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 | // https://github.com/llvm/llvm-project/issues/114425 |
35 | #if defined(_MSC_VER) && !defined(__clang__) && !defined(NDEBUG) |
36 | #pragma inline_depth(0) |
37 | #endif |
38 | |
39 | //===--------------------------------------------------------------------===// |
40 | /// AArch64DAGToDAGISel - AArch64 specific code to select AArch64 machine |
41 | /// instructions for SelectionDAG operations. |
42 | /// |
43 | namespace { |
44 | |
45 | class AArch64DAGToDAGISel : public SelectionDAGISel { |
46 | |
47 | /// Subtarget - Keep a pointer to the AArch64Subtarget around so that we can |
48 | /// make the right decision when generating code for different targets. |
49 | const AArch64Subtarget *Subtarget; |
50 | |
51 | public: |
52 | AArch64DAGToDAGISel() = delete; |
53 | |
54 | explicit AArch64DAGToDAGISel(AArch64TargetMachine &tm, |
55 | CodeGenOptLevel OptLevel) |
56 | : SelectionDAGISel(tm, OptLevel), Subtarget(nullptr) {} |
57 | |
58 | bool runOnMachineFunction(MachineFunction &MF) override { |
59 | Subtarget = &MF.getSubtarget<AArch64Subtarget>(); |
60 | return SelectionDAGISel::runOnMachineFunction(mf&: MF); |
61 | } |
62 | |
63 | void Select(SDNode *Node) override; |
64 | |
65 | /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for |
66 | /// inline asm expressions. |
67 | bool SelectInlineAsmMemoryOperand(const SDValue &Op, |
68 | InlineAsm::ConstraintCode ConstraintID, |
69 | std::vector<SDValue> &OutOps) override; |
70 | |
71 | template <signed Low, signed High, signed Scale> |
72 | bool SelectRDVLImm(SDValue N, SDValue &Imm); |
73 | |
74 | bool SelectArithExtendedRegister(SDValue N, SDValue &Reg, SDValue &Shift); |
75 | bool SelectArithUXTXRegister(SDValue N, SDValue &Reg, SDValue &Shift); |
76 | bool SelectArithImmed(SDValue N, SDValue &Val, SDValue &Shift); |
77 | bool SelectNegArithImmed(SDValue N, SDValue &Val, SDValue &Shift); |
78 | bool SelectArithShiftedRegister(SDValue N, SDValue &Reg, SDValue &Shift) { |
79 | return SelectShiftedRegister(N, AllowROR: false, Reg, Shift); |
80 | } |
81 | bool SelectLogicalShiftedRegister(SDValue N, SDValue &Reg, SDValue &Shift) { |
82 | return SelectShiftedRegister(N, AllowROR: true, Reg, Shift); |
83 | } |
84 | bool SelectAddrModeIndexed7S8(SDValue N, SDValue &Base, SDValue &OffImm) { |
85 | return SelectAddrModeIndexed7S(N, Size: 1, Base, OffImm); |
86 | } |
87 | bool SelectAddrModeIndexed7S16(SDValue N, SDValue &Base, SDValue &OffImm) { |
88 | return SelectAddrModeIndexed7S(N, Size: 2, Base, OffImm); |
89 | } |
90 | bool SelectAddrModeIndexed7S32(SDValue N, SDValue &Base, SDValue &OffImm) { |
91 | return SelectAddrModeIndexed7S(N, Size: 4, Base, OffImm); |
92 | } |
93 | bool SelectAddrModeIndexed7S64(SDValue N, SDValue &Base, SDValue &OffImm) { |
94 | return SelectAddrModeIndexed7S(N, Size: 8, Base, OffImm); |
95 | } |
96 | bool SelectAddrModeIndexed7S128(SDValue N, SDValue &Base, SDValue &OffImm) { |
97 | return SelectAddrModeIndexed7S(N, Size: 16, Base, OffImm); |
98 | } |
99 | bool SelectAddrModeIndexedS9S128(SDValue N, SDValue &Base, SDValue &OffImm) { |
100 | return SelectAddrModeIndexedBitWidth(N, IsSignedImm: true, BW: 9, Size: 16, Base, OffImm); |
101 | } |
102 | bool SelectAddrModeIndexedU6S128(SDValue N, SDValue &Base, SDValue &OffImm) { |
103 | return SelectAddrModeIndexedBitWidth(N, IsSignedImm: false, BW: 6, Size: 16, Base, OffImm); |
104 | } |
105 | bool SelectAddrModeIndexed8(SDValue N, SDValue &Base, SDValue &OffImm) { |
106 | return SelectAddrModeIndexed(N, Size: 1, Base, OffImm); |
107 | } |
108 | bool SelectAddrModeIndexed16(SDValue N, SDValue &Base, SDValue &OffImm) { |
109 | return SelectAddrModeIndexed(N, Size: 2, Base, OffImm); |
110 | } |
111 | bool SelectAddrModeIndexed32(SDValue N, SDValue &Base, SDValue &OffImm) { |
112 | return SelectAddrModeIndexed(N, Size: 4, Base, OffImm); |
113 | } |
114 | bool SelectAddrModeIndexed64(SDValue N, SDValue &Base, SDValue &OffImm) { |
115 | return SelectAddrModeIndexed(N, Size: 8, Base, OffImm); |
116 | } |
117 | bool SelectAddrModeIndexed128(SDValue N, SDValue &Base, SDValue &OffImm) { |
118 | return SelectAddrModeIndexed(N, Size: 16, Base, OffImm); |
119 | } |
120 | bool SelectAddrModeUnscaled8(SDValue N, SDValue &Base, SDValue &OffImm) { |
121 | return SelectAddrModeUnscaled(N, Size: 1, Base, OffImm); |
122 | } |
123 | bool SelectAddrModeUnscaled16(SDValue N, SDValue &Base, SDValue &OffImm) { |
124 | return SelectAddrModeUnscaled(N, Size: 2, Base, OffImm); |
125 | } |
126 | bool SelectAddrModeUnscaled32(SDValue N, SDValue &Base, SDValue &OffImm) { |
127 | return SelectAddrModeUnscaled(N, Size: 4, Base, OffImm); |
128 | } |
129 | bool SelectAddrModeUnscaled64(SDValue N, SDValue &Base, SDValue &OffImm) { |
130 | return SelectAddrModeUnscaled(N, Size: 8, Base, OffImm); |
131 | } |
132 | bool SelectAddrModeUnscaled128(SDValue N, SDValue &Base, SDValue &OffImm) { |
133 | return SelectAddrModeUnscaled(N, Size: 16, Base, OffImm); |
134 | } |
135 | template <unsigned Size, unsigned Max> |
136 | bool SelectAddrModeIndexedUImm(SDValue N, SDValue &Base, SDValue &OffImm) { |
137 | // Test if there is an appropriate addressing mode and check if the |
138 | // immediate fits. |
139 | bool Found = SelectAddrModeIndexed(N, Size, Base, OffImm); |
140 | if (Found) { |
141 | if (auto *CI = dyn_cast<ConstantSDNode>(Val&: OffImm)) { |
142 | int64_t C = CI->getSExtValue(); |
143 | if (C <= Max) |
144 | return true; |
145 | } |
146 | } |
147 | |
148 | // Otherwise, base only, materialize address in register. |
149 | Base = N; |
150 | OffImm = CurDAG->getTargetConstant(Val: 0, DL: SDLoc(N), VT: MVT::i64); |
151 | return true; |
152 | } |
153 | |
154 | template<int Width> |
155 | bool SelectAddrModeWRO(SDValue N, SDValue &Base, SDValue &Offset, |
156 | SDValue &SignExtend, SDValue &DoShift) { |
157 | return SelectAddrModeWRO(N, Size: Width / 8, Base, Offset, SignExtend, DoShift); |
158 | } |
159 | |
160 | template<int Width> |
161 | bool SelectAddrModeXRO(SDValue N, SDValue &Base, SDValue &Offset, |
162 | SDValue &SignExtend, SDValue &DoShift) { |
163 | return SelectAddrModeXRO(N, Size: Width / 8, Base, Offset, SignExtend, DoShift); |
164 | } |
165 | |
166 | bool (SDValue N, SDValue &Res) { |
167 | if (Subtarget->isLittleEndian() && N->getOpcode() == ISD::BITCAST) |
168 | N = N->getOperand(Num: 0); |
169 | if (N->getOpcode() != ISD::EXTRACT_SUBVECTOR || |
170 | !isa<ConstantSDNode>(Val: N->getOperand(Num: 1))) |
171 | return false; |
172 | EVT VT = N->getValueType(ResNo: 0); |
173 | EVT LVT = N->getOperand(Num: 0).getValueType(); |
174 | unsigned Index = N->getConstantOperandVal(Num: 1); |
175 | if (!VT.is64BitVector() || !LVT.is128BitVector() || |
176 | Index != VT.getVectorNumElements()) |
177 | return false; |
178 | Res = N->getOperand(Num: 0); |
179 | return true; |
180 | } |
181 | |
182 | bool SelectRoundingVLShr(SDValue N, SDValue &Res1, SDValue &Res2) { |
183 | if (N.getOpcode() != AArch64ISD::VLSHR) |
184 | return false; |
185 | SDValue Op = N->getOperand(Num: 0); |
186 | EVT VT = Op.getValueType(); |
187 | unsigned ShtAmt = N->getConstantOperandVal(Num: 1); |
188 | if (ShtAmt > VT.getScalarSizeInBits() / 2 || Op.getOpcode() != ISD::ADD) |
189 | return false; |
190 | |
191 | APInt Imm; |
192 | if (Op.getOperand(i: 1).getOpcode() == AArch64ISD::MOVIshift) |
193 | Imm = APInt(VT.getScalarSizeInBits(), |
194 | Op.getOperand(i: 1).getConstantOperandVal(i: 0) |
195 | << Op.getOperand(i: 1).getConstantOperandVal(i: 1)); |
196 | else if (Op.getOperand(i: 1).getOpcode() == AArch64ISD::DUP && |
197 | isa<ConstantSDNode>(Val: Op.getOperand(i: 1).getOperand(i: 0))) |
198 | Imm = APInt(VT.getScalarSizeInBits(), |
199 | Op.getOperand(i: 1).getConstantOperandVal(i: 0)); |
200 | else |
201 | return false; |
202 | |
203 | if (Imm != 1ULL << (ShtAmt - 1)) |
204 | return false; |
205 | |
206 | Res1 = Op.getOperand(i: 0); |
207 | Res2 = CurDAG->getTargetConstant(Val: ShtAmt, DL: SDLoc(N), VT: MVT::i32); |
208 | return true; |
209 | } |
210 | |
211 | bool SelectDupZeroOrUndef(SDValue N) { |
212 | switch(N->getOpcode()) { |
213 | case ISD::UNDEF: |
214 | return true; |
215 | case AArch64ISD::DUP: |
216 | case ISD::SPLAT_VECTOR: { |
217 | auto Opnd0 = N->getOperand(Num: 0); |
218 | if (isNullConstant(V: Opnd0)) |
219 | return true; |
220 | if (isNullFPConstant(V: Opnd0)) |
221 | return true; |
222 | break; |
223 | } |
224 | default: |
225 | break; |
226 | } |
227 | |
228 | return false; |
229 | } |
230 | |
231 | bool SelectAny(SDValue) { return true; } |
232 | |
233 | bool SelectDupZero(SDValue N) { |
234 | switch(N->getOpcode()) { |
235 | case AArch64ISD::DUP: |
236 | case ISD::SPLAT_VECTOR: { |
237 | auto Opnd0 = N->getOperand(Num: 0); |
238 | if (isNullConstant(V: Opnd0)) |
239 | return true; |
240 | if (isNullFPConstant(V: Opnd0)) |
241 | return true; |
242 | break; |
243 | } |
244 | } |
245 | |
246 | return false; |
247 | } |
248 | |
249 | template<MVT::SimpleValueType VT> |
250 | bool SelectSVEAddSubImm(SDValue N, SDValue &Imm, SDValue &Shift) { |
251 | return SelectSVEAddSubImm(N, VT, Imm, Shift); |
252 | } |
253 | |
254 | template <MVT::SimpleValueType VT, bool Negate> |
255 | bool SelectSVEAddSubSSatImm(SDValue N, SDValue &Imm, SDValue &Shift) { |
256 | return SelectSVEAddSubSSatImm(N, VT, Imm, Shift, Negate); |
257 | } |
258 | |
259 | template <MVT::SimpleValueType VT> |
260 | bool SelectSVECpyDupImm(SDValue N, SDValue &Imm, SDValue &Shift) { |
261 | return SelectSVECpyDupImm(N, VT, Imm, Shift); |
262 | } |
263 | |
264 | template <MVT::SimpleValueType VT, bool Invert = false> |
265 | bool SelectSVELogicalImm(SDValue N, SDValue &Imm) { |
266 | return SelectSVELogicalImm(N, VT, Imm, Invert); |
267 | } |
268 | |
269 | template <MVT::SimpleValueType VT> |
270 | bool SelectSVEArithImm(SDValue N, SDValue &Imm) { |
271 | return SelectSVEArithImm(N, VT, Imm); |
272 | } |
273 | |
274 | template <unsigned Low, unsigned High, bool AllowSaturation = false> |
275 | bool SelectSVEShiftImm(SDValue N, SDValue &Imm) { |
276 | return SelectSVEShiftImm(N, Low, High, AllowSaturation, Imm); |
277 | } |
278 | |
279 | bool SelectSVEShiftSplatImmR(SDValue N, SDValue &Imm) { |
280 | if (N->getOpcode() != ISD::SPLAT_VECTOR) |
281 | return false; |
282 | |
283 | EVT EltVT = N->getValueType(ResNo: 0).getVectorElementType(); |
284 | return SelectSVEShiftImm(N: N->getOperand(Num: 0), /* Low */ 1, |
285 | /* High */ EltVT.getFixedSizeInBits(), |
286 | /* AllowSaturation */ true, Imm); |
287 | } |
288 | |
289 | // Returns a suitable CNT/INC/DEC/RDVL multiplier to calculate VSCALE*N. |
290 | template<signed Min, signed Max, signed Scale, bool Shift> |
291 | bool SelectCntImm(SDValue N, SDValue &Imm) { |
292 | if (!isa<ConstantSDNode>(Val: N)) |
293 | return false; |
294 | |
295 | int64_t MulImm = cast<ConstantSDNode>(Val&: N)->getSExtValue(); |
296 | if (Shift) |
297 | MulImm = 1LL << MulImm; |
298 | |
299 | if ((MulImm % std::abs(x: Scale)) != 0) |
300 | return false; |
301 | |
302 | MulImm /= Scale; |
303 | if ((MulImm >= Min) && (MulImm <= Max)) { |
304 | Imm = CurDAG->getTargetConstant(Val: MulImm, DL: SDLoc(N), VT: MVT::i32); |
305 | return true; |
306 | } |
307 | |
308 | return false; |
309 | } |
310 | |
311 | template <signed Max, signed Scale> |
312 | bool SelectEXTImm(SDValue N, SDValue &Imm) { |
313 | if (!isa<ConstantSDNode>(Val: N)) |
314 | return false; |
315 | |
316 | int64_t MulImm = cast<ConstantSDNode>(Val&: N)->getSExtValue(); |
317 | |
318 | if (MulImm >= 0 && MulImm <= Max) { |
319 | MulImm *= Scale; |
320 | Imm = CurDAG->getTargetConstant(Val: MulImm, DL: SDLoc(N), VT: MVT::i32); |
321 | return true; |
322 | } |
323 | |
324 | return false; |
325 | } |
326 | |
327 | template <unsigned BaseReg, unsigned Max> |
328 | bool ImmToReg(SDValue N, SDValue &Imm) { |
329 | if (auto *CI = dyn_cast<ConstantSDNode>(Val&: N)) { |
330 | uint64_t C = CI->getZExtValue(); |
331 | |
332 | if (C > Max) |
333 | return false; |
334 | |
335 | Imm = CurDAG->getRegister(Reg: BaseReg + C, VT: MVT::Other); |
336 | return true; |
337 | } |
338 | return false; |
339 | } |
340 | |
341 | /// Form sequences of consecutive 64/128-bit registers for use in NEON |
342 | /// instructions making use of a vector-list (e.g. ldN, tbl). Vecs must have |
343 | /// between 1 and 4 elements. If it contains a single element that is returned |
344 | /// unchanged; otherwise a REG_SEQUENCE value is returned. |
345 | SDValue createDTuple(ArrayRef<SDValue> Vecs); |
346 | SDValue createQTuple(ArrayRef<SDValue> Vecs); |
347 | // Form a sequence of SVE registers for instructions using list of vectors, |
348 | // e.g. structured loads and stores (ldN, stN). |
349 | SDValue createZTuple(ArrayRef<SDValue> Vecs); |
350 | |
351 | // Similar to above, except the register must start at a multiple of the |
352 | // tuple, e.g. z2 for a 2-tuple, or z8 for a 4-tuple. |
353 | SDValue createZMulTuple(ArrayRef<SDValue> Regs); |
354 | |
355 | /// Generic helper for the createDTuple/createQTuple |
356 | /// functions. Those should almost always be called instead. |
357 | SDValue createTuple(ArrayRef<SDValue> Vecs, const unsigned RegClassIDs[], |
358 | const unsigned SubRegs[]); |
359 | |
360 | void SelectTable(SDNode *N, unsigned NumVecs, unsigned Opc, bool isExt); |
361 | |
362 | bool tryIndexedLoad(SDNode *N); |
363 | |
364 | void SelectPtrauthAuth(SDNode *N); |
365 | void SelectPtrauthResign(SDNode *N); |
366 | |
367 | bool trySelectStackSlotTagP(SDNode *N); |
368 | void SelectTagP(SDNode *N); |
369 | |
370 | void SelectLoad(SDNode *N, unsigned NumVecs, unsigned Opc, |
371 | unsigned SubRegIdx); |
372 | void SelectPostLoad(SDNode *N, unsigned NumVecs, unsigned Opc, |
373 | unsigned SubRegIdx); |
374 | void SelectLoadLane(SDNode *N, unsigned NumVecs, unsigned Opc); |
375 | void SelectPostLoadLane(SDNode *N, unsigned NumVecs, unsigned Opc); |
376 | void SelectPredicatedLoad(SDNode *N, unsigned NumVecs, unsigned Scale, |
377 | unsigned Opc_rr, unsigned Opc_ri, |
378 | bool IsIntr = false); |
379 | void SelectContiguousMultiVectorLoad(SDNode *N, unsigned NumVecs, |
380 | unsigned Scale, unsigned Opc_ri, |
381 | unsigned Opc_rr); |
382 | void SelectDestructiveMultiIntrinsic(SDNode *N, unsigned NumVecs, |
383 | bool IsZmMulti, unsigned Opcode, |
384 | bool HasPred = false); |
385 | void SelectPExtPair(SDNode *N, unsigned Opc); |
386 | void SelectWhilePair(SDNode *N, unsigned Opc); |
387 | void SelectCVTIntrinsic(SDNode *N, unsigned NumVecs, unsigned Opcode); |
388 | void SelectCVTIntrinsicFP8(SDNode *N, unsigned NumVecs, unsigned Opcode); |
389 | void SelectClamp(SDNode *N, unsigned NumVecs, unsigned Opcode); |
390 | void SelectUnaryMultiIntrinsic(SDNode *N, unsigned NumOutVecs, |
391 | bool IsTupleInput, unsigned Opc); |
392 | void SelectFrintFromVT(SDNode *N, unsigned NumVecs, unsigned Opcode); |
393 | |
394 | template <unsigned MaxIdx, unsigned Scale> |
395 | void SelectMultiVectorMove(SDNode *N, unsigned NumVecs, unsigned BaseReg, |
396 | unsigned Op); |
397 | void SelectMultiVectorMoveZ(SDNode *N, unsigned NumVecs, |
398 | unsigned Op, unsigned MaxIdx, unsigned Scale, |
399 | unsigned BaseReg = 0); |
400 | bool SelectAddrModeFrameIndexSVE(SDValue N, SDValue &Base, SDValue &OffImm); |
401 | /// SVE Reg+Imm addressing mode. |
402 | template <int64_t Min, int64_t Max> |
403 | bool SelectAddrModeIndexedSVE(SDNode *Root, SDValue N, SDValue &Base, |
404 | SDValue &OffImm); |
405 | /// SVE Reg+Reg address mode. |
406 | template <unsigned Scale> |
407 | bool SelectSVERegRegAddrMode(SDValue N, SDValue &Base, SDValue &Offset) { |
408 | return SelectSVERegRegAddrMode(N, Scale, Base, Offset); |
409 | } |
410 | |
411 | void SelectMultiVectorLutiLane(SDNode *Node, unsigned NumOutVecs, |
412 | unsigned Opc, uint32_t MaxImm); |
413 | |
414 | void SelectMultiVectorLuti(SDNode *Node, unsigned NumOutVecs, unsigned Opc); |
415 | |
416 | template <unsigned MaxIdx, unsigned Scale> |
417 | bool SelectSMETileSlice(SDValue N, SDValue &Vector, SDValue &Offset) { |
418 | return SelectSMETileSlice(N, MaxSize: MaxIdx, Vector, Offset, Scale); |
419 | } |
420 | |
421 | void SelectStore(SDNode *N, unsigned NumVecs, unsigned Opc); |
422 | void SelectPostStore(SDNode *N, unsigned NumVecs, unsigned Opc); |
423 | void SelectStoreLane(SDNode *N, unsigned NumVecs, unsigned Opc); |
424 | void SelectPostStoreLane(SDNode *N, unsigned NumVecs, unsigned Opc); |
425 | void SelectPredicatedStore(SDNode *N, unsigned NumVecs, unsigned Scale, |
426 | unsigned Opc_rr, unsigned Opc_ri); |
427 | std::tuple<unsigned, SDValue, SDValue> |
428 | findAddrModeSVELoadStore(SDNode *N, unsigned Opc_rr, unsigned Opc_ri, |
429 | const SDValue &OldBase, const SDValue &OldOffset, |
430 | unsigned Scale); |
431 | |
432 | bool tryBitfieldExtractOp(SDNode *N); |
433 | bool tryBitfieldExtractOpFromSExt(SDNode *N); |
434 | bool tryBitfieldInsertOp(SDNode *N); |
435 | bool tryBitfieldInsertInZeroOp(SDNode *N); |
436 | bool tryShiftAmountMod(SDNode *N); |
437 | |
438 | bool tryReadRegister(SDNode *N); |
439 | bool tryWriteRegister(SDNode *N); |
440 | |
441 | bool trySelectCastFixedLengthToScalableVector(SDNode *N); |
442 | bool trySelectCastScalableToFixedLengthVector(SDNode *N); |
443 | |
444 | bool trySelectXAR(SDNode *N); |
445 | |
446 | // Include the pieces autogenerated from the target description. |
447 | #include "AArch64GenDAGISel.inc" |
448 | |
449 | private: |
450 | bool SelectShiftedRegister(SDValue N, bool AllowROR, SDValue &Reg, |
451 | SDValue &Shift); |
452 | bool SelectShiftedRegisterFromAnd(SDValue N, SDValue &Reg, SDValue &Shift); |
453 | bool SelectAddrModeIndexed7S(SDValue N, unsigned Size, SDValue &Base, |
454 | SDValue &OffImm) { |
455 | return SelectAddrModeIndexedBitWidth(N, IsSignedImm: true, BW: 7, Size, Base, OffImm); |
456 | } |
457 | bool SelectAddrModeIndexedBitWidth(SDValue N, bool IsSignedImm, unsigned BW, |
458 | unsigned Size, SDValue &Base, |
459 | SDValue &OffImm); |
460 | bool SelectAddrModeIndexed(SDValue N, unsigned Size, SDValue &Base, |
461 | SDValue &OffImm); |
462 | bool SelectAddrModeUnscaled(SDValue N, unsigned Size, SDValue &Base, |
463 | SDValue &OffImm); |
464 | bool SelectAddrModeWRO(SDValue N, unsigned Size, SDValue &Base, |
465 | SDValue &Offset, SDValue &SignExtend, |
466 | SDValue &DoShift); |
467 | bool SelectAddrModeXRO(SDValue N, unsigned Size, SDValue &Base, |
468 | SDValue &Offset, SDValue &SignExtend, |
469 | SDValue &DoShift); |
470 | bool isWorthFoldingALU(SDValue V, bool LSL = false) const; |
471 | bool isWorthFoldingAddr(SDValue V, unsigned Size) const; |
472 | bool SelectExtendedSHL(SDValue N, unsigned Size, bool WantExtend, |
473 | SDValue &Offset, SDValue &SignExtend); |
474 | |
475 | template<unsigned RegWidth> |
476 | bool SelectCVTFixedPosOperand(SDValue N, SDValue &FixedPos) { |
477 | return SelectCVTFixedPosOperand(N, FixedPos, Width: RegWidth); |
478 | } |
479 | |
480 | bool SelectCVTFixedPosOperand(SDValue N, SDValue &FixedPos, unsigned Width); |
481 | |
482 | template<unsigned RegWidth> |
483 | bool SelectCVTFixedPosRecipOperand(SDValue N, SDValue &FixedPos) { |
484 | return SelectCVTFixedPosRecipOperand(N, FixedPos, Width: RegWidth); |
485 | } |
486 | |
487 | bool SelectCVTFixedPosRecipOperand(SDValue N, SDValue &FixedPos, |
488 | unsigned Width); |
489 | |
490 | bool SelectCMP_SWAP(SDNode *N); |
491 | |
492 | bool SelectSVEAddSubImm(SDValue N, MVT VT, SDValue &Imm, SDValue &Shift); |
493 | bool SelectSVEAddSubSSatImm(SDValue N, MVT VT, SDValue &Imm, SDValue &Shift, |
494 | bool Negate); |
495 | bool SelectSVECpyDupImm(SDValue N, MVT VT, SDValue &Imm, SDValue &Shift); |
496 | bool SelectSVELogicalImm(SDValue N, MVT VT, SDValue &Imm, bool Invert); |
497 | |
498 | bool SelectSVESignedArithImm(SDValue N, SDValue &Imm); |
499 | bool SelectSVEShiftImm(SDValue N, uint64_t Low, uint64_t High, |
500 | bool AllowSaturation, SDValue &Imm); |
501 | |
502 | bool SelectSVEArithImm(SDValue N, MVT VT, SDValue &Imm); |
503 | bool SelectSVERegRegAddrMode(SDValue N, unsigned Scale, SDValue &Base, |
504 | SDValue &Offset); |
505 | bool SelectSMETileSlice(SDValue N, unsigned MaxSize, SDValue &Vector, |
506 | SDValue &Offset, unsigned Scale = 1); |
507 | |
508 | bool SelectAllActivePredicate(SDValue N); |
509 | bool SelectAnyPredicate(SDValue N); |
510 | |
511 | bool SelectCmpBranchUImm6Operand(SDNode *P, SDValue N, SDValue &Imm); |
512 | }; |
513 | |
514 | class AArch64DAGToDAGISelLegacy : public SelectionDAGISelLegacy { |
515 | public: |
516 | static char ID; |
517 | explicit AArch64DAGToDAGISelLegacy(AArch64TargetMachine &tm, |
518 | CodeGenOptLevel OptLevel) |
519 | : SelectionDAGISelLegacy( |
520 | ID, std::make_unique<AArch64DAGToDAGISel>(args&: tm, args&: OptLevel)) {} |
521 | }; |
522 | } // end anonymous namespace |
523 | |
524 | char AArch64DAGToDAGISelLegacy::ID = 0; |
525 | |
526 | INITIALIZE_PASS(AArch64DAGToDAGISelLegacy, DEBUG_TYPE, PASS_NAME, false, false) |
527 | |
528 | /// isIntImmediate - This method tests to see if the node is a constant |
529 | /// operand. If so Imm will receive the 32-bit value. |
530 | static bool isIntImmediate(const SDNode *N, uint64_t &Imm) { |
531 | if (const ConstantSDNode *C = dyn_cast<const ConstantSDNode>(Val: N)) { |
532 | Imm = C->getZExtValue(); |
533 | return true; |
534 | } |
535 | return false; |
536 | } |
537 | |
538 | // isIntImmediate - This method tests to see if a constant operand. |
539 | // If so Imm will receive the value. |
540 | static bool isIntImmediate(SDValue N, uint64_t &Imm) { |
541 | return isIntImmediate(N: N.getNode(), Imm); |
542 | } |
543 | |
544 | // isOpcWithIntImmediate - This method tests to see if the node is a specific |
545 | // opcode and that it has a immediate integer right operand. |
546 | // If so Imm will receive the 32 bit value. |
547 | static bool isOpcWithIntImmediate(const SDNode *N, unsigned Opc, |
548 | uint64_t &Imm) { |
549 | return N->getOpcode() == Opc && |
550 | isIntImmediate(N: N->getOperand(Num: 1).getNode(), Imm); |
551 | } |
552 | |
553 | // isIntImmediateEq - This method tests to see if N is a constant operand that |
554 | // is equivalent to 'ImmExpected'. |
555 | #ifndef NDEBUG |
556 | static bool isIntImmediateEq(SDValue N, const uint64_t ImmExpected) { |
557 | uint64_t Imm; |
558 | if (!isIntImmediate(N.getNode(), Imm)) |
559 | return false; |
560 | return Imm == ImmExpected; |
561 | } |
562 | #endif |
563 | |
564 | bool AArch64DAGToDAGISel::SelectInlineAsmMemoryOperand( |
565 | const SDValue &Op, const InlineAsm::ConstraintCode ConstraintID, |
566 | std::vector<SDValue> &OutOps) { |
567 | switch(ConstraintID) { |
568 | default: |
569 | llvm_unreachable("Unexpected asm memory constraint" ); |
570 | case InlineAsm::ConstraintCode::m: |
571 | case InlineAsm::ConstraintCode::o: |
572 | case InlineAsm::ConstraintCode::Q: |
573 | // We need to make sure that this one operand does not end up in XZR, thus |
574 | // require the address to be in a PointerRegClass register. |
575 | const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo(); |
576 | const TargetRegisterClass *TRC = TRI->getPointerRegClass(MF: *MF); |
577 | SDLoc dl(Op); |
578 | SDValue RC = CurDAG->getTargetConstant(Val: TRC->getID(), DL: dl, VT: MVT::i64); |
579 | SDValue NewOp = |
580 | SDValue(CurDAG->getMachineNode(Opcode: TargetOpcode::COPY_TO_REGCLASS, |
581 | dl, VT: Op.getValueType(), |
582 | Op1: Op, Op2: RC), 0); |
583 | OutOps.push_back(x: NewOp); |
584 | return false; |
585 | } |
586 | return true; |
587 | } |
588 | |
589 | /// SelectArithImmed - Select an immediate value that can be represented as |
590 | /// a 12-bit value shifted left by either 0 or 12. If so, return true with |
591 | /// Val set to the 12-bit value and Shift set to the shifter operand. |
592 | bool AArch64DAGToDAGISel::SelectArithImmed(SDValue N, SDValue &Val, |
593 | SDValue &Shift) { |
594 | // This function is called from the addsub_shifted_imm ComplexPattern, |
595 | // which lists [imm] as the list of opcode it's interested in, however |
596 | // we still need to check whether the operand is actually an immediate |
597 | // here because the ComplexPattern opcode list is only used in |
598 | // root-level opcode matching. |
599 | if (!isa<ConstantSDNode>(Val: N.getNode())) |
600 | return false; |
601 | |
602 | uint64_t Immed = N.getNode()->getAsZExtVal(); |
603 | unsigned ShiftAmt; |
604 | |
605 | if (Immed >> 12 == 0) { |
606 | ShiftAmt = 0; |
607 | } else if ((Immed & 0xfff) == 0 && Immed >> 24 == 0) { |
608 | ShiftAmt = 12; |
609 | Immed = Immed >> 12; |
610 | } else |
611 | return false; |
612 | |
613 | unsigned ShVal = AArch64_AM::getShifterImm(ST: AArch64_AM::LSL, Imm: ShiftAmt); |
614 | SDLoc dl(N); |
615 | Val = CurDAG->getTargetConstant(Val: Immed, DL: dl, VT: MVT::i32); |
616 | Shift = CurDAG->getTargetConstant(Val: ShVal, DL: dl, VT: MVT::i32); |
617 | return true; |
618 | } |
619 | |
620 | /// SelectNegArithImmed - As above, but negates the value before trying to |
621 | /// select it. |
622 | bool AArch64DAGToDAGISel::SelectNegArithImmed(SDValue N, SDValue &Val, |
623 | SDValue &Shift) { |
624 | // This function is called from the addsub_shifted_imm ComplexPattern, |
625 | // which lists [imm] as the list of opcode it's interested in, however |
626 | // we still need to check whether the operand is actually an immediate |
627 | // here because the ComplexPattern opcode list is only used in |
628 | // root-level opcode matching. |
629 | if (!isa<ConstantSDNode>(Val: N.getNode())) |
630 | return false; |
631 | |
632 | // The immediate operand must be a 24-bit zero-extended immediate. |
633 | uint64_t Immed = N.getNode()->getAsZExtVal(); |
634 | |
635 | // This negation is almost always valid, but "cmp wN, #0" and "cmn wN, #0" |
636 | // have the opposite effect on the C flag, so this pattern mustn't match under |
637 | // those circumstances. |
638 | if (Immed == 0) |
639 | return false; |
640 | |
641 | if (N.getValueType() == MVT::i32) |
642 | Immed = ~((uint32_t)Immed) + 1; |
643 | else |
644 | Immed = ~Immed + 1ULL; |
645 | if (Immed & 0xFFFFFFFFFF000000ULL) |
646 | return false; |
647 | |
648 | Immed &= 0xFFFFFFULL; |
649 | return SelectArithImmed(N: CurDAG->getConstant(Val: Immed, DL: SDLoc(N), VT: MVT::i32), Val, |
650 | Shift); |
651 | } |
652 | |
653 | /// getShiftTypeForNode - Translate a shift node to the corresponding |
654 | /// ShiftType value. |
655 | static AArch64_AM::ShiftExtendType getShiftTypeForNode(SDValue N) { |
656 | switch (N.getOpcode()) { |
657 | default: |
658 | return AArch64_AM::InvalidShiftExtend; |
659 | case ISD::SHL: |
660 | return AArch64_AM::LSL; |
661 | case ISD::SRL: |
662 | return AArch64_AM::LSR; |
663 | case ISD::SRA: |
664 | return AArch64_AM::ASR; |
665 | case ISD::ROTR: |
666 | return AArch64_AM::ROR; |
667 | } |
668 | } |
669 | |
670 | static bool isMemOpOrPrefetch(SDNode *N) { |
671 | return isa<MemSDNode>(Val: *N) || N->getOpcode() == AArch64ISD::PREFETCH; |
672 | } |
673 | |
674 | /// Determine whether it is worth it to fold SHL into the addressing |
675 | /// mode. |
676 | static bool isWorthFoldingSHL(SDValue V) { |
677 | assert(V.getOpcode() == ISD::SHL && "invalid opcode" ); |
678 | // It is worth folding logical shift of up to three places. |
679 | auto *CSD = dyn_cast<ConstantSDNode>(Val: V.getOperand(i: 1)); |
680 | if (!CSD) |
681 | return false; |
682 | unsigned ShiftVal = CSD->getZExtValue(); |
683 | if (ShiftVal > 3) |
684 | return false; |
685 | |
686 | // Check if this particular node is reused in any non-memory related |
687 | // operation. If yes, do not try to fold this node into the address |
688 | // computation, since the computation will be kept. |
689 | const SDNode *Node = V.getNode(); |
690 | for (SDNode *UI : Node->users()) |
691 | if (!isMemOpOrPrefetch(N: UI)) |
692 | for (SDNode *UII : UI->users()) |
693 | if (!isMemOpOrPrefetch(N: UII)) |
694 | return false; |
695 | return true; |
696 | } |
697 | |
698 | /// Determine whether it is worth to fold V into an extended register addressing |
699 | /// mode. |
700 | bool AArch64DAGToDAGISel::isWorthFoldingAddr(SDValue V, unsigned Size) const { |
701 | // Trivial if we are optimizing for code size or if there is only |
702 | // one use of the value. |
703 | if (CurDAG->shouldOptForSize() || V.hasOneUse()) |
704 | return true; |
705 | |
706 | // If a subtarget has a slow shift, folding a shift into multiple loads |
707 | // costs additional micro-ops. |
708 | if (Subtarget->hasAddrLSLSlow14() && (Size == 2 || Size == 16)) |
709 | return false; |
710 | |
711 | // Check whether we're going to emit the address arithmetic anyway because |
712 | // it's used by a non-address operation. |
713 | if (V.getOpcode() == ISD::SHL && isWorthFoldingSHL(V)) |
714 | return true; |
715 | if (V.getOpcode() == ISD::ADD) { |
716 | const SDValue LHS = V.getOperand(i: 0); |
717 | const SDValue RHS = V.getOperand(i: 1); |
718 | if (LHS.getOpcode() == ISD::SHL && isWorthFoldingSHL(V: LHS)) |
719 | return true; |
720 | if (RHS.getOpcode() == ISD::SHL && isWorthFoldingSHL(V: RHS)) |
721 | return true; |
722 | } |
723 | |
724 | // It hurts otherwise, since the value will be reused. |
725 | return false; |
726 | } |
727 | |
728 | /// and (shl/srl/sra, x, c), mask --> shl (srl/sra, x, c1), c2 |
729 | /// to select more shifted register |
730 | bool AArch64DAGToDAGISel::SelectShiftedRegisterFromAnd(SDValue N, SDValue &Reg, |
731 | SDValue &Shift) { |
732 | EVT VT = N.getValueType(); |
733 | if (VT != MVT::i32 && VT != MVT::i64) |
734 | return false; |
735 | |
736 | if (N->getOpcode() != ISD::AND || !N->hasOneUse()) |
737 | return false; |
738 | SDValue LHS = N.getOperand(i: 0); |
739 | if (!LHS->hasOneUse()) |
740 | return false; |
741 | |
742 | unsigned LHSOpcode = LHS->getOpcode(); |
743 | if (LHSOpcode != ISD::SHL && LHSOpcode != ISD::SRL && LHSOpcode != ISD::SRA) |
744 | return false; |
745 | |
746 | ConstantSDNode *ShiftAmtNode = dyn_cast<ConstantSDNode>(Val: LHS.getOperand(i: 1)); |
747 | if (!ShiftAmtNode) |
748 | return false; |
749 | |
750 | uint64_t ShiftAmtC = ShiftAmtNode->getZExtValue(); |
751 | ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1)); |
752 | if (!RHSC) |
753 | return false; |
754 | |
755 | APInt AndMask = RHSC->getAPIntValue(); |
756 | unsigned LowZBits, MaskLen; |
757 | if (!AndMask.isShiftedMask(MaskIdx&: LowZBits, MaskLen)) |
758 | return false; |
759 | |
760 | unsigned BitWidth = N.getValueSizeInBits(); |
761 | SDLoc DL(LHS); |
762 | uint64_t NewShiftC; |
763 | unsigned NewShiftOp; |
764 | if (LHSOpcode == ISD::SHL) { |
765 | // LowZBits <= ShiftAmtC will fall into isBitfieldPositioningOp |
766 | // BitWidth != LowZBits + MaskLen doesn't match the pattern |
767 | if (LowZBits <= ShiftAmtC || (BitWidth != LowZBits + MaskLen)) |
768 | return false; |
769 | |
770 | NewShiftC = LowZBits - ShiftAmtC; |
771 | NewShiftOp = VT == MVT::i64 ? AArch64::UBFMXri : AArch64::UBFMWri; |
772 | } else { |
773 | if (LowZBits == 0) |
774 | return false; |
775 | |
776 | // NewShiftC >= BitWidth will fall into isBitfieldExtractOp |
777 | NewShiftC = LowZBits + ShiftAmtC; |
778 | if (NewShiftC >= BitWidth) |
779 | return false; |
780 | |
781 | // SRA need all high bits |
782 | if (LHSOpcode == ISD::SRA && (BitWidth != (LowZBits + MaskLen))) |
783 | return false; |
784 | |
785 | // SRL high bits can be 0 or 1 |
786 | if (LHSOpcode == ISD::SRL && (BitWidth > (NewShiftC + MaskLen))) |
787 | return false; |
788 | |
789 | if (LHSOpcode == ISD::SRL) |
790 | NewShiftOp = VT == MVT::i64 ? AArch64::UBFMXri : AArch64::UBFMWri; |
791 | else |
792 | NewShiftOp = VT == MVT::i64 ? AArch64::SBFMXri : AArch64::SBFMWri; |
793 | } |
794 | |
795 | assert(NewShiftC < BitWidth && "Invalid shift amount" ); |
796 | SDValue NewShiftAmt = CurDAG->getTargetConstant(Val: NewShiftC, DL, VT); |
797 | SDValue BitWidthMinus1 = CurDAG->getTargetConstant(Val: BitWidth - 1, DL, VT); |
798 | Reg = SDValue(CurDAG->getMachineNode(Opcode: NewShiftOp, dl: DL, VT, Op1: LHS->getOperand(Num: 0), |
799 | Op2: NewShiftAmt, Op3: BitWidthMinus1), |
800 | 0); |
801 | unsigned ShVal = AArch64_AM::getShifterImm(ST: AArch64_AM::LSL, Imm: LowZBits); |
802 | Shift = CurDAG->getTargetConstant(Val: ShVal, DL, VT: MVT::i32); |
803 | return true; |
804 | } |
805 | |
806 | /// getExtendTypeForNode - Translate an extend node to the corresponding |
807 | /// ExtendType value. |
808 | static AArch64_AM::ShiftExtendType |
809 | getExtendTypeForNode(SDValue N, bool IsLoadStore = false) { |
810 | if (N.getOpcode() == ISD::SIGN_EXTEND || |
811 | N.getOpcode() == ISD::SIGN_EXTEND_INREG) { |
812 | EVT SrcVT; |
813 | if (N.getOpcode() == ISD::SIGN_EXTEND_INREG) |
814 | SrcVT = cast<VTSDNode>(Val: N.getOperand(i: 1))->getVT(); |
815 | else |
816 | SrcVT = N.getOperand(i: 0).getValueType(); |
817 | |
818 | if (!IsLoadStore && SrcVT == MVT::i8) |
819 | return AArch64_AM::SXTB; |
820 | else if (!IsLoadStore && SrcVT == MVT::i16) |
821 | return AArch64_AM::SXTH; |
822 | else if (SrcVT == MVT::i32) |
823 | return AArch64_AM::SXTW; |
824 | assert(SrcVT != MVT::i64 && "extend from 64-bits?" ); |
825 | |
826 | return AArch64_AM::InvalidShiftExtend; |
827 | } else if (N.getOpcode() == ISD::ZERO_EXTEND || |
828 | N.getOpcode() == ISD::ANY_EXTEND) { |
829 | EVT SrcVT = N.getOperand(i: 0).getValueType(); |
830 | if (!IsLoadStore && SrcVT == MVT::i8) |
831 | return AArch64_AM::UXTB; |
832 | else if (!IsLoadStore && SrcVT == MVT::i16) |
833 | return AArch64_AM::UXTH; |
834 | else if (SrcVT == MVT::i32) |
835 | return AArch64_AM::UXTW; |
836 | assert(SrcVT != MVT::i64 && "extend from 64-bits?" ); |
837 | |
838 | return AArch64_AM::InvalidShiftExtend; |
839 | } else if (N.getOpcode() == ISD::AND) { |
840 | ConstantSDNode *CSD = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1)); |
841 | if (!CSD) |
842 | return AArch64_AM::InvalidShiftExtend; |
843 | uint64_t AndMask = CSD->getZExtValue(); |
844 | |
845 | switch (AndMask) { |
846 | default: |
847 | return AArch64_AM::InvalidShiftExtend; |
848 | case 0xFF: |
849 | return !IsLoadStore ? AArch64_AM::UXTB : AArch64_AM::InvalidShiftExtend; |
850 | case 0xFFFF: |
851 | return !IsLoadStore ? AArch64_AM::UXTH : AArch64_AM::InvalidShiftExtend; |
852 | case 0xFFFFFFFF: |
853 | return AArch64_AM::UXTW; |
854 | } |
855 | } |
856 | |
857 | return AArch64_AM::InvalidShiftExtend; |
858 | } |
859 | |
860 | /// Determine whether it is worth to fold V into an extended register of an |
861 | /// Add/Sub. LSL means we are folding into an `add w0, w1, w2, lsl #N` |
862 | /// instruction, and the shift should be treated as worth folding even if has |
863 | /// multiple uses. |
864 | bool AArch64DAGToDAGISel::isWorthFoldingALU(SDValue V, bool LSL) const { |
865 | // Trivial if we are optimizing for code size or if there is only |
866 | // one use of the value. |
867 | if (CurDAG->shouldOptForSize() || V.hasOneUse()) |
868 | return true; |
869 | |
870 | // If a subtarget has a fastpath LSL we can fold a logical shift into |
871 | // the add/sub and save a cycle. |
872 | if (LSL && Subtarget->hasALULSLFast() && V.getOpcode() == ISD::SHL && |
873 | V.getConstantOperandVal(i: 1) <= 4 && |
874 | getExtendTypeForNode(N: V.getOperand(i: 0)) == AArch64_AM::InvalidShiftExtend) |
875 | return true; |
876 | |
877 | // It hurts otherwise, since the value will be reused. |
878 | return false; |
879 | } |
880 | |
881 | /// SelectShiftedRegister - Select a "shifted register" operand. If the value |
882 | /// is not shifted, set the Shift operand to default of "LSL 0". The logical |
883 | /// instructions allow the shifted register to be rotated, but the arithmetic |
884 | /// instructions do not. The AllowROR parameter specifies whether ROR is |
885 | /// supported. |
886 | bool AArch64DAGToDAGISel::SelectShiftedRegister(SDValue N, bool AllowROR, |
887 | SDValue &Reg, SDValue &Shift) { |
888 | if (SelectShiftedRegisterFromAnd(N, Reg, Shift)) |
889 | return true; |
890 | |
891 | AArch64_AM::ShiftExtendType ShType = getShiftTypeForNode(N); |
892 | if (ShType == AArch64_AM::InvalidShiftExtend) |
893 | return false; |
894 | if (!AllowROR && ShType == AArch64_AM::ROR) |
895 | return false; |
896 | |
897 | if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1))) { |
898 | unsigned BitSize = N.getValueSizeInBits(); |
899 | unsigned Val = RHS->getZExtValue() & (BitSize - 1); |
900 | unsigned ShVal = AArch64_AM::getShifterImm(ST: ShType, Imm: Val); |
901 | |
902 | Reg = N.getOperand(i: 0); |
903 | Shift = CurDAG->getTargetConstant(Val: ShVal, DL: SDLoc(N), VT: MVT::i32); |
904 | return isWorthFoldingALU(V: N, LSL: true); |
905 | } |
906 | |
907 | return false; |
908 | } |
909 | |
910 | /// Instructions that accept extend modifiers like UXTW expect the register |
911 | /// being extended to be a GPR32, but the incoming DAG might be acting on a |
912 | /// GPR64 (either via SEXT_INREG or AND). Extract the appropriate low bits if |
913 | /// this is the case. |
914 | static SDValue narrowIfNeeded(SelectionDAG *CurDAG, SDValue N) { |
915 | if (N.getValueType() == MVT::i32) |
916 | return N; |
917 | |
918 | SDLoc dl(N); |
919 | return CurDAG->getTargetExtractSubreg(SRIdx: AArch64::sub_32, DL: dl, VT: MVT::i32, Operand: N); |
920 | } |
921 | |
922 | // Returns a suitable CNT/INC/DEC/RDVL multiplier to calculate VSCALE*N. |
923 | template<signed Low, signed High, signed Scale> |
924 | bool AArch64DAGToDAGISel::SelectRDVLImm(SDValue N, SDValue &Imm) { |
925 | if (!isa<ConstantSDNode>(Val: N)) |
926 | return false; |
927 | |
928 | int64_t MulImm = cast<ConstantSDNode>(Val&: N)->getSExtValue(); |
929 | if ((MulImm % std::abs(x: Scale)) == 0) { |
930 | int64_t RDVLImm = MulImm / Scale; |
931 | if ((RDVLImm >= Low) && (RDVLImm <= High)) { |
932 | Imm = CurDAG->getSignedTargetConstant(Val: RDVLImm, DL: SDLoc(N), VT: MVT::i32); |
933 | return true; |
934 | } |
935 | } |
936 | |
937 | return false; |
938 | } |
939 | |
940 | /// SelectArithExtendedRegister - Select a "extended register" operand. This |
941 | /// operand folds in an extend followed by an optional left shift. |
942 | bool AArch64DAGToDAGISel::SelectArithExtendedRegister(SDValue N, SDValue &Reg, |
943 | SDValue &Shift) { |
944 | unsigned ShiftVal = 0; |
945 | AArch64_AM::ShiftExtendType Ext; |
946 | |
947 | if (N.getOpcode() == ISD::SHL) { |
948 | ConstantSDNode *CSD = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1)); |
949 | if (!CSD) |
950 | return false; |
951 | ShiftVal = CSD->getZExtValue(); |
952 | if (ShiftVal > 4) |
953 | return false; |
954 | |
955 | Ext = getExtendTypeForNode(N: N.getOperand(i: 0)); |
956 | if (Ext == AArch64_AM::InvalidShiftExtend) |
957 | return false; |
958 | |
959 | Reg = N.getOperand(i: 0).getOperand(i: 0); |
960 | } else { |
961 | Ext = getExtendTypeForNode(N); |
962 | if (Ext == AArch64_AM::InvalidShiftExtend) |
963 | return false; |
964 | |
965 | Reg = N.getOperand(i: 0); |
966 | |
967 | // Don't match if free 32-bit -> 64-bit zext can be used instead. Use the |
968 | // isDef32 as a heuristic for when the operand is likely to be a 32bit def. |
969 | auto isDef32 = [](SDValue N) { |
970 | unsigned Opc = N.getOpcode(); |
971 | return Opc != ISD::TRUNCATE && Opc != TargetOpcode::EXTRACT_SUBREG && |
972 | Opc != ISD::CopyFromReg && Opc != ISD::AssertSext && |
973 | Opc != ISD::AssertZext && Opc != ISD::AssertAlign && |
974 | Opc != ISD::FREEZE; |
975 | }; |
976 | if (Ext == AArch64_AM::UXTW && Reg->getValueType(ResNo: 0).getSizeInBits() == 32 && |
977 | isDef32(Reg)) |
978 | return false; |
979 | } |
980 | |
981 | // AArch64 mandates that the RHS of the operation must use the smallest |
982 | // register class that could contain the size being extended from. Thus, |
983 | // if we're folding a (sext i8), we need the RHS to be a GPR32, even though |
984 | // there might not be an actual 32-bit value in the program. We can |
985 | // (harmlessly) synthesize one by injected an EXTRACT_SUBREG here. |
986 | assert(Ext != AArch64_AM::UXTX && Ext != AArch64_AM::SXTX); |
987 | Reg = narrowIfNeeded(CurDAG, N: Reg); |
988 | Shift = CurDAG->getTargetConstant(Val: getArithExtendImm(ET: Ext, Imm: ShiftVal), DL: SDLoc(N), |
989 | VT: MVT::i32); |
990 | return isWorthFoldingALU(V: N); |
991 | } |
992 | |
993 | /// SelectArithUXTXRegister - Select a "UXTX register" operand. This |
994 | /// operand is referred by the instructions have SP operand |
995 | bool AArch64DAGToDAGISel::SelectArithUXTXRegister(SDValue N, SDValue &Reg, |
996 | SDValue &Shift) { |
997 | unsigned ShiftVal = 0; |
998 | AArch64_AM::ShiftExtendType Ext; |
999 | |
1000 | if (N.getOpcode() != ISD::SHL) |
1001 | return false; |
1002 | |
1003 | ConstantSDNode *CSD = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1)); |
1004 | if (!CSD) |
1005 | return false; |
1006 | ShiftVal = CSD->getZExtValue(); |
1007 | if (ShiftVal > 4) |
1008 | return false; |
1009 | |
1010 | Ext = AArch64_AM::UXTX; |
1011 | Reg = N.getOperand(i: 0); |
1012 | Shift = CurDAG->getTargetConstant(Val: getArithExtendImm(ET: Ext, Imm: ShiftVal), DL: SDLoc(N), |
1013 | VT: MVT::i32); |
1014 | return isWorthFoldingALU(V: N); |
1015 | } |
1016 | |
1017 | /// If there's a use of this ADDlow that's not itself a load/store then we'll |
1018 | /// need to create a real ADD instruction from it anyway and there's no point in |
1019 | /// folding it into the mem op. Theoretically, it shouldn't matter, but there's |
1020 | /// a single pseudo-instruction for an ADRP/ADD pair so over-aggressive folding |
1021 | /// leads to duplicated ADRP instructions. |
1022 | static bool isWorthFoldingADDlow(SDValue N) { |
1023 | for (auto *User : N->users()) { |
1024 | if (User->getOpcode() != ISD::LOAD && User->getOpcode() != ISD::STORE && |
1025 | User->getOpcode() != ISD::ATOMIC_LOAD && |
1026 | User->getOpcode() != ISD::ATOMIC_STORE) |
1027 | return false; |
1028 | |
1029 | // ldar and stlr have much more restrictive addressing modes (just a |
1030 | // register). |
1031 | if (isStrongerThanMonotonic(AO: cast<MemSDNode>(Val: User)->getSuccessOrdering())) |
1032 | return false; |
1033 | } |
1034 | |
1035 | return true; |
1036 | } |
1037 | |
1038 | /// Check if the immediate offset is valid as a scaled immediate. |
1039 | static bool isValidAsScaledImmediate(int64_t Offset, unsigned Range, |
1040 | unsigned Size) { |
1041 | if ((Offset & (Size - 1)) == 0 && Offset >= 0 && |
1042 | Offset < (Range << Log2_32(Value: Size))) |
1043 | return true; |
1044 | return false; |
1045 | } |
1046 | |
1047 | /// SelectAddrModeIndexedBitWidth - Select a "register plus scaled (un)signed BW-bit |
1048 | /// immediate" address. The "Size" argument is the size in bytes of the memory |
1049 | /// reference, which determines the scale. |
1050 | bool AArch64DAGToDAGISel::SelectAddrModeIndexedBitWidth(SDValue N, bool IsSignedImm, |
1051 | unsigned BW, unsigned Size, |
1052 | SDValue &Base, |
1053 | SDValue &OffImm) { |
1054 | SDLoc dl(N); |
1055 | const DataLayout &DL = CurDAG->getDataLayout(); |
1056 | const TargetLowering *TLI = getTargetLowering(); |
1057 | if (N.getOpcode() == ISD::FrameIndex) { |
1058 | int FI = cast<FrameIndexSDNode>(Val&: N)->getIndex(); |
1059 | Base = CurDAG->getTargetFrameIndex(FI, VT: TLI->getPointerTy(DL)); |
1060 | OffImm = CurDAG->getTargetConstant(Val: 0, DL: dl, VT: MVT::i64); |
1061 | return true; |
1062 | } |
1063 | |
1064 | // As opposed to the (12-bit) Indexed addressing mode below, the 7/9-bit signed |
1065 | // selected here doesn't support labels/immediates, only base+offset. |
1066 | if (CurDAG->isBaseWithConstantOffset(Op: N)) { |
1067 | if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1))) { |
1068 | if (IsSignedImm) { |
1069 | int64_t RHSC = RHS->getSExtValue(); |
1070 | unsigned Scale = Log2_32(Value: Size); |
1071 | int64_t Range = 0x1LL << (BW - 1); |
1072 | |
1073 | if ((RHSC & (Size - 1)) == 0 && RHSC >= -(Range << Scale) && |
1074 | RHSC < (Range << Scale)) { |
1075 | Base = N.getOperand(i: 0); |
1076 | if (Base.getOpcode() == ISD::FrameIndex) { |
1077 | int FI = cast<FrameIndexSDNode>(Val&: Base)->getIndex(); |
1078 | Base = CurDAG->getTargetFrameIndex(FI, VT: TLI->getPointerTy(DL)); |
1079 | } |
1080 | OffImm = CurDAG->getTargetConstant(Val: RHSC >> Scale, DL: dl, VT: MVT::i64); |
1081 | return true; |
1082 | } |
1083 | } else { |
1084 | // unsigned Immediate |
1085 | uint64_t RHSC = RHS->getZExtValue(); |
1086 | unsigned Scale = Log2_32(Value: Size); |
1087 | uint64_t Range = 0x1ULL << BW; |
1088 | |
1089 | if ((RHSC & (Size - 1)) == 0 && RHSC < (Range << Scale)) { |
1090 | Base = N.getOperand(i: 0); |
1091 | if (Base.getOpcode() == ISD::FrameIndex) { |
1092 | int FI = cast<FrameIndexSDNode>(Val&: Base)->getIndex(); |
1093 | Base = CurDAG->getTargetFrameIndex(FI, VT: TLI->getPointerTy(DL)); |
1094 | } |
1095 | OffImm = CurDAG->getTargetConstant(Val: RHSC >> Scale, DL: dl, VT: MVT::i64); |
1096 | return true; |
1097 | } |
1098 | } |
1099 | } |
1100 | } |
1101 | // Base only. The address will be materialized into a register before |
1102 | // the memory is accessed. |
1103 | // add x0, Xbase, #offset |
1104 | // stp x1, x2, [x0] |
1105 | Base = N; |
1106 | OffImm = CurDAG->getTargetConstant(Val: 0, DL: dl, VT: MVT::i64); |
1107 | return true; |
1108 | } |
1109 | |
1110 | /// SelectAddrModeIndexed - Select a "register plus scaled unsigned 12-bit |
1111 | /// immediate" address. The "Size" argument is the size in bytes of the memory |
1112 | /// reference, which determines the scale. |
1113 | bool AArch64DAGToDAGISel::SelectAddrModeIndexed(SDValue N, unsigned Size, |
1114 | SDValue &Base, SDValue &OffImm) { |
1115 | SDLoc dl(N); |
1116 | const DataLayout &DL = CurDAG->getDataLayout(); |
1117 | const TargetLowering *TLI = getTargetLowering(); |
1118 | if (N.getOpcode() == ISD::FrameIndex) { |
1119 | int FI = cast<FrameIndexSDNode>(Val&: N)->getIndex(); |
1120 | Base = CurDAG->getTargetFrameIndex(FI, VT: TLI->getPointerTy(DL)); |
1121 | OffImm = CurDAG->getTargetConstant(Val: 0, DL: dl, VT: MVT::i64); |
1122 | return true; |
1123 | } |
1124 | |
1125 | if (N.getOpcode() == AArch64ISD::ADDlow && isWorthFoldingADDlow(N)) { |
1126 | GlobalAddressSDNode *GAN = |
1127 | dyn_cast<GlobalAddressSDNode>(Val: N.getOperand(i: 1).getNode()); |
1128 | Base = N.getOperand(i: 0); |
1129 | OffImm = N.getOperand(i: 1); |
1130 | if (!GAN) |
1131 | return true; |
1132 | |
1133 | if (GAN->getOffset() % Size == 0 && |
1134 | GAN->getGlobal()->getPointerAlignment(DL) >= Size) |
1135 | return true; |
1136 | } |
1137 | |
1138 | if (CurDAG->isBaseWithConstantOffset(Op: N)) { |
1139 | if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1))) { |
1140 | int64_t RHSC = (int64_t)RHS->getZExtValue(); |
1141 | unsigned Scale = Log2_32(Value: Size); |
1142 | if (isValidAsScaledImmediate(Offset: RHSC, Range: 0x1000, Size)) { |
1143 | Base = N.getOperand(i: 0); |
1144 | if (Base.getOpcode() == ISD::FrameIndex) { |
1145 | int FI = cast<FrameIndexSDNode>(Val&: Base)->getIndex(); |
1146 | Base = CurDAG->getTargetFrameIndex(FI, VT: TLI->getPointerTy(DL)); |
1147 | } |
1148 | OffImm = CurDAG->getTargetConstant(Val: RHSC >> Scale, DL: dl, VT: MVT::i64); |
1149 | return true; |
1150 | } |
1151 | } |
1152 | } |
1153 | |
1154 | // Before falling back to our general case, check if the unscaled |
1155 | // instructions can handle this. If so, that's preferable. |
1156 | if (SelectAddrModeUnscaled(N, Size, Base, OffImm)) |
1157 | return false; |
1158 | |
1159 | // Base only. The address will be materialized into a register before |
1160 | // the memory is accessed. |
1161 | // add x0, Xbase, #offset |
1162 | // ldr x0, [x0] |
1163 | Base = N; |
1164 | OffImm = CurDAG->getTargetConstant(Val: 0, DL: dl, VT: MVT::i64); |
1165 | return true; |
1166 | } |
1167 | |
1168 | /// SelectAddrModeUnscaled - Select a "register plus unscaled signed 9-bit |
1169 | /// immediate" address. This should only match when there is an offset that |
1170 | /// is not valid for a scaled immediate addressing mode. The "Size" argument |
1171 | /// is the size in bytes of the memory reference, which is needed here to know |
1172 | /// what is valid for a scaled immediate. |
1173 | bool AArch64DAGToDAGISel::SelectAddrModeUnscaled(SDValue N, unsigned Size, |
1174 | SDValue &Base, |
1175 | SDValue &OffImm) { |
1176 | if (!CurDAG->isBaseWithConstantOffset(Op: N)) |
1177 | return false; |
1178 | if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1))) { |
1179 | int64_t RHSC = RHS->getSExtValue(); |
1180 | if (RHSC >= -256 && RHSC < 256) { |
1181 | Base = N.getOperand(i: 0); |
1182 | if (Base.getOpcode() == ISD::FrameIndex) { |
1183 | int FI = cast<FrameIndexSDNode>(Val&: Base)->getIndex(); |
1184 | const TargetLowering *TLI = getTargetLowering(); |
1185 | Base = CurDAG->getTargetFrameIndex( |
1186 | FI, VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
1187 | } |
1188 | OffImm = CurDAG->getTargetConstant(Val: RHSC, DL: SDLoc(N), VT: MVT::i64); |
1189 | return true; |
1190 | } |
1191 | } |
1192 | return false; |
1193 | } |
1194 | |
1195 | static SDValue Widen(SelectionDAG *CurDAG, SDValue N) { |
1196 | SDLoc dl(N); |
1197 | SDValue ImpDef = SDValue( |
1198 | CurDAG->getMachineNode(Opcode: TargetOpcode::IMPLICIT_DEF, dl, VT: MVT::i64), 0); |
1199 | return CurDAG->getTargetInsertSubreg(SRIdx: AArch64::sub_32, DL: dl, VT: MVT::i64, Operand: ImpDef, |
1200 | Subreg: N); |
1201 | } |
1202 | |
1203 | /// Check if the given SHL node (\p N), can be used to form an |
1204 | /// extended register for an addressing mode. |
1205 | bool AArch64DAGToDAGISel::SelectExtendedSHL(SDValue N, unsigned Size, |
1206 | bool WantExtend, SDValue &Offset, |
1207 | SDValue &SignExtend) { |
1208 | assert(N.getOpcode() == ISD::SHL && "Invalid opcode." ); |
1209 | ConstantSDNode *CSD = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1)); |
1210 | if (!CSD || (CSD->getZExtValue() & 0x7) != CSD->getZExtValue()) |
1211 | return false; |
1212 | |
1213 | SDLoc dl(N); |
1214 | if (WantExtend) { |
1215 | AArch64_AM::ShiftExtendType Ext = |
1216 | getExtendTypeForNode(N: N.getOperand(i: 0), IsLoadStore: true); |
1217 | if (Ext == AArch64_AM::InvalidShiftExtend) |
1218 | return false; |
1219 | |
1220 | Offset = narrowIfNeeded(CurDAG, N: N.getOperand(i: 0).getOperand(i: 0)); |
1221 | SignExtend = CurDAG->getTargetConstant(Val: Ext == AArch64_AM::SXTW, DL: dl, |
1222 | VT: MVT::i32); |
1223 | } else { |
1224 | Offset = N.getOperand(i: 0); |
1225 | SignExtend = CurDAG->getTargetConstant(Val: 0, DL: dl, VT: MVT::i32); |
1226 | } |
1227 | |
1228 | unsigned LegalShiftVal = Log2_32(Value: Size); |
1229 | unsigned ShiftVal = CSD->getZExtValue(); |
1230 | |
1231 | if (ShiftVal != 0 && ShiftVal != LegalShiftVal) |
1232 | return false; |
1233 | |
1234 | return isWorthFoldingAddr(V: N, Size); |
1235 | } |
1236 | |
1237 | bool AArch64DAGToDAGISel::SelectAddrModeWRO(SDValue N, unsigned Size, |
1238 | SDValue &Base, SDValue &Offset, |
1239 | SDValue &SignExtend, |
1240 | SDValue &DoShift) { |
1241 | if (N.getOpcode() != ISD::ADD) |
1242 | return false; |
1243 | SDValue LHS = N.getOperand(i: 0); |
1244 | SDValue RHS = N.getOperand(i: 1); |
1245 | SDLoc dl(N); |
1246 | |
1247 | // We don't want to match immediate adds here, because they are better lowered |
1248 | // to the register-immediate addressing modes. |
1249 | if (isa<ConstantSDNode>(Val: LHS) || isa<ConstantSDNode>(Val: RHS)) |
1250 | return false; |
1251 | |
1252 | // Check if this particular node is reused in any non-memory related |
1253 | // operation. If yes, do not try to fold this node into the address |
1254 | // computation, since the computation will be kept. |
1255 | const SDNode *Node = N.getNode(); |
1256 | for (SDNode *UI : Node->users()) { |
1257 | if (!isMemOpOrPrefetch(N: UI)) |
1258 | return false; |
1259 | } |
1260 | |
1261 | // Remember if it is worth folding N when it produces extended register. |
1262 | bool IsExtendedRegisterWorthFolding = isWorthFoldingAddr(V: N, Size); |
1263 | |
1264 | // Try to match a shifted extend on the RHS. |
1265 | if (IsExtendedRegisterWorthFolding && RHS.getOpcode() == ISD::SHL && |
1266 | SelectExtendedSHL(N: RHS, Size, WantExtend: true, Offset, SignExtend)) { |
1267 | Base = LHS; |
1268 | DoShift = CurDAG->getTargetConstant(Val: true, DL: dl, VT: MVT::i32); |
1269 | return true; |
1270 | } |
1271 | |
1272 | // Try to match a shifted extend on the LHS. |
1273 | if (IsExtendedRegisterWorthFolding && LHS.getOpcode() == ISD::SHL && |
1274 | SelectExtendedSHL(N: LHS, Size, WantExtend: true, Offset, SignExtend)) { |
1275 | Base = RHS; |
1276 | DoShift = CurDAG->getTargetConstant(Val: true, DL: dl, VT: MVT::i32); |
1277 | return true; |
1278 | } |
1279 | |
1280 | // There was no shift, whatever else we find. |
1281 | DoShift = CurDAG->getTargetConstant(Val: false, DL: dl, VT: MVT::i32); |
1282 | |
1283 | AArch64_AM::ShiftExtendType Ext = AArch64_AM::InvalidShiftExtend; |
1284 | // Try to match an unshifted extend on the LHS. |
1285 | if (IsExtendedRegisterWorthFolding && |
1286 | (Ext = getExtendTypeForNode(N: LHS, IsLoadStore: true)) != |
1287 | AArch64_AM::InvalidShiftExtend) { |
1288 | Base = RHS; |
1289 | Offset = narrowIfNeeded(CurDAG, N: LHS.getOperand(i: 0)); |
1290 | SignExtend = CurDAG->getTargetConstant(Val: Ext == AArch64_AM::SXTW, DL: dl, |
1291 | VT: MVT::i32); |
1292 | if (isWorthFoldingAddr(V: LHS, Size)) |
1293 | return true; |
1294 | } |
1295 | |
1296 | // Try to match an unshifted extend on the RHS. |
1297 | if (IsExtendedRegisterWorthFolding && |
1298 | (Ext = getExtendTypeForNode(N: RHS, IsLoadStore: true)) != |
1299 | AArch64_AM::InvalidShiftExtend) { |
1300 | Base = LHS; |
1301 | Offset = narrowIfNeeded(CurDAG, N: RHS.getOperand(i: 0)); |
1302 | SignExtend = CurDAG->getTargetConstant(Val: Ext == AArch64_AM::SXTW, DL: dl, |
1303 | VT: MVT::i32); |
1304 | if (isWorthFoldingAddr(V: RHS, Size)) |
1305 | return true; |
1306 | } |
1307 | |
1308 | return false; |
1309 | } |
1310 | |
1311 | // Check if the given immediate is preferred by ADD. If an immediate can be |
1312 | // encoded in an ADD, or it can be encoded in an "ADD LSL #12" and can not be |
1313 | // encoded by one MOVZ, return true. |
1314 | static bool isPreferredADD(int64_t ImmOff) { |
1315 | // Constant in [0x0, 0xfff] can be encoded in ADD. |
1316 | if ((ImmOff & 0xfffffffffffff000LL) == 0x0LL) |
1317 | return true; |
1318 | // Check if it can be encoded in an "ADD LSL #12". |
1319 | if ((ImmOff & 0xffffffffff000fffLL) == 0x0LL) |
1320 | // As a single MOVZ is faster than a "ADD of LSL #12", ignore such constant. |
1321 | return (ImmOff & 0xffffffffff00ffffLL) != 0x0LL && |
1322 | (ImmOff & 0xffffffffffff0fffLL) != 0x0LL; |
1323 | return false; |
1324 | } |
1325 | |
1326 | bool AArch64DAGToDAGISel::SelectAddrModeXRO(SDValue N, unsigned Size, |
1327 | SDValue &Base, SDValue &Offset, |
1328 | SDValue &SignExtend, |
1329 | SDValue &DoShift) { |
1330 | if (N.getOpcode() != ISD::ADD) |
1331 | return false; |
1332 | SDValue LHS = N.getOperand(i: 0); |
1333 | SDValue RHS = N.getOperand(i: 1); |
1334 | SDLoc DL(N); |
1335 | |
1336 | // Check if this particular node is reused in any non-memory related |
1337 | // operation. If yes, do not try to fold this node into the address |
1338 | // computation, since the computation will be kept. |
1339 | const SDNode *Node = N.getNode(); |
1340 | for (SDNode *UI : Node->users()) { |
1341 | if (!isMemOpOrPrefetch(N: UI)) |
1342 | return false; |
1343 | } |
1344 | |
1345 | // Watch out if RHS is a wide immediate, it can not be selected into |
1346 | // [BaseReg+Imm] addressing mode. Also it may not be able to be encoded into |
1347 | // ADD/SUB. Instead it will use [BaseReg + 0] address mode and generate |
1348 | // instructions like: |
1349 | // MOV X0, WideImmediate |
1350 | // ADD X1, BaseReg, X0 |
1351 | // LDR X2, [X1, 0] |
1352 | // For such situation, using [BaseReg, XReg] addressing mode can save one |
1353 | // ADD/SUB: |
1354 | // MOV X0, WideImmediate |
1355 | // LDR X2, [BaseReg, X0] |
1356 | if (isa<ConstantSDNode>(Val: RHS)) { |
1357 | int64_t ImmOff = (int64_t)RHS->getAsZExtVal(); |
1358 | // Skip the immediate can be selected by load/store addressing mode. |
1359 | // Also skip the immediate can be encoded by a single ADD (SUB is also |
1360 | // checked by using -ImmOff). |
1361 | if (isValidAsScaledImmediate(Offset: ImmOff, Range: 0x1000, Size) || |
1362 | isPreferredADD(ImmOff) || isPreferredADD(ImmOff: -ImmOff)) |
1363 | return false; |
1364 | |
1365 | SDValue Ops[] = { RHS }; |
1366 | SDNode *MOVI = |
1367 | CurDAG->getMachineNode(Opcode: AArch64::MOVi64imm, dl: DL, VT: MVT::i64, Ops); |
1368 | SDValue MOVIV = SDValue(MOVI, 0); |
1369 | // This ADD of two X register will be selected into [Reg+Reg] mode. |
1370 | N = CurDAG->getNode(Opcode: ISD::ADD, DL, VT: MVT::i64, N1: LHS, N2: MOVIV); |
1371 | } |
1372 | |
1373 | // Remember if it is worth folding N when it produces extended register. |
1374 | bool IsExtendedRegisterWorthFolding = isWorthFoldingAddr(V: N, Size); |
1375 | |
1376 | // Try to match a shifted extend on the RHS. |
1377 | if (IsExtendedRegisterWorthFolding && RHS.getOpcode() == ISD::SHL && |
1378 | SelectExtendedSHL(N: RHS, Size, WantExtend: false, Offset, SignExtend)) { |
1379 | Base = LHS; |
1380 | DoShift = CurDAG->getTargetConstant(Val: true, DL, VT: MVT::i32); |
1381 | return true; |
1382 | } |
1383 | |
1384 | // Try to match a shifted extend on the LHS. |
1385 | if (IsExtendedRegisterWorthFolding && LHS.getOpcode() == ISD::SHL && |
1386 | SelectExtendedSHL(N: LHS, Size, WantExtend: false, Offset, SignExtend)) { |
1387 | Base = RHS; |
1388 | DoShift = CurDAG->getTargetConstant(Val: true, DL, VT: MVT::i32); |
1389 | return true; |
1390 | } |
1391 | |
1392 | // Match any non-shifted, non-extend, non-immediate add expression. |
1393 | Base = LHS; |
1394 | Offset = RHS; |
1395 | SignExtend = CurDAG->getTargetConstant(Val: false, DL, VT: MVT::i32); |
1396 | DoShift = CurDAG->getTargetConstant(Val: false, DL, VT: MVT::i32); |
1397 | // Reg1 + Reg2 is free: no check needed. |
1398 | return true; |
1399 | } |
1400 | |
1401 | SDValue AArch64DAGToDAGISel::createDTuple(ArrayRef<SDValue> Regs) { |
1402 | static const unsigned RegClassIDs[] = { |
1403 | AArch64::DDRegClassID, AArch64::DDDRegClassID, AArch64::DDDDRegClassID}; |
1404 | static const unsigned SubRegs[] = {AArch64::dsub0, AArch64::dsub1, |
1405 | AArch64::dsub2, AArch64::dsub3}; |
1406 | |
1407 | return createTuple(Vecs: Regs, RegClassIDs, SubRegs); |
1408 | } |
1409 | |
1410 | SDValue AArch64DAGToDAGISel::createQTuple(ArrayRef<SDValue> Regs) { |
1411 | static const unsigned RegClassIDs[] = { |
1412 | AArch64::QQRegClassID, AArch64::QQQRegClassID, AArch64::QQQQRegClassID}; |
1413 | static const unsigned SubRegs[] = {AArch64::qsub0, AArch64::qsub1, |
1414 | AArch64::qsub2, AArch64::qsub3}; |
1415 | |
1416 | return createTuple(Vecs: Regs, RegClassIDs, SubRegs); |
1417 | } |
1418 | |
1419 | SDValue AArch64DAGToDAGISel::createZTuple(ArrayRef<SDValue> Regs) { |
1420 | static const unsigned RegClassIDs[] = {AArch64::ZPR2RegClassID, |
1421 | AArch64::ZPR3RegClassID, |
1422 | AArch64::ZPR4RegClassID}; |
1423 | static const unsigned SubRegs[] = {AArch64::zsub0, AArch64::zsub1, |
1424 | AArch64::zsub2, AArch64::zsub3}; |
1425 | |
1426 | return createTuple(Vecs: Regs, RegClassIDs, SubRegs); |
1427 | } |
1428 | |
1429 | SDValue AArch64DAGToDAGISel::createZMulTuple(ArrayRef<SDValue> Regs) { |
1430 | assert(Regs.size() == 2 || Regs.size() == 4); |
1431 | |
1432 | // The createTuple interface requires 3 RegClassIDs for each possible |
1433 | // tuple type even though we only have them for ZPR2 and ZPR4. |
1434 | static const unsigned RegClassIDs[] = {AArch64::ZPR2Mul2RegClassID, 0, |
1435 | AArch64::ZPR4Mul4RegClassID}; |
1436 | static const unsigned SubRegs[] = {AArch64::zsub0, AArch64::zsub1, |
1437 | AArch64::zsub2, AArch64::zsub3}; |
1438 | return createTuple(Vecs: Regs, RegClassIDs, SubRegs); |
1439 | } |
1440 | |
1441 | SDValue AArch64DAGToDAGISel::createTuple(ArrayRef<SDValue> Regs, |
1442 | const unsigned RegClassIDs[], |
1443 | const unsigned SubRegs[]) { |
1444 | // There's no special register-class for a vector-list of 1 element: it's just |
1445 | // a vector. |
1446 | if (Regs.size() == 1) |
1447 | return Regs[0]; |
1448 | |
1449 | assert(Regs.size() >= 2 && Regs.size() <= 4); |
1450 | |
1451 | SDLoc DL(Regs[0]); |
1452 | |
1453 | SmallVector<SDValue, 4> Ops; |
1454 | |
1455 | // First operand of REG_SEQUENCE is the desired RegClass. |
1456 | Ops.push_back( |
1457 | Elt: CurDAG->getTargetConstant(Val: RegClassIDs[Regs.size() - 2], DL, VT: MVT::i32)); |
1458 | |
1459 | // Then we get pairs of source & subregister-position for the components. |
1460 | for (unsigned i = 0; i < Regs.size(); ++i) { |
1461 | Ops.push_back(Elt: Regs[i]); |
1462 | Ops.push_back(Elt: CurDAG->getTargetConstant(Val: SubRegs[i], DL, VT: MVT::i32)); |
1463 | } |
1464 | |
1465 | SDNode *N = |
1466 | CurDAG->getMachineNode(Opcode: TargetOpcode::REG_SEQUENCE, dl: DL, VT: MVT::Untyped, Ops); |
1467 | return SDValue(N, 0); |
1468 | } |
1469 | |
1470 | void AArch64DAGToDAGISel::SelectTable(SDNode *N, unsigned NumVecs, unsigned Opc, |
1471 | bool isExt) { |
1472 | SDLoc dl(N); |
1473 | EVT VT = N->getValueType(ResNo: 0); |
1474 | |
1475 | unsigned ExtOff = isExt; |
1476 | |
1477 | // Form a REG_SEQUENCE to force register allocation. |
1478 | unsigned Vec0Off = ExtOff + 1; |
1479 | SmallVector<SDValue, 4> Regs(N->ops().slice(N: Vec0Off, M: NumVecs)); |
1480 | SDValue RegSeq = createQTuple(Regs); |
1481 | |
1482 | SmallVector<SDValue, 6> Ops; |
1483 | if (isExt) |
1484 | Ops.push_back(Elt: N->getOperand(Num: 1)); |
1485 | Ops.push_back(Elt: RegSeq); |
1486 | Ops.push_back(Elt: N->getOperand(Num: NumVecs + ExtOff + 1)); |
1487 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: Opc, dl, VT, Ops)); |
1488 | } |
1489 | |
1490 | static std::tuple<SDValue, SDValue> |
1491 | (SDValue Disc, SelectionDAG *DAG) { |
1492 | SDLoc DL(Disc); |
1493 | SDValue AddrDisc; |
1494 | SDValue ConstDisc; |
1495 | |
1496 | // If this is a blend, remember the constant and address discriminators. |
1497 | // Otherwise, it's either a constant discriminator, or a non-blended |
1498 | // address discriminator. |
1499 | if (Disc->getOpcode() == ISD::INTRINSIC_WO_CHAIN && |
1500 | Disc->getConstantOperandVal(Num: 0) == Intrinsic::ptrauth_blend) { |
1501 | AddrDisc = Disc->getOperand(Num: 1); |
1502 | ConstDisc = Disc->getOperand(Num: 2); |
1503 | } else { |
1504 | ConstDisc = Disc; |
1505 | } |
1506 | |
1507 | // If the constant discriminator (either the blend RHS, or the entire |
1508 | // discriminator value) isn't a 16-bit constant, bail out, and let the |
1509 | // discriminator be computed separately. |
1510 | auto *ConstDiscN = dyn_cast<ConstantSDNode>(Val&: ConstDisc); |
1511 | if (!ConstDiscN || !isUInt<16>(x: ConstDiscN->getZExtValue())) |
1512 | return std::make_tuple(args: DAG->getTargetConstant(Val: 0, DL, VT: MVT::i64), args&: Disc); |
1513 | |
1514 | // If there's no address discriminator, use XZR directly. |
1515 | if (!AddrDisc) |
1516 | AddrDisc = DAG->getRegister(Reg: AArch64::XZR, VT: MVT::i64); |
1517 | |
1518 | return std::make_tuple( |
1519 | args: DAG->getTargetConstant(Val: ConstDiscN->getZExtValue(), DL, VT: MVT::i64), |
1520 | args&: AddrDisc); |
1521 | } |
1522 | |
1523 | void AArch64DAGToDAGISel::SelectPtrauthAuth(SDNode *N) { |
1524 | SDLoc DL(N); |
1525 | // IntrinsicID is operand #0 |
1526 | SDValue Val = N->getOperand(Num: 1); |
1527 | SDValue AUTKey = N->getOperand(Num: 2); |
1528 | SDValue AUTDisc = N->getOperand(Num: 3); |
1529 | |
1530 | unsigned AUTKeyC = cast<ConstantSDNode>(Val&: AUTKey)->getZExtValue(); |
1531 | AUTKey = CurDAG->getTargetConstant(Val: AUTKeyC, DL, VT: MVT::i64); |
1532 | |
1533 | SDValue AUTAddrDisc, AUTConstDisc; |
1534 | std::tie(args&: AUTConstDisc, args&: AUTAddrDisc) = |
1535 | extractPtrauthBlendDiscriminators(Disc: AUTDisc, DAG: CurDAG); |
1536 | |
1537 | SDValue X16Copy = CurDAG->getCopyToReg(Chain: CurDAG->getEntryNode(), dl: DL, |
1538 | Reg: AArch64::X16, N: Val, Glue: SDValue()); |
1539 | SDValue Ops[] = {AUTKey, AUTConstDisc, AUTAddrDisc, X16Copy.getValue(R: 1)}; |
1540 | |
1541 | SDNode *AUT = CurDAG->getMachineNode(Opcode: AArch64::AUT, dl: DL, VT: MVT::i64, Ops); |
1542 | ReplaceNode(F: N, T: AUT); |
1543 | } |
1544 | |
1545 | void AArch64DAGToDAGISel::SelectPtrauthResign(SDNode *N) { |
1546 | SDLoc DL(N); |
1547 | // IntrinsicID is operand #0 |
1548 | SDValue Val = N->getOperand(Num: 1); |
1549 | SDValue AUTKey = N->getOperand(Num: 2); |
1550 | SDValue AUTDisc = N->getOperand(Num: 3); |
1551 | SDValue PACKey = N->getOperand(Num: 4); |
1552 | SDValue PACDisc = N->getOperand(Num: 5); |
1553 | |
1554 | unsigned AUTKeyC = cast<ConstantSDNode>(Val&: AUTKey)->getZExtValue(); |
1555 | unsigned PACKeyC = cast<ConstantSDNode>(Val&: PACKey)->getZExtValue(); |
1556 | |
1557 | AUTKey = CurDAG->getTargetConstant(Val: AUTKeyC, DL, VT: MVT::i64); |
1558 | PACKey = CurDAG->getTargetConstant(Val: PACKeyC, DL, VT: MVT::i64); |
1559 | |
1560 | SDValue AUTAddrDisc, AUTConstDisc; |
1561 | std::tie(args&: AUTConstDisc, args&: AUTAddrDisc) = |
1562 | extractPtrauthBlendDiscriminators(Disc: AUTDisc, DAG: CurDAG); |
1563 | |
1564 | SDValue PACAddrDisc, PACConstDisc; |
1565 | std::tie(args&: PACConstDisc, args&: PACAddrDisc) = |
1566 | extractPtrauthBlendDiscriminators(Disc: PACDisc, DAG: CurDAG); |
1567 | |
1568 | SDValue X16Copy = CurDAG->getCopyToReg(Chain: CurDAG->getEntryNode(), dl: DL, |
1569 | Reg: AArch64::X16, N: Val, Glue: SDValue()); |
1570 | |
1571 | SDValue Ops[] = {AUTKey, AUTConstDisc, AUTAddrDisc, PACKey, |
1572 | PACConstDisc, PACAddrDisc, X16Copy.getValue(R: 1)}; |
1573 | |
1574 | SDNode *AUTPAC = CurDAG->getMachineNode(Opcode: AArch64::AUTPAC, dl: DL, VT: MVT::i64, Ops); |
1575 | ReplaceNode(F: N, T: AUTPAC); |
1576 | } |
1577 | |
1578 | bool AArch64DAGToDAGISel::tryIndexedLoad(SDNode *N) { |
1579 | LoadSDNode *LD = cast<LoadSDNode>(Val: N); |
1580 | if (LD->isUnindexed()) |
1581 | return false; |
1582 | EVT VT = LD->getMemoryVT(); |
1583 | EVT DstVT = N->getValueType(ResNo: 0); |
1584 | ISD::MemIndexedMode AM = LD->getAddressingMode(); |
1585 | bool IsPre = AM == ISD::PRE_INC || AM == ISD::PRE_DEC; |
1586 | ConstantSDNode *OffsetOp = cast<ConstantSDNode>(Val: LD->getOffset()); |
1587 | int OffsetVal = (int)OffsetOp->getZExtValue(); |
1588 | |
1589 | // We're not doing validity checking here. That was done when checking |
1590 | // if we should mark the load as indexed or not. We're just selecting |
1591 | // the right instruction. |
1592 | unsigned Opcode = 0; |
1593 | |
1594 | ISD::LoadExtType ExtType = LD->getExtensionType(); |
1595 | bool InsertTo64 = false; |
1596 | if (VT == MVT::i64) |
1597 | Opcode = IsPre ? AArch64::LDRXpre : AArch64::LDRXpost; |
1598 | else if (VT == MVT::i32) { |
1599 | if (ExtType == ISD::NON_EXTLOAD) |
1600 | Opcode = IsPre ? AArch64::LDRWpre : AArch64::LDRWpost; |
1601 | else if (ExtType == ISD::SEXTLOAD) |
1602 | Opcode = IsPre ? AArch64::LDRSWpre : AArch64::LDRSWpost; |
1603 | else { |
1604 | Opcode = IsPre ? AArch64::LDRWpre : AArch64::LDRWpost; |
1605 | InsertTo64 = true; |
1606 | // The result of the load is only i32. It's the subreg_to_reg that makes |
1607 | // it into an i64. |
1608 | DstVT = MVT::i32; |
1609 | } |
1610 | } else if (VT == MVT::i16) { |
1611 | if (ExtType == ISD::SEXTLOAD) { |
1612 | if (DstVT == MVT::i64) |
1613 | Opcode = IsPre ? AArch64::LDRSHXpre : AArch64::LDRSHXpost; |
1614 | else |
1615 | Opcode = IsPre ? AArch64::LDRSHWpre : AArch64::LDRSHWpost; |
1616 | } else { |
1617 | Opcode = IsPre ? AArch64::LDRHHpre : AArch64::LDRHHpost; |
1618 | InsertTo64 = DstVT == MVT::i64; |
1619 | // The result of the load is only i32. It's the subreg_to_reg that makes |
1620 | // it into an i64. |
1621 | DstVT = MVT::i32; |
1622 | } |
1623 | } else if (VT == MVT::i8) { |
1624 | if (ExtType == ISD::SEXTLOAD) { |
1625 | if (DstVT == MVT::i64) |
1626 | Opcode = IsPre ? AArch64::LDRSBXpre : AArch64::LDRSBXpost; |
1627 | else |
1628 | Opcode = IsPre ? AArch64::LDRSBWpre : AArch64::LDRSBWpost; |
1629 | } else { |
1630 | Opcode = IsPre ? AArch64::LDRBBpre : AArch64::LDRBBpost; |
1631 | InsertTo64 = DstVT == MVT::i64; |
1632 | // The result of the load is only i32. It's the subreg_to_reg that makes |
1633 | // it into an i64. |
1634 | DstVT = MVT::i32; |
1635 | } |
1636 | } else if (VT == MVT::f16) { |
1637 | Opcode = IsPre ? AArch64::LDRHpre : AArch64::LDRHpost; |
1638 | } else if (VT == MVT::bf16) { |
1639 | Opcode = IsPre ? AArch64::LDRHpre : AArch64::LDRHpost; |
1640 | } else if (VT == MVT::f32) { |
1641 | Opcode = IsPre ? AArch64::LDRSpre : AArch64::LDRSpost; |
1642 | } else if (VT == MVT::f64 || |
1643 | (VT.is64BitVector() && Subtarget->isLittleEndian())) { |
1644 | Opcode = IsPre ? AArch64::LDRDpre : AArch64::LDRDpost; |
1645 | } else if (VT.is128BitVector() && Subtarget->isLittleEndian()) { |
1646 | Opcode = IsPre ? AArch64::LDRQpre : AArch64::LDRQpost; |
1647 | } else if (VT.is64BitVector()) { |
1648 | if (IsPre || OffsetVal != 8) |
1649 | return false; |
1650 | switch (VT.getScalarSizeInBits()) { |
1651 | case 8: |
1652 | Opcode = AArch64::LD1Onev8b_POST; |
1653 | break; |
1654 | case 16: |
1655 | Opcode = AArch64::LD1Onev4h_POST; |
1656 | break; |
1657 | case 32: |
1658 | Opcode = AArch64::LD1Onev2s_POST; |
1659 | break; |
1660 | case 64: |
1661 | Opcode = AArch64::LD1Onev1d_POST; |
1662 | break; |
1663 | default: |
1664 | llvm_unreachable("Expected vector element to be a power of 2" ); |
1665 | } |
1666 | } else if (VT.is128BitVector()) { |
1667 | if (IsPre || OffsetVal != 16) |
1668 | return false; |
1669 | switch (VT.getScalarSizeInBits()) { |
1670 | case 8: |
1671 | Opcode = AArch64::LD1Onev16b_POST; |
1672 | break; |
1673 | case 16: |
1674 | Opcode = AArch64::LD1Onev8h_POST; |
1675 | break; |
1676 | case 32: |
1677 | Opcode = AArch64::LD1Onev4s_POST; |
1678 | break; |
1679 | case 64: |
1680 | Opcode = AArch64::LD1Onev2d_POST; |
1681 | break; |
1682 | default: |
1683 | llvm_unreachable("Expected vector element to be a power of 2" ); |
1684 | } |
1685 | } else |
1686 | return false; |
1687 | SDValue Chain = LD->getChain(); |
1688 | SDValue Base = LD->getBasePtr(); |
1689 | SDLoc dl(N); |
1690 | // LD1 encodes an immediate offset by using XZR as the offset register. |
1691 | SDValue Offset = (VT.isVector() && !Subtarget->isLittleEndian()) |
1692 | ? CurDAG->getRegister(Reg: AArch64::XZR, VT: MVT::i64) |
1693 | : CurDAG->getTargetConstant(Val: OffsetVal, DL: dl, VT: MVT::i64); |
1694 | SDValue Ops[] = { Base, Offset, Chain }; |
1695 | SDNode *Res = CurDAG->getMachineNode(Opcode, dl, VT1: MVT::i64, VT2: DstVT, |
1696 | VT3: MVT::Other, Ops); |
1697 | |
1698 | // Transfer memoperands. |
1699 | MachineMemOperand *MemOp = cast<MemSDNode>(Val: N)->getMemOperand(); |
1700 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: Res), NewMemRefs: {MemOp}); |
1701 | |
1702 | // Either way, we're replacing the node, so tell the caller that. |
1703 | SDValue LoadedVal = SDValue(Res, 1); |
1704 | if (InsertTo64) { |
1705 | SDValue SubReg = CurDAG->getTargetConstant(Val: AArch64::sub_32, DL: dl, VT: MVT::i32); |
1706 | LoadedVal = |
1707 | SDValue(CurDAG->getMachineNode( |
1708 | Opcode: AArch64::SUBREG_TO_REG, dl, VT: MVT::i64, |
1709 | Op1: CurDAG->getTargetConstant(Val: 0, DL: dl, VT: MVT::i64), Op2: LoadedVal, |
1710 | Op3: SubReg), |
1711 | 0); |
1712 | } |
1713 | |
1714 | ReplaceUses(F: SDValue(N, 0), T: LoadedVal); |
1715 | ReplaceUses(F: SDValue(N, 1), T: SDValue(Res, 0)); |
1716 | ReplaceUses(F: SDValue(N, 2), T: SDValue(Res, 2)); |
1717 | CurDAG->RemoveDeadNode(N); |
1718 | return true; |
1719 | } |
1720 | |
1721 | void AArch64DAGToDAGISel::SelectLoad(SDNode *N, unsigned NumVecs, unsigned Opc, |
1722 | unsigned SubRegIdx) { |
1723 | SDLoc dl(N); |
1724 | EVT VT = N->getValueType(ResNo: 0); |
1725 | SDValue Chain = N->getOperand(Num: 0); |
1726 | |
1727 | SDValue Ops[] = {N->getOperand(Num: 2), // Mem operand; |
1728 | Chain}; |
1729 | |
1730 | const EVT ResTys[] = {MVT::Untyped, MVT::Other}; |
1731 | |
1732 | SDNode *Ld = CurDAG->getMachineNode(Opcode: Opc, dl, ResultTys: ResTys, Ops); |
1733 | SDValue SuperReg = SDValue(Ld, 0); |
1734 | for (unsigned i = 0; i < NumVecs; ++i) |
1735 | ReplaceUses(F: SDValue(N, i), |
1736 | T: CurDAG->getTargetExtractSubreg(SRIdx: SubRegIdx + i, DL: dl, VT, Operand: SuperReg)); |
1737 | |
1738 | ReplaceUses(F: SDValue(N, NumVecs), T: SDValue(Ld, 1)); |
1739 | |
1740 | // Transfer memoperands. In the case of AArch64::LD64B, there won't be one, |
1741 | // because it's too simple to have needed special treatment during lowering. |
1742 | if (auto *MemIntr = dyn_cast<MemIntrinsicSDNode>(Val: N)) { |
1743 | MachineMemOperand *MemOp = MemIntr->getMemOperand(); |
1744 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: Ld), NewMemRefs: {MemOp}); |
1745 | } |
1746 | |
1747 | CurDAG->RemoveDeadNode(N); |
1748 | } |
1749 | |
1750 | void AArch64DAGToDAGISel::SelectPostLoad(SDNode *N, unsigned NumVecs, |
1751 | unsigned Opc, unsigned SubRegIdx) { |
1752 | SDLoc dl(N); |
1753 | EVT VT = N->getValueType(ResNo: 0); |
1754 | SDValue Chain = N->getOperand(Num: 0); |
1755 | |
1756 | SDValue Ops[] = {N->getOperand(Num: 1), // Mem operand |
1757 | N->getOperand(Num: 2), // Incremental |
1758 | Chain}; |
1759 | |
1760 | const EVT ResTys[] = {MVT::i64, // Type of the write back register |
1761 | MVT::Untyped, MVT::Other}; |
1762 | |
1763 | SDNode *Ld = CurDAG->getMachineNode(Opcode: Opc, dl, ResultTys: ResTys, Ops); |
1764 | |
1765 | // Update uses of write back register |
1766 | ReplaceUses(F: SDValue(N, NumVecs), T: SDValue(Ld, 0)); |
1767 | |
1768 | // Update uses of vector list |
1769 | SDValue SuperReg = SDValue(Ld, 1); |
1770 | if (NumVecs == 1) |
1771 | ReplaceUses(F: SDValue(N, 0), T: SuperReg); |
1772 | else |
1773 | for (unsigned i = 0; i < NumVecs; ++i) |
1774 | ReplaceUses(F: SDValue(N, i), |
1775 | T: CurDAG->getTargetExtractSubreg(SRIdx: SubRegIdx + i, DL: dl, VT, Operand: SuperReg)); |
1776 | |
1777 | // Update the chain |
1778 | ReplaceUses(F: SDValue(N, NumVecs + 1), T: SDValue(Ld, 2)); |
1779 | CurDAG->RemoveDeadNode(N); |
1780 | } |
1781 | |
1782 | /// Optimize \param OldBase and \param OldOffset selecting the best addressing |
1783 | /// mode. Returns a tuple consisting of an Opcode, an SDValue representing the |
1784 | /// new Base and an SDValue representing the new offset. |
1785 | std::tuple<unsigned, SDValue, SDValue> |
1786 | AArch64DAGToDAGISel::findAddrModeSVELoadStore(SDNode *N, unsigned Opc_rr, |
1787 | unsigned Opc_ri, |
1788 | const SDValue &OldBase, |
1789 | const SDValue &OldOffset, |
1790 | unsigned Scale) { |
1791 | SDValue NewBase = OldBase; |
1792 | SDValue NewOffset = OldOffset; |
1793 | // Detect a possible Reg+Imm addressing mode. |
1794 | const bool IsRegImm = SelectAddrModeIndexedSVE</*Min=*/-8, /*Max=*/7>( |
1795 | Root: N, N: OldBase, Base&: NewBase, OffImm&: NewOffset); |
1796 | |
1797 | // Detect a possible reg+reg addressing mode, but only if we haven't already |
1798 | // detected a Reg+Imm one. |
1799 | const bool IsRegReg = |
1800 | !IsRegImm && SelectSVERegRegAddrMode(N: OldBase, Scale, Base&: NewBase, Offset&: NewOffset); |
1801 | |
1802 | // Select the instruction. |
1803 | return std::make_tuple(args&: IsRegReg ? Opc_rr : Opc_ri, args&: NewBase, args&: NewOffset); |
1804 | } |
1805 | |
1806 | enum class SelectTypeKind { |
1807 | Int1 = 0, |
1808 | Int = 1, |
1809 | FP = 2, |
1810 | AnyType = 3, |
1811 | }; |
1812 | |
1813 | /// This function selects an opcode from a list of opcodes, which is |
1814 | /// expected to be the opcode for { 8-bit, 16-bit, 32-bit, 64-bit } |
1815 | /// element types, in this order. |
1816 | template <SelectTypeKind Kind> |
1817 | static unsigned SelectOpcodeFromVT(EVT VT, ArrayRef<unsigned> Opcodes) { |
1818 | // Only match scalable vector VTs |
1819 | if (!VT.isScalableVector()) |
1820 | return 0; |
1821 | |
1822 | EVT EltVT = VT.getVectorElementType(); |
1823 | unsigned Key = VT.getVectorMinNumElements(); |
1824 | switch (Kind) { |
1825 | case SelectTypeKind::AnyType: |
1826 | break; |
1827 | case SelectTypeKind::Int: |
1828 | if (EltVT != MVT::i8 && EltVT != MVT::i16 && EltVT != MVT::i32 && |
1829 | EltVT != MVT::i64) |
1830 | return 0; |
1831 | break; |
1832 | case SelectTypeKind::Int1: |
1833 | if (EltVT != MVT::i1) |
1834 | return 0; |
1835 | break; |
1836 | case SelectTypeKind::FP: |
1837 | if (EltVT == MVT::bf16) |
1838 | Key = 16; |
1839 | else if (EltVT != MVT::bf16 && EltVT != MVT::f16 && EltVT != MVT::f32 && |
1840 | EltVT != MVT::f64) |
1841 | return 0; |
1842 | break; |
1843 | } |
1844 | |
1845 | unsigned Offset; |
1846 | switch (Key) { |
1847 | case 16: // 8-bit or bf16 |
1848 | Offset = 0; |
1849 | break; |
1850 | case 8: // 16-bit |
1851 | Offset = 1; |
1852 | break; |
1853 | case 4: // 32-bit |
1854 | Offset = 2; |
1855 | break; |
1856 | case 2: // 64-bit |
1857 | Offset = 3; |
1858 | break; |
1859 | default: |
1860 | return 0; |
1861 | } |
1862 | |
1863 | return (Opcodes.size() <= Offset) ? 0 : Opcodes[Offset]; |
1864 | } |
1865 | |
1866 | // This function is almost identical to SelectWhilePair, but has an |
1867 | // extra check on the range of the immediate operand. |
1868 | // TODO: Merge these two functions together at some point? |
1869 | void AArch64DAGToDAGISel::SelectPExtPair(SDNode *N, unsigned Opc) { |
1870 | // Immediate can be either 0 or 1. |
1871 | if (ConstantSDNode *Imm = dyn_cast<ConstantSDNode>(Val: N->getOperand(Num: 2))) |
1872 | if (Imm->getZExtValue() > 1) |
1873 | return; |
1874 | |
1875 | SDLoc DL(N); |
1876 | EVT VT = N->getValueType(ResNo: 0); |
1877 | SDValue Ops[] = {N->getOperand(Num: 1), N->getOperand(Num: 2)}; |
1878 | SDNode *WhilePair = CurDAG->getMachineNode(Opcode: Opc, dl: DL, VT: MVT::Untyped, Ops); |
1879 | SDValue SuperReg = SDValue(WhilePair, 0); |
1880 | |
1881 | for (unsigned I = 0; I < 2; ++I) |
1882 | ReplaceUses(F: SDValue(N, I), T: CurDAG->getTargetExtractSubreg( |
1883 | SRIdx: AArch64::psub0 + I, DL, VT, Operand: SuperReg)); |
1884 | |
1885 | CurDAG->RemoveDeadNode(N); |
1886 | } |
1887 | |
1888 | void AArch64DAGToDAGISel::SelectWhilePair(SDNode *N, unsigned Opc) { |
1889 | SDLoc DL(N); |
1890 | EVT VT = N->getValueType(ResNo: 0); |
1891 | |
1892 | SDValue Ops[] = {N->getOperand(Num: 1), N->getOperand(Num: 2)}; |
1893 | |
1894 | SDNode *WhilePair = CurDAG->getMachineNode(Opcode: Opc, dl: DL, VT: MVT::Untyped, Ops); |
1895 | SDValue SuperReg = SDValue(WhilePair, 0); |
1896 | |
1897 | for (unsigned I = 0; I < 2; ++I) |
1898 | ReplaceUses(F: SDValue(N, I), T: CurDAG->getTargetExtractSubreg( |
1899 | SRIdx: AArch64::psub0 + I, DL, VT, Operand: SuperReg)); |
1900 | |
1901 | CurDAG->RemoveDeadNode(N); |
1902 | } |
1903 | |
1904 | void AArch64DAGToDAGISel::SelectCVTIntrinsic(SDNode *N, unsigned NumVecs, |
1905 | unsigned Opcode) { |
1906 | EVT VT = N->getValueType(ResNo: 0); |
1907 | SmallVector<SDValue, 4> Regs(N->ops().slice(N: 1, M: NumVecs)); |
1908 | SDValue Ops = createZTuple(Regs); |
1909 | SDLoc DL(N); |
1910 | SDNode *Intrinsic = CurDAG->getMachineNode(Opcode, dl: DL, VT: MVT::Untyped, Op1: Ops); |
1911 | SDValue SuperReg = SDValue(Intrinsic, 0); |
1912 | for (unsigned i = 0; i < NumVecs; ++i) |
1913 | ReplaceUses(F: SDValue(N, i), T: CurDAG->getTargetExtractSubreg( |
1914 | SRIdx: AArch64::zsub0 + i, DL, VT, Operand: SuperReg)); |
1915 | |
1916 | CurDAG->RemoveDeadNode(N); |
1917 | } |
1918 | |
1919 | void AArch64DAGToDAGISel::SelectCVTIntrinsicFP8(SDNode *N, unsigned NumVecs, |
1920 | unsigned Opcode) { |
1921 | SDLoc DL(N); |
1922 | EVT VT = N->getValueType(ResNo: 0); |
1923 | SmallVector<SDValue, 4> Ops(N->op_begin() + 2, N->op_end()); |
1924 | Ops.push_back(/*Chain*/ Elt: N->getOperand(Num: 0)); |
1925 | |
1926 | SDNode *Instruction = |
1927 | CurDAG->getMachineNode(Opcode, dl: DL, ResultTys: {MVT::Untyped, MVT::Other}, Ops); |
1928 | SDValue SuperReg = SDValue(Instruction, 0); |
1929 | |
1930 | for (unsigned i = 0; i < NumVecs; ++i) |
1931 | ReplaceUses(F: SDValue(N, i), T: CurDAG->getTargetExtractSubreg( |
1932 | SRIdx: AArch64::zsub0 + i, DL, VT, Operand: SuperReg)); |
1933 | |
1934 | // Copy chain |
1935 | unsigned ChainIdx = NumVecs; |
1936 | ReplaceUses(F: SDValue(N, ChainIdx), T: SDValue(Instruction, 1)); |
1937 | CurDAG->RemoveDeadNode(N); |
1938 | } |
1939 | |
1940 | void AArch64DAGToDAGISel::SelectDestructiveMultiIntrinsic(SDNode *N, |
1941 | unsigned NumVecs, |
1942 | bool IsZmMulti, |
1943 | unsigned Opcode, |
1944 | bool HasPred) { |
1945 | assert(Opcode != 0 && "Unexpected opcode" ); |
1946 | |
1947 | SDLoc DL(N); |
1948 | EVT VT = N->getValueType(ResNo: 0); |
1949 | unsigned FirstVecIdx = HasPred ? 2 : 1; |
1950 | |
1951 | auto GetMultiVecOperand = [=](unsigned StartIdx) { |
1952 | SmallVector<SDValue, 4> Regs(N->ops().slice(N: StartIdx, M: NumVecs)); |
1953 | return createZMulTuple(Regs); |
1954 | }; |
1955 | |
1956 | SDValue Zdn = GetMultiVecOperand(FirstVecIdx); |
1957 | |
1958 | SDValue Zm; |
1959 | if (IsZmMulti) |
1960 | Zm = GetMultiVecOperand(NumVecs + FirstVecIdx); |
1961 | else |
1962 | Zm = N->getOperand(Num: NumVecs + FirstVecIdx); |
1963 | |
1964 | SDNode *Intrinsic; |
1965 | if (HasPred) |
1966 | Intrinsic = CurDAG->getMachineNode(Opcode, dl: DL, VT: MVT::Untyped, |
1967 | Op1: N->getOperand(Num: 1), Op2: Zdn, Op3: Zm); |
1968 | else |
1969 | Intrinsic = CurDAG->getMachineNode(Opcode, dl: DL, VT: MVT::Untyped, Op1: Zdn, Op2: Zm); |
1970 | SDValue SuperReg = SDValue(Intrinsic, 0); |
1971 | for (unsigned i = 0; i < NumVecs; ++i) |
1972 | ReplaceUses(F: SDValue(N, i), T: CurDAG->getTargetExtractSubreg( |
1973 | SRIdx: AArch64::zsub0 + i, DL, VT, Operand: SuperReg)); |
1974 | |
1975 | CurDAG->RemoveDeadNode(N); |
1976 | } |
1977 | |
1978 | void AArch64DAGToDAGISel::SelectPredicatedLoad(SDNode *N, unsigned NumVecs, |
1979 | unsigned Scale, unsigned Opc_ri, |
1980 | unsigned Opc_rr, bool IsIntr) { |
1981 | assert(Scale < 5 && "Invalid scaling value." ); |
1982 | SDLoc DL(N); |
1983 | EVT VT = N->getValueType(ResNo: 0); |
1984 | SDValue Chain = N->getOperand(Num: 0); |
1985 | |
1986 | // Optimize addressing mode. |
1987 | SDValue Base, Offset; |
1988 | unsigned Opc; |
1989 | std::tie(args&: Opc, args&: Base, args&: Offset) = findAddrModeSVELoadStore( |
1990 | N, Opc_rr, Opc_ri, OldBase: N->getOperand(Num: IsIntr ? 3 : 2), |
1991 | OldOffset: CurDAG->getTargetConstant(Val: 0, DL, VT: MVT::i64), Scale); |
1992 | |
1993 | SDValue Ops[] = {N->getOperand(Num: IsIntr ? 2 : 1), // Predicate |
1994 | Base, // Memory operand |
1995 | Offset, Chain}; |
1996 | |
1997 | const EVT ResTys[] = {MVT::Untyped, MVT::Other}; |
1998 | |
1999 | SDNode *Load = CurDAG->getMachineNode(Opcode: Opc, dl: DL, ResultTys: ResTys, Ops); |
2000 | SDValue SuperReg = SDValue(Load, 0); |
2001 | for (unsigned i = 0; i < NumVecs; ++i) |
2002 | ReplaceUses(F: SDValue(N, i), T: CurDAG->getTargetExtractSubreg( |
2003 | SRIdx: AArch64::zsub0 + i, DL, VT, Operand: SuperReg)); |
2004 | |
2005 | // Copy chain |
2006 | unsigned ChainIdx = NumVecs; |
2007 | ReplaceUses(F: SDValue(N, ChainIdx), T: SDValue(Load, 1)); |
2008 | CurDAG->RemoveDeadNode(N); |
2009 | } |
2010 | |
2011 | void AArch64DAGToDAGISel::SelectContiguousMultiVectorLoad(SDNode *N, |
2012 | unsigned NumVecs, |
2013 | unsigned Scale, |
2014 | unsigned Opc_ri, |
2015 | unsigned Opc_rr) { |
2016 | assert(Scale < 4 && "Invalid scaling value." ); |
2017 | SDLoc DL(N); |
2018 | EVT VT = N->getValueType(ResNo: 0); |
2019 | SDValue Chain = N->getOperand(Num: 0); |
2020 | |
2021 | SDValue PNg = N->getOperand(Num: 2); |
2022 | SDValue Base = N->getOperand(Num: 3); |
2023 | SDValue Offset = CurDAG->getTargetConstant(Val: 0, DL, VT: MVT::i64); |
2024 | unsigned Opc; |
2025 | std::tie(args&: Opc, args&: Base, args&: Offset) = |
2026 | findAddrModeSVELoadStore(N, Opc_rr, Opc_ri, OldBase: Base, OldOffset: Offset, Scale); |
2027 | |
2028 | SDValue Ops[] = {PNg, // Predicate-as-counter |
2029 | Base, // Memory operand |
2030 | Offset, Chain}; |
2031 | |
2032 | const EVT ResTys[] = {MVT::Untyped, MVT::Other}; |
2033 | |
2034 | SDNode *Load = CurDAG->getMachineNode(Opcode: Opc, dl: DL, ResultTys: ResTys, Ops); |
2035 | SDValue SuperReg = SDValue(Load, 0); |
2036 | for (unsigned i = 0; i < NumVecs; ++i) |
2037 | ReplaceUses(F: SDValue(N, i), T: CurDAG->getTargetExtractSubreg( |
2038 | SRIdx: AArch64::zsub0 + i, DL, VT, Operand: SuperReg)); |
2039 | |
2040 | // Copy chain |
2041 | unsigned ChainIdx = NumVecs; |
2042 | ReplaceUses(F: SDValue(N, ChainIdx), T: SDValue(Load, 1)); |
2043 | CurDAG->RemoveDeadNode(N); |
2044 | } |
2045 | |
2046 | void AArch64DAGToDAGISel::SelectFrintFromVT(SDNode *N, unsigned NumVecs, |
2047 | unsigned Opcode) { |
2048 | if (N->getValueType(ResNo: 0) != MVT::nxv4f32) |
2049 | return; |
2050 | SelectUnaryMultiIntrinsic(N, NumOutVecs: NumVecs, IsTupleInput: true, Opc: Opcode); |
2051 | } |
2052 | |
2053 | void AArch64DAGToDAGISel::SelectMultiVectorLutiLane(SDNode *Node, |
2054 | unsigned NumOutVecs, |
2055 | unsigned Opc, |
2056 | uint32_t MaxImm) { |
2057 | if (ConstantSDNode *Imm = dyn_cast<ConstantSDNode>(Val: Node->getOperand(Num: 4))) |
2058 | if (Imm->getZExtValue() > MaxImm) |
2059 | return; |
2060 | |
2061 | SDValue ZtValue; |
2062 | if (!ImmToReg<AArch64::ZT0, 0>(N: Node->getOperand(Num: 2), Imm&: ZtValue)) |
2063 | return; |
2064 | |
2065 | SDValue Ops[] = {ZtValue, Node->getOperand(Num: 3), Node->getOperand(Num: 4)}; |
2066 | SDLoc DL(Node); |
2067 | EVT VT = Node->getValueType(ResNo: 0); |
2068 | |
2069 | SDNode *Instruction = |
2070 | CurDAG->getMachineNode(Opcode: Opc, dl: DL, ResultTys: {MVT::Untyped, MVT::Other}, Ops); |
2071 | SDValue SuperReg = SDValue(Instruction, 0); |
2072 | |
2073 | for (unsigned I = 0; I < NumOutVecs; ++I) |
2074 | ReplaceUses(F: SDValue(Node, I), T: CurDAG->getTargetExtractSubreg( |
2075 | SRIdx: AArch64::zsub0 + I, DL, VT, Operand: SuperReg)); |
2076 | |
2077 | // Copy chain |
2078 | unsigned ChainIdx = NumOutVecs; |
2079 | ReplaceUses(F: SDValue(Node, ChainIdx), T: SDValue(Instruction, 1)); |
2080 | CurDAG->RemoveDeadNode(N: Node); |
2081 | } |
2082 | |
2083 | void AArch64DAGToDAGISel::SelectMultiVectorLuti(SDNode *Node, |
2084 | unsigned NumOutVecs, |
2085 | unsigned Opc) { |
2086 | |
2087 | SDValue ZtValue; |
2088 | SmallVector<SDValue, 4> Ops; |
2089 | if (!ImmToReg<AArch64::ZT0, 0>(N: Node->getOperand(Num: 2), Imm&: ZtValue)) |
2090 | return; |
2091 | |
2092 | Ops.push_back(Elt: ZtValue); |
2093 | Ops.push_back(Elt: createZMulTuple(Regs: {Node->getOperand(Num: 3), Node->getOperand(Num: 4)})); |
2094 | SDLoc DL(Node); |
2095 | EVT VT = Node->getValueType(ResNo: 0); |
2096 | |
2097 | SDNode *Instruction = |
2098 | CurDAG->getMachineNode(Opcode: Opc, dl: DL, ResultTys: {MVT::Untyped, MVT::Other}, Ops); |
2099 | SDValue SuperReg = SDValue(Instruction, 0); |
2100 | |
2101 | for (unsigned I = 0; I < NumOutVecs; ++I) |
2102 | ReplaceUses(F: SDValue(Node, I), T: CurDAG->getTargetExtractSubreg( |
2103 | SRIdx: AArch64::zsub0 + I, DL, VT, Operand: SuperReg)); |
2104 | |
2105 | // Copy chain |
2106 | unsigned ChainIdx = NumOutVecs; |
2107 | ReplaceUses(F: SDValue(Node, ChainIdx), T: SDValue(Instruction, 1)); |
2108 | CurDAG->RemoveDeadNode(N: Node); |
2109 | } |
2110 | |
2111 | void AArch64DAGToDAGISel::SelectClamp(SDNode *N, unsigned NumVecs, |
2112 | unsigned Op) { |
2113 | SDLoc DL(N); |
2114 | EVT VT = N->getValueType(ResNo: 0); |
2115 | |
2116 | SmallVector<SDValue, 4> Regs(N->ops().slice(N: 1, M: NumVecs)); |
2117 | SDValue Zd = createZMulTuple(Regs); |
2118 | SDValue Zn = N->getOperand(Num: 1 + NumVecs); |
2119 | SDValue Zm = N->getOperand(Num: 2 + NumVecs); |
2120 | |
2121 | SDValue Ops[] = {Zd, Zn, Zm}; |
2122 | |
2123 | SDNode *Intrinsic = CurDAG->getMachineNode(Opcode: Op, dl: DL, VT: MVT::Untyped, Ops); |
2124 | SDValue SuperReg = SDValue(Intrinsic, 0); |
2125 | for (unsigned i = 0; i < NumVecs; ++i) |
2126 | ReplaceUses(F: SDValue(N, i), T: CurDAG->getTargetExtractSubreg( |
2127 | SRIdx: AArch64::zsub0 + i, DL, VT, Operand: SuperReg)); |
2128 | |
2129 | CurDAG->RemoveDeadNode(N); |
2130 | } |
2131 | |
2132 | bool SelectSMETile(unsigned &BaseReg, unsigned TileNum) { |
2133 | switch (BaseReg) { |
2134 | default: |
2135 | return false; |
2136 | case AArch64::ZA: |
2137 | case AArch64::ZAB0: |
2138 | if (TileNum == 0) |
2139 | break; |
2140 | return false; |
2141 | case AArch64::ZAH0: |
2142 | if (TileNum <= 1) |
2143 | break; |
2144 | return false; |
2145 | case AArch64::ZAS0: |
2146 | if (TileNum <= 3) |
2147 | break; |
2148 | return false; |
2149 | case AArch64::ZAD0: |
2150 | if (TileNum <= 7) |
2151 | break; |
2152 | return false; |
2153 | } |
2154 | |
2155 | BaseReg += TileNum; |
2156 | return true; |
2157 | } |
2158 | |
2159 | template <unsigned MaxIdx, unsigned Scale> |
2160 | void AArch64DAGToDAGISel::SelectMultiVectorMove(SDNode *N, unsigned NumVecs, |
2161 | unsigned BaseReg, unsigned Op) { |
2162 | unsigned TileNum = 0; |
2163 | if (BaseReg != AArch64::ZA) |
2164 | TileNum = N->getConstantOperandVal(Num: 2); |
2165 | |
2166 | if (!SelectSMETile(BaseReg, TileNum)) |
2167 | return; |
2168 | |
2169 | SDValue SliceBase, Base, Offset; |
2170 | if (BaseReg == AArch64::ZA) |
2171 | SliceBase = N->getOperand(Num: 2); |
2172 | else |
2173 | SliceBase = N->getOperand(Num: 3); |
2174 | |
2175 | if (!SelectSMETileSlice(N: SliceBase, MaxSize: MaxIdx, Vector&: Base, Offset, Scale)) |
2176 | return; |
2177 | |
2178 | SDLoc DL(N); |
2179 | SDValue SubReg = CurDAG->getRegister(Reg: BaseReg, VT: MVT::Other); |
2180 | SDValue Ops[] = {SubReg, Base, Offset, /*Chain*/ N->getOperand(Num: 0)}; |
2181 | SDNode *Mov = CurDAG->getMachineNode(Opcode: Op, dl: DL, ResultTys: {MVT::Untyped, MVT::Other}, Ops); |
2182 | |
2183 | EVT VT = N->getValueType(ResNo: 0); |
2184 | for (unsigned I = 0; I < NumVecs; ++I) |
2185 | ReplaceUses(F: SDValue(N, I), |
2186 | T: CurDAG->getTargetExtractSubreg(SRIdx: AArch64::zsub0 + I, DL, VT, |
2187 | Operand: SDValue(Mov, 0))); |
2188 | // Copy chain |
2189 | unsigned ChainIdx = NumVecs; |
2190 | ReplaceUses(F: SDValue(N, ChainIdx), T: SDValue(Mov, 1)); |
2191 | CurDAG->RemoveDeadNode(N); |
2192 | } |
2193 | |
2194 | void AArch64DAGToDAGISel::SelectMultiVectorMoveZ(SDNode *N, unsigned NumVecs, |
2195 | unsigned Op, unsigned MaxIdx, |
2196 | unsigned Scale, unsigned BaseReg) { |
2197 | // Slice can be in different positions |
2198 | // The array to vector: llvm.aarch64.sme.readz.<h/v>.<sz>(slice) |
2199 | // The tile to vector: llvm.aarch64.sme.readz.<h/v>.<sz>(tile, slice) |
2200 | SDValue SliceBase = N->getOperand(Num: 2); |
2201 | if (BaseReg != AArch64::ZA) |
2202 | SliceBase = N->getOperand(Num: 3); |
2203 | |
2204 | SDValue Base, Offset; |
2205 | if (!SelectSMETileSlice(N: SliceBase, MaxSize: MaxIdx, Vector&: Base, Offset, Scale)) |
2206 | return; |
2207 | // The correct Za tile number is computed in Machine Instruction |
2208 | // See EmitZAInstr |
2209 | // DAG cannot select Za tile as an output register with ZReg |
2210 | SDLoc DL(N); |
2211 | SmallVector<SDValue, 6> Ops; |
2212 | if (BaseReg != AArch64::ZA ) |
2213 | Ops.push_back(Elt: N->getOperand(Num: 2)); |
2214 | Ops.push_back(Elt: Base); |
2215 | Ops.push_back(Elt: Offset); |
2216 | Ops.push_back(Elt: N->getOperand(Num: 0)); //Chain |
2217 | SDNode *Mov = CurDAG->getMachineNode(Opcode: Op, dl: DL, ResultTys: {MVT::Untyped, MVT::Other}, Ops); |
2218 | |
2219 | EVT VT = N->getValueType(ResNo: 0); |
2220 | for (unsigned I = 0; I < NumVecs; ++I) |
2221 | ReplaceUses(F: SDValue(N, I), |
2222 | T: CurDAG->getTargetExtractSubreg(SRIdx: AArch64::zsub0 + I, DL, VT, |
2223 | Operand: SDValue(Mov, 0))); |
2224 | |
2225 | // Copy chain |
2226 | unsigned ChainIdx = NumVecs; |
2227 | ReplaceUses(F: SDValue(N, ChainIdx), T: SDValue(Mov, 1)); |
2228 | CurDAG->RemoveDeadNode(N); |
2229 | } |
2230 | |
2231 | void AArch64DAGToDAGISel::SelectUnaryMultiIntrinsic(SDNode *N, |
2232 | unsigned NumOutVecs, |
2233 | bool IsTupleInput, |
2234 | unsigned Opc) { |
2235 | SDLoc DL(N); |
2236 | EVT VT = N->getValueType(ResNo: 0); |
2237 | unsigned NumInVecs = N->getNumOperands() - 1; |
2238 | |
2239 | SmallVector<SDValue, 6> Ops; |
2240 | if (IsTupleInput) { |
2241 | assert((NumInVecs == 2 || NumInVecs == 4) && |
2242 | "Don't know how to handle multi-register input!" ); |
2243 | SmallVector<SDValue, 4> Regs(N->ops().slice(N: 1, M: NumInVecs)); |
2244 | Ops.push_back(Elt: createZMulTuple(Regs)); |
2245 | } else { |
2246 | // All intrinsic nodes have the ID as the first operand, hence the "1 + I". |
2247 | for (unsigned I = 0; I < NumInVecs; I++) |
2248 | Ops.push_back(Elt: N->getOperand(Num: 1 + I)); |
2249 | } |
2250 | |
2251 | SDNode *Res = CurDAG->getMachineNode(Opcode: Opc, dl: DL, VT: MVT::Untyped, Ops); |
2252 | SDValue SuperReg = SDValue(Res, 0); |
2253 | |
2254 | for (unsigned I = 0; I < NumOutVecs; I++) |
2255 | ReplaceUses(F: SDValue(N, I), T: CurDAG->getTargetExtractSubreg( |
2256 | SRIdx: AArch64::zsub0 + I, DL, VT, Operand: SuperReg)); |
2257 | CurDAG->RemoveDeadNode(N); |
2258 | } |
2259 | |
2260 | void AArch64DAGToDAGISel::SelectStore(SDNode *N, unsigned NumVecs, |
2261 | unsigned Opc) { |
2262 | SDLoc dl(N); |
2263 | EVT VT = N->getOperand(Num: 2)->getValueType(ResNo: 0); |
2264 | |
2265 | // Form a REG_SEQUENCE to force register allocation. |
2266 | bool Is128Bit = VT.getSizeInBits() == 128; |
2267 | SmallVector<SDValue, 4> Regs(N->ops().slice(N: 2, M: NumVecs)); |
2268 | SDValue RegSeq = Is128Bit ? createQTuple(Regs) : createDTuple(Regs); |
2269 | |
2270 | SDValue Ops[] = {RegSeq, N->getOperand(Num: NumVecs + 2), N->getOperand(Num: 0)}; |
2271 | SDNode *St = CurDAG->getMachineNode(Opcode: Opc, dl, VT: N->getValueType(ResNo: 0), Ops); |
2272 | |
2273 | // Transfer memoperands. |
2274 | MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(Val: N)->getMemOperand(); |
2275 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: St), NewMemRefs: {MemOp}); |
2276 | |
2277 | ReplaceNode(F: N, T: St); |
2278 | } |
2279 | |
2280 | void AArch64DAGToDAGISel::SelectPredicatedStore(SDNode *N, unsigned NumVecs, |
2281 | unsigned Scale, unsigned Opc_rr, |
2282 | unsigned Opc_ri) { |
2283 | SDLoc dl(N); |
2284 | |
2285 | // Form a REG_SEQUENCE to force register allocation. |
2286 | SmallVector<SDValue, 4> Regs(N->ops().slice(N: 2, M: NumVecs)); |
2287 | SDValue RegSeq = createZTuple(Regs); |
2288 | |
2289 | // Optimize addressing mode. |
2290 | unsigned Opc; |
2291 | SDValue Offset, Base; |
2292 | std::tie(args&: Opc, args&: Base, args&: Offset) = findAddrModeSVELoadStore( |
2293 | N, Opc_rr, Opc_ri, OldBase: N->getOperand(Num: NumVecs + 3), |
2294 | OldOffset: CurDAG->getTargetConstant(Val: 0, DL: dl, VT: MVT::i64), Scale); |
2295 | |
2296 | SDValue Ops[] = {RegSeq, N->getOperand(Num: NumVecs + 2), // predicate |
2297 | Base, // address |
2298 | Offset, // offset |
2299 | N->getOperand(Num: 0)}; // chain |
2300 | SDNode *St = CurDAG->getMachineNode(Opcode: Opc, dl, VT: N->getValueType(ResNo: 0), Ops); |
2301 | |
2302 | ReplaceNode(F: N, T: St); |
2303 | } |
2304 | |
2305 | bool AArch64DAGToDAGISel::SelectAddrModeFrameIndexSVE(SDValue N, SDValue &Base, |
2306 | SDValue &OffImm) { |
2307 | SDLoc dl(N); |
2308 | const DataLayout &DL = CurDAG->getDataLayout(); |
2309 | const TargetLowering *TLI = getTargetLowering(); |
2310 | |
2311 | // Try to match it for the frame address |
2312 | if (auto FINode = dyn_cast<FrameIndexSDNode>(Val&: N)) { |
2313 | int FI = FINode->getIndex(); |
2314 | Base = CurDAG->getTargetFrameIndex(FI, VT: TLI->getPointerTy(DL)); |
2315 | OffImm = CurDAG->getTargetConstant(Val: 0, DL: dl, VT: MVT::i64); |
2316 | return true; |
2317 | } |
2318 | |
2319 | return false; |
2320 | } |
2321 | |
2322 | void AArch64DAGToDAGISel::SelectPostStore(SDNode *N, unsigned NumVecs, |
2323 | unsigned Opc) { |
2324 | SDLoc dl(N); |
2325 | EVT VT = N->getOperand(Num: 2)->getValueType(ResNo: 0); |
2326 | const EVT ResTys[] = {MVT::i64, // Type of the write back register |
2327 | MVT::Other}; // Type for the Chain |
2328 | |
2329 | // Form a REG_SEQUENCE to force register allocation. |
2330 | bool Is128Bit = VT.getSizeInBits() == 128; |
2331 | SmallVector<SDValue, 4> Regs(N->ops().slice(N: 1, M: NumVecs)); |
2332 | SDValue RegSeq = Is128Bit ? createQTuple(Regs) : createDTuple(Regs); |
2333 | |
2334 | SDValue Ops[] = {RegSeq, |
2335 | N->getOperand(Num: NumVecs + 1), // base register |
2336 | N->getOperand(Num: NumVecs + 2), // Incremental |
2337 | N->getOperand(Num: 0)}; // Chain |
2338 | SDNode *St = CurDAG->getMachineNode(Opcode: Opc, dl, ResultTys: ResTys, Ops); |
2339 | |
2340 | ReplaceNode(F: N, T: St); |
2341 | } |
2342 | |
2343 | namespace { |
2344 | /// WidenVector - Given a value in the V64 register class, produce the |
2345 | /// equivalent value in the V128 register class. |
2346 | class WidenVector { |
2347 | SelectionDAG &DAG; |
2348 | |
2349 | public: |
2350 | WidenVector(SelectionDAG &DAG) : DAG(DAG) {} |
2351 | |
2352 | SDValue operator()(SDValue V64Reg) { |
2353 | EVT VT = V64Reg.getValueType(); |
2354 | unsigned NarrowSize = VT.getVectorNumElements(); |
2355 | MVT EltTy = VT.getVectorElementType().getSimpleVT(); |
2356 | MVT WideTy = MVT::getVectorVT(VT: EltTy, NumElements: 2 * NarrowSize); |
2357 | SDLoc DL(V64Reg); |
2358 | |
2359 | SDValue Undef = |
2360 | SDValue(DAG.getMachineNode(Opcode: TargetOpcode::IMPLICIT_DEF, dl: DL, VT: WideTy), 0); |
2361 | return DAG.getTargetInsertSubreg(SRIdx: AArch64::dsub, DL, VT: WideTy, Operand: Undef, Subreg: V64Reg); |
2362 | } |
2363 | }; |
2364 | } // namespace |
2365 | |
2366 | /// NarrowVector - Given a value in the V128 register class, produce the |
2367 | /// equivalent value in the V64 register class. |
2368 | static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) { |
2369 | EVT VT = V128Reg.getValueType(); |
2370 | unsigned WideSize = VT.getVectorNumElements(); |
2371 | MVT EltTy = VT.getVectorElementType().getSimpleVT(); |
2372 | MVT NarrowTy = MVT::getVectorVT(VT: EltTy, NumElements: WideSize / 2); |
2373 | |
2374 | return DAG.getTargetExtractSubreg(SRIdx: AArch64::dsub, DL: SDLoc(V128Reg), VT: NarrowTy, |
2375 | Operand: V128Reg); |
2376 | } |
2377 | |
2378 | void AArch64DAGToDAGISel::SelectLoadLane(SDNode *N, unsigned NumVecs, |
2379 | unsigned Opc) { |
2380 | SDLoc dl(N); |
2381 | EVT VT = N->getValueType(ResNo: 0); |
2382 | bool Narrow = VT.getSizeInBits() == 64; |
2383 | |
2384 | // Form a REG_SEQUENCE to force register allocation. |
2385 | SmallVector<SDValue, 4> Regs(N->ops().slice(N: 2, M: NumVecs)); |
2386 | |
2387 | if (Narrow) |
2388 | transform(Range&: Regs, d_first: Regs.begin(), |
2389 | F: WidenVector(*CurDAG)); |
2390 | |
2391 | SDValue RegSeq = createQTuple(Regs); |
2392 | |
2393 | const EVT ResTys[] = {MVT::Untyped, MVT::Other}; |
2394 | |
2395 | unsigned LaneNo = N->getConstantOperandVal(Num: NumVecs + 2); |
2396 | |
2397 | SDValue Ops[] = {RegSeq, CurDAG->getTargetConstant(Val: LaneNo, DL: dl, VT: MVT::i64), |
2398 | N->getOperand(Num: NumVecs + 3), N->getOperand(Num: 0)}; |
2399 | SDNode *Ld = CurDAG->getMachineNode(Opcode: Opc, dl, ResultTys: ResTys, Ops); |
2400 | SDValue SuperReg = SDValue(Ld, 0); |
2401 | |
2402 | EVT WideVT = RegSeq.getOperand(i: 1)->getValueType(ResNo: 0); |
2403 | static const unsigned QSubs[] = { AArch64::qsub0, AArch64::qsub1, |
2404 | AArch64::qsub2, AArch64::qsub3 }; |
2405 | for (unsigned i = 0; i < NumVecs; ++i) { |
2406 | SDValue NV = CurDAG->getTargetExtractSubreg(SRIdx: QSubs[i], DL: dl, VT: WideVT, Operand: SuperReg); |
2407 | if (Narrow) |
2408 | NV = NarrowVector(V128Reg: NV, DAG&: *CurDAG); |
2409 | ReplaceUses(F: SDValue(N, i), T: NV); |
2410 | } |
2411 | |
2412 | ReplaceUses(F: SDValue(N, NumVecs), T: SDValue(Ld, 1)); |
2413 | CurDAG->RemoveDeadNode(N); |
2414 | } |
2415 | |
2416 | void AArch64DAGToDAGISel::SelectPostLoadLane(SDNode *N, unsigned NumVecs, |
2417 | unsigned Opc) { |
2418 | SDLoc dl(N); |
2419 | EVT VT = N->getValueType(ResNo: 0); |
2420 | bool Narrow = VT.getSizeInBits() == 64; |
2421 | |
2422 | // Form a REG_SEQUENCE to force register allocation. |
2423 | SmallVector<SDValue, 4> Regs(N->ops().slice(N: 1, M: NumVecs)); |
2424 | |
2425 | if (Narrow) |
2426 | transform(Range&: Regs, d_first: Regs.begin(), |
2427 | F: WidenVector(*CurDAG)); |
2428 | |
2429 | SDValue RegSeq = createQTuple(Regs); |
2430 | |
2431 | const EVT ResTys[] = {MVT::i64, // Type of the write back register |
2432 | RegSeq->getValueType(ResNo: 0), MVT::Other}; |
2433 | |
2434 | unsigned LaneNo = N->getConstantOperandVal(Num: NumVecs + 1); |
2435 | |
2436 | SDValue Ops[] = {RegSeq, |
2437 | CurDAG->getTargetConstant(Val: LaneNo, DL: dl, |
2438 | VT: MVT::i64), // Lane Number |
2439 | N->getOperand(Num: NumVecs + 2), // Base register |
2440 | N->getOperand(Num: NumVecs + 3), // Incremental |
2441 | N->getOperand(Num: 0)}; |
2442 | SDNode *Ld = CurDAG->getMachineNode(Opcode: Opc, dl, ResultTys: ResTys, Ops); |
2443 | |
2444 | // Update uses of the write back register |
2445 | ReplaceUses(F: SDValue(N, NumVecs), T: SDValue(Ld, 0)); |
2446 | |
2447 | // Update uses of the vector list |
2448 | SDValue SuperReg = SDValue(Ld, 1); |
2449 | if (NumVecs == 1) { |
2450 | ReplaceUses(F: SDValue(N, 0), |
2451 | T: Narrow ? NarrowVector(V128Reg: SuperReg, DAG&: *CurDAG) : SuperReg); |
2452 | } else { |
2453 | EVT WideVT = RegSeq.getOperand(i: 1)->getValueType(ResNo: 0); |
2454 | static const unsigned QSubs[] = { AArch64::qsub0, AArch64::qsub1, |
2455 | AArch64::qsub2, AArch64::qsub3 }; |
2456 | for (unsigned i = 0; i < NumVecs; ++i) { |
2457 | SDValue NV = CurDAG->getTargetExtractSubreg(SRIdx: QSubs[i], DL: dl, VT: WideVT, |
2458 | Operand: SuperReg); |
2459 | if (Narrow) |
2460 | NV = NarrowVector(V128Reg: NV, DAG&: *CurDAG); |
2461 | ReplaceUses(F: SDValue(N, i), T: NV); |
2462 | } |
2463 | } |
2464 | |
2465 | // Update the Chain |
2466 | ReplaceUses(F: SDValue(N, NumVecs + 1), T: SDValue(Ld, 2)); |
2467 | CurDAG->RemoveDeadNode(N); |
2468 | } |
2469 | |
2470 | void AArch64DAGToDAGISel::SelectStoreLane(SDNode *N, unsigned NumVecs, |
2471 | unsigned Opc) { |
2472 | SDLoc dl(N); |
2473 | EVT VT = N->getOperand(Num: 2)->getValueType(ResNo: 0); |
2474 | bool Narrow = VT.getSizeInBits() == 64; |
2475 | |
2476 | // Form a REG_SEQUENCE to force register allocation. |
2477 | SmallVector<SDValue, 4> Regs(N->ops().slice(N: 2, M: NumVecs)); |
2478 | |
2479 | if (Narrow) |
2480 | transform(Range&: Regs, d_first: Regs.begin(), |
2481 | F: WidenVector(*CurDAG)); |
2482 | |
2483 | SDValue RegSeq = createQTuple(Regs); |
2484 | |
2485 | unsigned LaneNo = N->getConstantOperandVal(Num: NumVecs + 2); |
2486 | |
2487 | SDValue Ops[] = {RegSeq, CurDAG->getTargetConstant(Val: LaneNo, DL: dl, VT: MVT::i64), |
2488 | N->getOperand(Num: NumVecs + 3), N->getOperand(Num: 0)}; |
2489 | SDNode *St = CurDAG->getMachineNode(Opcode: Opc, dl, VT: MVT::Other, Ops); |
2490 | |
2491 | // Transfer memoperands. |
2492 | MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(Val: N)->getMemOperand(); |
2493 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: St), NewMemRefs: {MemOp}); |
2494 | |
2495 | ReplaceNode(F: N, T: St); |
2496 | } |
2497 | |
2498 | void AArch64DAGToDAGISel::SelectPostStoreLane(SDNode *N, unsigned NumVecs, |
2499 | unsigned Opc) { |
2500 | SDLoc dl(N); |
2501 | EVT VT = N->getOperand(Num: 2)->getValueType(ResNo: 0); |
2502 | bool Narrow = VT.getSizeInBits() == 64; |
2503 | |
2504 | // Form a REG_SEQUENCE to force register allocation. |
2505 | SmallVector<SDValue, 4> Regs(N->ops().slice(N: 1, M: NumVecs)); |
2506 | |
2507 | if (Narrow) |
2508 | transform(Range&: Regs, d_first: Regs.begin(), |
2509 | F: WidenVector(*CurDAG)); |
2510 | |
2511 | SDValue RegSeq = createQTuple(Regs); |
2512 | |
2513 | const EVT ResTys[] = {MVT::i64, // Type of the write back register |
2514 | MVT::Other}; |
2515 | |
2516 | unsigned LaneNo = N->getConstantOperandVal(Num: NumVecs + 1); |
2517 | |
2518 | SDValue Ops[] = {RegSeq, CurDAG->getTargetConstant(Val: LaneNo, DL: dl, VT: MVT::i64), |
2519 | N->getOperand(Num: NumVecs + 2), // Base Register |
2520 | N->getOperand(Num: NumVecs + 3), // Incremental |
2521 | N->getOperand(Num: 0)}; |
2522 | SDNode *St = CurDAG->getMachineNode(Opcode: Opc, dl, ResultTys: ResTys, Ops); |
2523 | |
2524 | // Transfer memoperands. |
2525 | MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(Val: N)->getMemOperand(); |
2526 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: St), NewMemRefs: {MemOp}); |
2527 | |
2528 | ReplaceNode(F: N, T: St); |
2529 | } |
2530 | |
2531 | static bool isBitfieldExtractOpFromAnd(SelectionDAG *CurDAG, SDNode *N, |
2532 | unsigned &Opc, SDValue &Opd0, |
2533 | unsigned &LSB, unsigned &MSB, |
2534 | unsigned NumberOfIgnoredLowBits, |
2535 | bool BiggerPattern) { |
2536 | assert(N->getOpcode() == ISD::AND && |
2537 | "N must be a AND operation to call this function" ); |
2538 | |
2539 | EVT VT = N->getValueType(ResNo: 0); |
2540 | |
2541 | // Here we can test the type of VT and return false when the type does not |
2542 | // match, but since it is done prior to that call in the current context |
2543 | // we turned that into an assert to avoid redundant code. |
2544 | assert((VT == MVT::i32 || VT == MVT::i64) && |
2545 | "Type checking must have been done before calling this function" ); |
2546 | |
2547 | // FIXME: simplify-demanded-bits in DAGCombine will probably have |
2548 | // changed the AND node to a 32-bit mask operation. We'll have to |
2549 | // undo that as part of the transform here if we want to catch all |
2550 | // the opportunities. |
2551 | // Currently the NumberOfIgnoredLowBits argument helps to recover |
2552 | // from these situations when matching bigger pattern (bitfield insert). |
2553 | |
2554 | // For unsigned extracts, check for a shift right and mask |
2555 | uint64_t AndImm = 0; |
2556 | if (!isOpcWithIntImmediate(N, Opc: ISD::AND, Imm&: AndImm)) |
2557 | return false; |
2558 | |
2559 | const SDNode *Op0 = N->getOperand(Num: 0).getNode(); |
2560 | |
2561 | // Because of simplify-demanded-bits in DAGCombine, the mask may have been |
2562 | // simplified. Try to undo that |
2563 | AndImm |= maskTrailingOnes<uint64_t>(N: NumberOfIgnoredLowBits); |
2564 | |
2565 | // The immediate is a mask of the low bits iff imm & (imm+1) == 0 |
2566 | if (AndImm & (AndImm + 1)) |
2567 | return false; |
2568 | |
2569 | bool ClampMSB = false; |
2570 | uint64_t SrlImm = 0; |
2571 | // Handle the SRL + ANY_EXTEND case. |
2572 | if (VT == MVT::i64 && Op0->getOpcode() == ISD::ANY_EXTEND && |
2573 | isOpcWithIntImmediate(N: Op0->getOperand(Num: 0).getNode(), Opc: ISD::SRL, Imm&: SrlImm)) { |
2574 | // Extend the incoming operand of the SRL to 64-bit. |
2575 | Opd0 = Widen(CurDAG, N: Op0->getOperand(Num: 0).getOperand(i: 0)); |
2576 | // Make sure to clamp the MSB so that we preserve the semantics of the |
2577 | // original operations. |
2578 | ClampMSB = true; |
2579 | } else if (VT == MVT::i32 && Op0->getOpcode() == ISD::TRUNCATE && |
2580 | isOpcWithIntImmediate(N: Op0->getOperand(Num: 0).getNode(), Opc: ISD::SRL, |
2581 | Imm&: SrlImm)) { |
2582 | // If the shift result was truncated, we can still combine them. |
2583 | Opd0 = Op0->getOperand(Num: 0).getOperand(i: 0); |
2584 | |
2585 | // Use the type of SRL node. |
2586 | VT = Opd0->getValueType(ResNo: 0); |
2587 | } else if (isOpcWithIntImmediate(N: Op0, Opc: ISD::SRL, Imm&: SrlImm)) { |
2588 | Opd0 = Op0->getOperand(Num: 0); |
2589 | ClampMSB = (VT == MVT::i32); |
2590 | } else if (BiggerPattern) { |
2591 | // Let's pretend a 0 shift right has been performed. |
2592 | // The resulting code will be at least as good as the original one |
2593 | // plus it may expose more opportunities for bitfield insert pattern. |
2594 | // FIXME: Currently we limit this to the bigger pattern, because |
2595 | // some optimizations expect AND and not UBFM. |
2596 | Opd0 = N->getOperand(Num: 0); |
2597 | } else |
2598 | return false; |
2599 | |
2600 | // Bail out on large immediates. This happens when no proper |
2601 | // combining/constant folding was performed. |
2602 | if (!BiggerPattern && (SrlImm <= 0 || SrlImm >= VT.getSizeInBits())) { |
2603 | LLVM_DEBUG( |
2604 | (dbgs() << N |
2605 | << ": Found large shift immediate, this should not happen\n" )); |
2606 | return false; |
2607 | } |
2608 | |
2609 | LSB = SrlImm; |
2610 | MSB = SrlImm + |
2611 | (VT == MVT::i32 ? llvm::countr_one<uint32_t>(Value: AndImm) |
2612 | : llvm::countr_one<uint64_t>(Value: AndImm)) - |
2613 | 1; |
2614 | if (ClampMSB) |
2615 | // Since we're moving the extend before the right shift operation, we need |
2616 | // to clamp the MSB to make sure we don't shift in undefined bits instead of |
2617 | // the zeros which would get shifted in with the original right shift |
2618 | // operation. |
2619 | MSB = MSB > 31 ? 31 : MSB; |
2620 | |
2621 | Opc = VT == MVT::i32 ? AArch64::UBFMWri : AArch64::UBFMXri; |
2622 | return true; |
2623 | } |
2624 | |
2625 | static bool (SDNode *N, unsigned &Opc, |
2626 | SDValue &Opd0, unsigned &Immr, |
2627 | unsigned &Imms) { |
2628 | assert(N->getOpcode() == ISD::SIGN_EXTEND_INREG); |
2629 | |
2630 | EVT VT = N->getValueType(ResNo: 0); |
2631 | unsigned BitWidth = VT.getSizeInBits(); |
2632 | assert((VT == MVT::i32 || VT == MVT::i64) && |
2633 | "Type checking must have been done before calling this function" ); |
2634 | |
2635 | SDValue Op = N->getOperand(Num: 0); |
2636 | if (Op->getOpcode() == ISD::TRUNCATE) { |
2637 | Op = Op->getOperand(Num: 0); |
2638 | VT = Op->getValueType(ResNo: 0); |
2639 | BitWidth = VT.getSizeInBits(); |
2640 | } |
2641 | |
2642 | uint64_t ShiftImm; |
2643 | if (!isOpcWithIntImmediate(N: Op.getNode(), Opc: ISD::SRL, Imm&: ShiftImm) && |
2644 | !isOpcWithIntImmediate(N: Op.getNode(), Opc: ISD::SRA, Imm&: ShiftImm)) |
2645 | return false; |
2646 | |
2647 | unsigned Width = cast<VTSDNode>(Val: N->getOperand(Num: 1))->getVT().getSizeInBits(); |
2648 | if (ShiftImm + Width > BitWidth) |
2649 | return false; |
2650 | |
2651 | Opc = (VT == MVT::i32) ? AArch64::SBFMWri : AArch64::SBFMXri; |
2652 | Opd0 = Op.getOperand(i: 0); |
2653 | Immr = ShiftImm; |
2654 | Imms = ShiftImm + Width - 1; |
2655 | return true; |
2656 | } |
2657 | |
2658 | static bool (SDNode *N, unsigned &Opc, |
2659 | SDValue &Opd0, unsigned &LSB, |
2660 | unsigned &MSB) { |
2661 | // We are looking for the following pattern which basically extracts several |
2662 | // continuous bits from the source value and places it from the LSB of the |
2663 | // destination value, all other bits of the destination value or set to zero: |
2664 | // |
2665 | // Value2 = AND Value, MaskImm |
2666 | // SRL Value2, ShiftImm |
2667 | // |
2668 | // with MaskImm >> ShiftImm to search for the bit width. |
2669 | // |
2670 | // This gets selected into a single UBFM: |
2671 | // |
2672 | // UBFM Value, ShiftImm, Log2_64(MaskImm) |
2673 | // |
2674 | |
2675 | if (N->getOpcode() != ISD::SRL) |
2676 | return false; |
2677 | |
2678 | uint64_t AndMask = 0; |
2679 | if (!isOpcWithIntImmediate(N: N->getOperand(Num: 0).getNode(), Opc: ISD::AND, Imm&: AndMask)) |
2680 | return false; |
2681 | |
2682 | Opd0 = N->getOperand(Num: 0).getOperand(i: 0); |
2683 | |
2684 | uint64_t SrlImm = 0; |
2685 | if (!isIntImmediate(N: N->getOperand(Num: 1), Imm&: SrlImm)) |
2686 | return false; |
2687 | |
2688 | // Check whether we really have several bits extract here. |
2689 | if (!isMask_64(Value: AndMask >> SrlImm)) |
2690 | return false; |
2691 | |
2692 | Opc = N->getValueType(ResNo: 0) == MVT::i32 ? AArch64::UBFMWri : AArch64::UBFMXri; |
2693 | LSB = SrlImm; |
2694 | MSB = llvm::Log2_64(Value: AndMask); |
2695 | return true; |
2696 | } |
2697 | |
2698 | static bool (SDNode *N, unsigned &Opc, SDValue &Opd0, |
2699 | unsigned &Immr, unsigned &Imms, |
2700 | bool BiggerPattern) { |
2701 | assert((N->getOpcode() == ISD::SRA || N->getOpcode() == ISD::SRL) && |
2702 | "N must be a SHR/SRA operation to call this function" ); |
2703 | |
2704 | EVT VT = N->getValueType(ResNo: 0); |
2705 | |
2706 | // Here we can test the type of VT and return false when the type does not |
2707 | // match, but since it is done prior to that call in the current context |
2708 | // we turned that into an assert to avoid redundant code. |
2709 | assert((VT == MVT::i32 || VT == MVT::i64) && |
2710 | "Type checking must have been done before calling this function" ); |
2711 | |
2712 | // Check for AND + SRL doing several bits extract. |
2713 | if (isSeveralBitsExtractOpFromShr(N, Opc, Opd0, LSB&: Immr, MSB&: Imms)) |
2714 | return true; |
2715 | |
2716 | // We're looking for a shift of a shift. |
2717 | uint64_t ShlImm = 0; |
2718 | uint64_t TruncBits = 0; |
2719 | if (isOpcWithIntImmediate(N: N->getOperand(Num: 0).getNode(), Opc: ISD::SHL, Imm&: ShlImm)) { |
2720 | Opd0 = N->getOperand(Num: 0).getOperand(i: 0); |
2721 | } else if (VT == MVT::i32 && N->getOpcode() == ISD::SRL && |
2722 | N->getOperand(Num: 0).getNode()->getOpcode() == ISD::TRUNCATE) { |
2723 | // We are looking for a shift of truncate. Truncate from i64 to i32 could |
2724 | // be considered as setting high 32 bits as zero. Our strategy here is to |
2725 | // always generate 64bit UBFM. This consistency will help the CSE pass |
2726 | // later find more redundancy. |
2727 | Opd0 = N->getOperand(Num: 0).getOperand(i: 0); |
2728 | TruncBits = Opd0->getValueType(ResNo: 0).getSizeInBits() - VT.getSizeInBits(); |
2729 | VT = Opd0.getValueType(); |
2730 | assert(VT == MVT::i64 && "the promoted type should be i64" ); |
2731 | } else if (BiggerPattern) { |
2732 | // Let's pretend a 0 shift left has been performed. |
2733 | // FIXME: Currently we limit this to the bigger pattern case, |
2734 | // because some optimizations expect AND and not UBFM |
2735 | Opd0 = N->getOperand(Num: 0); |
2736 | } else |
2737 | return false; |
2738 | |
2739 | // Missing combines/constant folding may have left us with strange |
2740 | // constants. |
2741 | if (ShlImm >= VT.getSizeInBits()) { |
2742 | LLVM_DEBUG( |
2743 | (dbgs() << N |
2744 | << ": Found large shift immediate, this should not happen\n" )); |
2745 | return false; |
2746 | } |
2747 | |
2748 | uint64_t SrlImm = 0; |
2749 | if (!isIntImmediate(N: N->getOperand(Num: 1), Imm&: SrlImm)) |
2750 | return false; |
2751 | |
2752 | assert(SrlImm > 0 && SrlImm < VT.getSizeInBits() && |
2753 | "bad amount in shift node!" ); |
2754 | int immr = SrlImm - ShlImm; |
2755 | Immr = immr < 0 ? immr + VT.getSizeInBits() : immr; |
2756 | Imms = VT.getSizeInBits() - ShlImm - TruncBits - 1; |
2757 | // SRA requires a signed extraction |
2758 | if (VT == MVT::i32) |
2759 | Opc = N->getOpcode() == ISD::SRA ? AArch64::SBFMWri : AArch64::UBFMWri; |
2760 | else |
2761 | Opc = N->getOpcode() == ISD::SRA ? AArch64::SBFMXri : AArch64::UBFMXri; |
2762 | return true; |
2763 | } |
2764 | |
2765 | bool AArch64DAGToDAGISel::(SDNode *N) { |
2766 | assert(N->getOpcode() == ISD::SIGN_EXTEND); |
2767 | |
2768 | EVT VT = N->getValueType(ResNo: 0); |
2769 | EVT NarrowVT = N->getOperand(Num: 0)->getValueType(ResNo: 0); |
2770 | if (VT != MVT::i64 || NarrowVT != MVT::i32) |
2771 | return false; |
2772 | |
2773 | uint64_t ShiftImm; |
2774 | SDValue Op = N->getOperand(Num: 0); |
2775 | if (!isOpcWithIntImmediate(N: Op.getNode(), Opc: ISD::SRA, Imm&: ShiftImm)) |
2776 | return false; |
2777 | |
2778 | SDLoc dl(N); |
2779 | // Extend the incoming operand of the shift to 64-bits. |
2780 | SDValue Opd0 = Widen(CurDAG, N: Op.getOperand(i: 0)); |
2781 | unsigned Immr = ShiftImm; |
2782 | unsigned Imms = NarrowVT.getSizeInBits() - 1; |
2783 | SDValue Ops[] = {Opd0, CurDAG->getTargetConstant(Val: Immr, DL: dl, VT), |
2784 | CurDAG->getTargetConstant(Val: Imms, DL: dl, VT)}; |
2785 | CurDAG->SelectNodeTo(N, MachineOpc: AArch64::SBFMXri, VT, Ops); |
2786 | return true; |
2787 | } |
2788 | |
2789 | static bool (SelectionDAG *CurDAG, SDNode *N, unsigned &Opc, |
2790 | SDValue &Opd0, unsigned &Immr, unsigned &Imms, |
2791 | unsigned NumberOfIgnoredLowBits = 0, |
2792 | bool BiggerPattern = false) { |
2793 | if (N->getValueType(ResNo: 0) != MVT::i32 && N->getValueType(ResNo: 0) != MVT::i64) |
2794 | return false; |
2795 | |
2796 | switch (N->getOpcode()) { |
2797 | default: |
2798 | if (!N->isMachineOpcode()) |
2799 | return false; |
2800 | break; |
2801 | case ISD::AND: |
2802 | return isBitfieldExtractOpFromAnd(CurDAG, N, Opc, Opd0, LSB&: Immr, MSB&: Imms, |
2803 | NumberOfIgnoredLowBits, BiggerPattern); |
2804 | case ISD::SRL: |
2805 | case ISD::SRA: |
2806 | return isBitfieldExtractOpFromShr(N, Opc, Opd0, Immr, Imms, BiggerPattern); |
2807 | |
2808 | case ISD::SIGN_EXTEND_INREG: |
2809 | return isBitfieldExtractOpFromSExtInReg(N, Opc, Opd0, Immr, Imms); |
2810 | } |
2811 | |
2812 | unsigned NOpc = N->getMachineOpcode(); |
2813 | switch (NOpc) { |
2814 | default: |
2815 | return false; |
2816 | case AArch64::SBFMWri: |
2817 | case AArch64::UBFMWri: |
2818 | case AArch64::SBFMXri: |
2819 | case AArch64::UBFMXri: |
2820 | Opc = NOpc; |
2821 | Opd0 = N->getOperand(Num: 0); |
2822 | Immr = N->getConstantOperandVal(Num: 1); |
2823 | Imms = N->getConstantOperandVal(Num: 2); |
2824 | return true; |
2825 | } |
2826 | // Unreachable |
2827 | return false; |
2828 | } |
2829 | |
2830 | bool AArch64DAGToDAGISel::(SDNode *N) { |
2831 | unsigned Opc, Immr, Imms; |
2832 | SDValue Opd0; |
2833 | if (!isBitfieldExtractOp(CurDAG, N, Opc, Opd0, Immr, Imms)) |
2834 | return false; |
2835 | |
2836 | EVT VT = N->getValueType(ResNo: 0); |
2837 | SDLoc dl(N); |
2838 | |
2839 | // If the bit extract operation is 64bit but the original type is 32bit, we |
2840 | // need to add one EXTRACT_SUBREG. |
2841 | if ((Opc == AArch64::SBFMXri || Opc == AArch64::UBFMXri) && VT == MVT::i32) { |
2842 | SDValue Ops64[] = {Opd0, CurDAG->getTargetConstant(Val: Immr, DL: dl, VT: MVT::i64), |
2843 | CurDAG->getTargetConstant(Val: Imms, DL: dl, VT: MVT::i64)}; |
2844 | |
2845 | SDNode *BFM = CurDAG->getMachineNode(Opcode: Opc, dl, VT: MVT::i64, Ops: Ops64); |
2846 | SDValue Inner = CurDAG->getTargetExtractSubreg(SRIdx: AArch64::sub_32, DL: dl, |
2847 | VT: MVT::i32, Operand: SDValue(BFM, 0)); |
2848 | ReplaceNode(F: N, T: Inner.getNode()); |
2849 | return true; |
2850 | } |
2851 | |
2852 | SDValue Ops[] = {Opd0, CurDAG->getTargetConstant(Val: Immr, DL: dl, VT), |
2853 | CurDAG->getTargetConstant(Val: Imms, DL: dl, VT)}; |
2854 | CurDAG->SelectNodeTo(N, MachineOpc: Opc, VT, Ops); |
2855 | return true; |
2856 | } |
2857 | |
2858 | /// Does DstMask form a complementary pair with the mask provided by |
2859 | /// BitsToBeInserted, suitable for use in a BFI instruction. Roughly speaking, |
2860 | /// this asks whether DstMask zeroes precisely those bits that will be set by |
2861 | /// the other half. |
2862 | static bool isBitfieldDstMask(uint64_t DstMask, const APInt &BitsToBeInserted, |
2863 | unsigned NumberOfIgnoredHighBits, EVT VT) { |
2864 | assert((VT == MVT::i32 || VT == MVT::i64) && |
2865 | "i32 or i64 mask type expected!" ); |
2866 | unsigned BitWidth = VT.getSizeInBits() - NumberOfIgnoredHighBits; |
2867 | |
2868 | // Enable implicitTrunc as we're intentionally ignoring high bits. |
2869 | APInt SignificantDstMask = |
2870 | APInt(BitWidth, DstMask, /*isSigned=*/false, /*implicitTrunc=*/true); |
2871 | APInt SignificantBitsToBeInserted = BitsToBeInserted.zextOrTrunc(width: BitWidth); |
2872 | |
2873 | return (SignificantDstMask & SignificantBitsToBeInserted) == 0 && |
2874 | (SignificantDstMask | SignificantBitsToBeInserted).isAllOnes(); |
2875 | } |
2876 | |
2877 | // Look for bits that will be useful for later uses. |
2878 | // A bit is consider useless as soon as it is dropped and never used |
2879 | // before it as been dropped. |
2880 | // E.g., looking for useful bit of x |
2881 | // 1. y = x & 0x7 |
2882 | // 2. z = y >> 2 |
2883 | // After #1, x useful bits are 0x7, then the useful bits of x, live through |
2884 | // y. |
2885 | // After #2, the useful bits of x are 0x4. |
2886 | // However, if x is used on an unpredictable instruction, then all its bits |
2887 | // are useful. |
2888 | // E.g. |
2889 | // 1. y = x & 0x7 |
2890 | // 2. z = y >> 2 |
2891 | // 3. str x, [@x] |
2892 | static void getUsefulBits(SDValue Op, APInt &UsefulBits, unsigned Depth = 0); |
2893 | |
2894 | static void getUsefulBitsFromAndWithImmediate(SDValue Op, APInt &UsefulBits, |
2895 | unsigned Depth) { |
2896 | uint64_t Imm = |
2897 | cast<const ConstantSDNode>(Val: Op.getOperand(i: 1).getNode())->getZExtValue(); |
2898 | Imm = AArch64_AM::decodeLogicalImmediate(val: Imm, regSize: UsefulBits.getBitWidth()); |
2899 | UsefulBits &= APInt(UsefulBits.getBitWidth(), Imm); |
2900 | getUsefulBits(Op, UsefulBits, Depth: Depth + 1); |
2901 | } |
2902 | |
2903 | static void getUsefulBitsFromBitfieldMoveOpd(SDValue Op, APInt &UsefulBits, |
2904 | uint64_t Imm, uint64_t MSB, |
2905 | unsigned Depth) { |
2906 | // inherit the bitwidth value |
2907 | APInt OpUsefulBits(UsefulBits); |
2908 | OpUsefulBits = 1; |
2909 | |
2910 | if (MSB >= Imm) { |
2911 | OpUsefulBits <<= MSB - Imm + 1; |
2912 | --OpUsefulBits; |
2913 | // The interesting part will be in the lower part of the result |
2914 | getUsefulBits(Op, UsefulBits&: OpUsefulBits, Depth: Depth + 1); |
2915 | // The interesting part was starting at Imm in the argument |
2916 | OpUsefulBits <<= Imm; |
2917 | } else { |
2918 | OpUsefulBits <<= MSB + 1; |
2919 | --OpUsefulBits; |
2920 | // The interesting part will be shifted in the result |
2921 | OpUsefulBits <<= OpUsefulBits.getBitWidth() - Imm; |
2922 | getUsefulBits(Op, UsefulBits&: OpUsefulBits, Depth: Depth + 1); |
2923 | // The interesting part was at zero in the argument |
2924 | OpUsefulBits.lshrInPlace(ShiftAmt: OpUsefulBits.getBitWidth() - Imm); |
2925 | } |
2926 | |
2927 | UsefulBits &= OpUsefulBits; |
2928 | } |
2929 | |
2930 | static void getUsefulBitsFromUBFM(SDValue Op, APInt &UsefulBits, |
2931 | unsigned Depth) { |
2932 | uint64_t Imm = |
2933 | cast<const ConstantSDNode>(Val: Op.getOperand(i: 1).getNode())->getZExtValue(); |
2934 | uint64_t MSB = |
2935 | cast<const ConstantSDNode>(Val: Op.getOperand(i: 2).getNode())->getZExtValue(); |
2936 | |
2937 | getUsefulBitsFromBitfieldMoveOpd(Op, UsefulBits, Imm, MSB, Depth); |
2938 | } |
2939 | |
2940 | static void getUsefulBitsFromOrWithShiftedReg(SDValue Op, APInt &UsefulBits, |
2941 | unsigned Depth) { |
2942 | uint64_t ShiftTypeAndValue = |
2943 | cast<const ConstantSDNode>(Val: Op.getOperand(i: 2).getNode())->getZExtValue(); |
2944 | APInt Mask(UsefulBits); |
2945 | Mask.clearAllBits(); |
2946 | Mask.flipAllBits(); |
2947 | |
2948 | if (AArch64_AM::getShiftType(Imm: ShiftTypeAndValue) == AArch64_AM::LSL) { |
2949 | // Shift Left |
2950 | uint64_t ShiftAmt = AArch64_AM::getShiftValue(Imm: ShiftTypeAndValue); |
2951 | Mask <<= ShiftAmt; |
2952 | getUsefulBits(Op, UsefulBits&: Mask, Depth: Depth + 1); |
2953 | Mask.lshrInPlace(ShiftAmt); |
2954 | } else if (AArch64_AM::getShiftType(Imm: ShiftTypeAndValue) == AArch64_AM::LSR) { |
2955 | // Shift Right |
2956 | // We do not handle AArch64_AM::ASR, because the sign will change the |
2957 | // number of useful bits |
2958 | uint64_t ShiftAmt = AArch64_AM::getShiftValue(Imm: ShiftTypeAndValue); |
2959 | Mask.lshrInPlace(ShiftAmt); |
2960 | getUsefulBits(Op, UsefulBits&: Mask, Depth: Depth + 1); |
2961 | Mask <<= ShiftAmt; |
2962 | } else |
2963 | return; |
2964 | |
2965 | UsefulBits &= Mask; |
2966 | } |
2967 | |
2968 | static void getUsefulBitsFromBFM(SDValue Op, SDValue Orig, APInt &UsefulBits, |
2969 | unsigned Depth) { |
2970 | uint64_t Imm = |
2971 | cast<const ConstantSDNode>(Val: Op.getOperand(i: 2).getNode())->getZExtValue(); |
2972 | uint64_t MSB = |
2973 | cast<const ConstantSDNode>(Val: Op.getOperand(i: 3).getNode())->getZExtValue(); |
2974 | |
2975 | APInt OpUsefulBits(UsefulBits); |
2976 | OpUsefulBits = 1; |
2977 | |
2978 | APInt ResultUsefulBits(UsefulBits.getBitWidth(), 0); |
2979 | ResultUsefulBits.flipAllBits(); |
2980 | APInt Mask(UsefulBits.getBitWidth(), 0); |
2981 | |
2982 | getUsefulBits(Op, UsefulBits&: ResultUsefulBits, Depth: Depth + 1); |
2983 | |
2984 | if (MSB >= Imm) { |
2985 | // The instruction is a BFXIL. |
2986 | uint64_t Width = MSB - Imm + 1; |
2987 | uint64_t LSB = Imm; |
2988 | |
2989 | OpUsefulBits <<= Width; |
2990 | --OpUsefulBits; |
2991 | |
2992 | if (Op.getOperand(i: 1) == Orig) { |
2993 | // Copy the low bits from the result to bits starting from LSB. |
2994 | Mask = ResultUsefulBits & OpUsefulBits; |
2995 | Mask <<= LSB; |
2996 | } |
2997 | |
2998 | if (Op.getOperand(i: 0) == Orig) |
2999 | // Bits starting from LSB in the input contribute to the result. |
3000 | Mask |= (ResultUsefulBits & ~OpUsefulBits); |
3001 | } else { |
3002 | // The instruction is a BFI. |
3003 | uint64_t Width = MSB + 1; |
3004 | uint64_t LSB = UsefulBits.getBitWidth() - Imm; |
3005 | |
3006 | OpUsefulBits <<= Width; |
3007 | --OpUsefulBits; |
3008 | OpUsefulBits <<= LSB; |
3009 | |
3010 | if (Op.getOperand(i: 1) == Orig) { |
3011 | // Copy the bits from the result to the zero bits. |
3012 | Mask = ResultUsefulBits & OpUsefulBits; |
3013 | Mask.lshrInPlace(ShiftAmt: LSB); |
3014 | } |
3015 | |
3016 | if (Op.getOperand(i: 0) == Orig) |
3017 | Mask |= (ResultUsefulBits & ~OpUsefulBits); |
3018 | } |
3019 | |
3020 | UsefulBits &= Mask; |
3021 | } |
3022 | |
3023 | static void getUsefulBitsForUse(SDNode *UserNode, APInt &UsefulBits, |
3024 | SDValue Orig, unsigned Depth) { |
3025 | |
3026 | // Users of this node should have already been instruction selected |
3027 | // FIXME: Can we turn that into an assert? |
3028 | if (!UserNode->isMachineOpcode()) |
3029 | return; |
3030 | |
3031 | switch (UserNode->getMachineOpcode()) { |
3032 | default: |
3033 | return; |
3034 | case AArch64::ANDSWri: |
3035 | case AArch64::ANDSXri: |
3036 | case AArch64::ANDWri: |
3037 | case AArch64::ANDXri: |
3038 | // We increment Depth only when we call the getUsefulBits |
3039 | return getUsefulBitsFromAndWithImmediate(Op: SDValue(UserNode, 0), UsefulBits, |
3040 | Depth); |
3041 | case AArch64::UBFMWri: |
3042 | case AArch64::UBFMXri: |
3043 | return getUsefulBitsFromUBFM(Op: SDValue(UserNode, 0), UsefulBits, Depth); |
3044 | |
3045 | case AArch64::ORRWrs: |
3046 | case AArch64::ORRXrs: |
3047 | if (UserNode->getOperand(Num: 0) != Orig && UserNode->getOperand(Num: 1) == Orig) |
3048 | getUsefulBitsFromOrWithShiftedReg(Op: SDValue(UserNode, 0), UsefulBits, |
3049 | Depth); |
3050 | return; |
3051 | case AArch64::BFMWri: |
3052 | case AArch64::BFMXri: |
3053 | return getUsefulBitsFromBFM(Op: SDValue(UserNode, 0), Orig, UsefulBits, Depth); |
3054 | |
3055 | case AArch64::STRBBui: |
3056 | case AArch64::STURBBi: |
3057 | if (UserNode->getOperand(Num: 0) != Orig) |
3058 | return; |
3059 | UsefulBits &= APInt(UsefulBits.getBitWidth(), 0xff); |
3060 | return; |
3061 | |
3062 | case AArch64::STRHHui: |
3063 | case AArch64::STURHHi: |
3064 | if (UserNode->getOperand(Num: 0) != Orig) |
3065 | return; |
3066 | UsefulBits &= APInt(UsefulBits.getBitWidth(), 0xffff); |
3067 | return; |
3068 | } |
3069 | } |
3070 | |
3071 | static void getUsefulBits(SDValue Op, APInt &UsefulBits, unsigned Depth) { |
3072 | if (Depth >= SelectionDAG::MaxRecursionDepth) |
3073 | return; |
3074 | // Initialize UsefulBits |
3075 | if (!Depth) { |
3076 | unsigned Bitwidth = Op.getScalarValueSizeInBits(); |
3077 | // At the beginning, assume every produced bits is useful |
3078 | UsefulBits = APInt(Bitwidth, 0); |
3079 | UsefulBits.flipAllBits(); |
3080 | } |
3081 | APInt UsersUsefulBits(UsefulBits.getBitWidth(), 0); |
3082 | |
3083 | for (SDNode *Node : Op.getNode()->users()) { |
3084 | // A use cannot produce useful bits |
3085 | APInt UsefulBitsForUse = APInt(UsefulBits); |
3086 | getUsefulBitsForUse(UserNode: Node, UsefulBits&: UsefulBitsForUse, Orig: Op, Depth); |
3087 | UsersUsefulBits |= UsefulBitsForUse; |
3088 | } |
3089 | // UsefulBits contains the produced bits that are meaningful for the |
3090 | // current definition, thus a user cannot make a bit meaningful at |
3091 | // this point |
3092 | UsefulBits &= UsersUsefulBits; |
3093 | } |
3094 | |
3095 | /// Create a machine node performing a notional SHL of Op by ShlAmount. If |
3096 | /// ShlAmount is negative, do a (logical) right-shift instead. If ShlAmount is |
3097 | /// 0, return Op unchanged. |
3098 | static SDValue getLeftShift(SelectionDAG *CurDAG, SDValue Op, int ShlAmount) { |
3099 | if (ShlAmount == 0) |
3100 | return Op; |
3101 | |
3102 | EVT VT = Op.getValueType(); |
3103 | SDLoc dl(Op); |
3104 | unsigned BitWidth = VT.getSizeInBits(); |
3105 | unsigned UBFMOpc = BitWidth == 32 ? AArch64::UBFMWri : AArch64::UBFMXri; |
3106 | |
3107 | SDNode *ShiftNode; |
3108 | if (ShlAmount > 0) { |
3109 | // LSL wD, wN, #Amt == UBFM wD, wN, #32-Amt, #31-Amt |
3110 | ShiftNode = CurDAG->getMachineNode( |
3111 | Opcode: UBFMOpc, dl, VT, Op1: Op, |
3112 | Op2: CurDAG->getTargetConstant(Val: BitWidth - ShlAmount, DL: dl, VT), |
3113 | Op3: CurDAG->getTargetConstant(Val: BitWidth - 1 - ShlAmount, DL: dl, VT)); |
3114 | } else { |
3115 | // LSR wD, wN, #Amt == UBFM wD, wN, #Amt, #32-1 |
3116 | assert(ShlAmount < 0 && "expected right shift" ); |
3117 | int ShrAmount = -ShlAmount; |
3118 | ShiftNode = CurDAG->getMachineNode( |
3119 | Opcode: UBFMOpc, dl, VT, Op1: Op, Op2: CurDAG->getTargetConstant(Val: ShrAmount, DL: dl, VT), |
3120 | Op3: CurDAG->getTargetConstant(Val: BitWidth - 1, DL: dl, VT)); |
3121 | } |
3122 | |
3123 | return SDValue(ShiftNode, 0); |
3124 | } |
3125 | |
3126 | // For bit-field-positioning pattern "(and (shl VAL, N), ShiftedMask)". |
3127 | static bool isBitfieldPositioningOpFromAnd(SelectionDAG *CurDAG, SDValue Op, |
3128 | bool BiggerPattern, |
3129 | const uint64_t NonZeroBits, |
3130 | SDValue &Src, int &DstLSB, |
3131 | int &Width); |
3132 | |
3133 | // For bit-field-positioning pattern "shl VAL, N)". |
3134 | static bool isBitfieldPositioningOpFromShl(SelectionDAG *CurDAG, SDValue Op, |
3135 | bool BiggerPattern, |
3136 | const uint64_t NonZeroBits, |
3137 | SDValue &Src, int &DstLSB, |
3138 | int &Width); |
3139 | |
3140 | /// Does this tree qualify as an attempt to move a bitfield into position, |
3141 | /// essentially "(and (shl VAL, N), Mask)" or (shl VAL, N). |
3142 | static bool isBitfieldPositioningOp(SelectionDAG *CurDAG, SDValue Op, |
3143 | bool BiggerPattern, SDValue &Src, |
3144 | int &DstLSB, int &Width) { |
3145 | EVT VT = Op.getValueType(); |
3146 | unsigned BitWidth = VT.getSizeInBits(); |
3147 | (void)BitWidth; |
3148 | assert(BitWidth == 32 || BitWidth == 64); |
3149 | |
3150 | KnownBits Known = CurDAG->computeKnownBits(Op); |
3151 | |
3152 | // Non-zero in the sense that they're not provably zero, which is the key |
3153 | // point if we want to use this value |
3154 | const uint64_t NonZeroBits = (~Known.Zero).getZExtValue(); |
3155 | if (!isShiftedMask_64(Value: NonZeroBits)) |
3156 | return false; |
3157 | |
3158 | switch (Op.getOpcode()) { |
3159 | default: |
3160 | break; |
3161 | case ISD::AND: |
3162 | return isBitfieldPositioningOpFromAnd(CurDAG, Op, BiggerPattern, |
3163 | NonZeroBits, Src, DstLSB, Width); |
3164 | case ISD::SHL: |
3165 | return isBitfieldPositioningOpFromShl(CurDAG, Op, BiggerPattern, |
3166 | NonZeroBits, Src, DstLSB, Width); |
3167 | } |
3168 | |
3169 | return false; |
3170 | } |
3171 | |
3172 | static bool isBitfieldPositioningOpFromAnd(SelectionDAG *CurDAG, SDValue Op, |
3173 | bool BiggerPattern, |
3174 | const uint64_t NonZeroBits, |
3175 | SDValue &Src, int &DstLSB, |
3176 | int &Width) { |
3177 | assert(isShiftedMask_64(NonZeroBits) && "Caller guaranteed" ); |
3178 | |
3179 | EVT VT = Op.getValueType(); |
3180 | assert((VT == MVT::i32 || VT == MVT::i64) && |
3181 | "Caller guarantees VT is one of i32 or i64" ); |
3182 | (void)VT; |
3183 | |
3184 | uint64_t AndImm; |
3185 | if (!isOpcWithIntImmediate(N: Op.getNode(), Opc: ISD::AND, Imm&: AndImm)) |
3186 | return false; |
3187 | |
3188 | // If (~AndImm & NonZeroBits) is not zero at POS, we know that |
3189 | // 1) (AndImm & (1 << POS) == 0) |
3190 | // 2) the result of AND is not zero at POS bit (according to NonZeroBits) |
3191 | // |
3192 | // 1) and 2) don't agree so something must be wrong (e.g., in |
3193 | // 'SelectionDAG::computeKnownBits') |
3194 | assert((~AndImm & NonZeroBits) == 0 && |
3195 | "Something must be wrong (e.g., in SelectionDAG::computeKnownBits)" ); |
3196 | |
3197 | SDValue AndOp0 = Op.getOperand(i: 0); |
3198 | |
3199 | uint64_t ShlImm; |
3200 | SDValue ShlOp0; |
3201 | if (isOpcWithIntImmediate(N: AndOp0.getNode(), Opc: ISD::SHL, Imm&: ShlImm)) { |
3202 | // For pattern "and(shl(val, N), shifted-mask)", 'ShlOp0' is set to 'val'. |
3203 | ShlOp0 = AndOp0.getOperand(i: 0); |
3204 | } else if (VT == MVT::i64 && AndOp0.getOpcode() == ISD::ANY_EXTEND && |
3205 | isOpcWithIntImmediate(N: AndOp0.getOperand(i: 0).getNode(), Opc: ISD::SHL, |
3206 | Imm&: ShlImm)) { |
3207 | // For pattern "and(any_extend(shl(val, N)), shifted-mask)" |
3208 | |
3209 | // ShlVal == shl(val, N), which is a left shift on a smaller type. |
3210 | SDValue ShlVal = AndOp0.getOperand(i: 0); |
3211 | |
3212 | // Since this is after type legalization and ShlVal is extended to MVT::i64, |
3213 | // expect VT to be MVT::i32. |
3214 | assert((ShlVal.getValueType() == MVT::i32) && "Expect VT to be MVT::i32." ); |
3215 | |
3216 | // Widens 'val' to MVT::i64 as the source of bit field positioning. |
3217 | ShlOp0 = Widen(CurDAG, N: ShlVal.getOperand(i: 0)); |
3218 | } else |
3219 | return false; |
3220 | |
3221 | // For !BiggerPattern, bail out if the AndOp0 has more than one use, since |
3222 | // then we'll end up generating AndOp0+UBFIZ instead of just keeping |
3223 | // AndOp0+AND. |
3224 | if (!BiggerPattern && !AndOp0.hasOneUse()) |
3225 | return false; |
3226 | |
3227 | DstLSB = llvm::countr_zero(Val: NonZeroBits); |
3228 | Width = llvm::countr_one(Value: NonZeroBits >> DstLSB); |
3229 | |
3230 | // Bail out on large Width. This happens when no proper combining / constant |
3231 | // folding was performed. |
3232 | if (Width >= (int)VT.getSizeInBits()) { |
3233 | // If VT is i64, Width > 64 is insensible since NonZeroBits is uint64_t, and |
3234 | // Width == 64 indicates a missed dag-combine from "(and val, AllOnes)" to |
3235 | // "val". |
3236 | // If VT is i32, what Width >= 32 means: |
3237 | // - For "(and (any_extend(shl val, N)), shifted-mask)", the`and` Op |
3238 | // demands at least 'Width' bits (after dag-combiner). This together with |
3239 | // `any_extend` Op (undefined higher bits) indicates missed combination |
3240 | // when lowering the 'and' IR instruction to an machine IR instruction. |
3241 | LLVM_DEBUG( |
3242 | dbgs() |
3243 | << "Found large Width in bit-field-positioning -- this indicates no " |
3244 | "proper combining / constant folding was performed\n" ); |
3245 | return false; |
3246 | } |
3247 | |
3248 | // BFI encompasses sufficiently many nodes that it's worth inserting an extra |
3249 | // LSL/LSR if the mask in NonZeroBits doesn't quite match up with the ISD::SHL |
3250 | // amount. BiggerPattern is true when this pattern is being matched for BFI, |
3251 | // BiggerPattern is false when this pattern is being matched for UBFIZ, in |
3252 | // which case it is not profitable to insert an extra shift. |
3253 | if (ShlImm != uint64_t(DstLSB) && !BiggerPattern) |
3254 | return false; |
3255 | |
3256 | Src = getLeftShift(CurDAG, Op: ShlOp0, ShlAmount: ShlImm - DstLSB); |
3257 | return true; |
3258 | } |
3259 | |
3260 | // For node (shl (and val, mask), N)), returns true if the node is equivalent to |
3261 | // UBFIZ. |
3262 | static bool isSeveralBitsPositioningOpFromShl(const uint64_t ShlImm, SDValue Op, |
3263 | SDValue &Src, int &DstLSB, |
3264 | int &Width) { |
3265 | // Caller should have verified that N is a left shift with constant shift |
3266 | // amount; asserts that. |
3267 | assert(Op.getOpcode() == ISD::SHL && |
3268 | "Op.getNode() should be a SHL node to call this function" ); |
3269 | assert(isIntImmediateEq(Op.getOperand(1), ShlImm) && |
3270 | "Op.getNode() should shift ShlImm to call this function" ); |
3271 | |
3272 | uint64_t AndImm = 0; |
3273 | SDValue Op0 = Op.getOperand(i: 0); |
3274 | if (!isOpcWithIntImmediate(N: Op0.getNode(), Opc: ISD::AND, Imm&: AndImm)) |
3275 | return false; |
3276 | |
3277 | const uint64_t ShiftedAndImm = ((AndImm << ShlImm) >> ShlImm); |
3278 | if (isMask_64(Value: ShiftedAndImm)) { |
3279 | // AndImm is a superset of (AllOnes >> ShlImm); in other words, AndImm |
3280 | // should end with Mask, and could be prefixed with random bits if those |
3281 | // bits are shifted out. |
3282 | // |
3283 | // For example, xyz11111 (with {x,y,z} being 0 or 1) is fine if ShlImm >= 3; |
3284 | // the AND result corresponding to those bits are shifted out, so it's fine |
3285 | // to not extract them. |
3286 | Width = llvm::countr_one(Value: ShiftedAndImm); |
3287 | DstLSB = ShlImm; |
3288 | Src = Op0.getOperand(i: 0); |
3289 | return true; |
3290 | } |
3291 | return false; |
3292 | } |
3293 | |
3294 | static bool isBitfieldPositioningOpFromShl(SelectionDAG *CurDAG, SDValue Op, |
3295 | bool BiggerPattern, |
3296 | const uint64_t NonZeroBits, |
3297 | SDValue &Src, int &DstLSB, |
3298 | int &Width) { |
3299 | assert(isShiftedMask_64(NonZeroBits) && "Caller guaranteed" ); |
3300 | |
3301 | EVT VT = Op.getValueType(); |
3302 | assert((VT == MVT::i32 || VT == MVT::i64) && |
3303 | "Caller guarantees that type is i32 or i64" ); |
3304 | (void)VT; |
3305 | |
3306 | uint64_t ShlImm; |
3307 | if (!isOpcWithIntImmediate(N: Op.getNode(), Opc: ISD::SHL, Imm&: ShlImm)) |
3308 | return false; |
3309 | |
3310 | if (!BiggerPattern && !Op.hasOneUse()) |
3311 | return false; |
3312 | |
3313 | if (isSeveralBitsPositioningOpFromShl(ShlImm, Op, Src, DstLSB, Width)) |
3314 | return true; |
3315 | |
3316 | DstLSB = llvm::countr_zero(Val: NonZeroBits); |
3317 | Width = llvm::countr_one(Value: NonZeroBits >> DstLSB); |
3318 | |
3319 | if (ShlImm != uint64_t(DstLSB) && !BiggerPattern) |
3320 | return false; |
3321 | |
3322 | Src = getLeftShift(CurDAG, Op: Op.getOperand(i: 0), ShlAmount: ShlImm - DstLSB); |
3323 | return true; |
3324 | } |
3325 | |
3326 | static bool isShiftedMask(uint64_t Mask, EVT VT) { |
3327 | assert(VT == MVT::i32 || VT == MVT::i64); |
3328 | if (VT == MVT::i32) |
3329 | return isShiftedMask_32(Value: Mask); |
3330 | return isShiftedMask_64(Value: Mask); |
3331 | } |
3332 | |
3333 | // Generate a BFI/BFXIL from 'or (and X, MaskImm), OrImm' iff the value being |
3334 | // inserted only sets known zero bits. |
3335 | static bool tryBitfieldInsertOpFromOrAndImm(SDNode *N, SelectionDAG *CurDAG) { |
3336 | assert(N->getOpcode() == ISD::OR && "Expect a OR operation" ); |
3337 | |
3338 | EVT VT = N->getValueType(ResNo: 0); |
3339 | if (VT != MVT::i32 && VT != MVT::i64) |
3340 | return false; |
3341 | |
3342 | unsigned BitWidth = VT.getSizeInBits(); |
3343 | |
3344 | uint64_t OrImm; |
3345 | if (!isOpcWithIntImmediate(N, Opc: ISD::OR, Imm&: OrImm)) |
3346 | return false; |
3347 | |
3348 | // Skip this transformation if the ORR immediate can be encoded in the ORR. |
3349 | // Otherwise, we'll trade an AND+ORR for ORR+BFI/BFXIL, which is most likely |
3350 | // performance neutral. |
3351 | if (AArch64_AM::isLogicalImmediate(imm: OrImm, regSize: BitWidth)) |
3352 | return false; |
3353 | |
3354 | uint64_t MaskImm; |
3355 | SDValue And = N->getOperand(Num: 0); |
3356 | // Must be a single use AND with an immediate operand. |
3357 | if (!And.hasOneUse() || |
3358 | !isOpcWithIntImmediate(N: And.getNode(), Opc: ISD::AND, Imm&: MaskImm)) |
3359 | return false; |
3360 | |
3361 | // Compute the Known Zero for the AND as this allows us to catch more general |
3362 | // cases than just looking for AND with imm. |
3363 | KnownBits Known = CurDAG->computeKnownBits(Op: And); |
3364 | |
3365 | // Non-zero in the sense that they're not provably zero, which is the key |
3366 | // point if we want to use this value. |
3367 | uint64_t NotKnownZero = (~Known.Zero).getZExtValue(); |
3368 | |
3369 | // The KnownZero mask must be a shifted mask (e.g., 1110..011, 11100..00). |
3370 | if (!isShiftedMask(Mask: Known.Zero.getZExtValue(), VT)) |
3371 | return false; |
3372 | |
3373 | // The bits being inserted must only set those bits that are known to be zero. |
3374 | if ((OrImm & NotKnownZero) != 0) { |
3375 | // FIXME: It's okay if the OrImm sets NotKnownZero bits to 1, but we don't |
3376 | // currently handle this case. |
3377 | return false; |
3378 | } |
3379 | |
3380 | // BFI/BFXIL dst, src, #lsb, #width. |
3381 | int LSB = llvm::countr_one(Value: NotKnownZero); |
3382 | int Width = BitWidth - APInt(BitWidth, NotKnownZero).popcount(); |
3383 | |
3384 | // BFI/BFXIL is an alias of BFM, so translate to BFM operands. |
3385 | unsigned ImmR = (BitWidth - LSB) % BitWidth; |
3386 | unsigned ImmS = Width - 1; |
3387 | |
3388 | // If we're creating a BFI instruction avoid cases where we need more |
3389 | // instructions to materialize the BFI constant as compared to the original |
3390 | // ORR. A BFXIL will use the same constant as the original ORR, so the code |
3391 | // should be no worse in this case. |
3392 | bool IsBFI = LSB != 0; |
3393 | uint64_t BFIImm = OrImm >> LSB; |
3394 | if (IsBFI && !AArch64_AM::isLogicalImmediate(imm: BFIImm, regSize: BitWidth)) { |
3395 | // We have a BFI instruction and we know the constant can't be materialized |
3396 | // with a ORR-immediate with the zero register. |
3397 | unsigned OrChunks = 0, BFIChunks = 0; |
3398 | for (unsigned Shift = 0; Shift < BitWidth; Shift += 16) { |
3399 | if (((OrImm >> Shift) & 0xFFFF) != 0) |
3400 | ++OrChunks; |
3401 | if (((BFIImm >> Shift) & 0xFFFF) != 0) |
3402 | ++BFIChunks; |
3403 | } |
3404 | if (BFIChunks > OrChunks) |
3405 | return false; |
3406 | } |
3407 | |
3408 | // Materialize the constant to be inserted. |
3409 | SDLoc DL(N); |
3410 | unsigned MOVIOpc = VT == MVT::i32 ? AArch64::MOVi32imm : AArch64::MOVi64imm; |
3411 | SDNode *MOVI = CurDAG->getMachineNode( |
3412 | Opcode: MOVIOpc, dl: DL, VT, Op1: CurDAG->getTargetConstant(Val: BFIImm, DL, VT)); |
3413 | |
3414 | // Create the BFI/BFXIL instruction. |
3415 | SDValue Ops[] = {And.getOperand(i: 0), SDValue(MOVI, 0), |
3416 | CurDAG->getTargetConstant(Val: ImmR, DL, VT), |
3417 | CurDAG->getTargetConstant(Val: ImmS, DL, VT)}; |
3418 | unsigned Opc = (VT == MVT::i32) ? AArch64::BFMWri : AArch64::BFMXri; |
3419 | CurDAG->SelectNodeTo(N, MachineOpc: Opc, VT, Ops); |
3420 | return true; |
3421 | } |
3422 | |
3423 | static bool isWorthFoldingIntoOrrWithShift(SDValue Dst, SelectionDAG *CurDAG, |
3424 | SDValue &ShiftedOperand, |
3425 | uint64_t &EncodedShiftImm) { |
3426 | // Avoid folding Dst into ORR-with-shift if Dst has other uses than ORR. |
3427 | if (!Dst.hasOneUse()) |
3428 | return false; |
3429 | |
3430 | EVT VT = Dst.getValueType(); |
3431 | assert((VT == MVT::i32 || VT == MVT::i64) && |
3432 | "Caller should guarantee that VT is one of i32 or i64" ); |
3433 | const unsigned SizeInBits = VT.getSizeInBits(); |
3434 | |
3435 | SDLoc DL(Dst.getNode()); |
3436 | uint64_t AndImm, ShlImm; |
3437 | if (isOpcWithIntImmediate(N: Dst.getNode(), Opc: ISD::AND, Imm&: AndImm) && |
3438 | isShiftedMask_64(Value: AndImm)) { |
3439 | // Avoid transforming 'DstOp0' if it has other uses than the AND node. |
3440 | SDValue DstOp0 = Dst.getOperand(i: 0); |
3441 | if (!DstOp0.hasOneUse()) |
3442 | return false; |
3443 | |
3444 | // An example to illustrate the transformation |
3445 | // From: |
3446 | // lsr x8, x1, #1 |
3447 | // and x8, x8, #0x3f80 |
3448 | // bfxil x8, x1, #0, #7 |
3449 | // To: |
3450 | // and x8, x23, #0x7f |
3451 | // ubfx x9, x23, #8, #7 |
3452 | // orr x23, x8, x9, lsl #7 |
3453 | // |
3454 | // The number of instructions remains the same, but ORR is faster than BFXIL |
3455 | // on many AArch64 processors (or as good as BFXIL if not faster). Besides, |
3456 | // the dependency chain is improved after the transformation. |
3457 | uint64_t SrlImm; |
3458 | if (isOpcWithIntImmediate(N: DstOp0.getNode(), Opc: ISD::SRL, Imm&: SrlImm)) { |
3459 | uint64_t NumTrailingZeroInShiftedMask = llvm::countr_zero(Val: AndImm); |
3460 | if ((SrlImm + NumTrailingZeroInShiftedMask) < SizeInBits) { |
3461 | unsigned MaskWidth = |
3462 | llvm::countr_one(Value: AndImm >> NumTrailingZeroInShiftedMask); |
3463 | unsigned UBFMOpc = |
3464 | (VT == MVT::i32) ? AArch64::UBFMWri : AArch64::UBFMXri; |
3465 | SDNode *UBFMNode = CurDAG->getMachineNode( |
3466 | Opcode: UBFMOpc, dl: DL, VT, Op1: DstOp0.getOperand(i: 0), |
3467 | Op2: CurDAG->getTargetConstant(Val: SrlImm + NumTrailingZeroInShiftedMask, DL, |
3468 | VT), |
3469 | Op3: CurDAG->getTargetConstant( |
3470 | Val: SrlImm + NumTrailingZeroInShiftedMask + MaskWidth - 1, DL, VT)); |
3471 | ShiftedOperand = SDValue(UBFMNode, 0); |
3472 | EncodedShiftImm = AArch64_AM::getShifterImm( |
3473 | ST: AArch64_AM::LSL, Imm: NumTrailingZeroInShiftedMask); |
3474 | return true; |
3475 | } |
3476 | } |
3477 | return false; |
3478 | } |
3479 | |
3480 | if (isOpcWithIntImmediate(N: Dst.getNode(), Opc: ISD::SHL, Imm&: ShlImm)) { |
3481 | ShiftedOperand = Dst.getOperand(i: 0); |
3482 | EncodedShiftImm = AArch64_AM::getShifterImm(ST: AArch64_AM::LSL, Imm: ShlImm); |
3483 | return true; |
3484 | } |
3485 | |
3486 | uint64_t SrlImm; |
3487 | if (isOpcWithIntImmediate(N: Dst.getNode(), Opc: ISD::SRL, Imm&: SrlImm)) { |
3488 | ShiftedOperand = Dst.getOperand(i: 0); |
3489 | EncodedShiftImm = AArch64_AM::getShifterImm(ST: AArch64_AM::LSR, Imm: SrlImm); |
3490 | return true; |
3491 | } |
3492 | return false; |
3493 | } |
3494 | |
3495 | // Given an 'ISD::OR' node that is going to be selected as BFM, analyze |
3496 | // the operands and select it to AArch64::ORR with shifted registers if |
3497 | // that's more efficient. Returns true iff selection to AArch64::ORR happens. |
3498 | static bool tryOrrWithShift(SDNode *N, SDValue OrOpd0, SDValue OrOpd1, |
3499 | SDValue Src, SDValue Dst, SelectionDAG *CurDAG, |
3500 | const bool BiggerPattern) { |
3501 | EVT VT = N->getValueType(ResNo: 0); |
3502 | assert(N->getOpcode() == ISD::OR && "Expect N to be an OR node" ); |
3503 | assert(((N->getOperand(0) == OrOpd0 && N->getOperand(1) == OrOpd1) || |
3504 | (N->getOperand(1) == OrOpd0 && N->getOperand(0) == OrOpd1)) && |
3505 | "Expect OrOpd0 and OrOpd1 to be operands of ISD::OR" ); |
3506 | assert((VT == MVT::i32 || VT == MVT::i64) && |
3507 | "Expect result type to be i32 or i64 since N is combinable to BFM" ); |
3508 | SDLoc DL(N); |
3509 | |
3510 | // Bail out if BFM simplifies away one node in BFM Dst. |
3511 | if (OrOpd1 != Dst) |
3512 | return false; |
3513 | |
3514 | const unsigned OrrOpc = (VT == MVT::i32) ? AArch64::ORRWrs : AArch64::ORRXrs; |
3515 | // For "BFM Rd, Rn, #immr, #imms", it's known that BFM simplifies away fewer |
3516 | // nodes from Rn (or inserts additional shift node) if BiggerPattern is true. |
3517 | if (BiggerPattern) { |
3518 | uint64_t SrcAndImm; |
3519 | if (isOpcWithIntImmediate(N: OrOpd0.getNode(), Opc: ISD::AND, Imm&: SrcAndImm) && |
3520 | isMask_64(Value: SrcAndImm) && OrOpd0.getOperand(i: 0) == Src) { |
3521 | // OrOpd0 = AND Src, #Mask |
3522 | // So BFM simplifies away one AND node from Src and doesn't simplify away |
3523 | // nodes from Dst. If ORR with left-shifted operand also simplifies away |
3524 | // one node (from Rd), ORR is better since it has higher throughput and |
3525 | // smaller latency than BFM on many AArch64 processors (and for the rest |
3526 | // ORR is at least as good as BFM). |
3527 | SDValue ShiftedOperand; |
3528 | uint64_t EncodedShiftImm; |
3529 | if (isWorthFoldingIntoOrrWithShift(Dst, CurDAG, ShiftedOperand, |
3530 | EncodedShiftImm)) { |
3531 | SDValue Ops[] = {OrOpd0, ShiftedOperand, |
3532 | CurDAG->getTargetConstant(Val: EncodedShiftImm, DL, VT)}; |
3533 | CurDAG->SelectNodeTo(N, MachineOpc: OrrOpc, VT, Ops); |
3534 | return true; |
3535 | } |
3536 | } |
3537 | return false; |
3538 | } |
3539 | |
3540 | assert((!BiggerPattern) && "BiggerPattern should be handled above" ); |
3541 | |
3542 | uint64_t ShlImm; |
3543 | if (isOpcWithIntImmediate(N: OrOpd0.getNode(), Opc: ISD::SHL, Imm&: ShlImm)) { |
3544 | if (OrOpd0.getOperand(i: 0) == Src && OrOpd0.hasOneUse()) { |
3545 | SDValue Ops[] = { |
3546 | Dst, Src, |
3547 | CurDAG->getTargetConstant( |
3548 | Val: AArch64_AM::getShifterImm(ST: AArch64_AM::LSL, Imm: ShlImm), DL, VT)}; |
3549 | CurDAG->SelectNodeTo(N, MachineOpc: OrrOpc, VT, Ops); |
3550 | return true; |
3551 | } |
3552 | |
3553 | // Select the following pattern to left-shifted operand rather than BFI. |
3554 | // %val1 = op .. |
3555 | // %val2 = shl %val1, #imm |
3556 | // %res = or %val1, %val2 |
3557 | // |
3558 | // If N is selected to be BFI, we know that |
3559 | // 1) OrOpd0 would be the operand from which extract bits (i.e., folded into |
3560 | // BFI) 2) OrOpd1 would be the destination operand (i.e., preserved) |
3561 | // |
3562 | // Instead of selecting N to BFI, fold OrOpd0 as a left shift directly. |
3563 | if (OrOpd0.getOperand(i: 0) == OrOpd1) { |
3564 | SDValue Ops[] = { |
3565 | OrOpd1, OrOpd1, |
3566 | CurDAG->getTargetConstant( |
3567 | Val: AArch64_AM::getShifterImm(ST: AArch64_AM::LSL, Imm: ShlImm), DL, VT)}; |
3568 | CurDAG->SelectNodeTo(N, MachineOpc: OrrOpc, VT, Ops); |
3569 | return true; |
3570 | } |
3571 | } |
3572 | |
3573 | uint64_t SrlImm; |
3574 | if (isOpcWithIntImmediate(N: OrOpd0.getNode(), Opc: ISD::SRL, Imm&: SrlImm)) { |
3575 | // Select the following pattern to right-shifted operand rather than BFXIL. |
3576 | // %val1 = op .. |
3577 | // %val2 = lshr %val1, #imm |
3578 | // %res = or %val1, %val2 |
3579 | // |
3580 | // If N is selected to be BFXIL, we know that |
3581 | // 1) OrOpd0 would be the operand from which extract bits (i.e., folded into |
3582 | // BFXIL) 2) OrOpd1 would be the destination operand (i.e., preserved) |
3583 | // |
3584 | // Instead of selecting N to BFXIL, fold OrOpd0 as a right shift directly. |
3585 | if (OrOpd0.getOperand(i: 0) == OrOpd1) { |
3586 | SDValue Ops[] = { |
3587 | OrOpd1, OrOpd1, |
3588 | CurDAG->getTargetConstant( |
3589 | Val: AArch64_AM::getShifterImm(ST: AArch64_AM::LSR, Imm: SrlImm), DL, VT)}; |
3590 | CurDAG->SelectNodeTo(N, MachineOpc: OrrOpc, VT, Ops); |
3591 | return true; |
3592 | } |
3593 | } |
3594 | |
3595 | return false; |
3596 | } |
3597 | |
3598 | static bool tryBitfieldInsertOpFromOr(SDNode *N, const APInt &UsefulBits, |
3599 | SelectionDAG *CurDAG) { |
3600 | assert(N->getOpcode() == ISD::OR && "Expect a OR operation" ); |
3601 | |
3602 | EVT VT = N->getValueType(ResNo: 0); |
3603 | if (VT != MVT::i32 && VT != MVT::i64) |
3604 | return false; |
3605 | |
3606 | unsigned BitWidth = VT.getSizeInBits(); |
3607 | |
3608 | // Because of simplify-demanded-bits in DAGCombine, involved masks may not |
3609 | // have the expected shape. Try to undo that. |
3610 | |
3611 | unsigned NumberOfIgnoredLowBits = UsefulBits.countr_zero(); |
3612 | unsigned NumberOfIgnoredHighBits = UsefulBits.countl_zero(); |
3613 | |
3614 | // Given a OR operation, check if we have the following pattern |
3615 | // ubfm c, b, imm, imm2 (or something that does the same jobs, see |
3616 | // isBitfieldExtractOp) |
3617 | // d = e & mask2 ; where mask is a binary sequence of 1..10..0 and |
3618 | // countTrailingZeros(mask2) == imm2 - imm + 1 |
3619 | // f = d | c |
3620 | // if yes, replace the OR instruction with: |
3621 | // f = BFM Opd0, Opd1, LSB, MSB ; where LSB = imm, and MSB = imm2 |
3622 | |
3623 | // OR is commutative, check all combinations of operand order and values of |
3624 | // BiggerPattern, i.e. |
3625 | // Opd0, Opd1, BiggerPattern=false |
3626 | // Opd1, Opd0, BiggerPattern=false |
3627 | // Opd0, Opd1, BiggerPattern=true |
3628 | // Opd1, Opd0, BiggerPattern=true |
3629 | // Several of these combinations may match, so check with BiggerPattern=false |
3630 | // first since that will produce better results by matching more instructions |
3631 | // and/or inserting fewer extra instructions. |
3632 | for (int I = 0; I < 4; ++I) { |
3633 | |
3634 | SDValue Dst, Src; |
3635 | unsigned ImmR, ImmS; |
3636 | bool BiggerPattern = I / 2; |
3637 | SDValue OrOpd0Val = N->getOperand(Num: I % 2); |
3638 | SDNode *OrOpd0 = OrOpd0Val.getNode(); |
3639 | SDValue OrOpd1Val = N->getOperand(Num: (I + 1) % 2); |
3640 | SDNode *OrOpd1 = OrOpd1Val.getNode(); |
3641 | |
3642 | unsigned BFXOpc; |
3643 | int DstLSB, Width; |
3644 | if (isBitfieldExtractOp(CurDAG, N: OrOpd0, Opc&: BFXOpc, Opd0&: Src, Immr&: ImmR, Imms&: ImmS, |
3645 | NumberOfIgnoredLowBits, BiggerPattern)) { |
3646 | // Check that the returned opcode is compatible with the pattern, |
3647 | // i.e., same type and zero extended (U and not S) |
3648 | if ((BFXOpc != AArch64::UBFMXri && VT == MVT::i64) || |
3649 | (BFXOpc != AArch64::UBFMWri && VT == MVT::i32)) |
3650 | continue; |
3651 | |
3652 | // Compute the width of the bitfield insertion |
3653 | DstLSB = 0; |
3654 | Width = ImmS - ImmR + 1; |
3655 | // FIXME: This constraint is to catch bitfield insertion we may |
3656 | // want to widen the pattern if we want to grab general bitfield |
3657 | // move case |
3658 | if (Width <= 0) |
3659 | continue; |
3660 | |
3661 | // If the mask on the insertee is correct, we have a BFXIL operation. We |
3662 | // can share the ImmR and ImmS values from the already-computed UBFM. |
3663 | } else if (isBitfieldPositioningOp(CurDAG, Op: OrOpd0Val, |
3664 | BiggerPattern, |
3665 | Src, DstLSB, Width)) { |
3666 | ImmR = (BitWidth - DstLSB) % BitWidth; |
3667 | ImmS = Width - 1; |
3668 | } else |
3669 | continue; |
3670 | |
3671 | // Check the second part of the pattern |
3672 | EVT VT = OrOpd1Val.getValueType(); |
3673 | assert((VT == MVT::i32 || VT == MVT::i64) && "unexpected OR operand" ); |
3674 | |
3675 | // Compute the Known Zero for the candidate of the first operand. |
3676 | // This allows to catch more general case than just looking for |
3677 | // AND with imm. Indeed, simplify-demanded-bits may have removed |
3678 | // the AND instruction because it proves it was useless. |
3679 | KnownBits Known = CurDAG->computeKnownBits(Op: OrOpd1Val); |
3680 | |
3681 | // Check if there is enough room for the second operand to appear |
3682 | // in the first one |
3683 | APInt BitsToBeInserted = |
3684 | APInt::getBitsSet(numBits: Known.getBitWidth(), loBit: DstLSB, hiBit: DstLSB + Width); |
3685 | |
3686 | if ((BitsToBeInserted & ~Known.Zero) != 0) |
3687 | continue; |
3688 | |
3689 | // Set the first operand |
3690 | uint64_t Imm; |
3691 | if (isOpcWithIntImmediate(N: OrOpd1, Opc: ISD::AND, Imm) && |
3692 | isBitfieldDstMask(DstMask: Imm, BitsToBeInserted, NumberOfIgnoredHighBits, VT)) |
3693 | // In that case, we can eliminate the AND |
3694 | Dst = OrOpd1->getOperand(Num: 0); |
3695 | else |
3696 | // Maybe the AND has been removed by simplify-demanded-bits |
3697 | // or is useful because it discards more bits |
3698 | Dst = OrOpd1Val; |
3699 | |
3700 | // Before selecting ISD::OR node to AArch64::BFM, see if an AArch64::ORR |
3701 | // with shifted operand is more efficient. |
3702 | if (tryOrrWithShift(N, OrOpd0: OrOpd0Val, OrOpd1: OrOpd1Val, Src, Dst, CurDAG, |
3703 | BiggerPattern)) |
3704 | return true; |
3705 | |
3706 | // both parts match |
3707 | SDLoc DL(N); |
3708 | SDValue Ops[] = {Dst, Src, CurDAG->getTargetConstant(Val: ImmR, DL, VT), |
3709 | CurDAG->getTargetConstant(Val: ImmS, DL, VT)}; |
3710 | unsigned Opc = (VT == MVT::i32) ? AArch64::BFMWri : AArch64::BFMXri; |
3711 | CurDAG->SelectNodeTo(N, MachineOpc: Opc, VT, Ops); |
3712 | return true; |
3713 | } |
3714 | |
3715 | // Generate a BFXIL from 'or (and X, Mask0Imm), (and Y, Mask1Imm)' iff |
3716 | // Mask0Imm and ~Mask1Imm are equivalent and one of the MaskImms is a shifted |
3717 | // mask (e.g., 0x000ffff0). |
3718 | uint64_t Mask0Imm, Mask1Imm; |
3719 | SDValue And0 = N->getOperand(Num: 0); |
3720 | SDValue And1 = N->getOperand(Num: 1); |
3721 | if (And0.hasOneUse() && And1.hasOneUse() && |
3722 | isOpcWithIntImmediate(N: And0.getNode(), Opc: ISD::AND, Imm&: Mask0Imm) && |
3723 | isOpcWithIntImmediate(N: And1.getNode(), Opc: ISD::AND, Imm&: Mask1Imm) && |
3724 | APInt(BitWidth, Mask0Imm) == ~APInt(BitWidth, Mask1Imm) && |
3725 | (isShiftedMask(Mask: Mask0Imm, VT) || isShiftedMask(Mask: Mask1Imm, VT))) { |
3726 | |
3727 | // ORR is commutative, so canonicalize to the form 'or (and X, Mask0Imm), |
3728 | // (and Y, Mask1Imm)' where Mask1Imm is the shifted mask masking off the |
3729 | // bits to be inserted. |
3730 | if (isShiftedMask(Mask: Mask0Imm, VT)) { |
3731 | std::swap(a&: And0, b&: And1); |
3732 | std::swap(a&: Mask0Imm, b&: Mask1Imm); |
3733 | } |
3734 | |
3735 | SDValue Src = And1->getOperand(Num: 0); |
3736 | SDValue Dst = And0->getOperand(Num: 0); |
3737 | unsigned LSB = llvm::countr_zero(Val: Mask1Imm); |
3738 | int Width = BitWidth - APInt(BitWidth, Mask0Imm).popcount(); |
3739 | |
3740 | // The BFXIL inserts the low-order bits from a source register, so right |
3741 | // shift the needed bits into place. |
3742 | SDLoc DL(N); |
3743 | unsigned ShiftOpc = (VT == MVT::i32) ? AArch64::UBFMWri : AArch64::UBFMXri; |
3744 | uint64_t LsrImm = LSB; |
3745 | if (Src->hasOneUse() && |
3746 | isOpcWithIntImmediate(N: Src.getNode(), Opc: ISD::SRL, Imm&: LsrImm) && |
3747 | (LsrImm + LSB) < BitWidth) { |
3748 | Src = Src->getOperand(Num: 0); |
3749 | LsrImm += LSB; |
3750 | } |
3751 | |
3752 | SDNode *LSR = CurDAG->getMachineNode( |
3753 | Opcode: ShiftOpc, dl: DL, VT, Op1: Src, Op2: CurDAG->getTargetConstant(Val: LsrImm, DL, VT), |
3754 | Op3: CurDAG->getTargetConstant(Val: BitWidth - 1, DL, VT)); |
3755 | |
3756 | // BFXIL is an alias of BFM, so translate to BFM operands. |
3757 | unsigned ImmR = (BitWidth - LSB) % BitWidth; |
3758 | unsigned ImmS = Width - 1; |
3759 | |
3760 | // Create the BFXIL instruction. |
3761 | SDValue Ops[] = {Dst, SDValue(LSR, 0), |
3762 | CurDAG->getTargetConstant(Val: ImmR, DL, VT), |
3763 | CurDAG->getTargetConstant(Val: ImmS, DL, VT)}; |
3764 | unsigned Opc = (VT == MVT::i32) ? AArch64::BFMWri : AArch64::BFMXri; |
3765 | CurDAG->SelectNodeTo(N, MachineOpc: Opc, VT, Ops); |
3766 | return true; |
3767 | } |
3768 | |
3769 | return false; |
3770 | } |
3771 | |
3772 | bool AArch64DAGToDAGISel::tryBitfieldInsertOp(SDNode *N) { |
3773 | if (N->getOpcode() != ISD::OR) |
3774 | return false; |
3775 | |
3776 | APInt NUsefulBits; |
3777 | getUsefulBits(Op: SDValue(N, 0), UsefulBits&: NUsefulBits); |
3778 | |
3779 | // If all bits are not useful, just return UNDEF. |
3780 | if (!NUsefulBits) { |
3781 | CurDAG->SelectNodeTo(N, MachineOpc: TargetOpcode::IMPLICIT_DEF, VT: N->getValueType(ResNo: 0)); |
3782 | return true; |
3783 | } |
3784 | |
3785 | if (tryBitfieldInsertOpFromOr(N, UsefulBits: NUsefulBits, CurDAG)) |
3786 | return true; |
3787 | |
3788 | return tryBitfieldInsertOpFromOrAndImm(N, CurDAG); |
3789 | } |
3790 | |
3791 | /// SelectBitfieldInsertInZeroOp - Match a UBFIZ instruction that is the |
3792 | /// equivalent of a left shift by a constant amount followed by an and masking |
3793 | /// out a contiguous set of bits. |
3794 | bool AArch64DAGToDAGISel::tryBitfieldInsertInZeroOp(SDNode *N) { |
3795 | if (N->getOpcode() != ISD::AND) |
3796 | return false; |
3797 | |
3798 | EVT VT = N->getValueType(ResNo: 0); |
3799 | if (VT != MVT::i32 && VT != MVT::i64) |
3800 | return false; |
3801 | |
3802 | SDValue Op0; |
3803 | int DstLSB, Width; |
3804 | if (!isBitfieldPositioningOp(CurDAG, Op: SDValue(N, 0), /*BiggerPattern=*/false, |
3805 | Src&: Op0, DstLSB, Width)) |
3806 | return false; |
3807 | |
3808 | // ImmR is the rotate right amount. |
3809 | unsigned ImmR = (VT.getSizeInBits() - DstLSB) % VT.getSizeInBits(); |
3810 | // ImmS is the most significant bit of the source to be moved. |
3811 | unsigned ImmS = Width - 1; |
3812 | |
3813 | SDLoc DL(N); |
3814 | SDValue Ops[] = {Op0, CurDAG->getTargetConstant(Val: ImmR, DL, VT), |
3815 | CurDAG->getTargetConstant(Val: ImmS, DL, VT)}; |
3816 | unsigned Opc = (VT == MVT::i32) ? AArch64::UBFMWri : AArch64::UBFMXri; |
3817 | CurDAG->SelectNodeTo(N, MachineOpc: Opc, VT, Ops); |
3818 | return true; |
3819 | } |
3820 | |
3821 | /// tryShiftAmountMod - Take advantage of built-in mod of shift amount in |
3822 | /// variable shift/rotate instructions. |
3823 | bool AArch64DAGToDAGISel::tryShiftAmountMod(SDNode *N) { |
3824 | EVT VT = N->getValueType(ResNo: 0); |
3825 | |
3826 | unsigned Opc; |
3827 | switch (N->getOpcode()) { |
3828 | case ISD::ROTR: |
3829 | Opc = (VT == MVT::i32) ? AArch64::RORVWr : AArch64::RORVXr; |
3830 | break; |
3831 | case ISD::SHL: |
3832 | Opc = (VT == MVT::i32) ? AArch64::LSLVWr : AArch64::LSLVXr; |
3833 | break; |
3834 | case ISD::SRL: |
3835 | Opc = (VT == MVT::i32) ? AArch64::LSRVWr : AArch64::LSRVXr; |
3836 | break; |
3837 | case ISD::SRA: |
3838 | Opc = (VT == MVT::i32) ? AArch64::ASRVWr : AArch64::ASRVXr; |
3839 | break; |
3840 | default: |
3841 | return false; |
3842 | } |
3843 | |
3844 | uint64_t Size; |
3845 | uint64_t Bits; |
3846 | if (VT == MVT::i32) { |
3847 | Bits = 5; |
3848 | Size = 32; |
3849 | } else if (VT == MVT::i64) { |
3850 | Bits = 6; |
3851 | Size = 64; |
3852 | } else |
3853 | return false; |
3854 | |
3855 | SDValue ShiftAmt = N->getOperand(Num: 1); |
3856 | SDLoc DL(N); |
3857 | SDValue NewShiftAmt; |
3858 | |
3859 | // Skip over an extend of the shift amount. |
3860 | if (ShiftAmt->getOpcode() == ISD::ZERO_EXTEND || |
3861 | ShiftAmt->getOpcode() == ISD::ANY_EXTEND) |
3862 | ShiftAmt = ShiftAmt->getOperand(Num: 0); |
3863 | |
3864 | if (ShiftAmt->getOpcode() == ISD::ADD || ShiftAmt->getOpcode() == ISD::SUB) { |
3865 | SDValue Add0 = ShiftAmt->getOperand(Num: 0); |
3866 | SDValue Add1 = ShiftAmt->getOperand(Num: 1); |
3867 | uint64_t Add0Imm; |
3868 | uint64_t Add1Imm; |
3869 | if (isIntImmediate(N: Add1, Imm&: Add1Imm) && (Add1Imm % Size == 0)) { |
3870 | // If we are shifting by X+/-N where N == 0 mod Size, then just shift by X |
3871 | // to avoid the ADD/SUB. |
3872 | NewShiftAmt = Add0; |
3873 | } else if (ShiftAmt->getOpcode() == ISD::SUB && |
3874 | isIntImmediate(N: Add0, Imm&: Add0Imm) && Add0Imm != 0 && |
3875 | (Add0Imm % Size == 0)) { |
3876 | // If we are shifting by N-X where N == 0 mod Size, then just shift by -X |
3877 | // to generate a NEG instead of a SUB from a constant. |
3878 | unsigned NegOpc; |
3879 | unsigned ZeroReg; |
3880 | EVT SubVT = ShiftAmt->getValueType(ResNo: 0); |
3881 | if (SubVT == MVT::i32) { |
3882 | NegOpc = AArch64::SUBWrr; |
3883 | ZeroReg = AArch64::WZR; |
3884 | } else { |
3885 | assert(SubVT == MVT::i64); |
3886 | NegOpc = AArch64::SUBXrr; |
3887 | ZeroReg = AArch64::XZR; |
3888 | } |
3889 | SDValue Zero = |
3890 | CurDAG->getCopyFromReg(Chain: CurDAG->getEntryNode(), dl: DL, Reg: ZeroReg, VT: SubVT); |
3891 | MachineSDNode *Neg = |
3892 | CurDAG->getMachineNode(Opcode: NegOpc, dl: DL, VT: SubVT, Op1: Zero, Op2: Add1); |
3893 | NewShiftAmt = SDValue(Neg, 0); |
3894 | } else if (ShiftAmt->getOpcode() == ISD::SUB && |
3895 | isIntImmediate(N: Add0, Imm&: Add0Imm) && (Add0Imm % Size == Size - 1)) { |
3896 | // If we are shifting by N-X where N == -1 mod Size, then just shift by ~X |
3897 | // to generate a NOT instead of a SUB from a constant. |
3898 | unsigned NotOpc; |
3899 | unsigned ZeroReg; |
3900 | EVT SubVT = ShiftAmt->getValueType(ResNo: 0); |
3901 | if (SubVT == MVT::i32) { |
3902 | NotOpc = AArch64::ORNWrr; |
3903 | ZeroReg = AArch64::WZR; |
3904 | } else { |
3905 | assert(SubVT == MVT::i64); |
3906 | NotOpc = AArch64::ORNXrr; |
3907 | ZeroReg = AArch64::XZR; |
3908 | } |
3909 | SDValue Zero = |
3910 | CurDAG->getCopyFromReg(Chain: CurDAG->getEntryNode(), dl: DL, Reg: ZeroReg, VT: SubVT); |
3911 | MachineSDNode *Not = |
3912 | CurDAG->getMachineNode(Opcode: NotOpc, dl: DL, VT: SubVT, Op1: Zero, Op2: Add1); |
3913 | NewShiftAmt = SDValue(Not, 0); |
3914 | } else |
3915 | return false; |
3916 | } else { |
3917 | // If the shift amount is masked with an AND, check that the mask covers the |
3918 | // bits that are implicitly ANDed off by the above opcodes and if so, skip |
3919 | // the AND. |
3920 | uint64_t MaskImm; |
3921 | if (!isOpcWithIntImmediate(N: ShiftAmt.getNode(), Opc: ISD::AND, Imm&: MaskImm) && |
3922 | !isOpcWithIntImmediate(N: ShiftAmt.getNode(), Opc: AArch64ISD::ANDS, Imm&: MaskImm)) |
3923 | return false; |
3924 | |
3925 | if ((unsigned)llvm::countr_one(Value: MaskImm) < Bits) |
3926 | return false; |
3927 | |
3928 | NewShiftAmt = ShiftAmt->getOperand(Num: 0); |
3929 | } |
3930 | |
3931 | // Narrow/widen the shift amount to match the size of the shift operation. |
3932 | if (VT == MVT::i32) |
3933 | NewShiftAmt = narrowIfNeeded(CurDAG, N: NewShiftAmt); |
3934 | else if (VT == MVT::i64 && NewShiftAmt->getValueType(ResNo: 0) == MVT::i32) { |
3935 | SDValue SubReg = CurDAG->getTargetConstant(Val: AArch64::sub_32, DL, VT: MVT::i32); |
3936 | MachineSDNode *Ext = CurDAG->getMachineNode( |
3937 | Opcode: AArch64::SUBREG_TO_REG, dl: DL, VT, |
3938 | Op1: CurDAG->getTargetConstant(Val: 0, DL, VT: MVT::i64), Op2: NewShiftAmt, Op3: SubReg); |
3939 | NewShiftAmt = SDValue(Ext, 0); |
3940 | } |
3941 | |
3942 | SDValue Ops[] = {N->getOperand(Num: 0), NewShiftAmt}; |
3943 | CurDAG->SelectNodeTo(N, MachineOpc: Opc, VT, Ops); |
3944 | return true; |
3945 | } |
3946 | |
3947 | static bool checkCVTFixedPointOperandWithFBits(SelectionDAG *CurDAG, SDValue N, |
3948 | SDValue &FixedPos, |
3949 | unsigned RegWidth, |
3950 | bool isReciprocal) { |
3951 | APFloat FVal(0.0); |
3952 | if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(Val&: N)) |
3953 | FVal = CN->getValueAPF(); |
3954 | else if (LoadSDNode *LN = dyn_cast<LoadSDNode>(Val&: N)) { |
3955 | // Some otherwise illegal constants are allowed in this case. |
3956 | if (LN->getOperand(Num: 1).getOpcode() != AArch64ISD::ADDlow || |
3957 | !isa<ConstantPoolSDNode>(Val: LN->getOperand(Num: 1)->getOperand(Num: 1))) |
3958 | return false; |
3959 | |
3960 | ConstantPoolSDNode *CN = |
3961 | dyn_cast<ConstantPoolSDNode>(Val: LN->getOperand(Num: 1)->getOperand(Num: 1)); |
3962 | FVal = cast<ConstantFP>(Val: CN->getConstVal())->getValueAPF(); |
3963 | } else |
3964 | return false; |
3965 | |
3966 | // An FCVT[SU] instruction performs: convertToInt(Val * 2^fbits) where fbits |
3967 | // is between 1 and 32 for a destination w-register, or 1 and 64 for an |
3968 | // x-register. |
3969 | // |
3970 | // By this stage, we've detected (fp_to_[su]int (fmul Val, THIS_NODE)) so we |
3971 | // want THIS_NODE to be 2^fbits. This is much easier to deal with using |
3972 | // integers. |
3973 | bool IsExact; |
3974 | |
3975 | if (isReciprocal) |
3976 | if (!FVal.getExactInverse(inv: &FVal)) |
3977 | return false; |
3978 | |
3979 | // fbits is between 1 and 64 in the worst-case, which means the fmul |
3980 | // could have 2^64 as an actual operand. Need 65 bits of precision. |
3981 | APSInt IntVal(65, true); |
3982 | FVal.convertToInteger(Result&: IntVal, RM: APFloat::rmTowardZero, IsExact: &IsExact); |
3983 | |
3984 | // N.b. isPowerOf2 also checks for > 0. |
3985 | if (!IsExact || !IntVal.isPowerOf2()) |
3986 | return false; |
3987 | unsigned FBits = IntVal.logBase2(); |
3988 | |
3989 | // Checks above should have guaranteed that we haven't lost information in |
3990 | // finding FBits, but it must still be in range. |
3991 | if (FBits == 0 || FBits > RegWidth) return false; |
3992 | |
3993 | FixedPos = CurDAG->getTargetConstant(Val: FBits, DL: SDLoc(N), VT: MVT::i32); |
3994 | return true; |
3995 | } |
3996 | |
3997 | bool AArch64DAGToDAGISel::SelectCVTFixedPosOperand(SDValue N, SDValue &FixedPos, |
3998 | unsigned RegWidth) { |
3999 | return checkCVTFixedPointOperandWithFBits(CurDAG, N, FixedPos, RegWidth, |
4000 | isReciprocal: false); |
4001 | } |
4002 | |
4003 | bool AArch64DAGToDAGISel::SelectCVTFixedPosRecipOperand(SDValue N, |
4004 | SDValue &FixedPos, |
4005 | unsigned RegWidth) { |
4006 | return checkCVTFixedPointOperandWithFBits(CurDAG, N, FixedPos, RegWidth, |
4007 | isReciprocal: true); |
4008 | } |
4009 | |
4010 | // Inspects a register string of the form o0:op1:CRn:CRm:op2 gets the fields |
4011 | // of the string and obtains the integer values from them and combines these |
4012 | // into a single value to be used in the MRS/MSR instruction. |
4013 | static int getIntOperandFromRegisterString(StringRef RegString) { |
4014 | SmallVector<StringRef, 5> Fields; |
4015 | RegString.split(A&: Fields, Separator: ':'); |
4016 | |
4017 | if (Fields.size() == 1) |
4018 | return -1; |
4019 | |
4020 | assert(Fields.size() == 5 |
4021 | && "Invalid number of fields in read register string" ); |
4022 | |
4023 | SmallVector<int, 5> Ops; |
4024 | bool AllIntFields = true; |
4025 | |
4026 | for (StringRef Field : Fields) { |
4027 | unsigned IntField; |
4028 | AllIntFields &= !Field.getAsInteger(Radix: 10, Result&: IntField); |
4029 | Ops.push_back(Elt: IntField); |
4030 | } |
4031 | |
4032 | assert(AllIntFields && |
4033 | "Unexpected non-integer value in special register string." ); |
4034 | (void)AllIntFields; |
4035 | |
4036 | // Need to combine the integer fields of the string into a single value |
4037 | // based on the bit encoding of MRS/MSR instruction. |
4038 | return (Ops[0] << 14) | (Ops[1] << 11) | (Ops[2] << 7) | |
4039 | (Ops[3] << 3) | (Ops[4]); |
4040 | } |
4041 | |
4042 | // Lower the read_register intrinsic to an MRS instruction node if the special |
4043 | // register string argument is either of the form detailed in the ALCE (the |
4044 | // form described in getIntOperandsFromRegisterString) or is a named register |
4045 | // known by the MRS SysReg mapper. |
4046 | bool AArch64DAGToDAGISel::tryReadRegister(SDNode *N) { |
4047 | const auto *MD = cast<MDNodeSDNode>(Val: N->getOperand(Num: 1)); |
4048 | const auto *RegString = cast<MDString>(Val: MD->getMD()->getOperand(I: 0)); |
4049 | SDLoc DL(N); |
4050 | |
4051 | bool ReadIs128Bit = N->getOpcode() == AArch64ISD::MRRS; |
4052 | |
4053 | unsigned Opcode64Bit = AArch64::MRS; |
4054 | int Imm = getIntOperandFromRegisterString(RegString: RegString->getString()); |
4055 | if (Imm == -1) { |
4056 | // No match, Use the sysreg mapper to map the remaining possible strings to |
4057 | // the value for the register to be used for the instruction operand. |
4058 | const auto *TheReg = |
4059 | AArch64SysReg::lookupSysRegByName(Name: RegString->getString()); |
4060 | if (TheReg && TheReg->Readable && |
4061 | TheReg->haveFeatures(ActiveFeatures: Subtarget->getFeatureBits())) |
4062 | Imm = TheReg->Encoding; |
4063 | else |
4064 | Imm = AArch64SysReg::parseGenericRegister(Name: RegString->getString()); |
4065 | |
4066 | if (Imm == -1) { |
4067 | // Still no match, see if this is "pc" or give up. |
4068 | if (!ReadIs128Bit && RegString->getString() == "pc" ) { |
4069 | Opcode64Bit = AArch64::ADR; |
4070 | Imm = 0; |
4071 | } else { |
4072 | return false; |
4073 | } |
4074 | } |
4075 | } |
4076 | |
4077 | SDValue InChain = N->getOperand(Num: 0); |
4078 | SDValue SysRegImm = CurDAG->getTargetConstant(Val: Imm, DL, VT: MVT::i32); |
4079 | if (!ReadIs128Bit) { |
4080 | CurDAG->SelectNodeTo(N, MachineOpc: Opcode64Bit, VT1: MVT::i64, VT2: MVT::Other /* Chain */, |
4081 | Ops: {SysRegImm, InChain}); |
4082 | } else { |
4083 | SDNode *MRRS = CurDAG->getMachineNode( |
4084 | Opcode: AArch64::MRRS, dl: DL, |
4085 | ResultTys: {MVT::Untyped /* XSeqPair */, MVT::Other /* Chain */}, |
4086 | Ops: {SysRegImm, InChain}); |
4087 | |
4088 | // Sysregs are not endian. The even register always contains the low half |
4089 | // of the register. |
4090 | SDValue Lo = CurDAG->getTargetExtractSubreg(SRIdx: AArch64::sube64, DL, VT: MVT::i64, |
4091 | Operand: SDValue(MRRS, 0)); |
4092 | SDValue Hi = CurDAG->getTargetExtractSubreg(SRIdx: AArch64::subo64, DL, VT: MVT::i64, |
4093 | Operand: SDValue(MRRS, 0)); |
4094 | SDValue OutChain = SDValue(MRRS, 1); |
4095 | |
4096 | ReplaceUses(F: SDValue(N, 0), T: Lo); |
4097 | ReplaceUses(F: SDValue(N, 1), T: Hi); |
4098 | ReplaceUses(F: SDValue(N, 2), T: OutChain); |
4099 | }; |
4100 | return true; |
4101 | } |
4102 | |
4103 | // Lower the write_register intrinsic to an MSR instruction node if the special |
4104 | // register string argument is either of the form detailed in the ALCE (the |
4105 | // form described in getIntOperandsFromRegisterString) or is a named register |
4106 | // known by the MSR SysReg mapper. |
4107 | bool AArch64DAGToDAGISel::tryWriteRegister(SDNode *N) { |
4108 | const auto *MD = cast<MDNodeSDNode>(Val: N->getOperand(Num: 1)); |
4109 | const auto *RegString = cast<MDString>(Val: MD->getMD()->getOperand(I: 0)); |
4110 | SDLoc DL(N); |
4111 | |
4112 | bool WriteIs128Bit = N->getOpcode() == AArch64ISD::MSRR; |
4113 | |
4114 | if (!WriteIs128Bit) { |
4115 | // Check if the register was one of those allowed as the pstatefield value |
4116 | // in the MSR (immediate) instruction. To accept the values allowed in the |
4117 | // pstatefield for the MSR (immediate) instruction, we also require that an |
4118 | // immediate value has been provided as an argument, we know that this is |
4119 | // the case as it has been ensured by semantic checking. |
4120 | auto trySelectPState = [&](auto PMapper, unsigned State) { |
4121 | if (PMapper) { |
4122 | assert(isa<ConstantSDNode>(N->getOperand(2)) && |
4123 | "Expected a constant integer expression." ); |
4124 | unsigned Reg = PMapper->Encoding; |
4125 | uint64_t Immed = N->getConstantOperandVal(Num: 2); |
4126 | CurDAG->SelectNodeTo( |
4127 | N, MachineOpc: State, VT: MVT::Other, Op1: CurDAG->getTargetConstant(Val: Reg, DL, VT: MVT::i32), |
4128 | Op2: CurDAG->getTargetConstant(Val: Immed, DL, VT: MVT::i16), Op3: N->getOperand(Num: 0)); |
4129 | return true; |
4130 | } |
4131 | return false; |
4132 | }; |
4133 | |
4134 | if (trySelectPState( |
4135 | AArch64PState::lookupPStateImm0_15ByName(Name: RegString->getString()), |
4136 | AArch64::MSRpstateImm4)) |
4137 | return true; |
4138 | if (trySelectPState( |
4139 | AArch64PState::lookupPStateImm0_1ByName(Name: RegString->getString()), |
4140 | AArch64::MSRpstateImm1)) |
4141 | return true; |
4142 | } |
4143 | |
4144 | int Imm = getIntOperandFromRegisterString(RegString: RegString->getString()); |
4145 | if (Imm == -1) { |
4146 | // Use the sysreg mapper to attempt to map the remaining possible strings |
4147 | // to the value for the register to be used for the MSR (register) |
4148 | // instruction operand. |
4149 | auto TheReg = AArch64SysReg::lookupSysRegByName(Name: RegString->getString()); |
4150 | if (TheReg && TheReg->Writeable && |
4151 | TheReg->haveFeatures(ActiveFeatures: Subtarget->getFeatureBits())) |
4152 | Imm = TheReg->Encoding; |
4153 | else |
4154 | Imm = AArch64SysReg::parseGenericRegister(Name: RegString->getString()); |
4155 | |
4156 | if (Imm == -1) |
4157 | return false; |
4158 | } |
4159 | |
4160 | SDValue InChain = N->getOperand(Num: 0); |
4161 | if (!WriteIs128Bit) { |
4162 | CurDAG->SelectNodeTo(N, MachineOpc: AArch64::MSR, VT: MVT::Other, |
4163 | Op1: CurDAG->getTargetConstant(Val: Imm, DL, VT: MVT::i32), |
4164 | Op2: N->getOperand(Num: 2), Op3: InChain); |
4165 | } else { |
4166 | // No endian swap. The lower half always goes into the even subreg, and the |
4167 | // higher half always into the odd supreg. |
4168 | SDNode *Pair = CurDAG->getMachineNode( |
4169 | Opcode: TargetOpcode::REG_SEQUENCE, dl: DL, VT: MVT::Untyped /* XSeqPair */, |
4170 | Ops: {CurDAG->getTargetConstant(Val: AArch64::XSeqPairsClassRegClass.getID(), DL, |
4171 | VT: MVT::i32), |
4172 | N->getOperand(Num: 2), |
4173 | CurDAG->getTargetConstant(Val: AArch64::sube64, DL, VT: MVT::i32), |
4174 | N->getOperand(Num: 3), |
4175 | CurDAG->getTargetConstant(Val: AArch64::subo64, DL, VT: MVT::i32)}); |
4176 | |
4177 | CurDAG->SelectNodeTo(N, MachineOpc: AArch64::MSRR, VT: MVT::Other, |
4178 | Op1: CurDAG->getTargetConstant(Val: Imm, DL, VT: MVT::i32), |
4179 | Op2: SDValue(Pair, 0), Op3: InChain); |
4180 | } |
4181 | |
4182 | return true; |
4183 | } |
4184 | |
4185 | /// We've got special pseudo-instructions for these |
4186 | bool AArch64DAGToDAGISel::SelectCMP_SWAP(SDNode *N) { |
4187 | unsigned Opcode; |
4188 | EVT MemTy = cast<MemSDNode>(Val: N)->getMemoryVT(); |
4189 | |
4190 | // Leave IR for LSE if subtarget supports it. |
4191 | if (Subtarget->hasLSE()) return false; |
4192 | |
4193 | if (MemTy == MVT::i8) |
4194 | Opcode = AArch64::CMP_SWAP_8; |
4195 | else if (MemTy == MVT::i16) |
4196 | Opcode = AArch64::CMP_SWAP_16; |
4197 | else if (MemTy == MVT::i32) |
4198 | Opcode = AArch64::CMP_SWAP_32; |
4199 | else if (MemTy == MVT::i64) |
4200 | Opcode = AArch64::CMP_SWAP_64; |
4201 | else |
4202 | llvm_unreachable("Unknown AtomicCmpSwap type" ); |
4203 | |
4204 | MVT RegTy = MemTy == MVT::i64 ? MVT::i64 : MVT::i32; |
4205 | SDValue Ops[] = {N->getOperand(Num: 1), N->getOperand(Num: 2), N->getOperand(Num: 3), |
4206 | N->getOperand(Num: 0)}; |
4207 | SDNode *CmpSwap = CurDAG->getMachineNode( |
4208 | Opcode, dl: SDLoc(N), |
4209 | VTs: CurDAG->getVTList(VT1: RegTy, VT2: MVT::i32, VT3: MVT::Other), Ops); |
4210 | |
4211 | MachineMemOperand *MemOp = cast<MemSDNode>(Val: N)->getMemOperand(); |
4212 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: CmpSwap), NewMemRefs: {MemOp}); |
4213 | |
4214 | ReplaceUses(F: SDValue(N, 0), T: SDValue(CmpSwap, 0)); |
4215 | ReplaceUses(F: SDValue(N, 1), T: SDValue(CmpSwap, 2)); |
4216 | CurDAG->RemoveDeadNode(N); |
4217 | |
4218 | return true; |
4219 | } |
4220 | |
4221 | bool AArch64DAGToDAGISel::SelectSVEAddSubImm(SDValue N, MVT VT, SDValue &Imm, |
4222 | SDValue &Shift) { |
4223 | if (!isa<ConstantSDNode>(Val: N)) |
4224 | return false; |
4225 | |
4226 | SDLoc DL(N); |
4227 | uint64_t Val = cast<ConstantSDNode>(Val&: N) |
4228 | ->getAPIntValue() |
4229 | .trunc(width: VT.getFixedSizeInBits()) |
4230 | .getZExtValue(); |
4231 | |
4232 | switch (VT.SimpleTy) { |
4233 | case MVT::i8: |
4234 | // All immediates are supported. |
4235 | Shift = CurDAG->getTargetConstant(Val: 0, DL, VT: MVT::i32); |
4236 | Imm = CurDAG->getTargetConstant(Val, DL, VT: MVT::i32); |
4237 | return true; |
4238 | case MVT::i16: |
4239 | case MVT::i32: |
4240 | case MVT::i64: |
4241 | // Support 8bit unsigned immediates. |
4242 | if (Val <= 255) { |
4243 | Shift = CurDAG->getTargetConstant(Val: 0, DL, VT: MVT::i32); |
4244 | Imm = CurDAG->getTargetConstant(Val, DL, VT: MVT::i32); |
4245 | return true; |
4246 | } |
4247 | // Support 16bit unsigned immediates that are a multiple of 256. |
4248 | if (Val <= 65280 && Val % 256 == 0) { |
4249 | Shift = CurDAG->getTargetConstant(Val: 8, DL, VT: MVT::i32); |
4250 | Imm = CurDAG->getTargetConstant(Val: Val >> 8, DL, VT: MVT::i32); |
4251 | return true; |
4252 | } |
4253 | break; |
4254 | default: |
4255 | break; |
4256 | } |
4257 | |
4258 | return false; |
4259 | } |
4260 | |
4261 | bool AArch64DAGToDAGISel::SelectSVEAddSubSSatImm(SDValue N, MVT VT, |
4262 | SDValue &Imm, SDValue &Shift, |
4263 | bool Negate) { |
4264 | if (!isa<ConstantSDNode>(Val: N)) |
4265 | return false; |
4266 | |
4267 | SDLoc DL(N); |
4268 | int64_t Val = cast<ConstantSDNode>(Val&: N) |
4269 | ->getAPIntValue() |
4270 | .trunc(width: VT.getFixedSizeInBits()) |
4271 | .getSExtValue(); |
4272 | |
4273 | if (Negate) |
4274 | Val = -Val; |
4275 | |
4276 | // Signed saturating instructions treat their immediate operand as unsigned, |
4277 | // whereas the related intrinsics define their operands to be signed. This |
4278 | // means we can only use the immediate form when the operand is non-negative. |
4279 | if (Val < 0) |
4280 | return false; |
4281 | |
4282 | switch (VT.SimpleTy) { |
4283 | case MVT::i8: |
4284 | // All positive immediates are supported. |
4285 | Shift = CurDAG->getTargetConstant(Val: 0, DL, VT: MVT::i32); |
4286 | Imm = CurDAG->getTargetConstant(Val, DL, VT: MVT::i32); |
4287 | return true; |
4288 | case MVT::i16: |
4289 | case MVT::i32: |
4290 | case MVT::i64: |
4291 | // Support 8bit positive immediates. |
4292 | if (Val <= 255) { |
4293 | Shift = CurDAG->getTargetConstant(Val: 0, DL, VT: MVT::i32); |
4294 | Imm = CurDAG->getTargetConstant(Val, DL, VT: MVT::i32); |
4295 | return true; |
4296 | } |
4297 | // Support 16bit positive immediates that are a multiple of 256. |
4298 | if (Val <= 65280 && Val % 256 == 0) { |
4299 | Shift = CurDAG->getTargetConstant(Val: 8, DL, VT: MVT::i32); |
4300 | Imm = CurDAG->getTargetConstant(Val: Val >> 8, DL, VT: MVT::i32); |
4301 | return true; |
4302 | } |
4303 | break; |
4304 | default: |
4305 | break; |
4306 | } |
4307 | |
4308 | return false; |
4309 | } |
4310 | |
4311 | bool AArch64DAGToDAGISel::SelectSVECpyDupImm(SDValue N, MVT VT, SDValue &Imm, |
4312 | SDValue &Shift) { |
4313 | if (!isa<ConstantSDNode>(Val: N)) |
4314 | return false; |
4315 | |
4316 | SDLoc DL(N); |
4317 | int64_t Val = cast<ConstantSDNode>(Val&: N) |
4318 | ->getAPIntValue() |
4319 | .trunc(width: VT.getFixedSizeInBits()) |
4320 | .getSExtValue(); |
4321 | |
4322 | switch (VT.SimpleTy) { |
4323 | case MVT::i8: |
4324 | // All immediates are supported. |
4325 | Shift = CurDAG->getTargetConstant(Val: 0, DL, VT: MVT::i32); |
4326 | Imm = CurDAG->getTargetConstant(Val: Val & 0xFF, DL, VT: MVT::i32); |
4327 | return true; |
4328 | case MVT::i16: |
4329 | case MVT::i32: |
4330 | case MVT::i64: |
4331 | // Support 8bit signed immediates. |
4332 | if (Val >= -128 && Val <= 127) { |
4333 | Shift = CurDAG->getTargetConstant(Val: 0, DL, VT: MVT::i32); |
4334 | Imm = CurDAG->getTargetConstant(Val: Val & 0xFF, DL, VT: MVT::i32); |
4335 | return true; |
4336 | } |
4337 | // Support 16bit signed immediates that are a multiple of 256. |
4338 | if (Val >= -32768 && Val <= 32512 && Val % 256 == 0) { |
4339 | Shift = CurDAG->getTargetConstant(Val: 8, DL, VT: MVT::i32); |
4340 | Imm = CurDAG->getTargetConstant(Val: (Val >> 8) & 0xFF, DL, VT: MVT::i32); |
4341 | return true; |
4342 | } |
4343 | break; |
4344 | default: |
4345 | break; |
4346 | } |
4347 | |
4348 | return false; |
4349 | } |
4350 | |
4351 | bool AArch64DAGToDAGISel::SelectSVESignedArithImm(SDValue N, SDValue &Imm) { |
4352 | if (auto CNode = dyn_cast<ConstantSDNode>(Val&: N)) { |
4353 | int64_t ImmVal = CNode->getSExtValue(); |
4354 | SDLoc DL(N); |
4355 | if (ImmVal >= -128 && ImmVal < 128) { |
4356 | Imm = CurDAG->getSignedTargetConstant(Val: ImmVal, DL, VT: MVT::i32); |
4357 | return true; |
4358 | } |
4359 | } |
4360 | return false; |
4361 | } |
4362 | |
4363 | bool AArch64DAGToDAGISel::SelectSVEArithImm(SDValue N, MVT VT, SDValue &Imm) { |
4364 | if (auto CNode = dyn_cast<ConstantSDNode>(Val&: N)) { |
4365 | uint64_t ImmVal = CNode->getZExtValue(); |
4366 | |
4367 | switch (VT.SimpleTy) { |
4368 | case MVT::i8: |
4369 | ImmVal &= 0xFF; |
4370 | break; |
4371 | case MVT::i16: |
4372 | ImmVal &= 0xFFFF; |
4373 | break; |
4374 | case MVT::i32: |
4375 | ImmVal &= 0xFFFFFFFF; |
4376 | break; |
4377 | case MVT::i64: |
4378 | break; |
4379 | default: |
4380 | llvm_unreachable("Unexpected type" ); |
4381 | } |
4382 | |
4383 | if (ImmVal < 256) { |
4384 | Imm = CurDAG->getTargetConstant(Val: ImmVal, DL: SDLoc(N), VT: MVT::i32); |
4385 | return true; |
4386 | } |
4387 | } |
4388 | return false; |
4389 | } |
4390 | |
4391 | bool AArch64DAGToDAGISel::SelectSVELogicalImm(SDValue N, MVT VT, SDValue &Imm, |
4392 | bool Invert) { |
4393 | if (auto CNode = dyn_cast<ConstantSDNode>(Val&: N)) { |
4394 | uint64_t ImmVal = CNode->getZExtValue(); |
4395 | SDLoc DL(N); |
4396 | |
4397 | if (Invert) |
4398 | ImmVal = ~ImmVal; |
4399 | |
4400 | // Shift mask depending on type size. |
4401 | switch (VT.SimpleTy) { |
4402 | case MVT::i8: |
4403 | ImmVal &= 0xFF; |
4404 | ImmVal |= ImmVal << 8; |
4405 | ImmVal |= ImmVal << 16; |
4406 | ImmVal |= ImmVal << 32; |
4407 | break; |
4408 | case MVT::i16: |
4409 | ImmVal &= 0xFFFF; |
4410 | ImmVal |= ImmVal << 16; |
4411 | ImmVal |= ImmVal << 32; |
4412 | break; |
4413 | case MVT::i32: |
4414 | ImmVal &= 0xFFFFFFFF; |
4415 | ImmVal |= ImmVal << 32; |
4416 | break; |
4417 | case MVT::i64: |
4418 | break; |
4419 | default: |
4420 | llvm_unreachable("Unexpected type" ); |
4421 | } |
4422 | |
4423 | uint64_t encoding; |
4424 | if (AArch64_AM::processLogicalImmediate(Imm: ImmVal, RegSize: 64, Encoding&: encoding)) { |
4425 | Imm = CurDAG->getTargetConstant(Val: encoding, DL, VT: MVT::i64); |
4426 | return true; |
4427 | } |
4428 | } |
4429 | return false; |
4430 | } |
4431 | |
4432 | // SVE shift intrinsics allow shift amounts larger than the element's bitwidth. |
4433 | // Rather than attempt to normalise everything we can sometimes saturate the |
4434 | // shift amount during selection. This function also allows for consistent |
4435 | // isel patterns by ensuring the resulting "Imm" node is of the i32 type |
4436 | // required by the instructions. |
4437 | bool AArch64DAGToDAGISel::SelectSVEShiftImm(SDValue N, uint64_t Low, |
4438 | uint64_t High, bool AllowSaturation, |
4439 | SDValue &Imm) { |
4440 | if (auto *CN = dyn_cast<ConstantSDNode>(Val&: N)) { |
4441 | uint64_t ImmVal = CN->getZExtValue(); |
4442 | |
4443 | // Reject shift amounts that are too small. |
4444 | if (ImmVal < Low) |
4445 | return false; |
4446 | |
4447 | // Reject or saturate shift amounts that are too big. |
4448 | if (ImmVal > High) { |
4449 | if (!AllowSaturation) |
4450 | return false; |
4451 | ImmVal = High; |
4452 | } |
4453 | |
4454 | Imm = CurDAG->getTargetConstant(Val: ImmVal, DL: SDLoc(N), VT: MVT::i32); |
4455 | return true; |
4456 | } |
4457 | |
4458 | return false; |
4459 | } |
4460 | |
4461 | bool AArch64DAGToDAGISel::trySelectStackSlotTagP(SDNode *N) { |
4462 | // tagp(FrameIndex, IRGstack, tag_offset): |
4463 | // since the offset between FrameIndex and IRGstack is a compile-time |
4464 | // constant, this can be lowered to a single ADDG instruction. |
4465 | if (!(isa<FrameIndexSDNode>(Val: N->getOperand(Num: 1)))) { |
4466 | return false; |
4467 | } |
4468 | |
4469 | SDValue IRG_SP = N->getOperand(Num: 2); |
4470 | if (IRG_SP->getOpcode() != ISD::INTRINSIC_W_CHAIN || |
4471 | IRG_SP->getConstantOperandVal(Num: 1) != Intrinsic::aarch64_irg_sp) { |
4472 | return false; |
4473 | } |
4474 | |
4475 | const TargetLowering *TLI = getTargetLowering(); |
4476 | SDLoc DL(N); |
4477 | int FI = cast<FrameIndexSDNode>(Val: N->getOperand(Num: 1))->getIndex(); |
4478 | SDValue FiOp = CurDAG->getTargetFrameIndex( |
4479 | FI, VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
4480 | int TagOffset = N->getConstantOperandVal(Num: 3); |
4481 | |
4482 | SDNode *Out = CurDAG->getMachineNode( |
4483 | Opcode: AArch64::TAGPstack, dl: DL, VT: MVT::i64, |
4484 | Ops: {FiOp, CurDAG->getTargetConstant(Val: 0, DL, VT: MVT::i64), N->getOperand(Num: 2), |
4485 | CurDAG->getTargetConstant(Val: TagOffset, DL, VT: MVT::i64)}); |
4486 | ReplaceNode(F: N, T: Out); |
4487 | return true; |
4488 | } |
4489 | |
4490 | void AArch64DAGToDAGISel::SelectTagP(SDNode *N) { |
4491 | assert(isa<ConstantSDNode>(N->getOperand(3)) && |
4492 | "llvm.aarch64.tagp third argument must be an immediate" ); |
4493 | if (trySelectStackSlotTagP(N)) |
4494 | return; |
4495 | // FIXME: above applies in any case when offset between Op1 and Op2 is a |
4496 | // compile-time constant, not just for stack allocations. |
4497 | |
4498 | // General case for unrelated pointers in Op1 and Op2. |
4499 | SDLoc DL(N); |
4500 | int TagOffset = N->getConstantOperandVal(Num: 3); |
4501 | SDNode *N1 = CurDAG->getMachineNode(Opcode: AArch64::SUBP, dl: DL, VT: MVT::i64, |
4502 | Ops: {N->getOperand(Num: 1), N->getOperand(Num: 2)}); |
4503 | SDNode *N2 = CurDAG->getMachineNode(Opcode: AArch64::ADDXrr, dl: DL, VT: MVT::i64, |
4504 | Ops: {SDValue(N1, 0), N->getOperand(Num: 2)}); |
4505 | SDNode *N3 = CurDAG->getMachineNode( |
4506 | Opcode: AArch64::ADDG, dl: DL, VT: MVT::i64, |
4507 | Ops: {SDValue(N2, 0), CurDAG->getTargetConstant(Val: 0, DL, VT: MVT::i64), |
4508 | CurDAG->getTargetConstant(Val: TagOffset, DL, VT: MVT::i64)}); |
4509 | ReplaceNode(F: N, T: N3); |
4510 | } |
4511 | |
4512 | bool AArch64DAGToDAGISel::trySelectCastFixedLengthToScalableVector(SDNode *N) { |
4513 | assert(N->getOpcode() == ISD::INSERT_SUBVECTOR && "Invalid Node!" ); |
4514 | |
4515 | // Bail when not a "cast" like insert_subvector. |
4516 | if (N->getConstantOperandVal(Num: 2) != 0) |
4517 | return false; |
4518 | if (!N->getOperand(Num: 0).isUndef()) |
4519 | return false; |
4520 | |
4521 | // Bail when normal isel should do the job. |
4522 | EVT VT = N->getValueType(ResNo: 0); |
4523 | EVT InVT = N->getOperand(Num: 1).getValueType(); |
4524 | if (VT.isFixedLengthVector() || InVT.isScalableVector()) |
4525 | return false; |
4526 | if (InVT.getSizeInBits() <= 128) |
4527 | return false; |
4528 | |
4529 | // NOTE: We can only get here when doing fixed length SVE code generation. |
4530 | // We do manual selection because the types involved are not linked to real |
4531 | // registers (despite being legal) and must be coerced into SVE registers. |
4532 | |
4533 | assert(VT.getSizeInBits().getKnownMinValue() == AArch64::SVEBitsPerBlock && |
4534 | "Expected to insert into a packed scalable vector!" ); |
4535 | |
4536 | SDLoc DL(N); |
4537 | auto RC = CurDAG->getTargetConstant(Val: AArch64::ZPRRegClassID, DL, VT: MVT::i64); |
4538 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: TargetOpcode::COPY_TO_REGCLASS, dl: DL, VT, |
4539 | Op1: N->getOperand(Num: 1), Op2: RC)); |
4540 | return true; |
4541 | } |
4542 | |
4543 | bool AArch64DAGToDAGISel::trySelectCastScalableToFixedLengthVector(SDNode *N) { |
4544 | assert(N->getOpcode() == ISD::EXTRACT_SUBVECTOR && "Invalid Node!" ); |
4545 | |
4546 | // Bail when not a "cast" like extract_subvector. |
4547 | if (N->getConstantOperandVal(Num: 1) != 0) |
4548 | return false; |
4549 | |
4550 | // Bail when normal isel can do the job. |
4551 | EVT VT = N->getValueType(ResNo: 0); |
4552 | EVT InVT = N->getOperand(Num: 0).getValueType(); |
4553 | if (VT.isScalableVector() || InVT.isFixedLengthVector()) |
4554 | return false; |
4555 | if (VT.getSizeInBits() <= 128) |
4556 | return false; |
4557 | |
4558 | // NOTE: We can only get here when doing fixed length SVE code generation. |
4559 | // We do manual selection because the types involved are not linked to real |
4560 | // registers (despite being legal) and must be coerced into SVE registers. |
4561 | |
4562 | assert(InVT.getSizeInBits().getKnownMinValue() == AArch64::SVEBitsPerBlock && |
4563 | "Expected to extract from a packed scalable vector!" ); |
4564 | |
4565 | SDLoc DL(N); |
4566 | auto RC = CurDAG->getTargetConstant(Val: AArch64::ZPRRegClassID, DL, VT: MVT::i64); |
4567 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: TargetOpcode::COPY_TO_REGCLASS, dl: DL, VT, |
4568 | Op1: N->getOperand(Num: 0), Op2: RC)); |
4569 | return true; |
4570 | } |
4571 | |
4572 | bool AArch64DAGToDAGISel::trySelectXAR(SDNode *N) { |
4573 | assert(N->getOpcode() == ISD::OR && "Expected OR instruction" ); |
4574 | |
4575 | SDValue N0 = N->getOperand(Num: 0); |
4576 | SDValue N1 = N->getOperand(Num: 1); |
4577 | |
4578 | EVT VT = N->getValueType(ResNo: 0); |
4579 | SDLoc DL(N); |
4580 | |
4581 | // Essentially: rotr (xor(x, y), imm) -> xar (x, y, imm) |
4582 | // Rotate by a constant is a funnel shift in IR which is exanded to |
4583 | // an OR with shifted operands. |
4584 | // We do the following transform: |
4585 | // OR N0, N1 -> xar (x, y, imm) |
4586 | // Where: |
4587 | // N1 = SRL_PRED true, V, splat(imm) --> rotr amount |
4588 | // N0 = SHL_PRED true, V, splat(bits-imm) |
4589 | // V = (xor x, y) |
4590 | if (VT.isScalableVector() && |
4591 | (Subtarget->hasSVE2() || |
4592 | (Subtarget->hasSME() && Subtarget->isStreaming()))) { |
4593 | if (N0.getOpcode() != AArch64ISD::SHL_PRED || |
4594 | N1.getOpcode() != AArch64ISD::SRL_PRED) |
4595 | std::swap(a&: N0, b&: N1); |
4596 | if (N0.getOpcode() != AArch64ISD::SHL_PRED || |
4597 | N1.getOpcode() != AArch64ISD::SRL_PRED) |
4598 | return false; |
4599 | |
4600 | auto *TLI = static_cast<const AArch64TargetLowering *>(getTargetLowering()); |
4601 | if (!TLI->isAllActivePredicate(DAG&: *CurDAG, N: N0.getOperand(i: 0)) || |
4602 | !TLI->isAllActivePredicate(DAG&: *CurDAG, N: N1.getOperand(i: 0))) |
4603 | return false; |
4604 | |
4605 | if (N0.getOperand(i: 1) != N1.getOperand(i: 1)) |
4606 | return false; |
4607 | |
4608 | SDValue R1, R2; |
4609 | bool IsXOROperand = true; |
4610 | if (N0.getOperand(i: 1).getOpcode() != ISD::XOR) { |
4611 | IsXOROperand = false; |
4612 | } else { |
4613 | R1 = N0.getOperand(i: 1).getOperand(i: 0); |
4614 | R2 = N1.getOperand(i: 1).getOperand(i: 1); |
4615 | } |
4616 | |
4617 | APInt ShlAmt, ShrAmt; |
4618 | if (!ISD::isConstantSplatVector(N: N0.getOperand(i: 2).getNode(), SplatValue&: ShlAmt) || |
4619 | !ISD::isConstantSplatVector(N: N1.getOperand(i: 2).getNode(), SplatValue&: ShrAmt)) |
4620 | return false; |
4621 | |
4622 | if (ShlAmt + ShrAmt != VT.getScalarSizeInBits()) |
4623 | return false; |
4624 | |
4625 | if (!IsXOROperand) { |
4626 | SDValue Zero = CurDAG->getTargetConstant(Val: 0, DL, VT: MVT::i64); |
4627 | SDNode *MOV = CurDAG->getMachineNode(Opcode: AArch64::MOVIv2d_ns, dl: DL, VT, Op1: Zero); |
4628 | SDValue MOVIV = SDValue(MOV, 0); |
4629 | |
4630 | SDValue ZSub = CurDAG->getTargetConstant(Val: AArch64::zsub, DL, VT: MVT::i32); |
4631 | SDNode *SubRegToReg = CurDAG->getMachineNode(Opcode: AArch64::SUBREG_TO_REG, dl: DL, |
4632 | VT, Op1: Zero, Op2: MOVIV, Op3: ZSub); |
4633 | |
4634 | R1 = N1->getOperand(Num: 1); |
4635 | R2 = SDValue(SubRegToReg, 0); |
4636 | } |
4637 | |
4638 | SDValue Imm = |
4639 | CurDAG->getTargetConstant(Val: ShrAmt.getZExtValue(), DL, VT: MVT::i32); |
4640 | |
4641 | SDValue Ops[] = {R1, R2, Imm}; |
4642 | if (auto Opc = SelectOpcodeFromVT<SelectTypeKind::Int>( |
4643 | VT, Opcodes: {AArch64::XAR_ZZZI_B, AArch64::XAR_ZZZI_H, AArch64::XAR_ZZZI_S, |
4644 | AArch64::XAR_ZZZI_D})) { |
4645 | CurDAG->SelectNodeTo(N, MachineOpc: Opc, VT, Ops); |
4646 | return true; |
4647 | } |
4648 | return false; |
4649 | } |
4650 | |
4651 | // We have Neon SHA3 XAR operation for v2i64 but for types |
4652 | // v4i32, v8i16, v16i8 we can use SVE operations when SVE2-SHA3 |
4653 | // is available. |
4654 | EVT SVT; |
4655 | switch (VT.getSimpleVT().SimpleTy) { |
4656 | case MVT::v4i32: |
4657 | case MVT::v2i32: |
4658 | SVT = MVT::nxv4i32; |
4659 | break; |
4660 | case MVT::v8i16: |
4661 | case MVT::v4i16: |
4662 | SVT = MVT::nxv8i16; |
4663 | break; |
4664 | case MVT::v16i8: |
4665 | case MVT::v8i8: |
4666 | SVT = MVT::nxv16i8; |
4667 | break; |
4668 | case MVT::v2i64: |
4669 | case MVT::v1i64: |
4670 | SVT = Subtarget->hasSHA3() ? MVT::v2i64 : MVT::nxv2i64; |
4671 | break; |
4672 | default: |
4673 | return false; |
4674 | } |
4675 | |
4676 | if ((!SVT.isScalableVector() && !Subtarget->hasSHA3()) || |
4677 | (SVT.isScalableVector() && !Subtarget->hasSVE2())) |
4678 | return false; |
4679 | |
4680 | if (N0->getOpcode() != AArch64ISD::VSHL || |
4681 | N1->getOpcode() != AArch64ISD::VLSHR) |
4682 | return false; |
4683 | |
4684 | if (N0->getOperand(Num: 0) != N1->getOperand(Num: 0)) |
4685 | return false; |
4686 | |
4687 | SDValue R1, R2; |
4688 | bool IsXOROperand = true; |
4689 | if (N1->getOperand(Num: 0)->getOpcode() != ISD::XOR) { |
4690 | IsXOROperand = false; |
4691 | } else { |
4692 | SDValue XOR = N0.getOperand(i: 0); |
4693 | R1 = XOR.getOperand(i: 0); |
4694 | R2 = XOR.getOperand(i: 1); |
4695 | } |
4696 | |
4697 | unsigned HsAmt = N0.getConstantOperandVal(i: 1); |
4698 | unsigned ShAmt = N1.getConstantOperandVal(i: 1); |
4699 | |
4700 | SDValue Imm = CurDAG->getTargetConstant( |
4701 | Val: ShAmt, DL, VT: N0.getOperand(i: 1).getValueType(), isOpaque: false); |
4702 | |
4703 | unsigned VTSizeInBits = VT.getScalarSizeInBits(); |
4704 | if (ShAmt + HsAmt != VTSizeInBits) |
4705 | return false; |
4706 | |
4707 | if (!IsXOROperand) { |
4708 | SDValue Zero = CurDAG->getTargetConstant(Val: 0, DL, VT: MVT::i64); |
4709 | SDNode *MOV = |
4710 | CurDAG->getMachineNode(Opcode: AArch64::MOVIv2d_ns, dl: DL, VT: MVT::v2i64, Op1: Zero); |
4711 | SDValue MOVIV = SDValue(MOV, 0); |
4712 | |
4713 | R1 = N1->getOperand(Num: 0); |
4714 | R2 = MOVIV; |
4715 | } |
4716 | |
4717 | if (SVT != VT) { |
4718 | SDValue Undef = |
4719 | SDValue(CurDAG->getMachineNode(Opcode: TargetOpcode::IMPLICIT_DEF, dl: DL, VT: SVT), 0); |
4720 | |
4721 | if (SVT.isScalableVector() && VT.is64BitVector()) { |
4722 | EVT QVT = VT.getDoubleNumVectorElementsVT(Context&: *CurDAG->getContext()); |
4723 | |
4724 | SDValue UndefQ = SDValue( |
4725 | CurDAG->getMachineNode(Opcode: TargetOpcode::IMPLICIT_DEF, dl: DL, VT: QVT), 0); |
4726 | SDValue DSub = CurDAG->getTargetConstant(Val: AArch64::dsub, DL, VT: MVT::i32); |
4727 | |
4728 | R1 = SDValue(CurDAG->getMachineNode(Opcode: AArch64::INSERT_SUBREG, dl: DL, VT: QVT, |
4729 | Op1: UndefQ, Op2: R1, Op3: DSub), |
4730 | 0); |
4731 | if (R2.getValueType() == VT) |
4732 | R2 = SDValue(CurDAG->getMachineNode(Opcode: AArch64::INSERT_SUBREG, dl: DL, VT: QVT, |
4733 | Op1: UndefQ, Op2: R2, Op3: DSub), |
4734 | 0); |
4735 | } |
4736 | |
4737 | SDValue SubReg = CurDAG->getTargetConstant( |
4738 | Val: (SVT.isScalableVector() ? AArch64::zsub : AArch64::dsub), DL, VT: MVT::i32); |
4739 | |
4740 | R1 = SDValue(CurDAG->getMachineNode(Opcode: AArch64::INSERT_SUBREG, dl: DL, VT: SVT, Op1: Undef, |
4741 | Op2: R1, Op3: SubReg), |
4742 | 0); |
4743 | |
4744 | if (SVT.isScalableVector() || R2.getValueType() != SVT) |
4745 | R2 = SDValue(CurDAG->getMachineNode(Opcode: AArch64::INSERT_SUBREG, dl: DL, VT: SVT, |
4746 | Op1: Undef, Op2: R2, Op3: SubReg), |
4747 | 0); |
4748 | } |
4749 | |
4750 | SDValue Ops[] = {R1, R2, Imm}; |
4751 | SDNode *XAR = nullptr; |
4752 | |
4753 | if (SVT.isScalableVector()) { |
4754 | if (auto Opc = SelectOpcodeFromVT<SelectTypeKind::Int>( |
4755 | VT: SVT, Opcodes: {AArch64::XAR_ZZZI_B, AArch64::XAR_ZZZI_H, AArch64::XAR_ZZZI_S, |
4756 | AArch64::XAR_ZZZI_D})) |
4757 | XAR = CurDAG->getMachineNode(Opcode: Opc, dl: DL, VT: SVT, Ops); |
4758 | } else { |
4759 | XAR = CurDAG->getMachineNode(Opcode: AArch64::XAR, dl: DL, VT: SVT, Ops); |
4760 | } |
4761 | |
4762 | assert(XAR && "Unexpected NULL value for XAR instruction in DAG" ); |
4763 | |
4764 | if (SVT != VT) { |
4765 | if (VT.is64BitVector() && SVT.isScalableVector()) { |
4766 | EVT QVT = VT.getDoubleNumVectorElementsVT(Context&: *CurDAG->getContext()); |
4767 | |
4768 | SDValue ZSub = CurDAG->getTargetConstant(Val: AArch64::zsub, DL, VT: MVT::i32); |
4769 | SDNode *Q = CurDAG->getMachineNode(Opcode: AArch64::EXTRACT_SUBREG, dl: DL, VT: QVT, |
4770 | Op1: SDValue(XAR, 0), Op2: ZSub); |
4771 | |
4772 | SDValue DSub = CurDAG->getTargetConstant(Val: AArch64::dsub, DL, VT: MVT::i32); |
4773 | XAR = CurDAG->getMachineNode(Opcode: AArch64::EXTRACT_SUBREG, dl: DL, VT, |
4774 | Op1: SDValue(Q, 0), Op2: DSub); |
4775 | } else { |
4776 | SDValue SubReg = CurDAG->getTargetConstant( |
4777 | Val: (SVT.isScalableVector() ? AArch64::zsub : AArch64::dsub), DL, |
4778 | VT: MVT::i32); |
4779 | XAR = CurDAG->getMachineNode(Opcode: AArch64::EXTRACT_SUBREG, dl: DL, VT, |
4780 | Op1: SDValue(XAR, 0), Op2: SubReg); |
4781 | } |
4782 | } |
4783 | ReplaceNode(F: N, T: XAR); |
4784 | return true; |
4785 | } |
4786 | |
4787 | void AArch64DAGToDAGISel::Select(SDNode *Node) { |
4788 | // If we have a custom node, we already have selected! |
4789 | if (Node->isMachineOpcode()) { |
4790 | LLVM_DEBUG(errs() << "== " ; Node->dump(CurDAG); errs() << "\n" ); |
4791 | Node->setNodeId(-1); |
4792 | return; |
4793 | } |
4794 | |
4795 | // Few custom selection stuff. |
4796 | EVT VT = Node->getValueType(ResNo: 0); |
4797 | |
4798 | switch (Node->getOpcode()) { |
4799 | default: |
4800 | break; |
4801 | |
4802 | case ISD::ATOMIC_CMP_SWAP: |
4803 | if (SelectCMP_SWAP(N: Node)) |
4804 | return; |
4805 | break; |
4806 | |
4807 | case ISD::READ_REGISTER: |
4808 | case AArch64ISD::MRRS: |
4809 | if (tryReadRegister(N: Node)) |
4810 | return; |
4811 | break; |
4812 | |
4813 | case ISD::WRITE_REGISTER: |
4814 | case AArch64ISD::MSRR: |
4815 | if (tryWriteRegister(N: Node)) |
4816 | return; |
4817 | break; |
4818 | |
4819 | case ISD::LOAD: { |
4820 | // Try to select as an indexed load. Fall through to normal processing |
4821 | // if we can't. |
4822 | if (tryIndexedLoad(N: Node)) |
4823 | return; |
4824 | break; |
4825 | } |
4826 | |
4827 | case ISD::SRL: |
4828 | case ISD::AND: |
4829 | case ISD::SRA: |
4830 | case ISD::SIGN_EXTEND_INREG: |
4831 | if (tryBitfieldExtractOp(N: Node)) |
4832 | return; |
4833 | if (tryBitfieldInsertInZeroOp(N: Node)) |
4834 | return; |
4835 | [[fallthrough]]; |
4836 | case ISD::ROTR: |
4837 | case ISD::SHL: |
4838 | if (tryShiftAmountMod(N: Node)) |
4839 | return; |
4840 | break; |
4841 | |
4842 | case ISD::SIGN_EXTEND: |
4843 | if (tryBitfieldExtractOpFromSExt(N: Node)) |
4844 | return; |
4845 | break; |
4846 | |
4847 | case ISD::OR: |
4848 | if (tryBitfieldInsertOp(N: Node)) |
4849 | return; |
4850 | if (trySelectXAR(N: Node)) |
4851 | return; |
4852 | break; |
4853 | |
4854 | case ISD::EXTRACT_SUBVECTOR: { |
4855 | if (trySelectCastScalableToFixedLengthVector(N: Node)) |
4856 | return; |
4857 | break; |
4858 | } |
4859 | |
4860 | case ISD::INSERT_SUBVECTOR: { |
4861 | if (trySelectCastFixedLengthToScalableVector(N: Node)) |
4862 | return; |
4863 | break; |
4864 | } |
4865 | |
4866 | case ISD::Constant: { |
4867 | // Materialize zero constants as copies from WZR/XZR. This allows |
4868 | // the coalescer to propagate these into other instructions. |
4869 | ConstantSDNode *ConstNode = cast<ConstantSDNode>(Val: Node); |
4870 | if (ConstNode->isZero()) { |
4871 | if (VT == MVT::i32) { |
4872 | SDValue New = CurDAG->getCopyFromReg( |
4873 | Chain: CurDAG->getEntryNode(), dl: SDLoc(Node), Reg: AArch64::WZR, VT: MVT::i32); |
4874 | ReplaceNode(F: Node, T: New.getNode()); |
4875 | return; |
4876 | } else if (VT == MVT::i64) { |
4877 | SDValue New = CurDAG->getCopyFromReg( |
4878 | Chain: CurDAG->getEntryNode(), dl: SDLoc(Node), Reg: AArch64::XZR, VT: MVT::i64); |
4879 | ReplaceNode(F: Node, T: New.getNode()); |
4880 | return; |
4881 | } |
4882 | } |
4883 | break; |
4884 | } |
4885 | |
4886 | case ISD::FrameIndex: { |
4887 | // Selects to ADDXri FI, 0 which in turn will become ADDXri SP, imm. |
4888 | int FI = cast<FrameIndexSDNode>(Val: Node)->getIndex(); |
4889 | unsigned Shifter = AArch64_AM::getShifterImm(ST: AArch64_AM::LSL, Imm: 0); |
4890 | const TargetLowering *TLI = getTargetLowering(); |
4891 | SDValue TFI = CurDAG->getTargetFrameIndex( |
4892 | FI, VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
4893 | SDLoc DL(Node); |
4894 | SDValue Ops[] = { TFI, CurDAG->getTargetConstant(Val: 0, DL, VT: MVT::i32), |
4895 | CurDAG->getTargetConstant(Val: Shifter, DL, VT: MVT::i32) }; |
4896 | CurDAG->SelectNodeTo(N: Node, MachineOpc: AArch64::ADDXri, VT: MVT::i64, Ops); |
4897 | return; |
4898 | } |
4899 | case ISD::INTRINSIC_W_CHAIN: { |
4900 | unsigned IntNo = Node->getConstantOperandVal(Num: 1); |
4901 | switch (IntNo) { |
4902 | default: |
4903 | break; |
4904 | case Intrinsic::aarch64_gcsss: { |
4905 | SDLoc DL(Node); |
4906 | SDValue Chain = Node->getOperand(Num: 0); |
4907 | SDValue Val = Node->getOperand(Num: 2); |
4908 | SDValue Zero = CurDAG->getCopyFromReg(Chain, dl: DL, Reg: AArch64::XZR, VT: MVT::i64); |
4909 | SDNode *SS1 = |
4910 | CurDAG->getMachineNode(Opcode: AArch64::GCSSS1, dl: DL, VT: MVT::Other, Op1: Val, Op2: Chain); |
4911 | SDNode *SS2 = CurDAG->getMachineNode(Opcode: AArch64::GCSSS2, dl: DL, VT1: MVT::i64, |
4912 | VT2: MVT::Other, Op1: Zero, Op2: SDValue(SS1, 0)); |
4913 | ReplaceNode(F: Node, T: SS2); |
4914 | return; |
4915 | } |
4916 | case Intrinsic::aarch64_ldaxp: |
4917 | case Intrinsic::aarch64_ldxp: { |
4918 | unsigned Op = |
4919 | IntNo == Intrinsic::aarch64_ldaxp ? AArch64::LDAXPX : AArch64::LDXPX; |
4920 | SDValue MemAddr = Node->getOperand(Num: 2); |
4921 | SDLoc DL(Node); |
4922 | SDValue Chain = Node->getOperand(Num: 0); |
4923 | |
4924 | SDNode *Ld = CurDAG->getMachineNode(Opcode: Op, dl: DL, VT1: MVT::i64, VT2: MVT::i64, |
4925 | VT3: MVT::Other, Op1: MemAddr, Op2: Chain); |
4926 | |
4927 | // Transfer memoperands. |
4928 | MachineMemOperand *MemOp = |
4929 | cast<MemIntrinsicSDNode>(Val: Node)->getMemOperand(); |
4930 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: Ld), NewMemRefs: {MemOp}); |
4931 | ReplaceNode(F: Node, T: Ld); |
4932 | return; |
4933 | } |
4934 | case Intrinsic::aarch64_stlxp: |
4935 | case Intrinsic::aarch64_stxp: { |
4936 | unsigned Op = |
4937 | IntNo == Intrinsic::aarch64_stlxp ? AArch64::STLXPX : AArch64::STXPX; |
4938 | SDLoc DL(Node); |
4939 | SDValue Chain = Node->getOperand(Num: 0); |
4940 | SDValue ValLo = Node->getOperand(Num: 2); |
4941 | SDValue ValHi = Node->getOperand(Num: 3); |
4942 | SDValue MemAddr = Node->getOperand(Num: 4); |
4943 | |
4944 | // Place arguments in the right order. |
4945 | SDValue Ops[] = {ValLo, ValHi, MemAddr, Chain}; |
4946 | |
4947 | SDNode *St = CurDAG->getMachineNode(Opcode: Op, dl: DL, VT1: MVT::i32, VT2: MVT::Other, Ops); |
4948 | // Transfer memoperands. |
4949 | MachineMemOperand *MemOp = |
4950 | cast<MemIntrinsicSDNode>(Val: Node)->getMemOperand(); |
4951 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: St), NewMemRefs: {MemOp}); |
4952 | |
4953 | ReplaceNode(F: Node, T: St); |
4954 | return; |
4955 | } |
4956 | case Intrinsic::aarch64_neon_ld1x2: |
4957 | if (VT == MVT::v8i8) { |
4958 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD1Twov8b, SubRegIdx: AArch64::dsub0); |
4959 | return; |
4960 | } else if (VT == MVT::v16i8) { |
4961 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD1Twov16b, SubRegIdx: AArch64::qsub0); |
4962 | return; |
4963 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
4964 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD1Twov4h, SubRegIdx: AArch64::dsub0); |
4965 | return; |
4966 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
4967 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD1Twov8h, SubRegIdx: AArch64::qsub0); |
4968 | return; |
4969 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
4970 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD1Twov2s, SubRegIdx: AArch64::dsub0); |
4971 | return; |
4972 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
4973 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD1Twov4s, SubRegIdx: AArch64::qsub0); |
4974 | return; |
4975 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
4976 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD1Twov1d, SubRegIdx: AArch64::dsub0); |
4977 | return; |
4978 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
4979 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD1Twov2d, SubRegIdx: AArch64::qsub0); |
4980 | return; |
4981 | } |
4982 | break; |
4983 | case Intrinsic::aarch64_neon_ld1x3: |
4984 | if (VT == MVT::v8i8) { |
4985 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD1Threev8b, SubRegIdx: AArch64::dsub0); |
4986 | return; |
4987 | } else if (VT == MVT::v16i8) { |
4988 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD1Threev16b, SubRegIdx: AArch64::qsub0); |
4989 | return; |
4990 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
4991 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD1Threev4h, SubRegIdx: AArch64::dsub0); |
4992 | return; |
4993 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
4994 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD1Threev8h, SubRegIdx: AArch64::qsub0); |
4995 | return; |
4996 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
4997 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD1Threev2s, SubRegIdx: AArch64::dsub0); |
4998 | return; |
4999 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
5000 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD1Threev4s, SubRegIdx: AArch64::qsub0); |
5001 | return; |
5002 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
5003 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD1Threev1d, SubRegIdx: AArch64::dsub0); |
5004 | return; |
5005 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
5006 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD1Threev2d, SubRegIdx: AArch64::qsub0); |
5007 | return; |
5008 | } |
5009 | break; |
5010 | case Intrinsic::aarch64_neon_ld1x4: |
5011 | if (VT == MVT::v8i8) { |
5012 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD1Fourv8b, SubRegIdx: AArch64::dsub0); |
5013 | return; |
5014 | } else if (VT == MVT::v16i8) { |
5015 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD1Fourv16b, SubRegIdx: AArch64::qsub0); |
5016 | return; |
5017 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
5018 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD1Fourv4h, SubRegIdx: AArch64::dsub0); |
5019 | return; |
5020 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
5021 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD1Fourv8h, SubRegIdx: AArch64::qsub0); |
5022 | return; |
5023 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
5024 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD1Fourv2s, SubRegIdx: AArch64::dsub0); |
5025 | return; |
5026 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
5027 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD1Fourv4s, SubRegIdx: AArch64::qsub0); |
5028 | return; |
5029 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
5030 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD1Fourv1d, SubRegIdx: AArch64::dsub0); |
5031 | return; |
5032 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
5033 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD1Fourv2d, SubRegIdx: AArch64::qsub0); |
5034 | return; |
5035 | } |
5036 | break; |
5037 | case Intrinsic::aarch64_neon_ld2: |
5038 | if (VT == MVT::v8i8) { |
5039 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Twov8b, SubRegIdx: AArch64::dsub0); |
5040 | return; |
5041 | } else if (VT == MVT::v16i8) { |
5042 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Twov16b, SubRegIdx: AArch64::qsub0); |
5043 | return; |
5044 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
5045 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Twov4h, SubRegIdx: AArch64::dsub0); |
5046 | return; |
5047 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
5048 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Twov8h, SubRegIdx: AArch64::qsub0); |
5049 | return; |
5050 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
5051 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Twov2s, SubRegIdx: AArch64::dsub0); |
5052 | return; |
5053 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
5054 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Twov4s, SubRegIdx: AArch64::qsub0); |
5055 | return; |
5056 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
5057 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD1Twov1d, SubRegIdx: AArch64::dsub0); |
5058 | return; |
5059 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
5060 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Twov2d, SubRegIdx: AArch64::qsub0); |
5061 | return; |
5062 | } |
5063 | break; |
5064 | case Intrinsic::aarch64_neon_ld3: |
5065 | if (VT == MVT::v8i8) { |
5066 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Threev8b, SubRegIdx: AArch64::dsub0); |
5067 | return; |
5068 | } else if (VT == MVT::v16i8) { |
5069 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Threev16b, SubRegIdx: AArch64::qsub0); |
5070 | return; |
5071 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
5072 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Threev4h, SubRegIdx: AArch64::dsub0); |
5073 | return; |
5074 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
5075 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Threev8h, SubRegIdx: AArch64::qsub0); |
5076 | return; |
5077 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
5078 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Threev2s, SubRegIdx: AArch64::dsub0); |
5079 | return; |
5080 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
5081 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Threev4s, SubRegIdx: AArch64::qsub0); |
5082 | return; |
5083 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
5084 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD1Threev1d, SubRegIdx: AArch64::dsub0); |
5085 | return; |
5086 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
5087 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Threev2d, SubRegIdx: AArch64::qsub0); |
5088 | return; |
5089 | } |
5090 | break; |
5091 | case Intrinsic::aarch64_neon_ld4: |
5092 | if (VT == MVT::v8i8) { |
5093 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Fourv8b, SubRegIdx: AArch64::dsub0); |
5094 | return; |
5095 | } else if (VT == MVT::v16i8) { |
5096 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Fourv16b, SubRegIdx: AArch64::qsub0); |
5097 | return; |
5098 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
5099 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Fourv4h, SubRegIdx: AArch64::dsub0); |
5100 | return; |
5101 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
5102 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Fourv8h, SubRegIdx: AArch64::qsub0); |
5103 | return; |
5104 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
5105 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Fourv2s, SubRegIdx: AArch64::dsub0); |
5106 | return; |
5107 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
5108 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Fourv4s, SubRegIdx: AArch64::qsub0); |
5109 | return; |
5110 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
5111 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD1Fourv1d, SubRegIdx: AArch64::dsub0); |
5112 | return; |
5113 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
5114 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Fourv2d, SubRegIdx: AArch64::qsub0); |
5115 | return; |
5116 | } |
5117 | break; |
5118 | case Intrinsic::aarch64_neon_ld2r: |
5119 | if (VT == MVT::v8i8) { |
5120 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Rv8b, SubRegIdx: AArch64::dsub0); |
5121 | return; |
5122 | } else if (VT == MVT::v16i8) { |
5123 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Rv16b, SubRegIdx: AArch64::qsub0); |
5124 | return; |
5125 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
5126 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Rv4h, SubRegIdx: AArch64::dsub0); |
5127 | return; |
5128 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
5129 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Rv8h, SubRegIdx: AArch64::qsub0); |
5130 | return; |
5131 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
5132 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Rv2s, SubRegIdx: AArch64::dsub0); |
5133 | return; |
5134 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
5135 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Rv4s, SubRegIdx: AArch64::qsub0); |
5136 | return; |
5137 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
5138 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Rv1d, SubRegIdx: AArch64::dsub0); |
5139 | return; |
5140 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
5141 | SelectLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Rv2d, SubRegIdx: AArch64::qsub0); |
5142 | return; |
5143 | } |
5144 | break; |
5145 | case Intrinsic::aarch64_neon_ld3r: |
5146 | if (VT == MVT::v8i8) { |
5147 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Rv8b, SubRegIdx: AArch64::dsub0); |
5148 | return; |
5149 | } else if (VT == MVT::v16i8) { |
5150 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Rv16b, SubRegIdx: AArch64::qsub0); |
5151 | return; |
5152 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
5153 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Rv4h, SubRegIdx: AArch64::dsub0); |
5154 | return; |
5155 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
5156 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Rv8h, SubRegIdx: AArch64::qsub0); |
5157 | return; |
5158 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
5159 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Rv2s, SubRegIdx: AArch64::dsub0); |
5160 | return; |
5161 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
5162 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Rv4s, SubRegIdx: AArch64::qsub0); |
5163 | return; |
5164 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
5165 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Rv1d, SubRegIdx: AArch64::dsub0); |
5166 | return; |
5167 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
5168 | SelectLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Rv2d, SubRegIdx: AArch64::qsub0); |
5169 | return; |
5170 | } |
5171 | break; |
5172 | case Intrinsic::aarch64_neon_ld4r: |
5173 | if (VT == MVT::v8i8) { |
5174 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Rv8b, SubRegIdx: AArch64::dsub0); |
5175 | return; |
5176 | } else if (VT == MVT::v16i8) { |
5177 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Rv16b, SubRegIdx: AArch64::qsub0); |
5178 | return; |
5179 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
5180 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Rv4h, SubRegIdx: AArch64::dsub0); |
5181 | return; |
5182 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
5183 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Rv8h, SubRegIdx: AArch64::qsub0); |
5184 | return; |
5185 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
5186 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Rv2s, SubRegIdx: AArch64::dsub0); |
5187 | return; |
5188 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
5189 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Rv4s, SubRegIdx: AArch64::qsub0); |
5190 | return; |
5191 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
5192 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Rv1d, SubRegIdx: AArch64::dsub0); |
5193 | return; |
5194 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
5195 | SelectLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Rv2d, SubRegIdx: AArch64::qsub0); |
5196 | return; |
5197 | } |
5198 | break; |
5199 | case Intrinsic::aarch64_neon_ld2lane: |
5200 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
5201 | SelectLoadLane(N: Node, NumVecs: 2, Opc: AArch64::LD2i8); |
5202 | return; |
5203 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
5204 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
5205 | SelectLoadLane(N: Node, NumVecs: 2, Opc: AArch64::LD2i16); |
5206 | return; |
5207 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
5208 | VT == MVT::v2f32) { |
5209 | SelectLoadLane(N: Node, NumVecs: 2, Opc: AArch64::LD2i32); |
5210 | return; |
5211 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
5212 | VT == MVT::v1f64) { |
5213 | SelectLoadLane(N: Node, NumVecs: 2, Opc: AArch64::LD2i64); |
5214 | return; |
5215 | } |
5216 | break; |
5217 | case Intrinsic::aarch64_neon_ld3lane: |
5218 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
5219 | SelectLoadLane(N: Node, NumVecs: 3, Opc: AArch64::LD3i8); |
5220 | return; |
5221 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
5222 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
5223 | SelectLoadLane(N: Node, NumVecs: 3, Opc: AArch64::LD3i16); |
5224 | return; |
5225 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
5226 | VT == MVT::v2f32) { |
5227 | SelectLoadLane(N: Node, NumVecs: 3, Opc: AArch64::LD3i32); |
5228 | return; |
5229 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
5230 | VT == MVT::v1f64) { |
5231 | SelectLoadLane(N: Node, NumVecs: 3, Opc: AArch64::LD3i64); |
5232 | return; |
5233 | } |
5234 | break; |
5235 | case Intrinsic::aarch64_neon_ld4lane: |
5236 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
5237 | SelectLoadLane(N: Node, NumVecs: 4, Opc: AArch64::LD4i8); |
5238 | return; |
5239 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
5240 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
5241 | SelectLoadLane(N: Node, NumVecs: 4, Opc: AArch64::LD4i16); |
5242 | return; |
5243 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
5244 | VT == MVT::v2f32) { |
5245 | SelectLoadLane(N: Node, NumVecs: 4, Opc: AArch64::LD4i32); |
5246 | return; |
5247 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
5248 | VT == MVT::v1f64) { |
5249 | SelectLoadLane(N: Node, NumVecs: 4, Opc: AArch64::LD4i64); |
5250 | return; |
5251 | } |
5252 | break; |
5253 | case Intrinsic::aarch64_ld64b: |
5254 | SelectLoad(N: Node, NumVecs: 8, Opc: AArch64::LD64B, SubRegIdx: AArch64::x8sub_0); |
5255 | return; |
5256 | case Intrinsic::aarch64_sve_ld2q_sret: { |
5257 | SelectPredicatedLoad(N: Node, NumVecs: 2, Scale: 4, Opc_ri: AArch64::LD2Q_IMM, Opc_rr: AArch64::LD2Q, IsIntr: true); |
5258 | return; |
5259 | } |
5260 | case Intrinsic::aarch64_sve_ld3q_sret: { |
5261 | SelectPredicatedLoad(N: Node, NumVecs: 3, Scale: 4, Opc_ri: AArch64::LD3Q_IMM, Opc_rr: AArch64::LD3Q, IsIntr: true); |
5262 | return; |
5263 | } |
5264 | case Intrinsic::aarch64_sve_ld4q_sret: { |
5265 | SelectPredicatedLoad(N: Node, NumVecs: 4, Scale: 4, Opc_ri: AArch64::LD4Q_IMM, Opc_rr: AArch64::LD4Q, IsIntr: true); |
5266 | return; |
5267 | } |
5268 | case Intrinsic::aarch64_sve_ld2_sret: { |
5269 | if (VT == MVT::nxv16i8) { |
5270 | SelectPredicatedLoad(N: Node, NumVecs: 2, Scale: 0, Opc_ri: AArch64::LD2B_IMM, Opc_rr: AArch64::LD2B, |
5271 | IsIntr: true); |
5272 | return; |
5273 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5274 | VT == MVT::nxv8bf16) { |
5275 | SelectPredicatedLoad(N: Node, NumVecs: 2, Scale: 1, Opc_ri: AArch64::LD2H_IMM, Opc_rr: AArch64::LD2H, |
5276 | IsIntr: true); |
5277 | return; |
5278 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5279 | SelectPredicatedLoad(N: Node, NumVecs: 2, Scale: 2, Opc_ri: AArch64::LD2W_IMM, Opc_rr: AArch64::LD2W, |
5280 | IsIntr: true); |
5281 | return; |
5282 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5283 | SelectPredicatedLoad(N: Node, NumVecs: 2, Scale: 3, Opc_ri: AArch64::LD2D_IMM, Opc_rr: AArch64::LD2D, |
5284 | IsIntr: true); |
5285 | return; |
5286 | } |
5287 | break; |
5288 | } |
5289 | case Intrinsic::aarch64_sve_ld1_pn_x2: { |
5290 | if (VT == MVT::nxv16i8) { |
5291 | if (Subtarget->hasSME2()) |
5292 | SelectContiguousMultiVectorLoad( |
5293 | N: Node, NumVecs: 2, Scale: 0, Opc_ri: AArch64::LD1B_2Z_IMM_PSEUDO, Opc_rr: AArch64::LD1B_2Z_PSEUDO); |
5294 | else if (Subtarget->hasSVE2p1()) |
5295 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 2, Scale: 0, Opc_ri: AArch64::LD1B_2Z_IMM, |
5296 | Opc_rr: AArch64::LD1B_2Z); |
5297 | else |
5298 | break; |
5299 | return; |
5300 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5301 | VT == MVT::nxv8bf16) { |
5302 | if (Subtarget->hasSME2()) |
5303 | SelectContiguousMultiVectorLoad( |
5304 | N: Node, NumVecs: 2, Scale: 1, Opc_ri: AArch64::LD1H_2Z_IMM_PSEUDO, Opc_rr: AArch64::LD1H_2Z_PSEUDO); |
5305 | else if (Subtarget->hasSVE2p1()) |
5306 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 2, Scale: 1, Opc_ri: AArch64::LD1H_2Z_IMM, |
5307 | Opc_rr: AArch64::LD1H_2Z); |
5308 | else |
5309 | break; |
5310 | return; |
5311 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5312 | if (Subtarget->hasSME2()) |
5313 | SelectContiguousMultiVectorLoad( |
5314 | N: Node, NumVecs: 2, Scale: 2, Opc_ri: AArch64::LD1W_2Z_IMM_PSEUDO, Opc_rr: AArch64::LD1W_2Z_PSEUDO); |
5315 | else if (Subtarget->hasSVE2p1()) |
5316 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 2, Scale: 2, Opc_ri: AArch64::LD1W_2Z_IMM, |
5317 | Opc_rr: AArch64::LD1W_2Z); |
5318 | else |
5319 | break; |
5320 | return; |
5321 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5322 | if (Subtarget->hasSME2()) |
5323 | SelectContiguousMultiVectorLoad( |
5324 | N: Node, NumVecs: 2, Scale: 3, Opc_ri: AArch64::LD1D_2Z_IMM_PSEUDO, Opc_rr: AArch64::LD1D_2Z_PSEUDO); |
5325 | else if (Subtarget->hasSVE2p1()) |
5326 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 2, Scale: 3, Opc_ri: AArch64::LD1D_2Z_IMM, |
5327 | Opc_rr: AArch64::LD1D_2Z); |
5328 | else |
5329 | break; |
5330 | return; |
5331 | } |
5332 | break; |
5333 | } |
5334 | case Intrinsic::aarch64_sve_ld1_pn_x4: { |
5335 | if (VT == MVT::nxv16i8) { |
5336 | if (Subtarget->hasSME2()) |
5337 | SelectContiguousMultiVectorLoad( |
5338 | N: Node, NumVecs: 4, Scale: 0, Opc_ri: AArch64::LD1B_4Z_IMM_PSEUDO, Opc_rr: AArch64::LD1B_4Z_PSEUDO); |
5339 | else if (Subtarget->hasSVE2p1()) |
5340 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 4, Scale: 0, Opc_ri: AArch64::LD1B_4Z_IMM, |
5341 | Opc_rr: AArch64::LD1B_4Z); |
5342 | else |
5343 | break; |
5344 | return; |
5345 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5346 | VT == MVT::nxv8bf16) { |
5347 | if (Subtarget->hasSME2()) |
5348 | SelectContiguousMultiVectorLoad( |
5349 | N: Node, NumVecs: 4, Scale: 1, Opc_ri: AArch64::LD1H_4Z_IMM_PSEUDO, Opc_rr: AArch64::LD1H_4Z_PSEUDO); |
5350 | else if (Subtarget->hasSVE2p1()) |
5351 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 4, Scale: 1, Opc_ri: AArch64::LD1H_4Z_IMM, |
5352 | Opc_rr: AArch64::LD1H_4Z); |
5353 | else |
5354 | break; |
5355 | return; |
5356 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5357 | if (Subtarget->hasSME2()) |
5358 | SelectContiguousMultiVectorLoad( |
5359 | N: Node, NumVecs: 4, Scale: 2, Opc_ri: AArch64::LD1W_4Z_IMM_PSEUDO, Opc_rr: AArch64::LD1W_4Z_PSEUDO); |
5360 | else if (Subtarget->hasSVE2p1()) |
5361 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 4, Scale: 2, Opc_ri: AArch64::LD1W_4Z_IMM, |
5362 | Opc_rr: AArch64::LD1W_4Z); |
5363 | else |
5364 | break; |
5365 | return; |
5366 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5367 | if (Subtarget->hasSME2()) |
5368 | SelectContiguousMultiVectorLoad( |
5369 | N: Node, NumVecs: 4, Scale: 3, Opc_ri: AArch64::LD1D_4Z_IMM_PSEUDO, Opc_rr: AArch64::LD1D_4Z_PSEUDO); |
5370 | else if (Subtarget->hasSVE2p1()) |
5371 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 4, Scale: 3, Opc_ri: AArch64::LD1D_4Z_IMM, |
5372 | Opc_rr: AArch64::LD1D_4Z); |
5373 | else |
5374 | break; |
5375 | return; |
5376 | } |
5377 | break; |
5378 | } |
5379 | case Intrinsic::aarch64_sve_ldnt1_pn_x2: { |
5380 | if (VT == MVT::nxv16i8) { |
5381 | if (Subtarget->hasSME2()) |
5382 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 2, Scale: 0, |
5383 | Opc_ri: AArch64::LDNT1B_2Z_IMM_PSEUDO, |
5384 | Opc_rr: AArch64::LDNT1B_2Z_PSEUDO); |
5385 | else if (Subtarget->hasSVE2p1()) |
5386 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 2, Scale: 0, Opc_ri: AArch64::LDNT1B_2Z_IMM, |
5387 | Opc_rr: AArch64::LDNT1B_2Z); |
5388 | else |
5389 | break; |
5390 | return; |
5391 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5392 | VT == MVT::nxv8bf16) { |
5393 | if (Subtarget->hasSME2()) |
5394 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 2, Scale: 1, |
5395 | Opc_ri: AArch64::LDNT1H_2Z_IMM_PSEUDO, |
5396 | Opc_rr: AArch64::LDNT1H_2Z_PSEUDO); |
5397 | else if (Subtarget->hasSVE2p1()) |
5398 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 2, Scale: 1, Opc_ri: AArch64::LDNT1H_2Z_IMM, |
5399 | Opc_rr: AArch64::LDNT1H_2Z); |
5400 | else |
5401 | break; |
5402 | return; |
5403 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5404 | if (Subtarget->hasSME2()) |
5405 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 2, Scale: 2, |
5406 | Opc_ri: AArch64::LDNT1W_2Z_IMM_PSEUDO, |
5407 | Opc_rr: AArch64::LDNT1W_2Z_PSEUDO); |
5408 | else if (Subtarget->hasSVE2p1()) |
5409 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 2, Scale: 2, Opc_ri: AArch64::LDNT1W_2Z_IMM, |
5410 | Opc_rr: AArch64::LDNT1W_2Z); |
5411 | else |
5412 | break; |
5413 | return; |
5414 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5415 | if (Subtarget->hasSME2()) |
5416 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 2, Scale: 3, |
5417 | Opc_ri: AArch64::LDNT1D_2Z_IMM_PSEUDO, |
5418 | Opc_rr: AArch64::LDNT1D_2Z_PSEUDO); |
5419 | else if (Subtarget->hasSVE2p1()) |
5420 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 2, Scale: 3, Opc_ri: AArch64::LDNT1D_2Z_IMM, |
5421 | Opc_rr: AArch64::LDNT1D_2Z); |
5422 | else |
5423 | break; |
5424 | return; |
5425 | } |
5426 | break; |
5427 | } |
5428 | case Intrinsic::aarch64_sve_ldnt1_pn_x4: { |
5429 | if (VT == MVT::nxv16i8) { |
5430 | if (Subtarget->hasSME2()) |
5431 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 4, Scale: 0, |
5432 | Opc_ri: AArch64::LDNT1B_4Z_IMM_PSEUDO, |
5433 | Opc_rr: AArch64::LDNT1B_4Z_PSEUDO); |
5434 | else if (Subtarget->hasSVE2p1()) |
5435 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 4, Scale: 0, Opc_ri: AArch64::LDNT1B_4Z_IMM, |
5436 | Opc_rr: AArch64::LDNT1B_4Z); |
5437 | else |
5438 | break; |
5439 | return; |
5440 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5441 | VT == MVT::nxv8bf16) { |
5442 | if (Subtarget->hasSME2()) |
5443 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 4, Scale: 1, |
5444 | Opc_ri: AArch64::LDNT1H_4Z_IMM_PSEUDO, |
5445 | Opc_rr: AArch64::LDNT1H_4Z_PSEUDO); |
5446 | else if (Subtarget->hasSVE2p1()) |
5447 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 4, Scale: 1, Opc_ri: AArch64::LDNT1H_4Z_IMM, |
5448 | Opc_rr: AArch64::LDNT1H_4Z); |
5449 | else |
5450 | break; |
5451 | return; |
5452 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5453 | if (Subtarget->hasSME2()) |
5454 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 4, Scale: 2, |
5455 | Opc_ri: AArch64::LDNT1W_4Z_IMM_PSEUDO, |
5456 | Opc_rr: AArch64::LDNT1W_4Z_PSEUDO); |
5457 | else if (Subtarget->hasSVE2p1()) |
5458 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 4, Scale: 2, Opc_ri: AArch64::LDNT1W_4Z_IMM, |
5459 | Opc_rr: AArch64::LDNT1W_4Z); |
5460 | else |
5461 | break; |
5462 | return; |
5463 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5464 | if (Subtarget->hasSME2()) |
5465 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 4, Scale: 3, |
5466 | Opc_ri: AArch64::LDNT1D_4Z_IMM_PSEUDO, |
5467 | Opc_rr: AArch64::LDNT1D_4Z_PSEUDO); |
5468 | else if (Subtarget->hasSVE2p1()) |
5469 | SelectContiguousMultiVectorLoad(N: Node, NumVecs: 4, Scale: 3, Opc_ri: AArch64::LDNT1D_4Z_IMM, |
5470 | Opc_rr: AArch64::LDNT1D_4Z); |
5471 | else |
5472 | break; |
5473 | return; |
5474 | } |
5475 | break; |
5476 | } |
5477 | case Intrinsic::aarch64_sve_ld3_sret: { |
5478 | if (VT == MVT::nxv16i8) { |
5479 | SelectPredicatedLoad(N: Node, NumVecs: 3, Scale: 0, Opc_ri: AArch64::LD3B_IMM, Opc_rr: AArch64::LD3B, |
5480 | IsIntr: true); |
5481 | return; |
5482 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5483 | VT == MVT::nxv8bf16) { |
5484 | SelectPredicatedLoad(N: Node, NumVecs: 3, Scale: 1, Opc_ri: AArch64::LD3H_IMM, Opc_rr: AArch64::LD3H, |
5485 | IsIntr: true); |
5486 | return; |
5487 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5488 | SelectPredicatedLoad(N: Node, NumVecs: 3, Scale: 2, Opc_ri: AArch64::LD3W_IMM, Opc_rr: AArch64::LD3W, |
5489 | IsIntr: true); |
5490 | return; |
5491 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5492 | SelectPredicatedLoad(N: Node, NumVecs: 3, Scale: 3, Opc_ri: AArch64::LD3D_IMM, Opc_rr: AArch64::LD3D, |
5493 | IsIntr: true); |
5494 | return; |
5495 | } |
5496 | break; |
5497 | } |
5498 | case Intrinsic::aarch64_sve_ld4_sret: { |
5499 | if (VT == MVT::nxv16i8) { |
5500 | SelectPredicatedLoad(N: Node, NumVecs: 4, Scale: 0, Opc_ri: AArch64::LD4B_IMM, Opc_rr: AArch64::LD4B, |
5501 | IsIntr: true); |
5502 | return; |
5503 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5504 | VT == MVT::nxv8bf16) { |
5505 | SelectPredicatedLoad(N: Node, NumVecs: 4, Scale: 1, Opc_ri: AArch64::LD4H_IMM, Opc_rr: AArch64::LD4H, |
5506 | IsIntr: true); |
5507 | return; |
5508 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5509 | SelectPredicatedLoad(N: Node, NumVecs: 4, Scale: 2, Opc_ri: AArch64::LD4W_IMM, Opc_rr: AArch64::LD4W, |
5510 | IsIntr: true); |
5511 | return; |
5512 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5513 | SelectPredicatedLoad(N: Node, NumVecs: 4, Scale: 3, Opc_ri: AArch64::LD4D_IMM, Opc_rr: AArch64::LD4D, |
5514 | IsIntr: true); |
5515 | return; |
5516 | } |
5517 | break; |
5518 | } |
5519 | case Intrinsic::aarch64_sme_read_hor_vg2: { |
5520 | if (VT == MVT::nxv16i8) { |
5521 | SelectMultiVectorMove<14, 2>(N: Node, NumVecs: 2, BaseReg: AArch64::ZAB0, |
5522 | Op: AArch64::MOVA_2ZMXI_H_B); |
5523 | return; |
5524 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5525 | VT == MVT::nxv8bf16) { |
5526 | SelectMultiVectorMove<6, 2>(N: Node, NumVecs: 2, BaseReg: AArch64::ZAH0, |
5527 | Op: AArch64::MOVA_2ZMXI_H_H); |
5528 | return; |
5529 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5530 | SelectMultiVectorMove<2, 2>(N: Node, NumVecs: 2, BaseReg: AArch64::ZAS0, |
5531 | Op: AArch64::MOVA_2ZMXI_H_S); |
5532 | return; |
5533 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5534 | SelectMultiVectorMove<0, 2>(N: Node, NumVecs: 2, BaseReg: AArch64::ZAD0, |
5535 | Op: AArch64::MOVA_2ZMXI_H_D); |
5536 | return; |
5537 | } |
5538 | break; |
5539 | } |
5540 | case Intrinsic::aarch64_sme_read_ver_vg2: { |
5541 | if (VT == MVT::nxv16i8) { |
5542 | SelectMultiVectorMove<14, 2>(N: Node, NumVecs: 2, BaseReg: AArch64::ZAB0, |
5543 | Op: AArch64::MOVA_2ZMXI_V_B); |
5544 | return; |
5545 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5546 | VT == MVT::nxv8bf16) { |
5547 | SelectMultiVectorMove<6, 2>(N: Node, NumVecs: 2, BaseReg: AArch64::ZAH0, |
5548 | Op: AArch64::MOVA_2ZMXI_V_H); |
5549 | return; |
5550 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5551 | SelectMultiVectorMove<2, 2>(N: Node, NumVecs: 2, BaseReg: AArch64::ZAS0, |
5552 | Op: AArch64::MOVA_2ZMXI_V_S); |
5553 | return; |
5554 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5555 | SelectMultiVectorMove<0, 2>(N: Node, NumVecs: 2, BaseReg: AArch64::ZAD0, |
5556 | Op: AArch64::MOVA_2ZMXI_V_D); |
5557 | return; |
5558 | } |
5559 | break; |
5560 | } |
5561 | case Intrinsic::aarch64_sme_read_hor_vg4: { |
5562 | if (VT == MVT::nxv16i8) { |
5563 | SelectMultiVectorMove<12, 4>(N: Node, NumVecs: 4, BaseReg: AArch64::ZAB0, |
5564 | Op: AArch64::MOVA_4ZMXI_H_B); |
5565 | return; |
5566 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5567 | VT == MVT::nxv8bf16) { |
5568 | SelectMultiVectorMove<4, 4>(N: Node, NumVecs: 4, BaseReg: AArch64::ZAH0, |
5569 | Op: AArch64::MOVA_4ZMXI_H_H); |
5570 | return; |
5571 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5572 | SelectMultiVectorMove<0, 2>(N: Node, NumVecs: 4, BaseReg: AArch64::ZAS0, |
5573 | Op: AArch64::MOVA_4ZMXI_H_S); |
5574 | return; |
5575 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5576 | SelectMultiVectorMove<0, 2>(N: Node, NumVecs: 4, BaseReg: AArch64::ZAD0, |
5577 | Op: AArch64::MOVA_4ZMXI_H_D); |
5578 | return; |
5579 | } |
5580 | break; |
5581 | } |
5582 | case Intrinsic::aarch64_sme_read_ver_vg4: { |
5583 | if (VT == MVT::nxv16i8) { |
5584 | SelectMultiVectorMove<12, 4>(N: Node, NumVecs: 4, BaseReg: AArch64::ZAB0, |
5585 | Op: AArch64::MOVA_4ZMXI_V_B); |
5586 | return; |
5587 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5588 | VT == MVT::nxv8bf16) { |
5589 | SelectMultiVectorMove<4, 4>(N: Node, NumVecs: 4, BaseReg: AArch64::ZAH0, |
5590 | Op: AArch64::MOVA_4ZMXI_V_H); |
5591 | return; |
5592 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5593 | SelectMultiVectorMove<0, 4>(N: Node, NumVecs: 4, BaseReg: AArch64::ZAS0, |
5594 | Op: AArch64::MOVA_4ZMXI_V_S); |
5595 | return; |
5596 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5597 | SelectMultiVectorMove<0, 4>(N: Node, NumVecs: 4, BaseReg: AArch64::ZAD0, |
5598 | Op: AArch64::MOVA_4ZMXI_V_D); |
5599 | return; |
5600 | } |
5601 | break; |
5602 | } |
5603 | case Intrinsic::aarch64_sme_read_vg1x2: { |
5604 | SelectMultiVectorMove<7, 1>(N: Node, NumVecs: 2, BaseReg: AArch64::ZA, |
5605 | Op: AArch64::MOVA_VG2_2ZMXI); |
5606 | return; |
5607 | } |
5608 | case Intrinsic::aarch64_sme_read_vg1x4: { |
5609 | SelectMultiVectorMove<7, 1>(N: Node, NumVecs: 4, BaseReg: AArch64::ZA, |
5610 | Op: AArch64::MOVA_VG4_4ZMXI); |
5611 | return; |
5612 | } |
5613 | case Intrinsic::aarch64_sme_readz_horiz_x2: { |
5614 | if (VT == MVT::nxv16i8) { |
5615 | SelectMultiVectorMoveZ(N: Node, NumVecs: 2, Op: AArch64::MOVAZ_2ZMI_H_B_PSEUDO, MaxIdx: 14, Scale: 2); |
5616 | return; |
5617 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5618 | VT == MVT::nxv8bf16) { |
5619 | SelectMultiVectorMoveZ(N: Node, NumVecs: 2, Op: AArch64::MOVAZ_2ZMI_H_H_PSEUDO, MaxIdx: 6, Scale: 2); |
5620 | return; |
5621 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5622 | SelectMultiVectorMoveZ(N: Node, NumVecs: 2, Op: AArch64::MOVAZ_2ZMI_H_S_PSEUDO, MaxIdx: 2, Scale: 2); |
5623 | return; |
5624 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5625 | SelectMultiVectorMoveZ(N: Node, NumVecs: 2, Op: AArch64::MOVAZ_2ZMI_H_D_PSEUDO, MaxIdx: 0, Scale: 2); |
5626 | return; |
5627 | } |
5628 | break; |
5629 | } |
5630 | case Intrinsic::aarch64_sme_readz_vert_x2: { |
5631 | if (VT == MVT::nxv16i8) { |
5632 | SelectMultiVectorMoveZ(N: Node, NumVecs: 2, Op: AArch64::MOVAZ_2ZMI_V_B_PSEUDO, MaxIdx: 14, Scale: 2); |
5633 | return; |
5634 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5635 | VT == MVT::nxv8bf16) { |
5636 | SelectMultiVectorMoveZ(N: Node, NumVecs: 2, Op: AArch64::MOVAZ_2ZMI_V_H_PSEUDO, MaxIdx: 6, Scale: 2); |
5637 | return; |
5638 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5639 | SelectMultiVectorMoveZ(N: Node, NumVecs: 2, Op: AArch64::MOVAZ_2ZMI_V_S_PSEUDO, MaxIdx: 2, Scale: 2); |
5640 | return; |
5641 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5642 | SelectMultiVectorMoveZ(N: Node, NumVecs: 2, Op: AArch64::MOVAZ_2ZMI_V_D_PSEUDO, MaxIdx: 0, Scale: 2); |
5643 | return; |
5644 | } |
5645 | break; |
5646 | } |
5647 | case Intrinsic::aarch64_sme_readz_horiz_x4: { |
5648 | if (VT == MVT::nxv16i8) { |
5649 | SelectMultiVectorMoveZ(N: Node, NumVecs: 4, Op: AArch64::MOVAZ_4ZMI_H_B_PSEUDO, MaxIdx: 12, Scale: 4); |
5650 | return; |
5651 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5652 | VT == MVT::nxv8bf16) { |
5653 | SelectMultiVectorMoveZ(N: Node, NumVecs: 4, Op: AArch64::MOVAZ_4ZMI_H_H_PSEUDO, MaxIdx: 4, Scale: 4); |
5654 | return; |
5655 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5656 | SelectMultiVectorMoveZ(N: Node, NumVecs: 4, Op: AArch64::MOVAZ_4ZMI_H_S_PSEUDO, MaxIdx: 0, Scale: 4); |
5657 | return; |
5658 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5659 | SelectMultiVectorMoveZ(N: Node, NumVecs: 4, Op: AArch64::MOVAZ_4ZMI_H_D_PSEUDO, MaxIdx: 0, Scale: 4); |
5660 | return; |
5661 | } |
5662 | break; |
5663 | } |
5664 | case Intrinsic::aarch64_sme_readz_vert_x4: { |
5665 | if (VT == MVT::nxv16i8) { |
5666 | SelectMultiVectorMoveZ(N: Node, NumVecs: 4, Op: AArch64::MOVAZ_4ZMI_V_B_PSEUDO, MaxIdx: 12, Scale: 4); |
5667 | return; |
5668 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
5669 | VT == MVT::nxv8bf16) { |
5670 | SelectMultiVectorMoveZ(N: Node, NumVecs: 4, Op: AArch64::MOVAZ_4ZMI_V_H_PSEUDO, MaxIdx: 4, Scale: 4); |
5671 | return; |
5672 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
5673 | SelectMultiVectorMoveZ(N: Node, NumVecs: 4, Op: AArch64::MOVAZ_4ZMI_V_S_PSEUDO, MaxIdx: 0, Scale: 4); |
5674 | return; |
5675 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
5676 | SelectMultiVectorMoveZ(N: Node, NumVecs: 4, Op: AArch64::MOVAZ_4ZMI_V_D_PSEUDO, MaxIdx: 0, Scale: 4); |
5677 | return; |
5678 | } |
5679 | break; |
5680 | } |
5681 | case Intrinsic::aarch64_sme_readz_x2: { |
5682 | SelectMultiVectorMoveZ(N: Node, NumVecs: 2, Op: AArch64::MOVAZ_VG2_2ZMXI_PSEUDO, MaxIdx: 7, Scale: 1, |
5683 | BaseReg: AArch64::ZA); |
5684 | return; |
5685 | } |
5686 | case Intrinsic::aarch64_sme_readz_x4: { |
5687 | SelectMultiVectorMoveZ(N: Node, NumVecs: 4, Op: AArch64::MOVAZ_VG4_4ZMXI_PSEUDO, MaxIdx: 7, Scale: 1, |
5688 | BaseReg: AArch64::ZA); |
5689 | return; |
5690 | } |
5691 | case Intrinsic::swift_async_context_addr: { |
5692 | SDLoc DL(Node); |
5693 | SDValue Chain = Node->getOperand(Num: 0); |
5694 | SDValue CopyFP = CurDAG->getCopyFromReg(Chain, dl: DL, Reg: AArch64::FP, VT: MVT::i64); |
5695 | SDValue Res = SDValue( |
5696 | CurDAG->getMachineNode(Opcode: AArch64::SUBXri, dl: DL, VT: MVT::i64, Op1: CopyFP, |
5697 | Op2: CurDAG->getTargetConstant(Val: 8, DL, VT: MVT::i32), |
5698 | Op3: CurDAG->getTargetConstant(Val: 0, DL, VT: MVT::i32)), |
5699 | 0); |
5700 | ReplaceUses(F: SDValue(Node, 0), T: Res); |
5701 | ReplaceUses(F: SDValue(Node, 1), T: CopyFP.getValue(R: 1)); |
5702 | CurDAG->RemoveDeadNode(N: Node); |
5703 | |
5704 | auto &MF = CurDAG->getMachineFunction(); |
5705 | MF.getFrameInfo().setFrameAddressIsTaken(true); |
5706 | MF.getInfo<AArch64FunctionInfo>()->setHasSwiftAsyncContext(true); |
5707 | return; |
5708 | } |
5709 | case Intrinsic::aarch64_sme_luti2_lane_zt_x4: { |
5710 | if (auto Opc = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
5711 | VT: Node->getValueType(ResNo: 0), |
5712 | Opcodes: {AArch64::LUTI2_4ZTZI_B, AArch64::LUTI2_4ZTZI_H, |
5713 | AArch64::LUTI2_4ZTZI_S})) |
5714 | // Second Immediate must be <= 3: |
5715 | SelectMultiVectorLutiLane(Node, NumOutVecs: 4, Opc, MaxImm: 3); |
5716 | return; |
5717 | } |
5718 | case Intrinsic::aarch64_sme_luti4_lane_zt_x4: { |
5719 | if (auto Opc = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
5720 | VT: Node->getValueType(ResNo: 0), |
5721 | Opcodes: {0, AArch64::LUTI4_4ZTZI_H, AArch64::LUTI4_4ZTZI_S})) |
5722 | // Second Immediate must be <= 1: |
5723 | SelectMultiVectorLutiLane(Node, NumOutVecs: 4, Opc, MaxImm: 1); |
5724 | return; |
5725 | } |
5726 | case Intrinsic::aarch64_sme_luti2_lane_zt_x2: { |
5727 | if (auto Opc = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
5728 | VT: Node->getValueType(ResNo: 0), |
5729 | Opcodes: {AArch64::LUTI2_2ZTZI_B, AArch64::LUTI2_2ZTZI_H, |
5730 | AArch64::LUTI2_2ZTZI_S})) |
5731 | // Second Immediate must be <= 7: |
5732 | SelectMultiVectorLutiLane(Node, NumOutVecs: 2, Opc, MaxImm: 7); |
5733 | return; |
5734 | } |
5735 | case Intrinsic::aarch64_sme_luti4_lane_zt_x2: { |
5736 | if (auto Opc = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
5737 | VT: Node->getValueType(ResNo: 0), |
5738 | Opcodes: {AArch64::LUTI4_2ZTZI_B, AArch64::LUTI4_2ZTZI_H, |
5739 | AArch64::LUTI4_2ZTZI_S})) |
5740 | // Second Immediate must be <= 3: |
5741 | SelectMultiVectorLutiLane(Node, NumOutVecs: 2, Opc, MaxImm: 3); |
5742 | return; |
5743 | } |
5744 | case Intrinsic::aarch64_sme_luti4_zt_x4: { |
5745 | SelectMultiVectorLuti(Node, NumOutVecs: 4, Opc: AArch64::LUTI4_4ZZT2Z); |
5746 | return; |
5747 | } |
5748 | case Intrinsic::aarch64_sve_fp8_cvtl1_x2: |
5749 | if (auto Opc = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5750 | VT: Node->getValueType(ResNo: 0), |
5751 | Opcodes: {AArch64::BF1CVTL_2ZZ_BtoH, AArch64::F1CVTL_2ZZ_BtoH})) |
5752 | SelectCVTIntrinsicFP8(N: Node, NumVecs: 2, Opcode: Opc); |
5753 | return; |
5754 | case Intrinsic::aarch64_sve_fp8_cvtl2_x2: |
5755 | if (auto Opc = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5756 | VT: Node->getValueType(ResNo: 0), |
5757 | Opcodes: {AArch64::BF2CVTL_2ZZ_BtoH, AArch64::F2CVTL_2ZZ_BtoH})) |
5758 | SelectCVTIntrinsicFP8(N: Node, NumVecs: 2, Opcode: Opc); |
5759 | return; |
5760 | case Intrinsic::aarch64_sve_fp8_cvt1_x2: |
5761 | if (auto Opc = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5762 | VT: Node->getValueType(ResNo: 0), |
5763 | Opcodes: {AArch64::BF1CVT_2ZZ_BtoH, AArch64::F1CVT_2ZZ_BtoH})) |
5764 | SelectCVTIntrinsicFP8(N: Node, NumVecs: 2, Opcode: Opc); |
5765 | return; |
5766 | case Intrinsic::aarch64_sve_fp8_cvt2_x2: |
5767 | if (auto Opc = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5768 | VT: Node->getValueType(ResNo: 0), |
5769 | Opcodes: {AArch64::BF2CVT_2ZZ_BtoH, AArch64::F2CVT_2ZZ_BtoH})) |
5770 | SelectCVTIntrinsicFP8(N: Node, NumVecs: 2, Opcode: Opc); |
5771 | return; |
5772 | } |
5773 | } break; |
5774 | case ISD::INTRINSIC_WO_CHAIN: { |
5775 | unsigned IntNo = Node->getConstantOperandVal(Num: 0); |
5776 | switch (IntNo) { |
5777 | default: |
5778 | break; |
5779 | case Intrinsic::aarch64_tagp: |
5780 | SelectTagP(N: Node); |
5781 | return; |
5782 | |
5783 | case Intrinsic::ptrauth_auth: |
5784 | SelectPtrauthAuth(N: Node); |
5785 | return; |
5786 | |
5787 | case Intrinsic::ptrauth_resign: |
5788 | SelectPtrauthResign(N: Node); |
5789 | return; |
5790 | |
5791 | case Intrinsic::aarch64_neon_tbl2: |
5792 | SelectTable(N: Node, NumVecs: 2, |
5793 | Opc: VT == MVT::v8i8 ? AArch64::TBLv8i8Two : AArch64::TBLv16i8Two, |
5794 | isExt: false); |
5795 | return; |
5796 | case Intrinsic::aarch64_neon_tbl3: |
5797 | SelectTable(N: Node, NumVecs: 3, Opc: VT == MVT::v8i8 ? AArch64::TBLv8i8Three |
5798 | : AArch64::TBLv16i8Three, |
5799 | isExt: false); |
5800 | return; |
5801 | case Intrinsic::aarch64_neon_tbl4: |
5802 | SelectTable(N: Node, NumVecs: 4, Opc: VT == MVT::v8i8 ? AArch64::TBLv8i8Four |
5803 | : AArch64::TBLv16i8Four, |
5804 | isExt: false); |
5805 | return; |
5806 | case Intrinsic::aarch64_neon_tbx2: |
5807 | SelectTable(N: Node, NumVecs: 2, |
5808 | Opc: VT == MVT::v8i8 ? AArch64::TBXv8i8Two : AArch64::TBXv16i8Two, |
5809 | isExt: true); |
5810 | return; |
5811 | case Intrinsic::aarch64_neon_tbx3: |
5812 | SelectTable(N: Node, NumVecs: 3, Opc: VT == MVT::v8i8 ? AArch64::TBXv8i8Three |
5813 | : AArch64::TBXv16i8Three, |
5814 | isExt: true); |
5815 | return; |
5816 | case Intrinsic::aarch64_neon_tbx4: |
5817 | SelectTable(N: Node, NumVecs: 4, Opc: VT == MVT::v8i8 ? AArch64::TBXv8i8Four |
5818 | : AArch64::TBXv16i8Four, |
5819 | isExt: true); |
5820 | return; |
5821 | case Intrinsic::aarch64_sve_srshl_single_x2: |
5822 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5823 | VT: Node->getValueType(ResNo: 0), |
5824 | Opcodes: {AArch64::SRSHL_VG2_2ZZ_B, AArch64::SRSHL_VG2_2ZZ_H, |
5825 | AArch64::SRSHL_VG2_2ZZ_S, AArch64::SRSHL_VG2_2ZZ_D})) |
5826 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
5827 | return; |
5828 | case Intrinsic::aarch64_sve_srshl_single_x4: |
5829 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5830 | VT: Node->getValueType(ResNo: 0), |
5831 | Opcodes: {AArch64::SRSHL_VG4_4ZZ_B, AArch64::SRSHL_VG4_4ZZ_H, |
5832 | AArch64::SRSHL_VG4_4ZZ_S, AArch64::SRSHL_VG4_4ZZ_D})) |
5833 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
5834 | return; |
5835 | case Intrinsic::aarch64_sve_urshl_single_x2: |
5836 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5837 | VT: Node->getValueType(ResNo: 0), |
5838 | Opcodes: {AArch64::URSHL_VG2_2ZZ_B, AArch64::URSHL_VG2_2ZZ_H, |
5839 | AArch64::URSHL_VG2_2ZZ_S, AArch64::URSHL_VG2_2ZZ_D})) |
5840 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
5841 | return; |
5842 | case Intrinsic::aarch64_sve_urshl_single_x4: |
5843 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5844 | VT: Node->getValueType(ResNo: 0), |
5845 | Opcodes: {AArch64::URSHL_VG4_4ZZ_B, AArch64::URSHL_VG4_4ZZ_H, |
5846 | AArch64::URSHL_VG4_4ZZ_S, AArch64::URSHL_VG4_4ZZ_D})) |
5847 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
5848 | return; |
5849 | case Intrinsic::aarch64_sve_srshl_x2: |
5850 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5851 | VT: Node->getValueType(ResNo: 0), |
5852 | Opcodes: {AArch64::SRSHL_VG2_2Z2Z_B, AArch64::SRSHL_VG2_2Z2Z_H, |
5853 | AArch64::SRSHL_VG2_2Z2Z_S, AArch64::SRSHL_VG2_2Z2Z_D})) |
5854 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
5855 | return; |
5856 | case Intrinsic::aarch64_sve_srshl_x4: |
5857 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5858 | VT: Node->getValueType(ResNo: 0), |
5859 | Opcodes: {AArch64::SRSHL_VG4_4Z4Z_B, AArch64::SRSHL_VG4_4Z4Z_H, |
5860 | AArch64::SRSHL_VG4_4Z4Z_S, AArch64::SRSHL_VG4_4Z4Z_D})) |
5861 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
5862 | return; |
5863 | case Intrinsic::aarch64_sve_urshl_x2: |
5864 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5865 | VT: Node->getValueType(ResNo: 0), |
5866 | Opcodes: {AArch64::URSHL_VG2_2Z2Z_B, AArch64::URSHL_VG2_2Z2Z_H, |
5867 | AArch64::URSHL_VG2_2Z2Z_S, AArch64::URSHL_VG2_2Z2Z_D})) |
5868 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
5869 | return; |
5870 | case Intrinsic::aarch64_sve_urshl_x4: |
5871 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5872 | VT: Node->getValueType(ResNo: 0), |
5873 | Opcodes: {AArch64::URSHL_VG4_4Z4Z_B, AArch64::URSHL_VG4_4Z4Z_H, |
5874 | AArch64::URSHL_VG4_4Z4Z_S, AArch64::URSHL_VG4_4Z4Z_D})) |
5875 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
5876 | return; |
5877 | case Intrinsic::aarch64_sve_sqdmulh_single_vgx2: |
5878 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5879 | VT: Node->getValueType(ResNo: 0), |
5880 | Opcodes: {AArch64::SQDMULH_VG2_2ZZ_B, AArch64::SQDMULH_VG2_2ZZ_H, |
5881 | AArch64::SQDMULH_VG2_2ZZ_S, AArch64::SQDMULH_VG2_2ZZ_D})) |
5882 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
5883 | return; |
5884 | case Intrinsic::aarch64_sve_sqdmulh_single_vgx4: |
5885 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5886 | VT: Node->getValueType(ResNo: 0), |
5887 | Opcodes: {AArch64::SQDMULH_VG4_4ZZ_B, AArch64::SQDMULH_VG4_4ZZ_H, |
5888 | AArch64::SQDMULH_VG4_4ZZ_S, AArch64::SQDMULH_VG4_4ZZ_D})) |
5889 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
5890 | return; |
5891 | case Intrinsic::aarch64_sve_sqdmulh_vgx2: |
5892 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5893 | VT: Node->getValueType(ResNo: 0), |
5894 | Opcodes: {AArch64::SQDMULH_VG2_2Z2Z_B, AArch64::SQDMULH_VG2_2Z2Z_H, |
5895 | AArch64::SQDMULH_VG2_2Z2Z_S, AArch64::SQDMULH_VG2_2Z2Z_D})) |
5896 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
5897 | return; |
5898 | case Intrinsic::aarch64_sve_sqdmulh_vgx4: |
5899 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5900 | VT: Node->getValueType(ResNo: 0), |
5901 | Opcodes: {AArch64::SQDMULH_VG4_4Z4Z_B, AArch64::SQDMULH_VG4_4Z4Z_H, |
5902 | AArch64::SQDMULH_VG4_4Z4Z_S, AArch64::SQDMULH_VG4_4Z4Z_D})) |
5903 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
5904 | return; |
5905 | case Intrinsic::aarch64_sme_fp8_scale_single_x2: |
5906 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5907 | VT: Node->getValueType(ResNo: 0), |
5908 | Opcodes: {0, AArch64::FSCALE_2ZZ_H, AArch64::FSCALE_2ZZ_S, |
5909 | AArch64::FSCALE_2ZZ_D})) |
5910 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
5911 | return; |
5912 | case Intrinsic::aarch64_sme_fp8_scale_single_x4: |
5913 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5914 | VT: Node->getValueType(ResNo: 0), |
5915 | Opcodes: {0, AArch64::FSCALE_4ZZ_H, AArch64::FSCALE_4ZZ_S, |
5916 | AArch64::FSCALE_4ZZ_D})) |
5917 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
5918 | return; |
5919 | case Intrinsic::aarch64_sme_fp8_scale_x2: |
5920 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5921 | VT: Node->getValueType(ResNo: 0), |
5922 | Opcodes: {0, AArch64::FSCALE_2Z2Z_H, AArch64::FSCALE_2Z2Z_S, |
5923 | AArch64::FSCALE_2Z2Z_D})) |
5924 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
5925 | return; |
5926 | case Intrinsic::aarch64_sme_fp8_scale_x4: |
5927 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
5928 | VT: Node->getValueType(ResNo: 0), |
5929 | Opcodes: {0, AArch64::FSCALE_4Z4Z_H, AArch64::FSCALE_4Z4Z_S, |
5930 | AArch64::FSCALE_4Z4Z_D})) |
5931 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
5932 | return; |
5933 | case Intrinsic::aarch64_sve_whilege_x2: |
5934 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int1>( |
5935 | VT: Node->getValueType(ResNo: 0), |
5936 | Opcodes: {AArch64::WHILEGE_2PXX_B, AArch64::WHILEGE_2PXX_H, |
5937 | AArch64::WHILEGE_2PXX_S, AArch64::WHILEGE_2PXX_D})) |
5938 | SelectWhilePair(N: Node, Opc: Op); |
5939 | return; |
5940 | case Intrinsic::aarch64_sve_whilegt_x2: |
5941 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int1>( |
5942 | VT: Node->getValueType(ResNo: 0), |
5943 | Opcodes: {AArch64::WHILEGT_2PXX_B, AArch64::WHILEGT_2PXX_H, |
5944 | AArch64::WHILEGT_2PXX_S, AArch64::WHILEGT_2PXX_D})) |
5945 | SelectWhilePair(N: Node, Opc: Op); |
5946 | return; |
5947 | case Intrinsic::aarch64_sve_whilehi_x2: |
5948 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int1>( |
5949 | VT: Node->getValueType(ResNo: 0), |
5950 | Opcodes: {AArch64::WHILEHI_2PXX_B, AArch64::WHILEHI_2PXX_H, |
5951 | AArch64::WHILEHI_2PXX_S, AArch64::WHILEHI_2PXX_D})) |
5952 | SelectWhilePair(N: Node, Opc: Op); |
5953 | return; |
5954 | case Intrinsic::aarch64_sve_whilehs_x2: |
5955 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int1>( |
5956 | VT: Node->getValueType(ResNo: 0), |
5957 | Opcodes: {AArch64::WHILEHS_2PXX_B, AArch64::WHILEHS_2PXX_H, |
5958 | AArch64::WHILEHS_2PXX_S, AArch64::WHILEHS_2PXX_D})) |
5959 | SelectWhilePair(N: Node, Opc: Op); |
5960 | return; |
5961 | case Intrinsic::aarch64_sve_whilele_x2: |
5962 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int1>( |
5963 | VT: Node->getValueType(ResNo: 0), |
5964 | Opcodes: {AArch64::WHILELE_2PXX_B, AArch64::WHILELE_2PXX_H, |
5965 | AArch64::WHILELE_2PXX_S, AArch64::WHILELE_2PXX_D})) |
5966 | SelectWhilePair(N: Node, Opc: Op); |
5967 | return; |
5968 | case Intrinsic::aarch64_sve_whilelo_x2: |
5969 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int1>( |
5970 | VT: Node->getValueType(ResNo: 0), |
5971 | Opcodes: {AArch64::WHILELO_2PXX_B, AArch64::WHILELO_2PXX_H, |
5972 | AArch64::WHILELO_2PXX_S, AArch64::WHILELO_2PXX_D})) |
5973 | SelectWhilePair(N: Node, Opc: Op); |
5974 | return; |
5975 | case Intrinsic::aarch64_sve_whilels_x2: |
5976 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int1>( |
5977 | VT: Node->getValueType(ResNo: 0), |
5978 | Opcodes: {AArch64::WHILELS_2PXX_B, AArch64::WHILELS_2PXX_H, |
5979 | AArch64::WHILELS_2PXX_S, AArch64::WHILELS_2PXX_D})) |
5980 | SelectWhilePair(N: Node, Opc: Op); |
5981 | return; |
5982 | case Intrinsic::aarch64_sve_whilelt_x2: |
5983 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int1>( |
5984 | VT: Node->getValueType(ResNo: 0), |
5985 | Opcodes: {AArch64::WHILELT_2PXX_B, AArch64::WHILELT_2PXX_H, |
5986 | AArch64::WHILELT_2PXX_S, AArch64::WHILELT_2PXX_D})) |
5987 | SelectWhilePair(N: Node, Opc: Op); |
5988 | return; |
5989 | case Intrinsic::aarch64_sve_smax_single_x2: |
5990 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5991 | VT: Node->getValueType(ResNo: 0), |
5992 | Opcodes: {AArch64::SMAX_VG2_2ZZ_B, AArch64::SMAX_VG2_2ZZ_H, |
5993 | AArch64::SMAX_VG2_2ZZ_S, AArch64::SMAX_VG2_2ZZ_D})) |
5994 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
5995 | return; |
5996 | case Intrinsic::aarch64_sve_umax_single_x2: |
5997 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
5998 | VT: Node->getValueType(ResNo: 0), |
5999 | Opcodes: {AArch64::UMAX_VG2_2ZZ_B, AArch64::UMAX_VG2_2ZZ_H, |
6000 | AArch64::UMAX_VG2_2ZZ_S, AArch64::UMAX_VG2_2ZZ_D})) |
6001 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
6002 | return; |
6003 | case Intrinsic::aarch64_sve_fmax_single_x2: |
6004 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6005 | VT: Node->getValueType(ResNo: 0), |
6006 | Opcodes: {AArch64::BFMAX_VG2_2ZZ_H, AArch64::FMAX_VG2_2ZZ_H, |
6007 | AArch64::FMAX_VG2_2ZZ_S, AArch64::FMAX_VG2_2ZZ_D})) |
6008 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
6009 | return; |
6010 | case Intrinsic::aarch64_sve_smax_single_x4: |
6011 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6012 | VT: Node->getValueType(ResNo: 0), |
6013 | Opcodes: {AArch64::SMAX_VG4_4ZZ_B, AArch64::SMAX_VG4_4ZZ_H, |
6014 | AArch64::SMAX_VG4_4ZZ_S, AArch64::SMAX_VG4_4ZZ_D})) |
6015 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
6016 | return; |
6017 | case Intrinsic::aarch64_sve_umax_single_x4: |
6018 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6019 | VT: Node->getValueType(ResNo: 0), |
6020 | Opcodes: {AArch64::UMAX_VG4_4ZZ_B, AArch64::UMAX_VG4_4ZZ_H, |
6021 | AArch64::UMAX_VG4_4ZZ_S, AArch64::UMAX_VG4_4ZZ_D})) |
6022 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
6023 | return; |
6024 | case Intrinsic::aarch64_sve_fmax_single_x4: |
6025 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6026 | VT: Node->getValueType(ResNo: 0), |
6027 | Opcodes: {AArch64::BFMAX_VG4_4ZZ_H, AArch64::FMAX_VG4_4ZZ_H, |
6028 | AArch64::FMAX_VG4_4ZZ_S, AArch64::FMAX_VG4_4ZZ_D})) |
6029 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
6030 | return; |
6031 | case Intrinsic::aarch64_sve_smin_single_x2: |
6032 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6033 | VT: Node->getValueType(ResNo: 0), |
6034 | Opcodes: {AArch64::SMIN_VG2_2ZZ_B, AArch64::SMIN_VG2_2ZZ_H, |
6035 | AArch64::SMIN_VG2_2ZZ_S, AArch64::SMIN_VG2_2ZZ_D})) |
6036 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
6037 | return; |
6038 | case Intrinsic::aarch64_sve_umin_single_x2: |
6039 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6040 | VT: Node->getValueType(ResNo: 0), |
6041 | Opcodes: {AArch64::UMIN_VG2_2ZZ_B, AArch64::UMIN_VG2_2ZZ_H, |
6042 | AArch64::UMIN_VG2_2ZZ_S, AArch64::UMIN_VG2_2ZZ_D})) |
6043 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
6044 | return; |
6045 | case Intrinsic::aarch64_sve_fmin_single_x2: |
6046 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6047 | VT: Node->getValueType(ResNo: 0), |
6048 | Opcodes: {AArch64::BFMIN_VG2_2ZZ_H, AArch64::FMIN_VG2_2ZZ_H, |
6049 | AArch64::FMIN_VG2_2ZZ_S, AArch64::FMIN_VG2_2ZZ_D})) |
6050 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
6051 | return; |
6052 | case Intrinsic::aarch64_sve_smin_single_x4: |
6053 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6054 | VT: Node->getValueType(ResNo: 0), |
6055 | Opcodes: {AArch64::SMIN_VG4_4ZZ_B, AArch64::SMIN_VG4_4ZZ_H, |
6056 | AArch64::SMIN_VG4_4ZZ_S, AArch64::SMIN_VG4_4ZZ_D})) |
6057 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
6058 | return; |
6059 | case Intrinsic::aarch64_sve_umin_single_x4: |
6060 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6061 | VT: Node->getValueType(ResNo: 0), |
6062 | Opcodes: {AArch64::UMIN_VG4_4ZZ_B, AArch64::UMIN_VG4_4ZZ_H, |
6063 | AArch64::UMIN_VG4_4ZZ_S, AArch64::UMIN_VG4_4ZZ_D})) |
6064 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
6065 | return; |
6066 | case Intrinsic::aarch64_sve_fmin_single_x4: |
6067 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6068 | VT: Node->getValueType(ResNo: 0), |
6069 | Opcodes: {AArch64::BFMIN_VG4_4ZZ_H, AArch64::FMIN_VG4_4ZZ_H, |
6070 | AArch64::FMIN_VG4_4ZZ_S, AArch64::FMIN_VG4_4ZZ_D})) |
6071 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
6072 | return; |
6073 | case Intrinsic::aarch64_sve_smax_x2: |
6074 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6075 | VT: Node->getValueType(ResNo: 0), |
6076 | Opcodes: {AArch64::SMAX_VG2_2Z2Z_B, AArch64::SMAX_VG2_2Z2Z_H, |
6077 | AArch64::SMAX_VG2_2Z2Z_S, AArch64::SMAX_VG2_2Z2Z_D})) |
6078 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
6079 | return; |
6080 | case Intrinsic::aarch64_sve_umax_x2: |
6081 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6082 | VT: Node->getValueType(ResNo: 0), |
6083 | Opcodes: {AArch64::UMAX_VG2_2Z2Z_B, AArch64::UMAX_VG2_2Z2Z_H, |
6084 | AArch64::UMAX_VG2_2Z2Z_S, AArch64::UMAX_VG2_2Z2Z_D})) |
6085 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
6086 | return; |
6087 | case Intrinsic::aarch64_sve_fmax_x2: |
6088 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6089 | VT: Node->getValueType(ResNo: 0), |
6090 | Opcodes: {AArch64::BFMAX_VG2_2Z2Z_H, AArch64::FMAX_VG2_2Z2Z_H, |
6091 | AArch64::FMAX_VG2_2Z2Z_S, AArch64::FMAX_VG2_2Z2Z_D})) |
6092 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
6093 | return; |
6094 | case Intrinsic::aarch64_sve_smax_x4: |
6095 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6096 | VT: Node->getValueType(ResNo: 0), |
6097 | Opcodes: {AArch64::SMAX_VG4_4Z4Z_B, AArch64::SMAX_VG4_4Z4Z_H, |
6098 | AArch64::SMAX_VG4_4Z4Z_S, AArch64::SMAX_VG4_4Z4Z_D})) |
6099 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
6100 | return; |
6101 | case Intrinsic::aarch64_sve_umax_x4: |
6102 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6103 | VT: Node->getValueType(ResNo: 0), |
6104 | Opcodes: {AArch64::UMAX_VG4_4Z4Z_B, AArch64::UMAX_VG4_4Z4Z_H, |
6105 | AArch64::UMAX_VG4_4Z4Z_S, AArch64::UMAX_VG4_4Z4Z_D})) |
6106 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
6107 | return; |
6108 | case Intrinsic::aarch64_sve_fmax_x4: |
6109 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6110 | VT: Node->getValueType(ResNo: 0), |
6111 | Opcodes: {AArch64::BFMAX_VG4_4Z2Z_H, AArch64::FMAX_VG4_4Z4Z_H, |
6112 | AArch64::FMAX_VG4_4Z4Z_S, AArch64::FMAX_VG4_4Z4Z_D})) |
6113 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
6114 | return; |
6115 | case Intrinsic::aarch64_sme_famax_x2: |
6116 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6117 | VT: Node->getValueType(ResNo: 0), |
6118 | Opcodes: {0, AArch64::FAMAX_2Z2Z_H, AArch64::FAMAX_2Z2Z_S, |
6119 | AArch64::FAMAX_2Z2Z_D})) |
6120 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
6121 | return; |
6122 | case Intrinsic::aarch64_sme_famax_x4: |
6123 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6124 | VT: Node->getValueType(ResNo: 0), |
6125 | Opcodes: {0, AArch64::FAMAX_4Z4Z_H, AArch64::FAMAX_4Z4Z_S, |
6126 | AArch64::FAMAX_4Z4Z_D})) |
6127 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
6128 | return; |
6129 | case Intrinsic::aarch64_sme_famin_x2: |
6130 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6131 | VT: Node->getValueType(ResNo: 0), |
6132 | Opcodes: {0, AArch64::FAMIN_2Z2Z_H, AArch64::FAMIN_2Z2Z_S, |
6133 | AArch64::FAMIN_2Z2Z_D})) |
6134 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
6135 | return; |
6136 | case Intrinsic::aarch64_sme_famin_x4: |
6137 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6138 | VT: Node->getValueType(ResNo: 0), |
6139 | Opcodes: {0, AArch64::FAMIN_4Z4Z_H, AArch64::FAMIN_4Z4Z_S, |
6140 | AArch64::FAMIN_4Z4Z_D})) |
6141 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
6142 | return; |
6143 | case Intrinsic::aarch64_sve_smin_x2: |
6144 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6145 | VT: Node->getValueType(ResNo: 0), |
6146 | Opcodes: {AArch64::SMIN_VG2_2Z2Z_B, AArch64::SMIN_VG2_2Z2Z_H, |
6147 | AArch64::SMIN_VG2_2Z2Z_S, AArch64::SMIN_VG2_2Z2Z_D})) |
6148 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
6149 | return; |
6150 | case Intrinsic::aarch64_sve_umin_x2: |
6151 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6152 | VT: Node->getValueType(ResNo: 0), |
6153 | Opcodes: {AArch64::UMIN_VG2_2Z2Z_B, AArch64::UMIN_VG2_2Z2Z_H, |
6154 | AArch64::UMIN_VG2_2Z2Z_S, AArch64::UMIN_VG2_2Z2Z_D})) |
6155 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
6156 | return; |
6157 | case Intrinsic::aarch64_sve_fmin_x2: |
6158 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6159 | VT: Node->getValueType(ResNo: 0), |
6160 | Opcodes: {AArch64::BFMIN_VG2_2Z2Z_H, AArch64::FMIN_VG2_2Z2Z_H, |
6161 | AArch64::FMIN_VG2_2Z2Z_S, AArch64::FMIN_VG2_2Z2Z_D})) |
6162 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
6163 | return; |
6164 | case Intrinsic::aarch64_sve_smin_x4: |
6165 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6166 | VT: Node->getValueType(ResNo: 0), |
6167 | Opcodes: {AArch64::SMIN_VG4_4Z4Z_B, AArch64::SMIN_VG4_4Z4Z_H, |
6168 | AArch64::SMIN_VG4_4Z4Z_S, AArch64::SMIN_VG4_4Z4Z_D})) |
6169 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
6170 | return; |
6171 | case Intrinsic::aarch64_sve_umin_x4: |
6172 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6173 | VT: Node->getValueType(ResNo: 0), |
6174 | Opcodes: {AArch64::UMIN_VG4_4Z4Z_B, AArch64::UMIN_VG4_4Z4Z_H, |
6175 | AArch64::UMIN_VG4_4Z4Z_S, AArch64::UMIN_VG4_4Z4Z_D})) |
6176 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
6177 | return; |
6178 | case Intrinsic::aarch64_sve_fmin_x4: |
6179 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6180 | VT: Node->getValueType(ResNo: 0), |
6181 | Opcodes: {AArch64::BFMIN_VG4_4Z2Z_H, AArch64::FMIN_VG4_4Z4Z_H, |
6182 | AArch64::FMIN_VG4_4Z4Z_S, AArch64::FMIN_VG4_4Z4Z_D})) |
6183 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
6184 | return; |
6185 | case Intrinsic::aarch64_sve_fmaxnm_single_x2 : |
6186 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6187 | VT: Node->getValueType(ResNo: 0), |
6188 | Opcodes: {AArch64::BFMAXNM_VG2_2ZZ_H, AArch64::FMAXNM_VG2_2ZZ_H, |
6189 | AArch64::FMAXNM_VG2_2ZZ_S, AArch64::FMAXNM_VG2_2ZZ_D})) |
6190 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
6191 | return; |
6192 | case Intrinsic::aarch64_sve_fmaxnm_single_x4 : |
6193 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6194 | VT: Node->getValueType(ResNo: 0), |
6195 | Opcodes: {AArch64::BFMAXNM_VG4_4ZZ_H, AArch64::FMAXNM_VG4_4ZZ_H, |
6196 | AArch64::FMAXNM_VG4_4ZZ_S, AArch64::FMAXNM_VG4_4ZZ_D})) |
6197 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
6198 | return; |
6199 | case Intrinsic::aarch64_sve_fminnm_single_x2: |
6200 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6201 | VT: Node->getValueType(ResNo: 0), |
6202 | Opcodes: {AArch64::BFMINNM_VG2_2ZZ_H, AArch64::FMINNM_VG2_2ZZ_H, |
6203 | AArch64::FMINNM_VG2_2ZZ_S, AArch64::FMINNM_VG2_2ZZ_D})) |
6204 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
6205 | return; |
6206 | case Intrinsic::aarch64_sve_fminnm_single_x4: |
6207 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6208 | VT: Node->getValueType(ResNo: 0), |
6209 | Opcodes: {AArch64::BFMINNM_VG4_4ZZ_H, AArch64::FMINNM_VG4_4ZZ_H, |
6210 | AArch64::FMINNM_VG4_4ZZ_S, AArch64::FMINNM_VG4_4ZZ_D})) |
6211 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
6212 | return; |
6213 | case Intrinsic::aarch64_sve_fmaxnm_x2: |
6214 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6215 | VT: Node->getValueType(ResNo: 0), |
6216 | Opcodes: {AArch64::BFMAXNM_VG2_2Z2Z_H, AArch64::FMAXNM_VG2_2Z2Z_H, |
6217 | AArch64::FMAXNM_VG2_2Z2Z_S, AArch64::FMAXNM_VG2_2Z2Z_D})) |
6218 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
6219 | return; |
6220 | case Intrinsic::aarch64_sve_fmaxnm_x4: |
6221 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6222 | VT: Node->getValueType(ResNo: 0), |
6223 | Opcodes: {AArch64::BFMAXNM_VG4_4Z2Z_H, AArch64::FMAXNM_VG4_4Z4Z_H, |
6224 | AArch64::FMAXNM_VG4_4Z4Z_S, AArch64::FMAXNM_VG4_4Z4Z_D})) |
6225 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
6226 | return; |
6227 | case Intrinsic::aarch64_sve_fminnm_x2: |
6228 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6229 | VT: Node->getValueType(ResNo: 0), |
6230 | Opcodes: {AArch64::BFMINNM_VG2_2Z2Z_H, AArch64::FMINNM_VG2_2Z2Z_H, |
6231 | AArch64::FMINNM_VG2_2Z2Z_S, AArch64::FMINNM_VG2_2Z2Z_D})) |
6232 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op); |
6233 | return; |
6234 | case Intrinsic::aarch64_sve_fminnm_x4: |
6235 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6236 | VT: Node->getValueType(ResNo: 0), |
6237 | Opcodes: {AArch64::BFMINNM_VG4_4Z2Z_H, AArch64::FMINNM_VG4_4Z4Z_H, |
6238 | AArch64::FMINNM_VG4_4Z4Z_S, AArch64::FMINNM_VG4_4Z4Z_D})) |
6239 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op); |
6240 | return; |
6241 | case Intrinsic::aarch64_sve_fcvtzs_x2: |
6242 | SelectCVTIntrinsic(N: Node, NumVecs: 2, Opcode: AArch64::FCVTZS_2Z2Z_StoS); |
6243 | return; |
6244 | case Intrinsic::aarch64_sve_scvtf_x2: |
6245 | SelectCVTIntrinsic(N: Node, NumVecs: 2, Opcode: AArch64::SCVTF_2Z2Z_StoS); |
6246 | return; |
6247 | case Intrinsic::aarch64_sve_fcvtzu_x2: |
6248 | SelectCVTIntrinsic(N: Node, NumVecs: 2, Opcode: AArch64::FCVTZU_2Z2Z_StoS); |
6249 | return; |
6250 | case Intrinsic::aarch64_sve_ucvtf_x2: |
6251 | SelectCVTIntrinsic(N: Node, NumVecs: 2, Opcode: AArch64::UCVTF_2Z2Z_StoS); |
6252 | return; |
6253 | case Intrinsic::aarch64_sve_fcvtzs_x4: |
6254 | SelectCVTIntrinsic(N: Node, NumVecs: 4, Opcode: AArch64::FCVTZS_4Z4Z_StoS); |
6255 | return; |
6256 | case Intrinsic::aarch64_sve_scvtf_x4: |
6257 | SelectCVTIntrinsic(N: Node, NumVecs: 4, Opcode: AArch64::SCVTF_4Z4Z_StoS); |
6258 | return; |
6259 | case Intrinsic::aarch64_sve_fcvtzu_x4: |
6260 | SelectCVTIntrinsic(N: Node, NumVecs: 4, Opcode: AArch64::FCVTZU_4Z4Z_StoS); |
6261 | return; |
6262 | case Intrinsic::aarch64_sve_ucvtf_x4: |
6263 | SelectCVTIntrinsic(N: Node, NumVecs: 4, Opcode: AArch64::UCVTF_4Z4Z_StoS); |
6264 | return; |
6265 | case Intrinsic::aarch64_sve_fcvt_widen_x2: |
6266 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 2, IsTupleInput: false, Opc: AArch64::FCVT_2ZZ_H_S); |
6267 | return; |
6268 | case Intrinsic::aarch64_sve_fcvtl_widen_x2: |
6269 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 2, IsTupleInput: false, Opc: AArch64::FCVTL_2ZZ_H_S); |
6270 | return; |
6271 | case Intrinsic::aarch64_sve_sclamp_single_x2: |
6272 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6273 | VT: Node->getValueType(ResNo: 0), |
6274 | Opcodes: {AArch64::SCLAMP_VG2_2Z2Z_B, AArch64::SCLAMP_VG2_2Z2Z_H, |
6275 | AArch64::SCLAMP_VG2_2Z2Z_S, AArch64::SCLAMP_VG2_2Z2Z_D})) |
6276 | SelectClamp(N: Node, NumVecs: 2, Op); |
6277 | return; |
6278 | case Intrinsic::aarch64_sve_uclamp_single_x2: |
6279 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6280 | VT: Node->getValueType(ResNo: 0), |
6281 | Opcodes: {AArch64::UCLAMP_VG2_2Z2Z_B, AArch64::UCLAMP_VG2_2Z2Z_H, |
6282 | AArch64::UCLAMP_VG2_2Z2Z_S, AArch64::UCLAMP_VG2_2Z2Z_D})) |
6283 | SelectClamp(N: Node, NumVecs: 2, Op); |
6284 | return; |
6285 | case Intrinsic::aarch64_sve_fclamp_single_x2: |
6286 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6287 | VT: Node->getValueType(ResNo: 0), |
6288 | Opcodes: {0, AArch64::FCLAMP_VG2_2Z2Z_H, AArch64::FCLAMP_VG2_2Z2Z_S, |
6289 | AArch64::FCLAMP_VG2_2Z2Z_D})) |
6290 | SelectClamp(N: Node, NumVecs: 2, Op); |
6291 | return; |
6292 | case Intrinsic::aarch64_sve_bfclamp_single_x2: |
6293 | SelectClamp(N: Node, NumVecs: 2, Op: AArch64::BFCLAMP_VG2_2ZZZ_H); |
6294 | return; |
6295 | case Intrinsic::aarch64_sve_sclamp_single_x4: |
6296 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6297 | VT: Node->getValueType(ResNo: 0), |
6298 | Opcodes: {AArch64::SCLAMP_VG4_4Z4Z_B, AArch64::SCLAMP_VG4_4Z4Z_H, |
6299 | AArch64::SCLAMP_VG4_4Z4Z_S, AArch64::SCLAMP_VG4_4Z4Z_D})) |
6300 | SelectClamp(N: Node, NumVecs: 4, Op); |
6301 | return; |
6302 | case Intrinsic::aarch64_sve_uclamp_single_x4: |
6303 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6304 | VT: Node->getValueType(ResNo: 0), |
6305 | Opcodes: {AArch64::UCLAMP_VG4_4Z4Z_B, AArch64::UCLAMP_VG4_4Z4Z_H, |
6306 | AArch64::UCLAMP_VG4_4Z4Z_S, AArch64::UCLAMP_VG4_4Z4Z_D})) |
6307 | SelectClamp(N: Node, NumVecs: 4, Op); |
6308 | return; |
6309 | case Intrinsic::aarch64_sve_fclamp_single_x4: |
6310 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::FP>( |
6311 | VT: Node->getValueType(ResNo: 0), |
6312 | Opcodes: {0, AArch64::FCLAMP_VG4_4Z4Z_H, AArch64::FCLAMP_VG4_4Z4Z_S, |
6313 | AArch64::FCLAMP_VG4_4Z4Z_D})) |
6314 | SelectClamp(N: Node, NumVecs: 4, Op); |
6315 | return; |
6316 | case Intrinsic::aarch64_sve_bfclamp_single_x4: |
6317 | SelectClamp(N: Node, NumVecs: 4, Op: AArch64::BFCLAMP_VG4_4ZZZ_H); |
6318 | return; |
6319 | case Intrinsic::aarch64_sve_add_single_x2: |
6320 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6321 | VT: Node->getValueType(ResNo: 0), |
6322 | Opcodes: {AArch64::ADD_VG2_2ZZ_B, AArch64::ADD_VG2_2ZZ_H, |
6323 | AArch64::ADD_VG2_2ZZ_S, AArch64::ADD_VG2_2ZZ_D})) |
6324 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: false, Opcode: Op); |
6325 | return; |
6326 | case Intrinsic::aarch64_sve_add_single_x4: |
6327 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6328 | VT: Node->getValueType(ResNo: 0), |
6329 | Opcodes: {AArch64::ADD_VG4_4ZZ_B, AArch64::ADD_VG4_4ZZ_H, |
6330 | AArch64::ADD_VG4_4ZZ_S, AArch64::ADD_VG4_4ZZ_D})) |
6331 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: false, Opcode: Op); |
6332 | return; |
6333 | case Intrinsic::aarch64_sve_zip_x2: |
6334 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
6335 | VT: Node->getValueType(ResNo: 0), |
6336 | Opcodes: {AArch64::ZIP_VG2_2ZZZ_B, AArch64::ZIP_VG2_2ZZZ_H, |
6337 | AArch64::ZIP_VG2_2ZZZ_S, AArch64::ZIP_VG2_2ZZZ_D})) |
6338 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 2, /*IsTupleInput=*/false, Opc: Op); |
6339 | return; |
6340 | case Intrinsic::aarch64_sve_zipq_x2: |
6341 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 2, /*IsTupleInput=*/false, |
6342 | Opc: AArch64::ZIP_VG2_2ZZZ_Q); |
6343 | return; |
6344 | case Intrinsic::aarch64_sve_zip_x4: |
6345 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
6346 | VT: Node->getValueType(ResNo: 0), |
6347 | Opcodes: {AArch64::ZIP_VG4_4Z4Z_B, AArch64::ZIP_VG4_4Z4Z_H, |
6348 | AArch64::ZIP_VG4_4Z4Z_S, AArch64::ZIP_VG4_4Z4Z_D})) |
6349 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 4, /*IsTupleInput=*/true, Opc: Op); |
6350 | return; |
6351 | case Intrinsic::aarch64_sve_zipq_x4: |
6352 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 4, /*IsTupleInput=*/true, |
6353 | Opc: AArch64::ZIP_VG4_4Z4Z_Q); |
6354 | return; |
6355 | case Intrinsic::aarch64_sve_uzp_x2: |
6356 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
6357 | VT: Node->getValueType(ResNo: 0), |
6358 | Opcodes: {AArch64::UZP_VG2_2ZZZ_B, AArch64::UZP_VG2_2ZZZ_H, |
6359 | AArch64::UZP_VG2_2ZZZ_S, AArch64::UZP_VG2_2ZZZ_D})) |
6360 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 2, /*IsTupleInput=*/false, Opc: Op); |
6361 | return; |
6362 | case Intrinsic::aarch64_sve_uzpq_x2: |
6363 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 2, /*IsTupleInput=*/false, |
6364 | Opc: AArch64::UZP_VG2_2ZZZ_Q); |
6365 | return; |
6366 | case Intrinsic::aarch64_sve_uzp_x4: |
6367 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
6368 | VT: Node->getValueType(ResNo: 0), |
6369 | Opcodes: {AArch64::UZP_VG4_4Z4Z_B, AArch64::UZP_VG4_4Z4Z_H, |
6370 | AArch64::UZP_VG4_4Z4Z_S, AArch64::UZP_VG4_4Z4Z_D})) |
6371 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 4, /*IsTupleInput=*/true, Opc: Op); |
6372 | return; |
6373 | case Intrinsic::aarch64_sve_uzpq_x4: |
6374 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 4, /*IsTupleInput=*/true, |
6375 | Opc: AArch64::UZP_VG4_4Z4Z_Q); |
6376 | return; |
6377 | case Intrinsic::aarch64_sve_sel_x2: |
6378 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
6379 | VT: Node->getValueType(ResNo: 0), |
6380 | Opcodes: {AArch64::SEL_VG2_2ZC2Z2Z_B, AArch64::SEL_VG2_2ZC2Z2Z_H, |
6381 | AArch64::SEL_VG2_2ZC2Z2Z_S, AArch64::SEL_VG2_2ZC2Z2Z_D})) |
6382 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 2, IsZmMulti: true, Opcode: Op, /*HasPred=*/true); |
6383 | return; |
6384 | case Intrinsic::aarch64_sve_sel_x4: |
6385 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
6386 | VT: Node->getValueType(ResNo: 0), |
6387 | Opcodes: {AArch64::SEL_VG4_4ZC4Z4Z_B, AArch64::SEL_VG4_4ZC4Z4Z_H, |
6388 | AArch64::SEL_VG4_4ZC4Z4Z_S, AArch64::SEL_VG4_4ZC4Z4Z_D})) |
6389 | SelectDestructiveMultiIntrinsic(N: Node, NumVecs: 4, IsZmMulti: true, Opcode: Op, /*HasPred=*/true); |
6390 | return; |
6391 | case Intrinsic::aarch64_sve_frinta_x2: |
6392 | SelectFrintFromVT(N: Node, NumVecs: 2, Opcode: AArch64::FRINTA_2Z2Z_S); |
6393 | return; |
6394 | case Intrinsic::aarch64_sve_frinta_x4: |
6395 | SelectFrintFromVT(N: Node, NumVecs: 4, Opcode: AArch64::FRINTA_4Z4Z_S); |
6396 | return; |
6397 | case Intrinsic::aarch64_sve_frintm_x2: |
6398 | SelectFrintFromVT(N: Node, NumVecs: 2, Opcode: AArch64::FRINTM_2Z2Z_S); |
6399 | return; |
6400 | case Intrinsic::aarch64_sve_frintm_x4: |
6401 | SelectFrintFromVT(N: Node, NumVecs: 4, Opcode: AArch64::FRINTM_4Z4Z_S); |
6402 | return; |
6403 | case Intrinsic::aarch64_sve_frintn_x2: |
6404 | SelectFrintFromVT(N: Node, NumVecs: 2, Opcode: AArch64::FRINTN_2Z2Z_S); |
6405 | return; |
6406 | case Intrinsic::aarch64_sve_frintn_x4: |
6407 | SelectFrintFromVT(N: Node, NumVecs: 4, Opcode: AArch64::FRINTN_4Z4Z_S); |
6408 | return; |
6409 | case Intrinsic::aarch64_sve_frintp_x2: |
6410 | SelectFrintFromVT(N: Node, NumVecs: 2, Opcode: AArch64::FRINTP_2Z2Z_S); |
6411 | return; |
6412 | case Intrinsic::aarch64_sve_frintp_x4: |
6413 | SelectFrintFromVT(N: Node, NumVecs: 4, Opcode: AArch64::FRINTP_4Z4Z_S); |
6414 | return; |
6415 | case Intrinsic::aarch64_sve_sunpk_x2: |
6416 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6417 | VT: Node->getValueType(ResNo: 0), |
6418 | Opcodes: {0, AArch64::SUNPK_VG2_2ZZ_H, AArch64::SUNPK_VG2_2ZZ_S, |
6419 | AArch64::SUNPK_VG2_2ZZ_D})) |
6420 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 2, /*IsTupleInput=*/false, Opc: Op); |
6421 | return; |
6422 | case Intrinsic::aarch64_sve_uunpk_x2: |
6423 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6424 | VT: Node->getValueType(ResNo: 0), |
6425 | Opcodes: {0, AArch64::UUNPK_VG2_2ZZ_H, AArch64::UUNPK_VG2_2ZZ_S, |
6426 | AArch64::UUNPK_VG2_2ZZ_D})) |
6427 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 2, /*IsTupleInput=*/false, Opc: Op); |
6428 | return; |
6429 | case Intrinsic::aarch64_sve_sunpk_x4: |
6430 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6431 | VT: Node->getValueType(ResNo: 0), |
6432 | Opcodes: {0, AArch64::SUNPK_VG4_4Z2Z_H, AArch64::SUNPK_VG4_4Z2Z_S, |
6433 | AArch64::SUNPK_VG4_4Z2Z_D})) |
6434 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 4, /*IsTupleInput=*/true, Opc: Op); |
6435 | return; |
6436 | case Intrinsic::aarch64_sve_uunpk_x4: |
6437 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::Int>( |
6438 | VT: Node->getValueType(ResNo: 0), |
6439 | Opcodes: {0, AArch64::UUNPK_VG4_4Z2Z_H, AArch64::UUNPK_VG4_4Z2Z_S, |
6440 | AArch64::UUNPK_VG4_4Z2Z_D})) |
6441 | SelectUnaryMultiIntrinsic(N: Node, NumOutVecs: 4, /*IsTupleInput=*/true, Opc: Op); |
6442 | return; |
6443 | case Intrinsic::aarch64_sve_pext_x2: { |
6444 | if (auto Op = SelectOpcodeFromVT<SelectTypeKind::AnyType>( |
6445 | VT: Node->getValueType(ResNo: 0), |
6446 | Opcodes: {AArch64::PEXT_2PCI_B, AArch64::PEXT_2PCI_H, AArch64::PEXT_2PCI_S, |
6447 | AArch64::PEXT_2PCI_D})) |
6448 | SelectPExtPair(N: Node, Opc: Op); |
6449 | return; |
6450 | } |
6451 | } |
6452 | break; |
6453 | } |
6454 | case ISD::INTRINSIC_VOID: { |
6455 | unsigned IntNo = Node->getConstantOperandVal(Num: 1); |
6456 | if (Node->getNumOperands() >= 3) |
6457 | VT = Node->getOperand(Num: 2)->getValueType(ResNo: 0); |
6458 | switch (IntNo) { |
6459 | default: |
6460 | break; |
6461 | case Intrinsic::aarch64_neon_st1x2: { |
6462 | if (VT == MVT::v8i8) { |
6463 | SelectStore(N: Node, NumVecs: 2, Opc: AArch64::ST1Twov8b); |
6464 | return; |
6465 | } else if (VT == MVT::v16i8) { |
6466 | SelectStore(N: Node, NumVecs: 2, Opc: AArch64::ST1Twov16b); |
6467 | return; |
6468 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || |
6469 | VT == MVT::v4bf16) { |
6470 | SelectStore(N: Node, NumVecs: 2, Opc: AArch64::ST1Twov4h); |
6471 | return; |
6472 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || |
6473 | VT == MVT::v8bf16) { |
6474 | SelectStore(N: Node, NumVecs: 2, Opc: AArch64::ST1Twov8h); |
6475 | return; |
6476 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6477 | SelectStore(N: Node, NumVecs: 2, Opc: AArch64::ST1Twov2s); |
6478 | return; |
6479 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6480 | SelectStore(N: Node, NumVecs: 2, Opc: AArch64::ST1Twov4s); |
6481 | return; |
6482 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6483 | SelectStore(N: Node, NumVecs: 2, Opc: AArch64::ST1Twov2d); |
6484 | return; |
6485 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6486 | SelectStore(N: Node, NumVecs: 2, Opc: AArch64::ST1Twov1d); |
6487 | return; |
6488 | } |
6489 | break; |
6490 | } |
6491 | case Intrinsic::aarch64_neon_st1x3: { |
6492 | if (VT == MVT::v8i8) { |
6493 | SelectStore(N: Node, NumVecs: 3, Opc: AArch64::ST1Threev8b); |
6494 | return; |
6495 | } else if (VT == MVT::v16i8) { |
6496 | SelectStore(N: Node, NumVecs: 3, Opc: AArch64::ST1Threev16b); |
6497 | return; |
6498 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || |
6499 | VT == MVT::v4bf16) { |
6500 | SelectStore(N: Node, NumVecs: 3, Opc: AArch64::ST1Threev4h); |
6501 | return; |
6502 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || |
6503 | VT == MVT::v8bf16) { |
6504 | SelectStore(N: Node, NumVecs: 3, Opc: AArch64::ST1Threev8h); |
6505 | return; |
6506 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6507 | SelectStore(N: Node, NumVecs: 3, Opc: AArch64::ST1Threev2s); |
6508 | return; |
6509 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6510 | SelectStore(N: Node, NumVecs: 3, Opc: AArch64::ST1Threev4s); |
6511 | return; |
6512 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6513 | SelectStore(N: Node, NumVecs: 3, Opc: AArch64::ST1Threev2d); |
6514 | return; |
6515 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6516 | SelectStore(N: Node, NumVecs: 3, Opc: AArch64::ST1Threev1d); |
6517 | return; |
6518 | } |
6519 | break; |
6520 | } |
6521 | case Intrinsic::aarch64_neon_st1x4: { |
6522 | if (VT == MVT::v8i8) { |
6523 | SelectStore(N: Node, NumVecs: 4, Opc: AArch64::ST1Fourv8b); |
6524 | return; |
6525 | } else if (VT == MVT::v16i8) { |
6526 | SelectStore(N: Node, NumVecs: 4, Opc: AArch64::ST1Fourv16b); |
6527 | return; |
6528 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || |
6529 | VT == MVT::v4bf16) { |
6530 | SelectStore(N: Node, NumVecs: 4, Opc: AArch64::ST1Fourv4h); |
6531 | return; |
6532 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || |
6533 | VT == MVT::v8bf16) { |
6534 | SelectStore(N: Node, NumVecs: 4, Opc: AArch64::ST1Fourv8h); |
6535 | return; |
6536 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6537 | SelectStore(N: Node, NumVecs: 4, Opc: AArch64::ST1Fourv2s); |
6538 | return; |
6539 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6540 | SelectStore(N: Node, NumVecs: 4, Opc: AArch64::ST1Fourv4s); |
6541 | return; |
6542 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6543 | SelectStore(N: Node, NumVecs: 4, Opc: AArch64::ST1Fourv2d); |
6544 | return; |
6545 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6546 | SelectStore(N: Node, NumVecs: 4, Opc: AArch64::ST1Fourv1d); |
6547 | return; |
6548 | } |
6549 | break; |
6550 | } |
6551 | case Intrinsic::aarch64_neon_st2: { |
6552 | if (VT == MVT::v8i8) { |
6553 | SelectStore(N: Node, NumVecs: 2, Opc: AArch64::ST2Twov8b); |
6554 | return; |
6555 | } else if (VT == MVT::v16i8) { |
6556 | SelectStore(N: Node, NumVecs: 2, Opc: AArch64::ST2Twov16b); |
6557 | return; |
6558 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || |
6559 | VT == MVT::v4bf16) { |
6560 | SelectStore(N: Node, NumVecs: 2, Opc: AArch64::ST2Twov4h); |
6561 | return; |
6562 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || |
6563 | VT == MVT::v8bf16) { |
6564 | SelectStore(N: Node, NumVecs: 2, Opc: AArch64::ST2Twov8h); |
6565 | return; |
6566 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6567 | SelectStore(N: Node, NumVecs: 2, Opc: AArch64::ST2Twov2s); |
6568 | return; |
6569 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6570 | SelectStore(N: Node, NumVecs: 2, Opc: AArch64::ST2Twov4s); |
6571 | return; |
6572 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6573 | SelectStore(N: Node, NumVecs: 2, Opc: AArch64::ST2Twov2d); |
6574 | return; |
6575 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6576 | SelectStore(N: Node, NumVecs: 2, Opc: AArch64::ST1Twov1d); |
6577 | return; |
6578 | } |
6579 | break; |
6580 | } |
6581 | case Intrinsic::aarch64_neon_st3: { |
6582 | if (VT == MVT::v8i8) { |
6583 | SelectStore(N: Node, NumVecs: 3, Opc: AArch64::ST3Threev8b); |
6584 | return; |
6585 | } else if (VT == MVT::v16i8) { |
6586 | SelectStore(N: Node, NumVecs: 3, Opc: AArch64::ST3Threev16b); |
6587 | return; |
6588 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || |
6589 | VT == MVT::v4bf16) { |
6590 | SelectStore(N: Node, NumVecs: 3, Opc: AArch64::ST3Threev4h); |
6591 | return; |
6592 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || |
6593 | VT == MVT::v8bf16) { |
6594 | SelectStore(N: Node, NumVecs: 3, Opc: AArch64::ST3Threev8h); |
6595 | return; |
6596 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6597 | SelectStore(N: Node, NumVecs: 3, Opc: AArch64::ST3Threev2s); |
6598 | return; |
6599 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6600 | SelectStore(N: Node, NumVecs: 3, Opc: AArch64::ST3Threev4s); |
6601 | return; |
6602 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6603 | SelectStore(N: Node, NumVecs: 3, Opc: AArch64::ST3Threev2d); |
6604 | return; |
6605 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6606 | SelectStore(N: Node, NumVecs: 3, Opc: AArch64::ST1Threev1d); |
6607 | return; |
6608 | } |
6609 | break; |
6610 | } |
6611 | case Intrinsic::aarch64_neon_st4: { |
6612 | if (VT == MVT::v8i8) { |
6613 | SelectStore(N: Node, NumVecs: 4, Opc: AArch64::ST4Fourv8b); |
6614 | return; |
6615 | } else if (VT == MVT::v16i8) { |
6616 | SelectStore(N: Node, NumVecs: 4, Opc: AArch64::ST4Fourv16b); |
6617 | return; |
6618 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || |
6619 | VT == MVT::v4bf16) { |
6620 | SelectStore(N: Node, NumVecs: 4, Opc: AArch64::ST4Fourv4h); |
6621 | return; |
6622 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || |
6623 | VT == MVT::v8bf16) { |
6624 | SelectStore(N: Node, NumVecs: 4, Opc: AArch64::ST4Fourv8h); |
6625 | return; |
6626 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6627 | SelectStore(N: Node, NumVecs: 4, Opc: AArch64::ST4Fourv2s); |
6628 | return; |
6629 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6630 | SelectStore(N: Node, NumVecs: 4, Opc: AArch64::ST4Fourv4s); |
6631 | return; |
6632 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6633 | SelectStore(N: Node, NumVecs: 4, Opc: AArch64::ST4Fourv2d); |
6634 | return; |
6635 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6636 | SelectStore(N: Node, NumVecs: 4, Opc: AArch64::ST1Fourv1d); |
6637 | return; |
6638 | } |
6639 | break; |
6640 | } |
6641 | case Intrinsic::aarch64_neon_st2lane: { |
6642 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
6643 | SelectStoreLane(N: Node, NumVecs: 2, Opc: AArch64::ST2i8); |
6644 | return; |
6645 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
6646 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
6647 | SelectStoreLane(N: Node, NumVecs: 2, Opc: AArch64::ST2i16); |
6648 | return; |
6649 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
6650 | VT == MVT::v2f32) { |
6651 | SelectStoreLane(N: Node, NumVecs: 2, Opc: AArch64::ST2i32); |
6652 | return; |
6653 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
6654 | VT == MVT::v1f64) { |
6655 | SelectStoreLane(N: Node, NumVecs: 2, Opc: AArch64::ST2i64); |
6656 | return; |
6657 | } |
6658 | break; |
6659 | } |
6660 | case Intrinsic::aarch64_neon_st3lane: { |
6661 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
6662 | SelectStoreLane(N: Node, NumVecs: 3, Opc: AArch64::ST3i8); |
6663 | return; |
6664 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
6665 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
6666 | SelectStoreLane(N: Node, NumVecs: 3, Opc: AArch64::ST3i16); |
6667 | return; |
6668 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
6669 | VT == MVT::v2f32) { |
6670 | SelectStoreLane(N: Node, NumVecs: 3, Opc: AArch64::ST3i32); |
6671 | return; |
6672 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
6673 | VT == MVT::v1f64) { |
6674 | SelectStoreLane(N: Node, NumVecs: 3, Opc: AArch64::ST3i64); |
6675 | return; |
6676 | } |
6677 | break; |
6678 | } |
6679 | case Intrinsic::aarch64_neon_st4lane: { |
6680 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
6681 | SelectStoreLane(N: Node, NumVecs: 4, Opc: AArch64::ST4i8); |
6682 | return; |
6683 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
6684 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
6685 | SelectStoreLane(N: Node, NumVecs: 4, Opc: AArch64::ST4i16); |
6686 | return; |
6687 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
6688 | VT == MVT::v2f32) { |
6689 | SelectStoreLane(N: Node, NumVecs: 4, Opc: AArch64::ST4i32); |
6690 | return; |
6691 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
6692 | VT == MVT::v1f64) { |
6693 | SelectStoreLane(N: Node, NumVecs: 4, Opc: AArch64::ST4i64); |
6694 | return; |
6695 | } |
6696 | break; |
6697 | } |
6698 | case Intrinsic::aarch64_sve_st2q: { |
6699 | SelectPredicatedStore(N: Node, NumVecs: 2, Scale: 4, Opc_rr: AArch64::ST2Q, Opc_ri: AArch64::ST2Q_IMM); |
6700 | return; |
6701 | } |
6702 | case Intrinsic::aarch64_sve_st3q: { |
6703 | SelectPredicatedStore(N: Node, NumVecs: 3, Scale: 4, Opc_rr: AArch64::ST3Q, Opc_ri: AArch64::ST3Q_IMM); |
6704 | return; |
6705 | } |
6706 | case Intrinsic::aarch64_sve_st4q: { |
6707 | SelectPredicatedStore(N: Node, NumVecs: 4, Scale: 4, Opc_rr: AArch64::ST4Q, Opc_ri: AArch64::ST4Q_IMM); |
6708 | return; |
6709 | } |
6710 | case Intrinsic::aarch64_sve_st2: { |
6711 | if (VT == MVT::nxv16i8) { |
6712 | SelectPredicatedStore(N: Node, NumVecs: 2, Scale: 0, Opc_rr: AArch64::ST2B, Opc_ri: AArch64::ST2B_IMM); |
6713 | return; |
6714 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
6715 | VT == MVT::nxv8bf16) { |
6716 | SelectPredicatedStore(N: Node, NumVecs: 2, Scale: 1, Opc_rr: AArch64::ST2H, Opc_ri: AArch64::ST2H_IMM); |
6717 | return; |
6718 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
6719 | SelectPredicatedStore(N: Node, NumVecs: 2, Scale: 2, Opc_rr: AArch64::ST2W, Opc_ri: AArch64::ST2W_IMM); |
6720 | return; |
6721 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
6722 | SelectPredicatedStore(N: Node, NumVecs: 2, Scale: 3, Opc_rr: AArch64::ST2D, Opc_ri: AArch64::ST2D_IMM); |
6723 | return; |
6724 | } |
6725 | break; |
6726 | } |
6727 | case Intrinsic::aarch64_sve_st3: { |
6728 | if (VT == MVT::nxv16i8) { |
6729 | SelectPredicatedStore(N: Node, NumVecs: 3, Scale: 0, Opc_rr: AArch64::ST3B, Opc_ri: AArch64::ST3B_IMM); |
6730 | return; |
6731 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
6732 | VT == MVT::nxv8bf16) { |
6733 | SelectPredicatedStore(N: Node, NumVecs: 3, Scale: 1, Opc_rr: AArch64::ST3H, Opc_ri: AArch64::ST3H_IMM); |
6734 | return; |
6735 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
6736 | SelectPredicatedStore(N: Node, NumVecs: 3, Scale: 2, Opc_rr: AArch64::ST3W, Opc_ri: AArch64::ST3W_IMM); |
6737 | return; |
6738 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
6739 | SelectPredicatedStore(N: Node, NumVecs: 3, Scale: 3, Opc_rr: AArch64::ST3D, Opc_ri: AArch64::ST3D_IMM); |
6740 | return; |
6741 | } |
6742 | break; |
6743 | } |
6744 | case Intrinsic::aarch64_sve_st4: { |
6745 | if (VT == MVT::nxv16i8) { |
6746 | SelectPredicatedStore(N: Node, NumVecs: 4, Scale: 0, Opc_rr: AArch64::ST4B, Opc_ri: AArch64::ST4B_IMM); |
6747 | return; |
6748 | } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 || |
6749 | VT == MVT::nxv8bf16) { |
6750 | SelectPredicatedStore(N: Node, NumVecs: 4, Scale: 1, Opc_rr: AArch64::ST4H, Opc_ri: AArch64::ST4H_IMM); |
6751 | return; |
6752 | } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) { |
6753 | SelectPredicatedStore(N: Node, NumVecs: 4, Scale: 2, Opc_rr: AArch64::ST4W, Opc_ri: AArch64::ST4W_IMM); |
6754 | return; |
6755 | } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) { |
6756 | SelectPredicatedStore(N: Node, NumVecs: 4, Scale: 3, Opc_rr: AArch64::ST4D, Opc_ri: AArch64::ST4D_IMM); |
6757 | return; |
6758 | } |
6759 | break; |
6760 | } |
6761 | } |
6762 | break; |
6763 | } |
6764 | case AArch64ISD::LD2post: { |
6765 | if (VT == MVT::v8i8) { |
6766 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Twov8b_POST, SubRegIdx: AArch64::dsub0); |
6767 | return; |
6768 | } else if (VT == MVT::v16i8) { |
6769 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Twov16b_POST, SubRegIdx: AArch64::qsub0); |
6770 | return; |
6771 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6772 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Twov4h_POST, SubRegIdx: AArch64::dsub0); |
6773 | return; |
6774 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6775 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Twov8h_POST, SubRegIdx: AArch64::qsub0); |
6776 | return; |
6777 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6778 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Twov2s_POST, SubRegIdx: AArch64::dsub0); |
6779 | return; |
6780 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6781 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Twov4s_POST, SubRegIdx: AArch64::qsub0); |
6782 | return; |
6783 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6784 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD1Twov1d_POST, SubRegIdx: AArch64::dsub0); |
6785 | return; |
6786 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6787 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Twov2d_POST, SubRegIdx: AArch64::qsub0); |
6788 | return; |
6789 | } |
6790 | break; |
6791 | } |
6792 | case AArch64ISD::LD3post: { |
6793 | if (VT == MVT::v8i8) { |
6794 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Threev8b_POST, SubRegIdx: AArch64::dsub0); |
6795 | return; |
6796 | } else if (VT == MVT::v16i8) { |
6797 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Threev16b_POST, SubRegIdx: AArch64::qsub0); |
6798 | return; |
6799 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6800 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Threev4h_POST, SubRegIdx: AArch64::dsub0); |
6801 | return; |
6802 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6803 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Threev8h_POST, SubRegIdx: AArch64::qsub0); |
6804 | return; |
6805 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6806 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Threev2s_POST, SubRegIdx: AArch64::dsub0); |
6807 | return; |
6808 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6809 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Threev4s_POST, SubRegIdx: AArch64::qsub0); |
6810 | return; |
6811 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6812 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD1Threev1d_POST, SubRegIdx: AArch64::dsub0); |
6813 | return; |
6814 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6815 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Threev2d_POST, SubRegIdx: AArch64::qsub0); |
6816 | return; |
6817 | } |
6818 | break; |
6819 | } |
6820 | case AArch64ISD::LD4post: { |
6821 | if (VT == MVT::v8i8) { |
6822 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Fourv8b_POST, SubRegIdx: AArch64::dsub0); |
6823 | return; |
6824 | } else if (VT == MVT::v16i8) { |
6825 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Fourv16b_POST, SubRegIdx: AArch64::qsub0); |
6826 | return; |
6827 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6828 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Fourv4h_POST, SubRegIdx: AArch64::dsub0); |
6829 | return; |
6830 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6831 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Fourv8h_POST, SubRegIdx: AArch64::qsub0); |
6832 | return; |
6833 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6834 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Fourv2s_POST, SubRegIdx: AArch64::dsub0); |
6835 | return; |
6836 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6837 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Fourv4s_POST, SubRegIdx: AArch64::qsub0); |
6838 | return; |
6839 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6840 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD1Fourv1d_POST, SubRegIdx: AArch64::dsub0); |
6841 | return; |
6842 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6843 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Fourv2d_POST, SubRegIdx: AArch64::qsub0); |
6844 | return; |
6845 | } |
6846 | break; |
6847 | } |
6848 | case AArch64ISD::LD1x2post: { |
6849 | if (VT == MVT::v8i8) { |
6850 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD1Twov8b_POST, SubRegIdx: AArch64::dsub0); |
6851 | return; |
6852 | } else if (VT == MVT::v16i8) { |
6853 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD1Twov16b_POST, SubRegIdx: AArch64::qsub0); |
6854 | return; |
6855 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6856 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD1Twov4h_POST, SubRegIdx: AArch64::dsub0); |
6857 | return; |
6858 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6859 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD1Twov8h_POST, SubRegIdx: AArch64::qsub0); |
6860 | return; |
6861 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6862 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD1Twov2s_POST, SubRegIdx: AArch64::dsub0); |
6863 | return; |
6864 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6865 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD1Twov4s_POST, SubRegIdx: AArch64::qsub0); |
6866 | return; |
6867 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6868 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD1Twov1d_POST, SubRegIdx: AArch64::dsub0); |
6869 | return; |
6870 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6871 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD1Twov2d_POST, SubRegIdx: AArch64::qsub0); |
6872 | return; |
6873 | } |
6874 | break; |
6875 | } |
6876 | case AArch64ISD::LD1x3post: { |
6877 | if (VT == MVT::v8i8) { |
6878 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD1Threev8b_POST, SubRegIdx: AArch64::dsub0); |
6879 | return; |
6880 | } else if (VT == MVT::v16i8) { |
6881 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD1Threev16b_POST, SubRegIdx: AArch64::qsub0); |
6882 | return; |
6883 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6884 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD1Threev4h_POST, SubRegIdx: AArch64::dsub0); |
6885 | return; |
6886 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6887 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD1Threev8h_POST, SubRegIdx: AArch64::qsub0); |
6888 | return; |
6889 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6890 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD1Threev2s_POST, SubRegIdx: AArch64::dsub0); |
6891 | return; |
6892 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6893 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD1Threev4s_POST, SubRegIdx: AArch64::qsub0); |
6894 | return; |
6895 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6896 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD1Threev1d_POST, SubRegIdx: AArch64::dsub0); |
6897 | return; |
6898 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6899 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD1Threev2d_POST, SubRegIdx: AArch64::qsub0); |
6900 | return; |
6901 | } |
6902 | break; |
6903 | } |
6904 | case AArch64ISD::LD1x4post: { |
6905 | if (VT == MVT::v8i8) { |
6906 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD1Fourv8b_POST, SubRegIdx: AArch64::dsub0); |
6907 | return; |
6908 | } else if (VT == MVT::v16i8) { |
6909 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD1Fourv16b_POST, SubRegIdx: AArch64::qsub0); |
6910 | return; |
6911 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6912 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD1Fourv4h_POST, SubRegIdx: AArch64::dsub0); |
6913 | return; |
6914 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6915 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD1Fourv8h_POST, SubRegIdx: AArch64::qsub0); |
6916 | return; |
6917 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6918 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD1Fourv2s_POST, SubRegIdx: AArch64::dsub0); |
6919 | return; |
6920 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6921 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD1Fourv4s_POST, SubRegIdx: AArch64::qsub0); |
6922 | return; |
6923 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6924 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD1Fourv1d_POST, SubRegIdx: AArch64::dsub0); |
6925 | return; |
6926 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6927 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD1Fourv2d_POST, SubRegIdx: AArch64::qsub0); |
6928 | return; |
6929 | } |
6930 | break; |
6931 | } |
6932 | case AArch64ISD::LD1DUPpost: { |
6933 | if (VT == MVT::v8i8) { |
6934 | SelectPostLoad(N: Node, NumVecs: 1, Opc: AArch64::LD1Rv8b_POST, SubRegIdx: AArch64::dsub0); |
6935 | return; |
6936 | } else if (VT == MVT::v16i8) { |
6937 | SelectPostLoad(N: Node, NumVecs: 1, Opc: AArch64::LD1Rv16b_POST, SubRegIdx: AArch64::qsub0); |
6938 | return; |
6939 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6940 | SelectPostLoad(N: Node, NumVecs: 1, Opc: AArch64::LD1Rv4h_POST, SubRegIdx: AArch64::dsub0); |
6941 | return; |
6942 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6943 | SelectPostLoad(N: Node, NumVecs: 1, Opc: AArch64::LD1Rv8h_POST, SubRegIdx: AArch64::qsub0); |
6944 | return; |
6945 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6946 | SelectPostLoad(N: Node, NumVecs: 1, Opc: AArch64::LD1Rv2s_POST, SubRegIdx: AArch64::dsub0); |
6947 | return; |
6948 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6949 | SelectPostLoad(N: Node, NumVecs: 1, Opc: AArch64::LD1Rv4s_POST, SubRegIdx: AArch64::qsub0); |
6950 | return; |
6951 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6952 | SelectPostLoad(N: Node, NumVecs: 1, Opc: AArch64::LD1Rv1d_POST, SubRegIdx: AArch64::dsub0); |
6953 | return; |
6954 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6955 | SelectPostLoad(N: Node, NumVecs: 1, Opc: AArch64::LD1Rv2d_POST, SubRegIdx: AArch64::qsub0); |
6956 | return; |
6957 | } |
6958 | break; |
6959 | } |
6960 | case AArch64ISD::LD2DUPpost: { |
6961 | if (VT == MVT::v8i8) { |
6962 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Rv8b_POST, SubRegIdx: AArch64::dsub0); |
6963 | return; |
6964 | } else if (VT == MVT::v16i8) { |
6965 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Rv16b_POST, SubRegIdx: AArch64::qsub0); |
6966 | return; |
6967 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6968 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Rv4h_POST, SubRegIdx: AArch64::dsub0); |
6969 | return; |
6970 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6971 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Rv8h_POST, SubRegIdx: AArch64::qsub0); |
6972 | return; |
6973 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
6974 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Rv2s_POST, SubRegIdx: AArch64::dsub0); |
6975 | return; |
6976 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
6977 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Rv4s_POST, SubRegIdx: AArch64::qsub0); |
6978 | return; |
6979 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
6980 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Rv1d_POST, SubRegIdx: AArch64::dsub0); |
6981 | return; |
6982 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
6983 | SelectPostLoad(N: Node, NumVecs: 2, Opc: AArch64::LD2Rv2d_POST, SubRegIdx: AArch64::qsub0); |
6984 | return; |
6985 | } |
6986 | break; |
6987 | } |
6988 | case AArch64ISD::LD3DUPpost: { |
6989 | if (VT == MVT::v8i8) { |
6990 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Rv8b_POST, SubRegIdx: AArch64::dsub0); |
6991 | return; |
6992 | } else if (VT == MVT::v16i8) { |
6993 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Rv16b_POST, SubRegIdx: AArch64::qsub0); |
6994 | return; |
6995 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
6996 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Rv4h_POST, SubRegIdx: AArch64::dsub0); |
6997 | return; |
6998 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
6999 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Rv8h_POST, SubRegIdx: AArch64::qsub0); |
7000 | return; |
7001 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
7002 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Rv2s_POST, SubRegIdx: AArch64::dsub0); |
7003 | return; |
7004 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
7005 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Rv4s_POST, SubRegIdx: AArch64::qsub0); |
7006 | return; |
7007 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
7008 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Rv1d_POST, SubRegIdx: AArch64::dsub0); |
7009 | return; |
7010 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
7011 | SelectPostLoad(N: Node, NumVecs: 3, Opc: AArch64::LD3Rv2d_POST, SubRegIdx: AArch64::qsub0); |
7012 | return; |
7013 | } |
7014 | break; |
7015 | } |
7016 | case AArch64ISD::LD4DUPpost: { |
7017 | if (VT == MVT::v8i8) { |
7018 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Rv8b_POST, SubRegIdx: AArch64::dsub0); |
7019 | return; |
7020 | } else if (VT == MVT::v16i8) { |
7021 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Rv16b_POST, SubRegIdx: AArch64::qsub0); |
7022 | return; |
7023 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
7024 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Rv4h_POST, SubRegIdx: AArch64::dsub0); |
7025 | return; |
7026 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
7027 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Rv8h_POST, SubRegIdx: AArch64::qsub0); |
7028 | return; |
7029 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
7030 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Rv2s_POST, SubRegIdx: AArch64::dsub0); |
7031 | return; |
7032 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
7033 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Rv4s_POST, SubRegIdx: AArch64::qsub0); |
7034 | return; |
7035 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
7036 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Rv1d_POST, SubRegIdx: AArch64::dsub0); |
7037 | return; |
7038 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
7039 | SelectPostLoad(N: Node, NumVecs: 4, Opc: AArch64::LD4Rv2d_POST, SubRegIdx: AArch64::qsub0); |
7040 | return; |
7041 | } |
7042 | break; |
7043 | } |
7044 | case AArch64ISD::LD1LANEpost: { |
7045 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
7046 | SelectPostLoadLane(N: Node, NumVecs: 1, Opc: AArch64::LD1i8_POST); |
7047 | return; |
7048 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
7049 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
7050 | SelectPostLoadLane(N: Node, NumVecs: 1, Opc: AArch64::LD1i16_POST); |
7051 | return; |
7052 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
7053 | VT == MVT::v2f32) { |
7054 | SelectPostLoadLane(N: Node, NumVecs: 1, Opc: AArch64::LD1i32_POST); |
7055 | return; |
7056 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
7057 | VT == MVT::v1f64) { |
7058 | SelectPostLoadLane(N: Node, NumVecs: 1, Opc: AArch64::LD1i64_POST); |
7059 | return; |
7060 | } |
7061 | break; |
7062 | } |
7063 | case AArch64ISD::LD2LANEpost: { |
7064 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
7065 | SelectPostLoadLane(N: Node, NumVecs: 2, Opc: AArch64::LD2i8_POST); |
7066 | return; |
7067 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
7068 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
7069 | SelectPostLoadLane(N: Node, NumVecs: 2, Opc: AArch64::LD2i16_POST); |
7070 | return; |
7071 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
7072 | VT == MVT::v2f32) { |
7073 | SelectPostLoadLane(N: Node, NumVecs: 2, Opc: AArch64::LD2i32_POST); |
7074 | return; |
7075 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
7076 | VT == MVT::v1f64) { |
7077 | SelectPostLoadLane(N: Node, NumVecs: 2, Opc: AArch64::LD2i64_POST); |
7078 | return; |
7079 | } |
7080 | break; |
7081 | } |
7082 | case AArch64ISD::LD3LANEpost: { |
7083 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
7084 | SelectPostLoadLane(N: Node, NumVecs: 3, Opc: AArch64::LD3i8_POST); |
7085 | return; |
7086 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
7087 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
7088 | SelectPostLoadLane(N: Node, NumVecs: 3, Opc: AArch64::LD3i16_POST); |
7089 | return; |
7090 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
7091 | VT == MVT::v2f32) { |
7092 | SelectPostLoadLane(N: Node, NumVecs: 3, Opc: AArch64::LD3i32_POST); |
7093 | return; |
7094 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
7095 | VT == MVT::v1f64) { |
7096 | SelectPostLoadLane(N: Node, NumVecs: 3, Opc: AArch64::LD3i64_POST); |
7097 | return; |
7098 | } |
7099 | break; |
7100 | } |
7101 | case AArch64ISD::LD4LANEpost: { |
7102 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
7103 | SelectPostLoadLane(N: Node, NumVecs: 4, Opc: AArch64::LD4i8_POST); |
7104 | return; |
7105 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
7106 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
7107 | SelectPostLoadLane(N: Node, NumVecs: 4, Opc: AArch64::LD4i16_POST); |
7108 | return; |
7109 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
7110 | VT == MVT::v2f32) { |
7111 | SelectPostLoadLane(N: Node, NumVecs: 4, Opc: AArch64::LD4i32_POST); |
7112 | return; |
7113 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
7114 | VT == MVT::v1f64) { |
7115 | SelectPostLoadLane(N: Node, NumVecs: 4, Opc: AArch64::LD4i64_POST); |
7116 | return; |
7117 | } |
7118 | break; |
7119 | } |
7120 | case AArch64ISD::ST2post: { |
7121 | VT = Node->getOperand(Num: 1).getValueType(); |
7122 | if (VT == MVT::v8i8) { |
7123 | SelectPostStore(N: Node, NumVecs: 2, Opc: AArch64::ST2Twov8b_POST); |
7124 | return; |
7125 | } else if (VT == MVT::v16i8) { |
7126 | SelectPostStore(N: Node, NumVecs: 2, Opc: AArch64::ST2Twov16b_POST); |
7127 | return; |
7128 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
7129 | SelectPostStore(N: Node, NumVecs: 2, Opc: AArch64::ST2Twov4h_POST); |
7130 | return; |
7131 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
7132 | SelectPostStore(N: Node, NumVecs: 2, Opc: AArch64::ST2Twov8h_POST); |
7133 | return; |
7134 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
7135 | SelectPostStore(N: Node, NumVecs: 2, Opc: AArch64::ST2Twov2s_POST); |
7136 | return; |
7137 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
7138 | SelectPostStore(N: Node, NumVecs: 2, Opc: AArch64::ST2Twov4s_POST); |
7139 | return; |
7140 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
7141 | SelectPostStore(N: Node, NumVecs: 2, Opc: AArch64::ST2Twov2d_POST); |
7142 | return; |
7143 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
7144 | SelectPostStore(N: Node, NumVecs: 2, Opc: AArch64::ST1Twov1d_POST); |
7145 | return; |
7146 | } |
7147 | break; |
7148 | } |
7149 | case AArch64ISD::ST3post: { |
7150 | VT = Node->getOperand(Num: 1).getValueType(); |
7151 | if (VT == MVT::v8i8) { |
7152 | SelectPostStore(N: Node, NumVecs: 3, Opc: AArch64::ST3Threev8b_POST); |
7153 | return; |
7154 | } else if (VT == MVT::v16i8) { |
7155 | SelectPostStore(N: Node, NumVecs: 3, Opc: AArch64::ST3Threev16b_POST); |
7156 | return; |
7157 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
7158 | SelectPostStore(N: Node, NumVecs: 3, Opc: AArch64::ST3Threev4h_POST); |
7159 | return; |
7160 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
7161 | SelectPostStore(N: Node, NumVecs: 3, Opc: AArch64::ST3Threev8h_POST); |
7162 | return; |
7163 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
7164 | SelectPostStore(N: Node, NumVecs: 3, Opc: AArch64::ST3Threev2s_POST); |
7165 | return; |
7166 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
7167 | SelectPostStore(N: Node, NumVecs: 3, Opc: AArch64::ST3Threev4s_POST); |
7168 | return; |
7169 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
7170 | SelectPostStore(N: Node, NumVecs: 3, Opc: AArch64::ST3Threev2d_POST); |
7171 | return; |
7172 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
7173 | SelectPostStore(N: Node, NumVecs: 3, Opc: AArch64::ST1Threev1d_POST); |
7174 | return; |
7175 | } |
7176 | break; |
7177 | } |
7178 | case AArch64ISD::ST4post: { |
7179 | VT = Node->getOperand(Num: 1).getValueType(); |
7180 | if (VT == MVT::v8i8) { |
7181 | SelectPostStore(N: Node, NumVecs: 4, Opc: AArch64::ST4Fourv8b_POST); |
7182 | return; |
7183 | } else if (VT == MVT::v16i8) { |
7184 | SelectPostStore(N: Node, NumVecs: 4, Opc: AArch64::ST4Fourv16b_POST); |
7185 | return; |
7186 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
7187 | SelectPostStore(N: Node, NumVecs: 4, Opc: AArch64::ST4Fourv4h_POST); |
7188 | return; |
7189 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
7190 | SelectPostStore(N: Node, NumVecs: 4, Opc: AArch64::ST4Fourv8h_POST); |
7191 | return; |
7192 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
7193 | SelectPostStore(N: Node, NumVecs: 4, Opc: AArch64::ST4Fourv2s_POST); |
7194 | return; |
7195 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
7196 | SelectPostStore(N: Node, NumVecs: 4, Opc: AArch64::ST4Fourv4s_POST); |
7197 | return; |
7198 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
7199 | SelectPostStore(N: Node, NumVecs: 4, Opc: AArch64::ST4Fourv2d_POST); |
7200 | return; |
7201 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
7202 | SelectPostStore(N: Node, NumVecs: 4, Opc: AArch64::ST1Fourv1d_POST); |
7203 | return; |
7204 | } |
7205 | break; |
7206 | } |
7207 | case AArch64ISD::ST1x2post: { |
7208 | VT = Node->getOperand(Num: 1).getValueType(); |
7209 | if (VT == MVT::v8i8) { |
7210 | SelectPostStore(N: Node, NumVecs: 2, Opc: AArch64::ST1Twov8b_POST); |
7211 | return; |
7212 | } else if (VT == MVT::v16i8) { |
7213 | SelectPostStore(N: Node, NumVecs: 2, Opc: AArch64::ST1Twov16b_POST); |
7214 | return; |
7215 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
7216 | SelectPostStore(N: Node, NumVecs: 2, Opc: AArch64::ST1Twov4h_POST); |
7217 | return; |
7218 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
7219 | SelectPostStore(N: Node, NumVecs: 2, Opc: AArch64::ST1Twov8h_POST); |
7220 | return; |
7221 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
7222 | SelectPostStore(N: Node, NumVecs: 2, Opc: AArch64::ST1Twov2s_POST); |
7223 | return; |
7224 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
7225 | SelectPostStore(N: Node, NumVecs: 2, Opc: AArch64::ST1Twov4s_POST); |
7226 | return; |
7227 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
7228 | SelectPostStore(N: Node, NumVecs: 2, Opc: AArch64::ST1Twov1d_POST); |
7229 | return; |
7230 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
7231 | SelectPostStore(N: Node, NumVecs: 2, Opc: AArch64::ST1Twov2d_POST); |
7232 | return; |
7233 | } |
7234 | break; |
7235 | } |
7236 | case AArch64ISD::ST1x3post: { |
7237 | VT = Node->getOperand(Num: 1).getValueType(); |
7238 | if (VT == MVT::v8i8) { |
7239 | SelectPostStore(N: Node, NumVecs: 3, Opc: AArch64::ST1Threev8b_POST); |
7240 | return; |
7241 | } else if (VT == MVT::v16i8) { |
7242 | SelectPostStore(N: Node, NumVecs: 3, Opc: AArch64::ST1Threev16b_POST); |
7243 | return; |
7244 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
7245 | SelectPostStore(N: Node, NumVecs: 3, Opc: AArch64::ST1Threev4h_POST); |
7246 | return; |
7247 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16 ) { |
7248 | SelectPostStore(N: Node, NumVecs: 3, Opc: AArch64::ST1Threev8h_POST); |
7249 | return; |
7250 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
7251 | SelectPostStore(N: Node, NumVecs: 3, Opc: AArch64::ST1Threev2s_POST); |
7252 | return; |
7253 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
7254 | SelectPostStore(N: Node, NumVecs: 3, Opc: AArch64::ST1Threev4s_POST); |
7255 | return; |
7256 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
7257 | SelectPostStore(N: Node, NumVecs: 3, Opc: AArch64::ST1Threev1d_POST); |
7258 | return; |
7259 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
7260 | SelectPostStore(N: Node, NumVecs: 3, Opc: AArch64::ST1Threev2d_POST); |
7261 | return; |
7262 | } |
7263 | break; |
7264 | } |
7265 | case AArch64ISD::ST1x4post: { |
7266 | VT = Node->getOperand(Num: 1).getValueType(); |
7267 | if (VT == MVT::v8i8) { |
7268 | SelectPostStore(N: Node, NumVecs: 4, Opc: AArch64::ST1Fourv8b_POST); |
7269 | return; |
7270 | } else if (VT == MVT::v16i8) { |
7271 | SelectPostStore(N: Node, NumVecs: 4, Opc: AArch64::ST1Fourv16b_POST); |
7272 | return; |
7273 | } else if (VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16) { |
7274 | SelectPostStore(N: Node, NumVecs: 4, Opc: AArch64::ST1Fourv4h_POST); |
7275 | return; |
7276 | } else if (VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v8bf16) { |
7277 | SelectPostStore(N: Node, NumVecs: 4, Opc: AArch64::ST1Fourv8h_POST); |
7278 | return; |
7279 | } else if (VT == MVT::v2i32 || VT == MVT::v2f32) { |
7280 | SelectPostStore(N: Node, NumVecs: 4, Opc: AArch64::ST1Fourv2s_POST); |
7281 | return; |
7282 | } else if (VT == MVT::v4i32 || VT == MVT::v4f32) { |
7283 | SelectPostStore(N: Node, NumVecs: 4, Opc: AArch64::ST1Fourv4s_POST); |
7284 | return; |
7285 | } else if (VT == MVT::v1i64 || VT == MVT::v1f64) { |
7286 | SelectPostStore(N: Node, NumVecs: 4, Opc: AArch64::ST1Fourv1d_POST); |
7287 | return; |
7288 | } else if (VT == MVT::v2i64 || VT == MVT::v2f64) { |
7289 | SelectPostStore(N: Node, NumVecs: 4, Opc: AArch64::ST1Fourv2d_POST); |
7290 | return; |
7291 | } |
7292 | break; |
7293 | } |
7294 | case AArch64ISD::ST2LANEpost: { |
7295 | VT = Node->getOperand(Num: 1).getValueType(); |
7296 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
7297 | SelectPostStoreLane(N: Node, NumVecs: 2, Opc: AArch64::ST2i8_POST); |
7298 | return; |
7299 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
7300 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
7301 | SelectPostStoreLane(N: Node, NumVecs: 2, Opc: AArch64::ST2i16_POST); |
7302 | return; |
7303 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
7304 | VT == MVT::v2f32) { |
7305 | SelectPostStoreLane(N: Node, NumVecs: 2, Opc: AArch64::ST2i32_POST); |
7306 | return; |
7307 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
7308 | VT == MVT::v1f64) { |
7309 | SelectPostStoreLane(N: Node, NumVecs: 2, Opc: AArch64::ST2i64_POST); |
7310 | return; |
7311 | } |
7312 | break; |
7313 | } |
7314 | case AArch64ISD::ST3LANEpost: { |
7315 | VT = Node->getOperand(Num: 1).getValueType(); |
7316 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
7317 | SelectPostStoreLane(N: Node, NumVecs: 3, Opc: AArch64::ST3i8_POST); |
7318 | return; |
7319 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
7320 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
7321 | SelectPostStoreLane(N: Node, NumVecs: 3, Opc: AArch64::ST3i16_POST); |
7322 | return; |
7323 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
7324 | VT == MVT::v2f32) { |
7325 | SelectPostStoreLane(N: Node, NumVecs: 3, Opc: AArch64::ST3i32_POST); |
7326 | return; |
7327 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
7328 | VT == MVT::v1f64) { |
7329 | SelectPostStoreLane(N: Node, NumVecs: 3, Opc: AArch64::ST3i64_POST); |
7330 | return; |
7331 | } |
7332 | break; |
7333 | } |
7334 | case AArch64ISD::ST4LANEpost: { |
7335 | VT = Node->getOperand(Num: 1).getValueType(); |
7336 | if (VT == MVT::v16i8 || VT == MVT::v8i8) { |
7337 | SelectPostStoreLane(N: Node, NumVecs: 4, Opc: AArch64::ST4i8_POST); |
7338 | return; |
7339 | } else if (VT == MVT::v8i16 || VT == MVT::v4i16 || VT == MVT::v4f16 || |
7340 | VT == MVT::v8f16 || VT == MVT::v4bf16 || VT == MVT::v8bf16) { |
7341 | SelectPostStoreLane(N: Node, NumVecs: 4, Opc: AArch64::ST4i16_POST); |
7342 | return; |
7343 | } else if (VT == MVT::v4i32 || VT == MVT::v2i32 || VT == MVT::v4f32 || |
7344 | VT == MVT::v2f32) { |
7345 | SelectPostStoreLane(N: Node, NumVecs: 4, Opc: AArch64::ST4i32_POST); |
7346 | return; |
7347 | } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 || |
7348 | VT == MVT::v1f64) { |
7349 | SelectPostStoreLane(N: Node, NumVecs: 4, Opc: AArch64::ST4i64_POST); |
7350 | return; |
7351 | } |
7352 | break; |
7353 | } |
7354 | } |
7355 | |
7356 | // Select the default instruction |
7357 | SelectCode(N: Node); |
7358 | } |
7359 | |
7360 | /// createAArch64ISelDag - This pass converts a legalized DAG into a |
7361 | /// AArch64-specific DAG, ready for instruction scheduling. |
7362 | FunctionPass *llvm::createAArch64ISelDag(AArch64TargetMachine &TM, |
7363 | CodeGenOptLevel OptLevel) { |
7364 | return new AArch64DAGToDAGISelLegacy(TM, OptLevel); |
7365 | } |
7366 | |
7367 | /// When \p PredVT is a scalable vector predicate in the form |
7368 | /// MVT::nx<M>xi1, it builds the correspondent scalable vector of |
7369 | /// integers MVT::nx<M>xi<bits> s.t. M x bits = 128. When targeting |
7370 | /// structured vectors (NumVec >1), the output data type is |
7371 | /// MVT::nx<M*NumVec>xi<bits> s.t. M x bits = 128. If the input |
7372 | /// PredVT is not in the form MVT::nx<M>xi1, it returns an invalid |
7373 | /// EVT. |
7374 | static EVT getPackedVectorTypeFromPredicateType(LLVMContext &Ctx, EVT PredVT, |
7375 | unsigned NumVec) { |
7376 | assert(NumVec > 0 && NumVec < 5 && "Invalid number of vectors." ); |
7377 | if (!PredVT.isScalableVector() || PredVT.getVectorElementType() != MVT::i1) |
7378 | return EVT(); |
7379 | |
7380 | if (PredVT != MVT::nxv16i1 && PredVT != MVT::nxv8i1 && |
7381 | PredVT != MVT::nxv4i1 && PredVT != MVT::nxv2i1) |
7382 | return EVT(); |
7383 | |
7384 | ElementCount EC = PredVT.getVectorElementCount(); |
7385 | EVT ScalarVT = |
7386 | EVT::getIntegerVT(Context&: Ctx, BitWidth: AArch64::SVEBitsPerBlock / EC.getKnownMinValue()); |
7387 | EVT MemVT = EVT::getVectorVT(Context&: Ctx, VT: ScalarVT, EC: EC * NumVec); |
7388 | |
7389 | return MemVT; |
7390 | } |
7391 | |
7392 | /// Return the EVT of the data associated to a memory operation in \p |
7393 | /// Root. If such EVT cannot be retrieved, it returns an invalid EVT. |
7394 | static EVT getMemVTFromNode(LLVMContext &Ctx, SDNode *Root) { |
7395 | if (auto *MemIntr = dyn_cast<MemIntrinsicSDNode>(Val: Root)) |
7396 | return MemIntr->getMemoryVT(); |
7397 | |
7398 | if (isa<MemSDNode>(Val: Root)) { |
7399 | EVT MemVT = cast<MemSDNode>(Val: Root)->getMemoryVT(); |
7400 | |
7401 | EVT DataVT; |
7402 | if (auto *Load = dyn_cast<LoadSDNode>(Val: Root)) |
7403 | DataVT = Load->getValueType(ResNo: 0); |
7404 | else if (auto *Load = dyn_cast<MaskedLoadSDNode>(Val: Root)) |
7405 | DataVT = Load->getValueType(ResNo: 0); |
7406 | else if (auto *Store = dyn_cast<StoreSDNode>(Val: Root)) |
7407 | DataVT = Store->getValue().getValueType(); |
7408 | else if (auto *Store = dyn_cast<MaskedStoreSDNode>(Val: Root)) |
7409 | DataVT = Store->getValue().getValueType(); |
7410 | else |
7411 | llvm_unreachable("Unexpected MemSDNode!" ); |
7412 | |
7413 | return DataVT.changeVectorElementType(EltVT: MemVT.getVectorElementType()); |
7414 | } |
7415 | |
7416 | const unsigned Opcode = Root->getOpcode(); |
7417 | // For custom ISD nodes, we have to look at them individually to extract the |
7418 | // type of the data moved to/from memory. |
7419 | switch (Opcode) { |
7420 | case AArch64ISD::LD1_MERGE_ZERO: |
7421 | case AArch64ISD::LD1S_MERGE_ZERO: |
7422 | case AArch64ISD::LDNF1_MERGE_ZERO: |
7423 | case AArch64ISD::LDNF1S_MERGE_ZERO: |
7424 | return cast<VTSDNode>(Val: Root->getOperand(Num: 3))->getVT(); |
7425 | case AArch64ISD::ST1_PRED: |
7426 | return cast<VTSDNode>(Val: Root->getOperand(Num: 4))->getVT(); |
7427 | default: |
7428 | break; |
7429 | } |
7430 | |
7431 | if (Opcode != ISD::INTRINSIC_VOID && Opcode != ISD::INTRINSIC_W_CHAIN) |
7432 | return EVT(); |
7433 | |
7434 | switch (Root->getConstantOperandVal(Num: 1)) { |
7435 | default: |
7436 | return EVT(); |
7437 | case Intrinsic::aarch64_sme_ldr: |
7438 | case Intrinsic::aarch64_sme_str: |
7439 | return MVT::nxv16i8; |
7440 | case Intrinsic::aarch64_sve_prf: |
7441 | // We are using an SVE prefetch intrinsic. Type must be inferred from the |
7442 | // width of the predicate. |
7443 | return getPackedVectorTypeFromPredicateType( |
7444 | Ctx, PredVT: Root->getOperand(Num: 2)->getValueType(ResNo: 0), /*NumVec=*/1); |
7445 | case Intrinsic::aarch64_sve_ld2_sret: |
7446 | case Intrinsic::aarch64_sve_ld2q_sret: |
7447 | return getPackedVectorTypeFromPredicateType( |
7448 | Ctx, PredVT: Root->getOperand(Num: 2)->getValueType(ResNo: 0), /*NumVec=*/2); |
7449 | case Intrinsic::aarch64_sve_st2q: |
7450 | return getPackedVectorTypeFromPredicateType( |
7451 | Ctx, PredVT: Root->getOperand(Num: 4)->getValueType(ResNo: 0), /*NumVec=*/2); |
7452 | case Intrinsic::aarch64_sve_ld3_sret: |
7453 | case Intrinsic::aarch64_sve_ld3q_sret: |
7454 | return getPackedVectorTypeFromPredicateType( |
7455 | Ctx, PredVT: Root->getOperand(Num: 2)->getValueType(ResNo: 0), /*NumVec=*/3); |
7456 | case Intrinsic::aarch64_sve_st3q: |
7457 | return getPackedVectorTypeFromPredicateType( |
7458 | Ctx, PredVT: Root->getOperand(Num: 5)->getValueType(ResNo: 0), /*NumVec=*/3); |
7459 | case Intrinsic::aarch64_sve_ld4_sret: |
7460 | case Intrinsic::aarch64_sve_ld4q_sret: |
7461 | return getPackedVectorTypeFromPredicateType( |
7462 | Ctx, PredVT: Root->getOperand(Num: 2)->getValueType(ResNo: 0), /*NumVec=*/4); |
7463 | case Intrinsic::aarch64_sve_st4q: |
7464 | return getPackedVectorTypeFromPredicateType( |
7465 | Ctx, PredVT: Root->getOperand(Num: 6)->getValueType(ResNo: 0), /*NumVec=*/4); |
7466 | case Intrinsic::aarch64_sve_ld1udq: |
7467 | case Intrinsic::aarch64_sve_st1dq: |
7468 | return EVT(MVT::nxv1i64); |
7469 | case Intrinsic::aarch64_sve_ld1uwq: |
7470 | case Intrinsic::aarch64_sve_st1wq: |
7471 | return EVT(MVT::nxv1i32); |
7472 | } |
7473 | } |
7474 | |
7475 | /// SelectAddrModeIndexedSVE - Attempt selection of the addressing mode: |
7476 | /// Base + OffImm * sizeof(MemVT) for Min >= OffImm <= Max |
7477 | /// where Root is the memory access using N for its address. |
7478 | template <int64_t Min, int64_t Max> |
7479 | bool AArch64DAGToDAGISel::SelectAddrModeIndexedSVE(SDNode *Root, SDValue N, |
7480 | SDValue &Base, |
7481 | SDValue &OffImm) { |
7482 | const EVT MemVT = getMemVTFromNode(Ctx&: *(CurDAG->getContext()), Root); |
7483 | const DataLayout &DL = CurDAG->getDataLayout(); |
7484 | const MachineFrameInfo &MFI = MF->getFrameInfo(); |
7485 | |
7486 | if (N.getOpcode() == ISD::FrameIndex) { |
7487 | int FI = cast<FrameIndexSDNode>(Val&: N)->getIndex(); |
7488 | // We can only encode VL scaled offsets, so only fold in frame indexes |
7489 | // referencing SVE objects. |
7490 | if (MFI.getStackID(ObjectIdx: FI) == TargetStackID::ScalableVector) { |
7491 | Base = CurDAG->getTargetFrameIndex(FI, VT: TLI->getPointerTy(DL)); |
7492 | OffImm = CurDAG->getTargetConstant(Val: 0, DL: SDLoc(N), VT: MVT::i64); |
7493 | return true; |
7494 | } |
7495 | |
7496 | return false; |
7497 | } |
7498 | |
7499 | if (MemVT == EVT()) |
7500 | return false; |
7501 | |
7502 | if (N.getOpcode() != ISD::ADD) |
7503 | return false; |
7504 | |
7505 | SDValue VScale = N.getOperand(i: 1); |
7506 | int64_t MulImm = std::numeric_limits<int64_t>::max(); |
7507 | if (VScale.getOpcode() == ISD::VSCALE) { |
7508 | MulImm = cast<ConstantSDNode>(Val: VScale.getOperand(i: 0))->getSExtValue(); |
7509 | } else if (auto C = dyn_cast<ConstantSDNode>(Val&: VScale)) { |
7510 | int64_t ByteOffset = C->getSExtValue(); |
7511 | const auto KnownVScale = |
7512 | Subtarget->getSVEVectorSizeInBits() / AArch64::SVEBitsPerBlock; |
7513 | |
7514 | if (!KnownVScale || ByteOffset % KnownVScale != 0) |
7515 | return false; |
7516 | |
7517 | MulImm = ByteOffset / KnownVScale; |
7518 | } else |
7519 | return false; |
7520 | |
7521 | TypeSize TS = MemVT.getSizeInBits(); |
7522 | int64_t MemWidthBytes = static_cast<int64_t>(TS.getKnownMinValue()) / 8; |
7523 | |
7524 | if ((MulImm % MemWidthBytes) != 0) |
7525 | return false; |
7526 | |
7527 | int64_t Offset = MulImm / MemWidthBytes; |
7528 | if (Offset < Min || Offset > Max) |
7529 | return false; |
7530 | |
7531 | Base = N.getOperand(i: 0); |
7532 | if (Base.getOpcode() == ISD::FrameIndex) { |
7533 | int FI = cast<FrameIndexSDNode>(Val&: Base)->getIndex(); |
7534 | // We can only encode VL scaled offsets, so only fold in frame indexes |
7535 | // referencing SVE objects. |
7536 | if (MFI.getStackID(ObjectIdx: FI) == TargetStackID::ScalableVector) |
7537 | Base = CurDAG->getTargetFrameIndex(FI, VT: TLI->getPointerTy(DL)); |
7538 | } |
7539 | |
7540 | OffImm = CurDAG->getTargetConstant(Val: Offset, DL: SDLoc(N), VT: MVT::i64); |
7541 | return true; |
7542 | } |
7543 | |
7544 | /// Select register plus register addressing mode for SVE, with scaled |
7545 | /// offset. |
7546 | bool AArch64DAGToDAGISel::SelectSVERegRegAddrMode(SDValue N, unsigned Scale, |
7547 | SDValue &Base, |
7548 | SDValue &Offset) { |
7549 | if (N.getOpcode() != ISD::ADD) |
7550 | return false; |
7551 | |
7552 | // Process an ADD node. |
7553 | const SDValue LHS = N.getOperand(i: 0); |
7554 | const SDValue RHS = N.getOperand(i: 1); |
7555 | |
7556 | // 8 bit data does not come with the SHL node, so it is treated |
7557 | // separately. |
7558 | if (Scale == 0) { |
7559 | Base = LHS; |
7560 | Offset = RHS; |
7561 | return true; |
7562 | } |
7563 | |
7564 | if (auto C = dyn_cast<ConstantSDNode>(Val: RHS)) { |
7565 | int64_t ImmOff = C->getSExtValue(); |
7566 | unsigned Size = 1 << Scale; |
7567 | |
7568 | // To use the reg+reg addressing mode, the immediate must be a multiple of |
7569 | // the vector element's byte size. |
7570 | if (ImmOff % Size) |
7571 | return false; |
7572 | |
7573 | SDLoc DL(N); |
7574 | Base = LHS; |
7575 | Offset = CurDAG->getTargetConstant(Val: ImmOff >> Scale, DL, VT: MVT::i64); |
7576 | SDValue Ops[] = {Offset}; |
7577 | SDNode *MI = CurDAG->getMachineNode(Opcode: AArch64::MOVi64imm, dl: DL, VT: MVT::i64, Ops); |
7578 | Offset = SDValue(MI, 0); |
7579 | return true; |
7580 | } |
7581 | |
7582 | // Check if the RHS is a shift node with a constant. |
7583 | if (RHS.getOpcode() != ISD::SHL) |
7584 | return false; |
7585 | |
7586 | const SDValue ShiftRHS = RHS.getOperand(i: 1); |
7587 | if (auto *C = dyn_cast<ConstantSDNode>(Val: ShiftRHS)) |
7588 | if (C->getZExtValue() == Scale) { |
7589 | Base = LHS; |
7590 | Offset = RHS.getOperand(i: 0); |
7591 | return true; |
7592 | } |
7593 | |
7594 | return false; |
7595 | } |
7596 | |
7597 | bool AArch64DAGToDAGISel::SelectAllActivePredicate(SDValue N) { |
7598 | const AArch64TargetLowering *TLI = |
7599 | static_cast<const AArch64TargetLowering *>(getTargetLowering()); |
7600 | |
7601 | return TLI->isAllActivePredicate(DAG&: *CurDAG, N); |
7602 | } |
7603 | |
7604 | bool AArch64DAGToDAGISel::SelectAnyPredicate(SDValue N) { |
7605 | EVT VT = N.getValueType(); |
7606 | return VT.isScalableVector() && VT.getVectorElementType() == MVT::i1; |
7607 | } |
7608 | |
7609 | bool AArch64DAGToDAGISel::SelectSMETileSlice(SDValue N, unsigned MaxSize, |
7610 | SDValue &Base, SDValue &Offset, |
7611 | unsigned Scale) { |
7612 | // Try to untangle an ADD node into a 'reg + offset' |
7613 | if (CurDAG->isBaseWithConstantOffset(Op: N)) |
7614 | if (auto C = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1))) { |
7615 | int64_t ImmOff = C->getSExtValue(); |
7616 | if ((ImmOff > 0 && ImmOff <= MaxSize && (ImmOff % Scale == 0))) { |
7617 | Base = N.getOperand(i: 0); |
7618 | Offset = CurDAG->getTargetConstant(Val: ImmOff / Scale, DL: SDLoc(N), VT: MVT::i64); |
7619 | return true; |
7620 | } |
7621 | } |
7622 | |
7623 | // By default, just match reg + 0. |
7624 | Base = N; |
7625 | Offset = CurDAG->getTargetConstant(Val: 0, DL: SDLoc(N), VT: MVT::i64); |
7626 | return true; |
7627 | } |
7628 | |
7629 | bool AArch64DAGToDAGISel::SelectCmpBranchUImm6Operand(SDNode *P, SDValue N, |
7630 | SDValue &Imm) { |
7631 | AArch64CC::CondCode CC = |
7632 | static_cast<AArch64CC::CondCode>(P->getConstantOperandVal(Num: 1)); |
7633 | if (auto *CN = dyn_cast<ConstantSDNode>(Val&: N)) { |
7634 | // Check conservatively if the immediate fits the valid range [0, 64). |
7635 | // Immediate variants for GE and HS definitely need to be decremented |
7636 | // when lowering the pseudos later, so an immediate of 1 would become 0. |
7637 | // For the inverse conditions LT and LO we don't know for sure if they |
7638 | // will need a decrement but should the decision be made to reverse the |
7639 | // branch condition, we again end up with the need to decrement. |
7640 | // The same argument holds for LE, LS, GT and HI and possibly |
7641 | // incremented immediates. This can lead to slightly less optimal |
7642 | // codegen, e.g. we never codegen the legal case |
7643 | // cblt w0, #63, A |
7644 | // because we could end up with the illegal case |
7645 | // cbge w0, #64, B |
7646 | // should the decision to reverse the branch direction be made. For the |
7647 | // lower bound cases this is no problem since we can express comparisons |
7648 | // against 0 with either tbz/tnbz or using wzr/xzr. |
7649 | uint64_t LowerBound = 0, UpperBound = 64; |
7650 | switch (CC) { |
7651 | case AArch64CC::GE: |
7652 | case AArch64CC::HS: |
7653 | case AArch64CC::LT: |
7654 | case AArch64CC::LO: |
7655 | LowerBound = 1; |
7656 | break; |
7657 | case AArch64CC::LE: |
7658 | case AArch64CC::LS: |
7659 | case AArch64CC::GT: |
7660 | case AArch64CC::HI: |
7661 | UpperBound = 63; |
7662 | break; |
7663 | default: |
7664 | break; |
7665 | } |
7666 | |
7667 | if (CN->getAPIntValue().uge(RHS: LowerBound) && |
7668 | CN->getAPIntValue().ult(RHS: UpperBound)) { |
7669 | SDLoc DL(N); |
7670 | Imm = CurDAG->getTargetConstant(Val: CN->getZExtValue(), DL, VT: N.getValueType()); |
7671 | return true; |
7672 | } |
7673 | } |
7674 | |
7675 | return false; |
7676 | } |
7677 | |