1 | //===-- AArch64ISelDAGToDAG.cpp - A dag to dag inst selector for AArch64 --===// |
2 | // |
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 |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file defines an instruction selector for the AArch64 target. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "AArch64MachineFunctionInfo.h" |
14 | #include "AArch64TargetMachine.h" |
15 | #include "MCTargetDesc/AArch64AddressingModes.h" |
16 | #include "llvm/ADT/APSInt.h" |
17 | #include "llvm/CodeGen/ISDOpcodes.h" |
18 | #include "llvm/CodeGen/SelectionDAGISel.h" |
19 | #include "llvm/IR/Function.h" // To access function attributes. |
20 | #include "llvm/IR/GlobalValue.h" |
21 | #include "llvm/IR/Intrinsics.h" |
22 | #include "llvm/IR/IntrinsicsAArch64.h" |
23 | #include "llvm/Support/Debug.h" |
24 | #include "llvm/Support/ErrorHandling.h" |
25 | #include "llvm/Support/KnownBits.h" |
26 | #include "llvm/Support/MathExtras.h" |
27 | #include "llvm/Support/raw_ostream.h" |
28 | |
29 | using namespace llvm; |
30 | |
31 | #define DEBUG_TYPE "aarch64-isel" |
32 | #define PASS_NAME "AArch64 Instruction Selection" |
33 | |
34 | //===--------------------------------------------------------------------===// |
35 | /// AArch64DAGToDAGISel - AArch64 specific code to select AArch64 machine |
36 | /// instructions for SelectionDAG operations. |
37 | /// |
38 | namespace { |
39 | |
40 | class AArch64DAGToDAGISel : public SelectionDAGISel { |
41 | |
42 | /// Subtarget - Keep a pointer to the AArch64Subtarget around so that we can |
43 | /// make the right decision when generating code for different targets. |
44 | const AArch64Subtarget *Subtarget; |
45 | |
46 | public: |
47 | static char ID; |
48 | |
49 | AArch64DAGToDAGISel() = delete; |
50 | |
51 | explicit AArch64DAGToDAGISel(AArch64TargetMachine &tm, |
52 | CodeGenOptLevel OptLevel) |
53 | : SelectionDAGISel(ID, tm, OptLevel), Subtarget(nullptr) {} |
54 | |
55 | bool runOnMachineFunction(MachineFunction &MF) override { |
56 | Subtarget = &MF.getSubtarget<AArch64Subtarget>(); |
57 | return SelectionDAGISel::runOnMachineFunction(MF); |
58 | } |
59 | |
60 | void Select(SDNode *Node) override; |
61 | |
62 | /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for |
63 | /// inline asm expressions. |
64 | bool SelectInlineAsmMemoryOperand(const SDValue &Op, |
65 | InlineAsm::ConstraintCode ConstraintID, |
66 | std::vector<SDValue> &OutOps) override; |
67 | |
68 | template <signed Low, signed High, signed Scale> |
69 | bool SelectRDVLImm(SDValue N, SDValue &Imm); |
70 | |
71 | bool SelectArithExtendedRegister(SDValue N, SDValue &Reg, SDValue &Shift); |
72 | bool SelectArithUXTXRegister(SDValue N, SDValue &Reg, SDValue &Shift); |
73 | bool SelectArithImmed(SDValue N, SDValue &Val, SDValue &Shift); |
74 | bool SelectNegArithImmed(SDValue N, SDValue &Val, SDValue &Shift); |
75 | bool SelectArithShiftedRegister(SDValue N, SDValue &Reg, SDValue &Shift) { |
76 | return SelectShiftedRegister(N, AllowROR: false, Reg, Shift); |
77 | } |
78 | bool SelectLogicalShiftedRegister(SDValue N, SDValue &Reg, SDValue &Shift) { |
79 | return SelectShiftedRegister(N, AllowROR: true, Reg, Shift); |
80 | } |
81 | bool SelectAddrModeIndexed7S8(SDValue N, SDValue &Base, SDValue &OffImm) { |
82 | return SelectAddrModeIndexed7S(N, Size: 1, Base, OffImm); |
83 | } |
84 | bool SelectAddrModeIndexed7S16(SDValue N, SDValue &Base, SDValue &OffImm) { |
85 | return SelectAddrModeIndexed7S(N, Size: 2, Base, OffImm); |
86 | } |
87 | bool SelectAddrModeIndexed7S32(SDValue N, SDValue &Base, SDValue &OffImm) { |
88 | return SelectAddrModeIndexed7S(N, Size: 4, Base, OffImm); |
89 | } |
90 | bool SelectAddrModeIndexed7S64(SDValue N, SDValue &Base, SDValue &OffImm) { |
91 | return SelectAddrModeIndexed7S(N, Size: 8, Base, OffImm); |
92 | } |
93 | bool SelectAddrModeIndexed7S128(SDValue N, SDValue &Base, SDValue &OffImm) { |
94 | return SelectAddrModeIndexed7S(N, Size: 16, Base, OffImm); |
95 | } |
96 | bool SelectAddrModeIndexedS9S128(SDValue N, SDValue &Base, SDValue &OffImm) { |
97 | return SelectAddrModeIndexedBitWidth(N, IsSignedImm: true, BW: 9, Size: 16, Base, OffImm); |
98 | } |
99 | bool SelectAddrModeIndexedU6S128(SDValue N, SDValue &Base, SDValue &OffImm) { |
100 | return SelectAddrModeIndexedBitWidth(N, IsSignedImm: false, BW: 6, Size: 16, Base, OffImm); |
101 | } |
102 | bool SelectAddrModeIndexed8(SDValue N, SDValue &Base, SDValue &OffImm) { |
103 | return SelectAddrModeIndexed(N, Size: 1, Base, OffImm); |
104 | } |
105 | bool SelectAddrModeIndexed16(SDValue N, SDValue &Base, SDValue &OffImm) { |
106 | return SelectAddrModeIndexed(N, Size: 2, Base, OffImm); |
107 | } |
108 | bool SelectAddrModeIndexed32(SDValue N, SDValue &Base, SDValue &OffImm) { |
109 | return SelectAddrModeIndexed(N, Size: 4, Base, OffImm); |
110 | } |
111 | bool SelectAddrModeIndexed64(SDValue N, SDValue &Base, SDValue &OffImm) { |
112 | return SelectAddrModeIndexed(N, Size: 8, Base, OffImm); |
113 | } |
114 | bool SelectAddrModeIndexed128(SDValue N, SDValue &Base, SDValue &OffImm) { |
115 | return SelectAddrModeIndexed(N, Size: 16, Base, OffImm); |
116 | } |
117 | bool SelectAddrModeUnscaled8(SDValue N, SDValue &Base, SDValue &OffImm) { |
118 | return SelectAddrModeUnscaled(N, Size: 1, Base, OffImm); |
119 | } |
120 | bool SelectAddrModeUnscaled16(SDValue N, SDValue &Base, SDValue &OffImm) { |
121 | return SelectAddrModeUnscaled(N, Size: 2, Base, OffImm); |
122 | } |
123 | bool SelectAddrModeUnscaled32(SDValue N, SDValue &Base, SDValue &OffImm) { |
124 | return SelectAddrModeUnscaled(N, Size: 4, Base, OffImm); |
125 | } |
126 | bool SelectAddrModeUnscaled64(SDValue N, SDValue &Base, SDValue &OffImm) { |
127 | return SelectAddrModeUnscaled(N, Size: 8, Base, OffImm); |
128 | } |
129 | bool SelectAddrModeUnscaled128(SDValue N, SDValue &Base, SDValue &OffImm) { |
130 | return SelectAddrModeUnscaled(N, Size: 16, Base, OffImm); |
131 | } |
132 | template <unsigned Size, unsigned Max> |
133 | bool SelectAddrModeIndexedUImm(SDValue N, SDValue &Base, SDValue &OffImm) { |
134 | // Test if there is an appropriate addressing mode and check if the |
135 | // immediate fits. |
136 | bool Found = SelectAddrModeIndexed(N, Size, Base, OffImm); |
137 | if (Found) { |
138 | if (auto *CI = dyn_cast<ConstantSDNode>(OffImm)) { |
139 | int64_t C = CI->getSExtValue(); |
140 | if (C <= Max) |
141 | return true; |
142 | } |
143 | } |
144 | |
145 | // Otherwise, base only, materialize address in register. |
146 | Base = N; |
147 | OffImm = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i64); |
148 | return true; |
149 | } |
150 | |
151 | template<int Width> |
152 | bool SelectAddrModeWRO(SDValue N, SDValue &Base, SDValue &Offset, |
153 | SDValue &SignExtend, SDValue &DoShift) { |
154 | return SelectAddrModeWRO(N, Size: Width / 8, Base, Offset, SignExtend, DoShift); |
155 | } |
156 | |
157 | template<int Width> |
158 | bool SelectAddrModeXRO(SDValue N, SDValue &Base, SDValue &Offset, |
159 | SDValue &SignExtend, SDValue &DoShift) { |
160 | return SelectAddrModeXRO(N, Size: Width / 8, Base, Offset, SignExtend, DoShift); |
161 | } |
162 | |
163 | bool (SDValue N, SDValue &Res) { |
164 | if (Subtarget->isLittleEndian() && N->getOpcode() == ISD::BITCAST) |
165 | N = N->getOperand(Num: 0); |
166 | if (N->getOpcode() != ISD::EXTRACT_SUBVECTOR || |
167 | !isa<ConstantSDNode>(N->getOperand(Num: 1))) |
168 | return false; |
169 | EVT VT = N->getValueType(ResNo: 0); |
170 | EVT LVT = N->getOperand(Num: 0).getValueType(); |
171 | unsigned Index = N->getConstantOperandVal(Num: 1); |
172 | if (!VT.is64BitVector() || !LVT.is128BitVector() || |
173 | Index != VT.getVectorNumElements()) |
174 | return false; |
175 | Res = N->getOperand(Num: 0); |
176 | return true; |
177 | } |
178 | |
179 | bool SelectRoundingVLShr(SDValue N, SDValue &Res1, SDValue &Res2) { |
180 | if (N.getOpcode() != AArch64ISD::VLSHR) |
181 | return false; |
182 | SDValue Op = N->getOperand(Num: 0); |
183 | EVT VT = Op.getValueType(); |
184 | unsigned ShtAmt = N->getConstantOperandVal(Num: 1); |
185 | if (ShtAmt > VT.getScalarSizeInBits() / 2 || Op.getOpcode() != ISD::ADD) |
186 | return false; |
187 | |
188 | APInt Imm; |
189 | if (Op.getOperand(i: 1).getOpcode() == AArch64ISD::MOVIshift) |
190 | Imm = APInt(VT.getScalarSizeInBits(), |
191 | Op.getOperand(i: 1).getConstantOperandVal(i: 0) |
192 | << Op.getOperand(i: 1).getConstantOperandVal(i: 1)); |
193 | else if (Op.getOperand(i: 1).getOpcode() == AArch64ISD::DUP && |
194 | isa<ConstantSDNode>(Op.getOperand(i: 1).getOperand(i: 0))) |
195 | Imm = APInt(VT.getScalarSizeInBits(), |
196 | Op.getOperand(i: 1).getConstantOperandVal(i: 0)); |
197 | else |
198 | return false; |
199 | |
200 | if (Imm != 1ULL << (ShtAmt - 1)) |
201 | return false; |
202 | |
203 | Res1 = Op.getOperand(i: 0); |
204 | Res2 = CurDAG->getTargetConstant(ShtAmt, SDLoc(N), MVT::i32); |
205 | return true; |
206 | } |
207 | |
208 | bool SelectDupZeroOrUndef(SDValue N) { |
209 | switch(N->getOpcode()) { |
210 | case ISD::UNDEF: |
211 | return true; |
212 | case AArch64ISD::DUP: |
213 | case ISD::SPLAT_VECTOR: { |
214 | auto Opnd0 = N->getOperand(Num: 0); |
215 | if (isNullConstant(V: Opnd0)) |
216 | return true; |
217 | if (isNullFPConstant(V: Opnd0)) |
218 | return true; |
219 | break; |
220 | } |
221 | default: |
222 | break; |
223 | } |
224 | |
225 | return false; |
226 | } |
227 | |
228 | bool SelectDupZero(SDValue N) { |
229 | switch(N->getOpcode()) { |
230 | case AArch64ISD::DUP: |
231 | case ISD::SPLAT_VECTOR: { |
232 | auto Opnd0 = N->getOperand(Num: 0); |
233 | if (isNullConstant(V: Opnd0)) |
234 | return true; |
235 | if (isNullFPConstant(V: Opnd0)) |
236 | return true; |
237 | break; |
238 | } |
239 | } |
240 | |
241 | return false; |
242 | } |
243 | |
244 | bool SelectDupNegativeZero(SDValue N) { |
245 | switch(N->getOpcode()) { |
246 | case AArch64ISD::DUP: |
247 | case ISD::SPLAT_VECTOR: { |
248 | ConstantFPSDNode *Const = dyn_cast<ConstantFPSDNode>(N->getOperand(Num: 0)); |
249 | return Const && Const->isZero() && Const->isNegative(); |
250 | } |
251 | } |
252 | |
253 | return false; |
254 | } |
255 | |
256 | template<MVT::SimpleValueType VT> |
257 | bool SelectSVEAddSubImm(SDValue N, SDValue &Imm, SDValue &Shift) { |
258 | return SelectSVEAddSubImm(N, VT, Imm, Shift); |
259 | } |
260 | |
261 | template <MVT::SimpleValueType VT, bool Negate> |
262 | bool SelectSVEAddSubSSatImm(SDValue N, SDValue &Imm, SDValue &Shift) { |
263 | return SelectSVEAddSubSSatImm(N, VT, Imm, Shift, Negate); |
264 | } |
265 | |
266 | template <MVT::SimpleValueType VT> |
267 | bool SelectSVECpyDupImm(SDValue N, SDValue &Imm, SDValue &Shift) { |
268 | return SelectSVECpyDupImm(N, VT, Imm, Shift); |
269 | } |
270 | |
271 | template <MVT::SimpleValueType VT, bool Invert = false> |
272 | bool SelectSVELogicalImm(SDValue N, SDValue &Imm) { |
273 | return SelectSVELogicalImm(N, VT, Imm, Invert); |
274 | } |
275 | |
276 | template <MVT::SimpleValueType VT> |
277 | bool SelectSVEArithImm(SDValue N, SDValue &Imm) { |
278 | return SelectSVEArithImm(N, VT, Imm); |
279 | } |
280 | |
281 | template <unsigned Low, unsigned High, bool AllowSaturation = false> |
282 | bool SelectSVEShiftImm(SDValue N, SDValue &Imm) { |
283 | return SelectSVEShiftImm(N, Low, High, AllowSaturation, Imm); |
284 | } |
285 | |
286 | bool SelectSVEShiftSplatImmR(SDValue N, SDValue &Imm) { |
287 | if (N->getOpcode() != ISD::SPLAT_VECTOR) |
288 | return false; |
289 | |
290 | EVT EltVT = N->getValueType(ResNo: 0).getVectorElementType(); |
291 | return SelectSVEShiftImm(N: N->getOperand(Num: 0), /* Low */ 1, |
292 | /* High */ EltVT.getFixedSizeInBits(), |
293 | /* AllowSaturation */ true, Imm); |
294 | } |
295 | |
296 | // Returns a suitable CNT/INC/DEC/RDVL multiplier to calculate VSCALE*N. |
297 | template<signed Min, signed Max, signed Scale, bool Shift> |
298 | bool SelectCntImm(SDValue N, SDValue &Imm) { |
299 | if (!isa<ConstantSDNode>(N)) |
300 | return false; |
301 | |
302 | int64_t MulImm = cast<ConstantSDNode>(N)->getSExtValue(); |
303 | if (Shift) |
304 | MulImm = 1LL << MulImm; |
305 | |
306 | if ((MulImm % std::abs(x: Scale)) != 0) |
307 | return false; |
308 | |
309 | MulImm /= Scale; |
310 | if ((MulImm >= Min) && (MulImm <= Max)) { |
311 | Imm = CurDAG->getTargetConstant(MulImm, SDLoc(N), MVT::i32); |
312 | return true; |
313 | } |
314 | |
315 | return false; |
316 | } |
317 | |
318 | template <signed Max, signed Scale> |
319 | bool SelectEXTImm(SDValue N, SDValue &Imm) { |
320 | if (!isa<ConstantSDNode>(N)) |
321 | return false; |
322 | |
323 | int64_t MulImm = cast<ConstantSDNode>(N)->getSExtValue(); |
324 | |
325 | if (MulImm >= 0 && MulImm <= Max) { |
326 | MulImm *= Scale; |
327 | Imm = CurDAG->getTargetConstant(MulImm, SDLoc(N), MVT::i32); |
328 | return true; |
329 | } |
330 | |
331 | return false; |
332 | } |
333 | |
334 | template <unsigned BaseReg, unsigned Max> |
335 | bool ImmToReg(SDValue N, SDValue &Imm) { |
336 | if (auto *CI = dyn_cast<ConstantSDNode>(N)) { |
337 | uint64_t C = CI->getZExtValue(); |
338 | |
339 | if (C > Max) |
340 | return false; |
341 | |
342 | Imm = CurDAG->getRegister(BaseReg + C, MVT::Other); |
343 | return true; |
344 | } |
345 | return false; |
346 | } |
347 | |
348 | /// Form sequences of consecutive 64/128-bit registers for use in NEON |
349 | /// instructions making use of a vector-list (e.g. ldN, tbl). Vecs must have |
350 | /// between 1 and 4 elements. If it contains a single element that is returned |
351 | /// unchanged; otherwise a REG_SEQUENCE value is returned. |
352 | SDValue createDTuple(ArrayRef<SDValue> Vecs); |
353 | SDValue createQTuple(ArrayRef<SDValue> Vecs); |
354 | // Form a sequence of SVE registers for instructions using list of vectors, |
355 | // e.g. structured loads and stores (ldN, stN). |
356 | SDValue createZTuple(ArrayRef<SDValue> Vecs); |
357 | |
358 | // Similar to above, except the register must start at a multiple of the |
359 | // tuple, e.g. z2 for a 2-tuple, or z8 for a 4-tuple. |
360 | SDValue createZMulTuple(ArrayRef<SDValue> Regs); |
361 | |
362 | /// Generic helper for the createDTuple/createQTuple |
363 | /// functions. Those should almost always be called instead. |
364 | SDValue createTuple(ArrayRef<SDValue> Vecs, const unsigned RegClassIDs[], |
365 | const unsigned SubRegs[]); |
366 | |
367 | void SelectTable(SDNode *N, unsigned NumVecs, unsigned Opc, bool isExt); |
368 | |
369 | bool tryIndexedLoad(SDNode *N); |
370 | |
371 | bool trySelectStackSlotTagP(SDNode *N); |
372 | void SelectTagP(SDNode *N); |
373 | |
374 | void SelectLoad(SDNode *N, unsigned NumVecs, unsigned Opc, |
375 | unsigned SubRegIdx); |
376 | void SelectPostLoad(SDNode *N, unsigned NumVecs, unsigned Opc, |
377 | unsigned SubRegIdx); |
378 | void SelectLoadLane(SDNode *N, unsigned NumVecs, unsigned Opc); |
379 | void SelectPostLoadLane(SDNode *N, unsigned NumVecs, unsigned Opc); |
380 | void SelectPredicatedLoad(SDNode *N, unsigned NumVecs, unsigned Scale, |
381 | unsigned Opc_rr, unsigned Opc_ri, |
382 | bool IsIntr = false); |
383 | void SelectContiguousMultiVectorLoad(SDNode *N, unsigned NumVecs, |
384 | unsigned Scale, unsigned Opc_ri, |
385 | unsigned Opc_rr); |
386 | void SelectDestructiveMultiIntrinsic(SDNode *N, unsigned NumVecs, |
387 | bool IsZmMulti, unsigned Opcode, |
388 | bool HasPred = false); |
389 | void SelectPExtPair(SDNode *N, unsigned Opc); |
390 | void SelectWhilePair(SDNode *N, unsigned Opc); |
391 | void SelectCVTIntrinsic(SDNode *N, unsigned NumVecs, unsigned Opcode); |
392 | void SelectClamp(SDNode *N, unsigned NumVecs, unsigned Opcode); |
393 | void SelectUnaryMultiIntrinsic(SDNode *N, unsigned NumOutVecs, |
394 | bool IsTupleInput, unsigned Opc); |
395 | void SelectFrintFromVT(SDNode *N, unsigned NumVecs, unsigned Opcode); |
396 | |
397 | template <unsigned MaxIdx, unsigned Scale> |
398 | void SelectMultiVectorMove(SDNode *N, unsigned NumVecs, unsigned BaseReg, |
399 | unsigned Op); |
400 | |
401 | bool SelectAddrModeFrameIndexSVE(SDValue N, SDValue &Base, SDValue &OffImm); |
402 | /// SVE Reg+Imm addressing mode. |
403 | template <int64_t Min, int64_t Max> |
404 | bool SelectAddrModeIndexedSVE(SDNode *Root, SDValue N, SDValue &Base, |
405 | SDValue &OffImm); |
406 | /// SVE Reg+Reg address mode. |
407 | template <unsigned Scale> |
408 | bool SelectSVERegRegAddrMode(SDValue N, SDValue &Base, SDValue &Offset) { |
409 | return SelectSVERegRegAddrMode(N, Scale, Base, Offset); |
410 | } |
411 | |
412 | void SelectMultiVectorLuti(SDNode *Node, unsigned NumOutVecs, unsigned Opc, |
413 | uint32_t MaxImm); |
414 | |
415 | template <unsigned MaxIdx, unsigned Scale> |
416 | bool SelectSMETileSlice(SDValue N, SDValue &Vector, SDValue &Offset) { |
417 | return SelectSMETileSlice(N, MaxSize: MaxIdx, Vector, Offset, Scale); |
418 | } |
419 | |
420 | void SelectStore(SDNode *N, unsigned NumVecs, unsigned Opc); |
421 | void SelectPostStore(SDNode *N, unsigned NumVecs, unsigned Opc); |
422 | void SelectStoreLane(SDNode *N, unsigned NumVecs, unsigned Opc); |
423 | void SelectPostStoreLane(SDNode *N, unsigned NumVecs, unsigned Opc); |
424 | void SelectPredicatedStore(SDNode *N, unsigned NumVecs, unsigned Scale, |
425 | unsigned Opc_rr, unsigned Opc_ri); |
426 | std::tuple<unsigned, SDValue, SDValue> |
427 | findAddrModeSVELoadStore(SDNode *N, unsigned Opc_rr, unsigned Opc_ri, |
428 | const SDValue &OldBase, const SDValue &OldOffset, |
429 | unsigned Scale); |
430 | |
431 | bool tryBitfieldExtractOp(SDNode *N); |
432 | bool tryBitfieldExtractOpFromSExt(SDNode *N); |
433 | bool tryBitfieldInsertOp(SDNode *N); |
434 | bool tryBitfieldInsertInZeroOp(SDNode *N); |
435 | bool tryShiftAmountMod(SDNode *N); |
436 | |
437 | bool tryReadRegister(SDNode *N); |
438 | bool tryWriteRegister(SDNode *N); |
439 | |
440 | bool trySelectCastFixedLengthToScalableVector(SDNode *N); |
441 | bool trySelectCastScalableToFixedLengthVector(SDNode *N); |
442 | |
443 | bool trySelectXAR(SDNode *N); |
444 | |
445 | // Include the pieces autogenerated from the target description. |
446 | #include "AArch64GenDAGISel.inc" |
447 | |
448 | private: |
449 | bool SelectShiftedRegister(SDValue N, bool AllowROR, SDValue &Reg, |
450 | SDValue &Shift); |
451 | bool SelectShiftedRegisterFromAnd(SDValue N, SDValue &Reg, SDValue &Shift); |
452 | bool SelectAddrModeIndexed7S(SDValue N, unsigned Size, SDValue &Base, |
453 | SDValue &OffImm) { |
454 | return SelectAddrModeIndexedBitWidth(N, IsSignedImm: true, BW: 7, Size, Base, OffImm); |
455 | } |
456 | bool SelectAddrModeIndexedBitWidth(SDValue N, bool IsSignedImm, unsigned BW, |
457 | unsigned Size, SDValue &Base, |
458 | SDValue &OffImm); |
459 | bool SelectAddrModeIndexed(SDValue N, unsigned Size, SDValue &Base, |
460 | SDValue &OffImm); |
461 | bool SelectAddrModeUnscaled(SDValue N, unsigned Size, SDValue &Base, |
462 | SDValue &OffImm); |
463 | bool SelectAddrModeWRO(SDValue N, unsigned Size, SDValue &Base, |
464 | SDValue &Offset, SDValue &SignExtend, |
465 | SDValue &DoShift); |
466 | bool SelectAddrModeXRO(SDValue N, unsigned Size, SDValue &Base, |
467 | SDValue &Offset, SDValue &SignExtend, |
468 | SDValue &DoShift); |
469 | bool isWorthFoldingALU(SDValue V, bool LSL = false) const; |
470 | bool isWorthFoldingAddr(SDValue V, unsigned Size) const; |
471 | bool SelectExtendedSHL(SDValue N, unsigned Size, bool WantExtend, |
472 | SDValue &Offset, SDValue &SignExtend); |
473 | |
474 | template<unsigned RegWidth> |
475 | bool SelectCVTFixedPosOperand(SDValue N, SDValue &FixedPos) { |
476 | return SelectCVTFixedPosOperand(N, FixedPos, Width: RegWidth); |
477 | } |
478 | |
479 | bool SelectCVTFixedPosOperand(SDValue N, SDValue &FixedPos, unsigned Width); |
480 | |
481 | template<unsigned RegWidth> |
482 | bool SelectCVTFixedPosRecipOperand(SDValue N, SDValue &FixedPos) { |
483 | return SelectCVTFixedPosRecipOperand(N, FixedPos, Width: RegWidth); |
484 | } |
485 | |
486 | bool SelectCVTFixedPosRecipOperand(SDValue N, SDValue &FixedPos, |
487 | unsigned Width); |
488 | |
489 | bool SelectCMP_SWAP(SDNode *N); |
490 | |
491 | bool SelectSVEAddSubImm(SDValue N, MVT VT, SDValue &Imm, SDValue &Shift); |
492 | bool SelectSVEAddSubSSatImm(SDValue N, MVT VT, SDValue &Imm, SDValue &Shift, |
493 | bool Negate); |
494 | bool SelectSVECpyDupImm(SDValue N, MVT VT, SDValue &Imm, SDValue &Shift); |
495 | bool SelectSVELogicalImm(SDValue N, MVT VT, SDValue &Imm, bool Invert); |
496 | |
497 | bool SelectSVESignedArithImm(SDValue N, SDValue &Imm); |
498 | bool SelectSVEShiftImm(SDValue N, uint64_t Low, uint64_t High, |
499 | bool AllowSaturation, SDValue &Imm); |
500 | |
501 | bool SelectSVEArithImm(SDValue N, MVT VT, SDValue &Imm); |
502 | bool SelectSVERegRegAddrMode(SDValue N, unsigned Scale, SDValue &Base, |
503 | SDValue &Offset); |
504 | bool SelectSMETileSlice(SDValue N, unsigned MaxSize, SDValue &Vector, |
505 | SDValue &Offset, unsigned Scale = 1); |
506 | |
507 | bool SelectAllActivePredicate(SDValue N); |
508 | bool SelectAnyPredicate(SDValue N); |
509 | }; |
510 | } // end anonymous namespace |
511 | |
512 | char AArch64DAGToDAGISel::ID = 0; |
513 | |
514 | INITIALIZE_PASS(AArch64DAGToDAGISel, DEBUG_TYPE, PASS_NAME, false, false) |
515 | |
516 | /// isIntImmediate - This method tests to see if the node is a constant |
517 | /// operand. If so Imm will receive the 32-bit value. |
518 | static bool isIntImmediate(const SDNode *N, uint64_t &Imm) { |
519 | if (const ConstantSDNode *C = dyn_cast<const ConstantSDNode>(Val: N)) { |
520 | Imm = C->getZExtValue(); |
521 | return true; |
522 | } |
523 | return false; |
524 | } |
525 | |
526 | // isIntImmediate - This method tests to see if a constant operand. |
527 | // If so Imm will receive the value. |
528 | static bool isIntImmediate(SDValue N, uint64_t &Imm) { |
529 | return isIntImmediate(N: N.getNode(), Imm); |
530 | } |
531 | |
532 | // isOpcWithIntImmediate - This method tests to see if the node is a specific |
533 | // opcode and that it has a immediate integer right operand. |
534 | // If so Imm will receive the 32 bit value. |
535 | static bool isOpcWithIntImmediate(const SDNode *N, unsigned Opc, |
536 | uint64_t &Imm) { |
537 | return N->getOpcode() == Opc && |
538 | isIntImmediate(N: N->getOperand(Num: 1).getNode(), Imm); |
539 | } |
540 | |
541 | // isIntImmediateEq - This method tests to see if N is a constant operand that |
542 | // is equivalent to 'ImmExpected'. |
543 | #ifndef NDEBUG |
544 | static bool isIntImmediateEq(SDValue N, const uint64_t ImmExpected) { |
545 | uint64_t Imm; |
546 | if (!isIntImmediate(N: N.getNode(), Imm)) |
547 | return false; |
548 | return Imm == ImmExpected; |
549 | } |
550 | #endif |
551 | |
552 | bool AArch64DAGToDAGISel::SelectInlineAsmMemoryOperand( |
553 | const SDValue &Op, const InlineAsm::ConstraintCode ConstraintID, |
554 | std::vector<SDValue> &OutOps) { |
555 | switch(ConstraintID) { |
556 | default: |
557 | llvm_unreachable("Unexpected asm memory constraint" ); |
558 | case InlineAsm::ConstraintCode::m: |
559 | case InlineAsm::ConstraintCode::o: |
560 | case InlineAsm::ConstraintCode::Q: |
561 | // We need to make sure that this one operand does not end up in XZR, thus |
562 | // require the address to be in a PointerRegClass register. |
563 | const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo(); |
564 | const TargetRegisterClass *TRC = TRI->getPointerRegClass(MF: *MF); |
565 | SDLoc dl(Op); |
566 | SDValue RC = CurDAG->getTargetConstant(TRC->getID(), dl, MVT::i64); |
567 | SDValue NewOp = |
568 | SDValue(CurDAG->getMachineNode(Opcode: TargetOpcode::COPY_TO_REGCLASS, |
569 | dl, VT: Op.getValueType(), |
570 | Op1: Op, Op2: RC), 0); |
571 | OutOps.push_back(x: NewOp); |
572 | return false; |
573 | } |
574 | return true; |
575 | } |
576 | |
577 | /// SelectArithImmed - Select an immediate value that can be represented as |
578 | /// a 12-bit value shifted left by either 0 or 12. If so, return true with |
579 | /// Val set to the 12-bit value and Shift set to the shifter operand. |
580 | bool AArch64DAGToDAGISel::SelectArithImmed(SDValue N, SDValue &Val, |
581 | SDValue &Shift) { |
582 | // This function is called from the addsub_shifted_imm ComplexPattern, |
583 | // which lists [imm] as the list of opcode it's interested in, however |
584 | // we still need to check whether the operand is actually an immediate |
585 | // here because the ComplexPattern opcode list is only used in |
586 | // root-level opcode matching. |
587 | if (!isa<ConstantSDNode>(Val: N.getNode())) |
588 | return false; |
589 | |
590 | uint64_t Immed = N.getNode()->getAsZExtVal(); |
591 | unsigned ShiftAmt; |
592 | |
593 | if (Immed >> 12 == 0) { |
594 | ShiftAmt = 0; |
595 | } else if ((Immed & 0xfff) == 0 && Immed >> 24 == 0) { |
596 | ShiftAmt = 12; |
597 | Immed = Immed >> 12; |
598 | } else |
599 | return false; |
600 | |
601 | unsigned ShVal = AArch64_AM::getShifterImm(ST: AArch64_AM::LSL, Imm: ShiftAmt); |
602 | SDLoc dl(N); |
603 | Val = CurDAG->getTargetConstant(Immed, dl, MVT::i32); |
604 | Shift = CurDAG->getTargetConstant(ShVal, dl, MVT::i32); |
605 | return true; |
606 | } |
607 | |
608 | /// SelectNegArithImmed - As above, but negates the value before trying to |
609 | /// select it. |
610 | bool AArch64DAGToDAGISel::SelectNegArithImmed(SDValue N, SDValue &Val, |
611 | SDValue &Shift) { |
612 | // This function is called from the addsub_shifted_imm ComplexPattern, |
613 | // which lists [imm] as the list of opcode it's interested in, however |
614 | // we still need to check whether the operand is actually an immediate |
615 | // here because the ComplexPattern opcode list is only used in |
616 | // root-level opcode matching. |
617 | if (!isa<ConstantSDNode>(Val: N.getNode())) |
618 | return false; |
619 | |
620 | // The immediate operand must be a 24-bit zero-extended immediate. |
621 | uint64_t Immed = N.getNode()->getAsZExtVal(); |
622 | |
623 | // This negation is almost always valid, but "cmp wN, #0" and "cmn wN, #0" |
624 | // have the opposite effect on the C flag, so this pattern mustn't match under |
625 | // those circumstances. |
626 | if (Immed == 0) |
627 | return false; |
628 | |
629 | if (N.getValueType() == MVT::i32) |
630 | Immed = ~((uint32_t)Immed) + 1; |
631 | else |
632 | Immed = ~Immed + 1ULL; |
633 | if (Immed & 0xFFFFFFFFFF000000ULL) |
634 | return false; |
635 | |
636 | Immed &= 0xFFFFFFULL; |
637 | return SelectArithImmed(N: CurDAG->getConstant(Immed, SDLoc(N), MVT::i32), Val, |
638 | Shift); |
639 | } |
640 | |
641 | /// getShiftTypeForNode - Translate a shift node to the corresponding |
642 | /// ShiftType value. |
643 | static AArch64_AM::ShiftExtendType getShiftTypeForNode(SDValue N) { |
644 | switch (N.getOpcode()) { |
645 | default: |
646 | return AArch64_AM::InvalidShiftExtend; |
647 | case ISD::SHL: |
648 | return AArch64_AM::LSL; |
649 | case ISD::SRL: |
650 | return AArch64_AM::LSR; |
651 | case ISD::SRA: |
652 | return AArch64_AM::ASR; |
653 | case ISD::ROTR: |
654 | return AArch64_AM::ROR; |
655 | } |
656 | } |
657 | |
658 | /// Determine whether it is worth it to fold SHL into the addressing |
659 | /// mode. |
660 | static bool isWorthFoldingSHL(SDValue V) { |
661 | assert(V.getOpcode() == ISD::SHL && "invalid opcode" ); |
662 | // It is worth folding logical shift of up to three places. |
663 | auto *CSD = dyn_cast<ConstantSDNode>(Val: V.getOperand(i: 1)); |
664 | if (!CSD) |
665 | return false; |
666 | unsigned ShiftVal = CSD->getZExtValue(); |
667 | if (ShiftVal > 3) |
668 | return false; |
669 | |
670 | // Check if this particular node is reused in any non-memory related |
671 | // operation. If yes, do not try to fold this node into the address |
672 | // computation, since the computation will be kept. |
673 | const SDNode *Node = V.getNode(); |
674 | for (SDNode *UI : Node->uses()) |
675 | if (!isa<MemSDNode>(Val: *UI)) |
676 | for (SDNode *UII : UI->uses()) |
677 | if (!isa<MemSDNode>(Val: *UII)) |
678 | return false; |
679 | return true; |
680 | } |
681 | |
682 | /// Determine whether it is worth to fold V into an extended register addressing |
683 | /// mode. |
684 | bool AArch64DAGToDAGISel::isWorthFoldingAddr(SDValue V, unsigned Size) const { |
685 | // Trivial if we are optimizing for code size or if there is only |
686 | // one use of the value. |
687 | if (CurDAG->shouldOptForSize() || V.hasOneUse()) |
688 | return true; |
689 | |
690 | // If a subtarget has a slow shift, folding a shift into multiple loads |
691 | // costs additional micro-ops. |
692 | if (Subtarget->hasAddrLSLSlow14() && (Size == 2 || Size == 16)) |
693 | return false; |
694 | |
695 | // Check whether we're going to emit the address arithmetic anyway because |
696 | // it's used by a non-address operation. |
697 | if (V.getOpcode() == ISD::SHL && isWorthFoldingSHL(V)) |
698 | return true; |
699 | if (V.getOpcode() == ISD::ADD) { |
700 | const SDValue LHS = V.getOperand(i: 0); |
701 | const SDValue RHS = V.getOperand(i: 1); |
702 | if (LHS.getOpcode() == ISD::SHL && isWorthFoldingSHL(V: LHS)) |
703 | return true; |
704 | if (RHS.getOpcode() == ISD::SHL && isWorthFoldingSHL(V: RHS)) |
705 | return true; |
706 | } |
707 | |
708 | // It hurts otherwise, since the value will be reused. |
709 | return false; |
710 | } |
711 | |
712 | /// and (shl/srl/sra, x, c), mask --> shl (srl/sra, x, c1), c2 |
713 | /// to select more shifted register |
714 | bool AArch64DAGToDAGISel::SelectShiftedRegisterFromAnd(SDValue N, SDValue &Reg, |
715 | SDValue &Shift) { |
716 | EVT VT = N.getValueType(); |
717 | if (VT != MVT::i32 && VT != MVT::i64) |
718 | return false; |
719 | |
720 | if (N->getOpcode() != ISD::AND || !N->hasOneUse()) |
721 | return false; |
722 | SDValue LHS = N.getOperand(i: 0); |
723 | if (!LHS->hasOneUse()) |
724 | return false; |
725 | |
726 | unsigned LHSOpcode = LHS->getOpcode(); |
727 | if (LHSOpcode != ISD::SHL && LHSOpcode != ISD::SRL && LHSOpcode != ISD::SRA) |
728 | return false; |
729 | |
730 | ConstantSDNode *ShiftAmtNode = dyn_cast<ConstantSDNode>(Val: LHS.getOperand(i: 1)); |
731 | if (!ShiftAmtNode) |
732 | return false; |
733 | |
734 | uint64_t ShiftAmtC = ShiftAmtNode->getZExtValue(); |
735 | ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1)); |
736 | if (!RHSC) |
737 | return false; |
738 | |
739 | APInt AndMask = RHSC->getAPIntValue(); |
740 | unsigned LowZBits, MaskLen; |
741 | if (!AndMask.isShiftedMask(MaskIdx&: LowZBits, MaskLen)) |
742 | return false; |
743 | |
744 | unsigned BitWidth = N.getValueSizeInBits(); |
745 | SDLoc DL(LHS); |
746 | uint64_t NewShiftC; |
747 | unsigned NewShiftOp; |
748 | if (LHSOpcode == ISD::SHL) { |
749 | // LowZBits <= ShiftAmtC will fall into isBitfieldPositioningOp |
750 | // BitWidth != LowZBits + MaskLen doesn't match the pattern |
751 | if (LowZBits <= ShiftAmtC || (BitWidth != LowZBits + MaskLen)) |
752 | return false; |
753 | |
754 | NewShiftC = LowZBits - ShiftAmtC; |
755 | NewShiftOp = VT == MVT::i64 ? AArch64::UBFMXri : AArch64::UBFMWri; |
756 | } else { |
757 | if (LowZBits == 0) |
758 | return false; |
759 | |
760 | // NewShiftC >= BitWidth will fall into isBitfieldExtractOp |
761 | NewShiftC = LowZBits + ShiftAmtC; |
762 | if (NewShiftC >= BitWidth) |
763 | return false; |
764 | |
765 | // SRA need all high bits |
766 | if (LHSOpcode == ISD::SRA && (BitWidth != (LowZBits + MaskLen))) |
767 | return false; |
768 | |
769 | // SRL high bits can be 0 or 1 |
770 | if (LHSOpcode == ISD::SRL && (BitWidth > (NewShiftC + MaskLen))) |
771 | return false; |
772 | |
773 | if (LHSOpcode == ISD::SRL) |
774 | NewShiftOp = VT == MVT::i64 ? AArch64::UBFMXri : AArch64::UBFMWri; |
775 | else |
776 | NewShiftOp = VT == MVT::i64 ? AArch64::SBFMXri : AArch64::SBFMWri; |
777 | } |
778 | |
779 | assert(NewShiftC < BitWidth && "Invalid shift amount" ); |
780 | SDValue NewShiftAmt = CurDAG->getTargetConstant(Val: NewShiftC, DL, VT); |
781 | SDValue BitWidthMinus1 = CurDAG->getTargetConstant(Val: BitWidth - 1, DL, VT); |
782 | Reg = SDValue(CurDAG->getMachineNode(Opcode: NewShiftOp, dl: DL, VT, Op1: LHS->getOperand(Num: 0), |
783 | Op2: NewShiftAmt, Op3: BitWidthMinus1), |
784 | 0); |
785 | unsigned ShVal = AArch64_AM::getShifterImm(ST: AArch64_AM::LSL, Imm: LowZBits); |
786 | Shift = CurDAG->getTargetConstant(ShVal, DL, MVT::i32); |
787 | return true; |
788 | } |
789 | |
790 | /// getExtendTypeForNode - Translate an extend node to the corresponding |
791 | /// ExtendType value. |
792 | static AArch64_AM::ShiftExtendType |
793 | getExtendTypeForNode(SDValue N, bool IsLoadStore = false) { |
794 | if (N.getOpcode() == ISD::SIGN_EXTEND || |
795 | N.getOpcode() == ISD::SIGN_EXTEND_INREG) { |
796 | EVT SrcVT; |
797 | if (N.getOpcode() == ISD::SIGN_EXTEND_INREG) |
798 | SrcVT = cast<VTSDNode>(Val: N.getOperand(i: 1))->getVT(); |
799 | else |
800 | SrcVT = N.getOperand(i: 0).getValueType(); |
801 | |
802 | if (!IsLoadStore && SrcVT == MVT::i8) |
803 | return AArch64_AM::SXTB; |
804 | else if (!IsLoadStore && SrcVT == MVT::i16) |
805 | return AArch64_AM::SXTH; |
806 | else if (SrcVT == MVT::i32) |
807 | return AArch64_AM::SXTW; |
808 | assert(SrcVT != MVT::i64 && "extend from 64-bits?" ); |
809 | |
810 | return AArch64_AM::InvalidShiftExtend; |
811 | } else if (N.getOpcode() == ISD::ZERO_EXTEND || |
812 | N.getOpcode() == ISD::ANY_EXTEND) { |
813 | EVT SrcVT = N.getOperand(i: 0).getValueType(); |
814 | if (!IsLoadStore && SrcVT == MVT::i8) |
815 | return AArch64_AM::UXTB; |
816 | else if (!IsLoadStore && SrcVT == MVT::i16) |
817 | return AArch64_AM::UXTH; |
818 | else if (SrcVT == MVT::i32) |
819 | return AArch64_AM::UXTW; |
820 | assert(SrcVT != MVT::i64 && "extend from 64-bits?" ); |
821 | |
822 | return AArch64_AM::InvalidShiftExtend; |
823 | } else if (N.getOpcode() == ISD::AND) { |
824 | ConstantSDNode *CSD = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1)); |
825 | if (!CSD) |
826 | return AArch64_AM::InvalidShiftExtend; |
827 | uint64_t AndMask = CSD->getZExtValue(); |
828 | |
829 | switch (AndMask) { |
830 | default: |
831 | return AArch64_AM::InvalidShiftExtend; |
832 | case 0xFF: |
833 | return !IsLoadStore ? AArch64_AM::UXTB : AArch64_AM::InvalidShiftExtend; |
834 | case 0xFFFF: |
835 | return !IsLoadStore ? AArch64_AM::UXTH : AArch64_AM::InvalidShiftExtend; |
836 | case 0xFFFFFFFF: |
837 | return AArch64_AM::UXTW; |
838 | } |
839 | } |
840 | |
841 | return AArch64_AM::InvalidShiftExtend; |
842 | } |
843 | |
844 | /// Determine whether it is worth to fold V into an extended register of an |
845 | /// Add/Sub. LSL means we are folding into an `add w0, w1, w2, lsl #N` |
846 | /// instruction, and the shift should be treated as worth folding even if has |
847 | /// multiple uses. |
848 | bool AArch64DAGToDAGISel::isWorthFoldingALU(SDValue V, bool LSL) const { |
849 | // Trivial if we are optimizing for code size or if there is only |
850 | // one use of the value. |
851 | if (CurDAG->shouldOptForSize() || V.hasOneUse()) |
852 | return true; |
853 | |
854 | // If a subtarget has a fastpath LSL we can fold a logical shift into |
855 | // the add/sub and save a cycle. |
856 | if (LSL && Subtarget->hasALULSLFast() && V.getOpcode() == ISD::SHL && |
857 | V.getConstantOperandVal(i: 1) <= 4 && |
858 | getExtendTypeForNode(N: V.getOperand(i: 0)) == AArch64_AM::InvalidShiftExtend) |
859 | return true; |
860 | |
861 | // It hurts otherwise, since the value will be reused. |
862 | return false; |
863 | } |
864 | |
865 | /// SelectShiftedRegister - Select a "shifted register" operand. If the value |
866 | /// is not shifted, set the Shift operand to default of "LSL 0". The logical |
867 | /// instructions allow the shifted register to be rotated, but the arithmetic |
868 | /// instructions do not. The AllowROR parameter specifies whether ROR is |
869 | /// supported. |
870 | bool AArch64DAGToDAGISel::SelectShiftedRegister(SDValue N, bool AllowROR, |
871 | SDValue &Reg, SDValue &Shift) { |
872 | if (SelectShiftedRegisterFromAnd(N, Reg, Shift)) |
873 | return true; |
874 | |
875 | AArch64_AM::ShiftExtendType ShType = getShiftTypeForNode(N); |
876 | if (ShType == AArch64_AM::InvalidShiftExtend) |
877 | return false; |
878 | if (!AllowROR && ShType == AArch64_AM::ROR) |
879 | return false; |
880 | |
881 | if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1))) { |
882 | unsigned BitSize = N.getValueSizeInBits(); |
883 | unsigned Val = RHS->getZExtValue() & (BitSize - 1); |
884 | unsigned ShVal = AArch64_AM::getShifterImm(ST: ShType, Imm: Val); |
885 | |
886 | Reg = N.getOperand(i: 0); |
887 | Shift = CurDAG->getTargetConstant(ShVal, SDLoc(N), MVT::i32); |
888 | return isWorthFoldingALU(V: N, LSL: true); |
889 | } |
890 | |
891 | return false; |
892 | } |
893 | |
894 | /// Instructions that accept extend modifiers like UXTW expect the register |
895 | /// being extended to be a GPR32, but the incoming DAG might be acting on a |
896 | /// GPR64 (either via SEXT_INREG or AND). Extract the appropriate low bits if |
897 | /// this is the case. |
898 | static SDValue narrowIfNeeded(SelectionDAG *CurDAG, SDValue N) { |
899 | if (N.getValueType() == MVT::i32) |
900 | return N; |
901 | |
902 | SDLoc dl(N); |
903 | return CurDAG->getTargetExtractSubreg(AArch64::SRIdx: sub_32, DL: dl, MVT::VT: i32, Operand: N); |
904 | } |
905 | |
906 | // Returns a suitable CNT/INC/DEC/RDVL multiplier to calculate VSCALE*N. |
907 | template<signed Low, signed High, signed Scale> |
908 | bool AArch64DAGToDAGISel::SelectRDVLImm(SDValue N, SDValue &Imm) { |
909 | if (!isa<ConstantSDNode>(Val: N)) |
910 | return false; |
911 | |
912 | int64_t MulImm = cast<ConstantSDNode>(Val&: N)->getSExtValue(); |
913 | if ((MulImm % std::abs(x: Scale)) == 0) { |
914 | int64_t RDVLImm = MulImm / Scale; |
915 | if ((RDVLImm >= Low) && (RDVLImm <= High)) { |
916 | Imm = CurDAG->getTargetConstant(RDVLImm, SDLoc(N), MVT::i32); |
917 | return true; |
918 | } |
919 | } |
920 | |
921 | return false; |
922 | } |
923 | |
924 | /// SelectArithExtendedRegister - Select a "extended register" operand. This |
925 | /// operand folds in an extend followed by an optional left shift. |
926 | bool AArch64DAGToDAGISel::SelectArithExtendedRegister(SDValue N, SDValue &Reg, |
927 | SDValue &Shift) { |
928 | unsigned ShiftVal = 0; |
929 | AArch64_AM::ShiftExtendType Ext; |
930 | |
931 | if (N.getOpcode() == ISD::SHL) { |
932 | ConstantSDNode *CSD = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1)); |
933 | if (!CSD) |
934 | return false; |
935 | ShiftVal = CSD->getZExtValue(); |
936 | if (ShiftVal > 4) |
937 | return false; |
938 | |
939 | Ext = getExtendTypeForNode(N: N.getOperand(i: 0)); |
940 | if (Ext == AArch64_AM::InvalidShiftExtend) |
941 | return false; |
942 | |
943 | Reg = N.getOperand(i: 0).getOperand(i: 0); |
944 | } else { |
945 | Ext = getExtendTypeForNode(N); |
946 | if (Ext == AArch64_AM::InvalidShiftExtend) |
947 | return false; |
948 | |
949 | Reg = N.getOperand(i: 0); |
950 | |
951 | // Don't match if free 32-bit -> 64-bit zext can be used instead. Use the |
952 | // isDef32 as a heuristic for when the operand is likely to be a 32bit def. |
953 | auto isDef32 = [](SDValue N) { |
954 | unsigned Opc = N.getOpcode(); |
955 | return Opc != ISD::TRUNCATE && Opc != TargetOpcode::EXTRACT_SUBREG && |
956 | Opc != ISD::CopyFromReg && Opc != ISD::AssertSext && |
957 | Opc != ISD::AssertZext && Opc != ISD::AssertAlign && |
958 | Opc != ISD::FREEZE; |
959 | }; |
960 | if (Ext == AArch64_AM::UXTW && Reg->getValueType(ResNo: 0).getSizeInBits() == 32 && |
961 | isDef32(Reg)) |
962 | return false; |
963 | } |
964 | |
965 | // AArch64 mandates that the RHS of the operation must use the smallest |
966 | // register class that could contain the size being extended from. Thus, |
967 | // if we're folding a (sext i8), we need the RHS to be a GPR32, even though |
968 | // there might not be an actual 32-bit value in the program. We can |
969 | // (harmlessly) synthesize one by injected an EXTRACT_SUBREG here. |
970 | assert(Ext != AArch64_AM::UXTX && Ext != AArch64_AM::SXTX); |
971 | Reg = narrowIfNeeded(CurDAG, N: Reg); |
972 | Shift = CurDAG->getTargetConstant(getArithExtendImm(ET: Ext, Imm: ShiftVal), SDLoc(N), |
973 | MVT::i32); |
974 | return isWorthFoldingALU(V: N); |
975 | } |
976 | |
977 | /// SelectArithUXTXRegister - Select a "UXTX register" operand. This |
978 | /// operand is refered by the instructions have SP operand |
979 | bool AArch64DAGToDAGISel::SelectArithUXTXRegister(SDValue N, SDValue &Reg, |
980 | SDValue &Shift) { |
981 | unsigned ShiftVal = 0; |
982 | AArch64_AM::ShiftExtendType Ext; |
983 | |
984 | if (N.getOpcode() != ISD::SHL) |
985 | return false; |
986 | |
987 | ConstantSDNode *CSD = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1)); |
988 | if (!CSD) |
989 | return false; |
990 | ShiftVal = CSD->getZExtValue(); |
991 | if (ShiftVal > 4) |
992 | return false; |
993 | |
994 | Ext = AArch64_AM::UXTX; |
995 | Reg = N.getOperand(i: 0); |
996 | Shift = CurDAG->getTargetConstant(getArithExtendImm(ET: Ext, Imm: ShiftVal), SDLoc(N), |
997 | MVT::i32); |
998 | return isWorthFoldingALU(V: N); |
999 | } |
1000 | |
1001 | /// If there's a use of this ADDlow that's not itself a load/store then we'll |
1002 | /// need to create a real ADD instruction from it anyway and there's no point in |
1003 | /// folding it into the mem op. Theoretically, it shouldn't matter, but there's |
1004 | /// a single pseudo-instruction for an ADRP/ADD pair so over-aggressive folding |
1005 | /// leads to duplicated ADRP instructions. |
1006 | static bool isWorthFoldingADDlow(SDValue N) { |
1007 | for (auto *Use : N->uses()) { |
1008 | if (Use->getOpcode() != ISD::LOAD && Use->getOpcode() != ISD::STORE && |
1009 | Use->getOpcode() != ISD::ATOMIC_LOAD && |
1010 | Use->getOpcode() != ISD::ATOMIC_STORE) |
1011 | return false; |
1012 | |
1013 | // ldar and stlr have much more restrictive addressing modes (just a |
1014 | // register). |
1015 | if (isStrongerThanMonotonic(AO: cast<MemSDNode>(Val: Use)->getSuccessOrdering())) |
1016 | return false; |
1017 | } |
1018 | |
1019 | return true; |
1020 | } |
1021 | |
1022 | /// Check if the immediate offset is valid as a scaled immediate. |
1023 | static bool isValidAsScaledImmediate(int64_t Offset, unsigned Range, |
1024 | unsigned Size) { |
1025 | if ((Offset & (Size - 1)) == 0 && Offset >= 0 && |
1026 | Offset < (Range << Log2_32(Value: Size))) |
1027 | return true; |
1028 | return false; |
1029 | } |
1030 | |
1031 | /// SelectAddrModeIndexedBitWidth - Select a "register plus scaled (un)signed BW-bit |
1032 | /// immediate" address. The "Size" argument is the size in bytes of the memory |
1033 | /// reference, which determines the scale. |
1034 | bool AArch64DAGToDAGISel::SelectAddrModeIndexedBitWidth(SDValue N, bool IsSignedImm, |
1035 | unsigned BW, unsigned Size, |
1036 | SDValue &Base, |
1037 | SDValue &OffImm) { |
1038 | SDLoc dl(N); |
1039 | const DataLayout &DL = CurDAG->getDataLayout(); |
1040 | const TargetLowering *TLI = getTargetLowering(); |
1041 | if (N.getOpcode() == ISD::FrameIndex) { |
1042 | int FI = cast<FrameIndexSDNode>(Val&: N)->getIndex(); |
1043 | Base = CurDAG->getTargetFrameIndex(FI, VT: TLI->getPointerTy(DL)); |
1044 | OffImm = CurDAG->getTargetConstant(0, dl, MVT::i64); |
1045 | return true; |
1046 | } |
1047 | |
1048 | // As opposed to the (12-bit) Indexed addressing mode below, the 7/9-bit signed |
1049 | // selected here doesn't support labels/immediates, only base+offset. |
1050 | if (CurDAG->isBaseWithConstantOffset(Op: N)) { |
1051 | if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1))) { |
1052 | if (IsSignedImm) { |
1053 | int64_t RHSC = RHS->getSExtValue(); |
1054 | unsigned Scale = Log2_32(Value: Size); |
1055 | int64_t Range = 0x1LL << (BW - 1); |
1056 | |
1057 | if ((RHSC & (Size - 1)) == 0 && RHSC >= -(Range << Scale) && |
1058 | RHSC < (Range << Scale)) { |
1059 | Base = N.getOperand(i: 0); |
1060 | if (Base.getOpcode() == ISD::FrameIndex) { |
1061 | int FI = cast<FrameIndexSDNode>(Val&: Base)->getIndex(); |
1062 | Base = CurDAG->getTargetFrameIndex(FI, VT: TLI->getPointerTy(DL)); |
1063 | } |
1064 | OffImm = CurDAG->getTargetConstant(RHSC >> Scale, dl, MVT::i64); |
1065 | return true; |
1066 | } |
1067 | } else { |
1068 | // unsigned Immediate |
1069 | uint64_t RHSC = RHS->getZExtValue(); |
1070 | unsigned Scale = Log2_32(Value: Size); |
1071 | uint64_t Range = 0x1ULL << BW; |
1072 | |
1073 | if ((RHSC & (Size - 1)) == 0 && RHSC < (Range << Scale)) { |
1074 | Base = N.getOperand(i: 0); |
1075 | if (Base.getOpcode() == ISD::FrameIndex) { |
1076 | int FI = cast<FrameIndexSDNode>(Val&: Base)->getIndex(); |
1077 | Base = CurDAG->getTargetFrameIndex(FI, VT: TLI->getPointerTy(DL)); |
1078 | } |
1079 | OffImm = CurDAG->getTargetConstant(RHSC >> Scale, dl, MVT::i64); |
1080 | return true; |
1081 | } |
1082 | } |
1083 | } |
1084 | } |
1085 | // Base only. The address will be materialized into a register before |
1086 | // the memory is accessed. |
1087 | // add x0, Xbase, #offset |
1088 | // stp x1, x2, [x0] |
1089 | Base = N; |
1090 | OffImm = CurDAG->getTargetConstant(0, dl, MVT::i64); |
1091 | return true; |
1092 | } |
1093 | |
1094 | /// SelectAddrModeIndexed - Select a "register plus scaled unsigned 12-bit |
1095 | /// immediate" address. The "Size" argument is the size in bytes of the memory |
1096 | /// reference, which determines the scale. |
1097 | bool AArch64DAGToDAGISel::SelectAddrModeIndexed(SDValue N, unsigned Size, |
1098 | SDValue &Base, SDValue &OffImm) { |
1099 | SDLoc dl(N); |
1100 | const DataLayout &DL = CurDAG->getDataLayout(); |
1101 | const TargetLowering *TLI = getTargetLowering(); |
1102 | if (N.getOpcode() == ISD::FrameIndex) { |
1103 | int FI = cast<FrameIndexSDNode>(Val&: N)->getIndex(); |
1104 | Base = CurDAG->getTargetFrameIndex(FI, VT: TLI->getPointerTy(DL)); |
1105 | OffImm = CurDAG->getTargetConstant(0, dl, MVT::i64); |
1106 | return true; |
1107 | } |
1108 | |
1109 | if (N.getOpcode() == AArch64ISD::ADDlow && isWorthFoldingADDlow(N)) { |
1110 | GlobalAddressSDNode *GAN = |
1111 | dyn_cast<GlobalAddressSDNode>(Val: N.getOperand(i: 1).getNode()); |
1112 | Base = N.getOperand(i: 0); |
1113 | OffImm = N.getOperand(i: 1); |
1114 | if (!GAN) |
1115 | return true; |
1116 | |
1117 | if (GAN->getOffset() % Size == 0 && |
1118 | GAN->getGlobal()->getPointerAlignment(DL) >= Size) |
1119 | return true; |
1120 | } |
1121 | |
1122 | if (CurDAG->isBaseWithConstantOffset(Op: N)) { |
1123 | if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1))) { |
1124 | int64_t RHSC = (int64_t)RHS->getZExtValue(); |
1125 | unsigned Scale = Log2_32(Value: Size); |
1126 | if (isValidAsScaledImmediate(Offset: RHSC, Range: 0x1000, Size)) { |
1127 | Base = N.getOperand(i: 0); |
1128 | if (Base.getOpcode() == ISD::FrameIndex) { |
1129 | int FI = cast<FrameIndexSDNode>(Val&: Base)->getIndex(); |
1130 | Base = CurDAG->getTargetFrameIndex(FI, VT: TLI->getPointerTy(DL)); |
1131 | } |
1132 | OffImm = CurDAG->getTargetConstant(RHSC >> Scale, dl, MVT::i64); |
1133 | return true; |
1134 | } |
1135 | } |
1136 | } |
1137 | |
1138 | // Before falling back to our general case, check if the unscaled |
1139 | // instructions can handle this. If so, that's preferable. |
1140 | if (SelectAddrModeUnscaled(N, Size, Base, OffImm)) |
1141 | return false; |
1142 | |
1143 | // Base only. The address will be materialized into a register before |
1144 | // the memory is accessed. |
1145 | // add x0, Xbase, #offset |
1146 | // ldr x0, [x0] |
1147 | Base = N; |
1148 | OffImm = CurDAG->getTargetConstant(0, dl, MVT::i64); |
1149 | return true; |
1150 | } |
1151 | |
1152 | /// SelectAddrModeUnscaled - Select a "register plus unscaled signed 9-bit |
1153 | /// immediate" address. This should only match when there is an offset that |
1154 | /// is not valid for a scaled immediate addressing mode. The "Size" argument |
1155 | /// is the size in bytes of the memory reference, which is needed here to know |
1156 | /// what is valid for a scaled immediate. |
1157 | bool AArch64DAGToDAGISel::SelectAddrModeUnscaled(SDValue N, unsigned Size, |
1158 | SDValue &Base, |
1159 | SDValue &OffImm) { |
1160 | if (!CurDAG->isBaseWithConstantOffset(Op: N)) |
1161 | return false; |
1162 | if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1))) { |
1163 | int64_t RHSC = RHS->getSExtValue(); |
1164 | if (RHSC >= -256 && RHSC < 256) { |
1165 | Base = N.getOperand(i: 0); |
1166 | if (Base.getOpcode() == ISD::FrameIndex) { |
1167 | int FI = cast<FrameIndexSDNode>(Val&: Base)->getIndex(); |
1168 | const TargetLowering *TLI = getTargetLowering(); |
1169 | Base = CurDAG->getTargetFrameIndex( |
1170 | FI, VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
1171 | } |
1172 | OffImm = CurDAG->getTargetConstant(RHSC, SDLoc(N), MVT::i64); |
1173 | return true; |
1174 | } |
1175 | } |
1176 | return false; |
1177 | } |
1178 | |
1179 | static SDValue Widen(SelectionDAG *CurDAG, SDValue N) { |
1180 | SDLoc dl(N); |
1181 | SDValue ImpDef = SDValue( |
1182 | CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, MVT::i64), 0); |
1183 | return CurDAG->getTargetInsertSubreg(AArch64::SRIdx: sub_32, DL: dl, MVT::VT: i64, Operand: ImpDef, |
1184 | Subreg: N); |
1185 | } |
1186 | |
1187 | /// Check if the given SHL node (\p N), can be used to form an |
1188 | /// extended register for an addressing mode. |
1189 | bool AArch64DAGToDAGISel::SelectExtendedSHL(SDValue N, unsigned Size, |
1190 | bool WantExtend, SDValue &Offset, |
1191 | SDValue &SignExtend) { |
1192 | assert(N.getOpcode() == ISD::SHL && "Invalid opcode." ); |
1193 | ConstantSDNode *CSD = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1)); |
1194 | if (!CSD || (CSD->getZExtValue() & 0x7) != CSD->getZExtValue()) |
1195 | return false; |
1196 | |
1197 | SDLoc dl(N); |
1198 | if (WantExtend) { |
1199 | AArch64_AM::ShiftExtendType Ext = |
1200 | getExtendTypeForNode(N: N.getOperand(i: 0), IsLoadStore: true); |
1201 | if (Ext == AArch64_AM::InvalidShiftExtend) |
1202 | return false; |
1203 | |
1204 | Offset = narrowIfNeeded(CurDAG, N: N.getOperand(i: 0).getOperand(i: 0)); |
1205 | SignExtend = CurDAG->getTargetConstant(Ext == AArch64_AM::SXTW, dl, |
1206 | MVT::i32); |
1207 | } else { |
1208 | Offset = N.getOperand(i: 0); |
1209 | SignExtend = CurDAG->getTargetConstant(0, dl, MVT::i32); |
1210 | } |
1211 | |
1212 | unsigned LegalShiftVal = Log2_32(Value: Size); |
1213 | unsigned ShiftVal = CSD->getZExtValue(); |
1214 | |
1215 | if (ShiftVal != 0 && ShiftVal != LegalShiftVal) |
1216 | return false; |
1217 | |
1218 | return isWorthFoldingAddr(V: N, Size); |
1219 | } |
1220 | |
1221 | bool AArch64DAGToDAGISel::SelectAddrModeWRO(SDValue N, unsigned Size, |
1222 | SDValue &Base, SDValue &Offset, |
1223 | SDValue &SignExtend, |
1224 | SDValue &DoShift) { |
1225 | if (N.getOpcode() != ISD::ADD) |
1226 | return false; |
1227 | SDValue LHS = N.getOperand(i: 0); |
1228 | SDValue RHS = N.getOperand(i: 1); |
1229 | SDLoc dl(N); |
1230 | |
1231 | // We don't want to match immediate adds here, because they are better lowered |
1232 | // to the register-immediate addressing modes. |
1233 | if (isa<ConstantSDNode>(Val: LHS) || isa<ConstantSDNode>(Val: RHS)) |
1234 | return false; |
1235 | |
1236 | // Check if this particular node is reused in any non-memory related |
1237 | // operation. If yes, do not try to fold this node into the address |
1238 | // computation, since the computation will be kept. |
1239 | const SDNode *Node = N.getNode(); |
1240 | for (SDNode *UI : Node->uses()) { |
1241 | if (!isa<MemSDNode>(Val: *UI)) |
1242 | return false; |
1243 | } |
1244 | |
1245 | // Remember if it is worth folding N when it produces extended register. |
1246 | bool IsExtendedRegisterWorthFolding = isWorthFoldingAddr(V: N, Size); |
1247 | |
1248 | // Try to match a shifted extend on the RHS. |
1249 | if (IsExtendedRegisterWorthFolding && RHS.getOpcode() == ISD::SHL && |
1250 | SelectExtendedSHL(N: RHS, Size, WantExtend: true, Offset, SignExtend)) { |
1251 | Base = LHS; |
1252 | DoShift = CurDAG->getTargetConstant(true, dl, MVT::i32); |
1253 | return true; |
1254 | } |
1255 | |
1256 | // Try to match a shifted extend on the LHS. |
1257 | if (IsExtendedRegisterWorthFolding && LHS.getOpcode() == ISD::SHL && |
1258 | SelectExtendedSHL(N: LHS, Size, WantExtend: true, Offset, SignExtend)) { |
1259 | Base = RHS; |
1260 | DoShift = CurDAG->getTargetConstant(true, dl, MVT::i32); |
1261 | return true; |
1262 | } |
1263 | |
1264 | // There was no shift, whatever else we find. |
1265 | DoShift = CurDAG->getTargetConstant(false, dl, MVT::i32); |
1266 | |
1267 | AArch64_AM::ShiftExtendType Ext = AArch64_AM::InvalidShiftExtend; |
1268 | // Try to match an unshifted extend on the LHS. |
1269 | if (IsExtendedRegisterWorthFolding && |
1270 | (Ext = getExtendTypeForNode(N: LHS, IsLoadStore: true)) != |
1271 | AArch64_AM::InvalidShiftExtend) { |
1272 | Base = RHS; |
1273 | Offset = narrowIfNeeded(CurDAG, N: LHS.getOperand(i: 0)); |
1274 | SignExtend = CurDAG->getTargetConstant(Ext == AArch64_AM::SXTW, dl, |
1275 | MVT::i32); |
1276 | if (isWorthFoldingAddr(V: LHS, Size)) |
1277 | return true; |
1278 | } |
1279 | |
1280 | // Try to match an unshifted extend on the RHS. |
1281 | if (IsExtendedRegisterWorthFolding && |
1282 | (Ext = getExtendTypeForNode(N: RHS, IsLoadStore: true)) != |
1283 | AArch64_AM::InvalidShiftExtend) { |
1284 | Base = LHS; |
1285 | Offset = narrowIfNeeded(CurDAG, N: RHS.getOperand(i: 0)); |
1286 | SignExtend = CurDAG->getTargetConstant(Ext == AArch64_AM::SXTW, dl, |
1287 | MVT::i32); |
1288 | if (isWorthFoldingAddr(V: RHS, Size)) |
1289 | return true; |
1290 | } |
1291 | |
1292 | return false; |
1293 | } |
1294 | |
1295 | // Check if the given immediate is preferred by ADD. If an immediate can be |
1296 | // encoded in an ADD, or it can be encoded in an "ADD LSL #12" and can not be |
1297 | // encoded by one MOVZ, return true. |
1298 | static bool isPreferredADD(int64_t ImmOff) { |
1299 | // Constant in [0x0, 0xfff] can be encoded in ADD. |
1300 | if ((ImmOff & 0xfffffffffffff000LL) == 0x0LL) |
1301 | return true; |
1302 | // Check if it can be encoded in an "ADD LSL #12". |
1303 | if ((ImmOff & 0xffffffffff000fffLL) == 0x0LL) |
1304 | // As a single MOVZ is faster than a "ADD of LSL #12", ignore such constant. |
1305 | return (ImmOff & 0xffffffffff00ffffLL) != 0x0LL && |
1306 | (ImmOff & 0xffffffffffff0fffLL) != 0x0LL; |
1307 | return false; |
1308 | } |
1309 | |
1310 | bool AArch64DAGToDAGISel::SelectAddrModeXRO(SDValue N, unsigned Size, |
1311 | SDValue &Base, SDValue &Offset, |
1312 | SDValue &SignExtend, |
1313 | SDValue &DoShift) { |
1314 | if (N.getOpcode() != ISD::ADD) |
1315 | return false; |
1316 | SDValue LHS = N.getOperand(i: 0); |
1317 | SDValue RHS = N.getOperand(i: 1); |
1318 | SDLoc DL(N); |
1319 | |
1320 | // Check if this particular node is reused in any non-memory related |
1321 | // operation. If yes, do not try to fold this node into the address |
1322 | // computation, since the computation will be kept. |
1323 | const SDNode *Node = N.getNode(); |
1324 | for (SDNode *UI : Node->uses()) { |
1325 | if (!isa<MemSDNode>(Val: *UI)) |
1326 | return false; |
1327 | } |
1328 | |
1329 | // Watch out if RHS is a wide immediate, it can not be selected into |
1330 | // [BaseReg+Imm] addressing mode. Also it may not be able to be encoded into |
1331 | // ADD/SUB. Instead it will use [BaseReg + 0] address mode and generate |
1332 | // instructions like: |
1333 | // MOV X0, WideImmediate |
1334 | // ADD X1, BaseReg, X0 |
1335 | // LDR X2, [X1, 0] |
1336 | // For such situation, using [BaseReg, XReg] addressing mode can save one |
1337 | // ADD/SUB: |
1338 | // MOV X0, WideImmediate |
1339 | // LDR X2, [BaseReg, X0] |
1340 | if (isa<ConstantSDNode>(Val: RHS)) { |
1341 | int64_t ImmOff = (int64_t)RHS->getAsZExtVal(); |
1342 | // Skip the immediate can be selected by load/store addressing mode. |
1343 | // Also skip the immediate can be encoded by a single ADD (SUB is also |
1344 | // checked by using -ImmOff). |
1345 | if (isValidAsScaledImmediate(Offset: ImmOff, Range: 0x1000, Size) || |
1346 | isPreferredADD(ImmOff) || isPreferredADD(ImmOff: -ImmOff)) |
1347 | return false; |
1348 | |
1349 | SDValue Ops[] = { RHS }; |
1350 | SDNode *MOVI = |
1351 | CurDAG->getMachineNode(AArch64::MOVi64imm, DL, MVT::i64, Ops); |
1352 | SDValue MOVIV = SDValue(MOVI, 0); |
1353 | // This ADD of two X register will be selected into [Reg+Reg] mode. |
1354 | N = CurDAG->getNode(ISD::ADD, DL, MVT::i64, LHS, MOVIV); |
1355 | } |
1356 | |
1357 | // Remember if it is worth folding N when it produces extended register. |
1358 | bool IsExtendedRegisterWorthFolding = isWorthFoldingAddr(V: N, Size); |
1359 | |
1360 | // Try to match a shifted extend on the RHS. |
1361 | if (IsExtendedRegisterWorthFolding && RHS.getOpcode() == ISD::SHL && |
1362 | SelectExtendedSHL(N: RHS, Size, WantExtend: false, Offset, SignExtend)) { |
1363 | Base = LHS; |
1364 | DoShift = CurDAG->getTargetConstant(true, DL, MVT::i32); |
1365 | return true; |
1366 | } |
1367 | |
1368 | // Try to match a shifted extend on the LHS. |
1369 | if (IsExtendedRegisterWorthFolding && LHS.getOpcode() == ISD::SHL && |
1370 | SelectExtendedSHL(N: LHS, Size, WantExtend: false, Offset, SignExtend)) { |
1371 | Base = RHS; |
1372 | DoShift = CurDAG->getTargetConstant(true, DL, MVT::i32); |
1373 | return true; |
1374 | } |
1375 | |
1376 | // Match any non-shifted, non-extend, non-immediate add expression. |
1377 | Base = LHS; |
1378 | Offset = RHS; |
1379 | SignExtend = CurDAG->getTargetConstant(false, DL, MVT::i32); |
1380 | DoShift = CurDAG->getTargetConstant(false, DL, MVT::i32); |
1381 | // Reg1 + Reg2 is free: no check needed. |
1382 | return true; |
1383 | } |
1384 | |
1385 | SDValue AArch64DAGToDAGISel::createDTuple(ArrayRef<SDValue> Regs) { |
1386 | static const unsigned RegClassIDs[] = { |
1387 | AArch64::DDRegClassID, AArch64::DDDRegClassID, AArch64::DDDDRegClassID}; |
1388 | static const unsigned SubRegs[] = {AArch64::dsub0, AArch64::dsub1, |
1389 | AArch64::dsub2, AArch64::dsub3}; |
1390 | |
1391 | return createTuple(Vecs: Regs, RegClassIDs, SubRegs); |
1392 | } |
1393 | |
1394 | SDValue AArch64DAGToDAGISel::createQTuple(ArrayRef<SDValue> Regs) { |
1395 | static const unsigned RegClassIDs[] = { |
1396 | AArch64::QQRegClassID, AArch64::QQQRegClassID, AArch64::QQQQRegClassID}; |
1397 | static const unsigned SubRegs[] = {AArch64::qsub0, AArch64::qsub1, |
1398 | AArch64::qsub2, AArch64::qsub3}; |
1399 | |
1400 | return createTuple(Vecs: Regs, RegClassIDs, SubRegs); |
1401 | } |
1402 | |
1403 | SDValue AArch64DAGToDAGISel::createZTuple(ArrayRef<SDValue> Regs) { |
1404 | static const unsigned RegClassIDs[] = {AArch64::ZPR2RegClassID, |
1405 | AArch64::ZPR3RegClassID, |
1406 | AArch64::ZPR4RegClassID}; |
1407 | static const unsigned SubRegs[] = {AArch64::zsub0, AArch64::zsub1, |
1408 | AArch64::zsub2, AArch64::zsub3}; |
1409 | |
1410 | return createTuple(Vecs: Regs, RegClassIDs, SubRegs); |
1411 | } |
1412 | |
1413 | SDValue AArch64DAGToDAGISel::createZMulTuple(ArrayRef<SDValue> Regs) { |
1414 | assert(Regs.size() == 2 || Regs.size() == 4); |
1415 | |
1416 | // The createTuple interface requires 3 RegClassIDs for each possible |
1417 | // tuple type even though we only have them for ZPR2 and ZPR4. |
1418 | static const unsigned RegClassIDs[] = {AArch64::ZPR2Mul2RegClassID, 0, |
1419 | AArch64::ZPR4Mul4RegClassID}; |
1420 | static const unsigned SubRegs[] = {AArch64::zsub0, AArch64::zsub1, |
1421 | AArch64::zsub2, AArch64::zsub3}; |
1422 | return createTuple(Vecs: Regs, RegClassIDs, SubRegs); |
1423 | } |
1424 | |
1425 | SDValue AArch64DAGToDAGISel::createTuple(ArrayRef<SDValue> Regs, |
1426 | const unsigned RegClassIDs[], |
1427 | const unsigned SubRegs[]) { |
1428 | // There's no special register-class for a vector-list of 1 element: it's just |
1429 | // a vector. |
1430 | if (Regs.size() == 1) |
1431 | return Regs[0]; |
1432 | |
1433 | assert(Regs.size() >= 2 && Regs.size() <= 4); |
1434 | |
1435 | SDLoc DL(Regs[0]); |
1436 | |
1437 | SmallVector<SDValue, 4> Ops; |
1438 | |
1439 | // First operand of REG_SEQUENCE is the desired RegClass. |
1440 | Ops.push_back( |
1441 | CurDAG->getTargetConstant(RegClassIDs[Regs.size() - 2], DL, MVT::i32)); |
1442 | |
1443 | // Then we get pairs of source & subregister-position for the components. |
1444 | for (unsigned i = 0; i < Regs.size(); ++i) { |
1445 | Ops.push_back(Elt: Regs[i]); |
1446 | Ops.push_back(CurDAG->getTargetConstant(SubRegs[i], DL, MVT::i32)); |
1447 | } |
1448 | |
1449 | SDNode *N = |
1450 | CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, DL, MVT::Untyped, Ops); |
1451 | return SDValue(N, 0); |
1452 | } |
1453 | |
1454 | void AArch64DAGToDAGISel::SelectTable(SDNode *N, unsigned NumVecs, unsigned Opc, |
1455 | bool isExt) { |
1456 | SDLoc dl(N); |
1457 | EVT VT = N->getValueType(ResNo: 0); |
1458 | |
1459 | unsigned ExtOff = isExt; |
1460 | |
1461 | // Form a REG_SEQUENCE to force register allocation. |
1462 | unsigned Vec0Off = ExtOff + 1; |
1463 | SmallVector<SDValue, 4> Regs(N->op_begin() + Vec0Off, |
1464 | N->op_begin() + Vec0Off + NumVecs); |
1465 | SDValue RegSeq = createQTuple(Regs); |
1466 | |
1467 | SmallVector<SDValue, 6> Ops; |
1468 | if (isExt) |
1469 | Ops.push_back(Elt: N->getOperand(Num: 1)); |
1470 | Ops.push_back(Elt: RegSeq); |
1471 | Ops.push_back(Elt: N->getOperand(Num: NumVecs + ExtOff + 1)); |
1472 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: Opc, dl, VT, Ops)); |
1473 | } |
1474 | |
1475 | bool AArch64DAGToDAGISel::tryIndexedLoad(SDNode *N) { |
1476 | LoadSDNode *LD = cast<LoadSDNode>(Val: N); |
1477 | if (LD->isUnindexed()) |
1478 | return false; |
1479 | EVT VT = LD->getMemoryVT(); |
1480 | EVT DstVT = N->getValueType(ResNo: 0); |
1481 | ISD::MemIndexedMode AM = LD->getAddressingMode(); |
1482 | bool IsPre = AM == ISD::PRE_INC || AM == ISD::PRE_DEC; |
1483 | |
1484 | // We're not doing validity checking here. That was done when checking |
1485 | // if we should mark the load as indexed or not. We're just selecting |
1486 | // the right instruction. |
1487 | unsigned Opcode = 0; |
1488 | |
1489 | ISD::LoadExtType ExtType = LD->getExtensionType(); |
1490 | bool InsertTo64 = false; |
1491 | if (VT == MVT::i64) |
1492 | Opcode = IsPre ? AArch64::LDRXpre : AArch64::LDRXpost; |
1493 | else if (VT == MVT::i32) { |
1494 | if (ExtType == ISD::NON_EXTLOAD) |
1495 | Opcode = IsPre ? AArch64::LDRWpre : AArch64::LDRWpost; |
1496 | else if (ExtType == ISD::SEXTLOAD) |
1497 | Opcode = IsPre ? AArch64::LDRSWpre : AArch64::LDRSWpost; |
1498 | else { |
1499 | Opcode = IsPre ? AArch64::LDRWpre : AArch64::LDRWpost; |
1500 | InsertTo64 = true; |
1501 | // The result of the load is only i32. It's the subreg_to_reg that makes |
1502 | // it into an i64. |
1503 | DstVT = MVT::i32; |
1504 | } |
1505 | } else if (VT == MVT::i16) { |
1506 | if (ExtType == ISD::SEXTLOAD) { |
1507 | if (DstVT == MVT::i64) |
1508 | Opcode = IsPre ? AArch64::LDRSHXpre : AArch64::LDRSHXpost; |
1509 | else |
1510 | Opcode = IsPre ? AArch64::LDRSHWpre : AArch64::LDRSHWpost; |
1511 | } else { |
1512 | Opcode = IsPre ? AArch64::LDRHHpre : AArch64::LDRHHpost; |
1513 | InsertTo64 = DstVT == MVT::i64; |
1514 | // The result of the load is only i32. It's the subreg_to_reg that makes |
1515 | // it into an i64. |
1516 | DstVT = MVT::i32; |
1517 | } |
1518 | } else if (VT == MVT::i8) { |
1519 | if (ExtType == ISD::SEXTLOAD) { |
1520 | if (DstVT == MVT::i64) |
1521 | Opcode = IsPre ? AArch64::LDRSBXpre : AArch64::LDRSBXpost; |
1522 | else |
1523 | Opcode = IsPre ? AArch64::LDRSBWpre : AArch64::LDRSBWpost; |
1524 | } else { |
1525 | Opcode = IsPre ? AArch64::LDRBBpre : AArch64::LDRBBpost; |
1526 | InsertTo64 = DstVT == MVT::i64; |
1527 | // The result of the load is only i32. It's the subreg_to_reg that makes |
1528 | // it into an i64. |
1529 | DstVT = MVT::i32; |
1530 | } |
1531 | } else if (VT == MVT::f16) { |
1532 | Opcode = IsPre ? AArch64::LDRHpre : AArch64::LDRHpost; |
1533 | } else if (VT == MVT::bf16) { |
1534 | Opcode = IsPre ? AArch64::LDRHpre : AArch64::LDRHpost; |
1535 | } else if (VT == MVT::f32) { |
1536 | Opcode = IsPre ? AArch64::LDRSpre : AArch64::LDRSpost; |
1537 | } else if (VT == MVT::f64 || VT.is64BitVector()) { |
1538 | Opcode = IsPre ? AArch64::LDRDpre : AArch64::LDRDpost; |
1539 | } else if (VT.is128BitVector()) { |
1540 | Opcode = IsPre ? AArch64::LDRQpre : AArch64::LDRQpost; |
1541 | } else |
1542 | return false; |
1543 | SDValue Chain = LD->getChain(); |
1544 | SDValue Base = LD->getBasePtr(); |
1545 | ConstantSDNode *OffsetOp = cast<ConstantSDNode>(Val: LD->getOffset()); |
1546 | int OffsetVal = (int)OffsetOp->getZExtValue(); |
1547 | SDLoc dl(N); |
1548 | SDValue Offset = CurDAG->getTargetConstant(OffsetVal, dl, MVT::i64); |
1549 | SDValue Ops[] = { Base, Offset, Chain }; |
1550 | SDNode *Res = CurDAG->getMachineNode(Opcode, dl, MVT::i64, DstVT, |
1551 | MVT::Other, Ops); |
1552 | |
1553 | // Transfer memoperands. |
1554 | MachineMemOperand *MemOp = cast<MemSDNode>(Val: N)->getMemOperand(); |
1555 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: Res), NewMemRefs: {MemOp}); |
1556 | |
1557 | // Either way, we're replacing the node, so tell the caller that. |
1558 | SDValue LoadedVal = SDValue(Res, 1); |
1559 | if (InsertTo64) { |
1560 | SDValue SubReg = CurDAG->getTargetConstant(AArch64::sub_32, dl, MVT::i32); |
1561 | LoadedVal = |
1562 | SDValue(CurDAG->getMachineNode( |
1563 | AArch64::SUBREG_TO_REG, dl, MVT::i64, |
1564 | CurDAG->getTargetConstant(0, dl, MVT::i64), LoadedVal, |
1565 | SubReg), |
1566 | 0); |
1567 | } |
1568 | |
1569 | ReplaceUses(F: SDValue(N, 0), T: LoadedVal); |
1570 | ReplaceUses(F: SDValue(N, 1), T: SDValue(Res, 0)); |
1571 | ReplaceUses(F: SDValue(N, 2), T: SDValue(Res, 2)); |
1572 | CurDAG->RemoveDeadNode(N); |
1573 | return true; |
1574 | } |
1575 | |
1576 | void AArch64DAGToDAGISel::SelectLoad(SDNode *N, unsigned NumVecs, unsigned Opc, |
1577 | unsigned SubRegIdx) { |
1578 | SDLoc dl(N); |
1579 | EVT VT = N->getValueType(ResNo: 0); |
1580 | SDValue Chain = N->getOperand(Num: 0); |
1581 | |
1582 | SDValue Ops[] = {N->getOperand(Num: 2), // Mem operand; |
1583 | Chain}; |
1584 | |
1585 | const EVT ResTys[] = {MVT::Untyped, MVT::Other}; |
1586 | |
1587 | SDNode *Ld = CurDAG->getMachineNode(Opc, dl, ResTys, Ops); |
1588 | SDValue SuperReg = SDValue(Ld, 0); |
1589 | for (unsigned i = 0; i < NumVecs; ++i) |
1590 | ReplaceUses(F: SDValue(N, i), |
1591 | T: CurDAG->getTargetExtractSubreg(SRIdx: SubRegIdx + i, DL: dl, VT, Operand: SuperReg)); |
1592 | |
1593 | ReplaceUses(F: SDValue(N, NumVecs), T: SDValue(Ld, 1)); |
1594 | |
1595 | // Transfer memoperands. In the case of AArch64::LD64B, there won't be one, |
1596 | // because it's too simple to have needed special treatment during lowering. |
1597 | if (auto *MemIntr = dyn_cast<MemIntrinsicSDNode>(Val: N)) { |
1598 | MachineMemOperand *MemOp = MemIntr->getMemOperand(); |
1599 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: Ld), NewMemRefs: {MemOp}); |
1600 | } |
1601 | |
1602 | CurDAG->RemoveDeadNode(N); |
1603 | } |
1604 | |
1605 | void AArch64DAGToDAGISel::SelectPostLoad(SDNode *N, unsigned NumVecs, |
1606 | unsigned Opc, unsigned SubRegIdx) { |
1607 | SDLoc dl(N); |
1608 | EVT VT = N->getValueType(ResNo: 0); |
1609 | SDValue Chain = N->getOperand(Num: 0); |
1610 | |
1611 | SDValue Ops[] = {N->getOperand(Num: 1), // Mem operand |
1612 | N->getOperand(Num: 2), // Incremental |
1613 | Chain}; |
1614 | |
1615 | const EVT ResTys[] = {MVT::i64, // Type of the write back register |
1616 | MVT::Untyped, MVT::Other}; |
1617 | |
1618 | SDNode *Ld = CurDAG->getMachineNode(Opc, dl, ResTys, Ops); |
1619 | |
1620 | // Update uses of write back register |
1621 | ReplaceUses(F: SDValue(N, NumVecs), T: SDValue(Ld, 0)); |
1622 | |
1623 | // Update uses of vector list |
1624 | SDValue SuperReg = SDValue(Ld, 1); |
1625 | if (NumVecs == 1) |
1626 | ReplaceUses(F: SDValue(N, 0), T: SuperReg); |
1627 | else |
1628 | for (unsigned i = 0; i < NumVecs; ++i) |
1629 | ReplaceUses(F: SDValue(N, i), |
1630 | T: CurDAG->getTargetExtractSubreg(SRIdx: SubRegIdx + i, DL: dl, VT, Operand: SuperReg)); |
1631 | |
1632 | // Update the chain |
1633 | ReplaceUses(F: SDValue(N, NumVecs + 1), T: SDValue(Ld, 2)); |
1634 | CurDAG->RemoveDeadNode(N); |
1635 | } |
1636 | |
1637 | /// Optimize \param OldBase and \param OldOffset selecting the best addressing |
1638 | /// mode. Returns a tuple consisting of an Opcode, an SDValue representing the |
1639 | /// new Base and an SDValue representing the new offset. |
1640 | std::tuple<unsigned, SDValue, SDValue> |
1641 | AArch64DAGToDAGISel::findAddrModeSVELoadStore(SDNode *N, unsigned Opc_rr, |
1642 | unsigned Opc_ri, |
1643 | const SDValue &OldBase, |
1644 | const SDValue &OldOffset, |
1645 | unsigned Scale) { |
1646 | SDValue NewBase = OldBase; |
1647 | SDValue NewOffset = OldOffset; |
1648 | // Detect a possible Reg+Imm addressing mode. |
1649 | const bool IsRegImm = SelectAddrModeIndexedSVE</*Min=*/-8, /*Max=*/7>( |
1650 | Root: N, N: OldBase, Base&: NewBase, OffImm&: NewOffset); |
1651 | |
1652 | // Detect a possible reg+reg addressing mode, but only if we haven't already |
1653 | // detected a Reg+Imm one. |
1654 | const bool IsRegReg = |
1655 | !IsRegImm && SelectSVERegRegAddrMode(N: OldBase, Scale, Base&: NewBase, Offset&: NewOffset); |
1656 | |
1657 | // Select the instruction. |
1658 | return std::make_tuple(args&: IsRegReg ? Opc_rr : Opc_ri, args&: NewBase, args&: NewOffset); |
1659 | } |
1660 | |
1661 | enum class SelectTypeKind { |
1662 | Int1 = 0, |
1663 | Int = 1, |
1664 | FP = 2, |
1665 | AnyType = 3, |
1666 | }; |
1667 | |
1668 | /// This function selects an opcode from a list of opcodes, which is |
1669 | /// expected to be the opcode for { 8-bit, 16-bit, 32-bit, 64-bit } |
1670 | /// element types, in this order. |
1671 | template <SelectTypeKind Kind> |
1672 | static unsigned SelectOpcodeFromVT(EVT VT, ArrayRef<unsigned> Opcodes) { |
1673 | // Only match scalable vector VTs |
1674 | if (!VT.isScalableVector()) |
1675 | return 0; |
1676 | |
1677 | EVT EltVT = VT.getVectorElementType(); |
1678 | switch (Kind) { |
1679 | case SelectTypeKind::AnyType: |
1680 | break; |
1681 | case SelectTypeKind::Int: |
1682 | if (EltVT != MVT::i8 && EltVT != MVT::i16 && EltVT != MVT::i32 && |
1683 | EltVT != MVT::i64) |
1684 | return 0; |
1685 | break; |
1686 | case SelectTypeKind::Int1: |
1687 | if (EltVT != MVT::i1) |
1688 | return 0; |
1689 | break; |
1690 | case SelectTypeKind::FP: |
1691 | if (EltVT != MVT::f16 && EltVT != MVT::f32 && EltVT != MVT::f64) |
1692 | return 0; |
1693 | break; |
1694 | } |
1695 | |
1696 | unsigned Offset; |
1697 | switch (VT.getVectorMinNumElements()) { |
1698 | case 16: // 8-bit |
1699 | Offset = 0; |
1700 | break; |
1701 | case 8: // 16-bit |
1702 | Offset = 1; |
1703 | break; |
1704 | case 4: // 32-bit |
1705 | Offset = 2; |
1706 | break; |
1707 | case 2: // 64-bit |
1708 | Offset = 3; |
1709 | break; |
1710 | default: |
1711 | return 0; |
1712 | } |
1713 | |
1714 | return (Opcodes.size() <= Offset) ? 0 : Opcodes[Offset]; |
1715 | } |
1716 | |
1717 | // This function is almost identical to SelectWhilePair, but has an |
1718 | // extra check on the range of the immediate operand. |
1719 | // TODO: Merge these two functions together at some point? |
1720 | void AArch64DAGToDAGISel::SelectPExtPair(SDNode *N, unsigned Opc) { |
1721 | // Immediate can be either 0 or 1. |
1722 | if (ConstantSDNode *Imm = dyn_cast<ConstantSDNode>(Val: N->getOperand(Num: 2))) |
1723 | if (Imm->getZExtValue() > 1) |
1724 | return; |
1725 | |
1726 | SDLoc DL(N); |
1727 | EVT VT = N->getValueType(ResNo: 0); |
1728 | SDValue Ops[] = {N->getOperand(Num: 1), N->getOperand(Num: 2)}; |
1729 | SDNode *WhilePair = CurDAG->getMachineNode(Opc, DL, MVT::Untyped, Ops); |
1730 | SDValue SuperReg = SDValue(WhilePair, 0); |
1731 | |
1732 | for (unsigned I = 0; I < 2; ++I) |
1733 | ReplaceUses(SDValue(N, I), CurDAG->getTargetExtractSubreg( |
1734 | AArch64::psub0 + I, DL, VT, SuperReg)); |
1735 | |
1736 | CurDAG->RemoveDeadNode(N); |
1737 | } |
1738 | |
1739 | void AArch64DAGToDAGISel::SelectWhilePair(SDNode *N, unsigned Opc) { |
1740 | SDLoc DL(N); |
1741 | EVT VT = N->getValueType(ResNo: 0); |
1742 | |
1743 | SDValue Ops[] = {N->getOperand(Num: 1), N->getOperand(Num: 2)}; |
1744 | |
1745 | SDNode *WhilePair = CurDAG->getMachineNode(Opc, DL, MVT::Untyped, Ops); |
1746 | SDValue SuperReg = SDValue(WhilePair, 0); |
1747 | |
1748 | for (unsigned I = 0; I < 2; ++I) |
1749 | ReplaceUses(SDValue(N, I), CurDAG->getTargetExtractSubreg( |
1750 | AArch64::psub0 + I, DL, VT, SuperReg)); |
1751 | |
1752 | CurDAG->RemoveDeadNode(N); |
1753 | } |
1754 | |
1755 | void AArch64DAGToDAGISel::SelectCVTIntrinsic(SDNode *N, unsigned NumVecs, |
1756 | unsigned Opcode) { |
1757 | EVT VT = N->getValueType(ResNo: 0); |
1758 | SmallVector<SDValue, 4> Regs(N->op_begin() + 1, N->op_begin() + 1 + NumVecs); |
1759 | SDValue Ops = createZTuple(Regs); |
1760 | SDLoc DL(N); |
1761 | SDNode *Intrinsic = CurDAG->getMachineNode(Opcode, DL, MVT::Untyped, Ops); |
1762 | SDValue SuperReg = SDValue(Intrinsic, 0); |
1763 | for (unsigned i = 0; i < NumVecs; ++i) |
1764 | ReplaceUses(SDValue(N, i), CurDAG->getTargetExtractSubreg( |
1765 | AArch64::zsub0 + i, DL, VT, SuperReg)); |
1766 | |
1767 | CurDAG->RemoveDeadNode(N); |
1768 | } |
1769 | |
1770 | void AArch64DAGToDAGISel::SelectDestructiveMultiIntrinsic(SDNode *N, |
1771 | unsigned NumVecs, |
1772 | bool IsZmMulti, |
1773 | unsigned Opcode, |
1774 | bool HasPred) { |
1775 | assert(Opcode != 0 && "Unexpected opcode" ); |
1776 | |
1777 | SDLoc DL(N); |
1778 | EVT VT = N->getValueType(ResNo: 0); |
1779 | unsigned FirstVecIdx = HasPred ? 2 : 1; |
1780 | |
1781 | auto GetMultiVecOperand = [=](unsigned StartIdx) { |
1782 | SmallVector<SDValue, 4> Regs(N->op_begin() + StartIdx, |
1783 | N->op_begin() + StartIdx + NumVecs); |
1784 | return createZMulTuple(Regs); |
1785 | }; |
1786 | |
1787 | SDValue Zdn = GetMultiVecOperand(FirstVecIdx); |
1788 | |
1789 | SDValue Zm; |
1790 | if (IsZmMulti) |
1791 | Zm = GetMultiVecOperand(NumVecs + FirstVecIdx); |
1792 | else |
1793 | Zm = N->getOperand(Num: NumVecs + FirstVecIdx); |
1794 | |
1795 | SDNode *Intrinsic; |
1796 | if (HasPred) |
1797 | Intrinsic = CurDAG->getMachineNode(Opcode, DL, MVT::Untyped, |
1798 | N->getOperand(1), Zdn, Zm); |
1799 | else |
1800 | Intrinsic = CurDAG->getMachineNode(Opcode, DL, MVT::Untyped, Zdn, Zm); |
1801 | SDValue SuperReg = SDValue(Intrinsic, 0); |
1802 | for (unsigned i = 0; i < NumVecs; ++i) |
1803 | ReplaceUses(SDValue(N, i), CurDAG->getTargetExtractSubreg( |
1804 | AArch64::zsub0 + i, DL, VT, SuperReg)); |
1805 | |
1806 | CurDAG->RemoveDeadNode(N); |
1807 | } |
1808 | |
1809 | void AArch64DAGToDAGISel::SelectPredicatedLoad(SDNode *N, unsigned NumVecs, |
1810 | unsigned Scale, unsigned Opc_ri, |
1811 | unsigned Opc_rr, bool IsIntr) { |
1812 | assert(Scale < 5 && "Invalid scaling value." ); |
1813 | SDLoc DL(N); |
1814 | EVT VT = N->getValueType(ResNo: 0); |
1815 | SDValue Chain = N->getOperand(Num: 0); |
1816 | |
1817 | // Optimize addressing mode. |
1818 | SDValue Base, Offset; |
1819 | unsigned Opc; |
1820 | std::tie(Opc, Base, Offset) = findAddrModeSVELoadStore( |
1821 | N, Opc_rr, Opc_ri, N->getOperand(IsIntr ? 3 : 2), |
1822 | CurDAG->getTargetConstant(0, DL, MVT::i64), Scale); |
1823 | |
1824 | SDValue Ops[] = {N->getOperand(Num: IsIntr ? 2 : 1), // Predicate |
1825 | Base, // Memory operand |
1826 | Offset, Chain}; |
1827 | |
1828 | const EVT ResTys[] = {MVT::Untyped, MVT::Other}; |
1829 | |
1830 | SDNode *Load = CurDAG->getMachineNode(Opc, DL, ResTys, Ops); |
1831 | SDValue SuperReg = SDValue(Load, 0); |
1832 | for (unsigned i = 0; i < NumVecs; ++i) |
1833 | ReplaceUses(SDValue(N, i), CurDAG->getTargetExtractSubreg( |
1834 | AArch64::zsub0 + i, DL, VT, SuperReg)); |
1835 | |
1836 | // Copy chain |
1837 | unsigned ChainIdx = NumVecs; |
1838 | ReplaceUses(F: SDValue(N, ChainIdx), T: SDValue(Load, 1)); |
1839 | CurDAG->RemoveDeadNode(N); |
1840 | } |
1841 | |
1842 | void AArch64DAGToDAGISel::SelectContiguousMultiVectorLoad(SDNode *N, |
1843 | unsigned NumVecs, |
1844 | unsigned Scale, |
1845 | unsigned Opc_ri, |
1846 | unsigned Opc_rr) { |
1847 | assert(Scale < 4 && "Invalid scaling value." ); |
1848 | SDLoc DL(N); |
1849 | EVT VT = N->getValueType(ResNo: 0); |
1850 | SDValue Chain = N->getOperand(Num: 0); |
1851 | |
1852 | SDValue PNg = N->getOperand(Num: 2); |
1853 | SDValue Base = N->getOperand(Num: 3); |
1854 | SDValue Offset = CurDAG->getTargetConstant(0, DL, MVT::i64); |
1855 | unsigned Opc; |
1856 | std::tie(args&: Opc, args&: Base, args&: Offset) = |
1857 | findAddrModeSVELoadStore(N, Opc_rr, Opc_ri, OldBase: Base, OldOffset: Offset, Scale); |
1858 | |
1859 | SDValue Ops[] = {PNg, // Predicate-as-counter |
1860 | Base, // Memory operand |
1861 | Offset, Chain}; |
1862 | |
1863 | const EVT ResTys[] = {MVT::Untyped, MVT::Other}; |
1864 | |
1865 | SDNode *Load = CurDAG->getMachineNode(Opc, DL, ResTys, Ops); |
1866 | SDValue SuperReg = SDValue(Load, 0); |
1867 | for (unsigned i = 0; i < NumVecs; ++i) |
1868 | ReplaceUses(SDValue(N, i), CurDAG->getTargetExtractSubreg( |
1869 | AArch64::zsub0 + i, DL, VT, SuperReg)); |
1870 | |
1871 | // Copy chain |
1872 | unsigned ChainIdx = NumVecs; |
1873 | ReplaceUses(F: SDValue(N, ChainIdx), T: SDValue(Load, 1)); |
1874 | CurDAG->RemoveDeadNode(N); |
1875 | } |
1876 | |
1877 | void AArch64DAGToDAGISel::SelectFrintFromVT(SDNode *N, unsigned NumVecs, |
1878 | unsigned Opcode) { |
1879 | if (N->getValueType(0) != MVT::nxv4f32) |
1880 | return; |
1881 | SelectUnaryMultiIntrinsic(N, NumOutVecs: NumVecs, IsTupleInput: true, Opc: Opcode); |
1882 | } |
1883 | |
1884 | void AArch64DAGToDAGISel::SelectMultiVectorLuti(SDNode *Node, |
1885 | unsigned NumOutVecs, |
1886 | unsigned Opc, uint32_t MaxImm) { |
1887 | if (ConstantSDNode *Imm = dyn_cast<ConstantSDNode>(Val: Node->getOperand(Num: 4))) |
1888 | if (Imm->getZExtValue() > MaxImm) |
1889 | return; |
1890 | |
1891 | SDValue ZtValue; |
1892 | if (!ImmToReg<AArch64::ZT0, 0>(Node->getOperand(2), ZtValue)) |
1893 | return; |
1894 | SDValue Ops[] = {ZtValue, Node->getOperand(Num: 3), Node->getOperand(Num: 4)}; |
1895 | SDLoc DL(Node); |
1896 | EVT VT = Node->getValueType(ResNo: 0); |
1897 | |
1898 | SDNode *Instruction = |
1899 | CurDAG->getMachineNode(Opc, DL, {MVT::Untyped, MVT::Other}, Ops); |
1900 | SDValue SuperReg = SDValue(Instruction, 0); |
1901 | |
1902 | for (unsigned I = 0; I < NumOutVecs; ++I) |
1903 | ReplaceUses(SDValue(Node, I), CurDAG->getTargetExtractSubreg( |
1904 | AArch64::zsub0 + I, DL, VT, SuperReg)); |
1905 | |
1906 | // Copy chain |
1907 | unsigned ChainIdx = NumOutVecs; |
1908 | ReplaceUses(F: SDValue(Node, ChainIdx), T: SDValue(Instruction, 1)); |
1909 | CurDAG->RemoveDeadNode(N: Node); |
1910 | } |
1911 | |
1912 | void AArch64DAGToDAGISel::SelectClamp(SDNode *N, unsigned NumVecs, |
1913 | unsigned Op) { |
1914 | SDLoc DL(N); |
1915 | EVT VT = N->getValueType(ResNo: 0); |
1916 | |
1917 | SmallVector<SDValue, 4> Regs(N->op_begin() + 1, N->op_begin() + 1 + NumVecs); |
1918 | SDValue Zd = createZMulTuple(Regs); |
1919 | SDValue Zn = N->getOperand(Num: 1 + NumVecs); |
1920 | SDValue Zm = N->getOperand(Num: 2 + NumVecs); |
1921 | |
1922 | SDValue Ops[] = {Zd, Zn, Zm}; |
1923 | |
1924 | SDNode *Intrinsic = CurDAG->getMachineNode(Op, DL, MVT::Untyped, Ops); |
1925 | SDValue SuperReg = SDValue(Intrinsic, 0); |
1926 | for (unsigned i = 0; i < NumVecs; ++i) |
1927 | ReplaceUses(SDValue(N, i), CurDAG->getTargetExtractSubreg( |
1928 | AArch64::zsub0 + i, DL, VT, SuperReg)); |
1929 | |
1930 | CurDAG->RemoveDeadNode(N); |
1931 | } |
1932 | |
1933 | bool SelectSMETile(unsigned &BaseReg, unsigned TileNum) { |
1934 | switch (BaseReg) { |
1935 | default: |
1936 | return false; |
1937 | case AArch64::ZA: |
1938 | case AArch64::ZAB0: |
1939 | if (TileNum == 0) |
1940 | break; |
1941 | return false; |
1942 | case AArch64::ZAH0: |
1943 | if (TileNum <= 1) |
1944 | break; |
1945 | return false; |
1946 | case AArch64::ZAS0: |
1947 | if (TileNum <= 3) |
1948 | break; |
1949 | return false; |
1950 | case AArch64::ZAD0: |
1951 | if (TileNum <= 7) |
1952 | break; |
1953 | return false; |
1954 | } |
1955 | |
1956 | BaseReg += TileNum; |
1957 | return true; |
1958 | } |
1959 | |
1960 | template <unsigned MaxIdx, unsigned Scale> |
1961 | void AArch64DAGToDAGISel::SelectMultiVectorMove(SDNode *N, unsigned NumVecs, |
1962 | unsigned BaseReg, unsigned Op) { |
1963 | unsigned TileNum = 0; |
1964 | if (BaseReg != AArch64::ZA) |
1965 | TileNum = N->getConstantOperandVal(Num: 2); |
1966 | |
1967 | if (!SelectSMETile(BaseReg, TileNum)) |
1968 | return; |
1969 | |
1970 | SDValue SliceBase, Base, Offset; |
1971 | if (BaseReg == AArch64::ZA) |
1972 | SliceBase = N->getOperand(Num: 2); |
1973 | else |
1974 | SliceBase = N->getOperand(Num: 3); |
1975 | |
1976 | if (!SelectSMETileSlice(N: SliceBase, MaxSize: MaxIdx, Vector&: Base, Offset, Scale)) |
1977 | return; |
1978 | |
1979 | SDLoc DL(N); |
1980 | SDValue SubReg = CurDAG->getRegister(BaseReg, MVT::Other); |
1981 | SDValue Ops[] = {SubReg, Base, Offset, /*Chain*/ N->getOperand(Num: 0)}; |
1982 | SDNode *Mov = CurDAG->getMachineNode(Op, DL, {MVT::Untyped, MVT::Other}, Ops); |
1983 | |
1984 | EVT VT = N->getValueType(ResNo: 0); |
1985 | for (unsigned I = 0; I < NumVecs; ++I) |
1986 | ReplaceUses(SDValue(N, I), |
1987 | CurDAG->getTargetExtractSubreg(AArch64::zsub0 + I, DL, VT, |
1988 | SDValue(Mov, 0))); |
1989 | // Copy chain |
1990 | unsigned ChainIdx = NumVecs; |
1991 | ReplaceUses(F: SDValue(N, ChainIdx), T: SDValue(Mov, 1)); |
1992 | CurDAG->RemoveDeadNode(N); |
1993 | } |
1994 | |
1995 | void AArch64DAGToDAGISel::SelectUnaryMultiIntrinsic(SDNode *N, |
1996 | unsigned NumOutVecs, |
1997 | bool IsTupleInput, |
1998 | unsigned Opc) { |
1999 | SDLoc DL(N); |
2000 | EVT VT = N->getValueType(ResNo: 0); |
2001 | unsigned NumInVecs = N->getNumOperands() - 1; |
2002 | |
2003 | SmallVector<SDValue, 6> Ops; |
2004 | if (IsTupleInput) { |
2005 | assert((NumInVecs == 2 || NumInVecs == 4) && |
2006 | "Don't know how to handle multi-register input!" ); |
2007 | SmallVector<SDValue, 4> Regs(N->op_begin() + 1, |
2008 | N->op_begin() + 1 + NumInVecs); |
2009 | Ops.push_back(Elt: createZMulTuple(Regs)); |
2010 | } else { |
2011 | // All intrinsic nodes have the ID as the first operand, hence the "1 + I". |
2012 | for (unsigned I = 0; I < NumInVecs; I++) |
2013 | Ops.push_back(Elt: N->getOperand(Num: 1 + I)); |
2014 | } |
2015 | |
2016 | SDNode *Res = CurDAG->getMachineNode(Opc, DL, MVT::Untyped, Ops); |
2017 | SDValue SuperReg = SDValue(Res, 0); |
2018 | |
2019 | for (unsigned I = 0; I < NumOutVecs; I++) |
2020 | ReplaceUses(SDValue(N, I), CurDAG->getTargetExtractSubreg( |
2021 | AArch64::zsub0 + I, DL, VT, SuperReg)); |
2022 | CurDAG->RemoveDeadNode(N); |
2023 | } |
2024 | |
2025 | void AArch64DAGToDAGISel::SelectStore(SDNode *N, unsigned NumVecs, |
2026 | unsigned Opc) { |
2027 | SDLoc dl(N); |
2028 | EVT VT = N->getOperand(Num: 2)->getValueType(ResNo: 0); |
2029 | |
2030 | // Form a REG_SEQUENCE to force register allocation. |
2031 | bool Is128Bit = VT.getSizeInBits() == 128; |
2032 | SmallVector<SDValue, 4> Regs(N->op_begin() + 2, N->op_begin() + 2 + NumVecs); |
2033 | SDValue RegSeq = Is128Bit ? createQTuple(Regs) : createDTuple(Regs); |
2034 | |
2035 | SDValue Ops[] = {RegSeq, N->getOperand(Num: NumVecs + 2), N->getOperand(Num: 0)}; |
2036 | SDNode *St = CurDAG->getMachineNode(Opcode: Opc, dl, VT: N->getValueType(ResNo: 0), Ops); |
2037 | |
2038 | // Transfer memoperands. |
2039 | MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(Val: N)->getMemOperand(); |
2040 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: St), NewMemRefs: {MemOp}); |
2041 | |
2042 | ReplaceNode(F: N, T: St); |
2043 | } |
2044 | |
2045 | void AArch64DAGToDAGISel::SelectPredicatedStore(SDNode *N, unsigned NumVecs, |
2046 | unsigned Scale, unsigned Opc_rr, |
2047 | unsigned Opc_ri) { |
2048 | SDLoc dl(N); |
2049 | |
2050 | // Form a REG_SEQUENCE to force register allocation. |
2051 | SmallVector<SDValue, 4> Regs(N->op_begin() + 2, N->op_begin() + 2 + NumVecs); |
2052 | SDValue RegSeq = createZTuple(Regs); |
2053 | |
2054 | // Optimize addressing mode. |
2055 | unsigned Opc; |
2056 | SDValue Offset, Base; |
2057 | std::tie(Opc, Base, Offset) = findAddrModeSVELoadStore( |
2058 | N, Opc_rr, Opc_ri, N->getOperand(NumVecs + 3), |
2059 | CurDAG->getTargetConstant(0, dl, MVT::i64), Scale); |
2060 | |
2061 | SDValue Ops[] = {RegSeq, N->getOperand(Num: NumVecs + 2), // predicate |
2062 | Base, // address |
2063 | Offset, // offset |
2064 | N->getOperand(Num: 0)}; // chain |
2065 | SDNode *St = CurDAG->getMachineNode(Opcode: Opc, dl, VT: N->getValueType(ResNo: 0), Ops); |
2066 | |
2067 | ReplaceNode(F: N, T: St); |
2068 | } |
2069 | |
2070 | bool AArch64DAGToDAGISel::SelectAddrModeFrameIndexSVE(SDValue N, SDValue &Base, |
2071 | SDValue &OffImm) { |
2072 | SDLoc dl(N); |
2073 | const DataLayout &DL = CurDAG->getDataLayout(); |
2074 | const TargetLowering *TLI = getTargetLowering(); |
2075 | |
2076 | // Try to match it for the frame address |
2077 | if (auto FINode = dyn_cast<FrameIndexSDNode>(Val&: N)) { |
2078 | int FI = FINode->getIndex(); |
2079 | Base = CurDAG->getTargetFrameIndex(FI, VT: TLI->getPointerTy(DL)); |
2080 | OffImm = CurDAG->getTargetConstant(0, dl, MVT::i64); |
2081 | return true; |
2082 | } |
2083 | |
2084 | return false; |
2085 | } |
2086 | |
2087 | void AArch64DAGToDAGISel::SelectPostStore(SDNode *N, unsigned NumVecs, |
2088 | unsigned Opc) { |
2089 | SDLoc dl(N); |
2090 | EVT VT = N->getOperand(Num: 2)->getValueType(ResNo: 0); |
2091 | const EVT ResTys[] = {MVT::i64, // Type of the write back register |
2092 | MVT::Other}; // Type for the Chain |
2093 | |
2094 | // Form a REG_SEQUENCE to force register allocation. |
2095 | bool Is128Bit = VT.getSizeInBits() == 128; |
2096 | SmallVector<SDValue, 4> Regs(N->op_begin() + 1, N->op_begin() + 1 + NumVecs); |
2097 | SDValue RegSeq = Is128Bit ? createQTuple(Regs) : createDTuple(Regs); |
2098 | |
2099 | SDValue Ops[] = {RegSeq, |
2100 | N->getOperand(Num: NumVecs + 1), // base register |
2101 | N->getOperand(Num: NumVecs + 2), // Incremental |
2102 | N->getOperand(Num: 0)}; // Chain |
2103 | SDNode *St = CurDAG->getMachineNode(Opc, dl, ResTys, Ops); |
2104 | |
2105 | ReplaceNode(F: N, T: St); |
2106 | } |
2107 | |
2108 | namespace { |
2109 | /// WidenVector - Given a value in the V64 register class, produce the |
2110 | /// equivalent value in the V128 register class. |
2111 | class WidenVector { |
2112 | SelectionDAG &DAG; |
2113 | |
2114 | public: |
2115 | WidenVector(SelectionDAG &DAG) : DAG(DAG) {} |
2116 | |
2117 | SDValue operator()(SDValue V64Reg) { |
2118 | EVT VT = V64Reg.getValueType(); |
2119 | unsigned NarrowSize = VT.getVectorNumElements(); |
2120 | MVT EltTy = VT.getVectorElementType().getSimpleVT(); |
2121 | MVT WideTy = MVT::getVectorVT(VT: EltTy, NumElements: 2 * NarrowSize); |
2122 | SDLoc DL(V64Reg); |
2123 | |
2124 | SDValue Undef = |
2125 | SDValue(DAG.getMachineNode(Opcode: TargetOpcode::IMPLICIT_DEF, dl: DL, VT: WideTy), 0); |
2126 | return DAG.getTargetInsertSubreg(AArch64::dsub, DL, WideTy, Undef, V64Reg); |
2127 | } |
2128 | }; |
2129 | } // namespace |
2130 | |
2131 | /// NarrowVector - Given a value in the V128 register class, produce the |
2132 | /// equivalent value in the V64 register class. |
2133 | static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) { |
2134 | EVT VT = V128Reg.getValueType(); |
2135 | unsigned WideSize = VT.getVectorNumElements(); |
2136 | MVT EltTy = VT.getVectorElementType().getSimpleVT(); |
2137 | MVT NarrowTy = MVT::getVectorVT(VT: EltTy, NumElements: WideSize / 2); |
2138 | |
2139 | return DAG.getTargetExtractSubreg(AArch64::dsub, SDLoc(V128Reg), NarrowTy, |
2140 | V128Reg); |
2141 | } |
2142 | |
2143 | void AArch64DAGToDAGISel::SelectLoadLane(SDNode *N, unsigned NumVecs, |
2144 | unsigned Opc) { |
2145 | SDLoc dl(N); |
2146 | EVT VT = N->getValueType(ResNo: 0); |
2147 | bool Narrow = VT.getSizeInBits() == 64; |
2148 | |
2149 | // Form a REG_SEQUENCE to force register allocation. |
2150 | SmallVector<SDValue, 4> Regs(N->op_begin() + 2, N->op_begin() + 2 + NumVecs); |
2151 | |
2152 | if (Narrow) |
2153 | transform(Range&: Regs, d_first: Regs.begin(), |
2154 | F: WidenVector(*CurDAG)); |
2155 | |
2156 | SDValue RegSeq = createQTuple(Regs); |
2157 | |
2158 | const EVT ResTys[] = {MVT::Untyped, MVT::Other}; |
2159 | |
2160 | unsigned LaneNo = N->getConstantOperandVal(Num: NumVecs + 2); |
2161 | |
2162 | SDValue Ops[] = {RegSeq, CurDAG->getTargetConstant(LaneNo, dl, MVT::i64), |
2163 | N->getOperand(NumVecs + 3), N->getOperand(0)}; |
2164 | SDNode *Ld = CurDAG->getMachineNode(Opc, dl, ResTys, Ops); |
2165 | SDValue SuperReg = SDValue(Ld, 0); |
2166 | |
2167 | EVT WideVT = RegSeq.getOperand(i: 1)->getValueType(ResNo: 0); |
2168 | static const unsigned QSubs[] = { AArch64::qsub0, AArch64::qsub1, |
2169 | AArch64::qsub2, AArch64::qsub3 }; |
2170 | for (unsigned i = 0; i < NumVecs; ++i) { |
2171 | SDValue NV = CurDAG->getTargetExtractSubreg(SRIdx: QSubs[i], DL: dl, VT: WideVT, Operand: SuperReg); |
2172 | if (Narrow) |
2173 | NV = NarrowVector(V128Reg: NV, DAG&: *CurDAG); |
2174 | ReplaceUses(F: SDValue(N, i), T: NV); |
2175 | } |
2176 | |
2177 | ReplaceUses(F: SDValue(N, NumVecs), T: SDValue(Ld, 1)); |
2178 | CurDAG->RemoveDeadNode(N); |
2179 | } |
2180 | |
2181 | void AArch64DAGToDAGISel::SelectPostLoadLane(SDNode *N, unsigned NumVecs, |
2182 | unsigned Opc) { |
2183 | SDLoc dl(N); |
2184 | EVT VT = N->getValueType(ResNo: 0); |
2185 | bool Narrow = VT.getSizeInBits() == 64; |
2186 | |
2187 | // Form a REG_SEQUENCE to force register allocation. |
2188 | SmallVector<SDValue, 4> Regs(N->op_begin() + 1, N->op_begin() + 1 + NumVecs); |
2189 | |
2190 | if (Narrow) |
2191 | transform(Range&: Regs, d_first: Regs.begin(), |
2192 | F: WidenVector(*CurDAG)); |
2193 | |
2194 | SDValue RegSeq = createQTuple(Regs); |
2195 | |
2196 | const EVT ResTys[] = {MVT::i64, // Type of the write back register |
2197 | RegSeq->getValueType(0), MVT::Other}; |
2198 | |
2199 | unsigned LaneNo = N->getConstantOperandVal(Num: NumVecs + 1); |
2200 | |
2201 | SDValue Ops[] = {RegSeq, |
2202 | CurDAG->getTargetConstant(LaneNo, dl, |
2203 | MVT::i64), // Lane Number |
2204 | N->getOperand(NumVecs + 2), // Base register |
2205 | N->getOperand(NumVecs + 3), // Incremental |
2206 | N->getOperand(0)}; |
2207 | SDNode *Ld = CurDAG->getMachineNode(Opc, dl, ResTys, Ops); |
2208 | |
2209 | // Update uses of the write back register |
2210 | ReplaceUses(F: SDValue(N, NumVecs), T: SDValue(Ld, 0)); |
2211 | |
2212 | // Update uses of the vector list |
2213 | SDValue SuperReg = SDValue(Ld, 1); |
2214 | if (NumVecs == 1) { |
2215 | ReplaceUses(F: SDValue(N, 0), |
2216 | T: Narrow ? NarrowVector(V128Reg: SuperReg, DAG&: *CurDAG) : SuperReg); |
2217 | } else { |
2218 | EVT WideVT = RegSeq.getOperand(i: 1)->getValueType(ResNo: 0); |
2219 | static const unsigned QSubs[] = { AArch64::qsub0, AArch64::qsub1, |
2220 | AArch64::qsub2, AArch64::qsub3 }; |
2221 | for (unsigned i = 0; i < NumVecs; ++i) { |
2222 | SDValue NV = CurDAG->getTargetExtractSubreg(SRIdx: QSubs[i], DL: dl, VT: WideVT, |
2223 | Operand: SuperReg); |
2224 | if (Narrow) |
2225 | NV = NarrowVector(V128Reg: NV, DAG&: *CurDAG); |
2226 | ReplaceUses(F: SDValue(N, i), T: NV); |
2227 | } |
2228 | } |
2229 | |
2230 | // Update the Chain |
2231 | ReplaceUses(F: SDValue(N, NumVecs + 1), T: SDValue(Ld, 2)); |
2232 | CurDAG->RemoveDeadNode(N); |
2233 | } |
2234 | |
2235 | void AArch64DAGToDAGISel::SelectStoreLane(SDNode *N, unsigned NumVecs, |
2236 | unsigned Opc) { |
2237 | SDLoc dl(N); |
2238 | EVT VT = N->getOperand(Num: 2)->getValueType(ResNo: 0); |
2239 | bool Narrow = VT.getSizeInBits() == 64; |
2240 | |
2241 | // Form a REG_SEQUENCE to force register allocation. |
2242 | SmallVector<SDValue, 4> Regs(N->op_begin() + 2, N->op_begin() + 2 + NumVecs); |
2243 | |
2244 | if (Narrow) |
2245 | transform(Range&: Regs, d_first: Regs.begin(), |
2246 | F: WidenVector(*CurDAG)); |
2247 | |
2248 | SDValue RegSeq = createQTuple(Regs); |
2249 | |
2250 | unsigned LaneNo = N->getConstantOperandVal(Num: NumVecs + 2); |
2251 | |
2252 | SDValue Ops[] = {RegSeq, CurDAG->getTargetConstant(LaneNo, dl, MVT::i64), |
2253 | N->getOperand(NumVecs + 3), N->getOperand(0)}; |
2254 | SDNode *St = CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops); |
2255 | |
2256 | // Transfer memoperands. |
2257 | MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(Val: N)->getMemOperand(); |
2258 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: St), NewMemRefs: {MemOp}); |
2259 | |
2260 | ReplaceNode(F: N, T: St); |
2261 | } |
2262 | |
2263 | void AArch64DAGToDAGISel::SelectPostStoreLane(SDNode *N, unsigned NumVecs, |
2264 | unsigned Opc) { |
2265 | SDLoc dl(N); |
2266 | EVT VT = N->getOperand(Num: 2)->getValueType(ResNo: 0); |
2267 | bool Narrow = VT.getSizeInBits() == 64; |
2268 | |
2269 | // Form a REG_SEQUENCE to force register allocation. |
2270 | SmallVector<SDValue, 4> Regs(N->op_begin() + 1, N->op_begin() + 1 + NumVecs); |
2271 | |
2272 | if (Narrow) |
2273 | transform(Range&: Regs, d_first: Regs.begin(), |
2274 | F: WidenVector(*CurDAG)); |
2275 | |
2276 | SDValue RegSeq = createQTuple(Regs); |
2277 | |
2278 | const EVT ResTys[] = {MVT::i64, // Type of the write back register |
2279 | MVT::Other}; |
2280 | |
2281 | unsigned LaneNo = N->getConstantOperandVal(Num: NumVecs + 1); |
2282 | |
2283 | SDValue Ops[] = {RegSeq, CurDAG->getTargetConstant(LaneNo, dl, MVT::i64), |
2284 | N->getOperand(NumVecs + 2), // Base Register |
2285 | N->getOperand(NumVecs + 3), // Incremental |
2286 | N->getOperand(0)}; |
2287 | SDNode *St = CurDAG->getMachineNode(Opc, dl, ResTys, Ops); |
2288 | |
2289 | // Transfer memoperands. |
2290 | MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(Val: N)->getMemOperand(); |
2291 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: St), NewMemRefs: {MemOp}); |
2292 | |
2293 | ReplaceNode(F: N, T: St); |
2294 | } |
2295 | |
2296 | static bool isBitfieldExtractOpFromAnd(SelectionDAG *CurDAG, SDNode *N, |
2297 | unsigned &Opc, SDValue &Opd0, |
2298 | unsigned &LSB, unsigned &MSB, |
2299 | unsigned NumberOfIgnoredLowBits, |
2300 | bool BiggerPattern) { |
2301 | assert(N->getOpcode() == ISD::AND && |
2302 | "N must be a AND operation to call this function" ); |
2303 | |
2304 | EVT VT = N->getValueType(ResNo: 0); |
2305 | |
2306 | // Here we can test the type of VT and return false when the type does not |
2307 | // match, but since it is done prior to that call in the current context |
2308 | // we turned that into an assert to avoid redundant code. |
2309 | assert((VT == MVT::i32 || VT == MVT::i64) && |
2310 | "Type checking must have been done before calling this function" ); |
2311 | |
2312 | // FIXME: simplify-demanded-bits in DAGCombine will probably have |
2313 | // changed the AND node to a 32-bit mask operation. We'll have to |
2314 | // undo that as part of the transform here if we want to catch all |
2315 | // the opportunities. |
2316 | // Currently the NumberOfIgnoredLowBits argument helps to recover |
2317 | // from these situations when matching bigger pattern (bitfield insert). |
2318 | |
2319 | // For unsigned extracts, check for a shift right and mask |
2320 | uint64_t AndImm = 0; |
2321 | if (!isOpcWithIntImmediate(N, Opc: ISD::AND, Imm&: AndImm)) |
2322 | return false; |
2323 | |
2324 | const SDNode *Op0 = N->getOperand(Num: 0).getNode(); |
2325 | |
2326 | // Because of simplify-demanded-bits in DAGCombine, the mask may have been |
2327 | // simplified. Try to undo that |
2328 | AndImm |= maskTrailingOnes<uint64_t>(N: NumberOfIgnoredLowBits); |
2329 | |
2330 | // The immediate is a mask of the low bits iff imm & (imm+1) == 0 |
2331 | if (AndImm & (AndImm + 1)) |
2332 | return false; |
2333 | |
2334 | bool ClampMSB = false; |
2335 | uint64_t SrlImm = 0; |
2336 | // Handle the SRL + ANY_EXTEND case. |
2337 | if (VT == MVT::i64 && Op0->getOpcode() == ISD::ANY_EXTEND && |
2338 | isOpcWithIntImmediate(Op0->getOperand(0).getNode(), ISD::SRL, SrlImm)) { |
2339 | // Extend the incoming operand of the SRL to 64-bit. |
2340 | Opd0 = Widen(CurDAG, N: Op0->getOperand(Num: 0).getOperand(i: 0)); |
2341 | // Make sure to clamp the MSB so that we preserve the semantics of the |
2342 | // original operations. |
2343 | ClampMSB = true; |
2344 | } else if (VT == MVT::i32 && Op0->getOpcode() == ISD::TRUNCATE && |
2345 | isOpcWithIntImmediate(Op0->getOperand(0).getNode(), ISD::SRL, |
2346 | SrlImm)) { |
2347 | // If the shift result was truncated, we can still combine them. |
2348 | Opd0 = Op0->getOperand(Num: 0).getOperand(i: 0); |
2349 | |
2350 | // Use the type of SRL node. |
2351 | VT = Opd0->getValueType(ResNo: 0); |
2352 | } else if (isOpcWithIntImmediate(N: Op0, Opc: ISD::SRL, Imm&: SrlImm)) { |
2353 | Opd0 = Op0->getOperand(Num: 0); |
2354 | ClampMSB = (VT == MVT::i32); |
2355 | } else if (BiggerPattern) { |
2356 | // Let's pretend a 0 shift right has been performed. |
2357 | // The resulting code will be at least as good as the original one |
2358 | // plus it may expose more opportunities for bitfield insert pattern. |
2359 | // FIXME: Currently we limit this to the bigger pattern, because |
2360 | // some optimizations expect AND and not UBFM. |
2361 | Opd0 = N->getOperand(Num: 0); |
2362 | } else |
2363 | return false; |
2364 | |
2365 | // Bail out on large immediates. This happens when no proper |
2366 | // combining/constant folding was performed. |
2367 | if (!BiggerPattern && (SrlImm <= 0 || SrlImm >= VT.getSizeInBits())) { |
2368 | LLVM_DEBUG( |
2369 | (dbgs() << N |
2370 | << ": Found large shift immediate, this should not happen\n" )); |
2371 | return false; |
2372 | } |
2373 | |
2374 | LSB = SrlImm; |
2375 | MSB = SrlImm + |
2376 | (VT == MVT::i32 ? llvm::countr_one<uint32_t>(AndImm) |
2377 | : llvm::countr_one<uint64_t>(AndImm)) - |
2378 | 1; |
2379 | if (ClampMSB) |
2380 | // Since we're moving the extend before the right shift operation, we need |
2381 | // to clamp the MSB to make sure we don't shift in undefined bits instead of |
2382 | // the zeros which would get shifted in with the original right shift |
2383 | // operation. |
2384 | MSB = MSB > 31 ? 31 : MSB; |
2385 | |
2386 | Opc = VT == MVT::i32 ? AArch64::UBFMWri : AArch64::UBFMXri; |
2387 | return true; |
2388 | } |
2389 | |
2390 | static bool (SDNode *N, unsigned &Opc, |
2391 | SDValue &Opd0, unsigned &Immr, |
2392 | unsigned &Imms) { |
2393 | assert(N->getOpcode() == ISD::SIGN_EXTEND_INREG); |
2394 | |
2395 | EVT VT = N->getValueType(ResNo: 0); |
2396 | unsigned BitWidth = VT.getSizeInBits(); |
2397 | assert((VT == MVT::i32 || VT == MVT::i64) && |
2398 | "Type checking must have been done before calling this function" ); |
2399 | |
2400 | SDValue Op = N->getOperand(Num: 0); |
2401 | if (Op->getOpcode() == ISD::TRUNCATE) { |
2402 | Op = Op->getOperand(Num: 0); |
2403 | VT = Op->getValueType(ResNo: 0); |
2404 | BitWidth = VT.getSizeInBits(); |
2405 | } |
2406 | |
2407 | uint64_t ShiftImm; |
2408 | if (!isOpcWithIntImmediate(N: Op.getNode(), Opc: ISD::SRL, Imm&: ShiftImm) && |
2409 | !isOpcWithIntImmediate(N: Op.getNode(), Opc: ISD::SRA, Imm&: ShiftImm)) |
2410 | return false; |
2411 | |
2412 | unsigned Width = cast<VTSDNode>(Val: N->getOperand(Num: 1))->getVT().getSizeInBits(); |
2413 | if (ShiftImm + Width > BitWidth) |
2414 | return false; |
2415 | |
2416 | Opc = (VT == MVT::i32) ? AArch64::SBFMWri : AArch64::SBFMXri; |
2417 | Opd0 = Op.getOperand(i: 0); |
2418 | Immr = ShiftImm; |
2419 | Imms = ShiftImm + Width - 1; |
2420 | return true; |
2421 | } |
2422 | |
2423 | static bool (SDNode *N, unsigned &Opc, |
2424 | SDValue &Opd0, unsigned &LSB, |
2425 | unsigned &MSB) { |
2426 | // We are looking for the following pattern which basically extracts several |
2427 | // continuous bits from the source value and places it from the LSB of the |
2428 | // destination value, all other bits of the destination value or set to zero: |
2429 | // |
2430 | // Value2 = AND Value, MaskImm |
2431 | // SRL Value2, ShiftImm |
2432 | // |
2433 | // with MaskImm >> ShiftImm to search for the bit width. |
2434 | // |
2435 | // This gets selected into a single UBFM: |
2436 | // |
2437 | // UBFM Value, ShiftImm, Log2_64(MaskImm) |
2438 | // |
2439 | |
2440 | if (N->getOpcode() != ISD::SRL) |
2441 | return false; |
2442 | |
2443 | uint64_t AndMask = 0; |
2444 | if (!isOpcWithIntImmediate(N: N->getOperand(Num: 0).getNode(), Opc: ISD::AND, Imm&: AndMask)) |
2445 | return false; |
2446 | |
2447 | Opd0 = N->getOperand(Num: 0).getOperand(i: 0); |
2448 | |
2449 | uint64_t SrlImm = 0; |
2450 | if (!isIntImmediate(N: N->getOperand(Num: 1), Imm&: SrlImm)) |
2451 | return false; |
2452 | |
2453 | // Check whether we really have several bits extract here. |
2454 | if (!isMask_64(Value: AndMask >> SrlImm)) |
2455 | return false; |
2456 | |
2457 | Opc = N->getValueType(0) == MVT::i32 ? AArch64::UBFMWri : AArch64::UBFMXri; |
2458 | LSB = SrlImm; |
2459 | MSB = llvm::Log2_64(Value: AndMask); |
2460 | return true; |
2461 | } |
2462 | |
2463 | static bool (SDNode *N, unsigned &Opc, SDValue &Opd0, |
2464 | unsigned &Immr, unsigned &Imms, |
2465 | bool BiggerPattern) { |
2466 | assert((N->getOpcode() == ISD::SRA || N->getOpcode() == ISD::SRL) && |
2467 | "N must be a SHR/SRA operation to call this function" ); |
2468 | |
2469 | EVT VT = N->getValueType(ResNo: 0); |
2470 | |
2471 | // Here we can test the type of VT and return false when the type does not |
2472 | // match, but since it is done prior to that call in the current context |
2473 | // we turned that into an assert to avoid redundant code. |
2474 | assert((VT == MVT::i32 || VT == MVT::i64) && |
2475 | "Type checking must have been done before calling this function" ); |
2476 | |
2477 | // Check for AND + SRL doing several bits extract. |
2478 | if (isSeveralBitsExtractOpFromShr(N, Opc, Opd0, LSB&: Immr, MSB&: Imms)) |
2479 | return true; |
2480 | |
2481 | // We're looking for a shift of a shift. |
2482 | uint64_t ShlImm = 0; |
2483 | uint64_t TruncBits = 0; |
2484 | if (isOpcWithIntImmediate(N: N->getOperand(Num: 0).getNode(), Opc: ISD::SHL, Imm&: ShlImm)) { |
2485 | Opd0 = N->getOperand(Num: 0).getOperand(i: 0); |
2486 | } else if (VT == MVT::i32 && N->getOpcode() == ISD::SRL && |
2487 | N->getOperand(0).getNode()->getOpcode() == ISD::TRUNCATE) { |
2488 | // We are looking for a shift of truncate. Truncate from i64 to i32 could |
2489 | // be considered as setting high 32 bits as zero. Our strategy here is to |
2490 | // always generate 64bit UBFM. This consistency will help the CSE pass |
2491 | // later find more redundancy. |
2492 | Opd0 = N->getOperand(Num: 0).getOperand(i: 0); |
2493 | TruncBits = Opd0->getValueType(ResNo: 0).getSizeInBits() - VT.getSizeInBits(); |
2494 | VT = Opd0.getValueType(); |
2495 | assert(VT == MVT::i64 && "the promoted type should be i64" ); |
2496 | } else if (BiggerPattern) { |
2497 | // Let's pretend a 0 shift left has been performed. |
2498 | // FIXME: Currently we limit this to the bigger pattern case, |
2499 | // because some optimizations expect AND and not UBFM |
2500 | Opd0 = N->getOperand(Num: 0); |
2501 | } else |
2502 | return false; |
2503 | |
2504 | // Missing combines/constant folding may have left us with strange |
2505 | // constants. |
2506 | if (ShlImm >= VT.getSizeInBits()) { |
2507 | LLVM_DEBUG( |
2508 | (dbgs() << N |
2509 | << ": Found large shift immediate, this should not happen\n" )); |
2510 | return false; |
2511 | } |
2512 | |
2513 | uint64_t SrlImm = 0; |
2514 | if (!isIntImmediate(N: N->getOperand(Num: 1), Imm&: SrlImm)) |
2515 | return false; |
2516 | |
2517 | assert(SrlImm > 0 && SrlImm < VT.getSizeInBits() && |
2518 | "bad amount in shift node!" ); |
2519 | int immr = SrlImm - ShlImm; |
2520 | Immr = immr < 0 ? immr + VT.getSizeInBits() : immr; |
2521 | Imms = VT.getSizeInBits() - ShlImm - TruncBits - 1; |
2522 | // SRA requires a signed extraction |
2523 | if (VT == MVT::i32) |
2524 | Opc = N->getOpcode() == ISD::SRA ? AArch64::SBFMWri : AArch64::UBFMWri; |
2525 | else |
2526 | Opc = N->getOpcode() == ISD::SRA ? AArch64::SBFMXri : AArch64::UBFMXri; |
2527 | return true; |
2528 | } |
2529 | |
2530 | bool AArch64DAGToDAGISel::(SDNode *N) { |
2531 | assert(N->getOpcode() == ISD::SIGN_EXTEND); |
2532 | |
2533 | EVT VT = N->getValueType(ResNo: 0); |
2534 | EVT NarrowVT = N->getOperand(Num: 0)->getValueType(ResNo: 0); |
2535 | if (VT != MVT::i64 || NarrowVT != MVT::i32) |
2536 | return false; |
2537 | |
2538 | uint64_t ShiftImm; |
2539 | SDValue Op = N->getOperand(Num: 0); |
2540 | if (!isOpcWithIntImmediate(N: Op.getNode(), Opc: ISD::SRA, Imm&: ShiftImm)) |
2541 | return false; |
2542 | |
2543 | SDLoc dl(N); |
2544 | // Extend the incoming operand of the shift to 64-bits. |
2545 | SDValue Opd0 = Widen(CurDAG, N: Op.getOperand(i: 0)); |
2546 | unsigned Immr = ShiftImm; |
2547 | unsigned Imms = NarrowVT.getSizeInBits() - 1; |
2548 | SDValue Ops[] = {Opd0, CurDAG->getTargetConstant(Val: Immr, DL: dl, VT), |
2549 | CurDAG->getTargetConstant(Val: Imms, DL: dl, VT)}; |
2550 | CurDAG->SelectNodeTo(N, AArch64::SBFMXri, VT, Ops); |
2551 | return true; |
2552 | } |
2553 | |
2554 | static bool (SelectionDAG *CurDAG, SDNode *N, unsigned &Opc, |
2555 | SDValue &Opd0, unsigned &Immr, unsigned &Imms, |
2556 | unsigned NumberOfIgnoredLowBits = 0, |
2557 | bool BiggerPattern = false) { |
2558 | if (N->getValueType(0) != MVT::i32 && N->getValueType(0) != MVT::i64) |
2559 | return false; |
2560 | |
2561 | switch (N->getOpcode()) { |
2562 | default: |
2563 | if (!N->isMachineOpcode()) |
2564 | return false; |
2565 | break; |
2566 | case ISD::AND: |
2567 | return isBitfieldExtractOpFromAnd(CurDAG, N, Opc, Opd0, LSB&: Immr, MSB&: Imms, |
2568 | NumberOfIgnoredLowBits, BiggerPattern); |
2569 | case ISD::SRL: |
2570 | case ISD::SRA: |
2571 | return isBitfieldExtractOpFromShr(N, Opc, Opd0, Immr, Imms, BiggerPattern); |
2572 | |
2573 | case ISD::SIGN_EXTEND_INREG: |
2574 | return isBitfieldExtractOpFromSExtInReg(N, Opc, Opd0, Immr, Imms); |
2575 | } |
2576 | |
2577 | unsigned NOpc = N->getMachineOpcode(); |
2578 | switch (NOpc) { |
2579 | default: |
2580 | return false; |
2581 | case AArch64::SBFMWri: |
2582 | case AArch64::UBFMWri: |
2583 | case AArch64::SBFMXri: |
2584 | case AArch64::UBFMXri: |
2585 | Opc = NOpc; |
2586 | Opd0 = N->getOperand(Num: 0); |
2587 | Immr = N->getConstantOperandVal(Num: 1); |
2588 | Imms = N->getConstantOperandVal(Num: 2); |
2589 | return true; |
2590 | } |
2591 | // Unreachable |
2592 | return false; |
2593 | } |
2594 | |
2595 | bool AArch64DAGToDAGISel::(SDNode *N) { |
2596 | unsigned Opc, Immr, Imms; |
2597 | SDValue Opd0; |
2598 | if (!isBitfieldExtractOp(CurDAG, N, Opc, Opd0, Immr, Imms)) |
2599 | return false; |
2600 | |
2601 | EVT VT = N->getValueType(ResNo: 0); |
2602 | SDLoc dl(N); |
2603 | |
2604 | // If the bit extract operation is 64bit but the original type is 32bit, we |
2605 | // need to add one EXTRACT_SUBREG. |
2606 | if ((Opc == AArch64::SBFMXri || Opc == AArch64::UBFMXri) && VT == MVT::i32) { |
2607 | SDValue Ops64[] = {Opd0, CurDAG->getTargetConstant(Immr, dl, MVT::i64), |
2608 | CurDAG->getTargetConstant(Imms, dl, MVT::i64)}; |
2609 | |
2610 | SDNode *BFM = CurDAG->getMachineNode(Opc, dl, MVT::i64, Ops64); |
2611 | SDValue Inner = CurDAG->getTargetExtractSubreg(AArch64::sub_32, dl, |
2612 | MVT::i32, SDValue(BFM, 0)); |
2613 | ReplaceNode(F: N, T: Inner.getNode()); |
2614 | return true; |
2615 | } |
2616 | |
2617 | SDValue Ops[] = {Opd0, CurDAG->getTargetConstant(Val: Immr, DL: dl, VT), |
2618 | CurDAG->getTargetConstant(Val: Imms, DL: dl, VT)}; |
2619 | CurDAG->SelectNodeTo(N, MachineOpc: Opc, VT, Ops); |
2620 | return true; |
2621 | } |
2622 | |
2623 | /// Does DstMask form a complementary pair with the mask provided by |
2624 | /// BitsToBeInserted, suitable for use in a BFI instruction. Roughly speaking, |
2625 | /// this asks whether DstMask zeroes precisely those bits that will be set by |
2626 | /// the other half. |
2627 | static bool isBitfieldDstMask(uint64_t DstMask, const APInt &BitsToBeInserted, |
2628 | unsigned NumberOfIgnoredHighBits, EVT VT) { |
2629 | assert((VT == MVT::i32 || VT == MVT::i64) && |
2630 | "i32 or i64 mask type expected!" ); |
2631 | unsigned BitWidth = VT.getSizeInBits() - NumberOfIgnoredHighBits; |
2632 | |
2633 | APInt SignificantDstMask = APInt(BitWidth, DstMask); |
2634 | APInt SignificantBitsToBeInserted = BitsToBeInserted.zextOrTrunc(width: BitWidth); |
2635 | |
2636 | return (SignificantDstMask & SignificantBitsToBeInserted) == 0 && |
2637 | (SignificantDstMask | SignificantBitsToBeInserted).isAllOnes(); |
2638 | } |
2639 | |
2640 | // Look for bits that will be useful for later uses. |
2641 | // A bit is consider useless as soon as it is dropped and never used |
2642 | // before it as been dropped. |
2643 | // E.g., looking for useful bit of x |
2644 | // 1. y = x & 0x7 |
2645 | // 2. z = y >> 2 |
2646 | // After #1, x useful bits are 0x7, then the useful bits of x, live through |
2647 | // y. |
2648 | // After #2, the useful bits of x are 0x4. |
2649 | // However, if x is used on an unpredicatable instruction, then all its bits |
2650 | // are useful. |
2651 | // E.g. |
2652 | // 1. y = x & 0x7 |
2653 | // 2. z = y >> 2 |
2654 | // 3. str x, [@x] |
2655 | static void getUsefulBits(SDValue Op, APInt &UsefulBits, unsigned Depth = 0); |
2656 | |
2657 | static void getUsefulBitsFromAndWithImmediate(SDValue Op, APInt &UsefulBits, |
2658 | unsigned Depth) { |
2659 | uint64_t Imm = |
2660 | cast<const ConstantSDNode>(Val: Op.getOperand(i: 1).getNode())->getZExtValue(); |
2661 | Imm = AArch64_AM::decodeLogicalImmediate(val: Imm, regSize: UsefulBits.getBitWidth()); |
2662 | UsefulBits &= APInt(UsefulBits.getBitWidth(), Imm); |
2663 | getUsefulBits(Op, UsefulBits, Depth: Depth + 1); |
2664 | } |
2665 | |
2666 | static void getUsefulBitsFromBitfieldMoveOpd(SDValue Op, APInt &UsefulBits, |
2667 | uint64_t Imm, uint64_t MSB, |
2668 | unsigned Depth) { |
2669 | // inherit the bitwidth value |
2670 | APInt OpUsefulBits(UsefulBits); |
2671 | OpUsefulBits = 1; |
2672 | |
2673 | if (MSB >= Imm) { |
2674 | OpUsefulBits <<= MSB - Imm + 1; |
2675 | --OpUsefulBits; |
2676 | // The interesting part will be in the lower part of the result |
2677 | getUsefulBits(Op, UsefulBits&: OpUsefulBits, Depth: Depth + 1); |
2678 | // The interesting part was starting at Imm in the argument |
2679 | OpUsefulBits <<= Imm; |
2680 | } else { |
2681 | OpUsefulBits <<= MSB + 1; |
2682 | --OpUsefulBits; |
2683 | // The interesting part will be shifted in the result |
2684 | OpUsefulBits <<= OpUsefulBits.getBitWidth() - Imm; |
2685 | getUsefulBits(Op, UsefulBits&: OpUsefulBits, Depth: Depth + 1); |
2686 | // The interesting part was at zero in the argument |
2687 | OpUsefulBits.lshrInPlace(ShiftAmt: OpUsefulBits.getBitWidth() - Imm); |
2688 | } |
2689 | |
2690 | UsefulBits &= OpUsefulBits; |
2691 | } |
2692 | |
2693 | static void getUsefulBitsFromUBFM(SDValue Op, APInt &UsefulBits, |
2694 | unsigned Depth) { |
2695 | uint64_t Imm = |
2696 | cast<const ConstantSDNode>(Val: Op.getOperand(i: 1).getNode())->getZExtValue(); |
2697 | uint64_t MSB = |
2698 | cast<const ConstantSDNode>(Val: Op.getOperand(i: 2).getNode())->getZExtValue(); |
2699 | |
2700 | getUsefulBitsFromBitfieldMoveOpd(Op, UsefulBits, Imm, MSB, Depth); |
2701 | } |
2702 | |
2703 | static void getUsefulBitsFromOrWithShiftedReg(SDValue Op, APInt &UsefulBits, |
2704 | unsigned Depth) { |
2705 | uint64_t ShiftTypeAndValue = |
2706 | cast<const ConstantSDNode>(Val: Op.getOperand(i: 2).getNode())->getZExtValue(); |
2707 | APInt Mask(UsefulBits); |
2708 | Mask.clearAllBits(); |
2709 | Mask.flipAllBits(); |
2710 | |
2711 | if (AArch64_AM::getShiftType(Imm: ShiftTypeAndValue) == AArch64_AM::LSL) { |
2712 | // Shift Left |
2713 | uint64_t ShiftAmt = AArch64_AM::getShiftValue(Imm: ShiftTypeAndValue); |
2714 | Mask <<= ShiftAmt; |
2715 | getUsefulBits(Op, UsefulBits&: Mask, Depth: Depth + 1); |
2716 | Mask.lshrInPlace(ShiftAmt); |
2717 | } else if (AArch64_AM::getShiftType(Imm: ShiftTypeAndValue) == AArch64_AM::LSR) { |
2718 | // Shift Right |
2719 | // We do not handle AArch64_AM::ASR, because the sign will change the |
2720 | // number of useful bits |
2721 | uint64_t ShiftAmt = AArch64_AM::getShiftValue(Imm: ShiftTypeAndValue); |
2722 | Mask.lshrInPlace(ShiftAmt); |
2723 | getUsefulBits(Op, UsefulBits&: Mask, Depth: Depth + 1); |
2724 | Mask <<= ShiftAmt; |
2725 | } else |
2726 | return; |
2727 | |
2728 | UsefulBits &= Mask; |
2729 | } |
2730 | |
2731 | static void getUsefulBitsFromBFM(SDValue Op, SDValue Orig, APInt &UsefulBits, |
2732 | unsigned Depth) { |
2733 | uint64_t Imm = |
2734 | cast<const ConstantSDNode>(Val: Op.getOperand(i: 2).getNode())->getZExtValue(); |
2735 | uint64_t MSB = |
2736 | cast<const ConstantSDNode>(Val: Op.getOperand(i: 3).getNode())->getZExtValue(); |
2737 | |
2738 | APInt OpUsefulBits(UsefulBits); |
2739 | OpUsefulBits = 1; |
2740 | |
2741 | APInt ResultUsefulBits(UsefulBits.getBitWidth(), 0); |
2742 | ResultUsefulBits.flipAllBits(); |
2743 | APInt Mask(UsefulBits.getBitWidth(), 0); |
2744 | |
2745 | getUsefulBits(Op, UsefulBits&: ResultUsefulBits, Depth: Depth + 1); |
2746 | |
2747 | if (MSB >= Imm) { |
2748 | // The instruction is a BFXIL. |
2749 | uint64_t Width = MSB - Imm + 1; |
2750 | uint64_t LSB = Imm; |
2751 | |
2752 | OpUsefulBits <<= Width; |
2753 | --OpUsefulBits; |
2754 | |
2755 | if (Op.getOperand(i: 1) == Orig) { |
2756 | // Copy the low bits from the result to bits starting from LSB. |
2757 | Mask = ResultUsefulBits & OpUsefulBits; |
2758 | Mask <<= LSB; |
2759 | } |
2760 | |
2761 | if (Op.getOperand(i: 0) == Orig) |
2762 | // Bits starting from LSB in the input contribute to the result. |
2763 | Mask |= (ResultUsefulBits & ~OpUsefulBits); |
2764 | } else { |
2765 | // The instruction is a BFI. |
2766 | uint64_t Width = MSB + 1; |
2767 | uint64_t LSB = UsefulBits.getBitWidth() - Imm; |
2768 | |
2769 | OpUsefulBits <<= Width; |
2770 | --OpUsefulBits; |
2771 | OpUsefulBits <<= LSB; |
2772 | |
2773 | if (Op.getOperand(i: 1) == Orig) { |
2774 | // Copy the bits from the result to the zero bits. |
2775 | Mask = ResultUsefulBits & OpUsefulBits; |
2776 | Mask.lshrInPlace(ShiftAmt: LSB); |
2777 | } |
2778 | |
2779 | if (Op.getOperand(i: 0) == Orig) |
2780 | Mask |= (ResultUsefulBits & ~OpUsefulBits); |
2781 | } |
2782 | |
2783 | UsefulBits &= Mask; |
2784 | } |
2785 | |
2786 | static void getUsefulBitsForUse(SDNode *UserNode, APInt &UsefulBits, |
2787 | SDValue Orig, unsigned Depth) { |
2788 | |
2789 | // Users of this node should have already been instruction selected |
2790 | // FIXME: Can we turn that into an assert? |
2791 | if (!UserNode->isMachineOpcode()) |
2792 | return; |
2793 | |
2794 | switch (UserNode->getMachineOpcode()) { |
2795 | default: |
2796 | return; |
2797 | case AArch64::ANDSWri: |
2798 | case AArch64::ANDSXri: |
2799 | case AArch64::ANDWri: |
2800 | case AArch64::ANDXri: |
2801 | // We increment Depth only when we call the getUsefulBits |
2802 | return getUsefulBitsFromAndWithImmediate(Op: SDValue(UserNode, 0), UsefulBits, |
2803 | Depth); |
2804 | case AArch64::UBFMWri: |
2805 | case AArch64::UBFMXri: |
2806 | return getUsefulBitsFromUBFM(Op: SDValue(UserNode, 0), UsefulBits, Depth); |
2807 | |
2808 | case AArch64::ORRWrs: |
2809 | case AArch64::ORRXrs: |
2810 | if (UserNode->getOperand(Num: 0) != Orig && UserNode->getOperand(Num: 1) == Orig) |
2811 | getUsefulBitsFromOrWithShiftedReg(Op: SDValue(UserNode, 0), UsefulBits, |
2812 | Depth); |
2813 | return; |
2814 | case AArch64::BFMWri: |
2815 | case AArch64::BFMXri: |
2816 | return getUsefulBitsFromBFM(Op: SDValue(UserNode, 0), Orig, UsefulBits, Depth); |
2817 | |
2818 | case AArch64::STRBBui: |
2819 | case AArch64::STURBBi: |
2820 | if (UserNode->getOperand(Num: 0) != Orig) |
2821 | return; |
2822 | UsefulBits &= APInt(UsefulBits.getBitWidth(), 0xff); |
2823 | return; |
2824 | |
2825 | case AArch64::STRHHui: |
2826 | case AArch64::STURHHi: |
2827 | if (UserNode->getOperand(Num: 0) != Orig) |
2828 | return; |
2829 | UsefulBits &= APInt(UsefulBits.getBitWidth(), 0xffff); |
2830 | return; |
2831 | } |
2832 | } |
2833 | |
2834 | static void getUsefulBits(SDValue Op, APInt &UsefulBits, unsigned Depth) { |
2835 | if (Depth >= SelectionDAG::MaxRecursionDepth) |
2836 | return; |
2837 | // Initialize UsefulBits |
2838 | if (!Depth) { |
2839 | unsigned Bitwidth = Op.getScalarValueSizeInBits(); |
2840 | // At the beginning, assume every produced bits is useful |
2841 | UsefulBits = APInt(Bitwidth, 0); |
2842 | UsefulBits.flipAllBits(); |
2843 | } |
2844 | APInt UsersUsefulBits(UsefulBits.getBitWidth(), 0); |
2845 | |
2846 | for (SDNode *Node : Op.getNode()->uses()) { |
2847 | // A use cannot produce useful bits |
2848 | APInt UsefulBitsForUse = APInt(UsefulBits); |
2849 | getUsefulBitsForUse(UserNode: Node, UsefulBits&: UsefulBitsForUse, Orig: Op, Depth); |
2850 | UsersUsefulBits |= UsefulBitsForUse; |
2851 | } |
2852 | // UsefulBits contains the produced bits that are meaningful for the |
2853 | // current definition, thus a user cannot make a bit meaningful at |
2854 | // this point |
2855 | UsefulBits &= UsersUsefulBits; |
2856 | } |
2857 | |
2858 | /// Create a machine node performing a notional SHL of Op by ShlAmount. If |
2859 | /// ShlAmount is negative, do a (logical) right-shift instead. If ShlAmount is |
2860 | /// 0, return Op unchanged. |
2861 | static SDValue getLeftShift(SelectionDAG *CurDAG, SDValue Op, int ShlAmount) { |
2862 | if (ShlAmount == 0) |
2863 | return Op; |
2864 | |
2865 | EVT VT = Op.getValueType(); |
2866 | SDLoc dl(Op); |
2867 | unsigned BitWidth = VT.getSizeInBits(); |
2868 | unsigned UBFMOpc = BitWidth == 32 ? AArch64::UBFMWri : AArch64::UBFMXri; |
2869 | |
2870 | SDNode *ShiftNode; |
2871 | if (ShlAmount > 0) { |
2872 | // LSL wD, wN, #Amt == UBFM wD, wN, #32-Amt, #31-Amt |
2873 | ShiftNode = CurDAG->getMachineNode( |
2874 | Opcode: UBFMOpc, dl, VT, Op1: Op, |
2875 | Op2: CurDAG->getTargetConstant(Val: BitWidth - ShlAmount, DL: dl, VT), |
2876 | Op3: CurDAG->getTargetConstant(Val: BitWidth - 1 - ShlAmount, DL: dl, VT)); |
2877 | } else { |
2878 | // LSR wD, wN, #Amt == UBFM wD, wN, #Amt, #32-1 |
2879 | assert(ShlAmount < 0 && "expected right shift" ); |
2880 | int ShrAmount = -ShlAmount; |
2881 | ShiftNode = CurDAG->getMachineNode( |
2882 | Opcode: UBFMOpc, dl, VT, Op1: Op, Op2: CurDAG->getTargetConstant(Val: ShrAmount, DL: dl, VT), |
2883 | Op3: CurDAG->getTargetConstant(Val: BitWidth - 1, DL: dl, VT)); |
2884 | } |
2885 | |
2886 | return SDValue(ShiftNode, 0); |
2887 | } |
2888 | |
2889 | // For bit-field-positioning pattern "(and (shl VAL, N), ShiftedMask)". |
2890 | static bool isBitfieldPositioningOpFromAnd(SelectionDAG *CurDAG, SDValue Op, |
2891 | bool BiggerPattern, |
2892 | const uint64_t NonZeroBits, |
2893 | SDValue &Src, int &DstLSB, |
2894 | int &Width); |
2895 | |
2896 | // For bit-field-positioning pattern "shl VAL, N)". |
2897 | static bool isBitfieldPositioningOpFromShl(SelectionDAG *CurDAG, SDValue Op, |
2898 | bool BiggerPattern, |
2899 | const uint64_t NonZeroBits, |
2900 | SDValue &Src, int &DstLSB, |
2901 | int &Width); |
2902 | |
2903 | /// Does this tree qualify as an attempt to move a bitfield into position, |
2904 | /// essentially "(and (shl VAL, N), Mask)" or (shl VAL, N). |
2905 | static bool isBitfieldPositioningOp(SelectionDAG *CurDAG, SDValue Op, |
2906 | bool BiggerPattern, SDValue &Src, |
2907 | int &DstLSB, int &Width) { |
2908 | EVT VT = Op.getValueType(); |
2909 | unsigned BitWidth = VT.getSizeInBits(); |
2910 | (void)BitWidth; |
2911 | assert(BitWidth == 32 || BitWidth == 64); |
2912 | |
2913 | KnownBits Known = CurDAG->computeKnownBits(Op); |
2914 | |
2915 | // Non-zero in the sense that they're not provably zero, which is the key |
2916 | // point if we want to use this value |
2917 | const uint64_t NonZeroBits = (~Known.Zero).getZExtValue(); |
2918 | if (!isShiftedMask_64(Value: NonZeroBits)) |
2919 | return false; |
2920 | |
2921 | switch (Op.getOpcode()) { |
2922 | default: |
2923 | break; |
2924 | case ISD::AND: |
2925 | return isBitfieldPositioningOpFromAnd(CurDAG, Op, BiggerPattern, |
2926 | NonZeroBits, Src, DstLSB, Width); |
2927 | case ISD::SHL: |
2928 | return isBitfieldPositioningOpFromShl(CurDAG, Op, BiggerPattern, |
2929 | NonZeroBits, Src, DstLSB, Width); |
2930 | } |
2931 | |
2932 | return false; |
2933 | } |
2934 | |
2935 | static bool isBitfieldPositioningOpFromAnd(SelectionDAG *CurDAG, SDValue Op, |
2936 | bool BiggerPattern, |
2937 | const uint64_t NonZeroBits, |
2938 | SDValue &Src, int &DstLSB, |
2939 | int &Width) { |
2940 | assert(isShiftedMask_64(NonZeroBits) && "Caller guaranteed" ); |
2941 | |
2942 | EVT VT = Op.getValueType(); |
2943 | assert((VT == MVT::i32 || VT == MVT::i64) && |
2944 | "Caller guarantees VT is one of i32 or i64" ); |
2945 | (void)VT; |
2946 | |
2947 | uint64_t AndImm; |
2948 | if (!isOpcWithIntImmediate(N: Op.getNode(), Opc: ISD::AND, Imm&: AndImm)) |
2949 | return false; |
2950 | |
2951 | // If (~AndImm & NonZeroBits) is not zero at POS, we know that |
2952 | // 1) (AndImm & (1 << POS) == 0) |
2953 | // 2) the result of AND is not zero at POS bit (according to NonZeroBits) |
2954 | // |
2955 | // 1) and 2) don't agree so something must be wrong (e.g., in |
2956 | // 'SelectionDAG::computeKnownBits') |
2957 | assert((~AndImm & NonZeroBits) == 0 && |
2958 | "Something must be wrong (e.g., in SelectionDAG::computeKnownBits)" ); |
2959 | |
2960 | SDValue AndOp0 = Op.getOperand(i: 0); |
2961 | |
2962 | uint64_t ShlImm; |
2963 | SDValue ShlOp0; |
2964 | if (isOpcWithIntImmediate(N: AndOp0.getNode(), Opc: ISD::SHL, Imm&: ShlImm)) { |
2965 | // For pattern "and(shl(val, N), shifted-mask)", 'ShlOp0' is set to 'val'. |
2966 | ShlOp0 = AndOp0.getOperand(i: 0); |
2967 | } else if (VT == MVT::i64 && AndOp0.getOpcode() == ISD::ANY_EXTEND && |
2968 | isOpcWithIntImmediate(AndOp0.getOperand(0).getNode(), ISD::SHL, |
2969 | ShlImm)) { |
2970 | // For pattern "and(any_extend(shl(val, N)), shifted-mask)" |
2971 | |
2972 | // ShlVal == shl(val, N), which is a left shift on a smaller type. |
2973 | SDValue ShlVal = AndOp0.getOperand(i: 0); |
2974 | |
2975 | // Since this is after type legalization and ShlVal is extended to MVT::i64, |
2976 | // expect VT to be MVT::i32. |
2977 | assert((ShlVal.getValueType() == MVT::i32) && "Expect VT to be MVT::i32." ); |
2978 | |
2979 | // Widens 'val' to MVT::i64 as the source of bit field positioning. |
2980 | ShlOp0 = Widen(CurDAG, N: ShlVal.getOperand(i: 0)); |
2981 | } else |
2982 | return false; |
2983 | |
2984 | // For !BiggerPattern, bail out if the AndOp0 has more than one use, since |
2985 | // then we'll end up generating AndOp0+UBFIZ instead of just keeping |
2986 | // AndOp0+AND. |
2987 | if (!BiggerPattern && !AndOp0.hasOneUse()) |
2988 | return false; |
2989 | |
2990 | DstLSB = llvm::countr_zero(Val: NonZeroBits); |
2991 | Width = llvm::countr_one(Value: NonZeroBits >> DstLSB); |
2992 | |
2993 | // Bail out on large Width. This happens when no proper combining / constant |
2994 | // folding was performed. |
2995 | if (Width >= (int)VT.getSizeInBits()) { |
2996 | // If VT is i64, Width > 64 is insensible since NonZeroBits is uint64_t, and |
2997 | // Width == 64 indicates a missed dag-combine from "(and val, AllOnes)" to |
2998 | // "val". |
2999 | // If VT is i32, what Width >= 32 means: |
3000 | // - For "(and (any_extend(shl val, N)), shifted-mask)", the`and` Op |
3001 | // demands at least 'Width' bits (after dag-combiner). This together with |
3002 | // `any_extend` Op (undefined higher bits) indicates missed combination |
3003 | // when lowering the 'and' IR instruction to an machine IR instruction. |
3004 | LLVM_DEBUG( |
3005 | dbgs() |
3006 | << "Found large Width in bit-field-positioning -- this indicates no " |
3007 | "proper combining / constant folding was performed\n" ); |
3008 | return false; |
3009 | } |
3010 | |
3011 | // BFI encompasses sufficiently many nodes that it's worth inserting an extra |
3012 | // LSL/LSR if the mask in NonZeroBits doesn't quite match up with the ISD::SHL |
3013 | // amount. BiggerPattern is true when this pattern is being matched for BFI, |
3014 | // BiggerPattern is false when this pattern is being matched for UBFIZ, in |
3015 | // which case it is not profitable to insert an extra shift. |
3016 | if (ShlImm != uint64_t(DstLSB) && !BiggerPattern) |
3017 | return false; |
3018 | |
3019 | Src = getLeftShift(CurDAG, Op: ShlOp0, ShlAmount: ShlImm - DstLSB); |
3020 | return true; |
3021 | } |
3022 | |
3023 | // For node (shl (and val, mask), N)), returns true if the node is equivalent to |
3024 | // UBFIZ. |
3025 | static bool isSeveralBitsPositioningOpFromShl(const uint64_t ShlImm, SDValue Op, |
3026 | SDValue &Src, int &DstLSB, |
3027 | int &Width) { |
3028 | // Caller should have verified that N is a left shift with constant shift |
3029 | // amount; asserts that. |
3030 | assert(Op.getOpcode() == ISD::SHL && |
3031 | "Op.getNode() should be a SHL node to call this function" ); |
3032 | assert(isIntImmediateEq(Op.getOperand(1), ShlImm) && |
3033 | "Op.getNode() should shift ShlImm to call this function" ); |
3034 | |
3035 | uint64_t AndImm = 0; |
3036 | SDValue Op0 = Op.getOperand(i: 0); |
3037 | if (!isOpcWithIntImmediate(N: Op0.getNode(), Opc: ISD::AND, Imm&: AndImm)) |
3038 | return false; |
3039 | |
3040 | const uint64_t ShiftedAndImm = ((AndImm << ShlImm) >> ShlImm); |
3041 | if (isMask_64(Value: ShiftedAndImm)) { |
3042 | // AndImm is a superset of (AllOnes >> ShlImm); in other words, AndImm |
3043 | // should end with Mask, and could be prefixed with random bits if those |
3044 | // bits are shifted out. |
3045 | // |
3046 | // For example, xyz11111 (with {x,y,z} being 0 or 1) is fine if ShlImm >= 3; |
3047 | // the AND result corresponding to those bits are shifted out, so it's fine |
3048 | // to not extract them. |
3049 | Width = llvm::countr_one(Value: ShiftedAndImm); |
3050 | DstLSB = ShlImm; |
3051 | Src = Op0.getOperand(i: 0); |
3052 | return true; |
3053 | } |
3054 | return false; |
3055 | } |
3056 | |
3057 | static bool isBitfieldPositioningOpFromShl(SelectionDAG *CurDAG, SDValue Op, |
3058 | bool BiggerPattern, |
3059 | const uint64_t NonZeroBits, |
3060 | SDValue &Src, int &DstLSB, |
3061 | int &Width) { |
3062 | assert(isShiftedMask_64(NonZeroBits) && "Caller guaranteed" ); |
3063 | |
3064 | EVT VT = Op.getValueType(); |
3065 | assert((VT == MVT::i32 || VT == MVT::i64) && |
3066 | "Caller guarantees that type is i32 or i64" ); |
3067 | (void)VT; |
3068 | |
3069 | uint64_t ShlImm; |
3070 | if (!isOpcWithIntImmediate(N: Op.getNode(), Opc: ISD::SHL, Imm&: ShlImm)) |
3071 | return false; |
3072 | |
3073 | if (!BiggerPattern && !Op.hasOneUse()) |
3074 | return false; |
3075 | |
3076 | if (isSeveralBitsPositioningOpFromShl(ShlImm, Op, Src, DstLSB, Width)) |
3077 | return true; |
3078 | |
3079 | DstLSB = llvm::countr_zero(Val: NonZeroBits); |
3080 | Width = llvm::countr_one(Value: NonZeroBits >> DstLSB); |
3081 | |
3082 | if (ShlImm != uint64_t(DstLSB) && !BiggerPattern) |
3083 | return false; |
3084 | |
3085 | Src = getLeftShift(CurDAG, Op: Op.getOperand(i: 0), ShlAmount: ShlImm - DstLSB); |
3086 | return true; |
3087 | } |
3088 | |
3089 | static bool isShiftedMask(uint64_t Mask, EVT VT) { |
3090 | assert(VT == MVT::i32 || VT == MVT::i64); |
3091 | if (VT == MVT::i32) |
3092 | return isShiftedMask_32(Value: Mask); |
3093 | return isShiftedMask_64(Value: Mask); |
3094 | } |
3095 | |
3096 | // Generate a BFI/BFXIL from 'or (and X, MaskImm), OrImm' iff the value being |
3097 | // inserted only sets known zero bits. |
3098 | static bool tryBitfieldInsertOpFromOrAndImm(SDNode *N, SelectionDAG *CurDAG) { |
3099 | assert(N->getOpcode() == ISD::OR && "Expect a OR operation" ); |
3100 | |
3101 | EVT VT = N->getValueType(ResNo: 0); |
3102 | if (VT != MVT::i32 && VT != MVT::i64) |
3103 | return false; |
3104 | |
3105 | unsigned BitWidth = VT.getSizeInBits(); |
3106 | |
3107 | uint64_t OrImm; |
3108 | if (!isOpcWithIntImmediate(N, Opc: ISD::OR, Imm&: OrImm)) |
3109 | return false; |
3110 | |
3111 | // Skip this transformation if the ORR immediate can be encoded in the ORR. |
3112 | // Otherwise, we'll trade an AND+ORR for ORR+BFI/BFXIL, which is most likely |
3113 | // performance neutral. |
3114 | if (AArch64_AM::isLogicalImmediate(imm: OrImm, regSize: BitWidth)) |
3115 | return false; |
3116 | |
3117 | uint64_t MaskImm; |
3118 | SDValue And = N->getOperand(Num: 0); |
3119 | // Must be a single use AND with an immediate operand. |
3120 | if (!And.hasOneUse() || |
3121 | !isOpcWithIntImmediate(N: And.getNode(), Opc: ISD::AND, Imm&: MaskImm)) |
3122 | return false; |
3123 | |
3124 | // Compute the Known Zero for the AND as this allows us to catch more general |
3125 | // cases than just looking for AND with imm. |
3126 | KnownBits Known = CurDAG->computeKnownBits(Op: And); |
3127 | |
3128 | // Non-zero in the sense that they're not provably zero, which is the key |
3129 | // point if we want to use this value. |
3130 | uint64_t NotKnownZero = (~Known.Zero).getZExtValue(); |
3131 | |
3132 | // The KnownZero mask must be a shifted mask (e.g., 1110..011, 11100..00). |
3133 | if (!isShiftedMask(Mask: Known.Zero.getZExtValue(), VT)) |
3134 | return false; |
3135 | |
3136 | // The bits being inserted must only set those bits that are known to be zero. |
3137 | if ((OrImm & NotKnownZero) != 0) { |
3138 | // FIXME: It's okay if the OrImm sets NotKnownZero bits to 1, but we don't |
3139 | // currently handle this case. |
3140 | return false; |
3141 | } |
3142 | |
3143 | // BFI/BFXIL dst, src, #lsb, #width. |
3144 | int LSB = llvm::countr_one(Value: NotKnownZero); |
3145 | int Width = BitWidth - APInt(BitWidth, NotKnownZero).popcount(); |
3146 | |
3147 | // BFI/BFXIL is an alias of BFM, so translate to BFM operands. |
3148 | unsigned ImmR = (BitWidth - LSB) % BitWidth; |
3149 | unsigned ImmS = Width - 1; |
3150 | |
3151 | // If we're creating a BFI instruction avoid cases where we need more |
3152 | // instructions to materialize the BFI constant as compared to the original |
3153 | // ORR. A BFXIL will use the same constant as the original ORR, so the code |
3154 | // should be no worse in this case. |
3155 | bool IsBFI = LSB != 0; |
3156 | uint64_t BFIImm = OrImm >> LSB; |
3157 | if (IsBFI && !AArch64_AM::isLogicalImmediate(imm: BFIImm, regSize: BitWidth)) { |
3158 | // We have a BFI instruction and we know the constant can't be materialized |
3159 | // with a ORR-immediate with the zero register. |
3160 | unsigned OrChunks = 0, BFIChunks = 0; |
3161 | for (unsigned Shift = 0; Shift < BitWidth; Shift += 16) { |
3162 | if (((OrImm >> Shift) & 0xFFFF) != 0) |
3163 | ++OrChunks; |
3164 | if (((BFIImm >> Shift) & 0xFFFF) != 0) |
3165 | ++BFIChunks; |
3166 | } |
3167 | if (BFIChunks > OrChunks) |
3168 | return false; |
3169 | } |
3170 | |
3171 | // Materialize the constant to be inserted. |
3172 | SDLoc DL(N); |
3173 | unsigned MOVIOpc = VT == MVT::i32 ? AArch64::MOVi32imm : AArch64::MOVi64imm; |
3174 | SDNode *MOVI = CurDAG->getMachineNode( |
3175 | Opcode: MOVIOpc, dl: DL, VT, Op1: CurDAG->getTargetConstant(Val: BFIImm, DL, VT)); |
3176 | |
3177 | // Create the BFI/BFXIL instruction. |
3178 | SDValue Ops[] = {And.getOperand(i: 0), SDValue(MOVI, 0), |
3179 | CurDAG->getTargetConstant(Val: ImmR, DL, VT), |
3180 | CurDAG->getTargetConstant(Val: ImmS, DL, VT)}; |
3181 | unsigned Opc = (VT == MVT::i32) ? AArch64::BFMWri : AArch64::BFMXri; |
3182 | CurDAG->SelectNodeTo(N, MachineOpc: Opc, VT, Ops); |
3183 | return true; |
3184 | } |
3185 | |
3186 | static bool isWorthFoldingIntoOrrWithShift(SDValue Dst, SelectionDAG *CurDAG, |
3187 | SDValue &ShiftedOperand, |
3188 | uint64_t &EncodedShiftImm) { |
3189 | // Avoid folding Dst into ORR-with-shift if Dst has other uses than ORR. |
3190 | if (!Dst.hasOneUse()) |
3191 | return false; |
3192 | |
3193 | EVT VT = Dst.getValueType(); |
3194 | assert((VT == MVT::i32 || VT == MVT::i64) && |
3195 | "Caller should guarantee that VT is one of i32 or i64" ); |
3196 | const unsigned SizeInBits = VT.getSizeInBits(); |
3197 | |
3198 | SDLoc DL(Dst.getNode()); |
3199 | uint64_t AndImm, ShlImm; |
3200 | if (isOpcWithIntImmediate(N: Dst.getNode(), Opc: ISD::AND, Imm&: AndImm) && |
3201 | isShiftedMask_64(Value: AndImm)) { |
3202 | // Avoid transforming 'DstOp0' if it has other uses than the AND node. |
3203 | SDValue DstOp0 = Dst.getOperand(i: 0); |
3204 | if (!DstOp0.hasOneUse()) |
3205 | return false; |
3206 | |
3207 | // An example to illustrate the transformation |
3208 | // From: |
3209 | // lsr x8, x1, #1 |
3210 | // and x8, x8, #0x3f80 |
3211 | // bfxil x8, x1, #0, #7 |
3212 | // To: |
3213 | // and x8, x23, #0x7f |
3214 | // ubfx x9, x23, #8, #7 |
3215 | // orr x23, x8, x9, lsl #7 |
3216 | // |
3217 | // The number of instructions remains the same, but ORR is faster than BFXIL |
3218 | // on many AArch64 processors (or as good as BFXIL if not faster). Besides, |
3219 | // the dependency chain is improved after the transformation. |
3220 | uint64_t SrlImm; |
3221 | if (isOpcWithIntImmediate(N: DstOp0.getNode(), Opc: ISD::SRL, Imm&: SrlImm)) { |
3222 | uint64_t NumTrailingZeroInShiftedMask = llvm::countr_zero(Val: AndImm); |
3223 | if ((SrlImm + NumTrailingZeroInShiftedMask) < SizeInBits) { |
3224 | unsigned MaskWidth = |
3225 | llvm::countr_one(Value: AndImm >> NumTrailingZeroInShiftedMask); |
3226 | unsigned UBFMOpc = |
3227 | (VT == MVT::i32) ? AArch64::UBFMWri : AArch64::UBFMXri; |
3228 | SDNode *UBFMNode = CurDAG->getMachineNode( |
3229 | Opcode: UBFMOpc, dl: DL, VT, Op1: DstOp0.getOperand(i: 0), |
3230 | Op2: CurDAG->getTargetConstant(Val: SrlImm + NumTrailingZeroInShiftedMask, DL, |
3231 | VT), |
3232 | Op3: CurDAG->getTargetConstant( |
3233 | Val: SrlImm + NumTrailingZeroInShiftedMask + MaskWidth - 1, DL, VT)); |
3234 | ShiftedOperand = SDValue(UBFMNode, 0); |
3235 | EncodedShiftImm = AArch64_AM::getShifterImm( |
3236 | ST: AArch64_AM::LSL, Imm: NumTrailingZeroInShiftedMask); |
3237 | return true; |
3238 | } |
3239 | } |
3240 | return false; |
3241 | } |
3242 | |
3243 | if (isOpcWithIntImmediate(N: Dst.getNode(), Opc: ISD::SHL, Imm&: ShlImm)) { |
3244 | ShiftedOperand = Dst.getOperand(i: 0); |
3245 | EncodedShiftImm = AArch64_AM::getShifterImm(ST: AArch64_AM::LSL, Imm: ShlImm); |
3246 | return true; |
3247 | } |
3248 | |
3249 | uint64_t SrlImm; |
3250 | if (isOpcWithIntImmediate(N: Dst.getNode(), Opc: ISD::SRL, Imm&: SrlImm)) { |
3251 | ShiftedOperand = Dst.getOperand(i: 0); |
3252 | EncodedShiftImm = AArch64_AM::getShifterImm(ST: AArch64_AM::LSR, Imm: SrlImm); |
3253 | return true; |
3254 | } |
3255 | return false; |
3256 | } |
3257 | |
3258 | // Given an 'ISD::OR' node that is going to be selected as BFM, analyze |
3259 | // the operands and select it to AArch64::ORR with shifted registers if |
3260 | // that's more efficient. Returns true iff selection to AArch64::ORR happens. |
3261 | static bool tryOrrWithShift(SDNode *N, SDValue OrOpd0, SDValue OrOpd1, |
3262 | SDValue Src, SDValue Dst, SelectionDAG *CurDAG, |
3263 | const bool BiggerPattern) { |
3264 | EVT VT = N->getValueType(ResNo: 0); |
3265 | assert(N->getOpcode() == ISD::OR && "Expect N to be an OR node" ); |
3266 | assert(((N->getOperand(0) == OrOpd0 && N->getOperand(1) == OrOpd1) || |
3267 | (N->getOperand(1) == OrOpd0 && N->getOperand(0) == OrOpd1)) && |
3268 | "Expect OrOpd0 and OrOpd1 to be operands of ISD::OR" ); |
3269 | assert((VT == MVT::i32 || VT == MVT::i64) && |
3270 | "Expect result type to be i32 or i64 since N is combinable to BFM" ); |
3271 | SDLoc DL(N); |
3272 | |
3273 | // Bail out if BFM simplifies away one node in BFM Dst. |
3274 | if (OrOpd1 != Dst) |
3275 | return false; |
3276 | |
3277 | const unsigned OrrOpc = (VT == MVT::i32) ? AArch64::ORRWrs : AArch64::ORRXrs; |
3278 | // For "BFM Rd, Rn, #immr, #imms", it's known that BFM simplifies away fewer |
3279 | // nodes from Rn (or inserts additional shift node) if BiggerPattern is true. |
3280 | if (BiggerPattern) { |
3281 | uint64_t SrcAndImm; |
3282 | if (isOpcWithIntImmediate(N: OrOpd0.getNode(), Opc: ISD::AND, Imm&: SrcAndImm) && |
3283 | isMask_64(Value: SrcAndImm) && OrOpd0.getOperand(i: 0) == Src) { |
3284 | // OrOpd0 = AND Src, #Mask |
3285 | // So BFM simplifies away one AND node from Src and doesn't simplify away |
3286 | // nodes from Dst. If ORR with left-shifted operand also simplifies away |
3287 | // one node (from Rd), ORR is better since it has higher throughput and |
3288 | // smaller latency than BFM on many AArch64 processors (and for the rest |
3289 | // ORR is at least as good as BFM). |
3290 | SDValue ShiftedOperand; |
3291 | uint64_t EncodedShiftImm; |
3292 | if (isWorthFoldingIntoOrrWithShift(Dst, CurDAG, ShiftedOperand, |
3293 | EncodedShiftImm)) { |
3294 | SDValue Ops[] = {OrOpd0, ShiftedOperand, |
3295 | CurDAG->getTargetConstant(Val: EncodedShiftImm, DL, VT)}; |
3296 | CurDAG->SelectNodeTo(N, MachineOpc: OrrOpc, VT, Ops); |
3297 | return true; |
3298 | } |
3299 | } |
3300 | return false; |
3301 | } |
3302 | |
3303 | assert((!BiggerPattern) && "BiggerPattern should be handled above" ); |
3304 | |
3305 | uint64_t ShlImm; |
3306 | if (isOpcWithIntImmediate(N: OrOpd0.getNode(), Opc: ISD::SHL, Imm&: ShlImm)) { |
3307 | if (OrOpd0.getOperand(i: 0) == Src && OrOpd0.hasOneUse()) { |
3308 | SDValue Ops[] = { |
3309 | Dst, Src, |
3310 | CurDAG->getTargetConstant( |
3311 | Val: AArch64_AM::getShifterImm(ST: AArch64_AM::LSL, Imm: ShlImm), DL, VT)}; |
3312 | CurDAG->SelectNodeTo(N, MachineOpc: OrrOpc, VT, Ops); |
3313 | return true; |
3314 | } |
3315 | |
3316 | // Select the following pattern to left-shifted operand rather than BFI. |
3317 | // %val1 = op .. |
3318 | // %val2 = shl %val1, #imm |
3319 | // %res = or %val1, %val2 |
3320 | // |
3321 | // If N is selected to be BFI, we know that |
3322 | // 1) OrOpd0 would be the operand from which extract bits (i.e., folded into |
3323 | // BFI) 2) OrOpd1 would be the destination operand (i.e., preserved) |
3324 | // |
3325 | // Instead of selecting N to BFI, fold OrOpd0 as a left shift directly. |
3326 | if (OrOpd0.getOperand(i: 0) == OrOpd1) { |
3327 | SDValue Ops[] = { |
3328 | OrOpd1, OrOpd1, |
3329 | CurDAG->getTargetConstant( |
3330 | Val: AArch64_AM::getShifterImm(ST: AArch64_AM::LSL, Imm: ShlImm), DL, VT)}; |
3331 | CurDAG->SelectNodeTo(N, MachineOpc: OrrOpc, VT, Ops); |
3332 | return true; |
3333 | } |
3334 | } |
3335 | |
3336 | uint64_t SrlImm; |
3337 | if (isOpcWithIntImmediate(N: OrOpd0.getNode(), Opc: ISD::SRL, Imm&: SrlImm)) { |
3338 | // Select the following pattern to right-shifted operand rather than BFXIL. |
3339 | // %val1 = op .. |
3340 | // %val2 = lshr %val1, #imm |
3341 | // %res = or %val1, %val2 |
3342 | // |
3343 | // If N is selected to be BFXIL, we know that |
3344 | // 1) OrOpd0 would be the operand from which extract bits (i.e., folded into |
3345 | // BFXIL) 2) OrOpd1 would be the destination operand (i.e., preserved) |
3346 | // |
3347 | // Instead of selecting N to BFXIL, fold OrOpd0 as a right shift directly. |
3348 | if (OrOpd0.getOperand(i: 0) == OrOpd1) { |
3349 | SDValue Ops[] = { |
3350 | OrOpd1, OrOpd1, |
3351 | CurDAG->getTargetConstant( |
3352 | Val: AArch64_AM::getShifterImm(ST: AArch64_AM::LSR, Imm: SrlImm), DL, VT)}; |
3353 | CurDAG->SelectNodeTo(N, MachineOpc: OrrOpc, VT, Ops); |
3354 | return true; |
3355 | } |
3356 | } |
3357 | |
3358 | return false; |
3359 | } |
3360 | |
3361 | static bool tryBitfieldInsertOpFromOr(SDNode *N, const APInt &UsefulBits, |
3362 | SelectionDAG *CurDAG) { |
3363 | assert(N->getOpcode() == ISD::OR && "Expect a OR operation" ); |
3364 | |
3365 | EVT VT = N->getValueType(ResNo: 0); |
3366 | if (VT != MVT::i32 && VT != MVT::i64) |
3367 | return false; |
3368 | |
3369 | unsigned BitWidth = VT.getSizeInBits(); |
3370 | |
3371 | // Because of simplify-demanded-bits in DAGCombine, involved masks may not |
3372 | // have the expected shape. Try to undo that. |
3373 | |
3374 | unsigned NumberOfIgnoredLowBits = UsefulBits.countr_zero(); |
3375 | unsigned NumberOfIgnoredHighBits = UsefulBits.countl_zero(); |
3376 | |
3377 | // Given a OR operation, check if we have the following pattern |
3378 | // ubfm c, b, imm, imm2 (or something that does the same jobs, see |
3379 | // isBitfieldExtractOp) |
3380 | // d = e & mask2 ; where mask is a binary sequence of 1..10..0 and |
3381 | // countTrailingZeros(mask2) == imm2 - imm + 1 |
3382 | // f = d | c |
3383 | // if yes, replace the OR instruction with: |
3384 | // f = BFM Opd0, Opd1, LSB, MSB ; where LSB = imm, and MSB = imm2 |
3385 | |
3386 | // OR is commutative, check all combinations of operand order and values of |
3387 | // BiggerPattern, i.e. |
3388 | // Opd0, Opd1, BiggerPattern=false |
3389 | // Opd1, Opd0, BiggerPattern=false |
3390 | // Opd0, Opd1, BiggerPattern=true |
3391 | // Opd1, Opd0, BiggerPattern=true |
3392 | // Several of these combinations may match, so check with BiggerPattern=false |
3393 | // first since that will produce better results by matching more instructions |
3394 | // and/or inserting fewer extra instructions. |
3395 | for (int I = 0; I < 4; ++I) { |
3396 | |
3397 | SDValue Dst, Src; |
3398 | unsigned ImmR, ImmS; |
3399 | bool BiggerPattern = I / 2; |
3400 | SDValue OrOpd0Val = N->getOperand(Num: I % 2); |
3401 | SDNode *OrOpd0 = OrOpd0Val.getNode(); |
3402 | SDValue OrOpd1Val = N->getOperand(Num: (I + 1) % 2); |
3403 | SDNode *OrOpd1 = OrOpd1Val.getNode(); |
3404 | |
3405 | unsigned BFXOpc; |
3406 | int DstLSB, Width; |
3407 | if (isBitfieldExtractOp(CurDAG, N: OrOpd0, Opc&: BFXOpc, Opd0&: Src, Immr&: ImmR, Imms&: ImmS, |
3408 | NumberOfIgnoredLowBits, BiggerPattern)) { |
3409 | // Check that the returned opcode is compatible with the pattern, |
3410 | // i.e., same type and zero extended (U and not S) |
3411 | if ((BFXOpc != AArch64::UBFMXri && VT == MVT::i64) || |
3412 | (BFXOpc != AArch64::UBFMWri && VT == MVT::i32)) |
3413 | continue; |
3414 | |
3415 | // Compute the width of the bitfield insertion |
3416 | DstLSB = 0; |
3417 | Width = ImmS - ImmR + 1; |
3418 | // FIXME: This constraint is to catch bitfield insertion we may |
3419 | // want to widen the pattern if we want to grab general bitfied |
3420 | // move case |
3421 | if (Width <= 0) |
3422 | continue; |
3423 | |
3424 | // If the mask on the insertee is correct, we have a BFXIL operation. We |
3425 | // can share the ImmR and ImmS values from the already-computed UBFM. |
3426 | } else if (isBitfieldPositioningOp(CurDAG, Op: OrOpd0Val, |
3427 | BiggerPattern, |
3428 | Src, DstLSB, Width)) { |
3429 | ImmR = (BitWidth - DstLSB) % BitWidth; |
3430 | ImmS = Width - 1; |
3431 | } else |
3432 | continue; |
3433 | |
3434 | // Check the second part of the pattern |
3435 | EVT VT = OrOpd1Val.getValueType(); |
3436 | assert((VT == MVT::i32 || VT == MVT::i64) && "unexpected OR operand" ); |
3437 | |
3438 | // Compute the Known Zero for the candidate of the first operand. |
3439 | // This allows to catch more general case than just looking for |
3440 | // AND with imm. Indeed, simplify-demanded-bits may have removed |
3441 | // the AND instruction because it proves it was useless. |
3442 | KnownBits Known = CurDAG->computeKnownBits(Op: OrOpd1Val); |
3443 | |
3444 | // Check if there is enough room for the second operand to appear |
3445 | // in the first one |
3446 | APInt BitsToBeInserted = |
3447 | APInt::getBitsSet(numBits: Known.getBitWidth(), loBit: DstLSB, hiBit: DstLSB + Width); |
3448 | |
3449 | if ((BitsToBeInserted & ~Known.Zero) != 0) |
3450 | continue; |
3451 | |
3452 | // Set the first operand |
3453 | uint64_t Imm; |
3454 | if (isOpcWithIntImmediate(N: OrOpd1, Opc: ISD::AND, Imm) && |
3455 | isBitfieldDstMask(DstMask: Imm, BitsToBeInserted, NumberOfIgnoredHighBits, VT)) |
3456 | // In that case, we can eliminate the AND |
3457 | Dst = OrOpd1->getOperand(Num: 0); |
3458 | else |
3459 | // Maybe the AND has been removed by simplify-demanded-bits |
3460 | // or is useful because it discards more bits |
3461 | Dst = OrOpd1Val; |
3462 | |
3463 | // Before selecting ISD::OR node to AArch64::BFM, see if an AArch64::ORR |
3464 | // with shifted operand is more efficient. |
3465 | if (tryOrrWithShift(N, OrOpd0: OrOpd0Val, OrOpd1: OrOpd1Val, Src, Dst, CurDAG, |
3466 | BiggerPattern)) |
3467 | return true; |
3468 | |
3469 | // both parts match |
3470 | SDLoc DL(N); |
3471 | SDValue Ops[] = {Dst, Src, CurDAG->getTargetConstant(Val: ImmR, DL, VT), |
3472 | CurDAG->getTargetConstant(Val: ImmS, DL, VT)}; |
3473 | unsigned Opc = (VT == MVT::i32) ? AArch64::BFMWri : AArch64::BFMXri; |
3474 | CurDAG->SelectNodeTo(N, MachineOpc: Opc, VT, Ops); |
3475 | return true; |
3476 | } |
3477 | |
3478 | // Generate a BFXIL from 'or (and X, Mask0Imm), (and Y, Mask1Imm)' iff |
3479 | // Mask0Imm and ~Mask1Imm are equivalent and one of the MaskImms is a shifted |
3480 | // mask (e.g., 0x000ffff0). |
3481 | uint64_t Mask0Imm, Mask1Imm; |
3482 | SDValue And0 = N->getOperand(Num: 0); |
3483 | SDValue And1 = N->getOperand(Num: 1); |
3484 | if (And0.hasOneUse() && And1.hasOneUse() && |
3485 | isOpcWithIntImmediate(N: And0.getNode(), Opc: ISD::AND, Imm&: Mask0Imm) && |
3486 | isOpcWithIntImmediate(N: And1.getNode(), Opc: ISD::AND, Imm&: Mask1Imm) && |
3487 | APInt(BitWidth, Mask0Imm) == ~APInt(BitWidth, Mask1Imm) && |
3488 | (isShiftedMask(Mask: Mask0Imm, VT) || isShiftedMask(Mask: Mask1Imm, VT))) { |
3489 | |
3490 | // ORR is commutative, so canonicalize to the form 'or (and X, Mask0Imm), |
3491 | // (and Y, Mask1Imm)' where Mask1Imm is the shifted mask masking off the |
3492 | // bits to be inserted. |
3493 | if (isShiftedMask(Mask: Mask0Imm, VT)) { |
3494 | std::swap(a&: And0, b&: And1); |
3495 | std::swap(a&: Mask0Imm, b&: Mask1Imm); |
3496 | } |
3497 | |
3498 | SDValue Src = And1->getOperand(Num: 0); |
3499 | SDValue Dst = And0->getOperand(Num: 0); |
3500 | unsigned LSB = llvm::countr_zero(Val: Mask1Imm); |
3501 | int Width = BitWidth - APInt(BitWidth, Mask0Imm).popcount(); |
3502 | |
3503 | // The BFXIL inserts the low-order bits from a source register, so right |
3504 | // shift the needed bits into place. |
3505 | SDLoc DL(N); |
3506 | unsigned ShiftOpc = (VT == MVT::i32) ? AArch64::UBFMWri : AArch64::UBFMXri; |
3507 | uint64_t LsrImm = LSB; |
3508 | if (Src->hasOneUse() && |
3509 | isOpcWithIntImmediate(N: Src.getNode(), Opc: ISD::SRL, Imm&: LsrImm) && |
3510 | (LsrImm + LSB) < BitWidth) { |
3511 | Src = Src->getOperand(Num: 0); |
3512 | LsrImm += LSB; |
3513 | } |
3514 | |
3515 | SDNode *LSR = CurDAG->getMachineNode( |
3516 | Opcode: ShiftOpc, dl: DL, VT, Op1: Src, Op2: CurDAG->getTargetConstant(Val: LsrImm, DL, VT), |
3517 | Op3: CurDAG->getTargetConstant(Val: BitWidth - 1, DL, VT)); |
3518 | |
3519 | // BFXIL is an alias of BFM, so translate to BFM operands. |
3520 | unsigned ImmR = (BitWidth - LSB) % BitWidth; |
3521 | unsigned ImmS = Width - 1; |
3522 | |
3523 | // Create the BFXIL instruction. |
3524 | SDValue Ops[] = {Dst, SDValue(LSR, 0), |
3525 | CurDAG->getTargetConstant(Val: ImmR, DL, VT), |
3526 | CurDAG->getTargetConstant(Val: ImmS, DL, VT)}; |
3527 | unsigned Opc = (VT == MVT::i32) ? AArch64::BFMWri : AArch64::BFMXri; |
3528 | CurDAG->SelectNodeTo(N, MachineOpc: Opc, VT, Ops); |
3529 | return true; |
3530 | } |
3531 | |
3532 | return false; |
3533 | } |
3534 | |
3535 | bool AArch64DAGToDAGISel::tryBitfieldInsertOp(SDNode *N) { |
3536 | if (N->getOpcode() != ISD::OR) |
3537 | return false; |
3538 | |
3539 | APInt NUsefulBits; |
3540 | getUsefulBits(Op: SDValue(N, 0), UsefulBits&: NUsefulBits); |
3541 | |
3542 | // If all bits are not useful, just return UNDEF. |
3543 | if (!NUsefulBits) { |
3544 | CurDAG->SelectNodeTo(N, MachineOpc: TargetOpcode::IMPLICIT_DEF, VT: N->getValueType(ResNo: 0)); |
3545 | return true; |
3546 | } |
3547 | |
3548 | if (tryBitfieldInsertOpFromOr(N, UsefulBits: NUsefulBits, CurDAG)) |
3549 | return true; |
3550 | |
3551 | return tryBitfieldInsertOpFromOrAndImm(N, CurDAG); |
3552 | } |
3553 | |
3554 | /// SelectBitfieldInsertInZeroOp - Match a UBFIZ instruction that is the |
3555 | /// equivalent of a left shift by a constant amount followed by an and masking |
3556 | /// out a contiguous set of bits. |
3557 | bool AArch64DAGToDAGISel::tryBitfieldInsertInZeroOp(SDNode *N) { |
3558 | if (N->getOpcode() != ISD::AND) |
3559 | return false; |
3560 | |
3561 | EVT VT = N->getValueType(ResNo: 0); |
3562 | if (VT != MVT::i32 && VT != MVT::i64) |
3563 | return false; |
3564 | |
3565 | SDValue Op0; |
3566 | int DstLSB, Width; |
3567 | if (!isBitfieldPositioningOp(CurDAG, Op: SDValue(N, 0), /*BiggerPattern=*/false, |
3568 | Src&: Op0, DstLSB, Width)) |
3569 | return false; |
3570 | |
3571 | // ImmR is the rotate right amount. |
3572 | unsigned ImmR = (VT.getSizeInBits() - DstLSB) % VT.getSizeInBits(); |
3573 | // ImmS is the most significant bit of the source to be moved. |
3574 | unsigned ImmS = Width - 1; |
3575 | |
3576 | SDLoc DL(N); |
3577 | SDValue Ops[] = {Op0, CurDAG->getTargetConstant(Val: ImmR, DL, VT), |
3578 | CurDAG->getTargetConstant(Val: ImmS, DL, VT)}; |
3579 | unsigned Opc = (VT == MVT::i32) ? AArch64::UBFMWri : AArch64::UBFMXri; |
3580 | CurDAG->SelectNodeTo(N, MachineOpc: Opc, VT, Ops); |
3581 | return true; |
3582 | } |
3583 | |
3584 | /// tryShiftAmountMod - Take advantage of built-in mod of shift amount in |
3585 | /// variable shift/rotate instructions. |
3586 | bool AArch64DAGToDAGISel::tryShiftAmountMod(SDNode *N) { |
3587 | EVT VT = N->getValueType(ResNo: 0); |
3588 | |
3589 | unsigned Opc; |
3590 | switch (N->getOpcode()) { |
3591 | case ISD::ROTR: |
3592 | Opc = (VT == MVT::i32) ? AArch64::RORVWr : AArch64::RORVXr; |
3593 | break; |
3594 | case ISD::SHL: |
3595 | Opc = (VT == MVT::i32) ? AArch64::LSLVWr : AArch64::LSLVXr; |
3596 | break; |
3597 | case ISD::SRL: |
3598 | Opc = (VT == MVT::i32) ? AArch64::LSRVWr : AArch64::LSRVXr; |
3599 | break; |
3600 | case ISD::SRA: |
3601 | Opc = (VT == MVT::i32) ? AArch64::ASRVWr : AArch64::ASRVXr; |
3602 | break; |
3603 | default: |
3604 | return false; |
3605 | } |
3606 | |
3607 | uint64_t Size; |
3608 | uint64_t Bits; |
3609 | if (VT == MVT::i32) { |
3610 | Bits = 5; |
3611 | Size = 32; |
3612 | } else if (VT == MVT::i64) { |
3613 | Bits = 6; |
3614 | Size = 64; |
3615 | } else |
3616 | return false; |
3617 | |
3618 | SDValue ShiftAmt = N->getOperand(Num: 1); |
3619 | SDLoc DL(N); |
3620 | SDValue NewShiftAmt; |
3621 | |
3622 | // Skip over an extend of the shift amount. |
3623 | if (ShiftAmt->getOpcode() == ISD::ZERO_EXTEND || |
3624 | ShiftAmt->getOpcode() == ISD::ANY_EXTEND) |
3625 | ShiftAmt = ShiftAmt->getOperand(Num: 0); |
3626 | |
3627 | if (ShiftAmt->getOpcode() == ISD::ADD || ShiftAmt->getOpcode() == ISD::SUB) { |
3628 | SDValue Add0 = ShiftAmt->getOperand(Num: 0); |
3629 | SDValue Add1 = ShiftAmt->getOperand(Num: 1); |
3630 | uint64_t Add0Imm; |
3631 | uint64_t Add1Imm; |
3632 | if (isIntImmediate(N: Add1, Imm&: Add1Imm) && (Add1Imm % Size == 0)) { |
3633 | // If we are shifting by X+/-N where N == 0 mod Size, then just shift by X |
3634 | // to avoid the ADD/SUB. |
3635 | NewShiftAmt = Add0; |
3636 | } else if (ShiftAmt->getOpcode() == ISD::SUB && |
3637 | isIntImmediate(N: Add0, Imm&: Add0Imm) && Add0Imm != 0 && |
3638 | (Add0Imm % Size == 0)) { |
3639 | // If we are shifting by N-X where N == 0 mod Size, then just shift by -X |
3640 | // to generate a NEG instead of a SUB from a constant. |
3641 | unsigned NegOpc; |
3642 | unsigned ZeroReg; |
3643 | EVT SubVT = ShiftAmt->getValueType(ResNo: 0); |
3644 | if (SubVT == MVT::i32) { |
3645 | NegOpc = AArch64::SUBWrr; |
3646 | ZeroReg = AArch64::WZR; |
3647 | } else { |
3648 | assert(SubVT == MVT::i64); |
3649 | NegOpc = AArch64::SUBXrr; |
3650 | ZeroReg = AArch64::XZR; |
3651 | } |
3652 | SDValue Zero = |
3653 | CurDAG->getCopyFromReg(Chain: CurDAG->getEntryNode(), dl: DL, Reg: ZeroReg, VT: SubVT); |
3654 | MachineSDNode *Neg = |
3655 | CurDAG->getMachineNode(Opcode: NegOpc, dl: DL, VT: SubVT, Op1: Zero, Op2: Add1); |
3656 | NewShiftAmt = SDValue(Neg, 0); |
3657 | } else if (ShiftAmt->getOpcode() == ISD::SUB && |
3658 | isIntImmediate(N: Add0, Imm&: Add0Imm) && (Add0Imm % Size == Size - 1)) { |
3659 | // If we are shifting by N-X where N == -1 mod Size, then just shift by ~X |
3660 | // to generate a NOT instead of a SUB from a constant. |
3661 | unsigned NotOpc; |
3662 | unsigned ZeroReg; |
3663 | EVT SubVT = ShiftAmt->getValueType(ResNo: 0); |
3664 | if (SubVT == MVT::i32) { |
3665 | NotOpc = AArch64::ORNWrr; |
3666 | ZeroReg = AArch64::WZR; |
3667 | } else { |
3668 | assert(SubVT == MVT::i64); |
3669 | NotOpc = AArch64::ORNXrr; |
3670 | ZeroReg = AArch64::XZR; |
3671 | } |
3672 | SDValue Zero = |
3673 | CurDAG->getCopyFromReg(Chain: CurDAG->getEntryNode(), dl: DL, Reg: ZeroReg, VT: SubVT); |
3674 | MachineSDNode *Not = |
3675 | CurDAG->getMachineNode(Opcode: NotOpc, dl: DL, VT: SubVT, Op1: Zero, Op2: Add1); |
3676 | NewShiftAmt = SDValue(Not, 0); |
3677 | } else |
3678 | return false; |
3679 | } else { |
3680 | // If the shift amount is masked with an AND, check that the mask covers the |
3681 | // bits that are implicitly ANDed off by the above opcodes and if so, skip |
3682 | // the AND. |
3683 | uint64_t MaskImm; |
3684 | if (!isOpcWithIntImmediate(N: ShiftAmt.getNode(), Opc: ISD::AND, Imm&: MaskImm) && |
3685 | !isOpcWithIntImmediate(N: ShiftAmt.getNode(), Opc: AArch64ISD::ANDS, Imm&: MaskImm)) |
3686 | return false; |
3687 | |
3688 | if ((unsigned)llvm::countr_one(Value: MaskImm) < Bits) |
3689 | return false; |
3690 | |
3691 | NewShiftAmt = ShiftAmt->getOperand(Num: 0); |
3692 | } |
3693 | |
3694 | // Narrow/widen the shift amount to match the size of the shift operation. |
3695 | if (VT == MVT::i32) |
3696 | NewShiftAmt = narrowIfNeeded(CurDAG, N: NewShiftAmt); |
3697 | else if (VT == MVT::i64 && NewShiftAmt->getValueType(0) == MVT::i32) { |
3698 | SDValue SubReg = CurDAG->getTargetConstant(AArch64::sub_32, DL, MVT::i32); |
3699 | MachineSDNode *Ext = CurDAG->getMachineNode( |
3700 | AArch64::SUBREG_TO_REG, DL, VT, |
3701 | CurDAG->getTargetConstant(0, DL, MVT::i64), NewShiftAmt, SubReg); |
3702 | NewShiftAmt = SDValue(Ext, 0); |
3703 | } |
3704 | |
3705 | SDValue Ops[] = {N->getOperand(Num: 0), NewShiftAmt}; |
3706 | CurDAG->SelectNodeTo(N, MachineOpc: Opc, VT, Ops); |
3707 | return true; |
3708 | } |
3709 | |
3710 | static bool checkCVTFixedPointOperandWithFBits(SelectionDAG *CurDAG, SDValue N, |
3711 | SDValue &FixedPos, |
3712 | unsigned RegWidth, |
3713 | bool isReciprocal) { |
3714 | APFloat FVal(0.0); |
3715 | if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(Val&: N)) |
3716 | FVal = CN->getValueAPF(); |
3717 | else if (LoadSDNode *LN = dyn_cast<LoadSDNode>(Val&: N)) { |
3718 | // Some otherwise illegal constants are allowed in this case. |
3719 | if (LN->getOperand(Num: 1).getOpcode() != AArch64ISD::ADDlow || |
3720 | !isa<ConstantPoolSDNode>(Val: LN->getOperand(Num: 1)->getOperand(Num: 1))) |
3721 | return false; |
3722 | |
3723 | ConstantPoolSDNode *CN = |
3724 | dyn_cast<ConstantPoolSDNode>(Val: LN->getOperand(Num: 1)->getOperand(Num: 1)); |
3725 | FVal = cast<ConstantFP>(Val: CN->getConstVal())->getValueAPF(); |
3726 | } else |
3727 | return false; |
3728 | |
3729 | // An FCVT[SU] instruction performs: convertToInt(Val * 2^fbits) where fbits |
3730 | // is between 1 and 32 for a destination w-register, or 1 and 64 for an |
3731 | // x-register. |
3732 | // |
3733 | // By this stage, we've detected (fp_to_[su]int (fmul Val, THIS_NODE)) so we |
3734 | // want THIS_NODE to be 2^fbits. This is much easier to deal with using |
3735 | // integers. |
3736 | bool IsExact; |
3737 | |
3738 | if (isReciprocal) |
3739 | if (!FVal.getExactInverse(inv: &FVal)) |
3740 | return false; |
3741 | |
3742 | // fbits is between 1 and 64 in the worst-case, which means the fmul |
3743 | // could have 2^64 as an actual operand. Need 65 bits of precision. |
3744 | APSInt IntVal(65, true); |
3745 | FVal.convertToInteger(Result&: IntVal, RM: APFloat::rmTowardZero, IsExact: &IsExact); |
3746 | |
3747 | // N.b. isPowerOf2 also checks for > 0. |
3748 | if (!IsExact || !IntVal.isPowerOf2()) |
3749 | return false; |
3750 | unsigned FBits = IntVal.logBase2(); |
3751 | |
3752 | // Checks above should have guaranteed that we haven't lost information in |
3753 | // finding FBits, but it must still be in range. |
3754 | if (FBits == 0 || FBits > RegWidth) return false; |
3755 | |
3756 | FixedPos = CurDAG->getTargetConstant(FBits, SDLoc(N), MVT::i32); |
3757 | return true; |
3758 | } |
3759 | |
3760 | bool AArch64DAGToDAGISel::SelectCVTFixedPosOperand(SDValue N, SDValue &FixedPos, |
3761 | unsigned RegWidth) { |
3762 | return checkCVTFixedPointOperandWithFBits(CurDAG, N, FixedPos, RegWidth, |
3763 | isReciprocal: false); |
3764 | } |
3765 | |
3766 | bool AArch64DAGToDAGISel::SelectCVTFixedPosRecipOperand(SDValue N, |
3767 | SDValue &FixedPos, |
3768 | unsigned RegWidth) { |
3769 | return checkCVTFixedPointOperandWithFBits(CurDAG, N, FixedPos, RegWidth, |
3770 | isReciprocal: true); |
3771 | } |
3772 | |
3773 | // Inspects a register string of the form o0:op1:CRn:CRm:op2 gets the fields |
3774 | // of the string and obtains the integer values from them and combines these |
3775 | // into a single value to be used in the MRS/MSR instruction. |
3776 | static int getIntOperandFromRegisterString(StringRef RegString) { |
3777 | SmallVector<StringRef, 5> Fields; |
3778 | RegString.split(A&: Fields, Separator: ':'); |
3779 | |
3780 | if (Fields.size() == 1) |
3781 | return -1; |
3782 | |
3783 | assert(Fields.size() == 5 |
3784 | && "Invalid number of fields in read register string" ); |
3785 | |
3786 | SmallVector<int, 5> Ops; |
3787 | bool AllIntFields = true; |
3788 | |
3789 | for (StringRef Field : Fields) { |
3790 | unsigned IntField; |
3791 | AllIntFields &= !Field.getAsInteger(Radix: 10, Result&: IntField); |
3792 | Ops.push_back(Elt: IntField); |
3793 | } |
3794 | |
3795 | assert(AllIntFields && |
3796 | "Unexpected non-integer value in special register string." ); |
3797 | (void)AllIntFields; |
3798 | |
3799 | // Need to combine the integer fields of the string into a single value |
3800 | // based on the bit encoding of MRS/MSR instruction. |
3801 | return (Ops[0] << 14) | (Ops[1] << 11) | (Ops[2] << 7) | |
3802 | (Ops[3] << 3) | (Ops[4]); |
3803 | } |
3804 | |
3805 | // Lower the read_register intrinsic to an MRS instruction node if the special |
3806 | // register string argument is either of the form detailed in the ALCE (the |
3807 | // form described in getIntOperandsFromRegsterString) or is a named register |
3808 | // known by the MRS SysReg mapper. |
3809 | bool AArch64DAGToDAGISel::tryReadRegister(SDNode *N) { |
3810 | const auto *MD = cast<MDNodeSDNode>(Val: N->getOperand(Num: 1)); |
3811 | const auto *RegString = cast<MDString>(Val: MD->getMD()->getOperand(I: 0)); |
3812 | SDLoc DL(N); |
3813 | |
3814 | bool ReadIs128Bit = N->getOpcode() == AArch64ISD::MRRS; |
3815 | |
3816 | unsigned Opcode64Bit = AArch64::MRS; |
3817 | int Imm = getIntOperandFromRegisterString(RegString: RegString->getString()); |
3818 | if (Imm == -1) { |
3819 | // No match, Use the sysreg mapper to map the remaining possible strings to |
3820 | // the value for the register to be used for the instruction operand. |
3821 | const auto *TheReg = |
3822 | AArch64SysReg::lookupSysRegByName(RegString->getString()); |
3823 | if (TheReg && TheReg->Readable && |
3824 | TheReg->haveFeatures(ActiveFeatures: Subtarget->getFeatureBits())) |
3825 | Imm = TheReg->Encoding; |
3826 | else |
3827 | Imm = AArch64SysReg::parseGenericRegister(Name: RegString->getString()); |
3828 | |
3829 | if (Imm == -1) { |
3830 | // Still no match, see if this is "pc" or give up. |
3831 | if (!ReadIs128Bit && RegString->getString() == "pc" ) { |
3832 | Opcode64Bit = AArch64::ADR; |
3833 | Imm = 0; |
3834 | } else { |
3835 | return false; |
3836 | } |
3837 | } |
3838 | } |
3839 | |
3840 | SDValue InChain = N->getOperand(Num: 0); |
3841 | SDValue SysRegImm = CurDAG->getTargetConstant(Imm, DL, MVT::i32); |
3842 | if (!ReadIs128Bit) { |
3843 | CurDAG->SelectNodeTo(N, Opcode64Bit, MVT::i64, MVT::Other /* Chain */, |
3844 | {SysRegImm, InChain}); |
3845 | } else { |
3846 | SDNode *MRRS = CurDAG->getMachineNode( |
3847 | AArch64::MRRS, DL, |
3848 | {MVT::Untyped /* XSeqPair */, MVT::Other /* Chain */}, |
3849 | {SysRegImm, InChain}); |
3850 | |
3851 | // Sysregs are not endian. The even register always contains the low half |
3852 | // of the register. |
3853 | SDValue Lo = CurDAG->getTargetExtractSubreg(AArch64::sube64, DL, MVT::i64, |
3854 | SDValue(MRRS, 0)); |
3855 | SDValue Hi = CurDAG->getTargetExtractSubreg(AArch64::subo64, DL, MVT::i64, |
3856 | SDValue(MRRS, 0)); |
3857 | SDValue OutChain = SDValue(MRRS, 1); |
3858 | |
3859 | ReplaceUses(F: SDValue(N, 0), T: Lo); |
3860 | ReplaceUses(F: SDValue(N, 1), T: Hi); |
3861 | ReplaceUses(F: SDValue(N, 2), T: OutChain); |
3862 | }; |
3863 | return true; |
3864 | } |
3865 | |
3866 | // Lower the write_register intrinsic to an MSR instruction node if the special |
3867 | // register string argument is either of the form detailed in the ALCE (the |
3868 | // form described in getIntOperandsFromRegsterString) or is a named register |
3869 | // known by the MSR SysReg mapper. |
3870 | bool AArch64DAGToDAGISel::tryWriteRegister(SDNode *N) { |
3871 | const auto *MD = cast<MDNodeSDNode>(Val: N->getOperand(Num: 1)); |
3872 | const auto *RegString = cast<MDString>(Val: MD->getMD()->getOperand(I: 0)); |
3873 | SDLoc DL(N); |
3874 | |
3875 | bool WriteIs128Bit = N->getOpcode() == AArch64ISD::MSRR; |
3876 | |
3877 | if (!WriteIs128Bit) { |
3878 | // Check if the register was one of those allowed as the pstatefield value |
3879 | // in the MSR (immediate) instruction. To accept the values allowed in the |
3880 | // pstatefield for the MSR (immediate) instruction, we also require that an |
3881 | // immediate value has been provided as an argument, we know that this is |
3882 | // the case as it has been ensured by semantic checking. |
3883 | auto trySelectPState = [&](auto PMapper, unsigned State) { |
3884 | if (PMapper) { |
3885 | assert(isa<ConstantSDNode>(N->getOperand(2)) && |
3886 | "Expected a constant integer expression." ); |
3887 | unsigned Reg = PMapper->Encoding; |
3888 | uint64_t Immed = N->getConstantOperandVal(Num: 2); |
3889 | CurDAG->SelectNodeTo( |
3890 | N, State, MVT::Other, CurDAG->getTargetConstant(Reg, DL, MVT::i32), |
3891 | CurDAG->getTargetConstant(Immed, DL, MVT::i16), N->getOperand(0)); |
3892 | return true; |
3893 | } |
3894 | return false; |
3895 | }; |
3896 | |
3897 | if (trySelectPState( |
3898 | AArch64PState::lookupPStateImm0_15ByName(RegString->getString()), |
3899 | AArch64::MSRpstateImm4)) |
3900 | return true; |
3901 | if (trySelectPState( |
3902 | AArch64PState::lookupPStateImm0_1ByName(RegString->getString()), |
3903 | AArch64::MSRpstateImm1)) |
3904 | return true; |
3905 | } |
3906 | |
3907 | int Imm = getIntOperandFromRegisterString(RegString: RegString->getString()); |
3908 | if (Imm == -1) { |
3909 | // Use the sysreg mapper to attempt to map the remaining possible strings |
3910 | // to the value for the register to be used for the MSR (register) |
3911 | // instruction operand. |
3912 | auto TheReg = AArch64SysReg::lookupSysRegByName(RegString->getString()); |
3913 | if (TheReg && TheReg->Writeable && |
3914 | TheReg->haveFeatures(ActiveFeatures: Subtarget->getFeatureBits())) |
3915 | Imm = TheReg->Encoding; |
3916 | else |
3917 | Imm = AArch64SysReg::parseGenericRegister(Name: RegString->getString()); |
3918 | |
3919 | if (Imm == -1) |
3920 | return false; |
3921 | } |
3922 | |
3923 | SDValue InChain = N->getOperand(Num: 0); |
3924 | if (!WriteIs128Bit) { |
3925 | CurDAG->SelectNodeTo(N, AArch64::MSR, MVT::Other, |
3926 | CurDAG->getTargetConstant(Imm, DL, MVT::i32), |
3927 | N->getOperand(2), InChain); |
3928 | } else { |
3929 | // No endian swap. The lower half always goes into the even subreg, and the |
3930 | // higher half always into the odd supreg. |
3931 | SDNode *Pair = CurDAG->getMachineNode( |
3932 | TargetOpcode::REG_SEQUENCE, DL, MVT::Untyped /* XSeqPair */, |
3933 | {CurDAG->getTargetConstant(AArch64::XSeqPairsClassRegClass.getID(), DL, |
3934 | MVT::i32), |
3935 | N->getOperand(2), |
3936 | CurDAG->getTargetConstant(AArch64::sube64, DL, MVT::i32), |
3937 | N->getOperand(3), |
3938 | CurDAG->getTargetConstant(AArch64::subo64, DL, MVT::i32)}); |
3939 | |
3940 | CurDAG->SelectNodeTo(N, AArch64::MSRR, MVT::Other, |
3941 | CurDAG->getTargetConstant(Imm, DL, MVT::i32), |
3942 | SDValue(Pair, 0), InChain); |
3943 | } |
3944 | |
3945 | return true; |
3946 | } |
3947 | |
3948 | /// We've got special pseudo-instructions for these |
3949 | bool AArch64DAGToDAGISel::SelectCMP_SWAP(SDNode *N) { |
3950 | unsigned Opcode; |
3951 | EVT MemTy = cast<MemSDNode>(Val: N)->getMemoryVT(); |
3952 | |
3953 | // Leave IR for LSE if subtarget supports it. |
3954 | if (Subtarget->hasLSE()) return false; |
3955 | |
3956 | if (MemTy == MVT::i8) |
3957 | Opcode = AArch64::CMP_SWAP_8; |
3958 | else if (MemTy == MVT::i16) |
3959 | Opcode = AArch64::CMP_SWAP_16; |
3960 | else if (MemTy == MVT::i32) |
3961 | Opcode = AArch64::CMP_SWAP_32; |
3962 | else if (MemTy == MVT::i64) |
3963 | Opcode = AArch64::CMP_SWAP_64; |
3964 | else |
3965 | llvm_unreachable("Unknown AtomicCmpSwap type" ); |
3966 | |
3967 | MVT RegTy = MemTy == MVT::i64 ? MVT::i64 : MVT::i32; |
3968 | SDValue Ops[] = {N->getOperand(Num: 1), N->getOperand(Num: 2), N->getOperand(Num: 3), |
3969 | N->getOperand(Num: 0)}; |
3970 | SDNode *CmpSwap = CurDAG->getMachineNode( |
3971 | Opcode, SDLoc(N), |
3972 | CurDAG->getVTList(RegTy, MVT::i32, MVT::Other), Ops); |
3973 | |
3974 | MachineMemOperand *MemOp = cast<MemSDNode>(Val: N)->getMemOperand(); |
3975 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: CmpSwap), NewMemRefs: {MemOp}); |
3976 | |
3977 | ReplaceUses(F: SDValue(N, 0), T: SDValue(CmpSwap, 0)); |
3978 | ReplaceUses(F: SDValue(N, 1), T: SDValue(CmpSwap, 2)); |
3979 | CurDAG->RemoveDeadNode(N); |
3980 | |
3981 | return true; |
3982 | } |
3983 | |
3984 | bool AArch64DAGToDAGISel::SelectSVEAddSubImm(SDValue N, MVT VT, SDValue &Imm, |
3985 | SDValue &Shift) { |
3986 | if (!isa<ConstantSDNode>(Val: N)) |
3987 | return false; |
3988 | |
3989 | SDLoc DL(N); |
3990 | uint64_t Val = cast<ConstantSDNode>(Val&: N) |
3991 | ->getAPIntValue() |
3992 | .trunc(width: VT.getFixedSizeInBits()) |
3993 | .getZExtValue(); |
3994 | |
3995 | switch (VT.SimpleTy) { |
3996 | case MVT::i8: |
3997 | // All immediates are supported. |
3998 | Shift = CurDAG->getTargetConstant(0, DL, MVT::i32); |
3999 | Imm = CurDAG->getTargetConstant(Val, DL, MVT::i32); |
4000 | return true; |
4001 | case MVT::i16: |
4002 | case MVT::i32: |
4003 | case MVT::i64: |
4004 | // Support 8bit unsigned immediates. |
4005 | if (Val <= 255) { |
4006 | Shift = CurDAG->getTargetConstant(0, DL, MVT::i32); |
4007 | Imm = CurDAG->getTargetConstant(Val, DL, MVT::i32); |
4008 | return true; |
4009 | } |
4010 | // Support 16bit unsigned immediates that are a multiple of 256. |
4011 | if (Val <= 65280 && Val % 256 == 0) { |
4012 | Shift = CurDAG->getTargetConstant(8, DL, MVT::i32); |
4013 | Imm = CurDAG->getTargetConstant(Val >> 8, DL, MVT::i32); |
4014 | return true; |
4015 | } |
4016 | break; |
4017 | default: |
4018 | break; |
4019 | } |
4020 | |
4021 | return false; |
4022 | } |
4023 | |
4024 | bool AArch64DAGToDAGISel::SelectSVEAddSubSSatImm(SDValue N, MVT VT, |
4025 | SDValue &Imm, SDValue &Shift, |
4026 | bool Negate) { |
4027 | if (!isa<ConstantSDNode>(Val: N)) |
4028 | return false; |
4029 | |
4030 | SDLoc DL(N); |
4031 | int64_t Val = cast<ConstantSDNode>(Val&: N) |
4032 | ->getAPIntValue() |
4033 | .trunc(width: VT.getFixedSizeInBits()) |
4034 | .getSExtValue(); |
4035 | |
4036 | if (Negate) |
4037 | Val = -Val; |
4038 | |
4039 | // Signed saturating instructions treat their immediate operand as unsigned, |
4040 | // whereas the related intrinsics define their operands to be signed. This |
4041 | // means we can only use the immediate form when the operand is non-negative. |
4042 | if (Val < 0) |
4043 | return false; |
4044 | |
4045 | switch (VT.SimpleTy) { |
4046 | case MVT::i8: |
4047 | // All positive immediates are supported. |
4048 | Shift = CurDAG->getTargetConstant(0, DL, MVT::i32); |
4049 | Imm = CurDAG->getTargetConstant(Val, DL, MVT::i32); |
4050 | return true; |
4051 | case MVT::i16: |
4052 | case MVT::i32: |
4053 | case MVT::i64: |
4054 | // Support 8bit positive immediates. |
4055 | if (Val <= 255) { |
4056 | Shift = CurDAG->getTargetConstant(0, DL, MVT::i32); |
4057 | Imm = CurDAG->getTargetConstant(Val, DL, MVT::i32); |
4058 | return true; |
4059 | } |
4060 | // Support 16bit positive immediates that are a multiple of 256. |
4061 | if (Val <= 65280 && Val % 256 == 0) { |
4062 | Shift = CurDAG->getTargetConstant(8, DL, MVT::i32); |
4063 | Imm = CurDAG->getTargetConstant(Val >> 8, DL, MVT::i32); |
4064 | return true; |
4065 | } |
4066 | break; |
4067 | default: |
4068 | break; |
4069 | } |
4070 | |
4071 | return false; |
4072 | } |
4073 | |
4074 | bool AArch64DAGToDAGISel::SelectSVECpyDupImm(SDValue N, MVT VT, SDValue &Imm, |
4075 | SDValue &Shift) { |
4076 | if (!isa<ConstantSDNode>(Val: N)) |
4077 | return false; |
4078 | |
4079 | SDLoc DL(N); |
4080 | int64_t Val = cast<ConstantSDNode>(Val&: N) |
4081 | ->getAPIntValue() |
4082 | .trunc(width: VT.getFixedSizeInBits()) |
4083 | .getSExtValue(); |
4084 | |
4085 | switch (VT.SimpleTy) { |
4086 | case MVT::i8: |
4087 | // All immediates are supported. |
4088 | Shift = CurDAG->getTargetConstant(0, DL, MVT::i32); |
4089 | Imm = CurDAG->getTargetConstant(Val & 0xFF, DL, MVT::i32); |
4090 | return true; |
4091 | case MVT::i16: |
4092 | case MVT::i32: |
4093 | case MVT::i64: |
4094 | // Support 8bit signed immediates. |
4095 | if (Val >= -128 && Val <= 127) { |
4096 | Shift = CurDAG->getTargetConstant(0, DL, MVT::i32); |
4097 | Imm = CurDAG->getTargetConstant(Val & 0xFF, DL, MVT::i32); |
4098 | return true; |
4099 | } |
4100 | // Support 16bit signed immediates that are a multiple of 256. |
4101 | if (Val >= -32768 && Val <= 32512 && Val % 256 == 0) { |
4102 | Shift = CurDAG->getTargetConstant(8, DL, MVT::i32); |
4103 | Imm = CurDAG->getTargetConstant((Val >> 8) & 0xFF, DL, MVT::i32); |
4104 | return true; |
4105 | } |
4106 | break; |
4107 | default: |
4108 | break; |
4109 | } |
4110 | |
4111 | return false; |
4112 | } |
4113 | |
4114 | bool AArch64DAGToDAGISel::SelectSVESignedArithImm(SDValue N, SDValue &Imm) { |
4115 | if (auto CNode = dyn_cast<ConstantSDNode>(Val&: N)) { |
4116 | int64_t ImmVal = CNode->getSExtValue(); |
4117 | SDLoc DL(N); |
4118 | if (ImmVal >= -128 && ImmVal < 128) { |
4119 | Imm = CurDAG->getTargetConstant(ImmVal, DL, MVT::i32); |
4120 | return true; |
4121 | } |
4122 | } |
4123 | return false; |
4124 | } |
4125 | |
4126 | bool AArch64DAGToDAGISel::SelectSVEArithImm(SDValue N, MVT VT, SDValue &Imm) { |
4127 | if (auto CNode = dyn_cast<ConstantSDNode>(Val&: N)) { |
4128 | uint64_t ImmVal = CNode->getZExtValue(); |
4129 | |
4130 | switch (VT.SimpleTy) { |
4131 | case MVT::i8: |
4132 | ImmVal &= 0xFF; |
4133 | break; |
4134 | case MVT::i16: |
4135 | ImmVal &= 0xFFFF; |
4136 | break; |
4137 | case MVT::i32: |
4138 | ImmVal &= 0xFFFFFFFF; |
4139 | break; |
4140 | case MVT::i64: |
4141 | break; |
4142 | default: |
4143 | llvm_unreachable("Unexpected type" ); |
4144 | } |
4145 | |
4146 | if (ImmVal < 256) { |
4147 | Imm = CurDAG->getTargetConstant(ImmVal, SDLoc(N), MVT::i32); |
4148 | return true; |
4149 | } |
4150 | } |
4151 | return false; |
4152 | } |
4153 | |
4154 | bool AArch64DAGToDAGISel::SelectSVELogicalImm(SDValue N, MVT VT, SDValue &Imm, |
4155 | bool Invert) { |
4156 | if (auto CNode = dyn_cast<ConstantSDNode>(Val&: N)) { |
4157 | uint64_t ImmVal = CNode->getZExtValue(); |
4158 | SDLoc DL(N); |
4159 | |
4160 | if (Invert) |
4161 | ImmVal = ~ImmVal; |
4162 | |
4163 | // Shift mask depending on type size. |
4164 | switch (VT.SimpleTy) { |
4165 | case MVT::i8: |
4166 | ImmVal &= 0xFF; |
4167 | ImmVal |= ImmVal << 8; |
4168 | ImmVal |= ImmVal << 16; |
4169 | ImmVal |= ImmVal << 32; |
4170 | break; |
4171 | case MVT::i16: |
4172 | ImmVal &= 0xFFFF; |
4173 | ImmVal |= ImmVal << 16; |
4174 | ImmVal |= ImmVal << 32; |
4175 | break; |
4176 | case MVT::i32: |
4177 | ImmVal &= 0xFFFFFFFF; |
4178 | ImmVal |= ImmVal << 32; |
4179 | break; |
4180 | case MVT::i64: |
4181 | break; |
4182 | default: |
4183 | llvm_unreachable("Unexpected type" ); |
4184 | } |
4185 | |
4186 | uint64_t encoding; |
4187 | if (AArch64_AM::processLogicalImmediate(Imm: ImmVal, RegSize: 64, Encoding&: encoding)) { |
4188 | Imm = CurDAG->getTargetConstant(encoding, DL, MVT::i64); |
4189 | return true; |
4190 | } |
4191 | } |
4192 | return false; |
4193 | } |
4194 | |
4195 | // SVE shift intrinsics allow shift amounts larger than the element's bitwidth. |
4196 | // Rather than attempt to normalise everything we can sometimes saturate the |
4197 | // shift amount during selection. This function also allows for consistent |
4198 | // isel patterns by ensuring the resulting "Imm" node is of the i32 type |
4199 | // required by the instructions. |
4200 | bool AArch64DAGToDAGISel::SelectSVEShiftImm(SDValue N, uint64_t Low, |
4201 | uint64_t High, bool AllowSaturation, |
4202 | SDValue &Imm) { |
4203 | if (auto *CN = dyn_cast<ConstantSDNode>(Val&: N)) { |
4204 | uint64_t ImmVal = CN->getZExtValue(); |
4205 | |
4206 | // Reject shift amounts that are too small. |
4207 | if (ImmVal < Low) |
4208 | return false; |
4209 | |
4210 | // Reject or saturate shift amounts that are too big. |
4211 | if (ImmVal > High) { |
4212 | if (!AllowSaturation) |
4213 | return false; |
4214 | ImmVal = High; |
4215 | } |
4216 | |
4217 | Imm = CurDAG->getTargetConstant(ImmVal, SDLoc(N), MVT::i32); |
4218 | return true; |
4219 | } |
4220 | |
4221 | return false; |
4222 | } |
4223 | |
4224 | bool AArch64DAGToDAGISel::trySelectStackSlotTagP(SDNode *N) { |
4225 | // tagp(FrameIndex, IRGstack, tag_offset): |
4226 | // since the offset between FrameIndex and IRGstack is a compile-time |
4227 | // constant, this can be lowered to a single ADDG instruction. |
4228 | if (!(isa<FrameIndexSDNode>(Val: N->getOperand(Num: 1)))) { |
4229 | return false; |
4230 | } |
4231 | |
4232 | SDValue IRG_SP = N->getOperand(Num: 2); |
4233 | if (IRG_SP->getOpcode() != ISD::INTRINSIC_W_CHAIN || |
4234 | IRG_SP->getConstantOperandVal(1) != Intrinsic::aarch64_irg_sp) { |
4235 | return false; |
4236 | } |
4237 | |
4238 | const TargetLowering *TLI = getTargetLowering(); |
4239 | SDLoc DL(N); |
4240 | int FI = cast<FrameIndexSDNode>(Val: N->getOperand(Num: 1))->getIndex(); |
4241 | SDValue FiOp = CurDAG->getTargetFrameIndex( |
4242 | FI, VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
4243 | int TagOffset = N->getConstantOperandVal(Num: 3); |
4244 | |
4245 | SDNode *Out = CurDAG->getMachineNode( |
4246 | AArch64::TAGPstack, DL, MVT::i64, |
4247 | {FiOp, CurDAG->getTargetConstant(0, DL, MVT::i64), N->getOperand(2), |
4248 | CurDAG->getTargetConstant(TagOffset, DL, MVT::i64)}); |
4249 | ReplaceNode(F: N, T: Out); |
4250 | return true; |
4251 | } |
4252 | |
4253 | void AArch64DAGToDAGISel::SelectTagP(SDNode *N) { |
4254 | assert(isa<ConstantSDNode>(N->getOperand(3)) && |
4255 | "llvm.aarch64.tagp third argument must be an immediate" ); |
4256 | if (trySelectStackSlotTagP(N)) |
4257 | return; |
4258 | // FIXME: above applies in any case when offset between Op1 and Op2 is a |
4259 | // compile-time constant, not just for stack allocations. |
4260 | |
4261 | // General case for unrelated pointers in Op1 and Op2. |
4262 | SDLoc DL(N); |
4263 | int TagOffset = N->getConstantOperandVal(Num: 3); |
4264 | SDNode *N1 = CurDAG->getMachineNode(AArch64::SUBP, DL, MVT::i64, |
4265 | {N->getOperand(1), N->getOperand(2)}); |
4266 | SDNode *N2 = CurDAG->getMachineNode(AArch64::ADDXrr, DL, MVT::i64, |
4267 | {SDValue(N1, 0), N->getOperand(2)}); |
4268 | SDNode *N3 = CurDAG->getMachineNode( |
4269 | AArch64::ADDG, DL, MVT::i64, |
4270 | {SDValue(N2, 0), CurDAG->getTargetConstant(0, DL, MVT::i64), |
4271 | CurDAG->getTargetConstant(TagOffset, DL, MVT::i64)}); |
4272 | ReplaceNode(F: N, T: N3); |
4273 | } |
4274 | |
4275 | bool AArch64DAGToDAGISel::trySelectCastFixedLengthToScalableVector(SDNode *N) { |
4276 | assert(N->getOpcode() == ISD::INSERT_SUBVECTOR && "Invalid Node!" ); |
4277 | |
4278 | // Bail when not a "cast" like insert_subvector. |
4279 | if (N->getConstantOperandVal(Num: 2) != 0) |
4280 | return false; |
4281 | if (!N->getOperand(Num: 0).isUndef()) |
4282 | return false; |
4283 | |
4284 | // Bail when normal isel should do the job. |
4285 | EVT VT = N->getValueType(ResNo: 0); |
4286 | EVT InVT = N->getOperand(Num: 1).getValueType(); |
4287 | if (VT.isFixedLengthVector() || InVT.isScalableVector()) |
4288 | return false; |
4289 | if (InVT.getSizeInBits() <= 128) |
4290 | return false; |
4291 | |
4292 | // NOTE: We can only get here when doing fixed length SVE code generation. |
4293 | // We do manual selection because the types involved are not linked to real |
4294 | // registers (despite being legal) and must be coerced into SVE registers. |
4295 | |
4296 | assert(VT.getSizeInBits().getKnownMinValue() == AArch64::SVEBitsPerBlock && |
4297 | "Expected to insert into a packed scalable vector!" ); |
4298 | |
4299 | SDLoc DL(N); |
4300 | auto RC = CurDAG->getTargetConstant(AArch64::ZPRRegClassID, DL, MVT::i64); |
4301 | ReplaceNode(F: N, T: CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS, DL, VT, |
4302 | N->getOperand(Num: 1), RC)); |
4303 | return true; |
4304 | } |
4305 | |
4306 | bool AArch64DAGToDAGISel::trySelectCastScalableToFixedLengthVector(SDNode *N) { |
4307 | assert(N->getOpcode() == ISD::EXTRACT_SUBVECTOR && "Invalid Node!" ); |
4308 | |
4309 | // Bail when not a "cast" like extract_subvector. |
4310 | if (N->getConstantOperandVal(Num: 1) != 0) |
4311 | return false; |
4312 | |
4313 | // Bail when normal isel can do the job. |
4314 | EVT VT = N->getValueType(ResNo: 0); |
4315 | EVT InVT = N->getOperand(Num: 0).getValueType(); |
4316 | if (VT.isScalableVector() || InVT.isFixedLengthVector()) |
4317 | return false; |
4318 | if (VT.getSizeInBits() <= 128) |
4319 | return false; |
4320 | |
4321 | // NOTE: We can only get here when doing fixed length SVE code generation. |
4322 | // We do manual selection because the types involved are not linked to real |
4323 | // registers (despite being legal) and must be coerced into SVE registers. |
4324 | |
4325 | assert(InVT.getSizeInBits().getKnownMinValue() == AArch64::SVEBitsPerBlock && |
4326 | "Expected to extract from a packed scalable vector!" ); |
4327 | |
4328 | SDLoc DL(N); |
4329 | auto RC = CurDAG->getTargetConstant(AArch64::ZPRRegClassID, DL, MVT::i64); |
4330 | ReplaceNode(F: N, T: CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS, DL, VT, |
4331 | N->getOperand(Num: 0), RC)); |
4332 | return true; |
4333 | } |
4334 | |
4335 | bool AArch64DAGToDAGISel::trySelectXAR(SDNode *N) { |
4336 | assert(N->getOpcode() == ISD::OR && "Expected OR instruction" ); |
4337 | |
4338 | SDValue N0 = N->getOperand(Num: 0); |
4339 | SDValue N1 = N->getOperand(Num: 1); |
4340 | EVT VT = N->getValueType(ResNo: 0); |
4341 | |
4342 | // Essentially: rotr (xor(x, y), imm) -> xar (x, y, imm) |
4343 | // Rotate by a constant is a funnel shift in IR which is exanded to |
4344 | // an OR with shifted operands. |
4345 | // We do the following transform: |
4346 | // OR N0, N1 -> xar (x, y, imm) |
4347 | // Where: |
4348 | // N1 = SRL_PRED true, V, splat(imm) --> rotr amount |
4349 | // N0 = SHL_PRED true, V, splat(bits-imm) |
4350 | // V = (xor x, y) |
4351 | if (VT.isScalableVector() && Subtarget->hasSVE2orSME()) { |
4352 | if (N0.getOpcode() != AArch64ISD::SHL_PRED || |
4353 | N1.getOpcode() != AArch64ISD::SRL_PRED) |
4354 | std::swap(a&: N0, b&: N1); |
4355 | if (N0.getOpcode() != AArch64ISD::SHL_PRED || |
4356 | N1.getOpcode() != AArch64ISD::SRL_PRED) |
4357 | return false; |
4358 | |
4359 | auto *TLI = static_cast<const AArch64TargetLowering *>(getTargetLowering()); |
4360 | if (!TLI->isAllActivePredicate(DAG&: *CurDAG, N: N0.getOperand(i: 0)) || |
4361 | !TLI->isAllActivePredicate(DAG&: *CurDAG, N: N1.getOperand(i: 0))) |
4362 | return false; |
4363 | |
4364 | SDValue XOR = N0.getOperand(i: 1); |
4365 | if (XOR.getOpcode() != ISD::XOR || XOR != N1.getOperand(i: 1)) |
4366 | return false; |
4367 | |
4368 | APInt ShlAmt, ShrAmt; |
4369 | if (!ISD::isConstantSplatVector(N: N0.getOperand(i: 2).getNode(), SplatValue&: ShlAmt) || |
4370 | !ISD::isConstantSplatVector(N: N1.getOperand(i: 2).getNode(), SplatValue&: ShrAmt)) |
4371 | return false; |
4372 | |
4373 | if (ShlAmt + ShrAmt != VT.getScalarSizeInBits()) |
4374 | return false; |
4375 | |
4376 | SDLoc DL(N); |
4377 | SDValue Imm = |
4378 | CurDAG->getTargetConstant(ShrAmt.getZExtValue(), DL, MVT::i32); |
4379 | |
4380 | SDValue Ops[] = {XOR.getOperand(i: 0), XOR.getOperand(i: 1), Imm}; |
4381 | if (auto Opc = SelectOpcodeFromVT<SelectTypeKind::Int>( |
4382 | VT, {AArch64::XAR_ZZZI_B, AArch64::XAR_ZZZI_H, AArch64::XAR_ZZZI_S, |
4383 | AArch64::XAR_ZZZI_D})) { |
4384 | CurDAG->SelectNodeTo(N, Opc, VT, Ops); |
4385 | return true; |
4386 | } |
4387 | return false; |
4388 | } |
4389 | |
4390 | if (!Subtarget->hasSHA3()) |
4391 | return false; |
4392 | |
4393 | if (N0->getOpcode() != AArch64ISD::VSHL || |
4394 | N1->getOpcode() != AArch64ISD::VLSHR) |
4395 | return false; |
4396 | |
4397 | if (N0->getOperand(Num: 0) != N1->getOperand(Num: 0) || |
4398 | N1->getOperand(Num: 0)->getOpcode() != ISD::XOR) |
4399 | return false; |
4400 | |
4401 | SDValue XOR = N0.getOperand(i: 0); |
4402 | SDValue R1 = XOR.getOperand(i: 0); |
4403 | SDValue R2 = XOR.getOperand(i: 1); |
4404 | |
4405 | unsigned HsAmt = N0.getConstantOperandVal(i: 1); |
4406 | unsigned ShAmt = N1.getConstantOperandVal(i: 1); |
4407 | |
4408 | SDLoc DL = SDLoc(N0.getOperand(i: 1)); |
4409 | SDValue Imm = CurDAG->getTargetConstant( |
4410 | Val: ShAmt, DL, VT: N0.getOperand(i: 1).getValueType(), isOpaque: false); |
4411 | |
4412 | if (ShAmt + HsAmt != 64) |
4413 | return false; |
4414 | |
4415 | SDValue Ops[] = {R1, R2, Imm}; |
4416 | CurDAG->SelectNodeTo(N, AArch64::XAR, N0.getValueType(), Ops); |
4417 | |
4418 | return true; |
4419 | } |
4420 | |
4421 | void AArch64DAGToDAGISel::Select(SDNode *Node) { |
4422 | // If we have a custom node, we already have selected! |
4423 | if (Node->isMachineOpcode()) { |
4424 | LLVM_DEBUG(errs() << "== " ; Node->dump(CurDAG); errs() << "\n" ); |
4425 | Node->setNodeId(-1); |
4426 | return; |
4427 | } |
4428 | |
4429 | // Few custom selection stuff. |
4430 | EVT VT = Node->getValueType(ResNo: 0); |
4431 | |
4432 | switch (Node->getOpcode()) { |
4433 | default: |
4434 | break; |
4435 | |
4436 | case ISD::ATOMIC_CMP_SWAP: |
4437 | if (SelectCMP_SWAP(N: Node)) |
4438 | return; |
4439 | break; |
4440 | |
4441 | case ISD::READ_REGISTER: |
4442 | case AArch64ISD::MRRS: |
4443 | if (tryReadRegister(N: Node)) |
4444 | return; |
4445 | break; |
4446 | |
4447 | case ISD::WRITE_REGISTER: |
4448 | case AArch64ISD::MSRR: |
4449 | if (tryWriteRegister(N: Node)) |
4450 | return; |
4451 | break; |
4452 | |
4453 | case ISD::LOAD: { |
4454 | // Try to select as an indexed load. Fall through to normal processing |
4455 | // if we can't. |
4456 | if (tryIndexedLoad(N: Node)) |
4457 | return; |
4458 | break; |
4459 | } |
4460 | |
4461 | case ISD::SRL: |
4462 | case ISD::AND: |
4463 | case ISD::SRA: |
4464 | case ISD::SIGN_EXTEND_INREG: |
4465 | if (tryBitfieldExtractOp(N: Node)) |
4466 | return; |
4467 | if (tryBitfieldInsertInZeroOp(N: Node)) |
4468 | return; |
4469 | [[fallthrough]]; |
4470 | case ISD::ROTR: |
4471 | case ISD::SHL: |
4472 | if (tryShiftAmountMod(N: Node)) |
4473 | return; |
4474 | break; |
4475 | |
4476 | case ISD::SIGN_EXTEND: |
4477 | if (tryBitfieldExtractOpFromSExt(N: Node)) |
4478 | return; |
4479 | break; |
4480 | |
4481 | case ISD::OR: |
4482 | if (tryBitfieldInsertOp(N: Node)) |
4483 | return; |
4484 | if (trySelectXAR(N: Node)) |
4485 | return; |
4486 | break; |
4487 | |
4488 | case ISD::EXTRACT_SUBVECTOR: { |
4489 | if (trySelectCastScalableToFixedLengthVector(N: Node)) |
4490 | return; |
4491 | break; |
4492 | } |
4493 | |
4494 | case ISD::INSERT_SUBVECTOR: { |
4495 | if (trySelectCastFixedLengthToScalableVector(N: Node)) |
4496 | return; |
4497 | break; |
4498 | } |
4499 | |
4500 | case ISD::Constant: { |
4501 | // Materialize zero constants as copies from WZR/XZR. This allows |
4502 | // the coalescer to propagate these into other instructions. |
4503 | ConstantSDNode *ConstNode = cast<ConstantSDNode>(Val: Node); |
4504 | if (ConstNode->isZero()) { |
4505 | if (VT == MVT::i32) { |
4506 | SDValue New = CurDAG->getCopyFromReg( |
4507 | CurDAG->getEntryNode(), SDLoc(Node), AArch64::WZR, MVT::i32); |
4508 | ReplaceNode(F: Node, T: New.getNode()); |
4509 | return; |
4510 | } else if (VT == MVT::i64) { |
4511 | SDValue New = CurDAG->getCopyFromReg( |
4512 | CurDAG->getEntryNode(), SDLoc(Node), AArch64::XZR, MVT::i64); |
4513 | ReplaceNode(F: Node, T: New.getNode()); |
4514 | return; |
4515 | } |
4516 | } |
4517 | break; |
4518 | } |
4519 | |
4520 | case ISD::FrameIndex: { |
4521 | // Selects to ADDXri FI, 0 which in turn will become ADDXri SP, imm. |
4522 | int FI = cast<FrameIndexSDNode>(Val: Node)->getIndex(); |
4523 | unsigned Shifter = AArch64_AM::getShifterImm(ST: AArch64_AM::LSL, Imm: 0); |
4524 | const TargetLowering *TLI = getTargetLowering(); |
4525 | SDValue TFI = CurDAG->getTargetFrameIndex( |
4526 | FI, VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
4527 | SDLoc DL(Node); |
4528 | SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, DL, MVT::i32), |
4529 | CurDAG->getTargetConstant(Shifter, DL, MVT::i32) }; |
4530 | CurDAG->SelectNodeTo(Node, AArch64::ADDXri, MVT::i64, Ops); |
4531 | return; |
4532 | } |
4533 | case ISD::INTRINSIC_W_CHAIN: { |
4534 | unsigned IntNo = Node->getConstantOperandVal(Num: 1); |
4535 | switch (IntNo) { |
4536 | default: |
4537 | break; |
4538 | case Intrinsic::aarch64_ldaxp: |
4539 | case Intrinsic::aarch64_ldxp: { |
4540 | unsigned Op = |
4541 | IntNo == Intrinsic::aarch64_ldaxp ? AArch64::LDAXPX : AArch64::LDXPX; |
4542 | SDValue MemAddr = Node->getOperand(Num: 2); |
4543 | SDLoc DL(Node); |
4544 | SDValue Chain = Node->getOperand(Num: 0); |
4545 | |
4546 | SDNode *Ld = CurDAG->getMachineNode(Op, DL, MVT::i64, MVT::i64, |
4547 | MVT::Other, MemAddr, Chain); |
4548 | |
4549 | // Transfer memoperands. |
4550 | MachineMemOperand *MemOp = |
4551 | cast<MemIntrinsicSDNode>(Val: Node)->getMemOperand(); |
4552 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: Ld), NewMemRefs: {MemOp}); |
4553 | ReplaceNode(F: Node, T: Ld); |
4554 | return; |
4555 | } |
4556 | case Intrinsic::aarch64_stlxp: |
4557 | case Intrinsic::aarch64_stxp: { |
4558 | unsigned Op = |
4559 | IntNo == Intrinsic::aarch64_stlxp ? AArch64::STLXPX : AArch64::STXPX; |
4560 | SDLoc DL(Node); |
4561 | SDValue Chain = Node->getOperand(Num: 0); |
4562 | SDValue ValLo = Node->getOperand(Num: 2); |
4563 | SDValue ValHi = Node->getOperand(Num: 3); |
4564 | SDValue MemAddr = Node->getOperand(Num: 4); |
4565 | |
4566 | // Place arguments in the right order. |
4567 | SDValue Ops[] = {ValLo, ValHi, MemAddr, Chain}; |
4568 | |
4569 | SDNode *St = CurDAG->getMachineNode(Op, DL, MVT::i32, MVT::Other, Ops); |
4570 | // Transfer memoperands. |
4571 | MachineMemOperand *MemOp = |
4572 | cast<MemIntrinsicSDNode>(Val: Node)->getMemOperand(); |
4573 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: St), NewMemRefs: {MemOp}); |
4574 | |
4575 | ReplaceNode(F: Node, T: St); |
4576 | return; |
4577 | } |
4578 | case Intrinsic::aarch64_neon_ld1x2: |
4579 | if (VT == MVT::v8i8) { |
4580 | SelectLoad(Node, 2, AArch64::LD1Twov8b, AArch64::dsub0); |
4581 | return; |
4582 | } else if (VT == MVT::v16i8) { |
4583 | SelectLoad(Node, 2, AArch64::LD1Twov16b, AArch64::qsub0); |
4584 | return; |
4585 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
4586 | SelectLoad(Node, 2, AArch64::LD1Twov4h, AArch64::dsub0); |
4587 | return; |
4588 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
4589 | SelectLoad(Node, 2, AArch64::LD1Twov8h, AArch64::qsub0); |
4590 | return; |
4591 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
4592 | SelectLoad(Node, 2, AArch64::LD1Twov2s, AArch64::dsub0); |
4593 | return; |
4594 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
4595 | SelectLoad(Node, 2, AArch64::LD1Twov4s, AArch64::qsub0); |
4596 | return; |
4597 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
4598 | SelectLoad(Node, 2, AArch64::LD1Twov1d, AArch64::dsub0); |
4599 | return; |
4600 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
4601 | SelectLoad(Node, 2, AArch64::LD1Twov2d, AArch64::qsub0); |
4602 | return; |
4603 | } |
4604 | break; |
4605 | case Intrinsic::aarch64_neon_ld1x3: |
4606 | if (VT == MVT::v8i8) { |
4607 | SelectLoad(Node, 3, AArch64::LD1Threev8b, AArch64::dsub0); |
4608 | return; |
4609 | } else if (VT == MVT::v16i8) { |
4610 | SelectLoad(Node, 3, AArch64::LD1Threev16b, AArch64::qsub0); |
4611 | return; |
4612 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
4613 | SelectLoad(Node, 3, AArch64::LD1Threev4h, AArch64::dsub0); |
4614 | return; |
4615 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
4616 | SelectLoad(Node, 3, AArch64::LD1Threev8h, AArch64::qsub0); |
4617 | return; |
4618 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
4619 | SelectLoad(Node, 3, AArch64::LD1Threev2s, AArch64::dsub0); |
4620 | return; |
4621 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
4622 | SelectLoad(Node, 3, AArch64::LD1Threev4s, AArch64::qsub0); |
4623 | return; |
4624 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
4625 | SelectLoad(Node, 3, AArch64::LD1Threev1d, AArch64::dsub0); |
4626 | return; |
4627 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
4628 | SelectLoad(Node, 3, AArch64::LD1Threev2d, AArch64::qsub0); |
4629 | return; |
4630 | } |
4631 | break; |
4632 | case Intrinsic::aarch64_neon_ld1x4: |
4633 | if (VT == MVT::v8i8) { |
4634 | SelectLoad(Node, 4, AArch64::LD1Fourv8b, AArch64::dsub0); |
4635 | return; |
4636 | } else if (VT == MVT::v16i8) { |
4637 | SelectLoad(Node, 4, AArch64::LD1Fourv16b, AArch64::qsub0); |
4638 | return; |
4639 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
4640 | SelectLoad(Node, 4, AArch64::LD1Fourv4h, AArch64::dsub0); |
4641 | return; |
4642 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
4643 | SelectLoad(Node, 4, AArch64::LD1Fourv8h, AArch64::qsub0); |
4644 | return; |
4645 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
4646 | SelectLoad(Node, 4, AArch64::LD1Fourv2s, AArch64::dsub0); |
4647 | return; |
4648 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
4649 | SelectLoad(Node, 4, AArch64::LD1Fourv4s, AArch64::qsub0); |
4650 | return; |
4651 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
4652 | SelectLoad(Node, 4, AArch64::LD1Fourv1d, AArch64::dsub0); |
4653 | return; |
4654 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
4655 | SelectLoad(Node, 4, AArch64::LD1Fourv2d, AArch64::qsub0); |
4656 | return; |
4657 | } |
4658 | break; |
4659 | case Intrinsic::aarch64_neon_ld2: |
4660 | if (VT == MVT::v8i8) { |
4661 | SelectLoad(Node, 2, AArch64::LD2Twov8b, AArch64::dsub0); |
4662 | return; |
4663 | } else if (VT == MVT::v16i8) { |
4664 | SelectLoad(Node, 2, AArch64::LD2Twov16b, AArch64::qsub0); |
4665 | return; |
4666 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
4667 | SelectLoad(Node, 2, AArch64::LD2Twov4h, AArch64::dsub0); |
4668 | return; |
4669 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
4670 | SelectLoad(Node, 2, AArch64::LD2Twov8h, AArch64::qsub0); |
4671 | return; |
4672 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
4673 | SelectLoad(Node, 2, AArch64::LD2Twov2s, AArch64::dsub0); |
4674 | return; |
4675 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
4676 | SelectLoad(Node, 2, AArch64::LD2Twov4s, AArch64::qsub0); |
4677 | return; |
4678 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
4679 | SelectLoad(Node, 2, AArch64::LD1Twov1d, AArch64::dsub0); |
4680 | return; |
4681 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
4682 | SelectLoad(Node, 2, AArch64::LD2Twov2d, AArch64::qsub0); |
4683 | return; |
4684 | } |
4685 | break; |
4686 | case Intrinsic::aarch64_neon_ld3: |
4687 | if (VT == MVT::v8i8) { |
4688 | SelectLoad(Node, 3, AArch64::LD3Threev8b, AArch64::dsub0); |
4689 | return; |
4690 | } else if (VT == MVT::v16i8) { |
4691 | SelectLoad(Node, 3, AArch64::LD3Threev16b, AArch64::qsub0); |
4692 | return; |
4693 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
4694 | SelectLoad(Node, 3, AArch64::LD3Threev4h, AArch64::dsub0); |
4695 | return; |
4696 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
4697 | SelectLoad(Node, 3, AArch64::LD3Threev8h, AArch64::qsub0); |
4698 | return; |
4699 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
4700 | SelectLoad(Node, 3, AArch64::LD3Threev2s, AArch64::dsub0); |
4701 | return; |
4702 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
4703 | SelectLoad(Node, 3, AArch64::LD3Threev4s, AArch64::qsub0); |
4704 | return; |
4705 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
4706 | SelectLoad(Node, 3, AArch64::LD1Threev1d, AArch64::dsub0); |
4707 | return; |
4708 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
4709 | SelectLoad(Node, 3, AArch64::LD3Threev2d, AArch64::qsub0); |
4710 | return; |
4711 | } |
4712 | break; |
4713 | case Intrinsic::aarch64_neon_ld4: |
4714 | if (VT == MVT::v8i8) { |
4715 | SelectLoad(Node, 4, AArch64::LD4Fourv8b, AArch64::dsub0); |
4716 | return; |
4717 | } else if (VT == MVT::v16i8) { |
4718 | SelectLoad(Node, 4, AArch64::LD4Fourv16b, AArch64::qsub0); |
4719 | return; |
4720 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
4721 | SelectLoad(Node, 4, AArch64::LD4Fourv4h, AArch64::dsub0); |
4722 | return; |
4723 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
4724 | SelectLoad(Node, 4, AArch64::LD4Fourv8h, AArch64::qsub0); |
4725 | return; |
4726 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
4727 | SelectLoad(Node, 4, AArch64::LD4Fourv2s, AArch64::dsub0); |
4728 | return; |
4729 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
4730 | SelectLoad(Node, 4, AArch64::LD4Fourv4s, AArch64::qsub0); |
4731 | return; |
4732 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
4733 | SelectLoad(Node, 4, AArch64::LD1Fourv1d, AArch64::dsub0); |
4734 | return; |
4735 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
4736 | SelectLoad(Node, 4, AArch64::LD4Fourv2d, AArch64::qsub0); |
4737 | return; |
4738 | } |
4739 | break; |
4740 | case Intrinsic::aarch64_neon_ld2r: |
4741 | if (VT == MVT::v8i8) { |
4742 | SelectLoad(Node, 2, AArch64::LD2Rv8b, AArch64::dsub0); |
4743 | return; |
4744 | } else if (VT == MVT::v16i8) { |
4745 | SelectLoad(Node, 2, AArch64::LD2Rv16b, AArch64::qsub0); |
4746 | return; |
4747 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
4748 | SelectLoad(Node, 2, AArch64::LD2Rv4h, AArch64::dsub0); |
4749 | return; |
4750 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
4751 | SelectLoad(Node, 2, AArch64::LD2Rv8h, AArch64::qsub0); |
4752 | return; |
4753 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
4754 | SelectLoad(Node, 2, AArch64::LD2Rv2s, AArch64::dsub0); |
4755 | return; |
4756 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
4757 | SelectLoad(Node, 2, AArch64::LD2Rv4s, AArch64::qsub0); |
4758 | return; |
4759 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
4760 | SelectLoad(Node, 2, AArch64::LD2Rv1d, AArch64::dsub0); |
4761 | return; |
4762 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
4763 | SelectLoad(Node, 2, AArch64::LD2Rv2d, AArch64::qsub0); |
4764 | return; |
4765 | } |
4766 | break; |
4767 | case Intrinsic::aarch64_neon_ld3r: |
4768 | if (VT == MVT::v8i8) { |
4769 | SelectLoad(Node, 3, AArch64::LD3Rv8b, AArch64::dsub0); |
4770 | return; |
4771 | } else if (VT == MVT::v16i8) { |
4772 | SelectLoad(Node, 3, AArch64::LD3Rv16b, AArch64::qsub0); |
4773 | return; |
4774 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
4775 | SelectLoad(Node, 3, AArch64::LD3Rv4h, AArch64::dsub0); |
4776 | return; |
4777 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
4778 | SelectLoad(Node, 3, AArch64::LD3Rv8h, AArch64::qsub0); |
4779 | return; |
4780 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
4781 | SelectLoad(Node, 3, AArch64::LD3Rv2s, AArch64::dsub0); |
4782 | return; |
4783 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
4784 | SelectLoad(Node, 3, AArch64::LD3Rv4s, AArch64::qsub0); |
4785 | return; |
4786 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
4787 | SelectLoad(Node, 3, AArch64::LD3Rv1d, AArch64::dsub0); |
4788 | return; |
4789 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
4790 | SelectLoad(Node, 3, AArch64::LD3Rv2d, AArch64::qsub0); |
4791 | return; |
4792 | } |
4793 | break; |
4794 | case Intrinsic::aarch64_neon_ld4r: |
4795 | if (VT == MVT::v8i8) { |
4796 | SelectLoad(Node, 4, AArch64::LD4Rv8b, AArch64::dsub0); |
4797 | return; |
4798 | } else if (VT == MVT::v16i8) { |
4799 | SelectLoad(Node, 4, AArch64::LD4Rv16b, AArch64::qsub0); |
4800 | return; |
4801 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
4802 | SelectLoad(Node, 4, AArch64::LD4Rv4h, AArch64::dsub0); |
4803 | return; |
4804 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
4805 | SelectLoad(Node, 4, AArch64::LD4Rv8h, AArch64::qsub0); |
4806 | return; |
4807 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
4808 | SelectLoad(Node, 4, AArch64::LD4Rv2s, AArch64::dsub0); |
4809 | return; |
4810 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
4811 | SelectLoad(Node, 4, AArch64::LD4Rv4s, AArch64::qsub0); |
4812 | return; |
4813 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
4814 | SelectLoad(Node, 4, AArch64::LD4Rv1d, AArch64::dsub0); |
4815 | return; |
4816 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
4817 | SelectLoad(Node, 4, AArch64::LD4Rv2d, AArch64::qsub0); |
4818 | return; |
4819 | } |
4820 | break; |
4821 | case Intrinsic::aarch64_neon_ld2lane: |
4822 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
4823 | SelectLoadLane(Node, 2, AArch64::LD2i8); |
4824 | return; |
4825 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
4826 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
4827 | SelectLoadLane(Node, 2, AArch64::LD2i16); |
4828 | return; |
4829 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
4830 | VT == MVT::v2f32) { |
4831 | SelectLoadLane(Node, 2, AArch64::LD2i32); |
4832 | return; |
4833 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
4834 | VT == MVT::v1f64) { |
4835 | SelectLoadLane(Node, 2, AArch64::LD2i64); |
4836 | return; |
4837 | } |
4838 | break; |
4839 | case Intrinsic::aarch64_neon_ld3lane: |
4840 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
4841 | SelectLoadLane(Node, 3, AArch64::LD3i8); |
4842 | return; |
4843 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
4844 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
4845 | SelectLoadLane(Node, 3, AArch64::LD3i16); |
4846 | return; |
4847 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
4848 | VT == MVT::v2f32) { |
4849 | SelectLoadLane(Node, 3, AArch64::LD3i32); |
4850 | return; |
4851 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
4852 | VT == MVT::v1f64) { |
4853 | SelectLoadLane(Node, 3, AArch64::LD3i64); |
4854 | return; |
4855 | } |
4856 | break; |
4857 | case Intrinsic::aarch64_neon_ld4lane: |
4858 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
4859 | SelectLoadLane(Node, 4, AArch64::LD4i8); |
4860 | return; |
4861 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
4862 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
4863 | SelectLoadLane(Node, 4, AArch64::LD4i16); |
4864 | return; |
4865 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
4866 | VT == MVT::v2f32) { |
4867 | SelectLoadLane(Node, 4, AArch64::LD4i32); |
4868 | return; |
4869 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
4870 | VT == MVT::v1f64) { |
4871 | SelectLoadLane(Node, 4, AArch64::LD4i64); |
4872 | return; |
4873 | } |
4874 | break; |
4875 | case Intrinsic::aarch64_ld64b: |
4876 | SelectLoad(Node, 8, AArch64::LD64B, AArch64::x8sub_0); |
4877 | return; |
4878 | case Intrinsic::aarch64_sve_ld2q_sret: { |
4879 | SelectPredicatedLoad(Node, 2, 4, AArch64::LD2Q_IMM, AArch64::LD2Q, true); |
4880 | return; |
4881 | } |
4882 | case Intrinsic::aarch64_sve_ld3q_sret: { |
4883 | SelectPredicatedLoad(Node, 3, 4, AArch64::LD3Q_IMM, AArch64::LD3Q, true); |
4884 | return; |
4885 | } |
4886 | case Intrinsic::aarch64_sve_ld4q_sret: { |
4887 | SelectPredicatedLoad(Node, 4, 4, AArch64::LD4Q_IMM, AArch64::LD4Q, true); |
4888 | return; |
4889 | } |
4890 | case Intrinsic::aarch64_sve_ld2_sret: { |
4891 | if (VT == MVT::nxv16i8) { |
4892 | SelectPredicatedLoad(Node, 2, 0, AArch64::LD2B_IMM, AArch64::LD2B, |
4893 | true); |
4894 | return; |
4895 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
4896 | VT == MVT::nxv8bf16) { |
4897 | SelectPredicatedLoad(Node, 2, 1, AArch64::LD2H_IMM, AArch64::LD2H, |
4898 | true); |
4899 | return; |
4900 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
4901 | SelectPredicatedLoad(Node, 2, 2, AArch64::LD2W_IMM, AArch64::LD2W, |
4902 | true); |
4903 | return; |
4904 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
4905 | SelectPredicatedLoad(Node, 2, 3, AArch64::LD2D_IMM, AArch64::LD2D, |
4906 | true); |
4907 | return; |
4908 | } |
4909 | break; |
4910 | } |
4911 | case Intrinsic::aarch64_sve_ld1_pn_x2: { |
4912 | if (VT == MVT::nxv16i8) { |
4913 | if (Subtarget->hasSME2()) |
4914 | SelectContiguousMultiVectorLoad( |
4915 | Node, 2, 0, AArch64::LD1B_2Z_IMM_PSEUDO, AArch64::LD1B_2Z_PSEUDO); |
4916 | else if (Subtarget->hasSVE2p1()) |
4917 | SelectContiguousMultiVectorLoad(Node, 2, 0, AArch64::LD1B_2Z_IMM, |
4918 | AArch64::LD1B_2Z); |
4919 | else |
4920 | break; |
4921 | return; |
4922 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
4923 | VT == MVT::nxv8bf16) { |
4924 | if (Subtarget->hasSME2()) |
4925 | SelectContiguousMultiVectorLoad( |
4926 | Node, 2, 1, AArch64::LD1H_2Z_IMM_PSEUDO, AArch64::LD1H_2Z_PSEUDO); |
4927 | else if (Subtarget->hasSVE2p1()) |
4928 | SelectContiguousMultiVectorLoad(Node, 2, 1, AArch64::LD1H_2Z_IMM, |
4929 | AArch64::LD1H_2Z); |
4930 | else |
4931 | break; |
4932 | return; |
4933 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
4934 | if (Subtarget->hasSME2()) |
4935 | SelectContiguousMultiVectorLoad( |
4936 | Node, 2, 2, AArch64::LD1W_2Z_IMM_PSEUDO, AArch64::LD1W_2Z_PSEUDO); |
4937 | else if (Subtarget->hasSVE2p1()) |
4938 | SelectContiguousMultiVectorLoad(Node, 2, 2, AArch64::LD1W_2Z_IMM, |
4939 | AArch64::LD1W_2Z); |
4940 | else |
4941 | break; |
4942 | return; |
4943 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
4944 | if (Subtarget->hasSME2()) |
4945 | SelectContiguousMultiVectorLoad( |
4946 | Node, 2, 3, AArch64::LD1D_2Z_IMM_PSEUDO, AArch64::LD1D_2Z_PSEUDO); |
4947 | else if (Subtarget->hasSVE2p1()) |
4948 | SelectContiguousMultiVectorLoad(Node, 2, 3, AArch64::LD1D_2Z_IMM, |
4949 | AArch64::LD1D_2Z); |
4950 | else |
4951 | break; |
4952 | return; |
4953 | } |
4954 | break; |
4955 | } |
4956 | case Intrinsic::aarch64_sve_ld1_pn_x4: { |
4957 | if (VT == MVT::nxv16i8) { |
4958 | if (Subtarget->hasSME2()) |
4959 | SelectContiguousMultiVectorLoad( |
4960 | Node, 4, 0, AArch64::LD1B_4Z_IMM_PSEUDO, AArch64::LD1B_4Z_PSEUDO); |
4961 | else if (Subtarget->hasSVE2p1()) |
4962 | SelectContiguousMultiVectorLoad(Node, 4, 0, AArch64::LD1B_4Z_IMM, |
4963 | AArch64::LD1B_4Z); |
4964 | else |
4965 | break; |
4966 | return; |
4967 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
4968 | VT == MVT::nxv8bf16) { |
4969 | if (Subtarget->hasSME2()) |
4970 | SelectContiguousMultiVectorLoad( |
4971 | Node, 4, 1, AArch64::LD1H_4Z_IMM_PSEUDO, AArch64::LD1H_4Z_PSEUDO); |
4972 | else if (Subtarget->hasSVE2p1()) |
4973 | SelectContiguousMultiVectorLoad(Node, 4, 1, AArch64::LD1H_4Z_IMM, |
4974 | AArch64::LD1H_4Z); |
4975 | else |
4976 | break; |
4977 | return; |
4978 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
4979 | if (Subtarget->hasSME2()) |
4980 | SelectContiguousMultiVectorLoad( |
4981 | Node, 4, 2, AArch64::LD1W_4Z_IMM_PSEUDO, AArch64::LD1W_4Z_PSEUDO); |
4982 | else if (Subtarget->hasSVE2p1()) |
4983 | SelectContiguousMultiVectorLoad(Node, 4, 2, AArch64::LD1W_4Z_IMM, |
4984 | AArch64::LD1W_4Z); |
4985 | else |
4986 | break; |
4987 | return; |
4988 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
4989 | if (Subtarget->hasSME2()) |
4990 | SelectContiguousMultiVectorLoad( |
4991 | Node, 4, 3, AArch64::LD1D_4Z_IMM_PSEUDO, AArch64::LD1D_4Z_PSEUDO); |
4992 | else if (Subtarget->hasSVE2p1()) |
4993 | SelectContiguousMultiVectorLoad(Node, 4, 3, AArch64::LD1D_4Z_IMM, |
4994 | AArch64::LD1D_4Z); |
4995 | else |
4996 | break; |
4997 | return; |
4998 | } |
4999 | break; |
5000 | } |
5001 | case Intrinsic::aarch64_sve_ldnt1_pn_x2: { |
5002 | if (VT == MVT::nxv16i8) { |
5003 | if (Subtarget->hasSME2()) |
5004 | SelectContiguousMultiVectorLoad(Node, 2, 0, |
5005 | AArch64::LDNT1B_2Z_IMM_PSEUDO, |
5006 | AArch64::LDNT1B_2Z_PSEUDO); |
5007 | else if (Subtarget->hasSVE2p1()) |
5008 | SelectContiguousMultiVectorLoad(Node, 2, 0, AArch64::LDNT1B_2Z_IMM, |
5009 | AArch64::LDNT1B_2Z); |
5010 | else |
5011 | break; |
5012 | return; |
5013 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5014 | VT == MVT::nxv8bf16) { |
5015 | if (Subtarget->hasSME2()) |
5016 | SelectContiguousMultiVectorLoad(Node, 2, 1, |
5017 | AArch64::LDNT1H_2Z_IMM_PSEUDO, |
5018 | AArch64::LDNT1H_2Z_PSEUDO); |
5019 | else if (Subtarget->hasSVE2p1()) |
5020 | SelectContiguousMultiVectorLoad(Node, 2, 1, AArch64::LDNT1H_2Z_IMM, |
5021 | AArch64::LDNT1H_2Z); |
5022 | else |
5023 | break; |
5024 | return; |
5025 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5026 | if (Subtarget->hasSME2()) |
5027 | SelectContiguousMultiVectorLoad(Node, 2, 2, |
5028 | AArch64::LDNT1W_2Z_IMM_PSEUDO, |
5029 | AArch64::LDNT1W_2Z_PSEUDO); |
5030 | else if (Subtarget->hasSVE2p1()) |
5031 | SelectContiguousMultiVectorLoad(Node, 2, 2, AArch64::LDNT1W_2Z_IMM, |
5032 | AArch64::LDNT1W_2Z); |
5033 | else |
5034 | break; |
5035 | return; |
5036 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5037 | if (Subtarget->hasSME2()) |
5038 | SelectContiguousMultiVectorLoad(Node, 2, 3, |
5039 | AArch64::LDNT1D_2Z_IMM_PSEUDO, |
5040 | AArch64::LDNT1D_2Z_PSEUDO); |
5041 | else if (Subtarget->hasSVE2p1()) |
5042 | SelectContiguousMultiVectorLoad(Node, 2, 3, AArch64::LDNT1D_2Z_IMM, |
5043 | AArch64::LDNT1D_2Z); |
5044 | else |
5045 | break; |
5046 | return; |
5047 | } |
5048 | break; |
5049 | } |
5050 | case Intrinsic::aarch64_sve_ldnt1_pn_x4: { |
5051 | if (VT == MVT::nxv16i8) { |
5052 | if (Subtarget->hasSME2()) |
5053 | SelectContiguousMultiVectorLoad(Node, 4, 0, |
5054 | AArch64::LDNT1B_4Z_IMM_PSEUDO, |
5055 | AArch64::LDNT1B_4Z_PSEUDO); |
5056 | else if (Subtarget->hasSVE2p1()) |
5057 | SelectContiguousMultiVectorLoad(Node, 4, 0, AArch64::LDNT1B_4Z_IMM, |
5058 | AArch64::LDNT1B_4Z); |
5059 | else |
5060 | break; |
5061 | return; |
5062 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5063 | VT == MVT::nxv8bf16) { |
5064 | if (Subtarget->hasSME2()) |
5065 | SelectContiguousMultiVectorLoad(Node, 4, 1, |
5066 | AArch64::LDNT1H_4Z_IMM_PSEUDO, |
5067 | AArch64::LDNT1H_4Z_PSEUDO); |
5068 | else if (Subtarget->hasSVE2p1()) |
5069 | SelectContiguousMultiVectorLoad(Node, 4, 1, AArch64::LDNT1H_4Z_IMM, |
5070 | AArch64::LDNT1H_4Z); |
5071 | else |
5072 | break; |
5073 | return; |
5074 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5075 | if (Subtarget->hasSME2()) |
5076 | SelectContiguousMultiVectorLoad(Node, 4, 2, |
5077 | AArch64::LDNT1W_4Z_IMM_PSEUDO, |
5078 | AArch64::LDNT1W_4Z_PSEUDO); |
5079 | else if (Subtarget->hasSVE2p1()) |
5080 | SelectContiguousMultiVectorLoad(Node, 4, 2, AArch64::LDNT1W_4Z_IMM, |
5081 | AArch64::LDNT1W_4Z); |
5082 | else |
5083 | break; |
5084 | return; |
5085 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5086 | if (Subtarget->hasSME2()) |
5087 | SelectContiguousMultiVectorLoad(Node, 4, 3, |
5088 | AArch64::LDNT1D_4Z_IMM_PSEUDO, |
5089 | AArch64::LDNT1D_4Z_PSEUDO); |
5090 | else if (Subtarget->hasSVE2p1()) |
5091 | SelectContiguousMultiVectorLoad(Node, 4, 3, AArch64::LDNT1D_4Z_IMM, |
5092 | AArch64::LDNT1D_4Z); |
5093 | else |
5094 | break; |
5095 | return; |
5096 | } |
5097 | break; |
5098 | } |
5099 | case Intrinsic::aarch64_sve_ld3_sret: { |
5100 | if (VT == MVT::nxv16i8) { |
5101 | SelectPredicatedLoad(Node, 3, 0, AArch64::LD3B_IMM, AArch64::LD3B, |
5102 | true); |
5103 | return; |
5104 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5105 | VT == MVT::nxv8bf16) { |
5106 | SelectPredicatedLoad(Node, 3, 1, AArch64::LD3H_IMM, AArch64::LD3H, |
5107 | true); |
5108 | return; |
5109 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5110 | SelectPredicatedLoad(Node, 3, 2, AArch64::LD3W_IMM, AArch64::LD3W, |
5111 | true); |
5112 | return; |
5113 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5114 | SelectPredicatedLoad(Node, 3, 3, AArch64::LD3D_IMM, AArch64::LD3D, |
5115 | true); |
5116 | return; |
5117 | } |
5118 | break; |
5119 | } |
5120 | case Intrinsic::aarch64_sve_ld4_sret: { |
5121 | if (VT == MVT::nxv16i8) { |
5122 | SelectPredicatedLoad(Node, 4, 0, AArch64::LD4B_IMM, AArch64::LD4B, |
5123 | true); |
5124 | return; |
5125 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5126 | VT == MVT::nxv8bf16) { |
5127 | SelectPredicatedLoad(Node, 4, 1, AArch64::LD4H_IMM, AArch64::LD4H, |
5128 | true); |
5129 | return; |
5130 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5131 | SelectPredicatedLoad(Node, 4, 2, AArch64::LD4W_IMM, AArch64::LD4W, |
5132 | true); |
5133 | return; |
5134 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5135 | SelectPredicatedLoad(Node, 4, 3, AArch64::LD4D_IMM, AArch64::LD4D, |
5136 | true); |
5137 | return; |
5138 | } |
5139 | break; |
5140 | } |
5141 | case Intrinsic::aarch64_sme_read_hor_vg2: { |
5142 | if (VT == MVT::nxv16i8) { |
5143 | SelectMultiVectorMove<14, 2>(Node, 2, AArch64::ZAB0, |
5144 | AArch64::MOVA_2ZMXI_H_B); |
5145 | return; |
5146 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5147 | VT == MVT::nxv8bf16) { |
5148 | SelectMultiVectorMove<6, 2>(Node, 2, AArch64::ZAH0, |
5149 | AArch64::MOVA_2ZMXI_H_H); |
5150 | return; |
5151 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5152 | SelectMultiVectorMove<2, 2>(Node, 2, AArch64::ZAS0, |
5153 | AArch64::MOVA_2ZMXI_H_S); |
5154 | return; |
5155 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5156 | SelectMultiVectorMove<0, 2>(Node, 2, AArch64::ZAD0, |
5157 | AArch64::MOVA_2ZMXI_H_D); |
5158 | return; |
5159 | } |
5160 | break; |
5161 | } |
5162 | case Intrinsic::aarch64_sme_read_ver_vg2: { |
5163 | if (VT == MVT::nxv16i8) { |
5164 | SelectMultiVectorMove<14, 2>(Node, 2, AArch64::ZAB0, |
5165 | AArch64::MOVA_2ZMXI_V_B); |
5166 | return; |
5167 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5168 | VT == MVT::nxv8bf16) { |
5169 | SelectMultiVectorMove<6, 2>(Node, 2, AArch64::ZAH0, |
5170 | AArch64::MOVA_2ZMXI_V_H); |
5171 | return; |
5172 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5173 | SelectMultiVectorMove<2, 2>(Node, 2, AArch64::ZAS0, |
5174 | AArch64::MOVA_2ZMXI_V_S); |
5175 | return; |
5176 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5177 | SelectMultiVectorMove<0, 2>(Node, 2, AArch64::ZAD0, |
5178 | AArch64::MOVA_2ZMXI_V_D); |
5179 | return; |
5180 | } |
5181 | break; |
5182 | } |
5183 | case Intrinsic::aarch64_sme_read_hor_vg4: { |
5184 | if (VT == MVT::nxv16i8) { |
5185 | SelectMultiVectorMove<12, 4>(Node, 4, AArch64::ZAB0, |
5186 | AArch64::MOVA_4ZMXI_H_B); |
5187 | return; |
5188 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5189 | VT == MVT::nxv8bf16) { |
5190 | SelectMultiVectorMove<4, 4>(Node, 4, AArch64::ZAH0, |
5191 | AArch64::MOVA_4ZMXI_H_H); |
5192 | return; |
5193 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5194 | SelectMultiVectorMove<0, 2>(Node, 4, AArch64::ZAS0, |
5195 | AArch64::MOVA_4ZMXI_H_S); |
5196 | return; |
5197 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5198 | SelectMultiVectorMove<0, 2>(Node, 4, AArch64::ZAD0, |
5199 | AArch64::MOVA_4ZMXI_H_D); |
5200 | return; |
5201 | } |
5202 | break; |
5203 | } |
5204 | case Intrinsic::aarch64_sme_read_ver_vg4: { |
5205 | if (VT == MVT::nxv16i8) { |
5206 | SelectMultiVectorMove<12, 4>(Node, 4, AArch64::ZAB0, |
5207 | AArch64::MOVA_4ZMXI_V_B); |
5208 | return; |
5209 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5210 | VT == MVT::nxv8bf16) { |
5211 | SelectMultiVectorMove<4, 4>(Node, 4, AArch64::ZAH0, |
5212 | AArch64::MOVA_4ZMXI_V_H); |
5213 | return; |
5214 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5215 | SelectMultiVectorMove<0, 4>(Node, 4, AArch64::ZAS0, |
5216 | AArch64::MOVA_4ZMXI_V_S); |
5217 | return; |
5218 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5219 | SelectMultiVectorMove<0, 4>(Node, 4, AArch64::ZAD0, |
5220 | AArch64::MOVA_4ZMXI_V_D); |
5221 | return; |
5222 | } |
5223 | break; |
5224 | } |
5225 | case Intrinsic::aarch64_sme_read_vg1x2: { |
5226 | SelectMultiVectorMove<7, 1>(Node, 2, AArch64::ZA, |
5227 | AArch64::MOVA_VG2_2ZMXI); |
5228 | return; |
5229 | } |
5230 | case Intrinsic::aarch64_sme_read_vg1x4: { |
5231 | SelectMultiVectorMove<7, 1>(Node, 4, AArch64::ZA, |
5232 | AArch64::MOVA_VG4_4ZMXI); |
5233 | return; |
5234 | } |
5235 | case Intrinsic::swift_async_context_addr: { |
5236 | SDLoc DL(Node); |
5237 | SDValue Chain = Node->getOperand(Num: 0); |
5238 | SDValue CopyFP = CurDAG->getCopyFromReg(Chain, DL, AArch64::FP, MVT::i64); |
5239 | SDValue Res = SDValue( |
5240 | CurDAG->getMachineNode(AArch64::SUBXri, DL, MVT::i64, CopyFP, |
5241 | CurDAG->getTargetConstant(8, DL, MVT::i32), |
5242 | CurDAG->getTargetConstant(0, DL, MVT::i32)), |
5243 | 0); |
5244 | ReplaceUses(F: SDValue(Node, 0), T: Res); |
5245 | ReplaceUses(F: SDValue(Node, 1), T: CopyFP.getValue(R: 1)); |
5246 | CurDAG->RemoveDeadNode(N: Node); |
5247 | |
5248 | auto &MF = CurDAG->getMachineFunction(); |
5249 | MF.getFrameInfo().setFrameAddressIsTaken(true); |
5250 | MF.getInfo<AArch64FunctionInfo>()->setHasSwiftAsyncContext(true); |
5251 | return; |
5252 | } |
5253 | case Intrinsic::aarch64_sme_luti2_lane_zt_x4: { |
5254 | if (auto Opc = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
5255 | Node->getValueType(0), |
5256 | {AArch64::LUTI2_4ZTZI_B, AArch64::LUTI2_4ZTZI_H, |
5257 | AArch64::LUTI2_4ZTZI_S})) |
5258 | // Second Immediate must be <= 3: |
5259 | SelectMultiVectorLuti(Node, NumOutVecs: 4, Opc: Opc, MaxImm: 3); |
5260 | return; |
5261 | } |
5262 | case Intrinsic::aarch64_sme_luti4_lane_zt_x4: { |
5263 | if (auto Opc = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
5264 | Node->getValueType(0), |
5265 | {0, AArch64::LUTI4_4ZTZI_H, AArch64::LUTI4_4ZTZI_S})) |
5266 | // Second Immediate must be <= 1: |
5267 | SelectMultiVectorLuti(Node, NumOutVecs: 4, Opc: Opc, MaxImm: 1); |
5268 | return; |
5269 | } |
5270 | case Intrinsic::aarch64_sme_luti2_lane_zt_x2: { |
5271 | if (auto Opc = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
5272 | Node->getValueType(0), |
5273 | {AArch64::LUTI2_2ZTZI_B, AArch64::LUTI2_2ZTZI_H, |
5274 | AArch64::LUTI2_2ZTZI_S})) |
5275 | // Second Immediate must be <= 7: |
5276 | SelectMultiVectorLuti(Node, NumOutVecs: 2, Opc: Opc, MaxImm: 7); |
5277 | return; |
5278 | } |
5279 | case Intrinsic::aarch64_sme_luti4_lane_zt_x2: { |
5280 | if (auto Opc = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
5281 | Node->getValueType(0), |
5282 | {AArch64::LUTI4_2ZTZI_B, AArch64::LUTI4_2ZTZI_H, |
5283 | AArch64::LUTI4_2ZTZI_S})) |
5284 | // Second Immediate must be <= 3: |
5285 | SelectMultiVectorLuti(Node, NumOutVecs: 2, Opc: Opc, MaxImm: 3); |
5286 | return; |
5287 | } |
5288 | } |
5289 | } break; |
5290 | case ISD::INTRINSIC_WO_CHAIN: { |
5291 | unsigned IntNo = Node->getConstantOperandVal(Num: 0); |
5292 | switch (IntNo) { |
5293 | default: |
5294 | break; |
5295 | case Intrinsic::aarch64_tagp: |
5296 | SelectTagP(N: Node); |
5297 | return; |
5298 | case Intrinsic::aarch64_neon_tbl2: |
5299 | SelectTable(Node, 2, |
5300 | VT == MVT::v8i8 ? AArch64::TBLv8i8Two : AArch64::TBLv16i8Two, |
5301 | false); |
5302 | return; |
5303 | case Intrinsic::aarch64_neon_tbl3: |
5304 | SelectTable(Node, 3, VT == MVT::v8i8 ? AArch64::TBLv8i8Three |
5305 | : AArch64::TBLv16i8Three, |
5306 | false); |
5307 | return; |
5308 | case Intrinsic::aarch64_neon_tbl4: |
5309 | SelectTable(Node, 4, VT == MVT::v8i8 ? AArch64::TBLv8i8Four |
5310 | : AArch64::TBLv16i8Four, |
5311 | false); |
5312 | return; |
5313 | case Intrinsic::aarch64_neon_tbx2: |
5314 | SelectTable(Node, 2, |
5315 | VT == MVT::v8i8 ? AArch64::TBXv8i8Two : AArch64::TBXv16i8Two, |
5316 | true); |
5317 | return; |
5318 | case Intrinsic::aarch64_neon_tbx3: |
5319 | SelectTable(Node, 3, VT == MVT::v8i8 ? AArch64::TBXv8i8Three |
5320 | : AArch64::TBXv16i8Three, |
5321 | true); |
5322 | return; |
5323 | case Intrinsic::aarch64_neon_tbx4: |
5324 | SelectTable(Node, 4, VT == MVT::v8i8 ? AArch64::TBXv8i8Four |
5325 | : AArch64::TBXv16i8Four, |
5326 | true); |
5327 | return; |
5328 | case Intrinsic::aarch64_sve_srshl_single_x2: |
5329 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5330 | Node->getValueType(0), |
5331 | {AArch64::SRSHL_VG2_2ZZ_B, AArch64::SRSHL_VG2_2ZZ_H, |
5332 | AArch64::SRSHL_VG2_2ZZ_S, AArch64::SRSHL_VG2_2ZZ_D})) |
5333 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
5334 | return; |
5335 | case Intrinsic::aarch64_sve_srshl_single_x4: |
5336 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5337 | Node->getValueType(0), |
5338 | {AArch64::SRSHL_VG4_4ZZ_B, AArch64::SRSHL_VG4_4ZZ_H, |
5339 | AArch64::SRSHL_VG4_4ZZ_S, AArch64::SRSHL_VG4_4ZZ_D})) |
5340 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
5341 | return; |
5342 | case Intrinsic::aarch64_sve_urshl_single_x2: |
5343 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5344 | Node->getValueType(0), |
5345 | {AArch64::URSHL_VG2_2ZZ_B, AArch64::URSHL_VG2_2ZZ_H, |
5346 | AArch64::URSHL_VG2_2ZZ_S, AArch64::URSHL_VG2_2ZZ_D})) |
5347 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
5348 | return; |
5349 | case Intrinsic::aarch64_sve_urshl_single_x4: |
5350 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5351 | Node->getValueType(0), |
5352 | {AArch64::URSHL_VG4_4ZZ_B, AArch64::URSHL_VG4_4ZZ_H, |
5353 | AArch64::URSHL_VG4_4ZZ_S, AArch64::URSHL_VG4_4ZZ_D})) |
5354 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
5355 | return; |
5356 | case Intrinsic::aarch64_sve_srshl_x2: |
5357 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5358 | Node->getValueType(0), |
5359 | {AArch64::SRSHL_VG2_2Z2Z_B, AArch64::SRSHL_VG2_2Z2Z_H, |
5360 | AArch64::SRSHL_VG2_2Z2Z_S, AArch64::SRSHL_VG2_2Z2Z_D})) |
5361 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
5362 | return; |
5363 | case Intrinsic::aarch64_sve_srshl_x4: |
5364 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5365 | Node->getValueType(0), |
5366 | {AArch64::SRSHL_VG4_4Z4Z_B, AArch64::SRSHL_VG4_4Z4Z_H, |
5367 | AArch64::SRSHL_VG4_4Z4Z_S, AArch64::SRSHL_VG4_4Z4Z_D})) |
5368 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
5369 | return; |
5370 | case Intrinsic::aarch64_sve_urshl_x2: |
5371 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5372 | Node->getValueType(0), |
5373 | {AArch64::URSHL_VG2_2Z2Z_B, AArch64::URSHL_VG2_2Z2Z_H, |
5374 | AArch64::URSHL_VG2_2Z2Z_S, AArch64::URSHL_VG2_2Z2Z_D})) |
5375 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
5376 | return; |
5377 | case Intrinsic::aarch64_sve_urshl_x4: |
5378 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5379 | Node->getValueType(0), |
5380 | {AArch64::URSHL_VG4_4Z4Z_B, AArch64::URSHL_VG4_4Z4Z_H, |
5381 | AArch64::URSHL_VG4_4Z4Z_S, AArch64::URSHL_VG4_4Z4Z_D})) |
5382 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
5383 | return; |
5384 | case Intrinsic::aarch64_sve_sqdmulh_single_vgx2: |
5385 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5386 | Node->getValueType(0), |
5387 | {AArch64::SQDMULH_VG2_2ZZ_B, AArch64::SQDMULH_VG2_2ZZ_H, |
5388 | AArch64::SQDMULH_VG2_2ZZ_S, AArch64::SQDMULH_VG2_2ZZ_D})) |
5389 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
5390 | return; |
5391 | case Intrinsic::aarch64_sve_sqdmulh_single_vgx4: |
5392 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5393 | Node->getValueType(0), |
5394 | {AArch64::SQDMULH_VG4_4ZZ_B, AArch64::SQDMULH_VG4_4ZZ_H, |
5395 | AArch64::SQDMULH_VG4_4ZZ_S, AArch64::SQDMULH_VG4_4ZZ_D})) |
5396 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
5397 | return; |
5398 | case Intrinsic::aarch64_sve_sqdmulh_vgx2: |
5399 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5400 | Node->getValueType(0), |
5401 | {AArch64::SQDMULH_VG2_2Z2Z_B, AArch64::SQDMULH_VG2_2Z2Z_H, |
5402 | AArch64::SQDMULH_VG2_2Z2Z_S, AArch64::SQDMULH_VG2_2Z2Z_D})) |
5403 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
5404 | return; |
5405 | case Intrinsic::aarch64_sve_sqdmulh_vgx4: |
5406 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5407 | Node->getValueType(0), |
5408 | {AArch64::SQDMULH_VG4_4Z4Z_B, AArch64::SQDMULH_VG4_4Z4Z_H, |
5409 | AArch64::SQDMULH_VG4_4Z4Z_S, AArch64::SQDMULH_VG4_4Z4Z_D})) |
5410 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
5411 | return; |
5412 | case Intrinsic::aarch64_sve_whilege_x2: |
5413 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int1>( |
5414 | Node->getValueType(0), |
5415 | {AArch64::WHILEGE_2PXX_B, AArch64::WHILEGE_2PXX_H, |
5416 | AArch64::WHILEGE_2PXX_S, AArch64::WHILEGE_2PXX_D})) |
5417 | SelectWhilePair(N: Node, Opc: Op); |
5418 | return; |
5419 | case Intrinsic::aarch64_sve_whilegt_x2: |
5420 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int1>( |
5421 | Node->getValueType(0), |
5422 | {AArch64::WHILEGT_2PXX_B, AArch64::WHILEGT_2PXX_H, |
5423 | AArch64::WHILEGT_2PXX_S, AArch64::WHILEGT_2PXX_D})) |
5424 | SelectWhilePair(N: Node, Opc: Op); |
5425 | return; |
5426 | case Intrinsic::aarch64_sve_whilehi_x2: |
5427 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int1>( |
5428 | Node->getValueType(0), |
5429 | {AArch64::WHILEHI_2PXX_B, AArch64::WHILEHI_2PXX_H, |
5430 | AArch64::WHILEHI_2PXX_S, AArch64::WHILEHI_2PXX_D})) |
5431 | SelectWhilePair(N: Node, Opc: Op); |
5432 | return; |
5433 | case Intrinsic::aarch64_sve_whilehs_x2: |
5434 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int1>( |
5435 | Node->getValueType(0), |
5436 | {AArch64::WHILEHS_2PXX_B, AArch64::WHILEHS_2PXX_H, |
5437 | AArch64::WHILEHS_2PXX_S, AArch64::WHILEHS_2PXX_D})) |
5438 | SelectWhilePair(N: Node, Opc: Op); |
5439 | return; |
5440 | case Intrinsic::aarch64_sve_whilele_x2: |
5441 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int1>( |
5442 | Node->getValueType(0), |
5443 | {AArch64::WHILELE_2PXX_B, AArch64::WHILELE_2PXX_H, |
5444 | AArch64::WHILELE_2PXX_S, AArch64::WHILELE_2PXX_D})) |
5445 | SelectWhilePair(N: Node, Opc: Op); |
5446 | return; |
5447 | case Intrinsic::aarch64_sve_whilelo_x2: |
5448 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int1>( |
5449 | Node->getValueType(0), |
5450 | {AArch64::WHILELO_2PXX_B, AArch64::WHILELO_2PXX_H, |
5451 | AArch64::WHILELO_2PXX_S, AArch64::WHILELO_2PXX_D})) |
5452 | SelectWhilePair(N: Node, Opc: Op); |
5453 | return; |
5454 | case Intrinsic::aarch64_sve_whilels_x2: |
5455 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int1>( |
5456 | Node->getValueType(0), |
5457 | {AArch64::WHILELS_2PXX_B, AArch64::WHILELS_2PXX_H, |
5458 | AArch64::WHILELS_2PXX_S, AArch64::WHILELS_2PXX_D})) |
5459 | SelectWhilePair(N: Node, Opc: Op); |
5460 | return; |
5461 | case Intrinsic::aarch64_sve_whilelt_x2: |
5462 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int1>( |
5463 | Node->getValueType(0), |
5464 | {AArch64::WHILELT_2PXX_B, AArch64::WHILELT_2PXX_H, |
5465 | AArch64::WHILELT_2PXX_S, AArch64::WHILELT_2PXX_D})) |
5466 | SelectWhilePair(N: Node, Opc: Op); |
5467 | return; |
5468 | case Intrinsic::aarch64_sve_smax_single_x2: |
5469 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5470 | Node->getValueType(0), |
5471 | {AArch64::SMAX_VG2_2ZZ_B, AArch64::SMAX_VG2_2ZZ_H, |
5472 | AArch64::SMAX_VG2_2ZZ_S, AArch64::SMAX_VG2_2ZZ_D})) |
5473 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
5474 | return; |
5475 | case Intrinsic::aarch64_sve_umax_single_x2: |
5476 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5477 | Node->getValueType(0), |
5478 | {AArch64::UMAX_VG2_2ZZ_B, AArch64::UMAX_VG2_2ZZ_H, |
5479 | AArch64::UMAX_VG2_2ZZ_S, AArch64::UMAX_VG2_2ZZ_D})) |
5480 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
5481 | return; |
5482 | case Intrinsic::aarch64_sve_fmax_single_x2: |
5483 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5484 | Node->getValueType(0), |
5485 | {0, AArch64::FMAX_VG2_2ZZ_H, AArch64::FMAX_VG2_2ZZ_S, |
5486 | AArch64::FMAX_VG2_2ZZ_D})) |
5487 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
5488 | return; |
5489 | case Intrinsic::aarch64_sve_smax_single_x4: |
5490 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5491 | Node->getValueType(0), |
5492 | {AArch64::SMAX_VG4_4ZZ_B, AArch64::SMAX_VG4_4ZZ_H, |
5493 | AArch64::SMAX_VG4_4ZZ_S, AArch64::SMAX_VG4_4ZZ_D})) |
5494 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
5495 | return; |
5496 | case Intrinsic::aarch64_sve_umax_single_x4: |
5497 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5498 | Node->getValueType(0), |
5499 | {AArch64::UMAX_VG4_4ZZ_B, AArch64::UMAX_VG4_4ZZ_H, |
5500 | AArch64::UMAX_VG4_4ZZ_S, AArch64::UMAX_VG4_4ZZ_D})) |
5501 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
5502 | return; |
5503 | case Intrinsic::aarch64_sve_fmax_single_x4: |
5504 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5505 | Node->getValueType(0), |
5506 | {0, AArch64::FMAX_VG4_4ZZ_H, AArch64::FMAX_VG4_4ZZ_S, |
5507 | AArch64::FMAX_VG4_4ZZ_D})) |
5508 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
5509 | return; |
5510 | case Intrinsic::aarch64_sve_smin_single_x2: |
5511 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5512 | Node->getValueType(0), |
5513 | {AArch64::SMIN_VG2_2ZZ_B, AArch64::SMIN_VG2_2ZZ_H, |
5514 | AArch64::SMIN_VG2_2ZZ_S, AArch64::SMIN_VG2_2ZZ_D})) |
5515 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
5516 | return; |
5517 | case Intrinsic::aarch64_sve_umin_single_x2: |
5518 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5519 | Node->getValueType(0), |
5520 | {AArch64::UMIN_VG2_2ZZ_B, AArch64::UMIN_VG2_2ZZ_H, |
5521 | AArch64::UMIN_VG2_2ZZ_S, AArch64::UMIN_VG2_2ZZ_D})) |
5522 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
5523 | return; |
5524 | case Intrinsic::aarch64_sve_fmin_single_x2: |
5525 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5526 | Node->getValueType(0), |
5527 | {0, AArch64::FMIN_VG2_2ZZ_H, AArch64::FMIN_VG2_2ZZ_S, |
5528 | AArch64::FMIN_VG2_2ZZ_D})) |
5529 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
5530 | return; |
5531 | case Intrinsic::aarch64_sve_smin_single_x4: |
5532 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5533 | Node->getValueType(0), |
5534 | {AArch64::SMIN_VG4_4ZZ_B, AArch64::SMIN_VG4_4ZZ_H, |
5535 | AArch64::SMIN_VG4_4ZZ_S, AArch64::SMIN_VG4_4ZZ_D})) |
5536 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
5537 | return; |
5538 | case Intrinsic::aarch64_sve_umin_single_x4: |
5539 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5540 | Node->getValueType(0), |
5541 | {AArch64::UMIN_VG4_4ZZ_B, AArch64::UMIN_VG4_4ZZ_H, |
5542 | AArch64::UMIN_VG4_4ZZ_S, AArch64::UMIN_VG4_4ZZ_D})) |
5543 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
5544 | return; |
5545 | case Intrinsic::aarch64_sve_fmin_single_x4: |
5546 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5547 | Node->getValueType(0), |
5548 | {0, AArch64::FMIN_VG4_4ZZ_H, AArch64::FMIN_VG4_4ZZ_S, |
5549 | AArch64::FMIN_VG4_4ZZ_D})) |
5550 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
5551 | return; |
5552 | case Intrinsic::aarch64_sve_smax_x2: |
5553 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5554 | Node->getValueType(0), |
5555 | {AArch64::SMAX_VG2_2Z2Z_B, AArch64::SMAX_VG2_2Z2Z_H, |
5556 | AArch64::SMAX_VG2_2Z2Z_S, AArch64::SMAX_VG2_2Z2Z_D})) |
5557 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
5558 | return; |
5559 | case Intrinsic::aarch64_sve_umax_x2: |
5560 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5561 | Node->getValueType(0), |
5562 | {AArch64::UMAX_VG2_2Z2Z_B, AArch64::UMAX_VG2_2Z2Z_H, |
5563 | AArch64::UMAX_VG2_2Z2Z_S, AArch64::UMAX_VG2_2Z2Z_D})) |
5564 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
5565 | return; |
5566 | case Intrinsic::aarch64_sve_fmax_x2: |
5567 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5568 | Node->getValueType(0), |
5569 | {0, AArch64::FMAX_VG2_2Z2Z_H, AArch64::FMAX_VG2_2Z2Z_S, |
5570 | AArch64::FMAX_VG2_2Z2Z_D})) |
5571 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
5572 | return; |
5573 | case Intrinsic::aarch64_sve_smax_x4: |
5574 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5575 | Node->getValueType(0), |
5576 | {AArch64::SMAX_VG4_4Z4Z_B, AArch64::SMAX_VG4_4Z4Z_H, |
5577 | AArch64::SMAX_VG4_4Z4Z_S, AArch64::SMAX_VG4_4Z4Z_D})) |
5578 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
5579 | return; |
5580 | case Intrinsic::aarch64_sve_umax_x4: |
5581 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5582 | Node->getValueType(0), |
5583 | {AArch64::UMAX_VG4_4Z4Z_B, AArch64::UMAX_VG4_4Z4Z_H, |
5584 | AArch64::UMAX_VG4_4Z4Z_S, AArch64::UMAX_VG4_4Z4Z_D})) |
5585 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
5586 | return; |
5587 | case Intrinsic::aarch64_sve_fmax_x4: |
5588 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5589 | Node->getValueType(0), |
5590 | {0, AArch64::FMAX_VG4_4Z4Z_H, AArch64::FMAX_VG4_4Z4Z_S, |
5591 | AArch64::FMAX_VG4_4Z4Z_D})) |
5592 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
5593 | return; |
5594 | case Intrinsic::aarch64_sve_smin_x2: |
5595 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5596 | Node->getValueType(0), |
5597 | {AArch64::SMIN_VG2_2Z2Z_B, AArch64::SMIN_VG2_2Z2Z_H, |
5598 | AArch64::SMIN_VG2_2Z2Z_S, AArch64::SMIN_VG2_2Z2Z_D})) |
5599 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
5600 | return; |
5601 | case Intrinsic::aarch64_sve_umin_x2: |
5602 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5603 | Node->getValueType(0), |
5604 | {AArch64::UMIN_VG2_2Z2Z_B, AArch64::UMIN_VG2_2Z2Z_H, |
5605 | AArch64::UMIN_VG2_2Z2Z_S, AArch64::UMIN_VG2_2Z2Z_D})) |
5606 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
5607 | return; |
5608 | case Intrinsic::aarch64_sve_fmin_x2: |
5609 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5610 | Node->getValueType(0), |
5611 | {0, AArch64::FMIN_VG2_2Z2Z_H, AArch64::FMIN_VG2_2Z2Z_S, |
5612 | AArch64::FMIN_VG2_2Z2Z_D})) |
5613 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
5614 | return; |
5615 | case Intrinsic::aarch64_sve_smin_x4: |
5616 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5617 | Node->getValueType(0), |
5618 | {AArch64::SMIN_VG4_4Z4Z_B, AArch64::SMIN_VG4_4Z4Z_H, |
5619 | AArch64::SMIN_VG4_4Z4Z_S, AArch64::SMIN_VG4_4Z4Z_D})) |
5620 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
5621 | return; |
5622 | case Intrinsic::aarch64_sve_umin_x4: |
5623 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5624 | Node->getValueType(0), |
5625 | {AArch64::UMIN_VG4_4Z4Z_B, AArch64::UMIN_VG4_4Z4Z_H, |
5626 | AArch64::UMIN_VG4_4Z4Z_S, AArch64::UMIN_VG4_4Z4Z_D})) |
5627 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
5628 | return; |
5629 | case Intrinsic::aarch64_sve_fmin_x4: |
5630 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5631 | Node->getValueType(0), |
5632 | {0, AArch64::FMIN_VG4_4Z4Z_H, AArch64::FMIN_VG4_4Z4Z_S, |
5633 | AArch64::FMIN_VG4_4Z4Z_D})) |
5634 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
5635 | return; |
5636 | case Intrinsic::aarch64_sve_fmaxnm_single_x2 : |
5637 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5638 | Node->getValueType(0), |
5639 | {0, AArch64::FMAXNM_VG2_2ZZ_H, AArch64::FMAXNM_VG2_2ZZ_S, |
5640 | AArch64::FMAXNM_VG2_2ZZ_D})) |
5641 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
5642 | return; |
5643 | case Intrinsic::aarch64_sve_fmaxnm_single_x4 : |
5644 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5645 | Node->getValueType(0), |
5646 | {0, AArch64::FMAXNM_VG4_4ZZ_H, AArch64::FMAXNM_VG4_4ZZ_S, |
5647 | AArch64::FMAXNM_VG4_4ZZ_D})) |
5648 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
5649 | return; |
5650 | case Intrinsic::aarch64_sve_fminnm_single_x2: |
5651 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5652 | Node->getValueType(0), |
5653 | {0, AArch64::FMINNM_VG2_2ZZ_H, AArch64::FMINNM_VG2_2ZZ_S, |
5654 | AArch64::FMINNM_VG2_2ZZ_D})) |
5655 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
5656 | return; |
5657 | case Intrinsic::aarch64_sve_fminnm_single_x4: |
5658 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5659 | Node->getValueType(0), |
5660 | {0, AArch64::FMINNM_VG4_4ZZ_H, AArch64::FMINNM_VG4_4ZZ_S, |
5661 | AArch64::FMINNM_VG4_4ZZ_D})) |
5662 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
5663 | return; |
5664 | case Intrinsic::aarch64_sve_fmaxnm_x2: |
5665 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5666 | Node->getValueType(0), |
5667 | {0, AArch64::FMAXNM_VG2_2Z2Z_H, AArch64::FMAXNM_VG2_2Z2Z_S, |
5668 | AArch64::FMAXNM_VG2_2Z2Z_D})) |
5669 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
5670 | return; |
5671 | case Intrinsic::aarch64_sve_fmaxnm_x4: |
5672 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5673 | Node->getValueType(0), |
5674 | {0, AArch64::FMAXNM_VG4_4Z4Z_H, AArch64::FMAXNM_VG4_4Z4Z_S, |
5675 | AArch64::FMAXNM_VG4_4Z4Z_D})) |
5676 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
5677 | return; |
5678 | case Intrinsic::aarch64_sve_fminnm_x2: |
5679 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5680 | Node->getValueType(0), |
5681 | {0, AArch64::FMINNM_VG2_2Z2Z_H, AArch64::FMINNM_VG2_2Z2Z_S, |
5682 | AArch64::FMINNM_VG2_2Z2Z_D})) |
5683 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
5684 | return; |
5685 | case Intrinsic::aarch64_sve_fminnm_x4: |
5686 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5687 | Node->getValueType(0), |
5688 | {0, AArch64::FMINNM_VG4_4Z4Z_H, AArch64::FMINNM_VG4_4Z4Z_S, |
5689 | AArch64::FMINNM_VG4_4Z4Z_D})) |
5690 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
5691 | return; |
5692 | case Intrinsic::aarch64_sve_fcvtzs_x2: |
5693 | SelectCVTIntrinsic(Node, 2, AArch64::FCVTZS_2Z2Z_StoS); |
5694 | return; |
5695 | case Intrinsic::aarch64_sve_scvtf_x2: |
5696 | SelectCVTIntrinsic(Node, 2, AArch64::SCVTF_2Z2Z_StoS); |
5697 | return; |
5698 | case Intrinsic::aarch64_sve_fcvtzu_x2: |
5699 | SelectCVTIntrinsic(Node, 2, AArch64::FCVTZU_2Z2Z_StoS); |
5700 | return; |
5701 | case Intrinsic::aarch64_sve_ucvtf_x2: |
5702 | SelectCVTIntrinsic(Node, 2, AArch64::UCVTF_2Z2Z_StoS); |
5703 | return; |
5704 | case Intrinsic::aarch64_sve_fcvtzs_x4: |
5705 | SelectCVTIntrinsic(Node, 4, AArch64::FCVTZS_4Z4Z_StoS); |
5706 | return; |
5707 | case Intrinsic::aarch64_sve_scvtf_x4: |
5708 | SelectCVTIntrinsic(Node, 4, AArch64::SCVTF_4Z4Z_StoS); |
5709 | return; |
5710 | case Intrinsic::aarch64_sve_fcvtzu_x4: |
5711 | SelectCVTIntrinsic(Node, 4, AArch64::FCVTZU_4Z4Z_StoS); |
5712 | return; |
5713 | case Intrinsic::aarch64_sve_ucvtf_x4: |
5714 | SelectCVTIntrinsic(Node, 4, AArch64::UCVTF_4Z4Z_StoS); |
5715 | return; |
5716 | case Intrinsic::aarch64_sve_sclamp_single_x2: |
5717 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5718 | Node->getValueType(0), |
5719 | {AArch64::SCLAMP_VG2_2Z2Z_B, AArch64::SCLAMP_VG2_2Z2Z_H, |
5720 | AArch64::SCLAMP_VG2_2Z2Z_S, AArch64::SCLAMP_VG2_2Z2Z_D})) |
5721 | SelectClamp(N: Node, NumVecs: 2, Op: Op); |
5722 | return; |
5723 | case Intrinsic::aarch64_sve_uclamp_single_x2: |
5724 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5725 | Node->getValueType(0), |
5726 | {AArch64::UCLAMP_VG2_2Z2Z_B, AArch64::UCLAMP_VG2_2Z2Z_H, |
5727 | AArch64::UCLAMP_VG2_2Z2Z_S, AArch64::UCLAMP_VG2_2Z2Z_D})) |
5728 | SelectClamp(N: Node, NumVecs: 2, Op: Op); |
5729 | return; |
5730 | case Intrinsic::aarch64_sve_fclamp_single_x2: |
5731 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5732 | Node->getValueType(0), |
5733 | {0, AArch64::FCLAMP_VG2_2Z2Z_H, AArch64::FCLAMP_VG2_2Z2Z_S, |
5734 | AArch64::FCLAMP_VG2_2Z2Z_D})) |
5735 | SelectClamp(N: Node, NumVecs: 2, Op: Op); |
5736 | return; |
5737 | case Intrinsic::aarch64_sve_sclamp_single_x4: |
5738 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5739 | Node->getValueType(0), |
5740 | {AArch64::SCLAMP_VG4_4Z4Z_B, AArch64::SCLAMP_VG4_4Z4Z_H, |
5741 | AArch64::SCLAMP_VG4_4Z4Z_S, AArch64::SCLAMP_VG4_4Z4Z_D})) |
5742 | SelectClamp(N: Node, NumVecs: 4, Op: Op); |
5743 | return; |
5744 | case Intrinsic::aarch64_sve_uclamp_single_x4: |
5745 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5746 | Node->getValueType(0), |
5747 | {AArch64::UCLAMP_VG4_4Z4Z_B, AArch64::UCLAMP_VG4_4Z4Z_H, |
5748 | AArch64::UCLAMP_VG4_4Z4Z_S, AArch64::UCLAMP_VG4_4Z4Z_D})) |
5749 | SelectClamp(N: Node, NumVecs: 4, Op: Op); |
5750 | return; |
5751 | case Intrinsic::aarch64_sve_fclamp_single_x4: |
5752 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5753 | Node->getValueType(0), |
5754 | {0, AArch64::FCLAMP_VG4_4Z4Z_H, AArch64::FCLAMP_VG4_4Z4Z_S, |
5755 | AArch64::FCLAMP_VG4_4Z4Z_D})) |
5756 | SelectClamp(N: Node, NumVecs: 4, Op: Op); |
5757 | return; |
5758 | case Intrinsic::aarch64_sve_add_single_x2: |
5759 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5760 | Node->getValueType(0), |
5761 | {AArch64::ADD_VG2_2ZZ_B, AArch64::ADD_VG2_2ZZ_H, |
5762 | AArch64::ADD_VG2_2ZZ_S, AArch64::ADD_VG2_2ZZ_D})) |
5763 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
5764 | return; |
5765 | case Intrinsic::aarch64_sve_add_single_x4: |
5766 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5767 | Node->getValueType(0), |
5768 | {AArch64::ADD_VG4_4ZZ_B, AArch64::ADD_VG4_4ZZ_H, |
5769 | AArch64::ADD_VG4_4ZZ_S, AArch64::ADD_VG4_4ZZ_D})) |
5770 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
5771 | return; |
5772 | case Intrinsic::aarch64_sve_zip_x2: |
5773 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
5774 | Node->getValueType(0), |
5775 | {AArch64::ZIP_VG2_2ZZZ_B, AArch64::ZIP_VG2_2ZZZ_H, |
5776 | AArch64::ZIP_VG2_2ZZZ_S, AArch64::ZIP_VG2_2ZZZ_D})) |
5777 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 2, /*IsTupleInput=*/false, Opc: Op); |
5778 | return; |
5779 | case Intrinsic::aarch64_sve_zipq_x2: |
5780 | SelectUnaryMultiIntrinsic(Node, 2, /*IsTupleInput=*/false, |
5781 | AArch64::ZIP_VG2_2ZZZ_Q); |
5782 | return; |
5783 | case Intrinsic::aarch64_sve_zip_x4: |
5784 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
5785 | Node->getValueType(0), |
5786 | {AArch64::ZIP_VG4_4Z4Z_B, AArch64::ZIP_VG4_4Z4Z_H, |
5787 | AArch64::ZIP_VG4_4Z4Z_S, AArch64::ZIP_VG4_4Z4Z_D})) |
5788 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 4, /*IsTupleInput=*/true, Opc: Op); |
5789 | return; |
5790 | case Intrinsic::aarch64_sve_zipq_x4: |
5791 | SelectUnaryMultiIntrinsic(Node, 4, /*IsTupleInput=*/true, |
5792 | AArch64::ZIP_VG4_4Z4Z_Q); |
5793 | return; |
5794 | case Intrinsic::aarch64_sve_uzp_x2: |
5795 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
5796 | Node->getValueType(0), |
5797 | {AArch64::UZP_VG2_2ZZZ_B, AArch64::UZP_VG2_2ZZZ_H, |
5798 | AArch64::UZP_VG2_2ZZZ_S, AArch64::UZP_VG2_2ZZZ_D})) |
5799 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 2, /*IsTupleInput=*/false, Opc: Op); |
5800 | return; |
5801 | case Intrinsic::aarch64_sve_uzpq_x2: |
5802 | SelectUnaryMultiIntrinsic(Node, 2, /*IsTupleInput=*/false, |
5803 | AArch64::UZP_VG2_2ZZZ_Q); |
5804 | return; |
5805 | case Intrinsic::aarch64_sve_uzp_x4: |
5806 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
5807 | Node->getValueType(0), |
5808 | {AArch64::UZP_VG4_4Z4Z_B, AArch64::UZP_VG4_4Z4Z_H, |
5809 | AArch64::UZP_VG4_4Z4Z_S, AArch64::UZP_VG4_4Z4Z_D})) |
5810 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 4, /*IsTupleInput=*/true, Opc: Op); |
5811 | return; |
5812 | case Intrinsic::aarch64_sve_uzpq_x4: |
5813 | SelectUnaryMultiIntrinsic(Node, 4, /*IsTupleInput=*/true, |
5814 | AArch64::UZP_VG4_4Z4Z_Q); |
5815 | return; |
5816 | case Intrinsic::aarch64_sve_sel_x2: |
5817 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
5818 | Node->getValueType(0), |
5819 | {AArch64::SEL_VG2_2ZC2Z2Z_B, AArch64::SEL_VG2_2ZC2Z2Z_H, |
5820 | AArch64::SEL_VG2_2ZC2Z2Z_S, AArch64::SEL_VG2_2ZC2Z2Z_D})) |
5821 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op, /*HasPred=*/true); |
5822 | return; |
5823 | case Intrinsic::aarch64_sve_sel_x4: |
5824 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
5825 | Node->getValueType(0), |
5826 | {AArch64::SEL_VG4_4ZC4Z4Z_B, AArch64::SEL_VG4_4ZC4Z4Z_H, |
5827 | AArch64::SEL_VG4_4ZC4Z4Z_S, AArch64::SEL_VG4_4ZC4Z4Z_D})) |
5828 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op, /*HasPred=*/true); |
5829 | return; |
5830 | case Intrinsic::aarch64_sve_frinta_x2: |
5831 | SelectFrintFromVT(Node, 2, AArch64::FRINTA_2Z2Z_S); |
5832 | return; |
5833 | case Intrinsic::aarch64_sve_frinta_x4: |
5834 | SelectFrintFromVT(Node, 4, AArch64::FRINTA_4Z4Z_S); |
5835 | return; |
5836 | case Intrinsic::aarch64_sve_frintm_x2: |
5837 | SelectFrintFromVT(Node, 2, AArch64::FRINTM_2Z2Z_S); |
5838 | return; |
5839 | case Intrinsic::aarch64_sve_frintm_x4: |
5840 | SelectFrintFromVT(Node, 4, AArch64::FRINTM_4Z4Z_S); |
5841 | return; |
5842 | case Intrinsic::aarch64_sve_frintn_x2: |
5843 | SelectFrintFromVT(Node, 2, AArch64::FRINTN_2Z2Z_S); |
5844 | return; |
5845 | case Intrinsic::aarch64_sve_frintn_x4: |
5846 | SelectFrintFromVT(Node, 4, AArch64::FRINTN_4Z4Z_S); |
5847 | return; |
5848 | case Intrinsic::aarch64_sve_frintp_x2: |
5849 | SelectFrintFromVT(Node, 2, AArch64::FRINTP_2Z2Z_S); |
5850 | return; |
5851 | case Intrinsic::aarch64_sve_frintp_x4: |
5852 | SelectFrintFromVT(Node, 4, AArch64::FRINTP_4Z4Z_S); |
5853 | return; |
5854 | case Intrinsic::aarch64_sve_sunpk_x2: |
5855 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5856 | Node->getValueType(0), |
5857 | {0, AArch64::SUNPK_VG2_2ZZ_H, AArch64::SUNPK_VG2_2ZZ_S, |
5858 | AArch64::SUNPK_VG2_2ZZ_D})) |
5859 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 2, /*IsTupleInput=*/false, Opc: Op); |
5860 | return; |
5861 | case Intrinsic::aarch64_sve_uunpk_x2: |
5862 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5863 | Node->getValueType(0), |
5864 | {0, AArch64::UUNPK_VG2_2ZZ_H, AArch64::UUNPK_VG2_2ZZ_S, |
5865 | AArch64::UUNPK_VG2_2ZZ_D})) |
5866 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 2, /*IsTupleInput=*/false, Opc: Op); |
5867 | return; |
5868 | case Intrinsic::aarch64_sve_sunpk_x4: |
5869 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5870 | Node->getValueType(0), |
5871 | {0, AArch64::SUNPK_VG4_4Z2Z_H, AArch64::SUNPK_VG4_4Z2Z_S, |
5872 | AArch64::SUNPK_VG4_4Z2Z_D})) |
5873 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 4, /*IsTupleInput=*/true, Opc: Op); |
5874 | return; |
5875 | case Intrinsic::aarch64_sve_uunpk_x4: |
5876 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5877 | Node->getValueType(0), |
5878 | {0, AArch64::UUNPK_VG4_4Z2Z_H, AArch64::UUNPK_VG4_4Z2Z_S, |
5879 | AArch64::UUNPK_VG4_4Z2Z_D})) |
5880 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 4, /*IsTupleInput=*/true, Opc: Op); |
5881 | return; |
5882 | case Intrinsic::aarch64_sve_pext_x2: { |
5883 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
5884 | Node->getValueType(0), |
5885 | {AArch64::PEXT_2PCI_B, AArch64::PEXT_2PCI_H, AArch64::PEXT_2PCI_S, |
5886 | AArch64::PEXT_2PCI_D})) |
5887 | SelectPExtPair(N: Node, Opc: Op); |
5888 | return; |
5889 | } |
5890 | } |
5891 | break; |
5892 | } |
5893 | case ISD::INTRINSIC_VOID: { |
5894 | unsigned IntNo = Node->getConstantOperandVal(Num: 1); |
5895 | if (Node->getNumOperands() >= 3) |
5896 | VT = Node->getOperand(Num: 2)->getValueType(ResNo: 0); |
5897 | switch (IntNo) { |
5898 | default: |
5899 | break; |
5900 | case Intrinsic::aarch64_neon_st1x2: { |
5901 | if (VT == MVT::v8i8) { |
5902 | SelectStore(Node, 2, AArch64::ST1Twov8b); |
5903 | return; |
5904 | } else if (VT == MVT::v16i8) { |
5905 | SelectStore(Node, 2, AArch64::ST1Twov16b); |
5906 | return; |
5907 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || |
5908 | VT == MVT::v4bf16) { |
5909 | SelectStore(Node, 2, AArch64::ST1Twov4h); |
5910 | return; |
5911 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || |
5912 | VT == MVT::v8bf16) { |
5913 | SelectStore(Node, 2, AArch64::ST1Twov8h); |
5914 | return; |
5915 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
5916 | SelectStore(Node, 2, AArch64::ST1Twov2s); |
5917 | return; |
5918 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
5919 | SelectStore(Node, 2, AArch64::ST1Twov4s); |
5920 | return; |
5921 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
5922 | SelectStore(Node, 2, AArch64::ST1Twov2d); |
5923 | return; |
5924 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
5925 | SelectStore(Node, 2, AArch64::ST1Twov1d); |
5926 | return; |
5927 | } |
5928 | break; |
5929 | } |
5930 | case Intrinsic::aarch64_neon_st1x3: { |
5931 | if (VT == MVT::v8i8) { |
5932 | SelectStore(Node, 3, AArch64::ST1Threev8b); |
5933 | return; |
5934 | } else if (VT == MVT::v16i8) { |
5935 | SelectStore(Node, 3, AArch64::ST1Threev16b); |
5936 | return; |
5937 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || |
5938 | VT == MVT::v4bf16) { |
5939 | SelectStore(Node, 3, AArch64::ST1Threev4h); |
5940 | return; |
5941 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || |
5942 | VT == MVT::v8bf16) { |
5943 | SelectStore(Node, 3, AArch64::ST1Threev8h); |
5944 | return; |
5945 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
5946 | SelectStore(Node, 3, AArch64::ST1Threev2s); |
5947 | return; |
5948 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
5949 | SelectStore(Node, 3, AArch64::ST1Threev4s); |
5950 | return; |
5951 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
5952 | SelectStore(Node, 3, AArch64::ST1Threev2d); |
5953 | return; |
5954 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
5955 | SelectStore(Node, 3, AArch64::ST1Threev1d); |
5956 | return; |
5957 | } |
5958 | break; |
5959 | } |
5960 | case Intrinsic::aarch64_neon_st1x4: { |
5961 | if (VT == MVT::v8i8) { |
5962 | SelectStore(Node, 4, AArch64::ST1Fourv8b); |
5963 | return; |
5964 | } else if (VT == MVT::v16i8) { |
5965 | SelectStore(Node, 4, AArch64::ST1Fourv16b); |
5966 | return; |
5967 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || |
5968 | VT == MVT::v4bf16) { |
5969 | SelectStore(Node, 4, AArch64::ST1Fourv4h); |
5970 | return; |
5971 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || |
5972 | VT == MVT::v8bf16) { |
5973 | SelectStore(Node, 4, AArch64::ST1Fourv8h); |
5974 | return; |
5975 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
5976 | SelectStore(Node, 4, AArch64::ST1Fourv2s); |
5977 | return; |
5978 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
5979 | SelectStore(Node, 4, AArch64::ST1Fourv4s); |
5980 | return; |
5981 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
5982 | SelectStore(Node, 4, AArch64::ST1Fourv2d); |
5983 | return; |
5984 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
5985 | SelectStore(Node, 4, AArch64::ST1Fourv1d); |
5986 | return; |
5987 | } |
5988 | break; |
5989 | } |
5990 | case Intrinsic::aarch64_neon_st2: { |
5991 | if (VT == MVT::v8i8) { |
5992 | SelectStore(Node, 2, AArch64::ST2Twov8b); |
5993 | return; |
5994 | } else if (VT == MVT::v16i8) { |
5995 | SelectStore(Node, 2, AArch64::ST2Twov16b); |
5996 | return; |
5997 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || |
5998 | VT == MVT::v4bf16) { |
5999 | SelectStore(Node, 2, AArch64::ST2Twov4h); |
6000 | return; |
6001 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || |
6002 | VT == MVT::v8bf16) { |
6003 | SelectStore(Node, 2, AArch64::ST2Twov8h); |
6004 | return; |
6005 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6006 | SelectStore(Node, 2, AArch64::ST2Twov2s); |
6007 | return; |
6008 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6009 | SelectStore(Node, 2, AArch64::ST2Twov4s); |
6010 | return; |
6011 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6012 | SelectStore(Node, 2, AArch64::ST2Twov2d); |
6013 | return; |
6014 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6015 | SelectStore(Node, 2, AArch64::ST1Twov1d); |
6016 | return; |
6017 | } |
6018 | break; |
6019 | } |
6020 | case Intrinsic::aarch64_neon_st3: { |
6021 | if (VT == MVT::v8i8) { |
6022 | SelectStore(Node, 3, AArch64::ST3Threev8b); |
6023 | return; |
6024 | } else if (VT == MVT::v16i8) { |
6025 | SelectStore(Node, 3, AArch64::ST3Threev16b); |
6026 | return; |
6027 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || |
6028 | VT == MVT::v4bf16) { |
6029 | SelectStore(Node, 3, AArch64::ST3Threev4h); |
6030 | return; |
6031 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || |
6032 | VT == MVT::v8bf16) { |
6033 | SelectStore(Node, 3, AArch64::ST3Threev8h); |
6034 | return; |
6035 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6036 | SelectStore(Node, 3, AArch64::ST3Threev2s); |
6037 | return; |
6038 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6039 | SelectStore(Node, 3, AArch64::ST3Threev4s); |
6040 | return; |
6041 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6042 | SelectStore(Node, 3, AArch64::ST3Threev2d); |
6043 | return; |
6044 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6045 | SelectStore(Node, 3, AArch64::ST1Threev1d); |
6046 | return; |
6047 | } |
6048 | break; |
6049 | } |
6050 | case Intrinsic::aarch64_neon_st4: { |
6051 | if (VT == MVT::v8i8) { |
6052 | SelectStore(Node, 4, AArch64::ST4Fourv8b); |
6053 | return; |
6054 | } else if (VT == MVT::v16i8) { |
6055 | SelectStore(Node, 4, AArch64::ST4Fourv16b); |
6056 | return; |
6057 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || |
6058 | VT == MVT::v4bf16) { |
6059 | SelectStore(Node, 4, AArch64::ST4Fourv4h); |
6060 | return; |
6061 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || |
6062 | VT == MVT::v8bf16) { |
6063 | SelectStore(Node, 4, AArch64::ST4Fourv8h); |
6064 | return; |
6065 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6066 | SelectStore(Node, 4, AArch64::ST4Fourv2s); |
6067 | return; |
6068 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6069 | SelectStore(Node, 4, AArch64::ST4Fourv4s); |
6070 | return; |
6071 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6072 | SelectStore(Node, 4, AArch64::ST4Fourv2d); |
6073 | return; |
6074 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6075 | SelectStore(Node, 4, AArch64::ST1Fourv1d); |
6076 | return; |
6077 | } |
6078 | break; |
6079 | } |
6080 | case Intrinsic::aarch64_neon_st2lane: { |
6081 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
6082 | SelectStoreLane(Node, 2, AArch64::ST2i8); |
6083 | return; |
6084 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
6085 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
6086 | SelectStoreLane(Node, 2, AArch64::ST2i16); |
6087 | return; |
6088 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
6089 | VT == MVT::v2f32) { |
6090 | SelectStoreLane(Node, 2, AArch64::ST2i32); |
6091 | return; |
6092 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
6093 | VT == MVT::v1f64) { |
6094 | SelectStoreLane(Node, 2, AArch64::ST2i64); |
6095 | return; |
6096 | } |
6097 | break; |
6098 | } |
6099 | case Intrinsic::aarch64_neon_st3lane: { |
6100 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
6101 | SelectStoreLane(Node, 3, AArch64::ST3i8); |
6102 | return; |
6103 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
6104 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
6105 | SelectStoreLane(Node, 3, AArch64::ST3i16); |
6106 | return; |
6107 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
6108 | VT == MVT::v2f32) { |
6109 | SelectStoreLane(Node, 3, AArch64::ST3i32); |
6110 | return; |
6111 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
6112 | VT == MVT::v1f64) { |
6113 | SelectStoreLane(Node, 3, AArch64::ST3i64); |
6114 | return; |
6115 | } |
6116 | break; |
6117 | } |
6118 | case Intrinsic::aarch64_neon_st4lane: { |
6119 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
6120 | SelectStoreLane(Node, 4, AArch64::ST4i8); |
6121 | return; |
6122 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
6123 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
6124 | SelectStoreLane(Node, 4, AArch64::ST4i16); |
6125 | return; |
6126 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
6127 | VT == MVT::v2f32) { |
6128 | SelectStoreLane(Node, 4, AArch64::ST4i32); |
6129 | return; |
6130 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
6131 | VT == MVT::v1f64) { |
6132 | SelectStoreLane(Node, 4, AArch64::ST4i64); |
6133 | return; |
6134 | } |
6135 | break; |
6136 | } |
6137 | case Intrinsic::aarch64_sve_st2q: { |
6138 | SelectPredicatedStore(Node, 2, 4, AArch64::ST2Q, AArch64::ST2Q_IMM); |
6139 | return; |
6140 | } |
6141 | case Intrinsic::aarch64_sve_st3q: { |
6142 | SelectPredicatedStore(Node, 3, 4, AArch64::ST3Q, AArch64::ST3Q_IMM); |
6143 | return; |
6144 | } |
6145 | case Intrinsic::aarch64_sve_st4q: { |
6146 | SelectPredicatedStore(Node, 4, 4, AArch64::ST4Q, AArch64::ST4Q_IMM); |
6147 | return; |
6148 | } |
6149 | case Intrinsic::aarch64_sve_st2: { |
6150 | if (VT == MVT::nxv16i8) { |
6151 | SelectPredicatedStore(Node, 2, 0, AArch64::ST2B, AArch64::ST2B_IMM); |
6152 | return; |
6153 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
6154 | VT == MVT::nxv8bf16) { |
6155 | SelectPredicatedStore(Node, 2, 1, AArch64::ST2H, AArch64::ST2H_IMM); |
6156 | return; |
6157 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
6158 | SelectPredicatedStore(Node, 2, 2, AArch64::ST2W, AArch64::ST2W_IMM); |
6159 | return; |
6160 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
6161 | SelectPredicatedStore(Node, 2, 3, AArch64::ST2D, AArch64::ST2D_IMM); |
6162 | return; |
6163 | } |
6164 | break; |
6165 | } |
6166 | case Intrinsic::aarch64_sve_st3: { |
6167 | if (VT == MVT::nxv16i8) { |
6168 | SelectPredicatedStore(Node, 3, 0, AArch64::ST3B, AArch64::ST3B_IMM); |
6169 | return; |
6170 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
6171 | VT == MVT::nxv8bf16) { |
6172 | SelectPredicatedStore(Node, 3, 1, AArch64::ST3H, AArch64::ST3H_IMM); |
6173 | return; |
6174 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
6175 | SelectPredicatedStore(Node, 3, 2, AArch64::ST3W, AArch64::ST3W_IMM); |
6176 | return; |
6177 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
6178 | SelectPredicatedStore(Node, 3, 3, AArch64::ST3D, AArch64::ST3D_IMM); |
6179 | return; |
6180 | } |
6181 | break; |
6182 | } |
6183 | case Intrinsic::aarch64_sve_st4: { |
6184 | if (VT == MVT::nxv16i8) { |
6185 | SelectPredicatedStore(Node, 4, 0, AArch64::ST4B, AArch64::ST4B_IMM); |
6186 | return; |
6187 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
6188 | VT == MVT::nxv8bf16) { |
6189 | SelectPredicatedStore(Node, 4, 1, AArch64::ST4H, AArch64::ST4H_IMM); |
6190 | return; |
6191 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
6192 | SelectPredicatedStore(Node, 4, 2, AArch64::ST4W, AArch64::ST4W_IMM); |
6193 | return; |
6194 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
6195 | SelectPredicatedStore(Node, 4, 3, AArch64::ST4D, AArch64::ST4D_IMM); |
6196 | return; |
6197 | } |
6198 | break; |
6199 | } |
6200 | } |
6201 | break; |
6202 | } |
6203 | case AArch64ISD::LD2post: { |
6204 | if (VT == MVT::v8i8) { |
6205 | SelectPostLoad(Node, 2, AArch64::LD2Twov8b_POST, AArch64::dsub0); |
6206 | return; |
6207 | } else if (VT == MVT::v16i8) { |
6208 | SelectPostLoad(Node, 2, AArch64::LD2Twov16b_POST, AArch64::qsub0); |
6209 | return; |
6210 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6211 | SelectPostLoad(Node, 2, AArch64::LD2Twov4h_POST, AArch64::dsub0); |
6212 | return; |
6213 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6214 | SelectPostLoad(Node, 2, AArch64::LD2Twov8h_POST, AArch64::qsub0); |
6215 | return; |
6216 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6217 | SelectPostLoad(Node, 2, AArch64::LD2Twov2s_POST, AArch64::dsub0); |
6218 | return; |
6219 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6220 | SelectPostLoad(Node, 2, AArch64::LD2Twov4s_POST, AArch64::qsub0); |
6221 | return; |
6222 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6223 | SelectPostLoad(Node, 2, AArch64::LD1Twov1d_POST, AArch64::dsub0); |
6224 | return; |
6225 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6226 | SelectPostLoad(Node, 2, AArch64::LD2Twov2d_POST, AArch64::qsub0); |
6227 | return; |
6228 | } |
6229 | break; |
6230 | } |
6231 | case AArch64ISD::LD3post: { |
6232 | if (VT == MVT::v8i8) { |
6233 | SelectPostLoad(Node, 3, AArch64::LD3Threev8b_POST, AArch64::dsub0); |
6234 | return; |
6235 | } else if (VT == MVT::v16i8) { |
6236 | SelectPostLoad(Node, 3, AArch64::LD3Threev16b_POST, AArch64::qsub0); |
6237 | return; |
6238 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6239 | SelectPostLoad(Node, 3, AArch64::LD3Threev4h_POST, AArch64::dsub0); |
6240 | return; |
6241 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6242 | SelectPostLoad(Node, 3, AArch64::LD3Threev8h_POST, AArch64::qsub0); |
6243 | return; |
6244 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6245 | SelectPostLoad(Node, 3, AArch64::LD3Threev2s_POST, AArch64::dsub0); |
6246 | return; |
6247 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6248 | SelectPostLoad(Node, 3, AArch64::LD3Threev4s_POST, AArch64::qsub0); |
6249 | return; |
6250 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6251 | SelectPostLoad(Node, 3, AArch64::LD1Threev1d_POST, AArch64::dsub0); |
6252 | return; |
6253 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6254 | SelectPostLoad(Node, 3, AArch64::LD3Threev2d_POST, AArch64::qsub0); |
6255 | return; |
6256 | } |
6257 | break; |
6258 | } |
6259 | case AArch64ISD::LD4post: { |
6260 | if (VT == MVT::v8i8) { |
6261 | SelectPostLoad(Node, 4, AArch64::LD4Fourv8b_POST, AArch64::dsub0); |
6262 | return; |
6263 | } else if (VT == MVT::v16i8) { |
6264 | SelectPostLoad(Node, 4, AArch64::LD4Fourv16b_POST, AArch64::qsub0); |
6265 | return; |
6266 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6267 | SelectPostLoad(Node, 4, AArch64::LD4Fourv4h_POST, AArch64::dsub0); |
6268 | return; |
6269 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6270 | SelectPostLoad(Node, 4, AArch64::LD4Fourv8h_POST, AArch64::qsub0); |
6271 | return; |
6272 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6273 | SelectPostLoad(Node, 4, AArch64::LD4Fourv2s_POST, AArch64::dsub0); |
6274 | return; |
6275 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6276 | SelectPostLoad(Node, 4, AArch64::LD4Fourv4s_POST, AArch64::qsub0); |
6277 | return; |
6278 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6279 | SelectPostLoad(Node, 4, AArch64::LD1Fourv1d_POST, AArch64::dsub0); |
6280 | return; |
6281 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6282 | SelectPostLoad(Node, 4, AArch64::LD4Fourv2d_POST, AArch64::qsub0); |
6283 | return; |
6284 | } |
6285 | break; |
6286 | } |
6287 | case AArch64ISD::LD1x2post: { |
6288 | if (VT == MVT::v8i8) { |
6289 | SelectPostLoad(Node, 2, AArch64::LD1Twov8b_POST, AArch64::dsub0); |
6290 | return; |
6291 | } else if (VT == MVT::v16i8) { |
6292 | SelectPostLoad(Node, 2, AArch64::LD1Twov16b_POST, AArch64::qsub0); |
6293 | return; |
6294 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6295 | SelectPostLoad(Node, 2, AArch64::LD1Twov4h_POST, AArch64::dsub0); |
6296 | return; |
6297 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6298 | SelectPostLoad(Node, 2, AArch64::LD1Twov8h_POST, AArch64::qsub0); |
6299 | return; |
6300 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6301 | SelectPostLoad(Node, 2, AArch64::LD1Twov2s_POST, AArch64::dsub0); |
6302 | return; |
6303 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6304 | SelectPostLoad(Node, 2, AArch64::LD1Twov4s_POST, AArch64::qsub0); |
6305 | return; |
6306 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6307 | SelectPostLoad(Node, 2, AArch64::LD1Twov1d_POST, AArch64::dsub0); |
6308 | return; |
6309 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6310 | SelectPostLoad(Node, 2, AArch64::LD1Twov2d_POST, AArch64::qsub0); |
6311 | return; |
6312 | } |
6313 | break; |
6314 | } |
6315 | case AArch64ISD::LD1x3post: { |
6316 | if (VT == MVT::v8i8) { |
6317 | SelectPostLoad(Node, 3, AArch64::LD1Threev8b_POST, AArch64::dsub0); |
6318 | return; |
6319 | } else if (VT == MVT::v16i8) { |
6320 | SelectPostLoad(Node, 3, AArch64::LD1Threev16b_POST, AArch64::qsub0); |
6321 | return; |
6322 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6323 | SelectPostLoad(Node, 3, AArch64::LD1Threev4h_POST, AArch64::dsub0); |
6324 | return; |
6325 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6326 | SelectPostLoad(Node, 3, AArch64::LD1Threev8h_POST, AArch64::qsub0); |
6327 | return; |
6328 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6329 | SelectPostLoad(Node, 3, AArch64::LD1Threev2s_POST, AArch64::dsub0); |
6330 | return; |
6331 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6332 | SelectPostLoad(Node, 3, AArch64::LD1Threev4s_POST, AArch64::qsub0); |
6333 | return; |
6334 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6335 | SelectPostLoad(Node, 3, AArch64::LD1Threev1d_POST, AArch64::dsub0); |
6336 | return; |
6337 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6338 | SelectPostLoad(Node, 3, AArch64::LD1Threev2d_POST, AArch64::qsub0); |
6339 | return; |
6340 | } |
6341 | break; |
6342 | } |
6343 | case AArch64ISD::LD1x4post: { |
6344 | if (VT == MVT::v8i8) { |
6345 | SelectPostLoad(Node, 4, AArch64::LD1Fourv8b_POST, AArch64::dsub0); |
6346 | return; |
6347 | } else if (VT == MVT::v16i8) { |
6348 | SelectPostLoad(Node, 4, AArch64::LD1Fourv16b_POST, AArch64::qsub0); |
6349 | return; |
6350 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6351 | SelectPostLoad(Node, 4, AArch64::LD1Fourv4h_POST, AArch64::dsub0); |
6352 | return; |
6353 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6354 | SelectPostLoad(Node, 4, AArch64::LD1Fourv8h_POST, AArch64::qsub0); |
6355 | return; |
6356 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6357 | SelectPostLoad(Node, 4, AArch64::LD1Fourv2s_POST, AArch64::dsub0); |
6358 | return; |
6359 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6360 | SelectPostLoad(Node, 4, AArch64::LD1Fourv4s_POST, AArch64::qsub0); |
6361 | return; |
6362 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6363 | SelectPostLoad(Node, 4, AArch64::LD1Fourv1d_POST, AArch64::dsub0); |
6364 | return; |
6365 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6366 | SelectPostLoad(Node, 4, AArch64::LD1Fourv2d_POST, AArch64::qsub0); |
6367 | return; |
6368 | } |
6369 | break; |
6370 | } |
6371 | case AArch64ISD::LD1DUPpost: { |
6372 | if (VT == MVT::v8i8) { |
6373 | SelectPostLoad(Node, 1, AArch64::LD1Rv8b_POST, AArch64::dsub0); |
6374 | return; |
6375 | } else if (VT == MVT::v16i8) { |
6376 | SelectPostLoad(Node, 1, AArch64::LD1Rv16b_POST, AArch64::qsub0); |
6377 | return; |
6378 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6379 | SelectPostLoad(Node, 1, AArch64::LD1Rv4h_POST, AArch64::dsub0); |
6380 | return; |
6381 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6382 | SelectPostLoad(Node, 1, AArch64::LD1Rv8h_POST, AArch64::qsub0); |
6383 | return; |
6384 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6385 | SelectPostLoad(Node, 1, AArch64::LD1Rv2s_POST, AArch64::dsub0); |
6386 | return; |
6387 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6388 | SelectPostLoad(Node, 1, AArch64::LD1Rv4s_POST, AArch64::qsub0); |
6389 | return; |
6390 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6391 | SelectPostLoad(Node, 1, AArch64::LD1Rv1d_POST, AArch64::dsub0); |
6392 | return; |
6393 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6394 | SelectPostLoad(Node, 1, AArch64::LD1Rv2d_POST, AArch64::qsub0); |
6395 | return; |
6396 | } |
6397 | break; |
6398 | } |
6399 | case AArch64ISD::LD2DUPpost: { |
6400 | if (VT == MVT::v8i8) { |
6401 | SelectPostLoad(Node, 2, AArch64::LD2Rv8b_POST, AArch64::dsub0); |
6402 | return; |
6403 | } else if (VT == MVT::v16i8) { |
6404 | SelectPostLoad(Node, 2, AArch64::LD2Rv16b_POST, AArch64::qsub0); |
6405 | return; |
6406 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6407 | SelectPostLoad(Node, 2, AArch64::LD2Rv4h_POST, AArch64::dsub0); |
6408 | return; |
6409 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6410 | SelectPostLoad(Node, 2, AArch64::LD2Rv8h_POST, AArch64::qsub0); |
6411 | return; |
6412 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6413 | SelectPostLoad(Node, 2, AArch64::LD2Rv2s_POST, AArch64::dsub0); |
6414 | return; |
6415 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6416 | SelectPostLoad(Node, 2, AArch64::LD2Rv4s_POST, AArch64::qsub0); |
6417 | return; |
6418 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6419 | SelectPostLoad(Node, 2, AArch64::LD2Rv1d_POST, AArch64::dsub0); |
6420 | return; |
6421 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6422 | SelectPostLoad(Node, 2, AArch64::LD2Rv2d_POST, AArch64::qsub0); |
6423 | return; |
6424 | } |
6425 | break; |
6426 | } |
6427 | case AArch64ISD::LD3DUPpost: { |
6428 | if (VT == MVT::v8i8) { |
6429 | SelectPostLoad(Node, 3, AArch64::LD3Rv8b_POST, AArch64::dsub0); |
6430 | return; |
6431 | } else if (VT == MVT::v16i8) { |
6432 | SelectPostLoad(Node, 3, AArch64::LD3Rv16b_POST, AArch64::qsub0); |
6433 | return; |
6434 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6435 | SelectPostLoad(Node, 3, AArch64::LD3Rv4h_POST, AArch64::dsub0); |
6436 | return; |
6437 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6438 | SelectPostLoad(Node, 3, AArch64::LD3Rv8h_POST, AArch64::qsub0); |
6439 | return; |
6440 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6441 | SelectPostLoad(Node, 3, AArch64::LD3Rv2s_POST, AArch64::dsub0); |
6442 | return; |
6443 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6444 | SelectPostLoad(Node, 3, AArch64::LD3Rv4s_POST, AArch64::qsub0); |
6445 | return; |
6446 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6447 | SelectPostLoad(Node, 3, AArch64::LD3Rv1d_POST, AArch64::dsub0); |
6448 | return; |
6449 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6450 | SelectPostLoad(Node, 3, AArch64::LD3Rv2d_POST, AArch64::qsub0); |
6451 | return; |
6452 | } |
6453 | break; |
6454 | } |
6455 | case AArch64ISD::LD4DUPpost: { |
6456 | if (VT == MVT::v8i8) { |
6457 | SelectPostLoad(Node, 4, AArch64::LD4Rv8b_POST, AArch64::dsub0); |
6458 | return; |
6459 | } else if (VT == MVT::v16i8) { |
6460 | SelectPostLoad(Node, 4, AArch64::LD4Rv16b_POST, AArch64::qsub0); |
6461 | return; |
6462 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6463 | SelectPostLoad(Node, 4, AArch64::LD4Rv4h_POST, AArch64::dsub0); |
6464 | return; |
6465 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6466 | SelectPostLoad(Node, 4, AArch64::LD4Rv8h_POST, AArch64::qsub0); |
6467 | return; |
6468 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6469 | SelectPostLoad(Node, 4, AArch64::LD4Rv2s_POST, AArch64::dsub0); |
6470 | return; |
6471 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6472 | SelectPostLoad(Node, 4, AArch64::LD4Rv4s_POST, AArch64::qsub0); |
6473 | return; |
6474 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6475 | SelectPostLoad(Node, 4, AArch64::LD4Rv1d_POST, AArch64::dsub0); |
6476 | return; |
6477 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6478 | SelectPostLoad(Node, 4, AArch64::LD4Rv2d_POST, AArch64::qsub0); |
6479 | return; |
6480 | } |
6481 | break; |
6482 | } |
6483 | case AArch64ISD::LD1LANEpost: { |
6484 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
6485 | SelectPostLoadLane(Node, 1, AArch64::LD1i8_POST); |
6486 | return; |
6487 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
6488 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
6489 | SelectPostLoadLane(Node, 1, AArch64::LD1i16_POST); |
6490 | return; |
6491 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
6492 | VT == MVT::v2f32) { |
6493 | SelectPostLoadLane(Node, 1, AArch64::LD1i32_POST); |
6494 | return; |
6495 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
6496 | VT == MVT::v1f64) { |
6497 | SelectPostLoadLane(Node, 1, AArch64::LD1i64_POST); |
6498 | return; |
6499 | } |
6500 | break; |
6501 | } |
6502 | case AArch64ISD::LD2LANEpost: { |
6503 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
6504 | SelectPostLoadLane(Node, 2, AArch64::LD2i8_POST); |
6505 | return; |
6506 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
6507 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
6508 | SelectPostLoadLane(Node, 2, AArch64::LD2i16_POST); |
6509 | return; |
6510 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
6511 | VT == MVT::v2f32) { |
6512 | SelectPostLoadLane(Node, 2, AArch64::LD2i32_POST); |
6513 | return; |
6514 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
6515 | VT == MVT::v1f64) { |
6516 | SelectPostLoadLane(Node, 2, AArch64::LD2i64_POST); |
6517 | return; |
6518 | } |
6519 | break; |
6520 | } |
6521 | case AArch64ISD::LD3LANEpost: { |
6522 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
6523 | SelectPostLoadLane(Node, 3, AArch64::LD3i8_POST); |
6524 | return; |
6525 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
6526 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
6527 | SelectPostLoadLane(Node, 3, AArch64::LD3i16_POST); |
6528 | return; |
6529 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
6530 | VT == MVT::v2f32) { |
6531 | SelectPostLoadLane(Node, 3, AArch64::LD3i32_POST); |
6532 | return; |
6533 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
6534 | VT == MVT::v1f64) { |
6535 | SelectPostLoadLane(Node, 3, AArch64::LD3i64_POST); |
6536 | return; |
6537 | } |
6538 | break; |
6539 | } |
6540 | case AArch64ISD::LD4LANEpost: { |
6541 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
6542 | SelectPostLoadLane(Node, 4, AArch64::LD4i8_POST); |
6543 | return; |
6544 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
6545 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
6546 | SelectPostLoadLane(Node, 4, AArch64::LD4i16_POST); |
6547 | return; |
6548 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
6549 | VT == MVT::v2f32) { |
6550 | SelectPostLoadLane(Node, 4, AArch64::LD4i32_POST); |
6551 | return; |
6552 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
6553 | VT == MVT::v1f64) { |
6554 | SelectPostLoadLane(Node, 4, AArch64::LD4i64_POST); |
6555 | return; |
6556 | } |
6557 | break; |
6558 | } |
6559 | case AArch64ISD::ST2post: { |
6560 | VT = Node->getOperand(Num: 1).getValueType(); |
6561 | if (VT == MVT::v8i8) { |
6562 | SelectPostStore(Node, 2, AArch64::ST2Twov8b_POST); |
6563 | return; |
6564 | } else if (VT == MVT::v16i8) { |
6565 | SelectPostStore(Node, 2, AArch64::ST2Twov16b_POST); |
6566 | return; |
6567 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6568 | SelectPostStore(Node, 2, AArch64::ST2Twov4h_POST); |
6569 | return; |
6570 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6571 | SelectPostStore(Node, 2, AArch64::ST2Twov8h_POST); |
6572 | return; |
6573 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6574 | SelectPostStore(Node, 2, AArch64::ST2Twov2s_POST); |
6575 | return; |
6576 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6577 | SelectPostStore(Node, 2, AArch64::ST2Twov4s_POST); |
6578 | return; |
6579 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6580 | SelectPostStore(Node, 2, AArch64::ST2Twov2d_POST); |
6581 | return; |
6582 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6583 | SelectPostStore(Node, 2, AArch64::ST1Twov1d_POST); |
6584 | return; |
6585 | } |
6586 | break; |
6587 | } |
6588 | case AArch64ISD::ST3post: { |
6589 | VT = Node->getOperand(Num: 1).getValueType(); |
6590 | if (VT == MVT::v8i8) { |
6591 | SelectPostStore(Node, 3, AArch64::ST3Threev8b_POST); |
6592 | return; |
6593 | } else if (VT == MVT::v16i8) { |
6594 | SelectPostStore(Node, 3, AArch64::ST3Threev16b_POST); |
6595 | return; |
6596 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6597 | SelectPostStore(Node, 3, AArch64::ST3Threev4h_POST); |
6598 | return; |
6599 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6600 | SelectPostStore(Node, 3, AArch64::ST3Threev8h_POST); |
6601 | return; |
6602 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6603 | SelectPostStore(Node, 3, AArch64::ST3Threev2s_POST); |
6604 | return; |
6605 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6606 | SelectPostStore(Node, 3, AArch64::ST3Threev4s_POST); |
6607 | return; |
6608 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6609 | SelectPostStore(Node, 3, AArch64::ST3Threev2d_POST); |
6610 | return; |
6611 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6612 | SelectPostStore(Node, 3, AArch64::ST1Threev1d_POST); |
6613 | return; |
6614 | } |
6615 | break; |
6616 | } |
6617 | case AArch64ISD::ST4post: { |
6618 | VT = Node->getOperand(Num: 1).getValueType(); |
6619 | if (VT == MVT::v8i8) { |
6620 | SelectPostStore(Node, 4, AArch64::ST4Fourv8b_POST); |
6621 | return; |
6622 | } else if (VT == MVT::v16i8) { |
6623 | SelectPostStore(Node, 4, AArch64::ST4Fourv16b_POST); |
6624 | return; |
6625 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6626 | SelectPostStore(Node, 4, AArch64::ST4Fourv4h_POST); |
6627 | return; |
6628 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6629 | SelectPostStore(Node, 4, AArch64::ST4Fourv8h_POST); |
6630 | return; |
6631 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6632 | SelectPostStore(Node, 4, AArch64::ST4Fourv2s_POST); |
6633 | return; |
6634 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6635 | SelectPostStore(Node, 4, AArch64::ST4Fourv4s_POST); |
6636 | return; |
6637 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6638 | SelectPostStore(Node, 4, AArch64::ST4Fourv2d_POST); |
6639 | return; |
6640 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6641 | SelectPostStore(Node, 4, AArch64::ST1Fourv1d_POST); |
6642 | return; |
6643 | } |
6644 | break; |
6645 | } |
6646 | case AArch64ISD::ST1x2post: { |
6647 | VT = Node->getOperand(Num: 1).getValueType(); |
6648 | if (VT == MVT::v8i8) { |
6649 | SelectPostStore(Node, 2, AArch64::ST1Twov8b_POST); |
6650 | return; |
6651 | } else if (VT == MVT::v16i8) { |
6652 | SelectPostStore(Node, 2, AArch64::ST1Twov16b_POST); |
6653 | return; |
6654 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6655 | SelectPostStore(Node, 2, AArch64::ST1Twov4h_POST); |
6656 | return; |
6657 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6658 | SelectPostStore(Node, 2, AArch64::ST1Twov8h_POST); |
6659 | return; |
6660 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6661 | SelectPostStore(Node, 2, AArch64::ST1Twov2s_POST); |
6662 | return; |
6663 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6664 | SelectPostStore(Node, 2, AArch64::ST1Twov4s_POST); |
6665 | return; |
6666 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6667 | SelectPostStore(Node, 2, AArch64::ST1Twov1d_POST); |
6668 | return; |
6669 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6670 | SelectPostStore(Node, 2, AArch64::ST1Twov2d_POST); |
6671 | return; |
6672 | } |
6673 | break; |
6674 | } |
6675 | case AArch64ISD::ST1x3post: { |
6676 | VT = Node->getOperand(Num: 1).getValueType(); |
6677 | if (VT == MVT::v8i8) { |
6678 | SelectPostStore(Node, 3, AArch64::ST1Threev8b_POST); |
6679 | return; |
6680 | } else if (VT == MVT::v16i8) { |
6681 | SelectPostStore(Node, 3, AArch64::ST1Threev16b_POST); |
6682 | return; |
6683 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6684 | SelectPostStore(Node, 3, AArch64::ST1Threev4h_POST); |
6685 | return; |
6686 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16 ) { |
6687 | SelectPostStore(Node, 3, AArch64::ST1Threev8h_POST); |
6688 | return; |
6689 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6690 | SelectPostStore(Node, 3, AArch64::ST1Threev2s_POST); |
6691 | return; |
6692 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6693 | SelectPostStore(Node, 3, AArch64::ST1Threev4s_POST); |
6694 | return; |
6695 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6696 | SelectPostStore(Node, 3, AArch64::ST1Threev1d_POST); |
6697 | return; |
6698 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6699 | SelectPostStore(Node, 3, AArch64::ST1Threev2d_POST); |
6700 | return; |
6701 | } |
6702 | break; |
6703 | } |
6704 | case AArch64ISD::ST1x4post: { |
6705 | VT = Node->getOperand(Num: 1).getValueType(); |
6706 | if (VT == MVT::v8i8) { |
6707 | SelectPostStore(Node, 4, AArch64::ST1Fourv8b_POST); |
6708 | return; |
6709 | } else if (VT == MVT::v16i8) { |
6710 | SelectPostStore(Node, 4, AArch64::ST1Fourv16b_POST); |
6711 | return; |
6712 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6713 | SelectPostStore(Node, 4, AArch64::ST1Fourv4h_POST); |
6714 | return; |
6715 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6716 | SelectPostStore(Node, 4, AArch64::ST1Fourv8h_POST); |
6717 | return; |
6718 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6719 | SelectPostStore(Node, 4, AArch64::ST1Fourv2s_POST); |
6720 | return; |
6721 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6722 | SelectPostStore(Node, 4, AArch64::ST1Fourv4s_POST); |
6723 | return; |
6724 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6725 | SelectPostStore(Node, 4, AArch64::ST1Fourv1d_POST); |
6726 | return; |
6727 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6728 | SelectPostStore(Node, 4, AArch64::ST1Fourv2d_POST); |
6729 | return; |
6730 | } |
6731 | break; |
6732 | } |
6733 | case AArch64ISD::ST2LANEpost: { |
6734 | VT = Node->getOperand(Num: 1).getValueType(); |
6735 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
6736 | SelectPostStoreLane(Node, 2, AArch64::ST2i8_POST); |
6737 | return; |
6738 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
6739 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
6740 | SelectPostStoreLane(Node, 2, AArch64::ST2i16_POST); |
6741 | return; |
6742 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
6743 | VT == MVT::v2f32) { |
6744 | SelectPostStoreLane(Node, 2, AArch64::ST2i32_POST); |
6745 | return; |
6746 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
6747 | VT == MVT::v1f64) { |
6748 | SelectPostStoreLane(Node, 2, AArch64::ST2i64_POST); |
6749 | return; |
6750 | } |
6751 | break; |
6752 | } |
6753 | case AArch64ISD::ST3LANEpost: { |
6754 | VT = Node->getOperand(Num: 1).getValueType(); |
6755 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
6756 | SelectPostStoreLane(Node, 3, AArch64::ST3i8_POST); |
6757 | return; |
6758 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
6759 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
6760 | SelectPostStoreLane(Node, 3, AArch64::ST3i16_POST); |
6761 | return; |
6762 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
6763 | VT == MVT::v2f32) { |
6764 | SelectPostStoreLane(Node, 3, AArch64::ST3i32_POST); |
6765 | return; |
6766 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
6767 | VT == MVT::v1f64) { |
6768 | SelectPostStoreLane(Node, 3, AArch64::ST3i64_POST); |
6769 | return; |
6770 | } |
6771 | break; |
6772 | } |
6773 | case AArch64ISD::ST4LANEpost: { |
6774 | VT = Node->getOperand(Num: 1).getValueType(); |
6775 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
6776 | SelectPostStoreLane(Node, 4, AArch64::ST4i8_POST); |
6777 | return; |
6778 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
6779 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
6780 | SelectPostStoreLane(Node, 4, AArch64::ST4i16_POST); |
6781 | return; |
6782 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
6783 | VT == MVT::v2f32) { |
6784 | SelectPostStoreLane(Node, 4, AArch64::ST4i32_POST); |
6785 | return; |
6786 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
6787 | VT == MVT::v1f64) { |
6788 | SelectPostStoreLane(Node, 4, AArch64::ST4i64_POST); |
6789 | return; |
6790 | } |
6791 | break; |
6792 | } |
6793 | case AArch64ISD::SVE_LD2_MERGE_ZERO: { |
6794 | if (VT == MVT::nxv16i8) { |
6795 | SelectPredicatedLoad(Node, 2, 0, AArch64::LD2B_IMM, AArch64::LD2B); |
6796 | return; |
6797 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
6798 | VT == MVT::nxv8bf16) { |
6799 | SelectPredicatedLoad(Node, 2, 1, AArch64::LD2H_IMM, AArch64::LD2H); |
6800 | return; |
6801 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
6802 | SelectPredicatedLoad(Node, 2, 2, AArch64::LD2W_IMM, AArch64::LD2W); |
6803 | return; |
6804 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
6805 | SelectPredicatedLoad(Node, 2, 3, AArch64::LD2D_IMM, AArch64::LD2D); |
6806 | return; |
6807 | } |
6808 | break; |
6809 | } |
6810 | case AArch64ISD::SVE_LD3_MERGE_ZERO: { |
6811 | if (VT == MVT::nxv16i8) { |
6812 | SelectPredicatedLoad(Node, 3, 0, AArch64::LD3B_IMM, AArch64::LD3B); |
6813 | return; |
6814 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
6815 | VT == MVT::nxv8bf16) { |
6816 | SelectPredicatedLoad(Node, 3, 1, AArch64::LD3H_IMM, AArch64::LD3H); |
6817 | return; |
6818 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
6819 | SelectPredicatedLoad(Node, 3, 2, AArch64::LD3W_IMM, AArch64::LD3W); |
6820 | return; |
6821 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
6822 | SelectPredicatedLoad(Node, 3, 3, AArch64::LD3D_IMM, AArch64::LD3D); |
6823 | return; |
6824 | } |
6825 | break; |
6826 | } |
6827 | case AArch64ISD::SVE_LD4_MERGE_ZERO: { |
6828 | if (VT == MVT::nxv16i8) { |
6829 | SelectPredicatedLoad(Node, 4, 0, AArch64::LD4B_IMM, AArch64::LD4B); |
6830 | return; |
6831 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
6832 | VT == MVT::nxv8bf16) { |
6833 | SelectPredicatedLoad(Node, 4, 1, AArch64::LD4H_IMM, AArch64::LD4H); |
6834 | return; |
6835 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
6836 | SelectPredicatedLoad(Node, 4, 2, AArch64::LD4W_IMM, AArch64::LD4W); |
6837 | return; |
6838 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
6839 | SelectPredicatedLoad(Node, 4, 3, AArch64::LD4D_IMM, AArch64::LD4D); |
6840 | return; |
6841 | } |
6842 | break; |
6843 | } |
6844 | } |
6845 | |
6846 | // Select the default instruction |
6847 | SelectCode(Node); |
6848 | } |
6849 | |
6850 | /// createAArch64ISelDag - This pass converts a legalized DAG into a |
6851 | /// AArch64-specific DAG, ready for instruction scheduling. |
6852 | FunctionPass *llvm::createAArch64ISelDag(AArch64TargetMachine &TM, |
6853 | CodeGenOptLevel OptLevel) { |
6854 | return new AArch64DAGToDAGISel(TM, OptLevel); |
6855 | } |
6856 | |
6857 | /// When \p PredVT is a scalable vector predicate in the form |
6858 | /// MVT::nx<M>xi1, it builds the correspondent scalable vector of |
6859 | /// integers MVT::nx<M>xi<bits> s.t. M x bits = 128. When targeting |
6860 | /// structured vectors (NumVec >1), the output data type is |
6861 | /// MVT::nx<M*NumVec>xi<bits> s.t. M x bits = 128. If the input |
6862 | /// PredVT is not in the form MVT::nx<M>xi1, it returns an invalid |
6863 | /// EVT. |
6864 | static EVT getPackedVectorTypeFromPredicateType(LLVMContext &Ctx, EVT PredVT, |
6865 | unsigned NumVec) { |
6866 | assert(NumVec > 0 && NumVec < 5 && "Invalid number of vectors." ); |
6867 | if (!PredVT.isScalableVector() || PredVT.getVectorElementType() != MVT::i1) |
6868 | return EVT(); |
6869 | |
6870 | if (PredVT != MVT::nxv16i1 && PredVT != MVT::nxv8i1 && |
6871 | PredVT != MVT::nxv4i1 && PredVT != MVT::nxv2i1) |
6872 | return EVT(); |
6873 | |
6874 | ElementCount EC = PredVT.getVectorElementCount(); |
6875 | EVT ScalarVT = |
6876 | EVT::getIntegerVT(Context&: Ctx, BitWidth: AArch64::SVEBitsPerBlock / EC.getKnownMinValue()); |
6877 | EVT MemVT = EVT::getVectorVT(Context&: Ctx, VT: ScalarVT, EC: EC * NumVec); |
6878 | |
6879 | return MemVT; |
6880 | } |
6881 | |
6882 | /// Return the EVT of the data associated to a memory operation in \p |
6883 | /// Root. If such EVT cannot be retrived, it returns an invalid EVT. |
6884 | static EVT getMemVTFromNode(LLVMContext &Ctx, SDNode *Root) { |
6885 | if (isa<MemSDNode>(Val: Root)) |
6886 | return cast<MemSDNode>(Val: Root)->getMemoryVT(); |
6887 | |
6888 | if (isa<MemIntrinsicSDNode>(Val: Root)) |
6889 | return cast<MemIntrinsicSDNode>(Val: Root)->getMemoryVT(); |
6890 | |
6891 | const unsigned Opcode = Root->getOpcode(); |
6892 | // For custom ISD nodes, we have to look at them individually to extract the |
6893 | // type of the data moved to/from memory. |
6894 | switch (Opcode) { |
6895 | case AArch64ISD::LD1_MERGE_ZERO: |
6896 | case AArch64ISD::LD1S_MERGE_ZERO: |
6897 | case AArch64ISD::LDNF1_MERGE_ZERO: |
6898 | case AArch64ISD::LDNF1S_MERGE_ZERO: |
6899 | return cast<VTSDNode>(Val: Root->getOperand(Num: 3))->getVT(); |
6900 | case AArch64ISD::ST1_PRED: |
6901 | return cast<VTSDNode>(Val: Root->getOperand(Num: 4))->getVT(); |
6902 | case AArch64ISD::SVE_LD2_MERGE_ZERO: |
6903 | return getPackedVectorTypeFromPredicateType( |
6904 | Ctx, PredVT: Root->getOperand(Num: 1)->getValueType(ResNo: 0), /*NumVec=*/2); |
6905 | case AArch64ISD::SVE_LD3_MERGE_ZERO: |
6906 | return getPackedVectorTypeFromPredicateType( |
6907 | Ctx, PredVT: Root->getOperand(Num: 1)->getValueType(ResNo: 0), /*NumVec=*/3); |
6908 | case AArch64ISD::SVE_LD4_MERGE_ZERO: |
6909 | return getPackedVectorTypeFromPredicateType( |
6910 | Ctx, PredVT: Root->getOperand(Num: 1)->getValueType(ResNo: 0), /*NumVec=*/4); |
6911 | default: |
6912 | break; |
6913 | } |
6914 | |
6915 | if (Opcode != ISD::INTRINSIC_VOID && Opcode != ISD::INTRINSIC_W_CHAIN) |
6916 | return EVT(); |
6917 | |
6918 | switch (Root->getConstantOperandVal(Num: 1)) { |
6919 | default: |
6920 | return EVT(); |
6921 | case Intrinsic::aarch64_sme_ldr: |
6922 | case Intrinsic::aarch64_sme_str: |
6923 | return MVT::nxv16i8; |
6924 | case Intrinsic::aarch64_sve_prf: |
6925 | // We are using an SVE prefetch intrinsic. Type must be inferred from the |
6926 | // width of the predicate. |
6927 | return getPackedVectorTypeFromPredicateType( |
6928 | Ctx, PredVT: Root->getOperand(Num: 2)->getValueType(ResNo: 0), /*NumVec=*/1); |
6929 | case Intrinsic::aarch64_sve_ld2_sret: |
6930 | case Intrinsic::aarch64_sve_ld2q_sret: |
6931 | return getPackedVectorTypeFromPredicateType( |
6932 | Ctx, PredVT: Root->getOperand(Num: 2)->getValueType(ResNo: 0), /*NumVec=*/2); |
6933 | case Intrinsic::aarch64_sve_st2q: |
6934 | return getPackedVectorTypeFromPredicateType( |
6935 | Ctx, PredVT: Root->getOperand(Num: 4)->getValueType(ResNo: 0), /*NumVec=*/2); |
6936 | case Intrinsic::aarch64_sve_ld3_sret: |
6937 | case Intrinsic::aarch64_sve_ld3q_sret: |
6938 | return getPackedVectorTypeFromPredicateType( |
6939 | Ctx, PredVT: Root->getOperand(Num: 2)->getValueType(ResNo: 0), /*NumVec=*/3); |
6940 | case Intrinsic::aarch64_sve_st3q: |
6941 | return getPackedVectorTypeFromPredicateType( |
6942 | Ctx, PredVT: Root->getOperand(Num: 5)->getValueType(ResNo: 0), /*NumVec=*/3); |
6943 | case Intrinsic::aarch64_sve_ld4_sret: |
6944 | case Intrinsic::aarch64_sve_ld4q_sret: |
6945 | return getPackedVectorTypeFromPredicateType( |
6946 | Ctx, PredVT: Root->getOperand(Num: 2)->getValueType(ResNo: 0), /*NumVec=*/4); |
6947 | case Intrinsic::aarch64_sve_st4q: |
6948 | return getPackedVectorTypeFromPredicateType( |
6949 | Ctx, PredVT: Root->getOperand(Num: 6)->getValueType(ResNo: 0), /*NumVec=*/4); |
6950 | case Intrinsic::aarch64_sve_ld1udq: |
6951 | case Intrinsic::aarch64_sve_st1dq: |
6952 | return EVT(MVT::nxv1i64); |
6953 | case Intrinsic::aarch64_sve_ld1uwq: |
6954 | case Intrinsic::aarch64_sve_st1wq: |
6955 | return EVT(MVT::nxv1i32); |
6956 | } |
6957 | } |
6958 | |
6959 | /// SelectAddrModeIndexedSVE - Attempt selection of the addressing mode: |
6960 | /// Base + OffImm * sizeof(MemVT) for Min >= OffImm <= Max |
6961 | /// where Root is the memory access using N for its address. |
6962 | template <int64_t Min, int64_t Max> |
6963 | bool AArch64DAGToDAGISel::SelectAddrModeIndexedSVE(SDNode *Root, SDValue N, |
6964 | SDValue &Base, |
6965 | SDValue &OffImm) { |
6966 | const EVT MemVT = getMemVTFromNode(Ctx&: *(CurDAG->getContext()), Root); |
6967 | const DataLayout &DL = CurDAG->getDataLayout(); |
6968 | const MachineFrameInfo &MFI = MF->getFrameInfo(); |
6969 | |
6970 | if (N.getOpcode() == ISD::FrameIndex) { |
6971 | int FI = cast<FrameIndexSDNode>(Val&: N)->getIndex(); |
6972 | // We can only encode VL scaled offsets, so only fold in frame indexes |
6973 | // referencing SVE objects. |
6974 | if (MFI.getStackID(ObjectIdx: FI) == TargetStackID::ScalableVector) { |
6975 | Base = CurDAG->getTargetFrameIndex(FI, VT: TLI->getPointerTy(DL)); |
6976 | OffImm = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i64); |
6977 | return true; |
6978 | } |
6979 | |
6980 | return false; |
6981 | } |
6982 | |
6983 | if (MemVT == EVT()) |
6984 | return false; |
6985 | |
6986 | if (N.getOpcode() != ISD::ADD) |
6987 | return false; |
6988 | |
6989 | SDValue VScale = N.getOperand(i: 1); |
6990 | if (VScale.getOpcode() != ISD::VSCALE) |
6991 | return false; |
6992 | |
6993 | TypeSize TS = MemVT.getSizeInBits(); |
6994 | int64_t MemWidthBytes = static_cast<int64_t>(TS.getKnownMinValue()) / 8; |
6995 | int64_t MulImm = cast<ConstantSDNode>(Val: VScale.getOperand(i: 0))->getSExtValue(); |
6996 | |
6997 | if ((MulImm % MemWidthBytes) != 0) |
6998 | return false; |
6999 | |
7000 | int64_t Offset = MulImm / MemWidthBytes; |
7001 | if (Offset < Min || Offset > Max) |
7002 | return false; |
7003 | |
7004 | Base = N.getOperand(i: 0); |
7005 | if (Base.getOpcode() == ISD::FrameIndex) { |
7006 | int FI = cast<FrameIndexSDNode>(Val&: Base)->getIndex(); |
7007 | // We can only encode VL scaled offsets, so only fold in frame indexes |
7008 | // referencing SVE objects. |
7009 | if (MFI.getStackID(ObjectIdx: FI) == TargetStackID::ScalableVector) |
7010 | Base = CurDAG->getTargetFrameIndex(FI, VT: TLI->getPointerTy(DL)); |
7011 | } |
7012 | |
7013 | OffImm = CurDAG->getTargetConstant(Offset, SDLoc(N), MVT::i64); |
7014 | return true; |
7015 | } |
7016 | |
7017 | /// Select register plus register addressing mode for SVE, with scaled |
7018 | /// offset. |
7019 | bool AArch64DAGToDAGISel::SelectSVERegRegAddrMode(SDValue N, unsigned Scale, |
7020 | SDValue &Base, |
7021 | SDValue &Offset) { |
7022 | if (N.getOpcode() != ISD::ADD) |
7023 | return false; |
7024 | |
7025 | // Process an ADD node. |
7026 | const SDValue LHS = N.getOperand(i: 0); |
7027 | const SDValue RHS = N.getOperand(i: 1); |
7028 | |
7029 | // 8 bit data does not come with the SHL node, so it is treated |
7030 | // separately. |
7031 | if (Scale == 0) { |
7032 | Base = LHS; |
7033 | Offset = RHS; |
7034 | return true; |
7035 | } |
7036 | |
7037 | if (auto C = dyn_cast<ConstantSDNode>(Val: RHS)) { |
7038 | int64_t ImmOff = C->getSExtValue(); |
7039 | unsigned Size = 1 << Scale; |
7040 | |
7041 | // To use the reg+reg addressing mode, the immediate must be a multiple of |
7042 | // the vector element's byte size. |
7043 | if (ImmOff % Size) |
7044 | return false; |
7045 | |
7046 | SDLoc DL(N); |
7047 | Base = LHS; |
7048 | Offset = CurDAG->getTargetConstant(ImmOff >> Scale, DL, MVT::i64); |
7049 | SDValue Ops[] = {Offset}; |
7050 | SDNode *MI = CurDAG->getMachineNode(AArch64::MOVi64imm, DL, MVT::i64, Ops); |
7051 | Offset = SDValue(MI, 0); |
7052 | return true; |
7053 | } |
7054 | |
7055 | // Check if the RHS is a shift node with a constant. |
7056 | if (RHS.getOpcode() != ISD::SHL) |
7057 | return false; |
7058 | |
7059 | const SDValue ShiftRHS = RHS.getOperand(i: 1); |
7060 | if (auto *C = dyn_cast<ConstantSDNode>(Val: ShiftRHS)) |
7061 | if (C->getZExtValue() == Scale) { |
7062 | Base = LHS; |
7063 | Offset = RHS.getOperand(i: 0); |
7064 | return true; |
7065 | } |
7066 | |
7067 | return false; |
7068 | } |
7069 | |
7070 | bool AArch64DAGToDAGISel::SelectAllActivePredicate(SDValue N) { |
7071 | const AArch64TargetLowering *TLI = |
7072 | static_cast<const AArch64TargetLowering *>(getTargetLowering()); |
7073 | |
7074 | return TLI->isAllActivePredicate(DAG&: *CurDAG, N); |
7075 | } |
7076 | |
7077 | bool AArch64DAGToDAGISel::SelectAnyPredicate(SDValue N) { |
7078 | EVT VT = N.getValueType(); |
7079 | return VT.isScalableVector() && VT.getVectorElementType() == MVT::i1; |
7080 | } |
7081 | |
7082 | bool AArch64DAGToDAGISel::SelectSMETileSlice(SDValue N, unsigned MaxSize, |
7083 | SDValue &Base, SDValue &Offset, |
7084 | unsigned Scale) { |
7085 | // Try to untangle an ADD node into a 'reg + offset' |
7086 | if (N.getOpcode() == ISD::ADD) |
7087 | if (auto C = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1))) { |
7088 | int64_t ImmOff = C->getSExtValue(); |
7089 | if ((ImmOff > 0 && ImmOff <= MaxSize && (ImmOff % Scale == 0))) { |
7090 | Base = N.getOperand(i: 0); |
7091 | Offset = CurDAG->getTargetConstant(ImmOff / Scale, SDLoc(N), MVT::i64); |
7092 | return true; |
7093 | } |
7094 | } |
7095 | |
7096 | // By default, just match reg + 0. |
7097 | Base = N; |
7098 | Offset = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i64); |
7099 | return true; |
7100 | } |
7101 | |