VACGThd DPR:$Vd, DPR:$Vm, DPR:$Vd, pred:$p)>;
5265 def: NEONInstAlias<"vaclt${p}.f16 $Vd, $Vm",
5266 (VACGThq QPR:$Vd, QPR:$Vm, QPR:$Vd, pred:$p)>;
5267 def: NEONInstAlias<"vacle${p}.f16 $Vd, $Vm",
5268 (VACGEhd DPR:$Vd, DPR:$Vm, DPR:$Vd, pred:$p)>;
5269 def: NEONInstAlias<"vacle${p}.f16 $Vd, $Vm",
5270 (VACGEhq QPR:$Vd, QPR:$Vm, QPR:$Vd, pred:$p)>;
5273 // Vector Bitwise Operations.
5275 def vnotd : PatFrag<(ops node:$in),
5276 (xor node:$in, (bitconvert (v8i8 NEONimmAllOnesV)))>;
5277 def vnotq : PatFrag<(ops node:$in),
5278 (xor node:$in, (bitconvert (v16i8 NEONimmAllOnesV)))>;
5281 // VAND : Vector Bitwise AND
5282 def VANDd : N3VDX<0, 0, 0b00, 0b0001, 1, IIC_VBINiD, "vand",
5283 v2i32, v2i32, and, 1>;
5284 def VANDq : N3VQX<0, 0, 0b00, 0b0001, 1, IIC_VBINiQ, "vand",
5285 v4i32, v4i32, and, 1>;
5287 // VEOR : Vector Bitwise Exclusive OR
5288 def VEORd : N3VDX<1, 0, 0b00, 0b0001, 1, IIC_VBINiD, "veor",
5289 v2i32, v2i32, xor, 1>;
5290 def VEORq : N3VQX<1, 0, 0b00, 0b0001, 1, IIC_VBINiQ, "veor",
5291 v4i32, v4i32, xor, 1>;
5293 // VORR : Vector Bitwise OR
5294 def VORRd : N3VDX<0, 0, 0b10, 0b0001, 1, IIC_VBINiD, "vorr",
5295 v2i32, v2i32, or, 1>;
5296 def VORRq : N3VQX<0, 0, 0b10, 0b0001, 1, IIC_VBINiQ, "vorr",
5297 v4i32, v4i32, or, 1>;
5299 def VORRiv4i16 : N1ModImm<1, 0b000, {1,0,?,1}, 0, 0, 0, 1,
5300 (outs DPR:$Vd), (ins nImmSplatI16:$SIMM, DPR:$src),
5302 "vorr", "i16", "$Vd, $SIMM", "$src = $Vd",
5304 (v4i16 (ARMvorrImm DPR:$src, timm:$SIMM)))]> {
5305 let Inst{9} = SIMM{9};
5308 def VORRiv2i32 : N1ModImm<1, 0b000, {0,?,?,1}, 0, 0, 0, 1,
5309 (outs DPR:$Vd), (ins nImmSplatI32:$SIMM, DPR:$src),
5311 "vorr", "i32", "$Vd, $SIMM", "$src = $Vd",
5313 (v2i32 (ARMvorrImm DPR:$src, timm:$SIMM)))]> {
5314 let Inst{10-9} = SIMM{10-9};
5317 def VORRiv8i16 : N1ModImm<1, 0b000, {1,0,?,1}, 0, 1, 0, 1,
5318 (outs QPR:$Vd), (ins nImmSplatI16:$SIMM, QPR:$src),
5320 "vorr", "i16", "$Vd, $SIMM", "$src = $Vd",
5322 (v8i16 (ARMvorrImm QPR:$src, timm:$SIMM)))]> {
5323 let Inst{9} = SIMM{9};
5326 def VORRiv4i32 : N1ModImm<1, 0b000, {0,?,?,1}, 0, 1, 0, 1,
5327 (outs QPR:$Vd), (ins nImmSplatI32:$SIMM, QPR:$src),
5329 "vorr", "i32", "$Vd, $SIMM", "$src = $Vd",
5331 (v4i32 (ARMvorrImm QPR:$src, timm:$SIMM)))]> {
5332 let Inst{10-9} = SIMM{10-9};
5336 // VBIC : Vector Bitwise Bit Clear (AND NOT)
5337 let TwoOperandAliasConstraint = "$Vn = $Vd" in {
5338 def VBICd : N3VX<0, 0, 0b01, 0b0001, 0, 1, (outs DPR:$Vd),
5339 (ins DPR:$Vn, DPR:$Vm), N3RegFrm, IIC_VBINiD,
5340 "vbic", "$Vd, $Vn, $Vm", "",
5341 [(set DPR:$Vd, (v2i32 (and DPR:$Vn,
5342 (vnotd DPR:$Vm))))]>;
5343 def VBICq : N3VX<0, 0, 0b01, 0b0001, 1, 1, (outs QPR:$Vd),
5344 (ins QPR:$Vn, QPR:$Vm), N3RegFrm, IIC_VBINiQ,
5345 "vbic", "$Vd, $Vn, $Vm", "",
5346 [(set QPR:$Vd, (v4i32 (and QPR:$Vn,
5347 (vnotq QPR:$Vm))))]>;
5350 def VBICiv4i16 : N1ModImm<1, 0b000, {1,0,?,1}, 0, 0, 1, 1,
5351 (outs DPR:$Vd), (ins nImmSplatI16:$SIMM, DPR:$src),
5353 "vbic", "i16", "$Vd, $SIMM", "$src = $Vd",
5355 (v4i16 (ARMvbicImm DPR:$src, timm:$SIMM)))]> {
5356 let Inst{9} = SIMM{9};
5359 def VBICiv2i32 : N1ModImm<1, 0b000, {0,?,?,1}, 0, 0, 1, 1,
5360 (outs DPR:$Vd), (ins nImmSplatI32:$SIMM, DPR:$src),
5362 "vbic", "i32", "$Vd, $SIMM", "$src = $Vd",
5364 (v2i32 (ARMvbicImm DPR:$src, timm:$SIMM)))]> {
5365 let Inst{10-9} = SIMM{10-9};
5368 def VBICiv8i16 : N1ModImm<1, 0b000, {1,0,?,1}, 0, 1, 1, 1,
5369 (outs QPR:$Vd), (ins nImmSplatI16:$SIMM, QPR:$src),
5371 "vbic", "i16", "$Vd, $SIMM", "$src = $Vd",
5373 (v8i16 (ARMvbicImm QPR:$src, timm:$SIMM)))]> {
5374 let Inst{9} = SIMM{9};
5377 def VBICiv4i32 : N1ModImm<1, 0b000, {0,?,?,1}, 0, 1, 1, 1,
5378 (outs QPR:$Vd), (ins nImmSplatI32:$SIMM, QPR:$src),
5380 "vbic", "i32", "$Vd, $SIMM", "$src = $Vd",
5382 (v4i32 (ARMvbicImm QPR:$src, timm:$SIMM)))]> {
5383 let Inst{10-9} = SIMM{10-9};
5386 // VORN : Vector Bitwise OR NOT
5387 def VORNd : N3VX<0, 0, 0b11, 0b0001, 0, 1, (outs DPR:$Vd),
5388 (ins DPR:$Vn, DPR:$Vm), N3RegFrm, IIC_VBINiD,
5389 "vorn", "$Vd, $Vn, $Vm", "",
5390 [(set DPR:$Vd, (v2i32 (or DPR:$Vn,
5391 (vnotd DPR:$Vm))))]>;
5392 def VORNq : N3VX<0, 0, 0b11, 0b0001, 1, 1, (outs QPR:$Vd),
5393 (ins QPR:$Vn, QPR:$Vm), N3RegFrm, IIC_VBINiQ,
5394 "vorn", "$Vd, $Vn, $Vm", "",
5395 [(set QPR:$Vd, (v4i32 (or QPR:$Vn,
5396 (vnotq QPR:$Vm))))]>;
5398 // VMVN : Vector Bitwise NOT (Immediate)
5400 let isReMaterializable = 1 in {
5402 def VMVNv4i16 : N1ModImm<1, 0b000, {1,0,?,0}, 0, 0, 1, 1, (outs DPR:$Vd),
5403 (ins nImmSplatI16:$SIMM), IIC_VMOVImm,
5404 "vmvn", "i16", "$Vd, $SIMM", "",
5405 [(set DPR:$Vd, (v4i16 (ARMvmvnImm timm:$SIMM)))]> {
5406 let Inst{9} = SIMM{9};
5409 def VMVNv8i16 : N1ModImm<1, 0b000, {1,0,?,0}, 0, 1, 1, 1, (outs QPR:$Vd),
5410 (ins nImmSplatI16:$SIMM), IIC_VMOVImm,
5411 "vmvn", "i16", "$Vd, $SIMM", "",
5412 [(set QPR:$Vd, (v8i16 (ARMvmvnImm timm:$SIMM)))]> {
5413 let Inst{9} = SIMM{9};
5416 def VMVNv2i32 : N1ModImm<1, 0b000, {?,?,?,?}, 0, 0, 1, 1, (outs DPR:$Vd),
5417 (ins nImmVMOVI32:$SIMM), IIC_VMOVImm,
5418 "vmvn", "i32", "$Vd, $SIMM", "",
5419 [(set DPR:$Vd, (v2i32 (ARMvmvnImm timm:$SIMM)))]> {
5420 let Inst{11-8} = SIMM{11-8};
5423 def VMVNv4i32 : N1ModImm<1, 0b000, {?,?,?,?}, 0, 1, 1, 1, (outs QPR:$Vd),
5424 (ins nImmVMOVI32:$SIMM), IIC_VMOVImm,
5425 "vmvn", "i32", "$Vd, $SIMM", "",
5426 [(set QPR:$Vd, (v4i32 (ARMvmvnImm timm:$SIMM)))]> {
5427 let Inst{11-8} = SIMM{11-8};
5431 // VMVN : Vector Bitwise NOT
5432 def VMVNd : N2VX<0b11, 0b11, 0b00, 0b00, 0b01011, 0, 0,
5433 (outs DPR:$Vd), (ins DPR:$Vm), IIC_VSUBiD,
5434 "vmvn", "$Vd, $Vm", "",
5435 [(set DPR:$Vd, (v2i32 (vnotd DPR:$Vm)))]>;
5436 def VMVNq : N2VX<0b11, 0b11, 0b00, 0b00, 0b01011, 1, 0,
5437 (outs QPR:$Vd), (ins QPR:$Vm), IIC_VSUBiD,
5438 "vmvn", "$Vd, $Vm", "",
5439 [(set QPR:$Vd, (v4i32 (vnotq QPR:$Vm)))]>;
5440 let Predicates = [HasNEON] in {
5441 def : Pat<(v2i32 (vnotd DPR:$src)), (VMVNd DPR:$src)>;
5442 def : Pat<(v4i32 (vnotq QPR:$src)), (VMVNq QPR:$src)>;
5445 // VBSL : Vector Bitwise Select
5446 def VBSLd : N3VX<1, 0, 0b01, 0b0001, 0, 1, (outs DPR:$Vd),
5447 (ins DPR:$src1, DPR:$Vn, DPR:$Vm),
5448 N3RegFrm, IIC_VCNTiD,
5449 "vbsl", "$Vd, $Vn, $Vm", "$src1 = $Vd",
5451 (v2i32 (NEONvbsl DPR:$src1, DPR:$Vn, DPR:$Vm)))]>;
5452 let Predicates = [HasNEON] in {
5453 def : Pat<(v8i8 (int_arm_neon_vbsl (v8i8 DPR:$src1),
5454 (v8i8 DPR:$Vn), (v8i8 DPR:$Vm))),
5455 (VBSLd DPR:$src1, DPR:$Vn, DPR:$Vm)>;
5456 def : Pat<(v4i16 (int_arm_neon_vbsl (v4i16 DPR:$src1),
5457 (v4i16 DPR:$Vn), (v4i16 DPR:$Vm))),
5458 (VBSLd DPR:$src1, DPR:$Vn, DPR:$Vm)>;
5459 def : Pat<(v2i32 (int_arm_neon_vbsl (v2i32 DPR:$src1),
5460 (v2i32 DPR:$Vn), (v2i32 DPR:$Vm))),
5461 (VBSLd DPR:$src1, DPR:$Vn, DPR:$Vm)>;
5462 def : Pat<(v2f32 (int_arm_neon_vbsl (v2f32 DPR:$src1),
5463 (v2f32 DPR:$Vn), (v2f32 DPR:$Vm))),
5464 (VBSLd DPR:$src1, DPR:$Vn, DPR:$Vm)>;
5465 def : Pat<(v1i64 (int_arm_neon_vbsl (v1i64 DPR:$src1),
5466 (v1i64 DPR:$Vn), (v1i64 DPR:$Vm))),
5467 (VBSLd DPR:$src1, DPR:$Vn, DPR:$Vm)>;
5469 def : Pat<(v2i32 (or (and DPR:$Vn, DPR:$Vd),
5470 (and DPR:$Vm, (vnotd DPR:$Vd)))),
5471 (VBSLd DPR:$Vd, DPR:$Vn, DPR:$Vm)>;
5473 def : Pat<(v1i64 (or (and DPR:$Vn, DPR:$Vd),
5474 (and DPR:$Vm, (vnotd DPR:$Vd)))),
5475 (VBSLd DPR:$Vd, DPR:$Vn, DPR:$Vm)>;
5478 def VBSLq : N3VX<1, 0, 0b01, 0b0001, 1, 1, (outs QPR:$Vd),
5479 (ins QPR:$src1, QPR:$Vn, QPR:$Vm),
5480 N3RegFrm, IIC_VCNTiQ,
5481 "vbsl", "$Vd, $Vn, $Vm", "$src1 = $Vd",
5483 (v4i32 (NEONvbsl QPR:$src1, QPR:$Vn, QPR:$Vm)))]>;
5485 let Predicates = [HasNEON] in {
5486 def : Pat<(v16i8 (int_arm_neon_vbsl (v16i8 QPR:$src1),
5487 (v16i8 QPR:$Vn), (v16i8 QPR:$Vm))),
5488 (VBSLq QPR:$src1, QPR:$Vn, QPR:$Vm)>;
5489 def : Pat<(v8i16 (int_arm_neon_vbsl (v8i16 QPR:$src1),
5490 (v8i16 QPR:$Vn), (v8i16 QPR:$Vm))),
5491 (VBSLq QPR:$src1, QPR:$Vn, QPR:$Vm)>;
5492 def : Pat<(v4i32 (int_arm_neon_vbsl (v4i32 QPR:$src1),
5493 (v4i32 QPR:$Vn), (v4i32 QPR:$Vm))),
5494 (VBSLq QPR:$src1, QPR:$Vn, QPR:$Vm)>;
5495 def : Pat<(v4f32 (int_arm_neon_vbsl (v4f32 QPR:$src1),
5496 (v4f32 QPR:$Vn), (v4f32 QPR:$Vm))),
5497 (VBSLq QPR:$src1, QPR:$Vn, QPR:$Vm)>;
5498 def : Pat<(v2i64 (int_arm_neon_vbsl (v2i64 QPR:$src1),
5499 (v2i64 QPR:$Vn), (v2i64 QPR:$Vm))),
5500 (VBSLq QPR:$src1, QPR:$Vn, QPR:$Vm)>;
5502 def : Pat<(v4i32 (or (and QPR:$Vn, QPR:$Vd),
5503 (and QPR:$Vm, (vnotq QPR:$Vd)))),
5504 (VBSLq QPR:$Vd, QPR:$Vn, QPR:$Vm)>;
5505 def : Pat<(v2i64 (or (and QPR:$Vn, QPR:$Vd),
5506 (and QPR:$Vm, (vnotq QPR:$Vd)))),
5507 (VBSLq QPR:$Vd, QPR:$Vn, QPR:$Vm)>;
5510 // VBIF : Vector Bitwise Insert if False
5511 // like VBSL but with: "vbif $dst, $src3, $src1", "$src2 = $dst",
5512 // FIXME: This instruction's encoding MAY NOT BE correct.
5513 def VBIFd : N3VX<1, 0, 0b11, 0b0001, 0, 1,
5514 (outs DPR:$Vd), (ins DPR:$src1, DPR:$Vn, DPR:$Vm),
5515 N3RegFrm, IIC_VBINiD,
5516 "vbif", "$Vd, $Vn, $Vm", "$src1 = $Vd",
5518 def VBIFq : N3VX<1, 0, 0b11, 0b0001, 1, 1,
5519 (outs QPR:$Vd), (ins QPR:$src1, QPR:$Vn, QPR:$Vm),
5520 N3RegFrm, IIC_VBINiQ,
5521 "vbif", "$Vd, $Vn, $Vm", "$src1 = $Vd",
5524 // VBIT : Vector Bitwise Insert if True
5525 // like VBSL but with: "vbit $dst, $src2, $src1", "$src3 = $dst",
5526 // FIXME: This instruction's encoding MAY NOT BE correct.
5527 def VBITd : N3VX<1, 0, 0b10, 0b0001, 0, 1,
5528 (outs DPR:$Vd), (ins DPR:$src1, DPR:$Vn, DPR:$Vm),
5529 N3RegFrm, IIC_VBINiD,
5530 "vbit", "$Vd, $Vn, $Vm", "$src1 = $Vd",
5532 def VBITq : N3VX<1, 0, 0b10, 0b0001, 1, 1,
5533 (outs QPR:$Vd), (ins QPR:$src1, QPR:$Vn, QPR:$Vm),
5534 N3RegFrm, IIC_VBINiQ,
5535 "vbit", "$Vd, $Vn, $Vm", "$src1 = $Vd",
5538 // VBIT/VBIF are not yet implemented. The TwoAddress pass will not go looking
5539 // for equivalent operations with different register constraints; it just
5542 // Vector Absolute Differences.
5544 // VABD : Vector Absolute Difference
5545 defm VABDs : N3VInt_QHS<0, 0, 0b0111, 0, N3RegFrm,
5546 IIC_VSUBi4D, IIC_VSUBi4D, IIC_VSUBi4Q, IIC_VSUBi4Q,
5547 "vabd", "s", int_arm_neon_vabds, 1>;
5548 defm VABDu : N3VInt_QHS<1, 0, 0b0111, 0, N3RegFrm,
5549 IIC_VSUBi4D, IIC_VSUBi4D, IIC_VSUBi4Q, IIC_VSUBi4Q,
5550 "vabd", "u", int_arm_neon_vabdu, 1>;
5551 def VABDfd : N3VDInt<1, 0, 0b10, 0b1101, 0, N3RegFrm, IIC_VBIND,
5552 "vabd", "f32", v2f32, v2f32, int_arm_neon_vabds, 1>;
5553 def VABDfq : N3VQInt<1, 0, 0b10, 0b1101, 0, N3RegFrm, IIC_VBINQ,
5554 "vabd", "f32", v4f32, v4f32, int_arm_neon_vabds, 1>;
5555 def VABDhd : N3VDInt<1, 0, 0b11, 0b1101, 0, N3RegFrm, IIC_VBIND,
5556 "vabd", "f16", v4f16, v4f16, int_arm_neon_vabds, 1>,
5557 Requires<[HasNEON, HasFullFP16]>;
5558 def VABDhq : N3VQInt<1, 0, 0b11, 0b1101, 0, N3RegFrm, IIC_VBINQ,
5559 "vabd", "f16", v8f16, v8f16, int_arm_neon_vabds, 1>,
5560 Requires<[HasNEON, HasFullFP16]>;
5562 // VABDL : Vector Absolute Difference Long (Q = | D - D |)
5563 defm VABDLs : N3VLIntExt_QHS<0,1,0b0111,0, IIC_VSUBi4Q,
5564 "vabdl", "s", int_arm_neon_vabds, zext, 1>;
5565 defm VABDLu : N3VLIntExt_QHS<1,1,0b0111,0, IIC_VSUBi4Q,
5566 "vabdl", "u", int_arm_neon_vabdu, zext, 1>;
5568 let Predicates = [HasNEON] in {
5569 def : Pat<(v8i16 (abs (sub (zext (v8i8 DPR:$opA)), (zext (v8i8 DPR:$opB))))),
5570 (VABDLuv8i16 DPR:$opA, DPR:$opB)>;
5571 def : Pat<(v4i32 (abs (sub (zext (v4i16 DPR:$opA)), (zext (v4i16 DPR:$opB))))),
5572 (VABDLuv4i32 DPR:$opA, DPR:$opB)>;
5575 // ISD::ABS is not legal for v2i64, so VABDL needs to be matched from the
5576 // shift/xor pattern for ABS.
5579 PatFrag<(ops node:$in1, node:$in2, node:$shift),
5580 (ARMvshrsImm (sub (zext node:$in1),
5581 (zext node:$in2)), (i32 $shift))>;
5583 let Predicates = [HasNEON] in {
5584 def : Pat<(xor (v4i32 (bitconvert (v2i64 (abd_shr (v2i32 DPR:$opA), (v2i32 DPR:$opB), 63)))),
5585 (v4i32 (bitconvert (v2i64 (add (sub (zext (v2i32 DPR:$opA)),
5586 (zext (v2i32 DPR:$opB))),
5587 (abd_shr (v2i32 DPR:$opA), (v2i32 DPR:$opB), 63)))))),
5588 (VABDLuv2i64 DPR:$opA, DPR:$opB)>;
5591 // VABA : Vector Absolute Difference and Accumulate
5592 defm VABAs : N3VIntOp_QHS<0,0,0b0111,1, IIC_VABAD, IIC_VABAQ,
5593 "vaba", "s", int_arm_neon_vabds, add>;
5594 defm VABAu : N3VIntOp_QHS<1,0,0b0111,1, IIC_VABAD, IIC_VABAQ,
5595 "vaba", "u", int_arm_neon_vabdu, add>;
5597 // VABAL : Vector Absolute Difference and Accumulate Long (Q += | D - D |)
5598 defm VABALs : N3VLIntExtOp_QHS<0,1,0b0101,0, IIC_VABAD,
5599 "vabal", "s", int_arm_neon_vabds, zext, add>;
5600 defm VABALu : N3VLIntExtOp_QHS<1,1,0b0101,0, IIC_VABAD,
5601 "vabal", "u", int_arm_neon_vabdu, zext, add>;
5603 // Vector Maximum and Minimum.
5605 // VMAX : Vector Maximum
5606 defm VMAXs : N3VInt_QHS<0, 0, 0b0110, 0, N3RegFrm,
5607 IIC_VSUBi4D, IIC_VSUBi4D, IIC_VSUBi4Q, IIC_VSUBi4Q,
5608 "vmax", "s", smax, 1>;
5609 defm VMAXu : N3VInt_QHS<1, 0, 0b0110, 0, N3RegFrm,
5610 IIC_VSUBi4D, IIC_VSUBi4D, IIC_VSUBi4Q, IIC_VSUBi4Q,
5611 "vmax", "u", umax, 1>;
5612 def VMAXfd : N3VDInt<0, 0, 0b00, 0b1111, 0, N3RegFrm, IIC_VBIND,
5614 v2f32, v2f32, fmaximum, 1>;
5615 def VMAXfq : N3VQInt<0, 0, 0b00, 0b1111, 0, N3RegFrm, IIC_VBINQ,
5617 v4f32, v4f32, fmaximum, 1>;
5618 def VMAXhd : N3VDInt<0, 0, 0b01, 0b1111, 0, N3RegFrm, IIC_VBIND,
5620 v4f16, v4f16, fmaximum, 1>,
5621 Requires<[HasNEON, HasFullFP16]>;
5622 def VMAXhq : N3VQInt<0, 0, 0b01, 0b1111, 0, N3RegFrm, IIC_VBINQ,
5624 v8f16, v8f16, fmaximum, 1>,
5625 Requires<[HasNEON, HasFullFP16]>;
5628 let PostEncoderMethod = "NEONThumb2V8PostEncoder", DecoderNamespace = "v8NEON" in {
5629 def NEON_VMAXNMNDf : N3VDIntnp<0b00110, 0b00, 0b1111, 0, 1,
5630 N3RegFrm, NoItinerary, "vmaxnm", "f32",
5631 v2f32, v2f32, fmaxnum, 1>,
5632 Requires<[HasV8, HasNEON]>;
5633 def NEON_VMAXNMNQf : N3VQIntnp<0b00110, 0b00, 0b1111, 1, 1,
5634 N3RegFrm, NoItinerary, "vmaxnm", "f32",
5635 v4f32, v4f32, fmaxnum, 1>,
5636 Requires<[HasV8, HasNEON]>;
5637 def NEON_VMAXNMNDh : N3VDIntnp<0b00110, 0b01, 0b1111, 0, 1,
5638 N3RegFrm, NoItinerary, "vmaxnm", "f16",
5639 v4f16, v4f16, fmaxnum, 1>,
5640 Requires<[HasV8, HasNEON, HasFullFP16]>;
5641 def NEON_VMAXNMNQh : N3VQIntnp<0b00110, 0b01, 0b1111, 1, 1,
5642 N3RegFrm, NoItinerary, "vmaxnm", "f16",
5643 v8f16, v8f16, fmaxnum, 1>,
5644 Requires<[HasV8, HasNEON, HasFullFP16]>;
5647 // VMIN : Vector Minimum
5648 defm VMINs : N3VInt_QHS<0, 0, 0b0110, 1, N3RegFrm,
5649 IIC_VSUBi4D, IIC_VSUBi4D, IIC_VSUBi4Q, IIC_VSUBi4Q,
5650 "vmin", "s", smin, 1>;
5651 defm VMINu : N3VInt_QHS<1, 0, 0b0110, 1, N3RegFrm,
5652 IIC_VSUBi4D, IIC_VSUBi4D, IIC_VSUBi4Q, IIC_VSUBi4Q,
5653 "vmin", "u", umin, 1>;
5654 def VMINfd : N3VDInt<0, 0, 0b10, 0b1111, 0, N3RegFrm, IIC_VBIND,
5656 v2f32, v2f32, fminimum, 1>;
5657 def VMINfq : N3VQInt<0, 0, 0b10, 0b1111, 0, N3RegFrm, IIC_VBINQ,
5659 v4f32, v4f32, fminimum, 1>;
5660 def VMINhd : N3VDInt<0, 0, 0b11, 0b1111, 0, N3RegFrm, IIC_VBIND,
5662 v4f16, v4f16, fminimum, 1>,
5663 Requires<[HasNEON, HasFullFP16]>;
5664 def VMINhq : N3VQInt<0, 0, 0b11, 0b1111, 0, N3RegFrm, IIC_VBINQ,
5666 v8f16, v8f16, fminimum, 1>,
5667 Requires<[HasNEON, HasFullFP16]>;
5670 let PostEncoderMethod = "NEONThumb2V8PostEncoder", DecoderNamespace = "v8NEON" in {
5671 def NEON_VMINNMNDf : N3VDIntnp<0b00110, 0b10, 0b1111, 0, 1,
5672 N3RegFrm, NoItinerary, "vminnm", "f32",
5673 v2f32, v2f32, fminnum, 1>,
5674 Requires<[HasV8, HasNEON]>;
5675 def NEON_VMINNMNQf : N3VQIntnp<0b00110, 0b10, 0b1111, 1, 1,
5676 N3RegFrm, NoItinerary, "vminnm", "f32",
5677 v4f32, v4f32, fminnum, 1>,
5678 Requires<[HasV8, HasNEON]>;
5679 def NEON_VMINNMNDh : N3VDIntnp<0b00110, 0b11, 0b1111, 0, 1,
5680 N3RegFrm, NoItinerary, "vminnm", "f16",
5681 v4f16, v4f16, fminnum, 1>,
5682 Requires<[HasV8, HasNEON, HasFullFP16]>;
5683 def NEON_VMINNMNQh : N3VQIntnp<0b00110, 0b11, 0b1111, 1, 1,
5684 N3RegFrm, NoItinerary, "vminnm", "f16",
5685 v8f16, v8f16, fminnum, 1>,
5686 Requires<[HasV8, HasNEON, HasFullFP16]>;
5689 // Vector Pairwise Operations.
5691 // VPADD : Vector Pairwise Add
5692 def VPADDi8 : N3VDInt<0, 0, 0b00, 0b1011, 1, N3RegFrm, IIC_VSHLiD,
5694 v8i8, v8i8, int_arm_neon_vpadd, 0>;
5695 def VPADDi16 : N3VDInt<0, 0, 0b01, 0b1011, 1, N3RegFrm, IIC_VSHLiD,
5697 v4i16, v4i16, int_arm_neon_vpadd, 0>;
5698 def VPADDi32 : N3VDInt<0, 0, 0b10, 0b1011, 1, N3RegFrm, IIC_VSHLiD,
5700 v2i32, v2i32, int_arm_neon_vpadd, 0>;
5701 def VPADDf : N3VDInt<1, 0, 0b00, 0b1101, 0, N3RegFrm,
5702 IIC_VPBIND, "vpadd", "f32",
5703 v2f32, v2f32, int_arm_neon_vpadd, 0>;
5704 def VPADDh : N3VDInt<1, 0, 0b01, 0b1101, 0, N3RegFrm,
5705 IIC_VPBIND, "vpadd", "f16",
5706 v4f16, v4f16, int_arm_neon_vpadd, 0>,
5707 Requires<[HasNEON, HasFullFP16]>;
5709 // VPADDL : Vector Pairwise Add Long
5710 defm VPADDLs : N2VPLInt_QHS<0b11, 0b11, 0b00, 0b00100, 0, "vpaddl", "s",
5711 int_arm_neon_vpaddls>;
5712 defm VPADDLu : N2VPLInt_QHS<0b11, 0b11, 0b00, 0b00101, 0, "vpaddl", "u",
5713 int_arm_neon_vpaddlu>;
5715 // VPADAL : Vector Pairwise Add and Accumulate Long
5716 defm VPADALs : N2VPLInt2_QHS<0b11, 0b11, 0b00, 0b01100, 0, "vpadal", "s",
5717 int_arm_neon_vpadals>;
5718 defm VPADALu : N2VPLInt2_QHS<0b11, 0b11, 0b00, 0b01101, 0, "vpadal", "u",
5719 int_arm_neon_vpadalu>;
5721 // VPMAX : Vector Pairwise Maximum
5722 def VPMAXs8 : N3VDInt<0, 0, 0b00, 0b1010, 0, N3RegFrm, IIC_VSUBi4D, "vpmax",
5723 "s8", v8i8, v8i8, int_arm_neon_vpmaxs, 0>;
5724 def VPMAXs16 : N3VDInt<0, 0, 0b01, 0b1010, 0, N3RegFrm, IIC_VSUBi4D, "vpmax",
5725 "s16", v4i16, v4i16, int_arm_neon_vpmaxs, 0>;
5726 def VPMAXs32 : N3VDInt<0, 0, 0b10, 0b1010, 0, N3RegFrm, IIC_VSUBi4D, "vpmax",
5727 "s32", v2i32, v2i32, int_arm_neon_vpmaxs, 0>;
5728 def VPMAXu8 : N3VDInt<1, 0, 0b00, 0b1010, 0, N3RegFrm, IIC_VSUBi4D, "vpmax",
5729 "u8", v8i8, v8i8, int_arm_neon_vpmaxu, 0>;
5730 def VPMAXu16 : N3VDInt<1, 0, 0b01, 0b1010, 0, N3RegFrm, IIC_VSUBi4D, "vpmax",
5731 "u16", v4i16, v4i16, int_arm_neon_vpmaxu, 0>;
5732 def VPMAXu32 : N3VDInt<1, 0, 0b10, 0b1010, 0, N3RegFrm, IIC_VSUBi4D, "vpmax",
5733 "u32", v2i32, v2i32, int_arm_neon_vpmaxu, 0>;
5734 def VPMAXf : N3VDInt<1, 0, 0b00, 0b1111, 0, N3RegFrm, IIC_VPBIND, "vpmax",
5735 "f32", v2f32, v2f32, int_arm_neon_vpmaxs, 0>;
5736 def VPMAXh : N3VDInt<1, 0, 0b01, 0b1111, 0, N3RegFrm, IIC_VPBIND, "vpmax",
5737 "f16", v4f16, v4f16, int_arm_neon_vpmaxs, 0>,
5738 Requires<[HasNEON, HasFullFP16]>;
5740 // VPMIN : Vector Pairwise Minimum
5741 def VPMINs8 : N3VDInt<0, 0, 0b00, 0b1010, 1, N3RegFrm, IIC_VSUBi4D, "vpmin",
5742 "s8", v8i8, v8i8, int_arm_neon_vpmins, 0>;
5743 def VPMINs16 : N3VDInt<0, 0, 0b01, 0b1010, 1, N3RegFrm, IIC_VSUBi4D, "vpmin",
5744 "s16", v4i16, v4i16, int_arm_neon_vpmins, 0>;
5745 def VPMINs32 : N3VDInt<0, 0, 0b10, 0b1010, 1, N3RegFrm, IIC_VSUBi4D, "vpmin",
5746 "s32", v2i32, v2i32, int_arm_neon_vpmins, 0>;
5747 def VPMINu8 : N3VDInt<1, 0, 0b00, 0b1010, 1, N3RegFrm, IIC_VSUBi4D, "vpmin",
5748 "u8", v8i8, v8i8, int_arm_neon_vpminu, 0>;
5749 def VPMINu16 : N3VDInt<1, 0, 0b01, 0b1010, 1, N3RegFrm, IIC_VSUBi4D, "vpmin",
5750 "u16", v4i16, v4i16, int_arm_neon_vpminu, 0>;
5751 def VPMINu32 : N3VDInt<1, 0, 0b10, 0b1010, 1, N3RegFrm, IIC_VSUBi4D, "vpmin",
5752 "u32", v2i32, v2i32, int_arm_neon_vpminu, 0>;
5753 def VPMINf : N3VDInt<1, 0, 0b10, 0b1111, 0, N3RegFrm, IIC_VPBIND, "vpmin",
5754 "f32", v2f32, v2f32, int_arm_neon_vpmins, 0>;
5755 def VPMINh : N3VDInt<1, 0, 0b11, 0b1111, 0, N3RegFrm, IIC_VPBIND, "vpmin",
5756 "f16", v4f16, v4f16, int_arm_neon_vpmins, 0>,
5757 Requires<[HasNEON, HasFullFP16]>;
5759 // Vector Reciprocal and Reciprocal Square Root Estimate and Step.
5761 // VRECPE : Vector Reciprocal Estimate
5762 def VRECPEd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01000, 0,
5763 IIC_VUNAD, "vrecpe", "u32",
5764 v2i32, v2i32, int_arm_neon_vrecpe>;
5765 def VRECPEq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01000, 0,
5766 IIC_VUNAQ, "vrecpe", "u32",
5767 v4i32, v4i32, int_arm_neon_vrecpe>;
5768 def VRECPEfd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01010, 0,
5769 IIC_VUNAD, "vrecpe", "f32",
5770 v2f32, v2f32, int_arm_neon_vrecpe>;
5771 def VRECPEfq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01010, 0,
5772 IIC_VUNAQ, "vrecpe", "f32",
5773 v4f32, v4f32, int_arm_neon_vrecpe>;
5774 def VRECPEhd : N2VDInt<0b11, 0b11, 0b01, 0b11, 0b01010, 0,
5775 IIC_VUNAD, "vrecpe", "f16",
5776 v4f16, v4f16, int_arm_neon_vrecpe>,
5777 Requires<[HasNEON, HasFullFP16]>;
5778 def VRECPEhq : N2VQInt<0b11, 0b11, 0b01, 0b11, 0b01010, 0,
5779 IIC_VUNAQ, "vrecpe", "f16",
5780 v8f16, v8f16, int_arm_neon_vrecpe>,
5781 Requires<[HasNEON, HasFullFP16]>;
5783 // VRECPS : Vector Reciprocal Step
5784 def VRECPSfd : N3VDInt<0, 0, 0b00, 0b1111, 1, N3RegFrm,
5785 IIC_VRECSD, "vrecps", "f32",
5786 v2f32, v2f32, int_arm_neon_vrecps, 1>;
5787 def VRECPSfq : N3VQInt<0, 0, 0b00, 0b1111, 1, N3RegFrm,
5788 IIC_VRECSQ, "vrecps", "f32",
5789 v4f32, v4f32, int_arm_neon_vrecps, 1>;
5790 def VRECPShd : N3VDInt<0, 0, 0b01, 0b1111, 1, N3RegFrm,
5791 IIC_VRECSD, "vrecps", "f16",
5792 v4f16, v4f16, int_arm_neon_vrecps, 1>,
5793 Requires<[HasNEON, HasFullFP16]>;
5794 def VRECPShq : N3VQInt<0, 0, 0b01, 0b1111, 1, N3RegFrm,
5795 IIC_VRECSQ, "vrecps", "f16",
5796 v8f16, v8f16, int_arm_neon_vrecps, 1>,
5797 Requires<[HasNEON, HasFullFP16]>;
5799 // VRSQRTE : Vector Reciprocal Square Root Estimate
5800 def VRSQRTEd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01001, 0,
5801 IIC_VUNAD, "vrsqrte", "u32",
5802 v2i32, v2i32, int_arm_neon_vrsqrte>;
5803 def VRSQRTEq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01001, 0,
5804 IIC_VUNAQ, "vrsqrte", "u32",
5805 v4i32, v4i32, int_arm_neon_vrsqrte>;
5806 def VRSQRTEfd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01011, 0,
5807 IIC_VUNAD, "vrsqrte", "f32",
5808 v2f32, v2f32, int_arm_neon_vrsqrte>;
5809 def VRSQRTEfq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01011, 0,
5810 IIC_VUNAQ, "vrsqrte", "f32",
5811 v4f32, v4f32, int_arm_neon_vrsqrte>;
5812 def VRSQRTEhd : N2VDInt<0b11, 0b11, 0b01, 0b11, 0b01011, 0,
5813 IIC_VUNAD, "vrsqrte", "f16",
5814 v4f16, v4f16, int_arm_neon_vrsqrte>,
5815 Requires<[HasNEON, HasFullFP16]>;
5816 def VRSQRTEhq : N2VQInt<0b11, 0b11, 0b01, 0b11, 0b01011, 0,
5817 IIC_VUNAQ, "vrsqrte", "f16",
5818 v8f16, v8f16, int_arm_neon_vrsqrte>,
5819 Requires<[HasNEON, HasFullFP16]>;
5821 // VRSQRTS : Vector Reciprocal Square Root Step
5822 def VRSQRTSfd : N3VDInt<0, 0, 0b10, 0b1111, 1, N3RegFrm,
5823 IIC_VRECSD, "vrsqrts", "f32",
5824 v2f32, v2f32, int_arm_neon_vrsqrts, 1>;
5825 def VRSQRTSfq : N3VQInt<0, 0, 0b10, 0b1111, 1, N3RegFrm,
5826 IIC_VRECSQ, "vrsqrts", "f32",
5827 v4f32, v4f32, int_arm_neon_vrsqrts, 1>;
5828 def VRSQRTShd : N3VDInt<0, 0, 0b11, 0b1111, 1, N3RegFrm,
5829 IIC_VRECSD, "vrsqrts", "f16",
5830 v4f16, v4f16, int_arm_neon_vrsqrts, 1>,
5831 Requires<[HasNEON, HasFullFP16]>;
5832 def VRSQRTShq : N3VQInt<0, 0, 0b11, 0b1111, 1, N3RegFrm,
5833 IIC_VRECSQ, "vrsqrts", "f16",
5834 v8f16, v8f16, int_arm_neon_vrsqrts, 1>,
5835 Requires<[HasNEON, HasFullFP16]>;
5839 // VSHL : Vector Shift
5840 defm VSHLs : N3VInt_QHSDSh<0, 0, 0b0100, 0, N3RegVShFrm,
5841 IIC_VSHLiD, IIC_VSHLiD, IIC_VSHLiQ, IIC_VSHLiQ,
5842 "vshl", "s", int_arm_neon_vshifts>;
5843 defm VSHLu : N3VInt_QHSDSh<1, 0, 0b0100, 0, N3RegVShFrm,
5844 IIC_VSHLiD, IIC_VSHLiD, IIC_VSHLiQ, IIC_VSHLiQ,
5845 "vshl", "u", int_arm_neon_vshiftu>;
5847 let Predicates = [HasNEON] in {
5848 def : Pat<(v8i8 (ARMvshls (v8i8 DPR:$Dn), (v8i8 DPR:$Dm))),
5849 (VSHLsv8i8 DPR:$Dn, DPR:$Dm)>;
5850 def : Pat<(v4i16 (ARMvshls (v4i16 DPR:$Dn), (v4i16 DPR:$Dm))),
5851 (VSHLsv4i16 DPR:$Dn, DPR:$Dm)>;
5852 def : Pat<(v2i32 (ARMvshls (v2i32 DPR:$Dn), (v2i32 DPR:$Dm))),
5853 (VSHLsv2i32 DPR:$Dn, DPR:$Dm)>;
5854 def : Pat<(v1i64 (ARMvshls (v1i64 DPR:$Dn), (v1i64 DPR:$Dm))),
5855 (VSHLsv1i64 DPR:$Dn, DPR:$Dm)>;
5856 def : Pat<(v16i8 (ARMvshls (v16i8 QPR:$Dn), (v16i8 QPR:$Dm))),
5857 (VSHLsv16i8 QPR:$Dn, QPR:$Dm)>;
5858 def : Pat<(v8i16 (ARMvshls (v8i16 QPR:$Dn), (v8i16 QPR:$Dm))),
5859 (VSHLsv8i16 QPR:$Dn, QPR:$Dm)>;
5860 def : Pat<(v4i32 (ARMvshls (v4i32 QPR:$Dn), (v4i32 QPR:$Dm))),
5861 (VSHLsv4i32 QPR:$Dn, QPR:$Dm)>;
5862 def : Pat<(v2i64 (ARMvshls (v2i64 QPR:$Dn), (v2i64 QPR:$Dm))),
5863 (VSHLsv2i64 QPR:$Dn, QPR:$Dm)>;
5865 def : Pat<(v8i8 (ARMvshlu (v8i8 DPR:$Dn), (v8i8 DPR:$Dm))),
5866 (VSHLuv8i8 DPR:$Dn, DPR:$Dm)>;
5867 def : Pat<(v4i16 (ARMvshlu (v4i16 DPR:$Dn), (v4i16 DPR:$Dm))),
5868 (VSHLuv4i16 DPR:$Dn, DPR:$Dm)>;
5869 def : Pat<(v2i32 (ARMvshlu (v2i32 DPR:$Dn), (v2i32 DPR:$Dm))),
5870 (VSHLuv2i32 DPR:$Dn, DPR:$Dm)>;
5871 def : Pat<(v1i64 (ARMvshlu (v1i64 DPR:$Dn), (v1i64 DPR:$Dm))),
5872 (VSHLuv1i64 DPR:$Dn, DPR:$Dm)>;
5873 def : Pat<(v16i8 (ARMvshlu (v16i8 QPR:$Dn), (v16i8 QPR:$Dm))),
5874 (VSHLuv16i8 QPR:$Dn, QPR:$Dm)>;
5875 def : Pat<(v8i16 (ARMvshlu (v8i16 QPR:$Dn), (v8i16 QPR:$Dm))),
5876 (VSHLuv8i16 QPR:$Dn, QPR:$Dm)>;
5877 def : Pat<(v4i32 (ARMvshlu (v4i32 QPR:$Dn), (v4i32 QPR:$Dm))),
5878 (VSHLuv4i32 QPR:$Dn, QPR:$Dm)>;
5879 def : Pat<(v2i64 (ARMvshlu (v2i64 QPR:$Dn), (v2i64 QPR:$Dm))),
5880 (VSHLuv2i64 QPR:$Dn, QPR:$Dm)>;
5884 // VSHL : Vector Shift Left (Immediate)
5885 defm VSHLi : N2VShL_QHSD<0, 1, 0b0101, 1, IIC_VSHLiD, "vshl", "i", ARMvshlImm>;
5887 // VSHR : Vector Shift Right (Immediate)
5888 defm VSHRs : N2VShR_QHSD<0, 1, 0b0000, 1, IIC_VSHLiD, "vshr", "s", "VSHRs",
5890 defm VSHRu : N2VShR_QHSD<1, 1, 0b0000, 1, IIC_VSHLiD, "vshr", "u", "VSHRu",
5893 // VSHLL : Vector Shift Left Long
5894 defm VSHLLs : N2VLSh_QHS<0, 1, 0b1010, 0, 0, 1, "vshll", "s",
5895 PatFrag<(ops node:$LHS, node:$RHS), (ARMvshlImm (sext node:$LHS), node:$RHS)>>;
5896 defm VSHLLu : N2VLSh_QHS<1, 1, 0b1010, 0, 0, 1, "vshll", "u",
5897 PatFrag<(ops node:$LHS, node:$RHS), (ARMvshlImm (zext node:$LHS), node:$RHS)>>;
5899 // VSHLL : Vector Shift Left Long (with maximum shift count)
5900 class N2VLShMax<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
5901 bit op6, bit op4, string OpcodeStr, string Dt, ValueType ResTy,
5902 ValueType OpTy, Operand ImmTy>
5903 : N2VLSh<op24, op23, op11_8, op7, op6, op4, OpcodeStr, Dt,
5904 ResTy, OpTy, ImmTy, null_frag> {
5905 let Inst{21-16} = op21_16;
5906 let DecoderMethod = "DecodeVSHLMaxInstruction";
5908 def VSHLLi8 : N2VLShMax<1, 1, 0b110010, 0b0011, 0, 0, 0, "vshll", "i8",
5910 def VSHLLi16 : N2VLShMax<1, 1, 0b110110, 0b0011, 0, 0, 0, "vshll", "i16",
5911 v4i32, v4i16, imm16>;
5912 def VSHLLi32 : N2VLShMax<1, 1, 0b111010, 0b0011, 0, 0, 0, "vshll", "i32",
5913 v2i64, v2i32, imm32>;
5915 let Predicates = [HasNEON] in {
5916 def : Pat<(v8i16 (ARMvshlImm (zext (v8i8 DPR:$Rn)), (i32 8))),
5917 (VSHLLi8 DPR:$Rn, 8)>;
5918 def : Pat<(v4i32 (ARMvshlImm (zext (v4i16 DPR:$Rn)), (i32 16))),
5919 (VSHLLi16 DPR:$Rn, 16)>;
5920 def : Pat<(v2i64 (ARMvshlImm (zext (v2i32 DPR:$Rn)), (i32 32))),
5921 (VSHLLi32 DPR:$Rn, 32)>;
5922 def : Pat<(v8i16 (ARMvshlImm (sext (v8i8 DPR:$Rn)), (i32 8))),
5923 (VSHLLi8 DPR:$Rn, 8)>;
5924 def : Pat<(v4i32 (ARMvshlImm (sext (v4i16 DPR:$Rn)), (i32 16))),
5925 (VSHLLi16 DPR:$Rn, 16)>;
5926 def : Pat<(v2i64 (ARMvshlImm (sext (v2i32 DPR:$Rn)), (i32 32))),
5927 (VSHLLi32 DPR:$Rn, 32)>;
5928 def : Pat<(v8i16 (ARMvshlImm (anyext (v8i8 DPR:$Rn)), (i32 8))),
5929 (VSHLLi8 DPR:$Rn, 8)>;
5930 def : Pat<(v4i32 (ARMvshlImm (anyext (v4i16 DPR:$Rn)), (i32 16))),
5931 (VSHLLi16 DPR:$Rn, 16)>;
5932 def : Pat<(v2i64 (ARMvshlImm (anyext (v2i32 DPR:$Rn)), (i32 32))),
5933 (VSHLLi32 DPR:$Rn, 32)>;
5936 // VSHRN : Vector Shift Right and Narrow
5937 defm VSHRN : N2VNSh_HSD<0,1,0b1000,0,0,1, IIC_VSHLiD, "vshrn", "i",
5938 PatFrag<(ops node:$Rn, node:$amt),
5939 (trunc (ARMvshrsImm node:$Rn, node:$amt))>>;
5941 let Predicates = [HasNEON] in {
5942 def : Pat<(v8i8 (trunc (ARMvshruImm (v8i16 QPR:$Vn), shr_imm8:$amt))),
5943 (VSHRNv8i8 QPR:$Vn, shr_imm8:$amt)>;
5944 def : Pat<(v4i16 (trunc (ARMvshruImm (v4i32 QPR:$Vn), shr_imm16:$amt))),
5945 (VSHRNv4i16 QPR:$Vn, shr_imm16:$amt)>;
5946 def : Pat<(v2i32 (trunc (ARMvshruImm (v2i64 QPR:$Vn), shr_imm32:$amt))),
5947 (VSHRNv2i32 QPR:$Vn, shr_imm32:$amt)>;
5950 // VRSHL : Vector Rounding Shift
5951 defm VRSHLs : N3VInt_QHSDSh<0, 0, 0b0101, 0, N3RegVShFrm,
5952 IIC_VSHLi4D, IIC_VSHLi4D, IIC_VSHLi4Q, IIC_VSHLi4Q,
5953 "vrshl", "s", int_arm_neon_vrshifts>;
5954 defm VRSHLu : N3VInt_QHSDSh<1, 0, 0b0101, 0, N3RegVShFrm,
5955 IIC_VSHLi4D, IIC_VSHLi4D, IIC_VSHLi4Q, IIC_VSHLi4Q,
5956 "vrshl", "u", int_arm_neon_vrshiftu>;
5957 // VRSHR : Vector Rounding Shift Right
5958 defm VRSHRs : N2VShR_QHSD<0,1,0b0010,1, IIC_VSHLi4D, "vrshr", "s", "VRSHRs",
5960 defm VRSHRu : N2VShR_QHSD<1,1,0b0010,1, IIC_VSHLi4D, "vrshr", "u", "VRSHRu",
5963 // VRSHRN : Vector Rounding Shift Right and Narrow
5964 defm VRSHRN : N2VNSh_HSD<0, 1, 0b1000, 0, 1, 1, IIC_VSHLi4D, "vrshrn", "i",
5967 // VQSHL : Vector Saturating Shift
5968 defm VQSHLs : N3VInt_QHSDSh<0, 0, 0b0100, 1, N3RegVShFrm,
5969 IIC_VSHLi4D, IIC_VSHLi4D, IIC_VSHLi4Q, IIC_VSHLi4Q,
5970 "vqshl", "s", int_arm_neon_vqshifts>;
5971 defm VQSHLu : N3VInt_QHSDSh<1, 0, 0b0100, 1, N3RegVShFrm,
5972 IIC_VSHLi4D, IIC_VSHLi4D, IIC_VSHLi4Q, IIC_VSHLi4Q,
5973 "vqshl", "u", int_arm_neon_vqshiftu>;
5974 // VQSHL : Vector Saturating Shift Left (Immediate)
5975 defm VQSHLsi : N2VShL_QHSD<0,1,0b0111,1, IIC_VSHLi4D, "vqshl", "s",NEONvqshlsImm>;
5976 defm VQSHLui : N2VShL_QHSD<1,1,0b0111,1, IIC_VSHLi4D, "vqshl", "u",NEONvqshluImm>;
5978 // VQSHLU : Vector Saturating Shift Left (Immediate, Unsigned)
5979 defm VQSHLsu : N2VShL_QHSD<1,1,0b0110,1, IIC_VSHLi4D,"vqshlu","s",NEONvqshlsuImm>;
5981 // VQSHRN : Vector Saturating Shift Right and Narrow
5982 defm VQSHRNs : N2VNSh_HSD<0, 1, 0b1001, 0, 0, 1, IIC_VSHLi4D, "vqshrn", "s",
5984 defm VQSHRNu : N2VNSh_HSD<1, 1, 0b1001, 0, 0, 1, IIC_VSHLi4D, "vqshrn", "u",
5987 // VQSHRUN : Vector Saturating Shift Right and Narrow (Unsigned)
5988 defm VQSHRUN : N2VNSh_HSD<1, 1, 0b1000, 0, 0, 1, IIC_VSHLi4D, "vqshrun", "s",
5991 // VQRSHL : Vector Saturating Rounding Shift
5992 defm VQRSHLs : N3VInt_QHSDSh<0, 0, 0b0101, 1, N3RegVShFrm,
5993 IIC_VSHLi4D, IIC_VSHLi4D, IIC_VSHLi4Q, IIC_VSHLi4Q,
5994 "vqrshl", "s", int_arm_neon_vqrshifts>;
5995 defm VQRSHLu : N3VInt_QHSDSh<1, 0, 0b0101, 1, N3RegVShFrm,
5996 IIC_VSHLi4D, IIC_VSHLi4D, IIC_VSHLi4Q, IIC_VSHLi4Q,
5997 "vqrshl", "u", int_arm_neon_vqrshiftu>;
5999 // VQRSHRN : Vector Saturating Rounding Shift Right and Narrow
6000 defm VQRSHRNs : N2VNSh_HSD<0, 1, 0b1001, 0, 1, 1, IIC_VSHLi4D, "vqrshrn", "s",
6002 defm VQRSHRNu : N2VNSh_HSD<1, 1, 0b1001, 0, 1, 1, IIC_VSHLi4D, "vqrshrn", "u",
6005 // VQRSHRUN : Vector Saturating Rounding Shift Right and Narrow (Unsigned)
6006 defm VQRSHRUN : N2VNSh_HSD<1, 1, 0b1000, 0, 1, 1, IIC_VSHLi4D, "vqrshrun", "s",
6009 // VSRA : Vector Shift Right and Accumulate
6010 defm VSRAs : N2VShAdd_QHSD<0, 1, 0b0001, 1, "vsra", "s", ARMvshrsImm>;
6011 defm VSRAu : N2VShAdd_QHSD<1, 1, 0b0001, 1, "vsra", "u", ARMvshruImm>;
6012 // VRSRA : Vector Rounding Shift Right and Accumulate
6013 defm VRSRAs : N2VShAdd_QHSD<0, 1, 0b0011, 1, "vrsra", "s", NEONvrshrsImm>;
6014 defm VRSRAu : N2VShAdd_QHSD<1, 1, 0b0011, 1, "vrsra", "u", NEONvrshruImm>;
6016 // VSLI : Vector Shift Left and Insert
6017 defm VSLI : N2VShInsL_QHSD<1, 1, 0b0101, 1, "vsli">;
6019 // VSRI : Vector Shift Right and Insert
6020 defm VSRI : N2VShInsR_QHSD<1, 1, 0b0100, 1, "vsri">;
6022 // Vector Absolute and Saturating Absolute.
6024 // VABS : Vector Absolute Value
6025 defm VABS : N2VInt_QHS<0b11, 0b11, 0b01, 0b00110, 0,
6026 IIC_VUNAiD, IIC_VUNAiQ, "vabs", "s", abs>;
6027 def VABSfd : N2VD<0b11, 0b11, 0b10, 0b01, 0b01110, 0,
6029 v2f32, v2f32, fabs>;
6030 def VABSfq : N2VQ<0b11, 0b11, 0b10, 0b01, 0b01110, 0,
6032 v4f32, v4f32, fabs>;
6033 def VABShd : N2VD<0b11, 0b11, 0b01, 0b01, 0b01110, 0,
6035 v4f16, v4f16, fabs>,
6036 Requires<[HasNEON, HasFullFP16]>;
6037 def VABShq : N2VQ<0b11, 0b11, 0b01, 0b01, 0b01110, 0,
6039 v8f16, v8f16, fabs>,
6040 Requires<[HasNEON, HasFullFP16]>;
6042 // VQABS : Vector Saturating Absolute Value
6043 defm VQABS : N2VInt_QHS<0b11, 0b11, 0b00, 0b01110, 0,
6044 IIC_VQUNAiD, IIC_VQUNAiQ, "vqabs", "s",
6045 int_arm_neon_vqabs>;
6049 def vnegd : PatFrag<(ops node:$in),
6050 (sub (bitconvert (v2i32 NEONimmAllZerosV)), node:$in)>;
6051 def vnegq : PatFrag<(ops node:$in),
6052 (sub (bitconvert (v4i32 NEONimmAllZerosV)), node:$in)>;
6054 class VNEGD<bits<2> size, string OpcodeStr, string Dt, ValueType Ty>
6055 : N2V<0b11, 0b11, size, 0b01, 0b00111, 0, 0, (outs DPR:$Vd), (ins DPR:$Vm),
6056 IIC_VSHLiD, OpcodeStr, Dt, "$Vd, $Vm", "",
6057 [(set DPR:$Vd, (Ty (vnegd DPR:$Vm)))]>;
6058 class VNEGQ<bits<2> size, string OpcodeStr, string Dt, ValueType Ty>
6059 : N2V<0b11, 0b11, size, 0b01, 0b00111, 1, 0, (outs QPR:$Vd), (ins QPR:$Vm),
6060 IIC_VSHLiQ, OpcodeStr, Dt, "$Vd, $Vm", "",
6061 [(set QPR:$Vd, (Ty (vnegq QPR:$Vm)))]>;
6063 // VNEG : Vector Negate (integer)
6064 def VNEGs8d : VNEGD<0b00, "vneg", "s8", v8i8>;
6065 def VNEGs16d : VNEGD<0b01, "vneg", "s16", v4i16>;
6066 def VNEGs32d : VNEGD<0b10, "vneg", "s32", v2i32>;
6067 def VNEGs8q : VNEGQ<0b00, "vneg", "s8", v16i8>;
6068 def VNEGs16q : VNEGQ<0b01, "vneg", "s16", v8i16>;
6069 def VNEGs32q : VNEGQ<0b10, "vneg", "s32", v4i32>;
6071 // VNEG : Vector Negate (floating-point)
6072 def VNEGfd : N2V<0b11, 0b11, 0b10, 0b01, 0b01111, 0, 0,
6073 (outs DPR:$Vd), (ins DPR:$Vm), IIC_VUNAD,
6074 "vneg", "f32", "$Vd, $Vm", "",
6075 [(set DPR:$Vd, (v2f32 (fneg DPR:$Vm)))]>;
6076 def VNEGf32q : N2V<0b11, 0b11, 0b10, 0b01, 0b01111, 1, 0,
6077 (outs QPR:$Vd), (ins QPR:$Vm), IIC_VUNAQ,
6078 "vneg", "f32", "$Vd, $Vm", "",
6079 [(set QPR:$Vd, (v4f32 (fneg QPR:$Vm)))]>;
6080 def VNEGhd : N2V<0b11, 0b11, 0b01, 0b01, 0b01111, 0, 0,
6081 (outs DPR:$Vd), (ins DPR:$Vm), IIC_VUNAD,
6082 "vneg", "f16", "$Vd, $Vm", "",
6083 [(set DPR:$Vd, (v4f16 (fneg DPR:$Vm)))]>,
6084 Requires<[HasNEON, HasFullFP16]>;
6085 def VNEGhq : N2V<0b11, 0b11, 0b01, 0b01, 0b01111, 1, 0,
6086 (outs QPR:$Vd), (ins QPR:$Vm), IIC_VUNAQ,
6087 "vneg", "f16", "$Vd, $Vm", "",
6088 [(set QPR:$Vd, (v8f16 (fneg QPR:$Vm)))]>,
6089 Requires<[HasNEON, HasFullFP16]>;
6091 let Predicates = [HasNEON] in {
6092 def : Pat<(v8i8 (vnegd DPR:$src)), (VNEGs8d DPR:$src)>;
6093 def : Pat<(v4i16 (vnegd DPR:$src)), (VNEGs16d DPR:$src)>;
6094 def : Pat<(v2i32 (vnegd DPR:$src)), (VNEGs32d DPR:$src)>;
6095 def : Pat<(v16i8 (vnegq QPR:$src)), (VNEGs8q QPR:$src)>;
6096 def : Pat<(v8i16 (vnegq QPR:$src)), (VNEGs16q QPR:$src)>;
6097 def : Pat<(v4i32 (vnegq QPR:$src)), (VNEGs32q QPR:$src)>;
6100 // VQNEG : Vector Saturating Negate
6101 defm VQNEG : N2VInt_QHS<0b11, 0b11, 0b00, 0b01111, 0,
6102 IIC_VQUNAiD, IIC_VQUNAiQ, "vqneg", "s",
6103 int_arm_neon_vqneg>;
6105 // Vector Bit Counting Operations.
6107 // VCLS : Vector Count Leading Sign Bits
6108 defm VCLS : N2VInt_QHS<0b11, 0b11, 0b00, 0b01000, 0,
6109 IIC_VCNTiD, IIC_VCNTiQ, "vcls", "s",
6111 // VCLZ : Vector Count Leading Zeros
6112 defm VCLZ : N2VInt_QHS<0b11, 0b11, 0b00, 0b01001, 0,
6113 IIC_VCNTiD, IIC_VCNTiQ, "vclz", "i",
6115 // VCNT : Vector Count One Bits
6116 def VCNTd : N2VDInt<0b11, 0b11, 0b00, 0b00, 0b01010, 0,
6117 IIC_VCNTiD, "vcnt", "8",
6119 def VCNTq : N2VQInt<0b11, 0b11, 0b00, 0b00, 0b01010, 0,
6120 IIC_VCNTiQ, "vcnt", "8",
6121 v16i8, v16i8, ctpop>;
6124 def VSWPd : N2VX<0b11, 0b11, 0b00, 0b10, 0b00000, 0, 0,
6125 (outs DPR:$Vd, DPR:$Vm), (ins DPR:$in1, DPR:$in2),
6126 NoItinerary, "vswp", "$Vd, $Vm", "$in1 = $Vd, $in2 = $Vm",
6128 def VSWPq : N2VX<0b11, 0b11, 0b00, 0b10, 0b00000, 1, 0,
6129 (outs QPR:$Vd, QPR:$Vm), (ins QPR:$in1, QPR:$in2),
6130 NoItinerary, "vswp", "$Vd, $Vm", "$in1 = $Vd, $in2 = $Vm",
6133 // Vector Move Operations.
6135 // VMOV : Vector Move (Register)
6136 def : NEONInstAlias<"vmov${p} $Vd, $Vm",
6137 (VORRd DPR:$Vd, DPR:$Vm, DPR:$Vm, pred:$p)>;
6138 def : NEONInstAlias<"vmov${p} $Vd, $Vm",
6139 (VORRq QPR:$Vd, QPR:$Vm, QPR:$Vm, pred:$p)>;
6141 // VMOV : Vector Move (Immediate)
6143 // Although VMOVs are not strictly speaking cheap, they are as expensive
6144 // as their copies counterpart (VORR), so we should prefer rematerialization
6145 // over splitting when it applies.
6146 let isReMaterializable = 1, isAsCheapAsAMove=1 in {
6147 def VMOVv8i8 : N1ModImm<1, 0b000, 0b1110, 0, 0, 0, 1, (outs DPR:$Vd),
6148 (ins nImmSplatI8:$SIMM), IIC_VMOVImm,
6149 "vmov", "i8", "$Vd, $SIMM", "",
6150 [(set DPR:$Vd, (v8i8 (ARMvmovImm timm:$SIMM)))]>;
6151 def VMOVv16i8 : N1ModImm<1, 0b000, 0b1110, 0, 1, 0, 1, (outs QPR:$Vd),
6152 (ins nImmSplatI8:$SIMM), IIC_VMOVImm,
6153 "vmov", "i8", "$Vd, $SIMM", "",
6154 [(set QPR:$Vd, (v16i8 (ARMvmovImm timm:$SIMM)))]>;
6156 def VMOVv4i16 : N1ModImm<1, 0b000, {1,0,?,0}, 0, 0, 0, 1, (outs DPR:$Vd),
6157 (ins nImmSplatI16:$SIMM), IIC_VMOVImm,
6158 "vmov", "i16", "$Vd, $SIMM", "",
6159 [(set DPR:$Vd, (v4i16 (ARMvmovImm timm:$SIMM)))]> {
6160 let Inst{9} = SIMM{9};
6163 def VMOVv8i16 : N1ModImm<1, 0b000, {1,0,?,0}, 0, 1, 0, 1, (outs QPR:$Vd),
6164 (ins nImmSplatI16:$SIMM), IIC_VMOVImm,
6165 "vmov", "i16", "$Vd, $SIMM", "",
6166 [(set QPR:$Vd, (v8i16 (ARMvmovImm timm:$SIMM)))]> {
6167 let Inst{9} = SIMM{9};
6170 def VMOVv2i32 : N1ModImm<1, 0b000, {?,?,?,?}, 0, 0, 0, 1, (outs DPR:$Vd),
6171 (ins nImmVMOVI32:$SIMM), IIC_VMOVImm,
6172 "vmov", "i32", "$Vd, $SIMM", "",
6173 [(set DPR:$Vd, (v2i32 (ARMvmovImm timm:$SIMM)))]> {
6174 let Inst{11-8} = SIMM{11-8};
6177 def VMOVv4i32 : N1ModImm<1, 0b000, {?,?,?,?}, 0, 1, 0, 1, (outs QPR:$Vd),
6178 (ins nImmVMOVI32:$SIMM), IIC_VMOVImm,
6179 "vmov", "i32", "$Vd, $SIMM", "",
6180 [(set QPR:$Vd, (v4i32 (ARMvmovImm timm:$SIMM)))]> {
6181 let Inst{11-8} = SIMM{11-8};
6184 def VMOVv1i64 : N1ModImm<1, 0b000, 0b1110, 0, 0, 1, 1, (outs DPR:$Vd),
6185 (ins nImmSplatI64:$SIMM), IIC_VMOVImm,
6186 "vmov", "i64", "$Vd, $SIMM", "",
6187 [(set DPR:$Vd, (v1i64 (ARMvmovImm timm:$SIMM)))]>;
6188 def VMOVv2i64 : N1ModImm<1, 0b000, 0b1110, 0, 1, 1, 1, (outs QPR:$Vd),
6189 (ins nImmSplatI64:$SIMM), IIC_VMOVImm,
6190 "vmov", "i64", "$Vd, $SIMM", "",
6191 [(set QPR:$Vd, (v2i64 (ARMvmovImm timm:$SIMM)))]>;
6193 def VMOVv2f32 : N1ModImm<1, 0b000, 0b1111, 0, 0, 0, 1, (outs DPR:$Vd),
6194 (ins nImmVMOVF32:$SIMM), IIC_VMOVImm,
6195 "vmov", "f32", "$Vd, $SIMM", "",
6196 [(set DPR:$Vd, (v2f32 (ARMvmovFPImm timm:$SIMM)))]>;
6197 def VMOVv4f32 : N1ModImm<1, 0b000, 0b1111, 0, 1, 0, 1, (outs QPR:$Vd),
6198 (ins nImmVMOVF32:$SIMM), IIC_VMOVImm,
6199 "vmov", "f32", "$Vd, $SIMM", "",
6200 [(set QPR:$Vd, (v4f32 (ARMvmovFPImm timm:$SIMM)))]>;
6201 } // isReMaterializable, isAsCheapAsAMove
6203 // Add support for bytes replication feature, so it could be GAS compatible.
6204 multiclass NEONImmReplicateI8InstAlias<ValueType To> {
6205 // E.g. instructions below:
6206 // "vmov.i32 d0, #0xffffffff"
6207 // "vmov.i32 d0, #0xabababab"
6208 // "vmov.i16 d0, #0xabab"
6209 // are incorrect, but we could deal with such cases.
6210 // For last two instructions, for example, it should emit:
6211 // "vmov.i8 d0, #0xab"
6212 def : NEONInstAlias<"vmov${p}.i" # To.Size # " $Vd, $Vm",
6213 (VMOVv8i8 DPR:$Vd, nImmVMOVIReplicate<i8, To>:$Vm, pred:$p)>;
6214 def : NEONInstAlias<"vmov${p}.i" # To.Size # " $Vd, $Vm",
6215 (VMOVv16i8 QPR:$Vd, nImmVMOVIReplicate<i8, To>:$Vm, pred:$p)>;
6216 // Also add same support for VMVN instructions. So instruction:
6217 // "vmvn.i32 d0, #0xabababab"
6219 // "vmov.i8 d0, #0x54"
6220 def : NEONInstAlias<"vmvn${p}.i" # To.Size # " $Vd, $Vm",
6221 (VMOVv8i8 DPR:$Vd, nImmVINVIReplicate<i8, To>:$Vm, pred:$p)>;
6222 def : NEONInstAlias<"vmvn${p}.i" # To.Size # " $Vd, $Vm",
6223 (VMOVv16i8 QPR:$Vd, nImmVINVIReplicate<i8, To>:$Vm, pred:$p)>;
6226 defm : NEONImmReplicateI8InstAlias<i16>;
6227 defm : NEONImmReplicateI8InstAlias<i32>;
6228 defm : NEONImmReplicateI8InstAlias<i64>;
6230 // Similar to above for types other than i8, e.g.:
6231 // "vmov.i32 d0, #0xab00ab00" -> "vmov.i16 d0, #0xab00"
6232 // "vmvn.i64 q0, #0xab000000ab000000" -> "vmvn.i32 q0, #0xab000000"
6233 // In this case we do not canonicalize VMVN to VMOV
6234 multiclass NEONImmReplicateInstAlias<ValueType From, NeonI V8, NeonI V16,
6235 NeonI NV8, NeonI NV16, ValueType To> {
6236 def : NEONInstAlias<"vmov${p}.i" # To.Size # " $Vd, $Vm",
6237 (V8 DPR:$Vd, nImmVMOVIReplicate<From, To>:$Vm, pred:$p)>;
6238 def : NEONInstAlias<"vmov${p}.i" # To.Size # " $Vd, $Vm",
6239 (V16 QPR:$Vd, nImmVMOVIReplicate<From, To>:$Vm, pred:$p)>;
6240 def : NEONInstAlias<"vmvn${p}.i" # To.Size # " $Vd, $Vm",
6241 (NV8 DPR:$Vd, nImmVMOVIReplicate<From, To>:$Vm, pred:$p)>;
6242 def : NEONInstAlias<"vmvn${p}.i" # To.Size # " $Vd, $Vm",
6243 (NV16 QPR:$Vd, nImmVMOVIReplicate<From, To>:$Vm, pred:$p)>;
6246 defm : NEONImmReplicateInstAlias<i16, VMOVv4i16, VMOVv8i16,
6247 VMVNv4i16, VMVNv8i16, i32>;
6248 defm : NEONImmReplicateInstAlias<i16, VMOVv4i16, VMOVv8i16,
6249 VMVNv4i16, VMVNv8i16, i64>;
6250 defm : NEONImmReplicateInstAlias<i32, VMOVv2i32, VMOVv4i32,
6251 VMVNv2i32, VMVNv4i32, i64>;
6252 // TODO: add "VMOV <-> VMVN" conversion for cases like
6253 // "vmov.i32 d0, #0xffaaffaa" -> "vmvn.i16 d0, #0x55"
6254 // "vmvn.i32 d0, #0xaaffaaff" -> "vmov.i16 d0, #0xff00"
6256 // On some CPUs the two instructions "vmov.i32 dD, #0" and "vmov.i32 qD, #0"
6257 // require zero cycles to execute so they should be used wherever possible for
6258 // setting a register to zero.
6260 // Even without these pseudo-insts we would probably end up with the correct
6261 // instruction, but we could not mark the general ones with "isAsCheapAsAMove"
6262 // since they are sometimes rather expensive (in general).
6264 let AddedComplexity = 50, isAsCheapAsAMove = 1, isReMaterializable = 1 in {
6265 def VMOVD0 : ARMPseudoExpand<(outs DPR:$Vd), (ins), 4, IIC_VMOVImm,
6266 [(set DPR:$Vd, (v2i32 NEONimmAllZerosV))],
6267 (VMOVv2i32 DPR:$Vd, 0, (ops 14, zero_reg))>,
6269 def VMOVQ0 : ARMPseudoExpand<(outs QPR:$Vd), (ins), 4, IIC_VMOVImm,
6270 [(set QPR:$Vd, (v4i32 NEONimmAllZerosV))],
6271 (VMOVv4i32 QPR:$Vd, 0, (ops 14, zero_reg))>,
6275 // VMOV : Vector Get Lane (move scalar to ARM core register)
6277 def VGETLNs8 : NVGetLane<{1,1,1,0,0,1,?,1}, 0b1011, {?,?},
6278 (outs GPR:$R), (ins DPR:$V, VectorIndex8:$lane),
6279 IIC_VMOVSI, "vmov", "s8", "$R, $V$lane",
6280 [(set GPR:$R, (ARMvgetlanes (v8i8 DPR:$V),
6282 let Inst{21} = lane{2};
6283 let Inst{6-5} = lane{1-0};
6285 def VGETLNs16 : NVGetLane<{1,1,1,0,0,0,?,1}, 0b1011, {?,1},
6286 (outs GPR:$R), (ins DPR:$V, VectorIndex16:$lane),
6287 IIC_VMOVSI, "vmov", "s16", "$R, $V$lane",
6288 [(set GPR:$R, (ARMvgetlanes (v4i16 DPR:$V),
6290 let Inst{21} = lane{1};
6291 let Inst{6} = lane{0};
6293 def VGETLNu8 : NVGetLane<{1,1,1,0,1,1,?,1}, 0b1011, {?,?},
6294 (outs GPR:$R), (ins DPR:$V, VectorIndex8:$lane),
6295 IIC_VMOVSI, "vmov", "u8", "$R, $V$lane",
6296 [(set GPR:$R, (ARMvgetlaneu (v8i8 DPR:$V),
6298 let Inst{21} = lane{2};
6299 let Inst{6-5} = lane{1-0};
6301 def VGETLNu16 : NVGetLane<{1,1,1,0,1,0,?,1}, 0b1011, {?,1},
6302 (outs GPR:$R), (ins DPR:$V, VectorIndex16:$lane),
6303 IIC_VMOVSI, "vmov", "u16", "$R, $V$lane",
6304 [(set GPR:$R, (ARMvgetlaneu (v4i16 DPR:$V),
6306 let Inst{21} = lane{1};
6307 let Inst{6} = lane{0};
6309 def VGETLNi32 : NVGetLane<{1,1,1,0,0,0,?,1}, 0b1011, 0b00,
6310 (outs GPR:$R), (ins DPR:$V, VectorIndex32:$lane),
6311 IIC_VMOVSI, "vmov", "32", "$R, $V$lane",
6312 [(set GPR:$R, (extractelt (v2i32 DPR:$V),
6314 Requires<[HasFPRegs, HasFastVGETLNi32]> {
6315 let Inst{21} = lane{0};
6317 let Predicates = [HasNEON] in {
6318 // def VGETLNf32: see FMRDH and FMRDL in ARMInstrVFP.td
6319 def : Pat<(ARMvgetlanes (v16i8 QPR:$src), imm:$lane),
6320 (VGETLNs8 (v8i8 (EXTRACT_SUBREG QPR:$src,
6321 (DSubReg_i8_reg imm:$lane))),
6322 (SubReg_i8_lane imm:$lane))>;
6323 def : Pat<(ARMvgetlanes (v8i16 QPR:$src), imm:$lane),
6324 (VGETLNs16 (v4i16 (EXTRACT_SUBREG QPR:$src,
6325 (DSubReg_i16_reg imm:$lane))),
6326 (SubReg_i16_lane imm:$lane))>;
6327 def : Pat<(ARMvgetlaneu (v16i8 QPR:$src), imm:$lane),
6328 (VGETLNu8 (v8i8 (EXTRACT_SUBREG QPR:$src,
6329 (DSubReg_i8_reg imm:$lane))),
6330 (SubReg_i8_lane imm:$lane))>;
6331 def : Pat<(ARMvgetlaneu (v8i16 QPR:$src), imm:$lane),
6332 (VGETLNu16 (v4i16 (EXTRACT_SUBREG QPR:$src,
6333 (DSubReg_i16_reg imm:$lane))),
6334 (SubReg_i16_lane imm:$lane))>;
6336 def : Pat<(extractelt (v4i32 QPR:$src), imm:$lane),
6337 (VGETLNi32 (v2i32 (EXTRACT_SUBREG QPR:$src,
6338 (DSubReg_i32_reg imm:$lane))),
6339 (SubReg_i32_lane imm:$lane))>,
6340 Requires<[HasNEON, HasFastVGETLNi32]>;
6341 def : Pat<(extractelt (v2i32 DPR:$src), imm:$lane),
6343 (i32 (EXTRACT_SUBREG DPR:$src, (SSubReg_f32_reg imm:$lane))), GPR)>,
6344 Requires<[HasNEON, HasSlowVGETLNi32]>;
6345 def : Pat<(extractelt (v4i32 QPR:$src), imm:$lane),
6347 (i32 (EXTRACT_SUBREG QPR:$src, (SSubReg_f32_reg imm:$lane))), GPR)>,
6348 Requires<[HasNEON, HasSlowVGETLNi32]>;
6349 let Predicates = [HasNEON] in {
6350 def : Pat<(extractelt (v2f32 DPR:$src1), imm:$src2),
6351 (EXTRACT_SUBREG (v2f32 (COPY_TO_REGCLASS (v2f32 DPR:$src1),DPR_VFP2)),
6352 (SSubReg_f32_reg imm:$src2))>;
6353 def : Pat<(extractelt (v4f32 QPR:$src1), imm:$src2),
6354 (EXTRACT_SUBREG (v4f32 (COPY_TO_REGCLASS (v4f32 QPR:$src1),QPR_VFP2)),
6355 (SSubReg_f32_reg imm:$src2))>;
6356 //def : Pat<(extractelt (v2i64 QPR:$src1), imm:$src2),
6357 // (EXTRACT_SUBREG QPR:$src1, (DSubReg_f64_reg imm:$src2))>;
6358 def : Pat<(extractelt (v2f64 QPR:$src1), imm:$src2),
6359 (EXTRACT_SUBREG QPR:$src1, (DSubReg_f64_reg imm:$src2))>;
6362 multiclass ExtractEltEvenF16<ValueType VT4, ValueType VT8> {
6363 def : Pat<(extractelt (VT4 DPR:$src), imm_even:$lane),
6365 (v2f32 (COPY_TO_REGCLASS (VT4 DPR:$src), DPR_VFP2)),
6366 (SSubReg_f16_reg imm_even:$lane))>;
6367 def : Pat<(extractelt (VT8 QPR:$src), imm_even:$lane),
6369 (v4f32 (COPY_TO_REGCLASS (VT8 QPR:$src), QPR_VFP2)),
6370 (SSubReg_f16_reg imm_even:$lane))>;
6373 multiclass ExtractEltOddF16VMOVH<ValueType VT4, ValueType VT8> {
6374 def : Pat<(extractelt (VT4 DPR:$src), imm_odd:$lane),
6376 (VMOVH (EXTRACT_SUBREG
6377 (v2f32 (COPY_TO_REGCLASS (VT4 DPR:$src), DPR_VFP2)),
6378 (SSubReg_f16_reg imm_odd:$lane))),
6380 def : Pat<(extractelt (VT8 QPR:$src), imm_odd:$lane),
6382 (VMOVH (EXTRACT_SUBREG
6383 (v4f32 (COPY_TO_REGCLASS (VT8 QPR:$src), QPR_VFP2)),
6384 (SSubReg_f16_reg imm_odd:$lane))),
6388 let Predicates = [HasNEON] in {
6389 defm : ExtractEltEvenF16<v4f16, v8f16>;
6390 defm : ExtractEltOddF16VMOVH<v4f16, v8f16>;
6393 let AddedComplexity = 1, Predicates = [HasNEON, HasBF16, HasFullFP16] in {
6394 // If VMOVH (vmovx.f16) is available use it to extract BF16 from the odd lanes
6395 defm : ExtractEltOddF16VMOVH<v4bf16, v8bf16>;
6398 let Predicates = [HasBF16, HasNEON] in {
6399 defm : ExtractEltEvenF16<v4bf16, v8bf16>;
6401 // Otherwise, if VMOVH is not available resort to extracting the odd lane
6402 // into a GPR and then moving to HPR
6403 def : Pat<(extractelt (v4bf16 DPR:$src), imm_odd:$lane),
6405 (VGETLNu16 (v4bf16 DPR:$src), imm:$lane),
6408 def : Pat<(extractelt (v8bf16 QPR:$src), imm_odd:$lane),
6410 (VGETLNu16 (v4i16 (EXTRACT_SUBREG QPR:$src,
6411 (DSubReg_i16_reg imm:$lane))),
6412 (SubReg_i16_lane imm:$lane)),
6416 // VMOV : Vector Set Lane (move ARM core register to scalar)
6418 let Constraints = "$src1 = $V" in {
6419 def VSETLNi8 : NVSetLane<{1,1,1,0,0,1,?,0}, 0b1011, {?,?}, (outs DPR:$V),
6420 (ins DPR:$src1, GPR:$R, VectorIndex8:$lane),
6421 IIC_VMOVISL, "vmov", "8", "$V$lane, $R",
6422 [(set DPR:$V, (vector_insert (v8i8 DPR:$src1),
6423 GPR:$R, imm:$lane))]> {
6424 let Inst{21} = lane{2};
6425 let Inst{6-5} = lane{1-0};
6427 def VSETLNi16 : NVSetLane<{1,1,1,0,0,0,?,0}, 0b1011, {?,1}, (outs DPR:$V),
6428 (ins DPR:$src1, GPR:$R, VectorIndex16:$lane),
6429 IIC_VMOVISL, "vmov", "16", "$V$lane, $R",
6430 [(set DPR:$V, (vector_insert (v4i16 DPR:$src1),
6431 GPR:$R, imm:$lane))]> {
6432 let Inst{21} = lane{1};
6433 let Inst{6} = lane{0};
6435 def VSETLNi32 : NVSetLane<{1,1,1,0,0,0,?,0}, 0b1011, 0b00, (outs DPR:$V),
6436 (ins DPR:$src1, GPR:$R, VectorIndex32:$lane),
6437 IIC_VMOVISL, "vmov", "32", "$V$lane, $R",
6438 [(set DPR:$V, (insertelt (v2i32 DPR:$src1),
6439 GPR:$R, imm:$lane))]>,
6440 Requires<[HasVFP2]> {
6441 let Inst{21} = lane{0};
6442 // This instruction is equivalent as
6443 // $V = INSERT_SUBREG $src1, $R, translateImmToSubIdx($imm)
6444 let isInsertSubreg = 1;
6448 // TODO: for odd lanes we could optimize this a bit by using the VINS
6449 // FullFP16 instruction when it is available
6450 multiclass InsertEltF16<ValueType VTScalar, ValueType VT4, ValueType VT8> {
6451 def : Pat<(insertelt (VT4 DPR:$src1), (VTScalar HPR:$src2), imm:$lane),
6452 (VT4 (VSETLNi16 DPR:$src1,
6453 (COPY_TO_REGCLASS HPR:$src2, GPR), imm:$lane))>;
6454 def : Pat<(insertelt (VT8 QPR:$src1), (VTScalar HPR:$src2), imm:$lane),
6455 (VT8 (INSERT_SUBREG QPR:$src1,
6456 (v4i16 (VSETLNi16 (v4i16 (EXTRACT_SUBREG QPR:$src1,
6457 (DSubReg_i16_reg imm:$lane))),
6458 (COPY_TO_REGCLASS HPR:$src2, GPR),
6459 (SubReg_i16_lane imm:$lane))),
6460 (DSubReg_i16_reg imm:$lane)))>;
6463 let Predicates = [HasNEON] in {
6464 def : Pat<(vector_insert (v16i8 QPR:$src1), GPR:$src2, imm:$lane),
6465 (v16i8 (INSERT_SUBREG QPR:$src1,
6466 (v8i8 (VSETLNi8 (v8i8 (EXTRACT_SUBREG QPR:$src1,
6467 (DSubReg_i8_reg imm:$lane))),
6468 GPR:$src2, (SubReg_i8_lane imm:$lane))),
6469 (DSubReg_i8_reg imm:$lane)))>;
6470 def : Pat<(vector_insert (v8i16 QPR:$src1), GPR:$src2, imm:$lane),
6471 (v8i16 (INSERT_SUBREG QPR:$src1,
6472 (v4i16 (VSETLNi16 (v4i16 (EXTRACT_SUBREG QPR:$src1,
6473 (DSubReg_i16_reg imm:$lane))),
6474 GPR:$src2, (SubReg_i16_lane imm:$lane))),
6475 (DSubReg_i16_reg imm:$lane)))>;
6476 def : Pat<(insertelt (v4i32 QPR:$src1), GPR:$src2, imm:$lane),
6477 (v4i32 (INSERT_SUBREG QPR:$src1,
6478 (v2i32 (VSETLNi32 (v2i32 (EXTRACT_SUBREG QPR:$src1,
6479 (DSubReg_i32_reg imm:$lane))),
6480 GPR:$src2, (SubReg_i32_lane imm:$lane))),
6481 (DSubReg_i32_reg imm:$lane)))>;
6483 def : Pat<(v2f32 (insertelt DPR:$src1, SPR:$src2, imm:$src3)),
6484 (INSERT_SUBREG (v2f32 (COPY_TO_REGCLASS DPR:$src1, DPR_VFP2)),
6485 SPR:$src2, (SSubReg_f32_reg imm:$src3))>;
6486 def : Pat<(v4f32 (insertelt QPR:$src1, SPR:$src2, imm:$src3)),
6487 (INSERT_SUBREG (v4f32 (COPY_TO_REGCLASS QPR:$src1, QPR_VFP2)),
6488 SPR:$src2, (SSubReg_f32_reg imm:$src3))>;
6490 defm : InsertEltF16<f16, v4f16, v8f16>;
6492 //def : Pat<(v2i64 (insertelt QPR:$src1, DPR:$src2, imm:$src3)),
6493 // (INSERT_SUBREG QPR:$src1, DPR:$src2, (DSubReg_f64_reg imm:$src3))>;
6494 def : Pat<(v2f64 (insertelt QPR:$src1, DPR:$src2, imm:$src3)),
6495 (INSERT_SUBREG QPR:$src1, DPR:$src2, (DSubReg_f64_reg imm:$src3))>;
6497 def : Pat<(v2f32 (scalar_to_vector SPR:$src)),
6498 (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$src, ssub_0)>;
6499 def : Pat<(v2f64 (scalar_to_vector (f64 DPR:$src))),
6500 (INSERT_SUBREG (v2f64 (IMPLICIT_DEF)), DPR:$src, dsub_0)>;
6501 def : Pat<(v4f32 (scalar_to_vector SPR:$src)),
6502 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), SPR:$src, ssub_0)>;
6504 def : Pat<(v8i8 (scalar_to_vector GPR:$src)),
6505 (VSETLNi8 (v8i8 (IMPLICIT_DEF)), GPR:$src, (i32 0))>;
6506 def : Pat<(v4i16 (scalar_to_vector GPR:$src)),
6507 (VSETLNi16 (v4i16 (IMPLICIT_DEF)), GPR:$src, (i32 0))>;
6508 def : Pat<(v2i32 (scalar_to_vector GPR:$src)),
6509 (VSETLNi32 (v2i32 (IMPLICIT_DEF)), GPR:$src, (i32 0))>;
6511 def : Pat<(v16i8 (scalar_to_vector GPR:$src)),
6512 (INSERT_SUBREG (v16i8 (IMPLICIT_DEF)),
6513 (VSETLNi8 (v8i8 (IMPLICIT_DEF)), GPR:$src, (i32 0)),
6515 def : Pat<(v8i16 (scalar_to_vector GPR:$src)),
6516 (INSERT_SUBREG (v8i16 (IMPLICIT_DEF)),
6517 (VSETLNi16 (v4i16 (IMPLICIT_DEF)), GPR:$src, (i32 0)),
6519 def : Pat<(v4i32 (scalar_to_vector GPR:$src)),
6520 (INSERT_SUBREG (v4i32 (IMPLICIT_DEF)),
6521 (VSETLNi32 (v2i32 (IMPLICIT_DEF)), GPR:$src, (i32 0)),
6525 let Predicates = [HasNEON, HasBF16] in
6526 defm : InsertEltF16<bf16, v4bf16, v8bf16>;
6528 // VDUP : Vector Duplicate (from ARM core register to all elements)
6530 class VDUPD<bits<8> opcod1, bits<2> opcod3, string Dt, ValueType Ty>
6531 : NVDup<opcod1, 0b1011, opcod3, (outs DPR:$V), (ins GPR:$R),
6532 IIC_VMOVIS, "vdup", Dt, "$V, $R",
6533 [(set DPR:$V, (Ty (ARMvdup (i32 GPR:$R))))]>;
6534 class VDUPQ<bits<8> opcod1, bits<2> opcod3, string Dt, ValueType Ty>
6535 : NVDup<opcod1, 0b1011, opcod3, (outs QPR:$V), (ins GPR:$R),
6536 IIC_VMOVIS, "vdup", Dt, "$V, $R",
6537 [(set QPR:$V, (Ty (ARMvdup (i32 GPR:$R))))]>;
6539 def VDUP8d : VDUPD<0b11101100, 0b00, "8", v8i8>;
6540 def VDUP16d : VDUPD<0b11101000, 0b01, "16", v4i16>;
6541 def VDUP32d : VDUPD<0b11101000, 0b00, "32", v2i32>,
6542 Requires<[HasNEON, HasFastVDUP32]>;
6543 def VDUP8q : VDUPQ<0b11101110, 0b00, "8", v16i8>;
6544 def VDUP16q : VDUPQ<0b11101010, 0b01, "16", v8i16>;
6545 def VDUP32q : VDUPQ<0b11101010, 0b00, "32", v4i32>;
6547 // ARMvdup patterns for uarchs with fast VDUP.32.
6548 def : Pat<(v2f32 (ARMvdup (f32 (bitconvert GPR:$R)))), (VDUP32d GPR:$R)>,
6549 Requires<[HasNEON,HasFastVDUP32]>;
6550 def : Pat<(v4f32 (ARMvdup (f32 (bitconvert GPR:$R)))), (VDUP32q GPR:$R)>,
6551 Requires<[HasNEON]>;
6553 // ARMvdup patterns for uarchs with slow VDUP.32 - use VMOVDRR instead.
6554 def : Pat<(v2i32 (ARMvdup (i32 GPR:$R))), (VMOVDRR GPR:$R, GPR:$R)>,
6555 Requires<[HasNEON,HasSlowVDUP32]>;
6556 def : Pat<(v2f32 (ARMvdup (f32 (bitconvert GPR:$R)))), (VMOVDRR GPR:$R, GPR:$R)>,
6557 Requires<[HasNEON,HasSlowVDUP32]>;
6559 // VDUP : Vector Duplicate Lane (from scalar to all elements)
6561 class VDUPLND<bits<4> op19_16, string OpcodeStr, string Dt,
6562 ValueType Ty, Operand IdxTy>
6563 : NVDupLane<op19_16, 0, (outs DPR:$Vd), (ins DPR:$Vm, IdxTy:$lane),
6564 IIC_VMOVD, OpcodeStr, Dt, "$Vd, $Vm$lane",
6565 [(set DPR:$Vd, (Ty (ARMvduplane (Ty DPR:$Vm), imm:$lane)))]>;
6567 class VDUPLNQ<bits<4> op19_16, string OpcodeStr, string Dt,
6568 ValueType ResTy, ValueType OpTy, Operand IdxTy>
6569 : NVDupLane<op19_16, 1, (outs QPR:$Vd), (ins DPR:$Vm, IdxTy:$lane),
6570 IIC_VMOVQ, OpcodeStr, Dt, "$Vd, $Vm$lane",
6571 [(set QPR:$Vd, (ResTy (ARMvduplane (OpTy DPR:$Vm),
6572 VectorIndex32:$lane)))]>;
6574 // Inst{19-16} is partially specified depending on the element size.
6576 def VDUPLN8d : VDUPLND<{?,?,?,1}, "vdup", "8", v8i8, VectorIndex8> {
6578 let Inst{19-17} = lane{2-0};
6580 def VDUPLN16d : VDUPLND<{?,?,1,0}, "vdup", "16", v4i16, VectorIndex16> {
6582 let Inst{19-18} = lane{1-0};
6584 def VDUPLN32d : VDUPLND<{?,1,0,0}, "vdup", "32", v2i32, VectorIndex32> {
6586 let Inst{19} = lane{0};
6588 def VDUPLN8q : VDUPLNQ<{?,?,?,1}, "vdup", "8", v16i8, v8i8, VectorIndex8> {
6590 let Inst{19-17} = lane{2-0};
6592 def VDUPLN16q : VDUPLNQ<{?,?,1,0}, "vdup", "16", v8i16, v4i16, VectorIndex16> {
6594 let Inst{19-18} = lane{1-0};
6596 def VDUPLN32q : VDUPLNQ<{?,1,0,0}, "vdup", "32", v4i32, v2i32, VectorIndex32> {
6598 let Inst{19} = lane{0};
6601 let Predicates = [HasNEON] in {
6602 def : Pat<(v4f16 (ARMvduplane (v4f16 DPR:$Vm), imm:$lane)),
6603 (VDUPLN32d DPR:$Vm, imm:$lane)>;
6605 def : Pat<(v2f32 (ARMvduplane (v2f32 DPR:$Vm), imm:$lane)),
6606 (VDUPLN32d DPR:$Vm, imm:$lane)>;
6608 def : Pat<(v4f32 (ARMvduplane (v2f32 DPR:$Vm), imm:$lane)),
6609 (VDUPLN32q DPR:$Vm, imm:$lane)>;
6611 def : Pat<(v16i8 (ARMvduplane (v16i8 QPR:$src), imm:$lane)),
6612 (v16i8 (VDUPLN8q (v8i8 (EXTRACT_SUBREG QPR:$src,
6613 (DSubReg_i8_reg imm:$lane))),
6614 (SubReg_i8_lane imm:$lane)))>;
6615 def : Pat<(v8i16 (ARMvduplane (v8i16 QPR:$src), imm:$lane)),
6616 (v8i16 (VDUPLN16q (v4i16 (EXTRACT_SUBREG QPR:$src,
6617 (DSubReg_i16_reg imm:$lane))),
6618 (SubReg_i16_lane imm:$lane)))>;
6619 def : Pat<(v8f16 (ARMvduplane (v8f16 QPR:$src), imm:$lane)),
6620 (v8f16 (VDUPLN16q (v4f16 (EXTRACT_SUBREG QPR:$src,
6621 (DSubReg_i16_reg imm:$lane))),
6622 (SubReg_i16_lane imm:$lane)))>;
6623 def : Pat<(v4i32 (ARMvduplane (v4i32 QPR:$src), imm:$lane)),
6624 (v4i32 (VDUPLN32q (v2i32 (EXTRACT_SUBREG QPR:$src,
6625 (DSubReg_i32_reg imm:$lane))),
6626 (SubReg_i32_lane imm:$lane)))>;
6627 def : Pat<(v4f32 (ARMvduplane (v4f32 QPR:$src), imm:$lane)),
6628 (v4f32 (VDUPLN32q (v2f32 (EXTRACT_SUBREG QPR:$src,
6629 (DSubReg_i32_reg imm:$lane))),
6630 (SubReg_i32_lane imm:$lane)))>;
6632 def : Pat<(v4f16 (ARMvdup (f16 HPR:$src))),
6633 (v4f16 (VDUPLN16d (INSERT_SUBREG (v4f16 (IMPLICIT_DEF)),
6634 (f16 HPR:$src), ssub_0), (i32 0)))>;
6635 def : Pat<(v2f32 (ARMvdup (f32 SPR:$src))),
6636 (v2f32 (VDUPLN32d (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)),
6637 SPR:$src, ssub_0), (i32 0)))>;
6638 def : Pat<(v4f32 (ARMvdup (f32 SPR:$src))),
6639 (v4f32 (VDUPLN32q (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)),
6640 SPR:$src, ssub_0), (i32 0)))>;
6641 def : Pat<(v8f16 (ARMvdup (f16 HPR:$src))),
6642 (v8f16 (VDUPLN16q (INSERT_SUBREG (v4f16 (IMPLICIT_DEF)),
6643 (f16 HPR:$src), ssub_0), (i32 0)))>;
6646 let Predicates = [HasNEON, HasBF16] in {
6647 def : Pat<(v4bf16 (ARMvduplane (v4bf16 DPR:$Vm), imm:$lane)),
6648 (VDUPLN16d DPR:$Vm, imm:$lane)>;
6650 def : Pat<(v8bf16 (ARMvduplane (v8bf16 QPR:$src), imm:$lane)),
6651 (v8bf16 (VDUPLN16q (v4bf16 (EXTRACT_SUBREG QPR:$src,
6652 (DSubReg_i16_reg imm:$lane))),
6653 (SubReg_i16_lane imm:$lane)))>;
6655 def : Pat<(v4bf16 (ARMvdup (bf16 HPR:$src))),
6656 (v4bf16 (VDUPLN16d (INSERT_SUBREG (v4bf16 (IMPLICIT_DEF)),
6657 (bf16 HPR:$src), ssub_0), (i32 0)))>;
6658 def : Pat<(v8bf16 (ARMvdup (bf16 HPR:$src))),
6659 (v8bf16 (VDUPLN16q (INSERT_SUBREG (v4bf16 (IMPLICIT_DEF)),
6660 (bf16 HPR:$src), ssub_0), (i32 0)))>;
6663 // VMOVN : Vector Narrowing Move
6664 defm VMOVN : N2VN_HSD<0b11,0b11,0b10,0b00100,0,0, IIC_VMOVN,
6665 "vmovn", "i", trunc>;
6666 // VQMOVN : Vector Saturating Narrowing Move
6667 defm VQMOVNs : N2VNInt_HSD<0b11,0b11,0b10,0b00101,0,0, IIC_VQUNAiD,
6668 "vqmovn", "s", int_arm_neon_vqmovns>;
6669 defm VQMOVNu : N2VNInt_HSD<0b11,0b11,0b10,0b00101,1,0, IIC_VQUNAiD,
6670 "vqmovn", "u", int_arm_neon_vqmovnu>;
6671 defm VQMOVNsu : N2VNInt_HSD<0b11,0b11,0b10,0b00100,1,0, IIC_VQUNAiD,
6672 "vqmovun", "s", int_arm_neon_vqmovnsu>;
6673 // VMOVL : Vector Lengthening Move
6674 defm VMOVLs : N2VL_QHS<0b01,0b10100,0,1, "vmovl", "s", sext>;
6675 defm VMOVLu : N2VL_QHS<0b11,0b10100,0,1, "vmovl", "u", zext>;
6677 let Predicates = [HasNEON] in {
6678 def : Pat<(v8i16 (anyext (v8i8 DPR:$Vm))), (VMOVLuv8i16 DPR:$Vm)>;
6679 def : Pat<(v4i32 (anyext (v4i16 DPR:$Vm))), (VMOVLuv4i32 DPR:$Vm)>;
6680 def : Pat<(v2i64 (anyext (v2i32 DPR:$Vm))), (VMOVLuv2i64 DPR:$Vm)>;
6683 // Vector Conversions.
6685 // VCVT : Vector Convert Between Floating-Point and Integers
6686 def VCVTf2sd : N2VD<0b11, 0b11, 0b10, 0b11, 0b01110, 0, "vcvt", "s32.f32",
6687 v2i32, v2f32, fp_to_sint>;
6688 def VCVTf2ud : N2VD<0b11, 0b11, 0b10, 0b11, 0b01111, 0, "vcvt", "u32.f32",
6689 v2i32, v2f32, fp_to_uint>;
6690 def VCVTs2fd : N2VD<0b11, 0b11, 0b10, 0b11, 0b01100, 0, "vcvt", "f32.s32",
6691 v2f32, v2i32, sint_to_fp>;
6692 def VCVTu2fd : N2VD<0b11, 0b11, 0b10, 0b11, 0b01101, 0, "vcvt", "f32.u32",
6693 v2f32, v2i32, uint_to_fp>;
6695 def VCVTf2sq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01110, 0, "vcvt", "s32.f32",
6696 v4i32, v4f32, fp_to_sint>;
6697 def VCVTf2uq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01111, 0, "vcvt", "u32.f32",
6698 v4i32, v4f32, fp_to_uint>;
6699 def VCVTs2fq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01100, 0, "vcvt", "f32.s32",
6700 v4f32, v4i32, sint_to_fp>;
6701 def VCVTu2fq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01101, 0, "vcvt", "f32.u32",
6702 v4f32, v4i32, uint_to_fp>;
6704 def VCVTh2sd : N2VD<0b11, 0b11, 0b01, 0b11, 0b01110, 0, "vcvt", "s16.f16",
6705 v4i16, v4f16, fp_to_sint>,
6706 Requires<[HasNEON, HasFullFP16]>;
6707 def VCVTh2ud : N2VD<0b11, 0b11, 0b01, 0b11, 0b01111, 0, "vcvt", "u16.f16",
6708 v4i16, v4f16, fp_to_uint>,
6709 Requires<[HasNEON, HasFullFP16]>;
6710 def VCVTs2hd : N2VD<0b11, 0b11, 0b01, 0b11, 0b01100, 0, "vcvt", "f16.s16",
6711 v4f16, v4i16, sint_to_fp>,
6712 Requires<[HasNEON, HasFullFP16]>;
6713 def VCVTu2hd : N2VD<0b11, 0b11, 0b01, 0b11, 0b01101, 0, "vcvt", "f16.u16",
6714 v4f16, v4i16, uint_to_fp>,
6715 Requires<[HasNEON, HasFullFP16]>;
6717 def VCVTh2sq : N2VQ<0b11, 0b11, 0b01, 0b11, 0b01110, 0, "vcvt", "s16.f16",
6718 v8i16, v8f16, fp_to_sint>,
6719 Requires<[HasNEON, HasFullFP16]>;
6720 def VCVTh2uq : N2VQ<0b11, 0b11, 0b01, 0b11, 0b01111, 0, "vcvt", "u16.f16",
6721 v8i16, v8f16, fp_to_uint>,
6722 Requires<[HasNEON, HasFullFP16]>;
6723 def VCVTs2hq : N2VQ<0b11, 0b11, 0b01, 0b11, 0b01100, 0, "vcvt", "f16.s16",
6724 v8f16, v8i16, sint_to_fp>,
6725 Requires<[HasNEON, HasFullFP16]>;
6726 def VCVTu2hq : N2VQ<0b11, 0b11, 0b01, 0b11, 0b01101, 0, "vcvt", "f16.u16",
6727 v8f16, v8i16, uint_to_fp>,
6728 Requires<[HasNEON, HasFullFP16]>;
6731 multiclass VCVT_FPI<string op, bits<3> op10_8, SDPatternOperator IntS,
6732 SDPatternOperator IntU> {
6733 let PostEncoderMethod = "NEONThumb2V8PostEncoder", DecoderNamespace = "v8NEON" in {
6734 def SDf : N2VDIntnp<0b10, 0b11, op10_8, 0, NoItinerary, !strconcat("vcvt", op),
6735 "s32.f32", v2i32, v2f32, IntS>, Requires<[HasV8, HasNEON]>;
6736 def SQf : N2VQIntnp<0b10, 0b11, op10_8, 0, NoItinerary, !strconcat("vcvt", op),
6737 "s32.f32", v4i32, v4f32, IntS>, Requires<[HasV8, HasNEON]>;
6738 def UDf : N2VDIntnp<0b10, 0b11, op10_8, 1, NoItinerary, !strconcat("vcvt", op),
6739 "u32.f32", v2i32, v2f32, IntU>, Requires<[HasV8, HasNEON]>;
6740 def UQf : N2VQIntnp<0b10, 0b11, op10_8, 1, NoItinerary, !strconcat("vcvt", op),
6741 "u32.f32", v4i32, v4f32, IntU>, Requires<[HasV8, HasNEON]>;
6742 def SDh : N2VDIntnp<0b01, 0b11, op10_8, 0, NoItinerary, !strconcat("vcvt", op),
6743 "s16.f16", v4i16, v4f16, IntS>,
6744 Requires<[HasV8, HasNEON, HasFullFP16]>;
6745 def SQh : N2VQIntnp<0b01, 0b11, op10_8, 0, NoItinerary, !strconcat("vcvt", op),
6746 "s16.f16", v8i16, v8f16, IntS>,
6747 Requires<[HasV8, HasNEON, HasFullFP16]>;
6748 def UDh : N2VDIntnp<0b01, 0b11, op10_8, 1, NoItinerary, !strconcat("vcvt", op),
6749 "u16.f16", v4i16, v4f16, IntU>,
6750 Requires<[HasV8, HasNEON, HasFullFP16]>;
6751 def UQh : N2VQIntnp<0b01, 0b11, op10_8, 1, NoItinerary, !strconcat("vcvt", op),
6752 "u16.f16", v8i16, v8f16, IntU>,
6753 Requires<[HasV8, HasNEON, HasFullFP16]>;
6757 defm VCVTAN : VCVT_FPI<"a", 0b000, int_arm_neon_vcvtas, int_arm_neon_vcvtau>;
6758 defm VCVTNN : VCVT_FPI<"n", 0b001, int_arm_neon_vcvtns, int_arm_neon_vcvtnu>;
6759 defm VCVTPN : VCVT_FPI<"p", 0b010, int_arm_neon_vcvtps, int_arm_neon_vcvtpu>;
6760 defm VCVTMN : VCVT_FPI<"m", 0b011, int_arm_neon_vcvtms, int_arm_neon_vcvtmu>;
6762 // VCVT : Vector Convert Between Floating-Point and Fixed-Point.
6763 let DecoderMethod = "DecodeVCVTD" in {
6764 def VCVTf2xsd : N2VCvtD<0, 1, 0b1111, 0, 1, "vcvt", "s32.f32",
6765 v2i32, v2f32, int_arm_neon_vcvtfp2fxs>;
6766 def VCVTf2xud : N2VCvtD<1, 1, 0b1111, 0, 1, "vcvt", "u32.f32",
6767 v2i32, v2f32, int_arm_neon_vcvtfp2fxu>;
6768 def VCVTxs2fd : N2VCvtD<0, 1, 0b1110, 0, 1, "vcvt", "f32.s32",
6769 v2f32, v2i32, int_arm_neon_vcvtfxs2fp>;
6770 def VCVTxu2fd : N2VCvtD<1, 1, 0b1110, 0, 1, "vcvt", "f32.u32",
6771 v2f32, v2i32, int_arm_neon_vcvtfxu2fp>;
6772 let Predicates = [HasNEON, HasFullFP16] in {
6773 def VCVTh2xsd : N2VCvtD<0, 1, 0b1101, 0, 1, "vcvt", "s16.f16",
6774 v4i16, v4f16, int_arm_neon_vcvtfp2fxs>;
6775 def VCVTh2xud : N2VCvtD<1, 1, 0b1101, 0, 1, "vcvt", "u16.f16",
6776 v4i16, v4f16, int_arm_neon_vcvtfp2fxu>;
6777 def VCVTxs2hd : N2VCvtD<0, 1, 0b1100, 0, 1, "vcvt", "f16.s16",
6778 v4f16, v4i16, int_arm_neon_vcvtfxs2fp>;
6779 def VCVTxu2hd : N2VCvtD<1, 1, 0b1100, 0, 1, "vcvt", "f16.u16",
6780 v4f16, v4i16, int_arm_neon_vcvtfxu2fp>;
6781 } // Predicates = [HasNEON, HasFullFP16]
6784 let DecoderMethod = "DecodeVCVTQ" in {
6785 def VCVTf2xsq : N2VCvtQ<0, 1, 0b1111, 0, 1, "vcvt", "s32.f32",
6786 v4i32, v4f32, int_arm_neon_vcvtfp2fxs>;
6787 def VCVTf2xuq : N2VCvtQ<1, 1, 0b1111, 0, 1, "vcvt", "u32.f32",
6788 v4i32, v4f32, int_arm_neon_vcvtfp2fxu>;
6789 def VCVTxs2fq : N2VCvtQ<0, 1, 0b1110, 0, 1, "vcvt", "f32.s32",
6790 v4f32, v4i32, int_arm_neon_vcvtfxs2fp>;
6791 def VCVTxu2fq : N2VCvtQ<1, 1, 0b1110, 0, 1, "vcvt", "f32.u32",
6792 v4f32, v4i32, int_arm_neon_vcvtfxu2fp>;
6793 let Predicates = [HasNEON, HasFullFP16] in {
6794 def VCVTh2xsq : N2VCvtQ<0, 1, 0b1101, 0, 1, "vcvt", "s16.f16",
6795 v8i16, v8f16, int_arm_neon_vcvtfp2fxs>;
6796 def VCVTh2xuq : N2VCvtQ<1, 1, 0b1101, 0, 1, "vcvt", "u16.f16",
6797 v8i16, v8f16, int_arm_neon_vcvtfp2fxu>;
6798 def VCVTxs2hq : N2VCvtQ<0, 1, 0b1100, 0, 1, "vcvt", "f16.s16",
6799 v8f16, v8i16, int_arm_neon_vcvtfxs2fp>;
6800 def VCVTxu2hq : N2VCvtQ<1, 1, 0b1100, 0, 1, "vcvt", "f16.u16",
6801 v8f16, v8i16, int_arm_neon_vcvtfxu2fp>;
6802 } // Predicates = [HasNEON, HasFullFP16]
6805 def : NEONInstAlias<"vcvt${p}.s32.f32 $Dd, $Dm, #0",
6806 (VCVTf2sd DPR:$Dd, DPR:$Dm, pred:$p)>;
6807 def : NEONInstAlias<"vcvt${p}.u32.f32 $Dd, $Dm, #0",
6808 (VCVTf2ud DPR:$Dd, DPR:$Dm, pred:$p)>;
6809 def : NEONInstAlias<"vcvt${p}.f32.s32 $Dd, $Dm, #0",
6810 (VCVTs2fd DPR:$Dd, DPR:$Dm, pred:$p)>;
6811 def : NEONInstAlias<"vcvt${p}.f32.u32 $Dd, $Dm, #0",
6812 (VCVTu2fd DPR:$Dd, DPR:$Dm, pred:$p)>;
6814 def : NEONInstAlias<"vcvt${p}.s32.f32 $Qd, $Qm, #0",
6815 (VCVTf2sq QPR:$Qd, QPR:$Qm, pred:$p)>;
6816 def : NEONInstAlias<"vcvt${p}.u32.f32 $Qd, $Qm, #0",
6817 (VCVTf2uq QPR:$Qd, QPR:$Qm, pred:$p)>;
6818 def : NEONInstAlias<"vcvt${p}.f32.s32 $Qd, $Qm, #0",
6819 (VCVTs2fq QPR:$Qd, QPR:$Qm, pred:$p)>;
6820 def : NEONInstAlias<"vcvt${p}.f32.u32 $Qd, $Qm, #0",
6821 (VCVTu2fq QPR:$Qd, QPR:$Qm, pred:$p)>;
6823 def : NEONInstAlias<"vcvt${p}.s16.f16 $Dd, $Dm, #0",
6824 (VCVTh2sd DPR:$Dd, DPR:$Dm, pred:$p)>;
6825 def : NEONInstAlias<"vcvt${p}.u16.f16 $Dd, $Dm, #0",
6826 (VCVTh2ud DPR:$Dd, DPR:$Dm, pred:$p)>;
6827 def : NEONInstAlias<"vcvt${p}.f16.s16 $Dd, $Dm, #0",
6828 (VCVTs2hd DPR:$Dd, DPR:$Dm, pred:$p)>;
6829 def : NEONInstAlias<"vcvt${p}.f16.u16 $Dd, $Dm, #0",
6830 (VCVTu2hd DPR:$Dd, DPR:$Dm, pred:$p)>;
6832 def : NEONInstAlias<"vcvt${p}.s16.f16 $Qd, $Qm, #0",
6833 (VCVTh2sq QPR:$Qd, QPR:$Qm, pred:$p)>;
6834 def : NEONInstAlias<"vcvt${p}.u16.f16 $Qd, $Qm, #0",
6835 (VCVTh2uq QPR:$Qd, QPR:$Qm, pred:$p)>;
6836 def : NEONInstAlias<"vcvt${p}.f16.s16 $Qd, $Qm, #0",
6837 (VCVTs2hq QPR:$Qd, QPR:$Qm, pred:$p)>;
6838 def : NEONInstAlias<"vcvt${p}.f16.u16 $Qd, $Qm, #0",
6839 (VCVTu2hq QPR:$Qd, QPR:$Qm, pred:$p)>;
6842 // VCVT : Vector Convert Between Half-Precision and Single-Precision.
6843 def VCVTf2h : N2VNInt<0b11, 0b11, 0b01, 0b10, 0b01100, 0, 0,
6844 IIC_VUNAQ, "vcvt", "f16.f32",
6845 v4i16, v4f32, int_arm_neon_vcvtfp2hf>,
6846 Requires<[HasNEON, HasFP16]>;
6847 def VCVTh2f : N2VLInt<0b11, 0b11, 0b01, 0b10, 0b01110, 0, 0,
6848 IIC_VUNAQ, "vcvt", "f32.f16",
6849 v4f32, v4i16, int_arm_neon_vcvthf2fp>,
6850 Requires<[HasNEON, HasFP16]>;
6854 // VREV64 : Vector Reverse elements within 64-bit doublewords
6856 class VREV64D<bits<2> op19_18, string OpcodeStr, string Dt, ValueType Ty>
6857 : N2V<0b11, 0b11, op19_18, 0b00, 0b00000, 0, 0, (outs DPR:$Vd),
6858 (ins DPR:$Vm), IIC_VMOVD,
6859 OpcodeStr, Dt, "$Vd, $Vm", "",
6860 [(set DPR:$Vd, (Ty (ARMvrev64 (Ty DPR:$Vm))))]>;
6861 class VREV64Q<bits<2> op19_18, string OpcodeStr, string Dt, ValueType Ty>
6862 : N2V<0b11, 0b11, op19_18, 0b00, 0b00000, 1, 0, (outs QPR:$Vd),
6863 (ins QPR:$Vm), IIC_VMOVQ,
6864 OpcodeStr, Dt, "$Vd, $Vm", "",
6865 [(set QPR:$Vd, (Ty (ARMvrev64 (Ty QPR:$Vm))))]>;
6867 def VREV64d8 : VREV64D<0b00, "vrev64", "8", v8i8>;
6868 def VREV64d16 : VREV64D<0b01, "vrev64", "16", v4i16>;
6869 def VREV64d32 : VREV64D<0b10, "vrev64", "32", v2i32>;
6870 let Predicates = [HasNEON] in {
6871 def : Pat<(v2f32 (ARMvrev64 (v2f32 DPR:$Vm))), (VREV64d32 DPR:$Vm)>;
6874 def VREV64q8 : VREV64Q<0b00, "vrev64", "8", v16i8>;
6875 def VREV64q16 : VREV64Q<0b01, "vrev64", "16", v8i16>;
6876 def VREV64q32 : VREV64Q<0b10, "vrev64", "32", v4i32>;
6878 let Predicates = [HasNEON] in {
6879 def : Pat<(v4f32 (ARMvrev64 (v4f32 QPR:$Vm))),
6880 (VREV64q32 QPR:$Vm)>;
6881 def : Pat<(v8f16 (ARMvrev64 (v8f16 QPR:$Vm))),
6882 (VREV64q16 QPR:$Vm)>;
6883 def : Pat<(v4f16 (ARMvrev64 (v4f16 DPR:$Vm))),
6884 (VREV64d16 DPR:$Vm)>;
6887 // VREV32 : Vector Reverse elements within 32-bit words
6889 class VREV32D<bits<2> op19_18, string OpcodeStr, string Dt, ValueType Ty>
6890 : N2V<0b11, 0b11, op19_18, 0b00, 0b00001, 0, 0, (outs DPR:$Vd),
6891 (ins DPR:$Vm), IIC_VMOVD,
6892 OpcodeStr, Dt, "$Vd, $Vm", "",
6893 [(set DPR:$Vd, (Ty (ARMvrev32 (Ty DPR:$Vm))))]>;
6894 class VREV32Q<bits<2> op19_18, string OpcodeStr, string Dt, ValueType Ty>
6895 : N2V<0b11, 0b11, op19_18, 0b00, 0b00001, 1, 0, (outs QPR:$Vd),
6896 (ins QPR:$Vm), IIC_VMOVQ,
6897 OpcodeStr, Dt, "$Vd, $Vm", "",
6898 [(set QPR:$Vd, (Ty (ARMvrev32 (Ty QPR:$Vm))))]>;
6900 def VREV32d8 : VREV32D<0b00, "vrev32", "8", v8i8>;
6901 def VREV32d16 : VREV32D<0b01, "vrev32", "16", v4i16>;
6903 def VREV32q8 : VREV32Q<0b00, "vrev32", "8", v16i8>;
6904 def VREV32q16 : VREV32Q<0b01, "vrev32", "16", v8i16>;
6906 let Predicates = [HasNEON] in {
6907 def : Pat<(v8f16 (ARMvrev32 (v8f16 QPR:$Vm))),
6908 (VREV32q16 QPR:$Vm)>;
6909 def : Pat<(v4f16 (ARMvrev32 (v4f16 DPR:$Vm))),
6910 (VREV32d16 DPR:$Vm)>;
6913 // VREV16 : Vector Reverse elements within 16-bit halfwords
6915 class VREV16D<bits<2> op19_18, string OpcodeStr, string Dt, ValueType Ty>
6916 : N2V<0b11, 0b11, op19_18, 0b00, 0b00010, 0, 0, (outs DPR:$Vd),
6917 (ins DPR:$Vm), IIC_VMOVD,
6918 OpcodeStr, Dt, "$Vd, $Vm", "",
6919 [(set DPR:$Vd, (Ty (ARMvrev16 (Ty DPR:$Vm))))]>;
6920 class VREV16Q<bits<2> op19_18, string OpcodeStr, string Dt, ValueType Ty>
6921 : N2V<0b11, 0b11, op19_18, 0b00, 0b00010, 1, 0, (outs QPR:$Vd),
6922 (ins QPR:$Vm), IIC_VMOVQ,
6923 OpcodeStr, Dt, "$Vd, $Vm", "",
6924 [(set QPR:$Vd, (Ty (ARMvrev16 (Ty QPR:$Vm))))]>;
6926 def VREV16d8 : VREV16D<0b00, "vrev16", "8", v8i8>;
6927 def VREV16q8 : VREV16Q<0b00, "vrev16", "8", v16i8>;
6929 // Other Vector Shuffles.
6931 // Aligned extractions: really just dropping registers
6933 class AlignedVEXTq<ValueType DestTy, ValueType SrcTy, SDNodeXForm LaneCVT>
6934 : Pat<(DestTy (vector_extract_subvec (SrcTy QPR:$src), (i32 imm:$start))),
6935 (EXTRACT_SUBREG (SrcTy QPR:$src), (LaneCVT imm:$start))>,
6936 Requires<[HasNEON]>;
6938 def : AlignedVEXTq<v8i8, v16i8, DSubReg_i8_reg>;
6940 def : AlignedVEXTq<v4i16, v8i16, DSubReg_i16_reg>;
6942 def : AlignedVEXTq<v2i32, v4i32, DSubReg_i32_reg>;
6944 def : AlignedVEXTq<v1i64, v2i64, DSubReg_f64_reg>;
6946 def : AlignedVEXTq<v2f32, v4f32, DSubReg_i32_reg>;
6948 def : AlignedVEXTq<v4f16, v8f16, DSubReg_i16_reg>; // v8f16 -> v4f16
6950 // VEXT : Vector Extract
6953 // All of these have a two-operand InstAlias.
6954 let TwoOperandAliasConstraint = "$Vn = $Vd" in {
6955 class VEXTd<string OpcodeStr, string Dt, ValueType Ty, Operand immTy>
6956 : N3V<0,1,0b11,{?,?,?,?},0,0, (outs DPR:$Vd),
6957 (ins DPR:$Vn, DPR:$Vm, immTy:$index), NVExtFrm,
6958 IIC_VEXTD, OpcodeStr, Dt, "$Vd, $Vn, $Vm, $index", "",
6959 [(set DPR:$Vd, (Ty (NEONvext (Ty DPR:$Vn),
6960 (Ty DPR:$Vm), imm:$index)))]> {
6963 let Inst{10-8} = index{2-0};
6966 class VEXTq<string OpcodeStr, string Dt, ValueType Ty, Operand immTy>
6967 : N3V<0,1,0b11,{?,?,?,?},1,0, (outs QPR:$Vd),
6968 (ins QPR:$Vn, QPR:$Vm, imm0_15:$index), NVExtFrm,
6969 IIC_VEXTQ, OpcodeStr, Dt, "$Vd, $Vn, $Vm, $index", "",
6970 [(set QPR:$Vd, (Ty (NEONvext (Ty QPR:$Vn),
6971 (Ty QPR:$Vm), imm:$index)))]> {
6973 let Inst{11-8} = index{3-0};
6977 def VEXTd8 : VEXTd<"vext", "8", v8i8, imm0_7> {
6978 let Inst{10-8} = index{2-0};
6980 def VEXTd16 : VEXTd<"vext", "16", v4i16, imm0_3> {
6981 let Inst{10-9} = index{1-0};
6984 let Predicates = [HasNEON] in {
6985 def : Pat<(v4f16 (NEONvext (v4f16 DPR:$Vn), (v4f16 DPR:$Vm), (i32 imm:$index))),
6986 (VEXTd16 DPR:$Vn, DPR:$Vm, imm:$index)>;
6989 def VEXTd32 : VEXTd<"vext", "32", v2i32, imm0_1> {
6990 let Inst{10} = index{0};
6991 let Inst{9-8} = 0b00;
6993 let Predicates = [HasNEON] in {
6994 def : Pat<(v2f32 (NEONvext (v2f32 DPR:$Vn), (v2f32 DPR:$Vm), (i32 imm:$index))),
6995 (VEXTd32 DPR:$Vn, DPR:$Vm, imm:$index)>;
6998 def VEXTq8 : VEXTq<"vext", "8", v16i8, imm0_15> {
6999 let Inst{11-8} = index{3-0};
7001 def VEXTq16 : VEXTq<"vext", "16", v8i16, imm0_7> {
7002 let Inst{11-9} = index{2-0};
7005 let Predicates = [HasNEON] in {
7006 def : Pat<(v8f16 (NEONvext (v8f16 QPR:$Vn), (v8f16 QPR:$Vm), (i32 imm:$index))),
7007 (VEXTq16 QPR:$Vn, QPR:$Vm, imm:$index)>;
7010 def VEXTq32 : VEXTq<"vext", "32", v4i32, imm0_3> {
7011 let Inst{11-10} = index{1-0};
7012 let Inst{9-8} = 0b00;
7014 def VEXTq64 : VEXTq<"vext", "64", v2i64, imm0_1> {
7015 let Inst{11} = index{0};
7016 let Inst{10-8} = 0b000;
7018 let Predicates = [HasNEON] in {
7019 def : Pat<(v4f32 (NEONvext (v4f32 QPR:$Vn), (v4f32 QPR:$Vm), (i32 imm:$index))),
7020 (VEXTq32 QPR:$Vn, QPR:$Vm, imm:$index)>;
7023 // VTRN : Vector Transpose
7025 def VTRNd8 : N2VDShuffle<0b00, 0b00001, "vtrn", "8">;
7026 def VTRNd16 : N2VDShuffle<0b01, 0b00001, "vtrn", "16">;
7027 def VTRNd32 : N2VDShuffle<0b10, 0b00001, "vtrn", "32">;
7029 def VTRNq8 : N2VQShuffle<0b00, 0b00001, IIC_VPERMQ, "vtrn", "8">;
7030 def VTRNq16 : N2VQShuffle<0b01, 0b00001, IIC_VPERMQ, "vtrn", "16">;
7031 def VTRNq32 : N2VQShuffle<0b10, 0b00001, IIC_VPERMQ, "vtrn", "32">;
7033 // VUZP : Vector Unzip (Deinterleave)
7035 def VUZPd8 : N2VDShuffle<0b00, 0b00010, "vuzp", "8">;
7036 def VUZPd16 : N2VDShuffle<0b01, 0b00010, "vuzp", "16">;
7037 // vuzp.32 Dd, Dm is a pseudo-instruction expanded to vtrn.32 Dd, Dm.
7038 def : NEONInstAlias<"vuzp${p}.32 $Dd, $Dm",
7039 (VTRNd32 DPR:$Dd, DPR:$Dm, pred:$p)>;
7041 def VUZPq8 : N2VQShuffle<0b00, 0b00010, IIC_VPERMQ3, "vuzp", "8">;
7042 def VUZPq16 : N2VQShuffle<0b01, 0b00010, IIC_VPERMQ3, "vuzp", "16">;
7043 def VUZPq32 : N2VQShuffle<0b10, 0b00010, IIC_VPERMQ3, "vuzp", "32">;
7045 // VZIP : Vector Zip (Interleave)
7047 def VZIPd8 : N2VDShuffle<0b00, 0b00011, "vzip", "8">;
7048 def VZIPd16 : N2VDShuffle<0b01, 0b00011, "vzip", "16">;
7049 // vzip.32 Dd, Dm is a pseudo-instruction expanded to vtrn.32 Dd, Dm.
7050 def : NEONInstAlias<"vzip${p}.32 $Dd, $Dm",
7051 (VTRNd32 DPR:$Dd, DPR:$Dm, pred:$p)>;
7053 def VZIPq8 : N2VQShuffle<0b00, 0b00011, IIC_VPERMQ3, "vzip", "8">;
7054 def VZIPq16 : N2VQShuffle<0b01, 0b00011, IIC_VPERMQ3, "vzip", "16">;
7055 def VZIPq32 : N2VQShuffle<0b10, 0b00011, IIC_VPERMQ3, "vzip", "32">;
7057 // Vector Table Lookup and Table Extension.
7059 // VTBL : Vector Table Lookup
7060 let DecoderMethod = "DecodeTBLInstruction" in {
7062 : N3V<1,1,0b11,0b1000,0,0, (outs DPR:$Vd),
7063 (ins VecListOneD:$Vn, DPR:$Vm), NVTBLFrm, IIC_VTB1,
7064 "vtbl", "8", "$Vd, $Vn, $Vm", "",
7065 [(set DPR:$Vd, (v8i8 (NEONvtbl1 VecListOneD:$Vn, DPR:$Vm)))]>;
7067 let hasExtraSrcRegAllocReq = 1 in {
7069 : N3V<1,1,0b11,0b1001,0,0, (outs DPR:$Vd),
7070 (ins VecListDPair:$Vn, DPR:$Vm), NVTBLFrm, IIC_VTB2,
7071 "vtbl", "8", "$Vd, $Vn, $Vm", "", []>;
7073 : N3V<1,1,0b11,0b1010,0,0, (outs DPR:$Vd),
7074 (ins VecListThreeD:$Vn, DPR:$Vm), NVTBLFrm, IIC_VTB3,
7075 "vtbl", "8", "$Vd, $Vn, $Vm", "", []>;
7077 : N3V<1,1,0b11,0b1011,0,0, (outs DPR:$Vd),
7078 (ins VecListFourD:$Vn, DPR:$Vm),
7080 "vtbl", "8", "$Vd, $Vn, $Vm", "", []>;
7081 } // hasExtraSrcRegAllocReq = 1
7084 : PseudoNeonI<(outs DPR:$dst), (ins QQPR:$tbl, DPR:$src), IIC_VTB3, "", []>;
7086 : PseudoNeonI<(outs DPR:$dst), (ins QQPR:$tbl, DPR:$src), IIC_VTB4, "", []>;
7088 // VTBX : Vector Table Extension
7090 : N3V<1,1,0b11,0b1000,1,0, (outs DPR:$Vd),
7091 (ins DPR:$orig, VecListOneD:$Vn, DPR:$Vm), NVTBLFrm, IIC_VTBX1,
7092 "vtbx", "8", "$Vd, $Vn, $Vm", "$orig = $Vd",
7093 [(set DPR:$Vd, (v8i8 (int_arm_neon_vtbx1
7094 DPR:$orig, VecListOneD:$Vn, DPR:$Vm)))]>;
7095 let hasExtraSrcRegAllocReq = 1 in {
7097 : N3V<1,1,0b11,0b1001,1,0, (outs DPR:$Vd),
7098 (ins DPR:$orig, VecListDPair:$Vn, DPR:$Vm), NVTBLFrm, IIC_VTBX2,
7099 "vtbx", "8", "$Vd, $Vn, $Vm", "$orig = $Vd", []>;
7101 : N3V<1,1,0b11,0b1010,1,0, (outs DPR:$Vd),
7102 (ins DPR:$orig, VecListThreeD:$Vn, DPR:$Vm),
7103 NVTBLFrm, IIC_VTBX3,
7104 "vtbx", "8", "$Vd, $Vn, $Vm",
7107 : N3V<1,1,0b11,0b1011,1,0, (outs DPR:$Vd),
7108 (ins DPR:$orig, VecListFourD:$Vn, DPR:$Vm), NVTBLFrm, IIC_VTBX4,
7109 "vtbx", "8", "$Vd, $Vn, $Vm",
7111 } // hasExtraSrcRegAllocReq = 1
7114 : PseudoNeonI<(outs DPR:$dst), (ins DPR:$orig, QQPR:$tbl, DPR:$src),
7115 IIC_VTBX3, "$orig = $dst", []>;
7117 : PseudoNeonI<(outs DPR:$dst), (ins DPR:$orig, QQPR:$tbl, DPR:$src),
7118 IIC_VTBX4, "$orig = $dst", []>;
7119 } // DecoderMethod = "DecodeTBLInstruction"
7121 let Predicates = [HasNEON] in {
7122 def : Pat<(v8i8 (NEONvtbl2 v8i8:$Vn0, v8i8:$Vn1, v8i8:$Vm)),
7123 (v8i8 (VTBL2 (REG_SEQUENCE DPair, v8i8:$Vn0, dsub_0,
7126 def : Pat<(v8i8 (int_arm_neon_vtbx2 v8i8:$orig, v8i8:$Vn0, v8i8:$Vn1,
7128 (v8i8 (VTBX2 v8i8:$orig,
7129 (REG_SEQUENCE DPair, v8i8:$Vn0, dsub_0,
7133 def : Pat<(v8i8 (int_arm_neon_vtbl3 v8i8:$Vn0, v8i8:$Vn1,
7134 v8i8:$Vn2, v8i8:$Vm)),
7135 (v8i8 (VTBL3Pseudo (REG_SEQUENCE QQPR, v8i8:$Vn0, dsub_0,
7138 (v8i8 (IMPLICIT_DEF)), dsub_3),
7140 def : Pat<(v8i8 (int_arm_neon_vtbx3 v8i8:$orig, v8i8:$Vn0, v8i8:$Vn1,
7141 v8i8:$Vn2, v8i8:$Vm)),
7142 (v8i8 (VTBX3Pseudo v8i8:$orig,
7143 (REG_SEQUENCE QQPR, v8i8:$Vn0, dsub_0,
7146 (v8i8 (IMPLICIT_DEF)), dsub_3),
7149 def : Pat<(v8i8 (int_arm_neon_vtbl4 v8i8:$Vn0, v8i8:$Vn1,
7150 v8i8:$Vn2, v8i8:$Vn3, v8i8:$Vm)),
7151 (v8i8 (VTBL4Pseudo (REG_SEQUENCE QQPR, v8i8:$Vn0, dsub_0,
7156 def : Pat<(v8i8 (int_arm_neon_vtbx4 v8i8:$orig, v8i8:$Vn0, v8i8:$Vn1,
7157 v8i8:$Vn2, v8i8:$Vn3, v8i8:$Vm)),
7158 (v8i8 (VTBX4Pseudo v8i8:$orig,
7159 (REG_SEQUENCE QQPR, v8i8:$Vn0, dsub_0,
7166 // VRINT : Vector Rounding
7167 multiclass VRINT_FPI<string op, bits<3> op9_7, SDPatternOperator Int> {
7168 let PostEncoderMethod = "NEONThumb2V8PostEncoder", DecoderNamespace = "v8NEON" in {
7169 def Df : N2VDIntnp<0b10, 0b10, 0b100, 0, NoItinerary,
7170 !strconcat("vrint", op), "f32",
7171 v2f32, v2f32, Int>, Requires<[HasV8, HasNEON]> {
7172 let Inst{9-7} = op9_7;
7174 def Qf : N2VQIntnp<0b10, 0b10, 0b100, 0, NoItinerary,
7175 !strconcat("vrint", op), "f32",
7176 v4f32, v4f32, Int>, Requires<[HasV8, HasNEON]> {
7177 let Inst{9-7} = op9_7;
7179 def Dh : N2VDIntnp<0b01, 0b10, 0b100, 0, NoItinerary,
7180 !strconcat("vrint", op), "f16",
7182 Requires<[HasV8, HasNEON, HasFullFP16]> {
7183 let Inst{9-7} = op9_7;
7185 def Qh : N2VQIntnp<0b01, 0b10, 0b100, 0, NoItinerary,
7186 !strconcat("vrint", op), "f16",
7188 Requires<[HasV8, HasNEON, HasFullFP16]> {
7189 let Inst{9-7} = op9_7;
7193 def : NEONInstAlias<!strconcat("vrint", op, ".f32.f32\t$Dd, $Dm"),
7194 (!cast<Instruction>(NAME#"Df") DPR:$Dd, DPR:$Dm)>;
7195 def : NEONInstAlias<!strconcat("vrint", op, ".f32.f32\t$Qd, $Qm"),
7196 (!cast<Instruction>(NAME#"Qf") QPR:$Qd, QPR:$Qm)>;
7197 let Predicates = [HasNEON, HasFullFP16] in {
7198 def : NEONInstAlias<!strconcat("vrint", op, ".f16.f16\t$Dd, $Dm"),
7199 (!cast<Instruction>(NAME#"Dh") DPR:$Dd, DPR:$Dm)>;
7200 def : NEONInstAlias<!strconcat("vrint", op, ".f16.f16\t$Qd, $Qm"),
7201 (!cast<Instruction>(NAME#"Qh") QPR:$Qd, QPR:$Qm)>;
7205 defm VRINTNN : VRINT_FPI<"n", 0b000, int_arm_neon_vrintn>;
7206 defm VRINTXN : VRINT_FPI<"x", 0b001, int_arm_neon_vrintx>;
7207 defm VRINTAN : VRINT_FPI<"a", 0b010, int_arm_neon_vrinta>;
7208 defm VRINTZN : VRINT_FPI<"z", 0b011, int_arm_neon_vrintz>;
7209 defm VRINTMN : VRINT_FPI<"m", 0b101, int_arm_neon_vrintm>;
7210 defm VRINTPN : VRINT_FPI<"p", 0b111, int_arm_neon_vrintp>;
7212 // Cryptography instructions
7213 let PostEncoderMethod = "NEONThumb2DataIPostEncoder",
7214 DecoderNamespace = "v8Crypto", hasSideEffects = 0 in {
7215 class AES<string op, bit op7, bit op6, SDPatternOperator Int>
7216 : N2VQIntXnp<0b00, 0b00, 0b011, op6, op7, NoItinerary,
7217 !strconcat("aes", op), "8", v16i8, v16i8, Int>,
7218 Requires<[HasV8, HasCrypto]>;
7219 class AES2Op<string op, bit op7, bit op6, SDPatternOperator Int>
7220 : N2VQIntX2np<0b00, 0b00, 0b011, op6, op7, NoItinerary,
7221 !strconcat("aes", op), "8", v16i8, v16i8, Int>,
7222 Requires<[HasV8, HasCrypto]>;
7223 class N2SHA<string op, bits<2> op17_16, bits<3> op10_8, bit op7, bit op6,
7224 SDPatternOperator Int>
7225 : N2VQIntXnp<0b10, op17_16, op10_8, op6, op7, NoItinerary,
7226 !strconcat("sha", op), "32", v4i32, v4i32, Int>,
7227 Requires<[HasV8, HasCrypto]>;
7228 class N2SHA2Op<string op, bits<2> op17_16, bits<3> op10_8, bit op7, bit op6,
7229 SDPatternOperator Int>
7230 : N2VQIntX2np<0b10, op17_16, op10_8, op6, op7, NoItinerary,
7231 !strconcat("sha", op), "32", v4i32, v4i32, Int>,
7232 Requires<[HasV8, HasCrypto]>;
7233 class N3SHA3Op<string op, bits<5> op27_23, bits<2> op21_20, SDPatternOperator Int>
7234 : N3VQInt3np<op27_23, op21_20, 0b1100, 1, 0, N3RegFrm, NoItinerary,
7235 !strconcat("sha", op), "32", v4i32, v4i32, Int, 0>,
7236 Requires<[HasV8, HasCrypto]>;
7239 def AESD : AES2Op<"d", 0, 1, int_arm_neon_aesd>;
7240 def AESE : AES2Op<"e", 0, 0, int_arm_neon_aese>;
7241 def AESIMC : AES<"imc", 1, 1, int_arm_neon_aesimc>;
7242 def AESMC : AES<"mc", 1, 0, int_arm_neon_aesmc>;
7244 def SHA1H : N2SHA<"1h", 0b01, 0b010, 1, 1, null_frag>;
7245 def SHA1SU1 : N2SHA2Op<"1su1", 0b10, 0b011, 1, 0, int_arm_neon_sha1su1>;
7246 def SHA256SU0 : N2SHA2Op<"256su0", 0b10, 0b011, 1, 1, int_arm_neon_sha256su0>;
7247 def SHA1C : N3SHA3Op<"1c", 0b00100, 0b00, null_frag>;
7248 def SHA1M : N3SHA3Op<"1m", 0b00100, 0b10, null_frag>;
7249 def SHA1P : N3SHA3Op<"1p", 0b00100, 0b01, null_frag>;
7250 def SHA1SU0 : N3SHA3Op<"1su0", 0b00100, 0b11, int_arm_neon_sha1su0>;
7251 def SHA256H : N3SHA3Op<"256h", 0b00110, 0b00, int_arm_neon_sha256h>;
7252 def SHA256H2 : N3SHA3Op<"256h2", 0b00110, 0b01, int_arm_neon_sha256h2>;
7253 def SHA256SU1 : N3SHA3Op<"256su1", 0b00110, 0b10, int_arm_neon_sha256su1>;
7255 let Predicates = [HasNEON] in {
7256 def : Pat<(i32 (int_arm_neon_sha1h i32:$Rn)),
7257 (COPY_TO_REGCLASS (f32 (EXTRACT_SUBREG
7258 (SHA1H (SUBREG_TO_REG (i64 0),
7259 (f32 (COPY_TO_REGCLASS i32:$Rn, SPR)),
7263 def : Pat<(v4i32 (int_arm_neon_sha1c v4i32:$hash_abcd, i32:$hash_e, v4i32:$wk)),
7264 (SHA1C v4i32:$hash_abcd,
7265 (SUBREG_TO_REG (i64 0),
7266 (f32 (COPY_TO_REGCLASS i32:$hash_e, SPR)),
7270 def : Pat<(v4i32 (int_arm_neon_sha1m v4i32:$hash_abcd, i32:$hash_e, v4i32:$wk)),
7271 (SHA1M v4i32:$hash_abcd,
7272 (SUBREG_TO_REG (i64 0),
7273 (f32 (COPY_TO_REGCLASS i32:$hash_e, SPR)),
7277 def : Pat<(v4i32 (int_arm_neon_sha1p v4i32:$hash_abcd, i32:$hash_e, v4i32:$wk)),
7278 (SHA1P v4i32:$hash_abcd,
7279 (SUBREG_TO_REG (i64 0),
7280 (f32 (COPY_TO_REGCLASS i32:$hash_e, SPR)),
7285 //===----------------------------------------------------------------------===//
7286 // NEON instructions for single-precision FP math
7287 //===----------------------------------------------------------------------===//
7289 class N2VSPat<SDNode OpNode, NeonI Inst>
7290 : NEONFPPat<(f32 (OpNode SPR:$a)),
7292 (v2f32 (COPY_TO_REGCLASS (Inst
7294 (v2f32 (COPY_TO_REGCLASS (v2f32 (IMPLICIT_DEF)), DPR_VFP2)),
7295 SPR:$a, ssub_0)), DPR_VFP2)), ssub_0)>;
7297 class N3VSPat<SDNode OpNode, NeonI Inst>
7298 : NEONFPPat<(f32 (OpNode SPR:$a, SPR:$b)),
7300 (v2f32 (COPY_TO_REGCLASS (Inst
7302 (v2f32 (COPY_TO_REGCLASS (v2f32 (IMPLICIT_DEF)), DPR_VFP2)),
7305 (v2f32 (COPY_TO_REGCLASS (v2f32 (IMPLICIT_DEF)), DPR_VFP2)),
7306 SPR:$b, ssub_0)), DPR_VFP2)), ssub_0)>;
7308 class N3VSPatFP16<SDNode OpNode, NeonI Inst>
7309 : NEONFPPat<(f16 (OpNode HPR:$a, HPR:$b)),
7311 (v4f16 (COPY_TO_REGCLASS (Inst
7313 (v4f16 (COPY_TO_REGCLASS (v4f16 (IMPLICIT_DEF)), DPR_VFP2)),
7316 (v4f16 (COPY_TO_REGCLASS (v4f16 (IMPLICIT_DEF)), DPR_VFP2)),
7317 HPR:$b, ssub_0)), DPR_VFP2)), ssub_0)>;
7319 class N3VSMulOpPat<SDNode MulNode, SDNode OpNode, NeonI Inst>
7320 : NEONFPPat<(f32 (OpNode SPR:$acc, (f32 (MulNode SPR:$a, SPR:$b)))),
7322 (v2f32 (COPY_TO_REGCLASS (Inst
7324 (v2f32 (COPY_TO_REGCLASS (v2f32 (IMPLICIT_DEF)), DPR_VFP2)),
7327 (v2f32 (COPY_TO_REGCLASS (v2f32 (IMPLICIT_DEF)), DPR_VFP2)),
7330 (v2f32 (COPY_TO_REGCLASS (v2f32 (IMPLICIT_DEF)), DPR_VFP2)),
7331 SPR:$b, ssub_0)), DPR_VFP2)), ssub_0)>;
7333 class NVCVTIFPat<SDNode OpNode, NeonI Inst>
7334 : NEONFPPat<(f32 (OpNode GPR:$a)),
7335 (f32 (EXTRACT_SUBREG
7338 (v2f32 (IMPLICIT_DEF)),
7339 (i32 (COPY_TO_REGCLASS GPR:$a, SPR)), ssub_0))),
7341 class NVCVTFIPat<SDNode OpNode, NeonI Inst>
7342 : NEONFPPat<(i32 (OpNode SPR:$a)),
7343 (i32 (EXTRACT_SUBREG
7344 (v2f32 (Inst (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)),
7348 def : N3VSPat<fadd, VADDfd>;
7349 def : N3VSPat<fsub, VSUBfd>;
7350 def : N3VSPat<fmul, VMULfd>;
7351 def : N3VSMulOpPat<fmul, fadd, VMLAfd>,
7352 Requires<[HasNEON, UseNEONForFP, UseFPVMLx]>;
7353 def : N3VSMulOpPat<fmul, fsub, VMLSfd>,
7354 Requires<[HasNEON, UseNEONForFP, UseFPVMLx]>;
7355 def : N3VSMulOpPat<fmul, fadd, VFMAfd>,
7356 Requires<[HasVFP4, UseNEONForFP, UseFusedMAC]>;
7357 def : N3VSMulOpPat<fmul, fsub, VFMSfd>,
7358 Requires<[HasVFP4, UseNEONForFP, UseFusedMAC]>;
7359 def : N2VSPat<fabs, VABSfd>;
7360 def : N2VSPat<fneg, VNEGfd>;
7361 def : N3VSPatFP16<fmaximum, VMAXhd>, Requires<[HasFullFP16]>;
7362 def : N3VSPatFP16<fminimum, VMINhd>, Requires<[HasFullFP16]>;
7363 def : N3VSPat<fmaximum, VMAXfd>, Requires<[HasNEON]>;
7364 def : N3VSPat<fminimum, VMINfd>, Requires<[HasNEON]>;
7365 def : NVCVTFIPat<fp_to_sint, VCVTf2sd>;
7366 def : NVCVTFIPat<fp_to_uint, VCVTf2ud>;
7367 def : NVCVTIFPat<sint_to_fp, VCVTs2fd>;
7368 def : NVCVTIFPat<uint_to_fp, VCVTu2fd>;
7370 // NEON doesn't have any f64 conversions, so provide patterns to make
7371 // sure the VFP conversions match when extracting from a vector.
7372 def : VFPPat<(f64 (sint_to_fp (extractelt (v2i32 DPR:$src), imm:$lane))),
7373 (VSITOD (EXTRACT_SUBREG DPR:$src, (SSubReg_f32_reg imm:$lane)))>;
7374 def : VFPPat<(f64 (sint_to_fp (extractelt (v4i32 QPR:$src), imm:$lane))),
7375 (VSITOD (EXTRACT_SUBREG QPR:$src, (SSubReg_f32_reg imm:$lane)))>;
7376 def : VFPPat<(f64 (uint_to_fp (extractelt (v2i32 DPR:$src), imm:$lane))),
7377 (VUITOD (EXTRACT_SUBREG DPR:$src, (SSubReg_f32_reg imm:$lane)))>;
7378 def : VFPPat<(f64 (uint_to_fp (extractelt (v4i32 QPR:$src), imm:$lane))),
7379 (VUITOD (EXTRACT_SUBREG QPR:$src, (SSubReg_f32_reg imm:$lane)))>;
7382 // Prefer VMOVDRR for i32 -> f32 bitcasts, it can write all DPR registers.
7383 def : Pat<(f32 (bitconvert GPR:$a)),
7384 (EXTRACT_SUBREG (VMOVDRR GPR:$a, GPR:$a), ssub_0)>,
7385 Requires<[HasNEON, DontUseVMOVSR]>;
7386 def : Pat<(arm_vmovsr GPR:$a),
7387 (EXTRACT_SUBREG (VMOVDRR GPR:$a, GPR:$a), ssub_0)>,
7388 Requires<[HasNEON, DontUseVMOVSR]>;
7390 //===----------------------------------------------------------------------===//
7391 // Non-Instruction Patterns or Endianess - Revert Patterns
7392 //===----------------------------------------------------------------------===//
7395 // 64 bit conversions
7396 let Predicates = [HasNEON] in {
7397 def : Pat<(f64 (bitconvert (v1i64 DPR:$src))), (f64 DPR:$src)>;
7398 def : Pat<(v1i64 (bitconvert (f64 DPR:$src))), (v1i64 DPR:$src)>;
7400 def : Pat<(v2f32 (bitconvert (v2i32 DPR:$src))), (v2f32 DPR:$src)>;
7401 def : Pat<(v2i32 (bitconvert (v2f32 DPR:$src))), (v2i32 DPR:$src)>;
7403 def : Pat<(v4i16 (bitconvert (v4f16 DPR:$src))), (v4i16 DPR:$src)>;
7404 def : Pat<(v4f16 (bitconvert (v4i16 DPR:$src))), (v4f16 DPR:$src)>;
7406 def : Pat<(v4i16 (bitconvert (v4bf16 DPR:$src))), (v4i16 DPR:$src)>;
7407 def : Pat<(v4bf16 (bitconvert (v4i16 DPR:$src))), (v4bf16 DPR:$src)>;
7409 // 128 bit conversions
7410 def : Pat<(v2f64 (bitconvert (v2i64 QPR:$src))), (v2f64 QPR:$src)>;
7411 def : Pat<(v2i64 (bitconvert (v2f64 QPR:$src))), (v2i64 QPR:$src)>;
7413 def : Pat<(v4i32 (bitconvert (v4f32 QPR:$src))), (v4i32 QPR:$src)>;
7414 def : Pat<(v4f32 (bitconvert (v4i32 QPR:$src))), (v4f32 QPR:$src)>;
7416 def : Pat<(v8i16 (bitconvert (v8f16 QPR:$src))), (v8i16 QPR:$src)>;
7417 def : Pat<(v8f16 (bitconvert (v8i16 QPR:$src))), (v8f16 QPR:$src)>;
7419 def : Pat<(v8i16 (bitconvert (v8bf16 QPR:$src))), (v8i16 QPR:$src)>;
7420 def : Pat<(v8bf16 (bitconvert (v8i16 QPR:$src))), (v8bf16 QPR:$src)>;
7423 let Predicates = [IsLE,HasNEON] in {
7424 // 64 bit conversions
7425 def : Pat<(f64 (bitconvert (v2f32 DPR:$src))), (f64 DPR:$src)>;
7426 def : Pat<(f64 (bitconvert (v2i32 DPR:$src))), (f64 DPR:$src)>;
7427 def : Pat<(f64 (bitconvert (v4f16 DPR:$src))), (f64 DPR:$src)>;
7428 def : Pat<(f64 (bitconvert (v4bf16 DPR:$src))), (f64 DPR:$src)>;
7429 def : Pat<(f64 (bitconvert (v4i16 DPR:$src))), (f64 DPR:$src)>;
7430 def : Pat<(f64 (bitconvert (v8i8 DPR:$src))), (f64 DPR:$src)>;
7432 def : Pat<(v1i64 (bitconvert (v2f32 DPR:$src))), (v1i64 DPR:$src)>;
7433 def : Pat<(v1i64 (bitconvert (v2i32 DPR:$src))), (v1i64 DPR:$src)>;
7434 def : Pat<(v1i64 (bitconvert (v4f16 DPR:$src))), (v1i64 DPR:$src)>;
7435 def : Pat<(v1i64 (bitconvert (v4bf16 DPR:$src))), (v1i64 DPR:$src)>;
7436 def : Pat<(v1i64 (bitconvert (v4i16 DPR:$src))), (v1i64 DPR:$src)>;
7437 def : Pat<(v1i64 (bitconvert (v8i8 DPR:$src))), (v1i64 DPR:$src)>;
7439 def : Pat<(v2f32 (bitconvert (f64 DPR:$src))), (v2f32 DPR:$src)>;
7440 def : Pat<(v2f32 (bitconvert (v1i64 DPR:$src))), (v2f32 DPR:$src)>;
7441 def : Pat<(v2f32 (bitconvert (v4f16 DPR:$src))), (v2f32 DPR:$src)>;
7442 def : Pat<(v2f32 (bitconvert (v4bf16 DPR:$src))), (v2f32 DPR:$src)>;
7443 def : Pat<(v2f32 (bitconvert (v4i16 DPR:$src))), (v2f32 DPR:$src)>;
7444 def : Pat<(v2f32 (bitconvert (v8i8 DPR:$src))), (v2f32 DPR:$src)>;
7446 def : Pat<(v2i32 (bitconvert (f64 DPR:$src))), (v2i32 DPR:$src)>;
7447 def : Pat<(v2i32 (bitconvert (v1i64 DPR:$src))), (v2i32 DPR:$src)>;
7448 def : Pat<(v2i32 (bitconvert (v4f16 DPR:$src))), (v2i32 DPR:$src)>;
7449 def : Pat<(v2i32 (bitconvert (v4bf16 DPR:$src))), (v2i32 DPR:$src)>;
7450 def : Pat<(v2i32 (bitconvert (v4i16 DPR:$src))), (v2i32 DPR:$src)>;
7451 def : Pat<(v2i32 (bitconvert (v8i8 DPR:$src))), (v2i32 DPR:$src)>;
7453 def : Pat<(v4f16 (bitconvert (f64 DPR:$src))), (v4f16 DPR:$src)>;
7454 def : Pat<(v4f16 (bitconvert (v1i64 DPR:$src))), (v4f16 DPR:$src)>;
7455 def : Pat<(v4f16 (bitconvert (v2f32 DPR:$src))), (v4f16 DPR:$src)>;
7456 def : Pat<(v4f16 (bitconvert (v2i32 DPR:$src))), (v4f16 DPR:$src)>;
7457 def : Pat<(v4f16 (bitconvert (v8i8 DPR:$src))), (v4f16 DPR:$src)>;
7459 def : Pat<(v4bf16 (bitconvert (f64 DPR:$src))), (v4bf16 DPR:$src)>;
7460 def : Pat<(v4bf16 (bitconvert (v1i64 DPR:$src))), (v4bf16 DPR:$src)>;
7461 def : Pat<(v4bf16 (bitconvert (v2f32 DPR:$src))), (v4bf16 DPR:$src)>;
7462 def : Pat<(v4bf16 (bitconvert (v2i32 DPR:$src))), (v4bf16 DPR:$src)>;
7463 def : Pat<(v4bf16 (bitconvert (v8i8 DPR:$src))), (v4bf16 DPR:$src)>;
7465 def : Pat<(v4i16 (bitconvert (f64 DPR:$src))), (v4i16 DPR:$src)>;
7466 def : Pat<(v4i16 (bitconvert (v1i64 DPR:$src))), (v4i16 DPR:$src)>;
7467 def : Pat<(v4i16 (bitconvert (v2f32 DPR:$src))), (v4i16 DPR:$src)>;
7468 def : Pat<(v4i16 (bitconvert (v2i32 DPR:$src))), (v4i16 DPR:$src)>;
7469 def : Pat<(v4i16 (bitconvert (v8i8 DPR:$src))), (v4i16 DPR:$src)>;
7471 def : Pat<(v8i8 (bitconvert (f64 DPR:$src))), (v8i8 DPR:$src)>;
7472 def : Pat<(v8i8 (bitconvert (v1i64 DPR:$src))), (v8i8 DPR:$src)>;
7473 def : Pat<(v8i8 (bitconvert (v2f32 DPR:$src))), (v8i8 DPR:$src)>;
7474 def : Pat<(v8i8 (bitconvert (v2i32 DPR:$src))), (v8i8 DPR:$src)>;
7475 def : Pat<(v8i8 (bitconvert (v4f16 DPR:$src))), (v8i8 DPR:$src)>;
7476 def : Pat<(v8i8 (bitconvert (v4bf16 DPR:$src))), (v8i8 DPR:$src)>;
7477 def : Pat<(v8i8 (bitconvert (v4i16 DPR:$src))), (v8i8 DPR:$src)>;
7479 // 128 bit conversions
7480 def : Pat<(v2f64 (bitconvert (v4f32 QPR:$src))), (v2f64 QPR:$src)>;
7481 def : Pat<(v2f64 (bitconvert (v4i32 QPR:$src))), (v2f64 QPR:$src)>;
7482 def : Pat<(v2f64 (bitconvert (v8f16 QPR:$src))), (v2f64 QPR:$src)>;
7483 def : Pat<(v2f64 (bitconvert (v8bf16 QPR:$src))), (v2f64 QPR:$src)>;
7484 def : Pat<(v2f64 (bitconvert (v8i16 QPR:$src))), (v2f64 QPR:$src)>;
7485 def : Pat<(v2f64 (bitconvert (v16i8 QPR:$src))), (v2f64 QPR:$src)>;
7487 def : Pat<(v2i64 (bitconvert (v4f32 QPR:$src))), (v2i64 QPR:$src)>;
7488 def : Pat<(v2i64 (bitconvert (v4i32 QPR:$src))), (v2i64 QPR:$src)>;
7489 def : Pat<(v2i64 (bitconvert (v8f16 QPR:$src))), (v2i64 QPR:$src)>;
7490 def : Pat<(v2i64 (bitconvert (v8bf16 QPR:$src))), (v2i64 QPR:$src)>;
7491 def : Pat<(v2i64 (bitconvert (v8i16 QPR:$src))), (v2i64 QPR:$src)>;
7492 def : Pat<(v2i64 (bitconvert (v16i8 QPR:$src))), (v2i64 QPR:$src)>;
7494 def : Pat<(v4f32 (bitconvert (v2f64 QPR:$src))), (v4f32 QPR:$src)>;
7495 def : Pat<(v4f32 (bitconvert (v2i64 QPR:$src))), (v4f32 QPR:$src)>;
7496 def : Pat<(v4f32 (bitconvert (v8f16 QPR:$src))), (v4f32 QPR:$src)>;
7497 def : Pat<(v4f32 (bitconvert (v8bf16 QPR:$src))), (v4f32 QPR:$src)>;
7498 def : Pat<(v4f32 (bitconvert (v8i16 QPR:$src))), (v4f32 QPR:$src)>;
7499 def : Pat<(v4f32 (bitconvert (v16i8 QPR:$src))), (v4f32 QPR:$src)>;
7501 def : Pat<(v4i32 (bitconvert (v2f64 QPR:$src))), (v4i32 QPR:$src)>;
7502 def : Pat<(v4i32 (bitconvert (v2i64 QPR:$src))), (v4i32 QPR:$src)>;
7503 def : Pat<(v4i32 (bitconvert (v8f16 QPR:$src))), (v4i32 QPR:$src)>;
7504 def : Pat<(v4i32 (bitconvert (v8bf16 QPR:$src))), (v4i32 QPR:$src)>;
7505 def : Pat<(v4i32 (bitconvert (v8i16 QPR:$src))), (v4i32 QPR:$src)>;
7506 def : Pat<(v4i32 (bitconvert (v16i8 QPR:$src))), (v4i32 QPR:$src)>;
7508 def : Pat<(v8f16 (bitconvert (v2f64 QPR:$src))), (v8f16 QPR:$src)>;
7509 def : Pat<(v8f16 (bitconvert (v2i64 QPR:$src))), (v8f16 QPR:$src)>;
7510 def : Pat<(v8f16 (bitconvert (v4f32 QPR:$src))), (v8f16 QPR:$src)>;
7511 def : Pat<(v8f16 (bitconvert (v4i32 QPR:$src))), (v8f16 QPR:$src)>;
7512 def : Pat<(v8f16 (bitconvert (v16i8 QPR:$src))), (v8f16 QPR:$src)>;
7514 def : Pat<(v8bf16 (bitconvert (v2f64 QPR:$src))), (v8bf16 QPR:$src)>;
7515 def : Pat<(v8bf16 (bitconvert (v2i64 QPR:$src))), (v8bf16 QPR:$src)>;
7516 def : Pat<(v8bf16 (bitconvert (v4f32 QPR:$src))), (v8bf16 QPR:$src)>;
7517 def : Pat<(v8bf16 (bitconvert (v4i32 QPR:$src))), (v8bf16 QPR:$src)>;
7518 def : Pat<(v8bf16 (bitconvert (v16i8 QPR:$src))), (v8bf16 QPR:$src)>;
7520 def : Pat<(v8i16 (bitconvert (v2f64 QPR:$src))), (v8i16 QPR:$src)>;
7521 def : Pat<(v8i16 (bitconvert (v2i64 QPR:$src))), (v8i16 QPR:$src)>;
7522 def : Pat<(v8i16 (bitconvert (v4f32 QPR:$src))), (v8i16 QPR:$src)>;
7523 def : Pat<(v8i16 (bitconvert (v4i32 QPR:$src))), (v8i16 QPR:$src)>;
7524 def : Pat<(v8i16 (bitconvert (v16i8 QPR:$src))), (v8i16 QPR:$src)>;
7526 def : Pat<(v16i8 (bitconvert (v2f64 QPR:$src))), (v16i8 QPR:$src)>;
7527 def : Pat<(v16i8 (bitconvert (v2i64 QPR:$src))), (v16i8 QPR:$src)>;
7528 def : Pat<(v16i8 (bitconvert (v4f32 QPR:$src))), (v16i8 QPR:$src)>;
7529 def : Pat<(v16i8 (bitconvert (v4i32 QPR:$src))), (v16i8 QPR:$src)>;
7530 def : Pat<(v16i8 (bitconvert (v8f16 QPR:$src))), (v16i8 QPR:$src)>;
7531 def : Pat<(v16i8 (bitconvert (v8bf16 QPR:$src))), (v16i8 QPR:$src)>;
7532 def : Pat<(v16i8 (bitconvert (v8i16 QPR:$src))), (v16i8 QPR:$src)>;
7535 let Predicates = [IsBE,HasNEON] in {
7536 // 64 bit conversions
7537 def : Pat<(f64 (bitconvert (v2f32 DPR:$src))), (VREV64d32 DPR:$src)>;
7538 def : Pat<(f64 (bitconvert (v2i32 DPR:$src))), (VREV64d32 DPR:$src)>;
7539 def : Pat<(f64 (bitconvert (v4f16 DPR:$src))), (VREV64d16 DPR:$src)>;
7540 def : Pat<(f64 (bitconvert (v4bf16 DPR:$src))), (VREV64d16 DPR:$src)>;
7541 def : Pat<(f64 (bitconvert (v4i16 DPR:$src))), (VREV64d16 DPR:$src)>;
7542 def : Pat<(f64 (bitconvert (v8i8 DPR:$src))), (VREV64d8 DPR:$src)>;
7544 def : Pat<(v1i64 (bitconvert (v2f32 DPR:$src))), (VREV64d32 DPR:$src)>;
7545 def : Pat<(v1i64 (bitconvert (v2i32 DPR:$src))), (VREV64d32 DPR:$src)>;
7546 def : Pat<(v1i64 (bitconvert (v4f16 DPR:$src))), (VREV64d16 DPR:$src)>;
7547 def : Pat<(v1i64 (bitconvert (v4bf16 DPR:$src))), (VREV64d16 DPR:$src)>;
7548 def : Pat<(v1i64 (bitconvert (v4i16 DPR:$src))), (VREV64d16 DPR:$src)>;
7549 def : Pat<(v1i64 (bitconvert (v8i8 DPR:$src))), (VREV64d8 DPR:$src)>;
7551 def : Pat<(v2f32 (bitconvert (f64 DPR:$src))), (VREV64d32 DPR:$src)>;
7552 def : Pat<(v2f32 (bitconvert (v1i64 DPR:$src))), (VREV64d32 DPR:$src)>;
7553 def : Pat<(v2f32 (bitconvert (v4f16 DPR:$src))), (VREV32d16 DPR:$src)>;
7554 def : Pat<(v2f32 (bitconvert (v4bf16 DPR:$src))), (VREV32d16 DPR:$src)>;
7555 def : Pat<(v2f32 (bitconvert (v4i16 DPR:$src))), (VREV32d16 DPR:$src)>;
7556 def : Pat<(v2f32 (bitconvert (v8i8 DPR:$src))), (VREV32d8 DPR:$src)>;
7558 def : Pat<(v2i32 (bitconvert (f64 DPR:$src))), (VREV64d32 DPR:$src)>;
7559 def : Pat<(v2i32 (bitconvert (v1i64 DPR:$src))), (VREV64d32 DPR:$src)>;
7560 def : Pat<(v2i32 (bitconvert (v4f16 DPR:$src))), (VREV32d16 DPR:$src)>;
7561 def : Pat<(v2i32 (bitconvert (v4bf16 DPR:$src))), (VREV32d16 DPR:$src)>;
7562 def : Pat<(v2i32 (bitconvert (v4i16 DPR:$src))), (VREV32d16 DPR:$src)>;
7563 def : Pat<(v2i32 (bitconvert (v8i8 DPR:$src))), (VREV32d8 DPR:$src)>;
7565 def : Pat<(v4f16 (bitconvert (f64 DPR:$src))), (VREV64d16 DPR:$src)>;
7566 def : Pat<(v4f16 (bitconvert (v1i64 DPR:$src))), (VREV64d16 DPR:$src)>;
7567 def : Pat<(v4f16 (bitconvert (v2f32 DPR:$src))), (VREV32d16 DPR:$src)>;
7568 def : Pat<(v4f16 (bitconvert (v2i32 DPR:$src))), (VREV32d16 DPR:$src)>;
7569 def : Pat<(v4f16 (bitconvert (v8i8 DPR:$src))), (VREV16d8 DPR:$src)>;
7571 def : Pat<(v4bf16 (bitconvert (f64 DPR:$src))), (VREV64d16 DPR:$src)>;
7572 def : Pat<(v4bf16 (bitconvert (v1i64 DPR:$src))), (VREV64d16 DPR:$src)>;
7573 def : Pat<(v4bf16 (bitconvert (v2f32 DPR:$src))), (VREV32d16 DPR:$src)>;
7574 def : Pat<(v4bf16 (bitconvert (v2i32 DPR:$src))), (VREV32d16 DPR:$src)>;
7575 def : Pat<(v4bf16 (bitconvert (v8i8 DPR:$src))), (VREV16d8 DPR:$src)>;
7577 def : Pat<(v4i16 (bitconvert (f64 DPR:$src))), (VREV64d16 DPR:$src)>;
7578 def : Pat<(v4i16 (bitconvert (v1i64 DPR:$src))), (VREV64d16 DPR:$src)>;
7579 def : Pat<(v4i16 (bitconvert (v2f32 DPR:$src))), (VREV32d16 DPR:$src)>;
7580 def : Pat<(v4i16 (bitconvert (v2i32 DPR:$src))), (VREV32d16 DPR:$src)>;
7581 def : Pat<(v4i16 (bitconvert (v8i8 DPR:$src))), (VREV16d8 DPR:$src)>;
7583 def : Pat<(v8i8 (bitconvert (f64 DPR:$src))), (VREV64d8 DPR:$src)>;
7584 def : Pat<(v8i8 (bitconvert (v1i64 DPR:$src))), (VREV64d8 DPR:$src)>;
7585 def : Pat<(v8i8 (bitconvert (v2f32 DPR:$src))), (VREV32d8 DPR:$src)>;
7586 def : Pat<(v8i8 (bitconvert (v2i32 DPR:$src))), (VREV32d8 DPR:$src)>;
7587 def : Pat<(v8i8 (bitconvert (v4f16 DPR:$src))), (VREV16d8 DPR:$src)>;
7588 def : Pat<(v8i8 (bitconvert (v4bf16 DPR:$src))), (VREV16d8 DPR:$src)>;
7589 def : Pat<(v8i8 (bitconvert (v4i16 DPR:$src))), (VREV16d8 DPR:$src)>;
7591 // 128 bit conversions
7592 def : Pat<(v2f64 (bitconvert (v4f32 QPR:$src))), (VREV64q32 QPR:$src)>;
7593 def : Pat<(v2f64 (bitconvert (v4i32 QPR:$src))), (VREV64q32 QPR:$src)>;
7594 def : Pat<(v2f64 (bitconvert (v8f16 QPR:$src))), (VREV64q16 QPR:$src)>;
7595 def : Pat<(v2f64 (bitconvert (v8bf16 QPR:$src))), (VREV64q16 QPR:$src)>;
7596 def : Pat<(v2f64 (bitconvert (v8i16 QPR:$src))), (VREV64q16 QPR:$src)>;
7597 def : Pat<(v2f64 (bitconvert (v16i8 QPR:$src))), (VREV64q8 QPR:$src)>;
7599 def : Pat<(v2i64 (bitconvert (v4f32 QPR:$src))), (VREV64q32 QPR:$src)>;
7600 def : Pat<(v2i64 (bitconvert (v4i32 QPR:$src))), (VREV64q32 QPR:$src)>;
7601 def : Pat<(v2i64 (bitconvert (v8f16 QPR:$src))), (VREV64q16 QPR:$src)>;
7602 def : Pat<(v2i64 (bitconvert (v8bf16 QPR:$src))), (VREV64q16 QPR:$src)>;
7603 def : Pat<(v2i64 (bitconvert (v8i16 QPR:$src))), (VREV64q16 QPR:$src)>;
7604 def : Pat<(v2i64 (bitconvert (v16i8 QPR:$src))), (VREV64q8 QPR:$src)>;
7606 def : Pat<(v4f32 (bitconvert (v2f64 QPR:$src))), (VREV64q32 QPR:$src)>;
7607 def : Pat<(v4f32 (bitconvert (v2i64 QPR:$src))), (VREV64q32 QPR:$src)>;
7608 def : Pat<(v4f32 (bitconvert (v8f16 QPR:$src))), (VREV32q16 QPR:$src)>;
7609 def : Pat<(v4f32 (bitconvert (v8bf16 QPR:$src))), (VREV32q16 QPR:$src)>;
7610 def : Pat<(v4f32 (bitconvert (v8i16 QPR:$src))), (VREV32q16 QPR:$src)>;
7611 def : Pat<(v4f32 (bitconvert (v16i8 QPR:$src))), (VREV32q8 QPR:$src)>;
7613 def : Pat<(v4i32 (bitconvert (v2f64 QPR:$src))), (VREV64q32 QPR:$src)>;
7614 def : Pat<(v4i32 (bitconvert (v2i64 QPR:$src))), (VREV64q32 QPR:$src)>;
7615 def : Pat<(v4i32 (bitconvert (v8f16 QPR:$src))), (VREV32q16 QPR:$src)>;
7616 def : Pat<(v4i32 (bitconvert (v8bf16 QPR:$src))), (VREV32q16 QPR:$src)>;
7617 def : Pat<(v4i32 (bitconvert (v8i16 QPR:$src))), (VREV32q16 QPR:$src)>;
7618 def : Pat<(v4i32 (bitconvert (v16i8 QPR:$src))), (VREV32q8 QPR:$src)>;
7620 def : Pat<(v8f16 (bitconvert (v2f64 QPR:$src))), (VREV64q16 QPR:$src)>;
7621 def : Pat<(v8f16 (bitconvert (v2i64 QPR:$src))), (VREV64q16 QPR:$src)>;
7622 def : Pat<(v8f16 (bitconvert (v4f32 QPR:$src))), (VREV32q16 QPR:$src)>;
7623 def : Pat<(v8f16 (bitconvert (v4i32 QPR:$src))), (VREV32q16 QPR:$src)>;
7624 def : Pat<(v8f16 (bitconvert (v16i8 QPR:$src))), (VREV16q8 QPR:$src)>;
7626 def : Pat<(v8bf16 (bitconvert (v2f64 QPR:$src))), (VREV64q16 QPR:$src)>;
7627 def : Pat<(v8bf16 (bitconvert (v2i64 QPR:$src))), (VREV64q16 QPR:$src)>;
7628 def : Pat<(v8bf16 (bitconvert (v4f32 QPR:$src))), (VREV32q16 QPR:$src)>;
7629 def : Pat<(v8bf16 (bitconvert (v4i32 QPR:$src))), (VREV32q16 QPR:$src)>;
7630 def : Pat<(v8bf16 (bitconvert (v16i8 QPR:$src))), (VREV16q8 QPR:$src)>;
7632 def : Pat<(v8i16 (bitconvert (v2f64 QPR:$src))), (VREV64q16 QPR:$src)>;
7633 def : Pat<(v8i16 (bitconvert (v2i64 QPR:$src))), (VREV64q16 QPR:$src)>;
7634 def : Pat<(v8i16 (bitconvert (v4f32 QPR:$src))), (VREV32q16 QPR:$src)>;
7635 def : Pat<(v8i16 (bitconvert (v4i32 QPR:$src))), (VREV32q16 QPR:$src)>;
7636 def : Pat<(v8i16 (bitconvert (v16i8 QPR:$src))), (VREV16q8 QPR:$src)>;
7638 def : Pat<(v16i8 (bitconvert (v2f64 QPR:$src))), (VREV64q8 QPR:$src)>;
7639 def : Pat<(v16i8 (bitconvert (v2i64 QPR:$src))), (VREV64q8 QPR:$src)>;
7640 def : Pat<(v16i8 (bitconvert (v4f32 QPR:$src))), (VREV32q8 QPR:$src)>;
7641 def : Pat<(v16i8 (bitconvert (v4i32 QPR:$src))), (VREV32q8 QPR:$src)>;
7642 def : Pat<(v16i8 (bitconvert (v8f16 QPR:$src))), (VREV16q8 QPR:$src)>;
7643 def : Pat<(v16i8 (bitconvert (v8bf16 QPR:$src))), (VREV16q8 QPR:$src)>;
7644 def : Pat<(v16i8 (bitconvert (v8i16 QPR:$src))), (VREV16q8 QPR:$src)>;
7647 let Predicates = [HasNEON] in {
7648 // Here we match the specific SDNode type 'ARMVectorRegCastImpl'
7649 // rather than the more general 'ARMVectorRegCast' which would also
7650 // match some bitconverts. If we use the latter in cases where the
7651 // input and output types are the same, the bitconvert gets elided
7652 // and we end up generating a nonsense match of nothing.
7654 foreach VT = [ v16i8, v8i16, v8f16, v8bf16, v4i32, v4f32, v2i64, v2f64 ] in
7655 foreach VT2 = [ v16i8, v8i16, v8f16, v8bf16, v4i32, v4f32, v2i64, v2f64 ] in
7656 def : Pat<(VT (ARMVectorRegCastImpl (VT2 QPR:$src))), (VT QPR:$src)>;
7658 foreach VT = [ v8i8, v4i16, v4f16, v4bf16, v2i32, v2f32, v1i64, f64 ] in
7659 foreach VT2 = [ v8i8, v4i16, v4f16, v4bf16, v2i32, v2f32, v1i64, f64 ] in
7660 def : Pat<(VT (ARMVectorRegCastImpl (VT2 DPR:$src))), (VT DPR:$src)>;
7663 // Use VLD1/VST1 + VREV for non-word-aligned v2f64 load/store on Big Endian
7664 let Predicates = [IsBE,HasNEON] in {
7665 def : Pat<(v2f64 (byte_alignedload addrmode6:$addr)),
7666 (VREV64q8 (VLD1q8 addrmode6:$addr))>;
7667 def : Pat<(byte_alignedstore (v2f64 QPR:$value), addrmode6:$addr),
7668 (VST1q8 addrmode6:$addr, (VREV64q8 QPR:$value))>;
7669 def : Pat<(v2f64 (hword_alignedload addrmode6:$addr)),
7670 (VREV64q16 (VLD1q16 addrmode6:$addr))>;
7671 def : Pat<(hword_alignedstore (v2f64 QPR:$value), addrmode6:$addr),
7672 (VST1q16 addrmode6:$addr, (VREV64q16 QPR:$value))>;
7675 // Fold extracting an element out of a v2i32 into a vfp register.
7676 def : Pat<(f32 (bitconvert (i32 (extractelt (v2i32 DPR:$src), imm:$lane)))),
7677 (f32 (EXTRACT_SUBREG DPR:$src, (SSubReg_f32_reg imm:$lane)))>,
7678 Requires<[HasNEON]>;
7680 // Vector lengthening move with load, matching extending loads.
7682 // extload, zextload and sextload for a standard lengthening load. Example:
7683 // Lengthen_Single<"8", "i16", "8"> =
7684 // Pat<(v8i16 (extloadvi8 addrmode6:$addr))
7685 // (VMOVLuv8i16 (VLD1d8 addrmode6:$addr,
7686 // (f64 (IMPLICIT_DEF)), (i32 0)))>;
7687 multiclass Lengthen_Single<string DestLanes, string DestTy, string SrcTy> {
7688 let AddedComplexity = 10 in {
7689 def _Any : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
7690 (!cast<PatFrag>("extloadvi" # SrcTy) addrmode6:$addr)),
7691 (!cast<Instruction>("VMOVLuv" # DestLanes # DestTy)
7692 (!cast<Instruction>("VLD1d" # SrcTy) addrmode6:$addr))>,
7693 Requires<[HasNEON]>;
7695 def _Z : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
7696 (!cast<PatFrag>("zextloadvi" # SrcTy) addrmode6:$addr)),
7697 (!cast<Instruction>("VMOVLuv" # DestLanes # DestTy)
7698 (!cast<Instruction>("VLD1d" # SrcTy) addrmode6:$addr))>,
7699 Requires<[HasNEON]>;
7701 def _S : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
7702 (!cast<PatFrag>("sextloadvi" # SrcTy) addrmode6:$addr)),
7703 (!cast<Instruction>("VMOVLsv" # DestLanes # DestTy)
7704 (!cast<Instruction>("VLD1d" # SrcTy) addrmode6:$addr))>,
7705 Requires<[HasNEON]>;
7709 // extload, zextload and sextload for a lengthening load which only uses
7710 // half the lanes available. Example:
7711 // Lengthen_HalfSingle<"4", "i16", "8", "i16", "i8"> =
7712 // Pat<(v4i16 (extloadvi8 addrmode6oneL32:$addr)),
7713 // (EXTRACT_SUBREG (VMOVLuv8i16 (VLD1LNd32 addrmode6oneL32:$addr,
7714 // (f64 (IMPLICIT_DEF)), (i32 0))),
7716 multiclass Lengthen_HalfSingle<string DestLanes, string DestTy, string SrcTy,
7717 string InsnLanes, string InsnTy> {
7718 def _Any : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
7719 (!cast<PatFrag>("extloadv" # SrcTy) addrmode6oneL32:$addr)),
7720 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # InsnLanes # InsnTy)
7721 (VLD1LNd32 addrmode6oneL32:$addr, (f64 (IMPLICIT_DEF)), (i32 0))),
7723 Requires<[HasNEON]>;
7724 def _Z : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
7725 (!cast<PatFrag>("zextloadv" # SrcTy) addrmode6oneL32:$addr)),
7726 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # InsnLanes # InsnTy)
7727 (VLD1LNd32 addrmode6oneL32:$addr, (f64 (IMPLICIT_DEF)), (i32 0))),
7729 Requires<[HasNEON]>;
7730 def _S : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
7731 (!cast<PatFrag>("sextloadv" # SrcTy) addrmode6oneL32:$addr)),
7732 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLsv" # InsnLanes # InsnTy)
7733 (VLD1LNd32 addrmode6oneL32:$addr, (f64 (IMPLICIT_DEF)), (i32 0))),
7735 Requires<[HasNEON]>;
7738 // The following class definition is basically a copy of the
7739 // Lengthen_HalfSingle definition above, however with an additional parameter
7740 // "RevLanes" to select the correct VREV32dXX instruction. This is to convert
7741 // data loaded by VLD1LN into proper vector format in big endian mode.
7742 multiclass Lengthen_HalfSingle_Big_Endian<string DestLanes, string DestTy, string SrcTy,
7743 string InsnLanes, string InsnTy, string RevLanes> {
7744 def _Any : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
7745 (!cast<PatFrag>("extloadv" # SrcTy) addrmode6oneL32:$addr)),
7746 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # InsnLanes # InsnTy)
7747 (!cast<Instruction>("VREV32d" # RevLanes)
7748 (VLD1LNd32 addrmode6oneL32:$addr, (f64 (IMPLICIT_DEF)), (i32 0)))),
7750 Requires<[HasNEON]>;
7751 def _Z : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
7752 (!cast<PatFrag>("zextloadv" # SrcTy) addrmode6oneL32:$addr)),
7753 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # InsnLanes # InsnTy)
7754 (!cast<Instruction>("VREV32d" # RevLanes)
7755 (VLD1LNd32 addrmode6oneL32:$addr, (f64 (IMPLICIT_DEF)), (i32 0)))),
7757 Requires<[HasNEON]>;
7758 def _S : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
7759 (!cast<PatFrag>("sextloadv" # SrcTy) addrmode6oneL32:$addr)),
7760 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLsv" # InsnLanes # InsnTy)
7761 (!cast<Instruction>("VREV32d" # RevLanes)
7762 (VLD1LNd32 addrmode6oneL32:$addr, (f64 (IMPLICIT_DEF)), (i32 0)))),
7764 Requires<[HasNEON]>;
7767 // extload, zextload and sextload for a lengthening load followed by another
7768 // lengthening load, to quadruple the initial length.
7770 // Lengthen_Double<"4", "i32", "i8", "8", "i16", "4", "i32"> =
7771 // Pat<(v4i32 (extloadvi8 addrmode6oneL32:$addr))
7772 // (EXTRACT_SUBREG (VMOVLuv4i32
7773 // (EXTRACT_SUBREG (VMOVLuv8i16 (VLD1LNd32 addrmode6oneL32:$addr,
7774 // (f64 (IMPLICIT_DEF)),
7778 multiclass Lengthen_Double<string DestLanes, string DestTy, string SrcTy,
7779 string Insn1Lanes, string Insn1Ty, string Insn2Lanes,
7781 def _Any : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
7782 (!cast<PatFrag>("extloadv" # SrcTy) addrmode6oneL32:$addr)),
7783 (!cast<Instruction>("VMOVLuv" # Insn2Lanes # Insn2Ty)
7784 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # Insn1Lanes # Insn1Ty)
7785 (VLD1LNd32 addrmode6oneL32:$addr, (f64 (IMPLICIT_DEF)), (i32 0))),
7787 Requires<[HasNEON]>;
7788 def _Z : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
7789 (!cast<PatFrag>("zextloadv" # SrcTy) addrmode6oneL32:$addr)),
7790 (!cast<Instruction>("VMOVLuv" # Insn2Lanes # Insn2Ty)
7791 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # Insn1Lanes # Insn1Ty)
7792 (VLD1LNd32 addrmode6oneL32:$addr, (f64 (IMPLICIT_DEF)), (i32 0))),
7794 Requires<[HasNEON]>;
7795 def _S : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
7796 (!cast<PatFrag>("sextloadv" # SrcTy) addrmode6oneL32:$addr)),
7797 (!cast<Instruction>("VMOVLsv" # Insn2Lanes # Insn2Ty)
7798 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLsv" # Insn1Lanes # Insn1Ty)
7799 (VLD1LNd32 addrmode6oneL32:$addr, (f64 (IMPLICIT_DEF)), (i32 0))),
7801 Requires<[HasNEON]>;
7804 // The following class definition is basically a copy of the
7805 // Lengthen_Double definition above, however with an additional parameter
7806 // "RevLanes" to select the correct VREV32dXX instruction. This is to convert
7807 // data loaded by VLD1LN into proper vector format in big endian mode.
7808 multiclass Lengthen_Double_Big_Endian<string DestLanes, string DestTy, string SrcTy,
7809 string Insn1Lanes, string Insn1Ty, string Insn2Lanes,
7810 string Insn2Ty, string RevLanes> {
7811 def _Any : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
7812 (!cast<PatFrag>("extloadv" # SrcTy) addrmode6oneL32:$addr)),
7813 (!cast<Instruction>("VMOVLuv" # Insn2Lanes # Insn2Ty)
7814 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # Insn1Lanes # Insn1Ty)
7815 (!cast<Instruction>("VREV32d" # RevLanes)
7816 (VLD1LNd32 addrmode6oneL32:$addr, (f64 (IMPLICIT_DEF)), (i32 0)))),
7818 Requires<[HasNEON]>;
7819 def _Z : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
7820 (!cast<PatFrag>("zextloadv" # SrcTy) addrmode6oneL32:$addr)),
7821 (!cast<Instruction>("VMOVLuv" # Insn2Lanes # Insn2Ty)
7822 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # Insn1Lanes # Insn1Ty)
7823 (!cast<Instruction>("VREV32d" # RevLanes)
7824 (VLD1LNd32 addrmode6oneL32:$addr, (f64 (IMPLICIT_DEF)), (i32 0)))),
7826 Requires<[HasNEON]>;
7827 def _S : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
7828 (!cast<PatFrag>("sextloadv" # SrcTy) addrmode6oneL32:$addr)),
7829 (!cast<Instruction>("VMOVLsv" # Insn2Lanes # Insn2Ty)
7830 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLsv" # Insn1Lanes # Insn1Ty)
7831 (!cast<Instruction>("VREV32d" # RevLanes)
7832 (VLD1LNd32 addrmode6oneL32:$addr, (f64 (IMPLICIT_DEF)), (i32 0)))),
7834 Requires<[HasNEON]>;
7837 // extload, zextload and sextload for a lengthening load followed by another
7838 // lengthening load, to quadruple the initial length, but which ends up only
7839 // requiring half the available lanes (a 64-bit outcome instead of a 128-bit).
7841 // Lengthen_HalfDouble<"2", "i32", "i8", "8", "i16", "4", "i32"> =
7842 // Pat<(v2i32 (extloadvi8 addrmode6:$addr))
7843 // (EXTRACT_SUBREG (VMOVLuv4i32
7844 // (EXTRACT_SUBREG (VMOVLuv8i16 (VLD1LNd16 addrmode6:$addr,
7845 // (f64 (IMPLICIT_DEF)), (i32 0))),
7848 multiclass Lengthen_HalfDouble<string DestLanes, string DestTy, string SrcTy,
7849 string Insn1Lanes, string Insn1Ty, string Insn2Lanes,
7851 def _Any : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
7852 (!cast<PatFrag>("extloadv" # SrcTy) addrmode6:$addr)),
7853 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # Insn2Lanes # Insn2Ty)
7854 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # Insn1Lanes # Insn1Ty)
7855 (VLD1LNd16 addrmode6:$addr, (f64 (IMPLICIT_DEF)), (i32 0))),
7858 Requires<[HasNEON]>;
7859 def _Z : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
7860 (!cast<PatFrag>("zextloadv" # SrcTy) addrmode6:$addr)),
7861 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # Insn2Lanes # Insn2Ty)
7862 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # Insn1Lanes # Insn1Ty)
7863 (VLD1LNd16 addrmode6:$addr, (f64 (IMPLICIT_DEF)), (i32 0))),
7866 Requires<[HasNEON]>;
7867 def _S : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
7868 (!cast<PatFrag>("sextloadv" # SrcTy) addrmode6:$addr)),
7869 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLsv" # Insn2Lanes # Insn2Ty)
7870 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLsv" # Insn1Lanes # Insn1Ty)
7871 (VLD1LNd16 addrmode6:$addr, (f64 (IMPLICIT_DEF)), (i32 0))),
7874 Requires<[HasNEON]>;
7877 // The following class definition is basically a copy of the
7878 // Lengthen_HalfDouble definition above, however with an additional VREV16d8
7879 // instruction to convert data loaded by VLD1LN into proper vector format
7880 // in big endian mode.
7881 multiclass Lengthen_HalfDouble_Big_Endian<string DestLanes, string DestTy, string SrcTy,
7882 string Insn1Lanes, string Insn1Ty, string Insn2Lanes,
7884 def _Any : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
7885 (!cast<PatFrag>("extloadv" # SrcTy) addrmode6:$addr)),
7886 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # Insn2Lanes # Insn2Ty)
7887 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # Insn1Lanes # Insn1Ty)
7888 (!cast<Instruction>("VREV16d8")
7889 (VLD1LNd16 addrmode6:$addr, (f64 (IMPLICIT_DEF)), (i32 0)))),
7892 Requires<[HasNEON]>;
7893 def _Z : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
7894 (!cast<PatFrag>("zextloadv" # SrcTy) addrmode6:$addr)),
7895 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # Insn2Lanes # Insn2Ty)
7896 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLuv" # Insn1Lanes # Insn1Ty)
7897 (!cast<Instruction>("VREV16d8")
7898 (VLD1LNd16 addrmode6:$addr, (f64 (IMPLICIT_DEF)), (i32 0)))),
7901 Requires<[HasNEON]>;
7902 def _S : Pat<(!cast<ValueType>("v" # DestLanes # DestTy)
7903 (!cast<PatFrag>("sextloadv" # SrcTy) addrmode6:$addr)),
7904 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLsv" # Insn2Lanes # Insn2Ty)
7905 (EXTRACT_SUBREG (!cast<Instruction>("VMOVLsv" # Insn1Lanes # Insn1Ty)
7906 (!cast<Instruction>("VREV16d8")
7907 (VLD1LNd16 addrmode6:$addr, (f64 (IMPLICIT_DEF)), (i32 0)))),
7910 Requires<[HasNEON]>;
7913 defm : Lengthen_Single<"8", "i16", "8">; // v8i8 -> v8i16
7914 defm : Lengthen_Single<"4", "i32", "16">; // v4i16 -> v4i32
7915 defm : Lengthen_Single<"2", "i64", "32">; // v2i32 -> v2i64
7917 let Predicates = [HasNEON,IsLE] in {
7918 defm : Lengthen_HalfSingle<"4", "i16", "i8", "8", "i16">; // v4i8 -> v4i16
7919 defm : Lengthen_HalfSingle<"2", "i32", "i16", "4", "i32">; // v2i16 -> v2i32
7921 // Double lengthening - v4i8 -> v4i16 -> v4i32
7922 defm : Lengthen_Double<"4", "i32", "i8", "8", "i16", "4", "i32">;
7923 // v2i8 -> v2i16 -> v2i32
7924 defm : Lengthen_HalfDouble<"2", "i32", "i8", "8", "i16", "4", "i32">;
7925 // v2i16 -> v2i32 -> v2i64
7926 defm : Lengthen_Double<"2", "i64", "i16", "4", "i32", "2", "i64">;
7929 let Predicates = [HasNEON,IsBE] in {
7930 defm : Lengthen_HalfSingle_Big_Endian<"4", "i16", "i8", "8", "i16", "8">; // v4i8 -> v4i16
7931 defm : Lengthen_HalfSingle_Big_Endian<"2", "i32", "i16", "4", "i32", "16">; // v2i16 -> v2i32
7933 // Double lengthening - v4i8 -> v4i16 -> v4i32
7934 defm : Lengthen_Double_Big_Endian<"4", "i32", "i8", "8", "i16", "4", "i32", "8">;
7935 // v2i8 -> v2i16 -> v2i32
7936 defm : Lengthen_HalfDouble_Big_Endian<"2", "i32", "i8", "8", "i16", "4", "i32">;
7937 // v2i16 -> v2i32 -> v2i64
7938 defm : Lengthen_Double_Big_Endian<"2", "i64", "i16", "4", "i32", "2", "i64", "16">;
7941 // Triple lengthening - v2i8 -> v2i16 -> v2i32 -> v2i64
7942 let Predicates = [HasNEON,IsLE] in {
7943 def : Pat<(v2i64 (extloadvi8 addrmode6:$addr)),
7944 (VMOVLuv2i64 (EXTRACT_SUBREG (VMOVLuv4i32 (EXTRACT_SUBREG (VMOVLuv8i16
7945 (VLD1LNd16 addrmode6:$addr,
7946 (f64 (IMPLICIT_DEF)), (i32 0))), dsub_0)), dsub_0))>;
7947 def : Pat<(v2i64 (zextloadvi8 addrmode6:$addr)),
7948 (VMOVLuv2i64 (EXTRACT_SUBREG (VMOVLuv4i32 (EXTRACT_SUBREG (VMOVLuv8i16
7949 (VLD1LNd16 addrmode6:$addr,
7950 (f64 (IMPLICIT_DEF)), (i32 0))), dsub_0)), dsub_0))>;
7951 def : Pat<(v2i64 (sextloadvi8 addrmode6:$addr)),
7952 (VMOVLsv2i64 (EXTRACT_SUBREG (VMOVLsv4i32 (EXTRACT_SUBREG (VMOVLsv8i16
7953 (VLD1LNd16 addrmode6:$addr,
7954 (f64 (IMPLICIT_DEF)), (i32 0))), dsub_0)), dsub_0))>;
7956 // The following patterns are basically a copy of the patterns above,
7957 // however with an additional VREV16d instruction to convert data
7958 // loaded by VLD1LN into proper vector format in big endian mode.
7959 let Predicates = [HasNEON,IsBE] in {
7960 def : Pat<(v2i64 (extloadvi8 addrmode6:$addr)),
7961 (VMOVLuv2i64 (EXTRACT_SUBREG (VMOVLuv4i32 (EXTRACT_SUBREG (VMOVLuv8i16
7962 (!cast<Instruction>("VREV16d8")
7963 (VLD1LNd16 addrmode6:$addr,
7964 (f64 (IMPLICIT_DEF)), (i32 0)))), dsub_0)), dsub_0))>;
7965 def : Pat<(v2i64 (zextloadvi8 addrmode6:$addr)),
7966 (VMOVLuv2i64 (EXTRACT_SUBREG (VMOVLuv4i32 (EXTRACT_SUBREG (VMOVLuv8i16
7967 (!cast<Instruction>("VREV16d8")
7968 (VLD1LNd16 addrmode6:$addr,
7969 (f64 (IMPLICIT_DEF)), (i32 0)))), dsub_0)), dsub_0))>;
7970 def : Pat<(v2i64 (sextloadvi8 addrmode6:$addr)),
7971 (VMOVLsv2i64 (EXTRACT_SUBREG (VMOVLsv4i32 (EXTRACT_SUBREG (VMOVLsv8i16
7972 (!cast<Instruction>("VREV16d8")
7973 (VLD1LNd16 addrmode6:$addr,
7974 (f64 (IMPLICIT_DEF)), (i32 0)))), dsub_0)), dsub_0))>;
7977 let Predicates = [HasNEON] in {
7978 def : Pat<(v2i64 (concat_vectors DPR:$Dn, DPR:$Dm)),
7979 (REG_SEQUENCE QPR, DPR:$Dn, dsub_0, DPR:$Dm, dsub_1)>;
7980 def : Pat<(v4i32 (concat_vectors DPR:$Dn, DPR:$Dm)),
7981 (REG_SEQUENCE QPR, DPR:$Dn, dsub_0, DPR:$Dm, dsub_1)>;
7982 def : Pat<(v8i16 (concat_vectors DPR:$Dn, DPR:$Dm)),
7983 (REG_SEQUENCE QPR, DPR:$Dn, dsub_0, DPR:$Dm, dsub_1)>;
7984 def : Pat<(v16i8 (concat_vectors DPR:$Dn, DPR:$Dm)),
7985 (REG_SEQUENCE QPR, DPR:$Dn, dsub_0, DPR:$Dm, dsub_1)>;
7986 def : Pat<(v4f32 (concat_vectors DPR:$Dn, DPR:$Dm)),
7987 (REG_SEQUENCE QPR, DPR:$Dn, dsub_0, DPR:$Dm, dsub_1)>;
7988 def : Pat<(v8f16 (concat_vectors DPR:$Dn, DPR:$Dm)),
7989 (REG_SEQUENCE QPR, DPR:$Dn, dsub_0, DPR:$Dm, dsub_1)>;
7990 def : Pat<(v8bf16 (concat_vectors DPR:$Dn, DPR:$Dm)),
7991 (REG_SEQUENCE QPR, DPR:$Dn, dsub_0, DPR:$Dm, dsub_1)>;
7994 //===----------------------------------------------------------------------===//
7995 // Assembler aliases
7998 def : VFP2InstAlias<"fmdhr${p} $Dd, $Rn",
7999 (VSETLNi32 DPR:$Dd, GPR:$Rn, 1, pred:$p)>;
8000 def : VFP2InstAlias<"fmdlr${p} $Dd, $Rn",
8001 (VSETLNi32 DPR:$Dd, GPR:$Rn, 0, pred:$p)>;
8003 // VAND/VBIC/VEOR/VORR accept but do not require a type suffix.
8004 defm : NEONDTAnyInstAlias<"vand${p}", "$Vd, $Vn, $Vm",
8005 (VANDd DPR:$Vd, DPR:$Vn, DPR:$Vm, pred:$p)>;
8006 defm : NEONDTAnyInstAlias<"vand${p}", "$Vd, $Vn, $Vm",
8007 (VANDq QPR:$Vd, QPR:$Vn, QPR:$Vm, pred:$p)>;
8008 defm : NEONDTAnyInstAlias<"vbic${p}", "$Vd, $Vn, $Vm",
8009 (VBICd DPR:$Vd, DPR:$Vn, DPR:$Vm, pred:$p)>;
8010 defm : NEONDTAnyInstAlias<"vbic${p}", "$Vd, $Vn, $Vm",
8011 (VBICq QPR:$Vd, QPR:$Vn, QPR:$Vm, pred:$p)>;
8012 defm : NEONDTAnyInstAlias<"veor${p}", "$Vd, $Vn, $Vm",
8013 (VEORd DPR:$Vd, DPR:$Vn, DPR:$Vm, pred:$p)>;
8014 defm : NEONDTAnyInstAlias<"veor${p}", "$Vd, $Vn, $Vm",
8015 (VEORq QPR:$Vd, QPR:$Vn, QPR:$Vm, pred:$p)>;
8016 defm : NEONDTAnyInstAlias<"vorr${p}", "$Vd, $Vn, $Vm",
8017 (VORRd DPR:$Vd, DPR:$Vn, DPR:$Vm, pred:$p)>;
8018 defm : NEONDTAnyInstAlias<"vorr${p}", "$Vd, $Vn, $Vm",
8019 (VORRq QPR:$Vd, QPR:$Vn, QPR:$Vm, pred:$p)>;
8020 // ... two-operand aliases
8021 defm : NEONDTAnyInstAlias<"vand${p}", "$Vdn, $Vm",
8022 (VANDd DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
8023 defm : NEONDTAnyInstAlias<"vand${p}", "$Vdn, $Vm",
8024 (VANDq QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
8025 defm : NEONDTAnyInstAlias<"veor${p}", "$Vdn, $Vm",
8026 (VEORd DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
8027 defm : NEONDTAnyInstAlias<"veor${p}", "$Vdn, $Vm",
8028 (VEORq QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
8029 defm : NEONDTAnyInstAlias<"vorr${p}", "$Vdn, $Vm",
8030 (VORRd DPR:$Vdn, DPR:$Vdn, DPR:$Vm, pred:$p)>;
8031 defm : NEONDTAnyInstAlias<"vorr${p}", "$Vdn, $Vm",
8032 (VORRq QPR:$Vdn, QPR:$Vdn, QPR:$Vm, pred:$p)>;
8034 def : NEONInstAlias<"vand${p}.i16 $Vd, $imm",
8035 (VBICiv4i16 DPR:$Vd, nImmSplatNotI16:$imm, pred:$p)>;
8036 def : NEONInstAlias<"vand${p}.i32 $Vd, $imm",
8037 (VBICiv2i32 DPR:$Vd, nImmSplatNotI32:$imm, pred:$p)>;
8038 def : NEONInstAlias<"vand${p}.i16 $Vd, $imm",
8039 (VBICiv8i16 QPR:$Vd, nImmSplatNotI16:$imm, pred:$p)>;
8040 def : NEONInstAlias<"vand${p}.i32 $Vd, $imm",
8041 (VBICiv4i32 QPR:$Vd, nImmSplatNotI32:$imm, pred:$p)>;
8044 // VLD1 single-lane pseudo-instructions. These need special handling for
8045 // the lane index that an InstAlias can't handle, so we use these instead.
8046 def VLD1LNdAsm_8 : NEONDataTypeAsmPseudoInst<"vld1${p}", ".8", "$list, $addr",
8047 (ins VecListOneDByteIndexed:$list, addrmode6alignNone:$addr,
8049 def VLD1LNdAsm_16 : NEONDataTypeAsmPseudoInst<"vld1${p}", ".16", "$list, $addr",
8050 (ins VecListOneDHWordIndexed:$list, addrmode6align16:$addr,
8052 def VLD1LNdAsm_32 : NEONDataTypeAsmPseudoInst<"vld1${p}", ".32", "$list, $addr",
8053 (ins VecListOneDWordIndexed:$list, addrmode6align32:$addr,
8056 def VLD1LNdWB_fixed_Asm_8 :
8057 NEONDataTypeAsmPseudoInst<"vld1${p}", ".8", "$list, $addr!",
8058 (ins VecListOneDByteIndexed:$list, addrmode6alignNone:$addr,
8060 def VLD1LNdWB_fixed_Asm_16 :
8061 NEONDataTypeAsmPseudoInst<"vld1${p}", ".16", "$list, $addr!",
8062 (ins VecListOneDHWordIndexed:$list, addrmode6align16:$addr,
8064 def VLD1LNdWB_fixed_Asm_32 :
8065 NEONDataTypeAsmPseudoInst<"vld1${p}", ".32", "$list, $addr!",
8066 (ins VecListOneDWordIndexed:$list, addrmode6align32:$addr,
8068 def VLD1LNdWB_register_Asm_8 :
8069 NEONDataTypeAsmPseudoInst<"vld1${p}", ".8", "$list, $addr, $Rm",
8070 (ins VecListOneDByteIndexed:$list, addrmode6alignNone:$addr,
8071 rGPR:$Rm, pred:$p)>;
8072 def VLD1LNdWB_register_Asm_16 :
8073 NEONDataTypeAsmPseudoInst<"vld1${p}", ".16", "$list, $addr, $Rm",
8074 (ins VecListOneDHWordIndexed:$list, addrmode6align16:$addr,
8075 rGPR:$Rm, pred:$p)>;
8076 def VLD1LNdWB_register_Asm_32 :
8077 NEONDataTypeAsmPseudoInst<"vld1${p}", ".32", "$list, $addr, $Rm",
8078 (ins VecListOneDWordIndexed:$list, addrmode6align32:$addr,
8079 rGPR:$Rm, pred:$p)>;
8082 // VST1 single-lane pseudo-instructions. These need special handling for
8083 // the lane index that an InstAlias can't handle, so we use these instead.
8084 def VST1LNdAsm_8 : NEONDataTypeAsmPseudoInst<"vst1${p}", ".8", "$list, $addr",
8085 (ins VecListOneDByteIndexed:$list, addrmode6alignNone:$addr,
8087 def VST1LNdAsm_16 : NEONDataTypeAsmPseudoInst<"vst1${p}", ".16", "$list, $addr",
8088 (ins VecListOneDHWordIndexed:$list, addrmode6align16:$addr,
8090 def VST1LNdAsm_32 : NEONDataTypeAsmPseudoInst<"vst1${p}", ".32", "$list, $addr",
8091 (ins VecListOneDWordIndexed:$list, addrmode6align32:$addr,
8094 def VST1LNdWB_fixed_Asm_8 :
8095 NEONDataTypeAsmPseudoInst<"vst1${p}", ".8", "$list, $addr!",
8096 (ins VecListOneDByteIndexed:$list, addrmode6alignNone:$addr,
8098 def VST1LNdWB_fixed_Asm_16 :
8099 NEONDataTypeAsmPseudoInst<"vst1${p}", ".16", "$list, $addr!",
8100 (ins VecListOneDHWordIndexed:$list, addrmode6align16:$addr,
8102 def VST1LNdWB_fixed_Asm_32 :
8103 NEONDataTypeAsmPseudoInst<"vst1${p}", ".32", "$list, $addr!",
8104 (ins VecListOneDWordIndexed:$list, addrmode6align32:$addr,
8106 def VST1LNdWB_register_Asm_8 :
8107 NEONDataTypeAsmPseudoInst<"vst1${p}", ".8", "$list, $addr, $Rm",
8108 (ins VecListOneDByteIndexed:$list, addrmode6alignNone:$addr,
8109 rGPR:$Rm, pred:$p)>;
8110 def VST1LNdWB_register_Asm_16 :
8111 NEONDataTypeAsmPseudoInst<"vst1${p}", ".16", "$list, $addr, $Rm",
8112 (ins VecListOneDHWordIndexed:$list, addrmode6align16:$addr,
8113 rGPR:$Rm, pred:$p)>;
8114 def VST1LNdWB_register_Asm_32 :
8115 NEONDataTypeAsmPseudoInst<"vst1${p}", ".32", "$list, $addr, $Rm",
8116 (ins VecListOneDWordIndexed:$list, addrmode6align32:$addr,
8117 rGPR:$Rm, pred:$p)>;
8119 // VLD2 single-lane pseudo-instructions. These need special handling for
8120 // the lane index that an InstAlias can't handle, so we use these instead.
8121 def VLD2LNdAsm_8 : NEONDataTypeAsmPseudoInst<"vld2${p}", ".8", "$list, $addr",
8122 (ins VecListTwoDByteIndexed:$list, addrmode6align16:$addr,
8124 def VLD2LNdAsm_16 : NEONDataTypeAsmPseudoInst<"vld2${p}", ".16", "$list, $addr",
8125 (ins VecListTwoDHWordIndexed:$list, addrmode6align32:$addr,
8127 def VLD2LNdAsm_32 : NEONDataTypeAsmPseudoInst<"vld2${p}", ".32", "$list, $addr",
8128 (ins VecListTwoDWordIndexed:$list, addrmode6align64:$addr, pred:$p)>;
8129 def VLD2LNqAsm_16 : NEONDataTypeAsmPseudoInst<"vld2${p}", ".16", "$list, $addr",
8130 (ins VecListTwoQHWordIndexed:$list, addrmode6align32:$addr,
8132 def VLD2LNqAsm_32 : NEONDataTypeAsmPseudoInst<"vld2${p}", ".32", "$list, $addr",
8133 (ins VecListTwoQWordIndexed:$list, addrmode6align64:$addr,
8136 def VLD2LNdWB_fixed_Asm_8 :
8137 NEONDataTypeAsmPseudoInst<"vld2${p}", ".8", "$list, $addr!",
8138 (ins VecListTwoDByteIndexed:$list, addrmode6align16:$addr,
8140 def VLD2LNdWB_fixed_Asm_16 :
8141 NEONDataTypeAsmPseudoInst<"vld2${p}", ".16", "$list, $addr!",
8142 (ins VecListTwoDHWordIndexed:$list, addrmode6align32:$addr,
8144 def VLD2LNdWB_fixed_Asm_32 :
8145 NEONDataTypeAsmPseudoInst<"vld2${p}", ".32", "$list, $addr!",
8146 (ins VecListTwoDWordIndexed:$list, addrmode6align64:$addr,
8148 def VLD2LNqWB_fixed_Asm_16 :
8149 NEONDataTypeAsmPseudoInst<"vld2${p}", ".16", "$list, $addr!",
8150 (ins VecListTwoQHWordIndexed:$list, addrmode6align32:$addr,
8152 def VLD2LNqWB_fixed_Asm_32 :
8153 NEONDataTypeAsmPseudoInst<"vld2${p}", ".32", "$list, $addr!",
8154 (ins VecListTwoQWordIndexed:$list, addrmode6align64:$addr,
8156 def VLD2LNdWB_register_Asm_8 :
8157 NEONDataTypeAsmPseudoInst<"vld2${p}", ".8", "$list, $addr, $Rm",
8158 (ins VecListTwoDByteIndexed:$list, addrmode6align16:$addr,
8159 rGPR:$Rm, pred:$p)>;
8160 def VLD2LNdWB_register_Asm_16 :
8161 NEONDataTypeAsmPseudoInst<"vld2${p}", ".16", "$list, $addr, $Rm",
8162 (ins VecListTwoDHWordIndexed:$list, addrmode6align32:$addr,
8163 rGPR:$Rm, pred:$p)>;
8164 def VLD2LNdWB_register_Asm_32 :
8165 NEONDataTypeAsmPseudoInst<"vld2${p}", ".32", "$list, $addr, $Rm",
8166 (ins VecListTwoDWordIndexed:$list, addrmode6align64:$addr,
8167 rGPR:$Rm, pred:$p)>;
8168 def VLD2LNqWB_register_Asm_16 :
8169 NEONDataTypeAsmPseudoInst<"vld2${p}", ".16", "$list, $addr, $Rm",
8170 (ins VecListTwoQHWordIndexed:$list, addrmode6align32:$addr,
8171 rGPR:$Rm, pred:$p)>;
8172 def VLD2LNqWB_register_Asm_32 :
8173 NEONDataTypeAsmPseudoInst<"vld2${p}", ".32", "$list, $addr, $Rm",
8174 (ins VecListTwoQWordIndexed:$list, addrmode6align64:$addr,
8175 rGPR:$Rm, pred:$p)>;
8178 // VST2 single-lane pseudo-instructions. These need special handling for
8179 // the lane index that an InstAlias can't handle, so we use these instead.
8180 def VST2LNdAsm_8 : NEONDataTypeAsmPseudoInst<"vst2${p}", ".8", "$list, $addr",
8181 (ins VecListTwoDByteIndexed:$list, addrmode6align16:$addr,
8183 def VST2LNdAsm_16 : NEONDataTypeAsmPseudoInst<"vst2${p}", ".16", "$list, $addr",
8184 (ins VecListTwoDHWordIndexed:$list, addrmode6align32:$addr,
8186 def VST2LNdAsm_32 : NEONDataTypeAsmPseudoInst<"vst2${p}", ".32", "$list, $addr",
8187 (ins VecListTwoDWordIndexed:$list, addrmode6align64:$addr,
8189 def VST2LNqAsm_16 : NEONDataTypeAsmPseudoInst<"vst2${p}", ".16", "$list, $addr",
8190 (ins VecListTwoQHWordIndexed:$list, addrmode6align32:$addr,
8192 def VST2LNqAsm_32 : NEONDataTypeAsmPseudoInst<"vst2${p}", ".32", "$list, $addr",
8193 (ins VecListTwoQWordIndexed:$list, addrmode6align64:$addr,
8196 def VST2LNdWB_fixed_Asm_8 :
8197 NEONDataTypeAsmPseudoInst<"vst2${p}", ".8", "$list, $addr!",
8198 (ins VecListTwoDByteIndexed:$list, addrmode6align16:$addr,
8200 def VST2LNdWB_fixed_Asm_16 :
8201 NEONDataTypeAsmPseudoInst<"vst2${p}", ".16", "$list, $addr!",
8202 (ins VecListTwoDHWordIndexed:$list, addrmode6align32:$addr,
8204 def VST2LNdWB_fixed_Asm_32 :
8205 NEONDataTypeAsmPseudoInst<"vst2${p}", ".32", "$list, $addr!",
8206 (ins VecListTwoDWordIndexed:$list, addrmode6align64:$addr,
8208 def VST2LNqWB_fixed_Asm_16 :
8209 NEONDataTypeAsmPseudoInst<"vst2${p}", ".16", "$list, $addr!",
8210 (ins VecListTwoQHWordIndexed:$list, addrmode6align32:$addr,
8212 def VST2LNqWB_fixed_Asm_32 :
8213 NEONDataTypeAsmPseudoInst<"vst2${p}", ".32", "$list, $addr!",
8214 (ins VecListTwoQWordIndexed:$list, addrmode6align64:$addr,
8216 def VST2LNdWB_register_Asm_8 :
8217 NEONDataTypeAsmPseudoInst<"vst2${p}", ".8", "$list, $addr, $Rm",
8218 (ins VecListTwoDByteIndexed:$list, addrmode6align16:$addr,
8219 rGPR:$Rm, pred:$p)>;
8220 def VST2LNdWB_register_Asm_16 :
8221 NEONDataTypeAsmPseudoInst<"vst2${p}", ".16","$list, $addr, $Rm",
8222 (ins VecListTwoDHWordIndexed:$list, addrmode6align32:$addr,
8223 rGPR:$Rm, pred:$p)>;
8224 def VST2LNdWB_register_Asm_32 :
8225 NEONDataTypeAsmPseudoInst<"vst2${p}", ".32", "$list, $addr, $Rm",
8226 (ins VecListTwoDWordIndexed:$list, addrmode6align64:$addr,
8227 rGPR:$Rm, pred:$p)>;
8228 def VST2LNqWB_register_Asm_16 :
8229 NEONDataTypeAsmPseudoInst<"vst2${p}", ".16","$list, $addr, $Rm",
8230 (ins VecListTwoQHWordIndexed:$list, addrmode6align32:$addr,
8231 rGPR:$Rm, pred:$p)>;
8232 def VST2LNqWB_register_Asm_32 :
8233 NEONDataTypeAsmPseudoInst<"vst2${p}", ".32", "$list, $addr, $Rm",
8234 (ins VecListTwoQWordIndexed:$list, addrmode6align64:$addr,
8235 rGPR:$Rm, pred:$p)>;
8237 // VLD3 all-lanes pseudo-instructions. These need special handling for
8238 // the lane index that an InstAlias can't handle, so we use these instead.
8239 def VLD3DUPdAsm_8 : NEONDataTypeAsmPseudoInst<"vld3${p}", ".8", "$list, $addr",
8240 (ins VecListThreeDAllLanes:$list, addrmode6dupalignNone:$addr,
8242 def VLD3DUPdAsm_16: NEONDataTypeAsmPseudoInst<"vld3${p}", ".16", "$list, $addr",
8243 (ins VecListThreeDAllLanes:$list, addrmode6dupalignNone:$addr,
8245 def VLD3DUPdAsm_32: NEONDataTypeAsmPseudoInst<"vld3${p}", ".32", "$list, $addr",
8246 (ins VecListThreeDAllLanes:$list, addrmode6dupalignNone:$addr,
8248 def VLD3DUPqAsm_8 : NEONDataTypeAsmPseudoInst<"vld3${p}", ".8", "$list, $addr",
8249 (ins VecListThreeQAllLanes:$list, addrmode6dupalignNone:$addr,
8251 def VLD3DUPqAsm_16: NEONDataTypeAsmPseudoInst<"vld3${p}", ".16", "$list, $addr",
8252 (ins VecListThreeQAllLanes:$list, addrmode6dupalignNone:$addr,
8254 def VLD3DUPqAsm_32: NEONDataTypeAsmPseudoInst<"vld3${p}", ".32", "$list, $addr",
8255 (ins VecListThreeQAllLanes:$list, addrmode6dupalignNone:$addr,
8258 def VLD3DUPdWB_fixed_Asm_8 :
8259 NEONDataTypeAsmPseudoInst<"vld3${p}", ".8", "$list, $addr!",
8260 (ins VecListThreeDAllLanes:$list, addrmode6dupalignNone:$addr,
8262 def VLD3DUPdWB_fixed_Asm_16 :
8263 NEONDataTypeAsmPseudoInst<"vld3${p}", ".16", "$list, $addr!",
8264 (ins VecListThreeDAllLanes:$list, addrmode6dupalignNone:$addr,
8266 def VLD3DUPdWB_fixed_Asm_32 :
8267 NEONDataTypeAsmPseudoInst<"vld3${p}", ".32", "$list, $addr!",
8268 (ins VecListThreeDAllLanes:$list, addrmode6dupalignNone:$addr,
8270 def VLD3DUPqWB_fixed_Asm_8 :
8271 NEONDataTypeAsmPseudoInst<"vld3${p}", ".8", "$list, $addr!",
8272 (ins VecListThreeQAllLanes:$list, addrmode6dupalignNone:$addr,
8274 def VLD3DUPqWB_fixed_Asm_16 :
8275 NEONDataTypeAsmPseudoInst<"vld3${p}", ".16", "$list, $addr!",
8276 (ins VecListThreeQAllLanes:$list, addrmode6dupalignNone:$addr,
8278 def VLD3DUPqWB_fixed_Asm_32 :
8279 NEONDataTypeAsmPseudoInst<"vld3${p}", ".32", "$list, $addr!",
8280 (ins VecListThreeQAllLanes:$list, addrmode6dupalignNone:$addr,
8282 def VLD3DUPdWB_register_Asm_8 :
8283 NEONDataTypeAsmPseudoInst<"vld3${p}", ".8", "$list, $addr, $Rm",
8284 (ins VecListThreeDAllLanes:$list, addrmode6dupalignNone:$addr,
8285 rGPR:$Rm, pred:$p)>;
8286 def VLD3DUPdWB_register_Asm_16 :
8287 NEONDataTypeAsmPseudoInst<"vld3${p}", ".16", "$list, $addr, $Rm",
8288 (ins VecListThreeDAllLanes:$list, addrmode6dupalignNone:$addr,
8289 rGPR:$Rm, pred:$p)>;
8290 def VLD3DUPdWB_register_Asm_32 :
8291 NEONDataTypeAsmPseudoInst<"vld3${p}", ".32", "$list, $addr, $Rm",
8292 (ins VecListThreeDAllLanes:$list, addrmode6dupalignNone:$addr,
8293 rGPR:$Rm, pred:$p)>;
8294 def VLD3DUPqWB_register_Asm_8 :
8295 NEONDataTypeAsmPseudoInst<"vld3${p}", ".8", "$list, $addr, $Rm",
8296 (ins VecListThreeQAllLanes:$list, addrmode6dupalignNone:$addr,
8297 rGPR:$Rm, pred:$p)>;
8298 def VLD3DUPqWB_register_Asm_16 :
8299 NEONDataTypeAsmPseudoInst<"vld3${p}", ".16", "$list, $addr, $Rm",
8300 (ins VecListThreeQAllLanes:$list, addrmode6dupalignNone:$addr,
8301 rGPR:$Rm, pred:$p)>;
8302 def VLD3DUPqWB_register_Asm_32 :
8303 NEONDataTypeAsmPseudoInst<"vld3${p}", ".32", "$list, $addr, $Rm",
8304 (ins VecListThreeQAllLanes:$list, addrmode6dupalignNone:$addr,
8305 rGPR:$Rm, pred:$p)>;
8308 // VLD3 single-lane pseudo-instructions. These need special handling for
8309 // the lane index that an InstAlias can't handle, so we use these instead.
8310 def VLD3LNdAsm_8 : NEONDataTypeAsmPseudoInst<"vld3${p}", ".8", "$list, $addr",
8311 (ins VecListThreeDByteIndexed:$list, addrmode6alignNone:$addr,
8313 def VLD3LNdAsm_16 : NEONDataTypeAsmPseudoInst<"vld3${p}", ".16", "$list, $addr",
8314 (ins VecListThreeDHWordIndexed:$list, addrmode6alignNone:$addr,
8316 def VLD3LNdAsm_32 : NEONDataTypeAsmPseudoInst<"vld3${p}", ".32", "$list, $addr",
8317 (ins VecListThreeDWordIndexed:$list, addrmode6alignNone:$addr,
8319 def VLD3LNqAsm_16 : NEONDataTypeAsmPseudoInst<"vld3${p}", ".16", "$list, $addr",
8320 (ins VecListThreeQHWordIndexed:$list, addrmode6alignNone:$addr,
8322 def VLD3LNqAsm_32 : NEONDataTypeAsmPseudoInst<"vld3${p}", ".32", "$list, $addr",
8323 (ins VecListThreeQWordIndexed:$list, addrmode6alignNone:$addr,
8326 def VLD3LNdWB_fixed_Asm_8 :
8327 NEONDataTypeAsmPseudoInst<"vld3${p}", ".8", "$list, $addr!",
8328 (ins VecListThreeDByteIndexed:$list, addrmode6alignNone:$addr,
8330 def VLD3LNdWB_fixed_Asm_16 :
8331 NEONDataTypeAsmPseudoInst<"vld3${p}", ".16", "$list, $addr!",
8332 (ins VecListThreeDHWordIndexed:$list, addrmode6alignNone:$addr,
8334 def VLD3LNdWB_fixed_Asm_32 :
8335 NEONDataTypeAsmPseudoInst<"vld3${p}", ".32", "$list, $addr!",
8336 (ins VecListThreeDWordIndexed:$list, addrmode6alignNone:$addr,
8338 def VLD3LNqWB_fixed_Asm_16 :
8339 NEONDataTypeAsmPseudoInst<"vld3${p}", ".16", "$list, $addr!",
8340 (ins VecListThreeQHWordIndexed:$list, addrmode6alignNone:$addr,
8342 def VLD3LNqWB_fixed_Asm_32 :
8343 NEONDataTypeAsmPseudoInst<"vld3${p}", ".32", "$list, $addr!",
8344 (ins VecListThreeQWordIndexed:$list, addrmode6alignNone:$addr,
8346 def VLD3LNdWB_register_Asm_8 :
8347 NEONDataTypeAsmPseudoInst<"vld3${p}", ".8", "$list, $addr, $Rm",
8348 (ins VecListThreeDByteIndexed:$list, addrmode6alignNone:$addr,
8349 rGPR:$Rm, pred:$p)>;
8350 def VLD3LNdWB_register_Asm_16 :
8351 NEONDataTypeAsmPseudoInst<"vld3${p}", ".16", "$list, $addr, $Rm",
8352 (ins VecListThreeDHWordIndexed:$list,
8353 addrmode6alignNone:$addr, rGPR:$Rm, pred:$p)>;
8354 def VLD3LNdWB_register_Asm_32 :
8355 NEONDataTypeAsmPseudoInst<"vld3${p}", ".32", "$list, $addr, $Rm",
8356 (ins VecListThreeDWordIndexed:$list, addrmode6alignNone:$addr,
8357 rGPR:$Rm, pred:$p)>;
8358 def VLD3LNqWB_register_Asm_16 :
8359 NEONDataTypeAsmPseudoInst<"vld3${p}", ".16", "$list, $addr, $Rm",
8360 (ins VecListThreeQHWordIndexed:$list,
8361 addrmode6alignNone:$addr, rGPR:$Rm, pred:$p)>;
8362 def VLD3LNqWB_register_Asm_32 :
8363 NEONDataTypeAsmPseudoInst<"vld3${p}", ".32", "$list, $addr, $Rm",
8364 (ins VecListThreeQWordIndexed:$list, addrmode6alignNone:$addr,
8365 rGPR:$Rm, pred:$p)>;
8367 // VLD3 multiple structure pseudo-instructions. These need special handling for
8368 // the vector operands that the normal instructions don't yet model.
8369 // FIXME: Remove these when the register classes and instructions are updated.
8370 def VLD3dAsm_8 : NEONDataTypeAsmPseudoInst<"vld3${p}", ".8", "$list, $addr",
8371 (ins VecListThreeD:$list, addrmode6align64:$addr, pred:$p)>;
8372 def VLD3dAsm_16 : NEONDataTypeAsmPseudoInst<"vld3${p}", ".16", "$list, $addr",
8373 (ins VecListThreeD:$list, addrmode6align64:$addr, pred:$p)>;
8374 def VLD3dAsm_32 : NEONDataTypeAsmPseudoInst<"vld3${p}", ".32", "$list, $addr",
8375 (ins VecListThreeD:$list, addrmode6align64:$addr, pred:$p)>;
8376 def VLD3qAsm_8 : NEONDataTypeAsmPseudoInst<"vld3${p}", ".8", "$list, $addr",
8377 (ins VecListThreeQ:$list, addrmode6align64:$addr, pred:$p)>;
8378 def VLD3qAsm_16 : NEONDataTypeAsmPseudoInst<"vld3${p}", ".16", "$list, $addr",
8379 (ins VecListThreeQ:$list, addrmode6align64:$addr, pred:$p)>;
8380 def VLD3qAsm_32 : NEONDataTypeAsmPseudoInst<"vld3${p}", ".32", "$list, $addr",
8381 (ins VecListThreeQ:$list, addrmode6align64:$addr, pred:$p)>;
8383 def VLD3dWB_fixed_Asm_8 :
8384 NEONDataTypeAsmPseudoInst<"vld3${p}", ".8", "$list, $addr!",
8385 (ins VecListThreeD:$list, addrmode6align64:$addr, pred:$p)>;
8386 def VLD3dWB_fixed_Asm_16 :
8387 NEONDataTypeAsmPseudoInst<"vld3${p}", ".16", "$list, $addr!",
8388 (ins VecListThreeD:$list, addrmode6align64:$addr, pred:$p)>;
8389 def VLD3dWB_fixed_Asm_32 :
8390 NEONDataTypeAsmPseudoInst<"vld3${p}", ".32", "$list, $addr!",
8391 (ins VecListThreeD:$list, addrmode6align64:$addr, pred:$p)>;
8392 def VLD3qWB_fixed_Asm_8 :
8393 NEONDataTypeAsmPseudoInst<"vld3${p}", ".8", "$list, $addr!",
8394 (ins VecListThreeQ:$list, addrmode6align64:$addr, pred:$p)>;
8395 def VLD3qWB_fixed_Asm_16 :
8396 NEONDataTypeAsmPseudoInst<"vld3${p}", ".16", "$list, $addr!",
8397 (ins VecListThreeQ:$list, addrmode6align64:$addr, pred:$p)>;
8398 def VLD3qWB_fixed_Asm_32 :
8399 NEONDataTypeAsmPseudoInst<"vld3${p}", ".32", "$list, $addr!",
8400 (ins VecListThreeQ:$list, addrmode6align64:$addr, pred:$p)>;
8401 def VLD3dWB_register_Asm_8 :
8402 NEONDataTypeAsmPseudoInst<"vld3${p}", ".8", "$list, $addr, $Rm",
8403 (ins VecListThreeD:$list, addrmode6align64:$addr,
8404 rGPR:$Rm, pred:$p)>;
8405 def VLD3dWB_register_Asm_16 :
8406 NEONDataTypeAsmPseudoInst<"vld3${p}", ".16", "$list, $addr, $Rm",
8407 (ins VecListThreeD:$list, addrmode6align64:$addr,
8408 rGPR:$Rm, pred:$p)>;
8409 def VLD3dWB_register_Asm_32 :
8410 NEONDataTypeAsmPseudoInst<"vld3${p}", ".32", "$list, $addr, $Rm",
8411 (ins VecListThreeD:$list, addrmode6align64:$addr,
8412 rGPR:$Rm, pred:$p)>;
8413 def VLD3qWB_register_Asm_8 :
8414 NEONDataTypeAsmPseudoInst<"vld3${p}", ".8", "$list, $addr, $Rm",
8415 (ins VecListThreeQ:$list, addrmode6align64:$addr,
8416 rGPR:$Rm, pred:$p)>;
8417 def VLD3qWB_register_Asm_16 :
8418 NEONDataTypeAsmPseudoInst<"vld3${p}", ".16", "$list, $addr, $Rm",
8419 (ins VecListThreeQ:$list, addrmode6align64:$addr,
8420 rGPR:$Rm, pred:$p)>;
8421 def VLD3qWB_register_Asm_32 :
8422 NEONDataTypeAsmPseudoInst<"vld3${p}", ".32", "$list, $addr, $Rm",
8423 (ins VecListThreeQ:$list, addrmode6align64:$addr,
8424 rGPR:$Rm, pred:$p)>;
8426 // VST3 single-lane pseudo-instructions. These need special handling for
8427 // the lane index that an InstAlias can't handle, so we use these instead.
8428 def VST3LNdAsm_8 : NEONDataTypeAsmPseudoInst<"vst3${p}", ".8", "$list, $addr",
8429 (ins VecListThreeDByteIndexed:$list, addrmode6alignNone:$addr,
8431 def VST3LNdAsm_16 : NEONDataTypeAsmPseudoInst<"vst3${p}", ".16", "$list, $addr",
8432 (ins VecListThreeDHWordIndexed:$list, addrmode6alignNone:$addr,
8434 def VST3LNdAsm_32 : NEONDataTypeAsmPseudoInst<"vst3${p}", ".32", "$list, $addr",
8435 (ins VecListThreeDWordIndexed:$list, addrmode6alignNone:$addr,
8437 def VST3LNqAsm_16 : NEONDataTypeAsmPseudoInst<"vst3${p}", ".16", "$list, $addr",
8438 (ins VecListThreeQHWordIndexed:$list, addrmode6alignNone:$addr,
8440 def VST3LNqAsm_32 : NEONDataTypeAsmPseudoInst<"vst3${p}", ".32", "$list, $addr",
8441 (ins VecListThreeQWordIndexed:$list, addrmode6alignNone:$addr,
8444 def VST3LNdWB_fixed_Asm_8 :
8445 NEONDataTypeAsmPseudoInst<"vst3${p}", ".8", "$list, $addr!",
8446 (ins VecListThreeDByteIndexed:$list, addrmode6alignNone:$addr,
8448 def VST3LNdWB_fixed_Asm_16 :
8449 NEONDataTypeAsmPseudoInst<"vst3${p}", ".16", "$list, $addr!",
8450 (ins VecListThreeDHWordIndexed:$list, addrmode6alignNone:$addr,
8452 def VST3LNdWB_fixed_Asm_32 :
8453 NEONDataTypeAsmPseudoInst<"vst3${p}", ".32", "$list, $addr!",
8454 (ins VecListThreeDWordIndexed:$list, addrmode6alignNone:$addr,
8456 def VST3LNqWB_fixed_Asm_16 :
8457 NEONDataTypeAsmPseudoInst<"vst3${p}", ".16", "$list, $addr!",
8458 (ins VecListThreeQHWordIndexed:$list, addrmode6alignNone:$addr,
8460 def VST3LNqWB_fixed_Asm_32 :
8461 NEONDataTypeAsmPseudoInst<"vst3${p}", ".32", "$list, $addr!",
8462 (ins VecListThreeQWordIndexed:$list, addrmode6alignNone:$addr,
8464 def VST3LNdWB_register_Asm_8 :
8465 NEONDataTypeAsmPseudoInst<"vst3${p}", ".8", "$list, $addr, $Rm",
8466 (ins VecListThreeDByteIndexed:$list, addrmode6alignNone:$addr,
8467 rGPR:$Rm, pred:$p)>;
8468 def VST3LNdWB_register_Asm_16 :
8469 NEONDataTypeAsmPseudoInst<"vst3${p}", ".16", "$list, $addr, $Rm",
8470 (ins VecListThreeDHWordIndexed:$list,
8471 addrmode6alignNone:$addr, rGPR:$Rm, pred:$p)>;
8472 def VST3LNdWB_register_Asm_32 :
8473 NEONDataTypeAsmPseudoInst<"vst3${p}", ".32", "$list, $addr, $Rm",
8474 (ins VecListThreeDWordIndexed:$list, addrmode6alignNone:$addr,
8475 rGPR:$Rm, pred:$p)>;
8476 def VST3LNqWB_register_Asm_16 :
8477 NEONDataTypeAsmPseudoInst<"vst3${p}", ".16", "$list, $addr, $Rm",
8478 (ins VecListThreeQHWordIndexed:$list,
8479 addrmode6alignNone:$addr, rGPR:$Rm, pred:$p)>;
8480 def VST3LNqWB_register_Asm_32 :
8481 NEONDataTypeAsmPseudoInst<"vst3${p}", ".32", "$list, $addr, $Rm",
8482 (ins VecListThreeQWordIndexed:$list, addrmode6alignNone:$addr,
8483 rGPR:$Rm, pred:$p)>;
8486 // VST3 multiple structure pseudo-instructions. These need special handling for
8487 // the vector operands that the normal instructions don't yet model.
8488 // FIXME: Remove these when the register classes and instructions are updated.
8489 def VST3dAsm_8 : NEONDataTypeAsmPseudoInst<"vst3${p}", ".8", "$list, $addr",
8490 (ins VecListThreeD:$list, addrmode6align64:$addr, pred:$p)>;
8491 def VST3dAsm_16 : NEONDataTypeAsmPseudoInst<"vst3${p}", ".16", "$list, $addr",
8492 (ins VecListThreeD:$list, addrmode6align64:$addr, pred:$p)>;
8493 def VST3dAsm_32 : NEONDataTypeAsmPseudoInst<"vst3${p}", ".32", "$list, $addr",
8494 (ins VecListThreeD:$list, addrmode6align64:$addr, pred:$p)>;
8495 def VST3qAsm_8 : NEONDataTypeAsmPseudoInst<"vst3${p}", ".8", "$list, $addr",
8496 (ins VecListThreeQ:$list, addrmode6align64:$addr, pred:$p)>;
8497 def VST3qAsm_16 : NEONDataTypeAsmPseudoInst<"vst3${p}", ".16", "$list, $addr",
8498 (ins VecListThreeQ:$list, addrmode6align64:$addr, pred:$p)>;
8499 def VST3qAsm_32 : NEONDataTypeAsmPseudoInst<"vst3${p}", ".32", "$list, $addr",
8500 (ins VecListThreeQ:$list, addrmode6align64:$addr, pred:$p)>;
8502 def VST3dWB_fixed_Asm_8 :
8503 NEONDataTypeAsmPseudoInst<"vst3${p}", ".8", "$list, $addr!",
8504 (ins VecListThreeD:$list, addrmode6align64:$addr, pred:$p)>;
8505 def VST3dWB_fixed_Asm_16 :
8506 NEONDataTypeAsmPseudoInst<"vst3${p}", ".16", "$list, $addr!",
8507 (ins VecListThreeD:$list, addrmode6align64:$addr, pred:$p)>;
8508 def VST3dWB_fixed_Asm_32 :
8509 NEONDataTypeAsmPseudoInst<"vst3${p}", ".32", "$list, $addr!",
8510 (ins VecListThreeD:$list, addrmode6align64:$addr, pred:$p)>;
8511 def VST3qWB_fixed_Asm_8 :
8512 NEONDataTypeAsmPseudoInst<"vst3${p}", ".8", "$list, $addr!",
8513 (ins VecListThreeQ:$list, addrmode6align64:$addr, pred:$p)>;
8514 def VST3qWB_fixed_Asm_16 :
8515 NEONDataTypeAsmPseudoInst<"vst3${p}", ".16", "$list, $addr!",
8516 (ins VecListThreeQ:$list, addrmode6align64:$addr, pred:$p)>;
8517 def VST3qWB_fixed_Asm_32 :
8518 NEONDataTypeAsmPseudoInst<"vst3${p}", ".32", "$list, $addr!",
8519 (ins VecListThreeQ:$list, addrmode6align64:$addr, pred:$p)>;
8520 def VST3dWB_register_Asm_8 :
8521 NEONDataTypeAsmPseudoInst<"vst3${p}", ".8", "$list, $addr, $Rm",
8522 (ins VecListThreeD:$list, addrmode6align64:$addr,
8523 rGPR:$Rm, pred:$p)>;
8524 def VST3dWB_register_Asm_16 :
8525 NEONDataTypeAsmPseudoInst<"vst3${p}", ".16", "$list, $addr, $Rm",
8526 (ins VecListThreeD:$list, addrmode6align64:$addr,
8527 rGPR:$Rm, pred:$p)>;
8528 def VST3dWB_register_Asm_32 :
8529 NEONDataTypeAsmPseudoInst<"vst3${p}", ".32", "$list, $addr, $Rm",
8530 (ins VecListThreeD:$list, addrmode6align64:$addr,
8531 rGPR:$Rm, pred:$p)>;
8532 def VST3qWB_register_Asm_8 :
8533 NEONDataTypeAsmPseudoInst<"vst3${p}", ".8", "$list, $addr, $Rm",
8534 (ins VecListThreeQ:$list, addrmode6align64:$addr,
8535 rGPR:$Rm, pred:$p)>;
8536 def VST3qWB_register_Asm_16 :
8537 NEONDataTypeAsmPseudoInst<"vst3${p}", ".16", "$list, $addr, $Rm",
8538 (ins VecListThreeQ:$list, addrmode6align64:$addr,
8539 rGPR:$Rm, pred:$p)>;
8540 def VST3qWB_register_Asm_32 :
8541 NEONDataTypeAsmPseudoInst<"vst3${p}", ".32", "$list, $addr, $Rm",
8542 (ins VecListThreeQ:$list, addrmode6align64:$addr,
8543 rGPR:$Rm, pred:$p)>;
8545 // VLD4 all-lanes pseudo-instructions. These need special handling for
8546 // the lane index that an InstAlias can't handle, so we use these instead.
8547 def VLD4DUPdAsm_8 : NEONDataTypeAsmPseudoInst<"vld4${p}", ".8", "$list, $addr",
8548 (ins VecListFourDAllLanes:$list, addrmode6dupalign32:$addr,
8550 def VLD4DUPdAsm_16: NEONDataTypeAsmPseudoInst<"vld4${p}", ".16", "$list, $addr",
8551 (ins VecListFourDAllLanes:$list, addrmode6dupalign64:$addr,
8553 def VLD4DUPdAsm_32: NEONDataTypeAsmPseudoInst<"vld4${p}", ".32", "$list, $addr",
8554 (ins VecListFourDAllLanes:$list, addrmode6dupalign64or128:$addr,
8556 def VLD4DUPqAsm_8 : NEONDataTypeAsmPseudoInst<"vld4${p}", ".8", "$list, $addr",
8557 (ins VecListFourQAllLanes:$list, addrmode6dupalign32:$addr,
8559 def VLD4DUPqAsm_16: NEONDataTypeAsmPseudoInst<"vld4${p}", ".16", "$list, $addr",
8560 (ins VecListFourQAllLanes:$list, addrmode6dupalign64:$addr,
8562 def VLD4DUPqAsm_32: NEONDataTypeAsmPseudoInst<"vld4${p}", ".32", "$list, $addr",
8563 (ins VecListFourQAllLanes:$list, addrmode6dupalign64or128:$addr,
8566 def VLD4DUPdWB_fixed_Asm_8 :
8567 NEONDataTypeAsmPseudoInst<"vld4${p}", ".8", "$list, $addr!",
8568 (ins VecListFourDAllLanes:$list, addrmode6dupalign32:$addr,
8570 def VLD4DUPdWB_fixed_Asm_16 :
8571 NEONDataTypeAsmPseudoInst<"vld4${p}", ".16", "$list, $addr!",
8572 (ins VecListFourDAllLanes:$list, addrmode6dupalign64:$addr,
8574 def VLD4DUPdWB_fixed_Asm_32 :
8575 NEONDataTypeAsmPseudoInst<"vld4${p}", ".32", "$list, $addr!",
8576 (ins VecListFourDAllLanes:$list, addrmode6dupalign64or128:$addr,
8578 def VLD4DUPqWB_fixed_Asm_8 :
8579 NEONDataTypeAsmPseudoInst<"vld4${p}", ".8", "$list, $addr!",
8580 (ins VecListFourQAllLanes:$list, addrmode6dupalign32:$addr,
8582 def VLD4DUPqWB_fixed_Asm_16 :
8583 NEONDataTypeAsmPseudoInst<"vld4${p}", ".16", "$list, $addr!",
8584 (ins VecListFourQAllLanes:$list, addrmode6dupalign64:$addr,
8586 def VLD4DUPqWB_fixed_Asm_32 :
8587 NEONDataTypeAsmPseudoInst<"vld4${p}", ".32", "$list, $addr!",
8588 (ins VecListFourQAllLanes:$list, addrmode6dupalign64or128:$addr,
8590 def VLD4DUPdWB_register_Asm_8 :
8591 NEONDataTypeAsmPseudoInst<"vld4${p}", ".8", "$list, $addr, $Rm",
8592 (ins VecListFourDAllLanes:$list, addrmode6dupalign32:$addr,
8593 rGPR:$Rm, pred:$p)>;
8594 def VLD4DUPdWB_register_Asm_16 :
8595 NEONDataTypeAsmPseudoInst<"vld4${p}", ".16", "$list, $addr, $Rm",
8596 (ins VecListFourDAllLanes:$list, addrmode6dupalign64:$addr,
8597 rGPR:$Rm, pred:$p)>;
8598 def VLD4DUPdWB_register_Asm_32 :
8599 NEONDataTypeAsmPseudoInst<"vld4${p}", ".32", "$list, $addr, $Rm",
8600 (ins VecListFourDAllLanes:$list,
8601 addrmode6dupalign64or128:$addr, rGPR:$Rm, pred:$p)>;
8602 def VLD4DUPqWB_register_Asm_8 :
8603 NEONDataTypeAsmPseudoInst<"vld4${p}", ".8", "$list, $addr, $Rm",
8604 (ins VecListFourQAllLanes:$list, addrmode6dupalign32:$addr,
8605 rGPR:$Rm, pred:$p)>;
8606 def VLD4DUPqWB_register_Asm_16 :
8607 NEONDataTypeAsmPseudoInst<"vld4${p}", ".16", "$list, $addr, $Rm",
8608 (ins VecListFourQAllLanes:$list, addrmode6dupalign64:$addr,
8609 rGPR:$Rm, pred:$p)>;
8610 def VLD4DUPqWB_register_Asm_32 :
8611 NEONDataTypeAsmPseudoInst<"vld4${p}", ".32", "$list, $addr, $Rm",
8612 (ins VecListFourQAllLanes:$list,
8613 addrmode6dupalign64or128:$addr, rGPR:$Rm, pred:$p)>;
8616 // VLD4 single-lane pseudo-instructions. These need special handling for
8617 // the lane index that an InstAlias can't handle, so we use these instead.
8618 def VLD4LNdAsm_8 : NEONDataTypeAsmPseudoInst<"vld4${p}", ".8", "$list, $addr",
8619 (ins VecListFourDByteIndexed:$list, addrmode6align32:$addr,
8621 def VLD4LNdAsm_16 : NEONDataTypeAsmPseudoInst<"vld4${p}", ".16", "$list, $addr",
8622 (ins VecListFourDHWordIndexed:$list, addrmode6align64:$addr,
8624 def VLD4LNdAsm_32 : NEONDataTypeAsmPseudoInst<"vld4${p}", ".32", "$list, $addr",
8625 (ins VecListFourDWordIndexed:$list, addrmode6align64or128:$addr,
8627 def VLD4LNqAsm_16 : NEONDataTypeAsmPseudoInst<"vld4${p}", ".16", "$list, $addr",
8628 (ins VecListFourQHWordIndexed:$list, addrmode6align64:$addr,
8630 def VLD4LNqAsm_32 : NEONDataTypeAsmPseudoInst<"vld4${p}", ".32", "$list, $addr",
8631 (ins VecListFourQWordIndexed:$list, addrmode6align64or128:$addr,
8634 def VLD4LNdWB_fixed_Asm_8 :
8635 NEONDataTypeAsmPseudoInst<"vld4${p}", ".8", "$list, $addr!",
8636 (ins VecListFourDByteIndexed:$list, addrmode6align32:$addr,
8638 def VLD4LNdWB_fixed_Asm_16 :
8639 NEONDataTypeAsmPseudoInst<"vld4${p}", ".16", "$list, $addr!",
8640 (ins VecListFourDHWordIndexed:$list, addrmode6align64:$addr,
8642 def VLD4LNdWB_fixed_Asm_32 :
8643 NEONDataTypeAsmPseudoInst<"vld4${p}", ".32", "$list, $addr!",
8644 (ins VecListFourDWordIndexed:$list, addrmode6align64or128:$addr,
8646 def VLD4LNqWB_fixed_Asm_16 :
8647 NEONDataTypeAsmPseudoInst<"vld4${p}", ".16", "$list, $addr!",
8648 (ins VecListFourQHWordIndexed:$list, addrmode6align64:$addr,
8650 def VLD4LNqWB_fixed_Asm_32 :
8651 NEONDataTypeAsmPseudoInst<"vld4${p}", ".32", "$list, $addr!",
8652 (ins VecListFourQWordIndexed:$list, addrmode6align64or128:$addr,
8654 def VLD4LNdWB_register_Asm_8 :
8655 NEONDataTypeAsmPseudoInst<"vld4${p}", ".8", "$list, $addr, $Rm",
8656 (ins VecListFourDByteIndexed:$list, addrmode6align32:$addr,
8657 rGPR:$Rm, pred:$p)>;
8658 def VLD4LNdWB_register_Asm_16 :
8659 NEONDataTypeAsmPseudoInst<"vld4${p}", ".16", "$list, $addr, $Rm",
8660 (ins VecListFourDHWordIndexed:$list, addrmode6align64:$addr,
8661 rGPR:$Rm, pred:$p)>;
8662 def VLD4LNdWB_register_Asm_32 :
8663 NEONDataTypeAsmPseudoInst<"vld4${p}", ".32", "$list, $addr, $Rm",
8664 (ins VecListFourDWordIndexed:$list,
8665 addrmode6align64or128:$addr, rGPR:$Rm, pred:$p)>;
8666 def VLD4LNqWB_register_Asm_16 :
8667 NEONDataTypeAsmPseudoInst<"vld4${p}", ".16", "$list, $addr, $Rm",
8668 (ins VecListFourQHWordIndexed:$list, addrmode6align64:$addr,
8669 rGPR:$Rm, pred:$p)>;
8670 def VLD4LNqWB_register_Asm_32 :
8671 NEONDataTypeAsmPseudoInst<"vld4${p}", ".32", "$list, $addr, $Rm",
8672 (ins VecListFourQWordIndexed:$list,
8673 addrmode6align64or128:$addr, rGPR:$Rm, pred:$p)>;
8677 // VLD4 multiple structure pseudo-instructions. These need special handling for
8678 // the vector operands that the normal instructions don't yet model.
8679 // FIXME: Remove these when the register classes and instructions are updated.
8680 def VLD4dAsm_8 : NEONDataTypeAsmPseudoInst<"vld4${p}", ".8", "$list, $addr",
8681 (ins VecListFourD:$list, addrmode6align64or128or256:$addr,
8683 def VLD4dAsm_16 : NEONDataTypeAsmPseudoInst<"vld4${p}", ".16", "$list, $addr",
8684 (ins VecListFourD:$list, addrmode6align64or128or256:$addr,
8686 def VLD4dAsm_32 : NEONDataTypeAsmPseudoInst<"vld4${p}", ".32", "$list, $addr",
8687 (ins VecListFourD:$list, addrmode6align64or128or256:$addr,
8689 def VLD4qAsm_8 : NEONDataTypeAsmPseudoInst<"vld4${p}", ".8", "$list, $addr",
8690 (ins VecListFourQ:$list, addrmode6align64or128or256:$addr,
8692 def VLD4qAsm_16 : NEONDataTypeAsmPseudoInst<"vld4${p}", ".16", "$list, $addr",
8693 (ins VecListFourQ:$list, addrmode6align64or128or256:$addr,
8695 def VLD4qAsm_32 : NEONDataTypeAsmPseudoInst<"vld4${p}", ".32", "$list, $addr",
8696 (ins VecListFourQ:$list, addrmode6align64or128or256:$addr,
8699 def VLD4dWB_fixed_Asm_8 :
8700 NEONDataTypeAsmPseudoInst<"vld4${p}", ".8", "$list, $addr!",
8701 (ins VecListFourD:$list, addrmode6align64or128or256:$addr,
8703 def VLD4dWB_fixed_Asm_16 :
8704 NEONDataTypeAsmPseudoInst<"vld4${p}", ".16", "$list, $addr!",
8705 (ins VecListFourD:$list, addrmode6align64or128or256:$addr,
8707 def VLD4dWB_fixed_Asm_32 :
8708 NEONDataTypeAsmPseudoInst<"vld4${p}", ".32", "$list, $addr!",
8709 (ins VecListFourD:$list, addrmode6align64or128or256:$addr,
8711 def VLD4qWB_fixed_Asm_8 :
8712 NEONDataTypeAsmPseudoInst<"vld4${p}", ".8", "$list, $addr!",
8713 (ins VecListFourQ:$list, addrmode6align64or128or256:$addr,
8715 def VLD4qWB_fixed_Asm_16 :
8716 NEONDataTypeAsmPseudoInst<"vld4${p}", ".16", "$list, $addr!",
8717 (ins VecListFourQ:$list, addrmode6align64or128or256:$addr,
8719 def VLD4qWB_fixed_Asm_32 :
8720 NEONDataTypeAsmPseudoInst<"vld4${p}", ".32", "$list, $addr!",
8721 (ins VecListFourQ:$list, addrmode6align64or128or256:$addr,
8723 def VLD4dWB_register_Asm_8 :
8724 NEONDataTypeAsmPseudoInst<"vld4${p}", ".8", "$list, $addr, $Rm",
8725 (ins VecListFourD:$list, addrmode6align64or128or256:$addr,
8726 rGPR:$Rm, pred:$p)>;
8727 def VLD4dWB_register_Asm_16 :
8728 NEONDataTypeAsmPseudoInst<"vld4${p}", ".16", "$list, $addr, $Rm",
8729 (ins VecListFourD:$list, addrmode6align64or128or256:$addr,
8730 rGPR:$Rm, pred:$p)>;
8731 def VLD4dWB_register_Asm_32 :
8732 NEONDataTypeAsmPseudoInst<"vld4${p}", ".32", "$list, $addr, $Rm",
8733 (ins VecListFourD:$list, addrmode6align64or128or256:$addr,
8734 rGPR:$Rm, pred:$p)>;
8735 def VLD4qWB_register_Asm_8 :
8736 NEONDataTypeAsmPseudoInst<"vld4${p}", ".8", "$list, $addr, $Rm",
8737 (ins VecListFourQ:$list, addrmode6align64or128or256:$addr,
8738 rGPR:$Rm, pred:$p)>;
8739 def VLD4qWB_register_Asm_16 :
8740 NEONDataTypeAsmPseudoInst<"vld4${p}", ".16", "$list, $addr, $Rm",
8741 (ins VecListFourQ:$list, addrmode6align64or128or256:$addr,
8742 rGPR:$Rm, pred:$p)>;
8743 def VLD4qWB_register_Asm_32 :
8744 NEONDataTypeAsmPseudoInst<"vld4${p}", ".32", "$list, $addr, $Rm",
8745 (ins VecListFourQ:$list, addrmode6align64or128or256:$addr,
8746 rGPR:$Rm, pred:$p)>;
8748 // VST4 single-lane pseudo-instructions. These need special handling for
8749 // the lane index that an InstAlias can't handle, so we use these instead.
8750 def VST4LNdAsm_8 : NEONDataTypeAsmPseudoInst<"vst4${p}", ".8", "$list, $addr",
8751 (ins VecListFourDByteIndexed:$list, addrmode6align32:$addr,
8753 def VST4LNdAsm_16 : NEONDataTypeAsmPseudoInst<"vst4${p}", ".16", "$list, $addr",
8754 (ins VecListFourDHWordIndexed:$list, addrmode6align64:$addr,
8756 def VST4LNdAsm_32 : NEONDataTypeAsmPseudoInst<"vst4${p}", ".32", "$list, $addr",
8757 (ins VecListFourDWordIndexed:$list, addrmode6align64or128:$addr,
8759 def VST4LNqAsm_16 : NEONDataTypeAsmPseudoInst<"vst4${p}", ".16", "$list, $addr",
8760 (ins VecListFourQHWordIndexed:$list, addrmode6align64:$addr,
8762 def VST4LNqAsm_32 : NEONDataTypeAsmPseudoInst<"vst4${p}", ".32", "$list, $addr",
8763 (ins VecListFourQWordIndexed:$list, addrmode6align64or128:$addr,
8766 def VST4LNdWB_fixed_Asm_8 :
8767 NEONDataTypeAsmPseudoInst<"vst4${p}", ".8", "$list, $addr!",
8768 (ins VecListFourDByteIndexed:$list, addrmode6align32:$addr,
8770 def VST4LNdWB_fixed_Asm_16 :
8771 NEONDataTypeAsmPseudoInst<"vst4${p}", ".16", "$list, $addr!",
8772 (ins VecListFourDHWordIndexed:$list, addrmode6align64:$addr,
8774 def VST4LNdWB_fixed_Asm_32 :
8775 NEONDataTypeAsmPseudoInst<"vst4${p}", ".32", "$list, $addr!",
8776 (ins VecListFourDWordIndexed:$list, addrmode6align64or128:$addr,
8778 def VST4LNqWB_fixed_Asm_16 :
8779 NEONDataTypeAsmPseudoInst<"vst4${p}", ".16", "$list, $addr!",
8780 (ins VecListFourQHWordIndexed:$list, addrmode6align64:$addr,
8782 def VST4LNqWB_fixed_Asm_32 :
8783 NEONDataTypeAsmPseudoInst<"vst4${p}", ".32", "$list, $addr!",
8784 (ins VecListFourQWordIndexed:$list, addrmode6align64or128:$addr,
8786 def VST4LNdWB_register_Asm_8 :
8787 NEONDataTypeAsmPseudoInst<"vst4${p}", ".8", "$list, $addr, $Rm",
8788 (ins VecListFourDByteIndexed:$list, addrmode6align32:$addr,
8789 rGPR:$Rm, pred:$p)>;
8790 def VST4LNdWB_register_Asm_16 :
8791 NEONDataTypeAsmPseudoInst<"vst4${p}", ".16", "$list, $addr, $Rm",
8792 (ins VecListFourDHWordIndexed:$list, addrmode6align64:$addr,
8793 rGPR:$Rm, pred:$p)>;
8794 def VST4LNdWB_register_Asm_32 :
8795 NEONDataTypeAsmPseudoInst<"vst4${p}", ".32", "$list, $addr, $Rm",
8796 (ins VecListFourDWordIndexed:$list,
8797 addrmode6align64or128:$addr, rGPR:$Rm, pred:$p)>;
8798 def VST4LNqWB_register_Asm_16 :
8799 NEONDataTypeAsmPseudoInst<"vst4${p}", ".16", "$list, $addr, $Rm",
8800 (ins VecListFourQHWordIndexed:$list, addrmode6align64:$addr,
8801 rGPR:$Rm, pred:$p)>;
8802 def VST4LNqWB_register_Asm_32 :
8803 NEONDataTypeAsmPseudoInst<"vst4${p}", ".32", "$list, $addr, $Rm",
8804 (ins VecListFourQWordIndexed:$list,
8805 addrmode6align64or128:$addr, rGPR:$Rm, pred:$p)>;
8808 // VST4 multiple structure pseudo-instructions. These need special handling for
8809 // the vector operands that the normal instructions don't yet model.
8810 // FIXME: Remove these when the register classes and instructions are updated.
8811 def VST4dAsm_8 : NEONDataTypeAsmPseudoInst<"vst4${p}", ".8", "$list, $addr",
8812 (ins VecListFourD:$list, addrmode6align64or128or256:$addr,
8814 def VST4dAsm_16 : NEONDataTypeAsmPseudoInst<"vst4${p}", ".16", "$list, $addr",
8815 (ins VecListFourD:$list, addrmode6align64or128or256:$addr,
8817 def VST4dAsm_32 : NEONDataTypeAsmPseudoInst<"vst4${p}", ".32", "$list, $addr",
8818 (ins VecListFourD:$list, addrmode6align64or128or256:$addr,
8820 def VST4qAsm_8 : NEONDataTypeAsmPseudoInst<"vst4${p}", ".8", "$list, $addr",
8821 (ins VecListFourQ:$list, addrmode6align64or128or256:$addr,
8823 def VST4qAsm_16 : NEONDataTypeAsmPseudoInst<"vst4${p}", ".16", "$list, $addr",
8824 (ins VecListFourQ:$list, addrmode6align64or128or256:$addr,
8826 def VST4qAsm_32 : NEONDataTypeAsmPseudoInst<"vst4${p}", ".32", "$list, $addr",
8827 (ins VecListFourQ:$list, addrmode6align64or128or256:$addr,
8830 def VST4dWB_fixed_Asm_8 :
8831 NEONDataTypeAsmPseudoInst<"vst4${p}", ".8", "$list, $addr!",
8832 (ins VecListFourD:$list, addrmode6align64or128or256:$addr,
8834 def VST4dWB_fixed_Asm_16 :
8835 NEONDataTypeAsmPseudoInst<"vst4${p}", ".16", "$list, $addr!",
8836 (ins VecListFourD:$list, addrmode6align64or128or256:$addr,
8838 def VST4dWB_fixed_Asm_32 :
8839 NEONDataTypeAsmPseudoInst<"vst4${p}", ".32", "$list, $addr!",
8840 (ins VecListFourD:$list, addrmode6align64or128or256:$addr,
8842 def VST4qWB_fixed_Asm_8 :
8843 NEONDataTypeAsmPseudoInst<"vst4${p}", ".8", "$list, $addr!",
8844 (ins VecListFourQ:$list, addrmode6align64or128or256:$addr,
8846 def VST4qWB_fixed_Asm_16 :
8847 NEONDataTypeAsmPseudoInst<"vst4${p}", ".16", "$list, $addr!",
8848 (ins VecListFourQ:$list, addrmode6align64or128or256:$addr,
8850 def VST4qWB_fixed_Asm_32 :
8851 NEONDataTypeAsmPseudoInst<"vst4${p}", ".32", "$list, $addr!",
8852 (ins VecListFourQ:$list, addrmode6align64or128or256:$addr,
8854 def VST4dWB_register_Asm_8 :
8855 NEONDataTypeAsmPseudoInst<"vst4${p}", ".8", "$list, $addr, $Rm",
8856 (ins VecListFourD:$list, addrmode6align64or128or256:$addr,
8857 rGPR:$Rm, pred:$p)>;
8858 def VST4dWB_register_Asm_16 :
8859 NEONDataTypeAsmPseudoInst<"vst4${p}", ".16", "$list, $addr, $Rm",
8860 (ins VecListFourD:$list, addrmode6align64or128or256:$addr,
8861 rGPR:$Rm, pred:$p)>;
8862 def VST4dWB_register_Asm_32 :
8863 NEONDataTypeAsmPseudoInst<"vst4${p}", ".32", "$list, $addr, $Rm",
8864 (ins VecListFourD:$list, addrmode6align64or128or256:$addr,
8865 rGPR:$Rm, pred:$p)>;
8866 def VST4qWB_register_Asm_8 :
8867 NEONDataTypeAsmPseudoInst<"vst4${p}", ".8", "$list, $addr, $Rm",
8868 (ins VecListFourQ:$list, addrmode6align64or128or256:$addr,
8869 rGPR:$Rm, pred:$p)>;
8870 def VST4qWB_register_Asm_16 :
8871 NEONDataTypeAsmPseudoInst<"vst4${p}", ".16", "$list, $addr, $Rm",
8872 (ins VecListFourQ:$list, addrmode6align64or128or256:$addr,
8873 rGPR:$Rm, pred:$p)>;
8874 def VST4qWB_register_Asm_32 :
8875 NEONDataTypeAsmPseudoInst<"vst4${p}", ".32", "$list, $addr, $Rm",
8876 (ins VecListFourQ:$list, addrmode6align64or128or256:$addr,
8877 rGPR:$Rm, pred:$p)>;
8879 // VMOV/VMVN takes an optional datatype suffix
8880 defm : NEONDTAnyInstAlias<"vmov${p}", "$Vd, $Vm",
8881 (VORRd DPR:$Vd, DPR:$Vm, DPR:$Vm, pred:$p)>;
8882 defm : NEONDTAnyInstAlias<"vmov${p}", "$Vd, $Vm",
8883 (VORRq QPR:$Vd, QPR:$Vm, QPR:$Vm, pred:$p)>;
8885 defm : NEONDTAnyInstAlias<"vmvn${p}", "$Vd, $Vm",
8886 (VMVNd DPR:$Vd, DPR:$Vm, pred:$p)>;
8887 defm : NEONDTAnyInstAlias<"vmvn${p}", "$Vd, $Vm",
8888 (VMVNq QPR:$Vd, QPR:$Vm, pred:$p)>;
8890 // VCLT (register) is an assembler alias for VCGT w/ the operands reversed.
8891 // D-register versions.
8892 def : NEONInstAlias<"vcle${p}.s8 $Dd, $Dn, $Dm",
8893 (VCGEsv8i8 DPR:$Dd, DPR:$Dm, DPR:$Dn, pred:$p)>;
8894 def : NEONInstAlias<"vcle${p}.s16 $Dd, $Dn, $Dm",
8895 (VCGEsv4i16 DPR:$Dd, DPR:$Dm, DPR:$Dn, pred:$p)>;
8896 def : NEONInstAlias<"vcle${p}.s32 $Dd, $Dn, $Dm",
8897 (VCGEsv2i32 DPR:$Dd, DPR:$Dm, DPR:$Dn, pred:$p)>;
8898 def : NEONInstAlias<"vcle${p}.u8 $Dd, $Dn, $Dm",
8899 (VCGEuv8i8 DPR:$Dd, DPR:$Dm, DPR:$Dn, pred:$p)>;
8900 def : NEONInstAlias<"vcle${p}.u16 $Dd, $Dn, $Dm",
8901 (VCGEuv4i16 DPR:$Dd, DPR:$Dm, DPR:$Dn, pred:$p)>;
8902 def : NEONInstAlias<"vcle${p}.u32 $Dd, $Dn, $Dm",
8903 (VCGEuv2i32 DPR:$Dd, DPR:$Dm, DPR:$Dn, pred:$p)>;
8904 def : NEONInstAlias<"vcle${p}.f32 $Dd, $Dn, $Dm",
8905 (VCGEfd DPR:$Dd, DPR:$Dm, DPR:$Dn, pred:$p)>;
8906 let Predicates = [HasNEON, HasFullFP16] in
8907 def : NEONInstAlias<"vcle${p}.f16 $Dd, $Dn, $Dm",
8908 (VCGEhd DPR:$Dd, DPR:$Dm, DPR:$Dn, pred:$p)>;
8909 // Q-register versions.
8910 def : NEONInstAlias<"vcle${p}.s8 $Qd, $Qn, $Qm",
8911 (VCGEsv16i8 QPR:$Qd, QPR:$Qm, QPR:$Qn, pred:$p)>;
8912 def : NEONInstAlias<"vcle${p}.s16 $Qd, $Qn, $Qm",
8913 (VCGEsv8i16 QPR:$Qd, QPR:$Qm, QPR:$Qn, pred:$p)>;
8914 def : NEONInstAlias<"vcle${p}.s32 $Qd, $Qn, $Qm",
8915 (VCGEsv4i32 QPR:$Qd, QPR:$Qm, QPR:$Qn, pred:$p)>;
8916 def : NEONInstAlias<"vcle${p}.u8 $Qd, $Qn, $Qm",
8917 (VCGEuv16i8 QPR:$Qd, QPR:$Qm, QPR:$Qn, pred:$p)>;
8918 def : NEONInstAlias<"vcle${p}.u16 $Qd, $Qn, $Qm",
8919 (VCGEuv8i16 QPR:$Qd, QPR:$Qm, QPR:$Qn, pred:$p)>;
8920 def : NEONInstAlias<"vcle${p}.u32 $Qd, $Qn, $Qm",
8921 (VCGEuv4i32 QPR:$Qd, QPR:$Qm, QPR:$Qn, pred:$p)>;
8922 def : NEONInstAlias<"vcle${p}.f32 $Qd, $Qn, $Qm",
8923 (VCGEfq QPR:$Qd, QPR:$Qm, QPR:$Qn, pred:$p)>;
8924 let Predicates = [HasNEON, HasFullFP16] in
8925 def : NEONInstAlias<"vcle${p}.f16 $Qd, $Qn, $Qm",
8926 (VCGEhq QPR:$Qd, QPR:$Qm, QPR:$Qn, pred:$p)>;
8928 // VCLT (register) is an assembler alias for VCGT w/ the operands reversed.
8929 // D-register versions.
8930 def : NEONInstAlias<"vclt${p}.s8 $Dd, $Dn, $Dm",
8931 (VCGTsv8i8 DPR:$Dd, DPR:$Dm, DPR:$Dn, pred:$p)>;
8932 def : NEONInstAlias<"vclt${p}.s16 $Dd, $Dn, $Dm",
8933 (VCGTsv4i16 DPR:$Dd, DPR:$Dm, DPR:$Dn, pred:$p)>;
8934 def : NEONInstAlias<"vclt${p}.s32 $Dd, $Dn, $Dm",
8935 (VCGTsv2i32 DPR:$Dd, DPR:$Dm, DPR:$Dn, pred:$p)>;
8936 def : NEONInstAlias<"vclt${p}.u8 $Dd, $Dn, $Dm",
8937 (VCGTuv8i8 DPR:$Dd, DPR:$Dm, DPR:$Dn, pred:$p)>;
8938 def : NEONInstAlias<"vclt${p}.u16 $Dd, $Dn, $Dm",
8939 (VCGTuv4i16 DPR:$Dd, DPR:$Dm, DPR:$Dn, pred:$p)>;
8940 def : NEONInstAlias<"vclt${p}.u32 $Dd, $Dn, $Dm",
8941 (VCGTuv2i32 DPR:$Dd, DPR:$Dm, DPR:$Dn, pred:$p)>;
8942 def : NEONInstAlias<"vclt${p}.f32 $Dd, $Dn, $Dm",
8943 (VCGTfd DPR:$Dd, DPR:$Dm, DPR:$Dn, pred:$p)>;
8944 let Predicates = [HasNEON, HasFullFP16] in
8945 def : NEONInstAlias<"vclt${p}.f16 $Dd, $Dn, $Dm",
8946 (VCGThd DPR:$Dd, DPR:$Dm, DPR:$Dn, pred:$p)>;
8947 // Q-register versions.
8948 def : NEONInstAlias<"vclt${p}.s8 $Qd, $Qn, $Qm",
8949 (VCGTsv16i8 QPR:$Qd, QPR:$Qm, QPR:$Qn, pred:$p)>;
8950 def : NEONInstAlias<"vclt${p}.s16 $Qd, $Qn, $Qm",
8951 (VCGTsv8i16 QPR:$Qd, QPR:$Qm, QPR:$Qn, pred:$p)>;
8952 def : NEONInstAlias<"vclt${p}.s32 $Qd, $Qn, $Qm",
8953 (VCGTsv4i32 QPR:$Qd, QPR:$Qm, QPR:$Qn, pred:$p)>;
8954 def : NEONInstAlias<"vclt${p}.u8 $Qd, $Qn, $Qm",
8955 (VCGTuv16i8 QPR:$Qd, QPR:$Qm, QPR:$Qn, pred:$p)>;
8956 def : NEONInstAlias<"vclt${p}.u16 $Qd, $Qn, $Qm",
8957 (VCGTuv8i16 QPR:$Qd, QPR:$Qm, QPR:$Qn, pred:$p)>;
8958 def : NEONInstAlias<"vclt${p}.u32 $Qd, $Qn, $Qm",
8959 (VCGTuv4i32 QPR:$Qd, QPR:$Qm, QPR:$Qn, pred:$p)>;
8960 def : NEONInstAlias<"vclt${p}.f32 $Qd, $Qn, $Qm",
8961 (VCGTfq QPR:$Qd, QPR:$Qm, QPR:$Qn, pred:$p)>;
8962 let Predicates = [HasNEON, HasFullFP16] in
8963 def : NEONInstAlias<"vclt${p}.f16 $Qd, $Qn, $Qm",
8964 (VCGThq QPR:$Qd, QPR:$Qm, QPR:$Qn, pred:$p)>;
8966 // VSWP allows, but does not require, a type suffix.
8967 defm : NEONDTAnyInstAlias<"vswp${p}", "$Vd, $Vm",
8968 (VSWPd DPR:$Vd, DPR:$Vm, pred:$p)>;
8969 defm : NEONDTAnyInstAlias<"vswp${p}", "$Vd, $Vm",
8970 (VSWPq QPR:$Vd, QPR:$Vm, pred:$p)>;
8972 // VBIF, VBIT, and VBSL allow, but do not require, a type suffix.
8973 defm : NEONDTAnyInstAlias<"vbif${p}", "$Vd, $Vn, $Vm",
8974 (VBIFd DPR:$Vd, DPR:$Vn, DPR:$Vm, pred:$p)>;
8975 defm : NEONDTAnyInstAlias<"vbit${p}", "$Vd, $Vn, $Vm",
8976 (VBITd DPR:$Vd, DPR:$Vn, DPR:$Vm, pred:$p)>;
8977 defm : NEONDTAnyInstAlias<"vbsl${p}", "$Vd, $Vn, $Vm",
8978 (VBSLd DPR:$Vd, DPR:$Vn, DPR:$Vm, pred:$p)>;
8979 defm : NEONDTAnyInstAlias<"vbif${p}", "$Vd, $Vn, $Vm",
8980 (VBIFq QPR:$Vd, QPR:$Vn, QPR:$Vm, pred:$p)>;
8981 defm : NEONDTAnyInstAlias<"vbit${p}", "$Vd, $Vn, $Vm",
8982 (VBITq QPR:$Vd, QPR:$Vn, QPR:$Vm, pred:$p)>;
8983 defm : NEONDTAnyInstAlias<"vbsl${p}", "$Vd, $Vn, $Vm",
8984 (VBSLq QPR:$Vd, QPR:$Vn, QPR:$Vm, pred:$p)>;
8986 // "vmov Rd, #-imm" can be handled via "vmvn".
8987 def : NEONInstAlias<"vmov${p}.i32 $Vd, $imm",
8988 (VMVNv2i32 DPR:$Vd, nImmVMOVI32Neg:$imm, pred:$p)>;
8989 def : NEONInstAlias<"vmov${p}.i32 $Vd, $imm",
8990 (VMVNv4i32 QPR:$Vd, nImmVMOVI32Neg:$imm, pred:$p)>;
8991 def : NEONInstAlias<"vmvn${p}.i32 $Vd, $imm",
8992 (VMOVv2i32 DPR:$Vd, nImmVMOVI32Neg:$imm, pred:$p)>;
8993 def : NEONInstAlias<"vmvn${p}.i32 $Vd, $imm",
8994 (VMOVv4i32 QPR:$Vd, nImmVMOVI32Neg:$imm, pred:$p)>;
8996 // 'gas' compatibility aliases for quad-word instructions. Strictly speaking,
8997 // these should restrict to just the Q register variants, but the register
8998 // classes are enough to match correctly regardless, so we keep it simple
8999 // and just use MnemonicAlias.
9000 def : NEONMnemonicAlias<"vbicq", "vbic">;
9001 def : NEONMnemonicAlias<"vandq", "vand">;
9002 def : NEONMnemonicAlias<"veorq", "veor">;
9003 def : NEONMnemonicAlias<"vorrq", "vorr">;
9005 def : NEONMnemonicAlias<"vmovq", "vmov">;
9006 def : NEONMnemonicAlias<"vmvnq", "vmvn">;
9007 // Explicit versions for floating point so that the FPImm variants get
9008 // handled early. The parser gets confused otherwise.
9009 def : NEONMnemonicAlias<"vmovq.f32", "vmov.f32">;
9010 def : NEONMnemonicAlias<"vmovq.f64", "vmov.f64">;
9012 def : NEONMnemonicAlias<"vaddq", "vadd">;
9013 def : NEONMnemonicAlias<"vsubq", "vsub">;
9015 def : NEONMnemonicAlias<"vminq", "vmin">;
9016 def : NEONMnemonicAlias<"vmaxq", "vmax">;
9018 def : NEONMnemonicAlias<"vmulq", "vmul">;
9020 def : NEONMnemonicAlias<"vabsq", "vabs">;
9022 def : NEONMnemonicAlias<"vshlq", "vshl">;
9023 def : NEONMnemonicAlias<"vshrq", "vshr">;
9025 def : NEONMnemonicAlias<"vcvtq", "vcvt">;
9027 def : NEONMnemonicAlias<"vcleq", "vcle">;
9028 def : NEONMnemonicAlias<"vceqq", "vceq">;
9030 def : NEONMnemonicAlias<"vzipq", "vzip">;
9031 def : NEONMnemonicAlias<"vswpq", "vswp">;
9033 def : NEONMnemonicAlias<"vrecpeq.f32", "vrecpe.f32">;
9034 def : NEONMnemonicAlias<"vrecpeq.u32", "vrecpe.u32">;
9037 // Alias for loading floating point immediates that aren't representable
9038 // using the vmov.f32 encoding but the bitpattern is representable using
9039 // the .i32 encoding.
9040 def : NEONInstAlias<"vmov${p}.f32 $Vd, $imm",
9041 (VMOVv4i32 QPR:$Vd, nImmVMOVI32:$imm, pred:$p)>;
9042 def : NEONInstAlias<"vmov${p}.f32 $Vd, $imm",
9043 (VMOVv2i32 DPR:$Vd, nImmVMOVI32:$imm, pred:$p)>;
9045 // ARMv8.6a BFloat16 instructions.
9046 let Predicates = [HasBF16, HasNEON] in {
9047 class BF16VDOT<bits<5> op27_23, bits<2> op21_20, bit op6,
9048 dag oops, dag iops, list<dag> pattern>
9049 : N3Vnp<op27_23, op21_20, 0b1101, op6, 0, oops, iops,
9050 N3RegFrm, IIC_VDOTPROD, "", "", pattern>
9052 let DecoderNamespace = "VFPV8";
9055 class BF16VDOTS<bit Q, RegisterClass RegTy, string opc, ValueType AccumTy, ValueType InputTy>
9056 : BF16VDOT<0b11000, 0b00, Q, (outs RegTy:$dst),
9057 (ins RegTy:$Vd, RegTy:$Vn, RegTy:$Vm),
9058 [(set (AccumTy RegTy:$dst),
9059 (int_arm_neon_bfdot (AccumTy RegTy:$Vd),
9060 (InputTy RegTy:$Vn),
9061 (InputTy RegTy:$Vm)))]> {
9062 let Constraints = "$dst = $Vd";
9063 let AsmString = !strconcat(opc, ".bf16", "\t$Vd, $Vn, $Vm");
9064 let DecoderNamespace = "VFPV8";
9067 multiclass BF16VDOTI<bit Q, RegisterClass RegTy, string opc, ValueType AccumTy,
9068 ValueType InputTy, dag RHS> {
9070 def "" : BF16VDOT<0b11100, 0b00, Q, (outs RegTy:$dst),
9071 (ins RegTy:$Vd, RegTy:$Vn,
9072 DPR_VFP2:$Vm, VectorIndex32:$lane), []> {
9075 let Constraints = "$dst = $Vd";
9076 let AsmString = !strconcat(opc, ".bf16", "\t$Vd, $Vn, $Vm$lane");
9077 let DecoderNamespace = "VFPV8";
9081 (AccumTy (int_arm_neon_bfdot (AccumTy RegTy:$Vd),
9082 (InputTy RegTy:$Vn),
9083 (InputTy (bitconvert (AccumTy
9084 (ARMvduplane (AccumTy RegTy:$Vm),
9085 VectorIndex32:$lane)))))),
9086 (!cast<Instruction>(NAME) RegTy:$Vd, RegTy:$Vn, RHS, VectorIndex32:$lane)>;
9089 def BF16VDOTS_VDOTD : BF16VDOTS<0, DPR, "vdot", v2f32, v8i8>;
9090 def BF16VDOTS_VDOTQ : BF16VDOTS<1, QPR, "vdot", v4f32, v16i8>;
9092 defm BF16VDOTI_VDOTD : BF16VDOTI<0, DPR, "vdot", v2f32, v8i8, (v2f32 DPR_VFP2:$Vm)>;
9093 defm BF16VDOTI_VDOTQ : BF16VDOTI<1, QPR, "vdot", v4f32, v16i8, (EXTRACT_SUBREG QPR:$Vm, dsub_0)>;
9095 class BF16MM<bit Q, RegisterClass RegTy,
9097 : N3Vnp<0b11000, 0b00, 0b1100, Q, 0,
9098 (outs RegTy:$dst), (ins RegTy:$Vd, RegTy:$Vn, RegTy:$Vm),
9099 N3RegFrm, IIC_VDOTPROD, "", "",
9100 [(set (v4f32 QPR:$dst), (int_arm_neon_bfmmla (v4f32 QPR:$Vd),
9102 (v16i8 QPR:$Vm)))]> {
9103 let Constraints = "$dst = $Vd";
9104 let AsmString = !strconcat(opc, ".bf16", "\t$Vd, $Vn, $Vm");
9105 let DecoderNamespace = "VFPV8";
9108 def VMMLA : BF16MM<1, QPR, "vmmla">;
9110 class VBF16MALQ<bit T, string suffix, SDPatternOperator OpNode>
9111 : N3VCP8<0b00, 0b11, T, 1,
9112 (outs QPR:$dst), (ins QPR:$Vd, QPR:$Vn, QPR:$Vm),
9113 NoItinerary, "vfma" # suffix, "bf16", "$Vd, $Vn, $Vm", "",
9114 [(set (v4f32 QPR:$dst),
9115 (OpNode (v4f32 QPR:$Vd),
9117 (v16i8 QPR:$Vm)))]> {
9118 let Constraints = "$dst = $Vd";
9119 let DecoderNamespace = "VFPV8";
9122 def VBF16MALTQ: VBF16MALQ<1, "t", int_arm_neon_bfmlalt>;
9123 def VBF16MALBQ: VBF16MALQ<0, "b", int_arm_neon_bfmlalb>;
9125 multiclass VBF16MALQI<bit T, string suffix, SDPatternOperator OpNode> {
9126 def "" : N3VLaneCP8<0, 0b11, T, 1, (outs QPR:$dst),
9127 (ins QPR:$Vd, QPR:$Vn, DPR_8:$Vm, VectorIndex16:$idx),
9128 IIC_VMACD, "vfma" # suffix, "bf16", "$Vd, $Vn, $Vm$idx", "", []> {
9130 let Inst{5} = idx{1};
9131 let Inst{3} = idx{0};
9132 let Constraints = "$dst = $Vd";
9133 let DecoderNamespace = "VFPV8";
9137 (v4f32 (OpNode (v4f32 QPR:$Vd),
9139 (v16i8 (bitconvert (v8bf16 (ARMvduplane (v8bf16 QPR:$Vm),
9140 VectorIndex16:$lane)))))),
9141 (!cast<Instruction>(NAME) QPR:$Vd,
9143 (EXTRACT_SUBREG QPR:$Vm,
9144 (DSubReg_i16_reg VectorIndex16:$lane)),
9145 (SubReg_i16_lane VectorIndex16:$lane))>;
9148 defm VBF16MALTQI: VBF16MALQI<1, "t", int_arm_neon_bfmlalt>;
9149 defm VBF16MALBQI: VBF16MALQI<0, "b", int_arm_neon_bfmlalb>;
9151 def BF16_VCVT : N2V<0b11, 0b11, 0b01, 0b10, 0b01100, 1, 0,
9152 (outs DPR:$Vd), (ins QPR:$Vm),
9153 NoItinerary, "vcvt", "bf16.f32", "$Vd, $Vm", "", []>;
9155 // End of BFloat16 instructions