1//===- SelectionDAG.cpp - Implement the SelectionDAG data structures ------===//
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 implements the SelectionDAG class.
10//
11//===----------------------------------------------------------------------===//
12
13#include "llvm/CodeGen/SelectionDAG.h"
14#include "SDNodeDbgValue.h"
15#include "llvm/ADT/APFloat.h"
16#include "llvm/ADT/APInt.h"
17#include "llvm/ADT/APSInt.h"
18#include "llvm/ADT/ArrayRef.h"
19#include "llvm/ADT/BitVector.h"
20#include "llvm/ADT/DenseSet.h"
21#include "llvm/ADT/FoldingSet.h"
22#include "llvm/ADT/STLExtras.h"
23#include "llvm/ADT/SmallPtrSet.h"
24#include "llvm/ADT/SmallVector.h"
25#include "llvm/ADT/Twine.h"
26#include "llvm/Analysis/AliasAnalysis.h"
27#include "llvm/Analysis/MemoryLocation.h"
28#include "llvm/Analysis/TargetLibraryInfo.h"
29#include "llvm/Analysis/ValueTracking.h"
30#include "llvm/Analysis/VectorUtils.h"
31#include "llvm/BinaryFormat/Dwarf.h"
32#include "llvm/CodeGen/Analysis.h"
33#include "llvm/CodeGen/FunctionLoweringInfo.h"
34#include "llvm/CodeGen/ISDOpcodes.h"
35#include "llvm/CodeGen/MachineBasicBlock.h"
36#include "llvm/CodeGen/MachineConstantPool.h"
37#include "llvm/CodeGen/MachineFrameInfo.h"
38#include "llvm/CodeGen/MachineFunction.h"
39#include "llvm/CodeGen/MachineMemOperand.h"
40#include "llvm/CodeGen/RuntimeLibcallUtil.h"
41#include "llvm/CodeGen/SDPatternMatch.h"
42#include "llvm/CodeGen/SelectionDAGAddressAnalysis.h"
43#include "llvm/CodeGen/SelectionDAGNodes.h"
44#include "llvm/CodeGen/SelectionDAGTargetInfo.h"
45#include "llvm/CodeGen/TargetFrameLowering.h"
46#include "llvm/CodeGen/TargetLowering.h"
47#include "llvm/CodeGen/TargetRegisterInfo.h"
48#include "llvm/CodeGen/TargetSubtargetInfo.h"
49#include "llvm/CodeGen/ValueTypes.h"
50#include "llvm/CodeGenTypes/MachineValueType.h"
51#include "llvm/IR/Constant.h"
52#include "llvm/IR/Constants.h"
53#include "llvm/IR/DataLayout.h"
54#include "llvm/IR/DebugInfoMetadata.h"
55#include "llvm/IR/DebugLoc.h"
56#include "llvm/IR/DerivedTypes.h"
57#include "llvm/IR/Function.h"
58#include "llvm/IR/GlobalValue.h"
59#include "llvm/IR/Metadata.h"
60#include "llvm/IR/Type.h"
61#include "llvm/Support/Casting.h"
62#include "llvm/Support/CodeGen.h"
63#include "llvm/Support/Compiler.h"
64#include "llvm/Support/Debug.h"
65#include "llvm/Support/ErrorHandling.h"
66#include "llvm/Support/KnownBits.h"
67#include "llvm/Support/MathExtras.h"
68#include "llvm/Support/raw_ostream.h"
69#include "llvm/Target/TargetMachine.h"
70#include "llvm/Target/TargetOptions.h"
71#include "llvm/TargetParser/Triple.h"
72#include "llvm/Transforms/Utils/SizeOpts.h"
73#include <algorithm>
74#include <cassert>
75#include <cstdint>
76#include <cstdlib>
77#include <limits>
78#include <optional>
79#include <set>
80#include <string>
81#include <utility>
82#include <vector>
83
84using namespace llvm;
85using namespace llvm::SDPatternMatch;
86
87/// makeVTList - Return an instance of the SDVTList struct initialized with the
88/// specified members.
89static SDVTList makeVTList(const EVT *VTs, unsigned NumVTs) {
90 SDVTList Res = {.VTs: VTs, .NumVTs: NumVTs};
91 return Res;
92}
93
94// Default null implementations of the callbacks.
95void SelectionDAG::DAGUpdateListener::NodeDeleted(SDNode*, SDNode*) {}
96void SelectionDAG::DAGUpdateListener::NodeUpdated(SDNode*) {}
97void SelectionDAG::DAGUpdateListener::NodeInserted(SDNode *) {}
98
99void SelectionDAG::DAGNodeDeletedListener::anchor() {}
100void SelectionDAG::DAGNodeInsertedListener::anchor() {}
101
102#define DEBUG_TYPE "selectiondag"
103
104static cl::opt<bool> EnableMemCpyDAGOpt("enable-memcpy-dag-opt",
105 cl::Hidden, cl::init(Val: true),
106 cl::desc("Gang up loads and stores generated by inlining of memcpy"));
107
108static cl::opt<int> MaxLdStGlue("ldstmemcpy-glue-max",
109 cl::desc("Number limit for gluing ld/st of memcpy."),
110 cl::Hidden, cl::init(Val: 0));
111
112static cl::opt<unsigned>
113 MaxSteps("has-predecessor-max-steps", cl::Hidden, cl::init(Val: 8192),
114 cl::desc("DAG combiner limit number of steps when searching DAG "
115 "for predecessor nodes"));
116
117static void NewSDValueDbgMsg(SDValue V, StringRef Msg, SelectionDAG *G) {
118 LLVM_DEBUG(dbgs() << Msg; V.getNode()->dump(G););
119}
120
121unsigned SelectionDAG::getHasPredecessorMaxSteps() { return MaxSteps; }
122
123//===----------------------------------------------------------------------===//
124// ConstantFPSDNode Class
125//===----------------------------------------------------------------------===//
126
127/// isExactlyValue - We don't rely on operator== working on double values, as
128/// it returns true for things that are clearly not equal, like -0.0 and 0.0.
129/// As such, this method can be used to do an exact bit-for-bit comparison of
130/// two floating point values.
131bool ConstantFPSDNode::isExactlyValue(const APFloat& V) const {
132 return getValueAPF().bitwiseIsEqual(RHS: V);
133}
134
135bool ConstantFPSDNode::isValueValidForType(EVT VT,
136 const APFloat& Val) {
137 assert(VT.isFloatingPoint() && "Can only convert between FP types");
138
139 // convert modifies in place, so make a copy.
140 APFloat Val2 = APFloat(Val);
141 bool losesInfo;
142 (void)Val2.convert(ToSemantics: VT.getFltSemantics(), RM: APFloat::rmNearestTiesToEven,
143 losesInfo: &losesInfo);
144 return !losesInfo;
145}
146
147//===----------------------------------------------------------------------===//
148// ISD Namespace
149//===----------------------------------------------------------------------===//
150
151bool ISD::isConstantSplatVector(const SDNode *N, APInt &SplatVal) {
152 if (N->getOpcode() == ISD::SPLAT_VECTOR) {
153 if (auto OptAPInt = N->getOperand(Num: 0)->bitcastToAPInt()) {
154 unsigned EltSize =
155 N->getValueType(ResNo: 0).getVectorElementType().getSizeInBits();
156 SplatVal = OptAPInt->trunc(width: EltSize);
157 return true;
158 }
159 }
160
161 auto *BV = dyn_cast<BuildVectorSDNode>(Val: N);
162 if (!BV)
163 return false;
164
165 APInt SplatUndef;
166 unsigned SplatBitSize;
167 bool HasUndefs;
168 unsigned EltSize = N->getValueType(ResNo: 0).getVectorElementType().getSizeInBits();
169 // Endianness does not matter here. We are checking for a splat given the
170 // element size of the vector, and if we find such a splat for little endian
171 // layout, then that should be valid also for big endian (as the full vector
172 // size is known to be a multiple of the element size).
173 const bool IsBigEndian = false;
174 return BV->isConstantSplat(SplatValue&: SplatVal, SplatUndef, SplatBitSize, HasAnyUndefs&: HasUndefs,
175 MinSplatBits: EltSize, isBigEndian: IsBigEndian) &&
176 EltSize == SplatBitSize;
177}
178
179// FIXME: AllOnes and AllZeros duplicate a lot of code. Could these be
180// specializations of the more general isConstantSplatVector()?
181
182bool ISD::isConstantSplatVectorAllOnes(const SDNode *N, bool BuildVectorOnly) {
183 // Look through a bit convert.
184 while (N->getOpcode() == ISD::BITCAST)
185 N = N->getOperand(Num: 0).getNode();
186
187 if (!BuildVectorOnly && N->getOpcode() == ISD::SPLAT_VECTOR) {
188 APInt SplatVal;
189 return isConstantSplatVector(N, SplatVal) && SplatVal.isAllOnes();
190 }
191
192 if (N->getOpcode() != ISD::BUILD_VECTOR) return false;
193
194 unsigned i = 0, e = N->getNumOperands();
195
196 // Skip over all of the undef values.
197 while (i != e && N->getOperand(Num: i).isUndef())
198 ++i;
199
200 // Do not accept an all-undef vector.
201 if (i == e) return false;
202
203 // Do not accept build_vectors that aren't all constants or which have non-~0
204 // elements. We have to be a bit careful here, as the type of the constant
205 // may not be the same as the type of the vector elements due to type
206 // legalization (the elements are promoted to a legal type for the target and
207 // a vector of a type may be legal when the base element type is not).
208 // We only want to check enough bits to cover the vector elements, because
209 // we care if the resultant vector is all ones, not whether the individual
210 // constants are.
211 SDValue NotZero = N->getOperand(Num: i);
212 if (auto OptAPInt = NotZero->bitcastToAPInt()) {
213 unsigned EltSize = N->getValueType(ResNo: 0).getScalarSizeInBits();
214 if (OptAPInt->countr_one() < EltSize)
215 return false;
216 } else
217 return false;
218
219 // Okay, we have at least one ~0 value, check to see if the rest match or are
220 // undefs. Even with the above element type twiddling, this should be OK, as
221 // the same type legalization should have applied to all the elements.
222 for (++i; i != e; ++i)
223 if (N->getOperand(Num: i) != NotZero && !N->getOperand(Num: i).isUndef())
224 return false;
225 return true;
226}
227
228bool ISD::isConstantSplatVectorAllZeros(const SDNode *N, bool BuildVectorOnly) {
229 // Look through a bit convert.
230 while (N->getOpcode() == ISD::BITCAST)
231 N = N->getOperand(Num: 0).getNode();
232
233 if (!BuildVectorOnly && N->getOpcode() == ISD::SPLAT_VECTOR) {
234 APInt SplatVal;
235 return isConstantSplatVector(N, SplatVal) && SplatVal.isZero();
236 }
237
238 if (N->getOpcode() != ISD::BUILD_VECTOR) return false;
239
240 bool IsAllUndef = true;
241 for (const SDValue &Op : N->op_values()) {
242 if (Op.isUndef())
243 continue;
244 IsAllUndef = false;
245 // Do not accept build_vectors that aren't all constants or which have non-0
246 // elements. We have to be a bit careful here, as the type of the constant
247 // may not be the same as the type of the vector elements due to type
248 // legalization (the elements are promoted to a legal type for the target
249 // and a vector of a type may be legal when the base element type is not).
250 // We only want to check enough bits to cover the vector elements, because
251 // we care if the resultant vector is all zeros, not whether the individual
252 // constants are.
253 if (auto OptAPInt = Op->bitcastToAPInt()) {
254 unsigned EltSize = N->getValueType(ResNo: 0).getScalarSizeInBits();
255 if (OptAPInt->countr_zero() < EltSize)
256 return false;
257 } else
258 return false;
259 }
260
261 // Do not accept an all-undef vector.
262 if (IsAllUndef)
263 return false;
264 return true;
265}
266
267bool ISD::isBuildVectorAllOnes(const SDNode *N) {
268 return isConstantSplatVectorAllOnes(N, /*BuildVectorOnly*/ true);
269}
270
271bool ISD::isBuildVectorAllZeros(const SDNode *N) {
272 return isConstantSplatVectorAllZeros(N, /*BuildVectorOnly*/ true);
273}
274
275bool ISD::isBuildVectorOfConstantSDNodes(const SDNode *N) {
276 if (N->getOpcode() != ISD::BUILD_VECTOR)
277 return false;
278
279 for (const SDValue &Op : N->op_values()) {
280 if (Op.isUndef())
281 continue;
282 if (!isa<ConstantSDNode>(Val: Op))
283 return false;
284 }
285 return true;
286}
287
288bool ISD::isBuildVectorOfConstantFPSDNodes(const SDNode *N) {
289 if (N->getOpcode() != ISD::BUILD_VECTOR)
290 return false;
291
292 for (const SDValue &Op : N->op_values()) {
293 if (Op.isUndef())
294 continue;
295 if (!isa<ConstantFPSDNode>(Val: Op))
296 return false;
297 }
298 return true;
299}
300
301bool ISD::isVectorShrinkable(const SDNode *N, unsigned NewEltSize,
302 bool Signed) {
303 assert(N->getValueType(0).isVector() && "Expected a vector!");
304
305 unsigned EltSize = N->getValueType(ResNo: 0).getScalarSizeInBits();
306 if (EltSize <= NewEltSize)
307 return false;
308
309 if (N->getOpcode() == ISD::ZERO_EXTEND) {
310 return (N->getOperand(Num: 0).getValueType().getScalarSizeInBits() <=
311 NewEltSize) &&
312 !Signed;
313 }
314 if (N->getOpcode() == ISD::SIGN_EXTEND) {
315 return (N->getOperand(Num: 0).getValueType().getScalarSizeInBits() <=
316 NewEltSize) &&
317 Signed;
318 }
319 if (N->getOpcode() != ISD::BUILD_VECTOR)
320 return false;
321
322 for (const SDValue &Op : N->op_values()) {
323 if (Op.isUndef())
324 continue;
325 if (!isa<ConstantSDNode>(Val: Op))
326 return false;
327
328 APInt C = Op->getAsAPIntVal().trunc(width: EltSize);
329 if (Signed && C.trunc(width: NewEltSize).sext(width: EltSize) != C)
330 return false;
331 if (!Signed && C.trunc(width: NewEltSize).zext(width: EltSize) != C)
332 return false;
333 }
334
335 return true;
336}
337
338bool ISD::allOperandsUndef(const SDNode *N) {
339 // Return false if the node has no operands.
340 // This is "logically inconsistent" with the definition of "all" but
341 // is probably the desired behavior.
342 if (N->getNumOperands() == 0)
343 return false;
344 return all_of(Range: N->op_values(), P: [](SDValue Op) { return Op.isUndef(); });
345}
346
347bool ISD::isFreezeUndef(const SDNode *N) {
348 return N->getOpcode() == ISD::FREEZE && N->getOperand(Num: 0).isUndef();
349}
350
351template <typename ConstNodeType>
352bool ISD::matchUnaryPredicateImpl(SDValue Op,
353 std::function<bool(ConstNodeType *)> Match,
354 bool AllowUndefs, bool AllowTruncation) {
355 // FIXME: Add support for scalar UNDEF cases?
356 if (auto *C = dyn_cast<ConstNodeType>(Op))
357 return Match(C);
358
359 // FIXME: Add support for vector UNDEF cases?
360 if (ISD::BUILD_VECTOR != Op.getOpcode() &&
361 ISD::SPLAT_VECTOR != Op.getOpcode())
362 return false;
363
364 EVT SVT = Op.getValueType().getScalarType();
365 for (unsigned i = 0, e = Op.getNumOperands(); i != e; ++i) {
366 if (AllowUndefs && Op.getOperand(i).isUndef()) {
367 if (!Match(nullptr))
368 return false;
369 continue;
370 }
371
372 auto *Cst = dyn_cast<ConstNodeType>(Op.getOperand(i));
373 if (!Cst || (!AllowTruncation && Cst->getValueType(0) != SVT) ||
374 !Match(Cst))
375 return false;
376 }
377 return true;
378}
379// Build used template types.
380template bool ISD::matchUnaryPredicateImpl<ConstantSDNode>(
381 SDValue, std::function<bool(ConstantSDNode *)>, bool, bool);
382template bool ISD::matchUnaryPredicateImpl<ConstantFPSDNode>(
383 SDValue, std::function<bool(ConstantFPSDNode *)>, bool, bool);
384
385bool ISD::matchBinaryPredicate(
386 SDValue LHS, SDValue RHS,
387 std::function<bool(ConstantSDNode *, ConstantSDNode *)> Match,
388 bool AllowUndefs, bool AllowTypeMismatch) {
389 if (!AllowTypeMismatch && LHS.getValueType() != RHS.getValueType())
390 return false;
391
392 // TODO: Add support for scalar UNDEF cases?
393 if (auto *LHSCst = dyn_cast<ConstantSDNode>(Val&: LHS))
394 if (auto *RHSCst = dyn_cast<ConstantSDNode>(Val&: RHS))
395 return Match(LHSCst, RHSCst);
396
397 // TODO: Add support for vector UNDEF cases?
398 if (LHS.getOpcode() != RHS.getOpcode() ||
399 (LHS.getOpcode() != ISD::BUILD_VECTOR &&
400 LHS.getOpcode() != ISD::SPLAT_VECTOR))
401 return false;
402
403 EVT SVT = LHS.getValueType().getScalarType();
404 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
405 SDValue LHSOp = LHS.getOperand(i);
406 SDValue RHSOp = RHS.getOperand(i);
407 bool LHSUndef = AllowUndefs && LHSOp.isUndef();
408 bool RHSUndef = AllowUndefs && RHSOp.isUndef();
409 auto *LHSCst = dyn_cast<ConstantSDNode>(Val&: LHSOp);
410 auto *RHSCst = dyn_cast<ConstantSDNode>(Val&: RHSOp);
411 if ((!LHSCst && !LHSUndef) || (!RHSCst && !RHSUndef))
412 return false;
413 if (!AllowTypeMismatch && (LHSOp.getValueType() != SVT ||
414 LHSOp.getValueType() != RHSOp.getValueType()))
415 return false;
416 if (!Match(LHSCst, RHSCst))
417 return false;
418 }
419 return true;
420}
421
422ISD::NodeType ISD::getInverseMinMaxOpcode(unsigned MinMaxOpc) {
423 switch (MinMaxOpc) {
424 default:
425 llvm_unreachable("unrecognized opcode");
426 case ISD::UMIN:
427 return ISD::UMAX;
428 case ISD::UMAX:
429 return ISD::UMIN;
430 case ISD::SMIN:
431 return ISD::SMAX;
432 case ISD::SMAX:
433 return ISD::SMIN;
434 }
435}
436
437ISD::NodeType ISD::getVecReduceBaseOpcode(unsigned VecReduceOpcode) {
438 switch (VecReduceOpcode) {
439 default:
440 llvm_unreachable("Expected VECREDUCE opcode");
441 case ISD::VECREDUCE_FADD:
442 case ISD::VECREDUCE_SEQ_FADD:
443 case ISD::VP_REDUCE_FADD:
444 case ISD::VP_REDUCE_SEQ_FADD:
445 return ISD::FADD;
446 case ISD::VECREDUCE_FMUL:
447 case ISD::VECREDUCE_SEQ_FMUL:
448 case ISD::VP_REDUCE_FMUL:
449 case ISD::VP_REDUCE_SEQ_FMUL:
450 return ISD::FMUL;
451 case ISD::VECREDUCE_ADD:
452 case ISD::VP_REDUCE_ADD:
453 return ISD::ADD;
454 case ISD::VECREDUCE_MUL:
455 case ISD::VP_REDUCE_MUL:
456 return ISD::MUL;
457 case ISD::VECREDUCE_AND:
458 case ISD::VP_REDUCE_AND:
459 return ISD::AND;
460 case ISD::VECREDUCE_OR:
461 case ISD::VP_REDUCE_OR:
462 return ISD::OR;
463 case ISD::VECREDUCE_XOR:
464 case ISD::VP_REDUCE_XOR:
465 return ISD::XOR;
466 case ISD::VECREDUCE_SMAX:
467 case ISD::VP_REDUCE_SMAX:
468 return ISD::SMAX;
469 case ISD::VECREDUCE_SMIN:
470 case ISD::VP_REDUCE_SMIN:
471 return ISD::SMIN;
472 case ISD::VECREDUCE_UMAX:
473 case ISD::VP_REDUCE_UMAX:
474 return ISD::UMAX;
475 case ISD::VECREDUCE_UMIN:
476 case ISD::VP_REDUCE_UMIN:
477 return ISD::UMIN;
478 case ISD::VECREDUCE_FMAX:
479 case ISD::VP_REDUCE_FMAX:
480 return ISD::FMAXNUM;
481 case ISD::VECREDUCE_FMIN:
482 case ISD::VP_REDUCE_FMIN:
483 return ISD::FMINNUM;
484 case ISD::VECREDUCE_FMAXIMUM:
485 case ISD::VP_REDUCE_FMAXIMUM:
486 return ISD::FMAXIMUM;
487 case ISD::VECREDUCE_FMINIMUM:
488 case ISD::VP_REDUCE_FMINIMUM:
489 return ISD::FMINIMUM;
490 }
491}
492
493bool ISD::isVPOpcode(unsigned Opcode) {
494 switch (Opcode) {
495 default:
496 return false;
497#define BEGIN_REGISTER_VP_SDNODE(VPSD, ...) \
498 case ISD::VPSD: \
499 return true;
500#include "llvm/IR/VPIntrinsics.def"
501 }
502}
503
504bool ISD::isVPBinaryOp(unsigned Opcode) {
505 switch (Opcode) {
506 default:
507 break;
508#define BEGIN_REGISTER_VP_SDNODE(VPSD, ...) case ISD::VPSD:
509#define VP_PROPERTY_BINARYOP return true;
510#define END_REGISTER_VP_SDNODE(VPSD) break;
511#include "llvm/IR/VPIntrinsics.def"
512 }
513 return false;
514}
515
516bool ISD::isVPReduction(unsigned Opcode) {
517 switch (Opcode) {
518 default:
519 return false;
520 case ISD::VP_REDUCE_ADD:
521 case ISD::VP_REDUCE_MUL:
522 case ISD::VP_REDUCE_AND:
523 case ISD::VP_REDUCE_OR:
524 case ISD::VP_REDUCE_XOR:
525 case ISD::VP_REDUCE_SMAX:
526 case ISD::VP_REDUCE_SMIN:
527 case ISD::VP_REDUCE_UMAX:
528 case ISD::VP_REDUCE_UMIN:
529 case ISD::VP_REDUCE_FMAX:
530 case ISD::VP_REDUCE_FMIN:
531 case ISD::VP_REDUCE_FMAXIMUM:
532 case ISD::VP_REDUCE_FMINIMUM:
533 case ISD::VP_REDUCE_FADD:
534 case ISD::VP_REDUCE_FMUL:
535 case ISD::VP_REDUCE_SEQ_FADD:
536 case ISD::VP_REDUCE_SEQ_FMUL:
537 return true;
538 }
539}
540
541/// The operand position of the vector mask.
542std::optional<unsigned> ISD::getVPMaskIdx(unsigned Opcode) {
543 switch (Opcode) {
544 default:
545 return std::nullopt;
546#define BEGIN_REGISTER_VP_SDNODE(VPSD, LEGALPOS, TDNAME, MASKPOS, ...) \
547 case ISD::VPSD: \
548 return MASKPOS;
549#include "llvm/IR/VPIntrinsics.def"
550 }
551}
552
553/// The operand position of the explicit vector length parameter.
554std::optional<unsigned> ISD::getVPExplicitVectorLengthIdx(unsigned Opcode) {
555 switch (Opcode) {
556 default:
557 return std::nullopt;
558#define BEGIN_REGISTER_VP_SDNODE(VPSD, LEGALPOS, TDNAME, MASKPOS, EVLPOS) \
559 case ISD::VPSD: \
560 return EVLPOS;
561#include "llvm/IR/VPIntrinsics.def"
562 }
563}
564
565std::optional<unsigned> ISD::getBaseOpcodeForVP(unsigned VPOpcode,
566 bool hasFPExcept) {
567 // FIXME: Return strict opcodes in case of fp exceptions.
568 switch (VPOpcode) {
569 default:
570 return std::nullopt;
571#define BEGIN_REGISTER_VP_SDNODE(VPOPC, ...) case ISD::VPOPC:
572#define VP_PROPERTY_FUNCTIONAL_SDOPC(SDOPC) return ISD::SDOPC;
573#define END_REGISTER_VP_SDNODE(VPOPC) break;
574#include "llvm/IR/VPIntrinsics.def"
575 }
576 return std::nullopt;
577}
578
579std::optional<unsigned> ISD::getVPForBaseOpcode(unsigned Opcode) {
580 switch (Opcode) {
581 default:
582 return std::nullopt;
583#define BEGIN_REGISTER_VP_SDNODE(VPOPC, ...) break;
584#define VP_PROPERTY_FUNCTIONAL_SDOPC(SDOPC) case ISD::SDOPC:
585#define END_REGISTER_VP_SDNODE(VPOPC) return ISD::VPOPC;
586#include "llvm/IR/VPIntrinsics.def"
587 }
588}
589
590ISD::NodeType ISD::getExtForLoadExtType(bool IsFP, ISD::LoadExtType ExtType) {
591 switch (ExtType) {
592 case ISD::EXTLOAD:
593 return IsFP ? ISD::FP_EXTEND : ISD::ANY_EXTEND;
594 case ISD::SEXTLOAD:
595 return ISD::SIGN_EXTEND;
596 case ISD::ZEXTLOAD:
597 return ISD::ZERO_EXTEND;
598 default:
599 break;
600 }
601
602 llvm_unreachable("Invalid LoadExtType");
603}
604
605ISD::CondCode ISD::getSetCCSwappedOperands(ISD::CondCode Operation) {
606 // To perform this operation, we just need to swap the L and G bits of the
607 // operation.
608 unsigned OldL = (Operation >> 2) & 1;
609 unsigned OldG = (Operation >> 1) & 1;
610 return ISD::CondCode((Operation & ~6) | // Keep the N, U, E bits
611 (OldL << 1) | // New G bit
612 (OldG << 2)); // New L bit.
613}
614
615static ISD::CondCode getSetCCInverseImpl(ISD::CondCode Op, bool isIntegerLike) {
616 unsigned Operation = Op;
617 if (isIntegerLike)
618 Operation ^= 7; // Flip L, G, E bits, but not U.
619 else
620 Operation ^= 15; // Flip all of the condition bits.
621
622 if (Operation > ISD::SETTRUE2)
623 Operation &= ~8; // Don't let N and U bits get set.
624
625 return ISD::CondCode(Operation);
626}
627
628ISD::CondCode ISD::getSetCCInverse(ISD::CondCode Op, EVT Type) {
629 return getSetCCInverseImpl(Op, isIntegerLike: Type.isInteger());
630}
631
632ISD::CondCode ISD::GlobalISel::getSetCCInverse(ISD::CondCode Op,
633 bool isIntegerLike) {
634 return getSetCCInverseImpl(Op, isIntegerLike);
635}
636
637/// For an integer comparison, return 1 if the comparison is a signed operation
638/// and 2 if the result is an unsigned comparison. Return zero if the operation
639/// does not depend on the sign of the input (setne and seteq).
640static int isSignedOp(ISD::CondCode Opcode) {
641 switch (Opcode) {
642 default: llvm_unreachable("Illegal integer setcc operation!");
643 case ISD::SETEQ:
644 case ISD::SETNE: return 0;
645 case ISD::SETLT:
646 case ISD::SETLE:
647 case ISD::SETGT:
648 case ISD::SETGE: return 1;
649 case ISD::SETULT:
650 case ISD::SETULE:
651 case ISD::SETUGT:
652 case ISD::SETUGE: return 2;
653 }
654}
655
656ISD::CondCode ISD::getSetCCOrOperation(ISD::CondCode Op1, ISD::CondCode Op2,
657 EVT Type) {
658 bool IsInteger = Type.isInteger();
659 if (IsInteger && (isSignedOp(Opcode: Op1) | isSignedOp(Opcode: Op2)) == 3)
660 // Cannot fold a signed integer setcc with an unsigned integer setcc.
661 return ISD::SETCC_INVALID;
662
663 unsigned Op = Op1 | Op2; // Combine all of the condition bits.
664
665 // If the N and U bits get set, then the resultant comparison DOES suddenly
666 // care about orderedness, and it is true when ordered.
667 if (Op > ISD::SETTRUE2)
668 Op &= ~16; // Clear the U bit if the N bit is set.
669
670 // Canonicalize illegal integer setcc's.
671 if (IsInteger && Op == ISD::SETUNE) // e.g. SETUGT | SETULT
672 Op = ISD::SETNE;
673
674 return ISD::CondCode(Op);
675}
676
677ISD::CondCode ISD::getSetCCAndOperation(ISD::CondCode Op1, ISD::CondCode Op2,
678 EVT Type) {
679 bool IsInteger = Type.isInteger();
680 if (IsInteger && (isSignedOp(Opcode: Op1) | isSignedOp(Opcode: Op2)) == 3)
681 // Cannot fold a signed setcc with an unsigned setcc.
682 return ISD::SETCC_INVALID;
683
684 // Combine all of the condition bits.
685 ISD::CondCode Result = ISD::CondCode(Op1 & Op2);
686
687 // Canonicalize illegal integer setcc's.
688 if (IsInteger) {
689 switch (Result) {
690 default: break;
691 case ISD::SETUO : Result = ISD::SETFALSE; break; // SETUGT & SETULT
692 case ISD::SETOEQ: // SETEQ & SETU[LG]E
693 case ISD::SETUEQ: Result = ISD::SETEQ ; break; // SETUGE & SETULE
694 case ISD::SETOLT: Result = ISD::SETULT ; break; // SETULT & SETNE
695 case ISD::SETOGT: Result = ISD::SETUGT ; break; // SETUGT & SETNE
696 }
697 }
698
699 return Result;
700}
701
702//===----------------------------------------------------------------------===//
703// SDNode Profile Support
704//===----------------------------------------------------------------------===//
705
706/// AddNodeIDOpcode - Add the node opcode to the NodeID data.
707static void AddNodeIDOpcode(FoldingSetNodeID &ID, unsigned OpC) {
708 ID.AddInteger(I: OpC);
709}
710
711/// AddNodeIDValueTypes - Value type lists are intern'd so we can represent them
712/// solely with their pointer.
713static void AddNodeIDValueTypes(FoldingSetNodeID &ID, SDVTList VTList) {
714 ID.AddPointer(Ptr: VTList.VTs);
715}
716
717/// AddNodeIDOperands - Various routines for adding operands to the NodeID data.
718static void AddNodeIDOperands(FoldingSetNodeID &ID,
719 ArrayRef<SDValue> Ops) {
720 for (const auto &Op : Ops) {
721 ID.AddPointer(Ptr: Op.getNode());
722 ID.AddInteger(I: Op.getResNo());
723 }
724}
725
726/// AddNodeIDOperands - Various routines for adding operands to the NodeID data.
727static void AddNodeIDOperands(FoldingSetNodeID &ID,
728 ArrayRef<SDUse> Ops) {
729 for (const auto &Op : Ops) {
730 ID.AddPointer(Ptr: Op.getNode());
731 ID.AddInteger(I: Op.getResNo());
732 }
733}
734
735static void AddNodeIDNode(FoldingSetNodeID &ID, unsigned OpC,
736 SDVTList VTList, ArrayRef<SDValue> OpList) {
737 AddNodeIDOpcode(ID, OpC);
738 AddNodeIDValueTypes(ID, VTList);
739 AddNodeIDOperands(ID, Ops: OpList);
740}
741
742/// If this is an SDNode with special info, add this info to the NodeID data.
743static void AddNodeIDCustom(FoldingSetNodeID &ID, const SDNode *N) {
744 switch (N->getOpcode()) {
745 case ISD::TargetExternalSymbol:
746 case ISD::ExternalSymbol:
747 case ISD::MCSymbol:
748 llvm_unreachable("Should only be used on nodes with operands");
749 default: break; // Normal nodes don't need extra info.
750 case ISD::TargetConstant:
751 case ISD::Constant: {
752 const ConstantSDNode *C = cast<ConstantSDNode>(Val: N);
753 ID.AddPointer(Ptr: C->getConstantIntValue());
754 ID.AddBoolean(B: C->isOpaque());
755 break;
756 }
757 case ISD::TargetConstantFP:
758 case ISD::ConstantFP:
759 ID.AddPointer(Ptr: cast<ConstantFPSDNode>(Val: N)->getConstantFPValue());
760 break;
761 case ISD::TargetGlobalAddress:
762 case ISD::GlobalAddress:
763 case ISD::TargetGlobalTLSAddress:
764 case ISD::GlobalTLSAddress: {
765 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Val: N);
766 ID.AddPointer(Ptr: GA->getGlobal());
767 ID.AddInteger(I: GA->getOffset());
768 ID.AddInteger(I: GA->getTargetFlags());
769 break;
770 }
771 case ISD::BasicBlock:
772 ID.AddPointer(Ptr: cast<BasicBlockSDNode>(Val: N)->getBasicBlock());
773 break;
774 case ISD::Register:
775 ID.AddInteger(I: cast<RegisterSDNode>(Val: N)->getReg().id());
776 break;
777 case ISD::RegisterMask:
778 ID.AddPointer(Ptr: cast<RegisterMaskSDNode>(Val: N)->getRegMask());
779 break;
780 case ISD::SRCVALUE:
781 ID.AddPointer(Ptr: cast<SrcValueSDNode>(Val: N)->getValue());
782 break;
783 case ISD::FrameIndex:
784 case ISD::TargetFrameIndex:
785 ID.AddInteger(I: cast<FrameIndexSDNode>(Val: N)->getIndex());
786 break;
787 case ISD::LIFETIME_START:
788 case ISD::LIFETIME_END:
789 if (cast<LifetimeSDNode>(Val: N)->hasOffset()) {
790 ID.AddInteger(I: cast<LifetimeSDNode>(Val: N)->getSize());
791 ID.AddInteger(I: cast<LifetimeSDNode>(Val: N)->getOffset());
792 }
793 break;
794 case ISD::PSEUDO_PROBE:
795 ID.AddInteger(I: cast<PseudoProbeSDNode>(Val: N)->getGuid());
796 ID.AddInteger(I: cast<PseudoProbeSDNode>(Val: N)->getIndex());
797 ID.AddInteger(I: cast<PseudoProbeSDNode>(Val: N)->getAttributes());
798 break;
799 case ISD::JumpTable:
800 case ISD::TargetJumpTable:
801 ID.AddInteger(I: cast<JumpTableSDNode>(Val: N)->getIndex());
802 ID.AddInteger(I: cast<JumpTableSDNode>(Val: N)->getTargetFlags());
803 break;
804 case ISD::ConstantPool:
805 case ISD::TargetConstantPool: {
806 const ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Val: N);
807 ID.AddInteger(I: CP->getAlign().value());
808 ID.AddInteger(I: CP->getOffset());
809 if (CP->isMachineConstantPoolEntry())
810 CP->getMachineCPVal()->addSelectionDAGCSEId(ID);
811 else
812 ID.AddPointer(Ptr: CP->getConstVal());
813 ID.AddInteger(I: CP->getTargetFlags());
814 break;
815 }
816 case ISD::TargetIndex: {
817 const TargetIndexSDNode *TI = cast<TargetIndexSDNode>(Val: N);
818 ID.AddInteger(I: TI->getIndex());
819 ID.AddInteger(I: TI->getOffset());
820 ID.AddInteger(I: TI->getTargetFlags());
821 break;
822 }
823 case ISD::LOAD: {
824 const LoadSDNode *LD = cast<LoadSDNode>(Val: N);
825 ID.AddInteger(I: LD->getMemoryVT().getRawBits());
826 ID.AddInteger(I: LD->getRawSubclassData());
827 ID.AddInteger(I: LD->getPointerInfo().getAddrSpace());
828 ID.AddInteger(I: LD->getMemOperand()->getFlags());
829 break;
830 }
831 case ISD::STORE: {
832 const StoreSDNode *ST = cast<StoreSDNode>(Val: N);
833 ID.AddInteger(I: ST->getMemoryVT().getRawBits());
834 ID.AddInteger(I: ST->getRawSubclassData());
835 ID.AddInteger(I: ST->getPointerInfo().getAddrSpace());
836 ID.AddInteger(I: ST->getMemOperand()->getFlags());
837 break;
838 }
839 case ISD::VP_LOAD: {
840 const VPLoadSDNode *ELD = cast<VPLoadSDNode>(Val: N);
841 ID.AddInteger(I: ELD->getMemoryVT().getRawBits());
842 ID.AddInteger(I: ELD->getRawSubclassData());
843 ID.AddInteger(I: ELD->getPointerInfo().getAddrSpace());
844 ID.AddInteger(I: ELD->getMemOperand()->getFlags());
845 break;
846 }
847 case ISD::VP_STORE: {
848 const VPStoreSDNode *EST = cast<VPStoreSDNode>(Val: N);
849 ID.AddInteger(I: EST->getMemoryVT().getRawBits());
850 ID.AddInteger(I: EST->getRawSubclassData());
851 ID.AddInteger(I: EST->getPointerInfo().getAddrSpace());
852 ID.AddInteger(I: EST->getMemOperand()->getFlags());
853 break;
854 }
855 case ISD::EXPERIMENTAL_VP_STRIDED_LOAD: {
856 const VPStridedLoadSDNode *SLD = cast<VPStridedLoadSDNode>(Val: N);
857 ID.AddInteger(I: SLD->getMemoryVT().getRawBits());
858 ID.AddInteger(I: SLD->getRawSubclassData());
859 ID.AddInteger(I: SLD->getPointerInfo().getAddrSpace());
860 break;
861 }
862 case ISD::EXPERIMENTAL_VP_STRIDED_STORE: {
863 const VPStridedStoreSDNode *SST = cast<VPStridedStoreSDNode>(Val: N);
864 ID.AddInteger(I: SST->getMemoryVT().getRawBits());
865 ID.AddInteger(I: SST->getRawSubclassData());
866 ID.AddInteger(I: SST->getPointerInfo().getAddrSpace());
867 break;
868 }
869 case ISD::VP_GATHER: {
870 const VPGatherSDNode *EG = cast<VPGatherSDNode>(Val: N);
871 ID.AddInteger(I: EG->getMemoryVT().getRawBits());
872 ID.AddInteger(I: EG->getRawSubclassData());
873 ID.AddInteger(I: EG->getPointerInfo().getAddrSpace());
874 ID.AddInteger(I: EG->getMemOperand()->getFlags());
875 break;
876 }
877 case ISD::VP_SCATTER: {
878 const VPScatterSDNode *ES = cast<VPScatterSDNode>(Val: N);
879 ID.AddInteger(I: ES->getMemoryVT().getRawBits());
880 ID.AddInteger(I: ES->getRawSubclassData());
881 ID.AddInteger(I: ES->getPointerInfo().getAddrSpace());
882 ID.AddInteger(I: ES->getMemOperand()->getFlags());
883 break;
884 }
885 case ISD::MLOAD: {
886 const MaskedLoadSDNode *MLD = cast<MaskedLoadSDNode>(Val: N);
887 ID.AddInteger(I: MLD->getMemoryVT().getRawBits());
888 ID.AddInteger(I: MLD->getRawSubclassData());
889 ID.AddInteger(I: MLD->getPointerInfo().getAddrSpace());
890 ID.AddInteger(I: MLD->getMemOperand()->getFlags());
891 break;
892 }
893 case ISD::MSTORE: {
894 const MaskedStoreSDNode *MST = cast<MaskedStoreSDNode>(Val: N);
895 ID.AddInteger(I: MST->getMemoryVT().getRawBits());
896 ID.AddInteger(I: MST->getRawSubclassData());
897 ID.AddInteger(I: MST->getPointerInfo().getAddrSpace());
898 ID.AddInteger(I: MST->getMemOperand()->getFlags());
899 break;
900 }
901 case ISD::MGATHER: {
902 const MaskedGatherSDNode *MG = cast<MaskedGatherSDNode>(Val: N);
903 ID.AddInteger(I: MG->getMemoryVT().getRawBits());
904 ID.AddInteger(I: MG->getRawSubclassData());
905 ID.AddInteger(I: MG->getPointerInfo().getAddrSpace());
906 ID.AddInteger(I: MG->getMemOperand()->getFlags());
907 break;
908 }
909 case ISD::MSCATTER: {
910 const MaskedScatterSDNode *MS = cast<MaskedScatterSDNode>(Val: N);
911 ID.AddInteger(I: MS->getMemoryVT().getRawBits());
912 ID.AddInteger(I: MS->getRawSubclassData());
913 ID.AddInteger(I: MS->getPointerInfo().getAddrSpace());
914 ID.AddInteger(I: MS->getMemOperand()->getFlags());
915 break;
916 }
917 case ISD::ATOMIC_CMP_SWAP:
918 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
919 case ISD::ATOMIC_SWAP:
920 case ISD::ATOMIC_LOAD_ADD:
921 case ISD::ATOMIC_LOAD_SUB:
922 case ISD::ATOMIC_LOAD_AND:
923 case ISD::ATOMIC_LOAD_CLR:
924 case ISD::ATOMIC_LOAD_OR:
925 case ISD::ATOMIC_LOAD_XOR:
926 case ISD::ATOMIC_LOAD_NAND:
927 case ISD::ATOMIC_LOAD_MIN:
928 case ISD::ATOMIC_LOAD_MAX:
929 case ISD::ATOMIC_LOAD_UMIN:
930 case ISD::ATOMIC_LOAD_UMAX:
931 case ISD::ATOMIC_LOAD:
932 case ISD::ATOMIC_STORE: {
933 const AtomicSDNode *AT = cast<AtomicSDNode>(Val: N);
934 ID.AddInteger(I: AT->getMemoryVT().getRawBits());
935 ID.AddInteger(I: AT->getRawSubclassData());
936 ID.AddInteger(I: AT->getPointerInfo().getAddrSpace());
937 ID.AddInteger(I: AT->getMemOperand()->getFlags());
938 break;
939 }
940 case ISD::VECTOR_SHUFFLE: {
941 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Val: N)->getMask();
942 for (int M : Mask)
943 ID.AddInteger(I: M);
944 break;
945 }
946 case ISD::ADDRSPACECAST: {
947 const AddrSpaceCastSDNode *ASC = cast<AddrSpaceCastSDNode>(Val: N);
948 ID.AddInteger(I: ASC->getSrcAddressSpace());
949 ID.AddInteger(I: ASC->getDestAddressSpace());
950 break;
951 }
952 case ISD::TargetBlockAddress:
953 case ISD::BlockAddress: {
954 const BlockAddressSDNode *BA = cast<BlockAddressSDNode>(Val: N);
955 ID.AddPointer(Ptr: BA->getBlockAddress());
956 ID.AddInteger(I: BA->getOffset());
957 ID.AddInteger(I: BA->getTargetFlags());
958 break;
959 }
960 case ISD::AssertAlign:
961 ID.AddInteger(I: cast<AssertAlignSDNode>(Val: N)->getAlign().value());
962 break;
963 case ISD::PREFETCH:
964 case ISD::INTRINSIC_VOID:
965 case ISD::INTRINSIC_W_CHAIN:
966 // Handled by MemIntrinsicSDNode check after the switch.
967 break;
968 case ISD::MDNODE_SDNODE:
969 ID.AddPointer(Ptr: cast<MDNodeSDNode>(Val: N)->getMD());
970 break;
971 } // end switch (N->getOpcode())
972
973 // MemIntrinsic nodes could also have subclass data, address spaces, and flags
974 // to check.
975 if (auto *MN = dyn_cast<MemIntrinsicSDNode>(Val: N)) {
976 ID.AddInteger(I: MN->getRawSubclassData());
977 ID.AddInteger(I: MN->getPointerInfo().getAddrSpace());
978 ID.AddInteger(I: MN->getMemOperand()->getFlags());
979 ID.AddInteger(I: MN->getMemoryVT().getRawBits());
980 }
981}
982
983/// AddNodeIDNode - Generic routine for adding a nodes info to the NodeID
984/// data.
985static void AddNodeIDNode(FoldingSetNodeID &ID, const SDNode *N) {
986 AddNodeIDOpcode(ID, OpC: N->getOpcode());
987 // Add the return value info.
988 AddNodeIDValueTypes(ID, VTList: N->getVTList());
989 // Add the operand info.
990 AddNodeIDOperands(ID, Ops: N->ops());
991
992 // Handle SDNode leafs with special info.
993 AddNodeIDCustom(ID, N);
994}
995
996//===----------------------------------------------------------------------===//
997// SelectionDAG Class
998//===----------------------------------------------------------------------===//
999
1000/// doNotCSE - Return true if CSE should not be performed for this node.
1001static bool doNotCSE(SDNode *N) {
1002 if (N->getValueType(ResNo: 0) == MVT::Glue)
1003 return true; // Never CSE anything that produces a glue result.
1004
1005 switch (N->getOpcode()) {
1006 default: break;
1007 case ISD::HANDLENODE:
1008 case ISD::EH_LABEL:
1009 return true; // Never CSE these nodes.
1010 }
1011
1012 // Check that remaining values produced are not flags.
1013 for (unsigned i = 1, e = N->getNumValues(); i != e; ++i)
1014 if (N->getValueType(ResNo: i) == MVT::Glue)
1015 return true; // Never CSE anything that produces a glue result.
1016
1017 return false;
1018}
1019
1020/// RemoveDeadNodes - This method deletes all unreachable nodes in the
1021/// SelectionDAG.
1022void SelectionDAG::RemoveDeadNodes() {
1023 // Create a dummy node (which is not added to allnodes), that adds a reference
1024 // to the root node, preventing it from being deleted.
1025 HandleSDNode Dummy(getRoot());
1026
1027 SmallVector<SDNode*, 128> DeadNodes;
1028
1029 // Add all obviously-dead nodes to the DeadNodes worklist.
1030 for (SDNode &Node : allnodes())
1031 if (Node.use_empty())
1032 DeadNodes.push_back(Elt: &Node);
1033
1034 RemoveDeadNodes(DeadNodes);
1035
1036 // If the root changed (e.g. it was a dead load, update the root).
1037 setRoot(Dummy.getValue());
1038}
1039
1040/// RemoveDeadNodes - This method deletes the unreachable nodes in the
1041/// given list, and any nodes that become unreachable as a result.
1042void SelectionDAG::RemoveDeadNodes(SmallVectorImpl<SDNode *> &DeadNodes) {
1043
1044 // Process the worklist, deleting the nodes and adding their uses to the
1045 // worklist.
1046 while (!DeadNodes.empty()) {
1047 SDNode *N = DeadNodes.pop_back_val();
1048 // Skip to next node if we've already managed to delete the node. This could
1049 // happen if replacing a node causes a node previously added to the node to
1050 // be deleted.
1051 if (N->getOpcode() == ISD::DELETED_NODE)
1052 continue;
1053
1054 for (DAGUpdateListener *DUL = UpdateListeners; DUL; DUL = DUL->Next)
1055 DUL->NodeDeleted(N, nullptr);
1056
1057 // Take the node out of the appropriate CSE map.
1058 RemoveNodeFromCSEMaps(N);
1059
1060 // Next, brutally remove the operand list. This is safe to do, as there are
1061 // no cycles in the graph.
1062 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) {
1063 SDUse &Use = *I++;
1064 SDNode *Operand = Use.getNode();
1065 Use.set(SDValue());
1066
1067 // Now that we removed this operand, see if there are no uses of it left.
1068 if (Operand->use_empty())
1069 DeadNodes.push_back(Elt: Operand);
1070 }
1071
1072 DeallocateNode(N);
1073 }
1074}
1075
1076void SelectionDAG::RemoveDeadNode(SDNode *N){
1077 SmallVector<SDNode*, 16> DeadNodes(1, N);
1078
1079 // Create a dummy node that adds a reference to the root node, preventing
1080 // it from being deleted. (This matters if the root is an operand of the
1081 // dead node.)
1082 HandleSDNode Dummy(getRoot());
1083
1084 RemoveDeadNodes(DeadNodes);
1085}
1086
1087void SelectionDAG::DeleteNode(SDNode *N) {
1088 // First take this out of the appropriate CSE map.
1089 RemoveNodeFromCSEMaps(N);
1090
1091 // Finally, remove uses due to operands of this node, remove from the
1092 // AllNodes list, and delete the node.
1093 DeleteNodeNotInCSEMaps(N);
1094}
1095
1096void SelectionDAG::DeleteNodeNotInCSEMaps(SDNode *N) {
1097 assert(N->getIterator() != AllNodes.begin() &&
1098 "Cannot delete the entry node!");
1099 assert(N->use_empty() && "Cannot delete a node that is not dead!");
1100
1101 // Drop all of the operands and decrement used node's use counts.
1102 N->DropOperands();
1103
1104 DeallocateNode(N);
1105}
1106
1107void SDDbgInfo::add(SDDbgValue *V, bool isParameter) {
1108 assert(!(V->isVariadic() && isParameter));
1109 if (isParameter)
1110 ByvalParmDbgValues.push_back(Elt: V);
1111 else
1112 DbgValues.push_back(Elt: V);
1113 for (const SDNode *Node : V->getSDNodes())
1114 if (Node)
1115 DbgValMap[Node].push_back(Elt: V);
1116}
1117
1118void SDDbgInfo::erase(const SDNode *Node) {
1119 DbgValMapType::iterator I = DbgValMap.find(Val: Node);
1120 if (I == DbgValMap.end())
1121 return;
1122 for (auto &Val: I->second)
1123 Val->setIsInvalidated();
1124 DbgValMap.erase(I);
1125}
1126
1127void SelectionDAG::DeallocateNode(SDNode *N) {
1128 // If we have operands, deallocate them.
1129 removeOperands(Node: N);
1130
1131 NodeAllocator.Deallocate(E: AllNodes.remove(IT: N));
1132
1133 // Set the opcode to DELETED_NODE to help catch bugs when node
1134 // memory is reallocated.
1135 // FIXME: There are places in SDag that have grown a dependency on the opcode
1136 // value in the released node.
1137 __asan_unpoison_memory_region(&N->NodeType, sizeof(N->NodeType));
1138 N->NodeType = ISD::DELETED_NODE;
1139
1140 // If any of the SDDbgValue nodes refer to this SDNode, invalidate
1141 // them and forget about that node.
1142 DbgInfo->erase(Node: N);
1143
1144 // Invalidate extra info.
1145 SDEI.erase(Val: N);
1146}
1147
1148#ifndef NDEBUG
1149/// VerifySDNode - Check the given SDNode. Aborts if it is invalid.
1150void SelectionDAG::verifyNode(SDNode *N) const {
1151 switch (N->getOpcode()) {
1152 default:
1153 if (N->isTargetOpcode())
1154 getSelectionDAGInfo().verifyTargetNode(DAG: *this, N);
1155 break;
1156 case ISD::BUILD_PAIR: {
1157 EVT VT = N->getValueType(ResNo: 0);
1158 assert(N->getNumValues() == 1 && "Too many results!");
1159 assert(!VT.isVector() && (VT.isInteger() || VT.isFloatingPoint()) &&
1160 "Wrong return type!");
1161 assert(N->getNumOperands() == 2 && "Wrong number of operands!");
1162 assert(N->getOperand(0).getValueType() == N->getOperand(1).getValueType() &&
1163 "Mismatched operand types!");
1164 assert(N->getOperand(0).getValueType().isInteger() == VT.isInteger() &&
1165 "Wrong operand type!");
1166 assert(VT.getSizeInBits() == 2 * N->getOperand(0).getValueSizeInBits() &&
1167 "Wrong return type size");
1168 break;
1169 }
1170 case ISD::BUILD_VECTOR: {
1171 assert(N->getNumValues() == 1 && "Too many results!");
1172 assert(N->getValueType(0).isVector() && "Wrong return type!");
1173 assert(N->getNumOperands() == N->getValueType(0).getVectorNumElements() &&
1174 "Wrong number of operands!");
1175 EVT EltVT = N->getValueType(ResNo: 0).getVectorElementType();
1176 for (const SDUse &Op : N->ops()) {
1177 assert((Op.getValueType() == EltVT ||
1178 (EltVT.isInteger() && Op.getValueType().isInteger() &&
1179 EltVT.bitsLE(Op.getValueType()))) &&
1180 "Wrong operand type!");
1181 assert(Op.getValueType() == N->getOperand(0).getValueType() &&
1182 "Operands must all have the same type");
1183 }
1184 break;
1185 }
1186 }
1187}
1188#endif // NDEBUG
1189
1190/// Insert a newly allocated node into the DAG.
1191///
1192/// Handles insertion into the all nodes list and CSE map, as well as
1193/// verification and other common operations when a new node is allocated.
1194void SelectionDAG::InsertNode(SDNode *N) {
1195 AllNodes.push_back(val: N);
1196#ifndef NDEBUG
1197 N->PersistentId = NextPersistentId++;
1198 verifyNode(N);
1199#endif
1200 for (DAGUpdateListener *DUL = UpdateListeners; DUL; DUL = DUL->Next)
1201 DUL->NodeInserted(N);
1202}
1203
1204/// RemoveNodeFromCSEMaps - Take the specified node out of the CSE map that
1205/// correspond to it. This is useful when we're about to delete or repurpose
1206/// the node. We don't want future request for structurally identical nodes
1207/// to return N anymore.
1208bool SelectionDAG::RemoveNodeFromCSEMaps(SDNode *N) {
1209 bool Erased = false;
1210 switch (N->getOpcode()) {
1211 case ISD::HANDLENODE: return false; // noop.
1212 case ISD::CONDCODE:
1213 assert(CondCodeNodes[cast<CondCodeSDNode>(N)->get()] &&
1214 "Cond code doesn't exist!");
1215 Erased = CondCodeNodes[cast<CondCodeSDNode>(Val: N)->get()] != nullptr;
1216 CondCodeNodes[cast<CondCodeSDNode>(Val: N)->get()] = nullptr;
1217 break;
1218 case ISD::ExternalSymbol:
1219 Erased = ExternalSymbols.erase(Key: cast<ExternalSymbolSDNode>(Val: N)->getSymbol());
1220 break;
1221 case ISD::TargetExternalSymbol: {
1222 ExternalSymbolSDNode *ESN = cast<ExternalSymbolSDNode>(Val: N);
1223 Erased = TargetExternalSymbols.erase(x: std::pair<std::string, unsigned>(
1224 ESN->getSymbol(), ESN->getTargetFlags()));
1225 break;
1226 }
1227 case ISD::MCSymbol: {
1228 auto *MCSN = cast<MCSymbolSDNode>(Val: N);
1229 Erased = MCSymbols.erase(Val: MCSN->getMCSymbol());
1230 break;
1231 }
1232 case ISD::VALUETYPE: {
1233 EVT VT = cast<VTSDNode>(Val: N)->getVT();
1234 if (VT.isExtended()) {
1235 Erased = ExtendedValueTypeNodes.erase(x: VT);
1236 } else {
1237 Erased = ValueTypeNodes[VT.getSimpleVT().SimpleTy] != nullptr;
1238 ValueTypeNodes[VT.getSimpleVT().SimpleTy] = nullptr;
1239 }
1240 break;
1241 }
1242 default:
1243 // Remove it from the CSE Map.
1244 assert(N->getOpcode() != ISD::DELETED_NODE && "DELETED_NODE in CSEMap!");
1245 assert(N->getOpcode() != ISD::EntryToken && "EntryToken in CSEMap!");
1246 Erased = CSEMap.RemoveNode(N);
1247 break;
1248 }
1249#ifndef NDEBUG
1250 // Verify that the node was actually in one of the CSE maps, unless it has a
1251 // glue result (which cannot be CSE'd) or is one of the special cases that are
1252 // not subject to CSE.
1253 if (!Erased && N->getValueType(ResNo: N->getNumValues()-1) != MVT::Glue &&
1254 !N->isMachineOpcode() && !doNotCSE(N)) {
1255 N->dump(G: this);
1256 dbgs() << "\n";
1257 llvm_unreachable("Node is not in map!");
1258 }
1259#endif
1260 return Erased;
1261}
1262
1263/// AddModifiedNodeToCSEMaps - The specified node has been removed from the CSE
1264/// maps and modified in place. Add it back to the CSE maps, unless an identical
1265/// node already exists, in which case transfer all its users to the existing
1266/// node. This transfer can potentially trigger recursive merging.
1267void
1268SelectionDAG::AddModifiedNodeToCSEMaps(SDNode *N) {
1269 // For node types that aren't CSE'd, just act as if no identical node
1270 // already exists.
1271 if (!doNotCSE(N)) {
1272 SDNode *Existing = CSEMap.GetOrInsertNode(N);
1273 if (Existing != N) {
1274 // If there was already an existing matching node, use ReplaceAllUsesWith
1275 // to replace the dead one with the existing one. This can cause
1276 // recursive merging of other unrelated nodes down the line.
1277 Existing->intersectFlagsWith(Flags: N->getFlags());
1278 ReplaceAllUsesWith(From: N, To: Existing);
1279
1280 // N is now dead. Inform the listeners and delete it.
1281 for (DAGUpdateListener *DUL = UpdateListeners; DUL; DUL = DUL->Next)
1282 DUL->NodeDeleted(N, Existing);
1283 DeleteNodeNotInCSEMaps(N);
1284 return;
1285 }
1286 }
1287
1288 // If the node doesn't already exist, we updated it. Inform listeners.
1289 for (DAGUpdateListener *DUL = UpdateListeners; DUL; DUL = DUL->Next)
1290 DUL->NodeUpdated(N);
1291}
1292
1293/// FindModifiedNodeSlot - Find a slot for the specified node if its operands
1294/// were replaced with those specified. If this node is never memoized,
1295/// return null, otherwise return a pointer to the slot it would take. If a
1296/// node already exists with these operands, the slot will be non-null.
1297SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, SDValue Op,
1298 void *&InsertPos) {
1299 if (doNotCSE(N))
1300 return nullptr;
1301
1302 SDValue Ops[] = { Op };
1303 FoldingSetNodeID ID;
1304 AddNodeIDNode(ID, OpC: N->getOpcode(), VTList: N->getVTList(), OpList: Ops);
1305 AddNodeIDCustom(ID, N);
1306 SDNode *Node = FindNodeOrInsertPos(ID, DL: SDLoc(N), InsertPos);
1307 if (Node)
1308 Node->intersectFlagsWith(Flags: N->getFlags());
1309 return Node;
1310}
1311
1312/// FindModifiedNodeSlot - Find a slot for the specified node if its operands
1313/// were replaced with those specified. If this node is never memoized,
1314/// return null, otherwise return a pointer to the slot it would take. If a
1315/// node already exists with these operands, the slot will be non-null.
1316SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N,
1317 SDValue Op1, SDValue Op2,
1318 void *&InsertPos) {
1319 if (doNotCSE(N))
1320 return nullptr;
1321
1322 SDValue Ops[] = { Op1, Op2 };
1323 FoldingSetNodeID ID;
1324 AddNodeIDNode(ID, OpC: N->getOpcode(), VTList: N->getVTList(), OpList: Ops);
1325 AddNodeIDCustom(ID, N);
1326 SDNode *Node = FindNodeOrInsertPos(ID, DL: SDLoc(N), InsertPos);
1327 if (Node)
1328 Node->intersectFlagsWith(Flags: N->getFlags());
1329 return Node;
1330}
1331
1332/// FindModifiedNodeSlot - Find a slot for the specified node if its operands
1333/// were replaced with those specified. If this node is never memoized,
1334/// return null, otherwise return a pointer to the slot it would take. If a
1335/// node already exists with these operands, the slot will be non-null.
1336SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, ArrayRef<SDValue> Ops,
1337 void *&InsertPos) {
1338 if (doNotCSE(N))
1339 return nullptr;
1340
1341 FoldingSetNodeID ID;
1342 AddNodeIDNode(ID, OpC: N->getOpcode(), VTList: N->getVTList(), OpList: Ops);
1343 AddNodeIDCustom(ID, N);
1344 SDNode *Node = FindNodeOrInsertPos(ID, DL: SDLoc(N), InsertPos);
1345 if (Node)
1346 Node->intersectFlagsWith(Flags: N->getFlags());
1347 return Node;
1348}
1349
1350Align SelectionDAG::getEVTAlign(EVT VT) const {
1351 Type *Ty = VT == MVT::iPTR ? PointerType::get(C&: *getContext(), AddressSpace: 0)
1352 : VT.getTypeForEVT(Context&: *getContext());
1353
1354 return getDataLayout().getABITypeAlign(Ty);
1355}
1356
1357// EntryNode could meaningfully have debug info if we can find it...
1358SelectionDAG::SelectionDAG(const TargetMachine &tm, CodeGenOptLevel OL)
1359 : TM(tm), OptLevel(OL), EntryNode(ISD::EntryToken, 0, DebugLoc(),
1360 getVTList(MVT::Other, MVT::Glue)),
1361 Root(getEntryNode()) {
1362 InsertNode(N: &EntryNode);
1363 DbgInfo = new SDDbgInfo();
1364}
1365
1366void SelectionDAG::init(MachineFunction &NewMF,
1367 OptimizationRemarkEmitter &NewORE, Pass *PassPtr,
1368 const TargetLibraryInfo *LibraryInfo,
1369 UniformityInfo *NewUA, ProfileSummaryInfo *PSIin,
1370 BlockFrequencyInfo *BFIin, MachineModuleInfo &MMIin,
1371 FunctionVarLocs const *VarLocs) {
1372 MF = &NewMF;
1373 SDAGISelPass = PassPtr;
1374 ORE = &NewORE;
1375 TLI = getSubtarget().getTargetLowering();
1376 TSI = getSubtarget().getSelectionDAGInfo();
1377 LibInfo = LibraryInfo;
1378 Context = &MF->getFunction().getContext();
1379 UA = NewUA;
1380 PSI = PSIin;
1381 BFI = BFIin;
1382 MMI = &MMIin;
1383 FnVarLocs = VarLocs;
1384}
1385
1386SelectionDAG::~SelectionDAG() {
1387 assert(!UpdateListeners && "Dangling registered DAGUpdateListeners");
1388 allnodes_clear();
1389 OperandRecycler.clear(OperandAllocator);
1390 delete DbgInfo;
1391}
1392
1393bool SelectionDAG::shouldOptForSize() const {
1394 return llvm::shouldOptimizeForSize(BB: FLI->MBB->getBasicBlock(), PSI, BFI);
1395}
1396
1397void SelectionDAG::allnodes_clear() {
1398 assert(&*AllNodes.begin() == &EntryNode);
1399 AllNodes.remove(IT: AllNodes.begin());
1400 while (!AllNodes.empty())
1401 DeallocateNode(N: &AllNodes.front());
1402#ifndef NDEBUG
1403 NextPersistentId = 0;
1404#endif
1405}
1406
1407SDNode *SelectionDAG::FindNodeOrInsertPos(const FoldingSetNodeID &ID,
1408 void *&InsertPos) {
1409 SDNode *N = CSEMap.FindNodeOrInsertPos(ID, InsertPos);
1410 if (N) {
1411 switch (N->getOpcode()) {
1412 default: break;
1413 case ISD::Constant:
1414 case ISD::ConstantFP:
1415 llvm_unreachable("Querying for Constant and ConstantFP nodes requires "
1416 "debug location. Use another overload.");
1417 }
1418 }
1419 return N;
1420}
1421
1422SDNode *SelectionDAG::FindNodeOrInsertPos(const FoldingSetNodeID &ID,
1423 const SDLoc &DL, void *&InsertPos) {
1424 SDNode *N = CSEMap.FindNodeOrInsertPos(ID, InsertPos);
1425 if (N) {
1426 switch (N->getOpcode()) {
1427 case ISD::Constant:
1428 case ISD::ConstantFP:
1429 // Erase debug location from the node if the node is used at several
1430 // different places. Do not propagate one location to all uses as it
1431 // will cause a worse single stepping debugging experience.
1432 if (N->getDebugLoc() != DL.getDebugLoc())
1433 N->setDebugLoc(DebugLoc());
1434 break;
1435 default:
1436 // When the node's point of use is located earlier in the instruction
1437 // sequence than its prior point of use, update its debug info to the
1438 // earlier location.
1439 if (DL.getIROrder() && DL.getIROrder() < N->getIROrder())
1440 N->setDebugLoc(DL.getDebugLoc());
1441 break;
1442 }
1443 }
1444 return N;
1445}
1446
1447void SelectionDAG::clear() {
1448 allnodes_clear();
1449 OperandRecycler.clear(OperandAllocator);
1450 OperandAllocator.Reset();
1451 CSEMap.clear();
1452
1453 ExtendedValueTypeNodes.clear();
1454 ExternalSymbols.clear();
1455 TargetExternalSymbols.clear();
1456 MCSymbols.clear();
1457 SDEI.clear();
1458 std::fill(first: CondCodeNodes.begin(), last: CondCodeNodes.end(), value: nullptr);
1459 std::fill(first: ValueTypeNodes.begin(), last: ValueTypeNodes.end(), value: nullptr);
1460
1461 EntryNode.UseList = nullptr;
1462 InsertNode(N: &EntryNode);
1463 Root = getEntryNode();
1464 DbgInfo->clear();
1465}
1466
1467SDValue SelectionDAG::getFPExtendOrRound(SDValue Op, const SDLoc &DL, EVT VT) {
1468 return VT.bitsGT(VT: Op.getValueType())
1469 ? getNode(Opcode: ISD::FP_EXTEND, DL, VT, Operand: Op)
1470 : getNode(Opcode: ISD::FP_ROUND, DL, VT, N1: Op,
1471 N2: getIntPtrConstant(Val: 0, DL, /*isTarget=*/true));
1472}
1473
1474std::pair<SDValue, SDValue>
1475SelectionDAG::getStrictFPExtendOrRound(SDValue Op, SDValue Chain,
1476 const SDLoc &DL, EVT VT) {
1477 assert(!VT.bitsEq(Op.getValueType()) &&
1478 "Strict no-op FP extend/round not allowed.");
1479 SDValue Res =
1480 VT.bitsGT(Op.getValueType())
1481 ? getNode(ISD::STRICT_FP_EXTEND, DL, {VT, MVT::Other}, {Chain, Op})
1482 : getNode(ISD::STRICT_FP_ROUND, DL, {VT, MVT::Other},
1483 {Chain, Op, getIntPtrConstant(0, DL, /*isTarget=*/true)});
1484
1485 return std::pair<SDValue, SDValue>(Res, SDValue(Res.getNode(), 1));
1486}
1487
1488SDValue SelectionDAG::getAnyExtOrTrunc(SDValue Op, const SDLoc &DL, EVT VT) {
1489 return VT.bitsGT(VT: Op.getValueType()) ?
1490 getNode(Opcode: ISD::ANY_EXTEND, DL, VT, Operand: Op) :
1491 getNode(Opcode: ISD::TRUNCATE, DL, VT, Operand: Op);
1492}
1493
1494SDValue SelectionDAG::getSExtOrTrunc(SDValue Op, const SDLoc &DL, EVT VT) {
1495 return VT.bitsGT(VT: Op.getValueType()) ?
1496 getNode(Opcode: ISD::SIGN_EXTEND, DL, VT, Operand: Op) :
1497 getNode(Opcode: ISD::TRUNCATE, DL, VT, Operand: Op);
1498}
1499
1500SDValue SelectionDAG::getZExtOrTrunc(SDValue Op, const SDLoc &DL, EVT VT) {
1501 return VT.bitsGT(VT: Op.getValueType()) ?
1502 getNode(Opcode: ISD::ZERO_EXTEND, DL, VT, Operand: Op) :
1503 getNode(Opcode: ISD::TRUNCATE, DL, VT, Operand: Op);
1504}
1505
1506SDValue SelectionDAG::getBitcastedAnyExtOrTrunc(SDValue Op, const SDLoc &DL,
1507 EVT VT) {
1508 assert(!VT.isVector());
1509 auto Type = Op.getValueType();
1510 SDValue DestOp;
1511 if (Type == VT)
1512 return Op;
1513 auto Size = Op.getValueSizeInBits();
1514 DestOp = getBitcast(VT: EVT::getIntegerVT(Context&: *Context, BitWidth: Size), V: Op);
1515 if (DestOp.getValueType() == VT)
1516 return DestOp;
1517
1518 return getAnyExtOrTrunc(Op: DestOp, DL, VT);
1519}
1520
1521SDValue SelectionDAG::getBitcastedSExtOrTrunc(SDValue Op, const SDLoc &DL,
1522 EVT VT) {
1523 assert(!VT.isVector());
1524 auto Type = Op.getValueType();
1525 SDValue DestOp;
1526 if (Type == VT)
1527 return Op;
1528 auto Size = Op.getValueSizeInBits();
1529 DestOp = getBitcast(VT: MVT::getIntegerVT(BitWidth: Size), V: Op);
1530 if (DestOp.getValueType() == VT)
1531 return DestOp;
1532
1533 return getSExtOrTrunc(Op: DestOp, DL, VT);
1534}
1535
1536SDValue SelectionDAG::getBitcastedZExtOrTrunc(SDValue Op, const SDLoc &DL,
1537 EVT VT) {
1538 assert(!VT.isVector());
1539 auto Type = Op.getValueType();
1540 SDValue DestOp;
1541 if (Type == VT)
1542 return Op;
1543 auto Size = Op.getValueSizeInBits();
1544 DestOp = getBitcast(VT: MVT::getIntegerVT(BitWidth: Size), V: Op);
1545 if (DestOp.getValueType() == VT)
1546 return DestOp;
1547
1548 return getZExtOrTrunc(Op: DestOp, DL, VT);
1549}
1550
1551SDValue SelectionDAG::getBoolExtOrTrunc(SDValue Op, const SDLoc &SL, EVT VT,
1552 EVT OpVT) {
1553 if (VT.bitsLE(VT: Op.getValueType()))
1554 return getNode(Opcode: ISD::TRUNCATE, DL: SL, VT, Operand: Op);
1555
1556 TargetLowering::BooleanContent BType = TLI->getBooleanContents(Type: OpVT);
1557 return getNode(Opcode: TLI->getExtendForContent(Content: BType), DL: SL, VT, Operand: Op);
1558}
1559
1560SDValue SelectionDAG::getZeroExtendInReg(SDValue Op, const SDLoc &DL, EVT VT) {
1561 EVT OpVT = Op.getValueType();
1562 assert(VT.isInteger() && OpVT.isInteger() &&
1563 "Cannot getZeroExtendInReg FP types");
1564 assert(VT.isVector() == OpVT.isVector() &&
1565 "getZeroExtendInReg type should be vector iff the operand "
1566 "type is vector!");
1567 assert((!VT.isVector() ||
1568 VT.getVectorElementCount() == OpVT.getVectorElementCount()) &&
1569 "Vector element counts must match in getZeroExtendInReg");
1570 assert(VT.bitsLE(OpVT) && "Not extending!");
1571 if (OpVT == VT)
1572 return Op;
1573 APInt Imm = APInt::getLowBitsSet(numBits: OpVT.getScalarSizeInBits(),
1574 loBitsSet: VT.getScalarSizeInBits());
1575 return getNode(Opcode: ISD::AND, DL, VT: OpVT, N1: Op, N2: getConstant(Val: Imm, DL, VT: OpVT));
1576}
1577
1578SDValue SelectionDAG::getVPZeroExtendInReg(SDValue Op, SDValue Mask,
1579 SDValue EVL, const SDLoc &DL,
1580 EVT VT) {
1581 EVT OpVT = Op.getValueType();
1582 assert(VT.isInteger() && OpVT.isInteger() &&
1583 "Cannot getVPZeroExtendInReg FP types");
1584 assert(VT.isVector() && OpVT.isVector() &&
1585 "getVPZeroExtendInReg type and operand type should be vector!");
1586 assert(VT.getVectorElementCount() == OpVT.getVectorElementCount() &&
1587 "Vector element counts must match in getZeroExtendInReg");
1588 assert(VT.bitsLE(OpVT) && "Not extending!");
1589 if (OpVT == VT)
1590 return Op;
1591 APInt Imm = APInt::getLowBitsSet(numBits: OpVT.getScalarSizeInBits(),
1592 loBitsSet: VT.getScalarSizeInBits());
1593 return getNode(Opcode: ISD::VP_AND, DL, VT: OpVT, N1: Op, N2: getConstant(Val: Imm, DL, VT: OpVT), N3: Mask,
1594 N4: EVL);
1595}
1596
1597SDValue SelectionDAG::getPtrExtOrTrunc(SDValue Op, const SDLoc &DL, EVT VT) {
1598 // Only unsigned pointer semantics are supported right now. In the future this
1599 // might delegate to TLI to check pointer signedness.
1600 return getZExtOrTrunc(Op, DL, VT);
1601}
1602
1603SDValue SelectionDAG::getPtrExtendInReg(SDValue Op, const SDLoc &DL, EVT VT) {
1604 // Only unsigned pointer semantics are supported right now. In the future this
1605 // might delegate to TLI to check pointer signedness.
1606 return getZeroExtendInReg(Op, DL, VT);
1607}
1608
1609SDValue SelectionDAG::getNegative(SDValue Val, const SDLoc &DL, EVT VT) {
1610 return getNode(Opcode: ISD::SUB, DL, VT, N1: getConstant(Val: 0, DL, VT), N2: Val);
1611}
1612
1613/// getNOT - Create a bitwise NOT operation as (XOR Val, -1).
1614SDValue SelectionDAG::getNOT(const SDLoc &DL, SDValue Val, EVT VT) {
1615 return getNode(Opcode: ISD::XOR, DL, VT, N1: Val, N2: getAllOnesConstant(DL, VT));
1616}
1617
1618SDValue SelectionDAG::getLogicalNOT(const SDLoc &DL, SDValue Val, EVT VT) {
1619 SDValue TrueValue = getBoolConstant(V: true, DL, VT, OpVT: VT);
1620 return getNode(Opcode: ISD::XOR, DL, VT, N1: Val, N2: TrueValue);
1621}
1622
1623SDValue SelectionDAG::getVPLogicalNOT(const SDLoc &DL, SDValue Val,
1624 SDValue Mask, SDValue EVL, EVT VT) {
1625 SDValue TrueValue = getBoolConstant(V: true, DL, VT, OpVT: VT);
1626 return getNode(Opcode: ISD::VP_XOR, DL, VT, N1: Val, N2: TrueValue, N3: Mask, N4: EVL);
1627}
1628
1629SDValue SelectionDAG::getVPPtrExtOrTrunc(const SDLoc &DL, EVT VT, SDValue Op,
1630 SDValue Mask, SDValue EVL) {
1631 return getVPZExtOrTrunc(DL, VT, Op, Mask, EVL);
1632}
1633
1634SDValue SelectionDAG::getVPZExtOrTrunc(const SDLoc &DL, EVT VT, SDValue Op,
1635 SDValue Mask, SDValue EVL) {
1636 if (VT.bitsGT(VT: Op.getValueType()))
1637 return getNode(Opcode: ISD::VP_ZERO_EXTEND, DL, VT, N1: Op, N2: Mask, N3: EVL);
1638 if (VT.bitsLT(VT: Op.getValueType()))
1639 return getNode(Opcode: ISD::VP_TRUNCATE, DL, VT, N1: Op, N2: Mask, N3: EVL);
1640 return Op;
1641}
1642
1643SDValue SelectionDAG::getBoolConstant(bool V, const SDLoc &DL, EVT VT,
1644 EVT OpVT) {
1645 if (!V)
1646 return getConstant(Val: 0, DL, VT);
1647
1648 switch (TLI->getBooleanContents(Type: OpVT)) {
1649 case TargetLowering::ZeroOrOneBooleanContent:
1650 case TargetLowering::UndefinedBooleanContent:
1651 return getConstant(Val: 1, DL, VT);
1652 case TargetLowering::ZeroOrNegativeOneBooleanContent:
1653 return getAllOnesConstant(DL, VT);
1654 }
1655 llvm_unreachable("Unexpected boolean content enum!");
1656}
1657
1658SDValue SelectionDAG::getConstant(uint64_t Val, const SDLoc &DL, EVT VT,
1659 bool isT, bool isO) {
1660 return getConstant(Val: APInt(VT.getScalarSizeInBits(), Val, /*isSigned=*/false),
1661 DL, VT, isTarget: isT, isOpaque: isO);
1662}
1663
1664SDValue SelectionDAG::getConstant(const APInt &Val, const SDLoc &DL, EVT VT,
1665 bool isT, bool isO) {
1666 return getConstant(Val: *ConstantInt::get(Context&: *Context, V: Val), DL, VT, isTarget: isT, isOpaque: isO);
1667}
1668
1669SDValue SelectionDAG::getConstant(const ConstantInt &Val, const SDLoc &DL,
1670 EVT VT, bool isT, bool isO) {
1671 assert(VT.isInteger() && "Cannot create FP integer constant!");
1672
1673 EVT EltVT = VT.getScalarType();
1674 const ConstantInt *Elt = &Val;
1675
1676 // Vector splats are explicit within the DAG, with ConstantSDNode holding the
1677 // to-be-splatted scalar ConstantInt.
1678 if (isa<VectorType>(Val: Elt->getType()))
1679 Elt = ConstantInt::get(Context&: *getContext(), V: Elt->getValue());
1680
1681 // In some cases the vector type is legal but the element type is illegal and
1682 // needs to be promoted, for example v8i8 on ARM. In this case, promote the
1683 // inserted value (the type does not need to match the vector element type).
1684 // Any extra bits introduced will be truncated away.
1685 if (VT.isVector() && TLI->getTypeAction(Context&: *getContext(), VT: EltVT) ==
1686 TargetLowering::TypePromoteInteger) {
1687 EltVT = TLI->getTypeToTransformTo(Context&: *getContext(), VT: EltVT);
1688 APInt NewVal;
1689 if (TLI->isSExtCheaperThanZExt(FromTy: VT.getScalarType(), ToTy: EltVT))
1690 NewVal = Elt->getValue().sextOrTrunc(width: EltVT.getSizeInBits());
1691 else
1692 NewVal = Elt->getValue().zextOrTrunc(width: EltVT.getSizeInBits());
1693 Elt = ConstantInt::get(Context&: *getContext(), V: NewVal);
1694 }
1695 // In other cases the element type is illegal and needs to be expanded, for
1696 // example v2i64 on MIPS32. In this case, find the nearest legal type, split
1697 // the value into n parts and use a vector type with n-times the elements.
1698 // Then bitcast to the type requested.
1699 // Legalizing constants too early makes the DAGCombiner's job harder so we
1700 // only legalize if the DAG tells us we must produce legal types.
1701 else if (NewNodesMustHaveLegalTypes && VT.isVector() &&
1702 TLI->getTypeAction(Context&: *getContext(), VT: EltVT) ==
1703 TargetLowering::TypeExpandInteger) {
1704 const APInt &NewVal = Elt->getValue();
1705 EVT ViaEltVT = TLI->getTypeToTransformTo(Context&: *getContext(), VT: EltVT);
1706 unsigned ViaEltSizeInBits = ViaEltVT.getSizeInBits();
1707
1708 // For scalable vectors, try to use a SPLAT_VECTOR_PARTS node.
1709 if (VT.isScalableVector() ||
1710 TLI->isOperationLegal(Op: ISD::SPLAT_VECTOR, VT)) {
1711 assert(EltVT.getSizeInBits() % ViaEltSizeInBits == 0 &&
1712 "Can only handle an even split!");
1713 unsigned Parts = EltVT.getSizeInBits() / ViaEltSizeInBits;
1714
1715 SmallVector<SDValue, 2> ScalarParts;
1716 for (unsigned i = 0; i != Parts; ++i)
1717 ScalarParts.push_back(Elt: getConstant(
1718 Val: NewVal.extractBits(numBits: ViaEltSizeInBits, bitPosition: i * ViaEltSizeInBits), DL,
1719 VT: ViaEltVT, isT, isO));
1720
1721 return getNode(Opcode: ISD::SPLAT_VECTOR_PARTS, DL, VT, Ops: ScalarParts);
1722 }
1723
1724 unsigned ViaVecNumElts = VT.getSizeInBits() / ViaEltSizeInBits;
1725 EVT ViaVecVT = EVT::getVectorVT(Context&: *getContext(), VT: ViaEltVT, NumElements: ViaVecNumElts);
1726
1727 // Check the temporary vector is the correct size. If this fails then
1728 // getTypeToTransformTo() probably returned a type whose size (in bits)
1729 // isn't a power-of-2 factor of the requested type size.
1730 assert(ViaVecVT.getSizeInBits() == VT.getSizeInBits());
1731
1732 SmallVector<SDValue, 2> EltParts;
1733 for (unsigned i = 0; i < ViaVecNumElts / VT.getVectorNumElements(); ++i)
1734 EltParts.push_back(Elt: getConstant(
1735 Val: NewVal.extractBits(numBits: ViaEltSizeInBits, bitPosition: i * ViaEltSizeInBits), DL,
1736 VT: ViaEltVT, isT, isO));
1737
1738 // EltParts is currently in little endian order. If we actually want
1739 // big-endian order then reverse it now.
1740 if (getDataLayout().isBigEndian())
1741 std::reverse(first: EltParts.begin(), last: EltParts.end());
1742
1743 // The elements must be reversed when the element order is different
1744 // to the endianness of the elements (because the BITCAST is itself a
1745 // vector shuffle in this situation). However, we do not need any code to
1746 // perform this reversal because getConstant() is producing a vector
1747 // splat.
1748 // This situation occurs in MIPS MSA.
1749
1750 SmallVector<SDValue, 8> Ops;
1751 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
1752 llvm::append_range(C&: Ops, R&: EltParts);
1753
1754 SDValue V =
1755 getNode(Opcode: ISD::BITCAST, DL, VT, Operand: getBuildVector(VT: ViaVecVT, DL, Ops));
1756 return V;
1757 }
1758
1759 assert(Elt->getBitWidth() == EltVT.getSizeInBits() &&
1760 "APInt size does not match type size!");
1761 unsigned Opc = isT ? ISD::TargetConstant : ISD::Constant;
1762 SDVTList VTs = getVTList(VT: EltVT);
1763 FoldingSetNodeID ID;
1764 AddNodeIDNode(ID, OpC: Opc, VTList: VTs, OpList: {});
1765 ID.AddPointer(Ptr: Elt);
1766 ID.AddBoolean(B: isO);
1767 void *IP = nullptr;
1768 SDNode *N = nullptr;
1769 if ((N = FindNodeOrInsertPos(ID, DL, InsertPos&: IP)))
1770 if (!VT.isVector())
1771 return SDValue(N, 0);
1772
1773 if (!N) {
1774 N = newSDNode<ConstantSDNode>(Args&: isT, Args&: isO, Args&: Elt, Args&: VTs);
1775 CSEMap.InsertNode(N, InsertPos: IP);
1776 InsertNode(N);
1777 NewSDValueDbgMsg(V: SDValue(N, 0), Msg: "Creating constant: ", G: this);
1778 }
1779
1780 SDValue Result(N, 0);
1781 if (VT.isVector())
1782 Result = getSplat(VT, DL, Op: Result);
1783 return Result;
1784}
1785
1786SDValue SelectionDAG::getSignedConstant(int64_t Val, const SDLoc &DL, EVT VT,
1787 bool isT, bool isO) {
1788 unsigned Size = VT.getScalarSizeInBits();
1789 return getConstant(Val: APInt(Size, Val, /*isSigned=*/true), DL, VT, isT, isO);
1790}
1791
1792SDValue SelectionDAG::getAllOnesConstant(const SDLoc &DL, EVT VT, bool IsTarget,
1793 bool IsOpaque) {
1794 return getConstant(Val: APInt::getAllOnes(numBits: VT.getScalarSizeInBits()), DL, VT,
1795 isT: IsTarget, isO: IsOpaque);
1796}
1797
1798SDValue SelectionDAG::getIntPtrConstant(uint64_t Val, const SDLoc &DL,
1799 bool isTarget) {
1800 return getConstant(Val, DL, VT: TLI->getPointerTy(DL: getDataLayout()), isT: isTarget);
1801}
1802
1803SDValue SelectionDAG::getShiftAmountConstant(uint64_t Val, EVT VT,
1804 const SDLoc &DL) {
1805 assert(VT.isInteger() && "Shift amount is not an integer type!");
1806 EVT ShiftVT = TLI->getShiftAmountTy(LHSTy: VT, DL: getDataLayout());
1807 return getConstant(Val, DL, VT: ShiftVT);
1808}
1809
1810SDValue SelectionDAG::getShiftAmountConstant(const APInt &Val, EVT VT,
1811 const SDLoc &DL) {
1812 assert(Val.ult(VT.getScalarSizeInBits()) && "Out of range shift");
1813 return getShiftAmountConstant(Val: Val.getZExtValue(), VT, DL);
1814}
1815
1816SDValue SelectionDAG::getVectorIdxConstant(uint64_t Val, const SDLoc &DL,
1817 bool isTarget) {
1818 return getConstant(Val, DL, VT: TLI->getVectorIdxTy(DL: getDataLayout()), isT: isTarget);
1819}
1820
1821SDValue SelectionDAG::getConstantFP(const APFloat &V, const SDLoc &DL, EVT VT,
1822 bool isTarget) {
1823 return getConstantFP(V: *ConstantFP::get(Context&: *getContext(), V), DL, VT, isTarget);
1824}
1825
1826SDValue SelectionDAG::getConstantFP(const ConstantFP &V, const SDLoc &DL,
1827 EVT VT, bool isTarget) {
1828 assert(VT.isFloatingPoint() && "Cannot create integer FP constant!");
1829
1830 EVT EltVT = VT.getScalarType();
1831 const ConstantFP *Elt = &V;
1832
1833 // Vector splats are explicit within the DAG, with ConstantFPSDNode holding
1834 // the to-be-splatted scalar ConstantFP.
1835 if (isa<VectorType>(Val: Elt->getType()))
1836 Elt = ConstantFP::get(Context&: *getContext(), V: Elt->getValue());
1837
1838 // Do the map lookup using the actual bit pattern for the floating point
1839 // value, so that we don't have problems with 0.0 comparing equal to -0.0, and
1840 // we don't have issues with SNANs.
1841 unsigned Opc = isTarget ? ISD::TargetConstantFP : ISD::ConstantFP;
1842 SDVTList VTs = getVTList(VT: EltVT);
1843 FoldingSetNodeID ID;
1844 AddNodeIDNode(ID, OpC: Opc, VTList: VTs, OpList: {});
1845 ID.AddPointer(Ptr: Elt);
1846 void *IP = nullptr;
1847 SDNode *N = nullptr;
1848 if ((N = FindNodeOrInsertPos(ID, DL, InsertPos&: IP)))
1849 if (!VT.isVector())
1850 return SDValue(N, 0);
1851
1852 if (!N) {
1853 N = newSDNode<ConstantFPSDNode>(Args&: isTarget, Args&: Elt, Args&: VTs);
1854 CSEMap.InsertNode(N, InsertPos: IP);
1855 InsertNode(N);
1856 }
1857
1858 SDValue Result(N, 0);
1859 if (VT.isVector())
1860 Result = getSplat(VT, DL, Op: Result);
1861 NewSDValueDbgMsg(V: Result, Msg: "Creating fp constant: ", G: this);
1862 return Result;
1863}
1864
1865SDValue SelectionDAG::getConstantFP(double Val, const SDLoc &DL, EVT VT,
1866 bool isTarget) {
1867 EVT EltVT = VT.getScalarType();
1868 if (EltVT == MVT::f32)
1869 return getConstantFP(V: APFloat((float)Val), DL, VT, isTarget);
1870 if (EltVT == MVT::f64)
1871 return getConstantFP(V: APFloat(Val), DL, VT, isTarget);
1872 if (EltVT == MVT::f80 || EltVT == MVT::f128 || EltVT == MVT::ppcf128 ||
1873 EltVT == MVT::f16 || EltVT == MVT::bf16) {
1874 bool Ignored;
1875 APFloat APF = APFloat(Val);
1876 APF.convert(ToSemantics: EltVT.getFltSemantics(), RM: APFloat::rmNearestTiesToEven,
1877 losesInfo: &Ignored);
1878 return getConstantFP(V: APF, DL, VT, isTarget);
1879 }
1880 llvm_unreachable("Unsupported type in getConstantFP");
1881}
1882
1883SDValue SelectionDAG::getGlobalAddress(const GlobalValue *GV, const SDLoc &DL,
1884 EVT VT, int64_t Offset, bool isTargetGA,
1885 unsigned TargetFlags) {
1886 assert((TargetFlags == 0 || isTargetGA) &&
1887 "Cannot set target flags on target-independent globals");
1888
1889 // Truncate (with sign-extension) the offset value to the pointer size.
1890 unsigned BitWidth = getDataLayout().getPointerTypeSizeInBits(GV->getType());
1891 if (BitWidth < 64)
1892 Offset = SignExtend64(X: Offset, B: BitWidth);
1893
1894 unsigned Opc;
1895 if (GV->isThreadLocal())
1896 Opc = isTargetGA ? ISD::TargetGlobalTLSAddress : ISD::GlobalTLSAddress;
1897 else
1898 Opc = isTargetGA ? ISD::TargetGlobalAddress : ISD::GlobalAddress;
1899
1900 SDVTList VTs = getVTList(VT);
1901 FoldingSetNodeID ID;
1902 AddNodeIDNode(ID, OpC: Opc, VTList: VTs, OpList: {});
1903 ID.AddPointer(Ptr: GV);
1904 ID.AddInteger(I: Offset);
1905 ID.AddInteger(I: TargetFlags);
1906 void *IP = nullptr;
1907 if (SDNode *E = FindNodeOrInsertPos(ID, DL, InsertPos&: IP))
1908 return SDValue(E, 0);
1909
1910 auto *N = newSDNode<GlobalAddressSDNode>(
1911 Args&: Opc, Args: DL.getIROrder(), Args: DL.getDebugLoc(), Args&: GV, Args&: VTs, Args&: Offset, Args&: TargetFlags);
1912 CSEMap.InsertNode(N, InsertPos: IP);
1913 InsertNode(N);
1914 return SDValue(N, 0);
1915}
1916
1917SDValue SelectionDAG::getFrameIndex(int FI, EVT VT, bool isTarget) {
1918 unsigned Opc = isTarget ? ISD::TargetFrameIndex : ISD::FrameIndex;
1919 SDVTList VTs = getVTList(VT);
1920 FoldingSetNodeID ID;
1921 AddNodeIDNode(ID, OpC: Opc, VTList: VTs, OpList: {});
1922 ID.AddInteger(I: FI);
1923 void *IP = nullptr;
1924 if (SDNode *E = FindNodeOrInsertPos(ID, InsertPos&: IP))
1925 return SDValue(E, 0);
1926
1927 auto *N = newSDNode<FrameIndexSDNode>(Args&: FI, Args&: VTs, Args&: isTarget);
1928 CSEMap.InsertNode(N, InsertPos: IP);
1929 InsertNode(N);
1930 return SDValue(N, 0);
1931}
1932
1933SDValue SelectionDAG::getJumpTable(int JTI, EVT VT, bool isTarget,
1934 unsigned TargetFlags) {
1935 assert((TargetFlags == 0 || isTarget) &&
1936 "Cannot set target flags on target-independent jump tables");
1937 unsigned Opc = isTarget ? ISD::TargetJumpTable : ISD::JumpTable;
1938 SDVTList VTs = getVTList(VT);
1939 FoldingSetNodeID ID;
1940 AddNodeIDNode(ID, OpC: Opc, VTList: VTs, OpList: {});
1941 ID.AddInteger(I: JTI);
1942 ID.AddInteger(I: TargetFlags);
1943 void *IP = nullptr;
1944 if (SDNode *E = FindNodeOrInsertPos(ID, InsertPos&: IP))
1945 return SDValue(E, 0);
1946
1947 auto *N = newSDNode<JumpTableSDNode>(Args&: JTI, Args&: VTs, Args&: isTarget, Args&: TargetFlags);
1948 CSEMap.InsertNode(N, InsertPos: IP);
1949 InsertNode(N);
1950 return SDValue(N, 0);
1951}
1952
1953SDValue SelectionDAG::getJumpTableDebugInfo(int JTI, SDValue Chain,
1954 const SDLoc &DL) {
1955 EVT PTy = getTargetLoweringInfo().getPointerTy(DL: getDataLayout());
1956 return getNode(ISD::JUMP_TABLE_DEBUG_INFO, DL, MVT::Glue, Chain,
1957 getTargetConstant(Val: static_cast<uint64_t>(JTI), DL, VT: PTy, isOpaque: true));
1958}
1959
1960SDValue SelectionDAG::getConstantPool(const Constant *C, EVT VT,
1961 MaybeAlign Alignment, int Offset,
1962 bool isTarget, unsigned TargetFlags) {
1963 assert((TargetFlags == 0 || isTarget) &&
1964 "Cannot set target flags on target-independent globals");
1965 if (!Alignment)
1966 Alignment = shouldOptForSize()
1967 ? getDataLayout().getABITypeAlign(Ty: C->getType())
1968 : getDataLayout().getPrefTypeAlign(Ty: C->getType());
1969 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool;
1970 SDVTList VTs = getVTList(VT);
1971 FoldingSetNodeID ID;
1972 AddNodeIDNode(ID, OpC: Opc, VTList: VTs, OpList: {});
1973 ID.AddInteger(I: Alignment->value());
1974 ID.AddInteger(I: Offset);
1975 ID.AddPointer(Ptr: C);
1976 ID.AddInteger(I: TargetFlags);
1977 void *IP = nullptr;
1978 if (SDNode *E = FindNodeOrInsertPos(ID, InsertPos&: IP))
1979 return SDValue(E, 0);
1980
1981 auto *N = newSDNode<ConstantPoolSDNode>(Args&: isTarget, Args&: C, Args&: VTs, Args&: Offset, Args&: *Alignment,
1982 Args&: TargetFlags);
1983 CSEMap.InsertNode(N, InsertPos: IP);
1984 InsertNode(N);
1985 SDValue V = SDValue(N, 0);
1986 NewSDValueDbgMsg(V, Msg: "Creating new constant pool: ", G: this);
1987 return V;
1988}
1989
1990SDValue SelectionDAG::getConstantPool(MachineConstantPoolValue *C, EVT VT,
1991 MaybeAlign Alignment, int Offset,
1992 bool isTarget, unsigned TargetFlags) {
1993 assert((TargetFlags == 0 || isTarget) &&
1994 "Cannot set target flags on target-independent globals");
1995 if (!Alignment)
1996 Alignment = getDataLayout().getPrefTypeAlign(Ty: C->getType());
1997 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool;
1998 SDVTList VTs = getVTList(VT);
1999 FoldingSetNodeID ID;
2000 AddNodeIDNode(ID, OpC: Opc, VTList: VTs, OpList: {});
2001 ID.AddInteger(I: Alignment->value());
2002 ID.AddInteger(I: Offset);
2003 C->addSelectionDAGCSEId(ID);
2004 ID.AddInteger(I: TargetFlags);
2005 void *IP = nullptr;
2006 if (SDNode *E = FindNodeOrInsertPos(ID, InsertPos&: IP))
2007 return SDValue(E, 0);
2008
2009 auto *N = newSDNode<ConstantPoolSDNode>(Args&: isTarget, Args&: C, Args&: VTs, Args&: Offset, Args&: *Alignment,
2010 Args&: TargetFlags);
2011 CSEMap.InsertNode(N, InsertPos: IP);
2012 InsertNode(N);
2013 return SDValue(N, 0);
2014}
2015
2016SDValue SelectionDAG::getBasicBlock(MachineBasicBlock *MBB) {
2017 FoldingSetNodeID ID;
2018 AddNodeIDNode(ID, ISD::BasicBlock, getVTList(MVT::Other), {});
2019 ID.AddPointer(Ptr: MBB);
2020 void *IP = nullptr;
2021 if (SDNode *E = FindNodeOrInsertPos(ID, InsertPos&: IP))
2022 return SDValue(E, 0);
2023
2024 auto *N = newSDNode<BasicBlockSDNode>(Args&: MBB);
2025 CSEMap.InsertNode(N, InsertPos: IP);
2026 InsertNode(N);
2027 return SDValue(N, 0);
2028}
2029
2030SDValue SelectionDAG::getValueType(EVT VT) {
2031 if (VT.isSimple() && (unsigned)VT.getSimpleVT().SimpleTy >=
2032 ValueTypeNodes.size())
2033 ValueTypeNodes.resize(new_size: VT.getSimpleVT().SimpleTy+1);
2034
2035 SDNode *&N = VT.isExtended() ?
2036 ExtendedValueTypeNodes[VT] : ValueTypeNodes[VT.getSimpleVT().SimpleTy];
2037
2038 if (N) return SDValue(N, 0);
2039 N = newSDNode<VTSDNode>(Args&: VT);
2040 InsertNode(N);
2041 return SDValue(N, 0);
2042}
2043
2044SDValue SelectionDAG::getExternalSymbol(const char *Sym, EVT VT) {
2045 SDNode *&N = ExternalSymbols[Sym];
2046 if (N) return SDValue(N, 0);
2047 N = newSDNode<ExternalSymbolSDNode>(Args: false, Args&: Sym, Args: 0, Args: getVTList(VT));
2048 InsertNode(N);
2049 return SDValue(N, 0);
2050}
2051
2052SDValue SelectionDAG::getMCSymbol(MCSymbol *Sym, EVT VT) {
2053 SDNode *&N = MCSymbols[Sym];
2054 if (N)
2055 return SDValue(N, 0);
2056 N = newSDNode<MCSymbolSDNode>(Args&: Sym, Args: getVTList(VT));
2057 InsertNode(N);
2058 return SDValue(N, 0);
2059}
2060
2061SDValue SelectionDAG::getTargetExternalSymbol(const char *Sym, EVT VT,
2062 unsigned TargetFlags) {
2063 SDNode *&N =
2064 TargetExternalSymbols[std::pair<std::string, unsigned>(Sym, TargetFlags)];
2065 if (N) return SDValue(N, 0);
2066 N = newSDNode<ExternalSymbolSDNode>(Args: true, Args&: Sym, Args&: TargetFlags, Args: getVTList(VT));
2067 InsertNode(N);
2068 return SDValue(N, 0);
2069}
2070
2071SDValue SelectionDAG::getCondCode(ISD::CondCode Cond) {
2072 if ((unsigned)Cond >= CondCodeNodes.size())
2073 CondCodeNodes.resize(new_size: Cond+1);
2074
2075 if (!CondCodeNodes[Cond]) {
2076 auto *N = newSDNode<CondCodeSDNode>(Args&: Cond);
2077 CondCodeNodes[Cond] = N;
2078 InsertNode(N);
2079 }
2080
2081 return SDValue(CondCodeNodes[Cond], 0);
2082}
2083
2084SDValue SelectionDAG::getVScale(const SDLoc &DL, EVT VT, APInt MulImm,
2085 bool ConstantFold) {
2086 assert(MulImm.getBitWidth() == VT.getSizeInBits() &&
2087 "APInt size does not match type size!");
2088
2089 if (MulImm == 0)
2090 return getConstant(Val: 0, DL, VT);
2091
2092 if (ConstantFold) {
2093 const MachineFunction &MF = getMachineFunction();
2094 const Function &F = MF.getFunction();
2095 ConstantRange CR = getVScaleRange(F: &F, BitWidth: 64);
2096 if (const APInt *C = CR.getSingleElement())
2097 return getConstant(Val: MulImm * C->getZExtValue(), DL, VT);
2098 }
2099
2100 return getNode(Opcode: ISD::VSCALE, DL, VT, Operand: getConstant(Val: MulImm, DL, VT));
2101}
2102
2103SDValue SelectionDAG::getElementCount(const SDLoc &DL, EVT VT, ElementCount EC,
2104 bool ConstantFold) {
2105 if (EC.isScalable())
2106 return getVScale(DL, VT,
2107 MulImm: APInt(VT.getSizeInBits(), EC.getKnownMinValue()));
2108
2109 return getConstant(Val: EC.getKnownMinValue(), DL, VT);
2110}
2111
2112SDValue SelectionDAG::getStepVector(const SDLoc &DL, EVT ResVT) {
2113 APInt One(ResVT.getScalarSizeInBits(), 1);
2114 return getStepVector(DL, ResVT, StepVal: One);
2115}
2116
2117SDValue SelectionDAG::getStepVector(const SDLoc &DL, EVT ResVT,
2118 const APInt &StepVal) {
2119 assert(ResVT.getScalarSizeInBits() == StepVal.getBitWidth());
2120 if (ResVT.isScalableVector())
2121 return getNode(
2122 Opcode: ISD::STEP_VECTOR, DL, VT: ResVT,
2123 Operand: getTargetConstant(Val: StepVal, DL, VT: ResVT.getVectorElementType()));
2124
2125 SmallVector<SDValue, 16> OpsStepConstants;
2126 for (uint64_t i = 0; i < ResVT.getVectorNumElements(); i++)
2127 OpsStepConstants.push_back(
2128 Elt: getConstant(Val: StepVal * i, DL, VT: ResVT.getVectorElementType()));
2129 return getBuildVector(VT: ResVT, DL, Ops: OpsStepConstants);
2130}
2131
2132/// Swaps the values of N1 and N2. Swaps all indices in the shuffle mask M that
2133/// point at N1 to point at N2 and indices that point at N2 to point at N1.
2134static void commuteShuffle(SDValue &N1, SDValue &N2, MutableArrayRef<int> M) {
2135 std::swap(a&: N1, b&: N2);
2136 ShuffleVectorSDNode::commuteMask(Mask: M);
2137}
2138
2139SDValue SelectionDAG::getVectorShuffle(EVT VT, const SDLoc &dl, SDValue N1,
2140 SDValue N2, ArrayRef<int> Mask) {
2141 assert(VT.getVectorNumElements() == Mask.size() &&
2142 "Must have the same number of vector elements as mask elements!");
2143 assert(VT == N1.getValueType() && VT == N2.getValueType() &&
2144 "Invalid VECTOR_SHUFFLE");
2145
2146 // Canonicalize shuffle undef, undef -> undef
2147 if (N1.isUndef() && N2.isUndef())
2148 return getUNDEF(VT);
2149
2150 // Validate that all indices in Mask are within the range of the elements
2151 // input to the shuffle.
2152 int NElts = Mask.size();
2153 assert(llvm::all_of(Mask,
2154 [&](int M) { return M < (NElts * 2) && M >= -1; }) &&
2155 "Index out of range");
2156
2157 // Copy the mask so we can do any needed cleanup.
2158 SmallVector<int, 8> MaskVec(Mask);
2159
2160 // Canonicalize shuffle v, v -> v, undef
2161 if (N1 == N2) {
2162 N2 = getUNDEF(VT);
2163 for (int i = 0; i != NElts; ++i)
2164 if (MaskVec[i] >= NElts) MaskVec[i] -= NElts;
2165 }
2166
2167 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
2168 if (N1.isUndef())
2169 commuteShuffle(N1, N2, M: MaskVec);
2170
2171 if (TLI->hasVectorBlend()) {
2172 // If shuffling a splat, try to blend the splat instead. We do this here so
2173 // that even when this arises during lowering we don't have to re-handle it.
2174 auto BlendSplat = [&](BuildVectorSDNode *BV, int Offset) {
2175 BitVector UndefElements;
2176 SDValue Splat = BV->getSplatValue(UndefElements: &UndefElements);
2177 if (!Splat)
2178 return;
2179
2180 for (int i = 0; i < NElts; ++i) {
2181 if (MaskVec[i] < Offset || MaskVec[i] >= (Offset + NElts))
2182 continue;
2183
2184 // If this input comes from undef, mark it as such.
2185 if (UndefElements[MaskVec[i] - Offset]) {
2186 MaskVec[i] = -1;
2187 continue;
2188 }
2189
2190 // If we can blend a non-undef lane, use that instead.
2191 if (!UndefElements[i])
2192 MaskVec[i] = i + Offset;
2193 }
2194 };
2195 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(Val&: N1))
2196 BlendSplat(N1BV, 0);
2197 if (auto *N2BV = dyn_cast<BuildVectorSDNode>(Val&: N2))
2198 BlendSplat(N2BV, NElts);
2199 }
2200
2201 // Canonicalize all index into lhs, -> shuffle lhs, undef
2202 // Canonicalize all index into rhs, -> shuffle rhs, undef
2203 bool AllLHS = true, AllRHS = true;
2204 bool N2Undef = N2.isUndef();
2205 for (int i = 0; i != NElts; ++i) {
2206 if (MaskVec[i] >= NElts) {
2207 if (N2Undef)
2208 MaskVec[i] = -1;
2209 else
2210 AllLHS = false;
2211 } else if (MaskVec[i] >= 0) {
2212 AllRHS = false;
2213 }
2214 }
2215 if (AllLHS && AllRHS)
2216 return getUNDEF(VT);
2217 if (AllLHS && !N2Undef)
2218 N2 = getUNDEF(VT);
2219 if (AllRHS) {
2220 N1 = getUNDEF(VT);
2221 commuteShuffle(N1, N2, M: MaskVec);
2222 }
2223 // Reset our undef status after accounting for the mask.
2224 N2Undef = N2.isUndef();
2225 // Re-check whether both sides ended up undef.
2226 if (N1.isUndef() && N2Undef)
2227 return getUNDEF(VT);
2228
2229 // If Identity shuffle return that node.
2230 bool Identity = true, AllSame = true;
2231 for (int i = 0; i != NElts; ++i) {
2232 if (MaskVec[i] >= 0 && MaskVec[i] != i) Identity = false;
2233 if (MaskVec[i] != MaskVec[0]) AllSame = false;
2234 }
2235 if (Identity && NElts)
2236 return N1;
2237
2238 // Shuffling a constant splat doesn't change the result.
2239 if (N2Undef) {
2240 SDValue V = N1;
2241
2242 // Look through any bitcasts. We check that these don't change the number
2243 // (and size) of elements and just changes their types.
2244 while (V.getOpcode() == ISD::BITCAST)
2245 V = V->getOperand(Num: 0);
2246
2247 // A splat should always show up as a build vector node.
2248 if (auto *BV = dyn_cast<BuildVectorSDNode>(Val&: V)) {
2249 BitVector UndefElements;
2250 SDValue Splat = BV->getSplatValue(UndefElements: &UndefElements);
2251 // If this is a splat of an undef, shuffling it is also undef.
2252 if (Splat && Splat.isUndef())
2253 return getUNDEF(VT);
2254
2255 bool SameNumElts =
2256 V.getValueType().getVectorNumElements() == VT.getVectorNumElements();
2257
2258 // We only have a splat which can skip shuffles if there is a splatted
2259 // value and no undef lanes rearranged by the shuffle.
2260 if (Splat && UndefElements.none()) {
2261 // Splat of <x, x, ..., x>, return <x, x, ..., x>, provided that the
2262 // number of elements match or the value splatted is a zero constant.
2263 if (SameNumElts || isNullConstant(V: Splat))
2264 return N1;
2265 }
2266
2267 // If the shuffle itself creates a splat, build the vector directly.
2268 if (AllSame && SameNumElts) {
2269 EVT BuildVT = BV->getValueType(ResNo: 0);
2270 const SDValue &Splatted = BV->getOperand(Num: MaskVec[0]);
2271 SDValue NewBV = getSplatBuildVector(VT: BuildVT, DL: dl, Op: Splatted);
2272
2273 // We may have jumped through bitcasts, so the type of the
2274 // BUILD_VECTOR may not match the type of the shuffle.
2275 if (BuildVT != VT)
2276 NewBV = getNode(Opcode: ISD::BITCAST, DL: dl, VT, Operand: NewBV);
2277 return NewBV;
2278 }
2279 }
2280 }
2281
2282 SDVTList VTs = getVTList(VT);
2283 FoldingSetNodeID ID;
2284 SDValue Ops[2] = { N1, N2 };
2285 AddNodeIDNode(ID, OpC: ISD::VECTOR_SHUFFLE, VTList: VTs, OpList: Ops);
2286 for (int i = 0; i != NElts; ++i)
2287 ID.AddInteger(I: MaskVec[i]);
2288
2289 void* IP = nullptr;
2290 if (SDNode *E = FindNodeOrInsertPos(ID, DL: dl, InsertPos&: IP))
2291 return SDValue(E, 0);
2292
2293 // Allocate the mask array for the node out of the BumpPtrAllocator, since
2294 // SDNode doesn't have access to it. This memory will be "leaked" when
2295 // the node is deallocated, but recovered when the NodeAllocator is released.
2296 int *MaskAlloc = OperandAllocator.Allocate<int>(Num: NElts);
2297 llvm::copy(Range&: MaskVec, Out: MaskAlloc);
2298
2299 auto *N = newSDNode<ShuffleVectorSDNode>(Args&: VTs, Args: dl.getIROrder(),
2300 Args: dl.getDebugLoc(), Args&: MaskAlloc);
2301 createOperands(Node: N, Vals: Ops);
2302
2303 CSEMap.InsertNode(N, InsertPos: IP);
2304 InsertNode(N);
2305 SDValue V = SDValue(N, 0);
2306 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
2307 return V;
2308}
2309
2310SDValue SelectionDAG::getCommutedVectorShuffle(const ShuffleVectorSDNode &SV) {
2311 EVT VT = SV.getValueType(ResNo: 0);
2312 SmallVector<int, 8> MaskVec(SV.getMask());
2313 ShuffleVectorSDNode::commuteMask(Mask: MaskVec);
2314
2315 SDValue Op0 = SV.getOperand(Num: 0);
2316 SDValue Op1 = SV.getOperand(Num: 1);
2317 return getVectorShuffle(VT, dl: SDLoc(&SV), N1: Op1, N2: Op0, Mask: MaskVec);
2318}
2319
2320SDValue SelectionDAG::getRegister(Register Reg, EVT VT) {
2321 SDVTList VTs = getVTList(VT);
2322 FoldingSetNodeID ID;
2323 AddNodeIDNode(ID, OpC: ISD::Register, VTList: VTs, OpList: {});
2324 ID.AddInteger(I: Reg.id());
2325 void *IP = nullptr;
2326 if (SDNode *E = FindNodeOrInsertPos(ID, InsertPos&: IP))
2327 return SDValue(E, 0);
2328
2329 auto *N = newSDNode<RegisterSDNode>(Args&: Reg, Args&: VTs);
2330 N->SDNodeBits.IsDivergent = TLI->isSDNodeSourceOfDivergence(N, FLI, UA);
2331 CSEMap.InsertNode(N, InsertPos: IP);
2332 InsertNode(N);
2333 return SDValue(N, 0);
2334}
2335
2336SDValue SelectionDAG::getRegisterMask(const uint32_t *RegMask) {
2337 FoldingSetNodeID ID;
2338 AddNodeIDNode(ID, ISD::RegisterMask, getVTList(MVT::Untyped), {});
2339 ID.AddPointer(Ptr: RegMask);
2340 void *IP = nullptr;
2341 if (SDNode *E = FindNodeOrInsertPos(ID, InsertPos&: IP))
2342 return SDValue(E, 0);
2343
2344 auto *N = newSDNode<RegisterMaskSDNode>(Args&: RegMask);
2345 CSEMap.InsertNode(N, InsertPos: IP);
2346 InsertNode(N);
2347 return SDValue(N, 0);
2348}
2349
2350SDValue SelectionDAG::getEHLabel(const SDLoc &dl, SDValue Root,
2351 MCSymbol *Label) {
2352 return getLabelNode(Opcode: ISD::EH_LABEL, dl, Root, Label);
2353}
2354
2355SDValue SelectionDAG::getLabelNode(unsigned Opcode, const SDLoc &dl,
2356 SDValue Root, MCSymbol *Label) {
2357 FoldingSetNodeID ID;
2358 SDValue Ops[] = { Root };
2359 AddNodeIDNode(ID, Opcode, getVTList(MVT::Other), Ops);
2360 ID.AddPointer(Ptr: Label);
2361 void *IP = nullptr;
2362 if (SDNode *E = FindNodeOrInsertPos(ID, InsertPos&: IP))
2363 return SDValue(E, 0);
2364
2365 auto *N =
2366 newSDNode<LabelSDNode>(Args&: Opcode, Args: dl.getIROrder(), Args: dl.getDebugLoc(), Args&: Label);
2367 createOperands(Node: N, Vals: Ops);
2368
2369 CSEMap.InsertNode(N, InsertPos: IP);
2370 InsertNode(N);
2371 return SDValue(N, 0);
2372}
2373
2374SDValue SelectionDAG::getBlockAddress(const BlockAddress *BA, EVT VT,
2375 int64_t Offset, bool isTarget,
2376 unsigned TargetFlags) {
2377 unsigned Opc = isTarget ? ISD::TargetBlockAddress : ISD::BlockAddress;
2378 SDVTList VTs = getVTList(VT);
2379
2380 FoldingSetNodeID ID;
2381 AddNodeIDNode(ID, OpC: Opc, VTList: VTs, OpList: {});
2382 ID.AddPointer(Ptr: BA);
2383 ID.AddInteger(I: Offset);
2384 ID.AddInteger(I: TargetFlags);
2385 void *IP = nullptr;
2386 if (SDNode *E = FindNodeOrInsertPos(ID, InsertPos&: IP))
2387 return SDValue(E, 0);
2388
2389 auto *N = newSDNode<BlockAddressSDNode>(Args&: Opc, Args&: VTs, Args&: BA, Args&: Offset, Args&: TargetFlags);
2390 CSEMap.InsertNode(N, InsertPos: IP);
2391 InsertNode(N);
2392 return SDValue(N, 0);
2393}
2394
2395SDValue SelectionDAG::getSrcValue(const Value *V) {
2396 FoldingSetNodeID ID;
2397 AddNodeIDNode(ID, ISD::SRCVALUE, getVTList(MVT::Other), {});
2398 ID.AddPointer(Ptr: V);
2399
2400 void *IP = nullptr;
2401 if (SDNode *E = FindNodeOrInsertPos(ID, InsertPos&: IP))
2402 return SDValue(E, 0);
2403
2404 auto *N = newSDNode<SrcValueSDNode>(Args&: V);
2405 CSEMap.InsertNode(N, InsertPos: IP);
2406 InsertNode(N);
2407 return SDValue(N, 0);
2408}
2409
2410SDValue SelectionDAG::getMDNode(const MDNode *MD) {
2411 FoldingSetNodeID ID;
2412 AddNodeIDNode(ID, ISD::MDNODE_SDNODE, getVTList(MVT::Other), {});
2413 ID.AddPointer(Ptr: MD);
2414
2415 void *IP = nullptr;
2416 if (SDNode *E = FindNodeOrInsertPos(ID, InsertPos&: IP))
2417 return SDValue(E, 0);
2418
2419 auto *N = newSDNode<MDNodeSDNode>(Args&: MD);
2420 CSEMap.InsertNode(N, InsertPos: IP);
2421 InsertNode(N);
2422 return SDValue(N, 0);
2423}
2424
2425SDValue SelectionDAG::getBitcast(EVT VT, SDValue V) {
2426 if (VT == V.getValueType())
2427 return V;
2428
2429 return getNode(Opcode: ISD::BITCAST, DL: SDLoc(V), VT, Operand: V);
2430}
2431
2432SDValue SelectionDAG::getAddrSpaceCast(const SDLoc &dl, EVT VT, SDValue Ptr,
2433 unsigned SrcAS, unsigned DestAS) {
2434 SDVTList VTs = getVTList(VT);
2435 SDValue Ops[] = {Ptr};
2436 FoldingSetNodeID ID;
2437 AddNodeIDNode(ID, OpC: ISD::ADDRSPACECAST, VTList: VTs, OpList: Ops);
2438 ID.AddInteger(I: SrcAS);
2439 ID.AddInteger(I: DestAS);
2440
2441 void *IP = nullptr;
2442 if (SDNode *E = FindNodeOrInsertPos(ID, DL: dl, InsertPos&: IP))
2443 return SDValue(E, 0);
2444
2445 auto *N = newSDNode<AddrSpaceCastSDNode>(Args: dl.getIROrder(), Args: dl.getDebugLoc(),
2446 Args&: VTs, Args&: SrcAS, Args&: DestAS);
2447 createOperands(Node: N, Vals: Ops);
2448
2449 CSEMap.InsertNode(N, InsertPos: IP);
2450 InsertNode(N);
2451 return SDValue(N, 0);
2452}
2453
2454SDValue SelectionDAG::getFreeze(SDValue V) {
2455 return getNode(Opcode: ISD::FREEZE, DL: SDLoc(V), VT: V.getValueType(), Operand: V);
2456}
2457
2458/// getShiftAmountOperand - Return the specified value casted to
2459/// the target's desired shift amount type.
2460SDValue SelectionDAG::getShiftAmountOperand(EVT LHSTy, SDValue Op) {
2461 EVT OpTy = Op.getValueType();
2462 EVT ShTy = TLI->getShiftAmountTy(LHSTy, DL: getDataLayout());
2463 if (OpTy == ShTy || OpTy.isVector()) return Op;
2464
2465 return getZExtOrTrunc(Op, DL: SDLoc(Op), VT: ShTy);
2466}
2467
2468/// Given a store node \p StoreNode, return true if it is safe to fold that node
2469/// into \p FPNode, which expands to a library call with output pointers.
2470static bool canFoldStoreIntoLibCallOutputPointers(StoreSDNode *StoreNode,
2471 SDNode *FPNode) {
2472 SmallVector<const SDNode *, 8> Worklist;
2473 SmallVector<const SDNode *, 8> DeferredNodes;
2474 SmallPtrSet<const SDNode *, 16> Visited;
2475
2476 // Skip FPNode use by StoreNode (that's the use we want to fold into FPNode).
2477 for (SDValue Op : StoreNode->ops())
2478 if (Op.getNode() != FPNode)
2479 Worklist.push_back(Elt: Op.getNode());
2480
2481 unsigned MaxSteps = SelectionDAG::getHasPredecessorMaxSteps();
2482 while (!Worklist.empty()) {
2483 const SDNode *Node = Worklist.pop_back_val();
2484 auto [_, Inserted] = Visited.insert(Ptr: Node);
2485 if (!Inserted)
2486 continue;
2487
2488 if (MaxSteps > 0 && Visited.size() >= MaxSteps)
2489 return false;
2490
2491 // Reached the FPNode (would result in a cycle).
2492 // OR Reached CALLSEQ_START (would result in nested call sequences).
2493 if (Node == FPNode || Node->getOpcode() == ISD::CALLSEQ_START)
2494 return false;
2495
2496 if (Node->getOpcode() == ISD::CALLSEQ_END) {
2497 // Defer looking into call sequences (so we can check we're outside one).
2498 // We still need to look through these for the predecessor check.
2499 DeferredNodes.push_back(Elt: Node);
2500 continue;
2501 }
2502
2503 for (SDValue Op : Node->ops())
2504 Worklist.push_back(Elt: Op.getNode());
2505 }
2506
2507 // True if we're outside a call sequence and don't have the FPNode as a
2508 // predecessor. No cycles or nested call sequences possible.
2509 return !SDNode::hasPredecessorHelper(N: FPNode, Visited, Worklist&: DeferredNodes,
2510 MaxSteps);
2511}
2512
2513bool SelectionDAG::expandMultipleResultFPLibCall(
2514 RTLIB::Libcall LC, SDNode *Node, SmallVectorImpl<SDValue> &Results,
2515 std::optional<unsigned> CallRetResNo) {
2516 LLVMContext &Ctx = *getContext();
2517 EVT VT = Node->getValueType(ResNo: 0);
2518 unsigned NumResults = Node->getNumValues();
2519
2520 const char *LCName = TLI->getLibcallName(Call: LC);
2521 if (!LC || !LCName)
2522 return false;
2523
2524 auto getVecDesc = [&]() -> VecDesc const * {
2525 for (bool Masked : {false, true}) {
2526 if (VecDesc const *VD = getLibInfo().getVectorMappingInfo(
2527 F: LCName, VF: VT.getVectorElementCount(), Masked)) {
2528 return VD;
2529 }
2530 }
2531 return nullptr;
2532 };
2533
2534 // For vector types, we must find a vector mapping for the libcall.
2535 VecDesc const *VD = nullptr;
2536 if (VT.isVector() && !(VD = getVecDesc()))
2537 return false;
2538
2539 // Find users of the node that store the results (and share input chains). The
2540 // destination pointers can be used instead of creating stack allocations.
2541 SDValue StoresInChain;
2542 SmallVector<StoreSDNode *, 2> ResultStores(NumResults);
2543 for (SDNode *User : Node->users()) {
2544 if (!ISD::isNormalStore(N: User))
2545 continue;
2546 auto *ST = cast<StoreSDNode>(Val: User);
2547 SDValue StoreValue = ST->getValue();
2548 unsigned ResNo = StoreValue.getResNo();
2549 // Ensure the store corresponds to an output pointer.
2550 if (CallRetResNo == ResNo)
2551 continue;
2552 // Ensure the store to the default address space and not atomic or volatile.
2553 if (!ST->isSimple() || ST->getAddressSpace() != 0)
2554 continue;
2555 // Ensure all store chains are the same (so they don't alias).
2556 if (StoresInChain && ST->getChain() != StoresInChain)
2557 continue;
2558 // Ensure the store is properly aligned.
2559 Type *StoreType = StoreValue.getValueType().getTypeForEVT(Context&: Ctx);
2560 if (ST->getAlign() <
2561 getDataLayout().getABITypeAlign(Ty: StoreType->getScalarType()))
2562 continue;
2563 // Avoid:
2564 // 1. Creating cyclic dependencies.
2565 // 2. Expanding the node to a call within a call sequence.
2566 if (!canFoldStoreIntoLibCallOutputPointers(StoreNode: ST, FPNode: Node))
2567 continue;
2568 ResultStores[ResNo] = ST;
2569 StoresInChain = ST->getChain();
2570 }
2571
2572 TargetLowering::ArgListTy Args;
2573 auto AddArgListEntry = [&](SDValue Node, Type *Ty) {
2574 TargetLowering::ArgListEntry Entry{};
2575 Entry.Ty = Ty;
2576 Entry.Node = Node;
2577 Args.push_back(x: Entry);
2578 };
2579
2580 // Pass the arguments.
2581 for (const SDValue &Op : Node->op_values()) {
2582 EVT ArgVT = Op.getValueType();
2583 Type *ArgTy = ArgVT.getTypeForEVT(Context&: Ctx);
2584 AddArgListEntry(Op, ArgTy);
2585 }
2586
2587 // Pass the output pointers.
2588 SmallVector<SDValue, 2> ResultPtrs(NumResults);
2589 Type *PointerTy = PointerType::getUnqual(C&: Ctx);
2590 for (auto [ResNo, ST] : llvm::enumerate(First&: ResultStores)) {
2591 if (ResNo == CallRetResNo)
2592 continue;
2593 EVT ResVT = Node->getValueType(ResNo);
2594 SDValue ResultPtr = ST ? ST->getBasePtr() : CreateStackTemporary(VT: ResVT);
2595 ResultPtrs[ResNo] = ResultPtr;
2596 AddArgListEntry(ResultPtr, PointerTy);
2597 }
2598
2599 SDLoc DL(Node);
2600
2601 // Pass the vector mask (if required).
2602 if (VD && VD->isMasked()) {
2603 EVT MaskVT = TLI->getSetCCResultType(DL: getDataLayout(), Context&: Ctx, VT);
2604 SDValue Mask = getBoolConstant(V: true, DL, VT: MaskVT, OpVT: VT);
2605 AddArgListEntry(Mask, MaskVT.getTypeForEVT(Context&: Ctx));
2606 }
2607
2608 Type *RetType = CallRetResNo.has_value()
2609 ? Node->getValueType(ResNo: *CallRetResNo).getTypeForEVT(Context&: Ctx)
2610 : Type::getVoidTy(C&: Ctx);
2611 SDValue InChain = StoresInChain ? StoresInChain : getEntryNode();
2612 SDValue Callee = getExternalSymbol(Sym: VD ? VD->getVectorFnName().data() : LCName,
2613 VT: TLI->getPointerTy(DL: getDataLayout()));
2614 TargetLowering::CallLoweringInfo CLI(*this);
2615 CLI.setDebugLoc(DL).setChain(InChain).setLibCallee(
2616 CC: TLI->getLibcallCallingConv(Call: LC), ResultType: RetType, Target: Callee, ArgsList: std::move(Args));
2617
2618 auto [Call, CallChain] = TLI->LowerCallTo(CLI);
2619
2620 for (auto [ResNo, ResultPtr] : llvm::enumerate(First&: ResultPtrs)) {
2621 if (ResNo == CallRetResNo) {
2622 Results.push_back(Elt: Call);
2623 continue;
2624 }
2625 MachinePointerInfo PtrInfo;
2626 SDValue LoadResult =
2627 getLoad(VT: Node->getValueType(ResNo), dl: DL, Chain: CallChain, Ptr: ResultPtr, PtrInfo);
2628 SDValue OutChain = LoadResult.getValue(R: 1);
2629
2630 if (StoreSDNode *ST = ResultStores[ResNo]) {
2631 // Replace store with the library call.
2632 ReplaceAllUsesOfValueWith(From: SDValue(ST, 0), To: OutChain);
2633 PtrInfo = ST->getPointerInfo();
2634 } else {
2635 PtrInfo = MachinePointerInfo::getFixedStack(
2636 MF&: getMachineFunction(), FI: cast<FrameIndexSDNode>(Val&: ResultPtr)->getIndex());
2637 }
2638
2639 Results.push_back(Elt: LoadResult);
2640 }
2641
2642 return true;
2643}
2644
2645SDValue SelectionDAG::expandVAArg(SDNode *Node) {
2646 SDLoc dl(Node);
2647 const TargetLowering &TLI = getTargetLoweringInfo();
2648 const Value *V = cast<SrcValueSDNode>(Val: Node->getOperand(Num: 2))->getValue();
2649 EVT VT = Node->getValueType(ResNo: 0);
2650 SDValue Tmp1 = Node->getOperand(Num: 0);
2651 SDValue Tmp2 = Node->getOperand(Num: 1);
2652 const MaybeAlign MA(Node->getConstantOperandVal(Num: 3));
2653
2654 SDValue VAListLoad = getLoad(VT: TLI.getPointerTy(DL: getDataLayout()), dl, Chain: Tmp1,
2655 Ptr: Tmp2, PtrInfo: MachinePointerInfo(V));
2656 SDValue VAList = VAListLoad;
2657
2658 if (MA && *MA > TLI.getMinStackArgumentAlignment()) {
2659 VAList = getNode(Opcode: ISD::ADD, DL: dl, VT: VAList.getValueType(), N1: VAList,
2660 N2: getConstant(Val: MA->value() - 1, DL: dl, VT: VAList.getValueType()));
2661
2662 VAList = getNode(
2663 Opcode: ISD::AND, DL: dl, VT: VAList.getValueType(), N1: VAList,
2664 N2: getSignedConstant(Val: -(int64_t)MA->value(), DL: dl, VT: VAList.getValueType()));
2665 }
2666
2667 // Increment the pointer, VAList, to the next vaarg
2668 Tmp1 = getNode(Opcode: ISD::ADD, DL: dl, VT: VAList.getValueType(), N1: VAList,
2669 N2: getConstant(Val: getDataLayout().getTypeAllocSize(
2670 Ty: VT.getTypeForEVT(Context&: *getContext())),
2671 DL: dl, VT: VAList.getValueType()));
2672 // Store the incremented VAList to the legalized pointer
2673 Tmp1 =
2674 getStore(Chain: VAListLoad.getValue(R: 1), dl, Val: Tmp1, Ptr: Tmp2, PtrInfo: MachinePointerInfo(V));
2675 // Load the actual argument out of the pointer VAList
2676 return getLoad(VT, dl, Chain: Tmp1, Ptr: VAList, PtrInfo: MachinePointerInfo());
2677}
2678
2679SDValue SelectionDAG::expandVACopy(SDNode *Node) {
2680 SDLoc dl(Node);
2681 const TargetLowering &TLI = getTargetLoweringInfo();
2682 // This defaults to loading a pointer from the input and storing it to the
2683 // output, returning the chain.
2684 const Value *VD = cast<SrcValueSDNode>(Val: Node->getOperand(Num: 3))->getValue();
2685 const Value *VS = cast<SrcValueSDNode>(Val: Node->getOperand(Num: 4))->getValue();
2686 SDValue Tmp1 =
2687 getLoad(VT: TLI.getPointerTy(DL: getDataLayout()), dl, Chain: Node->getOperand(Num: 0),
2688 Ptr: Node->getOperand(Num: 2), PtrInfo: MachinePointerInfo(VS));
2689 return getStore(Chain: Tmp1.getValue(R: 1), dl, Val: Tmp1, Ptr: Node->getOperand(Num: 1),
2690 PtrInfo: MachinePointerInfo(VD));
2691}
2692
2693Align SelectionDAG::getReducedAlign(EVT VT, bool UseABI) {
2694 const DataLayout &DL = getDataLayout();
2695 Type *Ty = VT.getTypeForEVT(Context&: *getContext());
2696 Align RedAlign = UseABI ? DL.getABITypeAlign(Ty) : DL.getPrefTypeAlign(Ty);
2697
2698 if (TLI->isTypeLegal(VT) || !VT.isVector())
2699 return RedAlign;
2700
2701 const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
2702 const Align StackAlign = TFI->getStackAlign();
2703
2704 // See if we can choose a smaller ABI alignment in cases where it's an
2705 // illegal vector type that will get broken down.
2706 if (RedAlign > StackAlign) {
2707 EVT IntermediateVT;
2708 MVT RegisterVT;
2709 unsigned NumIntermediates;
2710 TLI->getVectorTypeBreakdown(Context&: *getContext(), VT, IntermediateVT,
2711 NumIntermediates, RegisterVT);
2712 Ty = IntermediateVT.getTypeForEVT(Context&: *getContext());
2713 Align RedAlign2 = UseABI ? DL.getABITypeAlign(Ty) : DL.getPrefTypeAlign(Ty);
2714 if (RedAlign2 < RedAlign)
2715 RedAlign = RedAlign2;
2716
2717 if (!getMachineFunction().getFrameInfo().isStackRealignable())
2718 // If the stack is not realignable, the alignment should be limited to the
2719 // StackAlignment
2720 RedAlign = std::min(a: RedAlign, b: StackAlign);
2721 }
2722
2723 return RedAlign;
2724}
2725
2726SDValue SelectionDAG::CreateStackTemporary(TypeSize Bytes, Align Alignment) {
2727 MachineFrameInfo &MFI = MF->getFrameInfo();
2728 const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
2729 int StackID = 0;
2730 if (Bytes.isScalable())
2731 StackID = TFI->getStackIDForScalableVectors();
2732 // The stack id gives an indication of whether the object is scalable or
2733 // not, so it's safe to pass in the minimum size here.
2734 int FrameIdx = MFI.CreateStackObject(Size: Bytes.getKnownMinValue(), Alignment,
2735 isSpillSlot: false, Alloca: nullptr, ID: StackID);
2736 return getFrameIndex(FI: FrameIdx, VT: TLI->getFrameIndexTy(DL: getDataLayout()));
2737}
2738
2739SDValue SelectionDAG::CreateStackTemporary(EVT VT, unsigned minAlign) {
2740 Type *Ty = VT.getTypeForEVT(Context&: *getContext());
2741 Align StackAlign =
2742 std::max(a: getDataLayout().getPrefTypeAlign(Ty), b: Align(minAlign));
2743 return CreateStackTemporary(Bytes: VT.getStoreSize(), Alignment: StackAlign);
2744}
2745
2746SDValue SelectionDAG::CreateStackTemporary(EVT VT1, EVT VT2) {
2747 TypeSize VT1Size = VT1.getStoreSize();
2748 TypeSize VT2Size = VT2.getStoreSize();
2749 assert(VT1Size.isScalable() == VT2Size.isScalable() &&
2750 "Don't know how to choose the maximum size when creating a stack "
2751 "temporary");
2752 TypeSize Bytes = VT1Size.getKnownMinValue() > VT2Size.getKnownMinValue()
2753 ? VT1Size
2754 : VT2Size;
2755
2756 Type *Ty1 = VT1.getTypeForEVT(Context&: *getContext());
2757 Type *Ty2 = VT2.getTypeForEVT(Context&: *getContext());
2758 const DataLayout &DL = getDataLayout();
2759 Align Align = std::max(a: DL.getPrefTypeAlign(Ty: Ty1), b: DL.getPrefTypeAlign(Ty: Ty2));
2760 return CreateStackTemporary(Bytes, Alignment: Align);
2761}
2762
2763SDValue SelectionDAG::FoldSetCC(EVT VT, SDValue N1, SDValue N2,
2764 ISD::CondCode Cond, const SDLoc &dl) {
2765 EVT OpVT = N1.getValueType();
2766
2767 auto GetUndefBooleanConstant = [&]() {
2768 if (VT.getScalarType() == MVT::i1 ||
2769 TLI->getBooleanContents(Type: OpVT) ==
2770 TargetLowering::UndefinedBooleanContent)
2771 return getUNDEF(VT);
2772 // ZeroOrOne / ZeroOrNegative require specific values for the high bits,
2773 // so we cannot use getUNDEF(). Return zero instead.
2774 return getConstant(Val: 0, DL: dl, VT);
2775 };
2776
2777 // These setcc operations always fold.
2778 switch (Cond) {
2779 default: break;
2780 case ISD::SETFALSE:
2781 case ISD::SETFALSE2: return getBoolConstant(V: false, DL: dl, VT, OpVT);
2782 case ISD::SETTRUE:
2783 case ISD::SETTRUE2: return getBoolConstant(V: true, DL: dl, VT, OpVT);
2784
2785 case ISD::SETOEQ:
2786 case ISD::SETOGT:
2787 case ISD::SETOGE:
2788 case ISD::SETOLT:
2789 case ISD::SETOLE:
2790 case ISD::SETONE:
2791 case ISD::SETO:
2792 case ISD::SETUO:
2793 case ISD::SETUEQ:
2794 case ISD::SETUNE:
2795 assert(!OpVT.isInteger() && "Illegal setcc for integer!");
2796 break;
2797 }
2798
2799 if (OpVT.isInteger()) {
2800 // For EQ and NE, we can always pick a value for the undef to make the
2801 // predicate pass or fail, so we can return undef.
2802 // Matches behavior in llvm::ConstantFoldCompareInstruction.
2803 // icmp eq/ne X, undef -> undef.
2804 if ((N1.isUndef() || N2.isUndef()) &&
2805 (Cond == ISD::SETEQ || Cond == ISD::SETNE))
2806 return GetUndefBooleanConstant();
2807
2808 // If both operands are undef, we can return undef for int comparison.
2809 // icmp undef, undef -> undef.
2810 if (N1.isUndef() && N2.isUndef())
2811 return GetUndefBooleanConstant();
2812
2813 // icmp X, X -> true/false
2814 // icmp X, undef -> true/false because undef could be X.
2815 if (N1.isUndef() || N2.isUndef() || N1 == N2)
2816 return getBoolConstant(V: ISD::isTrueWhenEqual(Cond), DL: dl, VT, OpVT);
2817 }
2818
2819 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Val&: N2)) {
2820 const APInt &C2 = N2C->getAPIntValue();
2821 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Val&: N1)) {
2822 const APInt &C1 = N1C->getAPIntValue();
2823
2824 return getBoolConstant(V: ICmpInst::compare(LHS: C1, RHS: C2, Pred: getICmpCondCode(Pred: Cond)),
2825 DL: dl, VT, OpVT);
2826 }
2827 }
2828
2829 auto *N1CFP = dyn_cast<ConstantFPSDNode>(Val&: N1);
2830 auto *N2CFP = dyn_cast<ConstantFPSDNode>(Val&: N2);
2831
2832 if (N1CFP && N2CFP) {
2833 APFloat::cmpResult R = N1CFP->getValueAPF().compare(RHS: N2CFP->getValueAPF());
2834 switch (Cond) {
2835 default: break;
2836 case ISD::SETEQ: if (R==APFloat::cmpUnordered)
2837 return GetUndefBooleanConstant();
2838 [[fallthrough]];
2839 case ISD::SETOEQ: return getBoolConstant(V: R==APFloat::cmpEqual, DL: dl, VT,
2840 OpVT);
2841 case ISD::SETNE: if (R==APFloat::cmpUnordered)
2842 return GetUndefBooleanConstant();
2843 [[fallthrough]];
2844 case ISD::SETONE: return getBoolConstant(V: R==APFloat::cmpGreaterThan ||
2845 R==APFloat::cmpLessThan, DL: dl, VT,
2846 OpVT);
2847 case ISD::SETLT: if (R==APFloat::cmpUnordered)
2848 return GetUndefBooleanConstant();
2849 [[fallthrough]];
2850 case ISD::SETOLT: return getBoolConstant(V: R==APFloat::cmpLessThan, DL: dl, VT,
2851 OpVT);
2852 case ISD::SETGT: if (R==APFloat::cmpUnordered)
2853 return GetUndefBooleanConstant();
2854 [[fallthrough]];
2855 case ISD::SETOGT: return getBoolConstant(V: R==APFloat::cmpGreaterThan, DL: dl,
2856 VT, OpVT);
2857 case ISD::SETLE: if (R==APFloat::cmpUnordered)
2858 return GetUndefBooleanConstant();
2859 [[fallthrough]];
2860 case ISD::SETOLE: return getBoolConstant(V: R==APFloat::cmpLessThan ||
2861 R==APFloat::cmpEqual, DL: dl, VT,
2862 OpVT);
2863 case ISD::SETGE: if (R==APFloat::cmpUnordered)
2864 return GetUndefBooleanConstant();
2865 [[fallthrough]];
2866 case ISD::SETOGE: return getBoolConstant(V: R==APFloat::cmpGreaterThan ||
2867 R==APFloat::cmpEqual, DL: dl, VT, OpVT);
2868 case ISD::SETO: return getBoolConstant(V: R!=APFloat::cmpUnordered, DL: dl, VT,
2869 OpVT);
2870 case ISD::SETUO: return getBoolConstant(V: R==APFloat::cmpUnordered, DL: dl, VT,
2871 OpVT);
2872 case ISD::SETUEQ: return getBoolConstant(V: R==APFloat::cmpUnordered ||
2873 R==APFloat::cmpEqual, DL: dl, VT,
2874 OpVT);
2875 case ISD::SETUNE: return getBoolConstant(V: R!=APFloat::cmpEqual, DL: dl, VT,
2876 OpVT);
2877 case ISD::SETULT: return getBoolConstant(V: R==APFloat::cmpUnordered ||
2878 R==APFloat::cmpLessThan, DL: dl, VT,
2879 OpVT);
2880 case ISD::SETUGT: return getBoolConstant(V: R==APFloat::cmpGreaterThan ||
2881 R==APFloat::cmpUnordered, DL: dl, VT,
2882 OpVT);
2883 case ISD::SETULE: return getBoolConstant(V: R!=APFloat::cmpGreaterThan, DL: dl,
2884 VT, OpVT);
2885 case ISD::SETUGE: return getBoolConstant(V: R!=APFloat::cmpLessThan, DL: dl, VT,
2886 OpVT);
2887 }
2888 } else if (N1CFP && OpVT.isSimple() && !N2.isUndef()) {
2889 // Ensure that the constant occurs on the RHS.
2890 ISD::CondCode SwappedCond = ISD::getSetCCSwappedOperands(Operation: Cond);
2891 if (!TLI->isCondCodeLegal(CC: SwappedCond, VT: OpVT.getSimpleVT()))
2892 return SDValue();
2893 return getSetCC(DL: dl, VT, LHS: N2, RHS: N1, Cond: SwappedCond);
2894 } else if ((N2CFP && N2CFP->getValueAPF().isNaN()) ||
2895 (OpVT.isFloatingPoint() && (N1.isUndef() || N2.isUndef()))) {
2896 // If an operand is known to be a nan (or undef that could be a nan), we can
2897 // fold it.
2898 // Choosing NaN for the undef will always make unordered comparison succeed
2899 // and ordered comparison fails.
2900 // Matches behavior in llvm::ConstantFoldCompareInstruction.
2901 switch (ISD::getUnorderedFlavor(Cond)) {
2902 default:
2903 llvm_unreachable("Unknown flavor!");
2904 case 0: // Known false.
2905 return getBoolConstant(V: false, DL: dl, VT, OpVT);
2906 case 1: // Known true.
2907 return getBoolConstant(V: true, DL: dl, VT, OpVT);
2908 case 2: // Undefined.
2909 return GetUndefBooleanConstant();
2910 }
2911 }
2912
2913 // Could not fold it.
2914 return SDValue();
2915}
2916
2917/// SignBitIsZero - Return true if the sign bit of Op is known to be zero. We
2918/// use this predicate to simplify operations downstream.
2919bool SelectionDAG::SignBitIsZero(SDValue Op, unsigned Depth) const {
2920 unsigned BitWidth = Op.getScalarValueSizeInBits();
2921 return MaskedValueIsZero(Op, Mask: APInt::getSignMask(BitWidth), Depth);
2922}
2923
2924/// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
2925/// this predicate to simplify operations downstream. Mask is known to be zero
2926/// for bits that V cannot have.
2927bool SelectionDAG::MaskedValueIsZero(SDValue V, const APInt &Mask,
2928 unsigned Depth) const {
2929 return Mask.isSubsetOf(RHS: computeKnownBits(Op: V, Depth).Zero);
2930}
2931
2932/// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero in
2933/// DemandedElts. We use this predicate to simplify operations downstream.
2934/// Mask is known to be zero for bits that V cannot have.
2935bool SelectionDAG::MaskedValueIsZero(SDValue V, const APInt &Mask,
2936 const APInt &DemandedElts,
2937 unsigned Depth) const {
2938 return Mask.isSubsetOf(RHS: computeKnownBits(Op: V, DemandedElts, Depth).Zero);
2939}
2940
2941/// MaskedVectorIsZero - Return true if 'Op' is known to be zero in
2942/// DemandedElts. We use this predicate to simplify operations downstream.
2943bool SelectionDAG::MaskedVectorIsZero(SDValue V, const APInt &DemandedElts,
2944 unsigned Depth /* = 0 */) const {
2945 return computeKnownBits(Op: V, DemandedElts, Depth).isZero();
2946}
2947
2948/// MaskedValueIsAllOnes - Return true if '(Op & Mask) == Mask'.
2949bool SelectionDAG::MaskedValueIsAllOnes(SDValue V, const APInt &Mask,
2950 unsigned Depth) const {
2951 return Mask.isSubsetOf(RHS: computeKnownBits(Op: V, Depth).One);
2952}
2953
2954APInt SelectionDAG::computeVectorKnownZeroElements(SDValue Op,
2955 const APInt &DemandedElts,
2956 unsigned Depth) const {
2957 EVT VT = Op.getValueType();
2958 assert(VT.isVector() && !VT.isScalableVector() && "Only for fixed vectors!");
2959
2960 unsigned NumElts = VT.getVectorNumElements();
2961 assert(DemandedElts.getBitWidth() == NumElts && "Unexpected demanded mask.");
2962
2963 APInt KnownZeroElements = APInt::getZero(numBits: NumElts);
2964 for (unsigned EltIdx = 0; EltIdx != NumElts; ++EltIdx) {
2965 if (!DemandedElts[EltIdx])
2966 continue; // Don't query elements that are not demanded.
2967 APInt Mask = APInt::getOneBitSet(numBits: NumElts, BitNo: EltIdx);
2968 if (MaskedVectorIsZero(V: Op, DemandedElts: Mask, Depth))
2969 KnownZeroElements.setBit(EltIdx);
2970 }
2971 return KnownZeroElements;
2972}
2973
2974/// isSplatValue - Return true if the vector V has the same value
2975/// across all DemandedElts. For scalable vectors, we don't know the
2976/// number of lanes at compile time. Instead, we use a 1 bit APInt
2977/// to represent a conservative value for all lanes; that is, that
2978/// one bit value is implicitly splatted across all lanes.
2979bool SelectionDAG::isSplatValue(SDValue V, const APInt &DemandedElts,
2980 APInt &UndefElts, unsigned Depth) const {
2981 unsigned Opcode = V.getOpcode();
2982 EVT VT = V.getValueType();
2983 assert(VT.isVector() && "Vector type expected");
2984 assert((!VT.isScalableVector() || DemandedElts.getBitWidth() == 1) &&
2985 "scalable demanded bits are ignored");
2986
2987 if (!DemandedElts)
2988 return false; // No demanded elts, better to assume we don't know anything.
2989
2990 if (Depth >= MaxRecursionDepth)
2991 return false; // Limit search depth.
2992
2993 // Deal with some common cases here that work for both fixed and scalable
2994 // vector types.
2995 switch (Opcode) {
2996 case ISD::SPLAT_VECTOR:
2997 UndefElts = V.getOperand(i: 0).isUndef()
2998 ? APInt::getAllOnes(numBits: DemandedElts.getBitWidth())
2999 : APInt(DemandedElts.getBitWidth(), 0);
3000 return true;
3001 case ISD::ADD:
3002 case ISD::SUB:
3003 case ISD::AND:
3004 case ISD::XOR:
3005 case ISD::OR: {
3006 APInt UndefLHS, UndefRHS;
3007 SDValue LHS = V.getOperand(i: 0);
3008 SDValue RHS = V.getOperand(i: 1);
3009 // Only recognize splats with the same demanded undef elements for both
3010 // operands, otherwise we might fail to handle binop-specific undef
3011 // handling.
3012 // e.g. (and undef, 0) -> 0 etc.
3013 if (isSplatValue(V: LHS, DemandedElts, UndefElts&: UndefLHS, Depth: Depth + 1) &&
3014 isSplatValue(V: RHS, DemandedElts, UndefElts&: UndefRHS, Depth: Depth + 1) &&
3015 (DemandedElts & UndefLHS) == (DemandedElts & UndefRHS)) {
3016 UndefElts = UndefLHS | UndefRHS;
3017 return true;
3018 }
3019 return false;
3020 }
3021 case ISD::ABS:
3022 case ISD::TRUNCATE:
3023 case ISD::SIGN_EXTEND:
3024 case ISD::ZERO_EXTEND:
3025 return isSplatValue(V: V.getOperand(i: 0), DemandedElts, UndefElts, Depth: Depth + 1);
3026 default:
3027 if (Opcode >= ISD::BUILTIN_OP_END || Opcode == ISD::INTRINSIC_WO_CHAIN ||
3028 Opcode == ISD::INTRINSIC_W_CHAIN || Opcode == ISD::INTRINSIC_VOID)
3029 return TLI->isSplatValueForTargetNode(Op: V, DemandedElts, UndefElts, DAG: *this,
3030 Depth);
3031 break;
3032}
3033
3034 // We don't support other cases than those above for scalable vectors at
3035 // the moment.
3036 if (VT.isScalableVector())
3037 return false;
3038
3039 unsigned NumElts = VT.getVectorNumElements();
3040 assert(NumElts == DemandedElts.getBitWidth() && "Vector size mismatch");
3041 UndefElts = APInt::getZero(numBits: NumElts);
3042
3043 switch (Opcode) {
3044 case ISD::BUILD_VECTOR: {
3045 SDValue Scl;
3046 for (unsigned i = 0; i != NumElts; ++i) {
3047 SDValue Op = V.getOperand(i);
3048 if (Op.isUndef()) {
3049 UndefElts.setBit(i);
3050 continue;
3051 }
3052 if (!DemandedElts[i])
3053 continue;
3054 if (Scl && Scl != Op)
3055 return false;
3056 Scl = Op;
3057 }
3058 return true;
3059 }
3060 case ISD::VECTOR_SHUFFLE: {
3061 // Check if this is a shuffle node doing a splat or a shuffle of a splat.
3062 APInt DemandedLHS = APInt::getZero(numBits: NumElts);
3063 APInt DemandedRHS = APInt::getZero(numBits: NumElts);
3064 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Val&: V)->getMask();
3065 for (int i = 0; i != (int)NumElts; ++i) {
3066 int M = Mask[i];
3067 if (M < 0) {
3068 UndefElts.setBit(i);
3069 continue;
3070 }
3071 if (!DemandedElts[i])
3072 continue;
3073 if (M < (int)NumElts)
3074 DemandedLHS.setBit(M);
3075 else
3076 DemandedRHS.setBit(M - NumElts);
3077 }
3078
3079 // If we aren't demanding either op, assume there's no splat.
3080 // If we are demanding both ops, assume there's no splat.
3081 if ((DemandedLHS.isZero() && DemandedRHS.isZero()) ||
3082 (!DemandedLHS.isZero() && !DemandedRHS.isZero()))
3083 return false;
3084
3085 // See if the demanded elts of the source op is a splat or we only demand
3086 // one element, which should always be a splat.
3087 // TODO: Handle source ops splats with undefs.
3088 auto CheckSplatSrc = [&](SDValue Src, const APInt &SrcElts) {
3089 APInt SrcUndefs;
3090 return (SrcElts.popcount() == 1) ||
3091 (isSplatValue(V: Src, DemandedElts: SrcElts, UndefElts&: SrcUndefs, Depth: Depth + 1) &&
3092 (SrcElts & SrcUndefs).isZero());
3093 };
3094 if (!DemandedLHS.isZero())
3095 return CheckSplatSrc(V.getOperand(i: 0), DemandedLHS);
3096 return CheckSplatSrc(V.getOperand(i: 1), DemandedRHS);
3097 }
3098 case ISD::EXTRACT_SUBVECTOR: {
3099 // Offset the demanded elts by the subvector index.
3100 SDValue Src = V.getOperand(i: 0);
3101 // We don't support scalable vectors at the moment.
3102 if (Src.getValueType().isScalableVector())
3103 return false;
3104 uint64_t Idx = V.getConstantOperandVal(i: 1);
3105 unsigned NumSrcElts = Src.getValueType().getVectorNumElements();
3106 APInt UndefSrcElts;
3107 APInt DemandedSrcElts = DemandedElts.zext(width: NumSrcElts).shl(shiftAmt: Idx);
3108 if (isSplatValue(V: Src, DemandedElts: DemandedSrcElts, UndefElts&: UndefSrcElts, Depth: Depth + 1)) {
3109 UndefElts = UndefSrcElts.extractBits(numBits: NumElts, bitPosition: Idx);
3110 return true;
3111 }
3112 break;
3113 }
3114 case ISD::ANY_EXTEND_VECTOR_INREG:
3115 case ISD::SIGN_EXTEND_VECTOR_INREG:
3116 case ISD::ZERO_EXTEND_VECTOR_INREG: {
3117 // Widen the demanded elts by the src element count.
3118 SDValue Src = V.getOperand(i: 0);
3119 // We don't support scalable vectors at the moment.
3120 if (Src.getValueType().isScalableVector())
3121 return false;
3122 unsigned NumSrcElts = Src.getValueType().getVectorNumElements();
3123 APInt UndefSrcElts;
3124 APInt DemandedSrcElts = DemandedElts.zext(width: NumSrcElts);
3125 if (isSplatValue(V: Src, DemandedElts: DemandedSrcElts, UndefElts&: UndefSrcElts, Depth: Depth + 1)) {
3126 UndefElts = UndefSrcElts.trunc(width: NumElts);
3127 return true;
3128 }
3129 break;
3130 }
3131 case ISD::BITCAST: {
3132 SDValue Src = V.getOperand(i: 0);
3133 EVT SrcVT = Src.getValueType();
3134 unsigned SrcBitWidth = SrcVT.getScalarSizeInBits();
3135 unsigned BitWidth = VT.getScalarSizeInBits();
3136
3137 // Ignore bitcasts from unsupported types.
3138 // TODO: Add fp support?
3139 if (!SrcVT.isVector() || !SrcVT.isInteger() || !VT.isInteger())
3140 break;
3141
3142 // Bitcast 'small element' vector to 'large element' vector.
3143 if ((BitWidth % SrcBitWidth) == 0) {
3144 // See if each sub element is a splat.
3145 unsigned Scale = BitWidth / SrcBitWidth;
3146 unsigned NumSrcElts = SrcVT.getVectorNumElements();
3147 APInt ScaledDemandedElts =
3148 APIntOps::ScaleBitMask(A: DemandedElts, NewBitWidth: NumSrcElts);
3149 for (unsigned I = 0; I != Scale; ++I) {
3150 APInt SubUndefElts;
3151 APInt SubDemandedElt = APInt::getOneBitSet(numBits: Scale, BitNo: I);
3152 APInt SubDemandedElts = APInt::getSplat(NewLen: NumSrcElts, V: SubDemandedElt);
3153 SubDemandedElts &= ScaledDemandedElts;
3154 if (!isSplatValue(V: Src, DemandedElts: SubDemandedElts, UndefElts&: SubUndefElts, Depth: Depth + 1))
3155 return false;
3156 // TODO: Add support for merging sub undef elements.
3157 if (!SubUndefElts.isZero())
3158 return false;
3159 }
3160 return true;
3161 }
3162 break;
3163 }
3164 }
3165
3166 return false;
3167}
3168
3169/// Helper wrapper to main isSplatValue function.
3170bool SelectionDAG::isSplatValue(SDValue V, bool AllowUndefs) const {
3171 EVT VT = V.getValueType();
3172 assert(VT.isVector() && "Vector type expected");
3173
3174 APInt UndefElts;
3175 // Since the number of lanes in a scalable vector is unknown at compile time,
3176 // we track one bit which is implicitly broadcast to all lanes. This means
3177 // that all lanes in a scalable vector are considered demanded.
3178 APInt DemandedElts
3179 = APInt::getAllOnes(numBits: VT.isScalableVector() ? 1 : VT.getVectorNumElements());
3180 return isSplatValue(V, DemandedElts, UndefElts) &&
3181 (AllowUndefs || !UndefElts);
3182}
3183
3184SDValue SelectionDAG::getSplatSourceVector(SDValue V, int &SplatIdx) {
3185 V = peekThroughExtractSubvectors(V);
3186
3187 EVT VT = V.getValueType();
3188 unsigned Opcode = V.getOpcode();
3189 switch (Opcode) {
3190 default: {
3191 APInt UndefElts;
3192 // Since the number of lanes in a scalable vector is unknown at compile time,
3193 // we track one bit which is implicitly broadcast to all lanes. This means
3194 // that all lanes in a scalable vector are considered demanded.
3195 APInt DemandedElts
3196 = APInt::getAllOnes(numBits: VT.isScalableVector() ? 1 : VT.getVectorNumElements());
3197
3198 if (isSplatValue(V, DemandedElts, UndefElts)) {
3199 if (VT.isScalableVector()) {
3200 // DemandedElts and UndefElts are ignored for scalable vectors, since
3201 // the only supported cases are SPLAT_VECTOR nodes.
3202 SplatIdx = 0;
3203 } else {
3204 // Handle case where all demanded elements are UNDEF.
3205 if (DemandedElts.isSubsetOf(RHS: UndefElts)) {
3206 SplatIdx = 0;
3207 return getUNDEF(VT);
3208 }
3209 SplatIdx = (UndefElts & DemandedElts).countr_one();
3210 }
3211 return V;
3212 }
3213 break;
3214 }
3215 case ISD::SPLAT_VECTOR:
3216 SplatIdx = 0;
3217 return V;
3218 case ISD::VECTOR_SHUFFLE: {
3219 assert(!VT.isScalableVector());
3220 // Check if this is a shuffle node doing a splat.
3221 // TODO - remove this and rely purely on SelectionDAG::isSplatValue,
3222 // getTargetVShiftNode currently struggles without the splat source.
3223 auto *SVN = cast<ShuffleVectorSDNode>(Val&: V);
3224 if (!SVN->isSplat())
3225 break;
3226 int Idx = SVN->getSplatIndex();
3227 int NumElts = V.getValueType().getVectorNumElements();
3228 SplatIdx = Idx % NumElts;
3229 return V.getOperand(i: Idx / NumElts);
3230 }
3231 }
3232
3233 return SDValue();
3234}
3235
3236SDValue SelectionDAG::getSplatValue(SDValue V, bool LegalTypes) {
3237 int SplatIdx;
3238 if (SDValue SrcVector = getSplatSourceVector(V, SplatIdx)) {
3239 EVT SVT = SrcVector.getValueType().getScalarType();
3240 EVT LegalSVT = SVT;
3241 if (LegalTypes && !TLI->isTypeLegal(VT: SVT)) {
3242 if (!SVT.isInteger())
3243 return SDValue();
3244 LegalSVT = TLI->getTypeToTransformTo(Context&: *getContext(), VT: LegalSVT);
3245 if (LegalSVT.bitsLT(VT: SVT))
3246 return SDValue();
3247 }
3248 return getExtractVectorElt(DL: SDLoc(V), VT: LegalSVT, Vec: SrcVector, Idx: SplatIdx);
3249 }
3250 return SDValue();
3251}
3252
3253std::optional<ConstantRange>
3254SelectionDAG::getValidShiftAmountRange(SDValue V, const APInt &DemandedElts,
3255 unsigned Depth) const {
3256 assert((V.getOpcode() == ISD::SHL || V.getOpcode() == ISD::SRL ||
3257 V.getOpcode() == ISD::SRA) &&
3258 "Unknown shift node");
3259 // Shifting more than the bitwidth is not valid.
3260 unsigned BitWidth = V.getScalarValueSizeInBits();
3261
3262 if (auto *Cst = dyn_cast<ConstantSDNode>(Val: V.getOperand(i: 1))) {
3263 const APInt &ShAmt = Cst->getAPIntValue();
3264 if (ShAmt.uge(RHS: BitWidth))
3265 return std::nullopt;
3266 return ConstantRange(ShAmt);
3267 }
3268
3269 if (auto *BV = dyn_cast<BuildVectorSDNode>(Val: V.getOperand(i: 1))) {
3270 const APInt *MinAmt = nullptr, *MaxAmt = nullptr;
3271 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
3272 if (!DemandedElts[i])
3273 continue;
3274 auto *SA = dyn_cast<ConstantSDNode>(Val: BV->getOperand(Num: i));
3275 if (!SA) {
3276 MinAmt = MaxAmt = nullptr;
3277 break;
3278 }
3279 const APInt &ShAmt = SA->getAPIntValue();
3280 if (ShAmt.uge(RHS: BitWidth))
3281 return std::nullopt;
3282 if (!MinAmt || MinAmt->ugt(RHS: ShAmt))
3283 MinAmt = &ShAmt;
3284 if (!MaxAmt || MaxAmt->ult(RHS: ShAmt))
3285 MaxAmt = &ShAmt;
3286 }
3287 assert(((!MinAmt && !MaxAmt) || (MinAmt && MaxAmt)) &&
3288 "Failed to find matching min/max shift amounts");
3289 if (MinAmt && MaxAmt)
3290 return ConstantRange(*MinAmt, *MaxAmt + 1);
3291 }
3292
3293 // Use computeKnownBits to find a hidden constant/knownbits (usually type
3294 // legalized). e.g. Hidden behind multiple bitcasts/build_vector/casts etc.
3295 KnownBits KnownAmt = computeKnownBits(Op: V.getOperand(i: 1), DemandedElts, Depth);
3296 if (KnownAmt.getMaxValue().ult(RHS: BitWidth))
3297 return ConstantRange::fromKnownBits(Known: KnownAmt, /*IsSigned=*/false);
3298
3299 return std::nullopt;
3300}
3301
3302std::optional<uint64_t>
3303SelectionDAG::getValidShiftAmount(SDValue V, const APInt &DemandedElts,
3304 unsigned Depth) const {
3305 assert((V.getOpcode() == ISD::SHL || V.getOpcode() == ISD::SRL ||
3306 V.getOpcode() == ISD::SRA) &&
3307 "Unknown shift node");
3308 if (std::optional<ConstantRange> AmtRange =
3309 getValidShiftAmountRange(V, DemandedElts, Depth))
3310 if (const APInt *ShAmt = AmtRange->getSingleElement())
3311 return ShAmt->getZExtValue();
3312 return std::nullopt;
3313}
3314
3315std::optional<uint64_t>
3316SelectionDAG::getValidShiftAmount(SDValue V, unsigned Depth) const {
3317 EVT VT = V.getValueType();
3318 APInt DemandedElts = VT.isFixedLengthVector()
3319 ? APInt::getAllOnes(numBits: VT.getVectorNumElements())
3320 : APInt(1, 1);
3321 return getValidShiftAmount(V, DemandedElts, Depth);
3322}
3323
3324std::optional<uint64_t>
3325SelectionDAG::getValidMinimumShiftAmount(SDValue V, const APInt &DemandedElts,
3326 unsigned Depth) const {
3327 assert((V.getOpcode() == ISD::SHL || V.getOpcode() == ISD::SRL ||
3328 V.getOpcode() == ISD::SRA) &&
3329 "Unknown shift node");
3330 if (std::optional<ConstantRange> AmtRange =
3331 getValidShiftAmountRange(V, DemandedElts, Depth))
3332 return AmtRange->getUnsignedMin().getZExtValue();
3333 return std::nullopt;
3334}
3335
3336std::optional<uint64_t>
3337SelectionDAG::getValidMinimumShiftAmount(SDValue V, unsigned Depth) const {
3338 EVT VT = V.getValueType();
3339 APInt DemandedElts = VT.isFixedLengthVector()
3340 ? APInt::getAllOnes(numBits: VT.getVectorNumElements())
3341 : APInt(1, 1);
3342 return getValidMinimumShiftAmount(V, DemandedElts, Depth);
3343}
3344
3345std::optional<uint64_t>
3346SelectionDAG::getValidMaximumShiftAmount(SDValue V, const APInt &DemandedElts,
3347 unsigned Depth) const {
3348 assert((V.getOpcode() == ISD::SHL || V.getOpcode() == ISD::SRL ||
3349 V.getOpcode() == ISD::SRA) &&
3350 "Unknown shift node");
3351 if (std::optional<ConstantRange> AmtRange =
3352 getValidShiftAmountRange(V, DemandedElts, Depth))
3353 return AmtRange->getUnsignedMax().getZExtValue();
3354 return std::nullopt;
3355}
3356
3357std::optional<uint64_t>
3358SelectionDAG::getValidMaximumShiftAmount(SDValue V, unsigned Depth) const {
3359 EVT VT = V.getValueType();
3360 APInt DemandedElts = VT.isFixedLengthVector()
3361 ? APInt::getAllOnes(numBits: VT.getVectorNumElements())
3362 : APInt(1, 1);
3363 return getValidMaximumShiftAmount(V, DemandedElts, Depth);
3364}
3365
3366/// Determine which bits of Op are known to be either zero or one and return
3367/// them in Known. For vectors, the known bits are those that are shared by
3368/// every vector element.
3369KnownBits SelectionDAG::computeKnownBits(SDValue Op, unsigned Depth) const {
3370 EVT VT = Op.getValueType();
3371
3372 // Since the number of lanes in a scalable vector is unknown at compile time,
3373 // we track one bit which is implicitly broadcast to all lanes. This means
3374 // that all lanes in a scalable vector are considered demanded.
3375 APInt DemandedElts = VT.isFixedLengthVector()
3376 ? APInt::getAllOnes(numBits: VT.getVectorNumElements())
3377 : APInt(1, 1);
3378 return computeKnownBits(Op, DemandedElts, Depth);
3379}
3380
3381/// Determine which bits of Op are known to be either zero or one and return
3382/// them in Known. The DemandedElts argument allows us to only collect the known
3383/// bits that are shared by the requested vector elements.
3384KnownBits SelectionDAG::computeKnownBits(SDValue Op, const APInt &DemandedElts,
3385 unsigned Depth) const {
3386 unsigned BitWidth = Op.getScalarValueSizeInBits();
3387
3388 KnownBits Known(BitWidth); // Don't know anything.
3389
3390 if (auto OptAPInt = Op->bitcastToAPInt()) {
3391 // We know all of the bits for a constant!
3392 return KnownBits::makeConstant(C: *std::move(OptAPInt));
3393 }
3394
3395 if (Depth >= MaxRecursionDepth)
3396 return Known; // Limit search depth.
3397
3398 KnownBits Known2;
3399 unsigned NumElts = DemandedElts.getBitWidth();
3400 assert((!Op.getValueType().isFixedLengthVector() ||
3401 NumElts == Op.getValueType().getVectorNumElements()) &&
3402 "Unexpected vector size");
3403
3404 if (!DemandedElts)
3405 return Known; // No demanded elts, better to assume we don't know anything.
3406
3407 unsigned Opcode = Op.getOpcode();
3408 switch (Opcode) {
3409 case ISD::MERGE_VALUES:
3410 return computeKnownBits(Op: Op.getOperand(i: Op.getResNo()), DemandedElts,
3411 Depth: Depth + 1);
3412 case ISD::SPLAT_VECTOR: {
3413 SDValue SrcOp = Op.getOperand(i: 0);
3414 assert(SrcOp.getValueSizeInBits() >= BitWidth &&
3415 "Expected SPLAT_VECTOR implicit truncation");
3416 // Implicitly truncate the bits to match the official semantics of
3417 // SPLAT_VECTOR.
3418 Known = computeKnownBits(Op: SrcOp, Depth: Depth + 1).trunc(BitWidth);
3419 break;
3420 }
3421 case ISD::SPLAT_VECTOR_PARTS: {
3422 unsigned ScalarSize = Op.getOperand(i: 0).getScalarValueSizeInBits();
3423 assert(ScalarSize * Op.getNumOperands() == BitWidth &&
3424 "Expected SPLAT_VECTOR_PARTS scalars to cover element width");
3425 for (auto [I, SrcOp] : enumerate(First: Op->ops())) {
3426 Known.insertBits(SubBits: computeKnownBits(Op: SrcOp, Depth: Depth + 1), BitPosition: ScalarSize * I);
3427 }
3428 break;
3429 }
3430 case ISD::STEP_VECTOR: {
3431 const APInt &Step = Op.getConstantOperandAPInt(i: 0);
3432
3433 if (Step.isPowerOf2())
3434 Known.Zero.setLowBits(Step.logBase2());
3435
3436 const Function &F = getMachineFunction().getFunction();
3437
3438 if (!isUIntN(N: BitWidth, x: Op.getValueType().getVectorMinNumElements()))
3439 break;
3440 const APInt MinNumElts =
3441 APInt(BitWidth, Op.getValueType().getVectorMinNumElements());
3442
3443 bool Overflow;
3444 const APInt MaxNumElts = getVScaleRange(F: &F, BitWidth)
3445 .getUnsignedMax()
3446 .umul_ov(RHS: MinNumElts, Overflow);
3447 if (Overflow)
3448 break;
3449
3450 const APInt MaxValue = (MaxNumElts - 1).umul_ov(RHS: Step, Overflow);
3451 if (Overflow)
3452 break;
3453
3454 Known.Zero.setHighBits(MaxValue.countl_zero());
3455 break;
3456 }
3457 case ISD::BUILD_VECTOR:
3458 assert(!Op.getValueType().isScalableVector());
3459 // Collect the known bits that are shared by every demanded vector element.
3460 Known.Zero.setAllBits(); Known.One.setAllBits();
3461 for (unsigned i = 0, e = Op.getNumOperands(); i != e; ++i) {
3462 if (!DemandedElts[i])
3463 continue;
3464
3465 SDValue SrcOp = Op.getOperand(i);
3466 Known2 = computeKnownBits(Op: SrcOp, Depth: Depth + 1);
3467
3468 // BUILD_VECTOR can implicitly truncate sources, we must handle this.
3469 if (SrcOp.getValueSizeInBits() != BitWidth) {
3470 assert(SrcOp.getValueSizeInBits() > BitWidth &&
3471 "Expected BUILD_VECTOR implicit truncation");
3472 Known2 = Known2.trunc(BitWidth);
3473 }
3474
3475 // Known bits are the values that are shared by every demanded element.
3476 Known = Known.intersectWith(RHS: Known2);
3477
3478 // If we don't know any bits, early out.
3479 if (Known.isUnknown())
3480 break;
3481 }
3482 break;
3483 case ISD::VECTOR_SHUFFLE: {
3484 assert(!Op.getValueType().isScalableVector());
3485 // Collect the known bits that are shared by every vector element referenced
3486 // by the shuffle.
3487 APInt DemandedLHS, DemandedRHS;
3488 const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Val&: Op);
3489 assert(NumElts == SVN->getMask().size() && "Unexpected vector size");
3490 if (!getShuffleDemandedElts(SrcWidth: NumElts, Mask: SVN->getMask(), DemandedElts,
3491 DemandedLHS, DemandedRHS))
3492 break;
3493
3494 // Known bits are the values that are shared by every demanded element.
3495 Known.Zero.setAllBits(); Known.One.setAllBits();
3496 if (!!DemandedLHS) {
3497 SDValue LHS = Op.getOperand(i: 0);
3498 Known2 = computeKnownBits(Op: LHS, DemandedElts: DemandedLHS, Depth: Depth + 1);
3499 Known = Known.intersectWith(RHS: Known2);
3500 }
3501 // If we don't know any bits, early out.
3502 if (Known.isUnknown())
3503 break;
3504 if (!!DemandedRHS) {
3505 SDValue RHS = Op.getOperand(i: 1);
3506 Known2 = computeKnownBits(Op: RHS, DemandedElts: DemandedRHS, Depth: Depth + 1);
3507 Known = Known.intersectWith(RHS: Known2);
3508 }
3509 break;
3510 }
3511 case ISD::VSCALE: {
3512 const Function &F = getMachineFunction().getFunction();
3513 const APInt &Multiplier = Op.getConstantOperandAPInt(i: 0);
3514 Known = getVScaleRange(F: &F, BitWidth).multiply(Other: Multiplier).toKnownBits();
3515 break;
3516 }
3517 case ISD::CONCAT_VECTORS: {
3518 if (Op.getValueType().isScalableVector())
3519 break;
3520 // Split DemandedElts and test each of the demanded subvectors.
3521 Known.Zero.setAllBits(); Known.One.setAllBits();
3522 EVT SubVectorVT = Op.getOperand(i: 0).getValueType();
3523 unsigned NumSubVectorElts = SubVectorVT.getVectorNumElements();
3524 unsigned NumSubVectors = Op.getNumOperands();
3525 for (unsigned i = 0; i != NumSubVectors; ++i) {
3526 APInt DemandedSub =
3527 DemandedElts.extractBits(numBits: NumSubVectorElts, bitPosition: i * NumSubVectorElts);
3528 if (!!DemandedSub) {
3529 SDValue Sub = Op.getOperand(i);
3530 Known2 = computeKnownBits(Op: Sub, DemandedElts: DemandedSub, Depth: Depth + 1);
3531 Known = Known.intersectWith(RHS: Known2);
3532 }
3533 // If we don't know any bits, early out.
3534 if (Known.isUnknown())
3535 break;
3536 }
3537 break;
3538 }
3539 case ISD::INSERT_SUBVECTOR: {
3540 if (Op.getValueType().isScalableVector())
3541 break;
3542 // Demand any elements from the subvector and the remainder from the src its
3543 // inserted into.
3544 SDValue Src = Op.getOperand(i: 0);
3545 SDValue Sub = Op.getOperand(i: 1);
3546 uint64_t Idx = Op.getConstantOperandVal(i: 2);
3547 unsigned NumSubElts = Sub.getValueType().getVectorNumElements();
3548 APInt DemandedSubElts = DemandedElts.extractBits(numBits: NumSubElts, bitPosition: Idx);
3549 APInt DemandedSrcElts = DemandedElts;
3550 DemandedSrcElts.clearBits(LoBit: Idx, HiBit: Idx + NumSubElts);
3551
3552 Known.One.setAllBits();
3553 Known.Zero.setAllBits();
3554 if (!!DemandedSubElts) {
3555 Known = computeKnownBits(Op: Sub, DemandedElts: DemandedSubElts, Depth: Depth + 1);
3556 if (Known.isUnknown())
3557 break; // early-out.
3558 }
3559 if (!!DemandedSrcElts) {
3560 Known2 = computeKnownBits(Op: Src, DemandedElts: DemandedSrcElts, Depth: Depth + 1);
3561 Known = Known.intersectWith(RHS: Known2);
3562 }
3563 break;
3564 }
3565 case ISD::EXTRACT_SUBVECTOR: {
3566 // Offset the demanded elts by the subvector index.
3567 SDValue Src = Op.getOperand(i: 0);
3568 // Bail until we can represent demanded elements for scalable vectors.
3569 if (Op.getValueType().isScalableVector() || Src.getValueType().isScalableVector())
3570 break;
3571 uint64_t Idx = Op.getConstantOperandVal(i: 1);
3572 unsigned NumSrcElts = Src.getValueType().getVectorNumElements();
3573 APInt DemandedSrcElts = DemandedElts.zext(width: NumSrcElts).shl(shiftAmt: Idx);
3574 Known = computeKnownBits(Op: Src, DemandedElts: DemandedSrcElts, Depth: Depth + 1);
3575 break;
3576 }
3577 case ISD::SCALAR_TO_VECTOR: {
3578 if (Op.getValueType().isScalableVector())
3579 break;
3580 // We know about scalar_to_vector as much as we know about it source,
3581 // which becomes the first element of otherwise unknown vector.
3582 if (DemandedElts != 1)
3583 break;
3584
3585 SDValue N0 = Op.getOperand(i: 0);
3586 Known = computeKnownBits(Op: N0, Depth: Depth + 1);
3587 if (N0.getValueSizeInBits() != BitWidth)
3588 Known = Known.trunc(BitWidth);
3589
3590 break;
3591 }
3592 case ISD::BITCAST: {
3593 if (Op.getValueType().isScalableVector())
3594 break;
3595
3596 SDValue N0 = Op.getOperand(i: 0);
3597 EVT SubVT = N0.getValueType();
3598 unsigned SubBitWidth = SubVT.getScalarSizeInBits();
3599
3600 // Ignore bitcasts from unsupported types.
3601 if (!(SubVT.isInteger() || SubVT.isFloatingPoint()))
3602 break;
3603
3604 // Fast handling of 'identity' bitcasts.
3605 if (BitWidth == SubBitWidth) {
3606 Known = computeKnownBits(Op: N0, DemandedElts, Depth: Depth + 1);
3607 break;
3608 }
3609
3610 bool IsLE = getDataLayout().isLittleEndian();
3611
3612 // Bitcast 'small element' vector to 'large element' scalar/vector.
3613 if ((BitWidth % SubBitWidth) == 0) {
3614 assert(N0.getValueType().isVector() && "Expected bitcast from vector");
3615
3616 // Collect known bits for the (larger) output by collecting the known
3617 // bits from each set of sub elements and shift these into place.
3618 // We need to separately call computeKnownBits for each set of
3619 // sub elements as the knownbits for each is likely to be different.
3620 unsigned SubScale = BitWidth / SubBitWidth;
3621 APInt SubDemandedElts(NumElts * SubScale, 0);
3622 for (unsigned i = 0; i != NumElts; ++i)
3623 if (DemandedElts[i])
3624 SubDemandedElts.setBit(i * SubScale);
3625
3626 for (unsigned i = 0; i != SubScale; ++i) {
3627 Known2 = computeKnownBits(Op: N0, DemandedElts: SubDemandedElts.shl(shiftAmt: i),
3628 Depth: Depth + 1);
3629 unsigned Shifts = IsLE ? i : SubScale - 1 - i;
3630 Known.insertBits(SubBits: Known2, BitPosition: SubBitWidth * Shifts);
3631 }
3632 }
3633
3634 // Bitcast 'large element' scalar/vector to 'small element' vector.
3635 if ((SubBitWidth % BitWidth) == 0) {
3636 assert(Op.getValueType().isVector() && "Expected bitcast to vector");
3637
3638 // Collect known bits for the (smaller) output by collecting the known
3639 // bits from the overlapping larger input elements and extracting the
3640 // sub sections we actually care about.
3641 unsigned SubScale = SubBitWidth / BitWidth;
3642 APInt SubDemandedElts =
3643 APIntOps::ScaleBitMask(A: DemandedElts, NewBitWidth: NumElts / SubScale);
3644 Known2 = computeKnownBits(Op: N0, DemandedElts: SubDemandedElts, Depth: Depth + 1);
3645
3646 Known.Zero.setAllBits(); Known.One.setAllBits();
3647 for (unsigned i = 0; i != NumElts; ++i)
3648 if (DemandedElts[i]) {
3649 unsigned Shifts = IsLE ? i : NumElts - 1 - i;
3650 unsigned Offset = (Shifts % SubScale) * BitWidth;
3651 Known = Known.intersectWith(RHS: Known2.extractBits(NumBits: BitWidth, BitPosition: Offset));
3652 // If we don't know any bits, early out.
3653 if (Known.isUnknown())
3654 break;
3655 }
3656 }
3657 break;
3658 }
3659 case ISD::AND:
3660 Known = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
3661 Known2 = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3662
3663 Known &= Known2;
3664 break;
3665 case ISD::OR:
3666 Known = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
3667 Known2 = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3668
3669 Known |= Known2;
3670 break;
3671 case ISD::XOR:
3672 Known = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
3673 Known2 = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3674
3675 Known ^= Known2;
3676 break;
3677 case ISD::MUL: {
3678 Known = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
3679 Known2 = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3680 bool SelfMultiply = Op.getOperand(i: 0) == Op.getOperand(i: 1);
3681 // TODO: SelfMultiply can be poison, but not undef.
3682 if (SelfMultiply)
3683 SelfMultiply &= isGuaranteedNotToBeUndefOrPoison(
3684 Op: Op.getOperand(i: 0), DemandedElts, PoisonOnly: false, Depth: Depth + 1);
3685 Known = KnownBits::mul(LHS: Known, RHS: Known2, NoUndefSelfMultiply: SelfMultiply);
3686
3687 // If the multiplication is known not to overflow, the product of a number
3688 // with itself is non-negative. Only do this if we didn't already computed
3689 // the opposite value for the sign bit.
3690 if (Op->getFlags().hasNoSignedWrap() &&
3691 Op.getOperand(i: 0) == Op.getOperand(i: 1) &&
3692 !Known.isNegative())
3693 Known.makeNonNegative();
3694 break;
3695 }
3696 case ISD::MULHU: {
3697 Known = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
3698 Known2 = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3699 Known = KnownBits::mulhu(LHS: Known, RHS: Known2);
3700 break;
3701 }
3702 case ISD::MULHS: {
3703 Known = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
3704 Known2 = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3705 Known = KnownBits::mulhs(LHS: Known, RHS: Known2);
3706 break;
3707 }
3708 case ISD::ABDU: {
3709 Known = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
3710 Known2 = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3711 Known = KnownBits::abdu(LHS: Known, RHS: Known2);
3712 break;
3713 }
3714 case ISD::ABDS: {
3715 Known = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
3716 Known2 = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3717 Known = KnownBits::abds(LHS: Known, RHS: Known2);
3718 unsigned SignBits1 =
3719 ComputeNumSignBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
3720 if (SignBits1 == 1)
3721 break;
3722 unsigned SignBits0 =
3723 ComputeNumSignBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3724 Known.Zero.setHighBits(std::min(a: SignBits0, b: SignBits1) - 1);
3725 break;
3726 }
3727 case ISD::UMUL_LOHI: {
3728 assert((Op.getResNo() == 0 || Op.getResNo() == 1) && "Unknown result");
3729 Known = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
3730 Known2 = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3731 bool SelfMultiply = Op.getOperand(i: 0) == Op.getOperand(i: 1);
3732 if (Op.getResNo() == 0)
3733 Known = KnownBits::mul(LHS: Known, RHS: Known2, NoUndefSelfMultiply: SelfMultiply);
3734 else
3735 Known = KnownBits::mulhu(LHS: Known, RHS: Known2);
3736 break;
3737 }
3738 case ISD::SMUL_LOHI: {
3739 assert((Op.getResNo() == 0 || Op.getResNo() == 1) && "Unknown result");
3740 Known = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
3741 Known2 = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3742 bool SelfMultiply = Op.getOperand(i: 0) == Op.getOperand(i: 1);
3743 if (Op.getResNo() == 0)
3744 Known = KnownBits::mul(LHS: Known, RHS: Known2, NoUndefSelfMultiply: SelfMultiply);
3745 else
3746 Known = KnownBits::mulhs(LHS: Known, RHS: Known2);
3747 break;
3748 }
3749 case ISD::AVGFLOORU: {
3750 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3751 Known2 = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
3752 Known = KnownBits::avgFloorU(LHS: Known, RHS: Known2);
3753 break;
3754 }
3755 case ISD::AVGCEILU: {
3756 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3757 Known2 = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
3758 Known = KnownBits::avgCeilU(LHS: Known, RHS: Known2);
3759 break;
3760 }
3761 case ISD::AVGFLOORS: {
3762 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3763 Known2 = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
3764 Known = KnownBits::avgFloorS(LHS: Known, RHS: Known2);
3765 break;
3766 }
3767 case ISD::AVGCEILS: {
3768 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3769 Known2 = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
3770 Known = KnownBits::avgCeilS(LHS: Known, RHS: Known2);
3771 break;
3772 }
3773 case ISD::SELECT:
3774 case ISD::VSELECT:
3775 Known = computeKnownBits(Op: Op.getOperand(i: 2), DemandedElts, Depth: Depth+1);
3776 // If we don't know any bits, early out.
3777 if (Known.isUnknown())
3778 break;
3779 Known2 = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth+1);
3780
3781 // Only known if known in both the LHS and RHS.
3782 Known = Known.intersectWith(RHS: Known2);
3783 break;
3784 case ISD::SELECT_CC:
3785 Known = computeKnownBits(Op: Op.getOperand(i: 3), DemandedElts, Depth: Depth+1);
3786 // If we don't know any bits, early out.
3787 if (Known.isUnknown())
3788 break;
3789 Known2 = computeKnownBits(Op: Op.getOperand(i: 2), DemandedElts, Depth: Depth+1);
3790
3791 // Only known if known in both the LHS and RHS.
3792 Known = Known.intersectWith(RHS: Known2);
3793 break;
3794 case ISD::SMULO:
3795 case ISD::UMULO:
3796 if (Op.getResNo() != 1)
3797 break;
3798 // The boolean result conforms to getBooleanContents.
3799 // If we know the result of a setcc has the top bits zero, use this info.
3800 // We know that we have an integer-based boolean since these operations
3801 // are only available for integer.
3802 if (TLI->getBooleanContents(isVec: Op.getValueType().isVector(), isFloat: false) ==
3803 TargetLowering::ZeroOrOneBooleanContent &&
3804 BitWidth > 1)
3805 Known.Zero.setBitsFrom(1);
3806 break;
3807 case ISD::SETCC:
3808 case ISD::SETCCCARRY:
3809 case ISD::STRICT_FSETCC:
3810 case ISD::STRICT_FSETCCS: {
3811 unsigned OpNo = Op->isStrictFPOpcode() ? 1 : 0;
3812 // If we know the result of a setcc has the top bits zero, use this info.
3813 if (TLI->getBooleanContents(Type: Op.getOperand(i: OpNo).getValueType()) ==
3814 TargetLowering::ZeroOrOneBooleanContent &&
3815 BitWidth > 1)
3816 Known.Zero.setBitsFrom(1);
3817 break;
3818 }
3819 case ISD::SHL: {
3820 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3821 Known2 = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
3822
3823 bool NUW = Op->getFlags().hasNoUnsignedWrap();
3824 bool NSW = Op->getFlags().hasNoSignedWrap();
3825
3826 bool ShAmtNonZero = Known2.isNonZero();
3827
3828 Known = KnownBits::shl(LHS: Known, RHS: Known2, NUW, NSW, ShAmtNonZero);
3829
3830 // Minimum shift low bits are known zero.
3831 if (std::optional<uint64_t> ShMinAmt =
3832 getValidMinimumShiftAmount(V: Op, DemandedElts, Depth: Depth + 1))
3833 Known.Zero.setLowBits(*ShMinAmt);
3834 break;
3835 }
3836 case ISD::SRL:
3837 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3838 Known2 = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
3839 Known = KnownBits::lshr(LHS: Known, RHS: Known2, /*ShAmtNonZero=*/false,
3840 Exact: Op->getFlags().hasExact());
3841
3842 // Minimum shift high bits are known zero.
3843 if (std::optional<uint64_t> ShMinAmt =
3844 getValidMinimumShiftAmount(V: Op, DemandedElts, Depth: Depth + 1))
3845 Known.Zero.setHighBits(*ShMinAmt);
3846 break;
3847 case ISD::SRA:
3848 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3849 Known2 = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
3850 Known = KnownBits::ashr(LHS: Known, RHS: Known2, /*ShAmtNonZero=*/false,
3851 Exact: Op->getFlags().hasExact());
3852 break;
3853 case ISD::FSHL:
3854 case ISD::FSHR:
3855 if (ConstantSDNode *C = isConstOrConstSplat(N: Op.getOperand(i: 2), DemandedElts)) {
3856 unsigned Amt = C->getAPIntValue().urem(RHS: BitWidth);
3857
3858 // For fshl, 0-shift returns the 1st arg.
3859 // For fshr, 0-shift returns the 2nd arg.
3860 if (Amt == 0) {
3861 Known = computeKnownBits(Op: Op.getOperand(i: Opcode == ISD::FSHL ? 0 : 1),
3862 DemandedElts, Depth: Depth + 1);
3863 break;
3864 }
3865
3866 // fshl: (X << (Z % BW)) | (Y >> (BW - (Z % BW)))
3867 // fshr: (X << (BW - (Z % BW))) | (Y >> (Z % BW))
3868 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3869 Known2 = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
3870 if (Opcode == ISD::FSHL) {
3871 Known.One <<= Amt;
3872 Known.Zero <<= Amt;
3873 Known2.One.lshrInPlace(ShiftAmt: BitWidth - Amt);
3874 Known2.Zero.lshrInPlace(ShiftAmt: BitWidth - Amt);
3875 } else {
3876 Known.One <<= BitWidth - Amt;
3877 Known.Zero <<= BitWidth - Amt;
3878 Known2.One.lshrInPlace(ShiftAmt: Amt);
3879 Known2.Zero.lshrInPlace(ShiftAmt: Amt);
3880 }
3881 Known = Known.unionWith(RHS: Known2);
3882 }
3883 break;
3884 case ISD::SHL_PARTS:
3885 case ISD::SRA_PARTS:
3886 case ISD::SRL_PARTS: {
3887 assert((Op.getResNo() == 0 || Op.getResNo() == 1) && "Unknown result");
3888
3889 // Collect lo/hi source values and concatenate.
3890 unsigned LoBits = Op.getOperand(i: 0).getScalarValueSizeInBits();
3891 unsigned HiBits = Op.getOperand(i: 1).getScalarValueSizeInBits();
3892 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3893 Known2 = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
3894 Known = Known2.concat(Lo: Known);
3895
3896 // Collect shift amount.
3897 Known2 = computeKnownBits(Op: Op.getOperand(i: 2), DemandedElts, Depth: Depth + 1);
3898
3899 if (Opcode == ISD::SHL_PARTS)
3900 Known = KnownBits::shl(LHS: Known, RHS: Known2);
3901 else if (Opcode == ISD::SRA_PARTS)
3902 Known = KnownBits::ashr(LHS: Known, RHS: Known2);
3903 else // if (Opcode == ISD::SRL_PARTS)
3904 Known = KnownBits::lshr(LHS: Known, RHS: Known2);
3905
3906 // TODO: Minimum shift low/high bits are known zero.
3907
3908 if (Op.getResNo() == 0)
3909 Known = Known.extractBits(NumBits: LoBits, BitPosition: 0);
3910 else
3911 Known = Known.extractBits(NumBits: HiBits, BitPosition: LoBits);
3912 break;
3913 }
3914 case ISD::SIGN_EXTEND_INREG: {
3915 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3916 EVT EVT = cast<VTSDNode>(Val: Op.getOperand(i: 1))->getVT();
3917 Known = Known.sextInReg(SrcBitWidth: EVT.getScalarSizeInBits());
3918 break;
3919 }
3920 case ISD::CTTZ:
3921 case ISD::CTTZ_ZERO_UNDEF: {
3922 Known2 = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3923 // If we have a known 1, its position is our upper bound.
3924 unsigned PossibleTZ = Known2.countMaxTrailingZeros();
3925 unsigned LowBits = llvm::bit_width(Value: PossibleTZ);
3926 Known.Zero.setBitsFrom(LowBits);
3927 break;
3928 }
3929 case ISD::CTLZ:
3930 case ISD::CTLZ_ZERO_UNDEF: {
3931 Known2 = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3932 // If we have a known 1, its position is our upper bound.
3933 unsigned PossibleLZ = Known2.countMaxLeadingZeros();
3934 unsigned LowBits = llvm::bit_width(Value: PossibleLZ);
3935 Known.Zero.setBitsFrom(LowBits);
3936 break;
3937 }
3938 case ISD::CTPOP: {
3939 Known2 = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
3940 // If we know some of the bits are zero, they can't be one.
3941 unsigned PossibleOnes = Known2.countMaxPopulation();
3942 Known.Zero.setBitsFrom(llvm::bit_width(Value: PossibleOnes));
3943 break;
3944 }
3945 case ISD::PARITY: {
3946 // Parity returns 0 everywhere but the LSB.
3947 Known.Zero.setBitsFrom(1);
3948 break;
3949 }
3950 case ISD::MGATHER:
3951 case ISD::MLOAD: {
3952 ISD::LoadExtType ETy =
3953 (Opcode == ISD::MGATHER)
3954 ? cast<MaskedGatherSDNode>(Val&: Op)->getExtensionType()
3955 : cast<MaskedLoadSDNode>(Val&: Op)->getExtensionType();
3956 if (ETy == ISD::ZEXTLOAD) {
3957 EVT MemVT = cast<MemSDNode>(Val&: Op)->getMemoryVT();
3958 KnownBits Known0(MemVT.getScalarSizeInBits());
3959 return Known0.zext(BitWidth);
3960 }
3961 break;
3962 }
3963 case ISD::LOAD: {
3964 LoadSDNode *LD = cast<LoadSDNode>(Val&: Op);
3965 const Constant *Cst = TLI->getTargetConstantFromLoad(LD);
3966 if (ISD::isNON_EXTLoad(N: LD) && Cst) {
3967 // Determine any common known bits from the loaded constant pool value.
3968 Type *CstTy = Cst->getType();
3969 if ((NumElts * BitWidth) == CstTy->getPrimitiveSizeInBits() &&
3970 !Op.getValueType().isScalableVector()) {
3971 // If its a vector splat, then we can (quickly) reuse the scalar path.
3972 // NOTE: We assume all elements match and none are UNDEF.
3973 if (CstTy->isVectorTy()) {
3974 if (const Constant *Splat = Cst->getSplatValue()) {
3975 Cst = Splat;
3976 CstTy = Cst->getType();
3977 }
3978 }
3979 // TODO - do we need to handle different bitwidths?
3980 if (CstTy->isVectorTy() && BitWidth == CstTy->getScalarSizeInBits()) {
3981 // Iterate across all vector elements finding common known bits.
3982 Known.One.setAllBits();
3983 Known.Zero.setAllBits();
3984 for (unsigned i = 0; i != NumElts; ++i) {
3985 if (!DemandedElts[i])
3986 continue;
3987 if (Constant *Elt = Cst->getAggregateElement(Elt: i)) {
3988 if (auto *CInt = dyn_cast<ConstantInt>(Val: Elt)) {
3989 const APInt &Value = CInt->getValue();
3990 Known.One &= Value;
3991 Known.Zero &= ~Value;
3992 continue;
3993 }
3994 if (auto *CFP = dyn_cast<ConstantFP>(Val: Elt)) {
3995 APInt Value = CFP->getValueAPF().bitcastToAPInt();
3996 Known.One &= Value;
3997 Known.Zero &= ~Value;
3998 continue;
3999 }
4000 }
4001 Known.One.clearAllBits();
4002 Known.Zero.clearAllBits();
4003 break;
4004 }
4005 } else if (BitWidth == CstTy->getPrimitiveSizeInBits()) {
4006 if (auto *CInt = dyn_cast<ConstantInt>(Val: Cst)) {
4007 Known = KnownBits::makeConstant(C: CInt->getValue());
4008 } else if (auto *CFP = dyn_cast<ConstantFP>(Val: Cst)) {
4009 Known =
4010 KnownBits::makeConstant(C: CFP->getValueAPF().bitcastToAPInt());
4011 }
4012 }
4013 }
4014 } else if (Op.getResNo() == 0) {
4015 unsigned ScalarMemorySize = LD->getMemoryVT().getScalarSizeInBits();
4016 KnownBits KnownScalarMemory(ScalarMemorySize);
4017 if (const MDNode *MD = LD->getRanges())
4018 computeKnownBitsFromRangeMetadata(Ranges: *MD, Known&: KnownScalarMemory);
4019
4020 // Extend the Known bits from memory to the size of the scalar result.
4021 if (ISD::isZEXTLoad(N: Op.getNode()))
4022 Known = KnownScalarMemory.zext(BitWidth);
4023 else if (ISD::isSEXTLoad(N: Op.getNode()))
4024 Known = KnownScalarMemory.sext(BitWidth);
4025 else if (ISD::isEXTLoad(N: Op.getNode()))
4026 Known = KnownScalarMemory.anyext(BitWidth);
4027 else
4028 Known = KnownScalarMemory;
4029 assert(Known.getBitWidth() == BitWidth);
4030 return Known;
4031 }
4032 break;
4033 }
4034 case ISD::ZERO_EXTEND_VECTOR_INREG: {
4035 if (Op.getValueType().isScalableVector())
4036 break;
4037 EVT InVT = Op.getOperand(i: 0).getValueType();
4038 APInt InDemandedElts = DemandedElts.zext(width: InVT.getVectorNumElements());
4039 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts: InDemandedElts, Depth: Depth + 1);
4040 Known = Known.zext(BitWidth);
4041 break;
4042 }
4043 case ISD::ZERO_EXTEND: {
4044 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4045 Known = Known.zext(BitWidth);
4046 break;
4047 }
4048 case ISD::SIGN_EXTEND_VECTOR_INREG: {
4049 if (Op.getValueType().isScalableVector())
4050 break;
4051 EVT InVT = Op.getOperand(i: 0).getValueType();
4052 APInt InDemandedElts = DemandedElts.zext(width: InVT.getVectorNumElements());
4053 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts: InDemandedElts, Depth: Depth + 1);
4054 // If the sign bit is known to be zero or one, then sext will extend
4055 // it to the top bits, else it will just zext.
4056 Known = Known.sext(BitWidth);
4057 break;
4058 }
4059 case ISD::SIGN_EXTEND: {
4060 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4061 // If the sign bit is known to be zero or one, then sext will extend
4062 // it to the top bits, else it will just zext.
4063 Known = Known.sext(BitWidth);
4064 break;
4065 }
4066 case ISD::ANY_EXTEND_VECTOR_INREG: {
4067 if (Op.getValueType().isScalableVector())
4068 break;
4069 EVT InVT = Op.getOperand(i: 0).getValueType();
4070 APInt InDemandedElts = DemandedElts.zext(width: InVT.getVectorNumElements());
4071 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts: InDemandedElts, Depth: Depth + 1);
4072 Known = Known.anyext(BitWidth);
4073 break;
4074 }
4075 case ISD::ANY_EXTEND: {
4076 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4077 Known = Known.anyext(BitWidth);
4078 break;
4079 }
4080 case ISD::TRUNCATE: {
4081 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4082 Known = Known.trunc(BitWidth);
4083 break;
4084 }
4085 case ISD::AssertZext: {
4086 EVT VT = cast<VTSDNode>(Val: Op.getOperand(i: 1))->getVT();
4087 APInt InMask = APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: VT.getSizeInBits());
4088 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4089 Known.Zero |= (~InMask);
4090 Known.One &= (~Known.Zero);
4091 break;
4092 }
4093 case ISD::AssertAlign: {
4094 unsigned LogOfAlign = Log2(A: cast<AssertAlignSDNode>(Val&: Op)->getAlign());
4095 assert(LogOfAlign != 0);
4096
4097 // TODO: Should use maximum with source
4098 // If a node is guaranteed to be aligned, set low zero bits accordingly as
4099 // well as clearing one bits.
4100 Known.Zero.setLowBits(LogOfAlign);
4101 Known.One.clearLowBits(loBits: LogOfAlign);
4102 break;
4103 }
4104 case ISD::FGETSIGN:
4105 // All bits are zero except the low bit.
4106 Known.Zero.setBitsFrom(1);
4107 break;
4108 case ISD::ADD:
4109 case ISD::SUB: {
4110 SDNodeFlags Flags = Op.getNode()->getFlags();
4111 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4112 Known2 = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
4113 Known = KnownBits::computeForAddSub(
4114 Add: Op.getOpcode() == ISD::ADD, NSW: Flags.hasNoSignedWrap(),
4115 NUW: Flags.hasNoUnsignedWrap(), LHS: Known, RHS: Known2);
4116 break;
4117 }
4118 case ISD::USUBO:
4119 case ISD::SSUBO:
4120 case ISD::USUBO_CARRY:
4121 case ISD::SSUBO_CARRY:
4122 if (Op.getResNo() == 1) {
4123 // If we know the result of a setcc has the top bits zero, use this info.
4124 if (TLI->getBooleanContents(Type: Op.getOperand(i: 0).getValueType()) ==
4125 TargetLowering::ZeroOrOneBooleanContent &&
4126 BitWidth > 1)
4127 Known.Zero.setBitsFrom(1);
4128 break;
4129 }
4130 [[fallthrough]];
4131 case ISD::SUBC: {
4132 assert(Op.getResNo() == 0 &&
4133 "We only compute knownbits for the difference here.");
4134
4135 // With USUBO_CARRY and SSUBO_CARRY a borrow bit may be added in.
4136 KnownBits Borrow(1);
4137 if (Opcode == ISD::USUBO_CARRY || Opcode == ISD::SSUBO_CARRY) {
4138 Borrow = computeKnownBits(Op: Op.getOperand(i: 2), DemandedElts, Depth: Depth + 1);
4139 // Borrow has bit width 1
4140 Borrow = Borrow.trunc(BitWidth: 1);
4141 } else {
4142 Borrow.setAllZero();
4143 }
4144
4145 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4146 Known2 = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
4147 Known = KnownBits::computeForSubBorrow(LHS: Known, RHS: Known2, Borrow);
4148 break;
4149 }
4150 case ISD::UADDO:
4151 case ISD::SADDO:
4152 case ISD::UADDO_CARRY:
4153 case ISD::SADDO_CARRY:
4154 if (Op.getResNo() == 1) {
4155 // If we know the result of a setcc has the top bits zero, use this info.
4156 if (TLI->getBooleanContents(Type: Op.getOperand(i: 0).getValueType()) ==
4157 TargetLowering::ZeroOrOneBooleanContent &&
4158 BitWidth > 1)
4159 Known.Zero.setBitsFrom(1);
4160 break;
4161 }
4162 [[fallthrough]];
4163 case ISD::ADDC:
4164 case ISD::ADDE: {
4165 assert(Op.getResNo() == 0 && "We only compute knownbits for the sum here.");
4166
4167 // With ADDE and UADDO_CARRY, a carry bit may be added in.
4168 KnownBits Carry(1);
4169 if (Opcode == ISD::ADDE)
4170 // Can't track carry from glue, set carry to unknown.
4171 Carry.resetAll();
4172 else if (Opcode == ISD::UADDO_CARRY || Opcode == ISD::SADDO_CARRY) {
4173 Carry = computeKnownBits(Op: Op.getOperand(i: 2), DemandedElts, Depth: Depth + 1);
4174 // Carry has bit width 1
4175 Carry = Carry.trunc(BitWidth: 1);
4176 } else {
4177 Carry.setAllZero();
4178 }
4179
4180 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4181 Known2 = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
4182 Known = KnownBits::computeForAddCarry(LHS: Known, RHS: Known2, Carry);
4183 break;
4184 }
4185 case ISD::UDIV: {
4186 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4187 Known2 = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
4188 Known = KnownBits::udiv(LHS: Known, RHS: Known2, Exact: Op->getFlags().hasExact());
4189 break;
4190 }
4191 case ISD::SDIV: {
4192 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4193 Known2 = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
4194 Known = KnownBits::sdiv(LHS: Known, RHS: Known2, Exact: Op->getFlags().hasExact());
4195 break;
4196 }
4197 case ISD::SREM: {
4198 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4199 Known2 = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
4200 Known = KnownBits::srem(LHS: Known, RHS: Known2);
4201 break;
4202 }
4203 case ISD::UREM: {
4204 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4205 Known2 = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
4206 Known = KnownBits::urem(LHS: Known, RHS: Known2);
4207 break;
4208 }
4209 case ISD::EXTRACT_ELEMENT: {
4210 Known = computeKnownBits(Op: Op.getOperand(i: 0), Depth: Depth+1);
4211 const unsigned Index = Op.getConstantOperandVal(i: 1);
4212 const unsigned EltBitWidth = Op.getValueSizeInBits();
4213
4214 // Remove low part of known bits mask
4215 Known.Zero = Known.Zero.getHiBits(numBits: Known.getBitWidth() - Index * EltBitWidth);
4216 Known.One = Known.One.getHiBits(numBits: Known.getBitWidth() - Index * EltBitWidth);
4217
4218 // Remove high part of known bit mask
4219 Known = Known.trunc(BitWidth: EltBitWidth);
4220 break;
4221 }
4222 case ISD::EXTRACT_VECTOR_ELT: {
4223 SDValue InVec = Op.getOperand(i: 0);
4224 SDValue EltNo = Op.getOperand(i: 1);
4225 EVT VecVT = InVec.getValueType();
4226 // computeKnownBits not yet implemented for scalable vectors.
4227 if (VecVT.isScalableVector())
4228 break;
4229 const unsigned EltBitWidth = VecVT.getScalarSizeInBits();
4230 const unsigned NumSrcElts = VecVT.getVectorNumElements();
4231
4232 // If BitWidth > EltBitWidth the value is anyext:ed. So we do not know
4233 // anything about the extended bits.
4234 if (BitWidth > EltBitWidth)
4235 Known = Known.trunc(BitWidth: EltBitWidth);
4236
4237 // If we know the element index, just demand that vector element, else for
4238 // an unknown element index, ignore DemandedElts and demand them all.
4239 APInt DemandedSrcElts = APInt::getAllOnes(numBits: NumSrcElts);
4240 auto *ConstEltNo = dyn_cast<ConstantSDNode>(Val&: EltNo);
4241 if (ConstEltNo && ConstEltNo->getAPIntValue().ult(RHS: NumSrcElts))
4242 DemandedSrcElts =
4243 APInt::getOneBitSet(numBits: NumSrcElts, BitNo: ConstEltNo->getZExtValue());
4244
4245 Known = computeKnownBits(Op: InVec, DemandedElts: DemandedSrcElts, Depth: Depth + 1);
4246 if (BitWidth > EltBitWidth)
4247 Known = Known.anyext(BitWidth);
4248 break;
4249 }
4250 case ISD::INSERT_VECTOR_ELT: {
4251 if (Op.getValueType().isScalableVector())
4252 break;
4253
4254 // If we know the element index, split the demand between the
4255 // source vector and the inserted element, otherwise assume we need
4256 // the original demanded vector elements and the value.
4257 SDValue InVec = Op.getOperand(i: 0);
4258 SDValue InVal = Op.getOperand(i: 1);
4259 SDValue EltNo = Op.getOperand(i: 2);
4260 bool DemandedVal = true;
4261 APInt DemandedVecElts = DemandedElts;
4262 auto *CEltNo = dyn_cast<ConstantSDNode>(Val&: EltNo);
4263 if (CEltNo && CEltNo->getAPIntValue().ult(RHS: NumElts)) {
4264 unsigned EltIdx = CEltNo->getZExtValue();
4265 DemandedVal = !!DemandedElts[EltIdx];
4266 DemandedVecElts.clearBit(BitPosition: EltIdx);
4267 }
4268 Known.One.setAllBits();
4269 Known.Zero.setAllBits();
4270 if (DemandedVal) {
4271 Known2 = computeKnownBits(Op: InVal, Depth: Depth + 1);
4272 Known = Known.intersectWith(RHS: Known2.zextOrTrunc(BitWidth));
4273 }
4274 if (!!DemandedVecElts) {
4275 Known2 = computeKnownBits(Op: InVec, DemandedElts: DemandedVecElts, Depth: Depth + 1);
4276 Known = Known.intersectWith(RHS: Known2);
4277 }
4278 break;
4279 }
4280 case ISD::BITREVERSE: {
4281 Known2 = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4282 Known = Known2.reverseBits();
4283 break;
4284 }
4285 case ISD::BSWAP: {
4286 Known2 = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4287 Known = Known2.byteSwap();
4288 break;
4289 }
4290 case ISD::ABS: {
4291 Known2 = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4292 Known = Known2.abs();
4293 Known.Zero.setHighBits(
4294 ComputeNumSignBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1) - 1);
4295 break;
4296 }
4297 case ISD::USUBSAT: {
4298 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4299 Known2 = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
4300 Known = KnownBits::usub_sat(LHS: Known, RHS: Known2);
4301 break;
4302 }
4303 case ISD::UMIN: {
4304 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4305 Known2 = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
4306 Known = KnownBits::umin(LHS: Known, RHS: Known2);
4307 break;
4308 }
4309 case ISD::UMAX: {
4310 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4311 Known2 = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
4312 Known = KnownBits::umax(LHS: Known, RHS: Known2);
4313 break;
4314 }
4315 case ISD::SMIN:
4316 case ISD::SMAX: {
4317 // If we have a clamp pattern, we know that the number of sign bits will be
4318 // the minimum of the clamp min/max range.
4319 bool IsMax = (Opcode == ISD::SMAX);
4320 ConstantSDNode *CstLow = nullptr, *CstHigh = nullptr;
4321 if ((CstLow = isConstOrConstSplat(N: Op.getOperand(i: 1), DemandedElts)))
4322 if (Op.getOperand(i: 0).getOpcode() == (IsMax ? ISD::SMIN : ISD::SMAX))
4323 CstHigh =
4324 isConstOrConstSplat(N: Op.getOperand(i: 0).getOperand(i: 1), DemandedElts);
4325 if (CstLow && CstHigh) {
4326 if (!IsMax)
4327 std::swap(a&: CstLow, b&: CstHigh);
4328
4329 const APInt &ValueLow = CstLow->getAPIntValue();
4330 const APInt &ValueHigh = CstHigh->getAPIntValue();
4331 if (ValueLow.sle(RHS: ValueHigh)) {
4332 unsigned LowSignBits = ValueLow.getNumSignBits();
4333 unsigned HighSignBits = ValueHigh.getNumSignBits();
4334 unsigned MinSignBits = std::min(a: LowSignBits, b: HighSignBits);
4335 if (ValueLow.isNegative() && ValueHigh.isNegative()) {
4336 Known.One.setHighBits(MinSignBits);
4337 break;
4338 }
4339 if (ValueLow.isNonNegative() && ValueHigh.isNonNegative()) {
4340 Known.Zero.setHighBits(MinSignBits);
4341 break;
4342 }
4343 }
4344 }
4345
4346 Known = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4347 Known2 = computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
4348 if (IsMax)
4349 Known = KnownBits::smax(LHS: Known, RHS: Known2);
4350 else
4351 Known = KnownBits::smin(LHS: Known, RHS: Known2);
4352
4353 // For SMAX, if CstLow is non-negative we know the result will be
4354 // non-negative and thus all sign bits are 0.
4355 // TODO: There's an equivalent of this for smin with negative constant for
4356 // known ones.
4357 if (IsMax && CstLow) {
4358 const APInt &ValueLow = CstLow->getAPIntValue();
4359 if (ValueLow.isNonNegative()) {
4360 unsigned SignBits = ComputeNumSignBits(Op: Op.getOperand(i: 0), Depth: Depth + 1);
4361 Known.Zero.setHighBits(std::min(a: SignBits, b: ValueLow.getNumSignBits()));
4362 }
4363 }
4364
4365 break;
4366 }
4367 case ISD::UINT_TO_FP: {
4368 Known.makeNonNegative();
4369 break;
4370 }
4371 case ISD::SINT_TO_FP: {
4372 Known2 = computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4373 if (Known2.isNonNegative())
4374 Known.makeNonNegative();
4375 else if (Known2.isNegative())
4376 Known.makeNegative();
4377 break;
4378 }
4379 case ISD::FP_TO_UINT_SAT: {
4380 // FP_TO_UINT_SAT produces an unsigned value that fits in the saturating VT.
4381 EVT VT = cast<VTSDNode>(Val: Op.getOperand(i: 1))->getVT();
4382 Known.Zero |= APInt::getBitsSetFrom(numBits: BitWidth, loBit: VT.getScalarSizeInBits());
4383 break;
4384 }
4385 case ISD::ATOMIC_LOAD: {
4386 // If we are looking at the loaded value.
4387 if (Op.getResNo() == 0) {
4388 auto *AT = cast<AtomicSDNode>(Val&: Op);
4389 unsigned ScalarMemorySize = AT->getMemoryVT().getScalarSizeInBits();
4390 KnownBits KnownScalarMemory(ScalarMemorySize);
4391 if (const MDNode *MD = AT->getRanges())
4392 computeKnownBitsFromRangeMetadata(Ranges: *MD, Known&: KnownScalarMemory);
4393
4394 switch (AT->getExtensionType()) {
4395 case ISD::ZEXTLOAD:
4396 Known = KnownScalarMemory.zext(BitWidth);
4397 break;
4398 case ISD::SEXTLOAD:
4399 Known = KnownScalarMemory.sext(BitWidth);
4400 break;
4401 case ISD::EXTLOAD:
4402 switch (TLI->getExtendForAtomicOps()) {
4403 case ISD::ZERO_EXTEND:
4404 Known = KnownScalarMemory.zext(BitWidth);
4405 break;
4406 case ISD::SIGN_EXTEND:
4407 Known = KnownScalarMemory.sext(BitWidth);
4408 break;
4409 default:
4410 Known = KnownScalarMemory.anyext(BitWidth);
4411 break;
4412 }
4413 break;
4414 case ISD::NON_EXTLOAD:
4415 Known = KnownScalarMemory;
4416 break;
4417 }
4418 assert(Known.getBitWidth() == BitWidth);
4419 }
4420 break;
4421 }
4422 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
4423 if (Op.getResNo() == 1) {
4424 // The boolean result conforms to getBooleanContents.
4425 // If we know the result of a setcc has the top bits zero, use this info.
4426 // We know that we have an integer-based boolean since these operations
4427 // are only available for integer.
4428 if (TLI->getBooleanContents(isVec: Op.getValueType().isVector(), isFloat: false) ==
4429 TargetLowering::ZeroOrOneBooleanContent &&
4430 BitWidth > 1)
4431 Known.Zero.setBitsFrom(1);
4432 break;
4433 }
4434 [[fallthrough]];
4435 case ISD::ATOMIC_CMP_SWAP:
4436 case ISD::ATOMIC_SWAP:
4437 case ISD::ATOMIC_LOAD_ADD:
4438 case ISD::ATOMIC_LOAD_SUB:
4439 case ISD::ATOMIC_LOAD_AND:
4440 case ISD::ATOMIC_LOAD_CLR:
4441 case ISD::ATOMIC_LOAD_OR:
4442 case ISD::ATOMIC_LOAD_XOR:
4443 case ISD::ATOMIC_LOAD_NAND:
4444 case ISD::ATOMIC_LOAD_MIN:
4445 case ISD::ATOMIC_LOAD_MAX:
4446 case ISD::ATOMIC_LOAD_UMIN:
4447 case ISD::ATOMIC_LOAD_UMAX: {
4448 // If we are looking at the loaded value.
4449 if (Op.getResNo() == 0) {
4450 auto *AT = cast<AtomicSDNode>(Val&: Op);
4451 unsigned MemBits = AT->getMemoryVT().getScalarSizeInBits();
4452
4453 if (TLI->getExtendForAtomicOps() == ISD::ZERO_EXTEND)
4454 Known.Zero.setBitsFrom(MemBits);
4455 }
4456 break;
4457 }
4458 case ISD::FrameIndex:
4459 case ISD::TargetFrameIndex:
4460 TLI->computeKnownBitsForFrameIndex(FIOp: cast<FrameIndexSDNode>(Val&: Op)->getIndex(),
4461 Known, MF: getMachineFunction());
4462 break;
4463
4464 default:
4465 if (Opcode < ISD::BUILTIN_OP_END)
4466 break;
4467 [[fallthrough]];
4468 case ISD::INTRINSIC_WO_CHAIN:
4469 case ISD::INTRINSIC_W_CHAIN:
4470 case ISD::INTRINSIC_VOID:
4471 // TODO: Probably okay to remove after audit; here to reduce change size
4472 // in initial enablement patch for scalable vectors
4473 if (Op.getValueType().isScalableVector())
4474 break;
4475
4476 // Allow the target to implement this method for its nodes.
4477 TLI->computeKnownBitsForTargetNode(Op, Known, DemandedElts, DAG: *this, Depth);
4478 break;
4479 }
4480
4481 return Known;
4482}
4483
4484/// Convert ConstantRange OverflowResult into SelectionDAG::OverflowKind.
4485static SelectionDAG::OverflowKind mapOverflowResult(ConstantRange::OverflowResult OR) {
4486 switch (OR) {
4487 case ConstantRange::OverflowResult::MayOverflow:
4488 return SelectionDAG::OFK_Sometime;
4489 case ConstantRange::OverflowResult::AlwaysOverflowsLow:
4490 case ConstantRange::OverflowResult::AlwaysOverflowsHigh:
4491 return SelectionDAG::OFK_Always;
4492 case ConstantRange::OverflowResult::NeverOverflows:
4493 return SelectionDAG::OFK_Never;
4494 }
4495 llvm_unreachable("Unknown OverflowResult");
4496}
4497
4498SelectionDAG::OverflowKind
4499SelectionDAG::computeOverflowForSignedAdd(SDValue N0, SDValue N1) const {
4500 // X + 0 never overflow
4501 if (isNullConstant(V: N1))
4502 return OFK_Never;
4503
4504 // If both operands each have at least two sign bits, the addition
4505 // cannot overflow.
4506 if (ComputeNumSignBits(Op: N0) > 1 && ComputeNumSignBits(Op: N1) > 1)
4507 return OFK_Never;
4508
4509 // TODO: Add ConstantRange::signedAddMayOverflow handling.
4510 return OFK_Sometime;
4511}
4512
4513SelectionDAG::OverflowKind
4514SelectionDAG::computeOverflowForUnsignedAdd(SDValue N0, SDValue N1) const {
4515 // X + 0 never overflow
4516 if (isNullConstant(V: N1))
4517 return OFK_Never;
4518
4519 // mulhi + 1 never overflow
4520 KnownBits N1Known = computeKnownBits(Op: N1);
4521 if (N0.getOpcode() == ISD::UMUL_LOHI && N0.getResNo() == 1 &&
4522 N1Known.getMaxValue().ult(RHS: 2))
4523 return OFK_Never;
4524
4525 KnownBits N0Known = computeKnownBits(Op: N0);
4526 if (N1.getOpcode() == ISD::UMUL_LOHI && N1.getResNo() == 1 &&
4527 N0Known.getMaxValue().ult(RHS: 2))
4528 return OFK_Never;
4529
4530 // Fallback to ConstantRange::unsignedAddMayOverflow handling.
4531 ConstantRange N0Range = ConstantRange::fromKnownBits(Known: N0Known, IsSigned: false);
4532 ConstantRange N1Range = ConstantRange::fromKnownBits(Known: N1Known, IsSigned: false);
4533 return mapOverflowResult(OR: N0Range.unsignedAddMayOverflow(Other: N1Range));
4534}
4535
4536SelectionDAG::OverflowKind
4537SelectionDAG::computeOverflowForSignedSub(SDValue N0, SDValue N1) const {
4538 // X - 0 never overflow
4539 if (isNullConstant(V: N1))
4540 return OFK_Never;
4541
4542 // If both operands each have at least two sign bits, the subtraction
4543 // cannot overflow.
4544 if (ComputeNumSignBits(Op: N0) > 1 && ComputeNumSignBits(Op: N1) > 1)
4545 return OFK_Never;
4546
4547 KnownBits N0Known = computeKnownBits(Op: N0);
4548 KnownBits N1Known = computeKnownBits(Op: N1);
4549 ConstantRange N0Range = ConstantRange::fromKnownBits(Known: N0Known, IsSigned: true);
4550 ConstantRange N1Range = ConstantRange::fromKnownBits(Known: N1Known, IsSigned: true);
4551 return mapOverflowResult(OR: N0Range.signedSubMayOverflow(Other: N1Range));
4552}
4553
4554SelectionDAG::OverflowKind
4555SelectionDAG::computeOverflowForUnsignedSub(SDValue N0, SDValue N1) const {
4556 // X - 0 never overflow
4557 if (isNullConstant(V: N1))
4558 return OFK_Never;
4559
4560 KnownBits N0Known = computeKnownBits(Op: N0);
4561 KnownBits N1Known = computeKnownBits(Op: N1);
4562 ConstantRange N0Range = ConstantRange::fromKnownBits(Known: N0Known, IsSigned: false);
4563 ConstantRange N1Range = ConstantRange::fromKnownBits(Known: N1Known, IsSigned: false);
4564 return mapOverflowResult(OR: N0Range.unsignedSubMayOverflow(Other: N1Range));
4565}
4566
4567SelectionDAG::OverflowKind
4568SelectionDAG::computeOverflowForUnsignedMul(SDValue N0, SDValue N1) const {
4569 // X * 0 and X * 1 never overflow.
4570 if (isNullConstant(V: N1) || isOneConstant(V: N1))
4571 return OFK_Never;
4572
4573 KnownBits N0Known = computeKnownBits(Op: N0);
4574 KnownBits N1Known = computeKnownBits(Op: N1);
4575 ConstantRange N0Range = ConstantRange::fromKnownBits(Known: N0Known, IsSigned: false);
4576 ConstantRange N1Range = ConstantRange::fromKnownBits(Known: N1Known, IsSigned: false);
4577 return mapOverflowResult(OR: N0Range.unsignedMulMayOverflow(Other: N1Range));
4578}
4579
4580SelectionDAG::OverflowKind
4581SelectionDAG::computeOverflowForSignedMul(SDValue N0, SDValue N1) const {
4582 // X * 0 and X * 1 never overflow.
4583 if (isNullConstant(V: N1) || isOneConstant(V: N1))
4584 return OFK_Never;
4585
4586 // Get the size of the result.
4587 unsigned BitWidth = N0.getScalarValueSizeInBits();
4588
4589 // Sum of the sign bits.
4590 unsigned SignBits = ComputeNumSignBits(Op: N0) + ComputeNumSignBits(Op: N1);
4591
4592 // If we have enough sign bits, then there's no overflow.
4593 if (SignBits > BitWidth + 1)
4594 return OFK_Never;
4595
4596 if (SignBits == BitWidth + 1) {
4597 // The overflow occurs when the true multiplication of the
4598 // the operands is the minimum negative number.
4599 KnownBits N0Known = computeKnownBits(Op: N0);
4600 KnownBits N1Known = computeKnownBits(Op: N1);
4601 // If one of the operands is non-negative, then there's no
4602 // overflow.
4603 if (N0Known.isNonNegative() || N1Known.isNonNegative())
4604 return OFK_Never;
4605 }
4606
4607 return OFK_Sometime;
4608}
4609
4610bool SelectionDAG::isKnownToBeAPowerOfTwo(SDValue Val, unsigned Depth) const {
4611 if (Depth >= MaxRecursionDepth)
4612 return false; // Limit search depth.
4613
4614 EVT OpVT = Val.getValueType();
4615 unsigned BitWidth = OpVT.getScalarSizeInBits();
4616
4617 // Is the constant a known power of 2?
4618 if (ISD::matchUnaryPredicate(Op: Val, Match: [BitWidth](ConstantSDNode *C) {
4619 return C->getAPIntValue().zextOrTrunc(width: BitWidth).isPowerOf2();
4620 }))
4621 return true;
4622
4623 // A left-shift of a constant one will have exactly one bit set because
4624 // shifting the bit off the end is undefined.
4625 if (Val.getOpcode() == ISD::SHL) {
4626 auto *C = isConstOrConstSplat(N: Val.getOperand(i: 0));
4627 if (C && C->getAPIntValue() == 1)
4628 return true;
4629 return isKnownToBeAPowerOfTwo(Val: Val.getOperand(i: 0), Depth: Depth + 1) &&
4630 isKnownNeverZero(Op: Val, Depth);
4631 }
4632
4633 // Similarly, a logical right-shift of a constant sign-bit will have exactly
4634 // one bit set.
4635 if (Val.getOpcode() == ISD::SRL) {
4636 auto *C = isConstOrConstSplat(N: Val.getOperand(i: 0));
4637 if (C && C->getAPIntValue().isSignMask())
4638 return true;
4639 return isKnownToBeAPowerOfTwo(Val: Val.getOperand(i: 0), Depth: Depth + 1) &&
4640 isKnownNeverZero(Op: Val, Depth);
4641 }
4642
4643 if (Val.getOpcode() == ISD::ROTL || Val.getOpcode() == ISD::ROTR)
4644 return isKnownToBeAPowerOfTwo(Val: Val.getOperand(i: 0), Depth: Depth + 1);
4645
4646 // Are all operands of a build vector constant powers of two?
4647 if (Val.getOpcode() == ISD::BUILD_VECTOR)
4648 if (llvm::all_of(Range: Val->ops(), P: [BitWidth](SDValue E) {
4649 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: E))
4650 return C->getAPIntValue().zextOrTrunc(width: BitWidth).isPowerOf2();
4651 return false;
4652 }))
4653 return true;
4654
4655 // Is the operand of a splat vector a constant power of two?
4656 if (Val.getOpcode() == ISD::SPLAT_VECTOR)
4657 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val: Val->getOperand(Num: 0)))
4658 if (C->getAPIntValue().zextOrTrunc(width: BitWidth).isPowerOf2())
4659 return true;
4660
4661 // vscale(power-of-two) is a power-of-two for some targets
4662 if (Val.getOpcode() == ISD::VSCALE &&
4663 getTargetLoweringInfo().isVScaleKnownToBeAPowerOfTwo() &&
4664 isKnownToBeAPowerOfTwo(Val: Val.getOperand(i: 0), Depth: Depth + 1))
4665 return true;
4666
4667 if (Val.getOpcode() == ISD::SMIN || Val.getOpcode() == ISD::SMAX ||
4668 Val.getOpcode() == ISD::UMIN || Val.getOpcode() == ISD::UMAX)
4669 return isKnownToBeAPowerOfTwo(Val: Val.getOperand(i: 1), Depth: Depth + 1) &&
4670 isKnownToBeAPowerOfTwo(Val: Val.getOperand(i: 0), Depth: Depth + 1);
4671
4672 if (Val.getOpcode() == ISD::SELECT || Val.getOpcode() == ISD::VSELECT)
4673 return isKnownToBeAPowerOfTwo(Val: Val.getOperand(i: 2), Depth: Depth + 1) &&
4674 isKnownToBeAPowerOfTwo(Val: Val.getOperand(i: 1), Depth: Depth + 1);
4675
4676 // Looking for `x & -x` pattern:
4677 // If x == 0:
4678 // x & -x -> 0
4679 // If x != 0:
4680 // x & -x -> non-zero pow2
4681 // so if we find the pattern return whether we know `x` is non-zero.
4682 SDValue X;
4683 if (sd_match(N: Val, P: m_And(L: m_Value(N&: X), R: m_Neg(V: m_Deferred(V&: X)))))
4684 return isKnownNeverZero(Op: X, Depth);
4685
4686 if (Val.getOpcode() == ISD::ZERO_EXTEND)
4687 return isKnownToBeAPowerOfTwo(Val: Val.getOperand(i: 0), Depth: Depth + 1);
4688
4689 // More could be done here, though the above checks are enough
4690 // to handle some common cases.
4691 return false;
4692}
4693
4694bool SelectionDAG::isKnownToBeAPowerOfTwoFP(SDValue Val, unsigned Depth) const {
4695 if (ConstantFPSDNode *C1 = isConstOrConstSplatFP(N: Val, AllowUndefs: true))
4696 return C1->getValueAPF().getExactLog2Abs() >= 0;
4697
4698 if (Val.getOpcode() == ISD::UINT_TO_FP || Val.getOpcode() == ISD::SINT_TO_FP)
4699 return isKnownToBeAPowerOfTwo(Val: Val.getOperand(i: 0), Depth: Depth + 1);
4700
4701 return false;
4702}
4703
4704unsigned SelectionDAG::ComputeNumSignBits(SDValue Op, unsigned Depth) const {
4705 EVT VT = Op.getValueType();
4706
4707 // Since the number of lanes in a scalable vector is unknown at compile time,
4708 // we track one bit which is implicitly broadcast to all lanes. This means
4709 // that all lanes in a scalable vector are considered demanded.
4710 APInt DemandedElts = VT.isFixedLengthVector()
4711 ? APInt::getAllOnes(numBits: VT.getVectorNumElements())
4712 : APInt(1, 1);
4713 return ComputeNumSignBits(Op, DemandedElts, Depth);
4714}
4715
4716unsigned SelectionDAG::ComputeNumSignBits(SDValue Op, const APInt &DemandedElts,
4717 unsigned Depth) const {
4718 EVT VT = Op.getValueType();
4719 assert((VT.isInteger() || VT.isFloatingPoint()) && "Invalid VT!");
4720 unsigned VTBits = VT.getScalarSizeInBits();
4721 unsigned NumElts = DemandedElts.getBitWidth();
4722 unsigned Tmp, Tmp2;
4723 unsigned FirstAnswer = 1;
4724
4725 if (auto *C = dyn_cast<ConstantSDNode>(Val&: Op)) {
4726 const APInt &Val = C->getAPIntValue();
4727 return Val.getNumSignBits();
4728 }
4729
4730 if (Depth >= MaxRecursionDepth)
4731 return 1; // Limit search depth.
4732
4733 if (!DemandedElts)
4734 return 1; // No demanded elts, better to assume we don't know anything.
4735
4736 unsigned Opcode = Op.getOpcode();
4737 switch (Opcode) {
4738 default: break;
4739 case ISD::AssertSext:
4740 Tmp = cast<VTSDNode>(Val: Op.getOperand(i: 1))->getVT().getSizeInBits();
4741 return VTBits-Tmp+1;
4742 case ISD::AssertZext:
4743 Tmp = cast<VTSDNode>(Val: Op.getOperand(i: 1))->getVT().getSizeInBits();
4744 return VTBits-Tmp;
4745 case ISD::MERGE_VALUES:
4746 return ComputeNumSignBits(Op: Op.getOperand(i: Op.getResNo()), DemandedElts,
4747 Depth: Depth + 1);
4748 case ISD::SPLAT_VECTOR: {
4749 // Check if the sign bits of source go down as far as the truncated value.
4750 unsigned NumSrcBits = Op.getOperand(i: 0).getValueSizeInBits();
4751 unsigned NumSrcSignBits = ComputeNumSignBits(Op: Op.getOperand(i: 0), Depth: Depth + 1);
4752 if (NumSrcSignBits > (NumSrcBits - VTBits))
4753 return NumSrcSignBits - (NumSrcBits - VTBits);
4754 break;
4755 }
4756 case ISD::BUILD_VECTOR:
4757 assert(!VT.isScalableVector());
4758 Tmp = VTBits;
4759 for (unsigned i = 0, e = Op.getNumOperands(); (i < e) && (Tmp > 1); ++i) {
4760 if (!DemandedElts[i])
4761 continue;
4762
4763 SDValue SrcOp = Op.getOperand(i);
4764 // BUILD_VECTOR can implicitly truncate sources, we handle this specially
4765 // for constant nodes to ensure we only look at the sign bits.
4766 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: SrcOp)) {
4767 APInt T = C->getAPIntValue().trunc(width: VTBits);
4768 Tmp2 = T.getNumSignBits();
4769 } else {
4770 Tmp2 = ComputeNumSignBits(Op: SrcOp, Depth: Depth + 1);
4771
4772 if (SrcOp.getValueSizeInBits() != VTBits) {
4773 assert(SrcOp.getValueSizeInBits() > VTBits &&
4774 "Expected BUILD_VECTOR implicit truncation");
4775 unsigned ExtraBits = SrcOp.getValueSizeInBits() - VTBits;
4776 Tmp2 = (Tmp2 > ExtraBits ? Tmp2 - ExtraBits : 1);
4777 }
4778 }
4779 Tmp = std::min(a: Tmp, b: Tmp2);
4780 }
4781 return Tmp;
4782
4783 case ISD::VECTOR_SHUFFLE: {
4784 // Collect the minimum number of sign bits that are shared by every vector
4785 // element referenced by the shuffle.
4786 APInt DemandedLHS, DemandedRHS;
4787 const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Val&: Op);
4788 assert(NumElts == SVN->getMask().size() && "Unexpected vector size");
4789 if (!getShuffleDemandedElts(SrcWidth: NumElts, Mask: SVN->getMask(), DemandedElts,
4790 DemandedLHS, DemandedRHS))
4791 return 1;
4792
4793 Tmp = std::numeric_limits<unsigned>::max();
4794 if (!!DemandedLHS)
4795 Tmp = ComputeNumSignBits(Op: Op.getOperand(i: 0), DemandedElts: DemandedLHS, Depth: Depth + 1);
4796 if (!!DemandedRHS) {
4797 Tmp2 = ComputeNumSignBits(Op: Op.getOperand(i: 1), DemandedElts: DemandedRHS, Depth: Depth + 1);
4798 Tmp = std::min(a: Tmp, b: Tmp2);
4799 }
4800 // If we don't know anything, early out and try computeKnownBits fall-back.
4801 if (Tmp == 1)
4802 break;
4803 assert(Tmp <= VTBits && "Failed to determine minimum sign bits");
4804 return Tmp;
4805 }
4806
4807 case ISD::BITCAST: {
4808 if (VT.isScalableVector())
4809 break;
4810 SDValue N0 = Op.getOperand(i: 0);
4811 EVT SrcVT = N0.getValueType();
4812 unsigned SrcBits = SrcVT.getScalarSizeInBits();
4813
4814 // Ignore bitcasts from unsupported types..
4815 if (!(SrcVT.isInteger() || SrcVT.isFloatingPoint()))
4816 break;
4817
4818 // Fast handling of 'identity' bitcasts.
4819 if (VTBits == SrcBits)
4820 return ComputeNumSignBits(Op: N0, DemandedElts, Depth: Depth + 1);
4821
4822 bool IsLE = getDataLayout().isLittleEndian();
4823
4824 // Bitcast 'large element' scalar/vector to 'small element' vector.
4825 if ((SrcBits % VTBits) == 0) {
4826 assert(VT.isVector() && "Expected bitcast to vector");
4827
4828 unsigned Scale = SrcBits / VTBits;
4829 APInt SrcDemandedElts =
4830 APIntOps::ScaleBitMask(A: DemandedElts, NewBitWidth: NumElts / Scale);
4831
4832 // Fast case - sign splat can be simply split across the small elements.
4833 Tmp = ComputeNumSignBits(Op: N0, DemandedElts: SrcDemandedElts, Depth: Depth + 1);
4834 if (Tmp == SrcBits)
4835 return VTBits;
4836
4837 // Slow case - determine how far the sign extends into each sub-element.
4838 Tmp2 = VTBits;
4839 for (unsigned i = 0; i != NumElts; ++i)
4840 if (DemandedElts[i]) {
4841 unsigned SubOffset = i % Scale;
4842 SubOffset = (IsLE ? ((Scale - 1) - SubOffset) : SubOffset);
4843 SubOffset = SubOffset * VTBits;
4844 if (Tmp <= SubOffset)
4845 return 1;
4846 Tmp2 = std::min(a: Tmp2, b: Tmp - SubOffset);
4847 }
4848 return Tmp2;
4849 }
4850 break;
4851 }
4852
4853 case ISD::FP_TO_SINT_SAT:
4854 // FP_TO_SINT_SAT produces a signed value that fits in the saturating VT.
4855 Tmp = cast<VTSDNode>(Val: Op.getOperand(i: 1))->getVT().getScalarSizeInBits();
4856 return VTBits - Tmp + 1;
4857 case ISD::SIGN_EXTEND:
4858 Tmp = VTBits - Op.getOperand(i: 0).getScalarValueSizeInBits();
4859 return ComputeNumSignBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth+1) + Tmp;
4860 case ISD::SIGN_EXTEND_INREG:
4861 // Max of the input and what this extends.
4862 Tmp = cast<VTSDNode>(Val: Op.getOperand(i: 1))->getVT().getScalarSizeInBits();
4863 Tmp = VTBits-Tmp+1;
4864 Tmp2 = ComputeNumSignBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth+1);
4865 return std::max(a: Tmp, b: Tmp2);
4866 case ISD::SIGN_EXTEND_VECTOR_INREG: {
4867 if (VT.isScalableVector())
4868 break;
4869 SDValue Src = Op.getOperand(i: 0);
4870 EVT SrcVT = Src.getValueType();
4871 APInt DemandedSrcElts = DemandedElts.zext(width: SrcVT.getVectorNumElements());
4872 Tmp = VTBits - SrcVT.getScalarSizeInBits();
4873 return ComputeNumSignBits(Op: Src, DemandedElts: DemandedSrcElts, Depth: Depth+1) + Tmp;
4874 }
4875 case ISD::SRA:
4876 Tmp = ComputeNumSignBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4877 // SRA X, C -> adds C sign bits.
4878 if (std::optional<uint64_t> ShAmt =
4879 getValidMinimumShiftAmount(V: Op, DemandedElts, Depth: Depth + 1))
4880 Tmp = std::min<uint64_t>(a: Tmp + *ShAmt, b: VTBits);
4881 return Tmp;
4882 case ISD::SHL:
4883 if (std::optional<ConstantRange> ShAmtRange =
4884 getValidShiftAmountRange(V: Op, DemandedElts, Depth: Depth + 1)) {
4885 uint64_t MaxShAmt = ShAmtRange->getUnsignedMax().getZExtValue();
4886 uint64_t MinShAmt = ShAmtRange->getUnsignedMin().getZExtValue();
4887 // Try to look through ZERO/SIGN/ANY_EXTEND. If all extended bits are
4888 // shifted out, then we can compute the number of sign bits for the
4889 // operand being extended. A future improvement could be to pass along the
4890 // "shifted left by" information in the recursive calls to
4891 // ComputeKnownSignBits. Allowing us to handle this more generically.
4892 if (ISD::isExtOpcode(Opcode: Op.getOperand(i: 0).getOpcode())) {
4893 SDValue Ext = Op.getOperand(i: 0);
4894 EVT ExtVT = Ext.getValueType();
4895 SDValue Extendee = Ext.getOperand(i: 0);
4896 EVT ExtendeeVT = Extendee.getValueType();
4897 uint64_t SizeDifference =
4898 ExtVT.getScalarSizeInBits() - ExtendeeVT.getScalarSizeInBits();
4899 if (SizeDifference <= MinShAmt) {
4900 Tmp = SizeDifference +
4901 ComputeNumSignBits(Op: Extendee, DemandedElts, Depth: Depth + 1);
4902 if (MaxShAmt < Tmp)
4903 return Tmp - MaxShAmt;
4904 }
4905 }
4906 // shl destroys sign bits, ensure it doesn't shift out all sign bits.
4907 Tmp = ComputeNumSignBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4908 if (MaxShAmt < Tmp)
4909 return Tmp - MaxShAmt;
4910 }
4911 break;
4912 case ISD::AND:
4913 case ISD::OR:
4914 case ISD::XOR: // NOT is handled here.
4915 // Logical binary ops preserve the number of sign bits at the worst.
4916 Tmp = ComputeNumSignBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth+1);
4917 if (Tmp != 1) {
4918 Tmp2 = ComputeNumSignBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth+1);
4919 FirstAnswer = std::min(a: Tmp, b: Tmp2);
4920 // We computed what we know about the sign bits as our first
4921 // answer. Now proceed to the generic code that uses
4922 // computeKnownBits, and pick whichever answer is better.
4923 }
4924 break;
4925
4926 case ISD::SELECT:
4927 case ISD::VSELECT:
4928 Tmp = ComputeNumSignBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth+1);
4929 if (Tmp == 1) return 1; // Early out.
4930 Tmp2 = ComputeNumSignBits(Op: Op.getOperand(i: 2), DemandedElts, Depth: Depth+1);
4931 return std::min(a: Tmp, b: Tmp2);
4932 case ISD::SELECT_CC:
4933 Tmp = ComputeNumSignBits(Op: Op.getOperand(i: 2), DemandedElts, Depth: Depth+1);
4934 if (Tmp == 1) return 1; // Early out.
4935 Tmp2 = ComputeNumSignBits(Op: Op.getOperand(i: 3), DemandedElts, Depth: Depth+1);
4936 return std::min(a: Tmp, b: Tmp2);
4937
4938 case ISD::SMIN:
4939 case ISD::SMAX: {
4940 // If we have a clamp pattern, we know that the number of sign bits will be
4941 // the minimum of the clamp min/max range.
4942 bool IsMax = (Opcode == ISD::SMAX);
4943 ConstantSDNode *CstLow = nullptr, *CstHigh = nullptr;
4944 if ((CstLow = isConstOrConstSplat(N: Op.getOperand(i: 1), DemandedElts)))
4945 if (Op.getOperand(i: 0).getOpcode() == (IsMax ? ISD::SMIN : ISD::SMAX))
4946 CstHigh =
4947 isConstOrConstSplat(N: Op.getOperand(i: 0).getOperand(i: 1), DemandedElts);
4948 if (CstLow && CstHigh) {
4949 if (!IsMax)
4950 std::swap(a&: CstLow, b&: CstHigh);
4951 if (CstLow->getAPIntValue().sle(RHS: CstHigh->getAPIntValue())) {
4952 Tmp = CstLow->getAPIntValue().getNumSignBits();
4953 Tmp2 = CstHigh->getAPIntValue().getNumSignBits();
4954 return std::min(a: Tmp, b: Tmp2);
4955 }
4956 }
4957
4958 // Fallback - just get the minimum number of sign bits of the operands.
4959 Tmp = ComputeNumSignBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4960 if (Tmp == 1)
4961 return 1; // Early out.
4962 Tmp2 = ComputeNumSignBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
4963 return std::min(a: Tmp, b: Tmp2);
4964 }
4965 case ISD::UMIN:
4966 case ISD::UMAX:
4967 Tmp = ComputeNumSignBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
4968 if (Tmp == 1)
4969 return 1; // Early out.
4970 Tmp2 = ComputeNumSignBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
4971 return std::min(a: Tmp, b: Tmp2);
4972 case ISD::SSUBO_CARRY:
4973 case ISD::USUBO_CARRY:
4974 // sub_carry(x,x,c) -> 0/-1 (sext carry)
4975 if (Op.getResNo() == 0 && Op.getOperand(i: 0) == Op.getOperand(i: 1))
4976 return VTBits;
4977 [[fallthrough]];
4978 case ISD::SADDO:
4979 case ISD::UADDO:
4980 case ISD::SADDO_CARRY:
4981 case ISD::UADDO_CARRY:
4982 case ISD::SSUBO:
4983 case ISD::USUBO:
4984 case ISD::SMULO:
4985 case ISD::UMULO:
4986 if (Op.getResNo() != 1)
4987 break;
4988 // The boolean result conforms to getBooleanContents. Fall through.
4989 // If setcc returns 0/-1, all bits are sign bits.
4990 // We know that we have an integer-based boolean since these operations
4991 // are only available for integer.
4992 if (TLI->getBooleanContents(isVec: VT.isVector(), isFloat: false) ==
4993 TargetLowering::ZeroOrNegativeOneBooleanContent)
4994 return VTBits;
4995 break;
4996 case ISD::SETCC:
4997 case ISD::SETCCCARRY:
4998 case ISD::STRICT_FSETCC:
4999 case ISD::STRICT_FSETCCS: {
5000 unsigned OpNo = Op->isStrictFPOpcode() ? 1 : 0;
5001 // If setcc returns 0/-1, all bits are sign bits.
5002 if (TLI->getBooleanContents(Type: Op.getOperand(i: OpNo).getValueType()) ==
5003 TargetLowering::ZeroOrNegativeOneBooleanContent)
5004 return VTBits;
5005 break;
5006 }
5007 case ISD::ROTL:
5008 case ISD::ROTR:
5009 Tmp = ComputeNumSignBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
5010
5011 // If we're rotating an 0/-1 value, then it stays an 0/-1 value.
5012 if (Tmp == VTBits)
5013 return VTBits;
5014
5015 if (ConstantSDNode *C =
5016 isConstOrConstSplat(N: Op.getOperand(i: 1), DemandedElts)) {
5017 unsigned RotAmt = C->getAPIntValue().urem(RHS: VTBits);
5018
5019 // Handle rotate right by N like a rotate left by 32-N.
5020 if (Opcode == ISD::ROTR)
5021 RotAmt = (VTBits - RotAmt) % VTBits;
5022
5023 // If we aren't rotating out all of the known-in sign bits, return the
5024 // number that are left. This handles rotl(sext(x), 1) for example.
5025 if (Tmp > (RotAmt + 1)) return (Tmp - RotAmt);
5026 }
5027 break;
5028 case ISD::ADD:
5029 case ISD::ADDC:
5030 // Add can have at most one carry bit. Thus we know that the output
5031 // is, at worst, one more bit than the inputs.
5032 Tmp = ComputeNumSignBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
5033 if (Tmp == 1) return 1; // Early out.
5034
5035 // Special case decrementing a value (ADD X, -1):
5036 if (ConstantSDNode *CRHS =
5037 isConstOrConstSplat(N: Op.getOperand(i: 1), DemandedElts))
5038 if (CRHS->isAllOnes()) {
5039 KnownBits Known =
5040 computeKnownBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
5041
5042 // If the input is known to be 0 or 1, the output is 0/-1, which is all
5043 // sign bits set.
5044 if ((Known.Zero | 1).isAllOnes())
5045 return VTBits;
5046
5047 // If we are subtracting one from a positive number, there is no carry
5048 // out of the result.
5049 if (Known.isNonNegative())
5050 return Tmp;
5051 }
5052
5053 Tmp2 = ComputeNumSignBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
5054 if (Tmp2 == 1) return 1; // Early out.
5055 return std::min(a: Tmp, b: Tmp2) - 1;
5056 case ISD::SUB:
5057 Tmp2 = ComputeNumSignBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
5058 if (Tmp2 == 1) return 1; // Early out.
5059
5060 // Handle NEG.
5061 if (ConstantSDNode *CLHS =
5062 isConstOrConstSplat(N: Op.getOperand(i: 0), DemandedElts))
5063 if (CLHS->isZero()) {
5064 KnownBits Known =
5065 computeKnownBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
5066 // If the input is known to be 0 or 1, the output is 0/-1, which is all
5067 // sign bits set.
5068 if ((Known.Zero | 1).isAllOnes())
5069 return VTBits;
5070
5071 // If the input is known to be positive (the sign bit is known clear),
5072 // the output of the NEG has the same number of sign bits as the input.
5073 if (Known.isNonNegative())
5074 return Tmp2;
5075
5076 // Otherwise, we treat this like a SUB.
5077 }
5078
5079 // Sub can have at most one carry bit. Thus we know that the output
5080 // is, at worst, one more bit than the inputs.
5081 Tmp = ComputeNumSignBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
5082 if (Tmp == 1) return 1; // Early out.
5083 return std::min(a: Tmp, b: Tmp2) - 1;
5084 case ISD::MUL: {
5085 // The output of the Mul can be at most twice the valid bits in the inputs.
5086 unsigned SignBitsOp0 = ComputeNumSignBits(Op: Op.getOperand(i: 0), Depth: Depth + 1);
5087 if (SignBitsOp0 == 1)
5088 break;
5089 unsigned SignBitsOp1 = ComputeNumSignBits(Op: Op.getOperand(i: 1), Depth: Depth + 1);
5090 if (SignBitsOp1 == 1)
5091 break;
5092 unsigned OutValidBits =
5093 (VTBits - SignBitsOp0 + 1) + (VTBits - SignBitsOp1 + 1);
5094 return OutValidBits > VTBits ? 1 : VTBits - OutValidBits + 1;
5095 }
5096 case ISD::AVGCEILS:
5097 case ISD::AVGFLOORS:
5098 Tmp = ComputeNumSignBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
5099 if (Tmp == 1)
5100 return 1; // Early out.
5101 Tmp2 = ComputeNumSignBits(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1);
5102 return std::min(a: Tmp, b: Tmp2);
5103 case ISD::SREM:
5104 // The sign bit is the LHS's sign bit, except when the result of the
5105 // remainder is zero. The magnitude of the result should be less than or
5106 // equal to the magnitude of the LHS. Therefore, the result should have
5107 // at least as many sign bits as the left hand side.
5108 return ComputeNumSignBits(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1);
5109 case ISD::TRUNCATE: {
5110 // Check if the sign bits of source go down as far as the truncated value.
5111 unsigned NumSrcBits = Op.getOperand(i: 0).getScalarValueSizeInBits();
5112 unsigned NumSrcSignBits = ComputeNumSignBits(Op: Op.getOperand(i: 0), Depth: Depth + 1);
5113 if (NumSrcSignBits > (NumSrcBits - VTBits))
5114 return NumSrcSignBits - (NumSrcBits - VTBits);
5115 break;
5116 }
5117 case ISD::EXTRACT_ELEMENT: {
5118 if (VT.isScalableVector())
5119 break;
5120 const int KnownSign = ComputeNumSignBits(Op: Op.getOperand(i: 0), Depth: Depth+1);
5121 const int BitWidth = Op.getValueSizeInBits();
5122 const int Items = Op.getOperand(i: 0).getValueSizeInBits() / BitWidth;
5123
5124 // Get reverse index (starting from 1), Op1 value indexes elements from
5125 // little end. Sign starts at big end.
5126 const int rIndex = Items - 1 - Op.getConstantOperandVal(i: 1);
5127
5128 // If the sign portion ends in our element the subtraction gives correct
5129 // result. Otherwise it gives either negative or > bitwidth result
5130 return std::clamp(val: KnownSign - rIndex * BitWidth, lo: 0, hi: BitWidth);
5131 }
5132 case ISD::INSERT_VECTOR_ELT: {
5133 if (VT.isScalableVector())
5134 break;
5135 // If we know the element index, split the demand between the
5136 // source vector and the inserted element, otherwise assume we need
5137 // the original demanded vector elements and the value.
5138 SDValue InVec = Op.getOperand(i: 0);
5139 SDValue InVal = Op.getOperand(i: 1);
5140 SDValue EltNo = Op.getOperand(i: 2);
5141 bool DemandedVal = true;
5142 APInt DemandedVecElts = DemandedElts;
5143 auto *CEltNo = dyn_cast<ConstantSDNode>(Val&: EltNo);
5144 if (CEltNo && CEltNo->getAPIntValue().ult(RHS: NumElts)) {
5145 unsigned EltIdx = CEltNo->getZExtValue();
5146 DemandedVal = !!DemandedElts[EltIdx];
5147 DemandedVecElts.clearBit(BitPosition: EltIdx);
5148 }
5149 Tmp = std::numeric_limits<unsigned>::max();
5150 if (DemandedVal) {
5151 // TODO - handle implicit truncation of inserted elements.
5152 if (InVal.getScalarValueSizeInBits() != VTBits)
5153 break;
5154 Tmp2 = ComputeNumSignBits(Op: InVal, Depth: Depth + 1);
5155 Tmp = std::min(a: Tmp, b: Tmp2);
5156 }
5157 if (!!DemandedVecElts) {
5158 Tmp2 = ComputeNumSignBits(Op: InVec, DemandedElts: DemandedVecElts, Depth: Depth + 1);
5159 Tmp = std::min(a: Tmp, b: Tmp2);
5160 }
5161 assert(Tmp <= VTBits && "Failed to determine minimum sign bits");
5162 return Tmp;
5163 }
5164 case ISD::EXTRACT_VECTOR_ELT: {
5165 assert(!VT.isScalableVector());
5166 SDValue InVec = Op.getOperand(i: 0);
5167 SDValue EltNo = Op.getOperand(i: 1);
5168 EVT VecVT = InVec.getValueType();
5169 // ComputeNumSignBits not yet implemented for scalable vectors.
5170 if (VecVT.isScalableVector())
5171 break;
5172 const unsigned BitWidth = Op.getValueSizeInBits();
5173 const unsigned EltBitWidth = Op.getOperand(i: 0).getScalarValueSizeInBits();
5174 const unsigned NumSrcElts = VecVT.getVectorNumElements();
5175
5176 // If BitWidth > EltBitWidth the value is anyext:ed, and we do not know
5177 // anything about sign bits. But if the sizes match we can derive knowledge
5178 // about sign bits from the vector operand.
5179 if (BitWidth != EltBitWidth)
5180 break;
5181
5182 // If we know the element index, just demand that vector element, else for
5183 // an unknown element index, ignore DemandedElts and demand them all.
5184 APInt DemandedSrcElts = APInt::getAllOnes(numBits: NumSrcElts);
5185 auto *ConstEltNo = dyn_cast<ConstantSDNode>(Val&: EltNo);
5186 if (ConstEltNo && ConstEltNo->getAPIntValue().ult(RHS: NumSrcElts))
5187 DemandedSrcElts =
5188 APInt::getOneBitSet(numBits: NumSrcElts, BitNo: ConstEltNo->getZExtValue());
5189
5190 return ComputeNumSignBits(Op: InVec, DemandedElts: DemandedSrcElts, Depth: Depth + 1);
5191 }
5192 case ISD::EXTRACT_SUBVECTOR: {
5193 // Offset the demanded elts by the subvector index.
5194 SDValue Src = Op.getOperand(i: 0);
5195 // Bail until we can represent demanded elements for scalable vectors.
5196 if (Src.getValueType().isScalableVector())
5197 break;
5198 uint64_t Idx = Op.getConstantOperandVal(i: 1);
5199 unsigned NumSrcElts = Src.getValueType().getVectorNumElements();
5200 APInt DemandedSrcElts = DemandedElts.zext(width: NumSrcElts).shl(shiftAmt: Idx);
5201 return ComputeNumSignBits(Op: Src, DemandedElts: DemandedSrcElts, Depth: Depth + 1);
5202 }
5203 case ISD::CONCAT_VECTORS: {
5204 if (VT.isScalableVector())
5205 break;
5206 // Determine the minimum number of sign bits across all demanded
5207 // elts of the input vectors. Early out if the result is already 1.
5208 Tmp = std::numeric_limits<unsigned>::max();
5209 EVT SubVectorVT = Op.getOperand(i: 0).getValueType();
5210 unsigned NumSubVectorElts = SubVectorVT.getVectorNumElements();
5211 unsigned NumSubVectors = Op.getNumOperands();
5212 for (unsigned i = 0; (i < NumSubVectors) && (Tmp > 1); ++i) {
5213 APInt DemandedSub =
5214 DemandedElts.extractBits(numBits: NumSubVectorElts, bitPosition: i * NumSubVectorElts);
5215 if (!DemandedSub)
5216 continue;
5217 Tmp2 = ComputeNumSignBits(Op: Op.getOperand(i), DemandedElts: DemandedSub, Depth: Depth + 1);
5218 Tmp = std::min(a: Tmp, b: Tmp2);
5219 }
5220 assert(Tmp <= VTBits && "Failed to determine minimum sign bits");
5221 return Tmp;
5222 }
5223 case ISD::INSERT_SUBVECTOR: {
5224 if (VT.isScalableVector())
5225 break;
5226 // Demand any elements from the subvector and the remainder from the src its
5227 // inserted into.
5228 SDValue Src = Op.getOperand(i: 0);
5229 SDValue Sub = Op.getOperand(i: 1);
5230 uint64_t Idx = Op.getConstantOperandVal(i: 2);
5231 unsigned NumSubElts = Sub.getValueType().getVectorNumElements();
5232 APInt DemandedSubElts = DemandedElts.extractBits(numBits: NumSubElts, bitPosition: Idx);
5233 APInt DemandedSrcElts = DemandedElts;
5234 DemandedSrcElts.clearBits(LoBit: Idx, HiBit: Idx + NumSubElts);
5235
5236 Tmp = std::numeric_limits<unsigned>::max();
5237 if (!!DemandedSubElts) {
5238 Tmp = ComputeNumSignBits(Op: Sub, DemandedElts: DemandedSubElts, Depth: Depth + 1);
5239 if (Tmp == 1)
5240 return 1; // early-out
5241 }
5242 if (!!DemandedSrcElts) {
5243 Tmp2 = ComputeNumSignBits(Op: Src, DemandedElts: DemandedSrcElts, Depth: Depth + 1);
5244 Tmp = std::min(a: Tmp, b: Tmp2);
5245 }
5246 assert(Tmp <= VTBits && "Failed to determine minimum sign bits");
5247 return Tmp;
5248 }
5249 case ISD::LOAD: {
5250 LoadSDNode *LD = cast<LoadSDNode>(Val&: Op);
5251 if (const MDNode *Ranges = LD->getRanges()) {
5252 if (DemandedElts != 1)
5253 break;
5254
5255 ConstantRange CR = getConstantRangeFromMetadata(RangeMD: *Ranges);
5256 if (VTBits > CR.getBitWidth()) {
5257 switch (LD->getExtensionType()) {
5258 case ISD::SEXTLOAD:
5259 CR = CR.signExtend(BitWidth: VTBits);
5260 break;
5261 case ISD::ZEXTLOAD:
5262 CR = CR.zeroExtend(BitWidth: VTBits);
5263 break;
5264 default:
5265 break;
5266 }
5267 }
5268
5269 if (VTBits != CR.getBitWidth())
5270 break;
5271 return std::min(a: CR.getSignedMin().getNumSignBits(),
5272 b: CR.getSignedMax().getNumSignBits());
5273 }
5274
5275 break;
5276 }
5277 case ISD::ATOMIC_CMP_SWAP:
5278 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
5279 case ISD::ATOMIC_SWAP:
5280 case ISD::ATOMIC_LOAD_ADD:
5281 case ISD::ATOMIC_LOAD_SUB:
5282 case ISD::ATOMIC_LOAD_AND:
5283 case ISD::ATOMIC_LOAD_CLR:
5284 case ISD::ATOMIC_LOAD_OR:
5285 case ISD::ATOMIC_LOAD_XOR:
5286 case ISD::ATOMIC_LOAD_NAND:
5287 case ISD::ATOMIC_LOAD_MIN:
5288 case ISD::ATOMIC_LOAD_MAX:
5289 case ISD::ATOMIC_LOAD_UMIN:
5290 case ISD::ATOMIC_LOAD_UMAX:
5291 case ISD::ATOMIC_LOAD: {
5292 auto *AT = cast<AtomicSDNode>(Val&: Op);
5293 // If we are looking at the loaded value.
5294 if (Op.getResNo() == 0) {
5295 Tmp = AT->getMemoryVT().getScalarSizeInBits();
5296 if (Tmp == VTBits)
5297 return 1; // early-out
5298
5299 // For atomic_load, prefer to use the extension type.
5300 if (Op->getOpcode() == ISD::ATOMIC_LOAD) {
5301 switch (AT->getExtensionType()) {
5302 default:
5303 break;
5304 case ISD::SEXTLOAD:
5305 return VTBits - Tmp + 1;
5306 case ISD::ZEXTLOAD:
5307 return VTBits - Tmp;
5308 }
5309 }
5310
5311 if (TLI->getExtendForAtomicOps() == ISD::SIGN_EXTEND)
5312 return VTBits - Tmp + 1;
5313 if (TLI->getExtendForAtomicOps() == ISD::ZERO_EXTEND)
5314 return VTBits - Tmp;
5315 }
5316 break;
5317 }
5318 }
5319
5320 // If we are looking at the loaded value of the SDNode.
5321 if (Op.getResNo() == 0) {
5322 // Handle LOADX separately here. EXTLOAD case will fallthrough.
5323 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Val&: Op)) {
5324 unsigned ExtType = LD->getExtensionType();
5325 switch (ExtType) {
5326 default: break;
5327 case ISD::SEXTLOAD: // e.g. i16->i32 = '17' bits known.
5328 Tmp = LD->getMemoryVT().getScalarSizeInBits();
5329 return VTBits - Tmp + 1;
5330 case ISD::ZEXTLOAD: // e.g. i16->i32 = '16' bits known.
5331 Tmp = LD->getMemoryVT().getScalarSizeInBits();
5332 return VTBits - Tmp;
5333 case ISD::NON_EXTLOAD:
5334 if (const Constant *Cst = TLI->getTargetConstantFromLoad(LD)) {
5335 // We only need to handle vectors - computeKnownBits should handle
5336 // scalar cases.
5337 Type *CstTy = Cst->getType();
5338 if (CstTy->isVectorTy() && !VT.isScalableVector() &&
5339 (NumElts * VTBits) == CstTy->getPrimitiveSizeInBits() &&
5340 VTBits == CstTy->getScalarSizeInBits()) {
5341 Tmp = VTBits;
5342 for (unsigned i = 0; i != NumElts; ++i) {
5343 if (!DemandedElts[i])
5344 continue;
5345 if (Constant *Elt = Cst->getAggregateElement(Elt: i)) {
5346 if (auto *CInt = dyn_cast<ConstantInt>(Val: Elt)) {
5347 const APInt &Value = CInt->getValue();
5348 Tmp = std::min(a: Tmp, b: Value.getNumSignBits());
5349 continue;
5350 }
5351 if (auto *CFP = dyn_cast<ConstantFP>(Val: Elt)) {
5352 APInt Value = CFP->getValueAPF().bitcastToAPInt();
5353 Tmp = std::min(a: Tmp, b: Value.getNumSignBits());
5354 continue;
5355 }
5356 }
5357 // Unknown type. Conservatively assume no bits match sign bit.
5358 return 1;
5359 }
5360 return Tmp;
5361 }
5362 }
5363 break;
5364 }
5365 }
5366 }
5367
5368 // Allow the target to implement this method for its nodes.
5369 if (Opcode >= ISD::BUILTIN_OP_END ||
5370 Opcode == ISD::INTRINSIC_WO_CHAIN ||
5371 Opcode == ISD::INTRINSIC_W_CHAIN ||
5372 Opcode == ISD::INTRINSIC_VOID) {
5373 // TODO: This can probably be removed once target code is audited. This
5374 // is here purely to reduce patch size and review complexity.
5375 if (!VT.isScalableVector()) {
5376 unsigned NumBits =
5377 TLI->ComputeNumSignBitsForTargetNode(Op, DemandedElts, DAG: *this, Depth);
5378 if (NumBits > 1)
5379 FirstAnswer = std::max(a: FirstAnswer, b: NumBits);
5380 }
5381 }
5382
5383 // Finally, if we can prove that the top bits of the result are 0's or 1's,
5384 // use this information.
5385 KnownBits Known = computeKnownBits(Op, DemandedElts, Depth);
5386 return std::max(a: FirstAnswer, b: Known.countMinSignBits());
5387}
5388
5389unsigned SelectionDAG::ComputeMaxSignificantBits(SDValue Op,
5390 unsigned Depth) const {
5391 unsigned SignBits = ComputeNumSignBits(Op, Depth);
5392 return Op.getScalarValueSizeInBits() - SignBits + 1;
5393}
5394
5395unsigned SelectionDAG::ComputeMaxSignificantBits(SDValue Op,
5396 const APInt &DemandedElts,
5397 unsigned Depth) const {
5398 unsigned SignBits = ComputeNumSignBits(Op, DemandedElts, Depth);
5399 return Op.getScalarValueSizeInBits() - SignBits + 1;
5400}
5401
5402bool SelectionDAG::isGuaranteedNotToBeUndefOrPoison(SDValue Op, bool PoisonOnly,
5403 unsigned Depth) const {
5404 // Early out for FREEZE.
5405 if (Op.getOpcode() == ISD::FREEZE)
5406 return true;
5407
5408 EVT VT = Op.getValueType();
5409 APInt DemandedElts = VT.isFixedLengthVector()
5410 ? APInt::getAllOnes(numBits: VT.getVectorNumElements())
5411 : APInt(1, 1);
5412 return isGuaranteedNotToBeUndefOrPoison(Op, DemandedElts, PoisonOnly, Depth);
5413}
5414
5415bool SelectionDAG::isGuaranteedNotToBeUndefOrPoison(SDValue Op,
5416 const APInt &DemandedElts,
5417 bool PoisonOnly,
5418 unsigned Depth) const {
5419 unsigned Opcode = Op.getOpcode();
5420
5421 // Early out for FREEZE.
5422 if (Opcode == ISD::FREEZE)
5423 return true;
5424
5425 if (Depth >= MaxRecursionDepth)
5426 return false; // Limit search depth.
5427
5428 if (isIntOrFPConstant(V: Op))
5429 return true;
5430
5431 switch (Opcode) {
5432 case ISD::CONDCODE:
5433 case ISD::VALUETYPE:
5434 case ISD::FrameIndex:
5435 case ISD::TargetFrameIndex:
5436 case ISD::CopyFromReg:
5437 return true;
5438
5439 case ISD::POISON:
5440 return false;
5441
5442 case ISD::UNDEF:
5443 return PoisonOnly;
5444
5445 case ISD::BUILD_VECTOR:
5446 // NOTE: BUILD_VECTOR has implicit truncation of wider scalar elements -
5447 // this shouldn't affect the result.
5448 for (unsigned i = 0, e = Op.getNumOperands(); i < e; ++i) {
5449 if (!DemandedElts[i])
5450 continue;
5451 if (!isGuaranteedNotToBeUndefOrPoison(Op: Op.getOperand(i), PoisonOnly,
5452 Depth: Depth + 1))
5453 return false;
5454 }
5455 return true;
5456
5457 case ISD::SPLAT_VECTOR:
5458 return isGuaranteedNotToBeUndefOrPoison(Op: Op.getOperand(i: 0), PoisonOnly,
5459 Depth: Depth + 1);
5460
5461 case ISD::VECTOR_SHUFFLE: {
5462 APInt DemandedLHS, DemandedRHS;
5463 auto *SVN = cast<ShuffleVectorSDNode>(Val&: Op);
5464 if (!getShuffleDemandedElts(SrcWidth: DemandedElts.getBitWidth(), Mask: SVN->getMask(),
5465 DemandedElts, DemandedLHS, DemandedRHS,
5466 /*AllowUndefElts=*/false))
5467 return false;
5468 if (!DemandedLHS.isZero() &&
5469 !isGuaranteedNotToBeUndefOrPoison(Op: Op.getOperand(i: 0), DemandedElts: DemandedLHS,
5470 PoisonOnly, Depth: Depth + 1))
5471 return false;
5472 if (!DemandedRHS.isZero() &&
5473 !isGuaranteedNotToBeUndefOrPoison(Op: Op.getOperand(i: 1), DemandedElts: DemandedRHS,
5474 PoisonOnly, Depth: Depth + 1))
5475 return false;
5476 return true;
5477 }
5478
5479 // TODO: Search for noundef attributes from library functions.
5480
5481 // TODO: Pointers dereferenced by ISD::LOAD/STORE ops are noundef.
5482
5483 default:
5484 // Allow the target to implement this method for its nodes.
5485 if (Opcode >= ISD::BUILTIN_OP_END || Opcode == ISD::INTRINSIC_WO_CHAIN ||
5486 Opcode == ISD::INTRINSIC_W_CHAIN || Opcode == ISD::INTRINSIC_VOID)
5487 return TLI->isGuaranteedNotToBeUndefOrPoisonForTargetNode(
5488 Op, DemandedElts, DAG: *this, PoisonOnly, Depth);
5489 break;
5490 }
5491
5492 // If Op can't create undef/poison and none of its operands are undef/poison
5493 // then Op is never undef/poison.
5494 // NOTE: TargetNodes can handle this in themselves in
5495 // isGuaranteedNotToBeUndefOrPoisonForTargetNode or let
5496 // TargetLowering::isGuaranteedNotToBeUndefOrPoisonForTargetNode handle it.
5497 return !canCreateUndefOrPoison(Op, PoisonOnly, /*ConsiderFlags*/ true,
5498 Depth) &&
5499 all_of(Range: Op->ops(), P: [&](SDValue V) {
5500 return isGuaranteedNotToBeUndefOrPoison(Op: V, PoisonOnly, Depth: Depth + 1);
5501 });
5502}
5503
5504bool SelectionDAG::canCreateUndefOrPoison(SDValue Op, bool PoisonOnly,
5505 bool ConsiderFlags,
5506 unsigned Depth) const {
5507 EVT VT = Op.getValueType();
5508 APInt DemandedElts = VT.isFixedLengthVector()
5509 ? APInt::getAllOnes(numBits: VT.getVectorNumElements())
5510 : APInt(1, 1);
5511 return canCreateUndefOrPoison(Op, DemandedElts, PoisonOnly, ConsiderFlags,
5512 Depth);
5513}
5514
5515bool SelectionDAG::canCreateUndefOrPoison(SDValue Op, const APInt &DemandedElts,
5516 bool PoisonOnly, bool ConsiderFlags,
5517 unsigned Depth) const {
5518 if (ConsiderFlags && Op->hasPoisonGeneratingFlags())
5519 return true;
5520
5521 unsigned Opcode = Op.getOpcode();
5522 switch (Opcode) {
5523 case ISD::FREEZE:
5524 case ISD::CONCAT_VECTORS:
5525 case ISD::INSERT_SUBVECTOR:
5526 case ISD::EXTRACT_SUBVECTOR:
5527 case ISD::SADDSAT:
5528 case ISD::UADDSAT:
5529 case ISD::SSUBSAT:
5530 case ISD::USUBSAT:
5531 case ISD::MULHU:
5532 case ISD::MULHS:
5533 case ISD::SMIN:
5534 case ISD::SMAX:
5535 case ISD::UMIN:
5536 case ISD::UMAX:
5537 case ISD::AND:
5538 case ISD::XOR:
5539 case ISD::ROTL:
5540 case ISD::ROTR:
5541 case ISD::FSHL:
5542 case ISD::FSHR:
5543 case ISD::BSWAP:
5544 case ISD::CTPOP:
5545 case ISD::BITREVERSE:
5546 case ISD::PARITY:
5547 case ISD::SIGN_EXTEND:
5548 case ISD::TRUNCATE:
5549 case ISD::SIGN_EXTEND_INREG:
5550 case ISD::SIGN_EXTEND_VECTOR_INREG:
5551 case ISD::ZERO_EXTEND_VECTOR_INREG:
5552 case ISD::BITCAST:
5553 case ISD::BUILD_VECTOR:
5554 case ISD::BUILD_PAIR:
5555 case ISD::SPLAT_VECTOR:
5556 case ISD::VSELECT:
5557 return false;
5558
5559 case ISD::SELECT_CC:
5560 case ISD::SETCC: {
5561 // Integer setcc cannot create undef or poison.
5562 if (Op.getOperand(i: 0).getValueType().isInteger())
5563 return false;
5564
5565 // FP compares are more complicated. They can create poison for nan/infinity
5566 // based on options and flags. The options and flags also cause special
5567 // nonan condition codes to be used. Those condition codes may be preserved
5568 // even if the nonan flag is dropped somewhere.
5569 unsigned CCOp = Opcode == ISD::SETCC ? 2 : 4;
5570 ISD::CondCode CCCode = cast<CondCodeSDNode>(Val: Op.getOperand(i: CCOp))->get();
5571 if (((unsigned)CCCode & 0x10U))
5572 return true;
5573
5574 const TargetOptions &Options = getTarget().Options;
5575 return Options.NoNaNsFPMath || Options.NoInfsFPMath;
5576 }
5577
5578 case ISD::OR:
5579 case ISD::ZERO_EXTEND:
5580 case ISD::ADD:
5581 case ISD::SUB:
5582 case ISD::MUL:
5583 case ISD::FNEG:
5584 case ISD::FADD:
5585 case ISD::FSUB:
5586 case ISD::FMUL:
5587 case ISD::FDIV:
5588 case ISD::FREM:
5589 // No poison except from flags (which is handled above)
5590 return false;
5591
5592 case ISD::SHL:
5593 case ISD::SRL:
5594 case ISD::SRA:
5595 // If the max shift amount isn't in range, then the shift can
5596 // create poison.
5597 return !isGuaranteedNotToBeUndefOrPoison(Op: Op.getOperand(i: 1), DemandedElts,
5598 PoisonOnly, Depth: Depth + 1) ||
5599 !getValidMaximumShiftAmount(V: Op, DemandedElts, Depth: Depth + 1);
5600
5601 case ISD::SCALAR_TO_VECTOR:
5602 // Check if we demand any upper (undef) elements.
5603 return !PoisonOnly && DemandedElts.ugt(RHS: 1);
5604
5605 case ISD::INSERT_VECTOR_ELT:
5606 case ISD::EXTRACT_VECTOR_ELT: {
5607 // Ensure that the element index is in bounds.
5608 EVT VecVT = Op.getOperand(i: 0).getValueType();
5609 SDValue Idx = Op.getOperand(i: Opcode == ISD::INSERT_VECTOR_ELT ? 2 : 1);
5610 if (isGuaranteedNotToBeUndefOrPoison(Op: Idx, DemandedElts, PoisonOnly,
5611 Depth: Depth + 1)) {
5612 KnownBits KnownIdx = computeKnownBits(Op: Idx, Depth: Depth + 1);
5613 return KnownIdx.getMaxValue().uge(RHS: VecVT.getVectorMinNumElements());
5614 }
5615 return true;
5616 }
5617
5618 case ISD::VECTOR_SHUFFLE: {
5619 // Check for any demanded shuffle element that is undef.
5620 auto *SVN = cast<ShuffleVectorSDNode>(Val&: Op);
5621 for (auto [Idx, Elt] : enumerate(First: SVN->getMask()))
5622 if (Elt < 0 && DemandedElts[Idx])
5623 return true;
5624 return false;
5625 }
5626
5627 default:
5628 // Allow the target to implement this method for its nodes.
5629 if (Opcode >= ISD::BUILTIN_OP_END || Opcode == ISD::INTRINSIC_WO_CHAIN ||
5630 Opcode == ISD::INTRINSIC_W_CHAIN || Opcode == ISD::INTRINSIC_VOID)
5631 return TLI->canCreateUndefOrPoisonForTargetNode(
5632 Op, DemandedElts, DAG: *this, PoisonOnly, ConsiderFlags, Depth);
5633 break;
5634 }
5635
5636 // Be conservative and return true.
5637 return true;
5638}
5639
5640bool SelectionDAG::isADDLike(SDValue Op, bool NoWrap) const {
5641 unsigned Opcode = Op.getOpcode();
5642 if (Opcode == ISD::OR)
5643 return Op->getFlags().hasDisjoint() ||
5644 haveNoCommonBitsSet(A: Op.getOperand(i: 0), B: Op.getOperand(i: 1));
5645 if (Opcode == ISD::XOR)
5646 return !NoWrap && isMinSignedConstant(V: Op.getOperand(i: 1));
5647 return false;
5648}
5649
5650bool SelectionDAG::isBaseWithConstantOffset(SDValue Op) const {
5651 return Op.getNumOperands() == 2 && isa<ConstantSDNode>(Val: Op.getOperand(i: 1)) &&
5652 (Op.isAnyAdd() || isADDLike(Op));
5653}
5654
5655bool SelectionDAG::isKnownNeverNaN(SDValue Op, bool SNaN,
5656 unsigned Depth) const {
5657 EVT VT = Op.getValueType();
5658
5659 // Since the number of lanes in a scalable vector is unknown at compile time,
5660 // we track one bit which is implicitly broadcast to all lanes. This means
5661 // that all lanes in a scalable vector are considered demanded.
5662 APInt DemandedElts = VT.isFixedLengthVector()
5663 ? APInt::getAllOnes(numBits: VT.getVectorNumElements())
5664 : APInt(1, 1);
5665
5666 return isKnownNeverNaN(Op, DemandedElts, SNaN, Depth);
5667}
5668
5669bool SelectionDAG::isKnownNeverNaN(SDValue Op, const APInt &DemandedElts,
5670 bool SNaN, unsigned Depth) const {
5671 assert(!DemandedElts.isZero() && "No demanded elements");
5672
5673 // If we're told that NaNs won't happen, assume they won't.
5674 if (getTarget().Options.NoNaNsFPMath || Op->getFlags().hasNoNaNs())
5675 return true;
5676
5677 if (Depth >= MaxRecursionDepth)
5678 return false; // Limit search depth.
5679
5680 // If the value is a constant, we can obviously see if it is a NaN or not.
5681 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val&: Op)) {
5682 return !C->getValueAPF().isNaN() ||
5683 (SNaN && !C->getValueAPF().isSignaling());
5684 }
5685
5686 unsigned Opcode = Op.getOpcode();
5687 switch (Opcode) {
5688 case ISD::FADD:
5689 case ISD::FSUB:
5690 case ISD::FMUL:
5691 case ISD::FDIV:
5692 case ISD::FREM:
5693 case ISD::FSIN:
5694 case ISD::FCOS:
5695 case ISD::FTAN:
5696 case ISD::FASIN:
5697 case ISD::FACOS:
5698 case ISD::FATAN:
5699 case ISD::FATAN2:
5700 case ISD::FSINH:
5701 case ISD::FCOSH:
5702 case ISD::FTANH:
5703 case ISD::FMA:
5704 case ISD::FMAD: {
5705 if (SNaN)
5706 return true;
5707 // TODO: Need isKnownNeverInfinity
5708 return false;
5709 }
5710 case ISD::FCANONICALIZE:
5711 case ISD::FEXP:
5712 case ISD::FEXP2:
5713 case ISD::FEXP10:
5714 case ISD::FTRUNC:
5715 case ISD::FFLOOR:
5716 case ISD::FCEIL:
5717 case ISD::FROUND:
5718 case ISD::FROUNDEVEN:
5719 case ISD::LROUND:
5720 case ISD::LLROUND:
5721 case ISD::FRINT:
5722 case ISD::LRINT:
5723 case ISD::LLRINT:
5724 case ISD::FNEARBYINT:
5725 case ISD::FLDEXP: {
5726 if (SNaN)
5727 return true;
5728 return isKnownNeverNaN(Op: Op.getOperand(i: 0), DemandedElts, SNaN, Depth: Depth + 1);
5729 }
5730 case ISD::FABS:
5731 case ISD::FNEG:
5732 case ISD::FCOPYSIGN: {
5733 return isKnownNeverNaN(Op: Op.getOperand(i: 0), DemandedElts, SNaN, Depth: Depth + 1);
5734 }
5735 case ISD::SELECT:
5736 return isKnownNeverNaN(Op: Op.getOperand(i: 1), DemandedElts, SNaN, Depth: Depth + 1) &&
5737 isKnownNeverNaN(Op: Op.getOperand(i: 2), DemandedElts, SNaN, Depth: Depth + 1);
5738 case ISD::FP_EXTEND:
5739 case ISD::FP_ROUND: {
5740 if (SNaN)
5741 return true;
5742 return isKnownNeverNaN(Op: Op.getOperand(i: 0), DemandedElts, SNaN, Depth: Depth + 1);
5743 }
5744 case ISD::SINT_TO_FP:
5745 case ISD::UINT_TO_FP:
5746 return true;
5747 case ISD::FSQRT: // Need is known positive
5748 case ISD::FLOG:
5749 case ISD::FLOG2:
5750 case ISD::FLOG10:
5751 case ISD::FPOWI:
5752 case ISD::FPOW: {
5753 if (SNaN)
5754 return true;
5755 // TODO: Refine on operand
5756 return false;
5757 }
5758 case ISD::FMINNUM:
5759 case ISD::FMAXNUM:
5760 case ISD::FMINIMUMNUM:
5761 case ISD::FMAXIMUMNUM: {
5762 // Only one needs to be known not-nan, since it will be returned if the
5763 // other ends up being one.
5764 return isKnownNeverNaN(Op: Op.getOperand(i: 0), DemandedElts, SNaN, Depth: Depth + 1) ||
5765 isKnownNeverNaN(Op: Op.getOperand(i: 1), DemandedElts, SNaN, Depth: Depth + 1);
5766 }
5767 case ISD::FMINNUM_IEEE:
5768 case ISD::FMAXNUM_IEEE: {
5769 if (SNaN)
5770 return true;
5771 // This can return a NaN if either operand is an sNaN, or if both operands
5772 // are NaN.
5773 return (isKnownNeverNaN(Op: Op.getOperand(i: 0), DemandedElts, SNaN: false, Depth: Depth + 1) &&
5774 isKnownNeverSNaN(Op: Op.getOperand(i: 1), DemandedElts, Depth: Depth + 1)) ||
5775 (isKnownNeverNaN(Op: Op.getOperand(i: 1), DemandedElts, SNaN: false, Depth: Depth + 1) &&
5776 isKnownNeverSNaN(Op: Op.getOperand(i: 0), DemandedElts, Depth: Depth + 1));
5777 }
5778 case ISD::FMINIMUM:
5779 case ISD::FMAXIMUM: {
5780 // TODO: Does this quiet or return the origina NaN as-is?
5781 return isKnownNeverNaN(Op: Op.getOperand(i: 0), DemandedElts, SNaN, Depth: Depth + 1) &&
5782 isKnownNeverNaN(Op: Op.getOperand(i: 1), DemandedElts, SNaN, Depth: Depth + 1);
5783 }
5784 case ISD::EXTRACT_VECTOR_ELT: {
5785 SDValue Src = Op.getOperand(i: 0);
5786 auto *Idx = dyn_cast<ConstantSDNode>(Val: Op.getOperand(i: 1));
5787 EVT SrcVT = Src.getValueType();
5788 if (SrcVT.isFixedLengthVector() && Idx &&
5789 Idx->getAPIntValue().ult(RHS: SrcVT.getVectorNumElements())) {
5790 APInt DemandedSrcElts = APInt::getOneBitSet(numBits: SrcVT.getVectorNumElements(),
5791 BitNo: Idx->getZExtValue());
5792 return isKnownNeverNaN(Op: Src, DemandedElts: DemandedSrcElts, SNaN, Depth: Depth + 1);
5793 }
5794 return isKnownNeverNaN(Op: Src, SNaN, Depth: Depth + 1);
5795 }
5796 case ISD::EXTRACT_SUBVECTOR: {
5797 SDValue Src = Op.getOperand(i: 0);
5798 if (Src.getValueType().isFixedLengthVector()) {
5799 unsigned Idx = Op.getConstantOperandVal(i: 1);
5800 unsigned NumSrcElts = Src.getValueType().getVectorNumElements();
5801 APInt DemandedSrcElts = DemandedElts.zext(width: NumSrcElts).shl(shiftAmt: Idx);
5802 return isKnownNeverNaN(Op: Src, DemandedElts: DemandedSrcElts, SNaN, Depth: Depth + 1);
5803 }
5804 return isKnownNeverNaN(Op: Src, SNaN, Depth: Depth + 1);
5805 }
5806 case ISD::INSERT_SUBVECTOR: {
5807 SDValue BaseVector = Op.getOperand(i: 0);
5808 SDValue SubVector = Op.getOperand(i: 1);
5809 EVT BaseVectorVT = BaseVector.getValueType();
5810 if (BaseVectorVT.isFixedLengthVector()) {
5811 unsigned Idx = Op.getConstantOperandVal(i: 2);
5812 unsigned NumBaseElts = BaseVectorVT.getVectorNumElements();
5813 unsigned NumSubElts = SubVector.getValueType().getVectorNumElements();
5814
5815 // Clear/Extract the bits at the position where the subvector will be
5816 // inserted.
5817 APInt DemandedMask =
5818 APInt::getBitsSet(numBits: NumBaseElts, loBit: Idx, hiBit: Idx + NumSubElts);
5819 APInt DemandedSrcElts = DemandedElts & ~DemandedMask;
5820 APInt DemandedSubElts = DemandedElts.extractBits(numBits: NumSubElts, bitPosition: Idx);
5821
5822 bool NeverNaN = true;
5823 if (!DemandedSrcElts.isZero())
5824 NeverNaN &=
5825 isKnownNeverNaN(Op: BaseVector, DemandedElts: DemandedSrcElts, SNaN, Depth: Depth + 1);
5826 if (NeverNaN && !DemandedSubElts.isZero())
5827 NeverNaN &=
5828 isKnownNeverNaN(Op: SubVector, DemandedElts: DemandedSubElts, SNaN, Depth: Depth + 1);
5829 return NeverNaN;
5830 }
5831 return isKnownNeverNaN(Op: BaseVector, SNaN, Depth: Depth + 1) &&
5832 isKnownNeverNaN(Op: SubVector, SNaN, Depth: Depth + 1);
5833 }
5834 case ISD::BUILD_VECTOR: {
5835 unsigned NumElts = Op.getNumOperands();
5836 for (unsigned I = 0; I != NumElts; ++I)
5837 if (DemandedElts[I] &&
5838 !isKnownNeverNaN(Op: Op.getOperand(i: I), SNaN, Depth: Depth + 1))
5839 return false;
5840 return true;
5841 }
5842 case ISD::AssertNoFPClass: {
5843 FPClassTest NoFPClass =
5844 static_cast<FPClassTest>(Op.getConstantOperandVal(i: 1));
5845 if ((NoFPClass & fcNan) == fcNan)
5846 return true;
5847 if (SNaN && (NoFPClass & fcSNan) == fcSNan)
5848 return true;
5849 return isKnownNeverNaN(Op: Op.getOperand(i: 0), DemandedElts, SNaN, Depth: Depth + 1);
5850 }
5851 default:
5852 if (Opcode >= ISD::BUILTIN_OP_END || Opcode == ISD::INTRINSIC_WO_CHAIN ||
5853 Opcode == ISD::INTRINSIC_W_CHAIN || Opcode == ISD::INTRINSIC_VOID) {
5854 return TLI->isKnownNeverNaNForTargetNode(Op, DemandedElts, DAG: *this, SNaN,
5855 Depth);
5856 }
5857
5858 return false;
5859 }
5860}
5861
5862bool SelectionDAG::isKnownNeverZeroFloat(SDValue Op) const {
5863 assert(Op.getValueType().isFloatingPoint() &&
5864 "Floating point type expected");
5865
5866 // If the value is a constant, we can obviously see if it is a zero or not.
5867 return ISD::matchUnaryFpPredicate(
5868 Op, Match: [](ConstantFPSDNode *C) { return !C->isZero(); });
5869}
5870
5871bool SelectionDAG::isKnownNeverZero(SDValue Op, unsigned Depth) const {
5872 if (Depth >= MaxRecursionDepth)
5873 return false; // Limit search depth.
5874
5875 assert(!Op.getValueType().isFloatingPoint() &&
5876 "Floating point types unsupported - use isKnownNeverZeroFloat");
5877
5878 // If the value is a constant, we can obviously see if it is a zero or not.
5879 if (ISD::matchUnaryPredicate(Op,
5880 Match: [](ConstantSDNode *C) { return !C->isZero(); }))
5881 return true;
5882
5883 // TODO: Recognize more cases here. Most of the cases are also incomplete to
5884 // some degree.
5885 switch (Op.getOpcode()) {
5886 default:
5887 break;
5888
5889 case ISD::OR:
5890 return isKnownNeverZero(Op: Op.getOperand(i: 1), Depth: Depth + 1) ||
5891 isKnownNeverZero(Op: Op.getOperand(i: 0), Depth: Depth + 1);
5892
5893 case ISD::VSELECT:
5894 case ISD::SELECT:
5895 return isKnownNeverZero(Op: Op.getOperand(i: 1), Depth: Depth + 1) &&
5896 isKnownNeverZero(Op: Op.getOperand(i: 2), Depth: Depth + 1);
5897
5898 case ISD::SHL: {
5899 if (Op->getFlags().hasNoSignedWrap() || Op->getFlags().hasNoUnsignedWrap())
5900 return isKnownNeverZero(Op: Op.getOperand(i: 0), Depth: Depth + 1);
5901 KnownBits ValKnown = computeKnownBits(Op: Op.getOperand(i: 0), Depth: Depth + 1);
5902 // 1 << X is never zero.
5903 if (ValKnown.One[0])
5904 return true;
5905 // If max shift cnt of known ones is non-zero, result is non-zero.
5906 APInt MaxCnt = computeKnownBits(Op: Op.getOperand(i: 1), Depth: Depth + 1).getMaxValue();
5907 if (MaxCnt.ult(RHS: ValKnown.getBitWidth()) &&
5908 !ValKnown.One.shl(ShiftAmt: MaxCnt).isZero())
5909 return true;
5910 break;
5911 }
5912 case ISD::UADDSAT:
5913 case ISD::UMAX:
5914 return isKnownNeverZero(Op: Op.getOperand(i: 1), Depth: Depth + 1) ||
5915 isKnownNeverZero(Op: Op.getOperand(i: 0), Depth: Depth + 1);
5916
5917 // For smin/smax: If either operand is known negative/positive
5918 // respectively we don't need the other to be known at all.
5919 case ISD::SMAX: {
5920 KnownBits Op1 = computeKnownBits(Op: Op.getOperand(i: 1), Depth: Depth + 1);
5921 if (Op1.isStrictlyPositive())
5922 return true;
5923
5924 KnownBits Op0 = computeKnownBits(Op: Op.getOperand(i: 0), Depth: Depth + 1);
5925 if (Op0.isStrictlyPositive())
5926 return true;
5927
5928 if (Op1.isNonZero() && Op0.isNonZero())
5929 return true;
5930
5931 return isKnownNeverZero(Op: Op.getOperand(i: 1), Depth: Depth + 1) &&
5932 isKnownNeverZero(Op: Op.getOperand(i: 0), Depth: Depth + 1);
5933 }
5934 case ISD::SMIN: {
5935 KnownBits Op1 = computeKnownBits(Op: Op.getOperand(i: 1), Depth: Depth + 1);
5936 if (Op1.isNegative())
5937 return true;
5938
5939 KnownBits Op0 = computeKnownBits(Op: Op.getOperand(i: 0), Depth: Depth + 1);
5940 if (Op0.isNegative())
5941 return true;
5942
5943 if (Op1.isNonZero() && Op0.isNonZero())
5944 return true;
5945
5946 return isKnownNeverZero(Op: Op.getOperand(i: 1), Depth: Depth + 1) &&
5947 isKnownNeverZero(Op: Op.getOperand(i: 0), Depth: Depth + 1);
5948 }
5949 case ISD::UMIN:
5950 return isKnownNeverZero(Op: Op.getOperand(i: 1), Depth: Depth + 1) &&
5951 isKnownNeverZero(Op: Op.getOperand(i: 0), Depth: Depth + 1);
5952
5953 case ISD::ROTL:
5954 case ISD::ROTR:
5955 case ISD::BITREVERSE:
5956 case ISD::BSWAP:
5957 case ISD::CTPOP:
5958 case ISD::ABS:
5959 return isKnownNeverZero(Op: Op.getOperand(i: 0), Depth: Depth + 1);
5960
5961 case ISD::SRA:
5962 case ISD::SRL: {
5963 if (Op->getFlags().hasExact())
5964 return isKnownNeverZero(Op: Op.getOperand(i: 0), Depth: Depth + 1);
5965 KnownBits ValKnown = computeKnownBits(Op: Op.getOperand(i: 0), Depth: Depth + 1);
5966 if (ValKnown.isNegative())
5967 return true;
5968 // If max shift cnt of known ones is non-zero, result is non-zero.
5969 APInt MaxCnt = computeKnownBits(Op: Op.getOperand(i: 1), Depth: Depth + 1).getMaxValue();
5970 if (MaxCnt.ult(RHS: ValKnown.getBitWidth()) &&
5971 !ValKnown.One.lshr(ShiftAmt: MaxCnt).isZero())
5972 return true;
5973 break;
5974 }
5975 case ISD::UDIV:
5976 case ISD::SDIV:
5977 // div exact can only produce a zero if the dividend is zero.
5978 // TODO: For udiv this is also true if Op1 u<= Op0
5979 if (Op->getFlags().hasExact())
5980 return isKnownNeverZero(Op: Op.getOperand(i: 0), Depth: Depth + 1);
5981 break;
5982
5983 case ISD::ADD:
5984 if (Op->getFlags().hasNoUnsignedWrap())
5985 if (isKnownNeverZero(Op: Op.getOperand(i: 1), Depth: Depth + 1) ||
5986 isKnownNeverZero(Op: Op.getOperand(i: 0), Depth: Depth + 1))
5987 return true;
5988 // TODO: There are a lot more cases we can prove for add.
5989 break;
5990
5991 case ISD::SUB: {
5992 if (isNullConstant(V: Op.getOperand(i: 0)))
5993 return isKnownNeverZero(Op: Op.getOperand(i: 1), Depth: Depth + 1);
5994
5995 std::optional<bool> ne =
5996 KnownBits::ne(LHS: computeKnownBits(Op: Op.getOperand(i: 0), Depth: Depth + 1),
5997 RHS: computeKnownBits(Op: Op.getOperand(i: 1), Depth: Depth + 1));
5998 return ne && *ne;
5999 }
6000
6001 case ISD::MUL:
6002 if (Op->getFlags().hasNoSignedWrap() || Op->getFlags().hasNoUnsignedWrap())
6003 if (isKnownNeverZero(Op: Op.getOperand(i: 1), Depth: Depth + 1) &&
6004 isKnownNeverZero(Op: Op.getOperand(i: 0), Depth: Depth + 1))
6005 return true;
6006 break;
6007
6008 case ISD::ZERO_EXTEND:
6009 case ISD::SIGN_EXTEND:
6010 return isKnownNeverZero(Op: Op.getOperand(i: 0), Depth: Depth + 1);
6011 case ISD::VSCALE: {
6012 const Function &F = getMachineFunction().getFunction();
6013 const APInt &Multiplier = Op.getConstantOperandAPInt(i: 0);
6014 ConstantRange CR =
6015 getVScaleRange(F: &F, BitWidth: Op.getScalarValueSizeInBits()).multiply(Other: Multiplier);
6016 if (!CR.contains(Val: APInt(CR.getBitWidth(), 0)))
6017 return true;
6018 break;
6019 }
6020 }
6021
6022 return computeKnownBits(Op, Depth).isNonZero();
6023}
6024
6025bool SelectionDAG::cannotBeOrderedNegativeFP(SDValue Op) const {
6026 if (ConstantFPSDNode *C1 = isConstOrConstSplatFP(N: Op, AllowUndefs: true))
6027 return !C1->isNegative();
6028
6029 return Op.getOpcode() == ISD::FABS;
6030}
6031
6032bool SelectionDAG::isEqualTo(SDValue A, SDValue B) const {
6033 // Check the obvious case.
6034 if (A == B) return true;
6035
6036 // For negative and positive zero.
6037 if (const ConstantFPSDNode *CA = dyn_cast<ConstantFPSDNode>(Val&: A))
6038 if (const ConstantFPSDNode *CB = dyn_cast<ConstantFPSDNode>(Val&: B))
6039 if (CA->isZero() && CB->isZero()) return true;
6040
6041 // Otherwise they may not be equal.
6042 return false;
6043}
6044
6045// Only bits set in Mask must be negated, other bits may be arbitrary.
6046SDValue llvm::getBitwiseNotOperand(SDValue V, SDValue Mask, bool AllowUndefs) {
6047 if (isBitwiseNot(V, AllowUndefs))
6048 return V.getOperand(i: 0);
6049
6050 // Handle any_extend (not (truncate X)) pattern, where Mask only sets
6051 // bits in the non-extended part.
6052 ConstantSDNode *MaskC = isConstOrConstSplat(N: Mask);
6053 if (!MaskC || V.getOpcode() != ISD::ANY_EXTEND)
6054 return SDValue();
6055 SDValue ExtArg = V.getOperand(i: 0);
6056 if (ExtArg.getScalarValueSizeInBits() >=
6057 MaskC->getAPIntValue().getActiveBits() &&
6058 isBitwiseNot(V: ExtArg, AllowUndefs) &&
6059 ExtArg.getOperand(i: 0).getOpcode() == ISD::TRUNCATE &&
6060 ExtArg.getOperand(i: 0).getOperand(i: 0).getValueType() == V.getValueType())
6061 return ExtArg.getOperand(i: 0).getOperand(i: 0);
6062 return SDValue();
6063}
6064
6065static bool haveNoCommonBitsSetCommutative(SDValue A, SDValue B) {
6066 // Match masked merge pattern (X & ~M) op (Y & M)
6067 // Including degenerate case (X & ~M) op M
6068 auto MatchNoCommonBitsPattern = [&](SDValue Not, SDValue Mask,
6069 SDValue Other) {
6070 if (SDValue NotOperand =
6071 getBitwiseNotOperand(V: Not, Mask, /* AllowUndefs */ true)) {
6072 if (NotOperand->getOpcode() == ISD::ZERO_EXTEND ||
6073 NotOperand->getOpcode() == ISD::TRUNCATE)
6074 NotOperand = NotOperand->getOperand(Num: 0);
6075
6076 if (Other == NotOperand)
6077 return true;
6078 if (Other->getOpcode() == ISD::AND)
6079 return NotOperand == Other->getOperand(Num: 0) ||
6080 NotOperand == Other->getOperand(Num: 1);
6081 }
6082 return false;
6083 };
6084
6085 if (A->getOpcode() == ISD::ZERO_EXTEND || A->getOpcode() == ISD::TRUNCATE)
6086 A = A->getOperand(Num: 0);
6087
6088 if (B->getOpcode() == ISD::ZERO_EXTEND || B->getOpcode() == ISD::TRUNCATE)
6089 B = B->getOperand(Num: 0);
6090
6091 if (A->getOpcode() == ISD::AND)
6092 return MatchNoCommonBitsPattern(A->getOperand(Num: 0), A->getOperand(Num: 1), B) ||
6093 MatchNoCommonBitsPattern(A->getOperand(Num: 1), A->getOperand(Num: 0), B);
6094 return false;
6095}
6096
6097// FIXME: unify with llvm::haveNoCommonBitsSet.
6098bool SelectionDAG::haveNoCommonBitsSet(SDValue A, SDValue B) const {
6099 assert(A.getValueType() == B.getValueType() &&
6100 "Values must have the same type");
6101 if (haveNoCommonBitsSetCommutative(A, B) ||
6102 haveNoCommonBitsSetCommutative(A: B, B: A))
6103 return true;
6104 return KnownBits::haveNoCommonBitsSet(LHS: computeKnownBits(Op: A),
6105 RHS: computeKnownBits(Op: B));
6106}
6107
6108static SDValue FoldSTEP_VECTOR(const SDLoc &DL, EVT VT, SDValue Step,
6109 SelectionDAG &DAG) {
6110 if (cast<ConstantSDNode>(Val&: Step)->isZero())
6111 return DAG.getConstant(Val: 0, DL, VT);
6112
6113 return SDValue();
6114}
6115
6116static SDValue FoldBUILD_VECTOR(const SDLoc &DL, EVT VT,
6117 ArrayRef<SDValue> Ops,
6118 SelectionDAG &DAG) {
6119 int NumOps = Ops.size();
6120 assert(NumOps != 0 && "Can't build an empty vector!");
6121 assert(!VT.isScalableVector() &&
6122 "BUILD_VECTOR cannot be used with scalable types");
6123 assert(VT.getVectorNumElements() == (unsigned)NumOps &&
6124 "Incorrect element count in BUILD_VECTOR!");
6125
6126 // BUILD_VECTOR of UNDEFs is UNDEF.
6127 if (llvm::all_of(Range&: Ops, P: [](SDValue Op) { return Op.isUndef(); }))
6128 return DAG.getUNDEF(VT);
6129
6130 // BUILD_VECTOR of seq extract/insert from the same vector + type is Identity.
6131 SDValue IdentitySrc;
6132 bool IsIdentity = true;
6133 for (int i = 0; i != NumOps; ++i) {
6134 if (Ops[i].getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6135 Ops[i].getOperand(i: 0).getValueType() != VT ||
6136 (IdentitySrc && Ops[i].getOperand(i: 0) != IdentitySrc) ||
6137 !isa<ConstantSDNode>(Val: Ops[i].getOperand(i: 1)) ||
6138 Ops[i].getConstantOperandAPInt(i: 1) != i) {
6139 IsIdentity = false;
6140 break;
6141 }
6142 IdentitySrc = Ops[i].getOperand(i: 0);
6143 }
6144 if (IsIdentity)
6145 return IdentitySrc;
6146
6147 return SDValue();
6148}
6149
6150/// Try to simplify vector concatenation to an input value, undef, or build
6151/// vector.
6152static SDValue foldCONCAT_VECTORS(const SDLoc &DL, EVT VT,
6153 ArrayRef<SDValue> Ops,
6154 SelectionDAG &DAG) {
6155 assert(!Ops.empty() && "Can't concatenate an empty list of vectors!");
6156 assert(llvm::all_of(Ops,
6157 [Ops](SDValue Op) {
6158 return Ops[0].getValueType() == Op.getValueType();
6159 }) &&
6160 "Concatenation of vectors with inconsistent value types!");
6161 assert((Ops[0].getValueType().getVectorElementCount() * Ops.size()) ==
6162 VT.getVectorElementCount() &&
6163 "Incorrect element count in vector concatenation!");
6164
6165 if (Ops.size() == 1)
6166 return Ops[0];
6167
6168 // Concat of UNDEFs is UNDEF.
6169 if (llvm::all_of(Range&: Ops, P: [](SDValue Op) { return Op.isUndef(); }))
6170 return DAG.getUNDEF(VT);
6171
6172 // Scan the operands and look for extract operations from a single source
6173 // that correspond to insertion at the same location via this concatenation:
6174 // concat (extract X, 0*subvec_elts), (extract X, 1*subvec_elts), ...
6175 SDValue IdentitySrc;
6176 bool IsIdentity = true;
6177 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
6178 SDValue Op = Ops[i];
6179 unsigned IdentityIndex = i * Op.getValueType().getVectorMinNumElements();
6180 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR ||
6181 Op.getOperand(i: 0).getValueType() != VT ||
6182 (IdentitySrc && Op.getOperand(i: 0) != IdentitySrc) ||
6183 Op.getConstantOperandVal(i: 1) != IdentityIndex) {
6184 IsIdentity = false;
6185 break;
6186 }
6187 assert((!IdentitySrc || IdentitySrc == Op.getOperand(0)) &&
6188 "Unexpected identity source vector for concat of extracts");
6189 IdentitySrc = Op.getOperand(i: 0);
6190 }
6191 if (IsIdentity) {
6192 assert(IdentitySrc && "Failed to set source vector of extracts");
6193 return IdentitySrc;
6194 }
6195
6196 // The code below this point is only designed to work for fixed width
6197 // vectors, so we bail out for now.
6198 if (VT.isScalableVector())
6199 return SDValue();
6200
6201 // A CONCAT_VECTOR with all UNDEF/BUILD_VECTOR operands can be
6202 // simplified to one big BUILD_VECTOR.
6203 // FIXME: Add support for SCALAR_TO_VECTOR as well.
6204 EVT SVT = VT.getScalarType();
6205 SmallVector<SDValue, 16> Elts;
6206 for (SDValue Op : Ops) {
6207 EVT OpVT = Op.getValueType();
6208 if (Op.isUndef())
6209 Elts.append(NumInputs: OpVT.getVectorNumElements(), Elt: DAG.getUNDEF(VT: SVT));
6210 else if (Op.getOpcode() == ISD::BUILD_VECTOR)
6211 Elts.append(in_start: Op->op_begin(), in_end: Op->op_end());
6212 else
6213 return SDValue();
6214 }
6215
6216 // BUILD_VECTOR requires all inputs to be of the same type, find the
6217 // maximum type and extend them all.
6218 for (SDValue Op : Elts)
6219 SVT = (SVT.bitsLT(VT: Op.getValueType()) ? Op.getValueType() : SVT);
6220
6221 if (SVT.bitsGT(VT: VT.getScalarType())) {
6222 for (SDValue &Op : Elts) {
6223 if (Op.isUndef())
6224 Op = DAG.getUNDEF(VT: SVT);
6225 else
6226 Op = DAG.getTargetLoweringInfo().isZExtFree(FromTy: Op.getValueType(), ToTy: SVT)
6227 ? DAG.getZExtOrTrunc(Op, DL, VT: SVT)
6228 : DAG.getSExtOrTrunc(Op, DL, VT: SVT);
6229 }
6230 }
6231
6232 SDValue V = DAG.getBuildVector(VT, DL, Ops: Elts);
6233 NewSDValueDbgMsg(V, Msg: "New node fold concat vectors: ", G: &DAG);
6234 return V;
6235}
6236
6237/// Gets or creates the specified node.
6238SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT) {
6239 SDVTList VTs = getVTList(VT);
6240 FoldingSetNodeID ID;
6241 AddNodeIDNode(ID, OpC: Opcode, VTList: VTs, OpList: {});
6242 void *IP = nullptr;
6243 if (SDNode *E = FindNodeOrInsertPos(ID, DL, InsertPos&: IP))
6244 return SDValue(E, 0);
6245
6246 auto *N = newSDNode<SDNode>(Args&: Opcode, Args: DL.getIROrder(), Args: DL.getDebugLoc(), Args&: VTs);
6247 CSEMap.InsertNode(N, InsertPos: IP);
6248
6249 InsertNode(N);
6250 SDValue V = SDValue(N, 0);
6251 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
6252 return V;
6253}
6254
6255SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
6256 SDValue N1) {
6257 SDNodeFlags Flags;
6258 if (Inserter)
6259 Flags = Inserter->getFlags();
6260 return getNode(Opcode, DL, VT, Operand: N1, Flags);
6261}
6262
6263SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
6264 SDValue N1, const SDNodeFlags Flags) {
6265 assert(N1.getOpcode() != ISD::DELETED_NODE && "Operand is DELETED_NODE!");
6266
6267 // Constant fold unary operations with a vector integer or float operand.
6268 switch (Opcode) {
6269 default:
6270 // FIXME: Entirely reasonable to perform folding of other unary
6271 // operations here as the need arises.
6272 break;
6273 case ISD::FNEG:
6274 case ISD::FABS:
6275 case ISD::FCEIL:
6276 case ISD::FTRUNC:
6277 case ISD::FFLOOR:
6278 case ISD::FP_EXTEND:
6279 case ISD::FP_TO_SINT:
6280 case ISD::FP_TO_UINT:
6281 case ISD::FP_TO_FP16:
6282 case ISD::FP_TO_BF16:
6283 case ISD::TRUNCATE:
6284 case ISD::ANY_EXTEND:
6285 case ISD::ZERO_EXTEND:
6286 case ISD::SIGN_EXTEND:
6287 case ISD::UINT_TO_FP:
6288 case ISD::SINT_TO_FP:
6289 case ISD::FP16_TO_FP:
6290 case ISD::BF16_TO_FP:
6291 case ISD::BITCAST:
6292 case ISD::ABS:
6293 case ISD::BITREVERSE:
6294 case ISD::BSWAP:
6295 case ISD::CTLZ:
6296 case ISD::CTLZ_ZERO_UNDEF:
6297 case ISD::CTTZ:
6298 case ISD::CTTZ_ZERO_UNDEF:
6299 case ISD::CTPOP:
6300 case ISD::STEP_VECTOR: {
6301 SDValue Ops = {N1};
6302 if (SDValue Fold = FoldConstantArithmetic(Opcode, DL, VT, Ops))
6303 return Fold;
6304 }
6305 }
6306
6307 unsigned OpOpcode = N1.getNode()->getOpcode();
6308 switch (Opcode) {
6309 case ISD::STEP_VECTOR:
6310 assert(VT.isScalableVector() &&
6311 "STEP_VECTOR can only be used with scalable types");
6312 assert(OpOpcode == ISD::TargetConstant &&
6313 VT.getVectorElementType() == N1.getValueType() &&
6314 "Unexpected step operand");
6315 break;
6316 case ISD::FREEZE:
6317 assert(VT == N1.getValueType() && "Unexpected VT!");
6318 if (isGuaranteedNotToBeUndefOrPoison(Op: N1, /*PoisonOnly*/ false,
6319 /*Depth*/ 1))
6320 return N1;
6321 break;
6322 case ISD::TokenFactor:
6323 case ISD::MERGE_VALUES:
6324 case ISD::CONCAT_VECTORS:
6325 return N1; // Factor, merge or concat of one node? No need.
6326 case ISD::BUILD_VECTOR: {
6327 // Attempt to simplify BUILD_VECTOR.
6328 SDValue Ops[] = {N1};
6329 if (SDValue V = FoldBUILD_VECTOR(DL, VT, Ops, DAG&: *this))
6330 return V;
6331 break;
6332 }
6333 case ISD::FP_ROUND: llvm_unreachable("Invalid method to make FP_ROUND node");
6334 case ISD::FP_EXTEND:
6335 assert(VT.isFloatingPoint() && N1.getValueType().isFloatingPoint() &&
6336 "Invalid FP cast!");
6337 if (N1.getValueType() == VT) return N1; // noop conversion.
6338 assert((!VT.isVector() || VT.getVectorElementCount() ==
6339 N1.getValueType().getVectorElementCount()) &&
6340 "Vector element count mismatch!");
6341 assert(N1.getValueType().bitsLT(VT) && "Invalid fpext node, dst < src!");
6342 if (N1.isUndef())
6343 return getUNDEF(VT);
6344 break;
6345 case ISD::FP_TO_SINT:
6346 case ISD::FP_TO_UINT:
6347 if (N1.isUndef())
6348 return getUNDEF(VT);
6349 break;
6350 case ISD::SINT_TO_FP:
6351 case ISD::UINT_TO_FP:
6352 // [us]itofp(undef) = 0, because the result value is bounded.
6353 if (N1.isUndef())
6354 return getConstantFP(Val: 0.0, DL, VT);
6355 break;
6356 case ISD::SIGN_EXTEND:
6357 assert(VT.isInteger() && N1.getValueType().isInteger() &&
6358 "Invalid SIGN_EXTEND!");
6359 assert(VT.isVector() == N1.getValueType().isVector() &&
6360 "SIGN_EXTEND result type type should be vector iff the operand "
6361 "type is vector!");
6362 if (N1.getValueType() == VT) return N1; // noop extension
6363 assert((!VT.isVector() || VT.getVectorElementCount() ==
6364 N1.getValueType().getVectorElementCount()) &&
6365 "Vector element count mismatch!");
6366 assert(N1.getValueType().bitsLT(VT) && "Invalid sext node, dst < src!");
6367 if (OpOpcode == ISD::SIGN_EXTEND || OpOpcode == ISD::ZERO_EXTEND) {
6368 SDNodeFlags Flags;
6369 if (OpOpcode == ISD::ZERO_EXTEND)
6370 Flags.setNonNeg(N1->getFlags().hasNonNeg());
6371 SDValue NewVal = getNode(Opcode: OpOpcode, DL, VT, N1: N1.getOperand(i: 0), Flags);
6372 transferDbgValues(From: N1, To: NewVal);
6373 return NewVal;
6374 }
6375
6376 if (OpOpcode == ISD::POISON)
6377 return getPOISON(VT);
6378
6379 if (N1.isUndef())
6380 // sext(undef) = 0, because the top bits will all be the same.
6381 return getConstant(Val: 0, DL, VT);
6382 break;
6383 case ISD::ZERO_EXTEND:
6384 assert(VT.isInteger() && N1.getValueType().isInteger() &&
6385 "Invalid ZERO_EXTEND!");
6386 assert(VT.isVector() == N1.getValueType().isVector() &&
6387 "ZERO_EXTEND result type type should be vector iff the operand "
6388 "type is vector!");
6389 if (N1.getValueType() == VT) return N1; // noop extension
6390 assert((!VT.isVector() || VT.getVectorElementCount() ==
6391 N1.getValueType().getVectorElementCount()) &&
6392 "Vector element count mismatch!");
6393 assert(N1.getValueType().bitsLT(VT) && "Invalid zext node, dst < src!");
6394 if (OpOpcode == ISD::ZERO_EXTEND) { // (zext (zext x)) -> (zext x)
6395 SDNodeFlags Flags;
6396 Flags.setNonNeg(N1->getFlags().hasNonNeg());
6397 SDValue NewVal =
6398 getNode(Opcode: ISD::ZERO_EXTEND, DL, VT, N1: N1.getOperand(i: 0), Flags);
6399 transferDbgValues(From: N1, To: NewVal);
6400 return NewVal;
6401 }
6402
6403 if (OpOpcode == ISD::POISON)
6404 return getPOISON(VT);
6405
6406 if (N1.isUndef())
6407 // zext(undef) = 0, because the top bits will be zero.
6408 return getConstant(Val: 0, DL, VT);
6409
6410 // Skip unnecessary zext_inreg pattern:
6411 // (zext (trunc x)) -> x iff the upper bits are known zero.
6412 // TODO: Remove (zext (trunc (and x, c))) exception which some targets
6413 // use to recognise zext_inreg patterns.
6414 if (OpOpcode == ISD::TRUNCATE) {
6415 SDValue OpOp = N1.getOperand(i: 0);
6416 if (OpOp.getValueType() == VT) {
6417 if (OpOp.getOpcode() != ISD::AND) {
6418 APInt HiBits = APInt::getBitsSetFrom(numBits: VT.getScalarSizeInBits(),
6419 loBit: N1.getScalarValueSizeInBits());
6420 if (MaskedValueIsZero(V: OpOp, Mask: HiBits)) {
6421 transferDbgValues(From: N1, To: OpOp);
6422 return OpOp;
6423 }
6424 }
6425 }
6426 }
6427 break;
6428 case ISD::ANY_EXTEND:
6429 assert(VT.isInteger() && N1.getValueType().isInteger() &&
6430 "Invalid ANY_EXTEND!");
6431 assert(VT.isVector() == N1.getValueType().isVector() &&
6432 "ANY_EXTEND result type type should be vector iff the operand "
6433 "type is vector!");
6434 if (N1.getValueType() == VT) return N1; // noop extension
6435 assert((!VT.isVector() || VT.getVectorElementCount() ==
6436 N1.getValueType().getVectorElementCount()) &&
6437 "Vector element count mismatch!");
6438 assert(N1.getValueType().bitsLT(VT) && "Invalid anyext node, dst < src!");
6439
6440 if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND ||
6441 OpOpcode == ISD::ANY_EXTEND) {
6442 SDNodeFlags Flags;
6443 if (OpOpcode == ISD::ZERO_EXTEND)
6444 Flags.setNonNeg(N1->getFlags().hasNonNeg());
6445 // (ext (zext x)) -> (zext x) and (ext (sext x)) -> (sext x)
6446 return getNode(Opcode: OpOpcode, DL, VT, N1: N1.getOperand(i: 0), Flags);
6447 }
6448 if (N1.isUndef())
6449 return getUNDEF(VT);
6450
6451 // (ext (trunc x)) -> x
6452 if (OpOpcode == ISD::TRUNCATE) {
6453 SDValue OpOp = N1.getOperand(i: 0);
6454 if (OpOp.getValueType() == VT) {
6455 transferDbgValues(From: N1, To: OpOp);
6456 return OpOp;
6457 }
6458 }
6459 break;
6460 case ISD::TRUNCATE:
6461 assert(VT.isInteger() && N1.getValueType().isInteger() &&
6462 "Invalid TRUNCATE!");
6463 assert(VT.isVector() == N1.getValueType().isVector() &&
6464 "TRUNCATE result type type should be vector iff the operand "
6465 "type is vector!");
6466 if (N1.getValueType() == VT) return N1; // noop truncate
6467 assert((!VT.isVector() || VT.getVectorElementCount() ==
6468 N1.getValueType().getVectorElementCount()) &&
6469 "Vector element count mismatch!");
6470 assert(N1.getValueType().bitsGT(VT) && "Invalid truncate node, src < dst!");
6471 if (OpOpcode == ISD::TRUNCATE)
6472 return getNode(Opcode: ISD::TRUNCATE, DL, VT, N1: N1.getOperand(i: 0));
6473 if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND ||
6474 OpOpcode == ISD::ANY_EXTEND) {
6475 // If the source is smaller than the dest, we still need an extend.
6476 if (N1.getOperand(i: 0).getValueType().getScalarType().bitsLT(
6477 VT: VT.getScalarType()))
6478 return getNode(Opcode: OpOpcode, DL, VT, N1: N1.getOperand(i: 0));
6479 if (N1.getOperand(i: 0).getValueType().bitsGT(VT))
6480 return getNode(Opcode: ISD::TRUNCATE, DL, VT, N1: N1.getOperand(i: 0));
6481 return N1.getOperand(i: 0);
6482 }
6483 if (N1.isUndef())
6484 return getUNDEF(VT);
6485 if (OpOpcode == ISD::VSCALE && !NewNodesMustHaveLegalTypes)
6486 return getVScale(DL, VT,
6487 MulImm: N1.getConstantOperandAPInt(i: 0).trunc(width: VT.getSizeInBits()));
6488 break;
6489 case ISD::ANY_EXTEND_VECTOR_INREG:
6490 case ISD::ZERO_EXTEND_VECTOR_INREG:
6491 case ISD::SIGN_EXTEND_VECTOR_INREG:
6492 assert(VT.isVector() && "This DAG node is restricted to vector types.");
6493 assert(N1.getValueType().bitsLE(VT) &&
6494 "The input must be the same size or smaller than the result.");
6495 assert(VT.getVectorMinNumElements() <
6496 N1.getValueType().getVectorMinNumElements() &&
6497 "The destination vector type must have fewer lanes than the input.");
6498 break;
6499 case ISD::ABS:
6500 assert(VT.isInteger() && VT == N1.getValueType() && "Invalid ABS!");
6501 if (N1.isUndef())
6502 return getConstant(Val: 0, DL, VT);
6503 break;
6504 case ISD::BSWAP:
6505 assert(VT.isInteger() && VT == N1.getValueType() && "Invalid BSWAP!");
6506 assert((VT.getScalarSizeInBits() % 16 == 0) &&
6507 "BSWAP types must be a multiple of 16 bits!");
6508 if (N1.isUndef())
6509 return getUNDEF(VT);
6510 // bswap(bswap(X)) -> X.
6511 if (OpOpcode == ISD::BSWAP)
6512 return N1.getOperand(i: 0);
6513 break;
6514 case ISD::BITREVERSE:
6515 assert(VT.isInteger() && VT == N1.getValueType() && "Invalid BITREVERSE!");
6516 if (N1.isUndef())
6517 return getUNDEF(VT);
6518 break;
6519 case ISD::BITCAST:
6520 assert(VT.getSizeInBits() == N1.getValueSizeInBits() &&
6521 "Cannot BITCAST between types of different sizes!");
6522 if (VT == N1.getValueType()) return N1; // noop conversion.
6523 if (OpOpcode == ISD::BITCAST) // bitconv(bitconv(x)) -> bitconv(x)
6524 return getNode(Opcode: ISD::BITCAST, DL, VT, N1: N1.getOperand(i: 0));
6525 if (N1.isUndef())
6526 return getUNDEF(VT);
6527 break;
6528 case ISD::SCALAR_TO_VECTOR:
6529 assert(VT.isVector() && !N1.getValueType().isVector() &&
6530 (VT.getVectorElementType() == N1.getValueType() ||
6531 (VT.getVectorElementType().isInteger() &&
6532 N1.getValueType().isInteger() &&
6533 VT.getVectorElementType().bitsLE(N1.getValueType()))) &&
6534 "Illegal SCALAR_TO_VECTOR node!");
6535 if (N1.isUndef())
6536 return getUNDEF(VT);
6537 // scalar_to_vector(extract_vector_elt V, 0) -> V, top bits are undefined.
6538 if (OpOpcode == ISD::EXTRACT_VECTOR_ELT &&
6539 isa<ConstantSDNode>(Val: N1.getOperand(i: 1)) &&
6540 N1.getConstantOperandVal(i: 1) == 0 &&
6541 N1.getOperand(i: 0).getValueType() == VT)
6542 return N1.getOperand(i: 0);
6543 break;
6544 case ISD::FNEG:
6545 // Negation of an unknown bag of bits is still completely undefined.
6546 if (N1.isUndef())
6547 return getUNDEF(VT);
6548
6549 if (OpOpcode == ISD::FNEG) // --X -> X
6550 return N1.getOperand(i: 0);
6551 break;
6552 case ISD::FABS:
6553 if (OpOpcode == ISD::FNEG) // abs(-X) -> abs(X)
6554 return getNode(Opcode: ISD::FABS, DL, VT, N1: N1.getOperand(i: 0));
6555 break;
6556 case ISD::VSCALE:
6557 assert(VT == N1.getValueType() && "Unexpected VT!");
6558 break;
6559 case ISD::CTPOP:
6560 if (N1.getValueType().getScalarType() == MVT::i1)
6561 return N1;
6562 break;
6563 case ISD::CTLZ:
6564 case ISD::CTTZ:
6565 if (N1.getValueType().getScalarType() == MVT::i1)
6566 return getNOT(DL, Val: N1, VT: N1.getValueType());
6567 break;
6568 case ISD::VECREDUCE_ADD:
6569 if (N1.getValueType().getScalarType() == MVT::i1)
6570 return getNode(Opcode: ISD::VECREDUCE_XOR, DL, VT, N1);
6571 break;
6572 case ISD::VECREDUCE_SMIN:
6573 case ISD::VECREDUCE_UMAX:
6574 if (N1.getValueType().getScalarType() == MVT::i1)
6575 return getNode(Opcode: ISD::VECREDUCE_OR, DL, VT, N1);
6576 break;
6577 case ISD::VECREDUCE_SMAX:
6578 case ISD::VECREDUCE_UMIN:
6579 if (N1.getValueType().getScalarType() == MVT::i1)
6580 return getNode(Opcode: ISD::VECREDUCE_AND, DL, VT, N1);
6581 break;
6582 case ISD::SPLAT_VECTOR:
6583 assert(VT.isVector() && "Wrong return type!");
6584 // FIXME: Hexagon uses i32 scalar for a floating point zero vector so allow
6585 // that for now.
6586 assert((VT.getVectorElementType() == N1.getValueType() ||
6587 (VT.isFloatingPoint() && N1.getValueType() == MVT::i32) ||
6588 (VT.getVectorElementType().isInteger() &&
6589 N1.getValueType().isInteger() &&
6590 VT.getVectorElementType().bitsLE(N1.getValueType()))) &&
6591 "Wrong operand type!");
6592 break;
6593 }
6594
6595 SDNode *N;
6596 SDVTList VTs = getVTList(VT);
6597 SDValue Ops[] = {N1};
6598 if (VT != MVT::Glue) { // Don't CSE glue producing nodes
6599 FoldingSetNodeID ID;
6600 AddNodeIDNode(ID, OpC: Opcode, VTList: VTs, OpList: Ops);
6601 void *IP = nullptr;
6602 if (SDNode *E = FindNodeOrInsertPos(ID, DL, InsertPos&: IP)) {
6603 E->intersectFlagsWith(Flags);
6604 return SDValue(E, 0);
6605 }
6606
6607 N = newSDNode<SDNode>(Args&: Opcode, Args: DL.getIROrder(), Args: DL.getDebugLoc(), Args&: VTs);
6608 N->setFlags(Flags);
6609 createOperands(Node: N, Vals: Ops);
6610 CSEMap.InsertNode(N, InsertPos: IP);
6611 } else {
6612 N = newSDNode<SDNode>(Args&: Opcode, Args: DL.getIROrder(), Args: DL.getDebugLoc(), Args&: VTs);
6613 createOperands(Node: N, Vals: Ops);
6614 }
6615
6616 InsertNode(N);
6617 SDValue V = SDValue(N, 0);
6618 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
6619 return V;
6620}
6621
6622static std::optional<APInt> FoldValue(unsigned Opcode, const APInt &C1,
6623 const APInt &C2) {
6624 switch (Opcode) {
6625 case ISD::ADD: return C1 + C2;
6626 case ISD::SUB: return C1 - C2;
6627 case ISD::MUL: return C1 * C2;
6628 case ISD::AND: return C1 & C2;
6629 case ISD::OR: return C1 | C2;
6630 case ISD::XOR: return C1 ^ C2;
6631 case ISD::SHL: return C1 << C2;
6632 case ISD::SRL: return C1.lshr(ShiftAmt: C2);
6633 case ISD::SRA: return C1.ashr(ShiftAmt: C2);
6634 case ISD::ROTL: return C1.rotl(rotateAmt: C2);
6635 case ISD::ROTR: return C1.rotr(rotateAmt: C2);
6636 case ISD::SMIN: return C1.sle(RHS: C2) ? C1 : C2;
6637 case ISD::SMAX: return C1.sge(RHS: C2) ? C1 : C2;
6638 case ISD::UMIN: return C1.ule(RHS: C2) ? C1 : C2;
6639 case ISD::UMAX: return C1.uge(RHS: C2) ? C1 : C2;
6640 case ISD::SADDSAT: return C1.sadd_sat(RHS: C2);
6641 case ISD::UADDSAT: return C1.uadd_sat(RHS: C2);
6642 case ISD::SSUBSAT: return C1.ssub_sat(RHS: C2);
6643 case ISD::USUBSAT: return C1.usub_sat(RHS: C2);
6644 case ISD::SSHLSAT: return C1.sshl_sat(RHS: C2);
6645 case ISD::USHLSAT: return C1.ushl_sat(RHS: C2);
6646 case ISD::UDIV:
6647 if (!C2.getBoolValue())
6648 break;
6649 return C1.udiv(RHS: C2);
6650 case ISD::UREM:
6651 if (!C2.getBoolValue())
6652 break;
6653 return C1.urem(RHS: C2);
6654 case ISD::SDIV:
6655 if (!C2.getBoolValue())
6656 break;
6657 return C1.sdiv(RHS: C2);
6658 case ISD::SREM:
6659 if (!C2.getBoolValue())
6660 break;
6661 return C1.srem(RHS: C2);
6662 case ISD::AVGFLOORS:
6663 return APIntOps::avgFloorS(C1, C2);
6664 case ISD::AVGFLOORU:
6665 return APIntOps::avgFloorU(C1, C2);
6666 case ISD::AVGCEILS:
6667 return APIntOps::avgCeilS(C1, C2);
6668 case ISD::AVGCEILU:
6669 return APIntOps::avgCeilU(C1, C2);
6670 case ISD::ABDS:
6671 return APIntOps::abds(A: C1, B: C2);
6672 case ISD::ABDU:
6673 return APIntOps::abdu(A: C1, B: C2);
6674 case ISD::MULHS:
6675 return APIntOps::mulhs(C1, C2);
6676 case ISD::MULHU:
6677 return APIntOps::mulhu(C1, C2);
6678 }
6679 return std::nullopt;
6680}
6681// Handle constant folding with UNDEF.
6682// TODO: Handle more cases.
6683static std::optional<APInt> FoldValueWithUndef(unsigned Opcode, const APInt &C1,
6684 bool IsUndef1, const APInt &C2,
6685 bool IsUndef2) {
6686 if (!(IsUndef1 || IsUndef2))
6687 return FoldValue(Opcode, C1, C2);
6688
6689 // Fold and(x, undef) -> 0
6690 // Fold mul(x, undef) -> 0
6691 if (Opcode == ISD::AND || Opcode == ISD::MUL)
6692 return APInt::getZero(numBits: C1.getBitWidth());
6693
6694 return std::nullopt;
6695}
6696
6697SDValue SelectionDAG::FoldSymbolOffset(unsigned Opcode, EVT VT,
6698 const GlobalAddressSDNode *GA,
6699 const SDNode *N2) {
6700 if (GA->getOpcode() != ISD::GlobalAddress)
6701 return SDValue();
6702 if (!TLI->isOffsetFoldingLegal(GA))
6703 return SDValue();
6704 auto *C2 = dyn_cast<ConstantSDNode>(Val: N2);
6705 if (!C2)
6706 return SDValue();
6707 int64_t Offset = C2->getSExtValue();
6708 switch (Opcode) {
6709 case ISD::ADD: break;
6710 case ISD::SUB: Offset = -uint64_t(Offset); break;
6711 default: return SDValue();
6712 }
6713 return getGlobalAddress(GV: GA->getGlobal(), DL: SDLoc(C2), VT,
6714 Offset: GA->getOffset() + uint64_t(Offset));
6715}
6716
6717bool SelectionDAG::isUndef(unsigned Opcode, ArrayRef<SDValue> Ops) {
6718 switch (Opcode) {
6719 case ISD::SDIV:
6720 case ISD::UDIV:
6721 case ISD::SREM:
6722 case ISD::UREM: {
6723 // If a divisor is zero/undef or any element of a divisor vector is
6724 // zero/undef, the whole op is undef.
6725 assert(Ops.size() == 2 && "Div/rem should have 2 operands");
6726 SDValue Divisor = Ops[1];
6727 if (Divisor.isUndef() || isNullConstant(V: Divisor))
6728 return true;
6729
6730 return ISD::isBuildVectorOfConstantSDNodes(N: Divisor.getNode()) &&
6731 llvm::any_of(Range: Divisor->op_values(),
6732 P: [](SDValue V) { return V.isUndef() ||
6733 isNullConstant(V); });
6734 // TODO: Handle signed overflow.
6735 }
6736 // TODO: Handle oversized shifts.
6737 default:
6738 return false;
6739 }
6740}
6741
6742SDValue SelectionDAG::FoldConstantArithmetic(unsigned Opcode, const SDLoc &DL,
6743 EVT VT, ArrayRef<SDValue> Ops,
6744 SDNodeFlags Flags) {
6745 // If the opcode is a target-specific ISD node, there's nothing we can
6746 // do here and the operand rules may not line up with the below, so
6747 // bail early.
6748 // We can't create a scalar CONCAT_VECTORS so skip it. It will break
6749 // for concats involving SPLAT_VECTOR. Concats of BUILD_VECTORS are handled by
6750 // foldCONCAT_VECTORS in getNode before this is called.
6751 if (Opcode >= ISD::BUILTIN_OP_END || Opcode == ISD::CONCAT_VECTORS)
6752 return SDValue();
6753
6754 unsigned NumOps = Ops.size();
6755 if (NumOps == 0)
6756 return SDValue();
6757
6758 if (isUndef(Opcode, Ops))
6759 return getUNDEF(VT);
6760
6761 // Handle unary special cases.
6762 if (NumOps == 1) {
6763 SDValue N1 = Ops[0];
6764
6765 // Constant fold unary operations with an integer constant operand. Even
6766 // opaque constant will be folded, because the folding of unary operations
6767 // doesn't create new constants with different values. Nevertheless, the
6768 // opaque flag is preserved during folding to prevent future folding with
6769 // other constants.
6770 if (auto *C = dyn_cast<ConstantSDNode>(Val&: N1)) {
6771 const APInt &Val = C->getAPIntValue();
6772 switch (Opcode) {
6773 case ISD::SIGN_EXTEND:
6774 return getConstant(Val: Val.sextOrTrunc(width: VT.getSizeInBits()), DL, VT,
6775 isT: C->isTargetOpcode(), isO: C->isOpaque());
6776 case ISD::TRUNCATE:
6777 if (C->isOpaque())
6778 break;
6779 [[fallthrough]];
6780 case ISD::ZERO_EXTEND:
6781 return getConstant(Val: Val.zextOrTrunc(width: VT.getSizeInBits()), DL, VT,
6782 isT: C->isTargetOpcode(), isO: C->isOpaque());
6783 case ISD::ANY_EXTEND:
6784 // Some targets like RISCV prefer to sign extend some types.
6785 if (TLI->isSExtCheaperThanZExt(FromTy: N1.getValueType(), ToTy: VT))
6786 return getConstant(Val: Val.sextOrTrunc(width: VT.getSizeInBits()), DL, VT,
6787 isT: C->isTargetOpcode(), isO: C->isOpaque());
6788 return getConstant(Val: Val.zextOrTrunc(width: VT.getSizeInBits()), DL, VT,
6789 isT: C->isTargetOpcode(), isO: C->isOpaque());
6790 case ISD::ABS:
6791 return getConstant(Val: Val.abs(), DL, VT, isT: C->isTargetOpcode(),
6792 isO: C->isOpaque());
6793 case ISD::BITREVERSE:
6794 return getConstant(Val: Val.reverseBits(), DL, VT, isT: C->isTargetOpcode(),
6795 isO: C->isOpaque());
6796 case ISD::BSWAP:
6797 return getConstant(Val: Val.byteSwap(), DL, VT, isT: C->isTargetOpcode(),
6798 isO: C->isOpaque());
6799 case ISD::CTPOP:
6800 return getConstant(Val: Val.popcount(), DL, VT, isT: C->isTargetOpcode(),
6801 isO: C->isOpaque());
6802 case ISD::CTLZ:
6803 case ISD::CTLZ_ZERO_UNDEF:
6804 return getConstant(Val: Val.countl_zero(), DL, VT, isT: C->isTargetOpcode(),
6805 isO: C->isOpaque());
6806 case ISD::CTTZ:
6807 case ISD::CTTZ_ZERO_UNDEF:
6808 return getConstant(Val: Val.countr_zero(), DL, VT, isT: C->isTargetOpcode(),
6809 isO: C->isOpaque());
6810 case ISD::UINT_TO_FP:
6811 case ISD::SINT_TO_FP: {
6812 APFloat FPV(VT.getFltSemantics(), APInt::getZero(numBits: VT.getSizeInBits()));
6813 (void)FPV.convertFromAPInt(Input: Val, IsSigned: Opcode == ISD::SINT_TO_FP,
6814 RM: APFloat::rmNearestTiesToEven);
6815 return getConstantFP(V: FPV, DL, VT);
6816 }
6817 case ISD::FP16_TO_FP:
6818 case ISD::BF16_TO_FP: {
6819 bool Ignored;
6820 APFloat FPV(Opcode == ISD::FP16_TO_FP ? APFloat::IEEEhalf()
6821 : APFloat::BFloat(),
6822 (Val.getBitWidth() == 16) ? Val : Val.trunc(width: 16));
6823
6824 // This can return overflow, underflow, or inexact; we don't care.
6825 // FIXME need to be more flexible about rounding mode.
6826 (void)FPV.convert(ToSemantics: VT.getFltSemantics(), RM: APFloat::rmNearestTiesToEven,
6827 losesInfo: &Ignored);
6828 return getConstantFP(V: FPV, DL, VT);
6829 }
6830 case ISD::STEP_VECTOR:
6831 if (SDValue V = FoldSTEP_VECTOR(DL, VT, Step: N1, DAG&: *this))
6832 return V;
6833 break;
6834 case ISD::BITCAST:
6835 if (VT == MVT::f16 && C->getValueType(ResNo: 0) == MVT::i16)
6836 return getConstantFP(V: APFloat(APFloat::IEEEhalf(), Val), DL, VT);
6837 if (VT == MVT::f32 && C->getValueType(ResNo: 0) == MVT::i32)
6838 return getConstantFP(V: APFloat(APFloat::IEEEsingle(), Val), DL, VT);
6839 if (VT == MVT::f64 && C->getValueType(ResNo: 0) == MVT::i64)
6840 return getConstantFP(V: APFloat(APFloat::IEEEdouble(), Val), DL, VT);
6841 if (VT == MVT::f128 && C->getValueType(ResNo: 0) == MVT::i128)
6842 return getConstantFP(V: APFloat(APFloat::IEEEquad(), Val), DL, VT);
6843 break;
6844 }
6845 }
6846
6847 // Constant fold unary operations with a floating point constant operand.
6848 if (auto *C = dyn_cast<ConstantFPSDNode>(Val&: N1)) {
6849 APFloat V = C->getValueAPF(); // make copy
6850 switch (Opcode) {
6851 case ISD::FNEG:
6852 V.changeSign();
6853 return getConstantFP(V, DL, VT);
6854 case ISD::FABS:
6855 V.clearSign();
6856 return getConstantFP(V, DL, VT);
6857 case ISD::FCEIL: {
6858 APFloat::opStatus fs = V.roundToIntegral(RM: APFloat::rmTowardPositive);
6859 if (fs == APFloat::opOK || fs == APFloat::opInexact)
6860 return getConstantFP(V, DL, VT);
6861 return SDValue();
6862 }
6863 case ISD::FTRUNC: {
6864 APFloat::opStatus fs = V.roundToIntegral(RM: APFloat::rmTowardZero);
6865 if (fs == APFloat::opOK || fs == APFloat::opInexact)
6866 return getConstantFP(V, DL, VT);
6867 return SDValue();
6868 }
6869 case ISD::FFLOOR: {
6870 APFloat::opStatus fs = V.roundToIntegral(RM: APFloat::rmTowardNegative);
6871 if (fs == APFloat::opOK || fs == APFloat::opInexact)
6872 return getConstantFP(V, DL, VT);
6873 return SDValue();
6874 }
6875 case ISD::FP_EXTEND: {
6876 bool ignored;
6877 // This can return overflow, underflow, or inexact; we don't care.
6878 // FIXME need to be more flexible about rounding mode.
6879 (void)V.convert(ToSemantics: VT.getFltSemantics(), RM: APFloat::rmNearestTiesToEven,
6880 losesInfo: &ignored);
6881 return getConstantFP(V, DL, VT);
6882 }
6883 case ISD::FP_TO_SINT:
6884 case ISD::FP_TO_UINT: {
6885 bool ignored;
6886 APSInt IntVal(VT.getSizeInBits(), Opcode == ISD::FP_TO_UINT);
6887 // FIXME need to be more flexible about rounding mode.
6888 APFloat::opStatus s =
6889 V.convertToInteger(Result&: IntVal, RM: APFloat::rmTowardZero, IsExact: &ignored);
6890 if (s == APFloat::opInvalidOp) // inexact is OK, in fact usual
6891 break;
6892 return getConstant(Val: IntVal, DL, VT);
6893 }
6894 case ISD::FP_TO_FP16:
6895 case ISD::FP_TO_BF16: {
6896 bool Ignored;
6897 // This can return overflow, underflow, or inexact; we don't care.
6898 // FIXME need to be more flexible about rounding mode.
6899 (void)V.convert(ToSemantics: Opcode == ISD::FP_TO_FP16 ? APFloat::IEEEhalf()
6900 : APFloat::BFloat(),
6901 RM: APFloat::rmNearestTiesToEven, losesInfo: &Ignored);
6902 return getConstant(Val: V.bitcastToAPInt().getZExtValue(), DL, VT);
6903 }
6904 case ISD::BITCAST:
6905 if (VT == MVT::i16 && C->getValueType(ResNo: 0) == MVT::f16)
6906 return getConstant(Val: (uint16_t)V.bitcastToAPInt().getZExtValue(), DL,
6907 VT);
6908 if (VT == MVT::i16 && C->getValueType(ResNo: 0) == MVT::bf16)
6909 return getConstant(Val: (uint16_t)V.bitcastToAPInt().getZExtValue(), DL,
6910 VT);
6911 if (VT == MVT::i32 && C->getValueType(ResNo: 0) == MVT::f32)
6912 return getConstant(Val: (uint32_t)V.bitcastToAPInt().getZExtValue(), DL,
6913 VT);
6914 if (VT == MVT::i64 && C->getValueType(0) == MVT::f64)
6915 return getConstant(Val: V.bitcastToAPInt().getZExtValue(), DL, VT);
6916 break;
6917 }
6918 }
6919
6920 // Early-out if we failed to constant fold a bitcast.
6921 if (Opcode == ISD::BITCAST)
6922 return SDValue();
6923 }
6924
6925 // Handle binops special cases.
6926 if (NumOps == 2) {
6927 if (SDValue CFP = foldConstantFPMath(Opcode, DL, VT, Ops))
6928 return CFP;
6929
6930 if (auto *C1 = dyn_cast<ConstantSDNode>(Val: Ops[0])) {
6931 if (auto *C2 = dyn_cast<ConstantSDNode>(Val: Ops[1])) {
6932 if (C1->isOpaque() || C2->isOpaque())
6933 return SDValue();
6934
6935 std::optional<APInt> FoldAttempt =
6936 FoldValue(Opcode, C1: C1->getAPIntValue(), C2: C2->getAPIntValue());
6937 if (!FoldAttempt)
6938 return SDValue();
6939
6940 SDValue Folded = getConstant(Val: *FoldAttempt, DL, VT);
6941 assert((!Folded || !VT.isVector()) &&
6942 "Can't fold vectors ops with scalar operands");
6943 return Folded;
6944 }
6945 }
6946
6947 // fold (add Sym, c) -> Sym+c
6948 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Val: Ops[0]))
6949 return FoldSymbolOffset(Opcode, VT, GA, N2: Ops[1].getNode());
6950 if (TLI->isCommutativeBinOp(Opcode))
6951 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Val: Ops[1]))
6952 return FoldSymbolOffset(Opcode, VT, GA, N2: Ops[0].getNode());
6953
6954 // fold (sext_in_reg c1) -> c2
6955 if (Opcode == ISD::SIGN_EXTEND_INREG) {
6956 EVT EVT = cast<VTSDNode>(Val: Ops[1])->getVT();
6957
6958 auto SignExtendInReg = [&](APInt Val, llvm::EVT ConstantVT) {
6959 unsigned FromBits = EVT.getScalarSizeInBits();
6960 Val <<= Val.getBitWidth() - FromBits;
6961 Val.ashrInPlace(ShiftAmt: Val.getBitWidth() - FromBits);
6962 return getConstant(Val, DL, VT: ConstantVT);
6963 };
6964
6965 if (auto *C1 = dyn_cast<ConstantSDNode>(Val: Ops[0])) {
6966 const APInt &Val = C1->getAPIntValue();
6967 return SignExtendInReg(Val, VT);
6968 }
6969
6970 if (ISD::isBuildVectorOfConstantSDNodes(N: Ops[0].getNode())) {
6971 SmallVector<SDValue, 8> ScalarOps;
6972 llvm::EVT OpVT = Ops[0].getOperand(i: 0).getValueType();
6973 for (int I = 0, E = VT.getVectorNumElements(); I != E; ++I) {
6974 SDValue Op = Ops[0].getOperand(i: I);
6975 if (Op.isUndef()) {
6976 ScalarOps.push_back(Elt: getUNDEF(VT: OpVT));
6977 continue;
6978 }
6979 const APInt &Val = cast<ConstantSDNode>(Val&: Op)->getAPIntValue();
6980 ScalarOps.push_back(Elt: SignExtendInReg(Val, OpVT));
6981 }
6982 return getBuildVector(VT, DL, Ops: ScalarOps);
6983 }
6984
6985 if (Ops[0].getOpcode() == ISD::SPLAT_VECTOR &&
6986 isa<ConstantSDNode>(Val: Ops[0].getOperand(i: 0)))
6987 return getNode(Opcode: ISD::SPLAT_VECTOR, DL, VT,
6988 N1: SignExtendInReg(Ops[0].getConstantOperandAPInt(i: 0),
6989 Ops[0].getOperand(i: 0).getValueType()));
6990 }
6991 }
6992
6993 // This is for vector folding only from here on.
6994 if (!VT.isVector())
6995 return SDValue();
6996
6997 ElementCount NumElts = VT.getVectorElementCount();
6998
6999 // See if we can fold through any bitcasted integer ops.
7000 if (NumOps == 2 && VT.isFixedLengthVector() && VT.isInteger() &&
7001 Ops[0].getValueType() == VT && Ops[1].getValueType() == VT &&
7002 (Ops[0].getOpcode() == ISD::BITCAST ||
7003 Ops[1].getOpcode() == ISD::BITCAST)) {
7004 SDValue N1 = peekThroughBitcasts(V: Ops[0]);
7005 SDValue N2 = peekThroughBitcasts(V: Ops[1]);
7006 auto *BV1 = dyn_cast<BuildVectorSDNode>(Val&: N1);
7007 auto *BV2 = dyn_cast<BuildVectorSDNode>(Val&: N2);
7008 if (BV1 && BV2 && N1.getValueType().isInteger() &&
7009 N2.getValueType().isInteger()) {
7010 bool IsLE = getDataLayout().isLittleEndian();
7011 unsigned EltBits = VT.getScalarSizeInBits();
7012 SmallVector<APInt> RawBits1, RawBits2;
7013 BitVector UndefElts1, UndefElts2;
7014 if (BV1->getConstantRawBits(IsLittleEndian: IsLE, DstEltSizeInBits: EltBits, RawBitElements&: RawBits1, UndefElements&: UndefElts1) &&
7015 BV2->getConstantRawBits(IsLittleEndian: IsLE, DstEltSizeInBits: EltBits, RawBitElements&: RawBits2, UndefElements&: UndefElts2)) {
7016 SmallVector<APInt> RawBits;
7017 for (unsigned I = 0, E = NumElts.getFixedValue(); I != E; ++I) {
7018 std::optional<APInt> Fold = FoldValueWithUndef(
7019 Opcode, C1: RawBits1[I], IsUndef1: UndefElts1[I], C2: RawBits2[I], IsUndef2: UndefElts2[I]);
7020 if (!Fold)
7021 break;
7022 RawBits.push_back(Elt: *Fold);
7023 }
7024 if (RawBits.size() == NumElts.getFixedValue()) {
7025 // We have constant folded, but we might need to cast this again back
7026 // to the original (possibly legalized) type.
7027 EVT BVVT, BVEltVT;
7028 if (N1.getValueType() == VT) {
7029 BVVT = N1.getValueType();
7030 BVEltVT = BV1->getOperand(Num: 0).getValueType();
7031 } else {
7032 BVVT = N2.getValueType();
7033 BVEltVT = BV2->getOperand(Num: 0).getValueType();
7034 }
7035 unsigned BVEltBits = BVEltVT.getSizeInBits();
7036 SmallVector<APInt> DstBits;
7037 BitVector DstUndefs;
7038 BuildVectorSDNode::recastRawBits(IsLittleEndian: IsLE, DstEltSizeInBits: BVVT.getScalarSizeInBits(),
7039 DstBitElements&: DstBits, SrcBitElements: RawBits, DstUndefElements&: DstUndefs,
7040 SrcUndefElements: BitVector(RawBits.size(), false));
7041 SmallVector<SDValue> Ops(DstBits.size(), getUNDEF(VT: BVEltVT));
7042 for (unsigned I = 0, E = DstBits.size(); I != E; ++I) {
7043 if (DstUndefs[I])
7044 continue;
7045 Ops[I] = getConstant(Val: DstBits[I].sext(width: BVEltBits), DL, VT: BVEltVT);
7046 }
7047 return getBitcast(VT, V: getBuildVector(VT: BVVT, DL, Ops));
7048 }
7049 }
7050 }
7051 }
7052
7053 // Fold (mul step_vector(C0), C1) to (step_vector(C0 * C1)).
7054 // (shl step_vector(C0), C1) -> (step_vector(C0 << C1))
7055 if ((Opcode == ISD::MUL || Opcode == ISD::SHL) &&
7056 Ops[0].getOpcode() == ISD::STEP_VECTOR) {
7057 APInt RHSVal;
7058 if (ISD::isConstantSplatVector(N: Ops[1].getNode(), SplatVal&: RHSVal)) {
7059 APInt NewStep = Opcode == ISD::MUL
7060 ? Ops[0].getConstantOperandAPInt(i: 0) * RHSVal
7061 : Ops[0].getConstantOperandAPInt(i: 0) << RHSVal;
7062 return getStepVector(DL, ResVT: VT, StepVal: NewStep);
7063 }
7064 }
7065
7066 auto IsScalarOrSameVectorSize = [NumElts](const SDValue &Op) {
7067 return !Op.getValueType().isVector() ||
7068 Op.getValueType().getVectorElementCount() == NumElts;
7069 };
7070
7071 auto IsBuildVectorSplatVectorOrUndef = [](const SDValue &Op) {
7072 return Op.isUndef() || Op.getOpcode() == ISD::CONDCODE ||
7073 Op.getOpcode() == ISD::BUILD_VECTOR ||
7074 Op.getOpcode() == ISD::SPLAT_VECTOR;
7075 };
7076
7077 // All operands must be vector types with the same number of elements as
7078 // the result type and must be either UNDEF or a build/splat vector
7079 // or UNDEF scalars.
7080 if (!llvm::all_of(Range&: Ops, P: IsBuildVectorSplatVectorOrUndef) ||
7081 !llvm::all_of(Range&: Ops, P: IsScalarOrSameVectorSize))
7082 return SDValue();
7083
7084 // If we are comparing vectors, then the result needs to be a i1 boolean that
7085 // is then extended back to the legal result type depending on how booleans
7086 // are represented.
7087 EVT SVT = (Opcode == ISD::SETCC ? MVT::i1 : VT.getScalarType());
7088 ISD::NodeType ExtendCode =
7089 (Opcode == ISD::SETCC && SVT != VT.getScalarType())
7090 ? TargetLowering::getExtendForContent(Content: TLI->getBooleanContents(Type: VT))
7091 : ISD::SIGN_EXTEND;
7092
7093 // Find legal integer scalar type for constant promotion and
7094 // ensure that its scalar size is at least as large as source.
7095 EVT LegalSVT = VT.getScalarType();
7096 if (NewNodesMustHaveLegalTypes && LegalSVT.isInteger()) {
7097 LegalSVT = TLI->getTypeToTransformTo(Context&: *getContext(), VT: LegalSVT);
7098 if (LegalSVT.bitsLT(VT: VT.getScalarType()))
7099 return SDValue();
7100 }
7101
7102 // For scalable vector types we know we're dealing with SPLAT_VECTORs. We
7103 // only have one operand to check. For fixed-length vector types we may have
7104 // a combination of BUILD_VECTOR and SPLAT_VECTOR.
7105 unsigned NumVectorElts = NumElts.isScalable() ? 1 : NumElts.getFixedValue();
7106
7107 // Constant fold each scalar lane separately.
7108 SmallVector<SDValue, 4> ScalarResults;
7109 for (unsigned I = 0; I != NumVectorElts; I++) {
7110 SmallVector<SDValue, 4> ScalarOps;
7111 for (SDValue Op : Ops) {
7112 EVT InSVT = Op.getValueType().getScalarType();
7113 if (Op.getOpcode() != ISD::BUILD_VECTOR &&
7114 Op.getOpcode() != ISD::SPLAT_VECTOR) {
7115 if (Op.isUndef())
7116 ScalarOps.push_back(Elt: getUNDEF(VT: InSVT));
7117 else
7118 ScalarOps.push_back(Elt: Op);
7119 continue;
7120 }
7121
7122 SDValue ScalarOp =
7123 Op.getOperand(i: Op.getOpcode() == ISD::SPLAT_VECTOR ? 0 : I);
7124 EVT ScalarVT = ScalarOp.getValueType();
7125
7126 // Build vector (integer) scalar operands may need implicit
7127 // truncation - do this before constant folding.
7128 if (ScalarVT.isInteger() && ScalarVT.bitsGT(VT: InSVT)) {
7129 // Don't create illegally-typed nodes unless they're constants or undef
7130 // - if we fail to constant fold we can't guarantee the (dead) nodes
7131 // we're creating will be cleaned up before being visited for
7132 // legalization.
7133 if (NewNodesMustHaveLegalTypes && !ScalarOp.isUndef() &&
7134 !isa<ConstantSDNode>(Val: ScalarOp) &&
7135 TLI->getTypeAction(Context&: *getContext(), VT: InSVT) !=
7136 TargetLowering::TypeLegal)
7137 return SDValue();
7138 ScalarOp = getNode(Opcode: ISD::TRUNCATE, DL, VT: InSVT, N1: ScalarOp);
7139 }
7140
7141 ScalarOps.push_back(Elt: ScalarOp);
7142 }
7143
7144 // Constant fold the scalar operands.
7145 SDValue ScalarResult = getNode(Opcode, DL, VT: SVT, Ops: ScalarOps, Flags);
7146
7147 // Scalar folding only succeeded if the result is a constant or UNDEF.
7148 if (!ScalarResult.isUndef() && ScalarResult.getOpcode() != ISD::Constant &&
7149 ScalarResult.getOpcode() != ISD::ConstantFP)
7150 return SDValue();
7151
7152 // Legalize the (integer) scalar constant if necessary. We only do
7153 // this once we know the folding succeeded, since otherwise we would
7154 // get a node with illegal type which has a user.
7155 if (LegalSVT != SVT)
7156 ScalarResult = getNode(Opcode: ExtendCode, DL, VT: LegalSVT, N1: ScalarResult);
7157
7158 ScalarResults.push_back(Elt: ScalarResult);
7159 }
7160
7161 SDValue V = NumElts.isScalable() ? getSplatVector(VT, DL, Op: ScalarResults[0])
7162 : getBuildVector(VT, DL, Ops: ScalarResults);
7163 NewSDValueDbgMsg(V, Msg: "New node fold constant vector: ", G: this);
7164 return V;
7165}
7166
7167SDValue SelectionDAG::foldConstantFPMath(unsigned Opcode, const SDLoc &DL,
7168 EVT VT, ArrayRef<SDValue> Ops) {
7169 // TODO: Add support for unary/ternary fp opcodes.
7170 if (Ops.size() != 2)
7171 return SDValue();
7172
7173 // TODO: We don't do any constant folding for strict FP opcodes here, but we
7174 // should. That will require dealing with a potentially non-default
7175 // rounding mode, checking the "opStatus" return value from the APFloat
7176 // math calculations, and possibly other variations.
7177 SDValue N1 = Ops[0];
7178 SDValue N2 = Ops[1];
7179 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N: N1, /*AllowUndefs*/ false);
7180 ConstantFPSDNode *N2CFP = isConstOrConstSplatFP(N: N2, /*AllowUndefs*/ false);
7181 if (N1CFP && N2CFP) {
7182 APFloat C1 = N1CFP->getValueAPF(); // make copy
7183 const APFloat &C2 = N2CFP->getValueAPF();
7184 switch (Opcode) {
7185 case ISD::FADD:
7186 C1.add(RHS: C2, RM: APFloat::rmNearestTiesToEven);
7187 return getConstantFP(V: C1, DL, VT);
7188 case ISD::FSUB:
7189 C1.subtract(RHS: C2, RM: APFloat::rmNearestTiesToEven);
7190 return getConstantFP(V: C1, DL, VT);
7191 case ISD::FMUL:
7192 C1.multiply(RHS: C2, RM: APFloat::rmNearestTiesToEven);
7193 return getConstantFP(V: C1, DL, VT);
7194 case ISD::FDIV:
7195 C1.divide(RHS: C2, RM: APFloat::rmNearestTiesToEven);
7196 return getConstantFP(V: C1, DL, VT);
7197 case ISD::FREM:
7198 C1.mod(RHS: C2);
7199 return getConstantFP(V: C1, DL, VT);
7200 case ISD::FCOPYSIGN:
7201 C1.copySign(RHS: C2);
7202 return getConstantFP(V: C1, DL, VT);
7203 case ISD::FMINNUM:
7204 return getConstantFP(V: minnum(A: C1, B: C2), DL, VT);
7205 case ISD::FMAXNUM:
7206 return getConstantFP(V: maxnum(A: C1, B: C2), DL, VT);
7207 case ISD::FMINIMUM:
7208 return getConstantFP(V: minimum(A: C1, B: C2), DL, VT);
7209 case ISD::FMAXIMUM:
7210 return getConstantFP(V: maximum(A: C1, B: C2), DL, VT);
7211 case ISD::FMINIMUMNUM:
7212 return getConstantFP(V: minimumnum(A: C1, B: C2), DL, VT);
7213 case ISD::FMAXIMUMNUM:
7214 return getConstantFP(V: maximumnum(A: C1, B: C2), DL, VT);
7215 default: break;
7216 }
7217 }
7218 if (N1CFP && Opcode == ISD::FP_ROUND) {
7219 APFloat C1 = N1CFP->getValueAPF(); // make copy
7220 bool Unused;
7221 // This can return overflow, underflow, or inexact; we don't care.
7222 // FIXME need to be more flexible about rounding mode.
7223 (void)C1.convert(ToSemantics: VT.getFltSemantics(), RM: APFloat::rmNearestTiesToEven,
7224 losesInfo: &Unused);
7225 return getConstantFP(V: C1, DL, VT);
7226 }
7227
7228 switch (Opcode) {
7229 case ISD::FSUB:
7230 // -0.0 - undef --> undef (consistent with "fneg undef")
7231 if (ConstantFPSDNode *N1C = isConstOrConstSplatFP(N: N1, /*AllowUndefs*/ true))
7232 if (N1C && N1C->getValueAPF().isNegZero() && N2.isUndef())
7233 return getUNDEF(VT);
7234 [[fallthrough]];
7235
7236 case ISD::FADD:
7237 case ISD::FMUL:
7238 case ISD::FDIV:
7239 case ISD::FREM:
7240 // If both operands are undef, the result is undef. If 1 operand is undef,
7241 // the result is NaN. This should match the behavior of the IR optimizer.
7242 if (N1.isUndef() && N2.isUndef())
7243 return getUNDEF(VT);
7244 if (N1.isUndef() || N2.isUndef())
7245 return getConstantFP(V: APFloat::getNaN(Sem: VT.getFltSemantics()), DL, VT);
7246 }
7247 return SDValue();
7248}
7249
7250SDValue SelectionDAG::getAssertAlign(const SDLoc &DL, SDValue Val, Align A) {
7251 assert(Val.getValueType().isInteger() && "Invalid AssertAlign!");
7252
7253 // There's no need to assert on a byte-aligned pointer. All pointers are at
7254 // least byte aligned.
7255 if (A == Align(1))
7256 return Val;
7257
7258 SDVTList VTs = getVTList(VT: Val.getValueType());
7259 FoldingSetNodeID ID;
7260 AddNodeIDNode(ID, OpC: ISD::AssertAlign, VTList: VTs, OpList: {Val});
7261 ID.AddInteger(I: A.value());
7262
7263 void *IP = nullptr;
7264 if (SDNode *E = FindNodeOrInsertPos(ID, DL, InsertPos&: IP))
7265 return SDValue(E, 0);
7266
7267 auto *N =
7268 newSDNode<AssertAlignSDNode>(Args: DL.getIROrder(), Args: DL.getDebugLoc(), Args&: VTs, Args&: A);
7269 createOperands(Node: N, Vals: {Val});
7270
7271 CSEMap.InsertNode(N, InsertPos: IP);
7272 InsertNode(N);
7273
7274 SDValue V(N, 0);
7275 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
7276 return V;
7277}
7278
7279SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
7280 SDValue N1, SDValue N2) {
7281 SDNodeFlags Flags;
7282 if (Inserter)
7283 Flags = Inserter->getFlags();
7284 return getNode(Opcode, DL, VT, N1, N2, Flags);
7285}
7286
7287void SelectionDAG::canonicalizeCommutativeBinop(unsigned Opcode, SDValue &N1,
7288 SDValue &N2) const {
7289 if (!TLI->isCommutativeBinOp(Opcode))
7290 return;
7291
7292 // Canonicalize:
7293 // binop(const, nonconst) -> binop(nonconst, const)
7294 bool N1C = isConstantIntBuildVectorOrConstantInt(N: N1);
7295 bool N2C = isConstantIntBuildVectorOrConstantInt(N: N2);
7296 bool N1CFP = isConstantFPBuildVectorOrConstantFP(N: N1);
7297 bool N2CFP = isConstantFPBuildVectorOrConstantFP(N: N2);
7298 if ((N1C && !N2C) || (N1CFP && !N2CFP))
7299 std::swap(a&: N1, b&: N2);
7300
7301 // Canonicalize:
7302 // binop(splat(x), step_vector) -> binop(step_vector, splat(x))
7303 else if (N1.getOpcode() == ISD::SPLAT_VECTOR &&
7304 N2.getOpcode() == ISD::STEP_VECTOR)
7305 std::swap(a&: N1, b&: N2);
7306}
7307
7308SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
7309 SDValue N1, SDValue N2, const SDNodeFlags Flags) {
7310 assert(N1.getOpcode() != ISD::DELETED_NODE &&
7311 N2.getOpcode() != ISD::DELETED_NODE &&
7312 "Operand is DELETED_NODE!");
7313
7314 canonicalizeCommutativeBinop(Opcode, N1, N2);
7315
7316 auto *N1C = dyn_cast<ConstantSDNode>(Val&: N1);
7317 auto *N2C = dyn_cast<ConstantSDNode>(Val&: N2);
7318
7319 // Don't allow undefs in vector splats - we might be returning N2 when folding
7320 // to zero etc.
7321 ConstantSDNode *N2CV =
7322 isConstOrConstSplat(N: N2, /*AllowUndefs*/ false, /*AllowTruncation*/ true);
7323
7324 switch (Opcode) {
7325 default: break;
7326 case ISD::TokenFactor:
7327 assert(VT == MVT::Other && N1.getValueType() == MVT::Other &&
7328 N2.getValueType() == MVT::Other && "Invalid token factor!");
7329 // Fold trivial token factors.
7330 if (N1.getOpcode() == ISD::EntryToken) return N2;
7331 if (N2.getOpcode() == ISD::EntryToken) return N1;
7332 if (N1 == N2) return N1;
7333 break;
7334 case ISD::BUILD_VECTOR: {
7335 // Attempt to simplify BUILD_VECTOR.
7336 SDValue Ops[] = {N1, N2};
7337 if (SDValue V = FoldBUILD_VECTOR(DL, VT, Ops, DAG&: *this))
7338 return V;
7339 break;
7340 }
7341 case ISD::CONCAT_VECTORS: {
7342 SDValue Ops[] = {N1, N2};
7343 if (SDValue V = foldCONCAT_VECTORS(DL, VT, Ops, DAG&: *this))
7344 return V;
7345 break;
7346 }
7347 case ISD::AND:
7348 assert(VT.isInteger() && "This operator does not apply to FP types!");
7349 assert(N1.getValueType() == N2.getValueType() &&
7350 N1.getValueType() == VT && "Binary operator types must match!");
7351 // (X & 0) -> 0. This commonly occurs when legalizing i64 values, so it's
7352 // worth handling here.
7353 if (N2CV && N2CV->isZero())
7354 return N2;
7355 if (N2CV && N2CV->isAllOnes()) // X & -1 -> X
7356 return N1;
7357 break;
7358 case ISD::OR:
7359 case ISD::XOR:
7360 case ISD::ADD:
7361 case ISD::PTRADD:
7362 case ISD::SUB:
7363 assert(VT.isInteger() && "This operator does not apply to FP types!");
7364 assert(N1.getValueType() == N2.getValueType() &&
7365 N1.getValueType() == VT && "Binary operator types must match!");
7366 // The equal operand types requirement is unnecessarily strong for PTRADD.
7367 // However, the SelectionDAGBuilder does not generate PTRADDs with different
7368 // operand types, and we'd need to re-implement GEP's non-standard wrapping
7369 // logic everywhere where PTRADDs may be folded or combined to properly
7370 // support them. If/when we introduce pointer types to the SDAG, we will
7371 // need to relax this constraint.
7372
7373 // (X ^|+- 0) -> X. This commonly occurs when legalizing i64 values, so
7374 // it's worth handling here.
7375 if (N2CV && N2CV->isZero())
7376 return N1;
7377 if ((Opcode == ISD::ADD || Opcode == ISD::SUB) &&
7378 VT.getScalarType() == MVT::i1)
7379 return getNode(Opcode: ISD::XOR, DL, VT, N1, N2);
7380 break;
7381 case ISD::MUL:
7382 assert(VT.isInteger() && "This operator does not apply to FP types!");
7383 assert(N1.getValueType() == N2.getValueType() &&
7384 N1.getValueType() == VT && "Binary operator types must match!");
7385 if (VT.getScalarType() == MVT::i1)
7386 return getNode(Opcode: ISD::AND, DL, VT, N1, N2);
7387 if (N2C && (N1.getOpcode() == ISD::VSCALE) && Flags.hasNoSignedWrap()) {
7388 const APInt &MulImm = N1->getConstantOperandAPInt(Num: 0);
7389 const APInt &N2CImm = N2C->getAPIntValue();
7390 return getVScale(DL, VT, MulImm: MulImm * N2CImm);
7391 }
7392 break;
7393 case ISD::UDIV:
7394 case ISD::UREM:
7395 case ISD::MULHU:
7396 case ISD::MULHS:
7397 case ISD::SDIV:
7398 case ISD::SREM:
7399 case ISD::SADDSAT:
7400 case ISD::SSUBSAT:
7401 case ISD::UADDSAT:
7402 case ISD::USUBSAT:
7403 assert(VT.isInteger() && "This operator does not apply to FP types!");
7404 assert(N1.getValueType() == N2.getValueType() &&
7405 N1.getValueType() == VT && "Binary operator types must match!");
7406 if (VT.getScalarType() == MVT::i1) {
7407 // fold (add_sat x, y) -> (or x, y) for bool types.
7408 if (Opcode == ISD::SADDSAT || Opcode == ISD::UADDSAT)
7409 return getNode(Opcode: ISD::OR, DL, VT, N1, N2);
7410 // fold (sub_sat x, y) -> (and x, ~y) for bool types.
7411 if (Opcode == ISD::SSUBSAT || Opcode == ISD::USUBSAT)
7412 return getNode(Opcode: ISD::AND, DL, VT, N1, N2: getNOT(DL, Val: N2, VT));
7413 }
7414 break;
7415 case ISD::SCMP:
7416 case ISD::UCMP:
7417 assert(N1.getValueType() == N2.getValueType() &&
7418 "Types of operands of UCMP/SCMP must match");
7419 assert(N1.getValueType().isVector() == VT.isVector() &&
7420 "Operands and return type of must both be scalars or vectors");
7421 if (VT.isVector())
7422 assert(VT.getVectorElementCount() ==
7423 N1.getValueType().getVectorElementCount() &&
7424 "Result and operands must have the same number of elements");
7425 break;
7426 case ISD::AVGFLOORS:
7427 case ISD::AVGFLOORU:
7428 case ISD::AVGCEILS:
7429 case ISD::AVGCEILU:
7430 assert(VT.isInteger() && "This operator does not apply to FP types!");
7431 assert(N1.getValueType() == N2.getValueType() &&
7432 N1.getValueType() == VT && "Binary operator types must match!");
7433 break;
7434 case ISD::ABDS:
7435 case ISD::ABDU:
7436 assert(VT.isInteger() && "This operator does not apply to FP types!");
7437 assert(N1.getValueType() == N2.getValueType() &&
7438 N1.getValueType() == VT && "Binary operator types must match!");
7439 if (VT.getScalarType() == MVT::i1)
7440 return getNode(Opcode: ISD::XOR, DL, VT, N1, N2);
7441 break;
7442 case ISD::SMIN:
7443 case ISD::UMAX:
7444 assert(VT.isInteger() && "This operator does not apply to FP types!");
7445 assert(N1.getValueType() == N2.getValueType() &&
7446 N1.getValueType() == VT && "Binary operator types must match!");
7447 if (VT.getScalarType() == MVT::i1)
7448 return getNode(Opcode: ISD::OR, DL, VT, N1, N2);
7449 break;
7450 case ISD::SMAX:
7451 case ISD::UMIN:
7452 assert(VT.isInteger() && "This operator does not apply to FP types!");
7453 assert(N1.getValueType() == N2.getValueType() &&
7454 N1.getValueType() == VT && "Binary operator types must match!");
7455 if (VT.getScalarType() == MVT::i1)
7456 return getNode(Opcode: ISD::AND, DL, VT, N1, N2);
7457 break;
7458 case ISD::FADD:
7459 case ISD::FSUB:
7460 case ISD::FMUL:
7461 case ISD::FDIV:
7462 case ISD::FREM:
7463 assert(VT.isFloatingPoint() && "This operator only applies to FP types!");
7464 assert(N1.getValueType() == N2.getValueType() &&
7465 N1.getValueType() == VT && "Binary operator types must match!");
7466 if (SDValue V = simplifyFPBinop(Opcode, X: N1, Y: N2, Flags))
7467 return V;
7468 break;
7469 case ISD::FCOPYSIGN: // N1 and result must match. N1/N2 need not match.
7470 assert(N1.getValueType() == VT &&
7471 N1.getValueType().isFloatingPoint() &&
7472 N2.getValueType().isFloatingPoint() &&
7473 "Invalid FCOPYSIGN!");
7474 break;
7475 case ISD::SHL:
7476 if (N2C && (N1.getOpcode() == ISD::VSCALE) && Flags.hasNoSignedWrap()) {
7477 const APInt &MulImm = N1->getConstantOperandAPInt(Num: 0);
7478 const APInt &ShiftImm = N2C->getAPIntValue();
7479 return getVScale(DL, VT, MulImm: MulImm << ShiftImm);
7480 }
7481 [[fallthrough]];
7482 case ISD::SRA:
7483 case ISD::SRL:
7484 if (SDValue V = simplifyShift(X: N1, Y: N2))
7485 return V;
7486 [[fallthrough]];
7487 case ISD::ROTL:
7488 case ISD::ROTR:
7489 assert(VT == N1.getValueType() &&
7490 "Shift operators return type must be the same as their first arg");
7491 assert(VT.isInteger() && N2.getValueType().isInteger() &&
7492 "Shifts only work on integers");
7493 assert((!VT.isVector() || VT == N2.getValueType()) &&
7494 "Vector shift amounts must be in the same as their first arg");
7495 // Verify that the shift amount VT is big enough to hold valid shift
7496 // amounts. This catches things like trying to shift an i1024 value by an
7497 // i8, which is easy to fall into in generic code that uses
7498 // TLI.getShiftAmount().
7499 assert(N2.getValueType().getScalarSizeInBits() >=
7500 Log2_32_Ceil(VT.getScalarSizeInBits()) &&
7501 "Invalid use of small shift amount with oversized value!");
7502
7503 // Always fold shifts of i1 values so the code generator doesn't need to
7504 // handle them. Since we know the size of the shift has to be less than the
7505 // size of the value, the shift/rotate count is guaranteed to be zero.
7506 if (VT == MVT::i1)
7507 return N1;
7508 if (N2CV && N2CV->isZero())
7509 return N1;
7510 break;
7511 case ISD::FP_ROUND:
7512 assert(VT.isFloatingPoint() && N1.getValueType().isFloatingPoint() &&
7513 VT.bitsLE(N1.getValueType()) && N2C &&
7514 (N2C->getZExtValue() == 0 || N2C->getZExtValue() == 1) &&
7515 N2.getOpcode() == ISD::TargetConstant && "Invalid FP_ROUND!");
7516 if (N1.getValueType() == VT) return N1; // noop conversion.
7517 break;
7518 case ISD::AssertNoFPClass: {
7519 assert(N1.getValueType().isFloatingPoint() &&
7520 "AssertNoFPClass is used for a non-floating type");
7521 assert(isa<ConstantSDNode>(N2) && "NoFPClass is not Constant");
7522 FPClassTest NoFPClass = static_cast<FPClassTest>(N2->getAsZExtVal());
7523 assert(llvm::to_underlying(NoFPClass) <=
7524 BitmaskEnumDetail::Mask<FPClassTest>() &&
7525 "FPClassTest value too large");
7526 (void)NoFPClass;
7527 break;
7528 }
7529 case ISD::AssertSext:
7530 case ISD::AssertZext: {
7531 EVT EVT = cast<VTSDNode>(Val&: N2)->getVT();
7532 assert(VT == N1.getValueType() && "Not an inreg extend!");
7533 assert(VT.isInteger() && EVT.isInteger() &&
7534 "Cannot *_EXTEND_INREG FP types");
7535 assert(!EVT.isVector() &&
7536 "AssertSExt/AssertZExt type should be the vector element type "
7537 "rather than the vector type!");
7538 assert(EVT.bitsLE(VT.getScalarType()) && "Not extending!");
7539 if (VT.getScalarType() == EVT) return N1; // noop assertion.
7540 break;
7541 }
7542 case ISD::SIGN_EXTEND_INREG: {
7543 EVT EVT = cast<VTSDNode>(Val&: N2)->getVT();
7544 assert(VT == N1.getValueType() && "Not an inreg extend!");
7545 assert(VT.isInteger() && EVT.isInteger() &&
7546 "Cannot *_EXTEND_INREG FP types");
7547 assert(EVT.isVector() == VT.isVector() &&
7548 "SIGN_EXTEND_INREG type should be vector iff the operand "
7549 "type is vector!");
7550 assert((!EVT.isVector() ||
7551 EVT.getVectorElementCount() == VT.getVectorElementCount()) &&
7552 "Vector element counts must match in SIGN_EXTEND_INREG");
7553 assert(EVT.bitsLE(VT) && "Not extending!");
7554 if (EVT == VT) return N1; // Not actually extending
7555 break;
7556 }
7557 case ISD::FP_TO_SINT_SAT:
7558 case ISD::FP_TO_UINT_SAT: {
7559 assert(VT.isInteger() && cast<VTSDNode>(N2)->getVT().isInteger() &&
7560 N1.getValueType().isFloatingPoint() && "Invalid FP_TO_*INT_SAT");
7561 assert(N1.getValueType().isVector() == VT.isVector() &&
7562 "FP_TO_*INT_SAT type should be vector iff the operand type is "
7563 "vector!");
7564 assert((!VT.isVector() || VT.getVectorElementCount() ==
7565 N1.getValueType().getVectorElementCount()) &&
7566 "Vector element counts must match in FP_TO_*INT_SAT");
7567 assert(!cast<VTSDNode>(N2)->getVT().isVector() &&
7568 "Type to saturate to must be a scalar.");
7569 assert(cast<VTSDNode>(N2)->getVT().bitsLE(VT.getScalarType()) &&
7570 "Not extending!");
7571 break;
7572 }
7573 case ISD::EXTRACT_VECTOR_ELT:
7574 assert(VT.getSizeInBits() >= N1.getValueType().getScalarSizeInBits() &&
7575 "The result of EXTRACT_VECTOR_ELT must be at least as wide as the \
7576 element type of the vector.");
7577
7578 // Extract from an undefined value or using an undefined index is undefined.
7579 if (N1.isUndef() || N2.isUndef())
7580 return getUNDEF(VT);
7581
7582 // EXTRACT_VECTOR_ELT of out-of-bounds element is an UNDEF for fixed length
7583 // vectors. For scalable vectors we will provide appropriate support for
7584 // dealing with arbitrary indices.
7585 if (N2C && N1.getValueType().isFixedLengthVector() &&
7586 N2C->getAPIntValue().uge(RHS: N1.getValueType().getVectorNumElements()))
7587 return getUNDEF(VT);
7588
7589 // EXTRACT_VECTOR_ELT of CONCAT_VECTORS is often formed while lowering is
7590 // expanding copies of large vectors from registers. This only works for
7591 // fixed length vectors, since we need to know the exact number of
7592 // elements.
7593 if (N2C && N1.getOpcode() == ISD::CONCAT_VECTORS &&
7594 N1.getOperand(i: 0).getValueType().isFixedLengthVector()) {
7595 unsigned Factor = N1.getOperand(i: 0).getValueType().getVectorNumElements();
7596 return getExtractVectorElt(DL, VT,
7597 Vec: N1.getOperand(i: N2C->getZExtValue() / Factor),
7598 Idx: N2C->getZExtValue() % Factor);
7599 }
7600
7601 // EXTRACT_VECTOR_ELT of BUILD_VECTOR or SPLAT_VECTOR is often formed while
7602 // lowering is expanding large vector constants.
7603 if (N2C && (N1.getOpcode() == ISD::BUILD_VECTOR ||
7604 N1.getOpcode() == ISD::SPLAT_VECTOR)) {
7605 assert((N1.getOpcode() != ISD::BUILD_VECTOR ||
7606 N1.getValueType().isFixedLengthVector()) &&
7607 "BUILD_VECTOR used for scalable vectors");
7608 unsigned Index =
7609 N1.getOpcode() == ISD::BUILD_VECTOR ? N2C->getZExtValue() : 0;
7610 SDValue Elt = N1.getOperand(i: Index);
7611
7612 if (VT != Elt.getValueType())
7613 // If the vector element type is not legal, the BUILD_VECTOR operands
7614 // are promoted and implicitly truncated, and the result implicitly
7615 // extended. Make that explicit here.
7616 Elt = getAnyExtOrTrunc(Op: Elt, DL, VT);
7617
7618 return Elt;
7619 }
7620
7621 // EXTRACT_VECTOR_ELT of INSERT_VECTOR_ELT is often formed when vector
7622 // operations are lowered to scalars.
7623 if (N1.getOpcode() == ISD::INSERT_VECTOR_ELT) {
7624 // If the indices are the same, return the inserted element else
7625 // if the indices are known different, extract the element from
7626 // the original vector.
7627 SDValue N1Op2 = N1.getOperand(i: 2);
7628 ConstantSDNode *N1Op2C = dyn_cast<ConstantSDNode>(Val&: N1Op2);
7629
7630 if (N1Op2C && N2C) {
7631 if (N1Op2C->getZExtValue() == N2C->getZExtValue()) {
7632 if (VT == N1.getOperand(i: 1).getValueType())
7633 return N1.getOperand(i: 1);
7634 if (VT.isFloatingPoint()) {
7635 assert(VT.getSizeInBits() > N1.getOperand(1).getValueType().getSizeInBits());
7636 return getFPExtendOrRound(Op: N1.getOperand(i: 1), DL, VT);
7637 }
7638 return getSExtOrTrunc(Op: N1.getOperand(i: 1), DL, VT);
7639 }
7640 return getNode(Opcode: ISD::EXTRACT_VECTOR_ELT, DL, VT, N1: N1.getOperand(i: 0), N2);
7641 }
7642 }
7643
7644 // EXTRACT_VECTOR_ELT of v1iX EXTRACT_SUBVECTOR could be formed
7645 // when vector types are scalarized and v1iX is legal.
7646 // vextract (v1iX extract_subvector(vNiX, Idx)) -> vextract(vNiX,Idx).
7647 // Here we are completely ignoring the extract element index (N2),
7648 // which is fine for fixed width vectors, since any index other than 0
7649 // is undefined anyway. However, this cannot be ignored for scalable
7650 // vectors - in theory we could support this, but we don't want to do this
7651 // without a profitability check.
7652 if (N1.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
7653 N1.getValueType().isFixedLengthVector() &&
7654 N1.getValueType().getVectorNumElements() == 1) {
7655 return getNode(Opcode: ISD::EXTRACT_VECTOR_ELT, DL, VT, N1: N1.getOperand(i: 0),
7656 N2: N1.getOperand(i: 1));
7657 }
7658 break;
7659 case ISD::EXTRACT_ELEMENT:
7660 assert(N2C && (unsigned)N2C->getZExtValue() < 2 && "Bad EXTRACT_ELEMENT!");
7661 assert(!N1.getValueType().isVector() && !VT.isVector() &&
7662 (N1.getValueType().isInteger() == VT.isInteger()) &&
7663 N1.getValueType() != VT &&
7664 "Wrong types for EXTRACT_ELEMENT!");
7665
7666 // EXTRACT_ELEMENT of BUILD_PAIR is often formed while legalize is expanding
7667 // 64-bit integers into 32-bit parts. Instead of building the extract of
7668 // the BUILD_PAIR, only to have legalize rip it apart, just do it now.
7669 if (N1.getOpcode() == ISD::BUILD_PAIR)
7670 return N1.getOperand(i: N2C->getZExtValue());
7671
7672 // EXTRACT_ELEMENT of a constant int is also very common.
7673 if (N1C) {
7674 unsigned ElementSize = VT.getSizeInBits();
7675 unsigned Shift = ElementSize * N2C->getZExtValue();
7676 const APInt &Val = N1C->getAPIntValue();
7677 return getConstant(Val: Val.extractBits(numBits: ElementSize, bitPosition: Shift), DL, VT);
7678 }
7679 break;
7680 case ISD::EXTRACT_SUBVECTOR: {
7681 EVT N1VT = N1.getValueType();
7682 assert(VT.isVector() && N1VT.isVector() &&
7683 "Extract subvector VTs must be vectors!");
7684 assert(VT.getVectorElementType() == N1VT.getVectorElementType() &&
7685 "Extract subvector VTs must have the same element type!");
7686 assert((VT.isFixedLengthVector() || N1VT.isScalableVector()) &&
7687 "Cannot extract a scalable vector from a fixed length vector!");
7688 assert((VT.isScalableVector() != N1VT.isScalableVector() ||
7689 VT.getVectorMinNumElements() <= N1VT.getVectorMinNumElements()) &&
7690 "Extract subvector must be from larger vector to smaller vector!");
7691 assert(N2C && "Extract subvector index must be a constant");
7692 assert((VT.isScalableVector() != N1VT.isScalableVector() ||
7693 (VT.getVectorMinNumElements() + N2C->getZExtValue()) <=
7694 N1VT.getVectorMinNumElements()) &&
7695 "Extract subvector overflow!");
7696 assert(N2C->getAPIntValue().getBitWidth() ==
7697 TLI->getVectorIdxWidth(getDataLayout()) &&
7698 "Constant index for EXTRACT_SUBVECTOR has an invalid size");
7699
7700 // Trivial extraction.
7701 if (VT == N1VT)
7702 return N1;
7703
7704 // EXTRACT_SUBVECTOR of an UNDEF is an UNDEF.
7705 if (N1.isUndef())
7706 return getUNDEF(VT);
7707
7708 // EXTRACT_SUBVECTOR of CONCAT_VECTOR can be simplified if the pieces of
7709 // the concat have the same type as the extract.
7710 if (N1.getOpcode() == ISD::CONCAT_VECTORS &&
7711 VT == N1.getOperand(i: 0).getValueType()) {
7712 unsigned Factor = VT.getVectorMinNumElements();
7713 return N1.getOperand(i: N2C->getZExtValue() / Factor);
7714 }
7715
7716 // EXTRACT_SUBVECTOR of INSERT_SUBVECTOR is often created
7717 // during shuffle legalization.
7718 if (N1.getOpcode() == ISD::INSERT_SUBVECTOR && N2 == N1.getOperand(i: 2) &&
7719 VT == N1.getOperand(i: 1).getValueType())
7720 return N1.getOperand(i: 1);
7721 break;
7722 }
7723 }
7724
7725 // Perform trivial constant folding.
7726 if (SDValue SV = FoldConstantArithmetic(Opcode, DL, VT, Ops: {N1, N2}, Flags))
7727 return SV;
7728
7729 // Canonicalize an UNDEF to the RHS, even over a constant.
7730 if (N1.isUndef()) {
7731 if (TLI->isCommutativeBinOp(Opcode)) {
7732 std::swap(a&: N1, b&: N2);
7733 } else {
7734 switch (Opcode) {
7735 case ISD::PTRADD:
7736 case ISD::SUB:
7737 // fold op(undef, arg2) -> undef, fold op(poison, arg2) ->poison.
7738 return N1.getOpcode() == ISD::POISON ? getPOISON(VT) : getUNDEF(VT);
7739 case ISD::SIGN_EXTEND_INREG:
7740 case ISD::UDIV:
7741 case ISD::SDIV:
7742 case ISD::UREM:
7743 case ISD::SREM:
7744 case ISD::SSUBSAT:
7745 case ISD::USUBSAT:
7746 // fold op(undef, arg2) -> 0, fold op(poison, arg2) -> poison.
7747 return N1.getOpcode() == ISD::POISON ? getPOISON(VT)
7748 : getConstant(Val: 0, DL, VT);
7749 }
7750 }
7751 }
7752
7753 // Fold a bunch of operators when the RHS is undef.
7754 if (N2.isUndef()) {
7755 switch (Opcode) {
7756 case ISD::XOR:
7757 if (N1.isUndef())
7758 // Handle undef ^ undef -> 0 special case. This is a common
7759 // idiom (misuse).
7760 return getConstant(Val: 0, DL, VT);
7761 [[fallthrough]];
7762 case ISD::ADD:
7763 case ISD::PTRADD:
7764 case ISD::SUB:
7765 case ISD::UDIV:
7766 case ISD::SDIV:
7767 case ISD::UREM:
7768 case ISD::SREM:
7769 // fold op(arg1, undef) -> undef, fold op(arg1, poison) -> poison.
7770 return N2.getOpcode() == ISD::POISON ? getPOISON(VT) : getUNDEF(VT);
7771 case ISD::MUL:
7772 case ISD::AND:
7773 case ISD::SSUBSAT:
7774 case ISD::USUBSAT:
7775 // fold op(arg1, undef) -> 0, fold op(arg1, poison) -> poison.
7776 return N2.getOpcode() == ISD::POISON ? getPOISON(VT)
7777 : getConstant(Val: 0, DL, VT);
7778 case ISD::OR:
7779 case ISD::SADDSAT:
7780 case ISD::UADDSAT:
7781 // fold op(arg1, undef) -> an all-ones constant, fold op(arg1, poison) ->
7782 // poison.
7783 return N2.getOpcode() == ISD::POISON ? getPOISON(VT)
7784 : getAllOnesConstant(DL, VT);
7785 }
7786 }
7787
7788 // Memoize this node if possible.
7789 SDNode *N;
7790 SDVTList VTs = getVTList(VT);
7791 SDValue Ops[] = {N1, N2};
7792 if (VT != MVT::Glue) {
7793 FoldingSetNodeID ID;
7794 AddNodeIDNode(ID, OpC: Opcode, VTList: VTs, OpList: Ops);
7795 void *IP = nullptr;
7796 if (SDNode *E = FindNodeOrInsertPos(ID, DL, InsertPos&: IP)) {
7797 E->intersectFlagsWith(Flags);
7798 return SDValue(E, 0);
7799 }
7800
7801 N = newSDNode<SDNode>(Args&: Opcode, Args: DL.getIROrder(), Args: DL.getDebugLoc(), Args&: VTs);
7802 N->setFlags(Flags);
7803 createOperands(Node: N, Vals: Ops);
7804 CSEMap.InsertNode(N, InsertPos: IP);
7805 } else {
7806 N = newSDNode<SDNode>(Args&: Opcode, Args: DL.getIROrder(), Args: DL.getDebugLoc(), Args&: VTs);
7807 createOperands(Node: N, Vals: Ops);
7808 }
7809
7810 InsertNode(N);
7811 SDValue V = SDValue(N, 0);
7812 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
7813 return V;
7814}
7815
7816SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
7817 SDValue N1, SDValue N2, SDValue N3) {
7818 SDNodeFlags Flags;
7819 if (Inserter)
7820 Flags = Inserter->getFlags();
7821 return getNode(Opcode, DL, VT, N1, N2, N3, Flags);
7822}
7823
7824SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
7825 SDValue N1, SDValue N2, SDValue N3,
7826 const SDNodeFlags Flags) {
7827 assert(N1.getOpcode() != ISD::DELETED_NODE &&
7828 N2.getOpcode() != ISD::DELETED_NODE &&
7829 N3.getOpcode() != ISD::DELETED_NODE &&
7830 "Operand is DELETED_NODE!");
7831 // Perform various simplifications.
7832 switch (Opcode) {
7833 case ISD::FMA:
7834 case ISD::FMAD: {
7835 assert(VT.isFloatingPoint() && "This operator only applies to FP types!");
7836 assert(N1.getValueType() == VT && N2.getValueType() == VT &&
7837 N3.getValueType() == VT && "FMA types must match!");
7838 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(Val&: N1);
7839 ConstantFPSDNode *N2CFP = dyn_cast<ConstantFPSDNode>(Val&: N2);
7840 ConstantFPSDNode *N3CFP = dyn_cast<ConstantFPSDNode>(Val&: N3);
7841 if (N1CFP && N2CFP && N3CFP) {
7842 APFloat V1 = N1CFP->getValueAPF();
7843 const APFloat &V2 = N2CFP->getValueAPF();
7844 const APFloat &V3 = N3CFP->getValueAPF();
7845 if (Opcode == ISD::FMAD) {
7846 V1.multiply(RHS: V2, RM: APFloat::rmNearestTiesToEven);
7847 V1.add(RHS: V3, RM: APFloat::rmNearestTiesToEven);
7848 } else
7849 V1.fusedMultiplyAdd(Multiplicand: V2, Addend: V3, RM: APFloat::rmNearestTiesToEven);
7850 return getConstantFP(V: V1, DL, VT);
7851 }
7852 break;
7853 }
7854 case ISD::BUILD_VECTOR: {
7855 // Attempt to simplify BUILD_VECTOR.
7856 SDValue Ops[] = {N1, N2, N3};
7857 if (SDValue V = FoldBUILD_VECTOR(DL, VT, Ops, DAG&: *this))
7858 return V;
7859 break;
7860 }
7861 case ISD::CONCAT_VECTORS: {
7862 SDValue Ops[] = {N1, N2, N3};
7863 if (SDValue V = foldCONCAT_VECTORS(DL, VT, Ops, DAG&: *this))
7864 return V;
7865 break;
7866 }
7867 case ISD::SETCC: {
7868 assert(VT.isInteger() && "SETCC result type must be an integer!");
7869 assert(N1.getValueType() == N2.getValueType() &&
7870 "SETCC operands must have the same type!");
7871 assert(VT.isVector() == N1.getValueType().isVector() &&
7872 "SETCC type should be vector iff the operand type is vector!");
7873 assert((!VT.isVector() || VT.getVectorElementCount() ==
7874 N1.getValueType().getVectorElementCount()) &&
7875 "SETCC vector element counts must match!");
7876 // Use FoldSetCC to simplify SETCC's.
7877 if (SDValue V = FoldSetCC(VT, N1, N2, Cond: cast<CondCodeSDNode>(Val&: N3)->get(), dl: DL))
7878 return V;
7879 // Vector constant folding.
7880 SDValue Ops[] = {N1, N2, N3};
7881 if (SDValue V = FoldConstantArithmetic(Opcode, DL, VT, Ops)) {
7882 NewSDValueDbgMsg(V, Msg: "New node vector constant folding: ", G: this);
7883 return V;
7884 }
7885 break;
7886 }
7887 case ISD::SELECT:
7888 case ISD::VSELECT:
7889 if (SDValue V = simplifySelect(Cond: N1, TVal: N2, FVal: N3))
7890 return V;
7891 break;
7892 case ISD::VECTOR_SHUFFLE:
7893 llvm_unreachable("should use getVectorShuffle constructor!");
7894 case ISD::VECTOR_SPLICE: {
7895 if (cast<ConstantSDNode>(Val&: N3)->isZero())
7896 return N1;
7897 break;
7898 }
7899 case ISD::INSERT_VECTOR_ELT: {
7900 assert(VT.isVector() && VT == N1.getValueType() &&
7901 "INSERT_VECTOR_ELT vector type mismatch");
7902 assert(VT.isFloatingPoint() == N2.getValueType().isFloatingPoint() &&
7903 "INSERT_VECTOR_ELT scalar fp/int mismatch");
7904 assert((!VT.isFloatingPoint() ||
7905 VT.getVectorElementType() == N2.getValueType()) &&
7906 "INSERT_VECTOR_ELT fp scalar type mismatch");
7907 assert((!VT.isInteger() ||
7908 VT.getScalarSizeInBits() <= N2.getScalarValueSizeInBits()) &&
7909 "INSERT_VECTOR_ELT int scalar size mismatch");
7910
7911 auto *N3C = dyn_cast<ConstantSDNode>(Val&: N3);
7912 // INSERT_VECTOR_ELT into out-of-bounds element is an UNDEF, except
7913 // for scalable vectors where we will generate appropriate code to
7914 // deal with out-of-bounds cases correctly.
7915 if (N3C && N1.getValueType().isFixedLengthVector() &&
7916 N3C->getZExtValue() >= N1.getValueType().getVectorNumElements())
7917 return getUNDEF(VT);
7918
7919 // Undefined index can be assumed out-of-bounds, so that's UNDEF too.
7920 if (N3.isUndef())
7921 return getUNDEF(VT);
7922
7923 // If the inserted element is an UNDEF, just use the input vector.
7924 if (N2.isUndef())
7925 return N1;
7926
7927 break;
7928 }
7929 case ISD::INSERT_SUBVECTOR: {
7930 // Inserting undef into undef is still undef.
7931 if (N1.isUndef() && N2.isUndef())
7932 return getUNDEF(VT);
7933
7934 EVT N2VT = N2.getValueType();
7935 assert(VT == N1.getValueType() &&
7936 "Dest and insert subvector source types must match!");
7937 assert(VT.isVector() && N2VT.isVector() &&
7938 "Insert subvector VTs must be vectors!");
7939 assert(VT.getVectorElementType() == N2VT.getVectorElementType() &&
7940 "Insert subvector VTs must have the same element type!");
7941 assert((VT.isScalableVector() || N2VT.isFixedLengthVector()) &&
7942 "Cannot insert a scalable vector into a fixed length vector!");
7943 assert((VT.isScalableVector() != N2VT.isScalableVector() ||
7944 VT.getVectorMinNumElements() >= N2VT.getVectorMinNumElements()) &&
7945 "Insert subvector must be from smaller vector to larger vector!");
7946 assert(isa<ConstantSDNode>(N3) &&
7947 "Insert subvector index must be constant");
7948 assert((VT.isScalableVector() != N2VT.isScalableVector() ||
7949 (N2VT.getVectorMinNumElements() + N3->getAsZExtVal()) <=
7950 VT.getVectorMinNumElements()) &&
7951 "Insert subvector overflow!");
7952 assert(N3->getAsAPIntVal().getBitWidth() ==
7953 TLI->getVectorIdxWidth(getDataLayout()) &&
7954 "Constant index for INSERT_SUBVECTOR has an invalid size");
7955
7956 // Trivial insertion.
7957 if (VT == N2VT)
7958 return N2;
7959
7960 // If this is an insert of an extracted vector into an undef vector, we
7961 // can just use the input to the extract.
7962 if (N1.isUndef() && N2.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
7963 N2.getOperand(i: 1) == N3 && N2.getOperand(i: 0).getValueType() == VT)
7964 return N2.getOperand(i: 0);
7965 break;
7966 }
7967 case ISD::BITCAST:
7968 // Fold bit_convert nodes from a type to themselves.
7969 if (N1.getValueType() == VT)
7970 return N1;
7971 break;
7972 case ISD::VP_TRUNCATE:
7973 case ISD::VP_SIGN_EXTEND:
7974 case ISD::VP_ZERO_EXTEND:
7975 // Don't create noop casts.
7976 if (N1.getValueType() == VT)
7977 return N1;
7978 break;
7979 case ISD::VECTOR_COMPRESS: {
7980 [[maybe_unused]] EVT VecVT = N1.getValueType();
7981 [[maybe_unused]] EVT MaskVT = N2.getValueType();
7982 [[maybe_unused]] EVT PassthruVT = N3.getValueType();
7983 assert(VT == VecVT && "Vector and result type don't match.");
7984 assert(VecVT.isVector() && MaskVT.isVector() && PassthruVT.isVector() &&
7985 "All inputs must be vectors.");
7986 assert(VecVT == PassthruVT && "Vector and passthru types don't match.");
7987 assert(VecVT.getVectorElementCount() == MaskVT.getVectorElementCount() &&
7988 "Vector and mask must have same number of elements.");
7989
7990 if (N1.isUndef() || N2.isUndef())
7991 return N3;
7992
7993 break;
7994 }
7995 case ISD::PARTIAL_REDUCE_UMLA:
7996 case ISD::PARTIAL_REDUCE_SMLA:
7997 case ISD::PARTIAL_REDUCE_SUMLA: {
7998 [[maybe_unused]] EVT AccVT = N1.getValueType();
7999 [[maybe_unused]] EVT Input1VT = N2.getValueType();
8000 [[maybe_unused]] EVT Input2VT = N3.getValueType();
8001 assert(Input1VT.isVector() && Input1VT == Input2VT &&
8002 "Expected the second and third operands of the PARTIAL_REDUCE_MLA "
8003 "node to have the same type!");
8004 assert(VT.isVector() && VT == AccVT &&
8005 "Expected the first operand of the PARTIAL_REDUCE_MLA node to have "
8006 "the same type as its result!");
8007 assert(Input1VT.getVectorElementCount().hasKnownScalarFactor(
8008 AccVT.getVectorElementCount()) &&
8009 "Expected the element count of the second and third operands of the "
8010 "PARTIAL_REDUCE_MLA node to be a positive integer multiple of the "
8011 "element count of the first operand and the result!");
8012 assert(N2.getScalarValueSizeInBits() <= N1.getScalarValueSizeInBits() &&
8013 "Expected the second and third operands of the PARTIAL_REDUCE_MLA "
8014 "node to have an element type which is the same as or smaller than "
8015 "the element type of the first operand and result!");
8016 break;
8017 }
8018 }
8019
8020 // Memoize node if it doesn't produce a glue result.
8021 SDNode *N;
8022 SDVTList VTs = getVTList(VT);
8023 SDValue Ops[] = {N1, N2, N3};
8024 if (VT != MVT::Glue) {
8025 FoldingSetNodeID ID;
8026 AddNodeIDNode(ID, OpC: Opcode, VTList: VTs, OpList: Ops);
8027 void *IP = nullptr;
8028 if (SDNode *E = FindNodeOrInsertPos(ID, DL, InsertPos&: IP)) {
8029 E->intersectFlagsWith(Flags);
8030 return SDValue(E, 0);
8031 }
8032
8033 N = newSDNode<SDNode>(Args&: Opcode, Args: DL.getIROrder(), Args: DL.getDebugLoc(), Args&: VTs);
8034 N->setFlags(Flags);
8035 createOperands(Node: N, Vals: Ops);
8036 CSEMap.InsertNode(N, InsertPos: IP);
8037 } else {
8038 N = newSDNode<SDNode>(Args&: Opcode, Args: DL.getIROrder(), Args: DL.getDebugLoc(), Args&: VTs);
8039 createOperands(Node: N, Vals: Ops);
8040 }
8041
8042 InsertNode(N);
8043 SDValue V = SDValue(N, 0);
8044 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
8045 return V;
8046}
8047
8048SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
8049 SDValue N1, SDValue N2, SDValue N3, SDValue N4,
8050 const SDNodeFlags Flags) {
8051 SDValue Ops[] = { N1, N2, N3, N4 };
8052 return getNode(Opcode, DL, VT, Ops, Flags);
8053}
8054
8055SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
8056 SDValue N1, SDValue N2, SDValue N3, SDValue N4) {
8057 SDNodeFlags Flags;
8058 if (Inserter)
8059 Flags = Inserter->getFlags();
8060 return getNode(Opcode, DL, VT, N1, N2, N3, N4, Flags);
8061}
8062
8063SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
8064 SDValue N1, SDValue N2, SDValue N3, SDValue N4,
8065 SDValue N5, const SDNodeFlags Flags) {
8066 SDValue Ops[] = { N1, N2, N3, N4, N5 };
8067 return getNode(Opcode, DL, VT, Ops, Flags);
8068}
8069
8070SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
8071 SDValue N1, SDValue N2, SDValue N3, SDValue N4,
8072 SDValue N5) {
8073 SDNodeFlags Flags;
8074 if (Inserter)
8075 Flags = Inserter->getFlags();
8076 return getNode(Opcode, DL, VT, N1, N2, N3, N4, N5, Flags);
8077}
8078
8079/// getStackArgumentTokenFactor - Compute a TokenFactor to force all
8080/// the incoming stack arguments to be loaded from the stack.
8081SDValue SelectionDAG::getStackArgumentTokenFactor(SDValue Chain) {
8082 SmallVector<SDValue, 8> ArgChains;
8083
8084 // Include the original chain at the beginning of the list. When this is
8085 // used by target LowerCall hooks, this helps legalize find the
8086 // CALLSEQ_BEGIN node.
8087 ArgChains.push_back(Elt: Chain);
8088
8089 // Add a chain value for each stack argument.
8090 for (SDNode *U : getEntryNode().getNode()->users())
8091 if (LoadSDNode *L = dyn_cast<LoadSDNode>(Val: U))
8092 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Val: L->getBasePtr()))
8093 if (FI->getIndex() < 0)
8094 ArgChains.push_back(Elt: SDValue(L, 1));
8095
8096 // Build a tokenfactor for all the chains.
8097 return getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
8098}
8099
8100/// getMemsetValue - Vectorized representation of the memset value
8101/// operand.
8102static SDValue getMemsetValue(SDValue Value, EVT VT, SelectionDAG &DAG,
8103 const SDLoc &dl) {
8104 assert(!Value.isUndef());
8105
8106 unsigned NumBits = VT.getScalarSizeInBits();
8107 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: Value)) {
8108 assert(C->getAPIntValue().getBitWidth() == 8);
8109 APInt Val = APInt::getSplat(NewLen: NumBits, V: C->getAPIntValue());
8110 if (VT.isInteger()) {
8111 bool IsOpaque = VT.getSizeInBits() > 64 ||
8112 !DAG.getTargetLoweringInfo().isLegalStoreImmediate(Value: C->getSExtValue());
8113 return DAG.getConstant(Val, DL: dl, VT, isT: false, isO: IsOpaque);
8114 }
8115 return DAG.getConstantFP(V: APFloat(VT.getFltSemantics(), Val), DL: dl, VT);
8116 }
8117
8118 assert(Value.getValueType() == MVT::i8 && "memset with non-byte fill value?");
8119 EVT IntVT = VT.getScalarType();
8120 if (!IntVT.isInteger())
8121 IntVT = EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth: IntVT.getSizeInBits());
8122
8123 Value = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL: dl, VT: IntVT, N1: Value);
8124 if (NumBits > 8) {
8125 // Use a multiplication with 0x010101... to extend the input to the
8126 // required length.
8127 APInt Magic = APInt::getSplat(NewLen: NumBits, V: APInt(8, 0x01));
8128 Value = DAG.getNode(Opcode: ISD::MUL, DL: dl, VT: IntVT, N1: Value,
8129 N2: DAG.getConstant(Val: Magic, DL: dl, VT: IntVT));
8130 }
8131
8132 if (VT != Value.getValueType() && !VT.isInteger())
8133 Value = DAG.getBitcast(VT: VT.getScalarType(), V: Value);
8134 if (VT != Value.getValueType())
8135 Value = DAG.getSplatBuildVector(VT, DL: dl, Op: Value);
8136
8137 return Value;
8138}
8139
8140/// getMemsetStringVal - Similar to getMemsetValue. Except this is only
8141/// used when a memcpy is turned into a memset when the source is a constant
8142/// string ptr.
8143static SDValue getMemsetStringVal(EVT VT, const SDLoc &dl, SelectionDAG &DAG,
8144 const TargetLowering &TLI,
8145 const ConstantDataArraySlice &Slice) {
8146 // Handle vector with all elements zero.
8147 if (Slice.Array == nullptr) {
8148 if (VT.isInteger())
8149 return DAG.getConstant(Val: 0, DL: dl, VT);
8150 return DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT,
8151 N1: DAG.getConstant(Val: 0, DL: dl, VT: VT.changeTypeToInteger()));
8152 }
8153
8154 assert(!VT.isVector() && "Can't handle vector type here!");
8155 unsigned NumVTBits = VT.getSizeInBits();
8156 unsigned NumVTBytes = NumVTBits / 8;
8157 unsigned NumBytes = std::min(a: NumVTBytes, b: unsigned(Slice.Length));
8158
8159 APInt Val(NumVTBits, 0);
8160 if (DAG.getDataLayout().isLittleEndian()) {
8161 for (unsigned i = 0; i != NumBytes; ++i)
8162 Val |= (uint64_t)(unsigned char)Slice[i] << i*8;
8163 } else {
8164 for (unsigned i = 0; i != NumBytes; ++i)
8165 Val |= (uint64_t)(unsigned char)Slice[i] << (NumVTBytes-i-1)*8;
8166 }
8167
8168 // If the "cost" of materializing the integer immediate is less than the cost
8169 // of a load, then it is cost effective to turn the load into the immediate.
8170 Type *Ty = VT.getTypeForEVT(Context&: *DAG.getContext());
8171 if (TLI.shouldConvertConstantLoadToIntImm(Imm: Val, Ty))
8172 return DAG.getConstant(Val, DL: dl, VT);
8173 return SDValue();
8174}
8175
8176SDValue SelectionDAG::getMemBasePlusOffset(SDValue Base, TypeSize Offset,
8177 const SDLoc &DL,
8178 const SDNodeFlags Flags) {
8179 EVT VT = Base.getValueType();
8180 SDValue Index;
8181
8182 if (Offset.isScalable())
8183 Index = getVScale(DL, VT: Base.getValueType(),
8184 MulImm: APInt(Base.getValueSizeInBits().getFixedValue(),
8185 Offset.getKnownMinValue()));
8186 else
8187 Index = getConstant(Val: Offset.getFixedValue(), DL, VT);
8188
8189 return getMemBasePlusOffset(Base, Offset: Index, DL, Flags);
8190}
8191
8192SDValue SelectionDAG::getMemBasePlusOffset(SDValue Ptr, SDValue Offset,
8193 const SDLoc &DL,
8194 const SDNodeFlags Flags) {
8195 assert(Offset.getValueType().isInteger());
8196 EVT BasePtrVT = Ptr.getValueType();
8197 if (TLI->shouldPreservePtrArith(F: this->getMachineFunction().getFunction(),
8198 PtrVT: BasePtrVT))
8199 return getNode(Opcode: ISD::PTRADD, DL, VT: BasePtrVT, N1: Ptr, N2: Offset, Flags);
8200 return getNode(Opcode: ISD::ADD, DL, VT: BasePtrVT, N1: Ptr, N2: Offset, Flags);
8201}
8202
8203/// Returns true if memcpy source is constant data.
8204static bool isMemSrcFromConstant(SDValue Src, ConstantDataArraySlice &Slice) {
8205 uint64_t SrcDelta = 0;
8206 GlobalAddressSDNode *G = nullptr;
8207 if (Src.getOpcode() == ISD::GlobalAddress)
8208 G = cast<GlobalAddressSDNode>(Val&: Src);
8209 else if (Src.getOpcode() == ISD::ADD &&
8210 Src.getOperand(i: 0).getOpcode() == ISD::GlobalAddress &&
8211 Src.getOperand(i: 1).getOpcode() == ISD::Constant) {
8212 G = cast<GlobalAddressSDNode>(Val: Src.getOperand(i: 0));
8213 SrcDelta = Src.getConstantOperandVal(i: 1);
8214 }
8215 if (!G)
8216 return false;
8217
8218 return getConstantDataArrayInfo(V: G->getGlobal(), Slice, ElementSize: 8,
8219 Offset: SrcDelta + G->getOffset());
8220}
8221
8222static bool shouldLowerMemFuncForSize(const MachineFunction &MF,
8223 SelectionDAG &DAG) {
8224 // On Darwin, -Os means optimize for size without hurting performance, so
8225 // only really optimize for size when -Oz (MinSize) is used.
8226 if (MF.getTarget().getTargetTriple().isOSDarwin())
8227 return MF.getFunction().hasMinSize();
8228 return DAG.shouldOptForSize();
8229}
8230
8231static void chainLoadsAndStoresForMemcpy(SelectionDAG &DAG, const SDLoc &dl,
8232 SmallVector<SDValue, 32> &OutChains, unsigned From,
8233 unsigned To, SmallVector<SDValue, 16> &OutLoadChains,
8234 SmallVector<SDValue, 16> &OutStoreChains) {
8235 assert(OutLoadChains.size() && "Missing loads in memcpy inlining");
8236 assert(OutStoreChains.size() && "Missing stores in memcpy inlining");
8237 SmallVector<SDValue, 16> GluedLoadChains;
8238 for (unsigned i = From; i < To; ++i) {
8239 OutChains.push_back(Elt: OutLoadChains[i]);
8240 GluedLoadChains.push_back(Elt: OutLoadChains[i]);
8241 }
8242
8243 // Chain for all loads.
8244 SDValue LoadToken = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
8245 GluedLoadChains);
8246
8247 for (unsigned i = From; i < To; ++i) {
8248 StoreSDNode *ST = dyn_cast<StoreSDNode>(Val&: OutStoreChains[i]);
8249 SDValue NewStore = DAG.getTruncStore(Chain: LoadToken, dl, Val: ST->getValue(),
8250 Ptr: ST->getBasePtr(), SVT: ST->getMemoryVT(),
8251 MMO: ST->getMemOperand());
8252 OutChains.push_back(Elt: NewStore);
8253 }
8254}
8255
8256static SDValue getMemcpyLoadsAndStores(
8257 SelectionDAG &DAG, const SDLoc &dl, SDValue Chain, SDValue Dst, SDValue Src,
8258 uint64_t Size, Align Alignment, bool isVol, bool AlwaysInline,
8259 MachinePointerInfo DstPtrInfo, MachinePointerInfo SrcPtrInfo,
8260 const AAMDNodes &AAInfo, BatchAAResults *BatchAA) {
8261 // Turn a memcpy of undef to nop.
8262 // FIXME: We need to honor volatile even is Src is undef.
8263 if (Src.isUndef())
8264 return Chain;
8265
8266 // Expand memcpy to a series of load and store ops if the size operand falls
8267 // below a certain threshold.
8268 // TODO: In the AlwaysInline case, if the size is big then generate a loop
8269 // rather than maybe a humongous number of loads and stores.
8270 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8271 const DataLayout &DL = DAG.getDataLayout();
8272 LLVMContext &C = *DAG.getContext();
8273 std::vector<EVT> MemOps;
8274 bool DstAlignCanChange = false;
8275 MachineFunction &MF = DAG.getMachineFunction();
8276 MachineFrameInfo &MFI = MF.getFrameInfo();
8277 bool OptSize = shouldLowerMemFuncForSize(MF, DAG);
8278 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Val&: Dst);
8279 if (FI && !MFI.isFixedObjectIndex(ObjectIdx: FI->getIndex()))
8280 DstAlignCanChange = true;
8281 MaybeAlign SrcAlign = DAG.InferPtrAlign(Ptr: Src);
8282 if (!SrcAlign || Alignment > *SrcAlign)
8283 SrcAlign = Alignment;
8284 assert(SrcAlign && "SrcAlign must be set");
8285 ConstantDataArraySlice Slice;
8286 // If marked as volatile, perform a copy even when marked as constant.
8287 bool CopyFromConstant = !isVol && isMemSrcFromConstant(Src, Slice);
8288 bool isZeroConstant = CopyFromConstant && Slice.Array == nullptr;
8289 unsigned Limit = AlwaysInline ? ~0U : TLI.getMaxStoresPerMemcpy(OptSize);
8290 const MemOp Op = isZeroConstant
8291 ? MemOp::Set(Size, DstAlignCanChange, DstAlign: Alignment,
8292 /*IsZeroMemset*/ true, IsVolatile: isVol)
8293 : MemOp::Copy(Size, DstAlignCanChange, DstAlign: Alignment,
8294 SrcAlign: *SrcAlign, IsVolatile: isVol, MemcpyStrSrc: CopyFromConstant);
8295 if (!TLI.findOptimalMemOpLowering(
8296 MemOps, Limit, Op, DstAS: DstPtrInfo.getAddrSpace(),
8297 SrcAS: SrcPtrInfo.getAddrSpace(), FuncAttributes: MF.getFunction().getAttributes()))
8298 return SDValue();
8299
8300 if (DstAlignCanChange) {
8301 Type *Ty = MemOps[0].getTypeForEVT(Context&: C);
8302 Align NewAlign = DL.getABITypeAlign(Ty);
8303
8304 // Don't promote to an alignment that would require dynamic stack
8305 // realignment which may conflict with optimizations such as tail call
8306 // optimization.
8307 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
8308 if (!TRI->hasStackRealignment(MF))
8309 if (MaybeAlign StackAlign = DL.getStackAlignment())
8310 NewAlign = std::min(a: NewAlign, b: *StackAlign);
8311
8312 if (NewAlign > Alignment) {
8313 // Give the stack frame object a larger alignment if needed.
8314 if (MFI.getObjectAlign(ObjectIdx: FI->getIndex()) < NewAlign)
8315 MFI.setObjectAlignment(ObjectIdx: FI->getIndex(), Alignment: NewAlign);
8316 Alignment = NewAlign;
8317 }
8318 }
8319
8320 // Prepare AAInfo for loads/stores after lowering this memcpy.
8321 AAMDNodes NewAAInfo = AAInfo;
8322 NewAAInfo.TBAA = NewAAInfo.TBAAStruct = nullptr;
8323
8324 const Value *SrcVal = dyn_cast_if_present<const Value *>(Val&: SrcPtrInfo.V);
8325 bool isConstant =
8326 BatchAA && SrcVal &&
8327 BatchAA->pointsToConstantMemory(Loc: MemoryLocation(SrcVal, Size, AAInfo));
8328
8329 MachineMemOperand::Flags MMOFlags =
8330 isVol ? MachineMemOperand::MOVolatile : MachineMemOperand::MONone;
8331 SmallVector<SDValue, 16> OutLoadChains;
8332 SmallVector<SDValue, 16> OutStoreChains;
8333 SmallVector<SDValue, 32> OutChains;
8334 unsigned NumMemOps = MemOps.size();
8335 uint64_t SrcOff = 0, DstOff = 0;
8336 for (unsigned i = 0; i != NumMemOps; ++i) {
8337 EVT VT = MemOps[i];
8338 unsigned VTSize = VT.getSizeInBits() / 8;
8339 SDValue Value, Store;
8340
8341 if (VTSize > Size) {
8342 // Issuing an unaligned load / store pair that overlaps with the previous
8343 // pair. Adjust the offset accordingly.
8344 assert(i == NumMemOps-1 && i != 0);
8345 SrcOff -= VTSize - Size;
8346 DstOff -= VTSize - Size;
8347 }
8348
8349 if (CopyFromConstant &&
8350 (isZeroConstant || (VT.isInteger() && !VT.isVector()))) {
8351 // It's unlikely a store of a vector immediate can be done in a single
8352 // instruction. It would require a load from a constantpool first.
8353 // We only handle zero vectors here.
8354 // FIXME: Handle other cases where store of vector immediate is done in
8355 // a single instruction.
8356 ConstantDataArraySlice SubSlice;
8357 if (SrcOff < Slice.Length) {
8358 SubSlice = Slice;
8359 SubSlice.move(Delta: SrcOff);
8360 } else {
8361 // This is an out-of-bounds access and hence UB. Pretend we read zero.
8362 SubSlice.Array = nullptr;
8363 SubSlice.Offset = 0;
8364 SubSlice.Length = VTSize;
8365 }
8366 Value = getMemsetStringVal(VT, dl, DAG, TLI, Slice: SubSlice);
8367 if (Value.getNode()) {
8368 Store = DAG.getStore(
8369 Chain, dl, Val: Value,
8370 Ptr: DAG.getMemBasePlusOffset(Base: Dst, Offset: TypeSize::getFixed(ExactSize: DstOff), DL: dl),
8371 PtrInfo: DstPtrInfo.getWithOffset(O: DstOff), Alignment, MMOFlags, AAInfo: NewAAInfo);
8372 OutChains.push_back(Elt: Store);
8373 }
8374 }
8375
8376 if (!Store.getNode()) {
8377 // The type might not be legal for the target. This should only happen
8378 // if the type is smaller than a legal type, as on PPC, so the right
8379 // thing to do is generate a LoadExt/StoreTrunc pair. These simplify
8380 // to Load/Store if NVT==VT.
8381 // FIXME does the case above also need this?
8382 EVT NVT = TLI.getTypeToTransformTo(Context&: C, VT);
8383 assert(NVT.bitsGE(VT));
8384
8385 bool isDereferenceable =
8386 SrcPtrInfo.getWithOffset(O: SrcOff).isDereferenceable(Size: VTSize, C, DL);
8387 MachineMemOperand::Flags SrcMMOFlags = MMOFlags;
8388 if (isDereferenceable)
8389 SrcMMOFlags |= MachineMemOperand::MODereferenceable;
8390 if (isConstant)
8391 SrcMMOFlags |= MachineMemOperand::MOInvariant;
8392
8393 Value = DAG.getExtLoad(
8394 ExtType: ISD::EXTLOAD, dl, VT: NVT, Chain,
8395 Ptr: DAG.getMemBasePlusOffset(Base: Src, Offset: TypeSize::getFixed(ExactSize: SrcOff), DL: dl),
8396 PtrInfo: SrcPtrInfo.getWithOffset(O: SrcOff), MemVT: VT,
8397 Alignment: commonAlignment(A: *SrcAlign, Offset: SrcOff), MMOFlags: SrcMMOFlags, AAInfo: NewAAInfo);
8398 OutLoadChains.push_back(Elt: Value.getValue(R: 1));
8399
8400 Store = DAG.getTruncStore(
8401 Chain, dl, Val: Value,
8402 Ptr: DAG.getMemBasePlusOffset(Base: Dst, Offset: TypeSize::getFixed(ExactSize: DstOff), DL: dl),
8403 PtrInfo: DstPtrInfo.getWithOffset(O: DstOff), SVT: VT, Alignment, MMOFlags, AAInfo: NewAAInfo);
8404 OutStoreChains.push_back(Elt: Store);
8405 }
8406 SrcOff += VTSize;
8407 DstOff += VTSize;
8408 Size -= VTSize;
8409 }
8410
8411 unsigned GluedLdStLimit = MaxLdStGlue == 0 ?
8412 TLI.getMaxGluedStoresPerMemcpy() : MaxLdStGlue;
8413 unsigned NumLdStInMemcpy = OutStoreChains.size();
8414
8415 if (NumLdStInMemcpy) {
8416 // It may be that memcpy might be converted to memset if it's memcpy
8417 // of constants. In such a case, we won't have loads and stores, but
8418 // just stores. In the absence of loads, there is nothing to gang up.
8419 if ((GluedLdStLimit <= 1) || !EnableMemCpyDAGOpt) {
8420 // If target does not care, just leave as it.
8421 for (unsigned i = 0; i < NumLdStInMemcpy; ++i) {
8422 OutChains.push_back(Elt: OutLoadChains[i]);
8423 OutChains.push_back(Elt: OutStoreChains[i]);
8424 }
8425 } else {
8426 // Ld/St less than/equal limit set by target.
8427 if (NumLdStInMemcpy <= GluedLdStLimit) {
8428 chainLoadsAndStoresForMemcpy(DAG, dl, OutChains, From: 0,
8429 To: NumLdStInMemcpy, OutLoadChains,
8430 OutStoreChains);
8431 } else {
8432 unsigned NumberLdChain = NumLdStInMemcpy / GluedLdStLimit;
8433 unsigned RemainingLdStInMemcpy = NumLdStInMemcpy % GluedLdStLimit;
8434 unsigned GlueIter = 0;
8435
8436 for (unsigned cnt = 0; cnt < NumberLdChain; ++cnt) {
8437 unsigned IndexFrom = NumLdStInMemcpy - GlueIter - GluedLdStLimit;
8438 unsigned IndexTo = NumLdStInMemcpy - GlueIter;
8439
8440 chainLoadsAndStoresForMemcpy(DAG, dl, OutChains, From: IndexFrom, To: IndexTo,
8441 OutLoadChains, OutStoreChains);
8442 GlueIter += GluedLdStLimit;
8443 }
8444
8445 // Residual ld/st.
8446 if (RemainingLdStInMemcpy) {
8447 chainLoadsAndStoresForMemcpy(DAG, dl, OutChains, From: 0,
8448 To: RemainingLdStInMemcpy, OutLoadChains,
8449 OutStoreChains);
8450 }
8451 }
8452 }
8453 }
8454 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
8455}
8456
8457static SDValue getMemmoveLoadsAndStores(SelectionDAG &DAG, const SDLoc &dl,
8458 SDValue Chain, SDValue Dst, SDValue Src,
8459 uint64_t Size, Align Alignment,
8460 bool isVol, bool AlwaysInline,
8461 MachinePointerInfo DstPtrInfo,
8462 MachinePointerInfo SrcPtrInfo,
8463 const AAMDNodes &AAInfo) {
8464 // Turn a memmove of undef to nop.
8465 // FIXME: We need to honor volatile even is Src is undef.
8466 if (Src.isUndef())
8467 return Chain;
8468
8469 // Expand memmove to a series of load and store ops if the size operand falls
8470 // below a certain threshold.
8471 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8472 const DataLayout &DL = DAG.getDataLayout();
8473 LLVMContext &C = *DAG.getContext();
8474 std::vector<EVT> MemOps;
8475 bool DstAlignCanChange = false;
8476 MachineFunction &MF = DAG.getMachineFunction();
8477 MachineFrameInfo &MFI = MF.getFrameInfo();
8478 bool OptSize = shouldLowerMemFuncForSize(MF, DAG);
8479 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Val&: Dst);
8480 if (FI && !MFI.isFixedObjectIndex(ObjectIdx: FI->getIndex()))
8481 DstAlignCanChange = true;
8482 MaybeAlign SrcAlign = DAG.InferPtrAlign(Ptr: Src);
8483 if (!SrcAlign || Alignment > *SrcAlign)
8484 SrcAlign = Alignment;
8485 assert(SrcAlign && "SrcAlign must be set");
8486 unsigned Limit = AlwaysInline ? ~0U : TLI.getMaxStoresPerMemmove(OptSize);
8487 if (!TLI.findOptimalMemOpLowering(
8488 MemOps, Limit,
8489 Op: MemOp::Copy(Size, DstAlignCanChange, DstAlign: Alignment, SrcAlign: *SrcAlign,
8490 /*IsVolatile*/ true),
8491 DstAS: DstPtrInfo.getAddrSpace(), SrcAS: SrcPtrInfo.getAddrSpace(),
8492 FuncAttributes: MF.getFunction().getAttributes()))
8493 return SDValue();
8494
8495 if (DstAlignCanChange) {
8496 Type *Ty = MemOps[0].getTypeForEVT(Context&: C);
8497 Align NewAlign = DL.getABITypeAlign(Ty);
8498
8499 // Don't promote to an alignment that would require dynamic stack
8500 // realignment which may conflict with optimizations such as tail call
8501 // optimization.
8502 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
8503 if (!TRI->hasStackRealignment(MF))
8504 if (MaybeAlign StackAlign = DL.getStackAlignment())
8505 NewAlign = std::min(a: NewAlign, b: *StackAlign);
8506
8507 if (NewAlign > Alignment) {
8508 // Give the stack frame object a larger alignment if needed.
8509 if (MFI.getObjectAlign(ObjectIdx: FI->getIndex()) < NewAlign)
8510 MFI.setObjectAlignment(ObjectIdx: FI->getIndex(), Alignment: NewAlign);
8511 Alignment = NewAlign;
8512 }
8513 }
8514
8515 // Prepare AAInfo for loads/stores after lowering this memmove.
8516 AAMDNodes NewAAInfo = AAInfo;
8517 NewAAInfo.TBAA = NewAAInfo.TBAAStruct = nullptr;
8518
8519 MachineMemOperand::Flags MMOFlags =
8520 isVol ? MachineMemOperand::MOVolatile : MachineMemOperand::MONone;
8521 uint64_t SrcOff = 0, DstOff = 0;
8522 SmallVector<SDValue, 8> LoadValues;
8523 SmallVector<SDValue, 8> LoadChains;
8524 SmallVector<SDValue, 8> OutChains;
8525 unsigned NumMemOps = MemOps.size();
8526 for (unsigned i = 0; i < NumMemOps; i++) {
8527 EVT VT = MemOps[i];
8528 unsigned VTSize = VT.getSizeInBits() / 8;
8529 SDValue Value;
8530
8531 bool isDereferenceable =
8532 SrcPtrInfo.getWithOffset(O: SrcOff).isDereferenceable(Size: VTSize, C, DL);
8533 MachineMemOperand::Flags SrcMMOFlags = MMOFlags;
8534 if (isDereferenceable)
8535 SrcMMOFlags |= MachineMemOperand::MODereferenceable;
8536
8537 Value = DAG.getLoad(
8538 VT, dl, Chain,
8539 Ptr: DAG.getMemBasePlusOffset(Base: Src, Offset: TypeSize::getFixed(ExactSize: SrcOff), DL: dl),
8540 PtrInfo: SrcPtrInfo.getWithOffset(O: SrcOff), Alignment: *SrcAlign, MMOFlags: SrcMMOFlags, AAInfo: NewAAInfo);
8541 LoadValues.push_back(Elt: Value);
8542 LoadChains.push_back(Elt: Value.getValue(R: 1));
8543 SrcOff += VTSize;
8544 }
8545 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, LoadChains);
8546 OutChains.clear();
8547 for (unsigned i = 0; i < NumMemOps; i++) {
8548 EVT VT = MemOps[i];
8549 unsigned VTSize = VT.getSizeInBits() / 8;
8550 SDValue Store;
8551
8552 Store = DAG.getStore(
8553 Chain, dl, Val: LoadValues[i],
8554 Ptr: DAG.getMemBasePlusOffset(Base: Dst, Offset: TypeSize::getFixed(ExactSize: DstOff), DL: dl),
8555 PtrInfo: DstPtrInfo.getWithOffset(O: DstOff), Alignment, MMOFlags, AAInfo: NewAAInfo);
8556 OutChains.push_back(Elt: Store);
8557 DstOff += VTSize;
8558 }
8559
8560 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
8561}
8562
8563/// Lower the call to 'memset' intrinsic function into a series of store
8564/// operations.
8565///
8566/// \param DAG Selection DAG where lowered code is placed.
8567/// \param dl Link to corresponding IR location.
8568/// \param Chain Control flow dependency.
8569/// \param Dst Pointer to destination memory location.
8570/// \param Src Value of byte to write into the memory.
8571/// \param Size Number of bytes to write.
8572/// \param Alignment Alignment of the destination in bytes.
8573/// \param isVol True if destination is volatile.
8574/// \param AlwaysInline Makes sure no function call is generated.
8575/// \param DstPtrInfo IR information on the memory pointer.
8576/// \returns New head in the control flow, if lowering was successful, empty
8577/// SDValue otherwise.
8578///
8579/// The function tries to replace 'llvm.memset' intrinsic with several store
8580/// operations and value calculation code. This is usually profitable for small
8581/// memory size or when the semantic requires inlining.
8582static SDValue getMemsetStores(SelectionDAG &DAG, const SDLoc &dl,
8583 SDValue Chain, SDValue Dst, SDValue Src,
8584 uint64_t Size, Align Alignment, bool isVol,
8585 bool AlwaysInline, MachinePointerInfo DstPtrInfo,
8586 const AAMDNodes &AAInfo) {
8587 // Turn a memset of undef to nop.
8588 // FIXME: We need to honor volatile even is Src is undef.
8589 if (Src.isUndef())
8590 return Chain;
8591
8592 // Expand memset to a series of load/store ops if the size operand
8593 // falls below a certain threshold.
8594 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8595 std::vector<EVT> MemOps;
8596 bool DstAlignCanChange = false;
8597 MachineFunction &MF = DAG.getMachineFunction();
8598 MachineFrameInfo &MFI = MF.getFrameInfo();
8599 bool OptSize = shouldLowerMemFuncForSize(MF, DAG);
8600 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Val&: Dst);
8601 if (FI && !MFI.isFixedObjectIndex(ObjectIdx: FI->getIndex()))
8602 DstAlignCanChange = true;
8603 bool IsZeroVal = isNullConstant(V: Src);
8604 unsigned Limit = AlwaysInline ? ~0 : TLI.getMaxStoresPerMemset(OptSize);
8605
8606 if (!TLI.findOptimalMemOpLowering(
8607 MemOps, Limit,
8608 Op: MemOp::Set(Size, DstAlignCanChange, DstAlign: Alignment, IsZeroMemset: IsZeroVal, IsVolatile: isVol),
8609 DstAS: DstPtrInfo.getAddrSpace(), SrcAS: ~0u, FuncAttributes: MF.getFunction().getAttributes()))
8610 return SDValue();
8611
8612 if (DstAlignCanChange) {
8613 Type *Ty = MemOps[0].getTypeForEVT(Context&: *DAG.getContext());
8614 const DataLayout &DL = DAG.getDataLayout();
8615 Align NewAlign = DL.getABITypeAlign(Ty);
8616
8617 // Don't promote to an alignment that would require dynamic stack
8618 // realignment which may conflict with optimizations such as tail call
8619 // optimization.
8620 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
8621 if (!TRI->hasStackRealignment(MF))
8622 if (MaybeAlign StackAlign = DL.getStackAlignment())
8623 NewAlign = std::min(a: NewAlign, b: *StackAlign);
8624
8625 if (NewAlign > Alignment) {
8626 // Give the stack frame object a larger alignment if needed.
8627 if (MFI.getObjectAlign(ObjectIdx: FI->getIndex()) < NewAlign)
8628 MFI.setObjectAlignment(ObjectIdx: FI->getIndex(), Alignment: NewAlign);
8629 Alignment = NewAlign;
8630 }
8631 }
8632
8633 SmallVector<SDValue, 8> OutChains;
8634 uint64_t DstOff = 0;
8635 unsigned NumMemOps = MemOps.size();
8636
8637 // Find the largest store and generate the bit pattern for it.
8638 EVT LargestVT = MemOps[0];
8639 for (unsigned i = 1; i < NumMemOps; i++)
8640 if (MemOps[i].bitsGT(VT: LargestVT))
8641 LargestVT = MemOps[i];
8642 SDValue MemSetValue = getMemsetValue(Value: Src, VT: LargestVT, DAG, dl);
8643
8644 // Prepare AAInfo for loads/stores after lowering this memset.
8645 AAMDNodes NewAAInfo = AAInfo;
8646 NewAAInfo.TBAA = NewAAInfo.TBAAStruct = nullptr;
8647
8648 for (unsigned i = 0; i < NumMemOps; i++) {
8649 EVT VT = MemOps[i];
8650 unsigned VTSize = VT.getSizeInBits() / 8;
8651 if (VTSize > Size) {
8652 // Issuing an unaligned load / store pair that overlaps with the previous
8653 // pair. Adjust the offset accordingly.
8654 assert(i == NumMemOps-1 && i != 0);
8655 DstOff -= VTSize - Size;
8656 }
8657
8658 // If this store is smaller than the largest store see whether we can get
8659 // the smaller value for free with a truncate or extract vector element and
8660 // then store.
8661 SDValue Value = MemSetValue;
8662 if (VT.bitsLT(VT: LargestVT)) {
8663 unsigned Index;
8664 unsigned NElts = LargestVT.getSizeInBits() / VT.getSizeInBits();
8665 EVT SVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: VT.getScalarType(), NumElements: NElts);
8666 if (!LargestVT.isVector() && !VT.isVector() &&
8667 TLI.isTruncateFree(FromVT: LargestVT, ToVT: VT))
8668 Value = DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT, N1: MemSetValue);
8669 else if (LargestVT.isVector() && !VT.isVector() &&
8670 TLI.shallExtractConstSplatVectorElementToStore(
8671 VectorTy: LargestVT.getTypeForEVT(Context&: *DAG.getContext()),
8672 ElemSizeInBits: VT.getSizeInBits(), Index) &&
8673 TLI.isTypeLegal(VT: SVT) &&
8674 LargestVT.getSizeInBits() == SVT.getSizeInBits()) {
8675 // Target which can combine store(extractelement VectorTy, Idx) can get
8676 // the smaller value for free.
8677 SDValue TailValue = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: SVT, N1: MemSetValue);
8678 Value = DAG.getExtractVectorElt(DL: dl, VT, Vec: TailValue, Idx: Index);
8679 } else
8680 Value = getMemsetValue(Value: Src, VT, DAG, dl);
8681 }
8682 assert(Value.getValueType() == VT && "Value with wrong type.");
8683 SDValue Store = DAG.getStore(
8684 Chain, dl, Val: Value,
8685 Ptr: DAG.getMemBasePlusOffset(Base: Dst, Offset: TypeSize::getFixed(ExactSize: DstOff), DL: dl),
8686 PtrInfo: DstPtrInfo.getWithOffset(O: DstOff), Alignment,
8687 MMOFlags: isVol ? MachineMemOperand::MOVolatile : MachineMemOperand::MONone,
8688 AAInfo: NewAAInfo);
8689 OutChains.push_back(Elt: Store);
8690 DstOff += VT.getSizeInBits() / 8;
8691 Size -= VTSize;
8692 }
8693
8694 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
8695}
8696
8697static void checkAddrSpaceIsValidForLibcall(const TargetLowering *TLI,
8698 unsigned AS) {
8699 // Lowering memcpy / memset / memmove intrinsics to calls is only valid if all
8700 // pointer operands can be losslessly bitcasted to pointers of address space 0
8701 if (AS != 0 && !TLI->getTargetMachine().isNoopAddrSpaceCast(SrcAS: AS, DestAS: 0)) {
8702 report_fatal_error(reason: "cannot lower memory intrinsic in address space " +
8703 Twine(AS));
8704 }
8705}
8706
8707SDValue SelectionDAG::getMemcpy(
8708 SDValue Chain, const SDLoc &dl, SDValue Dst, SDValue Src, SDValue Size,
8709 Align Alignment, bool isVol, bool AlwaysInline, const CallInst *CI,
8710 std::optional<bool> OverrideTailCall, MachinePointerInfo DstPtrInfo,
8711 MachinePointerInfo SrcPtrInfo, const AAMDNodes &AAInfo,
8712 BatchAAResults *BatchAA) {
8713 // Check to see if we should lower the memcpy to loads and stores first.
8714 // For cases within the target-specified limits, this is the best choice.
8715 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Val&: Size);
8716 if (ConstantSize) {
8717 // Memcpy with size zero? Just return the original chain.
8718 if (ConstantSize->isZero())
8719 return Chain;
8720
8721 SDValue Result = getMemcpyLoadsAndStores(
8722 DAG&: *this, dl, Chain, Dst, Src, Size: ConstantSize->getZExtValue(), Alignment,
8723 isVol, AlwaysInline: false, DstPtrInfo, SrcPtrInfo, AAInfo, BatchAA);
8724 if (Result.getNode())
8725 return Result;
8726 }
8727
8728 // Then check to see if we should lower the memcpy with target-specific
8729 // code. If the target chooses to do this, this is the next best.
8730 if (TSI) {
8731 SDValue Result = TSI->EmitTargetCodeForMemcpy(
8732 DAG&: *this, dl, Chain, Op1: Dst, Op2: Src, Op3: Size, Alignment, isVolatile: isVol, AlwaysInline,
8733 DstPtrInfo, SrcPtrInfo);
8734 if (Result.getNode())
8735 return Result;
8736 }
8737
8738 // If we really need inline code and the target declined to provide it,
8739 // use a (potentially long) sequence of loads and stores.
8740 if (AlwaysInline) {
8741 assert(ConstantSize && "AlwaysInline requires a constant size!");
8742 return getMemcpyLoadsAndStores(
8743 DAG&: *this, dl, Chain, Dst, Src, Size: ConstantSize->getZExtValue(), Alignment,
8744 isVol, AlwaysInline: true, DstPtrInfo, SrcPtrInfo, AAInfo, BatchAA);
8745 }
8746
8747 checkAddrSpaceIsValidForLibcall(TLI, AS: DstPtrInfo.getAddrSpace());
8748 checkAddrSpaceIsValidForLibcall(TLI, AS: SrcPtrInfo.getAddrSpace());
8749
8750 // FIXME: If the memcpy is volatile (isVol), lowering it to a plain libc
8751 // memcpy is not guaranteed to be safe. libc memcpys aren't required to
8752 // respect volatile, so they may do things like read or write memory
8753 // beyond the given memory regions. But fixing this isn't easy, and most
8754 // people don't care.
8755
8756 // Emit a library call.
8757 TargetLowering::ArgListTy Args;
8758 TargetLowering::ArgListEntry Entry;
8759 Entry.Ty = PointerType::getUnqual(C&: *getContext());
8760 Entry.Node = Dst; Args.push_back(x: Entry);
8761 Entry.Node = Src; Args.push_back(x: Entry);
8762
8763 Entry.Ty = getDataLayout().getIntPtrType(C&: *getContext());
8764 Entry.Node = Size; Args.push_back(x: Entry);
8765 // FIXME: pass in SDLoc
8766 TargetLowering::CallLoweringInfo CLI(*this);
8767 bool IsTailCall = false;
8768 if (OverrideTailCall.has_value()) {
8769 IsTailCall = *OverrideTailCall;
8770 } else {
8771 bool LowersToMemcpy =
8772 TLI->getLibcallName(Call: RTLIB::MEMCPY) == StringRef("memcpy");
8773 bool ReturnsFirstArg = CI && funcReturnsFirstArgOfCall(CI: *CI);
8774 IsTailCall = CI && CI->isTailCall() &&
8775 isInTailCallPosition(Call: *CI, TM: getTarget(),
8776 ReturnsFirstArg: ReturnsFirstArg && LowersToMemcpy);
8777 }
8778
8779 CLI.setDebugLoc(dl)
8780 .setChain(Chain)
8781 .setLibCallee(CC: TLI->getLibcallCallingConv(Call: RTLIB::MEMCPY),
8782 ResultType: Dst.getValueType().getTypeForEVT(Context&: *getContext()),
8783 Target: getExternalSymbol(Sym: TLI->getLibcallName(Call: RTLIB::MEMCPY),
8784 VT: TLI->getPointerTy(DL: getDataLayout())),
8785 ArgsList: std::move(Args))
8786 .setDiscardResult()
8787 .setTailCall(IsTailCall);
8788
8789 std::pair<SDValue,SDValue> CallResult = TLI->LowerCallTo(CLI);
8790 return CallResult.second;
8791}
8792
8793SDValue SelectionDAG::getAtomicMemcpy(SDValue Chain, const SDLoc &dl,
8794 SDValue Dst, SDValue Src, SDValue Size,
8795 Type *SizeTy, unsigned ElemSz,
8796 bool isTailCall,
8797 MachinePointerInfo DstPtrInfo,
8798 MachinePointerInfo SrcPtrInfo) {
8799 // Emit a library call.
8800 TargetLowering::ArgListTy Args;
8801 TargetLowering::ArgListEntry Entry;
8802 Entry.Ty = getDataLayout().getIntPtrType(C&: *getContext());
8803 Entry.Node = Dst;
8804 Args.push_back(x: Entry);
8805
8806 Entry.Node = Src;
8807 Args.push_back(x: Entry);
8808
8809 Entry.Ty = SizeTy;
8810 Entry.Node = Size;
8811 Args.push_back(x: Entry);
8812
8813 RTLIB::Libcall LibraryCall =
8814 RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(ElementSize: ElemSz);
8815 if (LibraryCall == RTLIB::UNKNOWN_LIBCALL)
8816 report_fatal_error(reason: "Unsupported element size");
8817
8818 TargetLowering::CallLoweringInfo CLI(*this);
8819 CLI.setDebugLoc(dl)
8820 .setChain(Chain)
8821 .setLibCallee(CC: TLI->getLibcallCallingConv(Call: LibraryCall),
8822 ResultType: Type::getVoidTy(C&: *getContext()),
8823 Target: getExternalSymbol(Sym: TLI->getLibcallName(Call: LibraryCall),
8824 VT: TLI->getPointerTy(DL: getDataLayout())),
8825 ArgsList: std::move(Args))
8826 .setDiscardResult()
8827 .setTailCall(isTailCall);
8828
8829 std::pair<SDValue, SDValue> CallResult = TLI->LowerCallTo(CLI);
8830 return CallResult.second;
8831}
8832
8833SDValue SelectionDAG::getMemmove(SDValue Chain, const SDLoc &dl, SDValue Dst,
8834 SDValue Src, SDValue Size, Align Alignment,
8835 bool isVol, const CallInst *CI,
8836 std::optional<bool> OverrideTailCall,
8837 MachinePointerInfo DstPtrInfo,
8838 MachinePointerInfo SrcPtrInfo,
8839 const AAMDNodes &AAInfo,
8840 BatchAAResults *BatchAA) {
8841 // Check to see if we should lower the memmove to loads and stores first.
8842 // For cases within the target-specified limits, this is the best choice.
8843 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Val&: Size);
8844 if (ConstantSize) {
8845 // Memmove with size zero? Just return the original chain.
8846 if (ConstantSize->isZero())
8847 return Chain;
8848
8849 SDValue Result = getMemmoveLoadsAndStores(
8850 DAG&: *this, dl, Chain, Dst, Src, Size: ConstantSize->getZExtValue(), Alignment,
8851 isVol, AlwaysInline: false, DstPtrInfo, SrcPtrInfo, AAInfo);
8852 if (Result.getNode())
8853 return Result;
8854 }
8855
8856 // Then check to see if we should lower the memmove with target-specific
8857 // code. If the target chooses to do this, this is the next best.
8858 if (TSI) {
8859 SDValue Result =
8860 TSI->EmitTargetCodeForMemmove(DAG&: *this, dl, Chain, Op1: Dst, Op2: Src, Op3: Size,
8861 Alignment, isVolatile: isVol, DstPtrInfo, SrcPtrInfo);
8862 if (Result.getNode())
8863 return Result;
8864 }
8865
8866 checkAddrSpaceIsValidForLibcall(TLI, AS: DstPtrInfo.getAddrSpace());
8867 checkAddrSpaceIsValidForLibcall(TLI, AS: SrcPtrInfo.getAddrSpace());
8868
8869 // FIXME: If the memmove is volatile, lowering it to plain libc memmove may
8870 // not be safe. See memcpy above for more details.
8871
8872 // Emit a library call.
8873 TargetLowering::ArgListTy Args;
8874 TargetLowering::ArgListEntry Entry;
8875 Entry.Ty = PointerType::getUnqual(C&: *getContext());
8876 Entry.Node = Dst; Args.push_back(x: Entry);
8877 Entry.Node = Src; Args.push_back(x: Entry);
8878
8879 Entry.Ty = getDataLayout().getIntPtrType(C&: *getContext());
8880 Entry.Node = Size; Args.push_back(x: Entry);
8881 // FIXME: pass in SDLoc
8882 TargetLowering::CallLoweringInfo CLI(*this);
8883
8884 bool IsTailCall = false;
8885 if (OverrideTailCall.has_value()) {
8886 IsTailCall = *OverrideTailCall;
8887 } else {
8888 bool LowersToMemmove =
8889 TLI->getLibcallName(Call: RTLIB::MEMMOVE) == StringRef("memmove");
8890 bool ReturnsFirstArg = CI && funcReturnsFirstArgOfCall(CI: *CI);
8891 IsTailCall = CI && CI->isTailCall() &&
8892 isInTailCallPosition(Call: *CI, TM: getTarget(),
8893 ReturnsFirstArg: ReturnsFirstArg && LowersToMemmove);
8894 }
8895
8896 CLI.setDebugLoc(dl)
8897 .setChain(Chain)
8898 .setLibCallee(CC: TLI->getLibcallCallingConv(Call: RTLIB::MEMMOVE),
8899 ResultType: Dst.getValueType().getTypeForEVT(Context&: *getContext()),
8900 Target: getExternalSymbol(Sym: TLI->getLibcallName(Call: RTLIB::MEMMOVE),
8901 VT: TLI->getPointerTy(DL: getDataLayout())),
8902 ArgsList: std::move(Args))
8903 .setDiscardResult()
8904 .setTailCall(IsTailCall);
8905
8906 std::pair<SDValue,SDValue> CallResult = TLI->LowerCallTo(CLI);
8907 return CallResult.second;
8908}
8909
8910SDValue SelectionDAG::getAtomicMemmove(SDValue Chain, const SDLoc &dl,
8911 SDValue Dst, SDValue Src, SDValue Size,
8912 Type *SizeTy, unsigned ElemSz,
8913 bool isTailCall,
8914 MachinePointerInfo DstPtrInfo,
8915 MachinePointerInfo SrcPtrInfo) {
8916 // Emit a library call.
8917 TargetLowering::ArgListTy Args;
8918 TargetLowering::ArgListEntry Entry;
8919 Entry.Ty = getDataLayout().getIntPtrType(C&: *getContext());
8920 Entry.Node = Dst;
8921 Args.push_back(x: Entry);
8922
8923 Entry.Node = Src;
8924 Args.push_back(x: Entry);
8925
8926 Entry.Ty = SizeTy;
8927 Entry.Node = Size;
8928 Args.push_back(x: Entry);
8929
8930 RTLIB::Libcall LibraryCall =
8931 RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(ElementSize: ElemSz);
8932 if (LibraryCall == RTLIB::UNKNOWN_LIBCALL)
8933 report_fatal_error(reason: "Unsupported element size");
8934
8935 TargetLowering::CallLoweringInfo CLI(*this);
8936 CLI.setDebugLoc(dl)
8937 .setChain(Chain)
8938 .setLibCallee(CC: TLI->getLibcallCallingConv(Call: LibraryCall),
8939 ResultType: Type::getVoidTy(C&: *getContext()),
8940 Target: getExternalSymbol(Sym: TLI->getLibcallName(Call: LibraryCall),
8941 VT: TLI->getPointerTy(DL: getDataLayout())),
8942 ArgsList: std::move(Args))
8943 .setDiscardResult()
8944 .setTailCall(isTailCall);
8945
8946 std::pair<SDValue, SDValue> CallResult = TLI->LowerCallTo(CLI);
8947 return CallResult.second;
8948}
8949
8950SDValue SelectionDAG::getMemset(SDValue Chain, const SDLoc &dl, SDValue Dst,
8951 SDValue Src, SDValue Size, Align Alignment,
8952 bool isVol, bool AlwaysInline,
8953 const CallInst *CI,
8954 MachinePointerInfo DstPtrInfo,
8955 const AAMDNodes &AAInfo) {
8956 // Check to see if we should lower the memset to stores first.
8957 // For cases within the target-specified limits, this is the best choice.
8958 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Val&: Size);
8959 if (ConstantSize) {
8960 // Memset with size zero? Just return the original chain.
8961 if (ConstantSize->isZero())
8962 return Chain;
8963
8964 SDValue Result = getMemsetStores(DAG&: *this, dl, Chain, Dst, Src,
8965 Size: ConstantSize->getZExtValue(), Alignment,
8966 isVol, AlwaysInline: false, DstPtrInfo, AAInfo);
8967
8968 if (Result.getNode())
8969 return Result;
8970 }
8971
8972 // Then check to see if we should lower the memset with target-specific
8973 // code. If the target chooses to do this, this is the next best.
8974 if (TSI) {
8975 SDValue Result = TSI->EmitTargetCodeForMemset(
8976 DAG&: *this, dl, Chain, Op1: Dst, Op2: Src, Op3: Size, Alignment, isVolatile: isVol, AlwaysInline, DstPtrInfo);
8977 if (Result.getNode())
8978 return Result;
8979 }
8980
8981 // If we really need inline code and the target declined to provide it,
8982 // use a (potentially long) sequence of loads and stores.
8983 if (AlwaysInline) {
8984 assert(ConstantSize && "AlwaysInline requires a constant size!");
8985 SDValue Result = getMemsetStores(DAG&: *this, dl, Chain, Dst, Src,
8986 Size: ConstantSize->getZExtValue(), Alignment,
8987 isVol, AlwaysInline: true, DstPtrInfo, AAInfo);
8988 assert(Result &&
8989 "getMemsetStores must return a valid sequence when AlwaysInline");
8990 return Result;
8991 }
8992
8993 checkAddrSpaceIsValidForLibcall(TLI, AS: DstPtrInfo.getAddrSpace());
8994
8995 // Emit a library call.
8996 auto &Ctx = *getContext();
8997 const auto& DL = getDataLayout();
8998
8999 TargetLowering::CallLoweringInfo CLI(*this);
9000 // FIXME: pass in SDLoc
9001 CLI.setDebugLoc(dl).setChain(Chain);
9002
9003 const char *BzeroName = getTargetLoweringInfo().getLibcallName(Call: RTLIB::BZERO);
9004
9005 // Helper function to create an Entry from Node and Type.
9006 const auto CreateEntry = [](SDValue Node, Type *Ty) {
9007 TargetLowering::ArgListEntry Entry;
9008 Entry.Node = Node;
9009 Entry.Ty = Ty;
9010 return Entry;
9011 };
9012
9013 bool UseBZero = isNullConstant(V: Src) && BzeroName;
9014 // If zeroing out and bzero is present, use it.
9015 if (UseBZero) {
9016 TargetLowering::ArgListTy Args;
9017 Args.push_back(x: CreateEntry(Dst, PointerType::getUnqual(C&: Ctx)));
9018 Args.push_back(x: CreateEntry(Size, DL.getIntPtrType(C&: Ctx)));
9019 CLI.setLibCallee(
9020 CC: TLI->getLibcallCallingConv(Call: RTLIB::BZERO), ResultType: Type::getVoidTy(C&: Ctx),
9021 Target: getExternalSymbol(Sym: BzeroName, VT: TLI->getPointerTy(DL)), ArgsList: std::move(Args));
9022 } else {
9023 TargetLowering::ArgListTy Args;
9024 Args.push_back(x: CreateEntry(Dst, PointerType::getUnqual(C&: Ctx)));
9025 Args.push_back(x: CreateEntry(Src, Src.getValueType().getTypeForEVT(Context&: Ctx)));
9026 Args.push_back(x: CreateEntry(Size, DL.getIntPtrType(C&: Ctx)));
9027 CLI.setLibCallee(CC: TLI->getLibcallCallingConv(Call: RTLIB::MEMSET),
9028 ResultType: Dst.getValueType().getTypeForEVT(Context&: Ctx),
9029 Target: getExternalSymbol(Sym: TLI->getLibcallName(Call: RTLIB::MEMSET),
9030 VT: TLI->getPointerTy(DL)),
9031 ArgsList: std::move(Args));
9032 }
9033 bool LowersToMemset =
9034 TLI->getLibcallName(Call: RTLIB::MEMSET) == StringRef("memset");
9035 // If we're going to use bzero, make sure not to tail call unless the
9036 // subsequent return doesn't need a value, as bzero doesn't return the first
9037 // arg unlike memset.
9038 bool ReturnsFirstArg = CI && funcReturnsFirstArgOfCall(CI: *CI) && !UseBZero;
9039 bool IsTailCall =
9040 CI && CI->isTailCall() &&
9041 isInTailCallPosition(Call: *CI, TM: getTarget(), ReturnsFirstArg: ReturnsFirstArg && LowersToMemset);
9042 CLI.setDiscardResult().setTailCall(IsTailCall);
9043
9044 std::pair<SDValue, SDValue> CallResult = TLI->LowerCallTo(CLI);
9045 return CallResult.second;
9046}
9047
9048SDValue SelectionDAG::getAtomicMemset(SDValue Chain, const SDLoc &dl,
9049 SDValue Dst, SDValue Value, SDValue Size,
9050 Type *SizeTy, unsigned ElemSz,
9051 bool isTailCall,
9052 MachinePointerInfo DstPtrInfo) {
9053 // Emit a library call.
9054 TargetLowering::ArgListTy Args;
9055 TargetLowering::ArgListEntry Entry;
9056 Entry.Ty = getDataLayout().getIntPtrType(C&: *getContext());
9057 Entry.Node = Dst;
9058 Args.push_back(x: Entry);
9059
9060 Entry.Ty = Type::getInt8Ty(C&: *getContext());
9061 Entry.Node = Value;
9062 Args.push_back(x: Entry);
9063
9064 Entry.Ty = SizeTy;
9065 Entry.Node = Size;
9066 Args.push_back(x: Entry);
9067
9068 RTLIB::Libcall LibraryCall =
9069 RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(ElementSize: ElemSz);
9070 if (LibraryCall == RTLIB::UNKNOWN_LIBCALL)
9071 report_fatal_error(reason: "Unsupported element size");
9072
9073 TargetLowering::CallLoweringInfo CLI(*this);
9074 CLI.setDebugLoc(dl)
9075 .setChain(Chain)
9076 .setLibCallee(CC: TLI->getLibcallCallingConv(Call: LibraryCall),
9077 ResultType: Type::getVoidTy(C&: *getContext()),
9078 Target: getExternalSymbol(Sym: TLI->getLibcallName(Call: LibraryCall),
9079 VT: TLI->getPointerTy(DL: getDataLayout())),
9080 ArgsList: std::move(Args))
9081 .setDiscardResult()
9082 .setTailCall(isTailCall);
9083
9084 std::pair<SDValue, SDValue> CallResult = TLI->LowerCallTo(CLI);
9085 return CallResult.second;
9086}
9087
9088SDValue SelectionDAG::getAtomic(unsigned Opcode, const SDLoc &dl, EVT MemVT,
9089 SDVTList VTList, ArrayRef<SDValue> Ops,
9090 MachineMemOperand *MMO,
9091 ISD::LoadExtType ExtType) {
9092 FoldingSetNodeID ID;
9093 AddNodeIDNode(ID, OpC: Opcode, VTList, OpList: Ops);
9094 ID.AddInteger(I: MemVT.getRawBits());
9095 ID.AddInteger(I: getSyntheticNodeSubclassData<AtomicSDNode>(
9096 IROrder: dl.getIROrder(), Args&: Opcode, Args&: VTList, Args&: MemVT, Args&: MMO, Args&: ExtType));
9097 ID.AddInteger(I: MMO->getPointerInfo().getAddrSpace());
9098 ID.AddInteger(I: MMO->getFlags());
9099 void* IP = nullptr;
9100 if (SDNode *E = FindNodeOrInsertPos(ID, DL: dl, InsertPos&: IP)) {
9101 cast<AtomicSDNode>(Val: E)->refineAlignment(NewMMO: MMO);
9102 return SDValue(E, 0);
9103 }
9104
9105 auto *N = newSDNode<AtomicSDNode>(Args: dl.getIROrder(), Args: dl.getDebugLoc(), Args&: Opcode,
9106 Args&: VTList, Args&: MemVT, Args&: MMO, Args&: ExtType);
9107 createOperands(Node: N, Vals: Ops);
9108
9109 CSEMap.InsertNode(N, InsertPos: IP);
9110 InsertNode(N);
9111 SDValue V(N, 0);
9112 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
9113 return V;
9114}
9115
9116SDValue SelectionDAG::getAtomicCmpSwap(unsigned Opcode, const SDLoc &dl,
9117 EVT MemVT, SDVTList VTs, SDValue Chain,
9118 SDValue Ptr, SDValue Cmp, SDValue Swp,
9119 MachineMemOperand *MMO) {
9120 assert(Opcode == ISD::ATOMIC_CMP_SWAP ||
9121 Opcode == ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS);
9122 assert(Cmp.getValueType() == Swp.getValueType() && "Invalid Atomic Op Types");
9123
9124 SDValue Ops[] = {Chain, Ptr, Cmp, Swp};
9125 return getAtomic(Opcode, dl, MemVT, VTList: VTs, Ops, MMO);
9126}
9127
9128SDValue SelectionDAG::getAtomic(unsigned Opcode, const SDLoc &dl, EVT MemVT,
9129 SDValue Chain, SDValue Ptr, SDValue Val,
9130 MachineMemOperand *MMO) {
9131 assert((Opcode == ISD::ATOMIC_LOAD_ADD || Opcode == ISD::ATOMIC_LOAD_SUB ||
9132 Opcode == ISD::ATOMIC_LOAD_AND || Opcode == ISD::ATOMIC_LOAD_CLR ||
9133 Opcode == ISD::ATOMIC_LOAD_OR || Opcode == ISD::ATOMIC_LOAD_XOR ||
9134 Opcode == ISD::ATOMIC_LOAD_NAND || Opcode == ISD::ATOMIC_LOAD_MIN ||
9135 Opcode == ISD::ATOMIC_LOAD_MAX || Opcode == ISD::ATOMIC_LOAD_UMIN ||
9136 Opcode == ISD::ATOMIC_LOAD_UMAX || Opcode == ISD::ATOMIC_LOAD_FADD ||
9137 Opcode == ISD::ATOMIC_LOAD_FSUB || Opcode == ISD::ATOMIC_LOAD_FMAX ||
9138 Opcode == ISD::ATOMIC_LOAD_FMIN ||
9139 Opcode == ISD::ATOMIC_LOAD_FMINIMUM ||
9140 Opcode == ISD::ATOMIC_LOAD_FMAXIMUM ||
9141 Opcode == ISD::ATOMIC_LOAD_UINC_WRAP ||
9142 Opcode == ISD::ATOMIC_LOAD_UDEC_WRAP ||
9143 Opcode == ISD::ATOMIC_LOAD_USUB_COND ||
9144 Opcode == ISD::ATOMIC_LOAD_USUB_SAT || Opcode == ISD::ATOMIC_SWAP ||
9145 Opcode == ISD::ATOMIC_STORE) &&
9146 "Invalid Atomic Op");
9147
9148 EVT VT = Val.getValueType();
9149
9150 SDVTList VTs = Opcode == ISD::ATOMIC_STORE ? getVTList(MVT::Other) :
9151 getVTList(VT, MVT::Other);
9152 SDValue Ops[] = {Chain, Ptr, Val};
9153 return getAtomic(Opcode, dl, MemVT, VTList: VTs, Ops, MMO);
9154}
9155
9156SDValue SelectionDAG::getAtomicLoad(ISD::LoadExtType ExtType, const SDLoc &dl,
9157 EVT MemVT, EVT VT, SDValue Chain,
9158 SDValue Ptr, MachineMemOperand *MMO) {
9159 SDVTList VTs = getVTList(VT, MVT::Other);
9160 SDValue Ops[] = {Chain, Ptr};
9161 return getAtomic(Opcode: ISD::ATOMIC_LOAD, dl, MemVT, VTList: VTs, Ops, MMO, ExtType);
9162}
9163
9164/// getMergeValues - Create a MERGE_VALUES node from the given operands.
9165SDValue SelectionDAG::getMergeValues(ArrayRef<SDValue> Ops, const SDLoc &dl) {
9166 if (Ops.size() == 1)
9167 return Ops[0];
9168
9169 SmallVector<EVT, 4> VTs;
9170 VTs.reserve(N: Ops.size());
9171 for (const SDValue &Op : Ops)
9172 VTs.push_back(Elt: Op.getValueType());
9173 return getNode(Opcode: ISD::MERGE_VALUES, DL: dl, VTList: getVTList(VTs), Ops);
9174}
9175
9176SDValue SelectionDAG::getMemIntrinsicNode(
9177 unsigned Opcode, const SDLoc &dl, SDVTList VTList, ArrayRef<SDValue> Ops,
9178 EVT MemVT, MachinePointerInfo PtrInfo, Align Alignment,
9179 MachineMemOperand::Flags Flags, LocationSize Size,
9180 const AAMDNodes &AAInfo) {
9181 if (Size.hasValue() && !Size.getValue())
9182 Size = LocationSize::precise(Value: MemVT.getStoreSize());
9183
9184 MachineFunction &MF = getMachineFunction();
9185 MachineMemOperand *MMO =
9186 MF.getMachineMemOperand(PtrInfo, F: Flags, Size, BaseAlignment: Alignment, AAInfo);
9187
9188 return getMemIntrinsicNode(Opcode, dl, VTList, Ops, MemVT, MMO);
9189}
9190
9191SDValue SelectionDAG::getMemIntrinsicNode(unsigned Opcode, const SDLoc &dl,
9192 SDVTList VTList,
9193 ArrayRef<SDValue> Ops, EVT MemVT,
9194 MachineMemOperand *MMO) {
9195 assert(
9196 (Opcode == ISD::INTRINSIC_VOID || Opcode == ISD::INTRINSIC_W_CHAIN ||
9197 Opcode == ISD::PREFETCH ||
9198 (Opcode <= (unsigned)std::numeric_limits<int>::max() &&
9199 Opcode >= ISD::BUILTIN_OP_END && TSI->isTargetMemoryOpcode(Opcode))) &&
9200 "Opcode is not a memory-accessing opcode!");
9201
9202 // Memoize the node unless it returns a glue result.
9203 MemIntrinsicSDNode *N;
9204 if (VTList.VTs[VTList.NumVTs-1] != MVT::Glue) {
9205 FoldingSetNodeID ID;
9206 AddNodeIDNode(ID, OpC: Opcode, VTList, OpList: Ops);
9207 ID.AddInteger(I: getSyntheticNodeSubclassData<MemIntrinsicSDNode>(
9208 Opc: Opcode, Order: dl.getIROrder(), VTs: VTList, MemoryVT: MemVT, MMO));
9209 ID.AddInteger(I: MMO->getPointerInfo().getAddrSpace());
9210 ID.AddInteger(I: MMO->getFlags());
9211 ID.AddInteger(I: MemVT.getRawBits());
9212 void *IP = nullptr;
9213 if (SDNode *E = FindNodeOrInsertPos(ID, DL: dl, InsertPos&: IP)) {
9214 cast<MemIntrinsicSDNode>(Val: E)->refineAlignment(NewMMO: MMO);
9215 return SDValue(E, 0);
9216 }
9217
9218 N = newSDNode<MemIntrinsicSDNode>(Args&: Opcode, Args: dl.getIROrder(), Args: dl.getDebugLoc(),
9219 Args&: VTList, Args&: MemVT, Args&: MMO);
9220 createOperands(Node: N, Vals: Ops);
9221
9222 CSEMap.InsertNode(N, InsertPos: IP);
9223 } else {
9224 N = newSDNode<MemIntrinsicSDNode>(Args&: Opcode, Args: dl.getIROrder(), Args: dl.getDebugLoc(),
9225 Args&: VTList, Args&: MemVT, Args&: MMO);
9226 createOperands(Node: N, Vals: Ops);
9227 }
9228 InsertNode(N);
9229 SDValue V(N, 0);
9230 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
9231 return V;
9232}
9233
9234SDValue SelectionDAG::getLifetimeNode(bool IsStart, const SDLoc &dl,
9235 SDValue Chain, int FrameIndex,
9236 int64_t Size, int64_t Offset) {
9237 const unsigned Opcode = IsStart ? ISD::LIFETIME_START : ISD::LIFETIME_END;
9238 const auto VTs = getVTList(MVT::Other);
9239 SDValue Ops[2] = {
9240 Chain,
9241 getFrameIndex(FI: FrameIndex,
9242 VT: getTargetLoweringInfo().getFrameIndexTy(DL: getDataLayout()),
9243 isTarget: true)};
9244
9245 FoldingSetNodeID ID;
9246 AddNodeIDNode(ID, Opcode, VTs, Ops);
9247 ID.AddInteger(I: FrameIndex);
9248 ID.AddInteger(I: Size);
9249 ID.AddInteger(I: Offset);
9250 void *IP = nullptr;
9251 if (SDNode *E = FindNodeOrInsertPos(ID, DL: dl, InsertPos&: IP))
9252 return SDValue(E, 0);
9253
9254 LifetimeSDNode *N = newSDNode<LifetimeSDNode>(
9255 Opcode, dl.getIROrder(), dl.getDebugLoc(), VTs, Size, Offset);
9256 createOperands(Node: N, Vals: Ops);
9257 CSEMap.InsertNode(N, InsertPos: IP);
9258 InsertNode(N);
9259 SDValue V(N, 0);
9260 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
9261 return V;
9262}
9263
9264SDValue SelectionDAG::getPseudoProbeNode(const SDLoc &Dl, SDValue Chain,
9265 uint64_t Guid, uint64_t Index,
9266 uint32_t Attr) {
9267 const unsigned Opcode = ISD::PSEUDO_PROBE;
9268 const auto VTs = getVTList(MVT::Other);
9269 SDValue Ops[] = {Chain};
9270 FoldingSetNodeID ID;
9271 AddNodeIDNode(ID, Opcode, VTs, Ops);
9272 ID.AddInteger(I: Guid);
9273 ID.AddInteger(I: Index);
9274 void *IP = nullptr;
9275 if (SDNode *E = FindNodeOrInsertPos(ID, DL: Dl, InsertPos&: IP))
9276 return SDValue(E, 0);
9277
9278 auto *N = newSDNode<PseudoProbeSDNode>(
9279 Opcode, Dl.getIROrder(), Dl.getDebugLoc(), VTs, Guid, Index, Attr);
9280 createOperands(Node: N, Vals: Ops);
9281 CSEMap.InsertNode(N, IP);
9282 InsertNode(N: N);
9283 SDValue V(N, 0);
9284 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
9285 return V;
9286}
9287
9288/// InferPointerInfo - If the specified ptr/offset is a frame index, infer a
9289/// MachinePointerInfo record from it. This is particularly useful because the
9290/// code generator has many cases where it doesn't bother passing in a
9291/// MachinePointerInfo to getLoad or getStore when it has "FI+Cst".
9292static MachinePointerInfo InferPointerInfo(const MachinePointerInfo &Info,
9293 SelectionDAG &DAG, SDValue Ptr,
9294 int64_t Offset = 0) {
9295 // If this is FI+Offset, we can model it.
9296 if (const FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Val&: Ptr))
9297 return MachinePointerInfo::getFixedStack(MF&: DAG.getMachineFunction(),
9298 FI: FI->getIndex(), Offset);
9299
9300 // If this is (FI+Offset1)+Offset2, we can model it.
9301 if (Ptr.getOpcode() != ISD::ADD ||
9302 !isa<ConstantSDNode>(Val: Ptr.getOperand(i: 1)) ||
9303 !isa<FrameIndexSDNode>(Val: Ptr.getOperand(i: 0)))
9304 return Info;
9305
9306 int FI = cast<FrameIndexSDNode>(Val: Ptr.getOperand(i: 0))->getIndex();
9307 return MachinePointerInfo::getFixedStack(
9308 MF&: DAG.getMachineFunction(), FI,
9309 Offset: Offset + cast<ConstantSDNode>(Val: Ptr.getOperand(i: 1))->getSExtValue());
9310}
9311
9312/// InferPointerInfo - If the specified ptr/offset is a frame index, infer a
9313/// MachinePointerInfo record from it. This is particularly useful because the
9314/// code generator has many cases where it doesn't bother passing in a
9315/// MachinePointerInfo to getLoad or getStore when it has "FI+Cst".
9316static MachinePointerInfo InferPointerInfo(const MachinePointerInfo &Info,
9317 SelectionDAG &DAG, SDValue Ptr,
9318 SDValue OffsetOp) {
9319 // If the 'Offset' value isn't a constant, we can't handle this.
9320 if (ConstantSDNode *OffsetNode = dyn_cast<ConstantSDNode>(Val&: OffsetOp))
9321 return InferPointerInfo(Info, DAG, Ptr, Offset: OffsetNode->getSExtValue());
9322 if (OffsetOp.isUndef())
9323 return InferPointerInfo(Info, DAG, Ptr);
9324 return Info;
9325}
9326
9327SDValue SelectionDAG::getLoad(ISD::MemIndexedMode AM, ISD::LoadExtType ExtType,
9328 EVT VT, const SDLoc &dl, SDValue Chain,
9329 SDValue Ptr, SDValue Offset,
9330 MachinePointerInfo PtrInfo, EVT MemVT,
9331 Align Alignment,
9332 MachineMemOperand::Flags MMOFlags,
9333 const AAMDNodes &AAInfo, const MDNode *Ranges) {
9334 assert(Chain.getValueType() == MVT::Other &&
9335 "Invalid chain type");
9336
9337 MMOFlags |= MachineMemOperand::MOLoad;
9338 assert((MMOFlags & MachineMemOperand::MOStore) == 0);
9339 // If we don't have a PtrInfo, infer the trivial frame index case to simplify
9340 // clients.
9341 if (PtrInfo.V.isNull())
9342 PtrInfo = InferPointerInfo(Info: PtrInfo, DAG&: *this, Ptr, OffsetOp: Offset);
9343
9344 TypeSize Size = MemVT.getStoreSize();
9345 MachineFunction &MF = getMachineFunction();
9346 MachineMemOperand *MMO = MF.getMachineMemOperand(PtrInfo, F: MMOFlags, Size,
9347 BaseAlignment: Alignment, AAInfo, Ranges);
9348 return getLoad(AM, ExtType, VT, dl, Chain, Ptr, Offset, MemVT, MMO);
9349}
9350
9351SDValue SelectionDAG::getLoad(ISD::MemIndexedMode AM, ISD::LoadExtType ExtType,
9352 EVT VT, const SDLoc &dl, SDValue Chain,
9353 SDValue Ptr, SDValue Offset, EVT MemVT,
9354 MachineMemOperand *MMO) {
9355 if (VT == MemVT) {
9356 ExtType = ISD::NON_EXTLOAD;
9357 } else if (ExtType == ISD::NON_EXTLOAD) {
9358 assert(VT == MemVT && "Non-extending load from different memory type!");
9359 } else {
9360 // Extending load.
9361 assert(MemVT.getScalarType().bitsLT(VT.getScalarType()) &&
9362 "Should only be an extending load, not truncating!");
9363 assert(VT.isInteger() == MemVT.isInteger() &&
9364 "Cannot convert from FP to Int or Int -> FP!");
9365 assert(VT.isVector() == MemVT.isVector() &&
9366 "Cannot use an ext load to convert to or from a vector!");
9367 assert((!VT.isVector() ||
9368 VT.getVectorElementCount() == MemVT.getVectorElementCount()) &&
9369 "Cannot use an ext load to change the number of vector elements!");
9370 }
9371
9372 assert((!MMO->getRanges() ||
9373 (mdconst::extract<ConstantInt>(MMO->getRanges()->getOperand(0))
9374 ->getBitWidth() == MemVT.getScalarSizeInBits() &&
9375 MemVT.isInteger())) &&
9376 "Range metadata and load type must match!");
9377
9378 bool Indexed = AM != ISD::UNINDEXED;
9379 assert((Indexed || Offset.isUndef()) && "Unindexed load with an offset!");
9380
9381 SDVTList VTs = Indexed ?
9382 getVTList(VT, Ptr.getValueType(), MVT::Other) : getVTList(VT, MVT::Other);
9383 SDValue Ops[] = { Chain, Ptr, Offset };
9384 FoldingSetNodeID ID;
9385 AddNodeIDNode(ID, OpC: ISD::LOAD, VTList: VTs, OpList: Ops);
9386 ID.AddInteger(I: MemVT.getRawBits());
9387 ID.AddInteger(I: getSyntheticNodeSubclassData<LoadSDNode>(
9388 IROrder: dl.getIROrder(), Args&: VTs, Args&: AM, Args&: ExtType, Args&: MemVT, Args&: MMO));
9389 ID.AddInteger(I: MMO->getPointerInfo().getAddrSpace());
9390 ID.AddInteger(I: MMO->getFlags());
9391 void *IP = nullptr;
9392 if (SDNode *E = FindNodeOrInsertPos(ID, DL: dl, InsertPos&: IP)) {
9393 cast<LoadSDNode>(Val: E)->refineAlignment(NewMMO: MMO);
9394 return SDValue(E, 0);
9395 }
9396 auto *N = newSDNode<LoadSDNode>(Args: dl.getIROrder(), Args: dl.getDebugLoc(), Args&: VTs, Args&: AM,
9397 Args&: ExtType, Args&: MemVT, Args&: MMO);
9398 createOperands(Node: N, Vals: Ops);
9399
9400 CSEMap.InsertNode(N, InsertPos: IP);
9401 InsertNode(N);
9402 SDValue V(N, 0);
9403 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
9404 return V;
9405}
9406
9407SDValue SelectionDAG::getLoad(EVT VT, const SDLoc &dl, SDValue Chain,
9408 SDValue Ptr, MachinePointerInfo PtrInfo,
9409 MaybeAlign Alignment,
9410 MachineMemOperand::Flags MMOFlags,
9411 const AAMDNodes &AAInfo, const MDNode *Ranges) {
9412 SDValue Undef = getUNDEF(VT: Ptr.getValueType());
9413 return getLoad(AM: ISD::UNINDEXED, ExtType: ISD::NON_EXTLOAD, VT, dl, Chain, Ptr, Offset: Undef,
9414 PtrInfo, MemVT: VT, Alignment, MMOFlags, AAInfo, Ranges);
9415}
9416
9417SDValue SelectionDAG::getLoad(EVT VT, const SDLoc &dl, SDValue Chain,
9418 SDValue Ptr, MachineMemOperand *MMO) {
9419 SDValue Undef = getUNDEF(VT: Ptr.getValueType());
9420 return getLoad(AM: ISD::UNINDEXED, ExtType: ISD::NON_EXTLOAD, VT, dl, Chain, Ptr, Offset: Undef,
9421 MemVT: VT, MMO);
9422}
9423
9424SDValue SelectionDAG::getExtLoad(ISD::LoadExtType ExtType, const SDLoc &dl,
9425 EVT VT, SDValue Chain, SDValue Ptr,
9426 MachinePointerInfo PtrInfo, EVT MemVT,
9427 MaybeAlign Alignment,
9428 MachineMemOperand::Flags MMOFlags,
9429 const AAMDNodes &AAInfo) {
9430 SDValue Undef = getUNDEF(VT: Ptr.getValueType());
9431 return getLoad(AM: ISD::UNINDEXED, ExtType, VT, dl, Chain, Ptr, Offset: Undef, PtrInfo,
9432 MemVT, Alignment, MMOFlags, AAInfo);
9433}
9434
9435SDValue SelectionDAG::getExtLoad(ISD::LoadExtType ExtType, const SDLoc &dl,
9436 EVT VT, SDValue Chain, SDValue Ptr, EVT MemVT,
9437 MachineMemOperand *MMO) {
9438 SDValue Undef = getUNDEF(VT: Ptr.getValueType());
9439 return getLoad(AM: ISD::UNINDEXED, ExtType, VT, dl, Chain, Ptr, Offset: Undef,
9440 MemVT, MMO);
9441}
9442
9443SDValue SelectionDAG::getIndexedLoad(SDValue OrigLoad, const SDLoc &dl,
9444 SDValue Base, SDValue Offset,
9445 ISD::MemIndexedMode AM) {
9446 LoadSDNode *LD = cast<LoadSDNode>(Val&: OrigLoad);
9447 assert(LD->getOffset().isUndef() && "Load is already a indexed load!");
9448 // Don't propagate the invariant or dereferenceable flags.
9449 auto MMOFlags =
9450 LD->getMemOperand()->getFlags() &
9451 ~(MachineMemOperand::MOInvariant | MachineMemOperand::MODereferenceable);
9452 return getLoad(AM, ExtType: LD->getExtensionType(), VT: OrigLoad.getValueType(), dl,
9453 Chain: LD->getChain(), Ptr: Base, Offset, PtrInfo: LD->getPointerInfo(),
9454 MemVT: LD->getMemoryVT(), Alignment: LD->getAlign(), MMOFlags, AAInfo: LD->getAAInfo());
9455}
9456
9457SDValue SelectionDAG::getStore(SDValue Chain, const SDLoc &dl, SDValue Val,
9458 SDValue Ptr, MachinePointerInfo PtrInfo,
9459 Align Alignment,
9460 MachineMemOperand::Flags MMOFlags,
9461 const AAMDNodes &AAInfo) {
9462 assert(Chain.getValueType() == MVT::Other && "Invalid chain type");
9463
9464 MMOFlags |= MachineMemOperand::MOStore;
9465 assert((MMOFlags & MachineMemOperand::MOLoad) == 0);
9466
9467 if (PtrInfo.V.isNull())
9468 PtrInfo = InferPointerInfo(Info: PtrInfo, DAG&: *this, Ptr);
9469
9470 MachineFunction &MF = getMachineFunction();
9471 TypeSize Size = Val.getValueType().getStoreSize();
9472 MachineMemOperand *MMO =
9473 MF.getMachineMemOperand(PtrInfo, F: MMOFlags, Size, BaseAlignment: Alignment, AAInfo);
9474 return getStore(Chain, dl, Val, Ptr, MMO);
9475}
9476
9477SDValue SelectionDAG::getStore(SDValue Chain, const SDLoc &dl, SDValue Val,
9478 SDValue Ptr, MachineMemOperand *MMO) {
9479 SDValue Undef = getUNDEF(VT: Ptr.getValueType());
9480 return getStore(Chain, dl, Val, Ptr, Offset: Undef, SVT: Val.getValueType(), MMO,
9481 AM: ISD::UNINDEXED);
9482}
9483
9484SDValue SelectionDAG::getStore(SDValue Chain, const SDLoc &dl, SDValue Val,
9485 SDValue Ptr, SDValue Offset, EVT SVT,
9486 MachineMemOperand *MMO, ISD::MemIndexedMode AM,
9487 bool IsTruncating) {
9488 assert(Chain.getValueType() == MVT::Other && "Invalid chain type");
9489 EVT VT = Val.getValueType();
9490 if (VT == SVT) {
9491 IsTruncating = false;
9492 } else if (!IsTruncating) {
9493 assert(VT == SVT && "No-truncating store from different memory type!");
9494 } else {
9495 assert(SVT.getScalarType().bitsLT(VT.getScalarType()) &&
9496 "Should only be a truncating store, not extending!");
9497 assert(VT.isInteger() == SVT.isInteger() && "Can't do FP-INT conversion!");
9498 assert(VT.isVector() == SVT.isVector() &&
9499 "Cannot use trunc store to convert to or from a vector!");
9500 assert((!VT.isVector() ||
9501 VT.getVectorElementCount() == SVT.getVectorElementCount()) &&
9502 "Cannot use trunc store to change the number of vector elements!");
9503 }
9504
9505 bool Indexed = AM != ISD::UNINDEXED;
9506 assert((Indexed || Offset.isUndef()) && "Unindexed store with an offset!");
9507 SDVTList VTs = Indexed ? getVTList(Ptr.getValueType(), MVT::Other)
9508 : getVTList(MVT::Other);
9509 SDValue Ops[] = {Chain, Val, Ptr, Offset};
9510 FoldingSetNodeID ID;
9511 AddNodeIDNode(ID, OpC: ISD::STORE, VTList: VTs, OpList: Ops);
9512 ID.AddInteger(I: SVT.getRawBits());
9513 ID.AddInteger(I: getSyntheticNodeSubclassData<StoreSDNode>(
9514 IROrder: dl.getIROrder(), Args&: VTs, Args&: AM, Args&: IsTruncating, Args&: SVT, Args&: MMO));
9515 ID.AddInteger(I: MMO->getPointerInfo().getAddrSpace());
9516 ID.AddInteger(I: MMO->getFlags());
9517 void *IP = nullptr;
9518 if (SDNode *E = FindNodeOrInsertPos(ID, DL: dl, InsertPos&: IP)) {
9519 cast<StoreSDNode>(Val: E)->refineAlignment(NewMMO: MMO);
9520 return SDValue(E, 0);
9521 }
9522 auto *N = newSDNode<StoreSDNode>(Args: dl.getIROrder(), Args: dl.getDebugLoc(), Args&: VTs, Args&: AM,
9523 Args&: IsTruncating, Args&: SVT, Args&: MMO);
9524 createOperands(Node: N, Vals: Ops);
9525
9526 CSEMap.InsertNode(N, InsertPos: IP);
9527 InsertNode(N);
9528 SDValue V(N, 0);
9529 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
9530 return V;
9531}
9532
9533SDValue SelectionDAG::getTruncStore(SDValue Chain, const SDLoc &dl, SDValue Val,
9534 SDValue Ptr, MachinePointerInfo PtrInfo,
9535 EVT SVT, Align Alignment,
9536 MachineMemOperand::Flags MMOFlags,
9537 const AAMDNodes &AAInfo) {
9538 assert(Chain.getValueType() == MVT::Other &&
9539 "Invalid chain type");
9540
9541 MMOFlags |= MachineMemOperand::MOStore;
9542 assert((MMOFlags & MachineMemOperand::MOLoad) == 0);
9543
9544 if (PtrInfo.V.isNull())
9545 PtrInfo = InferPointerInfo(Info: PtrInfo, DAG&: *this, Ptr);
9546
9547 MachineFunction &MF = getMachineFunction();
9548 MachineMemOperand *MMO = MF.getMachineMemOperand(
9549 PtrInfo, F: MMOFlags, Size: SVT.getStoreSize(), BaseAlignment: Alignment, AAInfo);
9550 return getTruncStore(Chain, dl, Val, Ptr, SVT, MMO);
9551}
9552
9553SDValue SelectionDAG::getTruncStore(SDValue Chain, const SDLoc &dl, SDValue Val,
9554 SDValue Ptr, EVT SVT,
9555 MachineMemOperand *MMO) {
9556 SDValue Undef = getUNDEF(VT: Ptr.getValueType());
9557 return getStore(Chain, dl, Val, Ptr, Offset: Undef, SVT, MMO, AM: ISD::UNINDEXED, IsTruncating: true);
9558}
9559
9560SDValue SelectionDAG::getIndexedStore(SDValue OrigStore, const SDLoc &dl,
9561 SDValue Base, SDValue Offset,
9562 ISD::MemIndexedMode AM) {
9563 StoreSDNode *ST = cast<StoreSDNode>(Val&: OrigStore);
9564 assert(ST->getOffset().isUndef() && "Store is already a indexed store!");
9565 return getStore(Chain: ST->getChain(), dl, Val: ST->getValue(), Ptr: Base, Offset,
9566 SVT: ST->getMemoryVT(), MMO: ST->getMemOperand(), AM,
9567 IsTruncating: ST->isTruncatingStore());
9568}
9569
9570SDValue SelectionDAG::getLoadVP(
9571 ISD::MemIndexedMode AM, ISD::LoadExtType ExtType, EVT VT, const SDLoc &dl,
9572 SDValue Chain, SDValue Ptr, SDValue Offset, SDValue Mask, SDValue EVL,
9573 MachinePointerInfo PtrInfo, EVT MemVT, Align Alignment,
9574 MachineMemOperand::Flags MMOFlags, const AAMDNodes &AAInfo,
9575 const MDNode *Ranges, bool IsExpanding) {
9576 assert(Chain.getValueType() == MVT::Other && "Invalid chain type");
9577
9578 MMOFlags |= MachineMemOperand::MOLoad;
9579 assert((MMOFlags & MachineMemOperand::MOStore) == 0);
9580 // If we don't have a PtrInfo, infer the trivial frame index case to simplify
9581 // clients.
9582 if (PtrInfo.V.isNull())
9583 PtrInfo = InferPointerInfo(Info: PtrInfo, DAG&: *this, Ptr, OffsetOp: Offset);
9584
9585 TypeSize Size = MemVT.getStoreSize();
9586 MachineFunction &MF = getMachineFunction();
9587 MachineMemOperand *MMO = MF.getMachineMemOperand(PtrInfo, F: MMOFlags, Size,
9588 BaseAlignment: Alignment, AAInfo, Ranges);
9589 return getLoadVP(AM, ExtType, VT, dl, Chain, Ptr, Offset, Mask, EVL, MemVT,
9590 MMO, IsExpanding);
9591}
9592
9593SDValue SelectionDAG::getLoadVP(ISD::MemIndexedMode AM,
9594 ISD::LoadExtType ExtType, EVT VT,
9595 const SDLoc &dl, SDValue Chain, SDValue Ptr,
9596 SDValue Offset, SDValue Mask, SDValue EVL,
9597 EVT MemVT, MachineMemOperand *MMO,
9598 bool IsExpanding) {
9599 bool Indexed = AM != ISD::UNINDEXED;
9600 assert((Indexed || Offset.isUndef()) && "Unindexed load with an offset!");
9601
9602 SDVTList VTs = Indexed ? getVTList(VT, Ptr.getValueType(), MVT::Other)
9603 : getVTList(VT, MVT::Other);
9604 SDValue Ops[] = {Chain, Ptr, Offset, Mask, EVL};
9605 FoldingSetNodeID ID;
9606 AddNodeIDNode(ID, OpC: ISD::VP_LOAD, VTList: VTs, OpList: Ops);
9607 ID.AddInteger(I: MemVT.getRawBits());
9608 ID.AddInteger(I: getSyntheticNodeSubclassData<VPLoadSDNode>(
9609 IROrder: dl.getIROrder(), Args&: VTs, Args&: AM, Args&: ExtType, Args&: IsExpanding, Args&: MemVT, Args&: MMO));
9610 ID.AddInteger(I: MMO->getPointerInfo().getAddrSpace());
9611 ID.AddInteger(I: MMO->getFlags());
9612 void *IP = nullptr;
9613 if (SDNode *E = FindNodeOrInsertPos(ID, DL: dl, InsertPos&: IP)) {
9614 cast<VPLoadSDNode>(Val: E)->refineAlignment(NewMMO: MMO);
9615 return SDValue(E, 0);
9616 }
9617 auto *N = newSDNode<VPLoadSDNode>(Args: dl.getIROrder(), Args: dl.getDebugLoc(), Args&: VTs, Args&: AM,
9618 Args&: ExtType, Args&: IsExpanding, Args&: MemVT, Args&: MMO);
9619 createOperands(Node: N, Vals: Ops);
9620
9621 CSEMap.InsertNode(N, InsertPos: IP);
9622 InsertNode(N);
9623 SDValue V(N, 0);
9624 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
9625 return V;
9626}
9627
9628SDValue SelectionDAG::getLoadVP(EVT VT, const SDLoc &dl, SDValue Chain,
9629 SDValue Ptr, SDValue Mask, SDValue EVL,
9630 MachinePointerInfo PtrInfo,
9631 MaybeAlign Alignment,
9632 MachineMemOperand::Flags MMOFlags,
9633 const AAMDNodes &AAInfo, const MDNode *Ranges,
9634 bool IsExpanding) {
9635 SDValue Undef = getUNDEF(VT: Ptr.getValueType());
9636 return getLoadVP(AM: ISD::UNINDEXED, ExtType: ISD::NON_EXTLOAD, VT, dl, Chain, Ptr, Offset: Undef,
9637 Mask, EVL, PtrInfo, MemVT: VT, Alignment, MMOFlags, AAInfo, Ranges,
9638 IsExpanding);
9639}
9640
9641SDValue SelectionDAG::getLoadVP(EVT VT, const SDLoc &dl, SDValue Chain,
9642 SDValue Ptr, SDValue Mask, SDValue EVL,
9643 MachineMemOperand *MMO, bool IsExpanding) {
9644 SDValue Undef = getUNDEF(VT: Ptr.getValueType());
9645 return getLoadVP(AM: ISD::UNINDEXED, ExtType: ISD::NON_EXTLOAD, VT, dl, Chain, Ptr, Offset: Undef,
9646 Mask, EVL, MemVT: VT, MMO, IsExpanding);
9647}
9648
9649SDValue SelectionDAG::getExtLoadVP(ISD::LoadExtType ExtType, const SDLoc &dl,
9650 EVT VT, SDValue Chain, SDValue Ptr,
9651 SDValue Mask, SDValue EVL,
9652 MachinePointerInfo PtrInfo, EVT MemVT,
9653 MaybeAlign Alignment,
9654 MachineMemOperand::Flags MMOFlags,
9655 const AAMDNodes &AAInfo, bool IsExpanding) {
9656 SDValue Undef = getUNDEF(VT: Ptr.getValueType());
9657 return getLoadVP(AM: ISD::UNINDEXED, ExtType, VT, dl, Chain, Ptr, Offset: Undef, Mask,
9658 EVL, PtrInfo, MemVT, Alignment, MMOFlags, AAInfo, Ranges: nullptr,
9659 IsExpanding);
9660}
9661
9662SDValue SelectionDAG::getExtLoadVP(ISD::LoadExtType ExtType, const SDLoc &dl,
9663 EVT VT, SDValue Chain, SDValue Ptr,
9664 SDValue Mask, SDValue EVL, EVT MemVT,
9665 MachineMemOperand *MMO, bool IsExpanding) {
9666 SDValue Undef = getUNDEF(VT: Ptr.getValueType());
9667 return getLoadVP(AM: ISD::UNINDEXED, ExtType, VT, dl, Chain, Ptr, Offset: Undef, Mask,
9668 EVL, MemVT, MMO, IsExpanding);
9669}
9670
9671SDValue SelectionDAG::getIndexedLoadVP(SDValue OrigLoad, const SDLoc &dl,
9672 SDValue Base, SDValue Offset,
9673 ISD::MemIndexedMode AM) {
9674 auto *LD = cast<VPLoadSDNode>(Val&: OrigLoad);
9675 assert(LD->getOffset().isUndef() && "Load is already a indexed load!");
9676 // Don't propagate the invariant or dereferenceable flags.
9677 auto MMOFlags =
9678 LD->getMemOperand()->getFlags() &
9679 ~(MachineMemOperand::MOInvariant | MachineMemOperand::MODereferenceable);
9680 return getLoadVP(AM, ExtType: LD->getExtensionType(), VT: OrigLoad.getValueType(), dl,
9681 Chain: LD->getChain(), Ptr: Base, Offset, Mask: LD->getMask(),
9682 EVL: LD->getVectorLength(), PtrInfo: LD->getPointerInfo(),
9683 MemVT: LD->getMemoryVT(), Alignment: LD->getAlign(), MMOFlags, AAInfo: LD->getAAInfo(),
9684 Ranges: nullptr, IsExpanding: LD->isExpandingLoad());
9685}
9686
9687SDValue SelectionDAG::getStoreVP(SDValue Chain, const SDLoc &dl, SDValue Val,
9688 SDValue Ptr, SDValue Offset, SDValue Mask,
9689 SDValue EVL, EVT MemVT, MachineMemOperand *MMO,
9690 ISD::MemIndexedMode AM, bool IsTruncating,
9691 bool IsCompressing) {
9692 assert(Chain.getValueType() == MVT::Other && "Invalid chain type");
9693 bool Indexed = AM != ISD::UNINDEXED;
9694 assert((Indexed || Offset.isUndef()) && "Unindexed vp_store with an offset!");
9695 SDVTList VTs = Indexed ? getVTList(Ptr.getValueType(), MVT::Other)
9696 : getVTList(MVT::Other);
9697 SDValue Ops[] = {Chain, Val, Ptr, Offset, Mask, EVL};
9698 FoldingSetNodeID ID;
9699 AddNodeIDNode(ID, OpC: ISD::VP_STORE, VTList: VTs, OpList: Ops);
9700 ID.AddInteger(I: MemVT.getRawBits());
9701 ID.AddInteger(I: getSyntheticNodeSubclassData<VPStoreSDNode>(
9702 IROrder: dl.getIROrder(), Args&: VTs, Args&: AM, Args&: IsTruncating, Args&: IsCompressing, Args&: MemVT, Args&: MMO));
9703 ID.AddInteger(I: MMO->getPointerInfo().getAddrSpace());
9704 ID.AddInteger(I: MMO->getFlags());
9705 void *IP = nullptr;
9706 if (SDNode *E = FindNodeOrInsertPos(ID, DL: dl, InsertPos&: IP)) {
9707 cast<VPStoreSDNode>(Val: E)->refineAlignment(NewMMO: MMO);
9708 return SDValue(E, 0);
9709 }
9710 auto *N = newSDNode<VPStoreSDNode>(Args: dl.getIROrder(), Args: dl.getDebugLoc(), Args&: VTs, Args&: AM,
9711 Args&: IsTruncating, Args&: IsCompressing, Args&: MemVT, Args&: MMO);
9712 createOperands(Node: N, Vals: Ops);
9713
9714 CSEMap.InsertNode(N, InsertPos: IP);
9715 InsertNode(N);
9716 SDValue V(N, 0);
9717 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
9718 return V;
9719}
9720
9721SDValue SelectionDAG::getTruncStoreVP(SDValue Chain, const SDLoc &dl,
9722 SDValue Val, SDValue Ptr, SDValue Mask,
9723 SDValue EVL, MachinePointerInfo PtrInfo,
9724 EVT SVT, Align Alignment,
9725 MachineMemOperand::Flags MMOFlags,
9726 const AAMDNodes &AAInfo,
9727 bool IsCompressing) {
9728 assert(Chain.getValueType() == MVT::Other && "Invalid chain type");
9729
9730 MMOFlags |= MachineMemOperand::MOStore;
9731 assert((MMOFlags & MachineMemOperand::MOLoad) == 0);
9732
9733 if (PtrInfo.V.isNull())
9734 PtrInfo = InferPointerInfo(Info: PtrInfo, DAG&: *this, Ptr);
9735
9736 MachineFunction &MF = getMachineFunction();
9737 MachineMemOperand *MMO = MF.getMachineMemOperand(
9738 PtrInfo, F: MMOFlags, Size: SVT.getStoreSize(), BaseAlignment: Alignment, AAInfo);
9739 return getTruncStoreVP(Chain, dl, Val, Ptr, Mask, EVL, SVT, MMO,
9740 IsCompressing);
9741}
9742
9743SDValue SelectionDAG::getTruncStoreVP(SDValue Chain, const SDLoc &dl,
9744 SDValue Val, SDValue Ptr, SDValue Mask,
9745 SDValue EVL, EVT SVT,
9746 MachineMemOperand *MMO,
9747 bool IsCompressing) {
9748 EVT VT = Val.getValueType();
9749
9750 assert(Chain.getValueType() == MVT::Other && "Invalid chain type");
9751 if (VT == SVT)
9752 return getStoreVP(Chain, dl, Val, Ptr, Offset: getUNDEF(VT: Ptr.getValueType()), Mask,
9753 EVL, MemVT: VT, MMO, AM: ISD::UNINDEXED,
9754 /*IsTruncating*/ false, IsCompressing);
9755
9756 assert(SVT.getScalarType().bitsLT(VT.getScalarType()) &&
9757 "Should only be a truncating store, not extending!");
9758 assert(VT.isInteger() == SVT.isInteger() && "Can't do FP-INT conversion!");
9759 assert(VT.isVector() == SVT.isVector() &&
9760 "Cannot use trunc store to convert to or from a vector!");
9761 assert((!VT.isVector() ||
9762 VT.getVectorElementCount() == SVT.getVectorElementCount()) &&
9763 "Cannot use trunc store to change the number of vector elements!");
9764
9765 SDVTList VTs = getVTList(MVT::Other);
9766 SDValue Undef = getUNDEF(VT: Ptr.getValueType());
9767 SDValue Ops[] = {Chain, Val, Ptr, Undef, Mask, EVL};
9768 FoldingSetNodeID ID;
9769 AddNodeIDNode(ID, OpC: ISD::VP_STORE, VTList: VTs, OpList: Ops);
9770 ID.AddInteger(I: SVT.getRawBits());
9771 ID.AddInteger(I: getSyntheticNodeSubclassData<VPStoreSDNode>(
9772 IROrder: dl.getIROrder(), Args&: VTs, Args: ISD::UNINDEXED, Args: true, Args&: IsCompressing, Args&: SVT, Args&: MMO));
9773 ID.AddInteger(I: MMO->getPointerInfo().getAddrSpace());
9774 ID.AddInteger(I: MMO->getFlags());
9775 void *IP = nullptr;
9776 if (SDNode *E = FindNodeOrInsertPos(ID, DL: dl, InsertPos&: IP)) {
9777 cast<VPStoreSDNode>(Val: E)->refineAlignment(NewMMO: MMO);
9778 return SDValue(E, 0);
9779 }
9780 auto *N =
9781 newSDNode<VPStoreSDNode>(Args: dl.getIROrder(), Args: dl.getDebugLoc(), Args&: VTs,
9782 Args: ISD::UNINDEXED, Args: true, Args&: IsCompressing, Args&: SVT, Args&: MMO);
9783 createOperands(Node: N, Vals: Ops);
9784
9785 CSEMap.InsertNode(N, InsertPos: IP);
9786 InsertNode(N);
9787 SDValue V(N, 0);
9788 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
9789 return V;
9790}
9791
9792SDValue SelectionDAG::getIndexedStoreVP(SDValue OrigStore, const SDLoc &dl,
9793 SDValue Base, SDValue Offset,
9794 ISD::MemIndexedMode AM) {
9795 auto *ST = cast<VPStoreSDNode>(Val&: OrigStore);
9796 assert(ST->getOffset().isUndef() && "Store is already an indexed store!");
9797 SDVTList VTs = getVTList(Base.getValueType(), MVT::Other);
9798 SDValue Ops[] = {ST->getChain(), ST->getValue(), Base,
9799 Offset, ST->getMask(), ST->getVectorLength()};
9800 FoldingSetNodeID ID;
9801 AddNodeIDNode(ID, OpC: ISD::VP_STORE, VTList: VTs, OpList: Ops);
9802 ID.AddInteger(I: ST->getMemoryVT().getRawBits());
9803 ID.AddInteger(I: ST->getRawSubclassData());
9804 ID.AddInteger(I: ST->getPointerInfo().getAddrSpace());
9805 ID.AddInteger(I: ST->getMemOperand()->getFlags());
9806 void *IP = nullptr;
9807 if (SDNode *E = FindNodeOrInsertPos(ID, DL: dl, InsertPos&: IP))
9808 return SDValue(E, 0);
9809
9810 auto *N = newSDNode<VPStoreSDNode>(
9811 Args: dl.getIROrder(), Args: dl.getDebugLoc(), Args&: VTs, Args&: AM, Args: ST->isTruncatingStore(),
9812 Args: ST->isCompressingStore(), Args: ST->getMemoryVT(), Args: ST->getMemOperand());
9813 createOperands(Node: N, Vals: Ops);
9814
9815 CSEMap.InsertNode(N, InsertPos: IP);
9816 InsertNode(N);
9817 SDValue V(N, 0);
9818 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
9819 return V;
9820}
9821
9822SDValue SelectionDAG::getStridedLoadVP(
9823 ISD::MemIndexedMode AM, ISD::LoadExtType ExtType, EVT VT, const SDLoc &DL,
9824 SDValue Chain, SDValue Ptr, SDValue Offset, SDValue Stride, SDValue Mask,
9825 SDValue EVL, EVT MemVT, MachineMemOperand *MMO, bool IsExpanding) {
9826 bool Indexed = AM != ISD::UNINDEXED;
9827 assert((Indexed || Offset.isUndef()) && "Unindexed load with an offset!");
9828
9829 SDValue Ops[] = {Chain, Ptr, Offset, Stride, Mask, EVL};
9830 SDVTList VTs = Indexed ? getVTList(VT, Ptr.getValueType(), MVT::Other)
9831 : getVTList(VT, MVT::Other);
9832 FoldingSetNodeID ID;
9833 AddNodeIDNode(ID, OpC: ISD::EXPERIMENTAL_VP_STRIDED_LOAD, VTList: VTs, OpList: Ops);
9834 ID.AddInteger(I: VT.getRawBits());
9835 ID.AddInteger(I: getSyntheticNodeSubclassData<VPStridedLoadSDNode>(
9836 IROrder: DL.getIROrder(), Args&: VTs, Args&: AM, Args&: ExtType, Args&: IsExpanding, Args&: MemVT, Args&: MMO));
9837 ID.AddInteger(I: MMO->getPointerInfo().getAddrSpace());
9838
9839 void *IP = nullptr;
9840 if (SDNode *E = FindNodeOrInsertPos(ID, DL, InsertPos&: IP)) {
9841 cast<VPStridedLoadSDNode>(Val: E)->refineAlignment(NewMMO: MMO);
9842 return SDValue(E, 0);
9843 }
9844
9845 auto *N =
9846 newSDNode<VPStridedLoadSDNode>(Args: DL.getIROrder(), Args: DL.getDebugLoc(), Args&: VTs, Args&: AM,
9847 Args&: ExtType, Args&: IsExpanding, Args&: MemVT, Args&: MMO);
9848 createOperands(Node: N, Vals: Ops);
9849 CSEMap.InsertNode(N, InsertPos: IP);
9850 InsertNode(N);
9851 SDValue V(N, 0);
9852 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
9853 return V;
9854}
9855
9856SDValue SelectionDAG::getStridedLoadVP(EVT VT, const SDLoc &DL, SDValue Chain,
9857 SDValue Ptr, SDValue Stride,
9858 SDValue Mask, SDValue EVL,
9859 MachineMemOperand *MMO,
9860 bool IsExpanding) {
9861 SDValue Undef = getUNDEF(VT: Ptr.getValueType());
9862 return getStridedLoadVP(AM: ISD::UNINDEXED, ExtType: ISD::NON_EXTLOAD, VT, DL, Chain, Ptr,
9863 Offset: Undef, Stride, Mask, EVL, MemVT: VT, MMO, IsExpanding);
9864}
9865
9866SDValue SelectionDAG::getExtStridedLoadVP(
9867 ISD::LoadExtType ExtType, const SDLoc &DL, EVT VT, SDValue Chain,
9868 SDValue Ptr, SDValue Stride, SDValue Mask, SDValue EVL, EVT MemVT,
9869 MachineMemOperand *MMO, bool IsExpanding) {
9870 SDValue Undef = getUNDEF(VT: Ptr.getValueType());
9871 return getStridedLoadVP(AM: ISD::UNINDEXED, ExtType, VT, DL, Chain, Ptr, Offset: Undef,
9872 Stride, Mask, EVL, MemVT, MMO, IsExpanding);
9873}
9874
9875SDValue SelectionDAG::getStridedStoreVP(SDValue Chain, const SDLoc &DL,
9876 SDValue Val, SDValue Ptr,
9877 SDValue Offset, SDValue Stride,
9878 SDValue Mask, SDValue EVL, EVT MemVT,
9879 MachineMemOperand *MMO,
9880 ISD::MemIndexedMode AM,
9881 bool IsTruncating, bool IsCompressing) {
9882 assert(Chain.getValueType() == MVT::Other && "Invalid chain type");
9883 bool Indexed = AM != ISD::UNINDEXED;
9884 assert((Indexed || Offset.isUndef()) && "Unindexed vp_store with an offset!");
9885 SDVTList VTs = Indexed ? getVTList(Ptr.getValueType(), MVT::Other)
9886 : getVTList(MVT::Other);
9887 SDValue Ops[] = {Chain, Val, Ptr, Offset, Stride, Mask, EVL};
9888 FoldingSetNodeID ID;
9889 AddNodeIDNode(ID, OpC: ISD::EXPERIMENTAL_VP_STRIDED_STORE, VTList: VTs, OpList: Ops);
9890 ID.AddInteger(I: MemVT.getRawBits());
9891 ID.AddInteger(I: getSyntheticNodeSubclassData<VPStridedStoreSDNode>(
9892 IROrder: DL.getIROrder(), Args&: VTs, Args&: AM, Args&: IsTruncating, Args&: IsCompressing, Args&: MemVT, Args&: MMO));
9893 ID.AddInteger(I: MMO->getPointerInfo().getAddrSpace());
9894 void *IP = nullptr;
9895 if (SDNode *E = FindNodeOrInsertPos(ID, DL, InsertPos&: IP)) {
9896 cast<VPStridedStoreSDNode>(Val: E)->refineAlignment(NewMMO: MMO);
9897 return SDValue(E, 0);
9898 }
9899 auto *N = newSDNode<VPStridedStoreSDNode>(Args: DL.getIROrder(), Args: DL.getDebugLoc(),
9900 Args&: VTs, Args&: AM, Args&: IsTruncating,
9901 Args&: IsCompressing, Args&: MemVT, Args&: MMO);
9902 createOperands(Node: N, Vals: Ops);
9903
9904 CSEMap.InsertNode(N, InsertPos: IP);
9905 InsertNode(N);
9906 SDValue V(N, 0);
9907 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
9908 return V;
9909}
9910
9911SDValue SelectionDAG::getTruncStridedStoreVP(SDValue Chain, const SDLoc &DL,
9912 SDValue Val, SDValue Ptr,
9913 SDValue Stride, SDValue Mask,
9914 SDValue EVL, EVT SVT,
9915 MachineMemOperand *MMO,
9916 bool IsCompressing) {
9917 EVT VT = Val.getValueType();
9918
9919 assert(Chain.getValueType() == MVT::Other && "Invalid chain type");
9920 if (VT == SVT)
9921 return getStridedStoreVP(Chain, DL, Val, Ptr, Offset: getUNDEF(VT: Ptr.getValueType()),
9922 Stride, Mask, EVL, MemVT: VT, MMO, AM: ISD::UNINDEXED,
9923 /*IsTruncating*/ false, IsCompressing);
9924
9925 assert(SVT.getScalarType().bitsLT(VT.getScalarType()) &&
9926 "Should only be a truncating store, not extending!");
9927 assert(VT.isInteger() == SVT.isInteger() && "Can't do FP-INT conversion!");
9928 assert(VT.isVector() == SVT.isVector() &&
9929 "Cannot use trunc store to convert to or from a vector!");
9930 assert((!VT.isVector() ||
9931 VT.getVectorElementCount() == SVT.getVectorElementCount()) &&
9932 "Cannot use trunc store to change the number of vector elements!");
9933
9934 SDVTList VTs = getVTList(MVT::Other);
9935 SDValue Undef = getUNDEF(VT: Ptr.getValueType());
9936 SDValue Ops[] = {Chain, Val, Ptr, Undef, Stride, Mask, EVL};
9937 FoldingSetNodeID ID;
9938 AddNodeIDNode(ID, OpC: ISD::EXPERIMENTAL_VP_STRIDED_STORE, VTList: VTs, OpList: Ops);
9939 ID.AddInteger(I: SVT.getRawBits());
9940 ID.AddInteger(I: getSyntheticNodeSubclassData<VPStridedStoreSDNode>(
9941 IROrder: DL.getIROrder(), Args&: VTs, Args: ISD::UNINDEXED, Args: true, Args&: IsCompressing, Args&: SVT, Args&: MMO));
9942 ID.AddInteger(I: MMO->getPointerInfo().getAddrSpace());
9943 void *IP = nullptr;
9944 if (SDNode *E = FindNodeOrInsertPos(ID, DL, InsertPos&: IP)) {
9945 cast<VPStridedStoreSDNode>(Val: E)->refineAlignment(NewMMO: MMO);
9946 return SDValue(E, 0);
9947 }
9948 auto *N = newSDNode<VPStridedStoreSDNode>(Args: DL.getIROrder(), Args: DL.getDebugLoc(),
9949 Args&: VTs, Args: ISD::UNINDEXED, Args: true,
9950 Args&: IsCompressing, Args&: SVT, Args&: MMO);
9951 createOperands(Node: N, Vals: Ops);
9952
9953 CSEMap.InsertNode(N, InsertPos: IP);
9954 InsertNode(N);
9955 SDValue V(N, 0);
9956 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
9957 return V;
9958}
9959
9960SDValue SelectionDAG::getGatherVP(SDVTList VTs, EVT VT, const SDLoc &dl,
9961 ArrayRef<SDValue> Ops, MachineMemOperand *MMO,
9962 ISD::MemIndexType IndexType) {
9963 assert(Ops.size() == 6 && "Incompatible number of operands");
9964
9965 FoldingSetNodeID ID;
9966 AddNodeIDNode(ID, OpC: ISD::VP_GATHER, VTList: VTs, OpList: Ops);
9967 ID.AddInteger(I: VT.getRawBits());
9968 ID.AddInteger(I: getSyntheticNodeSubclassData<VPGatherSDNode>(
9969 IROrder: dl.getIROrder(), Args&: VTs, Args&: VT, Args&: MMO, Args&: IndexType));
9970 ID.AddInteger(I: MMO->getPointerInfo().getAddrSpace());
9971 ID.AddInteger(I: MMO->getFlags());
9972 void *IP = nullptr;
9973 if (SDNode *E = FindNodeOrInsertPos(ID, DL: dl, InsertPos&: IP)) {
9974 cast<VPGatherSDNode>(Val: E)->refineAlignment(NewMMO: MMO);
9975 return SDValue(E, 0);
9976 }
9977
9978 auto *N = newSDNode<VPGatherSDNode>(Args: dl.getIROrder(), Args: dl.getDebugLoc(), Args&: VTs,
9979 Args&: VT, Args&: MMO, Args&: IndexType);
9980 createOperands(Node: N, Vals: Ops);
9981
9982 assert(N->getMask().getValueType().getVectorElementCount() ==
9983 N->getValueType(0).getVectorElementCount() &&
9984 "Vector width mismatch between mask and data");
9985 assert(N->getIndex().getValueType().getVectorElementCount().isScalable() ==
9986 N->getValueType(0).getVectorElementCount().isScalable() &&
9987 "Scalable flags of index and data do not match");
9988 assert(ElementCount::isKnownGE(
9989 N->getIndex().getValueType().getVectorElementCount(),
9990 N->getValueType(0).getVectorElementCount()) &&
9991 "Vector width mismatch between index and data");
9992 assert(isa<ConstantSDNode>(N->getScale()) &&
9993 N->getScale()->getAsAPIntVal().isPowerOf2() &&
9994 "Scale should be a constant power of 2");
9995
9996 CSEMap.InsertNode(N, InsertPos: IP);
9997 InsertNode(N);
9998 SDValue V(N, 0);
9999 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
10000 return V;
10001}
10002
10003SDValue SelectionDAG::getScatterVP(SDVTList VTs, EVT VT, const SDLoc &dl,
10004 ArrayRef<SDValue> Ops,
10005 MachineMemOperand *MMO,
10006 ISD::MemIndexType IndexType) {
10007 assert(Ops.size() == 7 && "Incompatible number of operands");
10008
10009 FoldingSetNodeID ID;
10010 AddNodeIDNode(ID, OpC: ISD::VP_SCATTER, VTList: VTs, OpList: Ops);
10011 ID.AddInteger(I: VT.getRawBits());
10012 ID.AddInteger(I: getSyntheticNodeSubclassData<VPScatterSDNode>(
10013 IROrder: dl.getIROrder(), Args&: VTs, Args&: VT, Args&: MMO, Args&: IndexType));
10014 ID.AddInteger(I: MMO->getPointerInfo().getAddrSpace());
10015 ID.AddInteger(I: MMO->getFlags());
10016 void *IP = nullptr;
10017 if (SDNode *E = FindNodeOrInsertPos(ID, DL: dl, InsertPos&: IP)) {
10018 cast<VPScatterSDNode>(Val: E)->refineAlignment(NewMMO: MMO);
10019 return SDValue(E, 0);
10020 }
10021 auto *N = newSDNode<VPScatterSDNode>(Args: dl.getIROrder(), Args: dl.getDebugLoc(), Args&: VTs,
10022 Args&: VT, Args&: MMO, Args&: IndexType);
10023 createOperands(Node: N, Vals: Ops);
10024
10025 assert(N->getMask().getValueType().getVectorElementCount() ==
10026 N->getValue().getValueType().getVectorElementCount() &&
10027 "Vector width mismatch between mask and data");
10028 assert(
10029 N->getIndex().getValueType().getVectorElementCount().isScalable() ==
10030 N->getValue().getValueType().getVectorElementCount().isScalable() &&
10031 "Scalable flags of index and data do not match");
10032 assert(ElementCount::isKnownGE(
10033 N->getIndex().getValueType().getVectorElementCount(),
10034 N->getValue().getValueType().getVectorElementCount()) &&
10035 "Vector width mismatch between index and data");
10036 assert(isa<ConstantSDNode>(N->getScale()) &&
10037 N->getScale()->getAsAPIntVal().isPowerOf2() &&
10038 "Scale should be a constant power of 2");
10039
10040 CSEMap.InsertNode(N, InsertPos: IP);
10041 InsertNode(N);
10042 SDValue V(N, 0);
10043 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
10044 return V;
10045}
10046
10047SDValue SelectionDAG::getMaskedLoad(EVT VT, const SDLoc &dl, SDValue Chain,
10048 SDValue Base, SDValue Offset, SDValue Mask,
10049 SDValue PassThru, EVT MemVT,
10050 MachineMemOperand *MMO,
10051 ISD::MemIndexedMode AM,
10052 ISD::LoadExtType ExtTy, bool isExpanding) {
10053 bool Indexed = AM != ISD::UNINDEXED;
10054 assert((Indexed || Offset.isUndef()) &&
10055 "Unindexed masked load with an offset!");
10056 SDVTList VTs = Indexed ? getVTList(VT, Base.getValueType(), MVT::Other)
10057 : getVTList(VT, MVT::Other);
10058 SDValue Ops[] = {Chain, Base, Offset, Mask, PassThru};
10059 FoldingSetNodeID ID;
10060 AddNodeIDNode(ID, OpC: ISD::MLOAD, VTList: VTs, OpList: Ops);
10061 ID.AddInteger(I: MemVT.getRawBits());
10062 ID.AddInteger(I: getSyntheticNodeSubclassData<MaskedLoadSDNode>(
10063 IROrder: dl.getIROrder(), Args&: VTs, Args&: AM, Args&: ExtTy, Args&: isExpanding, Args&: MemVT, Args&: MMO));
10064 ID.AddInteger(I: MMO->getPointerInfo().getAddrSpace());
10065 ID.AddInteger(I: MMO->getFlags());
10066 void *IP = nullptr;
10067 if (SDNode *E = FindNodeOrInsertPos(ID, DL: dl, InsertPos&: IP)) {
10068 cast<MaskedLoadSDNode>(Val: E)->refineAlignment(NewMMO: MMO);
10069 return SDValue(E, 0);
10070 }
10071 auto *N = newSDNode<MaskedLoadSDNode>(Args: dl.getIROrder(), Args: dl.getDebugLoc(), Args&: VTs,
10072 Args&: AM, Args&: ExtTy, Args&: isExpanding, Args&: MemVT, Args&: MMO);
10073 createOperands(Node: N, Vals: Ops);
10074
10075 CSEMap.InsertNode(N, InsertPos: IP);
10076 InsertNode(N);
10077 SDValue V(N, 0);
10078 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
10079 return V;
10080}
10081
10082SDValue SelectionDAG::getIndexedMaskedLoad(SDValue OrigLoad, const SDLoc &dl,
10083 SDValue Base, SDValue Offset,
10084 ISD::MemIndexedMode AM) {
10085 MaskedLoadSDNode *LD = cast<MaskedLoadSDNode>(Val&: OrigLoad);
10086 assert(LD->getOffset().isUndef() && "Masked load is already a indexed load!");
10087 return getMaskedLoad(VT: OrigLoad.getValueType(), dl, Chain: LD->getChain(), Base,
10088 Offset, Mask: LD->getMask(), PassThru: LD->getPassThru(),
10089 MemVT: LD->getMemoryVT(), MMO: LD->getMemOperand(), AM,
10090 ExtTy: LD->getExtensionType(), isExpanding: LD->isExpandingLoad());
10091}
10092
10093SDValue SelectionDAG::getMaskedStore(SDValue Chain, const SDLoc &dl,
10094 SDValue Val, SDValue Base, SDValue Offset,
10095 SDValue Mask, EVT MemVT,
10096 MachineMemOperand *MMO,
10097 ISD::MemIndexedMode AM, bool IsTruncating,
10098 bool IsCompressing) {
10099 assert(Chain.getValueType() == MVT::Other &&
10100 "Invalid chain type");
10101 bool Indexed = AM != ISD::UNINDEXED;
10102 assert((Indexed || Offset.isUndef()) &&
10103 "Unindexed masked store with an offset!");
10104 SDVTList VTs = Indexed ? getVTList(Base.getValueType(), MVT::Other)
10105 : getVTList(MVT::Other);
10106 SDValue Ops[] = {Chain, Val, Base, Offset, Mask};
10107 FoldingSetNodeID ID;
10108 AddNodeIDNode(ID, OpC: ISD::MSTORE, VTList: VTs, OpList: Ops);
10109 ID.AddInteger(I: MemVT.getRawBits());
10110 ID.AddInteger(I: getSyntheticNodeSubclassData<MaskedStoreSDNode>(
10111 IROrder: dl.getIROrder(), Args&: VTs, Args&: AM, Args&: IsTruncating, Args&: IsCompressing, Args&: MemVT, Args&: MMO));
10112 ID.AddInteger(I: MMO->getPointerInfo().getAddrSpace());
10113 ID.AddInteger(I: MMO->getFlags());
10114 void *IP = nullptr;
10115 if (SDNode *E = FindNodeOrInsertPos(ID, DL: dl, InsertPos&: IP)) {
10116 cast<MaskedStoreSDNode>(Val: E)->refineAlignment(NewMMO: MMO);
10117 return SDValue(E, 0);
10118 }
10119 auto *N =
10120 newSDNode<MaskedStoreSDNode>(Args: dl.getIROrder(), Args: dl.getDebugLoc(), Args&: VTs, Args&: AM,
10121 Args&: IsTruncating, Args&: IsCompressing, Args&: MemVT, Args&: MMO);
10122 createOperands(Node: N, Vals: Ops);
10123
10124 CSEMap.InsertNode(N, InsertPos: IP);
10125 InsertNode(N);
10126 SDValue V(N, 0);
10127 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
10128 return V;
10129}
10130
10131SDValue SelectionDAG::getIndexedMaskedStore(SDValue OrigStore, const SDLoc &dl,
10132 SDValue Base, SDValue Offset,
10133 ISD::MemIndexedMode AM) {
10134 MaskedStoreSDNode *ST = cast<MaskedStoreSDNode>(Val&: OrigStore);
10135 assert(ST->getOffset().isUndef() &&
10136 "Masked store is already a indexed store!");
10137 return getMaskedStore(Chain: ST->getChain(), dl, Val: ST->getValue(), Base, Offset,
10138 Mask: ST->getMask(), MemVT: ST->getMemoryVT(), MMO: ST->getMemOperand(),
10139 AM, IsTruncating: ST->isTruncatingStore(), IsCompressing: ST->isCompressingStore());
10140}
10141
10142SDValue SelectionDAG::getMaskedGather(SDVTList VTs, EVT MemVT, const SDLoc &dl,
10143 ArrayRef<SDValue> Ops,
10144 MachineMemOperand *MMO,
10145 ISD::MemIndexType IndexType,
10146 ISD::LoadExtType ExtTy) {
10147 assert(Ops.size() == 6 && "Incompatible number of operands");
10148
10149 FoldingSetNodeID ID;
10150 AddNodeIDNode(ID, OpC: ISD::MGATHER, VTList: VTs, OpList: Ops);
10151 ID.AddInteger(I: MemVT.getRawBits());
10152 ID.AddInteger(I: getSyntheticNodeSubclassData<MaskedGatherSDNode>(
10153 IROrder: dl.getIROrder(), Args&: VTs, Args&: MemVT, Args&: MMO, Args&: IndexType, Args&: ExtTy));
10154 ID.AddInteger(I: MMO->getPointerInfo().getAddrSpace());
10155 ID.AddInteger(I: MMO->getFlags());
10156 void *IP = nullptr;
10157 if (SDNode *E = FindNodeOrInsertPos(ID, DL: dl, InsertPos&: IP)) {
10158 cast<MaskedGatherSDNode>(Val: E)->refineAlignment(NewMMO: MMO);
10159 return SDValue(E, 0);
10160 }
10161
10162 auto *N = newSDNode<MaskedGatherSDNode>(Args: dl.getIROrder(), Args: dl.getDebugLoc(),
10163 Args&: VTs, Args&: MemVT, Args&: MMO, Args&: IndexType, Args&: ExtTy);
10164 createOperands(Node: N, Vals: Ops);
10165
10166 assert(N->getPassThru().getValueType() == N->getValueType(0) &&
10167 "Incompatible type of the PassThru value in MaskedGatherSDNode");
10168 assert(N->getMask().getValueType().getVectorElementCount() ==
10169 N->getValueType(0).getVectorElementCount() &&
10170 "Vector width mismatch between mask and data");
10171 assert(N->getIndex().getValueType().getVectorElementCount().isScalable() ==
10172 N->getValueType(0).getVectorElementCount().isScalable() &&
10173 "Scalable flags of index and data do not match");
10174 assert(ElementCount::isKnownGE(
10175 N->getIndex().getValueType().getVectorElementCount(),
10176 N->getValueType(0).getVectorElementCount()) &&
10177 "Vector width mismatch between index and data");
10178 assert(isa<ConstantSDNode>(N->getScale()) &&
10179 N->getScale()->getAsAPIntVal().isPowerOf2() &&
10180 "Scale should be a constant power of 2");
10181
10182 CSEMap.InsertNode(N, InsertPos: IP);
10183 InsertNode(N);
10184 SDValue V(N, 0);
10185 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
10186 return V;
10187}
10188
10189SDValue SelectionDAG::getMaskedScatter(SDVTList VTs, EVT MemVT, const SDLoc &dl,
10190 ArrayRef<SDValue> Ops,
10191 MachineMemOperand *MMO,
10192 ISD::MemIndexType IndexType,
10193 bool IsTrunc) {
10194 assert(Ops.size() == 6 && "Incompatible number of operands");
10195
10196 FoldingSetNodeID ID;
10197 AddNodeIDNode(ID, OpC: ISD::MSCATTER, VTList: VTs, OpList: Ops);
10198 ID.AddInteger(I: MemVT.getRawBits());
10199 ID.AddInteger(I: getSyntheticNodeSubclassData<MaskedScatterSDNode>(
10200 IROrder: dl.getIROrder(), Args&: VTs, Args&: MemVT, Args&: MMO, Args&: IndexType, Args&: IsTrunc));
10201 ID.AddInteger(I: MMO->getPointerInfo().getAddrSpace());
10202 ID.AddInteger(I: MMO->getFlags());
10203 void *IP = nullptr;
10204 if (SDNode *E = FindNodeOrInsertPos(ID, DL: dl, InsertPos&: IP)) {
10205 cast<MaskedScatterSDNode>(Val: E)->refineAlignment(NewMMO: MMO);
10206 return SDValue(E, 0);
10207 }
10208
10209 auto *N = newSDNode<MaskedScatterSDNode>(Args: dl.getIROrder(), Args: dl.getDebugLoc(),
10210 Args&: VTs, Args&: MemVT, Args&: MMO, Args&: IndexType, Args&: IsTrunc);
10211 createOperands(Node: N, Vals: Ops);
10212
10213 assert(N->getMask().getValueType().getVectorElementCount() ==
10214 N->getValue().getValueType().getVectorElementCount() &&
10215 "Vector width mismatch between mask and data");
10216 assert(
10217 N->getIndex().getValueType().getVectorElementCount().isScalable() ==
10218 N->getValue().getValueType().getVectorElementCount().isScalable() &&
10219 "Scalable flags of index and data do not match");
10220 assert(ElementCount::isKnownGE(
10221 N->getIndex().getValueType().getVectorElementCount(),
10222 N->getValue().getValueType().getVectorElementCount()) &&
10223 "Vector width mismatch between index and data");
10224 assert(isa<ConstantSDNode>(N->getScale()) &&
10225 N->getScale()->getAsAPIntVal().isPowerOf2() &&
10226 "Scale should be a constant power of 2");
10227
10228 CSEMap.InsertNode(N, InsertPos: IP);
10229 InsertNode(N);
10230 SDValue V(N, 0);
10231 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
10232 return V;
10233}
10234
10235SDValue SelectionDAG::getMaskedHistogram(SDVTList VTs, EVT MemVT,
10236 const SDLoc &dl, ArrayRef<SDValue> Ops,
10237 MachineMemOperand *MMO,
10238 ISD::MemIndexType IndexType) {
10239 assert(Ops.size() == 7 && "Incompatible number of operands");
10240
10241 FoldingSetNodeID ID;
10242 AddNodeIDNode(ID, OpC: ISD::EXPERIMENTAL_VECTOR_HISTOGRAM, VTList: VTs, OpList: Ops);
10243 ID.AddInteger(I: MemVT.getRawBits());
10244 ID.AddInteger(I: getSyntheticNodeSubclassData<MaskedHistogramSDNode>(
10245 IROrder: dl.getIROrder(), Args&: VTs, Args&: MemVT, Args&: MMO, Args&: IndexType));
10246 ID.AddInteger(I: MMO->getPointerInfo().getAddrSpace());
10247 ID.AddInteger(I: MMO->getFlags());
10248 void *IP = nullptr;
10249 if (SDNode *E = FindNodeOrInsertPos(ID, DL: dl, InsertPos&: IP)) {
10250 cast<MaskedGatherSDNode>(Val: E)->refineAlignment(NewMMO: MMO);
10251 return SDValue(E, 0);
10252 }
10253
10254 auto *N = newSDNode<MaskedHistogramSDNode>(Args: dl.getIROrder(), Args: dl.getDebugLoc(),
10255 Args&: VTs, Args&: MemVT, Args&: MMO, Args&: IndexType);
10256 createOperands(Node: N, Vals: Ops);
10257
10258 assert(N->getMask().getValueType().getVectorElementCount() ==
10259 N->getIndex().getValueType().getVectorElementCount() &&
10260 "Vector width mismatch between mask and data");
10261 assert(isa<ConstantSDNode>(N->getScale()) &&
10262 N->getScale()->getAsAPIntVal().isPowerOf2() &&
10263 "Scale should be a constant power of 2");
10264 assert(N->getInc().getValueType().isInteger() && "Non integer update value");
10265
10266 CSEMap.InsertNode(N, InsertPos: IP);
10267 InsertNode(N);
10268 SDValue V(N, 0);
10269 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
10270 return V;
10271}
10272
10273SDValue SelectionDAG::getGetFPEnv(SDValue Chain, const SDLoc &dl, SDValue Ptr,
10274 EVT MemVT, MachineMemOperand *MMO) {
10275 assert(Chain.getValueType() == MVT::Other && "Invalid chain type");
10276 SDVTList VTs = getVTList(MVT::Other);
10277 SDValue Ops[] = {Chain, Ptr};
10278 FoldingSetNodeID ID;
10279 AddNodeIDNode(ID, OpC: ISD::GET_FPENV_MEM, VTList: VTs, OpList: Ops);
10280 ID.AddInteger(I: MemVT.getRawBits());
10281 ID.AddInteger(I: getSyntheticNodeSubclassData<FPStateAccessSDNode>(
10282 Opc: ISD::GET_FPENV_MEM, Order: dl.getIROrder(), VTs, MemoryVT: MemVT, MMO));
10283 ID.AddInteger(I: MMO->getPointerInfo().getAddrSpace());
10284 ID.AddInteger(I: MMO->getFlags());
10285 void *IP = nullptr;
10286 if (SDNode *E = FindNodeOrInsertPos(ID, DL: dl, InsertPos&: IP))
10287 return SDValue(E, 0);
10288
10289 auto *N = newSDNode<FPStateAccessSDNode>(Args: ISD::GET_FPENV_MEM, Args: dl.getIROrder(),
10290 Args: dl.getDebugLoc(), Args&: VTs, Args&: MemVT, Args&: MMO);
10291 createOperands(Node: N, Vals: Ops);
10292
10293 CSEMap.InsertNode(N, InsertPos: IP);
10294 InsertNode(N);
10295 SDValue V(N, 0);
10296 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
10297 return V;
10298}
10299
10300SDValue SelectionDAG::getSetFPEnv(SDValue Chain, const SDLoc &dl, SDValue Ptr,
10301 EVT MemVT, MachineMemOperand *MMO) {
10302 assert(Chain.getValueType() == MVT::Other && "Invalid chain type");
10303 SDVTList VTs = getVTList(MVT::Other);
10304 SDValue Ops[] = {Chain, Ptr};
10305 FoldingSetNodeID ID;
10306 AddNodeIDNode(ID, OpC: ISD::SET_FPENV_MEM, VTList: VTs, OpList: Ops);
10307 ID.AddInteger(I: MemVT.getRawBits());
10308 ID.AddInteger(I: getSyntheticNodeSubclassData<FPStateAccessSDNode>(
10309 Opc: ISD::SET_FPENV_MEM, Order: dl.getIROrder(), VTs, MemoryVT: MemVT, MMO));
10310 ID.AddInteger(I: MMO->getPointerInfo().getAddrSpace());
10311 ID.AddInteger(I: MMO->getFlags());
10312 void *IP = nullptr;
10313 if (SDNode *E = FindNodeOrInsertPos(ID, DL: dl, InsertPos&: IP))
10314 return SDValue(E, 0);
10315
10316 auto *N = newSDNode<FPStateAccessSDNode>(Args: ISD::SET_FPENV_MEM, Args: dl.getIROrder(),
10317 Args: dl.getDebugLoc(), Args&: VTs, Args&: MemVT, Args&: MMO);
10318 createOperands(Node: N, Vals: Ops);
10319
10320 CSEMap.InsertNode(N, InsertPos: IP);
10321 InsertNode(N);
10322 SDValue V(N, 0);
10323 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
10324 return V;
10325}
10326
10327SDValue SelectionDAG::simplifySelect(SDValue Cond, SDValue T, SDValue F) {
10328 // select undef, T, F --> T (if T is a constant), otherwise F
10329 // select, ?, undef, F --> F
10330 // select, ?, T, undef --> T
10331 if (Cond.isUndef())
10332 return isConstantValueOfAnyType(N: T) ? T : F;
10333 if (T.isUndef())
10334 return F;
10335 if (F.isUndef())
10336 return T;
10337
10338 // select true, T, F --> T
10339 // select false, T, F --> F
10340 if (auto C = isBoolConstant(N: Cond, /*AllowTruncation=*/true))
10341 return *C ? T : F;
10342
10343 // select ?, T, T --> T
10344 if (T == F)
10345 return T;
10346
10347 return SDValue();
10348}
10349
10350SDValue SelectionDAG::simplifyShift(SDValue X, SDValue Y) {
10351 // shift undef, Y --> 0 (can always assume that the undef value is 0)
10352 if (X.isUndef())
10353 return getConstant(Val: 0, DL: SDLoc(X.getNode()), VT: X.getValueType());
10354 // shift X, undef --> undef (because it may shift by the bitwidth)
10355 if (Y.isUndef())
10356 return getUNDEF(VT: X.getValueType());
10357
10358 // shift 0, Y --> 0
10359 // shift X, 0 --> X
10360 if (isNullOrNullSplat(V: X) || isNullOrNullSplat(V: Y))
10361 return X;
10362
10363 // shift X, C >= bitwidth(X) --> undef
10364 // All vector elements must be too big (or undef) to avoid partial undefs.
10365 auto isShiftTooBig = [X](ConstantSDNode *Val) {
10366 return !Val || Val->getAPIntValue().uge(RHS: X.getScalarValueSizeInBits());
10367 };
10368 if (ISD::matchUnaryPredicate(Op: Y, Match: isShiftTooBig, AllowUndefs: true))
10369 return getUNDEF(VT: X.getValueType());
10370
10371 // shift i1/vXi1 X, Y --> X (any non-zero shift amount is undefined).
10372 if (X.getValueType().getScalarType() == MVT::i1)
10373 return X;
10374
10375 return SDValue();
10376}
10377
10378SDValue SelectionDAG::simplifyFPBinop(unsigned Opcode, SDValue X, SDValue Y,
10379 SDNodeFlags Flags) {
10380 // If this operation has 'nnan' or 'ninf' and at least 1 disallowed operand
10381 // (an undef operand can be chosen to be Nan/Inf), then the result of this
10382 // operation is poison. That result can be relaxed to undef.
10383 ConstantFPSDNode *XC = isConstOrConstSplatFP(N: X, /* AllowUndefs */ true);
10384 ConstantFPSDNode *YC = isConstOrConstSplatFP(N: Y, /* AllowUndefs */ true);
10385 bool HasNan = (XC && XC->getValueAPF().isNaN()) ||
10386 (YC && YC->getValueAPF().isNaN());
10387 bool HasInf = (XC && XC->getValueAPF().isInfinity()) ||
10388 (YC && YC->getValueAPF().isInfinity());
10389
10390 if (Flags.hasNoNaNs() && (HasNan || X.isUndef() || Y.isUndef()))
10391 return getUNDEF(VT: X.getValueType());
10392
10393 if (Flags.hasNoInfs() && (HasInf || X.isUndef() || Y.isUndef()))
10394 return getUNDEF(VT: X.getValueType());
10395
10396 if (!YC)
10397 return SDValue();
10398
10399 // X + -0.0 --> X
10400 if (Opcode == ISD::FADD)
10401 if (YC->getValueAPF().isNegZero())
10402 return X;
10403
10404 // X - +0.0 --> X
10405 if (Opcode == ISD::FSUB)
10406 if (YC->getValueAPF().isPosZero())
10407 return X;
10408
10409 // X * 1.0 --> X
10410 // X / 1.0 --> X
10411 if (Opcode == ISD::FMUL || Opcode == ISD::FDIV)
10412 if (YC->getValueAPF().isExactlyValue(V: 1.0))
10413 return X;
10414
10415 // X * 0.0 --> 0.0
10416 if (Opcode == ISD::FMUL && Flags.hasNoNaNs() && Flags.hasNoSignedZeros())
10417 if (YC->getValueAPF().isZero())
10418 return getConstantFP(Val: 0.0, DL: SDLoc(Y), VT: Y.getValueType());
10419
10420 return SDValue();
10421}
10422
10423SDValue SelectionDAG::getVAArg(EVT VT, const SDLoc &dl, SDValue Chain,
10424 SDValue Ptr, SDValue SV, unsigned Align) {
10425 SDValue Ops[] = { Chain, Ptr, SV, getTargetConstant(Align, dl, MVT::i32) };
10426 return getNode(ISD::VAARG, dl, getVTList(VT, MVT::Other), Ops);
10427}
10428
10429SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
10430 ArrayRef<SDUse> Ops) {
10431 switch (Ops.size()) {
10432 case 0: return getNode(Opcode, DL, VT);
10433 case 1: return getNode(Opcode, DL, VT, N1: static_cast<const SDValue>(Ops[0]));
10434 case 2: return getNode(Opcode, DL, VT, N1: Ops[0], N2: Ops[1]);
10435 case 3: return getNode(Opcode, DL, VT, N1: Ops[0], N2: Ops[1], N3: Ops[2]);
10436 default: break;
10437 }
10438
10439 // Copy from an SDUse array into an SDValue array for use with
10440 // the regular getNode logic.
10441 SmallVector<SDValue, 8> NewOps(Ops);
10442 return getNode(Opcode, DL, VT, Ops: NewOps);
10443}
10444
10445SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
10446 ArrayRef<SDValue> Ops) {
10447 SDNodeFlags Flags;
10448 if (Inserter)
10449 Flags = Inserter->getFlags();
10450 return getNode(Opcode, DL, VT, Ops, Flags);
10451}
10452
10453SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
10454 ArrayRef<SDValue> Ops, const SDNodeFlags Flags) {
10455 unsigned NumOps = Ops.size();
10456 switch (NumOps) {
10457 case 0: return getNode(Opcode, DL, VT);
10458 case 1: return getNode(Opcode, DL, VT, N1: Ops[0], Flags);
10459 case 2: return getNode(Opcode, DL, VT, N1: Ops[0], N2: Ops[1], Flags);
10460 case 3: return getNode(Opcode, DL, VT, N1: Ops[0], N2: Ops[1], N3: Ops[2], Flags);
10461 default: break;
10462 }
10463
10464#ifndef NDEBUG
10465 for (const auto &Op : Ops)
10466 assert(Op.getOpcode() != ISD::DELETED_NODE &&
10467 "Operand is DELETED_NODE!");
10468#endif
10469
10470 switch (Opcode) {
10471 default: break;
10472 case ISD::BUILD_VECTOR:
10473 // Attempt to simplify BUILD_VECTOR.
10474 if (SDValue V = FoldBUILD_VECTOR(DL, VT, Ops, DAG&: *this))
10475 return V;
10476 break;
10477 case ISD::CONCAT_VECTORS:
10478 if (SDValue V = foldCONCAT_VECTORS(DL, VT, Ops, DAG&: *this))
10479 return V;
10480 break;
10481 case ISD::SELECT_CC:
10482 assert(NumOps == 5 && "SELECT_CC takes 5 operands!");
10483 assert(Ops[0].getValueType() == Ops[1].getValueType() &&
10484 "LHS and RHS of condition must have same type!");
10485 assert(Ops[2].getValueType() == Ops[3].getValueType() &&
10486 "True and False arms of SelectCC must have same type!");
10487 assert(Ops[2].getValueType() == VT &&
10488 "select_cc node must be of same type as true and false value!");
10489 assert((!Ops[0].getValueType().isVector() ||
10490 Ops[0].getValueType().getVectorElementCount() ==
10491 VT.getVectorElementCount()) &&
10492 "Expected select_cc with vector result to have the same sized "
10493 "comparison type!");
10494 break;
10495 case ISD::BR_CC:
10496 assert(NumOps == 5 && "BR_CC takes 5 operands!");
10497 assert(Ops[2].getValueType() == Ops[3].getValueType() &&
10498 "LHS/RHS of comparison should match types!");
10499 break;
10500 case ISD::VP_ADD:
10501 case ISD::VP_SUB:
10502 // If it is VP_ADD/VP_SUB mask operation then turn it to VP_XOR
10503 if (VT.getScalarType() == MVT::i1)
10504 Opcode = ISD::VP_XOR;
10505 break;
10506 case ISD::VP_MUL:
10507 // If it is VP_MUL mask operation then turn it to VP_AND
10508 if (VT.getScalarType() == MVT::i1)
10509 Opcode = ISD::VP_AND;
10510 break;
10511 case ISD::VP_REDUCE_MUL:
10512 // If it is VP_REDUCE_MUL mask operation then turn it to VP_REDUCE_AND
10513 if (VT == MVT::i1)
10514 Opcode = ISD::VP_REDUCE_AND;
10515 break;
10516 case ISD::VP_REDUCE_ADD:
10517 // If it is VP_REDUCE_ADD mask operation then turn it to VP_REDUCE_XOR
10518 if (VT == MVT::i1)
10519 Opcode = ISD::VP_REDUCE_XOR;
10520 break;
10521 case ISD::VP_REDUCE_SMAX:
10522 case ISD::VP_REDUCE_UMIN:
10523 // If it is VP_REDUCE_SMAX/VP_REDUCE_UMIN mask operation then turn it to
10524 // VP_REDUCE_AND.
10525 if (VT == MVT::i1)
10526 Opcode = ISD::VP_REDUCE_AND;
10527 break;
10528 case ISD::VP_REDUCE_SMIN:
10529 case ISD::VP_REDUCE_UMAX:
10530 // If it is VP_REDUCE_SMIN/VP_REDUCE_UMAX mask operation then turn it to
10531 // VP_REDUCE_OR.
10532 if (VT == MVT::i1)
10533 Opcode = ISD::VP_REDUCE_OR;
10534 break;
10535 }
10536
10537 // Memoize nodes.
10538 SDNode *N;
10539 SDVTList VTs = getVTList(VT);
10540
10541 if (VT != MVT::Glue) {
10542 FoldingSetNodeID ID;
10543 AddNodeIDNode(ID, OpC: Opcode, VTList: VTs, OpList: Ops);
10544 void *IP = nullptr;
10545
10546 if (SDNode *E = FindNodeOrInsertPos(ID, DL, InsertPos&: IP)) {
10547 E->intersectFlagsWith(Flags);
10548 return SDValue(E, 0);
10549 }
10550
10551 N = newSDNode<SDNode>(Args&: Opcode, Args: DL.getIROrder(), Args: DL.getDebugLoc(), Args&: VTs);
10552 createOperands(Node: N, Vals: Ops);
10553
10554 CSEMap.InsertNode(N, InsertPos: IP);
10555 } else {
10556 N = newSDNode<SDNode>(Args&: Opcode, Args: DL.getIROrder(), Args: DL.getDebugLoc(), Args&: VTs);
10557 createOperands(Node: N, Vals: Ops);
10558 }
10559
10560 N->setFlags(Flags);
10561 InsertNode(N);
10562 SDValue V(N, 0);
10563 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
10564 return V;
10565}
10566
10567SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL,
10568 ArrayRef<EVT> ResultTys, ArrayRef<SDValue> Ops) {
10569 return getNode(Opcode, DL, VTList: getVTList(VTs: ResultTys), Ops);
10570}
10571
10572SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList,
10573 ArrayRef<SDValue> Ops) {
10574 SDNodeFlags Flags;
10575 if (Inserter)
10576 Flags = Inserter->getFlags();
10577 return getNode(Opcode, DL, VTList, Ops, Flags);
10578}
10579
10580SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList,
10581 ArrayRef<SDValue> Ops, const SDNodeFlags Flags) {
10582 if (VTList.NumVTs == 1)
10583 return getNode(Opcode, DL, VT: VTList.VTs[0], Ops, Flags);
10584
10585#ifndef NDEBUG
10586 for (const auto &Op : Ops)
10587 assert(Op.getOpcode() != ISD::DELETED_NODE &&
10588 "Operand is DELETED_NODE!");
10589#endif
10590
10591 switch (Opcode) {
10592 case ISD::SADDO:
10593 case ISD::UADDO:
10594 case ISD::SSUBO:
10595 case ISD::USUBO: {
10596 assert(VTList.NumVTs == 2 && Ops.size() == 2 &&
10597 "Invalid add/sub overflow op!");
10598 assert(VTList.VTs[0].isInteger() && VTList.VTs[1].isInteger() &&
10599 Ops[0].getValueType() == Ops[1].getValueType() &&
10600 Ops[0].getValueType() == VTList.VTs[0] &&
10601 "Binary operator types must match!");
10602 SDValue N1 = Ops[0], N2 = Ops[1];
10603 canonicalizeCommutativeBinop(Opcode, N1, N2);
10604
10605 // (X +- 0) -> X with zero-overflow.
10606 ConstantSDNode *N2CV = isConstOrConstSplat(N: N2, /*AllowUndefs*/ false,
10607 /*AllowTruncation*/ true);
10608 if (N2CV && N2CV->isZero()) {
10609 SDValue ZeroOverFlow = getConstant(Val: 0, DL, VT: VTList.VTs[1]);
10610 return getNode(Opcode: ISD::MERGE_VALUES, DL, VTList, Ops: {N1, ZeroOverFlow}, Flags);
10611 }
10612
10613 if (VTList.VTs[0].getScalarType() == MVT::i1 &&
10614 VTList.VTs[1].getScalarType() == MVT::i1) {
10615 SDValue F1 = getFreeze(V: N1);
10616 SDValue F2 = getFreeze(V: N2);
10617 // {vXi1,vXi1} (u/s)addo(vXi1 x, vXi1y) -> {xor(x,y),and(x,y)}
10618 if (Opcode == ISD::UADDO || Opcode == ISD::SADDO)
10619 return getNode(Opcode: ISD::MERGE_VALUES, DL, VTList,
10620 Ops: {getNode(Opcode: ISD::XOR, DL, VT: VTList.VTs[0], N1: F1, N2: F2),
10621 getNode(Opcode: ISD::AND, DL, VT: VTList.VTs[1], N1: F1, N2: F2)},
10622 Flags);
10623 // {vXi1,vXi1} (u/s)subo(vXi1 x, vXi1y) -> {xor(x,y),and(~x,y)}
10624 if (Opcode == ISD::USUBO || Opcode == ISD::SSUBO) {
10625 SDValue NotF1 = getNOT(DL, Val: F1, VT: VTList.VTs[0]);
10626 return getNode(Opcode: ISD::MERGE_VALUES, DL, VTList,
10627 Ops: {getNode(Opcode: ISD::XOR, DL, VT: VTList.VTs[0], N1: F1, N2: F2),
10628 getNode(Opcode: ISD::AND, DL, VT: VTList.VTs[1], N1: NotF1, N2: F2)},
10629 Flags);
10630 }
10631 }
10632 break;
10633 }
10634 case ISD::SADDO_CARRY:
10635 case ISD::UADDO_CARRY:
10636 case ISD::SSUBO_CARRY:
10637 case ISD::USUBO_CARRY:
10638 assert(VTList.NumVTs == 2 && Ops.size() == 3 &&
10639 "Invalid add/sub overflow op!");
10640 assert(VTList.VTs[0].isInteger() && VTList.VTs[1].isInteger() &&
10641 Ops[0].getValueType() == Ops[1].getValueType() &&
10642 Ops[0].getValueType() == VTList.VTs[0] &&
10643 Ops[2].getValueType() == VTList.VTs[1] &&
10644 "Binary operator types must match!");
10645 break;
10646 case ISD::SMUL_LOHI:
10647 case ISD::UMUL_LOHI: {
10648 assert(VTList.NumVTs == 2 && Ops.size() == 2 && "Invalid mul lo/hi op!");
10649 assert(VTList.VTs[0].isInteger() && VTList.VTs[0] == VTList.VTs[1] &&
10650 VTList.VTs[0] == Ops[0].getValueType() &&
10651 VTList.VTs[0] == Ops[1].getValueType() &&
10652 "Binary operator types must match!");
10653 // Constant fold.
10654 ConstantSDNode *LHS = dyn_cast<ConstantSDNode>(Val: Ops[0]);
10655 ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Val: Ops[1]);
10656 if (LHS && RHS) {
10657 unsigned Width = VTList.VTs[0].getScalarSizeInBits();
10658 unsigned OutWidth = Width * 2;
10659 APInt Val = LHS->getAPIntValue();
10660 APInt Mul = RHS->getAPIntValue();
10661 if (Opcode == ISD::SMUL_LOHI) {
10662 Val = Val.sext(width: OutWidth);
10663 Mul = Mul.sext(width: OutWidth);
10664 } else {
10665 Val = Val.zext(width: OutWidth);
10666 Mul = Mul.zext(width: OutWidth);
10667 }
10668 Val *= Mul;
10669
10670 SDValue Hi =
10671 getConstant(Val: Val.extractBits(numBits: Width, bitPosition: Width), DL, VT: VTList.VTs[0]);
10672 SDValue Lo = getConstant(Val: Val.trunc(width: Width), DL, VT: VTList.VTs[0]);
10673 return getNode(Opcode: ISD::MERGE_VALUES, DL, VTList, Ops: {Lo, Hi}, Flags);
10674 }
10675 break;
10676 }
10677 case ISD::FFREXP: {
10678 assert(VTList.NumVTs == 2 && Ops.size() == 1 && "Invalid ffrexp op!");
10679 assert(VTList.VTs[0].isFloatingPoint() && VTList.VTs[1].isInteger() &&
10680 VTList.VTs[0] == Ops[0].getValueType() && "frexp type mismatch");
10681
10682 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val: Ops[0])) {
10683 int FrexpExp;
10684 APFloat FrexpMant =
10685 frexp(X: C->getValueAPF(), Exp&: FrexpExp, RM: APFloat::rmNearestTiesToEven);
10686 SDValue Result0 = getConstantFP(V: FrexpMant, DL, VT: VTList.VTs[0]);
10687 SDValue Result1 =
10688 getConstant(Val: FrexpMant.isFinite() ? FrexpExp : 0, DL, VT: VTList.VTs[1]);
10689 return getNode(Opcode: ISD::MERGE_VALUES, DL, VTList, Ops: {Result0, Result1}, Flags);
10690 }
10691
10692 break;
10693 }
10694 case ISD::STRICT_FP_EXTEND:
10695 assert(VTList.NumVTs == 2 && Ops.size() == 2 &&
10696 "Invalid STRICT_FP_EXTEND!");
10697 assert(VTList.VTs[0].isFloatingPoint() &&
10698 Ops[1].getValueType().isFloatingPoint() && "Invalid FP cast!");
10699 assert(VTList.VTs[0].isVector() == Ops[1].getValueType().isVector() &&
10700 "STRICT_FP_EXTEND result type should be vector iff the operand "
10701 "type is vector!");
10702 assert((!VTList.VTs[0].isVector() ||
10703 VTList.VTs[0].getVectorElementCount() ==
10704 Ops[1].getValueType().getVectorElementCount()) &&
10705 "Vector element count mismatch!");
10706 assert(Ops[1].getValueType().bitsLT(VTList.VTs[0]) &&
10707 "Invalid fpext node, dst <= src!");
10708 break;
10709 case ISD::STRICT_FP_ROUND:
10710 assert(VTList.NumVTs == 2 && Ops.size() == 3 && "Invalid STRICT_FP_ROUND!");
10711 assert(VTList.VTs[0].isVector() == Ops[1].getValueType().isVector() &&
10712 "STRICT_FP_ROUND result type should be vector iff the operand "
10713 "type is vector!");
10714 assert((!VTList.VTs[0].isVector() ||
10715 VTList.VTs[0].getVectorElementCount() ==
10716 Ops[1].getValueType().getVectorElementCount()) &&
10717 "Vector element count mismatch!");
10718 assert(VTList.VTs[0].isFloatingPoint() &&
10719 Ops[1].getValueType().isFloatingPoint() &&
10720 VTList.VTs[0].bitsLT(Ops[1].getValueType()) &&
10721 Ops[2].getOpcode() == ISD::TargetConstant &&
10722 (Ops[2]->getAsZExtVal() == 0 || Ops[2]->getAsZExtVal() == 1) &&
10723 "Invalid STRICT_FP_ROUND!");
10724 break;
10725#if 0
10726 // FIXME: figure out how to safely handle things like
10727 // int foo(int x) { return 1 << (x & 255); }
10728 // int bar() { return foo(256); }
10729 case ISD::SRA_PARTS:
10730 case ISD::SRL_PARTS:
10731 case ISD::SHL_PARTS:
10732 if (N3.getOpcode() == ISD::SIGN_EXTEND_INREG &&
10733 cast<VTSDNode>(N3.getOperand(1))->getVT() != MVT::i1)
10734 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0));
10735 else if (N3.getOpcode() == ISD::AND)
10736 if (ConstantSDNode *AndRHS = dyn_cast<ConstantSDNode>(N3.getOperand(1))) {
10737 // If the and is only masking out bits that cannot effect the shift,
10738 // eliminate the and.
10739 unsigned NumBits = VT.getScalarSizeInBits()*2;
10740 if ((AndRHS->getValue() & (NumBits-1)) == NumBits-1)
10741 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0));
10742 }
10743 break;
10744#endif
10745 }
10746
10747 // Memoize the node unless it returns a glue result.
10748 SDNode *N;
10749 if (VTList.VTs[VTList.NumVTs-1] != MVT::Glue) {
10750 FoldingSetNodeID ID;
10751 AddNodeIDNode(ID, OpC: Opcode, VTList, OpList: Ops);
10752 void *IP = nullptr;
10753 if (SDNode *E = FindNodeOrInsertPos(ID, DL, InsertPos&: IP)) {
10754 E->intersectFlagsWith(Flags);
10755 return SDValue(E, 0);
10756 }
10757
10758 N = newSDNode<SDNode>(Args&: Opcode, Args: DL.getIROrder(), Args: DL.getDebugLoc(), Args&: VTList);
10759 createOperands(Node: N, Vals: Ops);
10760 CSEMap.InsertNode(N, InsertPos: IP);
10761 } else {
10762 N = newSDNode<SDNode>(Args&: Opcode, Args: DL.getIROrder(), Args: DL.getDebugLoc(), Args&: VTList);
10763 createOperands(Node: N, Vals: Ops);
10764 }
10765
10766 N->setFlags(Flags);
10767 InsertNode(N);
10768 SDValue V(N, 0);
10769 NewSDValueDbgMsg(V, Msg: "Creating new node: ", G: this);
10770 return V;
10771}
10772
10773SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL,
10774 SDVTList VTList) {
10775 return getNode(Opcode, DL, VTList, Ops: ArrayRef<SDValue>());
10776}
10777
10778SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList,
10779 SDValue N1) {
10780 SDValue Ops[] = { N1 };
10781 return getNode(Opcode, DL, VTList, Ops);
10782}
10783
10784SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList,
10785 SDValue N1, SDValue N2) {
10786 SDValue Ops[] = { N1, N2 };
10787 return getNode(Opcode, DL, VTList, Ops);
10788}
10789
10790SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList,
10791 SDValue N1, SDValue N2, SDValue N3) {
10792 SDValue Ops[] = { N1, N2, N3 };
10793 return getNode(Opcode, DL, VTList, Ops);
10794}
10795
10796SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList,
10797 SDValue N1, SDValue N2, SDValue N3, SDValue N4) {
10798 SDValue Ops[] = { N1, N2, N3, N4 };
10799 return getNode(Opcode, DL, VTList, Ops);
10800}
10801
10802SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList,
10803 SDValue N1, SDValue N2, SDValue N3, SDValue N4,
10804 SDValue N5) {
10805 SDValue Ops[] = { N1, N2, N3, N4, N5 };
10806 return getNode(Opcode, DL, VTList, Ops);
10807}
10808
10809SDVTList SelectionDAG::getVTList(EVT VT) {
10810 if (!VT.isExtended())
10811 return makeVTList(VTs: SDNode::getValueTypeList(VT: VT.getSimpleVT()), NumVTs: 1);
10812
10813 return makeVTList(VTs: &(*EVTs.insert(x: VT).first), NumVTs: 1);
10814}
10815
10816SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2) {
10817 FoldingSetNodeID ID;
10818 ID.AddInteger(I: 2U);
10819 ID.AddInteger(I: VT1.getRawBits());
10820 ID.AddInteger(I: VT2.getRawBits());
10821
10822 void *IP = nullptr;
10823 SDVTListNode *Result = VTListMap.FindNodeOrInsertPos(ID, InsertPos&: IP);
10824 if (!Result) {
10825 EVT *Array = Allocator.Allocate<EVT>(Num: 2);
10826 Array[0] = VT1;
10827 Array[1] = VT2;
10828 Result = new (Allocator) SDVTListNode(ID.Intern(Allocator), Array, 2);
10829 VTListMap.InsertNode(N: Result, InsertPos: IP);
10830 }
10831 return Result->getSDVTList();
10832}
10833
10834SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3) {
10835 FoldingSetNodeID ID;
10836 ID.AddInteger(I: 3U);
10837 ID.AddInteger(I: VT1.getRawBits());
10838 ID.AddInteger(I: VT2.getRawBits());
10839 ID.AddInteger(I: VT3.getRawBits());
10840
10841 void *IP = nullptr;
10842 SDVTListNode *Result = VTListMap.FindNodeOrInsertPos(ID, InsertPos&: IP);
10843 if (!Result) {
10844 EVT *Array = Allocator.Allocate<EVT>(Num: 3);
10845 Array[0] = VT1;
10846 Array[1] = VT2;
10847 Array[2] = VT3;
10848 Result = new (Allocator) SDVTListNode(ID.Intern(Allocator), Array, 3);
10849 VTListMap.InsertNode(N: Result, InsertPos: IP);
10850 }
10851 return Result->getSDVTList();
10852}
10853
10854SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3, EVT VT4) {
10855 FoldingSetNodeID ID;
10856 ID.AddInteger(I: 4U);
10857 ID.AddInteger(I: VT1.getRawBits());
10858 ID.AddInteger(I: VT2.getRawBits());
10859 ID.AddInteger(I: VT3.getRawBits());
10860 ID.AddInteger(I: VT4.getRawBits());
10861
10862 void *IP = nullptr;
10863 SDVTListNode *Result = VTListMap.FindNodeOrInsertPos(ID, InsertPos&: IP);
10864 if (!Result) {
10865 EVT *Array = Allocator.Allocate<EVT>(Num: 4);
10866 Array[0] = VT1;
10867 Array[1] = VT2;
10868 Array[2] = VT3;
10869 Array[3] = VT4;
10870 Result = new (Allocator) SDVTListNode(ID.Intern(Allocator), Array, 4);
10871 VTListMap.InsertNode(N: Result, InsertPos: IP);
10872 }
10873 return Result->getSDVTList();
10874}
10875
10876SDVTList SelectionDAG::getVTList(ArrayRef<EVT> VTs) {
10877 unsigned NumVTs = VTs.size();
10878 FoldingSetNodeID ID;
10879 ID.AddInteger(I: NumVTs);
10880 for (unsigned index = 0; index < NumVTs; index++) {
10881 ID.AddInteger(I: VTs[index].getRawBits());
10882 }
10883
10884 void *IP = nullptr;
10885 SDVTListNode *Result = VTListMap.FindNodeOrInsertPos(ID, InsertPos&: IP);
10886 if (!Result) {
10887 EVT *Array = Allocator.Allocate<EVT>(Num: NumVTs);
10888 llvm::copy(Range&: VTs, Out: Array);
10889 Result = new (Allocator) SDVTListNode(ID.Intern(Allocator), Array, NumVTs);
10890 VTListMap.InsertNode(N: Result, InsertPos: IP);
10891 }
10892 return Result->getSDVTList();
10893}
10894
10895
10896/// UpdateNodeOperands - *Mutate* the specified node in-place to have the
10897/// specified operands. If the resultant node already exists in the DAG,
10898/// this does not modify the specified node, instead it returns the node that
10899/// already exists. If the resultant node does not exist in the DAG, the
10900/// input node is returned. As a degenerate case, if you specify the same
10901/// input operands as the node already has, the input node is returned.
10902SDNode *SelectionDAG::UpdateNodeOperands(SDNode *N, SDValue Op) {
10903 assert(N->getNumOperands() == 1 && "Update with wrong number of operands");
10904
10905 // Check to see if there is no change.
10906 if (Op == N->getOperand(Num: 0)) return N;
10907
10908 // See if the modified node already exists.
10909 void *InsertPos = nullptr;
10910 if (SDNode *Existing = FindModifiedNodeSlot(N, Op, InsertPos))
10911 return Existing;
10912
10913 // Nope it doesn't. Remove the node from its current place in the maps.
10914 if (InsertPos)
10915 if (!RemoveNodeFromCSEMaps(N))
10916 InsertPos = nullptr;
10917
10918 // Now we update the operands.
10919 N->OperandList[0].set(Op);
10920
10921 updateDivergence(N);
10922 // If this gets put into a CSE map, add it.
10923 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
10924 return N;
10925}
10926
10927SDNode *SelectionDAG::UpdateNodeOperands(SDNode *N, SDValue Op1, SDValue Op2) {
10928 assert(N->getNumOperands() == 2 && "Update with wrong number of operands");
10929
10930 // Check to see if there is no change.
10931 if (Op1 == N->getOperand(Num: 0) && Op2 == N->getOperand(Num: 1))
10932 return N; // No operands changed, just return the input node.
10933
10934 // See if the modified node already exists.
10935 void *InsertPos = nullptr;
10936 if (SDNode *Existing = FindModifiedNodeSlot(N, Op1, Op2, InsertPos))
10937 return Existing;
10938
10939 // Nope it doesn't. Remove the node from its current place in the maps.
10940 if (InsertPos)
10941 if (!RemoveNodeFromCSEMaps(N))
10942 InsertPos = nullptr;
10943
10944 // Now we update the operands.
10945 if (N->OperandList[0] != Op1)
10946 N->OperandList[0].set(Op1);
10947 if (N->OperandList[1] != Op2)
10948 N->OperandList[1].set(Op2);
10949
10950 updateDivergence(N);
10951 // If this gets put into a CSE map, add it.
10952 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
10953 return N;
10954}
10955
10956SDNode *SelectionDAG::
10957UpdateNodeOperands(SDNode *N, SDValue Op1, SDValue Op2, SDValue Op3) {
10958 SDValue Ops[] = { Op1, Op2, Op3 };
10959 return UpdateNodeOperands(N, Ops);
10960}
10961
10962SDNode *SelectionDAG::
10963UpdateNodeOperands(SDNode *N, SDValue Op1, SDValue Op2,
10964 SDValue Op3, SDValue Op4) {
10965 SDValue Ops[] = { Op1, Op2, Op3, Op4 };
10966 return UpdateNodeOperands(N, Ops);
10967}
10968
10969SDNode *SelectionDAG::
10970UpdateNodeOperands(SDNode *N, SDValue Op1, SDValue Op2,
10971 SDValue Op3, SDValue Op4, SDValue Op5) {
10972 SDValue Ops[] = { Op1, Op2, Op3, Op4, Op5 };
10973 return UpdateNodeOperands(N, Ops);
10974}
10975
10976SDNode *SelectionDAG::
10977UpdateNodeOperands(SDNode *N, ArrayRef<SDValue> Ops) {
10978 unsigned NumOps = Ops.size();
10979 assert(N->getNumOperands() == NumOps &&
10980 "Update with wrong number of operands");
10981
10982 // If no operands changed just return the input node.
10983 if (std::equal(first1: Ops.begin(), last1: Ops.end(), first2: N->op_begin()))
10984 return N;
10985
10986 // See if the modified node already exists.
10987 void *InsertPos = nullptr;
10988 if (SDNode *Existing = FindModifiedNodeSlot(N, Ops, InsertPos))
10989 return Existing;
10990
10991 // Nope it doesn't. Remove the node from its current place in the maps.
10992 if (InsertPos)
10993 if (!RemoveNodeFromCSEMaps(N))
10994 InsertPos = nullptr;
10995
10996 // Now we update the operands.
10997 for (unsigned i = 0; i != NumOps; ++i)
10998 if (N->OperandList[i] != Ops[i])
10999 N->OperandList[i].set(Ops[i]);
11000
11001 updateDivergence(N);
11002 // If this gets put into a CSE map, add it.
11003 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
11004 return N;
11005}
11006
11007/// DropOperands - Release the operands and set this node to have
11008/// zero operands.
11009void SDNode::DropOperands() {
11010 // Unlike the code in MorphNodeTo that does this, we don't need to
11011 // watch for dead nodes here.
11012 for (op_iterator I = op_begin(), E = op_end(); I != E; ) {
11013 SDUse &Use = *I++;
11014 Use.set(SDValue());
11015 }
11016}
11017
11018void SelectionDAG::setNodeMemRefs(MachineSDNode *N,
11019 ArrayRef<MachineMemOperand *> NewMemRefs) {
11020 if (NewMemRefs.empty()) {
11021 N->clearMemRefs();
11022 return;
11023 }
11024
11025 // Check if we can avoid allocating by storing a single reference directly.
11026 if (NewMemRefs.size() == 1) {
11027 N->MemRefs = NewMemRefs[0];
11028 N->NumMemRefs = 1;
11029 return;
11030 }
11031
11032 MachineMemOperand **MemRefsBuffer =
11033 Allocator.template Allocate<MachineMemOperand *>(Num: NewMemRefs.size());
11034 llvm::copy(Range&: NewMemRefs, Out: MemRefsBuffer);
11035 N->MemRefs = MemRefsBuffer;
11036 N->NumMemRefs = static_cast<int>(NewMemRefs.size());
11037}
11038
11039/// SelectNodeTo - These are wrappers around MorphNodeTo that accept a
11040/// machine opcode.
11041///
11042SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
11043 EVT VT) {
11044 SDVTList VTs = getVTList(VT);
11045 return SelectNodeTo(N, MachineOpc, VTs, Ops: {});
11046}
11047
11048SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
11049 EVT VT, SDValue Op1) {
11050 SDVTList VTs = getVTList(VT);
11051 SDValue Ops[] = { Op1 };
11052 return SelectNodeTo(N, MachineOpc, VTs, Ops);
11053}
11054
11055SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
11056 EVT VT, SDValue Op1,
11057 SDValue Op2) {
11058 SDVTList VTs = getVTList(VT);
11059 SDValue Ops[] = { Op1, Op2 };
11060 return SelectNodeTo(N, MachineOpc, VTs, Ops);
11061}
11062
11063SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
11064 EVT VT, SDValue Op1,
11065 SDValue Op2, SDValue Op3) {
11066 SDVTList VTs = getVTList(VT);
11067 SDValue Ops[] = { Op1, Op2, Op3 };
11068 return SelectNodeTo(N, MachineOpc, VTs, Ops);
11069}
11070
11071SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
11072 EVT VT, ArrayRef<SDValue> Ops) {
11073 SDVTList VTs = getVTList(VT);
11074 return SelectNodeTo(N, MachineOpc, VTs, Ops);
11075}
11076
11077SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
11078 EVT VT1, EVT VT2, ArrayRef<SDValue> Ops) {
11079 SDVTList VTs = getVTList(VT1, VT2);
11080 return SelectNodeTo(N, MachineOpc, VTs, Ops);
11081}
11082
11083SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
11084 EVT VT1, EVT VT2) {
11085 SDVTList VTs = getVTList(VT1, VT2);
11086 return SelectNodeTo(N, MachineOpc, VTs, Ops: {});
11087}
11088
11089SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
11090 EVT VT1, EVT VT2, EVT VT3,
11091 ArrayRef<SDValue> Ops) {
11092 SDVTList VTs = getVTList(VT1, VT2, VT3);
11093 return SelectNodeTo(N, MachineOpc, VTs, Ops);
11094}
11095
11096SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
11097 EVT VT1, EVT VT2,
11098 SDValue Op1, SDValue Op2) {
11099 SDVTList VTs = getVTList(VT1, VT2);
11100 SDValue Ops[] = { Op1, Op2 };
11101 return SelectNodeTo(N, MachineOpc, VTs, Ops);
11102}
11103
11104SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
11105 SDVTList VTs,ArrayRef<SDValue> Ops) {
11106 SDNode *New = MorphNodeTo(N, Opc: ~MachineOpc, VTs, Ops);
11107 // Reset the NodeID to -1.
11108 New->setNodeId(-1);
11109 if (New != N) {
11110 ReplaceAllUsesWith(From: N, To: New);
11111 RemoveDeadNode(N);
11112 }
11113 return New;
11114}
11115
11116/// UpdateSDLocOnMergeSDNode - If the opt level is -O0 then it throws away
11117/// the line number information on the merged node since it is not possible to
11118/// preserve the information that operation is associated with multiple lines.
11119/// This will make the debugger working better at -O0, were there is a higher
11120/// probability having other instructions associated with that line.
11121///
11122/// For IROrder, we keep the smaller of the two
11123SDNode *SelectionDAG::UpdateSDLocOnMergeSDNode(SDNode *N, const SDLoc &OLoc) {
11124 DebugLoc NLoc = N->getDebugLoc();
11125 if (NLoc && OptLevel == CodeGenOptLevel::None && OLoc.getDebugLoc() != NLoc) {
11126 N->setDebugLoc(DebugLoc());
11127 }
11128 unsigned Order = std::min(a: N->getIROrder(), b: OLoc.getIROrder());
11129 N->setIROrder(Order);
11130 return N;
11131}
11132
11133/// MorphNodeTo - This *mutates* the specified node to have the specified
11134/// return type, opcode, and operands.
11135///
11136/// Note that MorphNodeTo returns the resultant node. If there is already a
11137/// node of the specified opcode and operands, it returns that node instead of
11138/// the current one. Note that the SDLoc need not be the same.
11139///
11140/// Using MorphNodeTo is faster than creating a new node and swapping it in
11141/// with ReplaceAllUsesWith both because it often avoids allocating a new
11142/// node, and because it doesn't require CSE recalculation for any of
11143/// the node's users.
11144///
11145/// However, note that MorphNodeTo recursively deletes dead nodes from the DAG.
11146/// As a consequence it isn't appropriate to use from within the DAG combiner or
11147/// the legalizer which maintain worklists that would need to be updated when
11148/// deleting things.
11149SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
11150 SDVTList VTs, ArrayRef<SDValue> Ops) {
11151 // If an identical node already exists, use it.
11152 void *IP = nullptr;
11153 if (VTs.VTs[VTs.NumVTs-1] != MVT::Glue) {
11154 FoldingSetNodeID ID;
11155 AddNodeIDNode(ID, OpC: Opc, VTList: VTs, OpList: Ops);
11156 if (SDNode *ON = FindNodeOrInsertPos(ID, DL: SDLoc(N), InsertPos&: IP))
11157 return UpdateSDLocOnMergeSDNode(N: ON, OLoc: SDLoc(N));
11158 }
11159
11160 if (!RemoveNodeFromCSEMaps(N))
11161 IP = nullptr;
11162
11163 // Start the morphing.
11164 N->NodeType = Opc;
11165 N->ValueList = VTs.VTs;
11166 N->NumValues = VTs.NumVTs;
11167
11168 // Clear the operands list, updating used nodes to remove this from their
11169 // use list. Keep track of any operands that become dead as a result.
11170 SmallPtrSet<SDNode*, 16> DeadNodeSet;
11171 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) {
11172 SDUse &Use = *I++;
11173 SDNode *Used = Use.getNode();
11174 Use.set(SDValue());
11175 if (Used->use_empty())
11176 DeadNodeSet.insert(Ptr: Used);
11177 }
11178
11179 // For MachineNode, initialize the memory references information.
11180 if (MachineSDNode *MN = dyn_cast<MachineSDNode>(Val: N))
11181 MN->clearMemRefs();
11182
11183 // Swap for an appropriately sized array from the recycler.
11184 removeOperands(Node: N);
11185 createOperands(Node: N, Vals: Ops);
11186
11187 // Delete any nodes that are still dead after adding the uses for the
11188 // new operands.
11189 if (!DeadNodeSet.empty()) {
11190 SmallVector<SDNode *, 16> DeadNodes;
11191 for (SDNode *N : DeadNodeSet)
11192 if (N->use_empty())
11193 DeadNodes.push_back(Elt: N);
11194 RemoveDeadNodes(DeadNodes);
11195 }
11196
11197 if (IP)
11198 CSEMap.InsertNode(N, InsertPos: IP); // Memoize the new node.
11199 return N;
11200}
11201
11202SDNode* SelectionDAG::mutateStrictFPToFP(SDNode *Node) {
11203 unsigned OrigOpc = Node->getOpcode();
11204 unsigned NewOpc;
11205 switch (OrigOpc) {
11206 default:
11207 llvm_unreachable("mutateStrictFPToFP called with unexpected opcode!");
11208#define DAG_INSTRUCTION(NAME, NARG, ROUND_MODE, INTRINSIC, DAGN) \
11209 case ISD::STRICT_##DAGN: NewOpc = ISD::DAGN; break;
11210#define CMP_INSTRUCTION(NAME, NARG, ROUND_MODE, INTRINSIC, DAGN) \
11211 case ISD::STRICT_##DAGN: NewOpc = ISD::SETCC; break;
11212#include "llvm/IR/ConstrainedOps.def"
11213 }
11214
11215 assert(Node->getNumValues() == 2 && "Unexpected number of results!");
11216
11217 // We're taking this node out of the chain, so we need to re-link things.
11218 SDValue InputChain = Node->getOperand(Num: 0);
11219 SDValue OutputChain = SDValue(Node, 1);
11220 ReplaceAllUsesOfValueWith(From: OutputChain, To: InputChain);
11221
11222 SmallVector<SDValue, 3> Ops;
11223 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i)
11224 Ops.push_back(Elt: Node->getOperand(Num: i));
11225
11226 SDVTList VTs = getVTList(VT: Node->getValueType(ResNo: 0));
11227 SDNode *Res = MorphNodeTo(N: Node, Opc: NewOpc, VTs, Ops);
11228
11229 // MorphNodeTo can operate in two ways: if an existing node with the
11230 // specified operands exists, it can just return it. Otherwise, it
11231 // updates the node in place to have the requested operands.
11232 if (Res == Node) {
11233 // If we updated the node in place, reset the node ID. To the isel,
11234 // this should be just like a newly allocated machine node.
11235 Res->setNodeId(-1);
11236 } else {
11237 ReplaceAllUsesWith(From: Node, To: Res);
11238 RemoveDeadNode(N: Node);
11239 }
11240
11241 return Res;
11242}
11243
11244/// getMachineNode - These are used for target selectors to create a new node
11245/// with specified return type(s), MachineInstr opcode, and operands.
11246///
11247/// Note that getMachineNode returns the resultant node. If there is already a
11248/// node of the specified opcode and operands, it returns that node instead of
11249/// the current one.
11250MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
11251 EVT VT) {
11252 SDVTList VTs = getVTList(VT);
11253 return getMachineNode(Opcode, dl, VTs, Ops: {});
11254}
11255
11256MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
11257 EVT VT, SDValue Op1) {
11258 SDVTList VTs = getVTList(VT);
11259 SDValue Ops[] = { Op1 };
11260 return getMachineNode(Opcode, dl, VTs, Ops);
11261}
11262
11263MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
11264 EVT VT, SDValue Op1, SDValue Op2) {
11265 SDVTList VTs = getVTList(VT);
11266 SDValue Ops[] = { Op1, Op2 };
11267 return getMachineNode(Opcode, dl, VTs, Ops);
11268}
11269
11270MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
11271 EVT VT, SDValue Op1, SDValue Op2,
11272 SDValue Op3) {
11273 SDVTList VTs = getVTList(VT);
11274 SDValue Ops[] = { Op1, Op2, Op3 };
11275 return getMachineNode(Opcode, dl, VTs, Ops);
11276}
11277
11278MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
11279 EVT VT, ArrayRef<SDValue> Ops) {
11280 SDVTList VTs = getVTList(VT);
11281 return getMachineNode(Opcode, dl, VTs, Ops);
11282}
11283
11284MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
11285 EVT VT1, EVT VT2, SDValue Op1,
11286 SDValue Op2) {
11287 SDVTList VTs = getVTList(VT1, VT2);
11288 SDValue Ops[] = { Op1, Op2 };
11289 return getMachineNode(Opcode, dl, VTs, Ops);
11290}
11291
11292MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
11293 EVT VT1, EVT VT2, SDValue Op1,
11294 SDValue Op2, SDValue Op3) {
11295 SDVTList VTs = getVTList(VT1, VT2);
11296 SDValue Ops[] = { Op1, Op2, Op3 };
11297 return getMachineNode(Opcode, dl, VTs, Ops);
11298}
11299
11300MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
11301 EVT VT1, EVT VT2,
11302 ArrayRef<SDValue> Ops) {
11303 SDVTList VTs = getVTList(VT1, VT2);
11304 return getMachineNode(Opcode, dl, VTs, Ops);
11305}
11306
11307MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
11308 EVT VT1, EVT VT2, EVT VT3,
11309 SDValue Op1, SDValue Op2) {
11310 SDVTList VTs = getVTList(VT1, VT2, VT3);
11311 SDValue Ops[] = { Op1, Op2 };
11312 return getMachineNode(Opcode, dl, VTs, Ops);
11313}
11314
11315MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
11316 EVT VT1, EVT VT2, EVT VT3,
11317 SDValue Op1, SDValue Op2,
11318 SDValue Op3) {
11319 SDVTList VTs = getVTList(VT1, VT2, VT3);
11320 SDValue Ops[] = { Op1, Op2, Op3 };
11321 return getMachineNode(Opcode, dl, VTs, Ops);
11322}
11323
11324MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
11325 EVT VT1, EVT VT2, EVT VT3,
11326 ArrayRef<SDValue> Ops) {
11327 SDVTList VTs = getVTList(VT1, VT2, VT3);
11328 return getMachineNode(Opcode, dl, VTs, Ops);
11329}
11330
11331MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
11332 ArrayRef<EVT> ResultTys,
11333 ArrayRef<SDValue> Ops) {
11334 SDVTList VTs = getVTList(VTs: ResultTys);
11335 return getMachineNode(Opcode, dl, VTs, Ops);
11336}
11337
11338MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &DL,
11339 SDVTList VTs,
11340 ArrayRef<SDValue> Ops) {
11341 bool DoCSE = VTs.VTs[VTs.NumVTs-1] != MVT::Glue;
11342 MachineSDNode *N;
11343 void *IP = nullptr;
11344
11345 if (DoCSE) {
11346 FoldingSetNodeID ID;
11347 AddNodeIDNode(ID, OpC: ~Opcode, VTList: VTs, OpList: Ops);
11348 IP = nullptr;
11349 if (SDNode *E = FindNodeOrInsertPos(ID, DL, InsertPos&: IP)) {
11350 return cast<MachineSDNode>(Val: UpdateSDLocOnMergeSDNode(N: E, OLoc: DL));
11351 }
11352 }
11353
11354 // Allocate a new MachineSDNode.
11355 N = newSDNode<MachineSDNode>(Args: ~Opcode, Args: DL.getIROrder(), Args: DL.getDebugLoc(), Args&: VTs);
11356 createOperands(Node: N, Vals: Ops);
11357
11358 if (DoCSE)
11359 CSEMap.InsertNode(N, InsertPos: IP);
11360
11361 InsertNode(N);
11362 NewSDValueDbgMsg(V: SDValue(N, 0), Msg: "Creating new machine node: ", G: this);
11363 return N;
11364}
11365
11366/// getTargetExtractSubreg - A convenience function for creating
11367/// TargetOpcode::EXTRACT_SUBREG nodes.
11368SDValue SelectionDAG::getTargetExtractSubreg(int SRIdx, const SDLoc &DL, EVT VT,
11369 SDValue Operand) {
11370 SDValue SRIdxVal = getTargetConstant(SRIdx, DL, MVT::i32);
11371 SDNode *Subreg = getMachineNode(Opcode: TargetOpcode::EXTRACT_SUBREG, dl: DL,
11372 VT, Op1: Operand, Op2: SRIdxVal);
11373 return SDValue(Subreg, 0);
11374}
11375
11376/// getTargetInsertSubreg - A convenience function for creating
11377/// TargetOpcode::INSERT_SUBREG nodes.
11378SDValue SelectionDAG::getTargetInsertSubreg(int SRIdx, const SDLoc &DL, EVT VT,
11379 SDValue Operand, SDValue Subreg) {
11380 SDValue SRIdxVal = getTargetConstant(SRIdx, DL, MVT::i32);
11381 SDNode *Result = getMachineNode(Opcode: TargetOpcode::INSERT_SUBREG, dl: DL,
11382 VT, Op1: Operand, Op2: Subreg, Op3: SRIdxVal);
11383 return SDValue(Result, 0);
11384}
11385
11386/// getNodeIfExists - Get the specified node if it's already available, or
11387/// else return NULL.
11388SDNode *SelectionDAG::getNodeIfExists(unsigned Opcode, SDVTList VTList,
11389 ArrayRef<SDValue> Ops) {
11390 SDNodeFlags Flags;
11391 if (Inserter)
11392 Flags = Inserter->getFlags();
11393 return getNodeIfExists(Opcode, VTList, Ops, Flags);
11394}
11395
11396SDNode *SelectionDAG::getNodeIfExists(unsigned Opcode, SDVTList VTList,
11397 ArrayRef<SDValue> Ops,
11398 const SDNodeFlags Flags) {
11399 if (VTList.VTs[VTList.NumVTs - 1] != MVT::Glue) {
11400 FoldingSetNodeID ID;
11401 AddNodeIDNode(ID, OpC: Opcode, VTList, OpList: Ops);
11402 void *IP = nullptr;
11403 if (SDNode *E = FindNodeOrInsertPos(ID, DL: SDLoc(), InsertPos&: IP)) {
11404 E->intersectFlagsWith(Flags);
11405 return E;
11406 }
11407 }
11408 return nullptr;
11409}
11410
11411/// doesNodeExist - Check if a node exists without modifying its flags.
11412bool SelectionDAG::doesNodeExist(unsigned Opcode, SDVTList VTList,
11413 ArrayRef<SDValue> Ops) {
11414 if (VTList.VTs[VTList.NumVTs - 1] != MVT::Glue) {
11415 FoldingSetNodeID ID;
11416 AddNodeIDNode(ID, OpC: Opcode, VTList, OpList: Ops);
11417 void *IP = nullptr;
11418 if (FindNodeOrInsertPos(ID, DL: SDLoc(), InsertPos&: IP))
11419 return true;
11420 }
11421 return false;
11422}
11423
11424/// getDbgValue - Creates a SDDbgValue node.
11425///
11426/// SDNode
11427SDDbgValue *SelectionDAG::getDbgValue(DIVariable *Var, DIExpression *Expr,
11428 SDNode *N, unsigned R, bool IsIndirect,
11429 const DebugLoc &DL, unsigned O) {
11430 assert(cast<DILocalVariable>(Var)->isValidLocationForIntrinsic(DL) &&
11431 "Expected inlined-at fields to agree");
11432 return new (DbgInfo->getAlloc())
11433 SDDbgValue(DbgInfo->getAlloc(), Var, Expr, SDDbgOperand::fromNode(Node: N, ResNo: R),
11434 {}, IsIndirect, DL, O,
11435 /*IsVariadic=*/false);
11436}
11437
11438/// Constant
11439SDDbgValue *SelectionDAG::getConstantDbgValue(DIVariable *Var,
11440 DIExpression *Expr,
11441 const Value *C,
11442 const DebugLoc &DL, unsigned O) {
11443 assert(cast<DILocalVariable>(Var)->isValidLocationForIntrinsic(DL) &&
11444 "Expected inlined-at fields to agree");
11445 return new (DbgInfo->getAlloc())
11446 SDDbgValue(DbgInfo->getAlloc(), Var, Expr, SDDbgOperand::fromConst(Const: C), {},
11447 /*IsIndirect=*/false, DL, O,
11448 /*IsVariadic=*/false);
11449}
11450
11451/// FrameIndex
11452SDDbgValue *SelectionDAG::getFrameIndexDbgValue(DIVariable *Var,
11453 DIExpression *Expr, unsigned FI,
11454 bool IsIndirect,
11455 const DebugLoc &DL,
11456 unsigned O) {
11457 assert(cast<DILocalVariable>(Var)->isValidLocationForIntrinsic(DL) &&
11458 "Expected inlined-at fields to agree");
11459 return getFrameIndexDbgValue(Var, Expr, FI, Dependencies: {}, IsIndirect, DL, O);
11460}
11461
11462/// FrameIndex with dependencies
11463SDDbgValue *SelectionDAG::getFrameIndexDbgValue(DIVariable *Var,
11464 DIExpression *Expr, unsigned FI,
11465 ArrayRef<SDNode *> Dependencies,
11466 bool IsIndirect,
11467 const DebugLoc &DL,
11468 unsigned O) {
11469 assert(cast<DILocalVariable>(Var)->isValidLocationForIntrinsic(DL) &&
11470 "Expected inlined-at fields to agree");
11471 return new (DbgInfo->getAlloc())
11472 SDDbgValue(DbgInfo->getAlloc(), Var, Expr, SDDbgOperand::fromFrameIdx(FrameIdx: FI),
11473 Dependencies, IsIndirect, DL, O,
11474 /*IsVariadic=*/false);
11475}
11476
11477/// VReg
11478SDDbgValue *SelectionDAG::getVRegDbgValue(DIVariable *Var, DIExpression *Expr,
11479 Register VReg, bool IsIndirect,
11480 const DebugLoc &DL, unsigned O) {
11481 assert(cast<DILocalVariable>(Var)->isValidLocationForIntrinsic(DL) &&
11482 "Expected inlined-at fields to agree");
11483 return new (DbgInfo->getAlloc())
11484 SDDbgValue(DbgInfo->getAlloc(), Var, Expr, SDDbgOperand::fromVReg(VReg),
11485 {}, IsIndirect, DL, O,
11486 /*IsVariadic=*/false);
11487}
11488
11489SDDbgValue *SelectionDAG::getDbgValueList(DIVariable *Var, DIExpression *Expr,
11490 ArrayRef<SDDbgOperand> Locs,
11491 ArrayRef<SDNode *> Dependencies,
11492 bool IsIndirect, const DebugLoc &DL,
11493 unsigned O, bool IsVariadic) {
11494 assert(cast<DILocalVariable>(Var)->isValidLocationForIntrinsic(DL) &&
11495 "Expected inlined-at fields to agree");
11496 return new (DbgInfo->getAlloc())
11497 SDDbgValue(DbgInfo->getAlloc(), Var, Expr, Locs, Dependencies, IsIndirect,
11498 DL, O, IsVariadic);
11499}
11500
11501void SelectionDAG::transferDbgValues(SDValue From, SDValue To,
11502 unsigned OffsetInBits, unsigned SizeInBits,
11503 bool InvalidateDbg) {
11504 SDNode *FromNode = From.getNode();
11505 SDNode *ToNode = To.getNode();
11506 assert(FromNode && ToNode && "Can't modify dbg values");
11507
11508 // PR35338
11509 // TODO: assert(From != To && "Redundant dbg value transfer");
11510 // TODO: assert(FromNode != ToNode && "Intranode dbg value transfer");
11511 if (From == To || FromNode == ToNode)
11512 return;
11513
11514 if (!FromNode->getHasDebugValue())
11515 return;
11516
11517 SDDbgOperand FromLocOp =
11518 SDDbgOperand::fromNode(Node: From.getNode(), ResNo: From.getResNo());
11519 SDDbgOperand ToLocOp = SDDbgOperand::fromNode(Node: To.getNode(), ResNo: To.getResNo());
11520
11521 SmallVector<SDDbgValue *, 2> ClonedDVs;
11522 for (SDDbgValue *Dbg : GetDbgValues(SD: FromNode)) {
11523 if (Dbg->isInvalidated())
11524 continue;
11525
11526 // TODO: assert(!Dbg->isInvalidated() && "Transfer of invalid dbg value");
11527
11528 // Create a new location ops vector that is equal to the old vector, but
11529 // with each instance of FromLocOp replaced with ToLocOp.
11530 bool Changed = false;
11531 auto NewLocOps = Dbg->copyLocationOps();
11532 std::replace_if(
11533 first: NewLocOps.begin(), last: NewLocOps.end(),
11534 pred: [&Changed, FromLocOp](const SDDbgOperand &Op) {
11535 bool Match = Op == FromLocOp;
11536 Changed |= Match;
11537 return Match;
11538 },
11539 new_value: ToLocOp);
11540 // Ignore this SDDbgValue if we didn't find a matching location.
11541 if (!Changed)
11542 continue;
11543
11544 DIVariable *Var = Dbg->getVariable();
11545 auto *Expr = Dbg->getExpression();
11546 // If a fragment is requested, update the expression.
11547 if (SizeInBits) {
11548 // When splitting a larger (e.g., sign-extended) value whose
11549 // lower bits are described with an SDDbgValue, do not attempt
11550 // to transfer the SDDbgValue to the upper bits.
11551 if (auto FI = Expr->getFragmentInfo())
11552 if (OffsetInBits + SizeInBits > FI->SizeInBits)
11553 continue;
11554 auto Fragment = DIExpression::createFragmentExpression(Expr, OffsetInBits,
11555 SizeInBits);
11556 if (!Fragment)
11557 continue;
11558 Expr = *Fragment;
11559 }
11560
11561 auto AdditionalDependencies = Dbg->getAdditionalDependencies();
11562 // Clone the SDDbgValue and move it to To.
11563 SDDbgValue *Clone = getDbgValueList(
11564 Var, Expr, Locs: NewLocOps, Dependencies: AdditionalDependencies, IsIndirect: Dbg->isIndirect(),
11565 DL: Dbg->getDebugLoc(), O: std::max(a: ToNode->getIROrder(), b: Dbg->getOrder()),
11566 IsVariadic: Dbg->isVariadic());
11567 ClonedDVs.push_back(Elt: Clone);
11568
11569 if (InvalidateDbg) {
11570 // Invalidate value and indicate the SDDbgValue should not be emitted.
11571 Dbg->setIsInvalidated();
11572 Dbg->setIsEmitted();
11573 }
11574 }
11575
11576 for (SDDbgValue *Dbg : ClonedDVs) {
11577 assert(is_contained(Dbg->getSDNodes(), ToNode) &&
11578 "Transferred DbgValues should depend on the new SDNode");
11579 AddDbgValue(DB: Dbg, isParameter: false);
11580 }
11581}
11582
11583void SelectionDAG::salvageDebugInfo(SDNode &N) {
11584 if (!N.getHasDebugValue())
11585 return;
11586
11587 auto GetLocationOperand = [](SDNode *Node, unsigned ResNo) {
11588 if (auto *FISDN = dyn_cast<FrameIndexSDNode>(Val: Node))
11589 return SDDbgOperand::fromFrameIdx(FrameIdx: FISDN->getIndex());
11590 return SDDbgOperand::fromNode(Node, ResNo);
11591 };
11592
11593 SmallVector<SDDbgValue *, 2> ClonedDVs;
11594 for (auto *DV : GetDbgValues(SD: &N)) {
11595 if (DV->isInvalidated())
11596 continue;
11597 switch (N.getOpcode()) {
11598 default:
11599 break;
11600 case ISD::ADD: {
11601 SDValue N0 = N.getOperand(Num: 0);
11602 SDValue N1 = N.getOperand(Num: 1);
11603 if (!isa<ConstantSDNode>(Val: N0)) {
11604 bool RHSConstant = isa<ConstantSDNode>(Val: N1);
11605 uint64_t Offset;
11606 if (RHSConstant)
11607 Offset = N.getConstantOperandVal(Num: 1);
11608 // We are not allowed to turn indirect debug values variadic, so
11609 // don't salvage those.
11610 if (!RHSConstant && DV->isIndirect())
11611 continue;
11612
11613 // Rewrite an ADD constant node into a DIExpression. Since we are
11614 // performing arithmetic to compute the variable's *value* in the
11615 // DIExpression, we need to mark the expression with a
11616 // DW_OP_stack_value.
11617 auto *DIExpr = DV->getExpression();
11618 auto NewLocOps = DV->copyLocationOps();
11619 bool Changed = false;
11620 size_t OrigLocOpsSize = NewLocOps.size();
11621 for (size_t i = 0; i < OrigLocOpsSize; ++i) {
11622 // We're not given a ResNo to compare against because the whole
11623 // node is going away. We know that any ISD::ADD only has one
11624 // result, so we can assume any node match is using the result.
11625 if (NewLocOps[i].getKind() != SDDbgOperand::SDNODE ||
11626 NewLocOps[i].getSDNode() != &N)
11627 continue;
11628 NewLocOps[i] = GetLocationOperand(N0.getNode(), N0.getResNo());
11629 if (RHSConstant) {
11630 SmallVector<uint64_t, 3> ExprOps;
11631 DIExpression::appendOffset(Ops&: ExprOps, Offset);
11632 DIExpr = DIExpression::appendOpsToArg(Expr: DIExpr, Ops: ExprOps, ArgNo: i, StackValue: true);
11633 } else {
11634 // Convert to a variadic expression (if not already).
11635 // convertToVariadicExpression() returns a const pointer, so we use
11636 // a temporary const variable here.
11637 const auto *TmpDIExpr =
11638 DIExpression::convertToVariadicExpression(Expr: DIExpr);
11639 SmallVector<uint64_t, 3> ExprOps;
11640 ExprOps.push_back(Elt: dwarf::DW_OP_LLVM_arg);
11641 ExprOps.push_back(Elt: NewLocOps.size());
11642 ExprOps.push_back(Elt: dwarf::DW_OP_plus);
11643 SDDbgOperand RHS =
11644 SDDbgOperand::fromNode(Node: N1.getNode(), ResNo: N1.getResNo());
11645 NewLocOps.push_back(Elt: RHS);
11646 DIExpr = DIExpression::appendOpsToArg(Expr: TmpDIExpr, Ops: ExprOps, ArgNo: i, StackValue: true);
11647 }
11648 Changed = true;
11649 }
11650 (void)Changed;
11651 assert(Changed && "Salvage target doesn't use N");
11652
11653 bool IsVariadic =
11654 DV->isVariadic() || OrigLocOpsSize != NewLocOps.size();
11655
11656 auto AdditionalDependencies = DV->getAdditionalDependencies();
11657 SDDbgValue *Clone = getDbgValueList(
11658 Var: DV->getVariable(), Expr: DIExpr, Locs: NewLocOps, Dependencies: AdditionalDependencies,
11659 IsIndirect: DV->isIndirect(), DL: DV->getDebugLoc(), O: DV->getOrder(), IsVariadic);
11660 ClonedDVs.push_back(Elt: Clone);
11661 DV->setIsInvalidated();
11662 DV->setIsEmitted();
11663 LLVM_DEBUG(dbgs() << "SALVAGE: Rewriting";
11664 N0.getNode()->dumprFull(this);
11665 dbgs() << " into " << *DIExpr << '\n');
11666 }
11667 break;
11668 }
11669 case ISD::TRUNCATE: {
11670 SDValue N0 = N.getOperand(Num: 0);
11671 TypeSize FromSize = N0.getValueSizeInBits();
11672 TypeSize ToSize = N.getValueSizeInBits(ResNo: 0);
11673
11674 DIExpression *DbgExpression = DV->getExpression();
11675 auto ExtOps = DIExpression::getExtOps(FromSize, ToSize, Signed: false);
11676 auto NewLocOps = DV->copyLocationOps();
11677 bool Changed = false;
11678 for (size_t i = 0; i < NewLocOps.size(); ++i) {
11679 if (NewLocOps[i].getKind() != SDDbgOperand::SDNODE ||
11680 NewLocOps[i].getSDNode() != &N)
11681 continue;
11682
11683 NewLocOps[i] = GetLocationOperand(N0.getNode(), N0.getResNo());
11684 DbgExpression = DIExpression::appendOpsToArg(Expr: DbgExpression, Ops: ExtOps, ArgNo: i);
11685 Changed = true;
11686 }
11687 assert(Changed && "Salvage target doesn't use N");
11688 (void)Changed;
11689
11690 SDDbgValue *Clone =
11691 getDbgValueList(Var: DV->getVariable(), Expr: DbgExpression, Locs: NewLocOps,
11692 Dependencies: DV->getAdditionalDependencies(), IsIndirect: DV->isIndirect(),
11693 DL: DV->getDebugLoc(), O: DV->getOrder(), IsVariadic: DV->isVariadic());
11694
11695 ClonedDVs.push_back(Elt: Clone);
11696 DV->setIsInvalidated();
11697 DV->setIsEmitted();
11698 LLVM_DEBUG(dbgs() << "SALVAGE: Rewriting"; N0.getNode()->dumprFull(this);
11699 dbgs() << " into " << *DbgExpression << '\n');
11700 break;
11701 }
11702 }
11703 }
11704
11705 for (SDDbgValue *Dbg : ClonedDVs) {
11706 assert((!Dbg->getSDNodes().empty() ||
11707 llvm::any_of(Dbg->getLocationOps(),
11708 [&](const SDDbgOperand &Op) {
11709 return Op.getKind() == SDDbgOperand::FRAMEIX;
11710 })) &&
11711 "Salvaged DbgValue should depend on a new SDNode");
11712 AddDbgValue(DB: Dbg, isParameter: false);
11713 }
11714}
11715
11716/// Creates a SDDbgLabel node.
11717SDDbgLabel *SelectionDAG::getDbgLabel(DILabel *Label,
11718 const DebugLoc &DL, unsigned O) {
11719 assert(cast<DILabel>(Label)->isValidLocationForIntrinsic(DL) &&
11720 "Expected inlined-at fields to agree");
11721 return new (DbgInfo->getAlloc()) SDDbgLabel(Label, DL, O);
11722}
11723
11724namespace {
11725
11726/// RAUWUpdateListener - Helper for ReplaceAllUsesWith - When the node
11727/// pointed to by a use iterator is deleted, increment the use iterator
11728/// so that it doesn't dangle.
11729///
11730class RAUWUpdateListener : public SelectionDAG::DAGUpdateListener {
11731 SDNode::use_iterator &UI;
11732 SDNode::use_iterator &UE;
11733
11734 void NodeDeleted(SDNode *N, SDNode *E) override {
11735 // Increment the iterator as needed.
11736 while (UI != UE && N == UI->getUser())
11737 ++UI;
11738 }
11739
11740public:
11741 RAUWUpdateListener(SelectionDAG &d,
11742 SDNode::use_iterator &ui,
11743 SDNode::use_iterator &ue)
11744 : SelectionDAG::DAGUpdateListener(d), UI(ui), UE(ue) {}
11745};
11746
11747} // end anonymous namespace
11748
11749/// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
11750/// This can cause recursive merging of nodes in the DAG.
11751///
11752/// This version assumes From has a single result value.
11753///
11754void SelectionDAG::ReplaceAllUsesWith(SDValue FromN, SDValue To) {
11755 SDNode *From = FromN.getNode();
11756 assert(From->getNumValues() == 1 && FromN.getResNo() == 0 &&
11757 "Cannot replace with this method!");
11758 assert(From != To.getNode() && "Cannot replace uses of with self");
11759
11760 // Preserve Debug Values
11761 transferDbgValues(From: FromN, To);
11762 // Preserve extra info.
11763 copyExtraInfo(From, To: To.getNode());
11764
11765 // Iterate over all the existing uses of From. New uses will be added
11766 // to the beginning of the use list, which we avoid visiting.
11767 // This specifically avoids visiting uses of From that arise while the
11768 // replacement is happening, because any such uses would be the result
11769 // of CSE: If an existing node looks like From after one of its operands
11770 // is replaced by To, we don't want to replace of all its users with To
11771 // too. See PR3018 for more info.
11772 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
11773 RAUWUpdateListener Listener(*this, UI, UE);
11774 while (UI != UE) {
11775 SDNode *User = UI->getUser();
11776
11777 // This node is about to morph, remove its old self from the CSE maps.
11778 RemoveNodeFromCSEMaps(N: User);
11779
11780 // A user can appear in a use list multiple times, and when this
11781 // happens the uses are usually next to each other in the list.
11782 // To help reduce the number of CSE recomputations, process all
11783 // the uses of this user that we can find this way.
11784 do {
11785 SDUse &Use = *UI;
11786 ++UI;
11787 Use.set(To);
11788 if (To->isDivergent() != From->isDivergent())
11789 updateDivergence(N: User);
11790 } while (UI != UE && UI->getUser() == User);
11791 // Now that we have modified User, add it back to the CSE maps. If it
11792 // already exists there, recursively merge the results together.
11793 AddModifiedNodeToCSEMaps(N: User);
11794 }
11795
11796 // If we just RAUW'd the root, take note.
11797 if (FromN == getRoot())
11798 setRoot(To);
11799}
11800
11801/// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
11802/// This can cause recursive merging of nodes in the DAG.
11803///
11804/// This version assumes that for each value of From, there is a
11805/// corresponding value in To in the same position with the same type.
11806///
11807void SelectionDAG::ReplaceAllUsesWith(SDNode *From, SDNode *To) {
11808#ifndef NDEBUG
11809 for (unsigned i = 0, e = From->getNumValues(); i != e; ++i)
11810 assert((!From->hasAnyUseOfValue(i) ||
11811 From->getValueType(i) == To->getValueType(i)) &&
11812 "Cannot use this version of ReplaceAllUsesWith!");
11813#endif
11814
11815 // Handle the trivial case.
11816 if (From == To)
11817 return;
11818
11819 // Preserve Debug Info. Only do this if there's a use.
11820 for (unsigned i = 0, e = From->getNumValues(); i != e; ++i)
11821 if (From->hasAnyUseOfValue(Value: i)) {
11822 assert((i < To->getNumValues()) && "Invalid To location");
11823 transferDbgValues(From: SDValue(From, i), To: SDValue(To, i));
11824 }
11825 // Preserve extra info.
11826 copyExtraInfo(From, To);
11827
11828 // Iterate over just the existing users of From. See the comments in
11829 // the ReplaceAllUsesWith above.
11830 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
11831 RAUWUpdateListener Listener(*this, UI, UE);
11832 while (UI != UE) {
11833 SDNode *User = UI->getUser();
11834
11835 // This node is about to morph, remove its old self from the CSE maps.
11836 RemoveNodeFromCSEMaps(N: User);
11837
11838 // A user can appear in a use list multiple times, and when this
11839 // happens the uses are usually next to each other in the list.
11840 // To help reduce the number of CSE recomputations, process all
11841 // the uses of this user that we can find this way.
11842 do {
11843 SDUse &Use = *UI;
11844 ++UI;
11845 Use.setNode(To);
11846 if (To->isDivergent() != From->isDivergent())
11847 updateDivergence(N: User);
11848 } while (UI != UE && UI->getUser() == User);
11849
11850 // Now that we have modified User, add it back to the CSE maps. If it
11851 // already exists there, recursively merge the results together.
11852 AddModifiedNodeToCSEMaps(N: User);
11853 }
11854
11855 // If we just RAUW'd the root, take note.
11856 if (From == getRoot().getNode())
11857 setRoot(SDValue(To, getRoot().getResNo()));
11858}
11859
11860/// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
11861/// This can cause recursive merging of nodes in the DAG.
11862///
11863/// This version can replace From with any result values. To must match the
11864/// number and types of values returned by From.
11865void SelectionDAG::ReplaceAllUsesWith(SDNode *From, const SDValue *To) {
11866 if (From->getNumValues() == 1) // Handle the simple case efficiently.
11867 return ReplaceAllUsesWith(FromN: SDValue(From, 0), To: To[0]);
11868
11869 for (unsigned i = 0, e = From->getNumValues(); i != e; ++i) {
11870 // Preserve Debug Info.
11871 transferDbgValues(From: SDValue(From, i), To: To[i]);
11872 // Preserve extra info.
11873 copyExtraInfo(From, To: To[i].getNode());
11874 }
11875
11876 // Iterate over just the existing users of From. See the comments in
11877 // the ReplaceAllUsesWith above.
11878 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
11879 RAUWUpdateListener Listener(*this, UI, UE);
11880 while (UI != UE) {
11881 SDNode *User = UI->getUser();
11882
11883 // This node is about to morph, remove its old self from the CSE maps.
11884 RemoveNodeFromCSEMaps(N: User);
11885
11886 // A user can appear in a use list multiple times, and when this happens the
11887 // uses are usually next to each other in the list. To help reduce the
11888 // number of CSE and divergence recomputations, process all the uses of this
11889 // user that we can find this way.
11890 bool To_IsDivergent = false;
11891 do {
11892 SDUse &Use = *UI;
11893 const SDValue &ToOp = To[Use.getResNo()];
11894 ++UI;
11895 Use.set(ToOp);
11896 To_IsDivergent |= ToOp->isDivergent();
11897 } while (UI != UE && UI->getUser() == User);
11898
11899 if (To_IsDivergent != From->isDivergent())
11900 updateDivergence(N: User);
11901
11902 // Now that we have modified User, add it back to the CSE maps. If it
11903 // already exists there, recursively merge the results together.
11904 AddModifiedNodeToCSEMaps(N: User);
11905 }
11906
11907 // If we just RAUW'd the root, take note.
11908 if (From == getRoot().getNode())
11909 setRoot(SDValue(To[getRoot().getResNo()]));
11910}
11911
11912/// ReplaceAllUsesOfValueWith - Replace any uses of From with To, leaving
11913/// uses of other values produced by From.getNode() alone. The Deleted
11914/// vector is handled the same way as for ReplaceAllUsesWith.
11915void SelectionDAG::ReplaceAllUsesOfValueWith(SDValue From, SDValue To){
11916 // Handle the really simple, really trivial case efficiently.
11917 if (From == To) return;
11918
11919 // Handle the simple, trivial, case efficiently.
11920 if (From.getNode()->getNumValues() == 1) {
11921 ReplaceAllUsesWith(FromN: From, To);
11922 return;
11923 }
11924
11925 // Preserve Debug Info.
11926 transferDbgValues(From, To);
11927 copyExtraInfo(From: From.getNode(), To: To.getNode());
11928
11929 // Iterate over just the existing users of From. See the comments in
11930 // the ReplaceAllUsesWith above.
11931 SDNode::use_iterator UI = From.getNode()->use_begin(),
11932 UE = From.getNode()->use_end();
11933 RAUWUpdateListener Listener(*this, UI, UE);
11934 while (UI != UE) {
11935 SDNode *User = UI->getUser();
11936 bool UserRemovedFromCSEMaps = false;
11937
11938 // A user can appear in a use list multiple times, and when this
11939 // happens the uses are usually next to each other in the list.
11940 // To help reduce the number of CSE recomputations, process all
11941 // the uses of this user that we can find this way.
11942 do {
11943 SDUse &Use = *UI;
11944
11945 // Skip uses of different values from the same node.
11946 if (Use.getResNo() != From.getResNo()) {
11947 ++UI;
11948 continue;
11949 }
11950
11951 // If this node hasn't been modified yet, it's still in the CSE maps,
11952 // so remove its old self from the CSE maps.
11953 if (!UserRemovedFromCSEMaps) {
11954 RemoveNodeFromCSEMaps(N: User);
11955 UserRemovedFromCSEMaps = true;
11956 }
11957
11958 ++UI;
11959 Use.set(To);
11960 if (To->isDivergent() != From->isDivergent())
11961 updateDivergence(N: User);
11962 } while (UI != UE && UI->getUser() == User);
11963 // We are iterating over all uses of the From node, so if a use
11964 // doesn't use the specific value, no changes are made.
11965 if (!UserRemovedFromCSEMaps)
11966 continue;
11967
11968 // Now that we have modified User, add it back to the CSE maps. If it
11969 // already exists there, recursively merge the results together.
11970 AddModifiedNodeToCSEMaps(N: User);
11971 }
11972
11973 // If we just RAUW'd the root, take note.
11974 if (From == getRoot())
11975 setRoot(To);
11976}
11977
11978namespace {
11979
11980/// UseMemo - This class is used by SelectionDAG::ReplaceAllUsesOfValuesWith
11981/// to record information about a use.
11982struct UseMemo {
11983 SDNode *User;
11984 unsigned Index;
11985 SDUse *Use;
11986};
11987
11988/// operator< - Sort Memos by User.
11989bool operator<(const UseMemo &L, const UseMemo &R) {
11990 return (intptr_t)L.User < (intptr_t)R.User;
11991}
11992
11993/// RAUOVWUpdateListener - Helper for ReplaceAllUsesOfValuesWith - When the node
11994/// pointed to by a UseMemo is deleted, set the User to nullptr to indicate that
11995/// the node already has been taken care of recursively.
11996class RAUOVWUpdateListener : public SelectionDAG::DAGUpdateListener {
11997 SmallVectorImpl<UseMemo> &Uses;
11998
11999 void NodeDeleted(SDNode *N, SDNode *E) override {
12000 for (UseMemo &Memo : Uses)
12001 if (Memo.User == N)
12002 Memo.User = nullptr;
12003 }
12004
12005public:
12006 RAUOVWUpdateListener(SelectionDAG &d, SmallVectorImpl<UseMemo> &uses)
12007 : SelectionDAG::DAGUpdateListener(d), Uses(uses) {}
12008};
12009
12010} // end anonymous namespace
12011
12012/// Return true if a glue output should propagate divergence information.
12013static bool gluePropagatesDivergence(const SDNode *Node) {
12014 switch (Node->getOpcode()) {
12015 case ISD::CopyFromReg:
12016 case ISD::CopyToReg:
12017 return false;
12018 default:
12019 return true;
12020 }
12021
12022 llvm_unreachable("covered opcode switch");
12023}
12024
12025bool SelectionDAG::calculateDivergence(SDNode *N) {
12026 if (TLI->isSDNodeAlwaysUniform(N)) {
12027 assert(!TLI->isSDNodeSourceOfDivergence(N, FLI, UA) &&
12028 "Conflicting divergence information!");
12029 return false;
12030 }
12031 if (TLI->isSDNodeSourceOfDivergence(N, FLI, UA))
12032 return true;
12033 for (const auto &Op : N->ops()) {
12034 EVT VT = Op.getValueType();
12035
12036 // Skip Chain. It does not carry divergence.
12037 if (VT != MVT::Other && Op.getNode()->isDivergent() &&
12038 (VT != MVT::Glue || gluePropagatesDivergence(Op.getNode())))
12039 return true;
12040 }
12041 return false;
12042}
12043
12044void SelectionDAG::updateDivergence(SDNode *N) {
12045 SmallVector<SDNode *, 16> Worklist(1, N);
12046 do {
12047 N = Worklist.pop_back_val();
12048 bool IsDivergent = calculateDivergence(N);
12049 if (N->SDNodeBits.IsDivergent != IsDivergent) {
12050 N->SDNodeBits.IsDivergent = IsDivergent;
12051 llvm::append_range(C&: Worklist, R: N->users());
12052 }
12053 } while (!Worklist.empty());
12054}
12055
12056void SelectionDAG::CreateTopologicalOrder(std::vector<SDNode *> &Order) {
12057 DenseMap<SDNode *, unsigned> Degree;
12058 Order.reserve(n: AllNodes.size());
12059 for (auto &N : allnodes()) {
12060 unsigned NOps = N.getNumOperands();
12061 Degree[&N] = NOps;
12062 if (0 == NOps)
12063 Order.push_back(x: &N);
12064 }
12065 for (size_t I = 0; I != Order.size(); ++I) {
12066 SDNode *N = Order[I];
12067 for (auto *U : N->users()) {
12068 unsigned &UnsortedOps = Degree[U];
12069 if (0 == --UnsortedOps)
12070 Order.push_back(x: U);
12071 }
12072 }
12073}
12074
12075#if !defined(NDEBUG) && LLVM_ENABLE_ABI_BREAKING_CHECKS
12076void SelectionDAG::VerifyDAGDivergence() {
12077 std::vector<SDNode *> TopoOrder;
12078 CreateTopologicalOrder(Order&: TopoOrder);
12079 for (auto *N : TopoOrder) {
12080 assert(calculateDivergence(N) == N->isDivergent() &&
12081 "Divergence bit inconsistency detected");
12082 }
12083}
12084#endif
12085
12086/// ReplaceAllUsesOfValuesWith - Replace any uses of From with To, leaving
12087/// uses of other values produced by From.getNode() alone. The same value
12088/// may appear in both the From and To list. The Deleted vector is
12089/// handled the same way as for ReplaceAllUsesWith.
12090void SelectionDAG::ReplaceAllUsesOfValuesWith(const SDValue *From,
12091 const SDValue *To,
12092 unsigned Num){
12093 // Handle the simple, trivial case efficiently.
12094 if (Num == 1)
12095 return ReplaceAllUsesOfValueWith(From: *From, To: *To);
12096
12097 transferDbgValues(From: *From, To: *To);
12098 copyExtraInfo(From: From->getNode(), To: To->getNode());
12099
12100 // Read up all the uses and make records of them. This helps
12101 // processing new uses that are introduced during the
12102 // replacement process.
12103 SmallVector<UseMemo, 4> Uses;
12104 for (unsigned i = 0; i != Num; ++i) {
12105 unsigned FromResNo = From[i].getResNo();
12106 SDNode *FromNode = From[i].getNode();
12107 for (SDUse &Use : FromNode->uses()) {
12108 if (Use.getResNo() == FromResNo) {
12109 UseMemo Memo = {.User: Use.getUser(), .Index: i, .Use: &Use};
12110 Uses.push_back(Elt: Memo);
12111 }
12112 }
12113 }
12114
12115 // Sort the uses, so that all the uses from a given User are together.
12116 llvm::sort(C&: Uses);
12117 RAUOVWUpdateListener Listener(*this, Uses);
12118
12119 for (unsigned UseIndex = 0, UseIndexEnd = Uses.size();
12120 UseIndex != UseIndexEnd; ) {
12121 // We know that this user uses some value of From. If it is the right
12122 // value, update it.
12123 SDNode *User = Uses[UseIndex].User;
12124 // If the node has been deleted by recursive CSE updates when updating
12125 // another node, then just skip this entry.
12126 if (User == nullptr) {
12127 ++UseIndex;
12128 continue;
12129 }
12130
12131 // This node is about to morph, remove its old self from the CSE maps.
12132 RemoveNodeFromCSEMaps(N: User);
12133
12134 // The Uses array is sorted, so all the uses for a given User
12135 // are next to each other in the list.
12136 // To help reduce the number of CSE recomputations, process all
12137 // the uses of this user that we can find this way.
12138 do {
12139 unsigned i = Uses[UseIndex].Index;
12140 SDUse &Use = *Uses[UseIndex].Use;
12141 ++UseIndex;
12142
12143 Use.set(To[i]);
12144 } while (UseIndex != UseIndexEnd && Uses[UseIndex].User == User);
12145
12146 // Now that we have modified User, add it back to the CSE maps. If it
12147 // already exists there, recursively merge the results together.
12148 AddModifiedNodeToCSEMaps(N: User);
12149 }
12150}
12151
12152/// AssignTopologicalOrder - Assign a unique node id for each node in the DAG
12153/// based on their topological order. It returns the maximum id and a vector
12154/// of the SDNodes* in assigned order by reference.
12155unsigned SelectionDAG::AssignTopologicalOrder() {
12156 unsigned DAGSize = 0;
12157
12158 // SortedPos tracks the progress of the algorithm. Nodes before it are
12159 // sorted, nodes after it are unsorted. When the algorithm completes
12160 // it is at the end of the list.
12161 allnodes_iterator SortedPos = allnodes_begin();
12162
12163 // Visit all the nodes. Move nodes with no operands to the front of
12164 // the list immediately. Annotate nodes that do have operands with their
12165 // operand count. Before we do this, the Node Id fields of the nodes
12166 // may contain arbitrary values. After, the Node Id fields for nodes
12167 // before SortedPos will contain the topological sort index, and the
12168 // Node Id fields for nodes At SortedPos and after will contain the
12169 // count of outstanding operands.
12170 for (SDNode &N : llvm::make_early_inc_range(Range: allnodes())) {
12171 checkForCycles(N: &N, DAG: this);
12172 unsigned Degree = N.getNumOperands();
12173 if (Degree == 0) {
12174 // A node with no uses, add it to the result array immediately.
12175 N.setNodeId(DAGSize++);
12176 allnodes_iterator Q(&N);
12177 if (Q != SortedPos)
12178 SortedPos = AllNodes.insert(where: SortedPos, New: AllNodes.remove(IT&: Q));
12179 assert(SortedPos != AllNodes.end() && "Overran node list");
12180 ++SortedPos;
12181 } else {
12182 // Temporarily use the Node Id as scratch space for the degree count.
12183 N.setNodeId(Degree);
12184 }
12185 }
12186
12187 // Visit all the nodes. As we iterate, move nodes into sorted order,
12188 // such that by the time the end is reached all nodes will be sorted.
12189 for (SDNode &Node : allnodes()) {
12190 SDNode *N = &Node;
12191 checkForCycles(N, DAG: this);
12192 // N is in sorted position, so all its uses have one less operand
12193 // that needs to be sorted.
12194 for (SDNode *P : N->users()) {
12195 unsigned Degree = P->getNodeId();
12196 assert(Degree != 0 && "Invalid node degree");
12197 --Degree;
12198 if (Degree == 0) {
12199 // All of P's operands are sorted, so P may sorted now.
12200 P->setNodeId(DAGSize++);
12201 if (P->getIterator() != SortedPos)
12202 SortedPos = AllNodes.insert(where: SortedPos, New: AllNodes.remove(IT: P));
12203 assert(SortedPos != AllNodes.end() && "Overran node list");
12204 ++SortedPos;
12205 } else {
12206 // Update P's outstanding operand count.
12207 P->setNodeId(Degree);
12208 }
12209 }
12210 if (Node.getIterator() == SortedPos) {
12211#ifndef NDEBUG
12212 allnodes_iterator I(N);
12213 SDNode *S = &*++I;
12214 dbgs() << "Overran sorted position:\n";
12215 S->dumprFull(G: this); dbgs() << "\n";
12216 dbgs() << "Checking if this is due to cycles\n";
12217 checkForCycles(DAG: this, force: true);
12218#endif
12219 llvm_unreachable(nullptr);
12220 }
12221 }
12222
12223 assert(SortedPos == AllNodes.end() &&
12224 "Topological sort incomplete!");
12225 assert(AllNodes.front().getOpcode() == ISD::EntryToken &&
12226 "First node in topological sort is not the entry token!");
12227 assert(AllNodes.front().getNodeId() == 0 &&
12228 "First node in topological sort has non-zero id!");
12229 assert(AllNodes.front().getNumOperands() == 0 &&
12230 "First node in topological sort has operands!");
12231 assert(AllNodes.back().getNodeId() == (int)DAGSize-1 &&
12232 "Last node in topologic sort has unexpected id!");
12233 assert(AllNodes.back().use_empty() &&
12234 "Last node in topologic sort has users!");
12235 assert(DAGSize == allnodes_size() && "Node count mismatch!");
12236 return DAGSize;
12237}
12238
12239/// AddDbgValue - Add a dbg_value SDNode. If SD is non-null that means the
12240/// value is produced by SD.
12241void SelectionDAG::AddDbgValue(SDDbgValue *DB, bool isParameter) {
12242 for (SDNode *SD : DB->getSDNodes()) {
12243 if (!SD)
12244 continue;
12245 assert(DbgInfo->getSDDbgValues(SD).empty() || SD->getHasDebugValue());
12246 SD->setHasDebugValue(true);
12247 }
12248 DbgInfo->add(V: DB, isParameter);
12249}
12250
12251void SelectionDAG::AddDbgLabel(SDDbgLabel *DB) { DbgInfo->add(L: DB); }
12252
12253SDValue SelectionDAG::makeEquivalentMemoryOrdering(SDValue OldChain,
12254 SDValue NewMemOpChain) {
12255 assert(isa<MemSDNode>(NewMemOpChain) && "Expected a memop node");
12256 assert(NewMemOpChain.getValueType() == MVT::Other && "Expected a token VT");
12257 // The new memory operation must have the same position as the old load in
12258 // terms of memory dependency. Create a TokenFactor for the old load and new
12259 // memory operation and update uses of the old load's output chain to use that
12260 // TokenFactor.
12261 if (OldChain == NewMemOpChain || OldChain.use_empty())
12262 return NewMemOpChain;
12263
12264 SDValue TokenFactor = getNode(ISD::TokenFactor, SDLoc(OldChain), MVT::Other,
12265 OldChain, NewMemOpChain);
12266 ReplaceAllUsesOfValueWith(From: OldChain, To: TokenFactor);
12267 UpdateNodeOperands(N: TokenFactor.getNode(), Op1: OldChain, Op2: NewMemOpChain);
12268 return TokenFactor;
12269}
12270
12271SDValue SelectionDAG::makeEquivalentMemoryOrdering(LoadSDNode *OldLoad,
12272 SDValue NewMemOp) {
12273 assert(isa<MemSDNode>(NewMemOp.getNode()) && "Expected a memop node");
12274 SDValue OldChain = SDValue(OldLoad, 1);
12275 SDValue NewMemOpChain = NewMemOp.getValue(R: 1);
12276 return makeEquivalentMemoryOrdering(OldChain, NewMemOpChain);
12277}
12278
12279SDValue SelectionDAG::getSymbolFunctionGlobalAddress(SDValue Op,
12280 Function **OutFunction) {
12281 assert(isa<ExternalSymbolSDNode>(Op) && "Node should be an ExternalSymbol");
12282
12283 auto *Symbol = cast<ExternalSymbolSDNode>(Val&: Op)->getSymbol();
12284 auto *Module = MF->getFunction().getParent();
12285 auto *Function = Module->getFunction(Name: Symbol);
12286
12287 if (OutFunction != nullptr)
12288 *OutFunction = Function;
12289
12290 if (Function != nullptr) {
12291 auto PtrTy = TLI->getPointerTy(DL: getDataLayout(), AS: Function->getAddressSpace());
12292 return getGlobalAddress(GV: Function, DL: SDLoc(Op), VT: PtrTy);
12293 }
12294
12295 std::string ErrorStr;
12296 raw_string_ostream ErrorFormatter(ErrorStr);
12297 ErrorFormatter << "Undefined external symbol ";
12298 ErrorFormatter << '"' << Symbol << '"';
12299 report_fatal_error(reason: Twine(ErrorStr));
12300}
12301
12302//===----------------------------------------------------------------------===//
12303// SDNode Class
12304//===----------------------------------------------------------------------===//
12305
12306bool llvm::isNullConstant(SDValue V) {
12307 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(Val&: V);
12308 return Const != nullptr && Const->isZero();
12309}
12310
12311bool llvm::isNullConstantOrUndef(SDValue V) {
12312 return V.isUndef() || isNullConstant(V);
12313}
12314
12315bool llvm::isNullFPConstant(SDValue V) {
12316 ConstantFPSDNode *Const = dyn_cast<ConstantFPSDNode>(Val&: V);
12317 return Const != nullptr && Const->isZero() && !Const->isNegative();
12318}
12319
12320bool llvm::isAllOnesConstant(SDValue V) {
12321 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(Val&: V);
12322 return Const != nullptr && Const->isAllOnes();
12323}
12324
12325bool llvm::isOneConstant(SDValue V) {
12326 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(Val&: V);
12327 return Const != nullptr && Const->isOne();
12328}
12329
12330bool llvm::isMinSignedConstant(SDValue V) {
12331 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(Val&: V);
12332 return Const != nullptr && Const->isMinSignedValue();
12333}
12334
12335bool llvm::isNeutralConstant(unsigned Opcode, SDNodeFlags Flags, SDValue V,
12336 unsigned OperandNo) {
12337 // NOTE: The cases should match with IR's ConstantExpr::getBinOpIdentity().
12338 // TODO: Target-specific opcodes could be added.
12339 if (auto *ConstV = isConstOrConstSplat(N: V, /*AllowUndefs*/ false,
12340 /*AllowTruncation*/ true)) {
12341 APInt Const = ConstV->getAPIntValue().trunc(width: V.getScalarValueSizeInBits());
12342 switch (Opcode) {
12343 case ISD::ADD:
12344 case ISD::OR:
12345 case ISD::XOR:
12346 case ISD::UMAX:
12347 return Const.isZero();
12348 case ISD::MUL:
12349 return Const.isOne();
12350 case ISD::AND:
12351 case ISD::UMIN:
12352 return Const.isAllOnes();
12353 case ISD::SMAX:
12354 return Const.isMinSignedValue();
12355 case ISD::SMIN:
12356 return Const.isMaxSignedValue();
12357 case ISD::SUB:
12358 case ISD::SHL:
12359 case ISD::SRA:
12360 case ISD::SRL:
12361 return OperandNo == 1 && Const.isZero();
12362 case ISD::UDIV:
12363 case ISD::SDIV:
12364 return OperandNo == 1 && Const.isOne();
12365 }
12366 } else if (auto *ConstFP = isConstOrConstSplatFP(N: V)) {
12367 switch (Opcode) {
12368 case ISD::FADD:
12369 return ConstFP->isZero() &&
12370 (Flags.hasNoSignedZeros() || ConstFP->isNegative());
12371 case ISD::FSUB:
12372 return OperandNo == 1 && ConstFP->isZero() &&
12373 (Flags.hasNoSignedZeros() || !ConstFP->isNegative());
12374 case ISD::FMUL:
12375 return ConstFP->isExactlyValue(V: 1.0);
12376 case ISD::FDIV:
12377 return OperandNo == 1 && ConstFP->isExactlyValue(V: 1.0);
12378 case ISD::FMINNUM:
12379 case ISD::FMAXNUM: {
12380 // Neutral element for fminnum is NaN, Inf or FLT_MAX, depending on FMF.
12381 EVT VT = V.getValueType();
12382 const fltSemantics &Semantics = VT.getFltSemantics();
12383 APFloat NeutralAF = !Flags.hasNoNaNs()
12384 ? APFloat::getQNaN(Sem: Semantics)
12385 : !Flags.hasNoInfs()
12386 ? APFloat::getInf(Sem: Semantics)
12387 : APFloat::getLargest(Sem: Semantics);
12388 if (Opcode == ISD::FMAXNUM)
12389 NeutralAF.changeSign();
12390
12391 return ConstFP->isExactlyValue(V: NeutralAF);
12392 }
12393 }
12394 }
12395 return false;
12396}
12397
12398SDValue llvm::peekThroughBitcasts(SDValue V) {
12399 while (V.getOpcode() == ISD::BITCAST)
12400 V = V.getOperand(i: 0);
12401 return V;
12402}
12403
12404SDValue llvm::peekThroughOneUseBitcasts(SDValue V) {
12405 while (V.getOpcode() == ISD::BITCAST && V.getOperand(i: 0).hasOneUse())
12406 V = V.getOperand(i: 0);
12407 return V;
12408}
12409
12410SDValue llvm::peekThroughExtractSubvectors(SDValue V) {
12411 while (V.getOpcode() == ISD::EXTRACT_SUBVECTOR)
12412 V = V.getOperand(i: 0);
12413 return V;
12414}
12415
12416SDValue llvm::peekThroughTruncates(SDValue V) {
12417 while (V.getOpcode() == ISD::TRUNCATE)
12418 V = V.getOperand(i: 0);
12419 return V;
12420}
12421
12422bool llvm::isBitwiseNot(SDValue V, bool AllowUndefs) {
12423 if (V.getOpcode() != ISD::XOR)
12424 return false;
12425 V = peekThroughBitcasts(V: V.getOperand(i: 1));
12426 unsigned NumBits = V.getScalarValueSizeInBits();
12427 ConstantSDNode *C =
12428 isConstOrConstSplat(N: V, AllowUndefs, /*AllowTruncation*/ true);
12429 return C && (C->getAPIntValue().countr_one() >= NumBits);
12430}
12431
12432ConstantSDNode *llvm::isConstOrConstSplat(SDValue N, bool AllowUndefs,
12433 bool AllowTruncation) {
12434 EVT VT = N.getValueType();
12435 APInt DemandedElts = VT.isFixedLengthVector()
12436 ? APInt::getAllOnes(numBits: VT.getVectorMinNumElements())
12437 : APInt(1, 1);
12438 return isConstOrConstSplat(N, DemandedElts, AllowUndefs, AllowTruncation);
12439}
12440
12441ConstantSDNode *llvm::isConstOrConstSplat(SDValue N, const APInt &DemandedElts,
12442 bool AllowUndefs,
12443 bool AllowTruncation) {
12444 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val&: N))
12445 return CN;
12446
12447 // SplatVectors can truncate their operands. Ignore that case here unless
12448 // AllowTruncation is set.
12449 if (N->getOpcode() == ISD::SPLAT_VECTOR) {
12450 EVT VecEltVT = N->getValueType(ResNo: 0).getVectorElementType();
12451 if (auto *CN = dyn_cast<ConstantSDNode>(Val: N->getOperand(Num: 0))) {
12452 EVT CVT = CN->getValueType(ResNo: 0);
12453 assert(CVT.bitsGE(VecEltVT) && "Illegal splat_vector element extension");
12454 if (AllowTruncation || CVT == VecEltVT)
12455 return CN;
12456 }
12457 }
12458
12459 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Val&: N)) {
12460 BitVector UndefElements;
12461 ConstantSDNode *CN = BV->getConstantSplatNode(DemandedElts, UndefElements: &UndefElements);
12462
12463 // BuildVectors can truncate their operands. Ignore that case here unless
12464 // AllowTruncation is set.
12465 // TODO: Look into whether we should allow UndefElements in non-DemandedElts
12466 if (CN && (UndefElements.none() || AllowUndefs)) {
12467 EVT CVT = CN->getValueType(ResNo: 0);
12468 EVT NSVT = N.getValueType().getScalarType();
12469 assert(CVT.bitsGE(NSVT) && "Illegal build vector element extension");
12470 if (AllowTruncation || (CVT == NSVT))
12471 return CN;
12472 }
12473 }
12474
12475 return nullptr;
12476}
12477
12478ConstantFPSDNode *llvm::isConstOrConstSplatFP(SDValue N, bool AllowUndefs) {
12479 EVT VT = N.getValueType();
12480 APInt DemandedElts = VT.isFixedLengthVector()
12481 ? APInt::getAllOnes(numBits: VT.getVectorMinNumElements())
12482 : APInt(1, 1);
12483 return isConstOrConstSplatFP(N, DemandedElts, AllowUndefs);
12484}
12485
12486ConstantFPSDNode *llvm::isConstOrConstSplatFP(SDValue N,
12487 const APInt &DemandedElts,
12488 bool AllowUndefs) {
12489 if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(Val&: N))
12490 return CN;
12491
12492 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Val&: N)) {
12493 BitVector UndefElements;
12494 ConstantFPSDNode *CN =
12495 BV->getConstantFPSplatNode(DemandedElts, UndefElements: &UndefElements);
12496 // TODO: Look into whether we should allow UndefElements in non-DemandedElts
12497 if (CN && (UndefElements.none() || AllowUndefs))
12498 return CN;
12499 }
12500
12501 if (N.getOpcode() == ISD::SPLAT_VECTOR)
12502 if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(Val: N.getOperand(i: 0)))
12503 return CN;
12504
12505 return nullptr;
12506}
12507
12508bool llvm::isNullOrNullSplat(SDValue N, bool AllowUndefs) {
12509 // TODO: may want to use peekThroughBitcast() here.
12510 ConstantSDNode *C =
12511 isConstOrConstSplat(N, AllowUndefs, /*AllowTruncation=*/true);
12512 return C && C->isZero();
12513}
12514
12515bool llvm::isOneOrOneSplat(SDValue N, bool AllowUndefs) {
12516 ConstantSDNode *C =
12517 isConstOrConstSplat(N, AllowUndefs, /*AllowTruncation*/ true);
12518 return C && C->isOne();
12519}
12520
12521bool llvm::isAllOnesOrAllOnesSplat(SDValue N, bool AllowUndefs) {
12522 N = peekThroughBitcasts(V: N);
12523 unsigned BitWidth = N.getScalarValueSizeInBits();
12524 ConstantSDNode *C = isConstOrConstSplat(N, AllowUndefs);
12525 return C && C->isAllOnes() && C->getValueSizeInBits(ResNo: 0) == BitWidth;
12526}
12527
12528HandleSDNode::~HandleSDNode() {
12529 DropOperands();
12530}
12531
12532MemSDNode::MemSDNode(unsigned Opc, unsigned Order, const DebugLoc &dl,
12533 SDVTList VTs, EVT memvt, MachineMemOperand *mmo)
12534 : SDNode(Opc, Order, dl, VTs), MemoryVT(memvt), MMO(mmo) {
12535 MemSDNodeBits.IsVolatile = MMO->isVolatile();
12536 MemSDNodeBits.IsNonTemporal = MMO->isNonTemporal();
12537 MemSDNodeBits.IsDereferenceable = MMO->isDereferenceable();
12538 MemSDNodeBits.IsInvariant = MMO->isInvariant();
12539
12540 // We check here that the size of the memory operand fits within the size of
12541 // the MMO. This is because the MMO might indicate only a possible address
12542 // range instead of specifying the affected memory addresses precisely.
12543 assert(
12544 (!MMO->getType().isValid() ||
12545 TypeSize::isKnownLE(memvt.getStoreSize(), MMO->getSize().getValue())) &&
12546 "Size mismatch!");
12547}
12548
12549/// Profile - Gather unique data for the node.
12550///
12551void SDNode::Profile(FoldingSetNodeID &ID) const {
12552 AddNodeIDNode(ID, N: this);
12553}
12554
12555namespace {
12556
12557 struct EVTArray {
12558 std::vector<EVT> VTs;
12559
12560 EVTArray() {
12561 VTs.reserve(n: MVT::VALUETYPE_SIZE);
12562 for (unsigned i = 0; i < MVT::VALUETYPE_SIZE; ++i)
12563 VTs.push_back(x: MVT((MVT::SimpleValueType)i));
12564 }
12565 };
12566
12567} // end anonymous namespace
12568
12569/// getValueTypeList - Return a pointer to the specified value type.
12570///
12571const EVT *SDNode::getValueTypeList(MVT VT) {
12572 static EVTArray SimpleVTArray;
12573
12574 assert(VT < MVT::VALUETYPE_SIZE && "Value type out of range!");
12575 return &SimpleVTArray.VTs[VT.SimpleTy];
12576}
12577
12578/// hasAnyUseOfValue - Return true if there are any use of the indicated
12579/// value. This method ignores uses of other values defined by this operation.
12580bool SDNode::hasAnyUseOfValue(unsigned Value) const {
12581 assert(Value < getNumValues() && "Bad value!");
12582
12583 for (SDUse &U : uses())
12584 if (U.getResNo() == Value)
12585 return true;
12586
12587 return false;
12588}
12589
12590/// isOnlyUserOf - Return true if this node is the only use of N.
12591bool SDNode::isOnlyUserOf(const SDNode *N) const {
12592 bool Seen = false;
12593 for (const SDNode *User : N->users()) {
12594 if (User == this)
12595 Seen = true;
12596 else
12597 return false;
12598 }
12599
12600 return Seen;
12601}
12602
12603/// Return true if the only users of N are contained in Nodes.
12604bool SDNode::areOnlyUsersOf(ArrayRef<const SDNode *> Nodes, const SDNode *N) {
12605 bool Seen = false;
12606 for (const SDNode *User : N->users()) {
12607 if (llvm::is_contained(Range&: Nodes, Element: User))
12608 Seen = true;
12609 else
12610 return false;
12611 }
12612
12613 return Seen;
12614}
12615
12616/// isOperand - Return true if this node is an operand of N.
12617bool SDValue::isOperandOf(const SDNode *N) const {
12618 return is_contained(Range: N->op_values(), Element: *this);
12619}
12620
12621bool SDNode::isOperandOf(const SDNode *N) const {
12622 return any_of(Range: N->op_values(),
12623 P: [this](SDValue Op) { return this == Op.getNode(); });
12624}
12625
12626/// reachesChainWithoutSideEffects - Return true if this operand (which must
12627/// be a chain) reaches the specified operand without crossing any
12628/// side-effecting instructions on any chain path. In practice, this looks
12629/// through token factors and non-volatile loads. In order to remain efficient,
12630/// this only looks a couple of nodes in, it does not do an exhaustive search.
12631///
12632/// Note that we only need to examine chains when we're searching for
12633/// side-effects; SelectionDAG requires that all side-effects are represented
12634/// by chains, even if another operand would force a specific ordering. This
12635/// constraint is necessary to allow transformations like splitting loads.
12636bool SDValue::reachesChainWithoutSideEffects(SDValue Dest,
12637 unsigned Depth) const {
12638 if (*this == Dest) return true;
12639
12640 // Don't search too deeply, we just want to be able to see through
12641 // TokenFactor's etc.
12642 if (Depth == 0) return false;
12643
12644 // If this is a token factor, all inputs to the TF happen in parallel.
12645 if (getOpcode() == ISD::TokenFactor) {
12646 // First, try a shallow search.
12647 if (is_contained(Range: (*this)->ops(), Element: Dest)) {
12648 // We found the chain we want as an operand of this TokenFactor.
12649 // Essentially, we reach the chain without side-effects if we could
12650 // serialize the TokenFactor into a simple chain of operations with
12651 // Dest as the last operation. This is automatically true if the
12652 // chain has one use: there are no other ordering constraints.
12653 // If the chain has more than one use, we give up: some other
12654 // use of Dest might force a side-effect between Dest and the current
12655 // node.
12656 if (Dest.hasOneUse())
12657 return true;
12658 }
12659 // Next, try a deep search: check whether every operand of the TokenFactor
12660 // reaches Dest.
12661 return llvm::all_of(Range: (*this)->ops(), P: [=](SDValue Op) {
12662 return Op.reachesChainWithoutSideEffects(Dest, Depth: Depth - 1);
12663 });
12664 }
12665
12666 // Loads don't have side effects, look through them.
12667 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Val: *this)) {
12668 if (Ld->isUnordered())
12669 return Ld->getChain().reachesChainWithoutSideEffects(Dest, Depth: Depth-1);
12670 }
12671 return false;
12672}
12673
12674bool SDNode::hasPredecessor(const SDNode *N) const {
12675 SmallPtrSet<const SDNode *, 32> Visited;
12676 SmallVector<const SDNode *, 16> Worklist;
12677 Worklist.push_back(Elt: this);
12678 return hasPredecessorHelper(N, Visited, Worklist);
12679}
12680
12681void SDNode::intersectFlagsWith(const SDNodeFlags Flags) {
12682 this->Flags &= Flags;
12683}
12684
12685SDValue
12686SelectionDAG::matchBinOpReduction(SDNode *Extract, ISD::NodeType &BinOp,
12687 ArrayRef<ISD::NodeType> CandidateBinOps,
12688 bool AllowPartials) {
12689 // The pattern must end in an extract from index 0.
12690 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
12691 !isNullConstant(V: Extract->getOperand(Num: 1)))
12692 return SDValue();
12693
12694 // Match against one of the candidate binary ops.
12695 SDValue Op = Extract->getOperand(Num: 0);
12696 if (llvm::none_of(Range&: CandidateBinOps, P: [Op](ISD::NodeType BinOp) {
12697 return Op.getOpcode() == unsigned(BinOp);
12698 }))
12699 return SDValue();
12700
12701 // Floating-point reductions may require relaxed constraints on the final step
12702 // of the reduction because they may reorder intermediate operations.
12703 unsigned CandidateBinOp = Op.getOpcode();
12704 if (Op.getValueType().isFloatingPoint()) {
12705 SDNodeFlags Flags = Op->getFlags();
12706 switch (CandidateBinOp) {
12707 case ISD::FADD:
12708 if (!Flags.hasNoSignedZeros() || !Flags.hasAllowReassociation())
12709 return SDValue();
12710 break;
12711 default:
12712 llvm_unreachable("Unhandled FP opcode for binop reduction");
12713 }
12714 }
12715
12716 // Matching failed - attempt to see if we did enough stages that a partial
12717 // reduction from a subvector is possible.
12718 auto PartialReduction = [&](SDValue Op, unsigned NumSubElts) {
12719 if (!AllowPartials || !Op)
12720 return SDValue();
12721 EVT OpVT = Op.getValueType();
12722 EVT OpSVT = OpVT.getScalarType();
12723 EVT SubVT = EVT::getVectorVT(Context&: *getContext(), VT: OpSVT, NumElements: NumSubElts);
12724 if (!TLI->isExtractSubvectorCheap(ResVT: SubVT, SrcVT: OpVT, Index: 0))
12725 return SDValue();
12726 BinOp = (ISD::NodeType)CandidateBinOp;
12727 return getExtractSubvector(DL: SDLoc(Op), VT: SubVT, Vec: Op, Idx: 0);
12728 };
12729
12730 // At each stage, we're looking for something that looks like:
12731 // %s = shufflevector <8 x i32> %op, <8 x i32> undef,
12732 // <8 x i32> <i32 2, i32 3, i32 undef, i32 undef,
12733 // i32 undef, i32 undef, i32 undef, i32 undef>
12734 // %a = binop <8 x i32> %op, %s
12735 // Where the mask changes according to the stage. E.g. for a 3-stage pyramid,
12736 // we expect something like:
12737 // <4,5,6,7,u,u,u,u>
12738 // <2,3,u,u,u,u,u,u>
12739 // <1,u,u,u,u,u,u,u>
12740 // While a partial reduction match would be:
12741 // <2,3,u,u,u,u,u,u>
12742 // <1,u,u,u,u,u,u,u>
12743 unsigned Stages = Log2_32(Value: Op.getValueType().getVectorNumElements());
12744 SDValue PrevOp;
12745 for (unsigned i = 0; i < Stages; ++i) {
12746 unsigned MaskEnd = (1 << i);
12747
12748 if (Op.getOpcode() != CandidateBinOp)
12749 return PartialReduction(PrevOp, MaskEnd);
12750
12751 SDValue Op0 = Op.getOperand(i: 0);
12752 SDValue Op1 = Op.getOperand(i: 1);
12753
12754 ShuffleVectorSDNode *Shuffle = dyn_cast<ShuffleVectorSDNode>(Val&: Op0);
12755 if (Shuffle) {
12756 Op = Op1;
12757 } else {
12758 Shuffle = dyn_cast<ShuffleVectorSDNode>(Val&: Op1);
12759 Op = Op0;
12760 }
12761
12762 // The first operand of the shuffle should be the same as the other operand
12763 // of the binop.
12764 if (!Shuffle || Shuffle->getOperand(Num: 0) != Op)
12765 return PartialReduction(PrevOp, MaskEnd);
12766
12767 // Verify the shuffle has the expected (at this stage of the pyramid) mask.
12768 for (int Index = 0; Index < (int)MaskEnd; ++Index)
12769 if (Shuffle->getMaskElt(Idx: Index) != (int)(MaskEnd + Index))
12770 return PartialReduction(PrevOp, MaskEnd);
12771
12772 PrevOp = Op;
12773 }
12774
12775 // Handle subvector reductions, which tend to appear after the shuffle
12776 // reduction stages.
12777 while (Op.getOpcode() == CandidateBinOp) {
12778 unsigned NumElts = Op.getValueType().getVectorNumElements();
12779 SDValue Op0 = Op.getOperand(i: 0);
12780 SDValue Op1 = Op.getOperand(i: 1);
12781 if (Op0.getOpcode() != ISD::EXTRACT_SUBVECTOR ||
12782 Op1.getOpcode() != ISD::EXTRACT_SUBVECTOR ||
12783 Op0.getOperand(i: 0) != Op1.getOperand(i: 0))
12784 break;
12785 SDValue Src = Op0.getOperand(i: 0);
12786 unsigned NumSrcElts = Src.getValueType().getVectorNumElements();
12787 if (NumSrcElts != (2 * NumElts))
12788 break;
12789 if (!(Op0.getConstantOperandAPInt(i: 1) == 0 &&
12790 Op1.getConstantOperandAPInt(i: 1) == NumElts) &&
12791 !(Op1.getConstantOperandAPInt(i: 1) == 0 &&
12792 Op0.getConstantOperandAPInt(i: 1) == NumElts))
12793 break;
12794 Op = Src;
12795 }
12796
12797 BinOp = (ISD::NodeType)CandidateBinOp;
12798 return Op;
12799}
12800
12801SDValue SelectionDAG::UnrollVectorOp(SDNode *N, unsigned ResNE) {
12802 EVT VT = N->getValueType(ResNo: 0);
12803 EVT EltVT = VT.getVectorElementType();
12804 unsigned NE = VT.getVectorNumElements();
12805
12806 SDLoc dl(N);
12807
12808 // If ResNE is 0, fully unroll the vector op.
12809 if (ResNE == 0)
12810 ResNE = NE;
12811 else if (NE > ResNE)
12812 NE = ResNE;
12813
12814 if (N->getNumValues() == 2) {
12815 SmallVector<SDValue, 8> Scalars0, Scalars1;
12816 SmallVector<SDValue, 4> Operands(N->getNumOperands());
12817 EVT VT1 = N->getValueType(ResNo: 1);
12818 EVT EltVT1 = VT1.getVectorElementType();
12819
12820 unsigned i;
12821 for (i = 0; i != NE; ++i) {
12822 for (unsigned j = 0, e = N->getNumOperands(); j != e; ++j) {
12823 SDValue Operand = N->getOperand(Num: j);
12824 EVT OperandVT = Operand.getValueType();
12825
12826 // A vector operand; extract a single element.
12827 EVT OperandEltVT = OperandVT.getVectorElementType();
12828 Operands[j] = getExtractVectorElt(DL: dl, VT: OperandEltVT, Vec: Operand, Idx: i);
12829 }
12830
12831 SDValue EltOp = getNode(Opcode: N->getOpcode(), DL: dl, ResultTys: {EltVT, EltVT1}, Ops: Operands);
12832 Scalars0.push_back(Elt: EltOp);
12833 Scalars1.push_back(Elt: EltOp.getValue(R: 1));
12834 }
12835
12836 for (; i < ResNE; ++i) {
12837 Scalars0.push_back(Elt: getUNDEF(VT: EltVT));
12838 Scalars1.push_back(Elt: getUNDEF(VT: EltVT1));
12839 }
12840
12841 EVT VecVT = EVT::getVectorVT(Context&: *getContext(), VT: EltVT, NumElements: ResNE);
12842 EVT VecVT1 = EVT::getVectorVT(Context&: *getContext(), VT: EltVT1, NumElements: ResNE);
12843 SDValue Vec0 = getBuildVector(VT: VecVT, DL: dl, Ops: Scalars0);
12844 SDValue Vec1 = getBuildVector(VT: VecVT1, DL: dl, Ops: Scalars1);
12845 return getMergeValues(Ops: {Vec0, Vec1}, dl);
12846 }
12847
12848 assert(N->getNumValues() == 1 &&
12849 "Can't unroll a vector with multiple results!");
12850
12851 SmallVector<SDValue, 8> Scalars;
12852 SmallVector<SDValue, 4> Operands(N->getNumOperands());
12853
12854 unsigned i;
12855 for (i= 0; i != NE; ++i) {
12856 for (unsigned j = 0, e = N->getNumOperands(); j != e; ++j) {
12857 SDValue Operand = N->getOperand(Num: j);
12858 EVT OperandVT = Operand.getValueType();
12859 if (OperandVT.isVector()) {
12860 // A vector operand; extract a single element.
12861 EVT OperandEltVT = OperandVT.getVectorElementType();
12862 Operands[j] = getExtractVectorElt(DL: dl, VT: OperandEltVT, Vec: Operand, Idx: i);
12863 } else {
12864 // A scalar operand; just use it as is.
12865 Operands[j] = Operand;
12866 }
12867 }
12868
12869 switch (N->getOpcode()) {
12870 default: {
12871 Scalars.push_back(Elt: getNode(Opcode: N->getOpcode(), DL: dl, VT: EltVT, Ops: Operands,
12872 Flags: N->getFlags()));
12873 break;
12874 }
12875 case ISD::VSELECT:
12876 Scalars.push_back(Elt: getNode(Opcode: ISD::SELECT, DL: dl, VT: EltVT, Ops: Operands));
12877 break;
12878 case ISD::SHL:
12879 case ISD::SRA:
12880 case ISD::SRL:
12881 case ISD::ROTL:
12882 case ISD::ROTR:
12883 Scalars.push_back(Elt: getNode(Opcode: N->getOpcode(), DL: dl, VT: EltVT, N1: Operands[0],
12884 N2: getShiftAmountOperand(LHSTy: Operands[0].getValueType(),
12885 Op: Operands[1])));
12886 break;
12887 case ISD::SIGN_EXTEND_INREG: {
12888 EVT ExtVT = cast<VTSDNode>(Val&: Operands[1])->getVT().getVectorElementType();
12889 Scalars.push_back(Elt: getNode(Opcode: N->getOpcode(), DL: dl, VT: EltVT,
12890 N1: Operands[0],
12891 N2: getValueType(VT: ExtVT)));
12892 break;
12893 }
12894 case ISD::ADDRSPACECAST: {
12895 const auto *ASC = cast<AddrSpaceCastSDNode>(Val: N);
12896 Scalars.push_back(Elt: getAddrSpaceCast(dl, VT: EltVT, Ptr: Operands[0],
12897 SrcAS: ASC->getSrcAddressSpace(),
12898 DestAS: ASC->getDestAddressSpace()));
12899 break;
12900 }
12901 }
12902 }
12903
12904 for (; i < ResNE; ++i)
12905 Scalars.push_back(Elt: getUNDEF(VT: EltVT));
12906
12907 EVT VecVT = EVT::getVectorVT(Context&: *getContext(), VT: EltVT, NumElements: ResNE);
12908 return getBuildVector(VT: VecVT, DL: dl, Ops: Scalars);
12909}
12910
12911std::pair<SDValue, SDValue> SelectionDAG::UnrollVectorOverflowOp(
12912 SDNode *N, unsigned ResNE) {
12913 unsigned Opcode = N->getOpcode();
12914 assert((Opcode == ISD::UADDO || Opcode == ISD::SADDO ||
12915 Opcode == ISD::USUBO || Opcode == ISD::SSUBO ||
12916 Opcode == ISD::UMULO || Opcode == ISD::SMULO) &&
12917 "Expected an overflow opcode");
12918
12919 EVT ResVT = N->getValueType(ResNo: 0);
12920 EVT OvVT = N->getValueType(ResNo: 1);
12921 EVT ResEltVT = ResVT.getVectorElementType();
12922 EVT OvEltVT = OvVT.getVectorElementType();
12923 SDLoc dl(N);
12924
12925 // If ResNE is 0, fully unroll the vector op.
12926 unsigned NE = ResVT.getVectorNumElements();
12927 if (ResNE == 0)
12928 ResNE = NE;
12929 else if (NE > ResNE)
12930 NE = ResNE;
12931
12932 SmallVector<SDValue, 8> LHSScalars;
12933 SmallVector<SDValue, 8> RHSScalars;
12934 ExtractVectorElements(Op: N->getOperand(Num: 0), Args&: LHSScalars, Start: 0, Count: NE);
12935 ExtractVectorElements(Op: N->getOperand(Num: 1), Args&: RHSScalars, Start: 0, Count: NE);
12936
12937 EVT SVT = TLI->getSetCCResultType(DL: getDataLayout(), Context&: *getContext(), VT: ResEltVT);
12938 SDVTList VTs = getVTList(VT1: ResEltVT, VT2: SVT);
12939 SmallVector<SDValue, 8> ResScalars;
12940 SmallVector<SDValue, 8> OvScalars;
12941 for (unsigned i = 0; i < NE; ++i) {
12942 SDValue Res = getNode(Opcode, DL: dl, VTList: VTs, N1: LHSScalars[i], N2: RHSScalars[i]);
12943 SDValue Ov =
12944 getSelect(DL: dl, VT: OvEltVT, Cond: Res.getValue(R: 1),
12945 LHS: getBoolConstant(V: true, DL: dl, VT: OvEltVT, OpVT: ResVT),
12946 RHS: getConstant(Val: 0, DL: dl, VT: OvEltVT));
12947
12948 ResScalars.push_back(Elt: Res);
12949 OvScalars.push_back(Elt: Ov);
12950 }
12951
12952 ResScalars.append(NumInputs: ResNE - NE, Elt: getUNDEF(VT: ResEltVT));
12953 OvScalars.append(NumInputs: ResNE - NE, Elt: getUNDEF(VT: OvEltVT));
12954
12955 EVT NewResVT = EVT::getVectorVT(Context&: *getContext(), VT: ResEltVT, NumElements: ResNE);
12956 EVT NewOvVT = EVT::getVectorVT(Context&: *getContext(), VT: OvEltVT, NumElements: ResNE);
12957 return std::make_pair(x: getBuildVector(VT: NewResVT, DL: dl, Ops: ResScalars),
12958 y: getBuildVector(VT: NewOvVT, DL: dl, Ops: OvScalars));
12959}
12960
12961bool SelectionDAG::areNonVolatileConsecutiveLoads(LoadSDNode *LD,
12962 LoadSDNode *Base,
12963 unsigned Bytes,
12964 int Dist) const {
12965 if (LD->isVolatile() || Base->isVolatile())
12966 return false;
12967 // TODO: probably too restrictive for atomics, revisit
12968 if (!LD->isSimple())
12969 return false;
12970 if (LD->isIndexed() || Base->isIndexed())
12971 return false;
12972 if (LD->getChain() != Base->getChain())
12973 return false;
12974 EVT VT = LD->getMemoryVT();
12975 if (VT.getSizeInBits() / 8 != Bytes)
12976 return false;
12977
12978 auto BaseLocDecomp = BaseIndexOffset::match(N: Base, DAG: *this);
12979 auto LocDecomp = BaseIndexOffset::match(N: LD, DAG: *this);
12980
12981 int64_t Offset = 0;
12982 if (BaseLocDecomp.equalBaseIndex(Other: LocDecomp, DAG: *this, Off&: Offset))
12983 return (Dist * (int64_t)Bytes == Offset);
12984 return false;
12985}
12986
12987/// InferPtrAlignment - Infer alignment of a load / store address. Return
12988/// std::nullopt if it cannot be inferred.
12989MaybeAlign SelectionDAG::InferPtrAlign(SDValue Ptr) const {
12990 // If this is a GlobalAddress + cst, return the alignment.
12991 const GlobalValue *GV = nullptr;
12992 int64_t GVOffset = 0;
12993 if (TLI->isGAPlusOffset(N: Ptr.getNode(), GA&: GV, Offset&: GVOffset)) {
12994 unsigned PtrWidth = getDataLayout().getPointerTypeSizeInBits(GV->getType());
12995 KnownBits Known(PtrWidth);
12996 llvm::computeKnownBits(V: GV, Known, DL: getDataLayout());
12997 unsigned AlignBits = Known.countMinTrailingZeros();
12998 if (AlignBits)
12999 return commonAlignment(A: Align(1ull << std::min(a: 31U, b: AlignBits)), Offset: GVOffset);
13000 }
13001
13002 // If this is a direct reference to a stack slot, use information about the
13003 // stack slot's alignment.
13004 int FrameIdx = INT_MIN;
13005 int64_t FrameOffset = 0;
13006 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Val&: Ptr)) {
13007 FrameIdx = FI->getIndex();
13008 } else if (isBaseWithConstantOffset(Op: Ptr) &&
13009 isa<FrameIndexSDNode>(Val: Ptr.getOperand(i: 0))) {
13010 // Handle FI+Cst
13011 FrameIdx = cast<FrameIndexSDNode>(Val: Ptr.getOperand(i: 0))->getIndex();
13012 FrameOffset = Ptr.getConstantOperandVal(i: 1);
13013 }
13014
13015 if (FrameIdx != INT_MIN) {
13016 const MachineFrameInfo &MFI = getMachineFunction().getFrameInfo();
13017 return commonAlignment(A: MFI.getObjectAlign(ObjectIdx: FrameIdx), Offset: FrameOffset);
13018 }
13019
13020 return std::nullopt;
13021}
13022
13023/// Split the scalar node with EXTRACT_ELEMENT using the provided
13024/// VTs and return the low/high part.
13025std::pair<SDValue, SDValue> SelectionDAG::SplitScalar(const SDValue &N,
13026 const SDLoc &DL,
13027 const EVT &LoVT,
13028 const EVT &HiVT) {
13029 assert(!LoVT.isVector() && !HiVT.isVector() && !N.getValueType().isVector() &&
13030 "Split node must be a scalar type");
13031 SDValue Lo =
13032 getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: LoVT, N1: N, N2: getIntPtrConstant(Val: 0, DL));
13033 SDValue Hi =
13034 getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: HiVT, N1: N, N2: getIntPtrConstant(Val: 1, DL));
13035 return std::make_pair(x&: Lo, y&: Hi);
13036}
13037
13038/// GetSplitDestVTs - Compute the VTs needed for the low/hi parts of a type
13039/// which is split (or expanded) into two not necessarily identical pieces.
13040std::pair<EVT, EVT> SelectionDAG::GetSplitDestVTs(const EVT &VT) const {
13041 // Currently all types are split in half.
13042 EVT LoVT, HiVT;
13043 if (!VT.isVector())
13044 LoVT = HiVT = TLI->getTypeToTransformTo(Context&: *getContext(), VT);
13045 else
13046 LoVT = HiVT = VT.getHalfNumVectorElementsVT(Context&: *getContext());
13047
13048 return std::make_pair(x&: LoVT, y&: HiVT);
13049}
13050
13051/// GetDependentSplitDestVTs - Compute the VTs needed for the low/hi parts of a
13052/// type, dependent on an enveloping VT that has been split into two identical
13053/// pieces. Sets the HiIsEmpty flag when hi type has zero storage size.
13054std::pair<EVT, EVT>
13055SelectionDAG::GetDependentSplitDestVTs(const EVT &VT, const EVT &EnvVT,
13056 bool *HiIsEmpty) const {
13057 EVT EltTp = VT.getVectorElementType();
13058 // Examples:
13059 // custom VL=8 with enveloping VL=8/8 yields 8/0 (hi empty)
13060 // custom VL=9 with enveloping VL=8/8 yields 8/1
13061 // custom VL=10 with enveloping VL=8/8 yields 8/2
13062 // etc.
13063 ElementCount VTNumElts = VT.getVectorElementCount();
13064 ElementCount EnvNumElts = EnvVT.getVectorElementCount();
13065 assert(VTNumElts.isScalable() == EnvNumElts.isScalable() &&
13066 "Mixing fixed width and scalable vectors when enveloping a type");
13067 EVT LoVT, HiVT;
13068 if (VTNumElts.getKnownMinValue() > EnvNumElts.getKnownMinValue()) {
13069 LoVT = EVT::getVectorVT(Context&: *getContext(), VT: EltTp, EC: EnvNumElts);
13070 HiVT = EVT::getVectorVT(Context&: *getContext(), VT: EltTp, EC: VTNumElts - EnvNumElts);
13071 *HiIsEmpty = false;
13072 } else {
13073 // Flag that hi type has zero storage size, but return split envelop type
13074 // (this would be easier if vector types with zero elements were allowed).
13075 LoVT = EVT::getVectorVT(Context&: *getContext(), VT: EltTp, EC: VTNumElts);
13076 HiVT = EVT::getVectorVT(Context&: *getContext(), VT: EltTp, EC: EnvNumElts);
13077 *HiIsEmpty = true;
13078 }
13079 return std::make_pair(x&: LoVT, y&: HiVT);
13080}
13081
13082/// SplitVector - Split the vector with EXTRACT_SUBVECTOR and return the
13083/// low/high part.
13084std::pair<SDValue, SDValue>
13085SelectionDAG::SplitVector(const SDValue &N, const SDLoc &DL, const EVT &LoVT,
13086 const EVT &HiVT) {
13087 assert(LoVT.isScalableVector() == HiVT.isScalableVector() &&
13088 LoVT.isScalableVector() == N.getValueType().isScalableVector() &&
13089 "Splitting vector with an invalid mixture of fixed and scalable "
13090 "vector types");
13091 assert(LoVT.getVectorMinNumElements() + HiVT.getVectorMinNumElements() <=
13092 N.getValueType().getVectorMinNumElements() &&
13093 "More vector elements requested than available!");
13094 SDValue Lo, Hi;
13095 Lo = getExtractSubvector(DL, VT: LoVT, Vec: N, Idx: 0);
13096 // For scalable vectors it is safe to use LoVT.getVectorMinNumElements()
13097 // (rather than having to use ElementCount), because EXTRACT_SUBVECTOR scales
13098 // IDX with the runtime scaling factor of the result vector type. For
13099 // fixed-width result vectors, that runtime scaling factor is 1.
13100 Hi = getNode(Opcode: ISD::EXTRACT_SUBVECTOR, DL, VT: HiVT, N1: N,
13101 N2: getVectorIdxConstant(Val: LoVT.getVectorMinNumElements(), DL));
13102 return std::make_pair(x&: Lo, y&: Hi);
13103}
13104
13105std::pair<SDValue, SDValue> SelectionDAG::SplitEVL(SDValue N, EVT VecVT,
13106 const SDLoc &DL) {
13107 // Split the vector length parameter.
13108 // %evl -> umin(%evl, %halfnumelts) and usubsat(%evl - %halfnumelts).
13109 EVT VT = N.getValueType();
13110 assert(VecVT.getVectorElementCount().isKnownEven() &&
13111 "Expecting the mask to be an evenly-sized vector");
13112 unsigned HalfMinNumElts = VecVT.getVectorMinNumElements() / 2;
13113 SDValue HalfNumElts =
13114 VecVT.isFixedLengthVector()
13115 ? getConstant(Val: HalfMinNumElts, DL, VT)
13116 : getVScale(DL, VT, MulImm: APInt(VT.getScalarSizeInBits(), HalfMinNumElts));
13117 SDValue Lo = getNode(Opcode: ISD::UMIN, DL, VT, N1: N, N2: HalfNumElts);
13118 SDValue Hi = getNode(Opcode: ISD::USUBSAT, DL, VT, N1: N, N2: HalfNumElts);
13119 return std::make_pair(x&: Lo, y&: Hi);
13120}
13121
13122/// Widen the vector up to the next power of two using INSERT_SUBVECTOR.
13123SDValue SelectionDAG::WidenVector(const SDValue &N, const SDLoc &DL) {
13124 EVT VT = N.getValueType();
13125 EVT WideVT = EVT::getVectorVT(Context&: *getContext(), VT: VT.getVectorElementType(),
13126 NumElements: NextPowerOf2(A: VT.getVectorNumElements()));
13127 return getInsertSubvector(DL, Vec: getUNDEF(VT: WideVT), SubVec: N, Idx: 0);
13128}
13129
13130void SelectionDAG::ExtractVectorElements(SDValue Op,
13131 SmallVectorImpl<SDValue> &Args,
13132 unsigned Start, unsigned Count,
13133 EVT EltVT) {
13134 EVT VT = Op.getValueType();
13135 if (Count == 0)
13136 Count = VT.getVectorNumElements();
13137 if (EltVT == EVT())
13138 EltVT = VT.getVectorElementType();
13139 SDLoc SL(Op);
13140 for (unsigned i = Start, e = Start + Count; i != e; ++i) {
13141 Args.push_back(Elt: getExtractVectorElt(DL: SL, VT: EltVT, Vec: Op, Idx: i));
13142 }
13143}
13144
13145// getAddressSpace - Return the address space this GlobalAddress belongs to.
13146unsigned GlobalAddressSDNode::getAddressSpace() const {
13147 return getGlobal()->getType()->getAddressSpace();
13148}
13149
13150Type *ConstantPoolSDNode::getType() const {
13151 if (isMachineConstantPoolEntry())
13152 return Val.MachineCPVal->getType();
13153 return Val.ConstVal->getType();
13154}
13155
13156bool BuildVectorSDNode::isConstantSplat(APInt &SplatValue, APInt &SplatUndef,
13157 unsigned &SplatBitSize,
13158 bool &HasAnyUndefs,
13159 unsigned MinSplatBits,
13160 bool IsBigEndian) const {
13161 EVT VT = getValueType(ResNo: 0);
13162 assert(VT.isVector() && "Expected a vector type");
13163 unsigned VecWidth = VT.getSizeInBits();
13164 if (MinSplatBits > VecWidth)
13165 return false;
13166
13167 // FIXME: The widths are based on this node's type, but build vectors can
13168 // truncate their operands.
13169 SplatValue = APInt(VecWidth, 0);
13170 SplatUndef = APInt(VecWidth, 0);
13171
13172 // Get the bits. Bits with undefined values (when the corresponding element
13173 // of the vector is an ISD::UNDEF value) are set in SplatUndef and cleared
13174 // in SplatValue. If any of the values are not constant, give up and return
13175 // false.
13176 unsigned int NumOps = getNumOperands();
13177 assert(NumOps > 0 && "isConstantSplat has 0-size build vector");
13178 unsigned EltWidth = VT.getScalarSizeInBits();
13179
13180 for (unsigned j = 0; j < NumOps; ++j) {
13181 unsigned i = IsBigEndian ? NumOps - 1 - j : j;
13182 SDValue OpVal = getOperand(Num: i);
13183 unsigned BitPos = j * EltWidth;
13184
13185 if (OpVal.isUndef())
13186 SplatUndef.setBits(loBit: BitPos, hiBit: BitPos + EltWidth);
13187 else if (auto *CN = dyn_cast<ConstantSDNode>(Val&: OpVal))
13188 SplatValue.insertBits(SubBits: CN->getAPIntValue().zextOrTrunc(width: EltWidth), bitPosition: BitPos);
13189 else if (auto *CN = dyn_cast<ConstantFPSDNode>(Val&: OpVal))
13190 SplatValue.insertBits(SubBits: CN->getValueAPF().bitcastToAPInt(), bitPosition: BitPos);
13191 else
13192 return false;
13193 }
13194
13195 // The build_vector is all constants or undefs. Find the smallest element
13196 // size that splats the vector.
13197 HasAnyUndefs = (SplatUndef != 0);
13198
13199 // FIXME: This does not work for vectors with elements less than 8 bits.
13200 while (VecWidth > 8) {
13201 // If we can't split in half, stop here.
13202 if (VecWidth & 1)
13203 break;
13204
13205 unsigned HalfSize = VecWidth / 2;
13206 APInt HighValue = SplatValue.extractBits(numBits: HalfSize, bitPosition: HalfSize);
13207 APInt LowValue = SplatValue.extractBits(numBits: HalfSize, bitPosition: 0);
13208 APInt HighUndef = SplatUndef.extractBits(numBits: HalfSize, bitPosition: HalfSize);
13209 APInt LowUndef = SplatUndef.extractBits(numBits: HalfSize, bitPosition: 0);
13210
13211 // If the two halves do not match (ignoring undef bits), stop here.
13212 if ((HighValue & ~LowUndef) != (LowValue & ~HighUndef) ||
13213 MinSplatBits > HalfSize)
13214 break;
13215
13216 SplatValue = HighValue | LowValue;
13217 SplatUndef = HighUndef & LowUndef;
13218
13219 VecWidth = HalfSize;
13220 }
13221
13222 // FIXME: The loop above only tries to split in halves. But if the input
13223 // vector for example is <3 x i16> it wouldn't be able to detect a
13224 // SplatBitSize of 16. No idea if that is a design flaw currently limiting
13225 // optimizations. I guess that back in the days when this helper was created
13226 // vectors normally was power-of-2 sized.
13227
13228 SplatBitSize = VecWidth;
13229 return true;
13230}
13231
13232SDValue BuildVectorSDNode::getSplatValue(const APInt &DemandedElts,
13233 BitVector *UndefElements) const {
13234 unsigned NumOps = getNumOperands();
13235 if (UndefElements) {
13236 UndefElements->clear();
13237 UndefElements->resize(N: NumOps);
13238 }
13239 assert(NumOps == DemandedElts.getBitWidth() && "Unexpected vector size");
13240 if (!DemandedElts)
13241 return SDValue();
13242 SDValue Splatted;
13243 for (unsigned i = 0; i != NumOps; ++i) {
13244 if (!DemandedElts[i])
13245 continue;
13246 SDValue Op = getOperand(Num: i);
13247 if (Op.isUndef()) {
13248 if (UndefElements)
13249 (*UndefElements)[i] = true;
13250 } else if (!Splatted) {
13251 Splatted = Op;
13252 } else if (Splatted != Op) {
13253 return SDValue();
13254 }
13255 }
13256
13257 if (!Splatted) {
13258 unsigned FirstDemandedIdx = DemandedElts.countr_zero();
13259 assert(getOperand(FirstDemandedIdx).isUndef() &&
13260 "Can only have a splat without a constant for all undefs.");
13261 return getOperand(Num: FirstDemandedIdx);
13262 }
13263
13264 return Splatted;
13265}
13266
13267SDValue BuildVectorSDNode::getSplatValue(BitVector *UndefElements) const {
13268 APInt DemandedElts = APInt::getAllOnes(numBits: getNumOperands());
13269 return getSplatValue(DemandedElts, UndefElements);
13270}
13271
13272bool BuildVectorSDNode::getRepeatedSequence(const APInt &DemandedElts,
13273 SmallVectorImpl<SDValue> &Sequence,
13274 BitVector *UndefElements) const {
13275 unsigned NumOps = getNumOperands();
13276 Sequence.clear();
13277 if (UndefElements) {
13278 UndefElements->clear();
13279 UndefElements->resize(N: NumOps);
13280 }
13281 assert(NumOps == DemandedElts.getBitWidth() && "Unexpected vector size");
13282 if (!DemandedElts || NumOps < 2 || !isPowerOf2_32(Value: NumOps))
13283 return false;
13284
13285 // Set the undefs even if we don't find a sequence (like getSplatValue).
13286 if (UndefElements)
13287 for (unsigned I = 0; I != NumOps; ++I)
13288 if (DemandedElts[I] && getOperand(Num: I).isUndef())
13289 (*UndefElements)[I] = true;
13290
13291 // Iteratively widen the sequence length looking for repetitions.
13292 for (unsigned SeqLen = 1; SeqLen < NumOps; SeqLen *= 2) {
13293 Sequence.append(NumInputs: SeqLen, Elt: SDValue());
13294 for (unsigned I = 0; I != NumOps; ++I) {
13295 if (!DemandedElts[I])
13296 continue;
13297 SDValue &SeqOp = Sequence[I % SeqLen];
13298 SDValue Op = getOperand(Num: I);
13299 if (Op.isUndef()) {
13300 if (!SeqOp)
13301 SeqOp = Op;
13302 continue;
13303 }
13304 if (SeqOp && !SeqOp.isUndef() && SeqOp != Op) {
13305 Sequence.clear();
13306 break;
13307 }
13308 SeqOp = Op;
13309 }
13310 if (!Sequence.empty())
13311 return true;
13312 }
13313
13314 assert(Sequence.empty() && "Failed to empty non-repeating sequence pattern");
13315 return false;
13316}
13317
13318bool BuildVectorSDNode::getRepeatedSequence(SmallVectorImpl<SDValue> &Sequence,
13319 BitVector *UndefElements) const {
13320 APInt DemandedElts = APInt::getAllOnes(numBits: getNumOperands());
13321 return getRepeatedSequence(DemandedElts, Sequence, UndefElements);
13322}
13323
13324ConstantSDNode *
13325BuildVectorSDNode::getConstantSplatNode(const APInt &DemandedElts,
13326 BitVector *UndefElements) const {
13327 return dyn_cast_or_null<ConstantSDNode>(
13328 Val: getSplatValue(DemandedElts, UndefElements));
13329}
13330
13331ConstantSDNode *
13332BuildVectorSDNode::getConstantSplatNode(BitVector *UndefElements) const {
13333 return dyn_cast_or_null<ConstantSDNode>(Val: getSplatValue(UndefElements));
13334}
13335
13336ConstantFPSDNode *
13337BuildVectorSDNode::getConstantFPSplatNode(const APInt &DemandedElts,
13338 BitVector *UndefElements) const {
13339 return dyn_cast_or_null<ConstantFPSDNode>(
13340 Val: getSplatValue(DemandedElts, UndefElements));
13341}
13342
13343ConstantFPSDNode *
13344BuildVectorSDNode::getConstantFPSplatNode(BitVector *UndefElements) const {
13345 return dyn_cast_or_null<ConstantFPSDNode>(Val: getSplatValue(UndefElements));
13346}
13347
13348int32_t
13349BuildVectorSDNode::getConstantFPSplatPow2ToLog2Int(BitVector *UndefElements,
13350 uint32_t BitWidth) const {
13351 if (ConstantFPSDNode *CN =
13352 dyn_cast_or_null<ConstantFPSDNode>(Val: getSplatValue(UndefElements))) {
13353 bool IsExact;
13354 APSInt IntVal(BitWidth);
13355 const APFloat &APF = CN->getValueAPF();
13356 if (APF.convertToInteger(Result&: IntVal, RM: APFloat::rmTowardZero, IsExact: &IsExact) !=
13357 APFloat::opOK ||
13358 !IsExact)
13359 return -1;
13360
13361 return IntVal.exactLogBase2();
13362 }
13363 return -1;
13364}
13365
13366bool BuildVectorSDNode::getConstantRawBits(
13367 bool IsLittleEndian, unsigned DstEltSizeInBits,
13368 SmallVectorImpl<APInt> &RawBitElements, BitVector &UndefElements) const {
13369 // Early-out if this contains anything but Undef/Constant/ConstantFP.
13370 if (!isConstant())
13371 return false;
13372
13373 unsigned NumSrcOps = getNumOperands();
13374 unsigned SrcEltSizeInBits = getValueType(ResNo: 0).getScalarSizeInBits();
13375 assert(((NumSrcOps * SrcEltSizeInBits) % DstEltSizeInBits) == 0 &&
13376 "Invalid bitcast scale");
13377
13378 // Extract raw src bits.
13379 SmallVector<APInt> SrcBitElements(NumSrcOps,
13380 APInt::getZero(numBits: SrcEltSizeInBits));
13381 BitVector SrcUndeElements(NumSrcOps, false);
13382
13383 for (unsigned I = 0; I != NumSrcOps; ++I) {
13384 SDValue Op = getOperand(Num: I);
13385 if (Op.isUndef()) {
13386 SrcUndeElements.set(I);
13387 continue;
13388 }
13389 auto *CInt = dyn_cast<ConstantSDNode>(Val&: Op);
13390 auto *CFP = dyn_cast<ConstantFPSDNode>(Val&: Op);
13391 assert((CInt || CFP) && "Unknown constant");
13392 SrcBitElements[I] = CInt ? CInt->getAPIntValue().trunc(width: SrcEltSizeInBits)
13393 : CFP->getValueAPF().bitcastToAPInt();
13394 }
13395
13396 // Recast to dst width.
13397 recastRawBits(IsLittleEndian, DstEltSizeInBits, DstBitElements&: RawBitElements,
13398 SrcBitElements, DstUndefElements&: UndefElements, SrcUndefElements: SrcUndeElements);
13399 return true;
13400}
13401
13402void BuildVectorSDNode::recastRawBits(bool IsLittleEndian,
13403 unsigned DstEltSizeInBits,
13404 SmallVectorImpl<APInt> &DstBitElements,
13405 ArrayRef<APInt> SrcBitElements,
13406 BitVector &DstUndefElements,
13407 const BitVector &SrcUndefElements) {
13408 unsigned NumSrcOps = SrcBitElements.size();
13409 unsigned SrcEltSizeInBits = SrcBitElements[0].getBitWidth();
13410 assert(((NumSrcOps * SrcEltSizeInBits) % DstEltSizeInBits) == 0 &&
13411 "Invalid bitcast scale");
13412 assert(NumSrcOps == SrcUndefElements.size() &&
13413 "Vector size mismatch");
13414
13415 unsigned NumDstOps = (NumSrcOps * SrcEltSizeInBits) / DstEltSizeInBits;
13416 DstUndefElements.clear();
13417 DstUndefElements.resize(N: NumDstOps, t: false);
13418 DstBitElements.assign(NumElts: NumDstOps, Elt: APInt::getZero(numBits: DstEltSizeInBits));
13419
13420 // Concatenate src elements constant bits together into dst element.
13421 if (SrcEltSizeInBits <= DstEltSizeInBits) {
13422 unsigned Scale = DstEltSizeInBits / SrcEltSizeInBits;
13423 for (unsigned I = 0; I != NumDstOps; ++I) {
13424 DstUndefElements.set(I);
13425 APInt &DstBits = DstBitElements[I];
13426 for (unsigned J = 0; J != Scale; ++J) {
13427 unsigned Idx = (I * Scale) + (IsLittleEndian ? J : (Scale - J - 1));
13428 if (SrcUndefElements[Idx])
13429 continue;
13430 DstUndefElements.reset(Idx: I);
13431 const APInt &SrcBits = SrcBitElements[Idx];
13432 assert(SrcBits.getBitWidth() == SrcEltSizeInBits &&
13433 "Illegal constant bitwidths");
13434 DstBits.insertBits(SubBits: SrcBits, bitPosition: J * SrcEltSizeInBits);
13435 }
13436 }
13437 return;
13438 }
13439
13440 // Split src element constant bits into dst elements.
13441 unsigned Scale = SrcEltSizeInBits / DstEltSizeInBits;
13442 for (unsigned I = 0; I != NumSrcOps; ++I) {
13443 if (SrcUndefElements[I]) {
13444 DstUndefElements.set(I: I * Scale, E: (I + 1) * Scale);
13445 continue;
13446 }
13447 const APInt &SrcBits = SrcBitElements[I];
13448 for (unsigned J = 0; J != Scale; ++J) {
13449 unsigned Idx = (I * Scale) + (IsLittleEndian ? J : (Scale - J - 1));
13450 APInt &DstBits = DstBitElements[Idx];
13451 DstBits = SrcBits.extractBits(numBits: DstEltSizeInBits, bitPosition: J * DstEltSizeInBits);
13452 }
13453 }
13454}
13455
13456bool BuildVectorSDNode::isConstant() const {
13457 for (const SDValue &Op : op_values()) {
13458 unsigned Opc = Op.getOpcode();
13459 if (!Op.isUndef() && Opc != ISD::Constant && Opc != ISD::ConstantFP)
13460 return false;
13461 }
13462 return true;
13463}
13464
13465std::optional<std::pair<APInt, APInt>>
13466BuildVectorSDNode::isConstantSequence() const {
13467 unsigned NumOps = getNumOperands();
13468 if (NumOps < 2)
13469 return std::nullopt;
13470
13471 if (!isa<ConstantSDNode>(Val: getOperand(Num: 0)) ||
13472 !isa<ConstantSDNode>(Val: getOperand(Num: 1)))
13473 return std::nullopt;
13474
13475 unsigned EltSize = getValueType(ResNo: 0).getScalarSizeInBits();
13476 APInt Start = getConstantOperandAPInt(Num: 0).trunc(width: EltSize);
13477 APInt Stride = getConstantOperandAPInt(Num: 1).trunc(width: EltSize) - Start;
13478
13479 if (Stride.isZero())
13480 return std::nullopt;
13481
13482 for (unsigned i = 2; i < NumOps; ++i) {
13483 if (!isa<ConstantSDNode>(Val: getOperand(Num: i)))
13484 return std::nullopt;
13485
13486 APInt Val = getConstantOperandAPInt(Num: i).trunc(width: EltSize);
13487 if (Val != (Start + (Stride * i)))
13488 return std::nullopt;
13489 }
13490
13491 return std::make_pair(x&: Start, y&: Stride);
13492}
13493
13494bool ShuffleVectorSDNode::isSplatMask(ArrayRef<int> Mask) {
13495 // Find the first non-undef value in the shuffle mask.
13496 unsigned i, e;
13497 for (i = 0, e = Mask.size(); i != e && Mask[i] < 0; ++i)
13498 /* search */;
13499
13500 // If all elements are undefined, this shuffle can be considered a splat
13501 // (although it should eventually get simplified away completely).
13502 if (i == e)
13503 return true;
13504
13505 // Make sure all remaining elements are either undef or the same as the first
13506 // non-undef value.
13507 for (int Idx = Mask[i]; i != e; ++i)
13508 if (Mask[i] >= 0 && Mask[i] != Idx)
13509 return false;
13510 return true;
13511}
13512
13513// Returns true if it is a constant integer BuildVector or constant integer,
13514// possibly hidden by a bitcast.
13515bool SelectionDAG::isConstantIntBuildVectorOrConstantInt(
13516 SDValue N, bool AllowOpaques) const {
13517 N = peekThroughBitcasts(V: N);
13518
13519 if (auto *C = dyn_cast<ConstantSDNode>(Val&: N))
13520 return AllowOpaques || !C->isOpaque();
13521
13522 if (ISD::isBuildVectorOfConstantSDNodes(N: N.getNode()))
13523 return true;
13524
13525 // Treat a GlobalAddress supporting constant offset folding as a
13526 // constant integer.
13527 if (auto *GA = dyn_cast<GlobalAddressSDNode>(Val&: N))
13528 if (GA->getOpcode() == ISD::GlobalAddress &&
13529 TLI->isOffsetFoldingLegal(GA))
13530 return true;
13531
13532 if ((N.getOpcode() == ISD::SPLAT_VECTOR) &&
13533 isa<ConstantSDNode>(Val: N.getOperand(i: 0)))
13534 return true;
13535 return false;
13536}
13537
13538// Returns true if it is a constant float BuildVector or constant float.
13539bool SelectionDAG::isConstantFPBuildVectorOrConstantFP(SDValue N) const {
13540 if (isa<ConstantFPSDNode>(Val: N))
13541 return true;
13542
13543 if (ISD::isBuildVectorOfConstantFPSDNodes(N: N.getNode()))
13544 return true;
13545
13546 if ((N.getOpcode() == ISD::SPLAT_VECTOR) &&
13547 isa<ConstantFPSDNode>(Val: N.getOperand(i: 0)))
13548 return true;
13549
13550 return false;
13551}
13552
13553std::optional<bool> SelectionDAG::isBoolConstant(SDValue N,
13554 bool AllowTruncation) const {
13555 ConstantSDNode *Const = isConstOrConstSplat(N, AllowUndefs: false, AllowTruncation);
13556 if (!Const)
13557 return std::nullopt;
13558
13559 const APInt &CVal = Const->getAPIntValue();
13560 switch (TLI->getBooleanContents(Type: N.getValueType())) {
13561 case TargetLowering::ZeroOrOneBooleanContent:
13562 if (CVal.isOne())
13563 return true;
13564 if (CVal.isZero())
13565 return false;
13566 return std::nullopt;
13567 case TargetLowering::ZeroOrNegativeOneBooleanContent:
13568 if (CVal.isAllOnes())
13569 return true;
13570 if (CVal.isZero())
13571 return false;
13572 return std::nullopt;
13573 case TargetLowering::UndefinedBooleanContent:
13574 return CVal[0];
13575 }
13576 llvm_unreachable("Unknown BooleanContent enum");
13577}
13578
13579void SelectionDAG::createOperands(SDNode *Node, ArrayRef<SDValue> Vals) {
13580 assert(!Node->OperandList && "Node already has operands");
13581 assert(SDNode::getMaxNumOperands() >= Vals.size() &&
13582 "too many operands to fit into SDNode");
13583 SDUse *Ops = OperandRecycler.allocate(
13584 Cap: ArrayRecycler<SDUse>::Capacity::get(N: Vals.size()), Allocator&: OperandAllocator);
13585
13586 bool IsDivergent = false;
13587 for (unsigned I = 0; I != Vals.size(); ++I) {
13588 Ops[I].setUser(Node);
13589 Ops[I].setInitial(Vals[I]);
13590 EVT VT = Ops[I].getValueType();
13591
13592 // Skip Chain. It does not carry divergence.
13593 if (VT != MVT::Other &&
13594 (VT != MVT::Glue || gluePropagatesDivergence(Ops[I].getNode())) &&
13595 Ops[I].getNode()->isDivergent()) {
13596 IsDivergent = true;
13597 }
13598 }
13599 Node->NumOperands = Vals.size();
13600 Node->OperandList = Ops;
13601 if (!TLI->isSDNodeAlwaysUniform(N: Node)) {
13602 IsDivergent |= TLI->isSDNodeSourceOfDivergence(N: Node, FLI, UA);
13603 Node->SDNodeBits.IsDivergent = IsDivergent;
13604 }
13605 checkForCycles(N: Node);
13606}
13607
13608SDValue SelectionDAG::getTokenFactor(const SDLoc &DL,
13609 SmallVectorImpl<SDValue> &Vals) {
13610 size_t Limit = SDNode::getMaxNumOperands();
13611 while (Vals.size() > Limit) {
13612 unsigned SliceIdx = Vals.size() - Limit;
13613 auto ExtractedTFs = ArrayRef<SDValue>(Vals).slice(N: SliceIdx, M: Limit);
13614 SDValue NewTF = getNode(ISD::TokenFactor, DL, MVT::Other, ExtractedTFs);
13615 Vals.erase(CS: Vals.begin() + SliceIdx, CE: Vals.end());
13616 Vals.emplace_back(Args&: NewTF);
13617 }
13618 return getNode(ISD::TokenFactor, DL, MVT::Other, Vals);
13619}
13620
13621SDValue SelectionDAG::getNeutralElement(unsigned Opcode, const SDLoc &DL,
13622 EVT VT, SDNodeFlags Flags) {
13623 switch (Opcode) {
13624 default:
13625 return SDValue();
13626 case ISD::ADD:
13627 case ISD::OR:
13628 case ISD::XOR:
13629 case ISD::UMAX:
13630 return getConstant(Val: 0, DL, VT);
13631 case ISD::MUL:
13632 return getConstant(Val: 1, DL, VT);
13633 case ISD::AND:
13634 case ISD::UMIN:
13635 return getAllOnesConstant(DL, VT);
13636 case ISD::SMAX:
13637 return getConstant(Val: APInt::getSignedMinValue(numBits: VT.getSizeInBits()), DL, VT);
13638 case ISD::SMIN:
13639 return getConstant(Val: APInt::getSignedMaxValue(numBits: VT.getSizeInBits()), DL, VT);
13640 case ISD::FADD:
13641 // If flags allow, prefer positive zero since it's generally cheaper
13642 // to materialize on most targets.
13643 return getConstantFP(Val: Flags.hasNoSignedZeros() ? 0.0 : -0.0, DL, VT);
13644 case ISD::FMUL:
13645 return getConstantFP(Val: 1.0, DL, VT);
13646 case ISD::FMINNUM:
13647 case ISD::FMAXNUM: {
13648 // Neutral element for fminnum is NaN, Inf or FLT_MAX, depending on FMF.
13649 const fltSemantics &Semantics = VT.getFltSemantics();
13650 APFloat NeutralAF = !Flags.hasNoNaNs() ? APFloat::getQNaN(Sem: Semantics) :
13651 !Flags.hasNoInfs() ? APFloat::getInf(Sem: Semantics) :
13652 APFloat::getLargest(Sem: Semantics);
13653 if (Opcode == ISD::FMAXNUM)
13654 NeutralAF.changeSign();
13655
13656 return getConstantFP(V: NeutralAF, DL, VT);
13657 }
13658 case ISD::FMINIMUM:
13659 case ISD::FMAXIMUM: {
13660 // Neutral element for fminimum is Inf or FLT_MAX, depending on FMF.
13661 const fltSemantics &Semantics = VT.getFltSemantics();
13662 APFloat NeutralAF = !Flags.hasNoInfs() ? APFloat::getInf(Sem: Semantics)
13663 : APFloat::getLargest(Sem: Semantics);
13664 if (Opcode == ISD::FMAXIMUM)
13665 NeutralAF.changeSign();
13666
13667 return getConstantFP(V: NeutralAF, DL, VT);
13668 }
13669
13670 }
13671}
13672
13673/// Helper used to make a call to a library function that has one argument of
13674/// pointer type.
13675///
13676/// Such functions include 'fegetmode', 'fesetenv' and some others, which are
13677/// used to get or set floating-point state. They have one argument of pointer
13678/// type, which points to the memory region containing bits of the
13679/// floating-point state. The value returned by such function is ignored in the
13680/// created call.
13681///
13682/// \param LibFunc Reference to library function (value of RTLIB::Libcall).
13683/// \param Ptr Pointer used to save/load state.
13684/// \param InChain Ingoing token chain.
13685/// \returns Outgoing chain token.
13686SDValue SelectionDAG::makeStateFunctionCall(unsigned LibFunc, SDValue Ptr,
13687 SDValue InChain,
13688 const SDLoc &DLoc) {
13689 assert(InChain.getValueType() == MVT::Other && "Expected token chain");
13690 TargetLowering::ArgListTy Args;
13691 TargetLowering::ArgListEntry Entry;
13692 Entry.Node = Ptr;
13693 Entry.Ty = Ptr.getValueType().getTypeForEVT(Context&: *getContext());
13694 Args.push_back(x: Entry);
13695 RTLIB::Libcall LC = static_cast<RTLIB::Libcall>(LibFunc);
13696 SDValue Callee = getExternalSymbol(Sym: TLI->getLibcallName(Call: LC),
13697 VT: TLI->getPointerTy(DL: getDataLayout()));
13698 TargetLowering::CallLoweringInfo CLI(*this);
13699 CLI.setDebugLoc(DLoc).setChain(InChain).setLibCallee(
13700 CC: TLI->getLibcallCallingConv(Call: LC), ResultType: Type::getVoidTy(C&: *getContext()), Target: Callee,
13701 ArgsList: std::move(Args));
13702 return TLI->LowerCallTo(CLI).second;
13703}
13704
13705void SelectionDAG::copyExtraInfo(SDNode *From, SDNode *To) {
13706 assert(From && To && "Invalid SDNode; empty source SDValue?");
13707 auto I = SDEI.find(Val: From);
13708 if (I == SDEI.end())
13709 return;
13710
13711 // Use of operator[] on the DenseMap may cause an insertion, which invalidates
13712 // the iterator, hence the need to make a copy to prevent a use-after-free.
13713 NodeExtraInfo NEI = I->second;
13714 if (LLVM_LIKELY(!NEI.PCSections)) {
13715 // No deep copy required for the types of extra info set.
13716 //
13717 // FIXME: Investigate if other types of extra info also need deep copy. This
13718 // depends on the types of nodes they can be attached to: if some extra info
13719 // is only ever attached to nodes where a replacement To node is always the
13720 // node where later use and propagation of the extra info has the intended
13721 // semantics, no deep copy is required.
13722 SDEI[To] = std::move(NEI);
13723 return;
13724 }
13725
13726 // We need to copy NodeExtraInfo to all _new_ nodes that are being introduced
13727 // through the replacement of From with To. Otherwise, replacements of a node
13728 // (From) with more complex nodes (To and its operands) may result in lost
13729 // extra info where the root node (To) is insignificant in further propagating
13730 // and using extra info when further lowering to MIR.
13731 //
13732 // In the first step pre-populate the visited set with the nodes reachable
13733 // from the old From node. This avoids copying NodeExtraInfo to parts of the
13734 // DAG that is not new and should be left untouched.
13735 SmallVector<const SDNode *> Leafs{From}; // Leafs reachable with VisitFrom.
13736 DenseSet<const SDNode *> FromReach; // The set of nodes reachable from From.
13737 auto VisitFrom = [&](auto &&Self, const SDNode *N, int MaxDepth) {
13738 if (MaxDepth == 0) {
13739 // Remember this node in case we need to increase MaxDepth and continue
13740 // populating FromReach from this node.
13741 Leafs.emplace_back(Args&: N);
13742 return;
13743 }
13744 if (!FromReach.insert(V: N).second)
13745 return;
13746 for (const SDValue &Op : N->op_values())
13747 Self(Self, Op.getNode(), MaxDepth - 1);
13748 };
13749
13750 // Copy extra info to To and all its transitive operands (that are new).
13751 SmallPtrSet<const SDNode *, 8> Visited;
13752 auto DeepCopyTo = [&](auto &&Self, const SDNode *N) {
13753 if (FromReach.contains(V: N))
13754 return true;
13755 if (!Visited.insert(Ptr: N).second)
13756 return true;
13757 if (getEntryNode().getNode() == N)
13758 return false;
13759 for (const SDValue &Op : N->op_values()) {
13760 if (!Self(Self, Op.getNode()))
13761 return false;
13762 }
13763 // Copy only if entry node was not reached.
13764 SDEI[N] = NEI;
13765 return true;
13766 };
13767
13768 // We first try with a lower MaxDepth, assuming that the path to common
13769 // operands between From and To is relatively short. This significantly
13770 // improves performance in the common case. The initial MaxDepth is big
13771 // enough to avoid retry in the common case; the last MaxDepth is large
13772 // enough to avoid having to use the fallback below (and protects from
13773 // potential stack exhaustion from recursion).
13774 for (int PrevDepth = 0, MaxDepth = 16; MaxDepth <= 1024;
13775 PrevDepth = MaxDepth, MaxDepth *= 2, Visited.clear()) {
13776 // StartFrom is the previous (or initial) set of leafs reachable at the
13777 // previous maximum depth.
13778 SmallVector<const SDNode *> StartFrom;
13779 std::swap(LHS&: StartFrom, RHS&: Leafs);
13780 for (const SDNode *N : StartFrom)
13781 VisitFrom(VisitFrom, N, MaxDepth - PrevDepth);
13782 if (LLVM_LIKELY(DeepCopyTo(DeepCopyTo, To)))
13783 return;
13784 // This should happen very rarely (reached the entry node).
13785 LLVM_DEBUG(dbgs() << __func__ << ": MaxDepth=" << MaxDepth << " too low\n");
13786 assert(!Leafs.empty());
13787 }
13788
13789 // This should not happen - but if it did, that means the subgraph reachable
13790 // from From has depth greater or equal to maximum MaxDepth, and VisitFrom()
13791 // could not visit all reachable common operands. Consequently, we were able
13792 // to reach the entry node.
13793 errs() << "warning: incomplete propagation of SelectionDAG::NodeExtraInfo\n";
13794 assert(false && "From subgraph too complex - increase max. MaxDepth?");
13795 // Best-effort fallback if assertions disabled.
13796 SDEI[To] = std::move(NEI);
13797}
13798
13799#ifndef NDEBUG
13800static void checkForCyclesHelper(const SDNode *N,
13801 SmallPtrSetImpl<const SDNode*> &Visited,
13802 SmallPtrSetImpl<const SDNode*> &Checked,
13803 const llvm::SelectionDAG *DAG) {
13804 // If this node has already been checked, don't check it again.
13805 if (Checked.count(Ptr: N))
13806 return;
13807
13808 // If a node has already been visited on this depth-first walk, reject it as
13809 // a cycle.
13810 if (!Visited.insert(Ptr: N).second) {
13811 errs() << "Detected cycle in SelectionDAG\n";
13812 dbgs() << "Offending node:\n";
13813 N->dumprFull(G: DAG); dbgs() << "\n";
13814 abort();
13815 }
13816
13817 for (const SDValue &Op : N->op_values())
13818 checkForCyclesHelper(N: Op.getNode(), Visited, Checked, DAG);
13819
13820 Checked.insert(Ptr: N);
13821 Visited.erase(Ptr: N);
13822}
13823#endif
13824
13825void llvm::checkForCycles(const llvm::SDNode *N,
13826 const llvm::SelectionDAG *DAG,
13827 bool force) {
13828#ifndef NDEBUG
13829 bool check = force;
13830#ifdef EXPENSIVE_CHECKS
13831 check = true;
13832#endif // EXPENSIVE_CHECKS
13833 if (check) {
13834 assert(N && "Checking nonexistent SDNode");
13835 SmallPtrSet<const SDNode*, 32> visited;
13836 SmallPtrSet<const SDNode*, 32> checked;
13837 checkForCyclesHelper(N, Visited&: visited, Checked&: checked, DAG);
13838 }
13839#endif // !NDEBUG
13840}
13841
13842void llvm::checkForCycles(const llvm::SelectionDAG *DAG, bool force) {
13843 checkForCycles(N: DAG->getRoot().getNode(), DAG, force);
13844}
13845

Provided by KDAB

Privacy Policy
Learn to use CMake with our Intro Training
Find out more

source code of llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp