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/FoldingSet.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/Triple.h"
25 #include "llvm/ADT/Twine.h"
26 #include "llvm/Analysis/AliasAnalysis.h"
27 #include "llvm/Analysis/MemoryLocation.h"
28 #include "llvm/Analysis/ValueTracking.h"
29 #include "llvm/Analysis/VectorUtils.h"
30 #include "llvm/CodeGen/Analysis.h"
31 #include "llvm/CodeGen/FunctionLoweringInfo.h"
32 #include "llvm/CodeGen/ISDOpcodes.h"
33 #include "llvm/CodeGen/MachineBasicBlock.h"
34 #include "llvm/CodeGen/MachineConstantPool.h"
35 #include "llvm/CodeGen/MachineFrameInfo.h"
36 #include "llvm/CodeGen/MachineFunction.h"
37 #include "llvm/CodeGen/MachineMemOperand.h"
38 #include "llvm/CodeGen/RuntimeLibcalls.h"
39 #include "llvm/CodeGen/SelectionDAGAddressAnalysis.h"
40 #include "llvm/CodeGen/SelectionDAGNodes.h"
41 #include "llvm/CodeGen/SelectionDAGTargetInfo.h"
42 #include "llvm/CodeGen/TargetFrameLowering.h"
43 #include "llvm/CodeGen/TargetLowering.h"
44 #include "llvm/CodeGen/TargetRegisterInfo.h"
45 #include "llvm/CodeGen/TargetSubtargetInfo.h"
46 #include "llvm/CodeGen/ValueTypes.h"
47 #include "llvm/IR/Constant.h"
48 #include "llvm/IR/ConstantRange.h"
49 #include "llvm/IR/Constants.h"
50 #include "llvm/IR/DataLayout.h"
51 #include "llvm/IR/DebugInfoMetadata.h"
52 #include "llvm/IR/DebugLoc.h"
53 #include "llvm/IR/DerivedTypes.h"
54 #include "llvm/IR/Function.h"
55 #include "llvm/IR/GlobalValue.h"
56 #include "llvm/IR/Metadata.h"
57 #include "llvm/IR/Type.h"
58 #include "llvm/Support/Casting.h"
59 #include "llvm/Support/CodeGen.h"
60 #include "llvm/Support/Compiler.h"
61 #include "llvm/Support/Debug.h"
62 #include "llvm/Support/ErrorHandling.h"
63 #include "llvm/Support/KnownBits.h"
64 #include "llvm/Support/MachineValueType.h"
65 #include "llvm/Support/MathExtras.h"
66 #include "llvm/Support/Mutex.h"
67 #include "llvm/Support/raw_ostream.h"
68 #include "llvm/Target/TargetMachine.h"
69 #include "llvm/Target/TargetOptions.h"
70 #include "llvm/Transforms/Utils/SizeOpts.h"
71 #include <algorithm>
72 #include <cassert>
73 #include <cstdint>
74 #include <cstdlib>
75 #include <limits>
76 #include <set>
77 #include <string>
78 #include <utility>
79 #include <vector>
80
81 using namespace llvm;
82
83 /// makeVTList - Return an instance of the SDVTList struct initialized with the
84 /// specified members.
makeVTList(const EVT * VTs,unsigned NumVTs)85 static SDVTList makeVTList(const EVT *VTs, unsigned NumVTs) {
86 SDVTList Res = {VTs, NumVTs};
87 return Res;
88 }
89
90 // Default null implementations of the callbacks.
NodeDeleted(SDNode *,SDNode *)91 void SelectionDAG::DAGUpdateListener::NodeDeleted(SDNode*, SDNode*) {}
NodeUpdated(SDNode *)92 void SelectionDAG::DAGUpdateListener::NodeUpdated(SDNode*) {}
NodeInserted(SDNode *)93 void SelectionDAG::DAGUpdateListener::NodeInserted(SDNode *) {}
94
anchor()95 void SelectionDAG::DAGNodeDeletedListener::anchor() {}
anchor()96 void SelectionDAG::DAGNodeInsertedListener::anchor() {}
97
98 #define DEBUG_TYPE "selectiondag"
99
100 static cl::opt<bool> EnableMemCpyDAGOpt("enable-memcpy-dag-opt",
101 cl::Hidden, cl::init(true),
102 cl::desc("Gang up loads and stores generated by inlining of memcpy"));
103
104 static cl::opt<int> MaxLdStGlue("ldstmemcpy-glue-max",
105 cl::desc("Number limit for gluing ld/st of memcpy."),
106 cl::Hidden, cl::init(0));
107
NewSDValueDbgMsg(SDValue V,StringRef Msg,SelectionDAG * G)108 static void NewSDValueDbgMsg(SDValue V, StringRef Msg, SelectionDAG *G) {
109 LLVM_DEBUG(dbgs() << Msg; V.getNode()->dump(G););
110 }
111
112 //===----------------------------------------------------------------------===//
113 // ConstantFPSDNode Class
114 //===----------------------------------------------------------------------===//
115
116 /// isExactlyValue - We don't rely on operator== working on double values, as
117 /// it returns true for things that are clearly not equal, like -0.0 and 0.0.
118 /// As such, this method can be used to do an exact bit-for-bit comparison of
119 /// two floating point values.
isExactlyValue(const APFloat & V) const120 bool ConstantFPSDNode::isExactlyValue(const APFloat& V) const {
121 return getValueAPF().bitwiseIsEqual(V);
122 }
123
isValueValidForType(EVT VT,const APFloat & Val)124 bool ConstantFPSDNode::isValueValidForType(EVT VT,
125 const APFloat& Val) {
126 assert(VT.isFloatingPoint() && "Can only convert between FP types");
127
128 // convert modifies in place, so make a copy.
129 APFloat Val2 = APFloat(Val);
130 bool losesInfo;
131 (void) Val2.convert(SelectionDAG::EVTToAPFloatSemantics(VT),
132 APFloat::rmNearestTiesToEven,
133 &losesInfo);
134 return !losesInfo;
135 }
136
137 //===----------------------------------------------------------------------===//
138 // ISD Namespace
139 //===----------------------------------------------------------------------===//
140
isConstantSplatVector(const SDNode * N,APInt & SplatVal)141 bool ISD::isConstantSplatVector(const SDNode *N, APInt &SplatVal) {
142 if (N->getOpcode() == ISD::SPLAT_VECTOR) {
143 unsigned EltSize =
144 N->getValueType(0).getVectorElementType().getSizeInBits();
145 if (auto *Op0 = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
146 SplatVal = Op0->getAPIntValue().trunc(EltSize);
147 return true;
148 }
149 if (auto *Op0 = dyn_cast<ConstantFPSDNode>(N->getOperand(0))) {
150 SplatVal = Op0->getValueAPF().bitcastToAPInt().trunc(EltSize);
151 return true;
152 }
153 }
154
155 auto *BV = dyn_cast<BuildVectorSDNode>(N);
156 if (!BV)
157 return false;
158
159 APInt SplatUndef;
160 unsigned SplatBitSize;
161 bool HasUndefs;
162 unsigned EltSize = N->getValueType(0).getVectorElementType().getSizeInBits();
163 return BV->isConstantSplat(SplatVal, SplatUndef, SplatBitSize, HasUndefs,
164 EltSize) &&
165 EltSize == SplatBitSize;
166 }
167
168 // FIXME: AllOnes and AllZeros duplicate a lot of code. Could these be
169 // specializations of the more general isConstantSplatVector()?
170
isConstantSplatVectorAllOnes(const SDNode * N,bool BuildVectorOnly)171 bool ISD::isConstantSplatVectorAllOnes(const SDNode *N, bool BuildVectorOnly) {
172 // Look through a bit convert.
173 while (N->getOpcode() == ISD::BITCAST)
174 N = N->getOperand(0).getNode();
175
176 if (!BuildVectorOnly && N->getOpcode() == ISD::SPLAT_VECTOR) {
177 APInt SplatVal;
178 return isConstantSplatVector(N, SplatVal) && SplatVal.isAllOnes();
179 }
180
181 if (N->getOpcode() != ISD::BUILD_VECTOR) return false;
182
183 unsigned i = 0, e = N->getNumOperands();
184
185 // Skip over all of the undef values.
186 while (i != e && N->getOperand(i).isUndef())
187 ++i;
188
189 // Do not accept an all-undef vector.
190 if (i == e) return false;
191
192 // Do not accept build_vectors that aren't all constants or which have non-~0
193 // elements. We have to be a bit careful here, as the type of the constant
194 // may not be the same as the type of the vector elements due to type
195 // legalization (the elements are promoted to a legal type for the target and
196 // a vector of a type may be legal when the base element type is not).
197 // We only want to check enough bits to cover the vector elements, because
198 // we care if the resultant vector is all ones, not whether the individual
199 // constants are.
200 SDValue NotZero = N->getOperand(i);
201 unsigned EltSize = N->getValueType(0).getScalarSizeInBits();
202 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(NotZero)) {
203 if (CN->getAPIntValue().countTrailingOnes() < EltSize)
204 return false;
205 } else if (ConstantFPSDNode *CFPN = dyn_cast<ConstantFPSDNode>(NotZero)) {
206 if (CFPN->getValueAPF().bitcastToAPInt().countTrailingOnes() < EltSize)
207 return false;
208 } else
209 return false;
210
211 // Okay, we have at least one ~0 value, check to see if the rest match or are
212 // undefs. Even with the above element type twiddling, this should be OK, as
213 // the same type legalization should have applied to all the elements.
214 for (++i; i != e; ++i)
215 if (N->getOperand(i) != NotZero && !N->getOperand(i).isUndef())
216 return false;
217 return true;
218 }
219
isConstantSplatVectorAllZeros(const SDNode * N,bool BuildVectorOnly)220 bool ISD::isConstantSplatVectorAllZeros(const SDNode *N, bool BuildVectorOnly) {
221 // Look through a bit convert.
222 while (N->getOpcode() == ISD::BITCAST)
223 N = N->getOperand(0).getNode();
224
225 if (!BuildVectorOnly && N->getOpcode() == ISD::SPLAT_VECTOR) {
226 APInt SplatVal;
227 return isConstantSplatVector(N, SplatVal) && SplatVal.isZero();
228 }
229
230 if (N->getOpcode() != ISD::BUILD_VECTOR) return false;
231
232 bool IsAllUndef = true;
233 for (const SDValue &Op : N->op_values()) {
234 if (Op.isUndef())
235 continue;
236 IsAllUndef = false;
237 // Do not accept build_vectors that aren't all constants or which have non-0
238 // elements. We have to be a bit careful here, as the type of the constant
239 // may not be the same as the type of the vector elements due to type
240 // legalization (the elements are promoted to a legal type for the target
241 // and a vector of a type may be legal when the base element type is not).
242 // We only want to check enough bits to cover the vector elements, because
243 // we care if the resultant vector is all zeros, not whether the individual
244 // constants are.
245 unsigned EltSize = N->getValueType(0).getScalarSizeInBits();
246 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op)) {
247 if (CN->getAPIntValue().countTrailingZeros() < EltSize)
248 return false;
249 } else if (ConstantFPSDNode *CFPN = dyn_cast<ConstantFPSDNode>(Op)) {
250 if (CFPN->getValueAPF().bitcastToAPInt().countTrailingZeros() < EltSize)
251 return false;
252 } else
253 return false;
254 }
255
256 // Do not accept an all-undef vector.
257 if (IsAllUndef)
258 return false;
259 return true;
260 }
261
isBuildVectorAllOnes(const SDNode * N)262 bool ISD::isBuildVectorAllOnes(const SDNode *N) {
263 return isConstantSplatVectorAllOnes(N, /*BuildVectorOnly*/ true);
264 }
265
isBuildVectorAllZeros(const SDNode * N)266 bool ISD::isBuildVectorAllZeros(const SDNode *N) {
267 return isConstantSplatVectorAllZeros(N, /*BuildVectorOnly*/ true);
268 }
269
isBuildVectorOfConstantSDNodes(const SDNode * N)270 bool ISD::isBuildVectorOfConstantSDNodes(const SDNode *N) {
271 if (N->getOpcode() != ISD::BUILD_VECTOR)
272 return false;
273
274 for (const SDValue &Op : N->op_values()) {
275 if (Op.isUndef())
276 continue;
277 if (!isa<ConstantSDNode>(Op))
278 return false;
279 }
280 return true;
281 }
282
isBuildVectorOfConstantFPSDNodes(const SDNode * N)283 bool ISD::isBuildVectorOfConstantFPSDNodes(const SDNode *N) {
284 if (N->getOpcode() != ISD::BUILD_VECTOR)
285 return false;
286
287 for (const SDValue &Op : N->op_values()) {
288 if (Op.isUndef())
289 continue;
290 if (!isa<ConstantFPSDNode>(Op))
291 return false;
292 }
293 return true;
294 }
295
isVectorShrinkable(const SDNode * N,unsigned NewEltSize,bool Signed)296 bool ISD::isVectorShrinkable(const SDNode *N, unsigned NewEltSize,
297 bool Signed) {
298 assert(N->getValueType(0).isVector() && "Expected a vector!");
299
300 unsigned EltSize = N->getValueType(0).getScalarSizeInBits();
301 if (EltSize <= NewEltSize)
302 return false;
303
304 if (N->getOpcode() == ISD::ZERO_EXTEND) {
305 return (N->getOperand(0).getValueType().getScalarSizeInBits() <=
306 NewEltSize) &&
307 !Signed;
308 }
309 if (N->getOpcode() == ISD::SIGN_EXTEND) {
310 return (N->getOperand(0).getValueType().getScalarSizeInBits() <=
311 NewEltSize) &&
312 Signed;
313 }
314 if (N->getOpcode() != ISD::BUILD_VECTOR)
315 return false;
316
317 for (const SDValue &Op : N->op_values()) {
318 if (Op.isUndef())
319 continue;
320 if (!isa<ConstantSDNode>(Op))
321 return false;
322
323 APInt C = cast<ConstantSDNode>(Op)->getAPIntValue().trunc(EltSize);
324 if (Signed && C.trunc(NewEltSize).sext(EltSize) != C)
325 return false;
326 if (!Signed && C.trunc(NewEltSize).zext(EltSize) != C)
327 return false;
328 }
329
330 return true;
331 }
332
allOperandsUndef(const SDNode * N)333 bool ISD::allOperandsUndef(const SDNode *N) {
334 // Return false if the node has no operands.
335 // This is "logically inconsistent" with the definition of "all" but
336 // is probably the desired behavior.
337 if (N->getNumOperands() == 0)
338 return false;
339 return all_of(N->op_values(), [](SDValue Op) { return Op.isUndef(); });
340 }
341
isFreezeUndef(const SDNode * N)342 bool ISD::isFreezeUndef(const SDNode *N) {
343 return N->getOpcode() == ISD::FREEZE && N->getOperand(0).isUndef();
344 }
345
matchUnaryPredicate(SDValue Op,std::function<bool (ConstantSDNode *)> Match,bool AllowUndefs)346 bool ISD::matchUnaryPredicate(SDValue Op,
347 std::function<bool(ConstantSDNode *)> Match,
348 bool AllowUndefs) {
349 // FIXME: Add support for scalar UNDEF cases?
350 if (auto *Cst = dyn_cast<ConstantSDNode>(Op))
351 return Match(Cst);
352
353 // FIXME: Add support for vector UNDEF cases?
354 if (ISD::BUILD_VECTOR != Op.getOpcode() &&
355 ISD::SPLAT_VECTOR != Op.getOpcode())
356 return false;
357
358 EVT SVT = Op.getValueType().getScalarType();
359 for (unsigned i = 0, e = Op.getNumOperands(); i != e; ++i) {
360 if (AllowUndefs && Op.getOperand(i).isUndef()) {
361 if (!Match(nullptr))
362 return false;
363 continue;
364 }
365
366 auto *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(i));
367 if (!Cst || Cst->getValueType(0) != SVT || !Match(Cst))
368 return false;
369 }
370 return true;
371 }
372
matchBinaryPredicate(SDValue LHS,SDValue RHS,std::function<bool (ConstantSDNode *,ConstantSDNode *)> Match,bool AllowUndefs,bool AllowTypeMismatch)373 bool ISD::matchBinaryPredicate(
374 SDValue LHS, SDValue RHS,
375 std::function<bool(ConstantSDNode *, ConstantSDNode *)> Match,
376 bool AllowUndefs, bool AllowTypeMismatch) {
377 if (!AllowTypeMismatch && LHS.getValueType() != RHS.getValueType())
378 return false;
379
380 // TODO: Add support for scalar UNDEF cases?
381 if (auto *LHSCst = dyn_cast<ConstantSDNode>(LHS))
382 if (auto *RHSCst = dyn_cast<ConstantSDNode>(RHS))
383 return Match(LHSCst, RHSCst);
384
385 // TODO: Add support for vector UNDEF cases?
386 if (LHS.getOpcode() != RHS.getOpcode() ||
387 (LHS.getOpcode() != ISD::BUILD_VECTOR &&
388 LHS.getOpcode() != ISD::SPLAT_VECTOR))
389 return false;
390
391 EVT SVT = LHS.getValueType().getScalarType();
392 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
393 SDValue LHSOp = LHS.getOperand(i);
394 SDValue RHSOp = RHS.getOperand(i);
395 bool LHSUndef = AllowUndefs && LHSOp.isUndef();
396 bool RHSUndef = AllowUndefs && RHSOp.isUndef();
397 auto *LHSCst = dyn_cast<ConstantSDNode>(LHSOp);
398 auto *RHSCst = dyn_cast<ConstantSDNode>(RHSOp);
399 if ((!LHSCst && !LHSUndef) || (!RHSCst && !RHSUndef))
400 return false;
401 if (!AllowTypeMismatch && (LHSOp.getValueType() != SVT ||
402 LHSOp.getValueType() != RHSOp.getValueType()))
403 return false;
404 if (!Match(LHSCst, RHSCst))
405 return false;
406 }
407 return true;
408 }
409
getVecReduceBaseOpcode(unsigned VecReduceOpcode)410 ISD::NodeType ISD::getVecReduceBaseOpcode(unsigned VecReduceOpcode) {
411 switch (VecReduceOpcode) {
412 default:
413 llvm_unreachable("Expected VECREDUCE opcode");
414 case ISD::VECREDUCE_FADD:
415 case ISD::VECREDUCE_SEQ_FADD:
416 case ISD::VP_REDUCE_FADD:
417 case ISD::VP_REDUCE_SEQ_FADD:
418 return ISD::FADD;
419 case ISD::VECREDUCE_FMUL:
420 case ISD::VECREDUCE_SEQ_FMUL:
421 case ISD::VP_REDUCE_FMUL:
422 case ISD::VP_REDUCE_SEQ_FMUL:
423 return ISD::FMUL;
424 case ISD::VECREDUCE_ADD:
425 case ISD::VP_REDUCE_ADD:
426 return ISD::ADD;
427 case ISD::VECREDUCE_MUL:
428 case ISD::VP_REDUCE_MUL:
429 return ISD::MUL;
430 case ISD::VECREDUCE_AND:
431 case ISD::VP_REDUCE_AND:
432 return ISD::AND;
433 case ISD::VECREDUCE_OR:
434 case ISD::VP_REDUCE_OR:
435 return ISD::OR;
436 case ISD::VECREDUCE_XOR:
437 case ISD::VP_REDUCE_XOR:
438 return ISD::XOR;
439 case ISD::VECREDUCE_SMAX:
440 case ISD::VP_REDUCE_SMAX:
441 return ISD::SMAX;
442 case ISD::VECREDUCE_SMIN:
443 case ISD::VP_REDUCE_SMIN:
444 return ISD::SMIN;
445 case ISD::VECREDUCE_UMAX:
446 case ISD::VP_REDUCE_UMAX:
447 return ISD::UMAX;
448 case ISD::VECREDUCE_UMIN:
449 case ISD::VP_REDUCE_UMIN:
450 return ISD::UMIN;
451 case ISD::VECREDUCE_FMAX:
452 case ISD::VP_REDUCE_FMAX:
453 return ISD::FMAXNUM;
454 case ISD::VECREDUCE_FMIN:
455 case ISD::VP_REDUCE_FMIN:
456 return ISD::FMINNUM;
457 }
458 }
459
isVPOpcode(unsigned Opcode)460 bool ISD::isVPOpcode(unsigned Opcode) {
461 switch (Opcode) {
462 default:
463 return false;
464 #define BEGIN_REGISTER_VP_SDNODE(VPSD, ...) \
465 case ISD::VPSD: \
466 return true;
467 #include "llvm/IR/VPIntrinsics.def"
468 }
469 }
470
isVPBinaryOp(unsigned Opcode)471 bool ISD::isVPBinaryOp(unsigned Opcode) {
472 switch (Opcode) {
473 default:
474 break;
475 #define BEGIN_REGISTER_VP_SDNODE(VPSD, ...) case ISD::VPSD:
476 #define VP_PROPERTY_BINARYOP return true;
477 #define END_REGISTER_VP_SDNODE(VPSD) break;
478 #include "llvm/IR/VPIntrinsics.def"
479 }
480 return false;
481 }
482
isVPReduction(unsigned Opcode)483 bool ISD::isVPReduction(unsigned Opcode) {
484 switch (Opcode) {
485 default:
486 break;
487 #define BEGIN_REGISTER_VP_SDNODE(VPSD, ...) case ISD::VPSD:
488 #define VP_PROPERTY_REDUCTION(STARTPOS, ...) return true;
489 #define END_REGISTER_VP_SDNODE(VPSD) break;
490 #include "llvm/IR/VPIntrinsics.def"
491 }
492 return false;
493 }
494
495 /// The operand position of the vector mask.
getVPMaskIdx(unsigned Opcode)496 std::optional<unsigned> ISD::getVPMaskIdx(unsigned Opcode) {
497 switch (Opcode) {
498 default:
499 return std::nullopt;
500 #define BEGIN_REGISTER_VP_SDNODE(VPSD, LEGALPOS, TDNAME, MASKPOS, ...) \
501 case ISD::VPSD: \
502 return MASKPOS;
503 #include "llvm/IR/VPIntrinsics.def"
504 }
505 }
506
507 /// The operand position of the explicit vector length parameter.
getVPExplicitVectorLengthIdx(unsigned Opcode)508 std::optional<unsigned> ISD::getVPExplicitVectorLengthIdx(unsigned Opcode) {
509 switch (Opcode) {
510 default:
511 return std::nullopt;
512 #define BEGIN_REGISTER_VP_SDNODE(VPSD, LEGALPOS, TDNAME, MASKPOS, EVLPOS) \
513 case ISD::VPSD: \
514 return EVLPOS;
515 #include "llvm/IR/VPIntrinsics.def"
516 }
517 }
518
getExtForLoadExtType(bool IsFP,ISD::LoadExtType ExtType)519 ISD::NodeType ISD::getExtForLoadExtType(bool IsFP, ISD::LoadExtType ExtType) {
520 switch (ExtType) {
521 case ISD::EXTLOAD:
522 return IsFP ? ISD::FP_EXTEND : ISD::ANY_EXTEND;
523 case ISD::SEXTLOAD:
524 return ISD::SIGN_EXTEND;
525 case ISD::ZEXTLOAD:
526 return ISD::ZERO_EXTEND;
527 default:
528 break;
529 }
530
531 llvm_unreachable("Invalid LoadExtType");
532 }
533
getSetCCSwappedOperands(ISD::CondCode Operation)534 ISD::CondCode ISD::getSetCCSwappedOperands(ISD::CondCode Operation) {
535 // To perform this operation, we just need to swap the L and G bits of the
536 // operation.
537 unsigned OldL = (Operation >> 2) & 1;
538 unsigned OldG = (Operation >> 1) & 1;
539 return ISD::CondCode((Operation & ~6) | // Keep the N, U, E bits
540 (OldL << 1) | // New G bit
541 (OldG << 2)); // New L bit.
542 }
543
getSetCCInverseImpl(ISD::CondCode Op,bool isIntegerLike)544 static ISD::CondCode getSetCCInverseImpl(ISD::CondCode Op, bool isIntegerLike) {
545 unsigned Operation = Op;
546 if (isIntegerLike)
547 Operation ^= 7; // Flip L, G, E bits, but not U.
548 else
549 Operation ^= 15; // Flip all of the condition bits.
550
551 if (Operation > ISD::SETTRUE2)
552 Operation &= ~8; // Don't let N and U bits get set.
553
554 return ISD::CondCode(Operation);
555 }
556
getSetCCInverse(ISD::CondCode Op,EVT Type)557 ISD::CondCode ISD::getSetCCInverse(ISD::CondCode Op, EVT Type) {
558 return getSetCCInverseImpl(Op, Type.isInteger());
559 }
560
getSetCCInverse(ISD::CondCode Op,bool isIntegerLike)561 ISD::CondCode ISD::GlobalISel::getSetCCInverse(ISD::CondCode Op,
562 bool isIntegerLike) {
563 return getSetCCInverseImpl(Op, isIntegerLike);
564 }
565
566 /// For an integer comparison, return 1 if the comparison is a signed operation
567 /// and 2 if the result is an unsigned comparison. Return zero if the operation
568 /// does not depend on the sign of the input (setne and seteq).
isSignedOp(ISD::CondCode Opcode)569 static int isSignedOp(ISD::CondCode Opcode) {
570 switch (Opcode) {
571 default: llvm_unreachable("Illegal integer setcc operation!");
572 case ISD::SETEQ:
573 case ISD::SETNE: return 0;
574 case ISD::SETLT:
575 case ISD::SETLE:
576 case ISD::SETGT:
577 case ISD::SETGE: return 1;
578 case ISD::SETULT:
579 case ISD::SETULE:
580 case ISD::SETUGT:
581 case ISD::SETUGE: return 2;
582 }
583 }
584
getSetCCOrOperation(ISD::CondCode Op1,ISD::CondCode Op2,EVT Type)585 ISD::CondCode ISD::getSetCCOrOperation(ISD::CondCode Op1, ISD::CondCode Op2,
586 EVT Type) {
587 bool IsInteger = Type.isInteger();
588 if (IsInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3)
589 // Cannot fold a signed integer setcc with an unsigned integer setcc.
590 return ISD::SETCC_INVALID;
591
592 unsigned Op = Op1 | Op2; // Combine all of the condition bits.
593
594 // If the N and U bits get set, then the resultant comparison DOES suddenly
595 // care about orderedness, and it is true when ordered.
596 if (Op > ISD::SETTRUE2)
597 Op &= ~16; // Clear the U bit if the N bit is set.
598
599 // Canonicalize illegal integer setcc's.
600 if (IsInteger && Op == ISD::SETUNE) // e.g. SETUGT | SETULT
601 Op = ISD::SETNE;
602
603 return ISD::CondCode(Op);
604 }
605
getSetCCAndOperation(ISD::CondCode Op1,ISD::CondCode Op2,EVT Type)606 ISD::CondCode ISD::getSetCCAndOperation(ISD::CondCode Op1, ISD::CondCode Op2,
607 EVT Type) {
608 bool IsInteger = Type.isInteger();
609 if (IsInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3)
610 // Cannot fold a signed setcc with an unsigned setcc.
611 return ISD::SETCC_INVALID;
612
613 // Combine all of the condition bits.
614 ISD::CondCode Result = ISD::CondCode(Op1 & Op2);
615
616 // Canonicalize illegal integer setcc's.
617 if (IsInteger) {
618 switch (Result) {
619 default: break;
620 case ISD::SETUO : Result = ISD::SETFALSE; break; // SETUGT & SETULT
621 case ISD::SETOEQ: // SETEQ & SETU[LG]E
622 case ISD::SETUEQ: Result = ISD::SETEQ ; break; // SETUGE & SETULE
623 case ISD::SETOLT: Result = ISD::SETULT ; break; // SETULT & SETNE
624 case ISD::SETOGT: Result = ISD::SETUGT ; break; // SETUGT & SETNE
625 }
626 }
627
628 return Result;
629 }
630
631 //===----------------------------------------------------------------------===//
632 // SDNode Profile Support
633 //===----------------------------------------------------------------------===//
634
635 /// AddNodeIDOpcode - Add the node opcode to the NodeID data.
AddNodeIDOpcode(FoldingSetNodeID & ID,unsigned OpC)636 static void AddNodeIDOpcode(FoldingSetNodeID &ID, unsigned OpC) {
637 ID.AddInteger(OpC);
638 }
639
640 /// AddNodeIDValueTypes - Value type lists are intern'd so we can represent them
641 /// solely with their pointer.
AddNodeIDValueTypes(FoldingSetNodeID & ID,SDVTList VTList)642 static void AddNodeIDValueTypes(FoldingSetNodeID &ID, SDVTList VTList) {
643 ID.AddPointer(VTList.VTs);
644 }
645
646 /// AddNodeIDOperands - Various routines for adding operands to the NodeID data.
AddNodeIDOperands(FoldingSetNodeID & ID,ArrayRef<SDValue> Ops)647 static void AddNodeIDOperands(FoldingSetNodeID &ID,
648 ArrayRef<SDValue> Ops) {
649 for (const auto &Op : Ops) {
650 ID.AddPointer(Op.getNode());
651 ID.AddInteger(Op.getResNo());
652 }
653 }
654
655 /// AddNodeIDOperands - Various routines for adding operands to the NodeID data.
AddNodeIDOperands(FoldingSetNodeID & ID,ArrayRef<SDUse> Ops)656 static void AddNodeIDOperands(FoldingSetNodeID &ID,
657 ArrayRef<SDUse> Ops) {
658 for (const auto &Op : Ops) {
659 ID.AddPointer(Op.getNode());
660 ID.AddInteger(Op.getResNo());
661 }
662 }
663
AddNodeIDNode(FoldingSetNodeID & ID,unsigned OpC,SDVTList VTList,ArrayRef<SDValue> OpList)664 static void AddNodeIDNode(FoldingSetNodeID &ID, unsigned OpC,
665 SDVTList VTList, ArrayRef<SDValue> OpList) {
666 AddNodeIDOpcode(ID, OpC);
667 AddNodeIDValueTypes(ID, VTList);
668 AddNodeIDOperands(ID, OpList);
669 }
670
671 /// If this is an SDNode with special info, add this info to the NodeID data.
AddNodeIDCustom(FoldingSetNodeID & ID,const SDNode * N)672 static void AddNodeIDCustom(FoldingSetNodeID &ID, const SDNode *N) {
673 switch (N->getOpcode()) {
674 case ISD::TargetExternalSymbol:
675 case ISD::ExternalSymbol:
676 case ISD::MCSymbol:
677 llvm_unreachable("Should only be used on nodes with operands");
678 default: break; // Normal nodes don't need extra info.
679 case ISD::TargetConstant:
680 case ISD::Constant: {
681 const ConstantSDNode *C = cast<ConstantSDNode>(N);
682 ID.AddPointer(C->getConstantIntValue());
683 ID.AddBoolean(C->isOpaque());
684 break;
685 }
686 case ISD::TargetConstantFP:
687 case ISD::ConstantFP:
688 ID.AddPointer(cast<ConstantFPSDNode>(N)->getConstantFPValue());
689 break;
690 case ISD::TargetGlobalAddress:
691 case ISD::GlobalAddress:
692 case ISD::TargetGlobalTLSAddress:
693 case ISD::GlobalTLSAddress: {
694 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
695 ID.AddPointer(GA->getGlobal());
696 ID.AddInteger(GA->getOffset());
697 ID.AddInteger(GA->getTargetFlags());
698 break;
699 }
700 case ISD::BasicBlock:
701 ID.AddPointer(cast<BasicBlockSDNode>(N)->getBasicBlock());
702 break;
703 case ISD::Register:
704 ID.AddInteger(cast<RegisterSDNode>(N)->getReg());
705 break;
706 case ISD::RegisterMask:
707 ID.AddPointer(cast<RegisterMaskSDNode>(N)->getRegMask());
708 break;
709 case ISD::SRCVALUE:
710 ID.AddPointer(cast<SrcValueSDNode>(N)->getValue());
711 break;
712 case ISD::FrameIndex:
713 case ISD::TargetFrameIndex:
714 ID.AddInteger(cast<FrameIndexSDNode>(N)->getIndex());
715 break;
716 case ISD::LIFETIME_START:
717 case ISD::LIFETIME_END:
718 if (cast<LifetimeSDNode>(N)->hasOffset()) {
719 ID.AddInteger(cast<LifetimeSDNode>(N)->getSize());
720 ID.AddInteger(cast<LifetimeSDNode>(N)->getOffset());
721 }
722 break;
723 case ISD::PSEUDO_PROBE:
724 ID.AddInteger(cast<PseudoProbeSDNode>(N)->getGuid());
725 ID.AddInteger(cast<PseudoProbeSDNode>(N)->getIndex());
726 ID.AddInteger(cast<PseudoProbeSDNode>(N)->getAttributes());
727 break;
728 case ISD::JumpTable:
729 case ISD::TargetJumpTable:
730 ID.AddInteger(cast<JumpTableSDNode>(N)->getIndex());
731 ID.AddInteger(cast<JumpTableSDNode>(N)->getTargetFlags());
732 break;
733 case ISD::ConstantPool:
734 case ISD::TargetConstantPool: {
735 const ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(N);
736 ID.AddInteger(CP->getAlign().value());
737 ID.AddInteger(CP->getOffset());
738 if (CP->isMachineConstantPoolEntry())
739 CP->getMachineCPVal()->addSelectionDAGCSEId(ID);
740 else
741 ID.AddPointer(CP->getConstVal());
742 ID.AddInteger(CP->getTargetFlags());
743 break;
744 }
745 case ISD::TargetIndex: {
746 const TargetIndexSDNode *TI = cast<TargetIndexSDNode>(N);
747 ID.AddInteger(TI->getIndex());
748 ID.AddInteger(TI->getOffset());
749 ID.AddInteger(TI->getTargetFlags());
750 break;
751 }
752 case ISD::LOAD: {
753 const LoadSDNode *LD = cast<LoadSDNode>(N);
754 ID.AddInteger(LD->getMemoryVT().getRawBits());
755 ID.AddInteger(LD->getRawSubclassData());
756 ID.AddInteger(LD->getPointerInfo().getAddrSpace());
757 ID.AddInteger(LD->getMemOperand()->getFlags());
758 break;
759 }
760 case ISD::STORE: {
761 const StoreSDNode *ST = cast<StoreSDNode>(N);
762 ID.AddInteger(ST->getMemoryVT().getRawBits());
763 ID.AddInteger(ST->getRawSubclassData());
764 ID.AddInteger(ST->getPointerInfo().getAddrSpace());
765 ID.AddInteger(ST->getMemOperand()->getFlags());
766 break;
767 }
768 case ISD::VP_LOAD: {
769 const VPLoadSDNode *ELD = cast<VPLoadSDNode>(N);
770 ID.AddInteger(ELD->getMemoryVT().getRawBits());
771 ID.AddInteger(ELD->getRawSubclassData());
772 ID.AddInteger(ELD->getPointerInfo().getAddrSpace());
773 ID.AddInteger(ELD->getMemOperand()->getFlags());
774 break;
775 }
776 case ISD::VP_STORE: {
777 const VPStoreSDNode *EST = cast<VPStoreSDNode>(N);
778 ID.AddInteger(EST->getMemoryVT().getRawBits());
779 ID.AddInteger(EST->getRawSubclassData());
780 ID.AddInteger(EST->getPointerInfo().getAddrSpace());
781 ID.AddInteger(EST->getMemOperand()->getFlags());
782 break;
783 }
784 case ISD::EXPERIMENTAL_VP_STRIDED_LOAD: {
785 const VPStridedLoadSDNode *SLD = cast<VPStridedLoadSDNode>(N);
786 ID.AddInteger(SLD->getMemoryVT().getRawBits());
787 ID.AddInteger(SLD->getRawSubclassData());
788 ID.AddInteger(SLD->getPointerInfo().getAddrSpace());
789 break;
790 }
791 case ISD::EXPERIMENTAL_VP_STRIDED_STORE: {
792 const VPStridedStoreSDNode *SST = cast<VPStridedStoreSDNode>(N);
793 ID.AddInteger(SST->getMemoryVT().getRawBits());
794 ID.AddInteger(SST->getRawSubclassData());
795 ID.AddInteger(SST->getPointerInfo().getAddrSpace());
796 break;
797 }
798 case ISD::VP_GATHER: {
799 const VPGatherSDNode *EG = cast<VPGatherSDNode>(N);
800 ID.AddInteger(EG->getMemoryVT().getRawBits());
801 ID.AddInteger(EG->getRawSubclassData());
802 ID.AddInteger(EG->getPointerInfo().getAddrSpace());
803 ID.AddInteger(EG->getMemOperand()->getFlags());
804 break;
805 }
806 case ISD::VP_SCATTER: {
807 const VPScatterSDNode *ES = cast<VPScatterSDNode>(N);
808 ID.AddInteger(ES->getMemoryVT().getRawBits());
809 ID.AddInteger(ES->getRawSubclassData());
810 ID.AddInteger(ES->getPointerInfo().getAddrSpace());
811 ID.AddInteger(ES->getMemOperand()->getFlags());
812 break;
813 }
814 case ISD::MLOAD: {
815 const MaskedLoadSDNode *MLD = cast<MaskedLoadSDNode>(N);
816 ID.AddInteger(MLD->getMemoryVT().getRawBits());
817 ID.AddInteger(MLD->getRawSubclassData());
818 ID.AddInteger(MLD->getPointerInfo().getAddrSpace());
819 ID.AddInteger(MLD->getMemOperand()->getFlags());
820 break;
821 }
822 case ISD::MSTORE: {
823 const MaskedStoreSDNode *MST = cast<MaskedStoreSDNode>(N);
824 ID.AddInteger(MST->getMemoryVT().getRawBits());
825 ID.AddInteger(MST->getRawSubclassData());
826 ID.AddInteger(MST->getPointerInfo().getAddrSpace());
827 ID.AddInteger(MST->getMemOperand()->getFlags());
828 break;
829 }
830 case ISD::MGATHER: {
831 const MaskedGatherSDNode *MG = cast<MaskedGatherSDNode>(N);
832 ID.AddInteger(MG->getMemoryVT().getRawBits());
833 ID.AddInteger(MG->getRawSubclassData());
834 ID.AddInteger(MG->getPointerInfo().getAddrSpace());
835 ID.AddInteger(MG->getMemOperand()->getFlags());
836 break;
837 }
838 case ISD::MSCATTER: {
839 const MaskedScatterSDNode *MS = cast<MaskedScatterSDNode>(N);
840 ID.AddInteger(MS->getMemoryVT().getRawBits());
841 ID.AddInteger(MS->getRawSubclassData());
842 ID.AddInteger(MS->getPointerInfo().getAddrSpace());
843 ID.AddInteger(MS->getMemOperand()->getFlags());
844 break;
845 }
846 case ISD::ATOMIC_CMP_SWAP:
847 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
848 case ISD::ATOMIC_SWAP:
849 case ISD::ATOMIC_LOAD_ADD:
850 case ISD::ATOMIC_LOAD_SUB:
851 case ISD::ATOMIC_LOAD_AND:
852 case ISD::ATOMIC_LOAD_CLR:
853 case ISD::ATOMIC_LOAD_OR:
854 case ISD::ATOMIC_LOAD_XOR:
855 case ISD::ATOMIC_LOAD_NAND:
856 case ISD::ATOMIC_LOAD_MIN:
857 case ISD::ATOMIC_LOAD_MAX:
858 case ISD::ATOMIC_LOAD_UMIN:
859 case ISD::ATOMIC_LOAD_UMAX:
860 case ISD::ATOMIC_LOAD:
861 case ISD::ATOMIC_STORE: {
862 const AtomicSDNode *AT = cast<AtomicSDNode>(N);
863 ID.AddInteger(AT->getMemoryVT().getRawBits());
864 ID.AddInteger(AT->getRawSubclassData());
865 ID.AddInteger(AT->getPointerInfo().getAddrSpace());
866 ID.AddInteger(AT->getMemOperand()->getFlags());
867 break;
868 }
869 case ISD::PREFETCH: {
870 const MemSDNode *PF = cast<MemSDNode>(N);
871 ID.AddInteger(PF->getPointerInfo().getAddrSpace());
872 ID.AddInteger(PF->getMemOperand()->getFlags());
873 break;
874 }
875 case ISD::VECTOR_SHUFFLE: {
876 const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
877 for (unsigned i = 0, e = N->getValueType(0).getVectorNumElements();
878 i != e; ++i)
879 ID.AddInteger(SVN->getMaskElt(i));
880 break;
881 }
882 case ISD::TargetBlockAddress:
883 case ISD::BlockAddress: {
884 const BlockAddressSDNode *BA = cast<BlockAddressSDNode>(N);
885 ID.AddPointer(BA->getBlockAddress());
886 ID.AddInteger(BA->getOffset());
887 ID.AddInteger(BA->getTargetFlags());
888 break;
889 }
890 case ISD::AssertAlign:
891 ID.AddInteger(cast<AssertAlignSDNode>(N)->getAlign().value());
892 break;
893 } // end switch (N->getOpcode())
894
895 // Target specific memory nodes could also have address spaces and flags
896 // to check.
897 if (N->isTargetMemoryOpcode()) {
898 const MemSDNode *MN = cast<MemSDNode>(N);
899 ID.AddInteger(MN->getPointerInfo().getAddrSpace());
900 ID.AddInteger(MN->getMemOperand()->getFlags());
901 }
902 }
903
904 /// AddNodeIDNode - Generic routine for adding a nodes info to the NodeID
905 /// data.
AddNodeIDNode(FoldingSetNodeID & ID,const SDNode * N)906 static void AddNodeIDNode(FoldingSetNodeID &ID, const SDNode *N) {
907 AddNodeIDOpcode(ID, N->getOpcode());
908 // Add the return value info.
909 AddNodeIDValueTypes(ID, N->getVTList());
910 // Add the operand info.
911 AddNodeIDOperands(ID, N->ops());
912
913 // Handle SDNode leafs with special info.
914 AddNodeIDCustom(ID, N);
915 }
916
917 //===----------------------------------------------------------------------===//
918 // SelectionDAG Class
919 //===----------------------------------------------------------------------===//
920
921 /// doNotCSE - Return true if CSE should not be performed for this node.
doNotCSE(SDNode * N)922 static bool doNotCSE(SDNode *N) {
923 if (N->getValueType(0) == MVT::Glue)
924 return true; // Never CSE anything that produces a flag.
925
926 switch (N->getOpcode()) {
927 default: break;
928 case ISD::HANDLENODE:
929 case ISD::EH_LABEL:
930 return true; // Never CSE these nodes.
931 }
932
933 // Check that remaining values produced are not flags.
934 for (unsigned i = 1, e = N->getNumValues(); i != e; ++i)
935 if (N->getValueType(i) == MVT::Glue)
936 return true; // Never CSE anything that produces a flag.
937
938 return false;
939 }
940
941 /// RemoveDeadNodes - This method deletes all unreachable nodes in the
942 /// SelectionDAG.
RemoveDeadNodes()943 void SelectionDAG::RemoveDeadNodes() {
944 // Create a dummy node (which is not added to allnodes), that adds a reference
945 // to the root node, preventing it from being deleted.
946 HandleSDNode Dummy(getRoot());
947
948 SmallVector<SDNode*, 128> DeadNodes;
949
950 // Add all obviously-dead nodes to the DeadNodes worklist.
951 for (SDNode &Node : allnodes())
952 if (Node.use_empty())
953 DeadNodes.push_back(&Node);
954
955 RemoveDeadNodes(DeadNodes);
956
957 // If the root changed (e.g. it was a dead load, update the root).
958 setRoot(Dummy.getValue());
959 }
960
961 /// RemoveDeadNodes - This method deletes the unreachable nodes in the
962 /// given list, and any nodes that become unreachable as a result.
RemoveDeadNodes(SmallVectorImpl<SDNode * > & DeadNodes)963 void SelectionDAG::RemoveDeadNodes(SmallVectorImpl<SDNode *> &DeadNodes) {
964
965 // Process the worklist, deleting the nodes and adding their uses to the
966 // worklist.
967 while (!DeadNodes.empty()) {
968 SDNode *N = DeadNodes.pop_back_val();
969 // Skip to next node if we've already managed to delete the node. This could
970 // happen if replacing a node causes a node previously added to the node to
971 // be deleted.
972 if (N->getOpcode() == ISD::DELETED_NODE)
973 continue;
974
975 for (DAGUpdateListener *DUL = UpdateListeners; DUL; DUL = DUL->Next)
976 DUL->NodeDeleted(N, nullptr);
977
978 // Take the node out of the appropriate CSE map.
979 RemoveNodeFromCSEMaps(N);
980
981 // Next, brutally remove the operand list. This is safe to do, as there are
982 // no cycles in the graph.
983 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) {
984 SDUse &Use = *I++;
985 SDNode *Operand = Use.getNode();
986 Use.set(SDValue());
987
988 // Now that we removed this operand, see if there are no uses of it left.
989 if (Operand->use_empty())
990 DeadNodes.push_back(Operand);
991 }
992
993 DeallocateNode(N);
994 }
995 }
996
RemoveDeadNode(SDNode * N)997 void SelectionDAG::RemoveDeadNode(SDNode *N){
998 SmallVector<SDNode*, 16> DeadNodes(1, N);
999
1000 // Create a dummy node that adds a reference to the root node, preventing
1001 // it from being deleted. (This matters if the root is an operand of the
1002 // dead node.)
1003 HandleSDNode Dummy(getRoot());
1004
1005 RemoveDeadNodes(DeadNodes);
1006 }
1007
DeleteNode(SDNode * N)1008 void SelectionDAG::DeleteNode(SDNode *N) {
1009 // First take this out of the appropriate CSE map.
1010 RemoveNodeFromCSEMaps(N);
1011
1012 // Finally, remove uses due to operands of this node, remove from the
1013 // AllNodes list, and delete the node.
1014 DeleteNodeNotInCSEMaps(N);
1015 }
1016
DeleteNodeNotInCSEMaps(SDNode * N)1017 void SelectionDAG::DeleteNodeNotInCSEMaps(SDNode *N) {
1018 assert(N->getIterator() != AllNodes.begin() &&
1019 "Cannot delete the entry node!");
1020 assert(N->use_empty() && "Cannot delete a node that is not dead!");
1021
1022 // Drop all of the operands and decrement used node's use counts.
1023 N->DropOperands();
1024
1025 DeallocateNode(N);
1026 }
1027
add(SDDbgValue * V,bool isParameter)1028 void SDDbgInfo::add(SDDbgValue *V, bool isParameter) {
1029 assert(!(V->isVariadic() && isParameter));
1030 if (isParameter)
1031 ByvalParmDbgValues.push_back(V);
1032 else
1033 DbgValues.push_back(V);
1034 for (const SDNode *Node : V->getSDNodes())
1035 if (Node)
1036 DbgValMap[Node].push_back(V);
1037 }
1038
erase(const SDNode * Node)1039 void SDDbgInfo::erase(const SDNode *Node) {
1040 DbgValMapType::iterator I = DbgValMap.find(Node);
1041 if (I == DbgValMap.end())
1042 return;
1043 for (auto &Val: I->second)
1044 Val->setIsInvalidated();
1045 DbgValMap.erase(I);
1046 }
1047
DeallocateNode(SDNode * N)1048 void SelectionDAG::DeallocateNode(SDNode *N) {
1049 // If we have operands, deallocate them.
1050 removeOperands(N);
1051
1052 NodeAllocator.Deallocate(AllNodes.remove(N));
1053
1054 // Set the opcode to DELETED_NODE to help catch bugs when node
1055 // memory is reallocated.
1056 // FIXME: There are places in SDag that have grown a dependency on the opcode
1057 // value in the released node.
1058 __asan_unpoison_memory_region(&N->NodeType, sizeof(N->NodeType));
1059 N->NodeType = ISD::DELETED_NODE;
1060
1061 // If any of the SDDbgValue nodes refer to this SDNode, invalidate
1062 // them and forget about that node.
1063 DbgInfo->erase(N);
1064
1065 // Invalidate extra info.
1066 SDEI.erase(N);
1067 }
1068
1069 #ifndef NDEBUG
1070 /// VerifySDNode - Check the given SDNode. Aborts if it is invalid.
VerifySDNode(SDNode * N)1071 static void VerifySDNode(SDNode *N) {
1072 switch (N->getOpcode()) {
1073 default:
1074 break;
1075 case ISD::BUILD_PAIR: {
1076 EVT VT = N->getValueType(0);
1077 assert(N->getNumValues() == 1 && "Too many results!");
1078 assert(!VT.isVector() && (VT.isInteger() || VT.isFloatingPoint()) &&
1079 "Wrong return type!");
1080 assert(N->getNumOperands() == 2 && "Wrong number of operands!");
1081 assert(N->getOperand(0).getValueType() == N->getOperand(1).getValueType() &&
1082 "Mismatched operand types!");
1083 assert(N->getOperand(0).getValueType().isInteger() == VT.isInteger() &&
1084 "Wrong operand type!");
1085 assert(VT.getSizeInBits() == 2 * N->getOperand(0).getValueSizeInBits() &&
1086 "Wrong return type size");
1087 break;
1088 }
1089 case ISD::BUILD_VECTOR: {
1090 assert(N->getNumValues() == 1 && "Too many results!");
1091 assert(N->getValueType(0).isVector() && "Wrong return type!");
1092 assert(N->getNumOperands() == N->getValueType(0).getVectorNumElements() &&
1093 "Wrong number of operands!");
1094 EVT EltVT = N->getValueType(0).getVectorElementType();
1095 for (const SDUse &Op : N->ops()) {
1096 assert((Op.getValueType() == EltVT ||
1097 (EltVT.isInteger() && Op.getValueType().isInteger() &&
1098 EltVT.bitsLE(Op.getValueType()))) &&
1099 "Wrong operand type!");
1100 assert(Op.getValueType() == N->getOperand(0).getValueType() &&
1101 "Operands must all have the same type");
1102 }
1103 break;
1104 }
1105 }
1106 }
1107 #endif // NDEBUG
1108
1109 /// Insert a newly allocated node into the DAG.
1110 ///
1111 /// Handles insertion into the all nodes list and CSE map, as well as
1112 /// verification and other common operations when a new node is allocated.
InsertNode(SDNode * N)1113 void SelectionDAG::InsertNode(SDNode *N) {
1114 AllNodes.push_back(N);
1115 #ifndef NDEBUG
1116 N->PersistentId = NextPersistentId++;
1117 VerifySDNode(N);
1118 #endif
1119 for (DAGUpdateListener *DUL = UpdateListeners; DUL; DUL = DUL->Next)
1120 DUL->NodeInserted(N);
1121 }
1122
1123 /// RemoveNodeFromCSEMaps - Take the specified node out of the CSE map that
1124 /// correspond to it. This is useful when we're about to delete or repurpose
1125 /// the node. We don't want future request for structurally identical nodes
1126 /// to return N anymore.
RemoveNodeFromCSEMaps(SDNode * N)1127 bool SelectionDAG::RemoveNodeFromCSEMaps(SDNode *N) {
1128 bool Erased = false;
1129 switch (N->getOpcode()) {
1130 case ISD::HANDLENODE: return false; // noop.
1131 case ISD::CONDCODE:
1132 assert(CondCodeNodes[cast<CondCodeSDNode>(N)->get()] &&
1133 "Cond code doesn't exist!");
1134 Erased = CondCodeNodes[cast<CondCodeSDNode>(N)->get()] != nullptr;
1135 CondCodeNodes[cast<CondCodeSDNode>(N)->get()] = nullptr;
1136 break;
1137 case ISD::ExternalSymbol:
1138 Erased = ExternalSymbols.erase(cast<ExternalSymbolSDNode>(N)->getSymbol());
1139 break;
1140 case ISD::TargetExternalSymbol: {
1141 ExternalSymbolSDNode *ESN = cast<ExternalSymbolSDNode>(N);
1142 Erased = TargetExternalSymbols.erase(std::pair<std::string, unsigned>(
1143 ESN->getSymbol(), ESN->getTargetFlags()));
1144 break;
1145 }
1146 case ISD::MCSymbol: {
1147 auto *MCSN = cast<MCSymbolSDNode>(N);
1148 Erased = MCSymbols.erase(MCSN->getMCSymbol());
1149 break;
1150 }
1151 case ISD::VALUETYPE: {
1152 EVT VT = cast<VTSDNode>(N)->getVT();
1153 if (VT.isExtended()) {
1154 Erased = ExtendedValueTypeNodes.erase(VT);
1155 } else {
1156 Erased = ValueTypeNodes[VT.getSimpleVT().SimpleTy] != nullptr;
1157 ValueTypeNodes[VT.getSimpleVT().SimpleTy] = nullptr;
1158 }
1159 break;
1160 }
1161 default:
1162 // Remove it from the CSE Map.
1163 assert(N->getOpcode() != ISD::DELETED_NODE && "DELETED_NODE in CSEMap!");
1164 assert(N->getOpcode() != ISD::EntryToken && "EntryToken in CSEMap!");
1165 Erased = CSEMap.RemoveNode(N);
1166 break;
1167 }
1168 #ifndef NDEBUG
1169 // Verify that the node was actually in one of the CSE maps, unless it has a
1170 // flag result (which cannot be CSE'd) or is one of the special cases that are
1171 // not subject to CSE.
1172 if (!Erased && N->getValueType(N->getNumValues()-1) != MVT::Glue &&
1173 !N->isMachineOpcode() && !doNotCSE(N)) {
1174 N->dump(this);
1175 dbgs() << "\n";
1176 llvm_unreachable("Node is not in map!");
1177 }
1178 #endif
1179 return Erased;
1180 }
1181
1182 /// AddModifiedNodeToCSEMaps - The specified node has been removed from the CSE
1183 /// maps and modified in place. Add it back to the CSE maps, unless an identical
1184 /// node already exists, in which case transfer all its users to the existing
1185 /// node. This transfer can potentially trigger recursive merging.
1186 void
AddModifiedNodeToCSEMaps(SDNode * N)1187 SelectionDAG::AddModifiedNodeToCSEMaps(SDNode *N) {
1188 // For node types that aren't CSE'd, just act as if no identical node
1189 // already exists.
1190 if (!doNotCSE(N)) {
1191 SDNode *Existing = CSEMap.GetOrInsertNode(N);
1192 if (Existing != N) {
1193 // If there was already an existing matching node, use ReplaceAllUsesWith
1194 // to replace the dead one with the existing one. This can cause
1195 // recursive merging of other unrelated nodes down the line.
1196 ReplaceAllUsesWith(N, Existing);
1197
1198 // N is now dead. Inform the listeners and delete it.
1199 for (DAGUpdateListener *DUL = UpdateListeners; DUL; DUL = DUL->Next)
1200 DUL->NodeDeleted(N, Existing);
1201 DeleteNodeNotInCSEMaps(N);
1202 return;
1203 }
1204 }
1205
1206 // If the node doesn't already exist, we updated it. Inform listeners.
1207 for (DAGUpdateListener *DUL = UpdateListeners; DUL; DUL = DUL->Next)
1208 DUL->NodeUpdated(N);
1209 }
1210
1211 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
1212 /// were replaced with those specified. If this node is never memoized,
1213 /// return null, otherwise return a pointer to the slot it would take. If a
1214 /// node already exists with these operands, the slot will be non-null.
FindModifiedNodeSlot(SDNode * N,SDValue Op,void * & InsertPos)1215 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, SDValue Op,
1216 void *&InsertPos) {
1217 if (doNotCSE(N))
1218 return nullptr;
1219
1220 SDValue Ops[] = { Op };
1221 FoldingSetNodeID ID;
1222 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops);
1223 AddNodeIDCustom(ID, N);
1224 SDNode *Node = FindNodeOrInsertPos(ID, SDLoc(N), InsertPos);
1225 if (Node)
1226 Node->intersectFlagsWith(N->getFlags());
1227 return Node;
1228 }
1229
1230 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
1231 /// were replaced with those specified. If this node is never memoized,
1232 /// return null, otherwise return a pointer to the slot it would take. If a
1233 /// node already exists with these operands, the slot will be non-null.
FindModifiedNodeSlot(SDNode * N,SDValue Op1,SDValue Op2,void * & InsertPos)1234 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N,
1235 SDValue Op1, SDValue Op2,
1236 void *&InsertPos) {
1237 if (doNotCSE(N))
1238 return nullptr;
1239
1240 SDValue Ops[] = { Op1, Op2 };
1241 FoldingSetNodeID ID;
1242 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops);
1243 AddNodeIDCustom(ID, N);
1244 SDNode *Node = FindNodeOrInsertPos(ID, SDLoc(N), InsertPos);
1245 if (Node)
1246 Node->intersectFlagsWith(N->getFlags());
1247 return Node;
1248 }
1249
1250 /// FindModifiedNodeSlot - Find a slot for the specified node if its operands
1251 /// were replaced with those specified. If this node is never memoized,
1252 /// return null, otherwise return a pointer to the slot it would take. If a
1253 /// node already exists with these operands, the slot will be non-null.
FindModifiedNodeSlot(SDNode * N,ArrayRef<SDValue> Ops,void * & InsertPos)1254 SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, ArrayRef<SDValue> Ops,
1255 void *&InsertPos) {
1256 if (doNotCSE(N))
1257 return nullptr;
1258
1259 FoldingSetNodeID ID;
1260 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops);
1261 AddNodeIDCustom(ID, N);
1262 SDNode *Node = FindNodeOrInsertPos(ID, SDLoc(N), InsertPos);
1263 if (Node)
1264 Node->intersectFlagsWith(N->getFlags());
1265 return Node;
1266 }
1267
getEVTAlign(EVT VT) const1268 Align SelectionDAG::getEVTAlign(EVT VT) const {
1269 Type *Ty = VT == MVT::iPTR ?
1270 PointerType::get(Type::getInt8Ty(*getContext()), 0) :
1271 VT.getTypeForEVT(*getContext());
1272
1273 return getDataLayout().getABITypeAlign(Ty);
1274 }
1275
1276 // EntryNode could meaningfully have debug info if we can find it...
SelectionDAG(const TargetMachine & tm,CodeGenOpt::Level OL)1277 SelectionDAG::SelectionDAG(const TargetMachine &tm, CodeGenOpt::Level OL)
1278 : TM(tm), OptLevel(OL),
1279 EntryNode(ISD::EntryToken, 0, DebugLoc(), getVTList(MVT::Other, MVT::Glue)),
1280 Root(getEntryNode()) {
1281 InsertNode(&EntryNode);
1282 DbgInfo = new SDDbgInfo();
1283 }
1284
init(MachineFunction & NewMF,OptimizationRemarkEmitter & NewORE,Pass * PassPtr,const TargetLibraryInfo * LibraryInfo,LegacyDivergenceAnalysis * Divergence,ProfileSummaryInfo * PSIin,BlockFrequencyInfo * BFIin,FunctionVarLocs const * VarLocs)1285 void SelectionDAG::init(MachineFunction &NewMF,
1286 OptimizationRemarkEmitter &NewORE, Pass *PassPtr,
1287 const TargetLibraryInfo *LibraryInfo,
1288 LegacyDivergenceAnalysis *Divergence,
1289 ProfileSummaryInfo *PSIin, BlockFrequencyInfo *BFIin,
1290 FunctionVarLocs const *VarLocs) {
1291 MF = &NewMF;
1292 SDAGISelPass = PassPtr;
1293 ORE = &NewORE;
1294 TLI = getSubtarget().getTargetLowering();
1295 TSI = getSubtarget().getSelectionDAGInfo();
1296 LibInfo = LibraryInfo;
1297 Context = &MF->getFunction().getContext();
1298 DA = Divergence;
1299 PSI = PSIin;
1300 BFI = BFIin;
1301 FnVarLocs = VarLocs;
1302 }
1303
~SelectionDAG()1304 SelectionDAG::~SelectionDAG() {
1305 assert(!UpdateListeners && "Dangling registered DAGUpdateListeners");
1306 allnodes_clear();
1307 OperandRecycler.clear(OperandAllocator);
1308 delete DbgInfo;
1309 }
1310
shouldOptForSize() const1311 bool SelectionDAG::shouldOptForSize() const {
1312 return MF->getFunction().hasOptSize() ||
1313 llvm::shouldOptimizeForSize(FLI->MBB->getBasicBlock(), PSI, BFI);
1314 }
1315
allnodes_clear()1316 void SelectionDAG::allnodes_clear() {
1317 assert(&*AllNodes.begin() == &EntryNode);
1318 AllNodes.remove(AllNodes.begin());
1319 while (!AllNodes.empty())
1320 DeallocateNode(&AllNodes.front());
1321 #ifndef NDEBUG
1322 NextPersistentId = 0;
1323 #endif
1324 }
1325
FindNodeOrInsertPos(const FoldingSetNodeID & ID,void * & InsertPos)1326 SDNode *SelectionDAG::FindNodeOrInsertPos(const FoldingSetNodeID &ID,
1327 void *&InsertPos) {
1328 SDNode *N = CSEMap.FindNodeOrInsertPos(ID, InsertPos);
1329 if (N) {
1330 switch (N->getOpcode()) {
1331 default: break;
1332 case ISD::Constant:
1333 case ISD::ConstantFP:
1334 llvm_unreachable("Querying for Constant and ConstantFP nodes requires "
1335 "debug location. Use another overload.");
1336 }
1337 }
1338 return N;
1339 }
1340
FindNodeOrInsertPos(const FoldingSetNodeID & ID,const SDLoc & DL,void * & InsertPos)1341 SDNode *SelectionDAG::FindNodeOrInsertPos(const FoldingSetNodeID &ID,
1342 const SDLoc &DL, void *&InsertPos) {
1343 SDNode *N = CSEMap.FindNodeOrInsertPos(ID, InsertPos);
1344 if (N) {
1345 switch (N->getOpcode()) {
1346 case ISD::Constant:
1347 case ISD::ConstantFP:
1348 // Erase debug location from the node if the node is used at several
1349 // different places. Do not propagate one location to all uses as it
1350 // will cause a worse single stepping debugging experience.
1351 if (N->getDebugLoc() != DL.getDebugLoc())
1352 N->setDebugLoc(DebugLoc());
1353 break;
1354 default:
1355 // When the node's point of use is located earlier in the instruction
1356 // sequence than its prior point of use, update its debug info to the
1357 // earlier location.
1358 if (DL.getIROrder() && DL.getIROrder() < N->getIROrder())
1359 N->setDebugLoc(DL.getDebugLoc());
1360 break;
1361 }
1362 }
1363 return N;
1364 }
1365
clear()1366 void SelectionDAG::clear() {
1367 allnodes_clear();
1368 OperandRecycler.clear(OperandAllocator);
1369 OperandAllocator.Reset();
1370 CSEMap.clear();
1371
1372 ExtendedValueTypeNodes.clear();
1373 ExternalSymbols.clear();
1374 TargetExternalSymbols.clear();
1375 MCSymbols.clear();
1376 SDEI.clear();
1377 std::fill(CondCodeNodes.begin(), CondCodeNodes.end(),
1378 static_cast<CondCodeSDNode*>(nullptr));
1379 std::fill(ValueTypeNodes.begin(), ValueTypeNodes.end(),
1380 static_cast<SDNode*>(nullptr));
1381
1382 EntryNode.UseList = nullptr;
1383 InsertNode(&EntryNode);
1384 Root = getEntryNode();
1385 DbgInfo->clear();
1386 }
1387
getFPExtendOrRound(SDValue Op,const SDLoc & DL,EVT VT)1388 SDValue SelectionDAG::getFPExtendOrRound(SDValue Op, const SDLoc &DL, EVT VT) {
1389 return VT.bitsGT(Op.getValueType())
1390 ? getNode(ISD::FP_EXTEND, DL, VT, Op)
1391 : getNode(ISD::FP_ROUND, DL, VT, Op,
1392 getIntPtrConstant(0, DL, /*isTarget=*/true));
1393 }
1394
1395 std::pair<SDValue, SDValue>
getStrictFPExtendOrRound(SDValue Op,SDValue Chain,const SDLoc & DL,EVT VT)1396 SelectionDAG::getStrictFPExtendOrRound(SDValue Op, SDValue Chain,
1397 const SDLoc &DL, EVT VT) {
1398 assert(!VT.bitsEq(Op.getValueType()) &&
1399 "Strict no-op FP extend/round not allowed.");
1400 SDValue Res =
1401 VT.bitsGT(Op.getValueType())
1402 ? getNode(ISD::STRICT_FP_EXTEND, DL, {VT, MVT::Other}, {Chain, Op})
1403 : getNode(ISD::STRICT_FP_ROUND, DL, {VT, MVT::Other},
1404 {Chain, Op, getIntPtrConstant(0, DL)});
1405
1406 return std::pair<SDValue, SDValue>(Res, SDValue(Res.getNode(), 1));
1407 }
1408
getAnyExtOrTrunc(SDValue Op,const SDLoc & DL,EVT VT)1409 SDValue SelectionDAG::getAnyExtOrTrunc(SDValue Op, const SDLoc &DL, EVT VT) {
1410 return VT.bitsGT(Op.getValueType()) ?
1411 getNode(ISD::ANY_EXTEND, DL, VT, Op) :
1412 getNode(ISD::TRUNCATE, DL, VT, Op);
1413 }
1414
getSExtOrTrunc(SDValue Op,const SDLoc & DL,EVT VT)1415 SDValue SelectionDAG::getSExtOrTrunc(SDValue Op, const SDLoc &DL, EVT VT) {
1416 return VT.bitsGT(Op.getValueType()) ?
1417 getNode(ISD::SIGN_EXTEND, DL, VT, Op) :
1418 getNode(ISD::TRUNCATE, DL, VT, Op);
1419 }
1420
getZExtOrTrunc(SDValue Op,const SDLoc & DL,EVT VT)1421 SDValue SelectionDAG::getZExtOrTrunc(SDValue Op, const SDLoc &DL, EVT VT) {
1422 return VT.bitsGT(Op.getValueType()) ?
1423 getNode(ISD::ZERO_EXTEND, DL, VT, Op) :
1424 getNode(ISD::TRUNCATE, DL, VT, Op);
1425 }
1426
getBoolExtOrTrunc(SDValue Op,const SDLoc & SL,EVT VT,EVT OpVT)1427 SDValue SelectionDAG::getBoolExtOrTrunc(SDValue Op, const SDLoc &SL, EVT VT,
1428 EVT OpVT) {
1429 if (VT.bitsLE(Op.getValueType()))
1430 return getNode(ISD::TRUNCATE, SL, VT, Op);
1431
1432 TargetLowering::BooleanContent BType = TLI->getBooleanContents(OpVT);
1433 return getNode(TLI->getExtendForContent(BType), SL, VT, Op);
1434 }
1435
getZeroExtendInReg(SDValue Op,const SDLoc & DL,EVT VT)1436 SDValue SelectionDAG::getZeroExtendInReg(SDValue Op, const SDLoc &DL, EVT VT) {
1437 EVT OpVT = Op.getValueType();
1438 assert(VT.isInteger() && OpVT.isInteger() &&
1439 "Cannot getZeroExtendInReg FP types");
1440 assert(VT.isVector() == OpVT.isVector() &&
1441 "getZeroExtendInReg type should be vector iff the operand "
1442 "type is vector!");
1443 assert((!VT.isVector() ||
1444 VT.getVectorElementCount() == OpVT.getVectorElementCount()) &&
1445 "Vector element counts must match in getZeroExtendInReg");
1446 assert(VT.bitsLE(OpVT) && "Not extending!");
1447 if (OpVT == VT)
1448 return Op;
1449 APInt Imm = APInt::getLowBitsSet(OpVT.getScalarSizeInBits(),
1450 VT.getScalarSizeInBits());
1451 return getNode(ISD::AND, DL, OpVT, Op, getConstant(Imm, DL, OpVT));
1452 }
1453
getPtrExtOrTrunc(SDValue Op,const SDLoc & DL,EVT VT)1454 SDValue SelectionDAG::getPtrExtOrTrunc(SDValue Op, const SDLoc &DL, EVT VT) {
1455 // Only unsigned pointer semantics are supported right now. In the future this
1456 // might delegate to TLI to check pointer signedness.
1457 return getZExtOrTrunc(Op, DL, VT);
1458 }
1459
getPtrExtendInReg(SDValue Op,const SDLoc & DL,EVT VT)1460 SDValue SelectionDAG::getPtrExtendInReg(SDValue Op, const SDLoc &DL, EVT VT) {
1461 // Only unsigned pointer semantics are supported right now. In the future this
1462 // might delegate to TLI to check pointer signedness.
1463 return getZeroExtendInReg(Op, DL, VT);
1464 }
1465
getNegative(SDValue Val,const SDLoc & DL,EVT VT)1466 SDValue SelectionDAG::getNegative(SDValue Val, const SDLoc &DL, EVT VT) {
1467 return getNode(ISD::SUB, DL, VT, getConstant(0, DL, VT), Val);
1468 }
1469
1470 /// getNOT - Create a bitwise NOT operation as (XOR Val, -1).
getNOT(const SDLoc & DL,SDValue Val,EVT VT)1471 SDValue SelectionDAG::getNOT(const SDLoc &DL, SDValue Val, EVT VT) {
1472 return getNode(ISD::XOR, DL, VT, Val, getAllOnesConstant(DL, VT));
1473 }
1474
getLogicalNOT(const SDLoc & DL,SDValue Val,EVT VT)1475 SDValue SelectionDAG::getLogicalNOT(const SDLoc &DL, SDValue Val, EVT VT) {
1476 SDValue TrueValue = getBoolConstant(true, DL, VT, VT);
1477 return getNode(ISD::XOR, DL, VT, Val, TrueValue);
1478 }
1479
getVPLogicalNOT(const SDLoc & DL,SDValue Val,SDValue Mask,SDValue EVL,EVT VT)1480 SDValue SelectionDAG::getVPLogicalNOT(const SDLoc &DL, SDValue Val,
1481 SDValue Mask, SDValue EVL, EVT VT) {
1482 SDValue TrueValue = getBoolConstant(true, DL, VT, VT);
1483 return getNode(ISD::VP_XOR, DL, VT, Val, TrueValue, Mask, EVL);
1484 }
1485
getVPPtrExtOrTrunc(const SDLoc & DL,EVT VT,SDValue Op,SDValue Mask,SDValue EVL)1486 SDValue SelectionDAG::getVPPtrExtOrTrunc(const SDLoc &DL, EVT VT, SDValue Op,
1487 SDValue Mask, SDValue EVL) {
1488 return getVPZExtOrTrunc(DL, VT, Op, Mask, EVL);
1489 }
1490
getVPZExtOrTrunc(const SDLoc & DL,EVT VT,SDValue Op,SDValue Mask,SDValue EVL)1491 SDValue SelectionDAG::getVPZExtOrTrunc(const SDLoc &DL, EVT VT, SDValue Op,
1492 SDValue Mask, SDValue EVL) {
1493 if (VT.bitsGT(Op.getValueType()))
1494 return getNode(ISD::VP_ZERO_EXTEND, DL, VT, Op, Mask, EVL);
1495 if (VT.bitsLT(Op.getValueType()))
1496 return getNode(ISD::VP_TRUNCATE, DL, VT, Op, Mask, EVL);
1497 return Op;
1498 }
1499
getBoolConstant(bool V,const SDLoc & DL,EVT VT,EVT OpVT)1500 SDValue SelectionDAG::getBoolConstant(bool V, const SDLoc &DL, EVT VT,
1501 EVT OpVT) {
1502 if (!V)
1503 return getConstant(0, DL, VT);
1504
1505 switch (TLI->getBooleanContents(OpVT)) {
1506 case TargetLowering::ZeroOrOneBooleanContent:
1507 case TargetLowering::UndefinedBooleanContent:
1508 return getConstant(1, DL, VT);
1509 case TargetLowering::ZeroOrNegativeOneBooleanContent:
1510 return getAllOnesConstant(DL, VT);
1511 }
1512 llvm_unreachable("Unexpected boolean content enum!");
1513 }
1514
getConstant(uint64_t Val,const SDLoc & DL,EVT VT,bool isT,bool isO)1515 SDValue SelectionDAG::getConstant(uint64_t Val, const SDLoc &DL, EVT VT,
1516 bool isT, bool isO) {
1517 EVT EltVT = VT.getScalarType();
1518 assert((EltVT.getSizeInBits() >= 64 ||
1519 (uint64_t)((int64_t)Val >> EltVT.getSizeInBits()) + 1 < 2) &&
1520 "getConstant with a uint64_t value that doesn't fit in the type!");
1521 return getConstant(APInt(EltVT.getSizeInBits(), Val), DL, VT, isT, isO);
1522 }
1523
getConstant(const APInt & Val,const SDLoc & DL,EVT VT,bool isT,bool isO)1524 SDValue SelectionDAG::getConstant(const APInt &Val, const SDLoc &DL, EVT VT,
1525 bool isT, bool isO) {
1526 return getConstant(*ConstantInt::get(*Context, Val), DL, VT, isT, isO);
1527 }
1528
getConstant(const ConstantInt & Val,const SDLoc & DL,EVT VT,bool isT,bool isO)1529 SDValue SelectionDAG::getConstant(const ConstantInt &Val, const SDLoc &DL,
1530 EVT VT, bool isT, bool isO) {
1531 assert(VT.isInteger() && "Cannot create FP integer constant!");
1532
1533 EVT EltVT = VT.getScalarType();
1534 const ConstantInt *Elt = &Val;
1535
1536 // In some cases the vector type is legal but the element type is illegal and
1537 // needs to be promoted, for example v8i8 on ARM. In this case, promote the
1538 // inserted value (the type does not need to match the vector element type).
1539 // Any extra bits introduced will be truncated away.
1540 if (VT.isVector() && TLI->getTypeAction(*getContext(), EltVT) ==
1541 TargetLowering::TypePromoteInteger) {
1542 EltVT = TLI->getTypeToTransformTo(*getContext(), EltVT);
1543 APInt NewVal = Elt->getValue().zextOrTrunc(EltVT.getSizeInBits());
1544 Elt = ConstantInt::get(*getContext(), NewVal);
1545 }
1546 // In other cases the element type is illegal and needs to be expanded, for
1547 // example v2i64 on MIPS32. In this case, find the nearest legal type, split
1548 // the value into n parts and use a vector type with n-times the elements.
1549 // Then bitcast to the type requested.
1550 // Legalizing constants too early makes the DAGCombiner's job harder so we
1551 // only legalize if the DAG tells us we must produce legal types.
1552 else if (NewNodesMustHaveLegalTypes && VT.isVector() &&
1553 TLI->getTypeAction(*getContext(), EltVT) ==
1554 TargetLowering::TypeExpandInteger) {
1555 const APInt &NewVal = Elt->getValue();
1556 EVT ViaEltVT = TLI->getTypeToTransformTo(*getContext(), EltVT);
1557 unsigned ViaEltSizeInBits = ViaEltVT.getSizeInBits();
1558
1559 // For scalable vectors, try to use a SPLAT_VECTOR_PARTS node.
1560 if (VT.isScalableVector()) {
1561 assert(EltVT.getSizeInBits() % ViaEltSizeInBits == 0 &&
1562 "Can only handle an even split!");
1563 unsigned Parts = EltVT.getSizeInBits() / ViaEltSizeInBits;
1564
1565 SmallVector<SDValue, 2> ScalarParts;
1566 for (unsigned i = 0; i != Parts; ++i)
1567 ScalarParts.push_back(getConstant(
1568 NewVal.extractBits(ViaEltSizeInBits, i * ViaEltSizeInBits), DL,
1569 ViaEltVT, isT, isO));
1570
1571 return getNode(ISD::SPLAT_VECTOR_PARTS, DL, VT, ScalarParts);
1572 }
1573
1574 unsigned ViaVecNumElts = VT.getSizeInBits() / ViaEltSizeInBits;
1575 EVT ViaVecVT = EVT::getVectorVT(*getContext(), ViaEltVT, ViaVecNumElts);
1576
1577 // Check the temporary vector is the correct size. If this fails then
1578 // getTypeToTransformTo() probably returned a type whose size (in bits)
1579 // isn't a power-of-2 factor of the requested type size.
1580 assert(ViaVecVT.getSizeInBits() == VT.getSizeInBits());
1581
1582 SmallVector<SDValue, 2> EltParts;
1583 for (unsigned i = 0; i < ViaVecNumElts / VT.getVectorNumElements(); ++i)
1584 EltParts.push_back(getConstant(
1585 NewVal.extractBits(ViaEltSizeInBits, i * ViaEltSizeInBits), DL,
1586 ViaEltVT, isT, isO));
1587
1588 // EltParts is currently in little endian order. If we actually want
1589 // big-endian order then reverse it now.
1590 if (getDataLayout().isBigEndian())
1591 std::reverse(EltParts.begin(), EltParts.end());
1592
1593 // The elements must be reversed when the element order is different
1594 // to the endianness of the elements (because the BITCAST is itself a
1595 // vector shuffle in this situation). However, we do not need any code to
1596 // perform this reversal because getConstant() is producing a vector
1597 // splat.
1598 // This situation occurs in MIPS MSA.
1599
1600 SmallVector<SDValue, 8> Ops;
1601 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
1602 llvm::append_range(Ops, EltParts);
1603
1604 SDValue V =
1605 getNode(ISD::BITCAST, DL, VT, getBuildVector(ViaVecVT, DL, Ops));
1606 return V;
1607 }
1608
1609 assert(Elt->getBitWidth() == EltVT.getSizeInBits() &&
1610 "APInt size does not match type size!");
1611 unsigned Opc = isT ? ISD::TargetConstant : ISD::Constant;
1612 FoldingSetNodeID ID;
1613 AddNodeIDNode(ID, Opc, getVTList(EltVT), std::nullopt);
1614 ID.AddPointer(Elt);
1615 ID.AddBoolean(isO);
1616 void *IP = nullptr;
1617 SDNode *N = nullptr;
1618 if ((N = FindNodeOrInsertPos(ID, DL, IP)))
1619 if (!VT.isVector())
1620 return SDValue(N, 0);
1621
1622 if (!N) {
1623 N = newSDNode<ConstantSDNode>(isT, isO, Elt, EltVT);
1624 CSEMap.InsertNode(N, IP);
1625 InsertNode(N);
1626 NewSDValueDbgMsg(SDValue(N, 0), "Creating constant: ", this);
1627 }
1628
1629 SDValue Result(N, 0);
1630 if (VT.isVector())
1631 Result = getSplat(VT, DL, Result);
1632 return Result;
1633 }
1634
getIntPtrConstant(uint64_t Val,const SDLoc & DL,bool isTarget)1635 SDValue SelectionDAG::getIntPtrConstant(uint64_t Val, const SDLoc &DL,
1636 bool isTarget) {
1637 return getConstant(Val, DL, TLI->getPointerTy(getDataLayout()), isTarget);
1638 }
1639
getShiftAmountConstant(uint64_t Val,EVT VT,const SDLoc & DL,bool LegalTypes)1640 SDValue SelectionDAG::getShiftAmountConstant(uint64_t Val, EVT VT,
1641 const SDLoc &DL, bool LegalTypes) {
1642 assert(VT.isInteger() && "Shift amount is not an integer type!");
1643 EVT ShiftVT = TLI->getShiftAmountTy(VT, getDataLayout(), LegalTypes);
1644 return getConstant(Val, DL, ShiftVT);
1645 }
1646
getVectorIdxConstant(uint64_t Val,const SDLoc & DL,bool isTarget)1647 SDValue SelectionDAG::getVectorIdxConstant(uint64_t Val, const SDLoc &DL,
1648 bool isTarget) {
1649 return getConstant(Val, DL, TLI->getVectorIdxTy(getDataLayout()), isTarget);
1650 }
1651
getConstantFP(const APFloat & V,const SDLoc & DL,EVT VT,bool isTarget)1652 SDValue SelectionDAG::getConstantFP(const APFloat &V, const SDLoc &DL, EVT VT,
1653 bool isTarget) {
1654 return getConstantFP(*ConstantFP::get(*getContext(), V), DL, VT, isTarget);
1655 }
1656
getConstantFP(const ConstantFP & V,const SDLoc & DL,EVT VT,bool isTarget)1657 SDValue SelectionDAG::getConstantFP(const ConstantFP &V, const SDLoc &DL,
1658 EVT VT, bool isTarget) {
1659 assert(VT.isFloatingPoint() && "Cannot create integer FP constant!");
1660
1661 EVT EltVT = VT.getScalarType();
1662
1663 // Do the map lookup using the actual bit pattern for the floating point
1664 // value, so that we don't have problems with 0.0 comparing equal to -0.0, and
1665 // we don't have issues with SNANs.
1666 unsigned Opc = isTarget ? ISD::TargetConstantFP : ISD::ConstantFP;
1667 FoldingSetNodeID ID;
1668 AddNodeIDNode(ID, Opc, getVTList(EltVT), std::nullopt);
1669 ID.AddPointer(&V);
1670 void *IP = nullptr;
1671 SDNode *N = nullptr;
1672 if ((N = FindNodeOrInsertPos(ID, DL, IP)))
1673 if (!VT.isVector())
1674 return SDValue(N, 0);
1675
1676 if (!N) {
1677 N = newSDNode<ConstantFPSDNode>(isTarget, &V, EltVT);
1678 CSEMap.InsertNode(N, IP);
1679 InsertNode(N);
1680 }
1681
1682 SDValue Result(N, 0);
1683 if (VT.isVector())
1684 Result = getSplat(VT, DL, Result);
1685 NewSDValueDbgMsg(Result, "Creating fp constant: ", this);
1686 return Result;
1687 }
1688
getConstantFP(double Val,const SDLoc & DL,EVT VT,bool isTarget)1689 SDValue SelectionDAG::getConstantFP(double Val, const SDLoc &DL, EVT VT,
1690 bool isTarget) {
1691 EVT EltVT = VT.getScalarType();
1692 if (EltVT == MVT::f32)
1693 return getConstantFP(APFloat((float)Val), DL, VT, isTarget);
1694 if (EltVT == MVT::f64)
1695 return getConstantFP(APFloat(Val), DL, VT, isTarget);
1696 if (EltVT == MVT::f80 || EltVT == MVT::f128 || EltVT == MVT::ppcf128 ||
1697 EltVT == MVT::f16 || EltVT == MVT::bf16) {
1698 bool Ignored;
1699 APFloat APF = APFloat(Val);
1700 APF.convert(EVTToAPFloatSemantics(EltVT), APFloat::rmNearestTiesToEven,
1701 &Ignored);
1702 return getConstantFP(APF, DL, VT, isTarget);
1703 }
1704 llvm_unreachable("Unsupported type in getConstantFP");
1705 }
1706
getGlobalAddress(const GlobalValue * GV,const SDLoc & DL,EVT VT,int64_t Offset,bool isTargetGA,unsigned TargetFlags)1707 SDValue SelectionDAG::getGlobalAddress(const GlobalValue *GV, const SDLoc &DL,
1708 EVT VT, int64_t Offset, bool isTargetGA,
1709 unsigned TargetFlags) {
1710 assert((TargetFlags == 0 || isTargetGA) &&
1711 "Cannot set target flags on target-independent globals");
1712
1713 // Truncate (with sign-extension) the offset value to the pointer size.
1714 unsigned BitWidth = getDataLayout().getPointerTypeSizeInBits(GV->getType());
1715 if (BitWidth < 64)
1716 Offset = SignExtend64(Offset, BitWidth);
1717
1718 unsigned Opc;
1719 if (GV->isThreadLocal())
1720 Opc = isTargetGA ? ISD::TargetGlobalTLSAddress : ISD::GlobalTLSAddress;
1721 else
1722 Opc = isTargetGA ? ISD::TargetGlobalAddress : ISD::GlobalAddress;
1723
1724 FoldingSetNodeID ID;
1725 AddNodeIDNode(ID, Opc, getVTList(VT), std::nullopt);
1726 ID.AddPointer(GV);
1727 ID.AddInteger(Offset);
1728 ID.AddInteger(TargetFlags);
1729 void *IP = nullptr;
1730 if (SDNode *E = FindNodeOrInsertPos(ID, DL, IP))
1731 return SDValue(E, 0);
1732
1733 auto *N = newSDNode<GlobalAddressSDNode>(
1734 Opc, DL.getIROrder(), DL.getDebugLoc(), GV, VT, Offset, TargetFlags);
1735 CSEMap.InsertNode(N, IP);
1736 InsertNode(N);
1737 return SDValue(N, 0);
1738 }
1739
getFrameIndex(int FI,EVT VT,bool isTarget)1740 SDValue SelectionDAG::getFrameIndex(int FI, EVT VT, bool isTarget) {
1741 unsigned Opc = isTarget ? ISD::TargetFrameIndex : ISD::FrameIndex;
1742 FoldingSetNodeID ID;
1743 AddNodeIDNode(ID, Opc, getVTList(VT), std::nullopt);
1744 ID.AddInteger(FI);
1745 void *IP = nullptr;
1746 if (SDNode *E = FindNodeOrInsertPos(ID, IP))
1747 return SDValue(E, 0);
1748
1749 auto *N = newSDNode<FrameIndexSDNode>(FI, VT, isTarget);
1750 CSEMap.InsertNode(N, IP);
1751 InsertNode(N);
1752 return SDValue(N, 0);
1753 }
1754
getJumpTable(int JTI,EVT VT,bool isTarget,unsigned TargetFlags)1755 SDValue SelectionDAG::getJumpTable(int JTI, EVT VT, bool isTarget,
1756 unsigned TargetFlags) {
1757 assert((TargetFlags == 0 || isTarget) &&
1758 "Cannot set target flags on target-independent jump tables");
1759 unsigned Opc = isTarget ? ISD::TargetJumpTable : ISD::JumpTable;
1760 FoldingSetNodeID ID;
1761 AddNodeIDNode(ID, Opc, getVTList(VT), std::nullopt);
1762 ID.AddInteger(JTI);
1763 ID.AddInteger(TargetFlags);
1764 void *IP = nullptr;
1765 if (SDNode *E = FindNodeOrInsertPos(ID, IP))
1766 return SDValue(E, 0);
1767
1768 auto *N = newSDNode<JumpTableSDNode>(JTI, VT, isTarget, TargetFlags);
1769 CSEMap.InsertNode(N, IP);
1770 InsertNode(N);
1771 return SDValue(N, 0);
1772 }
1773
getConstantPool(const Constant * C,EVT VT,MaybeAlign Alignment,int Offset,bool isTarget,unsigned TargetFlags)1774 SDValue SelectionDAG::getConstantPool(const Constant *C, EVT VT,
1775 MaybeAlign Alignment, int Offset,
1776 bool isTarget, unsigned TargetFlags) {
1777 assert((TargetFlags == 0 || isTarget) &&
1778 "Cannot set target flags on target-independent globals");
1779 if (!Alignment)
1780 Alignment = shouldOptForSize()
1781 ? getDataLayout().getABITypeAlign(C->getType())
1782 : getDataLayout().getPrefTypeAlign(C->getType());
1783 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool;
1784 FoldingSetNodeID ID;
1785 AddNodeIDNode(ID, Opc, getVTList(VT), std::nullopt);
1786 ID.AddInteger(Alignment->value());
1787 ID.AddInteger(Offset);
1788 ID.AddPointer(C);
1789 ID.AddInteger(TargetFlags);
1790 void *IP = nullptr;
1791 if (SDNode *E = FindNodeOrInsertPos(ID, IP))
1792 return SDValue(E, 0);
1793
1794 auto *N = newSDNode<ConstantPoolSDNode>(isTarget, C, VT, Offset, *Alignment,
1795 TargetFlags);
1796 CSEMap.InsertNode(N, IP);
1797 InsertNode(N);
1798 SDValue V = SDValue(N, 0);
1799 NewSDValueDbgMsg(V, "Creating new constant pool: ", this);
1800 return V;
1801 }
1802
getConstantPool(MachineConstantPoolValue * C,EVT VT,MaybeAlign Alignment,int Offset,bool isTarget,unsigned TargetFlags)1803 SDValue SelectionDAG::getConstantPool(MachineConstantPoolValue *C, EVT VT,
1804 MaybeAlign Alignment, int Offset,
1805 bool isTarget, unsigned TargetFlags) {
1806 assert((TargetFlags == 0 || isTarget) &&
1807 "Cannot set target flags on target-independent globals");
1808 if (!Alignment)
1809 Alignment = getDataLayout().getPrefTypeAlign(C->getType());
1810 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool;
1811 FoldingSetNodeID ID;
1812 AddNodeIDNode(ID, Opc, getVTList(VT), std::nullopt);
1813 ID.AddInteger(Alignment->value());
1814 ID.AddInteger(Offset);
1815 C->addSelectionDAGCSEId(ID);
1816 ID.AddInteger(TargetFlags);
1817 void *IP = nullptr;
1818 if (SDNode *E = FindNodeOrInsertPos(ID, IP))
1819 return SDValue(E, 0);
1820
1821 auto *N = newSDNode<ConstantPoolSDNode>(isTarget, C, VT, Offset, *Alignment,
1822 TargetFlags);
1823 CSEMap.InsertNode(N, IP);
1824 InsertNode(N);
1825 return SDValue(N, 0);
1826 }
1827
getTargetIndex(int Index,EVT VT,int64_t Offset,unsigned TargetFlags)1828 SDValue SelectionDAG::getTargetIndex(int Index, EVT VT, int64_t Offset,
1829 unsigned TargetFlags) {
1830 FoldingSetNodeID ID;
1831 AddNodeIDNode(ID, ISD::TargetIndex, getVTList(VT), std::nullopt);
1832 ID.AddInteger(Index);
1833 ID.AddInteger(Offset);
1834 ID.AddInteger(TargetFlags);
1835 void *IP = nullptr;
1836 if (SDNode *E = FindNodeOrInsertPos(ID, IP))
1837 return SDValue(E, 0);
1838
1839 auto *N = newSDNode<TargetIndexSDNode>(Index, VT, Offset, TargetFlags);
1840 CSEMap.InsertNode(N, IP);
1841 InsertNode(N);
1842 return SDValue(N, 0);
1843 }
1844
getBasicBlock(MachineBasicBlock * MBB)1845 SDValue SelectionDAG::getBasicBlock(MachineBasicBlock *MBB) {
1846 FoldingSetNodeID ID;
1847 AddNodeIDNode(ID, ISD::BasicBlock, getVTList(MVT::Other), std::nullopt);
1848 ID.AddPointer(MBB);
1849 void *IP = nullptr;
1850 if (SDNode *E = FindNodeOrInsertPos(ID, IP))
1851 return SDValue(E, 0);
1852
1853 auto *N = newSDNode<BasicBlockSDNode>(MBB);
1854 CSEMap.InsertNode(N, IP);
1855 InsertNode(N);
1856 return SDValue(N, 0);
1857 }
1858
getValueType(EVT VT)1859 SDValue SelectionDAG::getValueType(EVT VT) {
1860 if (VT.isSimple() && (unsigned)VT.getSimpleVT().SimpleTy >=
1861 ValueTypeNodes.size())
1862 ValueTypeNodes.resize(VT.getSimpleVT().SimpleTy+1);
1863
1864 SDNode *&N = VT.isExtended() ?
1865 ExtendedValueTypeNodes[VT] : ValueTypeNodes[VT.getSimpleVT().SimpleTy];
1866
1867 if (N) return SDValue(N, 0);
1868 N = newSDNode<VTSDNode>(VT);
1869 InsertNode(N);
1870 return SDValue(N, 0);
1871 }
1872
getExternalSymbol(const char * Sym,EVT VT)1873 SDValue SelectionDAG::getExternalSymbol(const char *Sym, EVT VT) {
1874 SDNode *&N = ExternalSymbols[Sym];
1875 if (N) return SDValue(N, 0);
1876 N = newSDNode<ExternalSymbolSDNode>(false, Sym, 0, VT);
1877 InsertNode(N);
1878 return SDValue(N, 0);
1879 }
1880
getMCSymbol(MCSymbol * Sym,EVT VT)1881 SDValue SelectionDAG::getMCSymbol(MCSymbol *Sym, EVT VT) {
1882 SDNode *&N = MCSymbols[Sym];
1883 if (N)
1884 return SDValue(N, 0);
1885 N = newSDNode<MCSymbolSDNode>(Sym, VT);
1886 InsertNode(N);
1887 return SDValue(N, 0);
1888 }
1889
getTargetExternalSymbol(const char * Sym,EVT VT,unsigned TargetFlags)1890 SDValue SelectionDAG::getTargetExternalSymbol(const char *Sym, EVT VT,
1891 unsigned TargetFlags) {
1892 SDNode *&N =
1893 TargetExternalSymbols[std::pair<std::string, unsigned>(Sym, TargetFlags)];
1894 if (N) return SDValue(N, 0);
1895 N = newSDNode<ExternalSymbolSDNode>(true, Sym, TargetFlags, VT);
1896 InsertNode(N);
1897 return SDValue(N, 0);
1898 }
1899
getCondCode(ISD::CondCode Cond)1900 SDValue SelectionDAG::getCondCode(ISD::CondCode Cond) {
1901 if ((unsigned)Cond >= CondCodeNodes.size())
1902 CondCodeNodes.resize(Cond+1);
1903
1904 if (!CondCodeNodes[Cond]) {
1905 auto *N = newSDNode<CondCodeSDNode>(Cond);
1906 CondCodeNodes[Cond] = N;
1907 InsertNode(N);
1908 }
1909
1910 return SDValue(CondCodeNodes[Cond], 0);
1911 }
1912
getStepVector(const SDLoc & DL,EVT ResVT)1913 SDValue SelectionDAG::getStepVector(const SDLoc &DL, EVT ResVT) {
1914 APInt One(ResVT.getScalarSizeInBits(), 1);
1915 return getStepVector(DL, ResVT, One);
1916 }
1917
getStepVector(const SDLoc & DL,EVT ResVT,APInt StepVal)1918 SDValue SelectionDAG::getStepVector(const SDLoc &DL, EVT ResVT, APInt StepVal) {
1919 assert(ResVT.getScalarSizeInBits() == StepVal.getBitWidth());
1920 if (ResVT.isScalableVector())
1921 return getNode(
1922 ISD::STEP_VECTOR, DL, ResVT,
1923 getTargetConstant(StepVal, DL, ResVT.getVectorElementType()));
1924
1925 SmallVector<SDValue, 16> OpsStepConstants;
1926 for (uint64_t i = 0; i < ResVT.getVectorNumElements(); i++)
1927 OpsStepConstants.push_back(
1928 getConstant(StepVal * i, DL, ResVT.getVectorElementType()));
1929 return getBuildVector(ResVT, DL, OpsStepConstants);
1930 }
1931
1932 /// Swaps the values of N1 and N2. Swaps all indices in the shuffle mask M that
1933 /// point at N1 to point at N2 and indices that point at N2 to point at N1.
commuteShuffle(SDValue & N1,SDValue & N2,MutableArrayRef<int> M)1934 static void commuteShuffle(SDValue &N1, SDValue &N2, MutableArrayRef<int> M) {
1935 std::swap(N1, N2);
1936 ShuffleVectorSDNode::commuteMask(M);
1937 }
1938
getVectorShuffle(EVT VT,const SDLoc & dl,SDValue N1,SDValue N2,ArrayRef<int> Mask)1939 SDValue SelectionDAG::getVectorShuffle(EVT VT, const SDLoc &dl, SDValue N1,
1940 SDValue N2, ArrayRef<int> Mask) {
1941 assert(VT.getVectorNumElements() == Mask.size() &&
1942 "Must have the same number of vector elements as mask elements!");
1943 assert(VT == N1.getValueType() && VT == N2.getValueType() &&
1944 "Invalid VECTOR_SHUFFLE");
1945
1946 // Canonicalize shuffle undef, undef -> undef
1947 if (N1.isUndef() && N2.isUndef())
1948 return getUNDEF(VT);
1949
1950 // Validate that all indices in Mask are within the range of the elements
1951 // input to the shuffle.
1952 int NElts = Mask.size();
1953 assert(llvm::all_of(Mask,
1954 [&](int M) { return M < (NElts * 2) && M >= -1; }) &&
1955 "Index out of range");
1956
1957 // Copy the mask so we can do any needed cleanup.
1958 SmallVector<int, 8> MaskVec(Mask);
1959
1960 // Canonicalize shuffle v, v -> v, undef
1961 if (N1 == N2) {
1962 N2 = getUNDEF(VT);
1963 for (int i = 0; i != NElts; ++i)
1964 if (MaskVec[i] >= NElts) MaskVec[i] -= NElts;
1965 }
1966
1967 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
1968 if (N1.isUndef())
1969 commuteShuffle(N1, N2, MaskVec);
1970
1971 if (TLI->hasVectorBlend()) {
1972 // If shuffling a splat, try to blend the splat instead. We do this here so
1973 // that even when this arises during lowering we don't have to re-handle it.
1974 auto BlendSplat = [&](BuildVectorSDNode *BV, int Offset) {
1975 BitVector UndefElements;
1976 SDValue Splat = BV->getSplatValue(&UndefElements);
1977 if (!Splat)
1978 return;
1979
1980 for (int i = 0; i < NElts; ++i) {
1981 if (MaskVec[i] < Offset || MaskVec[i] >= (Offset + NElts))
1982 continue;
1983
1984 // If this input comes from undef, mark it as such.
1985 if (UndefElements[MaskVec[i] - Offset]) {
1986 MaskVec[i] = -1;
1987 continue;
1988 }
1989
1990 // If we can blend a non-undef lane, use that instead.
1991 if (!UndefElements[i])
1992 MaskVec[i] = i + Offset;
1993 }
1994 };
1995 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
1996 BlendSplat(N1BV, 0);
1997 if (auto *N2BV = dyn_cast<BuildVectorSDNode>(N2))
1998 BlendSplat(N2BV, NElts);
1999 }
2000
2001 // Canonicalize all index into lhs, -> shuffle lhs, undef
2002 // Canonicalize all index into rhs, -> shuffle rhs, undef
2003 bool AllLHS = true, AllRHS = true;
2004 bool N2Undef = N2.isUndef();
2005 for (int i = 0; i != NElts; ++i) {
2006 if (MaskVec[i] >= NElts) {
2007 if (N2Undef)
2008 MaskVec[i] = -1;
2009 else
2010 AllLHS = false;
2011 } else if (MaskVec[i] >= 0) {
2012 AllRHS = false;
2013 }
2014 }
2015 if (AllLHS && AllRHS)
2016 return getUNDEF(VT);
2017 if (AllLHS && !N2Undef)
2018 N2 = getUNDEF(VT);
2019 if (AllRHS) {
2020 N1 = getUNDEF(VT);
2021 commuteShuffle(N1, N2, MaskVec);
2022 }
2023 // Reset our undef status after accounting for the mask.
2024 N2Undef = N2.isUndef();
2025 // Re-check whether both sides ended up undef.
2026 if (N1.isUndef() && N2Undef)
2027 return getUNDEF(VT);
2028
2029 // If Identity shuffle return that node.
2030 bool Identity = true, AllSame = true;
2031 for (int i = 0; i != NElts; ++i) {
2032 if (MaskVec[i] >= 0 && MaskVec[i] != i) Identity = false;
2033 if (MaskVec[i] != MaskVec[0]) AllSame = false;
2034 }
2035 if (Identity && NElts)
2036 return N1;
2037
2038 // Shuffling a constant splat doesn't change the result.
2039 if (N2Undef) {
2040 SDValue V = N1;
2041
2042 // Look through any bitcasts. We check that these don't change the number
2043 // (and size) of elements and just changes their types.
2044 while (V.getOpcode() == ISD::BITCAST)
2045 V = V->getOperand(0);
2046
2047 // A splat should always show up as a build vector node.
2048 if (auto *BV = dyn_cast<BuildVectorSDNode>(V)) {
2049 BitVector UndefElements;
2050 SDValue Splat = BV->getSplatValue(&UndefElements);
2051 // If this is a splat of an undef, shuffling it is also undef.
2052 if (Splat && Splat.isUndef())
2053 return getUNDEF(VT);
2054
2055 bool SameNumElts =
2056 V.getValueType().getVectorNumElements() == VT.getVectorNumElements();
2057
2058 // We only have a splat which can skip shuffles if there is a splatted
2059 // value and no undef lanes rearranged by the shuffle.
2060 if (Splat && UndefElements.none()) {
2061 // Splat of <x, x, ..., x>, return <x, x, ..., x>, provided that the
2062 // number of elements match or the value splatted is a zero constant.
2063 if (SameNumElts)
2064 return N1;
2065 if (auto *C = dyn_cast<ConstantSDNode>(Splat))
2066 if (C->isZero())
2067 return N1;
2068 }
2069
2070 // If the shuffle itself creates a splat, build the vector directly.
2071 if (AllSame && SameNumElts) {
2072 EVT BuildVT = BV->getValueType(0);
2073 const SDValue &Splatted = BV->getOperand(MaskVec[0]);
2074 SDValue NewBV = getSplatBuildVector(BuildVT, dl, Splatted);
2075
2076 // We may have jumped through bitcasts, so the type of the
2077 // BUILD_VECTOR may not match the type of the shuffle.
2078 if (BuildVT != VT)
2079 NewBV = getNode(ISD::BITCAST, dl, VT, NewBV);
2080 return NewBV;
2081 }
2082 }
2083 }
2084
2085 FoldingSetNodeID ID;
2086 SDValue Ops[2] = { N1, N2 };
2087 AddNodeIDNode(ID, ISD::VECTOR_SHUFFLE, getVTList(VT), Ops);
2088 for (int i = 0; i != NElts; ++i)
2089 ID.AddInteger(MaskVec[i]);
2090
2091 void* IP = nullptr;
2092 if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP))
2093 return SDValue(E, 0);
2094
2095 // Allocate the mask array for the node out of the BumpPtrAllocator, since
2096 // SDNode doesn't have access to it. This memory will be "leaked" when
2097 // the node is deallocated, but recovered when the NodeAllocator is released.
2098 int *MaskAlloc = OperandAllocator.Allocate<int>(NElts);
2099 llvm::copy(MaskVec, MaskAlloc);
2100
2101 auto *N = newSDNode<ShuffleVectorSDNode>(VT, dl.getIROrder(),
2102 dl.getDebugLoc(), MaskAlloc);
2103 createOperands(N, Ops);
2104
2105 CSEMap.InsertNode(N, IP);
2106 InsertNode(N);
2107 SDValue V = SDValue(N, 0);
2108 NewSDValueDbgMsg(V, "Creating new node: ", this);
2109 return V;
2110 }
2111
getCommutedVectorShuffle(const ShuffleVectorSDNode & SV)2112 SDValue SelectionDAG::getCommutedVectorShuffle(const ShuffleVectorSDNode &SV) {
2113 EVT VT = SV.getValueType(0);
2114 SmallVector<int, 8> MaskVec(SV.getMask());
2115 ShuffleVectorSDNode::commuteMask(MaskVec);
2116
2117 SDValue Op0 = SV.getOperand(0);
2118 SDValue Op1 = SV.getOperand(1);
2119 return getVectorShuffle(VT, SDLoc(&SV), Op1, Op0, MaskVec);
2120 }
2121
getRegister(unsigned RegNo,EVT VT)2122 SDValue SelectionDAG::getRegister(unsigned RegNo, EVT VT) {
2123 FoldingSetNodeID ID;
2124 AddNodeIDNode(ID, ISD::Register, getVTList(VT), std::nullopt);
2125 ID.AddInteger(RegNo);
2126 void *IP = nullptr;
2127 if (SDNode *E = FindNodeOrInsertPos(ID, IP))
2128 return SDValue(E, 0);
2129
2130 auto *N = newSDNode<RegisterSDNode>(RegNo, VT);
2131 N->SDNodeBits.IsDivergent = TLI->isSDNodeSourceOfDivergence(N, FLI, DA);
2132 CSEMap.InsertNode(N, IP);
2133 InsertNode(N);
2134 return SDValue(N, 0);
2135 }
2136
getRegisterMask(const uint32_t * RegMask)2137 SDValue SelectionDAG::getRegisterMask(const uint32_t *RegMask) {
2138 FoldingSetNodeID ID;
2139 AddNodeIDNode(ID, ISD::RegisterMask, getVTList(MVT::Untyped), std::nullopt);
2140 ID.AddPointer(RegMask);
2141 void *IP = nullptr;
2142 if (SDNode *E = FindNodeOrInsertPos(ID, IP))
2143 return SDValue(E, 0);
2144
2145 auto *N = newSDNode<RegisterMaskSDNode>(RegMask);
2146 CSEMap.InsertNode(N, IP);
2147 InsertNode(N);
2148 return SDValue(N, 0);
2149 }
2150
getEHLabel(const SDLoc & dl,SDValue Root,MCSymbol * Label)2151 SDValue SelectionDAG::getEHLabel(const SDLoc &dl, SDValue Root,
2152 MCSymbol *Label) {
2153 return getLabelNode(ISD::EH_LABEL, dl, Root, Label);
2154 }
2155
getLabelNode(unsigned Opcode,const SDLoc & dl,SDValue Root,MCSymbol * Label)2156 SDValue SelectionDAG::getLabelNode(unsigned Opcode, const SDLoc &dl,
2157 SDValue Root, MCSymbol *Label) {
2158 FoldingSetNodeID ID;
2159 SDValue Ops[] = { Root };
2160 AddNodeIDNode(ID, Opcode, getVTList(MVT::Other), Ops);
2161 ID.AddPointer(Label);
2162 void *IP = nullptr;
2163 if (SDNode *E = FindNodeOrInsertPos(ID, IP))
2164 return SDValue(E, 0);
2165
2166 auto *N =
2167 newSDNode<LabelSDNode>(Opcode, dl.getIROrder(), dl.getDebugLoc(), Label);
2168 createOperands(N, Ops);
2169
2170 CSEMap.InsertNode(N, IP);
2171 InsertNode(N);
2172 return SDValue(N, 0);
2173 }
2174
getBlockAddress(const BlockAddress * BA,EVT VT,int64_t Offset,bool isTarget,unsigned TargetFlags)2175 SDValue SelectionDAG::getBlockAddress(const BlockAddress *BA, EVT VT,
2176 int64_t Offset, bool isTarget,
2177 unsigned TargetFlags) {
2178 unsigned Opc = isTarget ? ISD::TargetBlockAddress : ISD::BlockAddress;
2179
2180 FoldingSetNodeID ID;
2181 AddNodeIDNode(ID, Opc, getVTList(VT), std::nullopt);
2182 ID.AddPointer(BA);
2183 ID.AddInteger(Offset);
2184 ID.AddInteger(TargetFlags);
2185 void *IP = nullptr;
2186 if (SDNode *E = FindNodeOrInsertPos(ID, IP))
2187 return SDValue(E, 0);
2188
2189 auto *N = newSDNode<BlockAddressSDNode>(Opc, VT, BA, Offset, TargetFlags);
2190 CSEMap.InsertNode(N, IP);
2191 InsertNode(N);
2192 return SDValue(N, 0);
2193 }
2194
getSrcValue(const Value * V)2195 SDValue SelectionDAG::getSrcValue(const Value *V) {
2196 FoldingSetNodeID ID;
2197 AddNodeIDNode(ID, ISD::SRCVALUE, getVTList(MVT::Other), std::nullopt);
2198 ID.AddPointer(V);
2199
2200 void *IP = nullptr;
2201 if (SDNode *E = FindNodeOrInsertPos(ID, IP))
2202 return SDValue(E, 0);
2203
2204 auto *N = newSDNode<SrcValueSDNode>(V);
2205 CSEMap.InsertNode(N, IP);
2206 InsertNode(N);
2207 return SDValue(N, 0);
2208 }
2209
getMDNode(const MDNode * MD)2210 SDValue SelectionDAG::getMDNode(const MDNode *MD) {
2211 FoldingSetNodeID ID;
2212 AddNodeIDNode(ID, ISD::MDNODE_SDNODE, getVTList(MVT::Other), std::nullopt);
2213 ID.AddPointer(MD);
2214
2215 void *IP = nullptr;
2216 if (SDNode *E = FindNodeOrInsertPos(ID, IP))
2217 return SDValue(E, 0);
2218
2219 auto *N = newSDNode<MDNodeSDNode>(MD);
2220 CSEMap.InsertNode(N, IP);
2221 InsertNode(N);
2222 return SDValue(N, 0);
2223 }
2224
getBitcast(EVT VT,SDValue V)2225 SDValue SelectionDAG::getBitcast(EVT VT, SDValue V) {
2226 if (VT == V.getValueType())
2227 return V;
2228
2229 return getNode(ISD::BITCAST, SDLoc(V), VT, V);
2230 }
2231
getAddrSpaceCast(const SDLoc & dl,EVT VT,SDValue Ptr,unsigned SrcAS,unsigned DestAS)2232 SDValue SelectionDAG::getAddrSpaceCast(const SDLoc &dl, EVT VT, SDValue Ptr,
2233 unsigned SrcAS, unsigned DestAS) {
2234 SDValue Ops[] = {Ptr};
2235 FoldingSetNodeID ID;
2236 AddNodeIDNode(ID, ISD::ADDRSPACECAST, getVTList(VT), Ops);
2237 ID.AddInteger(SrcAS);
2238 ID.AddInteger(DestAS);
2239
2240 void *IP = nullptr;
2241 if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP))
2242 return SDValue(E, 0);
2243
2244 auto *N = newSDNode<AddrSpaceCastSDNode>(dl.getIROrder(), dl.getDebugLoc(),
2245 VT, SrcAS, DestAS);
2246 createOperands(N, Ops);
2247
2248 CSEMap.InsertNode(N, IP);
2249 InsertNode(N);
2250 return SDValue(N, 0);
2251 }
2252
getFreeze(SDValue V)2253 SDValue SelectionDAG::getFreeze(SDValue V) {
2254 return getNode(ISD::FREEZE, SDLoc(V), V.getValueType(), V);
2255 }
2256
2257 /// getShiftAmountOperand - Return the specified value casted to
2258 /// the target's desired shift amount type.
getShiftAmountOperand(EVT LHSTy,SDValue Op)2259 SDValue SelectionDAG::getShiftAmountOperand(EVT LHSTy, SDValue Op) {
2260 EVT OpTy = Op.getValueType();
2261 EVT ShTy = TLI->getShiftAmountTy(LHSTy, getDataLayout());
2262 if (OpTy == ShTy || OpTy.isVector()) return Op;
2263
2264 return getZExtOrTrunc(Op, SDLoc(Op), ShTy);
2265 }
2266
expandVAArg(SDNode * Node)2267 SDValue SelectionDAG::expandVAArg(SDNode *Node) {
2268 SDLoc dl(Node);
2269 const TargetLowering &TLI = getTargetLoweringInfo();
2270 const Value *V = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
2271 EVT VT = Node->getValueType(0);
2272 SDValue Tmp1 = Node->getOperand(0);
2273 SDValue Tmp2 = Node->getOperand(1);
2274 const MaybeAlign MA(Node->getConstantOperandVal(3));
2275
2276 SDValue VAListLoad = getLoad(TLI.getPointerTy(getDataLayout()), dl, Tmp1,
2277 Tmp2, MachinePointerInfo(V));
2278 SDValue VAList = VAListLoad;
2279
2280 if (MA && *MA > TLI.getMinStackArgumentAlignment()) {
2281 VAList = getNode(ISD::ADD, dl, VAList.getValueType(), VAList,
2282 getConstant(MA->value() - 1, dl, VAList.getValueType()));
2283
2284 VAList =
2285 getNode(ISD::AND, dl, VAList.getValueType(), VAList,
2286 getConstant(-(int64_t)MA->value(), dl, VAList.getValueType()));
2287 }
2288
2289 // Increment the pointer, VAList, to the next vaarg
2290 Tmp1 = getNode(ISD::ADD, dl, VAList.getValueType(), VAList,
2291 getConstant(getDataLayout().getTypeAllocSize(
2292 VT.getTypeForEVT(*getContext())),
2293 dl, VAList.getValueType()));
2294 // Store the incremented VAList to the legalized pointer
2295 Tmp1 =
2296 getStore(VAListLoad.getValue(1), dl, Tmp1, Tmp2, MachinePointerInfo(V));
2297 // Load the actual argument out of the pointer VAList
2298 return getLoad(VT, dl, Tmp1, VAList, MachinePointerInfo());
2299 }
2300
expandVACopy(SDNode * Node)2301 SDValue SelectionDAG::expandVACopy(SDNode *Node) {
2302 SDLoc dl(Node);
2303 const TargetLowering &TLI = getTargetLoweringInfo();
2304 // This defaults to loading a pointer from the input and storing it to the
2305 // output, returning the chain.
2306 const Value *VD = cast<SrcValueSDNode>(Node->getOperand(3))->getValue();
2307 const Value *VS = cast<SrcValueSDNode>(Node->getOperand(4))->getValue();
2308 SDValue Tmp1 =
2309 getLoad(TLI.getPointerTy(getDataLayout()), dl, Node->getOperand(0),
2310 Node->getOperand(2), MachinePointerInfo(VS));
2311 return getStore(Tmp1.getValue(1), dl, Tmp1, Node->getOperand(1),
2312 MachinePointerInfo(VD));
2313 }
2314
getReducedAlign(EVT VT,bool UseABI)2315 Align SelectionDAG::getReducedAlign(EVT VT, bool UseABI) {
2316 const DataLayout &DL = getDataLayout();
2317 Type *Ty = VT.getTypeForEVT(*getContext());
2318 Align RedAlign = UseABI ? DL.getABITypeAlign(Ty) : DL.getPrefTypeAlign(Ty);
2319
2320 if (TLI->isTypeLegal(VT) || !VT.isVector())
2321 return RedAlign;
2322
2323 const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
2324 const Align StackAlign = TFI->getStackAlign();
2325
2326 // See if we can choose a smaller ABI alignment in cases where it's an
2327 // illegal vector type that will get broken down.
2328 if (RedAlign > StackAlign) {
2329 EVT IntermediateVT;
2330 MVT RegisterVT;
2331 unsigned NumIntermediates;
2332 TLI->getVectorTypeBreakdown(*getContext(), VT, IntermediateVT,
2333 NumIntermediates, RegisterVT);
2334 Ty = IntermediateVT.getTypeForEVT(*getContext());
2335 Align RedAlign2 = UseABI ? DL.getABITypeAlign(Ty) : DL.getPrefTypeAlign(Ty);
2336 if (RedAlign2 < RedAlign)
2337 RedAlign = RedAlign2;
2338 }
2339
2340 return RedAlign;
2341 }
2342
CreateStackTemporary(TypeSize Bytes,Align Alignment)2343 SDValue SelectionDAG::CreateStackTemporary(TypeSize Bytes, Align Alignment) {
2344 MachineFrameInfo &MFI = MF->getFrameInfo();
2345 const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
2346 int StackID = 0;
2347 if (Bytes.isScalable())
2348 StackID = TFI->getStackIDForScalableVectors();
2349 // The stack id gives an indication of whether the object is scalable or
2350 // not, so it's safe to pass in the minimum size here.
2351 int FrameIdx = MFI.CreateStackObject(Bytes.getKnownMinValue(), Alignment,
2352 false, nullptr, StackID);
2353 return getFrameIndex(FrameIdx, TLI->getFrameIndexTy(getDataLayout()));
2354 }
2355
CreateStackTemporary(EVT VT,unsigned minAlign)2356 SDValue SelectionDAG::CreateStackTemporary(EVT VT, unsigned minAlign) {
2357 Type *Ty = VT.getTypeForEVT(*getContext());
2358 Align StackAlign =
2359 std::max(getDataLayout().getPrefTypeAlign(Ty), Align(minAlign));
2360 return CreateStackTemporary(VT.getStoreSize(), StackAlign);
2361 }
2362
CreateStackTemporary(EVT VT1,EVT VT2)2363 SDValue SelectionDAG::CreateStackTemporary(EVT VT1, EVT VT2) {
2364 TypeSize VT1Size = VT1.getStoreSize();
2365 TypeSize VT2Size = VT2.getStoreSize();
2366 assert(VT1Size.isScalable() == VT2Size.isScalable() &&
2367 "Don't know how to choose the maximum size when creating a stack "
2368 "temporary");
2369 TypeSize Bytes = VT1Size.getKnownMinValue() > VT2Size.getKnownMinValue()
2370 ? VT1Size
2371 : VT2Size;
2372
2373 Type *Ty1 = VT1.getTypeForEVT(*getContext());
2374 Type *Ty2 = VT2.getTypeForEVT(*getContext());
2375 const DataLayout &DL = getDataLayout();
2376 Align Align = std::max(DL.getPrefTypeAlign(Ty1), DL.getPrefTypeAlign(Ty2));
2377 return CreateStackTemporary(Bytes, Align);
2378 }
2379
FoldSetCC(EVT VT,SDValue N1,SDValue N2,ISD::CondCode Cond,const SDLoc & dl)2380 SDValue SelectionDAG::FoldSetCC(EVT VT, SDValue N1, SDValue N2,
2381 ISD::CondCode Cond, const SDLoc &dl) {
2382 EVT OpVT = N1.getValueType();
2383
2384 // These setcc operations always fold.
2385 switch (Cond) {
2386 default: break;
2387 case ISD::SETFALSE:
2388 case ISD::SETFALSE2: return getBoolConstant(false, dl, VT, OpVT);
2389 case ISD::SETTRUE:
2390 case ISD::SETTRUE2: return getBoolConstant(true, dl, VT, OpVT);
2391
2392 case ISD::SETOEQ:
2393 case ISD::SETOGT:
2394 case ISD::SETOGE:
2395 case ISD::SETOLT:
2396 case ISD::SETOLE:
2397 case ISD::SETONE:
2398 case ISD::SETO:
2399 case ISD::SETUO:
2400 case ISD::SETUEQ:
2401 case ISD::SETUNE:
2402 assert(!OpVT.isInteger() && "Illegal setcc for integer!");
2403 break;
2404 }
2405
2406 if (OpVT.isInteger()) {
2407 // For EQ and NE, we can always pick a value for the undef to make the
2408 // predicate pass or fail, so we can return undef.
2409 // Matches behavior in llvm::ConstantFoldCompareInstruction.
2410 // icmp eq/ne X, undef -> undef.
2411 if ((N1.isUndef() || N2.isUndef()) &&
2412 (Cond == ISD::SETEQ || Cond == ISD::SETNE))
2413 return getUNDEF(VT);
2414
2415 // If both operands are undef, we can return undef for int comparison.
2416 // icmp undef, undef -> undef.
2417 if (N1.isUndef() && N2.isUndef())
2418 return getUNDEF(VT);
2419
2420 // icmp X, X -> true/false
2421 // icmp X, undef -> true/false because undef could be X.
2422 if (N1 == N2)
2423 return getBoolConstant(ISD::isTrueWhenEqual(Cond), dl, VT, OpVT);
2424 }
2425
2426 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2)) {
2427 const APInt &C2 = N2C->getAPIntValue();
2428 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1)) {
2429 const APInt &C1 = N1C->getAPIntValue();
2430
2431 return getBoolConstant(ICmpInst::compare(C1, C2, getICmpCondCode(Cond)),
2432 dl, VT, OpVT);
2433 }
2434 }
2435
2436 auto *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2437 auto *N2CFP = dyn_cast<ConstantFPSDNode>(N2);
2438
2439 if (N1CFP && N2CFP) {
2440 APFloat::cmpResult R = N1CFP->getValueAPF().compare(N2CFP->getValueAPF());
2441 switch (Cond) {
2442 default: break;
2443 case ISD::SETEQ: if (R==APFloat::cmpUnordered)
2444 return getUNDEF(VT);
2445 [[fallthrough]];
2446 case ISD::SETOEQ: return getBoolConstant(R==APFloat::cmpEqual, dl, VT,
2447 OpVT);
2448 case ISD::SETNE: if (R==APFloat::cmpUnordered)
2449 return getUNDEF(VT);
2450 [[fallthrough]];
2451 case ISD::SETONE: return getBoolConstant(R==APFloat::cmpGreaterThan ||
2452 R==APFloat::cmpLessThan, dl, VT,
2453 OpVT);
2454 case ISD::SETLT: if (R==APFloat::cmpUnordered)
2455 return getUNDEF(VT);
2456 [[fallthrough]];
2457 case ISD::SETOLT: return getBoolConstant(R==APFloat::cmpLessThan, dl, VT,
2458 OpVT);
2459 case ISD::SETGT: if (R==APFloat::cmpUnordered)
2460 return getUNDEF(VT);
2461 [[fallthrough]];
2462 case ISD::SETOGT: return getBoolConstant(R==APFloat::cmpGreaterThan, dl,
2463 VT, OpVT);
2464 case ISD::SETLE: if (R==APFloat::cmpUnordered)
2465 return getUNDEF(VT);
2466 [[fallthrough]];
2467 case ISD::SETOLE: return getBoolConstant(R==APFloat::cmpLessThan ||
2468 R==APFloat::cmpEqual, dl, VT,
2469 OpVT);
2470 case ISD::SETGE: if (R==APFloat::cmpUnordered)
2471 return getUNDEF(VT);
2472 [[fallthrough]];
2473 case ISD::SETOGE: return getBoolConstant(R==APFloat::cmpGreaterThan ||
2474 R==APFloat::cmpEqual, dl, VT, OpVT);
2475 case ISD::SETO: return getBoolConstant(R!=APFloat::cmpUnordered, dl, VT,
2476 OpVT);
2477 case ISD::SETUO: return getBoolConstant(R==APFloat::cmpUnordered, dl, VT,
2478 OpVT);
2479 case ISD::SETUEQ: return getBoolConstant(R==APFloat::cmpUnordered ||
2480 R==APFloat::cmpEqual, dl, VT,
2481 OpVT);
2482 case ISD::SETUNE: return getBoolConstant(R!=APFloat::cmpEqual, dl, VT,
2483 OpVT);
2484 case ISD::SETULT: return getBoolConstant(R==APFloat::cmpUnordered ||
2485 R==APFloat::cmpLessThan, dl, VT,
2486 OpVT);
2487 case ISD::SETUGT: return getBoolConstant(R==APFloat::cmpGreaterThan ||
2488 R==APFloat::cmpUnordered, dl, VT,
2489 OpVT);
2490 case ISD::SETULE: return getBoolConstant(R!=APFloat::cmpGreaterThan, dl,
2491 VT, OpVT);
2492 case ISD::SETUGE: return getBoolConstant(R!=APFloat::cmpLessThan, dl, VT,
2493 OpVT);
2494 }
2495 } else if (N1CFP && OpVT.isSimple() && !N2.isUndef()) {
2496 // Ensure that the constant occurs on the RHS.
2497 ISD::CondCode SwappedCond = ISD::getSetCCSwappedOperands(Cond);
2498 if (!TLI->isCondCodeLegal(SwappedCond, OpVT.getSimpleVT()))
2499 return SDValue();
2500 return getSetCC(dl, VT, N2, N1, SwappedCond);
2501 } else if ((N2CFP && N2CFP->getValueAPF().isNaN()) ||
2502 (OpVT.isFloatingPoint() && (N1.isUndef() || N2.isUndef()))) {
2503 // If an operand is known to be a nan (or undef that could be a nan), we can
2504 // fold it.
2505 // Choosing NaN for the undef will always make unordered comparison succeed
2506 // and ordered comparison fails.
2507 // Matches behavior in llvm::ConstantFoldCompareInstruction.
2508 switch (ISD::getUnorderedFlavor(Cond)) {
2509 default:
2510 llvm_unreachable("Unknown flavor!");
2511 case 0: // Known false.
2512 return getBoolConstant(false, dl, VT, OpVT);
2513 case 1: // Known true.
2514 return getBoolConstant(true, dl, VT, OpVT);
2515 case 2: // Undefined.
2516 return getUNDEF(VT);
2517 }
2518 }
2519
2520 // Could not fold it.
2521 return SDValue();
2522 }
2523
2524 /// SignBitIsZero - Return true if the sign bit of Op is known to be zero. We
2525 /// use this predicate to simplify operations downstream.
SignBitIsZero(SDValue Op,unsigned Depth) const2526 bool SelectionDAG::SignBitIsZero(SDValue Op, unsigned Depth) const {
2527 unsigned BitWidth = Op.getScalarValueSizeInBits();
2528 return MaskedValueIsZero(Op, APInt::getSignMask(BitWidth), Depth);
2529 }
2530
2531 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
2532 /// this predicate to simplify operations downstream. Mask is known to be zero
2533 /// for bits that V cannot have.
MaskedValueIsZero(SDValue V,const APInt & Mask,unsigned Depth) const2534 bool SelectionDAG::MaskedValueIsZero(SDValue V, const APInt &Mask,
2535 unsigned Depth) const {
2536 return Mask.isSubsetOf(computeKnownBits(V, Depth).Zero);
2537 }
2538
2539 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero in
2540 /// DemandedElts. We use this predicate to simplify operations downstream.
2541 /// Mask is known to be zero for bits that V cannot have.
MaskedValueIsZero(SDValue V,const APInt & Mask,const APInt & DemandedElts,unsigned Depth) const2542 bool SelectionDAG::MaskedValueIsZero(SDValue V, const APInt &Mask,
2543 const APInt &DemandedElts,
2544 unsigned Depth) const {
2545 return Mask.isSubsetOf(computeKnownBits(V, DemandedElts, Depth).Zero);
2546 }
2547
2548 /// MaskedVectorIsZero - Return true if 'Op' is known to be zero in
2549 /// DemandedElts. We use this predicate to simplify operations downstream.
MaskedVectorIsZero(SDValue V,const APInt & DemandedElts,unsigned Depth) const2550 bool SelectionDAG::MaskedVectorIsZero(SDValue V, const APInt &DemandedElts,
2551 unsigned Depth /* = 0 */) const {
2552 return computeKnownBits(V, DemandedElts, Depth).isZero();
2553 }
2554
2555 /// MaskedValueIsAllOnes - Return true if '(Op & Mask) == Mask'.
MaskedValueIsAllOnes(SDValue V,const APInt & Mask,unsigned Depth) const2556 bool SelectionDAG::MaskedValueIsAllOnes(SDValue V, const APInt &Mask,
2557 unsigned Depth) const {
2558 return Mask.isSubsetOf(computeKnownBits(V, Depth).One);
2559 }
2560
computeVectorKnownZeroElements(SDValue Op,const APInt & DemandedElts,unsigned Depth) const2561 APInt SelectionDAG::computeVectorKnownZeroElements(SDValue Op,
2562 const APInt &DemandedElts,
2563 unsigned Depth) const {
2564 EVT VT = Op.getValueType();
2565 assert(VT.isVector() && !VT.isScalableVector() && "Only for fixed vectors!");
2566
2567 unsigned NumElts = VT.getVectorNumElements();
2568 assert(DemandedElts.getBitWidth() == NumElts && "Unexpected demanded mask.");
2569
2570 APInt KnownZeroElements = APInt::getNullValue(NumElts);
2571 for (unsigned EltIdx = 0; EltIdx != NumElts; ++EltIdx) {
2572 if (!DemandedElts[EltIdx])
2573 continue; // Don't query elements that are not demanded.
2574 APInt Mask = APInt::getOneBitSet(NumElts, EltIdx);
2575 if (MaskedVectorIsZero(Op, Mask, Depth))
2576 KnownZeroElements.setBit(EltIdx);
2577 }
2578 return KnownZeroElements;
2579 }
2580
2581 /// isSplatValue - Return true if the vector V has the same value
2582 /// across all DemandedElts. For scalable vectors, we don't know the
2583 /// number of lanes at compile time. Instead, we use a 1 bit APInt
2584 /// to represent a conservative value for all lanes; that is, that
2585 /// one bit value is implicitly splatted across all lanes.
isSplatValue(SDValue V,const APInt & DemandedElts,APInt & UndefElts,unsigned Depth) const2586 bool SelectionDAG::isSplatValue(SDValue V, const APInt &DemandedElts,
2587 APInt &UndefElts, unsigned Depth) const {
2588 unsigned Opcode = V.getOpcode();
2589 EVT VT = V.getValueType();
2590 assert(VT.isVector() && "Vector type expected");
2591 assert((!VT.isScalableVector() || DemandedElts.getBitWidth() == 1) &&
2592 "scalable demanded bits are ignored");
2593
2594 if (!DemandedElts)
2595 return false; // No demanded elts, better to assume we don't know anything.
2596
2597 if (Depth >= MaxRecursionDepth)
2598 return false; // Limit search depth.
2599
2600 // Deal with some common cases here that work for both fixed and scalable
2601 // vector types.
2602 switch (Opcode) {
2603 case ISD::SPLAT_VECTOR:
2604 UndefElts = V.getOperand(0).isUndef()
2605 ? APInt::getAllOnes(DemandedElts.getBitWidth())
2606 : APInt(DemandedElts.getBitWidth(), 0);
2607 return true;
2608 case ISD::ADD:
2609 case ISD::SUB:
2610 case ISD::AND:
2611 case ISD::XOR:
2612 case ISD::OR: {
2613 APInt UndefLHS, UndefRHS;
2614 SDValue LHS = V.getOperand(0);
2615 SDValue RHS = V.getOperand(1);
2616 if (isSplatValue(LHS, DemandedElts, UndefLHS, Depth + 1) &&
2617 isSplatValue(RHS, DemandedElts, UndefRHS, Depth + 1)) {
2618 UndefElts = UndefLHS | UndefRHS;
2619 return true;
2620 }
2621 return false;
2622 }
2623 case ISD::ABS:
2624 case ISD::TRUNCATE:
2625 case ISD::SIGN_EXTEND:
2626 case ISD::ZERO_EXTEND:
2627 return isSplatValue(V.getOperand(0), DemandedElts, UndefElts, Depth + 1);
2628 default:
2629 if (Opcode >= ISD::BUILTIN_OP_END || Opcode == ISD::INTRINSIC_WO_CHAIN ||
2630 Opcode == ISD::INTRINSIC_W_CHAIN || Opcode == ISD::INTRINSIC_VOID)
2631 return TLI->isSplatValueForTargetNode(V, DemandedElts, UndefElts, *this,
2632 Depth);
2633 break;
2634 }
2635
2636 // We don't support other cases than those above for scalable vectors at
2637 // the moment.
2638 if (VT.isScalableVector())
2639 return false;
2640
2641 unsigned NumElts = VT.getVectorNumElements();
2642 assert(NumElts == DemandedElts.getBitWidth() && "Vector size mismatch");
2643 UndefElts = APInt::getZero(NumElts);
2644
2645 switch (Opcode) {
2646 case ISD::BUILD_VECTOR: {
2647 SDValue Scl;
2648 for (unsigned i = 0; i != NumElts; ++i) {
2649 SDValue Op = V.getOperand(i);
2650 if (Op.isUndef()) {
2651 UndefElts.setBit(i);
2652 continue;
2653 }
2654 if (!DemandedElts[i])
2655 continue;
2656 if (Scl && Scl != Op)
2657 return false;
2658 Scl = Op;
2659 }
2660 return true;
2661 }
2662 case ISD::VECTOR_SHUFFLE: {
2663 // Check if this is a shuffle node doing a splat or a shuffle of a splat.
2664 APInt DemandedLHS = APInt::getNullValue(NumElts);
2665 APInt DemandedRHS = APInt::getNullValue(NumElts);
2666 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(V)->getMask();
2667 for (int i = 0; i != (int)NumElts; ++i) {
2668 int M = Mask[i];
2669 if (M < 0) {
2670 UndefElts.setBit(i);
2671 continue;
2672 }
2673 if (!DemandedElts[i])
2674 continue;
2675 if (M < (int)NumElts)
2676 DemandedLHS.setBit(M);
2677 else
2678 DemandedRHS.setBit(M - NumElts);
2679 }
2680
2681 // If we aren't demanding either op, assume there's no splat.
2682 // If we are demanding both ops, assume there's no splat.
2683 if ((DemandedLHS.isZero() && DemandedRHS.isZero()) ||
2684 (!DemandedLHS.isZero() && !DemandedRHS.isZero()))
2685 return false;
2686
2687 // See if the demanded elts of the source op is a splat or we only demand
2688 // one element, which should always be a splat.
2689 // TODO: Handle source ops splats with undefs.
2690 auto CheckSplatSrc = [&](SDValue Src, const APInt &SrcElts) {
2691 APInt SrcUndefs;
2692 return (SrcElts.countPopulation() == 1) ||
2693 (isSplatValue(Src, SrcElts, SrcUndefs, Depth + 1) &&
2694 (SrcElts & SrcUndefs).isZero());
2695 };
2696 if (!DemandedLHS.isZero())
2697 return CheckSplatSrc(V.getOperand(0), DemandedLHS);
2698 return CheckSplatSrc(V.getOperand(1), DemandedRHS);
2699 }
2700 case ISD::EXTRACT_SUBVECTOR: {
2701 // Offset the demanded elts by the subvector index.
2702 SDValue Src = V.getOperand(0);
2703 // We don't support scalable vectors at the moment.
2704 if (Src.getValueType().isScalableVector())
2705 return false;
2706 uint64_t Idx = V.getConstantOperandVal(1);
2707 unsigned NumSrcElts = Src.getValueType().getVectorNumElements();
2708 APInt UndefSrcElts;
2709 APInt DemandedSrcElts = DemandedElts.zext(NumSrcElts).shl(Idx);
2710 if (isSplatValue(Src, DemandedSrcElts, UndefSrcElts, Depth + 1)) {
2711 UndefElts = UndefSrcElts.extractBits(NumElts, Idx);
2712 return true;
2713 }
2714 break;
2715 }
2716 case ISD::ANY_EXTEND_VECTOR_INREG:
2717 case ISD::SIGN_EXTEND_VECTOR_INREG:
2718 case ISD::ZERO_EXTEND_VECTOR_INREG: {
2719 // Widen the demanded elts by the src element count.
2720 SDValue Src = V.getOperand(0);
2721 // We don't support scalable vectors at the moment.
2722 if (Src.getValueType().isScalableVector())
2723 return false;
2724 unsigned NumSrcElts = Src.getValueType().getVectorNumElements();
2725 APInt UndefSrcElts;
2726 APInt DemandedSrcElts = DemandedElts.zext(NumSrcElts);
2727 if (isSplatValue(Src, DemandedSrcElts, UndefSrcElts, Depth + 1)) {
2728 UndefElts = UndefSrcElts.trunc(NumElts);
2729 return true;
2730 }
2731 break;
2732 }
2733 case ISD::BITCAST: {
2734 SDValue Src = V.getOperand(0);
2735 EVT SrcVT = Src.getValueType();
2736 unsigned SrcBitWidth = SrcVT.getScalarSizeInBits();
2737 unsigned BitWidth = VT.getScalarSizeInBits();
2738
2739 // Ignore bitcasts from unsupported types.
2740 // TODO: Add fp support?
2741 if (!SrcVT.isVector() || !SrcVT.isInteger() || !VT.isInteger())
2742 break;
2743
2744 // Bitcast 'small element' vector to 'large element' vector.
2745 if ((BitWidth % SrcBitWidth) == 0) {
2746 // See if each sub element is a splat.
2747 unsigned Scale = BitWidth / SrcBitWidth;
2748 unsigned NumSrcElts = SrcVT.getVectorNumElements();
2749 APInt ScaledDemandedElts =
2750 APIntOps::ScaleBitMask(DemandedElts, NumSrcElts);
2751 for (unsigned I = 0; I != Scale; ++I) {
2752 APInt SubUndefElts;
2753 APInt SubDemandedElt = APInt::getOneBitSet(Scale, I);
2754 APInt SubDemandedElts = APInt::getSplat(NumSrcElts, SubDemandedElt);
2755 SubDemandedElts &= ScaledDemandedElts;
2756 if (!isSplatValue(Src, SubDemandedElts, SubUndefElts, Depth + 1))
2757 return false;
2758 // TODO: Add support for merging sub undef elements.
2759 if (!SubUndefElts.isZero())
2760 return false;
2761 }
2762 return true;
2763 }
2764 break;
2765 }
2766 }
2767
2768 return false;
2769 }
2770
2771 /// Helper wrapper to main isSplatValue function.
isSplatValue(SDValue V,bool AllowUndefs) const2772 bool SelectionDAG::isSplatValue(SDValue V, bool AllowUndefs) const {
2773 EVT VT = V.getValueType();
2774 assert(VT.isVector() && "Vector type expected");
2775
2776 APInt UndefElts;
2777 // Since the number of lanes in a scalable vector is unknown at compile time,
2778 // we track one bit which is implicitly broadcast to all lanes. This means
2779 // that all lanes in a scalable vector are considered demanded.
2780 APInt DemandedElts
2781 = APInt::getAllOnes(VT.isScalableVector() ? 1 : VT.getVectorNumElements());
2782 return isSplatValue(V, DemandedElts, UndefElts) &&
2783 (AllowUndefs || !UndefElts);
2784 }
2785
getSplatSourceVector(SDValue V,int & SplatIdx)2786 SDValue SelectionDAG::getSplatSourceVector(SDValue V, int &SplatIdx) {
2787 V = peekThroughExtractSubvectors(V);
2788
2789 EVT VT = V.getValueType();
2790 unsigned Opcode = V.getOpcode();
2791 switch (Opcode) {
2792 default: {
2793 APInt UndefElts;
2794 // Since the number of lanes in a scalable vector is unknown at compile time,
2795 // we track one bit which is implicitly broadcast to all lanes. This means
2796 // that all lanes in a scalable vector are considered demanded.
2797 APInt DemandedElts
2798 = APInt::getAllOnes(VT.isScalableVector() ? 1 : VT.getVectorNumElements());
2799
2800 if (isSplatValue(V, DemandedElts, UndefElts)) {
2801 if (VT.isScalableVector()) {
2802 // DemandedElts and UndefElts are ignored for scalable vectors, since
2803 // the only supported cases are SPLAT_VECTOR nodes.
2804 SplatIdx = 0;
2805 } else {
2806 // Handle case where all demanded elements are UNDEF.
2807 if (DemandedElts.isSubsetOf(UndefElts)) {
2808 SplatIdx = 0;
2809 return getUNDEF(VT);
2810 }
2811 SplatIdx = (UndefElts & DemandedElts).countTrailingOnes();
2812 }
2813 return V;
2814 }
2815 break;
2816 }
2817 case ISD::SPLAT_VECTOR:
2818 SplatIdx = 0;
2819 return V;
2820 case ISD::VECTOR_SHUFFLE: {
2821 assert(!VT.isScalableVector());
2822 // Check if this is a shuffle node doing a splat.
2823 // TODO - remove this and rely purely on SelectionDAG::isSplatValue,
2824 // getTargetVShiftNode currently struggles without the splat source.
2825 auto *SVN = cast<ShuffleVectorSDNode>(V);
2826 if (!SVN->isSplat())
2827 break;
2828 int Idx = SVN->getSplatIndex();
2829 int NumElts = V.getValueType().getVectorNumElements();
2830 SplatIdx = Idx % NumElts;
2831 return V.getOperand(Idx / NumElts);
2832 }
2833 }
2834
2835 return SDValue();
2836 }
2837
getSplatValue(SDValue V,bool LegalTypes)2838 SDValue SelectionDAG::getSplatValue(SDValue V, bool LegalTypes) {
2839 int SplatIdx;
2840 if (SDValue SrcVector = getSplatSourceVector(V, SplatIdx)) {
2841 EVT SVT = SrcVector.getValueType().getScalarType();
2842 EVT LegalSVT = SVT;
2843 if (LegalTypes && !TLI->isTypeLegal(SVT)) {
2844 if (!SVT.isInteger())
2845 return SDValue();
2846 LegalSVT = TLI->getTypeToTransformTo(*getContext(), LegalSVT);
2847 if (LegalSVT.bitsLT(SVT))
2848 return SDValue();
2849 }
2850 return getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(V), LegalSVT, SrcVector,
2851 getVectorIdxConstant(SplatIdx, SDLoc(V)));
2852 }
2853 return SDValue();
2854 }
2855
2856 const APInt *
getValidShiftAmountConstant(SDValue V,const APInt & DemandedElts) const2857 SelectionDAG::getValidShiftAmountConstant(SDValue V,
2858 const APInt &DemandedElts) const {
2859 assert((V.getOpcode() == ISD::SHL || V.getOpcode() == ISD::SRL ||
2860 V.getOpcode() == ISD::SRA) &&
2861 "Unknown shift node");
2862 unsigned BitWidth = V.getScalarValueSizeInBits();
2863 if (ConstantSDNode *SA = isConstOrConstSplat(V.getOperand(1), DemandedElts)) {
2864 // Shifting more than the bitwidth is not valid.
2865 const APInt &ShAmt = SA->getAPIntValue();
2866 if (ShAmt.ult(BitWidth))
2867 return &ShAmt;
2868 }
2869 return nullptr;
2870 }
2871
getValidMinimumShiftAmountConstant(SDValue V,const APInt & DemandedElts) const2872 const APInt *SelectionDAG::getValidMinimumShiftAmountConstant(
2873 SDValue V, const APInt &DemandedElts) const {
2874 assert((V.getOpcode() == ISD::SHL || V.getOpcode() == ISD::SRL ||
2875 V.getOpcode() == ISD::SRA) &&
2876 "Unknown shift node");
2877 if (const APInt *ValidAmt = getValidShiftAmountConstant(V, DemandedElts))
2878 return ValidAmt;
2879 unsigned BitWidth = V.getScalarValueSizeInBits();
2880 auto *BV = dyn_cast<BuildVectorSDNode>(V.getOperand(1));
2881 if (!BV)
2882 return nullptr;
2883 const APInt *MinShAmt = nullptr;
2884 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
2885 if (!DemandedElts[i])
2886 continue;
2887 auto *SA = dyn_cast<ConstantSDNode>(BV->getOperand(i));
2888 if (!SA)
2889 return nullptr;
2890 // Shifting more than the bitwidth is not valid.
2891 const APInt &ShAmt = SA->getAPIntValue();
2892 if (ShAmt.uge(BitWidth))
2893 return nullptr;
2894 if (MinShAmt && MinShAmt->ule(ShAmt))
2895 continue;
2896 MinShAmt = &ShAmt;
2897 }
2898 return MinShAmt;
2899 }
2900
getValidMaximumShiftAmountConstant(SDValue V,const APInt & DemandedElts) const2901 const APInt *SelectionDAG::getValidMaximumShiftAmountConstant(
2902 SDValue V, const APInt &DemandedElts) const {
2903 assert((V.getOpcode() == ISD::SHL || V.getOpcode() == ISD::SRL ||
2904 V.getOpcode() == ISD::SRA) &&
2905 "Unknown shift node");
2906 if (const APInt *ValidAmt = getValidShiftAmountConstant(V, DemandedElts))
2907 return ValidAmt;
2908 unsigned BitWidth = V.getScalarValueSizeInBits();
2909 auto *BV = dyn_cast<BuildVectorSDNode>(V.getOperand(1));
2910 if (!BV)
2911 return nullptr;
2912 const APInt *MaxShAmt = nullptr;
2913 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
2914 if (!DemandedElts[i])
2915 continue;
2916 auto *SA = dyn_cast<ConstantSDNode>(BV->getOperand(i));
2917 if (!SA)
2918 return nullptr;
2919 // Shifting more than the bitwidth is not valid.
2920 const APInt &ShAmt = SA->getAPIntValue();
2921 if (ShAmt.uge(BitWidth))
2922 return nullptr;
2923 if (MaxShAmt && MaxShAmt->uge(ShAmt))
2924 continue;
2925 MaxShAmt = &ShAmt;
2926 }
2927 return MaxShAmt;
2928 }
2929
2930 /// Determine which bits of Op are known to be either zero or one and return
2931 /// them in Known. For vectors, the known bits are those that are shared by
2932 /// every vector element.
computeKnownBits(SDValue Op,unsigned Depth) const2933 KnownBits SelectionDAG::computeKnownBits(SDValue Op, unsigned Depth) const {
2934 EVT VT = Op.getValueType();
2935
2936 // Since the number of lanes in a scalable vector is unknown at compile time,
2937 // we track one bit which is implicitly broadcast to all lanes. This means
2938 // that all lanes in a scalable vector are considered demanded.
2939 APInt DemandedElts = VT.isFixedLengthVector()
2940 ? APInt::getAllOnes(VT.getVectorNumElements())
2941 : APInt(1, 1);
2942 return computeKnownBits(Op, DemandedElts, Depth);
2943 }
2944
2945 /// Determine which bits of Op are known to be either zero or one and return
2946 /// them in Known. The DemandedElts argument allows us to only collect the known
2947 /// bits that are shared by the requested vector elements.
computeKnownBits(SDValue Op,const APInt & DemandedElts,unsigned Depth) const2948 KnownBits SelectionDAG::computeKnownBits(SDValue Op, const APInt &DemandedElts,
2949 unsigned Depth) const {
2950 unsigned BitWidth = Op.getScalarValueSizeInBits();
2951
2952 KnownBits Known(BitWidth); // Don't know anything.
2953
2954 if (auto *C = dyn_cast<ConstantSDNode>(Op)) {
2955 // We know all of the bits for a constant!
2956 return KnownBits::makeConstant(C->getAPIntValue());
2957 }
2958 if (auto *C = dyn_cast<ConstantFPSDNode>(Op)) {
2959 // We know all of the bits for a constant fp!
2960 return KnownBits::makeConstant(C->getValueAPF().bitcastToAPInt());
2961 }
2962
2963 if (Depth >= MaxRecursionDepth)
2964 return Known; // Limit search depth.
2965
2966 KnownBits Known2;
2967 unsigned NumElts = DemandedElts.getBitWidth();
2968 assert((!Op.getValueType().isFixedLengthVector() ||
2969 NumElts == Op.getValueType().getVectorNumElements()) &&
2970 "Unexpected vector size");
2971
2972 if (!DemandedElts)
2973 return Known; // No demanded elts, better to assume we don't know anything.
2974
2975 unsigned Opcode = Op.getOpcode();
2976 switch (Opcode) {
2977 case ISD::MERGE_VALUES:
2978 return computeKnownBits(Op.getOperand(Op.getResNo()), DemandedElts,
2979 Depth + 1);
2980 case ISD::SPLAT_VECTOR: {
2981 SDValue SrcOp = Op.getOperand(0);
2982 assert(SrcOp.getValueSizeInBits() >= BitWidth &&
2983 "Expected SPLAT_VECTOR implicit truncation");
2984 // Implicitly truncate the bits to match the official semantics of
2985 // SPLAT_VECTOR.
2986 Known = computeKnownBits(SrcOp, Depth + 1).trunc(BitWidth);
2987 break;
2988 }
2989 case ISD::BUILD_VECTOR:
2990 assert(!Op.getValueType().isScalableVector());
2991 // Collect the known bits that are shared by every demanded vector element.
2992 Known.Zero.setAllBits(); Known.One.setAllBits();
2993 for (unsigned i = 0, e = Op.getNumOperands(); i != e; ++i) {
2994 if (!DemandedElts[i])
2995 continue;
2996
2997 SDValue SrcOp = Op.getOperand(i);
2998 Known2 = computeKnownBits(SrcOp, Depth + 1);
2999
3000 // BUILD_VECTOR can implicitly truncate sources, we must handle this.
3001 if (SrcOp.getValueSizeInBits() != BitWidth) {
3002 assert(SrcOp.getValueSizeInBits() > BitWidth &&
3003 "Expected BUILD_VECTOR implicit truncation");
3004 Known2 = Known2.trunc(BitWidth);
3005 }
3006
3007 // Known bits are the values that are shared by every demanded element.
3008 Known = KnownBits::commonBits(Known, Known2);
3009
3010 // If we don't know any bits, early out.
3011 if (Known.isUnknown())
3012 break;
3013 }
3014 break;
3015 case ISD::VECTOR_SHUFFLE: {
3016 assert(!Op.getValueType().isScalableVector());
3017 // Collect the known bits that are shared by every vector element referenced
3018 // by the shuffle.
3019 APInt DemandedLHS, DemandedRHS;
3020 const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op);
3021 assert(NumElts == SVN->getMask().size() && "Unexpected vector size");
3022 if (!getShuffleDemandedElts(NumElts, SVN->getMask(), DemandedElts,
3023 DemandedLHS, DemandedRHS))
3024 break;
3025
3026 // Known bits are the values that are shared by every demanded element.
3027 Known.Zero.setAllBits(); Known.One.setAllBits();
3028 if (!!DemandedLHS) {
3029 SDValue LHS = Op.getOperand(0);
3030 Known2 = computeKnownBits(LHS, DemandedLHS, Depth + 1);
3031 Known = KnownBits::commonBits(Known, Known2);
3032 }
3033 // If we don't know any bits, early out.
3034 if (Known.isUnknown())
3035 break;
3036 if (!!DemandedRHS) {
3037 SDValue RHS = Op.getOperand(1);
3038 Known2 = computeKnownBits(RHS, DemandedRHS, Depth + 1);
3039 Known = KnownBits::commonBits(Known, Known2);
3040 }
3041 break;
3042 }
3043 case ISD::CONCAT_VECTORS: {
3044 if (Op.getValueType().isScalableVector())
3045 break;
3046 // Split DemandedElts and test each of the demanded subvectors.
3047 Known.Zero.setAllBits(); Known.One.setAllBits();
3048 EVT SubVectorVT = Op.getOperand(0).getValueType();
3049 unsigned NumSubVectorElts = SubVectorVT.getVectorNumElements();
3050 unsigned NumSubVectors = Op.getNumOperands();
3051 for (unsigned i = 0; i != NumSubVectors; ++i) {
3052 APInt DemandedSub =
3053 DemandedElts.extractBits(NumSubVectorElts, i * NumSubVectorElts);
3054 if (!!DemandedSub) {
3055 SDValue Sub = Op.getOperand(i);
3056 Known2 = computeKnownBits(Sub, DemandedSub, Depth + 1);
3057 Known = KnownBits::commonBits(Known, Known2);
3058 }
3059 // If we don't know any bits, early out.
3060 if (Known.isUnknown())
3061 break;
3062 }
3063 break;
3064 }
3065 case ISD::INSERT_SUBVECTOR: {
3066 if (Op.getValueType().isScalableVector())
3067 break;
3068 // Demand any elements from the subvector and the remainder from the src its
3069 // inserted into.
3070 SDValue Src = Op.getOperand(0);
3071 SDValue Sub = Op.getOperand(1);
3072 uint64_t Idx = Op.getConstantOperandVal(2);
3073 unsigned NumSubElts = Sub.getValueType().getVectorNumElements();
3074 APInt DemandedSubElts = DemandedElts.extractBits(NumSubElts, Idx);
3075 APInt DemandedSrcElts = DemandedElts;
3076 DemandedSrcElts.insertBits(APInt::getZero(NumSubElts), Idx);
3077
3078 Known.One.setAllBits();
3079 Known.Zero.setAllBits();
3080 if (!!DemandedSubElts) {
3081 Known = computeKnownBits(Sub, DemandedSubElts, Depth + 1);
3082 if (Known.isUnknown())
3083 break; // early-out.
3084 }
3085 if (!!DemandedSrcElts) {
3086 Known2 = computeKnownBits(Src, DemandedSrcElts, Depth + 1);
3087 Known = KnownBits::commonBits(Known, Known2);
3088 }
3089 break;
3090 }
3091 case ISD::EXTRACT_SUBVECTOR: {
3092 // Offset the demanded elts by the subvector index.
3093 SDValue Src = Op.getOperand(0);
3094 // Bail until we can represent demanded elements for scalable vectors.
3095 if (Op.getValueType().isScalableVector() || Src.getValueType().isScalableVector())
3096 break;
3097 uint64_t Idx = Op.getConstantOperandVal(1);
3098 unsigned NumSrcElts = Src.getValueType().getVectorNumElements();
3099 APInt DemandedSrcElts = DemandedElts.zext(NumSrcElts).shl(Idx);
3100 Known = computeKnownBits(Src, DemandedSrcElts, Depth + 1);
3101 break;
3102 }
3103 case ISD::SCALAR_TO_VECTOR: {
3104 if (Op.getValueType().isScalableVector())
3105 break;
3106 // We know about scalar_to_vector as much as we know about it source,
3107 // which becomes the first element of otherwise unknown vector.
3108 if (DemandedElts != 1)
3109 break;
3110
3111 SDValue N0 = Op.getOperand(0);
3112 Known = computeKnownBits(N0, Depth + 1);
3113 if (N0.getValueSizeInBits() != BitWidth)
3114 Known = Known.trunc(BitWidth);
3115
3116 break;
3117 }
3118 case ISD::BITCAST: {
3119 if (Op.getValueType().isScalableVector())
3120 break;
3121
3122 SDValue N0 = Op.getOperand(0);
3123 EVT SubVT = N0.getValueType();
3124 unsigned SubBitWidth = SubVT.getScalarSizeInBits();
3125
3126 // Ignore bitcasts from unsupported types.
3127 if (!(SubVT.isInteger() || SubVT.isFloatingPoint()))
3128 break;
3129
3130 // Fast handling of 'identity' bitcasts.
3131 if (BitWidth == SubBitWidth) {
3132 Known = computeKnownBits(N0, DemandedElts, Depth + 1);
3133 break;
3134 }
3135
3136 bool IsLE = getDataLayout().isLittleEndian();
3137
3138 // Bitcast 'small element' vector to 'large element' scalar/vector.
3139 if ((BitWidth % SubBitWidth) == 0) {
3140 assert(N0.getValueType().isVector() && "Expected bitcast from vector");
3141
3142 // Collect known bits for the (larger) output by collecting the known
3143 // bits from each set of sub elements and shift these into place.
3144 // We need to separately call computeKnownBits for each set of
3145 // sub elements as the knownbits for each is likely to be different.
3146 unsigned SubScale = BitWidth / SubBitWidth;
3147 APInt SubDemandedElts(NumElts * SubScale, 0);
3148 for (unsigned i = 0; i != NumElts; ++i)
3149 if (DemandedElts[i])
3150 SubDemandedElts.setBit(i * SubScale);
3151
3152 for (unsigned i = 0; i != SubScale; ++i) {
3153 Known2 = computeKnownBits(N0, SubDemandedElts.shl(i),
3154 Depth + 1);
3155 unsigned Shifts = IsLE ? i : SubScale - 1 - i;
3156 Known.insertBits(Known2, SubBitWidth * Shifts);
3157 }
3158 }
3159
3160 // Bitcast 'large element' scalar/vector to 'small element' vector.
3161 if ((SubBitWidth % BitWidth) == 0) {
3162 assert(Op.getValueType().isVector() && "Expected bitcast to vector");
3163
3164 // Collect known bits for the (smaller) output by collecting the known
3165 // bits from the overlapping larger input elements and extracting the
3166 // sub sections we actually care about.
3167 unsigned SubScale = SubBitWidth / BitWidth;
3168 APInt SubDemandedElts =
3169 APIntOps::ScaleBitMask(DemandedElts, NumElts / SubScale);
3170 Known2 = computeKnownBits(N0, SubDemandedElts, Depth + 1);
3171
3172 Known.Zero.setAllBits(); Known.One.setAllBits();
3173 for (unsigned i = 0; i != NumElts; ++i)
3174 if (DemandedElts[i]) {
3175 unsigned Shifts = IsLE ? i : NumElts - 1 - i;
3176 unsigned Offset = (Shifts % SubScale) * BitWidth;
3177 Known = KnownBits::commonBits(Known,
3178 Known2.extractBits(BitWidth, Offset));
3179 // If we don't know any bits, early out.
3180 if (Known.isUnknown())
3181 break;
3182 }
3183 }
3184 break;
3185 }
3186 case ISD::AND:
3187 Known = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3188 Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3189
3190 Known &= Known2;
3191 break;
3192 case ISD::OR:
3193 Known = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3194 Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3195
3196 Known |= Known2;
3197 break;
3198 case ISD::XOR:
3199 Known = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3200 Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3201
3202 Known ^= Known2;
3203 break;
3204 case ISD::MUL: {
3205 Known = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3206 Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3207 bool SelfMultiply = Op.getOperand(0) == Op.getOperand(1);
3208 // TODO: SelfMultiply can be poison, but not undef.
3209 if (SelfMultiply)
3210 SelfMultiply &= isGuaranteedNotToBeUndefOrPoison(
3211 Op.getOperand(0), DemandedElts, false, Depth + 1);
3212 Known = KnownBits::mul(Known, Known2, SelfMultiply);
3213
3214 // If the multiplication is known not to overflow, the product of a number
3215 // with itself is non-negative. Only do this if we didn't already computed
3216 // the opposite value for the sign bit.
3217 if (Op->getFlags().hasNoSignedWrap() &&
3218 Op.getOperand(0) == Op.getOperand(1) &&
3219 !Known.isNegative())
3220 Known.makeNonNegative();
3221 break;
3222 }
3223 case ISD::MULHU: {
3224 Known = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3225 Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3226 Known = KnownBits::mulhu(Known, Known2);
3227 break;
3228 }
3229 case ISD::MULHS: {
3230 Known = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3231 Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3232 Known = KnownBits::mulhs(Known, Known2);
3233 break;
3234 }
3235 case ISD::UMUL_LOHI: {
3236 assert((Op.getResNo() == 0 || Op.getResNo() == 1) && "Unknown result");
3237 Known = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3238 Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3239 bool SelfMultiply = Op.getOperand(0) == Op.getOperand(1);
3240 if (Op.getResNo() == 0)
3241 Known = KnownBits::mul(Known, Known2, SelfMultiply);
3242 else
3243 Known = KnownBits::mulhu(Known, Known2);
3244 break;
3245 }
3246 case ISD::SMUL_LOHI: {
3247 assert((Op.getResNo() == 0 || Op.getResNo() == 1) && "Unknown result");
3248 Known = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3249 Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3250 bool SelfMultiply = Op.getOperand(0) == Op.getOperand(1);
3251 if (Op.getResNo() == 0)
3252 Known = KnownBits::mul(Known, Known2, SelfMultiply);
3253 else
3254 Known = KnownBits::mulhs(Known, Known2);
3255 break;
3256 }
3257 case ISD::AVGCEILU: {
3258 Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3259 Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3260 Known = Known.zext(BitWidth + 1);
3261 Known2 = Known2.zext(BitWidth + 1);
3262 KnownBits One = KnownBits::makeConstant(APInt(1, 1));
3263 Known = KnownBits::computeForAddCarry(Known, Known2, One);
3264 Known = Known.extractBits(BitWidth, 1);
3265 break;
3266 }
3267 case ISD::SELECT:
3268 case ISD::VSELECT:
3269 Known = computeKnownBits(Op.getOperand(2), DemandedElts, Depth+1);
3270 // If we don't know any bits, early out.
3271 if (Known.isUnknown())
3272 break;
3273 Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth+1);
3274
3275 // Only known if known in both the LHS and RHS.
3276 Known = KnownBits::commonBits(Known, Known2);
3277 break;
3278 case ISD::SELECT_CC:
3279 Known = computeKnownBits(Op.getOperand(3), DemandedElts, Depth+1);
3280 // If we don't know any bits, early out.
3281 if (Known.isUnknown())
3282 break;
3283 Known2 = computeKnownBits(Op.getOperand(2), DemandedElts, Depth+1);
3284
3285 // Only known if known in both the LHS and RHS.
3286 Known = KnownBits::commonBits(Known, Known2);
3287 break;
3288 case ISD::SMULO:
3289 case ISD::UMULO:
3290 if (Op.getResNo() != 1)
3291 break;
3292 // The boolean result conforms to getBooleanContents.
3293 // If we know the result of a setcc has the top bits zero, use this info.
3294 // We know that we have an integer-based boolean since these operations
3295 // are only available for integer.
3296 if (TLI->getBooleanContents(Op.getValueType().isVector(), false) ==
3297 TargetLowering::ZeroOrOneBooleanContent &&
3298 BitWidth > 1)
3299 Known.Zero.setBitsFrom(1);
3300 break;
3301 case ISD::SETCC:
3302 case ISD::SETCCCARRY:
3303 case ISD::STRICT_FSETCC:
3304 case ISD::STRICT_FSETCCS: {
3305 unsigned OpNo = Op->isStrictFPOpcode() ? 1 : 0;
3306 // If we know the result of a setcc has the top bits zero, use this info.
3307 if (TLI->getBooleanContents(Op.getOperand(OpNo).getValueType()) ==
3308 TargetLowering::ZeroOrOneBooleanContent &&
3309 BitWidth > 1)
3310 Known.Zero.setBitsFrom(1);
3311 break;
3312 }
3313 case ISD::SHL:
3314 Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3315 Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3316 Known = KnownBits::shl(Known, Known2);
3317
3318 // Minimum shift low bits are known zero.
3319 if (const APInt *ShMinAmt =
3320 getValidMinimumShiftAmountConstant(Op, DemandedElts))
3321 Known.Zero.setLowBits(ShMinAmt->getZExtValue());
3322 break;
3323 case ISD::SRL:
3324 Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3325 Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3326 Known = KnownBits::lshr(Known, Known2);
3327
3328 // Minimum shift high bits are known zero.
3329 if (const APInt *ShMinAmt =
3330 getValidMinimumShiftAmountConstant(Op, DemandedElts))
3331 Known.Zero.setHighBits(ShMinAmt->getZExtValue());
3332 break;
3333 case ISD::SRA:
3334 Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3335 Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3336 Known = KnownBits::ashr(Known, Known2);
3337 // TODO: Add minimum shift high known sign bits.
3338 break;
3339 case ISD::FSHL:
3340 case ISD::FSHR:
3341 if (ConstantSDNode *C = isConstOrConstSplat(Op.getOperand(2), DemandedElts)) {
3342 unsigned Amt = C->getAPIntValue().urem(BitWidth);
3343
3344 // For fshl, 0-shift returns the 1st arg.
3345 // For fshr, 0-shift returns the 2nd arg.
3346 if (Amt == 0) {
3347 Known = computeKnownBits(Op.getOperand(Opcode == ISD::FSHL ? 0 : 1),
3348 DemandedElts, Depth + 1);
3349 break;
3350 }
3351
3352 // fshl: (X << (Z % BW)) | (Y >> (BW - (Z % BW)))
3353 // fshr: (X << (BW - (Z % BW))) | (Y >> (Z % BW))
3354 Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3355 Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3356 if (Opcode == ISD::FSHL) {
3357 Known.One <<= Amt;
3358 Known.Zero <<= Amt;
3359 Known2.One.lshrInPlace(BitWidth - Amt);
3360 Known2.Zero.lshrInPlace(BitWidth - Amt);
3361 } else {
3362 Known.One <<= BitWidth - Amt;
3363 Known.Zero <<= BitWidth - Amt;
3364 Known2.One.lshrInPlace(Amt);
3365 Known2.Zero.lshrInPlace(Amt);
3366 }
3367 Known.One |= Known2.One;
3368 Known.Zero |= Known2.Zero;
3369 }
3370 break;
3371 case ISD::SHL_PARTS:
3372 case ISD::SRA_PARTS:
3373 case ISD::SRL_PARTS: {
3374 assert((Op.getResNo() == 0 || Op.getResNo() == 1) && "Unknown result");
3375
3376 // Collect lo/hi source values and concatenate.
3377 unsigned LoBits = Op.getOperand(0).getScalarValueSizeInBits();
3378 unsigned HiBits = Op.getOperand(1).getScalarValueSizeInBits();
3379 Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3380 Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3381 Known = Known2.concat(Known);
3382
3383 // Collect shift amount.
3384 Known2 = computeKnownBits(Op.getOperand(2), DemandedElts, Depth + 1);
3385
3386 if (Opcode == ISD::SHL_PARTS)
3387 Known = KnownBits::shl(Known, Known2);
3388 else if (Opcode == ISD::SRA_PARTS)
3389 Known = KnownBits::ashr(Known, Known2);
3390 else // if (Opcode == ISD::SRL_PARTS)
3391 Known = KnownBits::lshr(Known, Known2);
3392
3393 // TODO: Minimum shift low/high bits are known zero.
3394
3395 if (Op.getResNo() == 0)
3396 Known = Known.extractBits(LoBits, 0);
3397 else
3398 Known = Known.extractBits(HiBits, LoBits);
3399 break;
3400 }
3401 case ISD::SIGN_EXTEND_INREG: {
3402 Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3403 EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
3404 Known = Known.sextInReg(EVT.getScalarSizeInBits());
3405 break;
3406 }
3407 case ISD::CTTZ:
3408 case ISD::CTTZ_ZERO_UNDEF: {
3409 Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3410 // If we have a known 1, its position is our upper bound.
3411 unsigned PossibleTZ = Known2.countMaxTrailingZeros();
3412 unsigned LowBits = llvm::bit_width(PossibleTZ);
3413 Known.Zero.setBitsFrom(LowBits);
3414 break;
3415 }
3416 case ISD::CTLZ:
3417 case ISD::CTLZ_ZERO_UNDEF: {
3418 Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3419 // If we have a known 1, its position is our upper bound.
3420 unsigned PossibleLZ = Known2.countMaxLeadingZeros();
3421 unsigned LowBits = llvm::bit_width(PossibleLZ);
3422 Known.Zero.setBitsFrom(LowBits);
3423 break;
3424 }
3425 case ISD::CTPOP: {
3426 Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3427 // If we know some of the bits are zero, they can't be one.
3428 unsigned PossibleOnes = Known2.countMaxPopulation();
3429 Known.Zero.setBitsFrom(llvm::bit_width(PossibleOnes));
3430 break;
3431 }
3432 case ISD::PARITY: {
3433 // Parity returns 0 everywhere but the LSB.
3434 Known.Zero.setBitsFrom(1);
3435 break;
3436 }
3437 case ISD::LOAD: {
3438 LoadSDNode *LD = cast<LoadSDNode>(Op);
3439 const Constant *Cst = TLI->getTargetConstantFromLoad(LD);
3440 if (ISD::isNON_EXTLoad(LD) && Cst) {
3441 // Determine any common known bits from the loaded constant pool value.
3442 Type *CstTy = Cst->getType();
3443 if ((NumElts * BitWidth) == CstTy->getPrimitiveSizeInBits() &&
3444 !Op.getValueType().isScalableVector()) {
3445 // If its a vector splat, then we can (quickly) reuse the scalar path.
3446 // NOTE: We assume all elements match and none are UNDEF.
3447 if (CstTy->isVectorTy()) {
3448 if (const Constant *Splat = Cst->getSplatValue()) {
3449 Cst = Splat;
3450 CstTy = Cst->getType();
3451 }
3452 }
3453 // TODO - do we need to handle different bitwidths?
3454 if (CstTy->isVectorTy() && BitWidth == CstTy->getScalarSizeInBits()) {
3455 // Iterate across all vector elements finding common known bits.
3456 Known.One.setAllBits();
3457 Known.Zero.setAllBits();
3458 for (unsigned i = 0; i != NumElts; ++i) {
3459 if (!DemandedElts[i])
3460 continue;
3461 if (Constant *Elt = Cst->getAggregateElement(i)) {
3462 if (auto *CInt = dyn_cast<ConstantInt>(Elt)) {
3463 const APInt &Value = CInt->getValue();
3464 Known.One &= Value;
3465 Known.Zero &= ~Value;
3466 continue;
3467 }
3468 if (auto *CFP = dyn_cast<ConstantFP>(Elt)) {
3469 APInt Value = CFP->getValueAPF().bitcastToAPInt();
3470 Known.One &= Value;
3471 Known.Zero &= ~Value;
3472 continue;
3473 }
3474 }
3475 Known.One.clearAllBits();
3476 Known.Zero.clearAllBits();
3477 break;
3478 }
3479 } else if (BitWidth == CstTy->getPrimitiveSizeInBits()) {
3480 if (auto *CInt = dyn_cast<ConstantInt>(Cst)) {
3481 Known = KnownBits::makeConstant(CInt->getValue());
3482 } else if (auto *CFP = dyn_cast<ConstantFP>(Cst)) {
3483 Known =
3484 KnownBits::makeConstant(CFP->getValueAPF().bitcastToAPInt());
3485 }
3486 }
3487 }
3488 } else if (ISD::isZEXTLoad(Op.getNode()) && Op.getResNo() == 0) {
3489 // If this is a ZEXTLoad and we are looking at the loaded value.
3490 EVT VT = LD->getMemoryVT();
3491 unsigned MemBits = VT.getScalarSizeInBits();
3492 Known.Zero.setBitsFrom(MemBits);
3493 } else if (const MDNode *Ranges = LD->getRanges()) {
3494 EVT VT = LD->getValueType(0);
3495
3496 // TODO: Handle for extending loads
3497 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
3498 if (VT.isVector()) {
3499 // Handle truncation to the first demanded element.
3500 // TODO: Figure out which demanded elements are covered
3501 if (DemandedElts != 1 || !getDataLayout().isLittleEndian())
3502 break;
3503
3504 // Handle the case where a load has a vector type, but scalar memory
3505 // with an attached range.
3506 EVT MemVT = LD->getMemoryVT();
3507 KnownBits KnownFull(MemVT.getSizeInBits());
3508
3509 computeKnownBitsFromRangeMetadata(*Ranges, KnownFull);
3510 Known = KnownFull.trunc(BitWidth);
3511 } else
3512 computeKnownBitsFromRangeMetadata(*Ranges, Known);
3513 }
3514 }
3515 break;
3516 }
3517 case ISD::ZERO_EXTEND_VECTOR_INREG: {
3518 if (Op.getValueType().isScalableVector())
3519 break;
3520 EVT InVT = Op.getOperand(0).getValueType();
3521 APInt InDemandedElts = DemandedElts.zext(InVT.getVectorNumElements());
3522 Known = computeKnownBits(Op.getOperand(0), InDemandedElts, Depth + 1);
3523 Known = Known.zext(BitWidth);
3524 break;
3525 }
3526 case ISD::ZERO_EXTEND: {
3527 Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3528 Known = Known.zext(BitWidth);
3529 break;
3530 }
3531 case ISD::SIGN_EXTEND_VECTOR_INREG: {
3532 if (Op.getValueType().isScalableVector())
3533 break;
3534 EVT InVT = Op.getOperand(0).getValueType();
3535 APInt InDemandedElts = DemandedElts.zext(InVT.getVectorNumElements());
3536 Known = computeKnownBits(Op.getOperand(0), InDemandedElts, Depth + 1);
3537 // If the sign bit is known to be zero or one, then sext will extend
3538 // it to the top bits, else it will just zext.
3539 Known = Known.sext(BitWidth);
3540 break;
3541 }
3542 case ISD::SIGN_EXTEND: {
3543 Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3544 // If the sign bit is known to be zero or one, then sext will extend
3545 // it to the top bits, else it will just zext.
3546 Known = Known.sext(BitWidth);
3547 break;
3548 }
3549 case ISD::ANY_EXTEND_VECTOR_INREG: {
3550 if (Op.getValueType().isScalableVector())
3551 break;
3552 EVT InVT = Op.getOperand(0).getValueType();
3553 APInt InDemandedElts = DemandedElts.zext(InVT.getVectorNumElements());
3554 Known = computeKnownBits(Op.getOperand(0), InDemandedElts, Depth + 1);
3555 Known = Known.anyext(BitWidth);
3556 break;
3557 }
3558 case ISD::ANY_EXTEND: {
3559 Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3560 Known = Known.anyext(BitWidth);
3561 break;
3562 }
3563 case ISD::TRUNCATE: {
3564 Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3565 Known = Known.trunc(BitWidth);
3566 break;
3567 }
3568 case ISD::AssertZext: {
3569 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
3570 APInt InMask = APInt::getLowBitsSet(BitWidth, VT.getSizeInBits());
3571 Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3572 Known.Zero |= (~InMask);
3573 Known.One &= (~Known.Zero);
3574 break;
3575 }
3576 case ISD::AssertAlign: {
3577 unsigned LogOfAlign = Log2(cast<AssertAlignSDNode>(Op)->getAlign());
3578 assert(LogOfAlign != 0);
3579
3580 // TODO: Should use maximum with source
3581 // If a node is guaranteed to be aligned, set low zero bits accordingly as
3582 // well as clearing one bits.
3583 Known.Zero.setLowBits(LogOfAlign);
3584 Known.One.clearLowBits(LogOfAlign);
3585 break;
3586 }
3587 case ISD::FGETSIGN:
3588 // All bits are zero except the low bit.
3589 Known.Zero.setBitsFrom(1);
3590 break;
3591 case ISD::USUBO:
3592 case ISD::SSUBO:
3593 case ISD::SUBCARRY:
3594 case ISD::SSUBO_CARRY:
3595 if (Op.getResNo() == 1) {
3596 // If we know the result of a setcc has the top bits zero, use this info.
3597 if (TLI->getBooleanContents(Op.getOperand(0).getValueType()) ==
3598 TargetLowering::ZeroOrOneBooleanContent &&
3599 BitWidth > 1)
3600 Known.Zero.setBitsFrom(1);
3601 break;
3602 }
3603 [[fallthrough]];
3604 case ISD::SUB:
3605 case ISD::SUBC: {
3606 assert(Op.getResNo() == 0 &&
3607 "We only compute knownbits for the difference here.");
3608
3609 // TODO: Compute influence of the carry operand.
3610 if (Opcode == ISD::SUBCARRY || Opcode == ISD::SSUBO_CARRY)
3611 break;
3612
3613 Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3614 Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3615 Known = KnownBits::computeForAddSub(/* Add */ false, /* NSW */ false,
3616 Known, Known2);
3617 break;
3618 }
3619 case ISD::UADDO:
3620 case ISD::SADDO:
3621 case ISD::ADDCARRY:
3622 case ISD::SADDO_CARRY:
3623 if (Op.getResNo() == 1) {
3624 // If we know the result of a setcc has the top bits zero, use this info.
3625 if (TLI->getBooleanContents(Op.getOperand(0).getValueType()) ==
3626 TargetLowering::ZeroOrOneBooleanContent &&
3627 BitWidth > 1)
3628 Known.Zero.setBitsFrom(1);
3629 break;
3630 }
3631 [[fallthrough]];
3632 case ISD::ADD:
3633 case ISD::ADDC:
3634 case ISD::ADDE: {
3635 assert(Op.getResNo() == 0 && "We only compute knownbits for the sum here.");
3636
3637 // With ADDE and ADDCARRY, a carry bit may be added in.
3638 KnownBits Carry(1);
3639 if (Opcode == ISD::ADDE)
3640 // Can't track carry from glue, set carry to unknown.
3641 Carry.resetAll();
3642 else if (Opcode == ISD::ADDCARRY || Opcode == ISD::SADDO_CARRY)
3643 // TODO: Compute known bits for the carry operand. Not sure if it is worth
3644 // the trouble (how often will we find a known carry bit). And I haven't
3645 // tested this very much yet, but something like this might work:
3646 // Carry = computeKnownBits(Op.getOperand(2), DemandedElts, Depth + 1);
3647 // Carry = Carry.zextOrTrunc(1, false);
3648 Carry.resetAll();
3649 else
3650 Carry.setAllZero();
3651
3652 Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3653 Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3654 Known = KnownBits::computeForAddCarry(Known, Known2, Carry);
3655 break;
3656 }
3657 case ISD::UDIV: {
3658 Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3659 Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3660 Known = KnownBits::udiv(Known, Known2);
3661 break;
3662 }
3663 case ISD::SREM: {
3664 Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3665 Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3666 Known = KnownBits::srem(Known, Known2);
3667 break;
3668 }
3669 case ISD::UREM: {
3670 Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3671 Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3672 Known = KnownBits::urem(Known, Known2);
3673 break;
3674 }
3675 case ISD::EXTRACT_ELEMENT: {
3676 Known = computeKnownBits(Op.getOperand(0), Depth+1);
3677 const unsigned Index = Op.getConstantOperandVal(1);
3678 const unsigned EltBitWidth = Op.getValueSizeInBits();
3679
3680 // Remove low part of known bits mask
3681 Known.Zero = Known.Zero.getHiBits(Known.getBitWidth() - Index * EltBitWidth);
3682 Known.One = Known.One.getHiBits(Known.getBitWidth() - Index * EltBitWidth);
3683
3684 // Remove high part of known bit mask
3685 Known = Known.trunc(EltBitWidth);
3686 break;
3687 }
3688 case ISD::EXTRACT_VECTOR_ELT: {
3689 SDValue InVec = Op.getOperand(0);
3690 SDValue EltNo = Op.getOperand(1);
3691 EVT VecVT = InVec.getValueType();
3692 // computeKnownBits not yet implemented for scalable vectors.
3693 if (VecVT.isScalableVector())
3694 break;
3695 const unsigned EltBitWidth = VecVT.getScalarSizeInBits();
3696 const unsigned NumSrcElts = VecVT.getVectorNumElements();
3697
3698 // If BitWidth > EltBitWidth the value is anyext:ed. So we do not know
3699 // anything about the extended bits.
3700 if (BitWidth > EltBitWidth)
3701 Known = Known.trunc(EltBitWidth);
3702
3703 // If we know the element index, just demand that vector element, else for
3704 // an unknown element index, ignore DemandedElts and demand them all.
3705 APInt DemandedSrcElts = APInt::getAllOnes(NumSrcElts);
3706 auto *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo);
3707 if (ConstEltNo && ConstEltNo->getAPIntValue().ult(NumSrcElts))
3708 DemandedSrcElts =
3709 APInt::getOneBitSet(NumSrcElts, ConstEltNo->getZExtValue());
3710
3711 Known = computeKnownBits(InVec, DemandedSrcElts, Depth + 1);
3712 if (BitWidth > EltBitWidth)
3713 Known = Known.anyext(BitWidth);
3714 break;
3715 }
3716 case ISD::INSERT_VECTOR_ELT: {
3717 if (Op.getValueType().isScalableVector())
3718 break;
3719
3720 // If we know the element index, split the demand between the
3721 // source vector and the inserted element, otherwise assume we need
3722 // the original demanded vector elements and the value.
3723 SDValue InVec = Op.getOperand(0);
3724 SDValue InVal = Op.getOperand(1);
3725 SDValue EltNo = Op.getOperand(2);
3726 bool DemandedVal = true;
3727 APInt DemandedVecElts = DemandedElts;
3728 auto *CEltNo = dyn_cast<ConstantSDNode>(EltNo);
3729 if (CEltNo && CEltNo->getAPIntValue().ult(NumElts)) {
3730 unsigned EltIdx = CEltNo->getZExtValue();
3731 DemandedVal = !!DemandedElts[EltIdx];
3732 DemandedVecElts.clearBit(EltIdx);
3733 }
3734 Known.One.setAllBits();
3735 Known.Zero.setAllBits();
3736 if (DemandedVal) {
3737 Known2 = computeKnownBits(InVal, Depth + 1);
3738 Known = KnownBits::commonBits(Known, Known2.zextOrTrunc(BitWidth));
3739 }
3740 if (!!DemandedVecElts) {
3741 Known2 = computeKnownBits(InVec, DemandedVecElts, Depth + 1);
3742 Known = KnownBits::commonBits(Known, Known2);
3743 }
3744 break;
3745 }
3746 case ISD::BITREVERSE: {
3747 Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3748 Known = Known2.reverseBits();
3749 break;
3750 }
3751 case ISD::BSWAP: {
3752 Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3753 Known = Known2.byteSwap();
3754 break;
3755 }
3756 case ISD::ABS: {
3757 Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3758 Known = Known2.abs();
3759 break;
3760 }
3761 case ISD::USUBSAT: {
3762 // The result of usubsat will never be larger than the LHS.
3763 Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3764 Known.Zero.setHighBits(Known2.countMinLeadingZeros());
3765 break;
3766 }
3767 case ISD::UMIN: {
3768 Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3769 Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3770 Known = KnownBits::umin(Known, Known2);
3771 break;
3772 }
3773 case ISD::UMAX: {
3774 Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3775 Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3776 Known = KnownBits::umax(Known, Known2);
3777 break;
3778 }
3779 case ISD::SMIN:
3780 case ISD::SMAX: {
3781 // If we have a clamp pattern, we know that the number of sign bits will be
3782 // the minimum of the clamp min/max range.
3783 bool IsMax = (Opcode == ISD::SMAX);
3784 ConstantSDNode *CstLow = nullptr, *CstHigh = nullptr;
3785 if ((CstLow = isConstOrConstSplat(Op.getOperand(1), DemandedElts)))
3786 if (Op.getOperand(0).getOpcode() == (IsMax ? ISD::SMIN : ISD::SMAX))
3787 CstHigh =
3788 isConstOrConstSplat(Op.getOperand(0).getOperand(1), DemandedElts);
3789 if (CstLow && CstHigh) {
3790 if (!IsMax)
3791 std::swap(CstLow, CstHigh);
3792
3793 const APInt &ValueLow = CstLow->getAPIntValue();
3794 const APInt &ValueHigh = CstHigh->getAPIntValue();
3795 if (ValueLow.sle(ValueHigh)) {
3796 unsigned LowSignBits = ValueLow.getNumSignBits();
3797 unsigned HighSignBits = ValueHigh.getNumSignBits();
3798 unsigned MinSignBits = std::min(LowSignBits, HighSignBits);
3799 if (ValueLow.isNegative() && ValueHigh.isNegative()) {
3800 Known.One.setHighBits(MinSignBits);
3801 break;
3802 }
3803 if (ValueLow.isNonNegative() && ValueHigh.isNonNegative()) {
3804 Known.Zero.setHighBits(MinSignBits);
3805 break;
3806 }
3807 }
3808 }
3809
3810 Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
3811 Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
3812 if (IsMax)
3813 Known = KnownBits::smax(Known, Known2);
3814 else
3815 Known = KnownBits::smin(Known, Known2);
3816
3817 // For SMAX, if CstLow is non-negative we know the result will be
3818 // non-negative and thus all sign bits are 0.
3819 // TODO: There's an equivalent of this for smin with negative constant for
3820 // known ones.
3821 if (IsMax && CstLow) {
3822 const APInt &ValueLow = CstLow->getAPIntValue();
3823 if (ValueLow.isNonNegative()) {
3824 unsigned SignBits = ComputeNumSignBits(Op.getOperand(0), Depth + 1);
3825 Known.Zero.setHighBits(std::min(SignBits, ValueLow.getNumSignBits()));
3826 }
3827 }
3828
3829 break;
3830 }
3831 case ISD::FP_TO_UINT_SAT: {
3832 // FP_TO_UINT_SAT produces an unsigned value that fits in the saturating VT.
3833 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
3834 Known.Zero |= APInt::getBitsSetFrom(BitWidth, VT.getScalarSizeInBits());
3835 break;
3836 }
3837 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
3838 if (Op.getResNo() == 1) {
3839 // The boolean result conforms to getBooleanContents.
3840 // If we know the result of a setcc has the top bits zero, use this info.
3841 // We know that we have an integer-based boolean since these operations
3842 // are only available for integer.
3843 if (TLI->getBooleanContents(Op.getValueType().isVector(), false) ==
3844 TargetLowering::ZeroOrOneBooleanContent &&
3845 BitWidth > 1)
3846 Known.Zero.setBitsFrom(1);
3847 break;
3848 }
3849 [[fallthrough]];
3850 case ISD::ATOMIC_CMP_SWAP:
3851 case ISD::ATOMIC_SWAP:
3852 case ISD::ATOMIC_LOAD_ADD:
3853 case ISD::ATOMIC_LOAD_SUB:
3854 case ISD::ATOMIC_LOAD_AND:
3855 case ISD::ATOMIC_LOAD_CLR:
3856 case ISD::ATOMIC_LOAD_OR:
3857 case ISD::ATOMIC_LOAD_XOR:
3858 case ISD::ATOMIC_LOAD_NAND:
3859 case ISD::ATOMIC_LOAD_MIN:
3860 case ISD::ATOMIC_LOAD_MAX:
3861 case ISD::ATOMIC_LOAD_UMIN:
3862 case ISD::ATOMIC_LOAD_UMAX:
3863 case ISD::ATOMIC_LOAD: {
3864 unsigned MemBits =
3865 cast<AtomicSDNode>(Op)->getMemoryVT().getScalarSizeInBits();
3866 // If we are looking at the loaded value.
3867 if (Op.getResNo() == 0) {
3868 if (TLI->getExtendForAtomicOps() == ISD::ZERO_EXTEND)
3869 Known.Zero.setBitsFrom(MemBits);
3870 }
3871 break;
3872 }
3873 case ISD::FrameIndex:
3874 case ISD::TargetFrameIndex:
3875 TLI->computeKnownBitsForFrameIndex(cast<FrameIndexSDNode>(Op)->getIndex(),
3876 Known, getMachineFunction());
3877 break;
3878
3879 default:
3880 if (Opcode < ISD::BUILTIN_OP_END)
3881 break;
3882 [[fallthrough]];
3883 case ISD::INTRINSIC_WO_CHAIN:
3884 case ISD::INTRINSIC_W_CHAIN:
3885 case ISD::INTRINSIC_VOID:
3886 // TODO: Probably okay to remove after audit; here to reduce change size
3887 // in initial enablement patch for scalable vectors
3888 if (Op.getValueType().isScalableVector())
3889 break;
3890
3891 // Allow the target to implement this method for its nodes.
3892 TLI->computeKnownBitsForTargetNode(Op, Known, DemandedElts, *this, Depth);
3893 break;
3894 }
3895
3896 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
3897 return Known;
3898 }
3899
computeOverflowKind(SDValue N0,SDValue N1) const3900 SelectionDAG::OverflowKind SelectionDAG::computeOverflowKind(SDValue N0,
3901 SDValue N1) const {
3902 // X + 0 never overflow
3903 if (isNullConstant(N1))
3904 return OFK_Never;
3905
3906 KnownBits N1Known = computeKnownBits(N1);
3907 if (N1Known.Zero.getBoolValue()) {
3908 KnownBits N0Known = computeKnownBits(N0);
3909
3910 bool overflow;
3911 (void)N0Known.getMaxValue().uadd_ov(N1Known.getMaxValue(), overflow);
3912 if (!overflow)
3913 return OFK_Never;
3914 }
3915
3916 // mulhi + 1 never overflow
3917 if (N0.getOpcode() == ISD::UMUL_LOHI && N0.getResNo() == 1 &&
3918 (N1Known.getMaxValue() & 0x01) == N1Known.getMaxValue())
3919 return OFK_Never;
3920
3921 if (N1.getOpcode() == ISD::UMUL_LOHI && N1.getResNo() == 1) {
3922 KnownBits N0Known = computeKnownBits(N0);
3923
3924 if ((N0Known.getMaxValue() & 0x01) == N0Known.getMaxValue())
3925 return OFK_Never;
3926 }
3927
3928 return OFK_Sometime;
3929 }
3930
isKnownToBeAPowerOfTwo(SDValue Val) const3931 bool SelectionDAG::isKnownToBeAPowerOfTwo(SDValue Val) const {
3932 EVT OpVT = Val.getValueType();
3933 unsigned BitWidth = OpVT.getScalarSizeInBits();
3934
3935 // Is the constant a known power of 2?
3936 if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(Val))
3937 return Const->getAPIntValue().zextOrTrunc(BitWidth).isPowerOf2();
3938
3939 // A left-shift of a constant one will have exactly one bit set because
3940 // shifting the bit off the end is undefined.
3941 if (Val.getOpcode() == ISD::SHL) {
3942 auto *C = isConstOrConstSplat(Val.getOperand(0));
3943 if (C && C->getAPIntValue() == 1)
3944 return true;
3945 }
3946
3947 // Similarly, a logical right-shift of a constant sign-bit will have exactly
3948 // one bit set.
3949 if (Val.getOpcode() == ISD::SRL) {
3950 auto *C = isConstOrConstSplat(Val.getOperand(0));
3951 if (C && C->getAPIntValue().isSignMask())
3952 return true;
3953 }
3954
3955 // Are all operands of a build vector constant powers of two?
3956 if (Val.getOpcode() == ISD::BUILD_VECTOR)
3957 if (llvm::all_of(Val->ops(), [BitWidth](SDValue E) {
3958 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(E))
3959 return C->getAPIntValue().zextOrTrunc(BitWidth).isPowerOf2();
3960 return false;
3961 }))
3962 return true;
3963
3964 // Is the operand of a splat vector a constant power of two?
3965 if (Val.getOpcode() == ISD::SPLAT_VECTOR)
3966 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val->getOperand(0)))
3967 if (C->getAPIntValue().zextOrTrunc(BitWidth).isPowerOf2())
3968 return true;
3969
3970 // vscale(power-of-two) is a power-of-two for some targets
3971 if (Val.getOpcode() == ISD::VSCALE &&
3972 getTargetLoweringInfo().isVScaleKnownToBeAPowerOfTwo() &&
3973 isKnownToBeAPowerOfTwo(Val.getOperand(0)))
3974 return true;
3975
3976 // More could be done here, though the above checks are enough
3977 // to handle some common cases.
3978
3979 // Fall back to computeKnownBits to catch other known cases.
3980 KnownBits Known = computeKnownBits(Val);
3981 return (Known.countMaxPopulation() == 1) && (Known.countMinPopulation() == 1);
3982 }
3983
ComputeNumSignBits(SDValue Op,unsigned Depth) const3984 unsigned SelectionDAG::ComputeNumSignBits(SDValue Op, unsigned Depth) const {
3985 EVT VT = Op.getValueType();
3986
3987 // Since the number of lanes in a scalable vector is unknown at compile time,
3988 // we track one bit which is implicitly broadcast to all lanes. This means
3989 // that all lanes in a scalable vector are considered demanded.
3990 APInt DemandedElts = VT.isFixedLengthVector()
3991 ? APInt::getAllOnes(VT.getVectorNumElements())
3992 : APInt(1, 1);
3993 return ComputeNumSignBits(Op, DemandedElts, Depth);
3994 }
3995
ComputeNumSignBits(SDValue Op,const APInt & DemandedElts,unsigned Depth) const3996 unsigned SelectionDAG::ComputeNumSignBits(SDValue Op, const APInt &DemandedElts,
3997 unsigned Depth) const {
3998 EVT VT = Op.getValueType();
3999 assert((VT.isInteger() || VT.isFloatingPoint()) && "Invalid VT!");
4000 unsigned VTBits = VT.getScalarSizeInBits();
4001 unsigned NumElts = DemandedElts.getBitWidth();
4002 unsigned Tmp, Tmp2;
4003 unsigned FirstAnswer = 1;
4004
4005 if (auto *C = dyn_cast<ConstantSDNode>(Op)) {
4006 const APInt &Val = C->getAPIntValue();
4007 return Val.getNumSignBits();
4008 }
4009
4010 if (Depth >= MaxRecursionDepth)
4011 return 1; // Limit search depth.
4012
4013 if (!DemandedElts)
4014 return 1; // No demanded elts, better to assume we don't know anything.
4015
4016 unsigned Opcode = Op.getOpcode();
4017 switch (Opcode) {
4018 default: break;
4019 case ISD::AssertSext:
4020 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits();
4021 return VTBits-Tmp+1;
4022 case ISD::AssertZext:
4023 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits();
4024 return VTBits-Tmp;
4025 case ISD::MERGE_VALUES:
4026 return ComputeNumSignBits(Op.getOperand(Op.getResNo()), DemandedElts,
4027 Depth + 1);
4028 case ISD::SPLAT_VECTOR: {
4029 // Check if the sign bits of source go down as far as the truncated value.
4030 unsigned NumSrcBits = Op.getOperand(0).getValueSizeInBits();
4031 unsigned NumSrcSignBits = ComputeNumSignBits(Op.getOperand(0), Depth + 1);
4032 if (NumSrcSignBits > (NumSrcBits - VTBits))
4033 return NumSrcSignBits - (NumSrcBits - VTBits);
4034 break;
4035 }
4036 case ISD::BUILD_VECTOR:
4037 assert(!VT.isScalableVector());
4038 Tmp = VTBits;
4039 for (unsigned i = 0, e = Op.getNumOperands(); (i < e) && (Tmp > 1); ++i) {
4040 if (!DemandedElts[i])
4041 continue;
4042
4043 SDValue SrcOp = Op.getOperand(i);
4044 Tmp2 = ComputeNumSignBits(SrcOp, Depth + 1);
4045
4046 // BUILD_VECTOR can implicitly truncate sources, we must handle this.
4047 if (SrcOp.getValueSizeInBits() != VTBits) {
4048 assert(SrcOp.getValueSizeInBits() > VTBits &&
4049 "Expected BUILD_VECTOR implicit truncation");
4050 unsigned ExtraBits = SrcOp.getValueSizeInBits() - VTBits;
4051 Tmp2 = (Tmp2 > ExtraBits ? Tmp2 - ExtraBits : 1);
4052 }
4053 Tmp = std::min(Tmp, Tmp2);
4054 }
4055 return Tmp;
4056
4057 case ISD::VECTOR_SHUFFLE: {
4058 // Collect the minimum number of sign bits that are shared by every vector
4059 // element referenced by the shuffle.
4060 APInt DemandedLHS, DemandedRHS;
4061 const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op);
4062 assert(NumElts == SVN->getMask().size() && "Unexpected vector size");
4063 if (!getShuffleDemandedElts(NumElts, SVN->getMask(), DemandedElts,
4064 DemandedLHS, DemandedRHS))
4065 return 1;
4066
4067 Tmp = std::numeric_limits<unsigned>::max();
4068 if (!!DemandedLHS)
4069 Tmp = ComputeNumSignBits(Op.getOperand(0), DemandedLHS, Depth + 1);
4070 if (!!DemandedRHS) {
4071 Tmp2 = ComputeNumSignBits(Op.getOperand(1), DemandedRHS, Depth + 1);
4072 Tmp = std::min(Tmp, Tmp2);
4073 }
4074 // If we don't know anything, early out and try computeKnownBits fall-back.
4075 if (Tmp == 1)
4076 break;
4077 assert(Tmp <= VTBits && "Failed to determine minimum sign bits");
4078 return Tmp;
4079 }
4080
4081 case ISD::BITCAST: {
4082 if (VT.isScalableVector())
4083 break;
4084 SDValue N0 = Op.getOperand(0);
4085 EVT SrcVT = N0.getValueType();
4086 unsigned SrcBits = SrcVT.getScalarSizeInBits();
4087
4088 // Ignore bitcasts from unsupported types..
4089 if (!(SrcVT.isInteger() || SrcVT.isFloatingPoint()))
4090 break;
4091
4092 // Fast handling of 'identity' bitcasts.
4093 if (VTBits == SrcBits)
4094 return ComputeNumSignBits(N0, DemandedElts, Depth + 1);
4095
4096 bool IsLE = getDataLayout().isLittleEndian();
4097
4098 // Bitcast 'large element' scalar/vector to 'small element' vector.
4099 if ((SrcBits % VTBits) == 0) {
4100 assert(VT.isVector() && "Expected bitcast to vector");
4101
4102 unsigned Scale = SrcBits / VTBits;
4103 APInt SrcDemandedElts =
4104 APIntOps::ScaleBitMask(DemandedElts, NumElts / Scale);
4105
4106 // Fast case - sign splat can be simply split across the small elements.
4107 Tmp = ComputeNumSignBits(N0, SrcDemandedElts, Depth + 1);
4108 if (Tmp == SrcBits)
4109 return VTBits;
4110
4111 // Slow case - determine how far the sign extends into each sub-element.
4112 Tmp2 = VTBits;
4113 for (unsigned i = 0; i != NumElts; ++i)
4114 if (DemandedElts[i]) {
4115 unsigned SubOffset = i % Scale;
4116 SubOffset = (IsLE ? ((Scale - 1) - SubOffset) : SubOffset);
4117 SubOffset = SubOffset * VTBits;
4118 if (Tmp <= SubOffset)
4119 return 1;
4120 Tmp2 = std::min(Tmp2, Tmp - SubOffset);
4121 }
4122 return Tmp2;
4123 }
4124 break;
4125 }
4126
4127 case ISD::FP_TO_SINT_SAT:
4128 // FP_TO_SINT_SAT produces a signed value that fits in the saturating VT.
4129 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getScalarSizeInBits();
4130 return VTBits - Tmp + 1;
4131 case ISD::SIGN_EXTEND:
4132 Tmp = VTBits - Op.getOperand(0).getScalarValueSizeInBits();
4133 return ComputeNumSignBits(Op.getOperand(0), DemandedElts, Depth+1) + Tmp;
4134 case ISD::SIGN_EXTEND_INREG:
4135 // Max of the input and what this extends.
4136 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getScalarSizeInBits();
4137 Tmp = VTBits-Tmp+1;
4138 Tmp2 = ComputeNumSignBits(Op.getOperand(0), DemandedElts, Depth+1);
4139 return std::max(Tmp, Tmp2);
4140 case ISD::SIGN_EXTEND_VECTOR_INREG: {
4141 if (VT.isScalableVector())
4142 break;
4143 SDValue Src = Op.getOperand(0);
4144 EVT SrcVT = Src.getValueType();
4145 APInt DemandedSrcElts = DemandedElts.zext(SrcVT.getVectorNumElements());
4146 Tmp = VTBits - SrcVT.getScalarSizeInBits();
4147 return ComputeNumSignBits(Src, DemandedSrcElts, Depth+1) + Tmp;
4148 }
4149 case ISD::SRA:
4150 Tmp = ComputeNumSignBits(Op.getOperand(0), DemandedElts, Depth + 1);
4151 // SRA X, C -> adds C sign bits.
4152 if (const APInt *ShAmt =
4153 getValidMinimumShiftAmountConstant(Op, DemandedElts))
4154 Tmp = std::min<uint64_t>(Tmp + ShAmt->getZExtValue(), VTBits);
4155 return Tmp;
4156 case ISD::SHL:
4157 if (const APInt *ShAmt =
4158 getValidMaximumShiftAmountConstant(Op, DemandedElts)) {
4159 // shl destroys sign bits, ensure it doesn't shift out all sign bits.
4160 Tmp = ComputeNumSignBits(Op.getOperand(0), DemandedElts, Depth + 1);
4161 if (ShAmt->ult(Tmp))
4162 return Tmp - ShAmt->getZExtValue();
4163 }
4164 break;
4165 case ISD::AND:
4166 case ISD::OR:
4167 case ISD::XOR: // NOT is handled here.
4168 // Logical binary ops preserve the number of sign bits at the worst.
4169 Tmp = ComputeNumSignBits(Op.getOperand(0), DemandedElts, Depth+1);
4170 if (Tmp != 1) {
4171 Tmp2 = ComputeNumSignBits(Op.getOperand(1), DemandedElts, Depth+1);
4172 FirstAnswer = std::min(Tmp, Tmp2);
4173 // We computed what we know about the sign bits as our first
4174 // answer. Now proceed to the generic code that uses
4175 // computeKnownBits, and pick whichever answer is better.
4176 }
4177 break;
4178
4179 case ISD::SELECT:
4180 case ISD::VSELECT:
4181 Tmp = ComputeNumSignBits(Op.getOperand(1), DemandedElts, Depth+1);
4182 if (Tmp == 1) return 1; // Early out.
4183 Tmp2 = ComputeNumSignBits(Op.getOperand(2), DemandedElts, Depth+1);
4184 return std::min(Tmp, Tmp2);
4185 case ISD::SELECT_CC:
4186 Tmp = ComputeNumSignBits(Op.getOperand(2), DemandedElts, Depth+1);
4187 if (Tmp == 1) return 1; // Early out.
4188 Tmp2 = ComputeNumSignBits(Op.getOperand(3), DemandedElts, Depth+1);
4189 return std::min(Tmp, Tmp2);
4190
4191 case ISD::SMIN:
4192 case ISD::SMAX: {
4193 // If we have a clamp pattern, we know that the number of sign bits will be
4194 // the minimum of the clamp min/max range.
4195 bool IsMax = (Opcode == ISD::SMAX);
4196 ConstantSDNode *CstLow = nullptr, *CstHigh = nullptr;
4197 if ((CstLow = isConstOrConstSplat(Op.getOperand(1), DemandedElts)))
4198 if (Op.getOperand(0).getOpcode() == (IsMax ? ISD::SMIN : ISD::SMAX))
4199 CstHigh =
4200 isConstOrConstSplat(Op.getOperand(0).getOperand(1), DemandedElts);
4201 if (CstLow && CstHigh) {
4202 if (!IsMax)
4203 std::swap(CstLow, CstHigh);
4204 if (CstLow->getAPIntValue().sle(CstHigh->getAPIntValue())) {
4205 Tmp = CstLow->getAPIntValue().getNumSignBits();
4206 Tmp2 = CstHigh->getAPIntValue().getNumSignBits();
4207 return std::min(Tmp, Tmp2);
4208 }
4209 }
4210
4211 // Fallback - just get the minimum number of sign bits of the operands.
4212 Tmp = ComputeNumSignBits(Op.getOperand(0), DemandedElts, Depth + 1);
4213 if (Tmp == 1)
4214 return 1; // Early out.
4215 Tmp2 = ComputeNumSignBits(Op.getOperand(1), DemandedElts, Depth + 1);
4216 return std::min(Tmp, Tmp2);
4217 }
4218 case ISD::UMIN:
4219 case ISD::UMAX:
4220 Tmp = ComputeNumSignBits(Op.getOperand(0), DemandedElts, Depth + 1);
4221 if (Tmp == 1)
4222 return 1; // Early out.
4223 Tmp2 = ComputeNumSignBits(Op.getOperand(1), DemandedElts, Depth + 1);
4224 return std::min(Tmp, Tmp2);
4225 case ISD::SADDO:
4226 case ISD::UADDO:
4227 case ISD::SADDO_CARRY:
4228 case ISD::ADDCARRY:
4229 case ISD::SSUBO:
4230 case ISD::USUBO:
4231 case ISD::SSUBO_CARRY:
4232 case ISD::SUBCARRY:
4233 case ISD::SMULO:
4234 case ISD::UMULO:
4235 if (Op.getResNo() != 1)
4236 break;
4237 // The boolean result conforms to getBooleanContents. Fall through.
4238 // If setcc returns 0/-1, all bits are sign bits.
4239 // We know that we have an integer-based boolean since these operations
4240 // are only available for integer.
4241 if (TLI->getBooleanContents(VT.isVector(), false) ==
4242 TargetLowering::ZeroOrNegativeOneBooleanContent)
4243 return VTBits;
4244 break;
4245 case ISD::SETCC:
4246 case ISD::SETCCCARRY:
4247 case ISD::STRICT_FSETCC:
4248 case ISD::STRICT_FSETCCS: {
4249 unsigned OpNo = Op->isStrictFPOpcode() ? 1 : 0;
4250 // If setcc returns 0/-1, all bits are sign bits.
4251 if (TLI->getBooleanContents(Op.getOperand(OpNo).getValueType()) ==
4252 TargetLowering::ZeroOrNegativeOneBooleanContent)
4253 return VTBits;
4254 break;
4255 }
4256 case ISD::ROTL:
4257 case ISD::ROTR:
4258 Tmp = ComputeNumSignBits(Op.getOperand(0), DemandedElts, Depth + 1);
4259
4260 // If we're rotating an 0/-1 value, then it stays an 0/-1 value.
4261 if (Tmp == VTBits)
4262 return VTBits;
4263
4264 if (ConstantSDNode *C =
4265 isConstOrConstSplat(Op.getOperand(1), DemandedElts)) {
4266 unsigned RotAmt = C->getAPIntValue().urem(VTBits);
4267
4268 // Handle rotate right by N like a rotate left by 32-N.
4269 if (Opcode == ISD::ROTR)
4270 RotAmt = (VTBits - RotAmt) % VTBits;
4271
4272 // If we aren't rotating out all of the known-in sign bits, return the
4273 // number that are left. This handles rotl(sext(x), 1) for example.
4274 if (Tmp > (RotAmt + 1)) return (Tmp - RotAmt);
4275 }
4276 break;
4277 case ISD::ADD:
4278 case ISD::ADDC:
4279 // Add can have at most one carry bit. Thus we know that the output
4280 // is, at worst, one more bit than the inputs.
4281 Tmp = ComputeNumSignBits(Op.getOperand(0), DemandedElts, Depth + 1);
4282 if (Tmp == 1) return 1; // Early out.
4283
4284 // Special case decrementing a value (ADD X, -1):
4285 if (ConstantSDNode *CRHS =
4286 isConstOrConstSplat(Op.getOperand(1), DemandedElts))
4287 if (CRHS->isAllOnes()) {
4288 KnownBits Known =
4289 computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
4290
4291 // If the input is known to be 0 or 1, the output is 0/-1, which is all
4292 // sign bits set.
4293 if ((Known.Zero | 1).isAllOnes())
4294 return VTBits;
4295
4296 // If we are subtracting one from a positive number, there is no carry
4297 // out of the result.
4298 if (Known.isNonNegative())
4299 return Tmp;
4300 }
4301
4302 Tmp2 = ComputeNumSignBits(Op.getOperand(1), DemandedElts, Depth + 1);
4303 if (Tmp2 == 1) return 1; // Early out.
4304 return std::min(Tmp, Tmp2) - 1;
4305 case ISD::SUB:
4306 Tmp2 = ComputeNumSignBits(Op.getOperand(1), DemandedElts, Depth + 1);
4307 if (Tmp2 == 1) return 1; // Early out.
4308
4309 // Handle NEG.
4310 if (ConstantSDNode *CLHS =
4311 isConstOrConstSplat(Op.getOperand(0), DemandedElts))
4312 if (CLHS->isZero()) {
4313 KnownBits Known =
4314 computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
4315 // If the input is known to be 0 or 1, the output is 0/-1, which is all
4316 // sign bits set.
4317 if ((Known.Zero | 1).isAllOnes())
4318 return VTBits;
4319
4320 // If the input is known to be positive (the sign bit is known clear),
4321 // the output of the NEG has the same number of sign bits as the input.
4322 if (Known.isNonNegative())
4323 return Tmp2;
4324
4325 // Otherwise, we treat this like a SUB.
4326 }
4327
4328 // Sub can have at most one carry bit. Thus we know that the output
4329 // is, at worst, one more bit than the inputs.
4330 Tmp = ComputeNumSignBits(Op.getOperand(0), DemandedElts, Depth + 1);
4331 if (Tmp == 1) return 1; // Early out.
4332 return std::min(Tmp, Tmp2) - 1;
4333 case ISD::MUL: {
4334 // The output of the Mul can be at most twice the valid bits in the inputs.
4335 unsigned SignBitsOp0 = ComputeNumSignBits(Op.getOperand(0), Depth + 1);
4336 if (SignBitsOp0 == 1)
4337 break;
4338 unsigned SignBitsOp1 = ComputeNumSignBits(Op.getOperand(1), Depth + 1);
4339 if (SignBitsOp1 == 1)
4340 break;
4341 unsigned OutValidBits =
4342 (VTBits - SignBitsOp0 + 1) + (VTBits - SignBitsOp1 + 1);
4343 return OutValidBits > VTBits ? 1 : VTBits - OutValidBits + 1;
4344 }
4345 case ISD::SREM:
4346 // The sign bit is the LHS's sign bit, except when the result of the
4347 // remainder is zero. The magnitude of the result should be less than or
4348 // equal to the magnitude of the LHS. Therefore, the result should have
4349 // at least as many sign bits as the left hand side.
4350 return ComputeNumSignBits(Op.getOperand(0), DemandedElts, Depth + 1);
4351 case ISD::TRUNCATE: {
4352 // Check if the sign bits of source go down as far as the truncated value.
4353 unsigned NumSrcBits = Op.getOperand(0).getScalarValueSizeInBits();
4354 unsigned NumSrcSignBits = ComputeNumSignBits(Op.getOperand(0), Depth + 1);
4355 if (NumSrcSignBits > (NumSrcBits - VTBits))
4356 return NumSrcSignBits - (NumSrcBits - VTBits);
4357 break;
4358 }
4359 case ISD::EXTRACT_ELEMENT: {
4360 if (VT.isScalableVector())
4361 break;
4362 const int KnownSign = ComputeNumSignBits(Op.getOperand(0), Depth+1);
4363 const int BitWidth = Op.getValueSizeInBits();
4364 const int Items = Op.getOperand(0).getValueSizeInBits() / BitWidth;
4365
4366 // Get reverse index (starting from 1), Op1 value indexes elements from
4367 // little end. Sign starts at big end.
4368 const int rIndex = Items - 1 - Op.getConstantOperandVal(1);
4369
4370 // If the sign portion ends in our element the subtraction gives correct
4371 // result. Otherwise it gives either negative or > bitwidth result
4372 return std::clamp(KnownSign - rIndex * BitWidth, 0, BitWidth);
4373 }
4374 case ISD::INSERT_VECTOR_ELT: {
4375 if (VT.isScalableVector())
4376 break;
4377 // If we know the element index, split the demand between the
4378 // source vector and the inserted element, otherwise assume we need
4379 // the original demanded vector elements and the value.
4380 SDValue InVec = Op.getOperand(0);
4381 SDValue InVal = Op.getOperand(1);
4382 SDValue EltNo = Op.getOperand(2);
4383 bool DemandedVal = true;
4384 APInt DemandedVecElts = DemandedElts;
4385 auto *CEltNo = dyn_cast<ConstantSDNode>(EltNo);
4386 if (CEltNo && CEltNo->getAPIntValue().ult(NumElts)) {
4387 unsigned EltIdx = CEltNo->getZExtValue();
4388 DemandedVal = !!DemandedElts[EltIdx];
4389 DemandedVecElts.clearBit(EltIdx);
4390 }
4391 Tmp = std::numeric_limits<unsigned>::max();
4392 if (DemandedVal) {
4393 // TODO - handle implicit truncation of inserted elements.
4394 if (InVal.getScalarValueSizeInBits() != VTBits)
4395 break;
4396 Tmp2 = ComputeNumSignBits(InVal, Depth + 1);
4397 Tmp = std::min(Tmp, Tmp2);
4398 }
4399 if (!!DemandedVecElts) {
4400 Tmp2 = ComputeNumSignBits(InVec, DemandedVecElts, Depth + 1);
4401 Tmp = std::min(Tmp, Tmp2);
4402 }
4403 assert(Tmp <= VTBits && "Failed to determine minimum sign bits");
4404 return Tmp;
4405 }
4406 case ISD::EXTRACT_VECTOR_ELT: {
4407 assert(!VT.isScalableVector());
4408 SDValue InVec = Op.getOperand(0);
4409 SDValue EltNo = Op.getOperand(1);
4410 EVT VecVT = InVec.getValueType();
4411 // ComputeNumSignBits not yet implemented for scalable vectors.
4412 if (VecVT.isScalableVector())
4413 break;
4414 const unsigned BitWidth = Op.getValueSizeInBits();
4415 const unsigned EltBitWidth = Op.getOperand(0).getScalarValueSizeInBits();
4416 const unsigned NumSrcElts = VecVT.getVectorNumElements();
4417
4418 // If BitWidth > EltBitWidth the value is anyext:ed, and we do not know
4419 // anything about sign bits. But if the sizes match we can derive knowledge
4420 // about sign bits from the vector operand.
4421 if (BitWidth != EltBitWidth)
4422 break;
4423
4424 // If we know the element index, just demand that vector element, else for
4425 // an unknown element index, ignore DemandedElts and demand them all.
4426 APInt DemandedSrcElts = APInt::getAllOnes(NumSrcElts);
4427 auto *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo);
4428 if (ConstEltNo && ConstEltNo->getAPIntValue().ult(NumSrcElts))
4429 DemandedSrcElts =
4430 APInt::getOneBitSet(NumSrcElts, ConstEltNo->getZExtValue());
4431
4432 return ComputeNumSignBits(InVec, DemandedSrcElts, Depth + 1);
4433 }
4434 case ISD::EXTRACT_SUBVECTOR: {
4435 // Offset the demanded elts by the subvector index.
4436 SDValue Src = Op.getOperand(0);
4437 // Bail until we can represent demanded elements for scalable vectors.
4438 if (Src.getValueType().isScalableVector())
4439 break;
4440 uint64_t Idx = Op.getConstantOperandVal(1);
4441 unsigned NumSrcElts = Src.getValueType().getVectorNumElements();
4442 APInt DemandedSrcElts = DemandedElts.zext(NumSrcElts).shl(Idx);
4443 return ComputeNumSignBits(Src, DemandedSrcElts, Depth + 1);
4444 }
4445 case ISD::CONCAT_VECTORS: {
4446 if (VT.isScalableVector())
4447 break;
4448 // Determine the minimum number of sign bits across all demanded
4449 // elts of the input vectors. Early out if the result is already 1.
4450 Tmp = std::numeric_limits<unsigned>::max();
4451 EVT SubVectorVT = Op.getOperand(0).getValueType();
4452 unsigned NumSubVectorElts = SubVectorVT.getVectorNumElements();
4453 unsigned NumSubVectors = Op.getNumOperands();
4454 for (unsigned i = 0; (i < NumSubVectors) && (Tmp > 1); ++i) {
4455 APInt DemandedSub =
4456 DemandedElts.extractBits(NumSubVectorElts, i * NumSubVectorElts);
4457 if (!DemandedSub)
4458 continue;
4459 Tmp2 = ComputeNumSignBits(Op.getOperand(i), DemandedSub, Depth + 1);
4460 Tmp = std::min(Tmp, Tmp2);
4461 }
4462 assert(Tmp <= VTBits && "Failed to determine minimum sign bits");
4463 return Tmp;
4464 }
4465 case ISD::INSERT_SUBVECTOR: {
4466 if (VT.isScalableVector())
4467 break;
4468 // Demand any elements from the subvector and the remainder from the src its
4469 // inserted into.
4470 SDValue Src = Op.getOperand(0);
4471 SDValue Sub = Op.getOperand(1);
4472 uint64_t Idx = Op.getConstantOperandVal(2);
4473 unsigned NumSubElts = Sub.getValueType().getVectorNumElements();
4474 APInt DemandedSubElts = DemandedElts.extractBits(NumSubElts, Idx);
4475 APInt DemandedSrcElts = DemandedElts;
4476 DemandedSrcElts.insertBits(APInt::getZero(NumSubElts), Idx);
4477
4478 Tmp = std::numeric_limits<unsigned>::max();
4479 if (!!DemandedSubElts) {
4480 Tmp = ComputeNumSignBits(Sub, DemandedSubElts, Depth + 1);
4481 if (Tmp == 1)
4482 return 1; // early-out
4483 }
4484 if (!!DemandedSrcElts) {
4485 Tmp2 = ComputeNumSignBits(Src, DemandedSrcElts, Depth + 1);
4486 Tmp = std::min(Tmp, Tmp2);
4487 }
4488 assert(Tmp <= VTBits && "Failed to determine minimum sign bits");
4489 return Tmp;
4490 }
4491 case ISD::LOAD: {
4492 LoadSDNode *LD = cast<LoadSDNode>(Op);
4493 if (const MDNode *Ranges = LD->getRanges()) {
4494 if (DemandedElts != 1)
4495 break;
4496
4497 ConstantRange CR = getConstantRangeFromMetadata(*Ranges);
4498 if (VTBits > CR.getBitWidth()) {
4499 switch (LD->getExtensionType()) {
4500 case ISD::SEXTLOAD:
4501 CR = CR.signExtend(VTBits);
4502 break;
4503 case ISD::ZEXTLOAD:
4504 CR = CR.zeroExtend(VTBits);
4505 break;
4506 default:
4507 break;
4508 }
4509 }
4510
4511 if (VTBits != CR.getBitWidth())
4512 break;
4513 return std::min(CR.getSignedMin().getNumSignBits(),
4514 CR.getSignedMax().getNumSignBits());
4515 }
4516
4517 break;
4518 }
4519 case ISD::ATOMIC_CMP_SWAP:
4520 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
4521 case ISD::ATOMIC_SWAP:
4522 case ISD::ATOMIC_LOAD_ADD:
4523 case ISD::ATOMIC_LOAD_SUB:
4524 case ISD::ATOMIC_LOAD_AND:
4525 case ISD::ATOMIC_LOAD_CLR:
4526 case ISD::ATOMIC_LOAD_OR:
4527 case ISD::ATOMIC_LOAD_XOR:
4528 case ISD::ATOMIC_LOAD_NAND:
4529 case ISD::ATOMIC_LOAD_MIN:
4530 case ISD::ATOMIC_LOAD_MAX:
4531 case ISD::ATOMIC_LOAD_UMIN:
4532 case ISD::ATOMIC_LOAD_UMAX:
4533 case ISD::ATOMIC_LOAD: {
4534 Tmp = cast<AtomicSDNode>(Op)->getMemoryVT().getScalarSizeInBits();
4535 // If we are looking at the loaded value.
4536 if (Op.getResNo() == 0) {
4537 if (Tmp == VTBits)
4538 return 1; // early-out
4539 if (TLI->getExtendForAtomicOps() == ISD::SIGN_EXTEND)
4540 return VTBits - Tmp + 1;
4541 if (TLI->getExtendForAtomicOps() == ISD::ZERO_EXTEND)
4542 return VTBits - Tmp;
4543 }
4544 break;
4545 }
4546 }
4547
4548 // If we are looking at the loaded value of the SDNode.
4549 if (Op.getResNo() == 0) {
4550 // Handle LOADX separately here. EXTLOAD case will fallthrough.
4551 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
4552 unsigned ExtType = LD->getExtensionType();
4553 switch (ExtType) {
4554 default: break;
4555 case ISD::SEXTLOAD: // e.g. i16->i32 = '17' bits known.
4556 Tmp = LD->getMemoryVT().getScalarSizeInBits();
4557 return VTBits - Tmp + 1;
4558 case ISD::ZEXTLOAD: // e.g. i16->i32 = '16' bits known.
4559 Tmp = LD->getMemoryVT().getScalarSizeInBits();
4560 return VTBits - Tmp;
4561 case ISD::NON_EXTLOAD:
4562 if (const Constant *Cst = TLI->getTargetConstantFromLoad(LD)) {
4563 // We only need to handle vectors - computeKnownBits should handle
4564 // scalar cases.
4565 Type *CstTy = Cst->getType();
4566 if (CstTy->isVectorTy() && !VT.isScalableVector() &&
4567 (NumElts * VTBits) == CstTy->getPrimitiveSizeInBits() &&
4568 VTBits == CstTy->getScalarSizeInBits()) {
4569 Tmp = VTBits;
4570 for (unsigned i = 0; i != NumElts; ++i) {
4571 if (!DemandedElts[i])
4572 continue;
4573 if (Constant *Elt = Cst->getAggregateElement(i)) {
4574 if (auto *CInt = dyn_cast<ConstantInt>(Elt)) {
4575 const APInt &Value = CInt->getValue();
4576 Tmp = std::min(Tmp, Value.getNumSignBits());
4577 continue;
4578 }
4579 if (auto *CFP = dyn_cast<ConstantFP>(Elt)) {
4580 APInt Value = CFP->getValueAPF().bitcastToAPInt();
4581 Tmp = std::min(Tmp, Value.getNumSignBits());
4582 continue;
4583 }
4584 }
4585 // Unknown type. Conservatively assume no bits match sign bit.
4586 return 1;
4587 }
4588 return Tmp;
4589 }
4590 }
4591 break;
4592 }
4593 }
4594 }
4595
4596 // Allow the target to implement this method for its nodes.
4597 if (Opcode >= ISD::BUILTIN_OP_END ||
4598 Opcode == ISD::INTRINSIC_WO_CHAIN ||
4599 Opcode == ISD::INTRINSIC_W_CHAIN ||
4600 Opcode == ISD::INTRINSIC_VOID) {
4601 // TODO: This can probably be removed once target code is audited. This
4602 // is here purely to reduce patch size and review complexity.
4603 if (!VT.isScalableVector()) {
4604 unsigned NumBits =
4605 TLI->ComputeNumSignBitsForTargetNode(Op, DemandedElts, *this, Depth);
4606 if (NumBits > 1)
4607 FirstAnswer = std::max(FirstAnswer, NumBits);
4608 }
4609 }
4610
4611 // Finally, if we can prove that the top bits of the result are 0's or 1's,
4612 // use this information.
4613 KnownBits Known = computeKnownBits(Op, DemandedElts, Depth);
4614 return std::max(FirstAnswer, Known.countMinSignBits());
4615 }
4616
ComputeMaxSignificantBits(SDValue Op,unsigned Depth) const4617 unsigned SelectionDAG::ComputeMaxSignificantBits(SDValue Op,
4618 unsigned Depth) const {
4619 unsigned SignBits = ComputeNumSignBits(Op, Depth);
4620 return Op.getScalarValueSizeInBits() - SignBits + 1;
4621 }
4622
ComputeMaxSignificantBits(SDValue Op,const APInt & DemandedElts,unsigned Depth) const4623 unsigned SelectionDAG::ComputeMaxSignificantBits(SDValue Op,
4624 const APInt &DemandedElts,
4625 unsigned Depth) const {
4626 unsigned SignBits = ComputeNumSignBits(Op, DemandedElts, Depth);
4627 return Op.getScalarValueSizeInBits() - SignBits + 1;
4628 }
4629
isGuaranteedNotToBeUndefOrPoison(SDValue Op,bool PoisonOnly,unsigned Depth) const4630 bool SelectionDAG::isGuaranteedNotToBeUndefOrPoison(SDValue Op, bool PoisonOnly,
4631 unsigned Depth) const {
4632 // Early out for FREEZE.
4633 if (Op.getOpcode() == ISD::FREEZE)
4634 return true;
4635
4636 // TODO: Assume we don't know anything for now.
4637 EVT VT = Op.getValueType();
4638 if (VT.isScalableVector())
4639 return false;
4640
4641 APInt DemandedElts = VT.isVector()
4642 ? APInt::getAllOnes(VT.getVectorNumElements())
4643 : APInt(1, 1);
4644 return isGuaranteedNotToBeUndefOrPoison(Op, DemandedElts, PoisonOnly, Depth);
4645 }
4646
isGuaranteedNotToBeUndefOrPoison(SDValue Op,const APInt & DemandedElts,bool PoisonOnly,unsigned Depth) const4647 bool SelectionDAG::isGuaranteedNotToBeUndefOrPoison(SDValue Op,
4648 const APInt &DemandedElts,
4649 bool PoisonOnly,
4650 unsigned Depth) const {
4651 unsigned Opcode = Op.getOpcode();
4652
4653 // Early out for FREEZE.
4654 if (Opcode == ISD::FREEZE)
4655 return true;
4656
4657 if (Depth >= MaxRecursionDepth)
4658 return false; // Limit search depth.
4659
4660 if (isIntOrFPConstant(Op))
4661 return true;
4662
4663 switch (Opcode) {
4664 case ISD::VALUETYPE:
4665 case ISD::FrameIndex:
4666 case ISD::TargetFrameIndex:
4667 return true;
4668
4669 case ISD::UNDEF:
4670 return PoisonOnly;
4671
4672 case ISD::BUILD_VECTOR:
4673 // NOTE: BUILD_VECTOR has implicit truncation of wider scalar elements -
4674 // this shouldn't affect the result.
4675 for (unsigned i = 0, e = Op.getNumOperands(); i < e; ++i) {
4676 if (!DemandedElts[i])
4677 continue;
4678 if (!isGuaranteedNotToBeUndefOrPoison(Op.getOperand(i), PoisonOnly,
4679 Depth + 1))
4680 return false;
4681 }
4682 return true;
4683
4684 // TODO: Search for noundef attributes from library functions.
4685
4686 // TODO: Pointers dereferenced by ISD::LOAD/STORE ops are noundef.
4687
4688 default:
4689 // Allow the target to implement this method for its nodes.
4690 if (Opcode >= ISD::BUILTIN_OP_END || Opcode == ISD::INTRINSIC_WO_CHAIN ||
4691 Opcode == ISD::INTRINSIC_W_CHAIN || Opcode == ISD::INTRINSIC_VOID)
4692 return TLI->isGuaranteedNotToBeUndefOrPoisonForTargetNode(
4693 Op, DemandedElts, *this, PoisonOnly, Depth);
4694 break;
4695 }
4696
4697 // If Op can't create undef/poison and none of its operands are undef/poison
4698 // then Op is never undef/poison.
4699 // NOTE: TargetNodes should handle this in themselves in
4700 // isGuaranteedNotToBeUndefOrPoisonForTargetNode.
4701 return !canCreateUndefOrPoison(Op, PoisonOnly, /*ConsiderFlags*/ true,
4702 Depth) &&
4703 all_of(Op->ops(), [&](SDValue V) {
4704 return isGuaranteedNotToBeUndefOrPoison(V, PoisonOnly, Depth + 1);
4705 });
4706 }
4707
canCreateUndefOrPoison(SDValue Op,bool PoisonOnly,bool ConsiderFlags,unsigned Depth) const4708 bool SelectionDAG::canCreateUndefOrPoison(SDValue Op, bool PoisonOnly,
4709 bool ConsiderFlags,
4710 unsigned Depth) const {
4711 // TODO: Assume we don't know anything for now.
4712 EVT VT = Op.getValueType();
4713 if (VT.isScalableVector())
4714 return true;
4715
4716 APInt DemandedElts = VT.isVector()
4717 ? APInt::getAllOnes(VT.getVectorNumElements())
4718 : APInt(1, 1);
4719 return canCreateUndefOrPoison(Op, DemandedElts, PoisonOnly, ConsiderFlags,
4720 Depth);
4721 }
4722
canCreateUndefOrPoison(SDValue Op,const APInt & DemandedElts,bool PoisonOnly,bool ConsiderFlags,unsigned Depth) const4723 bool SelectionDAG::canCreateUndefOrPoison(SDValue Op, const APInt &DemandedElts,
4724 bool PoisonOnly, bool ConsiderFlags,
4725 unsigned Depth) const {
4726 // TODO: Assume we don't know anything for now.
4727 EVT VT = Op.getValueType();
4728 if (VT.isScalableVector())
4729 return true;
4730
4731 unsigned Opcode = Op.getOpcode();
4732 switch (Opcode) {
4733 case ISD::AssertSext:
4734 case ISD::AssertZext:
4735 case ISD::FREEZE:
4736 case ISD::INSERT_SUBVECTOR:
4737 case ISD::AND:
4738 case ISD::OR:
4739 case ISD::XOR:
4740 case ISD::ROTL:
4741 case ISD::ROTR:
4742 case ISD::FSHL:
4743 case ISD::FSHR:
4744 case ISD::BSWAP:
4745 case ISD::CTPOP:
4746 case ISD::BITREVERSE:
4747 case ISD::PARITY:
4748 case ISD::SIGN_EXTEND:
4749 case ISD::ZERO_EXTEND:
4750 case ISD::TRUNCATE:
4751 case ISD::SIGN_EXTEND_INREG:
4752 case ISD::SIGN_EXTEND_VECTOR_INREG:
4753 case ISD::ZERO_EXTEND_VECTOR_INREG:
4754 case ISD::BITCAST:
4755 case ISD::BUILD_VECTOR:
4756 return false;
4757
4758 case ISD::ADD:
4759 case ISD::SUB:
4760 case ISD::MUL:
4761 // Matches hasPoisonGeneratingFlags().
4762 return ConsiderFlags && (Op->getFlags().hasNoSignedWrap() ||
4763 Op->getFlags().hasNoUnsignedWrap());
4764
4765 case ISD::SHL:
4766 // If the max shift amount isn't in range, then the shift can create poison.
4767 if (!getValidMaximumShiftAmountConstant(Op, DemandedElts))
4768 return true;
4769
4770 // Matches hasPoisonGeneratingFlags().
4771 return ConsiderFlags && (Op->getFlags().hasNoSignedWrap() ||
4772 Op->getFlags().hasNoUnsignedWrap());
4773
4774 default:
4775 // Allow the target to implement this method for its nodes.
4776 if (Opcode >= ISD::BUILTIN_OP_END || Opcode == ISD::INTRINSIC_WO_CHAIN ||
4777 Opcode == ISD::INTRINSIC_W_CHAIN || Opcode == ISD::INTRINSIC_VOID)
4778 return TLI->canCreateUndefOrPoisonForTargetNode(
4779 Op, DemandedElts, *this, PoisonOnly, ConsiderFlags, Depth);
4780 break;
4781 }
4782
4783 // Be conservative and return true.
4784 return true;
4785 }
4786
isBaseWithConstantOffset(SDValue Op) const4787 bool SelectionDAG::isBaseWithConstantOffset(SDValue Op) const {
4788 if ((Op.getOpcode() != ISD::ADD && Op.getOpcode() != ISD::OR) ||
4789 !isa<ConstantSDNode>(Op.getOperand(1)))
4790 return false;
4791
4792 if (Op.getOpcode() == ISD::OR &&
4793 !MaskedValueIsZero(Op.getOperand(0), Op.getConstantOperandAPInt(1)))
4794 return false;
4795
4796 return true;
4797 }
4798
isKnownNeverNaN(SDValue Op,bool SNaN,unsigned Depth) const4799 bool SelectionDAG::isKnownNeverNaN(SDValue Op, bool SNaN, unsigned Depth) const {
4800 // If we're told that NaNs won't happen, assume they won't.
4801 if (getTarget().Options.NoNaNsFPMath || Op->getFlags().hasNoNaNs())
4802 return true;
4803
4804 if (Depth >= MaxRecursionDepth)
4805 return false; // Limit search depth.
4806
4807 // If the value is a constant, we can obviously see if it is a NaN or not.
4808 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op)) {
4809 return !C->getValueAPF().isNaN() ||
4810 (SNaN && !C->getValueAPF().isSignaling());
4811 }
4812
4813 unsigned Opcode = Op.getOpcode();
4814 switch (Opcode) {
4815 case ISD::FADD:
4816 case ISD::FSUB:
4817 case ISD::FMUL:
4818 case ISD::FDIV:
4819 case ISD::FREM:
4820 case ISD::FSIN:
4821 case ISD::FCOS:
4822 case ISD::FMA:
4823 case ISD::FMAD: {
4824 if (SNaN)
4825 return true;
4826 // TODO: Need isKnownNeverInfinity
4827 return false;
4828 }
4829 case ISD::FCANONICALIZE:
4830 case ISD::FEXP:
4831 case ISD::FEXP2:
4832 case ISD::FTRUNC:
4833 case ISD::FFLOOR:
4834 case ISD::FCEIL:
4835 case ISD::FROUND:
4836 case ISD::FROUNDEVEN:
4837 case ISD::FRINT:
4838 case ISD::FNEARBYINT: {
4839 if (SNaN)
4840 return true;
4841 return isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1);
4842 }
4843 case ISD::FABS:
4844 case ISD::FNEG:
4845 case ISD::FCOPYSIGN: {
4846 return isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1);
4847 }
4848 case ISD::SELECT:
4849 return isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1) &&
4850 isKnownNeverNaN(Op.getOperand(2), SNaN, Depth + 1);
4851 case ISD::FP_EXTEND:
4852 case ISD::FP_ROUND: {
4853 if (SNaN)
4854 return true;
4855 return isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1);
4856 }
4857 case ISD::SINT_TO_FP:
4858 case ISD::UINT_TO_FP:
4859 return true;
4860 case ISD::FSQRT: // Need is known positive
4861 case ISD::FLOG:
4862 case ISD::FLOG2:
4863 case ISD::FLOG10:
4864 case ISD::FPOWI:
4865 case ISD::FPOW: {
4866 if (SNaN)
4867 return true;
4868 // TODO: Refine on operand
4869 return false;
4870 }
4871 case ISD::FMINNUM:
4872 case ISD::FMAXNUM: {
4873 // Only one needs to be known not-nan, since it will be returned if the
4874 // other ends up being one.
4875 return isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1) ||
4876 isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1);
4877 }
4878 case ISD::FMINNUM_IEEE:
4879 case ISD::FMAXNUM_IEEE: {
4880 if (SNaN)
4881 return true;
4882 // This can return a NaN if either operand is an sNaN, or if both operands
4883 // are NaN.
4884 return (isKnownNeverNaN(Op.getOperand(0), false, Depth + 1) &&
4885 isKnownNeverSNaN(Op.getOperand(1), Depth + 1)) ||
4886 (isKnownNeverNaN(Op.getOperand(1), false, Depth + 1) &&
4887 isKnownNeverSNaN(Op.getOperand(0), Depth + 1));
4888 }
4889 case ISD::FMINIMUM:
4890 case ISD::FMAXIMUM: {
4891 // TODO: Does this quiet or return the origina NaN as-is?
4892 return isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1) &&
4893 isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1);
4894 }
4895 case ISD::EXTRACT_VECTOR_ELT: {
4896 return isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1);
4897 }
4898 case ISD::BUILD_VECTOR: {
4899 for (const SDValue &Opnd : Op->ops())
4900 if (!isKnownNeverNaN(Opnd, SNaN, Depth + 1))
4901 return false;
4902 return true;
4903 }
4904 default:
4905 if (Opcode >= ISD::BUILTIN_OP_END ||
4906 Opcode == ISD::INTRINSIC_WO_CHAIN ||
4907 Opcode == ISD::INTRINSIC_W_CHAIN ||
4908 Opcode == ISD::INTRINSIC_VOID) {
4909 return TLI->isKnownNeverNaNForTargetNode(Op, *this, SNaN, Depth);
4910 }
4911
4912 return false;
4913 }
4914 }
4915
isKnownNeverZeroFloat(SDValue Op) const4916 bool SelectionDAG::isKnownNeverZeroFloat(SDValue Op) const {
4917 assert(Op.getValueType().isFloatingPoint() &&
4918 "Floating point type expected");
4919
4920 // If the value is a constant, we can obviously see if it is a zero or not.
4921 // TODO: Add BuildVector support.
4922 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op))
4923 return !C->isZero();
4924 return false;
4925 }
4926
isKnownNeverZero(SDValue Op) const4927 bool SelectionDAG::isKnownNeverZero(SDValue Op) const {
4928 assert(!Op.getValueType().isFloatingPoint() &&
4929 "Floating point types unsupported - use isKnownNeverZeroFloat");
4930
4931 // If the value is a constant, we can obviously see if it is a zero or not.
4932 if (ISD::matchUnaryPredicate(Op,
4933 [](ConstantSDNode *C) { return !C->isZero(); }))
4934 return true;
4935
4936 // TODO: Recognize more cases here.
4937 switch (Op.getOpcode()) {
4938 default: break;
4939 case ISD::OR:
4940 if (isKnownNeverZero(Op.getOperand(1)) ||
4941 isKnownNeverZero(Op.getOperand(0)))
4942 return true;
4943 break;
4944 }
4945
4946 return false;
4947 }
4948
isEqualTo(SDValue A,SDValue B) const4949 bool SelectionDAG::isEqualTo(SDValue A, SDValue B) const {
4950 // Check the obvious case.
4951 if (A == B) return true;
4952
4953 // For for negative and positive zero.
4954 if (const ConstantFPSDNode *CA = dyn_cast<ConstantFPSDNode>(A))
4955 if (const ConstantFPSDNode *CB = dyn_cast<ConstantFPSDNode>(B))
4956 if (CA->isZero() && CB->isZero()) return true;
4957
4958 // Otherwise they may not be equal.
4959 return false;
4960 }
4961
4962 // Only bits set in Mask must be negated, other bits may be arbitrary.
getBitwiseNotOperand(SDValue V,SDValue Mask,bool AllowUndefs)4963 SDValue llvm::getBitwiseNotOperand(SDValue V, SDValue Mask, bool AllowUndefs) {
4964 if (isBitwiseNot(V, AllowUndefs))
4965 return V.getOperand(0);
4966
4967 // Handle any_extend (not (truncate X)) pattern, where Mask only sets
4968 // bits in the non-extended part.
4969 ConstantSDNode *MaskC = isConstOrConstSplat(Mask);
4970 if (!MaskC || V.getOpcode() != ISD::ANY_EXTEND)
4971 return SDValue();
4972 SDValue ExtArg = V.getOperand(0);
4973 if (ExtArg.getScalarValueSizeInBits() >=
4974 MaskC->getAPIntValue().getActiveBits() &&
4975 isBitwiseNot(ExtArg, AllowUndefs) &&
4976 ExtArg.getOperand(0).getOpcode() == ISD::TRUNCATE &&
4977 ExtArg.getOperand(0).getOperand(0).getValueType() == V.getValueType())
4978 return ExtArg.getOperand(0).getOperand(0);
4979 return SDValue();
4980 }
4981
haveNoCommonBitsSetCommutative(SDValue A,SDValue B)4982 static bool haveNoCommonBitsSetCommutative(SDValue A, SDValue B) {
4983 // Match masked merge pattern (X & ~M) op (Y & M)
4984 // Including degenerate case (X & ~M) op M
4985 auto MatchNoCommonBitsPattern = [&](SDValue Not, SDValue Mask,
4986 SDValue Other) {
4987 if (SDValue NotOperand =
4988 getBitwiseNotOperand(Not, Mask, /* AllowUndefs */ true)) {
4989 if (Other == NotOperand)
4990 return true;
4991 if (Other->getOpcode() == ISD::AND)
4992 return NotOperand == Other->getOperand(0) ||
4993 NotOperand == Other->getOperand(1);
4994 }
4995 return false;
4996 };
4997 if (A->getOpcode() == ISD::AND)
4998 return MatchNoCommonBitsPattern(A->getOperand(0), A->getOperand(1), B) ||
4999 MatchNoCommonBitsPattern(A->getOperand(1), A->getOperand(0), B);
5000 return false;
5001 }
5002
5003 // FIXME: unify with llvm::haveNoCommonBitsSet.
haveNoCommonBitsSet(SDValue A,SDValue B) const5004 bool SelectionDAG::haveNoCommonBitsSet(SDValue A, SDValue B) const {
5005 assert(A.getValueType() == B.getValueType() &&
5006 "Values must have the same type");
5007 if (haveNoCommonBitsSetCommutative(A, B) ||
5008 haveNoCommonBitsSetCommutative(B, A))
5009 return true;
5010 return KnownBits::haveNoCommonBitsSet(computeKnownBits(A),
5011 computeKnownBits(B));
5012 }
5013
FoldSTEP_VECTOR(const SDLoc & DL,EVT VT,SDValue Step,SelectionDAG & DAG)5014 static SDValue FoldSTEP_VECTOR(const SDLoc &DL, EVT VT, SDValue Step,
5015 SelectionDAG &DAG) {
5016 if (cast<ConstantSDNode>(Step)->isZero())
5017 return DAG.getConstant(0, DL, VT);
5018
5019 return SDValue();
5020 }
5021
FoldBUILD_VECTOR(const SDLoc & DL,EVT VT,ArrayRef<SDValue> Ops,SelectionDAG & DAG)5022 static SDValue FoldBUILD_VECTOR(const SDLoc &DL, EVT VT,
5023 ArrayRef<SDValue> Ops,
5024 SelectionDAG &DAG) {
5025 int NumOps = Ops.size();
5026 assert(NumOps != 0 && "Can't build an empty vector!");
5027 assert(!VT.isScalableVector() &&
5028 "BUILD_VECTOR cannot be used with scalable types");
5029 assert(VT.getVectorNumElements() == (unsigned)NumOps &&
5030 "Incorrect element count in BUILD_VECTOR!");
5031
5032 // BUILD_VECTOR of UNDEFs is UNDEF.
5033 if (llvm::all_of(Ops, [](SDValue Op) { return Op.isUndef(); }))
5034 return DAG.getUNDEF(VT);
5035
5036 // BUILD_VECTOR of seq extract/insert from the same vector + type is Identity.
5037 SDValue IdentitySrc;
5038 bool IsIdentity = true;
5039 for (int i = 0; i != NumOps; ++i) {
5040 if (Ops[i].getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5041 Ops[i].getOperand(0).getValueType() != VT ||
5042 (IdentitySrc && Ops[i].getOperand(0) != IdentitySrc) ||
5043 !isa<ConstantSDNode>(Ops[i].getOperand(1)) ||
5044 cast<ConstantSDNode>(Ops[i].getOperand(1))->getAPIntValue() != i) {
5045 IsIdentity = false;
5046 break;
5047 }
5048 IdentitySrc = Ops[i].getOperand(0);
5049 }
5050 if (IsIdentity)
5051 return IdentitySrc;
5052
5053 return SDValue();
5054 }
5055
5056 /// Try to simplify vector concatenation to an input value, undef, or build
5057 /// vector.
foldCONCAT_VECTORS(const SDLoc & DL,EVT VT,ArrayRef<SDValue> Ops,SelectionDAG & DAG)5058 static SDValue foldCONCAT_VECTORS(const SDLoc &DL, EVT VT,
5059 ArrayRef<SDValue> Ops,
5060 SelectionDAG &DAG) {
5061 assert(!Ops.empty() && "Can't concatenate an empty list of vectors!");
5062 assert(llvm::all_of(Ops,
5063 [Ops](SDValue Op) {
5064 return Ops[0].getValueType() == Op.getValueType();
5065 }) &&
5066 "Concatenation of vectors with inconsistent value types!");
5067 assert((Ops[0].getValueType().getVectorElementCount() * Ops.size()) ==
5068 VT.getVectorElementCount() &&
5069 "Incorrect element count in vector concatenation!");
5070
5071 if (Ops.size() == 1)
5072 return Ops[0];
5073
5074 // Concat of UNDEFs is UNDEF.
5075 if (llvm::all_of(Ops, [](SDValue Op) { return Op.isUndef(); }))
5076 return DAG.getUNDEF(VT);
5077
5078 // Scan the operands and look for extract operations from a single source
5079 // that correspond to insertion at the same location via this concatenation:
5080 // concat (extract X, 0*subvec_elts), (extract X, 1*subvec_elts), ...
5081 SDValue IdentitySrc;
5082 bool IsIdentity = true;
5083 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
5084 SDValue Op = Ops[i];
5085 unsigned IdentityIndex = i * Op.getValueType().getVectorMinNumElements();
5086 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR ||
5087 Op.getOperand(0).getValueType() != VT ||
5088 (IdentitySrc && Op.getOperand(0) != IdentitySrc) ||
5089 Op.getConstantOperandVal(1) != IdentityIndex) {
5090 IsIdentity = false;
5091 break;
5092 }
5093 assert((!IdentitySrc || IdentitySrc == Op.getOperand(0)) &&
5094 "Unexpected identity source vector for concat of extracts");
5095 IdentitySrc = Op.getOperand(0);
5096 }
5097 if (IsIdentity) {
5098 assert(IdentitySrc && "Failed to set source vector of extracts");
5099 return IdentitySrc;
5100 }
5101
5102 // The code below this point is only designed to work for fixed width
5103 // vectors, so we bail out for now.
5104 if (VT.isScalableVector())
5105 return SDValue();
5106
5107 // A CONCAT_VECTOR with all UNDEF/BUILD_VECTOR operands can be
5108 // simplified to one big BUILD_VECTOR.
5109 // FIXME: Add support for SCALAR_TO_VECTOR as well.
5110 EVT SVT = VT.getScalarType();
5111 SmallVector<SDValue, 16> Elts;
5112 for (SDValue Op : Ops) {
5113 EVT OpVT = Op.getValueType();
5114 if (Op.isUndef())
5115 Elts.append(OpVT.getVectorNumElements(), DAG.getUNDEF(SVT));
5116 else if (Op.getOpcode() == ISD::BUILD_VECTOR)
5117 Elts.append(Op->op_begin(), Op->op_end());
5118 else
5119 return SDValue();
5120 }
5121
5122 // BUILD_VECTOR requires all inputs to be of the same type, find the
5123 // maximum type and extend them all.
5124 for (SDValue Op : Elts)
5125 SVT = (SVT.bitsLT(Op.getValueType()) ? Op.getValueType() : SVT);
5126
5127 if (SVT.bitsGT(VT.getScalarType())) {
5128 for (SDValue &Op : Elts) {
5129 if (Op.isUndef())
5130 Op = DAG.getUNDEF(SVT);
5131 else
5132 Op = DAG.getTargetLoweringInfo().isZExtFree(Op.getValueType(), SVT)
5133 ? DAG.getZExtOrTrunc(Op, DL, SVT)
5134 : DAG.getSExtOrTrunc(Op, DL, SVT);
5135 }
5136 }
5137
5138 SDValue V = DAG.getBuildVector(VT, DL, Elts);
5139 NewSDValueDbgMsg(V, "New node fold concat vectors: ", &DAG);
5140 return V;
5141 }
5142
5143 /// Gets or creates the specified node.
getNode(unsigned Opcode,const SDLoc & DL,EVT VT)5144 SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT) {
5145 FoldingSetNodeID ID;
5146 AddNodeIDNode(ID, Opcode, getVTList(VT), std::nullopt);
5147 void *IP = nullptr;
5148 if (SDNode *E = FindNodeOrInsertPos(ID, DL, IP))
5149 return SDValue(E, 0);
5150
5151 auto *N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(),
5152 getVTList(VT));
5153 CSEMap.InsertNode(N, IP);
5154
5155 InsertNode(N);
5156 SDValue V = SDValue(N, 0);
5157 NewSDValueDbgMsg(V, "Creating new node: ", this);
5158 return V;
5159 }
5160
getNode(unsigned Opcode,const SDLoc & DL,EVT VT,SDValue Operand)5161 SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
5162 SDValue Operand) {
5163 SDNodeFlags Flags;
5164 if (Inserter)
5165 Flags = Inserter->getFlags();
5166 return getNode(Opcode, DL, VT, Operand, Flags);
5167 }
5168
getNode(unsigned Opcode,const SDLoc & DL,EVT VT,SDValue Operand,const SDNodeFlags Flags)5169 SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
5170 SDValue Operand, const SDNodeFlags Flags) {
5171 assert(Operand.getOpcode() != ISD::DELETED_NODE &&
5172 "Operand is DELETED_NODE!");
5173 // Constant fold unary operations with an integer constant operand. Even
5174 // opaque constant will be folded, because the folding of unary operations
5175 // doesn't create new constants with different values. Nevertheless, the
5176 // opaque flag is preserved during folding to prevent future folding with
5177 // other constants.
5178 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Operand)) {
5179 const APInt &Val = C->getAPIntValue();
5180 switch (Opcode) {
5181 default: break;
5182 case ISD::SIGN_EXTEND:
5183 return getConstant(Val.sextOrTrunc(VT.getSizeInBits()), DL, VT,
5184 C->isTargetOpcode(), C->isOpaque());
5185 case ISD::TRUNCATE:
5186 if (C->isOpaque())
5187 break;
5188 [[fallthrough]];
5189 case ISD::ZERO_EXTEND:
5190 return getConstant(Val.zextOrTrunc(VT.getSizeInBits()), DL, VT,
5191 C->isTargetOpcode(), C->isOpaque());
5192 case ISD::ANY_EXTEND:
5193 // Some targets like RISCV prefer to sign extend some types.
5194 if (TLI->isSExtCheaperThanZExt(Operand.getValueType(), VT))
5195 return getConstant(Val.sextOrTrunc(VT.getSizeInBits()), DL, VT,
5196 C->isTargetOpcode(), C->isOpaque());
5197 return getConstant(Val.zextOrTrunc(VT.getSizeInBits()), DL, VT,
5198 C->isTargetOpcode(), C->isOpaque());
5199 case ISD::UINT_TO_FP:
5200 case ISD::SINT_TO_FP: {
5201 APFloat apf(EVTToAPFloatSemantics(VT),
5202 APInt::getZero(VT.getSizeInBits()));
5203 (void)apf.convertFromAPInt(Val,
5204 Opcode==ISD::SINT_TO_FP,
5205 APFloat::rmNearestTiesToEven);
5206 return getConstantFP(apf, DL, VT);
5207 }
5208 case ISD::BITCAST:
5209 if (VT == MVT::f16 && C->getValueType(0) == MVT::i16)
5210 return getConstantFP(APFloat(APFloat::IEEEhalf(), Val), DL, VT);
5211 if (VT == MVT::f32 && C->getValueType(0) == MVT::i32)
5212 return getConstantFP(APFloat(APFloat::IEEEsingle(), Val), DL, VT);
5213 if (VT == MVT::f64 && C->getValueType(0) == MVT::i64)
5214 return getConstantFP(APFloat(APFloat::IEEEdouble(), Val), DL, VT);
5215 if (VT == MVT::f128 && C->getValueType(0) == MVT::i128)
5216 return getConstantFP(APFloat(APFloat::IEEEquad(), Val), DL, VT);
5217 break;
5218 case ISD::ABS:
5219 return getConstant(Val.abs(), DL, VT, C->isTargetOpcode(),
5220 C->isOpaque());
5221 case ISD::BITREVERSE:
5222 return getConstant(Val.reverseBits(), DL, VT, C->isTargetOpcode(),
5223 C->isOpaque());
5224 case ISD::BSWAP:
5225 return getConstant(Val.byteSwap(), DL, VT, C->isTargetOpcode(),
5226 C->isOpaque());
5227 case ISD::CTPOP:
5228 return getConstant(Val.countPopulation(), DL, VT, C->isTargetOpcode(),
5229 C->isOpaque());
5230 case ISD::CTLZ:
5231 case ISD::CTLZ_ZERO_UNDEF:
5232 return getConstant(Val.countLeadingZeros(), DL, VT, C->isTargetOpcode(),
5233 C->isOpaque());
5234 case ISD::CTTZ:
5235 case ISD::CTTZ_ZERO_UNDEF:
5236 return getConstant(Val.countTrailingZeros(), DL, VT, C->isTargetOpcode(),
5237 C->isOpaque());
5238 case ISD::FP16_TO_FP:
5239 case ISD::BF16_TO_FP: {
5240 bool Ignored;
5241 APFloat FPV(Opcode == ISD::FP16_TO_FP ? APFloat::IEEEhalf()
5242 : APFloat::BFloat(),
5243 (Val.getBitWidth() == 16) ? Val : Val.trunc(16));
5244
5245 // This can return overflow, underflow, or inexact; we don't care.
5246 // FIXME need to be more flexible about rounding mode.
5247 (void)FPV.convert(EVTToAPFloatSemantics(VT),
5248 APFloat::rmNearestTiesToEven, &Ignored);
5249 return getConstantFP(FPV, DL, VT);
5250 }
5251 case ISD::STEP_VECTOR: {
5252 if (SDValue V = FoldSTEP_VECTOR(DL, VT, Operand, *this))
5253 return V;
5254 break;
5255 }
5256 }
5257 }
5258
5259 // Constant fold unary operations with a floating point constant operand.
5260 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Operand)) {
5261 APFloat V = C->getValueAPF(); // make copy
5262 switch (Opcode) {
5263 case ISD::FNEG:
5264 V.changeSign();
5265 return getConstantFP(V, DL, VT);
5266 case ISD::FABS:
5267 V.clearSign();
5268 return getConstantFP(V, DL, VT);
5269 case ISD::FCEIL: {
5270 APFloat::opStatus fs = V.roundToIntegral(APFloat::rmTowardPositive);
5271 if (fs == APFloat::opOK || fs == APFloat::opInexact)
5272 return getConstantFP(V, DL, VT);
5273 break;
5274 }
5275 case ISD::FTRUNC: {
5276 APFloat::opStatus fs = V.roundToIntegral(APFloat::rmTowardZero);
5277 if (fs == APFloat::opOK || fs == APFloat::opInexact)
5278 return getConstantFP(V, DL, VT);
5279 break;
5280 }
5281 case ISD::FFLOOR: {
5282 APFloat::opStatus fs = V.roundToIntegral(APFloat::rmTowardNegative);
5283 if (fs == APFloat::opOK || fs == APFloat::opInexact)
5284 return getConstantFP(V, DL, VT);
5285 break;
5286 }
5287 case ISD::FP_EXTEND: {
5288 bool ignored;
5289 // This can return overflow, underflow, or inexact; we don't care.
5290 // FIXME need to be more flexible about rounding mode.
5291 (void)V.convert(EVTToAPFloatSemantics(VT),
5292 APFloat::rmNearestTiesToEven, &ignored);
5293 return getConstantFP(V, DL, VT);
5294 }
5295 case ISD::FP_TO_SINT:
5296 case ISD::FP_TO_UINT: {
5297 bool ignored;
5298 APSInt IntVal(VT.getSizeInBits(), Opcode == ISD::FP_TO_UINT);
5299 // FIXME need to be more flexible about rounding mode.
5300 APFloat::opStatus s =
5301 V.convertToInteger(IntVal, APFloat::rmTowardZero, &ignored);
5302 if (s == APFloat::opInvalidOp) // inexact is OK, in fact usual
5303 break;
5304 return getConstant(IntVal, DL, VT);
5305 }
5306 case ISD::BITCAST:
5307 if (VT == MVT::i16 && C->getValueType(0) == MVT::f16)
5308 return getConstant((uint16_t)V.bitcastToAPInt().getZExtValue(), DL, VT);
5309 if (VT == MVT::i16 && C->getValueType(0) == MVT::bf16)
5310 return getConstant((uint16_t)V.bitcastToAPInt().getZExtValue(), DL, VT);
5311 if (VT == MVT::i32 && C->getValueType(0) == MVT::f32)
5312 return getConstant((uint32_t)V.bitcastToAPInt().getZExtValue(), DL, VT);
5313 if (VT == MVT::i64 && C->getValueType(0) == MVT::f64)
5314 return getConstant(V.bitcastToAPInt().getZExtValue(), DL, VT);
5315 break;
5316 case ISD::FP_TO_FP16:
5317 case ISD::FP_TO_BF16: {
5318 bool Ignored;
5319 // This can return overflow, underflow, or inexact; we don't care.
5320 // FIXME need to be more flexible about rounding mode.
5321 (void)V.convert(Opcode == ISD::FP_TO_FP16 ? APFloat::IEEEhalf()
5322 : APFloat::BFloat(),
5323 APFloat::rmNearestTiesToEven, &Ignored);
5324 return getConstant(V.bitcastToAPInt().getZExtValue(), DL, VT);
5325 }
5326 }
5327 }
5328
5329 // Constant fold unary operations with a vector integer or float operand.
5330 switch (Opcode) {
5331 default:
5332 // FIXME: Entirely reasonable to perform folding of other unary
5333 // operations here as the need arises.
5334 break;
5335 case ISD::FNEG:
5336 case ISD::FABS:
5337 case ISD::FCEIL:
5338 case ISD::FTRUNC:
5339 case ISD::FFLOOR:
5340 case ISD::FP_EXTEND:
5341 case ISD::FP_TO_SINT:
5342 case ISD::FP_TO_UINT:
5343 case ISD::TRUNCATE:
5344 case ISD::ANY_EXTEND:
5345 case ISD::ZERO_EXTEND:
5346 case ISD::SIGN_EXTEND:
5347 case ISD::UINT_TO_FP:
5348 case ISD::SINT_TO_FP:
5349 case ISD::ABS:
5350 case ISD::BITREVERSE:
5351 case ISD::BSWAP:
5352 case ISD::CTLZ:
5353 case ISD::CTLZ_ZERO_UNDEF:
5354 case ISD::CTTZ:
5355 case ISD::CTTZ_ZERO_UNDEF:
5356 case ISD::CTPOP: {
5357 SDValue Ops = {Operand};
5358 if (SDValue Fold = FoldConstantArithmetic(Opcode, DL, VT, Ops))
5359 return Fold;
5360 }
5361 }
5362
5363 unsigned OpOpcode = Operand.getNode()->getOpcode();
5364 switch (Opcode) {
5365 case ISD::STEP_VECTOR:
5366 assert(VT.isScalableVector() &&
5367 "STEP_VECTOR can only be used with scalable types");
5368 assert(OpOpcode == ISD::TargetConstant &&
5369 VT.getVectorElementType() == Operand.getValueType() &&
5370 "Unexpected step operand");
5371 break;
5372 case ISD::FREEZE:
5373 assert(VT == Operand.getValueType() && "Unexpected VT!");
5374 if (isGuaranteedNotToBeUndefOrPoison(Operand, /*PoisonOnly*/ false,
5375 /*Depth*/ 1))
5376 return Operand;
5377 break;
5378 case ISD::TokenFactor:
5379 case ISD::MERGE_VALUES:
5380 case ISD::CONCAT_VECTORS:
5381 return Operand; // Factor, merge or concat of one node? No need.
5382 case ISD::BUILD_VECTOR: {
5383 // Attempt to simplify BUILD_VECTOR.
5384 SDValue Ops[] = {Operand};
5385 if (SDValue V = FoldBUILD_VECTOR(DL, VT, Ops, *this))
5386 return V;
5387 break;
5388 }
5389 case ISD::FP_ROUND: llvm_unreachable("Invalid method to make FP_ROUND node");
5390 case ISD::FP_EXTEND:
5391 assert(VT.isFloatingPoint() &&
5392 Operand.getValueType().isFloatingPoint() && "Invalid FP cast!");
5393 if (Operand.getValueType() == VT) return Operand; // noop conversion.
5394 assert((!VT.isVector() ||
5395 VT.getVectorElementCount() ==
5396 Operand.getValueType().getVectorElementCount()) &&
5397 "Vector element count mismatch!");
5398 assert(Operand.getValueType().bitsLT(VT) &&
5399 "Invalid fpext node, dst < src!");
5400 if (Operand.isUndef())
5401 return getUNDEF(VT);
5402 break;
5403 case ISD::FP_TO_SINT:
5404 case ISD::FP_TO_UINT:
5405 if (Operand.isUndef())
5406 return getUNDEF(VT);
5407 break;
5408 case ISD::SINT_TO_FP:
5409 case ISD::UINT_TO_FP:
5410 // [us]itofp(undef) = 0, because the result value is bounded.
5411 if (Operand.isUndef())
5412 return getConstantFP(0.0, DL, VT);
5413 break;
5414 case ISD::SIGN_EXTEND:
5415 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
5416 "Invalid SIGN_EXTEND!");
5417 assert(VT.isVector() == Operand.getValueType().isVector() &&
5418 "SIGN_EXTEND result type type should be vector iff the operand "
5419 "type is vector!");
5420 if (Operand.getValueType() == VT) return Operand; // noop extension
5421 assert((!VT.isVector() ||
5422 VT.getVectorElementCount() ==
5423 Operand.getValueType().getVectorElementCount()) &&
5424 "Vector element count mismatch!");
5425 assert(Operand.getValueType().bitsLT(VT) &&
5426 "Invalid sext node, dst < src!");
5427 if (OpOpcode == ISD::SIGN_EXTEND || OpOpcode == ISD::ZERO_EXTEND)
5428 return getNode(OpOpcode, DL, VT, Operand.getOperand(0));
5429 if (OpOpcode == ISD::UNDEF)
5430 // sext(undef) = 0, because the top bits will all be the same.
5431 return getConstant(0, DL, VT);
5432 break;
5433 case ISD::ZERO_EXTEND:
5434 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
5435 "Invalid ZERO_EXTEND!");
5436 assert(VT.isVector() == Operand.getValueType().isVector() &&
5437 "ZERO_EXTEND result type type should be vector iff the operand "
5438 "type is vector!");
5439 if (Operand.getValueType() == VT) return Operand; // noop extension
5440 assert((!VT.isVector() ||
5441 VT.getVectorElementCount() ==
5442 Operand.getValueType().getVectorElementCount()) &&
5443 "Vector element count mismatch!");
5444 assert(Operand.getValueType().bitsLT(VT) &&
5445 "Invalid zext node, dst < src!");
5446 if (OpOpcode == ISD::ZERO_EXTEND) // (zext (zext x)) -> (zext x)
5447 return getNode(ISD::ZERO_EXTEND, DL, VT, Operand.getOperand(0));
5448 if (OpOpcode == ISD::UNDEF)
5449 // zext(undef) = 0, because the top bits will be zero.
5450 return getConstant(0, DL, VT);
5451 break;
5452 case ISD::ANY_EXTEND:
5453 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
5454 "Invalid ANY_EXTEND!");
5455 assert(VT.isVector() == Operand.getValueType().isVector() &&
5456 "ANY_EXTEND result type type should be vector iff the operand "
5457 "type is vector!");
5458 if (Operand.getValueType() == VT) return Operand; // noop extension
5459 assert((!VT.isVector() ||
5460 VT.getVectorElementCount() ==
5461 Operand.getValueType().getVectorElementCount()) &&
5462 "Vector element count mismatch!");
5463 assert(Operand.getValueType().bitsLT(VT) &&
5464 "Invalid anyext node, dst < src!");
5465
5466 if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND ||
5467 OpOpcode == ISD::ANY_EXTEND)
5468 // (ext (zext x)) -> (zext x) and (ext (sext x)) -> (sext x)
5469 return getNode(OpOpcode, DL, VT, Operand.getOperand(0));
5470 if (OpOpcode == ISD::UNDEF)
5471 return getUNDEF(VT);
5472
5473 // (ext (trunc x)) -> x
5474 if (OpOpcode == ISD::TRUNCATE) {
5475 SDValue OpOp = Operand.getOperand(0);
5476 if (OpOp.getValueType() == VT) {
5477 transferDbgValues(Operand, OpOp);
5478 return OpOp;
5479 }
5480 }
5481 break;
5482 case ISD::TRUNCATE:
5483 assert(VT.isInteger() && Operand.getValueType().isInteger() &&
5484 "Invalid TRUNCATE!");
5485 assert(VT.isVector() == Operand.getValueType().isVector() &&
5486 "TRUNCATE result type type should be vector iff the operand "
5487 "type is vector!");
5488 if (Operand.getValueType() == VT) return Operand; // noop truncate
5489 assert((!VT.isVector() ||
5490 VT.getVectorElementCount() ==
5491 Operand.getValueType().getVectorElementCount()) &&
5492 "Vector element count mismatch!");
5493 assert(Operand.getValueType().bitsGT(VT) &&
5494 "Invalid truncate node, src < dst!");
5495 if (OpOpcode == ISD::TRUNCATE)
5496 return getNode(ISD::TRUNCATE, DL, VT, Operand.getOperand(0));
5497 if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND ||
5498 OpOpcode == ISD::ANY_EXTEND) {
5499 // If the source is smaller than the dest, we still need an extend.
5500 if (Operand.getOperand(0).getValueType().getScalarType()
5501 .bitsLT(VT.getScalarType()))
5502 return getNode(OpOpcode, DL, VT, Operand.getOperand(0));
5503 if (Operand.getOperand(0).getValueType().bitsGT(VT))
5504 return getNode(ISD::TRUNCATE, DL, VT, Operand.getOperand(0));
5505 return Operand.getOperand(0);
5506 }
5507 if (OpOpcode == ISD::UNDEF)
5508 return getUNDEF(VT);
5509 if (OpOpcode == ISD::VSCALE && !NewNodesMustHaveLegalTypes)
5510 return getVScale(DL, VT, Operand.getConstantOperandAPInt(0));
5511 break;
5512 case ISD::ANY_EXTEND_VECTOR_INREG:
5513 case ISD::ZERO_EXTEND_VECTOR_INREG:
5514 case ISD::SIGN_EXTEND_VECTOR_INREG:
5515 assert(VT.isVector() && "This DAG node is restricted to vector types.");
5516 assert(Operand.getValueType().bitsLE(VT) &&
5517 "The input must be the same size or smaller than the result.");
5518 assert(VT.getVectorMinNumElements() <
5519 Operand.getValueType().getVectorMinNumElements() &&
5520 "The destination vector type must have fewer lanes than the input.");
5521 break;
5522 case ISD::ABS:
5523 assert(VT.isInteger() && VT == Operand.getValueType() &&
5524 "Invalid ABS!");
5525 if (OpOpcode == ISD::UNDEF)
5526 return getConstant(0, DL, VT);
5527 break;
5528 case ISD::BSWAP:
5529 assert(VT.isInteger() && VT == Operand.getValueType() &&
5530 "Invalid BSWAP!");
5531 assert((VT.getScalarSizeInBits() % 16 == 0) &&
5532 "BSWAP types must be a multiple of 16 bits!");
5533 if (OpOpcode == ISD::UNDEF)
5534 return getUNDEF(VT);
5535 // bswap(bswap(X)) -> X.
5536 if (OpOpcode == ISD::BSWAP)
5537 return Operand.getOperand(0);
5538 break;
5539 case ISD::BITREVERSE:
5540 assert(VT.isInteger() && VT == Operand.getValueType() &&
5541 "Invalid BITREVERSE!");
5542 if (OpOpcode == ISD::UNDEF)
5543 return getUNDEF(VT);
5544 break;
5545 case ISD::BITCAST:
5546 assert(VT.getSizeInBits() == Operand.getValueSizeInBits() &&
5547 "Cannot BITCAST between types of different sizes!");
5548 if (VT == Operand.getValueType()) return Operand; // noop conversion.
5549 if (OpOpcode == ISD::BITCAST) // bitconv(bitconv(x)) -> bitconv(x)
5550 return getNode(ISD::BITCAST, DL, VT, Operand.getOperand(0));
5551 if (OpOpcode == ISD::UNDEF)
5552 return getUNDEF(VT);
5553 break;
5554 case ISD::SCALAR_TO_VECTOR:
5555 assert(VT.isVector() && !Operand.getValueType().isVector() &&
5556 (VT.getVectorElementType() == Operand.getValueType() ||
5557 (VT.getVectorElementType().isInteger() &&
5558 Operand.getValueType().isInteger() &&
5559 VT.getVectorElementType().bitsLE(Operand.getValueType()))) &&
5560 "Illegal SCALAR_TO_VECTOR node!");
5561 if (OpOpcode == ISD::UNDEF)
5562 return getUNDEF(VT);
5563 // scalar_to_vector(extract_vector_elt V, 0) -> V, top bits are undefined.
5564 if (OpOpcode == ISD::EXTRACT_VECTOR_ELT &&
5565 isa<ConstantSDNode>(Operand.getOperand(1)) &&
5566 Operand.getConstantOperandVal(1) == 0 &&
5567 Operand.getOperand(0).getValueType() == VT)
5568 return Operand.getOperand(0);
5569 break;
5570 case ISD::FNEG:
5571 // Negation of an unknown bag of bits is still completely undefined.
5572 if (OpOpcode == ISD::UNDEF)
5573 return getUNDEF(VT);
5574
5575 if (OpOpcode == ISD::FNEG) // --X -> X
5576 return Operand.getOperand(0);
5577 break;
5578 case ISD::FABS:
5579 if (OpOpcode == ISD::FNEG) // abs(-X) -> abs(X)
5580 return getNode(ISD::FABS, DL, VT, Operand.getOperand(0));
5581 break;
5582 case ISD::VSCALE:
5583 assert(VT == Operand.getValueType() && "Unexpected VT!");
5584 break;
5585 case ISD::CTPOP:
5586 if (Operand.getValueType().getScalarType() == MVT::i1)
5587 return Operand;
5588 break;
5589 case ISD::CTLZ:
5590 case ISD::CTTZ:
5591 if (Operand.getValueType().getScalarType() == MVT::i1)
5592 return getNOT(DL, Operand, Operand.getValueType());
5593 break;
5594 case ISD::VECREDUCE_ADD:
5595 if (Operand.getValueType().getScalarType() == MVT::i1)
5596 return getNode(ISD::VECREDUCE_XOR, DL, VT, Operand);
5597 break;
5598 case ISD::VECREDUCE_SMIN:
5599 case ISD::VECREDUCE_UMAX:
5600 if (Operand.getValueType().getScalarType() == MVT::i1)
5601 return getNode(ISD::VECREDUCE_OR, DL, VT, Operand);
5602 break;
5603 case ISD::VECREDUCE_SMAX:
5604 case ISD::VECREDUCE_UMIN:
5605 if (Operand.getValueType().getScalarType() == MVT::i1)
5606 return getNode(ISD::VECREDUCE_AND, DL, VT, Operand);
5607 break;
5608 }
5609
5610 SDNode *N;
5611 SDVTList VTs = getVTList(VT);
5612 SDValue Ops[] = {Operand};
5613 if (VT != MVT::Glue) { // Don't CSE flag producing nodes
5614 FoldingSetNodeID ID;
5615 AddNodeIDNode(ID, Opcode, VTs, Ops);
5616 void *IP = nullptr;
5617 if (SDNode *E = FindNodeOrInsertPos(ID, DL, IP)) {
5618 E->intersectFlagsWith(Flags);
5619 return SDValue(E, 0);
5620 }
5621
5622 N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(), VTs);
5623 N->setFlags(Flags);
5624 createOperands(N, Ops);
5625 CSEMap.InsertNode(N, IP);
5626 } else {
5627 N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(), VTs);
5628 createOperands(N, Ops);
5629 }
5630
5631 InsertNode(N);
5632 SDValue V = SDValue(N, 0);
5633 NewSDValueDbgMsg(V, "Creating new node: ", this);
5634 return V;
5635 }
5636
FoldValue(unsigned Opcode,const APInt & C1,const APInt & C2)5637 static std::optional<APInt> FoldValue(unsigned Opcode, const APInt &C1,
5638 const APInt &C2) {
5639 switch (Opcode) {
5640 case ISD::ADD: return C1 + C2;
5641 case ISD::SUB: return C1 - C2;
5642 case ISD::MUL: return C1 * C2;
5643 case ISD::AND: return C1 & C2;
5644 case ISD::OR: return C1 | C2;
5645 case ISD::XOR: return C1 ^ C2;
5646 case ISD::SHL: return C1 << C2;
5647 case ISD::SRL: return C1.lshr(C2);
5648 case ISD::SRA: return C1.ashr(C2);
5649 case ISD::ROTL: return C1.rotl(C2);
5650 case ISD::ROTR: return C1.rotr(C2);
5651 case ISD::SMIN: return C1.sle(C2) ? C1 : C2;
5652 case ISD::SMAX: return C1.sge(C2) ? C1 : C2;
5653 case ISD::UMIN: return C1.ule(C2) ? C1 : C2;
5654 case ISD::UMAX: return C1.uge(C2) ? C1 : C2;
5655 case ISD::SADDSAT: return C1.sadd_sat(C2);
5656 case ISD::UADDSAT: return C1.uadd_sat(C2);
5657 case ISD::SSUBSAT: return C1.ssub_sat(C2);
5658 case ISD::USUBSAT: return C1.usub_sat(C2);
5659 case ISD::SSHLSAT: return C1.sshl_sat(C2);
5660 case ISD::USHLSAT: return C1.ushl_sat(C2);
5661 case ISD::UDIV:
5662 if (!C2.getBoolValue())
5663 break;
5664 return C1.udiv(C2);
5665 case ISD::UREM:
5666 if (!C2.getBoolValue())
5667 break;
5668 return C1.urem(C2);
5669 case ISD::SDIV:
5670 if (!C2.getBoolValue())
5671 break;
5672 return C1.sdiv(C2);
5673 case ISD::SREM:
5674 if (!C2.getBoolValue())
5675 break;
5676 return C1.srem(C2);
5677 case ISD::MULHS: {
5678 unsigned FullWidth = C1.getBitWidth() * 2;
5679 APInt C1Ext = C1.sext(FullWidth);
5680 APInt C2Ext = C2.sext(FullWidth);
5681 return (C1Ext * C2Ext).extractBits(C1.getBitWidth(), C1.getBitWidth());
5682 }
5683 case ISD::MULHU: {
5684 unsigned FullWidth = C1.getBitWidth() * 2;
5685 APInt C1Ext = C1.zext(FullWidth);
5686 APInt C2Ext = C2.zext(FullWidth);
5687 return (C1Ext * C2Ext).extractBits(C1.getBitWidth(), C1.getBitWidth());
5688 }
5689 case ISD::AVGFLOORS: {
5690 unsigned FullWidth = C1.getBitWidth() + 1;
5691 APInt C1Ext = C1.sext(FullWidth);
5692 APInt C2Ext = C2.sext(FullWidth);
5693 return (C1Ext + C2Ext).extractBits(C1.getBitWidth(), 1);
5694 }
5695 case ISD::AVGFLOORU: {
5696 unsigned FullWidth = C1.getBitWidth() + 1;
5697 APInt C1Ext = C1.zext(FullWidth);
5698 APInt C2Ext = C2.zext(FullWidth);
5699 return (C1Ext + C2Ext).extractBits(C1.getBitWidth(), 1);
5700 }
5701 case ISD::AVGCEILS: {
5702 unsigned FullWidth = C1.getBitWidth() + 1;
5703 APInt C1Ext = C1.sext(FullWidth);
5704 APInt C2Ext = C2.sext(FullWidth);
5705 return (C1Ext + C2Ext + 1).extractBits(C1.getBitWidth(), 1);
5706 }
5707 case ISD::AVGCEILU: {
5708 unsigned FullWidth = C1.getBitWidth() + 1;
5709 APInt C1Ext = C1.zext(FullWidth);
5710 APInt C2Ext = C2.zext(FullWidth);
5711 return (C1Ext + C2Ext + 1).extractBits(C1.getBitWidth(), 1);
5712 }
5713 }
5714 return std::nullopt;
5715 }
5716
5717 // Handle constant folding with UNDEF.
5718 // TODO: Handle more cases.
FoldValueWithUndef(unsigned Opcode,const APInt & C1,bool IsUndef1,const APInt & C2,bool IsUndef2)5719 static std::optional<APInt> FoldValueWithUndef(unsigned Opcode, const APInt &C1,
5720 bool IsUndef1, const APInt &C2,
5721 bool IsUndef2) {
5722 if (!(IsUndef1 || IsUndef2))
5723 return FoldValue(Opcode, C1, C2);
5724
5725 // Fold and(x, undef) -> 0
5726 // Fold mul(x, undef) -> 0
5727 if (Opcode == ISD::AND || Opcode == ISD::MUL)
5728 return APInt::getZero(C1.getBitWidth());
5729
5730 return std::nullopt;
5731 }
5732
FoldSymbolOffset(unsigned Opcode,EVT VT,const GlobalAddressSDNode * GA,const SDNode * N2)5733 SDValue SelectionDAG::FoldSymbolOffset(unsigned Opcode, EVT VT,
5734 const GlobalAddressSDNode *GA,
5735 const SDNode *N2) {
5736 if (GA->getOpcode() != ISD::GlobalAddress)
5737 return SDValue();
5738 if (!TLI->isOffsetFoldingLegal(GA))
5739 return SDValue();
5740 auto *C2 = dyn_cast<ConstantSDNode>(N2);
5741 if (!C2)
5742 return SDValue();
5743 int64_t Offset = C2->getSExtValue();
5744 switch (Opcode) {
5745 case ISD::ADD: break;
5746 case ISD::SUB: Offset = -uint64_t(Offset); break;
5747 default: return SDValue();
5748 }
5749 return getGlobalAddress(GA->getGlobal(), SDLoc(C2), VT,
5750 GA->getOffset() + uint64_t(Offset));
5751 }
5752
isUndef(unsigned Opcode,ArrayRef<SDValue> Ops)5753 bool SelectionDAG::isUndef(unsigned Opcode, ArrayRef<SDValue> Ops) {
5754 switch (Opcode) {
5755 case ISD::SDIV:
5756 case ISD::UDIV:
5757 case ISD::SREM:
5758 case ISD::UREM: {
5759 // If a divisor is zero/undef or any element of a divisor vector is
5760 // zero/undef, the whole op is undef.
5761 assert(Ops.size() == 2 && "Div/rem should have 2 operands");
5762 SDValue Divisor = Ops[1];
5763 if (Divisor.isUndef() || isNullConstant(Divisor))
5764 return true;
5765
5766 return ISD::isBuildVectorOfConstantSDNodes(Divisor.getNode()) &&
5767 llvm::any_of(Divisor->op_values(),
5768 [](SDValue V) { return V.isUndef() ||
5769 isNullConstant(V); });
5770 // TODO: Handle signed overflow.
5771 }
5772 // TODO: Handle oversized shifts.
5773 default:
5774 return false;
5775 }
5776 }
5777
FoldConstantArithmetic(unsigned Opcode,const SDLoc & DL,EVT VT,ArrayRef<SDValue> Ops)5778 SDValue SelectionDAG::FoldConstantArithmetic(unsigned Opcode, const SDLoc &DL,
5779 EVT VT, ArrayRef<SDValue> Ops) {
5780 // If the opcode is a target-specific ISD node, there's nothing we can
5781 // do here and the operand rules may not line up with the below, so
5782 // bail early.
5783 // We can't create a scalar CONCAT_VECTORS so skip it. It will break
5784 // for concats involving SPLAT_VECTOR. Concats of BUILD_VECTORS are handled by
5785 // foldCONCAT_VECTORS in getNode before this is called.
5786 if (Opcode >= ISD::BUILTIN_OP_END || Opcode == ISD::CONCAT_VECTORS)
5787 return SDValue();
5788
5789 unsigned NumOps = Ops.size();
5790 if (NumOps == 0)
5791 return SDValue();
5792
5793 if (isUndef(Opcode, Ops))
5794 return getUNDEF(VT);
5795
5796 // Handle binops special cases.
5797 if (NumOps == 2) {
5798 if (SDValue CFP = foldConstantFPMath(Opcode, DL, VT, Ops[0], Ops[1]))
5799 return CFP;
5800
5801 if (auto *C1 = dyn_cast<ConstantSDNode>(Ops[0])) {
5802 if (auto *C2 = dyn_cast<ConstantSDNode>(Ops[1])) {
5803 if (C1->isOpaque() || C2->isOpaque())
5804 return SDValue();
5805
5806 std::optional<APInt> FoldAttempt =
5807 FoldValue(Opcode, C1->getAPIntValue(), C2->getAPIntValue());
5808 if (!FoldAttempt)
5809 return SDValue();
5810
5811 SDValue Folded = getConstant(*FoldAttempt, DL, VT);
5812 assert((!Folded || !VT.isVector()) &&
5813 "Can't fold vectors ops with scalar operands");
5814 return Folded;
5815 }
5816 }
5817
5818 // fold (add Sym, c) -> Sym+c
5819 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Ops[0]))
5820 return FoldSymbolOffset(Opcode, VT, GA, Ops[1].getNode());
5821 if (TLI->isCommutativeBinOp(Opcode))
5822 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Ops[1]))
5823 return FoldSymbolOffset(Opcode, VT, GA, Ops[0].getNode());
5824 }
5825
5826 // This is for vector folding only from here on.
5827 if (!VT.isVector())
5828 return SDValue();
5829
5830 ElementCount NumElts = VT.getVectorElementCount();
5831
5832 // See if we can fold through bitcasted integer ops.
5833 if (NumOps == 2 && VT.isFixedLengthVector() && VT.isInteger() &&
5834 Ops[0].getValueType() == VT && Ops[1].getValueType() == VT &&
5835 Ops[0].getOpcode() == ISD::BITCAST &&
5836 Ops[1].getOpcode() == ISD::BITCAST) {
5837 SDValue N1 = peekThroughBitcasts(Ops[0]);
5838 SDValue N2 = peekThroughBitcasts(Ops[1]);
5839 auto *BV1 = dyn_cast<BuildVectorSDNode>(N1);
5840 auto *BV2 = dyn_cast<BuildVectorSDNode>(N2);
5841 EVT BVVT = N1.getValueType();
5842 if (BV1 && BV2 && BVVT.isInteger() && BVVT == N2.getValueType()) {
5843 bool IsLE = getDataLayout().isLittleEndian();
5844 unsigned EltBits = VT.getScalarSizeInBits();
5845 SmallVector<APInt> RawBits1, RawBits2;
5846 BitVector UndefElts1, UndefElts2;
5847 if (BV1->getConstantRawBits(IsLE, EltBits, RawBits1, UndefElts1) &&
5848 BV2->getConstantRawBits(IsLE, EltBits, RawBits2, UndefElts2)) {
5849 SmallVector<APInt> RawBits;
5850 for (unsigned I = 0, E = NumElts.getFixedValue(); I != E; ++I) {
5851 std::optional<APInt> Fold = FoldValueWithUndef(
5852 Opcode, RawBits1[I], UndefElts1[I], RawBits2[I], UndefElts2[I]);
5853 if (!Fold)
5854 break;
5855 RawBits.push_back(*Fold);
5856 }
5857 if (RawBits.size() == NumElts.getFixedValue()) {
5858 // We have constant folded, but we need to cast this again back to
5859 // the original (possibly legalized) type.
5860 SmallVector<APInt> DstBits;
5861 BitVector DstUndefs;
5862 BuildVectorSDNode::recastRawBits(IsLE, BVVT.getScalarSizeInBits(),
5863 DstBits, RawBits, DstUndefs,
5864 BitVector(RawBits.size(), false));
5865 EVT BVEltVT = BV1->getOperand(0).getValueType();
5866 unsigned BVEltBits = BVEltVT.getSizeInBits();
5867 SmallVector<SDValue> Ops(DstBits.size(), getUNDEF(BVEltVT));
5868 for (unsigned I = 0, E = DstBits.size(); I != E; ++I) {
5869 if (DstUndefs[I])
5870 continue;
5871 Ops[I] = getConstant(DstBits[I].sext(BVEltBits), DL, BVEltVT);
5872 }
5873 return getBitcast(VT, getBuildVector(BVVT, DL, Ops));
5874 }
5875 }
5876 }
5877 }
5878
5879 // Fold (mul step_vector(C0), C1) to (step_vector(C0 * C1)).
5880 // (shl step_vector(C0), C1) -> (step_vector(C0 << C1))
5881 if ((Opcode == ISD::MUL || Opcode == ISD::SHL) &&
5882 Ops[0].getOpcode() == ISD::STEP_VECTOR) {
5883 APInt RHSVal;
5884 if (ISD::isConstantSplatVector(Ops[1].getNode(), RHSVal)) {
5885 APInt NewStep = Opcode == ISD::MUL
5886 ? Ops[0].getConstantOperandAPInt(0) * RHSVal
5887 : Ops[0].getConstantOperandAPInt(0) << RHSVal;
5888 return getStepVector(DL, VT, NewStep);
5889 }
5890 }
5891
5892 auto IsScalarOrSameVectorSize = [NumElts](const SDValue &Op) {
5893 return !Op.getValueType().isVector() ||
5894 Op.getValueType().getVectorElementCount() == NumElts;
5895 };
5896
5897 auto IsBuildVectorSplatVectorOrUndef = [](const SDValue &Op) {
5898 return Op.isUndef() || Op.getOpcode() == ISD::CONDCODE ||
5899 Op.getOpcode() == ISD::BUILD_VECTOR ||
5900 Op.getOpcode() == ISD::SPLAT_VECTOR;
5901 };
5902
5903 // All operands must be vector types with the same number of elements as
5904 // the result type and must be either UNDEF or a build/splat vector
5905 // or UNDEF scalars.
5906 if (!llvm::all_of(Ops, IsBuildVectorSplatVectorOrUndef) ||
5907 !llvm::all_of(Ops, IsScalarOrSameVectorSize))
5908 return SDValue();
5909
5910 // If we are comparing vectors, then the result needs to be a i1 boolean that
5911 // is then extended back to the legal result type depending on how booleans
5912 // are represented.
5913 EVT SVT = (Opcode == ISD::SETCC ? MVT::i1 : VT.getScalarType());
5914 ISD::NodeType ExtendCode =
5915 (Opcode == ISD::SETCC && SVT != VT.getScalarType())
5916 ? TargetLowering::getExtendForContent(TLI->getBooleanContents(VT))
5917 : ISD::SIGN_EXTEND;
5918
5919 // Find legal integer scalar type for constant promotion and
5920 // ensure that its scalar size is at least as large as source.
5921 EVT LegalSVT = VT.getScalarType();
5922 if (NewNodesMustHaveLegalTypes && LegalSVT.isInteger()) {
5923 LegalSVT = TLI->getTypeToTransformTo(*getContext(), LegalSVT);
5924 if (LegalSVT.bitsLT(VT.getScalarType()))
5925 return SDValue();
5926 }
5927
5928 // For scalable vector types we know we're dealing with SPLAT_VECTORs. We
5929 // only have one operand to check. For fixed-length vector types we may have
5930 // a combination of BUILD_VECTOR and SPLAT_VECTOR.
5931 unsigned NumVectorElts = NumElts.isScalable() ? 1 : NumElts.getFixedValue();
5932
5933 // Constant fold each scalar lane separately.
5934 SmallVector<SDValue, 4> ScalarResults;
5935 for (unsigned I = 0; I != NumVectorElts; I++) {
5936 SmallVector<SDValue, 4> ScalarOps;
5937 for (SDValue Op : Ops) {
5938 EVT InSVT = Op.getValueType().getScalarType();
5939 if (Op.getOpcode() != ISD::BUILD_VECTOR &&
5940 Op.getOpcode() != ISD::SPLAT_VECTOR) {
5941 if (Op.isUndef())
5942 ScalarOps.push_back(getUNDEF(InSVT));
5943 else
5944 ScalarOps.push_back(Op);
5945 continue;
5946 }
5947
5948 SDValue ScalarOp =
5949 Op.getOperand(Op.getOpcode() == ISD::SPLAT_VECTOR ? 0 : I);
5950 EVT ScalarVT = ScalarOp.getValueType();
5951
5952 // Build vector (integer) scalar operands may need implicit
5953 // truncation - do this before constant folding.
5954 if (ScalarVT.isInteger() && ScalarVT.bitsGT(InSVT)) {
5955 // Don't create illegally-typed nodes unless they're constants or undef
5956 // - if we fail to constant fold we can't guarantee the (dead) nodes
5957 // we're creating will be cleaned up before being visited for
5958 // legalization.
5959 if (NewNodesMustHaveLegalTypes && !ScalarOp.isUndef() &&
5960 !isa<ConstantSDNode>(ScalarOp) &&
5961 TLI->getTypeAction(*getContext(), InSVT) !=
5962 TargetLowering::TypeLegal)
5963 return SDValue();
5964 ScalarOp = getNode(ISD::TRUNCATE, DL, InSVT, ScalarOp);
5965 }
5966
5967 ScalarOps.push_back(ScalarOp);
5968 }
5969
5970 // Constant fold the scalar operands.
5971 SDValue ScalarResult = getNode(Opcode, DL, SVT, ScalarOps);
5972
5973 // Legalize the (integer) scalar constant if necessary.
5974 if (LegalSVT != SVT)
5975 ScalarResult = getNode(ExtendCode, DL, LegalSVT, ScalarResult);
5976
5977 // Scalar folding only succeeded if the result is a constant or UNDEF.
5978 if (!ScalarResult.isUndef() && ScalarResult.getOpcode() != ISD::Constant &&
5979 ScalarResult.getOpcode() != ISD::ConstantFP)
5980 return SDValue();
5981 ScalarResults.push_back(ScalarResult);
5982 }
5983
5984 SDValue V = NumElts.isScalable() ? getSplatVector(VT, DL, ScalarResults[0])
5985 : getBuildVector(VT, DL, ScalarResults);
5986 NewSDValueDbgMsg(V, "New node fold constant vector: ", this);
5987 return V;
5988 }
5989
foldConstantFPMath(unsigned Opcode,const SDLoc & DL,EVT VT,SDValue N1,SDValue N2)5990 SDValue SelectionDAG::foldConstantFPMath(unsigned Opcode, const SDLoc &DL,
5991 EVT VT, SDValue N1, SDValue N2) {
5992 // TODO: We don't do any constant folding for strict FP opcodes here, but we
5993 // should. That will require dealing with a potentially non-default
5994 // rounding mode, checking the "opStatus" return value from the APFloat
5995 // math calculations, and possibly other variations.
5996 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1, /*AllowUndefs*/ false);
5997 ConstantFPSDNode *N2CFP = isConstOrConstSplatFP(N2, /*AllowUndefs*/ false);
5998 if (N1CFP && N2CFP) {
5999 APFloat C1 = N1CFP->getValueAPF(); // make copy
6000 const APFloat &C2 = N2CFP->getValueAPF();
6001 switch (Opcode) {
6002 case ISD::FADD:
6003 C1.add(C2, APFloat::rmNearestTiesToEven);
6004 return getConstantFP(C1, DL, VT);
6005 case ISD::FSUB:
6006 C1.subtract(C2, APFloat::rmNearestTiesToEven);
6007 return getConstantFP(C1, DL, VT);
6008 case ISD::FMUL:
6009 C1.multiply(C2, APFloat::rmNearestTiesToEven);
6010 return getConstantFP(C1, DL, VT);
6011 case ISD::FDIV:
6012 C1.divide(C2, APFloat::rmNearestTiesToEven);
6013 return getConstantFP(C1, DL, VT);
6014 case ISD::FREM:
6015 C1.mod(C2);
6016 return getConstantFP(C1, DL, VT);
6017 case ISD::FCOPYSIGN:
6018 C1.copySign(C2);
6019 return getConstantFP(C1, DL, VT);
6020 case ISD::FMINNUM:
6021 return getConstantFP(minnum(C1, C2), DL, VT);
6022 case ISD::FMAXNUM:
6023 return getConstantFP(maxnum(C1, C2), DL, VT);
6024 case ISD::FMINIMUM:
6025 return getConstantFP(minimum(C1, C2), DL, VT);
6026 case ISD::FMAXIMUM:
6027 return getConstantFP(maximum(C1, C2), DL, VT);
6028 default: break;
6029 }
6030 }
6031 if (N1CFP && Opcode == ISD::FP_ROUND) {
6032 APFloat C1 = N1CFP->getValueAPF(); // make copy
6033 bool Unused;
6034 // This can return overflow, underflow, or inexact; we don't care.
6035 // FIXME need to be more flexible about rounding mode.
6036 (void) C1.convert(EVTToAPFloatSemantics(VT), APFloat::rmNearestTiesToEven,
6037 &Unused);
6038 return getConstantFP(C1, DL, VT);
6039 }
6040
6041 switch (Opcode) {
6042 case ISD::FSUB:
6043 // -0.0 - undef --> undef (consistent with "fneg undef")
6044 if (ConstantFPSDNode *N1C = isConstOrConstSplatFP(N1, /*AllowUndefs*/ true))
6045 if (N1C && N1C->getValueAPF().isNegZero() && N2.isUndef())
6046 return getUNDEF(VT);
6047 [[fallthrough]];
6048
6049 case ISD::FADD:
6050 case ISD::FMUL:
6051 case ISD::FDIV:
6052 case ISD::FREM:
6053 // If both operands are undef, the result is undef. If 1 operand is undef,
6054 // the result is NaN. This should match the behavior of the IR optimizer.
6055 if (N1.isUndef() && N2.isUndef())
6056 return getUNDEF(VT);
6057 if (N1.isUndef() || N2.isUndef())
6058 return getConstantFP(APFloat::getNaN(EVTToAPFloatSemantics(VT)), DL, VT);
6059 }
6060 return SDValue();
6061 }
6062
getAssertAlign(const SDLoc & DL,SDValue Val,Align A)6063 SDValue SelectionDAG::getAssertAlign(const SDLoc &DL, SDValue Val, Align A) {
6064 assert(Val.getValueType().isInteger() && "Invalid AssertAlign!");
6065
6066 // There's no need to assert on a byte-aligned pointer. All pointers are at
6067 // least byte aligned.
6068 if (A == Align(1))
6069 return Val;
6070
6071 FoldingSetNodeID ID;
6072 AddNodeIDNode(ID, ISD::AssertAlign, getVTList(Val.getValueType()), {Val});
6073 ID.AddInteger(A.value());
6074
6075 void *IP = nullptr;
6076 if (SDNode *E = FindNodeOrInsertPos(ID, DL, IP))
6077 return SDValue(E, 0);
6078
6079 auto *N = newSDNode<AssertAlignSDNode>(DL.getIROrder(), DL.getDebugLoc(),
6080 Val.getValueType(), A);
6081 createOperands(N, {Val});
6082
6083 CSEMap.InsertNode(N, IP);
6084 InsertNode(N);
6085
6086 SDValue V(N, 0);
6087 NewSDValueDbgMsg(V, "Creating new node: ", this);
6088 return V;
6089 }
6090
getNode(unsigned Opcode,const SDLoc & DL,EVT VT,SDValue N1,SDValue N2)6091 SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
6092 SDValue N1, SDValue N2) {
6093 SDNodeFlags Flags;
6094 if (Inserter)
6095 Flags = Inserter->getFlags();
6096 return getNode(Opcode, DL, VT, N1, N2, Flags);
6097 }
6098
canonicalizeCommutativeBinop(unsigned Opcode,SDValue & N1,SDValue & N2) const6099 void SelectionDAG::canonicalizeCommutativeBinop(unsigned Opcode, SDValue &N1,
6100 SDValue &N2) const {
6101 if (!TLI->isCommutativeBinOp(Opcode))
6102 return;
6103
6104 // Canonicalize:
6105 // binop(const, nonconst) -> binop(nonconst, const)
6106 SDNode *N1C = isConstantIntBuildVectorOrConstantInt(N1);
6107 SDNode *N2C = isConstantIntBuildVectorOrConstantInt(N2);
6108 SDNode *N1CFP = isConstantFPBuildVectorOrConstantFP(N1);
6109 SDNode *N2CFP = isConstantFPBuildVectorOrConstantFP(N2);
6110 if ((N1C && !N2C) || (N1CFP && !N2CFP))
6111 std::swap(N1, N2);
6112
6113 // Canonicalize:
6114 // binop(splat(x), step_vector) -> binop(step_vector, splat(x))
6115 else if (N1.getOpcode() == ISD::SPLAT_VECTOR &&
6116 N2.getOpcode() == ISD::STEP_VECTOR)
6117 std::swap(N1, N2);
6118 }
6119
getNode(unsigned Opcode,const SDLoc & DL,EVT VT,SDValue N1,SDValue N2,const SDNodeFlags Flags)6120 SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
6121 SDValue N1, SDValue N2, const SDNodeFlags Flags) {
6122 assert(N1.getOpcode() != ISD::DELETED_NODE &&
6123 N2.getOpcode() != ISD::DELETED_NODE &&
6124 "Operand is DELETED_NODE!");
6125
6126 canonicalizeCommutativeBinop(Opcode, N1, N2);
6127
6128 auto *N1C = dyn_cast<ConstantSDNode>(N1);
6129 auto *N2C = dyn_cast<ConstantSDNode>(N2);
6130
6131 // Don't allow undefs in vector splats - we might be returning N2 when folding
6132 // to zero etc.
6133 ConstantSDNode *N2CV =
6134 isConstOrConstSplat(N2, /*AllowUndefs*/ false, /*AllowTruncation*/ true);
6135
6136 switch (Opcode) {
6137 default: break;
6138 case ISD::TokenFactor:
6139 assert(VT == MVT::Other && N1.getValueType() == MVT::Other &&
6140 N2.getValueType() == MVT::Other && "Invalid token factor!");
6141 // Fold trivial token factors.
6142 if (N1.getOpcode() == ISD::EntryToken) return N2;
6143 if (N2.getOpcode() == ISD::EntryToken) return N1;
6144 if (N1 == N2) return N1;
6145 break;
6146 case ISD::BUILD_VECTOR: {
6147 // Attempt to simplify BUILD_VECTOR.
6148 SDValue Ops[] = {N1, N2};
6149 if (SDValue V = FoldBUILD_VECTOR(DL, VT, Ops, *this))
6150 return V;
6151 break;
6152 }
6153 case ISD::CONCAT_VECTORS: {
6154 SDValue Ops[] = {N1, N2};
6155 if (SDValue V = foldCONCAT_VECTORS(DL, VT, Ops, *this))
6156 return V;
6157 break;
6158 }
6159 case ISD::AND:
6160 assert(VT.isInteger() && "This operator does not apply to FP types!");
6161 assert(N1.getValueType() == N2.getValueType() &&
6162 N1.getValueType() == VT && "Binary operator types must match!");
6163 // (X & 0) -> 0. This commonly occurs when legalizing i64 values, so it's
6164 // worth handling here.
6165 if (N2CV && N2CV->isZero())
6166 return N2;
6167 if (N2CV && N2CV->isAllOnes()) // X & -1 -> X
6168 return N1;
6169 break;
6170 case ISD::OR:
6171 case ISD::XOR:
6172 case ISD::ADD:
6173 case ISD::SUB:
6174 assert(VT.isInteger() && "This operator does not apply to FP types!");
6175 assert(N1.getValueType() == N2.getValueType() &&
6176 N1.getValueType() == VT && "Binary operator types must match!");
6177 // (X ^|+- 0) -> X. This commonly occurs when legalizing i64 values, so
6178 // it's worth handling here.
6179 if (N2CV && N2CV->isZero())
6180 return N1;
6181 if ((Opcode == ISD::ADD || Opcode == ISD::SUB) && VT.isVector() &&
6182 VT.getVectorElementType() == MVT::i1)
6183 return getNode(ISD::XOR, DL, VT, N1, N2);
6184 break;
6185 case ISD::MUL:
6186 assert(VT.isInteger() && "This operator does not apply to FP types!");
6187 assert(N1.getValueType() == N2.getValueType() &&
6188 N1.getValueType() == VT && "Binary operator types must match!");
6189 if (VT.isVector() && VT.getVectorElementType() == MVT::i1)
6190 return getNode(ISD::AND, DL, VT, N1, N2);
6191 if (N2C && (N1.getOpcode() == ISD::VSCALE) && Flags.hasNoSignedWrap()) {
6192 const APInt &MulImm = N1->getConstantOperandAPInt(0);
6193 const APInt &N2CImm = N2C->getAPIntValue();
6194 return getVScale(DL, VT, MulImm * N2CImm);
6195 }
6196 break;
6197 case ISD::UDIV:
6198 case ISD::UREM:
6199 case ISD::MULHU:
6200 case ISD::MULHS:
6201 case ISD::SDIV:
6202 case ISD::SREM:
6203 case ISD::SADDSAT:
6204 case ISD::SSUBSAT:
6205 case ISD::UADDSAT:
6206 case ISD::USUBSAT:
6207 assert(VT.isInteger() && "This operator does not apply to FP types!");
6208 assert(N1.getValueType() == N2.getValueType() &&
6209 N1.getValueType() == VT && "Binary operator types must match!");
6210 if (VT.isVector() && VT.getVectorElementType() == MVT::i1) {
6211 // fold (add_sat x, y) -> (or x, y) for bool types.
6212 if (Opcode == ISD::SADDSAT || Opcode == ISD::UADDSAT)
6213 return getNode(ISD::OR, DL, VT, N1, N2);
6214 // fold (sub_sat x, y) -> (and x, ~y) for bool types.
6215 if (Opcode == ISD::SSUBSAT || Opcode == ISD::USUBSAT)
6216 return getNode(ISD::AND, DL, VT, N1, getNOT(DL, N2, VT));
6217 }
6218 break;
6219 case ISD::ABDS:
6220 case ISD::ABDU:
6221 assert(VT.isInteger() && "This operator does not apply to FP types!");
6222 assert(N1.getValueType() == N2.getValueType() &&
6223 N1.getValueType() == VT && "Binary operator types must match!");
6224 break;
6225 case ISD::SMIN:
6226 case ISD::UMAX:
6227 assert(VT.isInteger() && "This operator does not apply to FP types!");
6228 assert(N1.getValueType() == N2.getValueType() &&
6229 N1.getValueType() == VT && "Binary operator types must match!");
6230 if (VT.isVector() && VT.getVectorElementType() == MVT::i1)
6231 return getNode(ISD::OR, DL, VT, N1, N2);
6232 break;
6233 case ISD::SMAX:
6234 case ISD::UMIN:
6235 assert(VT.isInteger() && "This operator does not apply to FP types!");
6236 assert(N1.getValueType() == N2.getValueType() &&
6237 N1.getValueType() == VT && "Binary operator types must match!");
6238 if (VT.isVector() && VT.getVectorElementType() == MVT::i1)
6239 return getNode(ISD::AND, DL, VT, N1, N2);
6240 break;
6241 case ISD::FADD:
6242 case ISD::FSUB:
6243 case ISD::FMUL:
6244 case ISD::FDIV:
6245 case ISD::FREM:
6246 assert(VT.isFloatingPoint() && "This operator only applies to FP types!");
6247 assert(N1.getValueType() == N2.getValueType() &&
6248 N1.getValueType() == VT && "Binary operator types must match!");
6249 if (SDValue V = simplifyFPBinop(Opcode, N1, N2, Flags))
6250 return V;
6251 break;
6252 case ISD::FCOPYSIGN: // N1 and result must match. N1/N2 need not match.
6253 assert(N1.getValueType() == VT &&
6254 N1.getValueType().isFloatingPoint() &&
6255 N2.getValueType().isFloatingPoint() &&
6256 "Invalid FCOPYSIGN!");
6257 break;
6258 case ISD::SHL:
6259 if (N2C && (N1.getOpcode() == ISD::VSCALE) && Flags.hasNoSignedWrap()) {
6260 const APInt &MulImm = N1->getConstantOperandAPInt(0);
6261 const APInt &ShiftImm = N2C->getAPIntValue();
6262 return getVScale(DL, VT, MulImm << ShiftImm);
6263 }
6264 [[fallthrough]];
6265 case ISD::SRA:
6266 case ISD::SRL:
6267 if (SDValue V = simplifyShift(N1, N2))
6268 return V;
6269 [[fallthrough]];
6270 case ISD::ROTL:
6271 case ISD::ROTR:
6272 assert(VT == N1.getValueType() &&
6273 "Shift operators return type must be the same as their first arg");
6274 assert(VT.isInteger() && N2.getValueType().isInteger() &&
6275 "Shifts only work on integers");
6276 assert((!VT.isVector() || VT == N2.getValueType()) &&
6277 "Vector shift amounts must be in the same as their first arg");
6278 // Verify that the shift amount VT is big enough to hold valid shift
6279 // amounts. This catches things like trying to shift an i1024 value by an
6280 // i8, which is easy to fall into in generic code that uses
6281 // TLI.getShiftAmount().
6282 assert(N2.getValueType().getScalarSizeInBits() >=
6283 Log2_32_Ceil(VT.getScalarSizeInBits()) &&
6284 "Invalid use of small shift amount with oversized value!");
6285
6286 // Always fold shifts of i1 values so the code generator doesn't need to
6287 // handle them. Since we know the size of the shift has to be less than the
6288 // size of the value, the shift/rotate count is guaranteed to be zero.
6289 if (VT == MVT::i1)
6290 return N1;
6291 if (N2CV && N2CV->isZero())
6292 return N1;
6293 break;
6294 case ISD::FP_ROUND:
6295 assert(VT.isFloatingPoint() &&
6296 N1.getValueType().isFloatingPoint() &&
6297 VT.bitsLE(N1.getValueType()) &&
6298 N2C && (N2C->getZExtValue() == 0 || N2C->getZExtValue() == 1) &&
6299 "Invalid FP_ROUND!");
6300 if (N1.getValueType() == VT) return N1; // noop conversion.
6301 break;
6302 case ISD::AssertSext:
6303 case ISD::AssertZext: {
6304 EVT EVT = cast<VTSDNode>(N2)->getVT();
6305 assert(VT == N1.getValueType() && "Not an inreg extend!");
6306 assert(VT.isInteger() && EVT.isInteger() &&
6307 "Cannot *_EXTEND_INREG FP types");
6308 assert(!EVT.isVector() &&
6309 "AssertSExt/AssertZExt type should be the vector element type "
6310 "rather than the vector type!");
6311 assert(EVT.bitsLE(VT.getScalarType()) && "Not extending!");
6312 if (VT.getScalarType() == EVT) return N1; // noop assertion.
6313 break;
6314 }
6315 case ISD::SIGN_EXTEND_INREG: {
6316 EVT EVT = cast<VTSDNode>(N2)->getVT();
6317 assert(VT == N1.getValueType() && "Not an inreg extend!");
6318 assert(VT.isInteger() && EVT.isInteger() &&
6319 "Cannot *_EXTEND_INREG FP types");
6320 assert(EVT.isVector() == VT.isVector() &&
6321 "SIGN_EXTEND_INREG type should be vector iff the operand "
6322 "type is vector!");
6323 assert((!EVT.isVector() ||
6324 EVT.getVectorElementCount() == VT.getVectorElementCount()) &&
6325 "Vector element counts must match in SIGN_EXTEND_INREG");
6326 assert(EVT.bitsLE(VT) && "Not extending!");
6327 if (EVT == VT) return N1; // Not actually extending
6328
6329 auto SignExtendInReg = [&](APInt Val, llvm::EVT ConstantVT) {
6330 unsigned FromBits = EVT.getScalarSizeInBits();
6331 Val <<= Val.getBitWidth() - FromBits;
6332 Val.ashrInPlace(Val.getBitWidth() - FromBits);
6333 return getConstant(Val, DL, ConstantVT);
6334 };
6335
6336 if (N1C) {
6337 const APInt &Val = N1C->getAPIntValue();
6338 return SignExtendInReg(Val, VT);
6339 }
6340
6341 if (ISD::isBuildVectorOfConstantSDNodes(N1.getNode())) {
6342 SmallVector<SDValue, 8> Ops;
6343 llvm::EVT OpVT = N1.getOperand(0).getValueType();
6344 for (int i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
6345 SDValue Op = N1.getOperand(i);
6346 if (Op.isUndef()) {
6347 Ops.push_back(getUNDEF(OpVT));
6348 continue;
6349 }
6350 ConstantSDNode *C = cast<ConstantSDNode>(Op);
6351 APInt Val = C->getAPIntValue();
6352 Ops.push_back(SignExtendInReg(Val, OpVT));
6353 }
6354 return getBuildVector(VT, DL, Ops);
6355 }
6356 break;
6357 }
6358 case ISD::FP_TO_SINT_SAT:
6359 case ISD::FP_TO_UINT_SAT: {
6360 assert(VT.isInteger() && cast<VTSDNode>(N2)->getVT().isInteger() &&
6361 N1.getValueType().isFloatingPoint() && "Invalid FP_TO_*INT_SAT");
6362 assert(N1.getValueType().isVector() == VT.isVector() &&
6363 "FP_TO_*INT_SAT type should be vector iff the operand type is "
6364 "vector!");
6365 assert((!VT.isVector() || VT.getVectorElementCount() ==
6366 N1.getValueType().getVectorElementCount()) &&
6367 "Vector element counts must match in FP_TO_*INT_SAT");
6368 assert(!cast<VTSDNode>(N2)->getVT().isVector() &&
6369 "Type to saturate to must be a scalar.");
6370 assert(cast<VTSDNode>(N2)->getVT().bitsLE(VT.getScalarType()) &&
6371 "Not extending!");
6372 break;
6373 }
6374 case ISD::EXTRACT_VECTOR_ELT:
6375 assert(VT.getSizeInBits() >= N1.getValueType().getScalarSizeInBits() &&
6376 "The result of EXTRACT_VECTOR_ELT must be at least as wide as the \
6377 element type of the vector.");
6378
6379 // Extract from an undefined value or using an undefined index is undefined.
6380 if (N1.isUndef() || N2.isUndef())
6381 return getUNDEF(VT);
6382
6383 // EXTRACT_VECTOR_ELT of out-of-bounds element is an UNDEF for fixed length
6384 // vectors. For scalable vectors we will provide appropriate support for
6385 // dealing with arbitrary indices.
6386 if (N2C && N1.getValueType().isFixedLengthVector() &&
6387 N2C->getAPIntValue().uge(N1.getValueType().getVectorNumElements()))
6388 return getUNDEF(VT);
6389
6390 // EXTRACT_VECTOR_ELT of CONCAT_VECTORS is often formed while lowering is
6391 // expanding copies of large vectors from registers. This only works for
6392 // fixed length vectors, since we need to know the exact number of
6393 // elements.
6394 if (N2C && N1.getOperand(0).getValueType().isFixedLengthVector() &&
6395 N1.getOpcode() == ISD::CONCAT_VECTORS && N1.getNumOperands() > 0) {
6396 unsigned Factor =
6397 N1.getOperand(0).getValueType().getVectorNumElements();
6398 return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
6399 N1.getOperand(N2C->getZExtValue() / Factor),
6400 getVectorIdxConstant(N2C->getZExtValue() % Factor, DL));
6401 }
6402
6403 // EXTRACT_VECTOR_ELT of BUILD_VECTOR or SPLAT_VECTOR is often formed while
6404 // lowering is expanding large vector constants.
6405 if (N2C && (N1.getOpcode() == ISD::BUILD_VECTOR ||
6406 N1.getOpcode() == ISD::SPLAT_VECTOR)) {
6407 assert((N1.getOpcode() != ISD::BUILD_VECTOR ||
6408 N1.getValueType().isFixedLengthVector()) &&
6409 "BUILD_VECTOR used for scalable vectors");
6410 unsigned Index =
6411 N1.getOpcode() == ISD::BUILD_VECTOR ? N2C->getZExtValue() : 0;
6412 SDValue Elt = N1.getOperand(Index);
6413
6414 if (VT != Elt.getValueType())
6415 // If the vector element type is not legal, the BUILD_VECTOR operands
6416 // are promoted and implicitly truncated, and the result implicitly
6417 // extended. Make that explicit here.
6418 Elt = getAnyExtOrTrunc(Elt, DL, VT);
6419
6420 return Elt;
6421 }
6422
6423 // EXTRACT_VECTOR_ELT of INSERT_VECTOR_ELT is often formed when vector
6424 // operations are lowered to scalars.
6425 if (N1.getOpcode() == ISD::INSERT_VECTOR_ELT) {
6426 // If the indices are the same, return the inserted element else
6427 // if the indices are known different, extract the element from
6428 // the original vector.
6429 SDValue N1Op2 = N1.getOperand(2);
6430 ConstantSDNode *N1Op2C = dyn_cast<ConstantSDNode>(N1Op2);
6431
6432 if (N1Op2C && N2C) {
6433 if (N1Op2C->getZExtValue() == N2C->getZExtValue()) {
6434 if (VT == N1.getOperand(1).getValueType())
6435 return N1.getOperand(1);
6436 if (VT.isFloatingPoint()) {
6437 assert(VT.getSizeInBits() > N1.getOperand(1).getValueType().getSizeInBits());
6438 return getFPExtendOrRound(N1.getOperand(1), DL, VT);
6439 }
6440 return getSExtOrTrunc(N1.getOperand(1), DL, VT);
6441 }
6442 return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, N1.getOperand(0), N2);
6443 }
6444 }
6445
6446 // EXTRACT_VECTOR_ELT of v1iX EXTRACT_SUBVECTOR could be formed
6447 // when vector types are scalarized and v1iX is legal.
6448 // vextract (v1iX extract_subvector(vNiX, Idx)) -> vextract(vNiX,Idx).
6449 // Here we are completely ignoring the extract element index (N2),
6450 // which is fine for fixed width vectors, since any index other than 0
6451 // is undefined anyway. However, this cannot be ignored for scalable
6452 // vectors - in theory we could support this, but we don't want to do this
6453 // without a profitability check.
6454 if (N1.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
6455 N1.getValueType().isFixedLengthVector() &&
6456 N1.getValueType().getVectorNumElements() == 1) {
6457 return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, N1.getOperand(0),
6458 N1.getOperand(1));
6459 }
6460 break;
6461 case ISD::EXTRACT_ELEMENT:
6462 assert(N2C && (unsigned)N2C->getZExtValue() < 2 && "Bad EXTRACT_ELEMENT!");
6463 assert(!N1.getValueType().isVector() && !VT.isVector() &&
6464 (N1.getValueType().isInteger() == VT.isInteger()) &&
6465 N1.getValueType() != VT &&
6466 "Wrong types for EXTRACT_ELEMENT!");
6467
6468 // EXTRACT_ELEMENT of BUILD_PAIR is often formed while legalize is expanding
6469 // 64-bit integers into 32-bit parts. Instead of building the extract of
6470 // the BUILD_PAIR, only to have legalize rip it apart, just do it now.
6471 if (N1.getOpcode() == ISD::BUILD_PAIR)
6472 return N1.getOperand(N2C->getZExtValue());
6473
6474 // EXTRACT_ELEMENT of a constant int is also very common.
6475 if (N1C) {
6476 unsigned ElementSize = VT.getSizeInBits();
6477 unsigned Shift = ElementSize * N2C->getZExtValue();
6478 const APInt &Val = N1C->getAPIntValue();
6479 return getConstant(Val.extractBits(ElementSize, Shift), DL, VT);
6480 }
6481 break;
6482 case ISD::EXTRACT_SUBVECTOR: {
6483 EVT N1VT = N1.getValueType();
6484 assert(VT.isVector() && N1VT.isVector() &&
6485 "Extract subvector VTs must be vectors!");
6486 assert(VT.getVectorElementType() == N1VT.getVectorElementType() &&
6487 "Extract subvector VTs must have the same element type!");
6488 assert((VT.isFixedLengthVector() || N1VT.isScalableVector()) &&
6489 "Cannot extract a scalable vector from a fixed length vector!");
6490 assert((VT.isScalableVector() != N1VT.isScalableVector() ||
6491 VT.getVectorMinNumElements() <= N1VT.getVectorMinNumElements()) &&
6492 "Extract subvector must be from larger vector to smaller vector!");
6493 assert(N2C && "Extract subvector index must be a constant");
6494 assert((VT.isScalableVector() != N1VT.isScalableVector() ||
6495 (VT.getVectorMinNumElements() + N2C->getZExtValue()) <=
6496 N1VT.getVectorMinNumElements()) &&
6497 "Extract subvector overflow!");
6498 assert(N2C->getAPIntValue().getBitWidth() ==
6499 TLI->getVectorIdxTy(getDataLayout()).getFixedSizeInBits() &&
6500 "Constant index for EXTRACT_SUBVECTOR has an invalid size");
6501
6502 // Trivial extraction.
6503 if (VT == N1VT)
6504 return N1;
6505
6506 // EXTRACT_SUBVECTOR of an UNDEF is an UNDEF.
6507 if (N1.isUndef())
6508 return getUNDEF(VT);
6509
6510 // EXTRACT_SUBVECTOR of CONCAT_VECTOR can be simplified if the pieces of
6511 // the concat have the same type as the extract.
6512 if (N1.getOpcode() == ISD::CONCAT_VECTORS && N1.getNumOperands() > 0 &&
6513 VT == N1.getOperand(0).getValueType()) {
6514 unsigned Factor = VT.getVectorMinNumElements();
6515 return N1.getOperand(N2C->getZExtValue() / Factor);
6516 }
6517
6518 // EXTRACT_SUBVECTOR of INSERT_SUBVECTOR is often created
6519 // during shuffle legalization.
6520 if (N1.getOpcode() == ISD::INSERT_SUBVECTOR && N2 == N1.getOperand(2) &&
6521 VT == N1.getOperand(1).getValueType())
6522 return N1.getOperand(1);
6523 break;
6524 }
6525 }
6526
6527 // Perform trivial constant folding.
6528 if (SDValue SV = FoldConstantArithmetic(Opcode, DL, VT, {N1, N2}))
6529 return SV;
6530
6531 // Canonicalize an UNDEF to the RHS, even over a constant.
6532 if (N1.isUndef()) {
6533 if (TLI->isCommutativeBinOp(Opcode)) {
6534 std::swap(N1, N2);
6535 } else {
6536 switch (Opcode) {
6537 case ISD::SUB:
6538 return getUNDEF(VT); // fold op(undef, arg2) -> undef
6539 case ISD::SIGN_EXTEND_INREG:
6540 case ISD::UDIV:
6541 case ISD::SDIV:
6542 case ISD::UREM:
6543 case ISD::SREM:
6544 case ISD::SSUBSAT:
6545 case ISD::USUBSAT:
6546 return getConstant(0, DL, VT); // fold op(undef, arg2) -> 0
6547 }
6548 }
6549 }
6550
6551 // Fold a bunch of operators when the RHS is undef.
6552 if (N2.isUndef()) {
6553 switch (Opcode) {
6554 case ISD::XOR:
6555 if (N1.isUndef())
6556 // Handle undef ^ undef -> 0 special case. This is a common
6557 // idiom (misuse).
6558 return getConstant(0, DL, VT);
6559 [[fallthrough]];
6560 case ISD::ADD:
6561 case ISD::SUB:
6562 case ISD::UDIV:
6563 case ISD::SDIV:
6564 case ISD::UREM:
6565 case ISD::SREM:
6566 return getUNDEF(VT); // fold op(arg1, undef) -> undef
6567 case ISD::MUL:
6568 case ISD::AND:
6569 case ISD::SSUBSAT:
6570 case ISD::USUBSAT:
6571 return getConstant(0, DL, VT); // fold op(arg1, undef) -> 0
6572 case ISD::OR:
6573 case ISD::SADDSAT:
6574 case ISD::UADDSAT:
6575 return getAllOnesConstant(DL, VT);
6576 }
6577 }
6578
6579 // Memoize this node if possible.
6580 SDNode *N;
6581 SDVTList VTs = getVTList(VT);
6582 SDValue Ops[] = {N1, N2};
6583 if (VT != MVT::Glue) {
6584 FoldingSetNodeID ID;
6585 AddNodeIDNode(ID, Opcode, VTs, Ops);
6586 void *IP = nullptr;
6587 if (SDNode *E = FindNodeOrInsertPos(ID, DL, IP)) {
6588 E->intersectFlagsWith(Flags);
6589 return SDValue(E, 0);
6590 }
6591
6592 N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(), VTs);
6593 N->setFlags(Flags);
6594 createOperands(N, Ops);
6595 CSEMap.InsertNode(N, IP);
6596 } else {
6597 N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(), VTs);
6598 createOperands(N, Ops);
6599 }
6600
6601 InsertNode(N);
6602 SDValue V = SDValue(N, 0);
6603 NewSDValueDbgMsg(V, "Creating new node: ", this);
6604 return V;
6605 }
6606
getNode(unsigned Opcode,const SDLoc & DL,EVT VT,SDValue N1,SDValue N2,SDValue N3)6607 SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
6608 SDValue N1, SDValue N2, SDValue N3) {
6609 SDNodeFlags Flags;
6610 if (Inserter)
6611 Flags = Inserter->getFlags();
6612 return getNode(Opcode, DL, VT, N1, N2, N3, Flags);
6613 }
6614
getNode(unsigned Opcode,const SDLoc & DL,EVT VT,SDValue N1,SDValue N2,SDValue N3,const SDNodeFlags Flags)6615 SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
6616 SDValue N1, SDValue N2, SDValue N3,
6617 const SDNodeFlags Flags) {
6618 assert(N1.getOpcode() != ISD::DELETED_NODE &&
6619 N2.getOpcode() != ISD::DELETED_NODE &&
6620 N3.getOpcode() != ISD::DELETED_NODE &&
6621 "Operand is DELETED_NODE!");
6622 // Perform various simplifications.
6623 switch (Opcode) {
6624 case ISD::FMA: {
6625 assert(VT.isFloatingPoint() && "This operator only applies to FP types!");
6626 assert(N1.getValueType() == VT && N2.getValueType() == VT &&
6627 N3.getValueType() == VT && "FMA types must match!");
6628 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6629 ConstantFPSDNode *N2CFP = dyn_cast<ConstantFPSDNode>(N2);
6630 ConstantFPSDNode *N3CFP = dyn_cast<ConstantFPSDNode>(N3);
6631 if (N1CFP && N2CFP && N3CFP) {
6632 APFloat V1 = N1CFP->getValueAPF();
6633 const APFloat &V2 = N2CFP->getValueAPF();
6634 const APFloat &V3 = N3CFP->getValueAPF();
6635 V1.fusedMultiplyAdd(V2, V3, APFloat::rmNearestTiesToEven);
6636 return getConstantFP(V1, DL, VT);
6637 }
6638 break;
6639 }
6640 case ISD::BUILD_VECTOR: {
6641 // Attempt to simplify BUILD_VECTOR.
6642 SDValue Ops[] = {N1, N2, N3};
6643 if (SDValue V = FoldBUILD_VECTOR(DL, VT, Ops, *this))
6644 return V;
6645 break;
6646 }
6647 case ISD::CONCAT_VECTORS: {
6648 SDValue Ops[] = {N1, N2, N3};
6649 if (SDValue V = foldCONCAT_VECTORS(DL, VT, Ops, *this))
6650 return V;
6651 break;
6652 }
6653 case ISD::SETCC: {
6654 assert(VT.isInteger() && "SETCC result type must be an integer!");
6655 assert(N1.getValueType() == N2.getValueType() &&
6656 "SETCC operands must have the same type!");
6657 assert(VT.isVector() == N1.getValueType().isVector() &&
6658 "SETCC type should be vector iff the operand type is vector!");
6659 assert((!VT.isVector() || VT.getVectorElementCount() ==
6660 N1.getValueType().getVectorElementCount()) &&
6661 "SETCC vector element counts must match!");
6662 // Use FoldSetCC to simplify SETCC's.
6663 if (SDValue V = FoldSetCC(VT, N1, N2, cast<CondCodeSDNode>(N3)->get(), DL))
6664 return V;
6665 // Vector constant folding.
6666 SDValue Ops[] = {N1, N2, N3};
6667 if (SDValue V = FoldConstantArithmetic(Opcode, DL, VT, Ops)) {
6668 NewSDValueDbgMsg(V, "New node vector constant folding: ", this);
6669 return V;
6670 }
6671 break;
6672 }
6673 case ISD::SELECT:
6674 case ISD::VSELECT:
6675 if (SDValue V = simplifySelect(N1, N2, N3))
6676 return V;
6677 break;
6678 case ISD::VECTOR_SHUFFLE:
6679 llvm_unreachable("should use getVectorShuffle constructor!");
6680 case ISD::VECTOR_SPLICE: {
6681 if (cast<ConstantSDNode>(N3)->isNullValue())
6682 return N1;
6683 break;
6684 }
6685 case ISD::INSERT_VECTOR_ELT: {
6686 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3);
6687 // INSERT_VECTOR_ELT into out-of-bounds element is an UNDEF, except
6688 // for scalable vectors where we will generate appropriate code to
6689 // deal with out-of-bounds cases correctly.
6690 if (N3C && N1.getValueType().isFixedLengthVector() &&
6691 N3C->getZExtValue() >= N1.getValueType().getVectorNumElements())
6692 return getUNDEF(VT);
6693
6694 // Undefined index can be assumed out-of-bounds, so that's UNDEF too.
6695 if (N3.isUndef())
6696 return getUNDEF(VT);
6697
6698 // If the inserted element is an UNDEF, just use the input vector.
6699 if (N2.isUndef())
6700 return N1;
6701
6702 break;
6703 }
6704 case ISD::INSERT_SUBVECTOR: {
6705 // Inserting undef into undef is still undef.
6706 if (N1.isUndef() && N2.isUndef())
6707 return getUNDEF(VT);
6708
6709 EVT N2VT = N2.getValueType();
6710 assert(VT == N1.getValueType() &&
6711 "Dest and insert subvector source types must match!");
6712 assert(VT.isVector() && N2VT.isVector() &&
6713 "Insert subvector VTs must be vectors!");
6714 assert(VT.getVectorElementType() == N2VT.getVectorElementType() &&
6715 "Insert subvector VTs must have the same element type!");
6716 assert((VT.isScalableVector() || N2VT.isFixedLengthVector()) &&
6717 "Cannot insert a scalable vector into a fixed length vector!");
6718 assert((VT.isScalableVector() != N2VT.isScalableVector() ||
6719 VT.getVectorMinNumElements() >= N2VT.getVectorMinNumElements()) &&
6720 "Insert subvector must be from smaller vector to larger vector!");
6721 assert(isa<ConstantSDNode>(N3) &&
6722 "Insert subvector index must be constant");
6723 assert((VT.isScalableVector() != N2VT.isScalableVector() ||
6724 (N2VT.getVectorMinNumElements() +
6725 cast<ConstantSDNode>(N3)->getZExtValue()) <=
6726 VT.getVectorMinNumElements()) &&
6727 "Insert subvector overflow!");
6728 assert(cast<ConstantSDNode>(N3)->getAPIntValue().getBitWidth() ==
6729 TLI->getVectorIdxTy(getDataLayout()).getFixedSizeInBits() &&
6730 "Constant index for INSERT_SUBVECTOR has an invalid size");
6731
6732 // Trivial insertion.
6733 if (VT == N2VT)
6734 return N2;
6735
6736 // If this is an insert of an extracted vector into an undef vector, we
6737 // can just use the input to the extract.
6738 if (N1.isUndef() && N2.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
6739 N2.getOperand(1) == N3 && N2.getOperand(0).getValueType() == VT)
6740 return N2.getOperand(0);
6741 break;
6742 }
6743 case ISD::BITCAST:
6744 // Fold bit_convert nodes from a type to themselves.
6745 if (N1.getValueType() == VT)
6746 return N1;
6747 break;
6748 }
6749
6750 // Memoize node if it doesn't produce a flag.
6751 SDNode *N;
6752 SDVTList VTs = getVTList(VT);
6753 SDValue Ops[] = {N1, N2, N3};
6754 if (VT != MVT::Glue) {
6755 FoldingSetNodeID ID;
6756 AddNodeIDNode(ID, Opcode, VTs, Ops);
6757 void *IP = nullptr;
6758 if (SDNode *E = FindNodeOrInsertPos(ID, DL, IP)) {
6759 E->intersectFlagsWith(Flags);
6760 return SDValue(E, 0);
6761 }
6762
6763 N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(), VTs);
6764 N->setFlags(Flags);
6765 createOperands(N, Ops);
6766 CSEMap.InsertNode(N, IP);
6767 } else {
6768 N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(), VTs);
6769 createOperands(N, Ops);
6770 }
6771
6772 InsertNode(N);
6773 SDValue V = SDValue(N, 0);
6774 NewSDValueDbgMsg(V, "Creating new node: ", this);
6775 return V;
6776 }
6777
getNode(unsigned Opcode,const SDLoc & DL,EVT VT,SDValue N1,SDValue N2,SDValue N3,SDValue N4)6778 SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
6779 SDValue N1, SDValue N2, SDValue N3, SDValue N4) {
6780 SDValue Ops[] = { N1, N2, N3, N4 };
6781 return getNode(Opcode, DL, VT, Ops);
6782 }
6783
getNode(unsigned Opcode,const SDLoc & DL,EVT VT,SDValue N1,SDValue N2,SDValue N3,SDValue N4,SDValue N5)6784 SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
6785 SDValue N1, SDValue N2, SDValue N3, SDValue N4,
6786 SDValue N5) {
6787 SDValue Ops[] = { N1, N2, N3, N4, N5 };
6788 return getNode(Opcode, DL, VT, Ops);
6789 }
6790
6791 /// getStackArgumentTokenFactor - Compute a TokenFactor to force all
6792 /// the incoming stack arguments to be loaded from the stack.
getStackArgumentTokenFactor(SDValue Chain)6793 SDValue SelectionDAG::getStackArgumentTokenFactor(SDValue Chain) {
6794 SmallVector<SDValue, 8> ArgChains;
6795
6796 // Include the original chain at the beginning of the list. When this is
6797 // used by target LowerCall hooks, this helps legalize find the
6798 // CALLSEQ_BEGIN node.
6799 ArgChains.push_back(Chain);
6800
6801 // Add a chain value for each stack argument.
6802 for (SDNode *U : getEntryNode().getNode()->uses())
6803 if (LoadSDNode *L = dyn_cast<LoadSDNode>(U))
6804 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
6805 if (FI->getIndex() < 0)
6806 ArgChains.push_back(SDValue(L, 1));
6807
6808 // Build a tokenfactor for all the chains.
6809 return getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
6810 }
6811
6812 /// getMemsetValue - Vectorized representation of the memset value
6813 /// operand.
getMemsetValue(SDValue Value,EVT VT,SelectionDAG & DAG,const SDLoc & dl)6814 static SDValue getMemsetValue(SDValue Value, EVT VT, SelectionDAG &DAG,
6815 const SDLoc &dl) {
6816 assert(!Value.isUndef());
6817
6818 unsigned NumBits = VT.getScalarSizeInBits();
6819 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Value)) {
6820 assert(C->getAPIntValue().getBitWidth() == 8);
6821 APInt Val = APInt::getSplat(NumBits, C->getAPIntValue());
6822 if (VT.isInteger()) {
6823 bool IsOpaque = VT.getSizeInBits() > 64 ||
6824 !DAG.getTargetLoweringInfo().isLegalStoreImmediate(C->getSExtValue());
6825 return DAG.getConstant(Val, dl, VT, false, IsOpaque);
6826 }
6827 return DAG.getConstantFP(APFloat(DAG.EVTToAPFloatSemantics(VT), Val), dl,
6828 VT);
6829 }
6830
6831 assert(Value.getValueType() == MVT::i8 && "memset with non-byte fill value?");
6832 EVT IntVT = VT.getScalarType();
6833 if (!IntVT.isInteger())
6834 IntVT = EVT::getIntegerVT(*DAG.getContext(), IntVT.getSizeInBits());
6835
6836 Value = DAG.getNode(ISD::ZERO_EXTEND, dl, IntVT, Value);
6837 if (NumBits > 8) {
6838 // Use a multiplication with 0x010101... to extend the input to the
6839 // required length.
6840 APInt Magic = APInt::getSplat(NumBits, APInt(8, 0x01));
6841 Value = DAG.getNode(ISD::MUL, dl, IntVT, Value,
6842 DAG.getConstant(Magic, dl, IntVT));
6843 }
6844
6845 if (VT != Value.getValueType() && !VT.isInteger())
6846 Value = DAG.getBitcast(VT.getScalarType(), Value);
6847 if (VT != Value.getValueType())
6848 Value = DAG.getSplatBuildVector(VT, dl, Value);
6849
6850 return Value;
6851 }
6852
6853 /// getMemsetStringVal - Similar to getMemsetValue. Except this is only
6854 /// used when a memcpy is turned into a memset when the source is a constant
6855 /// string ptr.
getMemsetStringVal(EVT VT,const SDLoc & dl,SelectionDAG & DAG,const TargetLowering & TLI,const ConstantDataArraySlice & Slice)6856 static SDValue getMemsetStringVal(EVT VT, const SDLoc &dl, SelectionDAG &DAG,
6857 const TargetLowering &TLI,
6858 const ConstantDataArraySlice &Slice) {
6859 // Handle vector with all elements zero.
6860 if (Slice.Array == nullptr) {
6861 if (VT.isInteger())
6862 return DAG.getConstant(0, dl, VT);
6863 if (VT == MVT::f32 || VT == MVT::f64 || VT == MVT::f128)
6864 return DAG.getConstantFP(0.0, dl, VT);
6865 if (VT.isVector()) {
6866 unsigned NumElts = VT.getVectorNumElements();
6867 MVT EltVT = (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
6868 return DAG.getNode(ISD::BITCAST, dl, VT,
6869 DAG.getConstant(0, dl,
6870 EVT::getVectorVT(*DAG.getContext(),
6871 EltVT, NumElts)));
6872 }
6873 llvm_unreachable("Expected type!");
6874 }
6875
6876 assert(!VT.isVector() && "Can't handle vector type here!");
6877 unsigned NumVTBits = VT.getSizeInBits();
6878 unsigned NumVTBytes = NumVTBits / 8;
6879 unsigned NumBytes = std::min(NumVTBytes, unsigned(Slice.Length));
6880
6881 APInt Val(NumVTBits, 0);
6882 if (DAG.getDataLayout().isLittleEndian()) {
6883 for (unsigned i = 0; i != NumBytes; ++i)
6884 Val |= (uint64_t)(unsigned char)Slice[i] << i*8;
6885 } else {
6886 for (unsigned i = 0; i != NumBytes; ++i)
6887 Val |= (uint64_t)(unsigned char)Slice[i] << (NumVTBytes-i-1)*8;
6888 }
6889
6890 // If the "cost" of materializing the integer immediate is less than the cost
6891 // of a load, then it is cost effective to turn the load into the immediate.
6892 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
6893 if (TLI.shouldConvertConstantLoadToIntImm(Val, Ty))
6894 return DAG.getConstant(Val, dl, VT);
6895 return SDValue();
6896 }
6897
getMemBasePlusOffset(SDValue Base,TypeSize Offset,const SDLoc & DL,const SDNodeFlags Flags)6898 SDValue SelectionDAG::getMemBasePlusOffset(SDValue Base, TypeSize Offset,
6899 const SDLoc &DL,
6900 const SDNodeFlags Flags) {
6901 EVT VT = Base.getValueType();
6902 SDValue Index;
6903
6904 if (Offset.isScalable())
6905 Index = getVScale(DL, Base.getValueType(),
6906 APInt(Base.getValueSizeInBits().getFixedValue(),
6907 Offset.getKnownMinValue()));
6908 else
6909 Index = getConstant(Offset.getFixedValue(), DL, VT);
6910
6911 return getMemBasePlusOffset(Base, Index, DL, Flags);
6912 }
6913
getMemBasePlusOffset(SDValue Ptr,SDValue Offset,const SDLoc & DL,const SDNodeFlags Flags)6914 SDValue SelectionDAG::getMemBasePlusOffset(SDValue Ptr, SDValue Offset,
6915 const SDLoc &DL,
6916 const SDNodeFlags Flags) {
6917 assert(Offset.getValueType().isInteger());
6918 EVT BasePtrVT = Ptr.getValueType();
6919 return getNode(ISD::ADD, DL, BasePtrVT, Ptr, Offset, Flags);
6920 }
6921
6922 /// Returns true if memcpy source is constant data.
isMemSrcFromConstant(SDValue Src,ConstantDataArraySlice & Slice)6923 static bool isMemSrcFromConstant(SDValue Src, ConstantDataArraySlice &Slice) {
6924 uint64_t SrcDelta = 0;
6925 GlobalAddressSDNode *G = nullptr;
6926 if (Src.getOpcode() == ISD::GlobalAddress)
6927 G = cast<GlobalAddressSDNode>(Src);
6928 else if (Src.getOpcode() == ISD::ADD &&
6929 Src.getOperand(0).getOpcode() == ISD::GlobalAddress &&
6930 Src.getOperand(1).getOpcode() == ISD::Constant) {
6931 G = cast<GlobalAddressSDNode>(Src.getOperand(0));
6932 SrcDelta = cast<ConstantSDNode>(Src.getOperand(1))->getZExtValue();
6933 }
6934 if (!G)
6935 return false;
6936
6937 return getConstantDataArrayInfo(G->getGlobal(), Slice, 8,
6938 SrcDelta + G->getOffset());
6939 }
6940
shouldLowerMemFuncForSize(const MachineFunction & MF,SelectionDAG & DAG)6941 static bool shouldLowerMemFuncForSize(const MachineFunction &MF,
6942 SelectionDAG &DAG) {
6943 // On Darwin, -Os means optimize for size without hurting performance, so
6944 // only really optimize for size when -Oz (MinSize) is used.
6945 if (MF.getTarget().getTargetTriple().isOSDarwin())
6946 return MF.getFunction().hasMinSize();
6947 return DAG.shouldOptForSize();
6948 }
6949
chainLoadsAndStoresForMemcpy(SelectionDAG & DAG,const SDLoc & dl,SmallVector<SDValue,32> & OutChains,unsigned From,unsigned To,SmallVector<SDValue,16> & OutLoadChains,SmallVector<SDValue,16> & OutStoreChains)6950 static void chainLoadsAndStoresForMemcpy(SelectionDAG &DAG, const SDLoc &dl,
6951 SmallVector<SDValue, 32> &OutChains, unsigned From,
6952 unsigned To, SmallVector<SDValue, 16> &OutLoadChains,
6953 SmallVector<SDValue, 16> &OutStoreChains) {
6954 assert(OutLoadChains.size() && "Missing loads in memcpy inlining");
6955 assert(OutStoreChains.size() && "Missing stores in memcpy inlining");
6956 SmallVector<SDValue, 16> GluedLoadChains;
6957 for (unsigned i = From; i < To; ++i) {
6958 OutChains.push_back(OutLoadChains[i]);
6959 GluedLoadChains.push_back(OutLoadChains[i]);
6960 }
6961
6962 // Chain for all loads.
6963 SDValue LoadToken = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6964 GluedLoadChains);
6965
6966 for (unsigned i = From; i < To; ++i) {
6967 StoreSDNode *ST = dyn_cast<StoreSDNode>(OutStoreChains[i]);
6968 SDValue NewStore = DAG.getTruncStore(LoadToken, dl, ST->getValue(),
6969 ST->getBasePtr(), ST->getMemoryVT(),
6970 ST->getMemOperand());
6971 OutChains.push_back(NewStore);
6972 }
6973 }
6974
getMemcpyLoadsAndStores(SelectionDAG & DAG,const SDLoc & dl,SDValue Chain,SDValue Dst,SDValue Src,uint64_t Size,Align Alignment,bool isVol,bool AlwaysInline,MachinePointerInfo DstPtrInfo,MachinePointerInfo SrcPtrInfo,const AAMDNodes & AAInfo,AAResults * AA)6975 static SDValue getMemcpyLoadsAndStores(SelectionDAG &DAG, const SDLoc &dl,
6976 SDValue Chain, SDValue Dst, SDValue Src,
6977 uint64_t Size, Align Alignment,
6978 bool isVol, bool AlwaysInline,
6979 MachinePointerInfo DstPtrInfo,
6980 MachinePointerInfo SrcPtrInfo,
6981 const AAMDNodes &AAInfo, AAResults *AA) {
6982 // Turn a memcpy of undef to nop.
6983 // FIXME: We need to honor volatile even is Src is undef.
6984 if (Src.isUndef())
6985 return Chain;
6986
6987 // Expand memcpy to a series of load and store ops if the size operand falls
6988 // below a certain threshold.
6989 // TODO: In the AlwaysInline case, if the size is big then generate a loop
6990 // rather than maybe a humongous number of loads and stores.
6991 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6992 const DataLayout &DL = DAG.getDataLayout();
6993 LLVMContext &C = *DAG.getContext();
6994 std::vector<EVT> MemOps;
6995 bool DstAlignCanChange = false;
6996 MachineFunction &MF = DAG.getMachineFunction();
6997 MachineFrameInfo &MFI = MF.getFrameInfo();
6998 bool OptSize = shouldLowerMemFuncForSize(MF, DAG);
6999 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Dst);
7000 if (FI && !MFI.isFixedObjectIndex(FI->getIndex()))
7001 DstAlignCanChange = true;
7002 MaybeAlign SrcAlign = DAG.InferPtrAlign(Src);
7003 if (!SrcAlign || Alignment > *SrcAlign)
7004 SrcAlign = Alignment;
7005 assert(SrcAlign && "SrcAlign must be set");
7006 ConstantDataArraySlice Slice;
7007 // If marked as volatile, perform a copy even when marked as constant.
7008 bool CopyFromConstant = !isVol && isMemSrcFromConstant(Src, Slice);
7009 bool isZeroConstant = CopyFromConstant && Slice.Array == nullptr;
7010 unsigned Limit = AlwaysInline ? ~0U : TLI.getMaxStoresPerMemcpy(OptSize);
7011 const MemOp Op = isZeroConstant
7012 ? MemOp::Set(Size, DstAlignCanChange, Alignment,
7013 /*IsZeroMemset*/ true, isVol)
7014 : MemOp::Copy(Size, DstAlignCanChange, Alignment,
7015 *SrcAlign, isVol, CopyFromConstant);
7016 if (!TLI.findOptimalMemOpLowering(
7017 MemOps, Limit, Op, DstPtrInfo.getAddrSpace(),
7018 SrcPtrInfo.getAddrSpace(), MF.getFunction().getAttributes()))
7019 return SDValue();
7020
7021 if (DstAlignCanChange) {
7022 Type *Ty = MemOps[0].getTypeForEVT(C);
7023 Align NewAlign = DL.getABITypeAlign(Ty);
7024
7025 // Don't promote to an alignment that would require dynamic stack
7026 // realignment which may conflict with optimizations such as tail call
7027 // optimization.
7028 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
7029 if (!TRI->hasStackRealignment(MF))
7030 while (NewAlign > Alignment && DL.exceedsNaturalStackAlignment(NewAlign))
7031 NewAlign = NewAlign.previous();
7032
7033 if (NewAlign > Alignment) {
7034 // Give the stack frame object a larger alignment if needed.
7035 if (MFI.getObjectAlign(FI->getIndex()) < NewAlign)
7036 MFI.setObjectAlignment(FI->getIndex(), NewAlign);
7037 Alignment = NewAlign;
7038 }
7039 }
7040
7041 // Prepare AAInfo for loads/stores after lowering this memcpy.
7042 AAMDNodes NewAAInfo = AAInfo;
7043 NewAAInfo.TBAA = NewAAInfo.TBAAStruct = nullptr;
7044
7045 const Value *SrcVal = SrcPtrInfo.V.dyn_cast<const Value *>();
7046 bool isConstant =
7047 AA && SrcVal &&
7048 AA->pointsToConstantMemory(MemoryLocation(SrcVal, Size, AAInfo));
7049
7050 MachineMemOperand::Flags MMOFlags =
7051 isVol ? MachineMemOperand::MOVolatile : MachineMemOperand::MONone;
7052 SmallVector<SDValue, 16> OutLoadChains;
7053 SmallVector<SDValue, 16> OutStoreChains;
7054 SmallVector<SDValue, 32> OutChains;
7055 unsigned NumMemOps = MemOps.size();
7056 uint64_t SrcOff = 0, DstOff = 0;
7057 for (unsigned i = 0; i != NumMemOps; ++i) {
7058 EVT VT = MemOps[i];
7059 unsigned VTSize = VT.getSizeInBits() / 8;
7060 SDValue Value, Store;
7061
7062 if (VTSize > Size) {
7063 // Issuing an unaligned load / store pair that overlaps with the previous
7064 // pair. Adjust the offset accordingly.
7065 assert(i == NumMemOps-1 && i != 0);
7066 SrcOff -= VTSize - Size;
7067 DstOff -= VTSize - Size;
7068 }
7069
7070 if (CopyFromConstant &&
7071 (isZeroConstant || (VT.isInteger() && !VT.isVector()))) {
7072 // It's unlikely a store of a vector immediate can be done in a single
7073 // instruction. It would require a load from a constantpool first.
7074 // We only handle zero vectors here.
7075 // FIXME: Handle other cases where store of vector immediate is done in
7076 // a single instruction.
7077 ConstantDataArraySlice SubSlice;
7078 if (SrcOff < Slice.Length) {
7079 SubSlice = Slice;
7080 SubSlice.move(SrcOff);
7081 } else {
7082 // This is an out-of-bounds access and hence UB. Pretend we read zero.
7083 SubSlice.Array = nullptr;
7084 SubSlice.Offset = 0;
7085 SubSlice.Length = VTSize;
7086 }
7087 Value = getMemsetStringVal(VT, dl, DAG, TLI, SubSlice);
7088 if (Value.getNode()) {
7089 Store = DAG.getStore(
7090 Chain, dl, Value,
7091 DAG.getMemBasePlusOffset(Dst, TypeSize::Fixed(DstOff), dl),
7092 DstPtrInfo.getWithOffset(DstOff), Alignment, MMOFlags, NewAAInfo);
7093 OutChains.push_back(Store);
7094 }
7095 }
7096
7097 if (!Store.getNode()) {
7098 // The type might not be legal for the target. This should only happen
7099 // if the type is smaller than a legal type, as on PPC, so the right
7100 // thing to do is generate a LoadExt/StoreTrunc pair. These simplify
7101 // to Load/Store if NVT==VT.
7102 // FIXME does the case above also need this?
7103 EVT NVT = TLI.getTypeToTransformTo(C, VT);
7104 assert(NVT.bitsGE(VT));
7105
7106 bool isDereferenceable =
7107 SrcPtrInfo.getWithOffset(SrcOff).isDereferenceable(VTSize, C, DL);
7108 MachineMemOperand::Flags SrcMMOFlags = MMOFlags;
7109 if (isDereferenceable)
7110 SrcMMOFlags |= MachineMemOperand::MODereferenceable;
7111 if (isConstant)
7112 SrcMMOFlags |= MachineMemOperand::MOInvariant;
7113
7114 Value = DAG.getExtLoad(
7115 ISD::EXTLOAD, dl, NVT, Chain,
7116 DAG.getMemBasePlusOffset(Src, TypeSize::Fixed(SrcOff), dl),
7117 SrcPtrInfo.getWithOffset(SrcOff), VT,
7118 commonAlignment(*SrcAlign, SrcOff), SrcMMOFlags, NewAAInfo);
7119 OutLoadChains.push_back(Value.getValue(1));
7120
7121 Store = DAG.getTruncStore(
7122 Chain, dl, Value,
7123 DAG.getMemBasePlusOffset(Dst, TypeSize::Fixed(DstOff), dl),
7124 DstPtrInfo.getWithOffset(DstOff), VT, Alignment, MMOFlags, NewAAInfo);
7125 OutStoreChains.push_back(Store);
7126 }
7127 SrcOff += VTSize;
7128 DstOff += VTSize;
7129 Size -= VTSize;
7130 }
7131
7132 unsigned GluedLdStLimit = MaxLdStGlue == 0 ?
7133 TLI.getMaxGluedStoresPerMemcpy() : MaxLdStGlue;
7134 unsigned NumLdStInMemcpy = OutStoreChains.size();
7135
7136 if (NumLdStInMemcpy) {
7137 // It may be that memcpy might be converted to memset if it's memcpy
7138 // of constants. In such a case, we won't have loads and stores, but
7139 // just stores. In the absence of loads, there is nothing to gang up.
7140 if ((GluedLdStLimit <= 1) || !EnableMemCpyDAGOpt) {
7141 // If target does not care, just leave as it.
7142 for (unsigned i = 0; i < NumLdStInMemcpy; ++i) {
7143 OutChains.push_back(OutLoadChains[i]);
7144 OutChains.push_back(OutStoreChains[i]);
7145 }
7146 } else {
7147 // Ld/St less than/equal limit set by target.
7148 if (NumLdStInMemcpy <= GluedLdStLimit) {
7149 chainLoadsAndStoresForMemcpy(DAG, dl, OutChains, 0,
7150 NumLdStInMemcpy, OutLoadChains,
7151 OutStoreChains);
7152 } else {
7153 unsigned NumberLdChain = NumLdStInMemcpy / GluedLdStLimit;
7154 unsigned RemainingLdStInMemcpy = NumLdStInMemcpy % GluedLdStLimit;
7155 unsigned GlueIter = 0;
7156
7157 for (unsigned cnt = 0; cnt < NumberLdChain; ++cnt) {
7158 unsigned IndexFrom = NumLdStInMemcpy - GlueIter - GluedLdStLimit;
7159 unsigned IndexTo = NumLdStInMemcpy - GlueIter;
7160
7161 chainLoadsAndStoresForMemcpy(DAG, dl, OutChains, IndexFrom, IndexTo,
7162 OutLoadChains, OutStoreChains);
7163 GlueIter += GluedLdStLimit;
7164 }
7165
7166 // Residual ld/st.
7167 if (RemainingLdStInMemcpy) {
7168 chainLoadsAndStoresForMemcpy(DAG, dl, OutChains, 0,
7169 RemainingLdStInMemcpy, OutLoadChains,
7170 OutStoreChains);
7171 }
7172 }
7173 }
7174 }
7175 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
7176 }
7177
getMemmoveLoadsAndStores(SelectionDAG & DAG,const SDLoc & dl,SDValue Chain,SDValue Dst,SDValue Src,uint64_t Size,Align Alignment,bool isVol,bool AlwaysInline,MachinePointerInfo DstPtrInfo,MachinePointerInfo SrcPtrInfo,const AAMDNodes & AAInfo)7178 static SDValue getMemmoveLoadsAndStores(SelectionDAG &DAG, const SDLoc &dl,
7179 SDValue Chain, SDValue Dst, SDValue Src,
7180 uint64_t Size, Align Alignment,
7181 bool isVol, bool AlwaysInline,
7182 MachinePointerInfo DstPtrInfo,
7183 MachinePointerInfo SrcPtrInfo,
7184 const AAMDNodes &AAInfo) {
7185 // Turn a memmove of undef to nop.
7186 // FIXME: We need to honor volatile even is Src is undef.
7187 if (Src.isUndef())
7188 return Chain;
7189
7190 // Expand memmove to a series of load and store ops if the size operand falls
7191 // below a certain threshold.
7192 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7193 const DataLayout &DL = DAG.getDataLayout();
7194 LLVMContext &C = *DAG.getContext();
7195 std::vector<EVT> MemOps;
7196 bool DstAlignCanChange = false;
7197 MachineFunction &MF = DAG.getMachineFunction();
7198 MachineFrameInfo &MFI = MF.getFrameInfo();
7199 bool OptSize = shouldLowerMemFuncForSize(MF, DAG);
7200 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Dst);
7201 if (FI && !MFI.isFixedObjectIndex(FI->getIndex()))
7202 DstAlignCanChange = true;
7203 MaybeAlign SrcAlign = DAG.InferPtrAlign(Src);
7204 if (!SrcAlign || Alignment > *SrcAlign)
7205 SrcAlign = Alignment;
7206 assert(SrcAlign && "SrcAlign must be set");
7207 unsigned Limit = AlwaysInline ? ~0U : TLI.getMaxStoresPerMemmove(OptSize);
7208 if (!TLI.findOptimalMemOpLowering(
7209 MemOps, Limit,
7210 MemOp::Copy(Size, DstAlignCanChange, Alignment, *SrcAlign,
7211 /*IsVolatile*/ true),
7212 DstPtrInfo.getAddrSpace(), SrcPtrInfo.getAddrSpace(),
7213 MF.getFunction().getAttributes()))
7214 return SDValue();
7215
7216 if (DstAlignCanChange) {
7217 Type *Ty = MemOps[0].getTypeForEVT(C);
7218 Align NewAlign = DL.getABITypeAlign(Ty);
7219
7220 // Don't promote to an alignment that would require dynamic stack
7221 // realignment which may conflict with optimizations such as tail call
7222 // optimization.
7223 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
7224 if (!TRI->hasStackRealignment(MF))
7225 while (NewAlign > Alignment && DL.exceedsNaturalStackAlignment(NewAlign))
7226 NewAlign = NewAlign.previous();
7227
7228 if (NewAlign > Alignment) {
7229 // Give the stack frame object a larger alignment if needed.
7230 if (MFI.getObjectAlign(FI->getIndex()) < NewAlign)
7231 MFI.setObjectAlignment(FI->getIndex(), NewAlign);
7232 Alignment = NewAlign;
7233 }
7234 }
7235
7236 // Prepare AAInfo for loads/stores after lowering this memmove.
7237 AAMDNodes NewAAInfo = AAInfo;
7238 NewAAInfo.TBAA = NewAAInfo.TBAAStruct = nullptr;
7239
7240 MachineMemOperand::Flags MMOFlags =
7241 isVol ? MachineMemOperand::MOVolatile : MachineMemOperand::MONone;
7242 uint64_t SrcOff = 0, DstOff = 0;
7243 SmallVector<SDValue, 8> LoadValues;
7244 SmallVector<SDValue, 8> LoadChains;
7245 SmallVector<SDValue, 8> OutChains;
7246 unsigned NumMemOps = MemOps.size();
7247 for (unsigned i = 0; i < NumMemOps; i++) {
7248 EVT VT = MemOps[i];
7249 unsigned VTSize = VT.getSizeInBits() / 8;
7250 SDValue Value;
7251
7252 bool isDereferenceable =
7253 SrcPtrInfo.getWithOffset(SrcOff).isDereferenceable(VTSize, C, DL);
7254 MachineMemOperand::Flags SrcMMOFlags = MMOFlags;
7255 if (isDereferenceable)
7256 SrcMMOFlags |= MachineMemOperand::MODereferenceable;
7257
7258 Value = DAG.getLoad(
7259 VT, dl, Chain,
7260 DAG.getMemBasePlusOffset(Src, TypeSize::Fixed(SrcOff), dl),
7261 SrcPtrInfo.getWithOffset(SrcOff), *SrcAlign, SrcMMOFlags, NewAAInfo);
7262 LoadValues.push_back(Value);
7263 LoadChains.push_back(Value.getValue(1));
7264 SrcOff += VTSize;
7265 }
7266 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, LoadChains);
7267 OutChains.clear();
7268 for (unsigned i = 0; i < NumMemOps; i++) {
7269 EVT VT = MemOps[i];
7270 unsigned VTSize = VT.getSizeInBits() / 8;
7271 SDValue Store;
7272
7273 Store = DAG.getStore(
7274 Chain, dl, LoadValues[i],
7275 DAG.getMemBasePlusOffset(Dst, TypeSize::Fixed(DstOff), dl),
7276 DstPtrInfo.getWithOffset(DstOff), Alignment, MMOFlags, NewAAInfo);
7277 OutChains.push_back(Store);
7278 DstOff += VTSize;
7279 }
7280
7281 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
7282 }
7283
7284 /// Lower the call to 'memset' intrinsic function into a series of store
7285 /// operations.
7286 ///
7287 /// \param DAG Selection DAG where lowered code is placed.
7288 /// \param dl Link to corresponding IR location.
7289 /// \param Chain Control flow dependency.
7290 /// \param Dst Pointer to destination memory location.
7291 /// \param Src Value of byte to write into the memory.
7292 /// \param Size Number of bytes to write.
7293 /// \param Alignment Alignment of the destination in bytes.
7294 /// \param isVol True if destination is volatile.
7295 /// \param AlwaysInline Makes sure no function call is generated.
7296 /// \param DstPtrInfo IR information on the memory pointer.
7297 /// \returns New head in the control flow, if lowering was successful, empty
7298 /// SDValue otherwise.
7299 ///
7300 /// The function tries to replace 'llvm.memset' intrinsic with several store
7301 /// operations and value calculation code. This is usually profitable for small
7302 /// memory size or when the semantic requires inlining.
getMemsetStores(SelectionDAG & DAG,const SDLoc & dl,SDValue Chain,SDValue Dst,SDValue Src,uint64_t Size,Align Alignment,bool isVol,bool AlwaysInline,MachinePointerInfo DstPtrInfo,const AAMDNodes & AAInfo)7303 static SDValue getMemsetStores(SelectionDAG &DAG, const SDLoc &dl,
7304 SDValue Chain, SDValue Dst, SDValue Src,
7305 uint64_t Size, Align Alignment, bool isVol,
7306 bool AlwaysInline, MachinePointerInfo DstPtrInfo,
7307 const AAMDNodes &AAInfo) {
7308 // Turn a memset of undef to nop.
7309 // FIXME: We need to honor volatile even is Src is undef.
7310 if (Src.isUndef())
7311 return Chain;
7312
7313 // Expand memset to a series of load/store ops if the size operand
7314 // falls below a certain threshold.
7315 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7316 std::vector<EVT> MemOps;
7317 bool DstAlignCanChange = false;
7318 MachineFunction &MF = DAG.getMachineFunction();
7319 MachineFrameInfo &MFI = MF.getFrameInfo();
7320 bool OptSize = shouldLowerMemFuncForSize(MF, DAG);
7321 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Dst);
7322 if (FI && !MFI.isFixedObjectIndex(FI->getIndex()))
7323 DstAlignCanChange = true;
7324 bool IsZeroVal =
7325 isa<ConstantSDNode>(Src) && cast<ConstantSDNode>(Src)->isZero();
7326 unsigned Limit = AlwaysInline ? ~0 : TLI.getMaxStoresPerMemset(OptSize);
7327
7328 if (!TLI.findOptimalMemOpLowering(
7329 MemOps, Limit,
7330 MemOp::Set(Size, DstAlignCanChange, Alignment, IsZeroVal, isVol),
7331 DstPtrInfo.getAddrSpace(), ~0u, MF.getFunction().getAttributes()))
7332 return SDValue();
7333
7334 if (DstAlignCanChange) {
7335 Type *Ty = MemOps[0].getTypeForEVT(*DAG.getContext());
7336 const DataLayout &DL = DAG.getDataLayout();
7337 Align NewAlign = DL.getABITypeAlign(Ty);
7338
7339 // Don't promote to an alignment that would require dynamic stack
7340 // realignment which may conflict with optimizations such as tail call
7341 // optimization.
7342 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
7343 if (!TRI->hasStackRealignment(MF))
7344 while (NewAlign > Alignment && DL.exceedsNaturalStackAlignment(NewAlign))
7345 NewAlign = NewAlign.previous();
7346
7347 if (NewAlign > Alignment) {
7348 // Give the stack frame object a larger alignment if needed.
7349 if (MFI.getObjectAlign(FI->getIndex()) < NewAlign)
7350 MFI.setObjectAlignment(FI->getIndex(), NewAlign);
7351 Alignment = NewAlign;
7352 }
7353 }
7354
7355 SmallVector<SDValue, 8> OutChains;
7356 uint64_t DstOff = 0;
7357 unsigned NumMemOps = MemOps.size();
7358
7359 // Find the largest store and generate the bit pattern for it.
7360 EVT LargestVT = MemOps[0];
7361 for (unsigned i = 1; i < NumMemOps; i++)
7362 if (MemOps[i].bitsGT(LargestVT))
7363 LargestVT = MemOps[i];
7364 SDValue MemSetValue = getMemsetValue(Src, LargestVT, DAG, dl);
7365
7366 // Prepare AAInfo for loads/stores after lowering this memset.
7367 AAMDNodes NewAAInfo = AAInfo;
7368 NewAAInfo.TBAA = NewAAInfo.TBAAStruct = nullptr;
7369
7370 for (unsigned i = 0; i < NumMemOps; i++) {
7371 EVT VT = MemOps[i];
7372 unsigned VTSize = VT.getSizeInBits() / 8;
7373 if (VTSize > Size) {
7374 // Issuing an unaligned load / store pair that overlaps with the previous
7375 // pair. Adjust the offset accordingly.
7376 assert(i == NumMemOps-1 && i != 0);
7377 DstOff -= VTSize - Size;
7378 }
7379
7380 // If this store is smaller than the largest store see whether we can get
7381 // the smaller value for free with a truncate.
7382 SDValue Value = MemSetValue;
7383 if (VT.bitsLT(LargestVT)) {
7384 if (!LargestVT.isVector() && !VT.isVector() &&
7385 TLI.isTruncateFree(LargestVT, VT))
7386 Value = DAG.getNode(ISD::TRUNCATE, dl, VT, MemSetValue);
7387 else
7388 Value = getMemsetValue(Src, VT, DAG, dl);
7389 }
7390 assert(Value.getValueType() == VT && "Value with wrong type.");
7391 SDValue Store = DAG.getStore(
7392 Chain, dl, Value,
7393 DAG.getMemBasePlusOffset(Dst, TypeSize::Fixed(DstOff), dl),
7394 DstPtrInfo.getWithOffset(DstOff), Alignment,
7395 isVol ? MachineMemOperand::MOVolatile : MachineMemOperand::MONone,
7396 NewAAInfo);
7397 OutChains.push_back(Store);
7398 DstOff += VT.getSizeInBits() / 8;
7399 Size -= VTSize;
7400 }
7401
7402 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
7403 }
7404
checkAddrSpaceIsValidForLibcall(const TargetLowering * TLI,unsigned AS)7405 static void checkAddrSpaceIsValidForLibcall(const TargetLowering *TLI,
7406 unsigned AS) {
7407 // Lowering memcpy / memset / memmove intrinsics to calls is only valid if all
7408 // pointer operands can be losslessly bitcasted to pointers of address space 0
7409 if (AS != 0 && !TLI->getTargetMachine().isNoopAddrSpaceCast(AS, 0)) {
7410 report_fatal_error("cannot lower memory intrinsic in address space " +
7411 Twine(AS));
7412 }
7413 }
7414
getMemcpy(SDValue Chain,const SDLoc & dl,SDValue Dst,SDValue Src,SDValue Size,Align Alignment,bool isVol,bool AlwaysInline,bool isTailCall,MachinePointerInfo DstPtrInfo,MachinePointerInfo SrcPtrInfo,const AAMDNodes & AAInfo,AAResults * AA)7415 SDValue SelectionDAG::getMemcpy(SDValue Chain, const SDLoc &dl, SDValue Dst,
7416 SDValue Src, SDValue Size, Align Alignment,
7417 bool isVol, bool AlwaysInline, bool isTailCall,
7418 MachinePointerInfo DstPtrInfo,
7419 MachinePointerInfo SrcPtrInfo,
7420 const AAMDNodes &AAInfo, AAResults *AA) {
7421 // Check to see if we should lower the memcpy to loads and stores first.
7422 // For cases within the target-specified limits, this is the best choice.
7423 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
7424 if (ConstantSize) {
7425 // Memcpy with size zero? Just return the original chain.
7426 if (ConstantSize->isZero())
7427 return Chain;
7428
7429 SDValue Result = getMemcpyLoadsAndStores(
7430 *this, dl, Chain, Dst, Src, ConstantSize->getZExtValue(), Alignment,
7431 isVol, false, DstPtrInfo, SrcPtrInfo, AAInfo, AA);
7432 if (Result.getNode())
7433 return Result;
7434 }
7435
7436 // Then check to see if we should lower the memcpy with target-specific
7437 // code. If the target chooses to do this, this is the next best.
7438 if (TSI) {
7439 SDValue Result = TSI->EmitTargetCodeForMemcpy(
7440 *this, dl, Chain, Dst, Src, Size, Alignment, isVol, AlwaysInline,
7441 DstPtrInfo, SrcPtrInfo);
7442 if (Result.getNode())
7443 return Result;
7444 }
7445
7446 // If we really need inline code and the target declined to provide it,
7447 // use a (potentially long) sequence of loads and stores.
7448 if (AlwaysInline) {
7449 assert(ConstantSize && "AlwaysInline requires a constant size!");
7450 return getMemcpyLoadsAndStores(
7451 *this, dl, Chain, Dst, Src, ConstantSize->getZExtValue(), Alignment,
7452 isVol, true, DstPtrInfo, SrcPtrInfo, AAInfo, AA);
7453 }
7454
7455 checkAddrSpaceIsValidForLibcall(TLI, DstPtrInfo.getAddrSpace());
7456 checkAddrSpaceIsValidForLibcall(TLI, SrcPtrInfo.getAddrSpace());
7457
7458 // FIXME: If the memcpy is volatile (isVol), lowering it to a plain libc
7459 // memcpy is not guaranteed to be safe. libc memcpys aren't required to
7460 // respect volatile, so they may do things like read or write memory
7461 // beyond the given memory regions. But fixing this isn't easy, and most
7462 // people don't care.
7463
7464 // Emit a library call.
7465 TargetLowering::ArgListTy Args;
7466 TargetLowering::ArgListEntry Entry;
7467 Entry.Ty = Type::getInt8PtrTy(*getContext());
7468 Entry.Node = Dst; Args.push_back(Entry);
7469 Entry.Node = Src; Args.push_back(Entry);
7470
7471 Entry.Ty = getDataLayout().getIntPtrType(*getContext());
7472 Entry.Node = Size; Args.push_back(Entry);
7473 // FIXME: pass in SDLoc
7474 TargetLowering::CallLoweringInfo CLI(*this);
7475 CLI.setDebugLoc(dl)
7476 .setChain(Chain)
7477 .setLibCallee(TLI->getLibcallCallingConv(RTLIB::MEMCPY),
7478 Dst.getValueType().getTypeForEVT(*getContext()),
7479 getExternalSymbol(TLI->getLibcallName(RTLIB::MEMCPY),
7480 TLI->getPointerTy(getDataLayout())),
7481 std::move(Args))
7482 .setDiscardResult()
7483 .setTailCall(isTailCall);
7484
7485 std::pair<SDValue,SDValue> CallResult = TLI->LowerCallTo(CLI);
7486 return CallResult.second;
7487 }
7488
getAtomicMemcpy(SDValue Chain,const SDLoc & dl,SDValue Dst,SDValue Src,SDValue Size,Type * SizeTy,unsigned ElemSz,bool isTailCall,MachinePointerInfo DstPtrInfo,MachinePointerInfo SrcPtrInfo)7489 SDValue SelectionDAG::getAtomicMemcpy(SDValue Chain, const SDLoc &dl,
7490 SDValue Dst, SDValue Src, SDValue Size,
7491 Type *SizeTy, unsigned ElemSz,
7492 bool isTailCall,
7493 MachinePointerInfo DstPtrInfo,
7494 MachinePointerInfo SrcPtrInfo) {
7495 // Emit a library call.
7496 TargetLowering::ArgListTy Args;
7497 TargetLowering::ArgListEntry Entry;
7498 Entry.Ty = getDataLayout().getIntPtrType(*getContext());
7499 Entry.Node = Dst;
7500 Args.push_back(Entry);
7501
7502 Entry.Node = Src;
7503 Args.push_back(Entry);
7504
7505 Entry.Ty = SizeTy;
7506 Entry.Node = Size;
7507 Args.push_back(Entry);
7508
7509 RTLIB::Libcall LibraryCall =
7510 RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(ElemSz);
7511 if (LibraryCall == RTLIB::UNKNOWN_LIBCALL)
7512 report_fatal_error("Unsupported element size");
7513
7514 TargetLowering::CallLoweringInfo CLI(*this);
7515 CLI.setDebugLoc(dl)
7516 .setChain(Chain)
7517 .setLibCallee(TLI->getLibcallCallingConv(LibraryCall),
7518 Type::getVoidTy(*getContext()),
7519 getExternalSymbol(TLI->getLibcallName(LibraryCall),
7520 TLI->getPointerTy(getDataLayout())),
7521 std::move(Args))
7522 .setDiscardResult()
7523 .setTailCall(isTailCall);
7524
7525 std::pair<SDValue, SDValue> CallResult = TLI->LowerCallTo(CLI);
7526 return CallResult.second;
7527 }
7528
getMemmove(SDValue Chain,const SDLoc & dl,SDValue Dst,SDValue Src,SDValue Size,Align Alignment,bool isVol,bool isTailCall,MachinePointerInfo DstPtrInfo,MachinePointerInfo SrcPtrInfo,const AAMDNodes & AAInfo,AAResults * AA)7529 SDValue SelectionDAG::getMemmove(SDValue Chain, const SDLoc &dl, SDValue Dst,
7530 SDValue Src, SDValue Size, Align Alignment,
7531 bool isVol, bool isTailCall,
7532 MachinePointerInfo DstPtrInfo,
7533 MachinePointerInfo SrcPtrInfo,
7534 const AAMDNodes &AAInfo, AAResults *AA) {
7535 // Check to see if we should lower the memmove to loads and stores first.
7536 // For cases within the target-specified limits, this is the best choice.
7537 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
7538 if (ConstantSize) {
7539 // Memmove with size zero? Just return the original chain.
7540 if (ConstantSize->isZero())
7541 return Chain;
7542
7543 SDValue Result = getMemmoveLoadsAndStores(
7544 *this, dl, Chain, Dst, Src, ConstantSize->getZExtValue(), Alignment,
7545 isVol, false, DstPtrInfo, SrcPtrInfo, AAInfo);
7546 if (Result.getNode())
7547 return Result;
7548 }
7549
7550 // Then check to see if we should lower the memmove with target-specific
7551 // code. If the target chooses to do this, this is the next best.
7552 if (TSI) {
7553 SDValue Result =
7554 TSI->EmitTargetCodeForMemmove(*this, dl, Chain, Dst, Src, Size,
7555 Alignment, isVol, DstPtrInfo, SrcPtrInfo);
7556 if (Result.getNode())
7557 return Result;
7558 }
7559
7560 checkAddrSpaceIsValidForLibcall(TLI, DstPtrInfo.getAddrSpace());
7561 checkAddrSpaceIsValidForLibcall(TLI, SrcPtrInfo.getAddrSpace());
7562
7563 // FIXME: If the memmove is volatile, lowering it to plain libc memmove may
7564 // not be safe. See memcpy above for more details.
7565
7566 // Emit a library call.
7567 TargetLowering::ArgListTy Args;
7568 TargetLowering::ArgListEntry Entry;
7569 Entry.Ty = Type::getInt8PtrTy(*getContext());
7570 Entry.Node = Dst; Args.push_back(Entry);
7571 Entry.Node = Src; Args.push_back(Entry);
7572
7573 Entry.Ty = getDataLayout().getIntPtrType(*getContext());
7574 Entry.Node = Size; Args.push_back(Entry);
7575 // FIXME: pass in SDLoc
7576 TargetLowering::CallLoweringInfo CLI(*this);
7577 CLI.setDebugLoc(dl)
7578 .setChain(Chain)
7579 .setLibCallee(TLI->getLibcallCallingConv(RTLIB::MEMMOVE),
7580 Dst.getValueType().getTypeForEVT(*getContext()),
7581 getExternalSymbol(TLI->getLibcallName(RTLIB::MEMMOVE),
7582 TLI->getPointerTy(getDataLayout())),
7583 std::move(Args))
7584 .setDiscardResult()
7585 .setTailCall(isTailCall);
7586
7587 std::pair<SDValue,SDValue> CallResult = TLI->LowerCallTo(CLI);
7588 return CallResult.second;
7589 }
7590
getAtomicMemmove(SDValue Chain,const SDLoc & dl,SDValue Dst,SDValue Src,SDValue Size,Type * SizeTy,unsigned ElemSz,bool isTailCall,MachinePointerInfo DstPtrInfo,MachinePointerInfo SrcPtrInfo)7591 SDValue SelectionDAG::getAtomicMemmove(SDValue Chain, const SDLoc &dl,
7592 SDValue Dst, SDValue Src, SDValue Size,
7593 Type *SizeTy, unsigned ElemSz,
7594 bool isTailCall,
7595 MachinePointerInfo DstPtrInfo,
7596 MachinePointerInfo SrcPtrInfo) {
7597 // Emit a library call.
7598 TargetLowering::ArgListTy Args;
7599 TargetLowering::ArgListEntry Entry;
7600 Entry.Ty = getDataLayout().getIntPtrType(*getContext());
7601 Entry.Node = Dst;
7602 Args.push_back(Entry);
7603
7604 Entry.Node = Src;
7605 Args.push_back(Entry);
7606
7607 Entry.Ty = SizeTy;
7608 Entry.Node = Size;
7609 Args.push_back(Entry);
7610
7611 RTLIB::Libcall LibraryCall =
7612 RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(ElemSz);
7613 if (LibraryCall == RTLIB::UNKNOWN_LIBCALL)
7614 report_fatal_error("Unsupported element size");
7615
7616 TargetLowering::CallLoweringInfo CLI(*this);
7617 CLI.setDebugLoc(dl)
7618 .setChain(Chain)
7619 .setLibCallee(TLI->getLibcallCallingConv(LibraryCall),
7620 Type::getVoidTy(*getContext()),
7621 getExternalSymbol(TLI->getLibcallName(LibraryCall),
7622 TLI->getPointerTy(getDataLayout())),
7623 std::move(Args))
7624 .setDiscardResult()
7625 .setTailCall(isTailCall);
7626
7627 std::pair<SDValue, SDValue> CallResult = TLI->LowerCallTo(CLI);
7628 return CallResult.second;
7629 }
7630
getMemset(SDValue Chain,const SDLoc & dl,SDValue Dst,SDValue Src,SDValue Size,Align Alignment,bool isVol,bool AlwaysInline,bool isTailCall,MachinePointerInfo DstPtrInfo,const AAMDNodes & AAInfo)7631 SDValue SelectionDAG::getMemset(SDValue Chain, const SDLoc &dl, SDValue Dst,
7632 SDValue Src, SDValue Size, Align Alignment,
7633 bool isVol, bool AlwaysInline, bool isTailCall,
7634 MachinePointerInfo DstPtrInfo,
7635 const AAMDNodes &AAInfo) {
7636 // Check to see if we should lower the memset to stores first.
7637 // For cases within the target-specified limits, this is the best choice.
7638 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
7639 if (ConstantSize) {
7640 // Memset with size zero? Just return the original chain.
7641 if (ConstantSize->isZero())
7642 return Chain;
7643
7644 SDValue Result = getMemsetStores(*this, dl, Chain, Dst, Src,
7645 ConstantSize->getZExtValue(), Alignment,
7646 isVol, false, DstPtrInfo, AAInfo);
7647
7648 if (Result.getNode())
7649 return Result;
7650 }
7651
7652 // Then check to see if we should lower the memset with target-specific
7653 // code. If the target chooses to do this, this is the next best.
7654 if (TSI) {
7655 SDValue Result = TSI->EmitTargetCodeForMemset(
7656 *this, dl, Chain, Dst, Src, Size, Alignment, isVol, AlwaysInline, DstPtrInfo);
7657 if (Result.getNode())
7658 return Result;
7659 }
7660
7661 // If we really need inline code and the target declined to provide it,
7662 // use a (potentially long) sequence of loads and stores.
7663 if (AlwaysInline) {
7664 assert(ConstantSize && "AlwaysInline requires a constant size!");
7665 SDValue Result = getMemsetStores(*this, dl, Chain, Dst, Src,
7666 ConstantSize->getZExtValue(), Alignment,
7667 isVol, true, DstPtrInfo, AAInfo);
7668 assert(Result &&
7669 "getMemsetStores must return a valid sequence when AlwaysInline");
7670 return Result;
7671 }
7672
7673 checkAddrSpaceIsValidForLibcall(TLI, DstPtrInfo.getAddrSpace());
7674
7675 // Emit a library call.
7676 auto &Ctx = *getContext();
7677 const auto& DL = getDataLayout();
7678
7679 TargetLowering::CallLoweringInfo CLI(*this);
7680 // FIXME: pass in SDLoc
7681 CLI.setDebugLoc(dl).setChain(Chain);
7682
7683 ConstantSDNode *ConstantSrc = dyn_cast<ConstantSDNode>(Src);
7684 const bool SrcIsZero = ConstantSrc && ConstantSrc->isZero();
7685 const char *BzeroName = getTargetLoweringInfo().getLibcallName(RTLIB::BZERO);
7686
7687 // Helper function to create an Entry from Node and Type.
7688 const auto CreateEntry = [](SDValue Node, Type *Ty) {
7689 TargetLowering::ArgListEntry Entry;
7690 Entry.Node = Node;
7691 Entry.Ty = Ty;
7692 return Entry;
7693 };
7694
7695 // If zeroing out and bzero is present, use it.
7696 if (SrcIsZero && BzeroName) {
7697 TargetLowering::ArgListTy Args;
7698 Args.push_back(CreateEntry(Dst, Type::getInt8PtrTy(Ctx)));
7699 Args.push_back(CreateEntry(Size, DL.getIntPtrType(Ctx)));
7700 CLI.setLibCallee(
7701 TLI->getLibcallCallingConv(RTLIB::BZERO), Type::getVoidTy(Ctx),
7702 getExternalSymbol(BzeroName, TLI->getPointerTy(DL)), std::move(Args));
7703 } else {
7704 TargetLowering::ArgListTy Args;
7705 Args.push_back(CreateEntry(Dst, Type::getInt8PtrTy(Ctx)));
7706 Args.push_back(CreateEntry(Src, Src.getValueType().getTypeForEVT(Ctx)));
7707 Args.push_back(CreateEntry(Size, DL.getIntPtrType(Ctx)));
7708 CLI.setLibCallee(TLI->getLibcallCallingConv(RTLIB::MEMSET),
7709 Dst.getValueType().getTypeForEVT(Ctx),
7710 getExternalSymbol(TLI->getLibcallName(RTLIB::MEMSET),
7711 TLI->getPointerTy(DL)),
7712 std::move(Args));
7713 }
7714
7715 CLI.setDiscardResult().setTailCall(isTailCall);
7716
7717 std::pair<SDValue, SDValue> CallResult = TLI->LowerCallTo(CLI);
7718 return CallResult.second;
7719 }
7720
getAtomicMemset(SDValue Chain,const SDLoc & dl,SDValue Dst,SDValue Value,SDValue Size,Type * SizeTy,unsigned ElemSz,bool isTailCall,MachinePointerInfo DstPtrInfo)7721 SDValue SelectionDAG::getAtomicMemset(SDValue Chain, const SDLoc &dl,
7722 SDValue Dst, SDValue Value, SDValue Size,
7723 Type *SizeTy, unsigned ElemSz,
7724 bool isTailCall,
7725 MachinePointerInfo DstPtrInfo) {
7726 // Emit a library call.
7727 TargetLowering::ArgListTy Args;
7728 TargetLowering::ArgListEntry Entry;
7729 Entry.Ty = getDataLayout().getIntPtrType(*getContext());
7730 Entry.Node = Dst;
7731 Args.push_back(Entry);
7732
7733 Entry.Ty = Type::getInt8Ty(*getContext());
7734 Entry.Node = Value;
7735 Args.push_back(Entry);
7736
7737 Entry.Ty = SizeTy;
7738 Entry.Node = Size;
7739 Args.push_back(Entry);
7740
7741 RTLIB::Libcall LibraryCall =
7742 RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(ElemSz);
7743 if (LibraryCall == RTLIB::UNKNOWN_LIBCALL)
7744 report_fatal_error("Unsupported element size");
7745
7746 TargetLowering::CallLoweringInfo CLI(*this);
7747 CLI.setDebugLoc(dl)
7748 .setChain(Chain)
7749 .setLibCallee(TLI->getLibcallCallingConv(LibraryCall),
7750 Type::getVoidTy(*getContext()),
7751 getExternalSymbol(TLI->getLibcallName(LibraryCall),
7752 TLI->getPointerTy(getDataLayout())),
7753 std::move(Args))
7754 .setDiscardResult()
7755 .setTailCall(isTailCall);
7756
7757 std::pair<SDValue, SDValue> CallResult = TLI->LowerCallTo(CLI);
7758 return CallResult.second;
7759 }
7760
getAtomic(unsigned Opcode,const SDLoc & dl,EVT MemVT,SDVTList VTList,ArrayRef<SDValue> Ops,MachineMemOperand * MMO)7761 SDValue SelectionDAG::getAtomic(unsigned Opcode, const SDLoc &dl, EVT MemVT,
7762 SDVTList VTList, ArrayRef<SDValue> Ops,
7763 MachineMemOperand *MMO) {
7764 FoldingSetNodeID ID;
7765 ID.AddInteger(MemVT.getRawBits());
7766 AddNodeIDNode(ID, Opcode, VTList, Ops);
7767 ID.AddInteger(MMO->getPointerInfo().getAddrSpace());
7768 ID.AddInteger(MMO->getFlags());
7769 void* IP = nullptr;
7770 if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) {
7771 cast<AtomicSDNode>(E)->refineAlignment(MMO);
7772 return SDValue(E, 0);
7773 }
7774
7775 auto *N = newSDNode<AtomicSDNode>(Opcode, dl.getIROrder(), dl.getDebugLoc(),
7776 VTList, MemVT, MMO);
7777 createOperands(N, Ops);
7778
7779 CSEMap.InsertNode(N, IP);
7780 InsertNode(N);
7781 return SDValue(N, 0);
7782 }
7783
getAtomicCmpSwap(unsigned Opcode,const SDLoc & dl,EVT MemVT,SDVTList VTs,SDValue Chain,SDValue Ptr,SDValue Cmp,SDValue Swp,MachineMemOperand * MMO)7784 SDValue SelectionDAG::getAtomicCmpSwap(unsigned Opcode, const SDLoc &dl,
7785 EVT MemVT, SDVTList VTs, SDValue Chain,
7786 SDValue Ptr, SDValue Cmp, SDValue Swp,
7787 MachineMemOperand *MMO) {
7788 assert(Opcode == ISD::ATOMIC_CMP_SWAP ||
7789 Opcode == ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS);
7790 assert(Cmp.getValueType() == Swp.getValueType() && "Invalid Atomic Op Types");
7791
7792 SDValue Ops[] = {Chain, Ptr, Cmp, Swp};
7793 return getAtomic(Opcode, dl, MemVT, VTs, Ops, MMO);
7794 }
7795
getAtomic(unsigned Opcode,const SDLoc & dl,EVT MemVT,SDValue Chain,SDValue Ptr,SDValue Val,MachineMemOperand * MMO)7796 SDValue SelectionDAG::getAtomic(unsigned Opcode, const SDLoc &dl, EVT MemVT,
7797 SDValue Chain, SDValue Ptr, SDValue Val,
7798 MachineMemOperand *MMO) {
7799 assert((Opcode == ISD::ATOMIC_LOAD_ADD ||
7800 Opcode == ISD::ATOMIC_LOAD_SUB ||
7801 Opcode == ISD::ATOMIC_LOAD_AND ||
7802 Opcode == ISD::ATOMIC_LOAD_CLR ||
7803 Opcode == ISD::ATOMIC_LOAD_OR ||
7804 Opcode == ISD::ATOMIC_LOAD_XOR ||
7805 Opcode == ISD::ATOMIC_LOAD_NAND ||
7806 Opcode == ISD::ATOMIC_LOAD_MIN ||
7807 Opcode == ISD::ATOMIC_LOAD_MAX ||
7808 Opcode == ISD::ATOMIC_LOAD_UMIN ||
7809 Opcode == ISD::ATOMIC_LOAD_UMAX ||
7810 Opcode == ISD::ATOMIC_LOAD_FADD ||
7811 Opcode == ISD::ATOMIC_LOAD_FSUB ||
7812 Opcode == ISD::ATOMIC_LOAD_FMAX ||
7813 Opcode == ISD::ATOMIC_LOAD_FMIN ||
7814 Opcode == ISD::ATOMIC_LOAD_UINC_WRAP ||
7815 Opcode == ISD::ATOMIC_LOAD_UDEC_WRAP ||
7816 Opcode == ISD::ATOMIC_SWAP ||
7817 Opcode == ISD::ATOMIC_STORE) &&
7818 "Invalid Atomic Op");
7819
7820 EVT VT = Val.getValueType();
7821
7822 SDVTList VTs = Opcode == ISD::ATOMIC_STORE ? getVTList(MVT::Other) :
7823 getVTList(VT, MVT::Other);
7824 SDValue Ops[] = {Chain, Ptr, Val};
7825 return getAtomic(Opcode, dl, MemVT, VTs, Ops, MMO);
7826 }
7827
getAtomic(unsigned Opcode,const SDLoc & dl,EVT MemVT,EVT VT,SDValue Chain,SDValue Ptr,MachineMemOperand * MMO)7828 SDValue SelectionDAG::getAtomic(unsigned Opcode, const SDLoc &dl, EVT MemVT,
7829 EVT VT, SDValue Chain, SDValue Ptr,
7830 MachineMemOperand *MMO) {
7831 assert(Opcode == ISD::ATOMIC_LOAD && "Invalid Atomic Op");
7832
7833 SDVTList VTs = getVTList(VT, MVT::Other);
7834 SDValue Ops[] = {Chain, Ptr};
7835 return getAtomic(Opcode, dl, MemVT, VTs, Ops, MMO);
7836 }
7837
7838 /// getMergeValues - Create a MERGE_VALUES node from the given operands.
getMergeValues(ArrayRef<SDValue> Ops,const SDLoc & dl)7839 SDValue SelectionDAG::getMergeValues(ArrayRef<SDValue> Ops, const SDLoc &dl) {
7840 if (Ops.size() == 1)
7841 return Ops[0];
7842
7843 SmallVector<EVT, 4> VTs;
7844 VTs.reserve(Ops.size());
7845 for (const SDValue &Op : Ops)
7846 VTs.push_back(Op.getValueType());
7847 return getNode(ISD::MERGE_VALUES, dl, getVTList(VTs), Ops);
7848 }
7849
getMemIntrinsicNode(unsigned Opcode,const SDLoc & dl,SDVTList VTList,ArrayRef<SDValue> Ops,EVT MemVT,MachinePointerInfo PtrInfo,Align Alignment,MachineMemOperand::Flags Flags,uint64_t Size,const AAMDNodes & AAInfo)7850 SDValue SelectionDAG::getMemIntrinsicNode(
7851 unsigned Opcode, const SDLoc &dl, SDVTList VTList, ArrayRef<SDValue> Ops,
7852 EVT MemVT, MachinePointerInfo PtrInfo, Align Alignment,
7853 MachineMemOperand::Flags Flags, uint64_t Size, const AAMDNodes &AAInfo) {
7854 if (!Size && MemVT.isScalableVector())
7855 Size = MemoryLocation::UnknownSize;
7856 else if (!Size)
7857 Size = MemVT.getStoreSize();
7858
7859 MachineFunction &MF = getMachineFunction();
7860 MachineMemOperand *MMO =
7861 MF.getMachineMemOperand(PtrInfo, Flags, Size, Alignment, AAInfo);
7862
7863 return getMemIntrinsicNode(Opcode, dl, VTList, Ops, MemVT, MMO);
7864 }
7865
getMemIntrinsicNode(unsigned Opcode,const SDLoc & dl,SDVTList VTList,ArrayRef<SDValue> Ops,EVT MemVT,MachineMemOperand * MMO)7866 SDValue SelectionDAG::getMemIntrinsicNode(unsigned Opcode, const SDLoc &dl,
7867 SDVTList VTList,
7868 ArrayRef<SDValue> Ops, EVT MemVT,
7869 MachineMemOperand *MMO) {
7870 assert((Opcode == ISD::INTRINSIC_VOID ||
7871 Opcode == ISD::INTRINSIC_W_CHAIN ||
7872 Opcode == ISD::PREFETCH ||
7873 ((int)Opcode <= std::numeric_limits<int>::max() &&
7874 (int)Opcode >= ISD::FIRST_TARGET_MEMORY_OPCODE)) &&
7875 "Opcode is not a memory-accessing opcode!");
7876
7877 // Memoize the node unless it returns a flag.
7878 MemIntrinsicSDNode *N;
7879 if (VTList.VTs[VTList.NumVTs-1] != MVT::Glue) {
7880 FoldingSetNodeID ID;
7881 AddNodeIDNode(ID, Opcode, VTList, Ops);
7882 ID.AddInteger(getSyntheticNodeSubclassData<MemIntrinsicSDNode>(
7883 Opcode, dl.getIROrder(), VTList, MemVT, MMO));
7884 ID.AddInteger(MMO->getPointerInfo().getAddrSpace());
7885 ID.AddInteger(MMO->getFlags());
7886 void *IP = nullptr;
7887 if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) {
7888 cast<MemIntrinsicSDNode>(E)->refineAlignment(MMO);
7889 return SDValue(E, 0);
7890 }
7891
7892 N = newSDNode<MemIntrinsicSDNode>(Opcode, dl.getIROrder(), dl.getDebugLoc(),
7893 VTList, MemVT, MMO);
7894 createOperands(N, Ops);
7895
7896 CSEMap.InsertNode(N, IP);
7897 } else {
7898 N = newSDNode<MemIntrinsicSDNode>(Opcode, dl.getIROrder(), dl.getDebugLoc(),
7899 VTList, MemVT, MMO);
7900 createOperands(N, Ops);
7901 }
7902 InsertNode(N);
7903 SDValue V(N, 0);
7904 NewSDValueDbgMsg(V, "Creating new node: ", this);
7905 return V;
7906 }
7907
getLifetimeNode(bool IsStart,const SDLoc & dl,SDValue Chain,int FrameIndex,int64_t Size,int64_t Offset)7908 SDValue SelectionDAG::getLifetimeNode(bool IsStart, const SDLoc &dl,
7909 SDValue Chain, int FrameIndex,
7910 int64_t Size, int64_t Offset) {
7911 const unsigned Opcode = IsStart ? ISD::LIFETIME_START : ISD::LIFETIME_END;
7912 const auto VTs = getVTList(MVT::Other);
7913 SDValue Ops[2] = {
7914 Chain,
7915 getFrameIndex(FrameIndex,
7916 getTargetLoweringInfo().getFrameIndexTy(getDataLayout()),
7917 true)};
7918
7919 FoldingSetNodeID ID;
7920 AddNodeIDNode(ID, Opcode, VTs, Ops);
7921 ID.AddInteger(FrameIndex);
7922 ID.AddInteger(Size);
7923 ID.AddInteger(Offset);
7924 void *IP = nullptr;
7925 if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP))
7926 return SDValue(E, 0);
7927
7928 LifetimeSDNode *N = newSDNode<LifetimeSDNode>(
7929 Opcode, dl.getIROrder(), dl.getDebugLoc(), VTs, Size, Offset);
7930 createOperands(N, Ops);
7931 CSEMap.InsertNode(N, IP);
7932 InsertNode(N);
7933 SDValue V(N, 0);
7934 NewSDValueDbgMsg(V, "Creating new node: ", this);
7935 return V;
7936 }
7937
getPseudoProbeNode(const SDLoc & Dl,SDValue Chain,uint64_t Guid,uint64_t Index,uint32_t Attr)7938 SDValue SelectionDAG::getPseudoProbeNode(const SDLoc &Dl, SDValue Chain,
7939 uint64_t Guid, uint64_t Index,
7940 uint32_t Attr) {
7941 const unsigned Opcode = ISD::PSEUDO_PROBE;
7942 const auto VTs = getVTList(MVT::Other);
7943 SDValue Ops[] = {Chain};
7944 FoldingSetNodeID ID;
7945 AddNodeIDNode(ID, Opcode, VTs, Ops);
7946 ID.AddInteger(Guid);
7947 ID.AddInteger(Index);
7948 void *IP = nullptr;
7949 if (SDNode *E = FindNodeOrInsertPos(ID, Dl, IP))
7950 return SDValue(E, 0);
7951
7952 auto *N = newSDNode<PseudoProbeSDNode>(
7953 Opcode, Dl.getIROrder(), Dl.getDebugLoc(), VTs, Guid, Index, Attr);
7954 createOperands(N, Ops);
7955 CSEMap.InsertNode(N, IP);
7956 InsertNode(N);
7957 SDValue V(N, 0);
7958 NewSDValueDbgMsg(V, "Creating new node: ", this);
7959 return V;
7960 }
7961
7962 /// InferPointerInfo - If the specified ptr/offset is a frame index, infer a
7963 /// MachinePointerInfo record from it. This is particularly useful because the
7964 /// code generator has many cases where it doesn't bother passing in a
7965 /// MachinePointerInfo to getLoad or getStore when it has "FI+Cst".
InferPointerInfo(const MachinePointerInfo & Info,SelectionDAG & DAG,SDValue Ptr,int64_t Offset=0)7966 static MachinePointerInfo InferPointerInfo(const MachinePointerInfo &Info,
7967 SelectionDAG &DAG, SDValue Ptr,
7968 int64_t Offset = 0) {
7969 // If this is FI+Offset, we can model it.
7970 if (const FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Ptr))
7971 return MachinePointerInfo::getFixedStack(DAG.getMachineFunction(),
7972 FI->getIndex(), Offset);
7973
7974 // If this is (FI+Offset1)+Offset2, we can model it.
7975 if (Ptr.getOpcode() != ISD::ADD ||
7976 !isa<ConstantSDNode>(Ptr.getOperand(1)) ||
7977 !isa<FrameIndexSDNode>(Ptr.getOperand(0)))
7978 return Info;
7979
7980 int FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
7981 return MachinePointerInfo::getFixedStack(
7982 DAG.getMachineFunction(), FI,
7983 Offset + cast<ConstantSDNode>(Ptr.getOperand(1))->getSExtValue());
7984 }
7985
7986 /// InferPointerInfo - If the specified ptr/offset is a frame index, infer a
7987 /// MachinePointerInfo record from it. This is particularly useful because the
7988 /// code generator has many cases where it doesn't bother passing in a
7989 /// MachinePointerInfo to getLoad or getStore when it has "FI+Cst".
InferPointerInfo(const MachinePointerInfo & Info,SelectionDAG & DAG,SDValue Ptr,SDValue OffsetOp)7990 static MachinePointerInfo InferPointerInfo(const MachinePointerInfo &Info,
7991 SelectionDAG &DAG, SDValue Ptr,
7992 SDValue OffsetOp) {
7993 // If the 'Offset' value isn't a constant, we can't handle this.
7994 if (ConstantSDNode *OffsetNode = dyn_cast<ConstantSDNode>(OffsetOp))
7995 return InferPointerInfo(Info, DAG, Ptr, OffsetNode->getSExtValue());
7996 if (OffsetOp.isUndef())
7997 return InferPointerInfo(Info, DAG, Ptr);
7998 return Info;
7999 }
8000
getLoad(ISD::MemIndexedMode AM,ISD::LoadExtType ExtType,EVT VT,const SDLoc & dl,SDValue Chain,SDValue Ptr,SDValue Offset,MachinePointerInfo PtrInfo,EVT MemVT,Align Alignment,MachineMemOperand::Flags MMOFlags,const AAMDNodes & AAInfo,const MDNode * Ranges)8001 SDValue SelectionDAG::getLoad(ISD::MemIndexedMode AM, ISD::LoadExtType ExtType,
8002 EVT VT, const SDLoc &dl, SDValue Chain,
8003 SDValue Ptr, SDValue Offset,
8004 MachinePointerInfo PtrInfo, EVT MemVT,
8005 Align Alignment,
8006 MachineMemOperand::Flags MMOFlags,
8007 const AAMDNodes &AAInfo, const MDNode *Ranges) {
8008 assert(Chain.getValueType() == MVT::Other &&
8009 "Invalid chain type");
8010
8011 MMOFlags |= MachineMemOperand::MOLoad;
8012 assert((MMOFlags & MachineMemOperand::MOStore) == 0);
8013 // If we don't have a PtrInfo, infer the trivial frame index case to simplify
8014 // clients.
8015 if (PtrInfo.V.isNull())
8016 PtrInfo = InferPointerInfo(PtrInfo, *this, Ptr, Offset);
8017
8018 uint64_t Size = MemoryLocation::getSizeOrUnknown(MemVT.getStoreSize());
8019 MachineFunction &MF = getMachineFunction();
8020 MachineMemOperand *MMO = MF.getMachineMemOperand(PtrInfo, MMOFlags, Size,
8021 Alignment, AAInfo, Ranges);
8022 return getLoad(AM, ExtType, VT, dl, Chain, Ptr, Offset, MemVT, MMO);
8023 }
8024
getLoad(ISD::MemIndexedMode AM,ISD::LoadExtType ExtType,EVT VT,const SDLoc & dl,SDValue Chain,SDValue Ptr,SDValue Offset,EVT MemVT,MachineMemOperand * MMO)8025 SDValue SelectionDAG::getLoad(ISD::MemIndexedMode AM, ISD::LoadExtType ExtType,
8026 EVT VT, const SDLoc &dl, SDValue Chain,
8027 SDValue Ptr, SDValue Offset, EVT MemVT,
8028 MachineMemOperand *MMO) {
8029 if (VT == MemVT) {
8030 ExtType = ISD::NON_EXTLOAD;
8031 } else if (ExtType == ISD::NON_EXTLOAD) {
8032 assert(VT == MemVT && "Non-extending load from different memory type!");
8033 } else {
8034 // Extending load.
8035 assert(MemVT.getScalarType().bitsLT(VT.getScalarType()) &&
8036 "Should only be an extending load, not truncating!");
8037 assert(VT.isInteger() == MemVT.isInteger() &&
8038 "Cannot convert from FP to Int or Int -> FP!");
8039 assert(VT.isVector() == MemVT.isVector() &&
8040 "Cannot use an ext load to convert to or from a vector!");
8041 assert((!VT.isVector() ||
8042 VT.getVectorElementCount() == MemVT.getVectorElementCount()) &&
8043 "Cannot use an ext load to change the number of vector elements!");
8044 }
8045
8046 bool Indexed = AM != ISD::UNINDEXED;
8047 assert((Indexed || Offset.isUndef()) && "Unindexed load with an offset!");
8048
8049 SDVTList VTs = Indexed ?
8050 getVTList(VT, Ptr.getValueType(), MVT::Other) : getVTList(VT, MVT::Other);
8051 SDValue Ops[] = { Chain, Ptr, Offset };
8052 FoldingSetNodeID ID;
8053 AddNodeIDNode(ID, ISD::LOAD, VTs, Ops);
8054 ID.AddInteger(MemVT.getRawBits());
8055 ID.AddInteger(getSyntheticNodeSubclassData<LoadSDNode>(
8056 dl.getIROrder(), VTs, AM, ExtType, MemVT, MMO));
8057 ID.AddInteger(MMO->getPointerInfo().getAddrSpace());
8058 ID.AddInteger(MMO->getFlags());
8059 void *IP = nullptr;
8060 if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) {
8061 cast<LoadSDNode>(E)->refineAlignment(MMO);
8062 return SDValue(E, 0);
8063 }
8064 auto *N = newSDNode<LoadSDNode>(dl.getIROrder(), dl.getDebugLoc(), VTs, AM,
8065 ExtType, MemVT, MMO);
8066 createOperands(N, Ops);
8067
8068 CSEMap.InsertNode(N, IP);
8069 InsertNode(N);
8070 SDValue V(N, 0);
8071 NewSDValueDbgMsg(V, "Creating new node: ", this);
8072 return V;
8073 }
8074
getLoad(EVT VT,const SDLoc & dl,SDValue Chain,SDValue Ptr,MachinePointerInfo PtrInfo,MaybeAlign Alignment,MachineMemOperand::Flags MMOFlags,const AAMDNodes & AAInfo,const MDNode * Ranges)8075 SDValue SelectionDAG::getLoad(EVT VT, const SDLoc &dl, SDValue Chain,
8076 SDValue Ptr, MachinePointerInfo PtrInfo,
8077 MaybeAlign Alignment,
8078 MachineMemOperand::Flags MMOFlags,
8079 const AAMDNodes &AAInfo, const MDNode *Ranges) {
8080 SDValue Undef = getUNDEF(Ptr.getValueType());
8081 return getLoad(ISD::UNINDEXED, ISD::NON_EXTLOAD, VT, dl, Chain, Ptr, Undef,
8082 PtrInfo, VT, Alignment, MMOFlags, AAInfo, Ranges);
8083 }
8084
getLoad(EVT VT,const SDLoc & dl,SDValue Chain,SDValue Ptr,MachineMemOperand * MMO)8085 SDValue SelectionDAG::getLoad(EVT VT, const SDLoc &dl, SDValue Chain,
8086 SDValue Ptr, MachineMemOperand *MMO) {
8087 SDValue Undef = getUNDEF(Ptr.getValueType());
8088 return getLoad(ISD::UNINDEXED, ISD::NON_EXTLOAD, VT, dl, Chain, Ptr, Undef,
8089 VT, MMO);
8090 }
8091
getExtLoad(ISD::LoadExtType ExtType,const SDLoc & dl,EVT VT,SDValue Chain,SDValue Ptr,MachinePointerInfo PtrInfo,EVT MemVT,MaybeAlign Alignment,MachineMemOperand::Flags MMOFlags,const AAMDNodes & AAInfo)8092 SDValue SelectionDAG::getExtLoad(ISD::LoadExtType ExtType, const SDLoc &dl,
8093 EVT VT, SDValue Chain, SDValue Ptr,
8094 MachinePointerInfo PtrInfo, EVT MemVT,
8095 MaybeAlign Alignment,
8096 MachineMemOperand::Flags MMOFlags,
8097 const AAMDNodes &AAInfo) {
8098 SDValue Undef = getUNDEF(Ptr.getValueType());
8099 return getLoad(ISD::UNINDEXED, ExtType, VT, dl, Chain, Ptr, Undef, PtrInfo,
8100 MemVT, Alignment, MMOFlags, AAInfo);
8101 }
8102
getExtLoad(ISD::LoadExtType ExtType,const SDLoc & dl,EVT VT,SDValue Chain,SDValue Ptr,EVT MemVT,MachineMemOperand * MMO)8103 SDValue SelectionDAG::getExtLoad(ISD::LoadExtType ExtType, const SDLoc &dl,
8104 EVT VT, SDValue Chain, SDValue Ptr, EVT MemVT,
8105 MachineMemOperand *MMO) {
8106 SDValue Undef = getUNDEF(Ptr.getValueType());
8107 return getLoad(ISD::UNINDEXED, ExtType, VT, dl, Chain, Ptr, Undef,
8108 MemVT, MMO);
8109 }
8110
getIndexedLoad(SDValue OrigLoad,const SDLoc & dl,SDValue Base,SDValue Offset,ISD::MemIndexedMode AM)8111 SDValue SelectionDAG::getIndexedLoad(SDValue OrigLoad, const SDLoc &dl,
8112 SDValue Base, SDValue Offset,
8113 ISD::MemIndexedMode AM) {
8114 LoadSDNode *LD = cast<LoadSDNode>(OrigLoad);
8115 assert(LD->getOffset().isUndef() && "Load is already a indexed load!");
8116 // Don't propagate the invariant or dereferenceable flags.
8117 auto MMOFlags =
8118 LD->getMemOperand()->getFlags() &
8119 ~(MachineMemOperand::MOInvariant | MachineMemOperand::MODereferenceable);
8120 return getLoad(AM, LD->getExtensionType(), OrigLoad.getValueType(), dl,
8121 LD->getChain(), Base, Offset, LD->getPointerInfo(),
8122 LD->getMemoryVT(), LD->getAlign(), MMOFlags, LD->getAAInfo());
8123 }
8124
getStore(SDValue Chain,const SDLoc & dl,SDValue Val,SDValue Ptr,MachinePointerInfo PtrInfo,Align Alignment,MachineMemOperand::Flags MMOFlags,const AAMDNodes & AAInfo)8125 SDValue SelectionDAG::getStore(SDValue Chain, const SDLoc &dl, SDValue Val,
8126 SDValue Ptr, MachinePointerInfo PtrInfo,
8127 Align Alignment,
8128 MachineMemOperand::Flags MMOFlags,
8129 const AAMDNodes &AAInfo) {
8130 assert(Chain.getValueType() == MVT::Other && "Invalid chain type");
8131
8132 MMOFlags |= MachineMemOperand::MOStore;
8133 assert((MMOFlags & MachineMemOperand::MOLoad) == 0);
8134
8135 if (PtrInfo.V.isNull())
8136 PtrInfo = InferPointerInfo(PtrInfo, *this, Ptr);
8137
8138 MachineFunction &MF = getMachineFunction();
8139 uint64_t Size =
8140 MemoryLocation::getSizeOrUnknown(Val.getValueType().getStoreSize());
8141 MachineMemOperand *MMO =
8142 MF.getMachineMemOperand(PtrInfo, MMOFlags, Size, Alignment, AAInfo);
8143 return getStore(Chain, dl, Val, Ptr, MMO);
8144 }
8145
getStore(SDValue Chain,const SDLoc & dl,SDValue Val,SDValue Ptr,MachineMemOperand * MMO)8146 SDValue SelectionDAG::getStore(SDValue Chain, const SDLoc &dl, SDValue Val,
8147 SDValue Ptr, MachineMemOperand *MMO) {
8148 assert(Chain.getValueType() == MVT::Other &&
8149 "Invalid chain type");
8150 EVT VT = Val.getValueType();
8151 SDVTList VTs = getVTList(MVT::Other);
8152 SDValue Undef = getUNDEF(Ptr.getValueType());
8153 SDValue Ops[] = { Chain, Val, Ptr, Undef };
8154 FoldingSetNodeID ID;
8155 AddNodeIDNode(ID, ISD::STORE, VTs, Ops);
8156 ID.AddInteger(VT.getRawBits());
8157 ID.AddInteger(getSyntheticNodeSubclassData<StoreSDNode>(
8158 dl.getIROrder(), VTs, ISD::UNINDEXED, false, VT, MMO));
8159 ID.AddInteger(MMO->getPointerInfo().getAddrSpace());
8160 ID.AddInteger(MMO->getFlags());
8161 void *IP = nullptr;
8162 if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) {
8163 cast<StoreSDNode>(E)->refineAlignment(MMO);
8164 return SDValue(E, 0);
8165 }
8166 auto *N = newSDNode<StoreSDNode>(dl.getIROrder(), dl.getDebugLoc(), VTs,
8167 ISD::UNINDEXED, false, VT, MMO);
8168 createOperands(N, Ops);
8169
8170 CSEMap.InsertNode(N, IP);
8171 InsertNode(N);
8172 SDValue V(N, 0);
8173 NewSDValueDbgMsg(V, "Creating new node: ", this);
8174 return V;
8175 }
8176
getTruncStore(SDValue Chain,const SDLoc & dl,SDValue Val,SDValue Ptr,MachinePointerInfo PtrInfo,EVT SVT,Align Alignment,MachineMemOperand::Flags MMOFlags,const AAMDNodes & AAInfo)8177 SDValue SelectionDAG::getTruncStore(SDValue Chain, const SDLoc &dl, SDValue Val,
8178 SDValue Ptr, MachinePointerInfo PtrInfo,
8179 EVT SVT, Align Alignment,
8180 MachineMemOperand::Flags MMOFlags,
8181 const AAMDNodes &AAInfo) {
8182 assert(Chain.getValueType() == MVT::Other &&
8183 "Invalid chain type");
8184
8185 MMOFlags |= MachineMemOperand::MOStore;
8186 assert((MMOFlags & MachineMemOperand::MOLoad) == 0);
8187
8188 if (PtrInfo.V.isNull())
8189 PtrInfo = InferPointerInfo(PtrInfo, *this, Ptr);
8190
8191 MachineFunction &MF = getMachineFunction();
8192 MachineMemOperand *MMO = MF.getMachineMemOperand(
8193 PtrInfo, MMOFlags, MemoryLocation::getSizeOrUnknown(SVT.getStoreSize()),
8194 Alignment, AAInfo);
8195 return getTruncStore(Chain, dl, Val, Ptr, SVT, MMO);
8196 }
8197
getTruncStore(SDValue Chain,const SDLoc & dl,SDValue Val,SDValue Ptr,EVT SVT,MachineMemOperand * MMO)8198 SDValue SelectionDAG::getTruncStore(SDValue Chain, const SDLoc &dl, SDValue Val,
8199 SDValue Ptr, EVT SVT,
8200 MachineMemOperand *MMO) {
8201 EVT VT = Val.getValueType();
8202
8203 assert(Chain.getValueType() == MVT::Other &&
8204 "Invalid chain type");
8205 if (VT == SVT)
8206 return getStore(Chain, dl, Val, Ptr, MMO);
8207
8208 assert(SVT.getScalarType().bitsLT(VT.getScalarType()) &&
8209 "Should only be a truncating store, not extending!");
8210 assert(VT.isInteger() == SVT.isInteger() &&
8211 "Can't do FP-INT conversion!");
8212 assert(VT.isVector() == SVT.isVector() &&
8213 "Cannot use trunc store to convert to or from a vector!");
8214 assert((!VT.isVector() ||
8215 VT.getVectorElementCount() == SVT.getVectorElementCount()) &&
8216 "Cannot use trunc store to change the number of vector elements!");
8217
8218 SDVTList VTs = getVTList(MVT::Other);
8219 SDValue Undef = getUNDEF(Ptr.getValueType());
8220 SDValue Ops[] = { Chain, Val, Ptr, Undef };
8221 FoldingSetNodeID ID;
8222 AddNodeIDNode(ID, ISD::STORE, VTs, Ops);
8223 ID.AddInteger(SVT.getRawBits());
8224 ID.AddInteger(getSyntheticNodeSubclassData<StoreSDNode>(
8225 dl.getIROrder(), VTs, ISD::UNINDEXED, true, SVT, MMO));
8226 ID.AddInteger(MMO->getPointerInfo().getAddrSpace());
8227 ID.AddInteger(MMO->getFlags());
8228 void *IP = nullptr;
8229 if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) {
8230 cast<StoreSDNode>(E)->refineAlignment(MMO);
8231 return SDValue(E, 0);
8232 }
8233 auto *N = newSDNode<StoreSDNode>(dl.getIROrder(), dl.getDebugLoc(), VTs,
8234 ISD::UNINDEXED, true, SVT, MMO);
8235 createOperands(N, Ops);
8236
8237 CSEMap.InsertNode(N, IP);
8238 InsertNode(N);
8239 SDValue V(N, 0);
8240 NewSDValueDbgMsg(V, "Creating new node: ", this);
8241 return V;
8242 }
8243
getIndexedStore(SDValue OrigStore,const SDLoc & dl,SDValue Base,SDValue Offset,ISD::MemIndexedMode AM)8244 SDValue SelectionDAG::getIndexedStore(SDValue OrigStore, const SDLoc &dl,
8245 SDValue Base, SDValue Offset,
8246 ISD::MemIndexedMode AM) {
8247 StoreSDNode *ST = cast<StoreSDNode>(OrigStore);
8248 assert(ST->getOffset().isUndef() && "Store is already a indexed store!");
8249 SDVTList VTs = getVTList(Base.getValueType(), MVT::Other);
8250 SDValue Ops[] = { ST->getChain(), ST->getValue(), Base, Offset };
8251 FoldingSetNodeID ID;
8252 AddNodeIDNode(ID, ISD::STORE, VTs, Ops);
8253 ID.AddInteger(ST->getMemoryVT().getRawBits());
8254 ID.AddInteger(ST->getRawSubclassData());
8255 ID.AddInteger(ST->getPointerInfo().getAddrSpace());
8256 ID.AddInteger(ST->getMemOperand()->getFlags());
8257 void *IP = nullptr;
8258 if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP))
8259 return SDValue(E, 0);
8260
8261 auto *N = newSDNode<StoreSDNode>(dl.getIROrder(), dl.getDebugLoc(), VTs, AM,
8262 ST->isTruncatingStore(), ST->getMemoryVT(),
8263 ST->getMemOperand());
8264 createOperands(N, Ops);
8265
8266 CSEMap.InsertNode(N, IP);
8267 InsertNode(N);
8268 SDValue V(N, 0);
8269 NewSDValueDbgMsg(V, "Creating new node: ", this);
8270 return V;
8271 }
8272
getLoadVP(ISD::MemIndexedMode AM,ISD::LoadExtType ExtType,EVT VT,const SDLoc & dl,SDValue Chain,SDValue Ptr,SDValue Offset,SDValue Mask,SDValue EVL,MachinePointerInfo PtrInfo,EVT MemVT,Align Alignment,MachineMemOperand::Flags MMOFlags,const AAMDNodes & AAInfo,const MDNode * Ranges,bool IsExpanding)8273 SDValue SelectionDAG::getLoadVP(
8274 ISD::MemIndexedMode AM, ISD::LoadExtType ExtType, EVT VT, const SDLoc &dl,
8275 SDValue Chain, SDValue Ptr, SDValue Offset, SDValue Mask, SDValue EVL,
8276 MachinePointerInfo PtrInfo, EVT MemVT, Align Alignment,
8277 MachineMemOperand::Flags MMOFlags, const AAMDNodes &AAInfo,
8278 const MDNode *Ranges, bool IsExpanding) {
8279 assert(Chain.getValueType() == MVT::Other && "Invalid chain type");
8280
8281 MMOFlags |= MachineMemOperand::MOLoad;
8282 assert((MMOFlags & MachineMemOperand::MOStore) == 0);
8283 // If we don't have a PtrInfo, infer the trivial frame index case to simplify
8284 // clients.
8285 if (PtrInfo.V.isNull())
8286 PtrInfo = InferPointerInfo(PtrInfo, *this, Ptr, Offset);
8287
8288 uint64_t Size = MemoryLocation::getSizeOrUnknown(MemVT.getStoreSize());
8289 MachineFunction &MF = getMachineFunction();
8290 MachineMemOperand *MMO = MF.getMachineMemOperand(PtrInfo, MMOFlags, Size,
8291 Alignment, AAInfo, Ranges);
8292 return getLoadVP(AM, ExtType, VT, dl, Chain, Ptr, Offset, Mask, EVL, MemVT,
8293 MMO, IsExpanding);
8294 }
8295
getLoadVP(ISD::MemIndexedMode AM,ISD::LoadExtType ExtType,EVT VT,const SDLoc & dl,SDValue Chain,SDValue Ptr,SDValue Offset,SDValue Mask,SDValue EVL,EVT MemVT,MachineMemOperand * MMO,bool IsExpanding)8296 SDValue SelectionDAG::getLoadVP(ISD::MemIndexedMode AM,
8297 ISD::LoadExtType ExtType, EVT VT,
8298 const SDLoc &dl, SDValue Chain, SDValue Ptr,
8299 SDValue Offset, SDValue Mask, SDValue EVL,
8300 EVT MemVT, MachineMemOperand *MMO,
8301 bool IsExpanding) {
8302 bool Indexed = AM != ISD::UNINDEXED;
8303 assert((Indexed || Offset.isUndef()) && "Unindexed load with an offset!");
8304
8305 SDVTList VTs = Indexed ? getVTList(VT, Ptr.getValueType(), MVT::Other)
8306 : getVTList(VT, MVT::Other);
8307 SDValue Ops[] = {Chain, Ptr, Offset, Mask, EVL};
8308 FoldingSetNodeID ID;
8309 AddNodeIDNode(ID, ISD::VP_LOAD, VTs, Ops);
8310 ID.AddInteger(VT.getRawBits());
8311 ID.AddInteger(getSyntheticNodeSubclassData<VPLoadSDNode>(
8312 dl.getIROrder(), VTs, AM, ExtType, IsExpanding, MemVT, MMO));
8313 ID.AddInteger(MMO->getPointerInfo().getAddrSpace());
8314 ID.AddInteger(MMO->getFlags());
8315 void *IP = nullptr;
8316 if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) {
8317 cast<VPLoadSDNode>(E)->refineAlignment(MMO);
8318 return SDValue(E, 0);
8319 }
8320 auto *N = newSDNode<VPLoadSDNode>(dl.getIROrder(), dl.getDebugLoc(), VTs, AM,
8321 ExtType, IsExpanding, MemVT, MMO);
8322 createOperands(N, Ops);
8323
8324 CSEMap.InsertNode(N, IP);
8325 InsertNode(N);
8326 SDValue V(N, 0);
8327 NewSDValueDbgMsg(V, "Creating new node: ", this);
8328 return V;
8329 }
8330
getLoadVP(EVT VT,const SDLoc & dl,SDValue Chain,SDValue Ptr,SDValue Mask,SDValue EVL,MachinePointerInfo PtrInfo,MaybeAlign Alignment,MachineMemOperand::Flags MMOFlags,const AAMDNodes & AAInfo,const MDNode * Ranges,bool IsExpanding)8331 SDValue SelectionDAG::getLoadVP(EVT VT, const SDLoc &dl, SDValue Chain,
8332 SDValue Ptr, SDValue Mask, SDValue EVL,
8333 MachinePointerInfo PtrInfo,
8334 MaybeAlign Alignment,
8335 MachineMemOperand::Flags MMOFlags,
8336 const AAMDNodes &AAInfo, const MDNode *Ranges,
8337 bool IsExpanding) {
8338 SDValue Undef = getUNDEF(Ptr.getValueType());
8339 return getLoadVP(ISD::UNINDEXED, ISD::NON_EXTLOAD, VT, dl, Chain, Ptr, Undef,
8340 Mask, EVL, PtrInfo, VT, Alignment, MMOFlags, AAInfo, Ranges,
8341 IsExpanding);
8342 }
8343
getLoadVP(EVT VT,const SDLoc & dl,SDValue Chain,SDValue Ptr,SDValue Mask,SDValue EVL,MachineMemOperand * MMO,bool IsExpanding)8344 SDValue SelectionDAG::getLoadVP(EVT VT, const SDLoc &dl, SDValue Chain,
8345 SDValue Ptr, SDValue Mask, SDValue EVL,
8346 MachineMemOperand *MMO, bool IsExpanding) {
8347 SDValue Undef = getUNDEF(Ptr.getValueType());
8348 return getLoadVP(ISD::UNINDEXED, ISD::NON_EXTLOAD, VT, dl, Chain, Ptr, Undef,
8349 Mask, EVL, VT, MMO, IsExpanding);
8350 }
8351
getExtLoadVP(ISD::LoadExtType ExtType,const SDLoc & dl,EVT VT,SDValue Chain,SDValue Ptr,SDValue Mask,SDValue EVL,MachinePointerInfo PtrInfo,EVT MemVT,MaybeAlign Alignment,MachineMemOperand::Flags MMOFlags,const AAMDNodes & AAInfo,bool IsExpanding)8352 SDValue SelectionDAG::getExtLoadVP(ISD::LoadExtType ExtType, const SDLoc &dl,
8353 EVT VT, SDValue Chain, SDValue Ptr,
8354 SDValue Mask, SDValue EVL,
8355 MachinePointerInfo PtrInfo, EVT MemVT,
8356 MaybeAlign Alignment,
8357 MachineMemOperand::Flags MMOFlags,
8358 const AAMDNodes &AAInfo, bool IsExpanding) {
8359 SDValue Undef = getUNDEF(Ptr.getValueType());
8360 return getLoadVP(ISD::UNINDEXED, ExtType, VT, dl, Chain, Ptr, Undef, Mask,
8361 EVL, PtrInfo, MemVT, Alignment, MMOFlags, AAInfo, nullptr,
8362 IsExpanding);
8363 }
8364
getExtLoadVP(ISD::LoadExtType ExtType,const SDLoc & dl,EVT VT,SDValue Chain,SDValue Ptr,SDValue Mask,SDValue EVL,EVT MemVT,MachineMemOperand * MMO,bool IsExpanding)8365 SDValue SelectionDAG::getExtLoadVP(ISD::LoadExtType ExtType, const SDLoc &dl,
8366 EVT VT, SDValue Chain, SDValue Ptr,
8367 SDValue Mask, SDValue EVL, EVT MemVT,
8368 MachineMemOperand *MMO, bool IsExpanding) {
8369 SDValue Undef = getUNDEF(Ptr.getValueType());
8370 return getLoadVP(ISD::UNINDEXED, ExtType, VT, dl, Chain, Ptr, Undef, Mask,
8371 EVL, MemVT, MMO, IsExpanding);
8372 }
8373
getIndexedLoadVP(SDValue OrigLoad,const SDLoc & dl,SDValue Base,SDValue Offset,ISD::MemIndexedMode AM)8374 SDValue SelectionDAG::getIndexedLoadVP(SDValue OrigLoad, const SDLoc &dl,
8375 SDValue Base, SDValue Offset,
8376 ISD::MemIndexedMode AM) {
8377 auto *LD = cast<VPLoadSDNode>(OrigLoad);
8378 assert(LD->getOffset().isUndef() && "Load is already a indexed load!");
8379 // Don't propagate the invariant or dereferenceable flags.
8380 auto MMOFlags =
8381 LD->getMemOperand()->getFlags() &
8382 ~(MachineMemOperand::MOInvariant | MachineMemOperand::MODereferenceable);
8383 return getLoadVP(AM, LD->getExtensionType(), OrigLoad.getValueType(), dl,
8384 LD->getChain(), Base, Offset, LD->getMask(),
8385 LD->getVectorLength(), LD->getPointerInfo(),
8386 LD->getMemoryVT(), LD->getAlign(), MMOFlags, LD->getAAInfo(),
8387 nullptr, LD->isExpandingLoad());
8388 }
8389
getStoreVP(SDValue Chain,const SDLoc & dl,SDValue Val,SDValue Ptr,SDValue Offset,SDValue Mask,SDValue EVL,EVT MemVT,MachineMemOperand * MMO,ISD::MemIndexedMode AM,bool IsTruncating,bool IsCompressing)8390 SDValue SelectionDAG::getStoreVP(SDValue Chain, const SDLoc &dl, SDValue Val,
8391 SDValue Ptr, SDValue Offset, SDValue Mask,
8392 SDValue EVL, EVT MemVT, MachineMemOperand *MMO,
8393 ISD::MemIndexedMode AM, bool IsTruncating,
8394 bool IsCompressing) {
8395 assert(Chain.getValueType() == MVT::Other && "Invalid chain type");
8396 bool Indexed = AM != ISD::UNINDEXED;
8397 assert((Indexed || Offset.isUndef()) && "Unindexed vp_store with an offset!");
8398 SDVTList VTs = Indexed ? getVTList(Ptr.getValueType(), MVT::Other)
8399 : getVTList(MVT::Other);
8400 SDValue Ops[] = {Chain, Val, Ptr, Offset, Mask, EVL};
8401 FoldingSetNodeID ID;
8402 AddNodeIDNode(ID, ISD::VP_STORE, VTs, Ops);
8403 ID.AddInteger(MemVT.getRawBits());
8404 ID.AddInteger(getSyntheticNodeSubclassData<VPStoreSDNode>(
8405 dl.getIROrder(), VTs, AM, IsTruncating, IsCompressing, MemVT, MMO));
8406 ID.AddInteger(MMO->getPointerInfo().getAddrSpace());
8407 ID.AddInteger(MMO->getFlags());
8408 void *IP = nullptr;
8409 if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) {
8410 cast<VPStoreSDNode>(E)->refineAlignment(MMO);
8411 return SDValue(E, 0);
8412 }
8413 auto *N = newSDNode<VPStoreSDNode>(dl.getIROrder(), dl.getDebugLoc(), VTs, AM,
8414 IsTruncating, IsCompressing, MemVT, MMO);
8415 createOperands(N, Ops);
8416
8417 CSEMap.InsertNode(N, IP);
8418 InsertNode(N);
8419 SDValue V(N, 0);
8420 NewSDValueDbgMsg(V, "Creating new node: ", this);
8421 return V;
8422 }
8423
getTruncStoreVP(SDValue Chain,const SDLoc & dl,SDValue Val,SDValue Ptr,SDValue Mask,SDValue EVL,MachinePointerInfo PtrInfo,EVT SVT,Align Alignment,MachineMemOperand::Flags MMOFlags,const AAMDNodes & AAInfo,bool IsCompressing)8424 SDValue SelectionDAG::getTruncStoreVP(SDValue Chain, const SDLoc &dl,
8425 SDValue Val, SDValue Ptr, SDValue Mask,
8426 SDValue EVL, MachinePointerInfo PtrInfo,
8427 EVT SVT, Align Alignment,
8428 MachineMemOperand::Flags MMOFlags,
8429 const AAMDNodes &AAInfo,
8430 bool IsCompressing) {
8431 assert(Chain.getValueType() == MVT::Other && "Invalid chain type");
8432
8433 MMOFlags |= MachineMemOperand::MOStore;
8434 assert((MMOFlags & MachineMemOperand::MOLoad) == 0);
8435
8436 if (PtrInfo.V.isNull())
8437 PtrInfo = InferPointerInfo(PtrInfo, *this, Ptr);
8438
8439 MachineFunction &MF = getMachineFunction();
8440 MachineMemOperand *MMO = MF.getMachineMemOperand(
8441 PtrInfo, MMOFlags, MemoryLocation::getSizeOrUnknown(SVT.getStoreSize()),
8442 Alignment, AAInfo);
8443 return getTruncStoreVP(Chain, dl, Val, Ptr, Mask, EVL, SVT, MMO,
8444 IsCompressing);
8445 }
8446
getTruncStoreVP(SDValue Chain,const SDLoc & dl,SDValue Val,SDValue Ptr,SDValue Mask,SDValue EVL,EVT SVT,MachineMemOperand * MMO,bool IsCompressing)8447 SDValue SelectionDAG::getTruncStoreVP(SDValue Chain, const SDLoc &dl,
8448 SDValue Val, SDValue Ptr, SDValue Mask,
8449 SDValue EVL, EVT SVT,
8450 MachineMemOperand *MMO,
8451 bool IsCompressing) {
8452 EVT VT = Val.getValueType();
8453
8454 assert(Chain.getValueType() == MVT::Other && "Invalid chain type");
8455 if (VT == SVT)
8456 return getStoreVP(Chain, dl, Val, Ptr, getUNDEF(Ptr.getValueType()), Mask,
8457 EVL, VT, MMO, ISD::UNINDEXED,
8458 /*IsTruncating*/ false, IsCompressing);
8459
8460 assert(SVT.getScalarType().bitsLT(VT.getScalarType()) &&
8461 "Should only be a truncating store, not extending!");
8462 assert(VT.isInteger() == SVT.isInteger() && "Can't do FP-INT conversion!");
8463 assert(VT.isVector() == SVT.isVector() &&
8464 "Cannot use trunc store to convert to or from a vector!");
8465 assert((!VT.isVector() ||
8466 VT.getVectorElementCount() == SVT.getVectorElementCount()) &&
8467 "Cannot use trunc store to change the number of vector elements!");
8468
8469 SDVTList VTs = getVTList(MVT::Other);
8470 SDValue Undef = getUNDEF(Ptr.getValueType());
8471 SDValue Ops[] = {Chain, Val, Ptr, Undef, Mask, EVL};
8472 FoldingSetNodeID ID;
8473 AddNodeIDNode(ID, ISD::VP_STORE, VTs, Ops);
8474 ID.AddInteger(SVT.getRawBits());
8475 ID.AddInteger(getSyntheticNodeSubclassData<VPStoreSDNode>(
8476 dl.getIROrder(), VTs, ISD::UNINDEXED, true, IsCompressing, SVT, MMO));
8477 ID.AddInteger(MMO->getPointerInfo().getAddrSpace());
8478 ID.AddInteger(MMO->getFlags());
8479 void *IP = nullptr;
8480 if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) {
8481 cast<VPStoreSDNode>(E)->refineAlignment(MMO);
8482 return SDValue(E, 0);
8483 }
8484 auto *N =
8485 newSDNode<VPStoreSDNode>(dl.getIROrder(), dl.getDebugLoc(), VTs,
8486 ISD::UNINDEXED, true, IsCompressing, SVT, MMO);
8487 createOperands(N, Ops);
8488
8489 CSEMap.InsertNode(N, IP);
8490 InsertNode(N);
8491 SDValue V(N, 0);
8492 NewSDValueDbgMsg(V, "Creating new node: ", this);
8493 return V;
8494 }
8495
getIndexedStoreVP(SDValue OrigStore,const SDLoc & dl,SDValue Base,SDValue Offset,ISD::MemIndexedMode AM)8496 SDValue SelectionDAG::getIndexedStoreVP(SDValue OrigStore, const SDLoc &dl,
8497 SDValue Base, SDValue Offset,
8498 ISD::MemIndexedMode AM) {
8499 auto *ST = cast<VPStoreSDNode>(OrigStore);
8500 assert(ST->getOffset().isUndef() && "Store is already an indexed store!");
8501 SDVTList VTs = getVTList(Base.getValueType(), MVT::Other);
8502 SDValue Ops[] = {ST->getChain(), ST->getValue(), Base,
8503 Offset, ST->getMask(), ST->getVectorLength()};
8504 FoldingSetNodeID ID;
8505 AddNodeIDNode(ID, ISD::VP_STORE, VTs, Ops);
8506 ID.AddInteger(ST->getMemoryVT().getRawBits());
8507 ID.AddInteger(ST->getRawSubclassData());
8508 ID.AddInteger(ST->getPointerInfo().getAddrSpace());
8509 ID.AddInteger(ST->getMemOperand()->getFlags());
8510 void *IP = nullptr;
8511 if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP))
8512 return SDValue(E, 0);
8513
8514 auto *N = newSDNode<VPStoreSDNode>(
8515 dl.getIROrder(), dl.getDebugLoc(), VTs, AM, ST->isTruncatingStore(),
8516 ST->isCompressingStore(), ST->getMemoryVT(), ST->getMemOperand());
8517 createOperands(N, Ops);
8518
8519 CSEMap.InsertNode(N, IP);
8520 InsertNode(N);
8521 SDValue V(N, 0);
8522 NewSDValueDbgMsg(V, "Creating new node: ", this);
8523 return V;
8524 }
8525
getStridedLoadVP(ISD::MemIndexedMode AM,ISD::LoadExtType ExtType,EVT VT,const SDLoc & DL,SDValue Chain,SDValue Ptr,SDValue Offset,SDValue Stride,SDValue Mask,SDValue EVL,MachinePointerInfo PtrInfo,EVT MemVT,Align Alignment,MachineMemOperand::Flags MMOFlags,const AAMDNodes & AAInfo,const MDNode * Ranges,bool IsExpanding)8526 SDValue SelectionDAG::getStridedLoadVP(
8527 ISD::MemIndexedMode AM, ISD::LoadExtType ExtType, EVT VT, const SDLoc &DL,
8528 SDValue Chain, SDValue Ptr, SDValue Offset, SDValue Stride, SDValue Mask,
8529 SDValue EVL, MachinePointerInfo PtrInfo, EVT MemVT, Align Alignment,
8530 MachineMemOperand::Flags MMOFlags, const AAMDNodes &AAInfo,
8531 const MDNode *Ranges, bool IsExpanding) {
8532 assert(Chain.getValueType() == MVT::Other && "Invalid chain type");
8533
8534 MMOFlags |= MachineMemOperand::MOLoad;
8535 assert((MMOFlags & MachineMemOperand::MOStore) == 0);
8536 // If we don't have a PtrInfo, infer the trivial frame index case to simplify
8537 // clients.
8538 if (PtrInfo.V.isNull())
8539 PtrInfo = InferPointerInfo(PtrInfo, *this, Ptr, Offset);
8540
8541 uint64_t Size = MemoryLocation::UnknownSize;
8542 MachineFunction &MF = getMachineFunction();
8543 MachineMemOperand *MMO = MF.getMachineMemOperand(PtrInfo, MMOFlags, Size,
8544 Alignment, AAInfo, Ranges);
8545 return getStridedLoadVP(AM, ExtType, VT, DL, Chain, Ptr, Offset, Stride, Mask,
8546 EVL, MemVT, MMO, IsExpanding);
8547 }
8548
getStridedLoadVP(ISD::MemIndexedMode AM,ISD::LoadExtType ExtType,EVT VT,const SDLoc & DL,SDValue Chain,SDValue Ptr,SDValue Offset,SDValue Stride,SDValue Mask,SDValue EVL,EVT MemVT,MachineMemOperand * MMO,bool IsExpanding)8549 SDValue SelectionDAG::getStridedLoadVP(
8550 ISD::MemIndexedMode AM, ISD::LoadExtType ExtType, EVT VT, const SDLoc &DL,
8551 SDValue Chain, SDValue Ptr, SDValue Offset, SDValue Stride, SDValue Mask,
8552 SDValue EVL, EVT MemVT, MachineMemOperand *MMO, bool IsExpanding) {
8553 bool Indexed = AM != ISD::UNINDEXED;
8554 assert((Indexed || Offset.isUndef()) && "Unindexed load with an offset!");
8555
8556 SDValue Ops[] = {Chain, Ptr, Offset, Stride, Mask, EVL};
8557 SDVTList VTs = Indexed ? getVTList(VT, Ptr.getValueType(), MVT::Other)
8558 : getVTList(VT, MVT::Other);
8559 FoldingSetNodeID ID;
8560 AddNodeIDNode(ID, ISD::EXPERIMENTAL_VP_STRIDED_LOAD, VTs, Ops);
8561 ID.AddInteger(VT.getRawBits());
8562 ID.AddInteger(getSyntheticNodeSubclassData<VPStridedLoadSDNode>(
8563 DL.getIROrder(), VTs, AM, ExtType, IsExpanding, MemVT, MMO));
8564 ID.AddInteger(MMO->getPointerInfo().getAddrSpace());
8565
8566 void *IP = nullptr;
8567 if (SDNode *E = FindNodeOrInsertPos(ID, DL, IP)) {
8568 cast<VPStridedLoadSDNode>(E)->refineAlignment(MMO);
8569 return SDValue(E, 0);
8570 }
8571
8572 auto *N =
8573 newSDNode<VPStridedLoadSDNode>(DL.getIROrder(), DL.getDebugLoc(), VTs, AM,
8574 ExtType, IsExpanding, MemVT, MMO);
8575 createOperands(N, Ops);
8576 CSEMap.InsertNode(N, IP);
8577 InsertNode(N);
8578 SDValue V(N, 0);
8579 NewSDValueDbgMsg(V, "Creating new node: ", this);
8580 return V;
8581 }
8582
getStridedLoadVP(EVT VT,const SDLoc & DL,SDValue Chain,SDValue Ptr,SDValue Stride,SDValue Mask,SDValue EVL,MachinePointerInfo PtrInfo,MaybeAlign Alignment,MachineMemOperand::Flags MMOFlags,const AAMDNodes & AAInfo,const MDNode * Ranges,bool IsExpanding)8583 SDValue SelectionDAG::getStridedLoadVP(
8584 EVT VT, const SDLoc &DL, SDValue Chain, SDValue Ptr, SDValue Stride,
8585 SDValue Mask, SDValue EVL, MachinePointerInfo PtrInfo, MaybeAlign Alignment,
8586 MachineMemOperand::Flags MMOFlags, const AAMDNodes &AAInfo,
8587 const MDNode *Ranges, bool IsExpanding) {
8588 SDValue Undef = getUNDEF(Ptr.getValueType());
8589 return getStridedLoadVP(ISD::UNINDEXED, ISD::NON_EXTLOAD, VT, DL, Chain, Ptr,
8590 Undef, Stride, Mask, EVL, PtrInfo, VT, Alignment,
8591 MMOFlags, AAInfo, Ranges, IsExpanding);
8592 }
8593
getStridedLoadVP(EVT VT,const SDLoc & DL,SDValue Chain,SDValue Ptr,SDValue Stride,SDValue Mask,SDValue EVL,MachineMemOperand * MMO,bool IsExpanding)8594 SDValue SelectionDAG::getStridedLoadVP(EVT VT, const SDLoc &DL, SDValue Chain,
8595 SDValue Ptr, SDValue Stride,
8596 SDValue Mask, SDValue EVL,
8597 MachineMemOperand *MMO,
8598 bool IsExpanding) {
8599 SDValue Undef = getUNDEF(Ptr.getValueType());
8600 return getStridedLoadVP(ISD::UNINDEXED, ISD::NON_EXTLOAD, VT, DL, Chain, Ptr,
8601 Undef, Stride, Mask, EVL, VT, MMO, IsExpanding);
8602 }
8603
getExtStridedLoadVP(ISD::LoadExtType ExtType,const SDLoc & DL,EVT VT,SDValue Chain,SDValue Ptr,SDValue Stride,SDValue Mask,SDValue EVL,MachinePointerInfo PtrInfo,EVT MemVT,MaybeAlign Alignment,MachineMemOperand::Flags MMOFlags,const AAMDNodes & AAInfo,bool IsExpanding)8604 SDValue SelectionDAG::getExtStridedLoadVP(
8605 ISD::LoadExtType ExtType, const SDLoc &DL, EVT VT, SDValue Chain,
8606 SDValue Ptr, SDValue Stride, SDValue Mask, SDValue EVL,
8607 MachinePointerInfo PtrInfo, EVT MemVT, MaybeAlign Alignment,
8608 MachineMemOperand::Flags MMOFlags, const AAMDNodes &AAInfo,
8609 bool IsExpanding) {
8610 SDValue Undef = getUNDEF(Ptr.getValueType());
8611 return getStridedLoadVP(ISD::UNINDEXED, ExtType, VT, DL, Chain, Ptr, Undef,
8612 Stride, Mask, EVL, PtrInfo, MemVT, Alignment,
8613 MMOFlags, AAInfo, nullptr, IsExpanding);
8614 }
8615
getExtStridedLoadVP(ISD::LoadExtType ExtType,const SDLoc & DL,EVT VT,SDValue Chain,SDValue Ptr,SDValue Stride,SDValue Mask,SDValue EVL,EVT MemVT,MachineMemOperand * MMO,bool IsExpanding)8616 SDValue SelectionDAG::getExtStridedLoadVP(
8617 ISD::LoadExtType ExtType, const SDLoc &DL, EVT VT, SDValue Chain,
8618 SDValue Ptr, SDValue Stride, SDValue Mask, SDValue EVL, EVT MemVT,
8619 MachineMemOperand *MMO, bool IsExpanding) {
8620 SDValue Undef = getUNDEF(Ptr.getValueType());
8621 return getStridedLoadVP(ISD::UNINDEXED, ExtType, VT, DL, Chain, Ptr, Undef,
8622 Stride, Mask, EVL, MemVT, MMO, IsExpanding);
8623 }
8624
getIndexedStridedLoadVP(SDValue OrigLoad,const SDLoc & DL,SDValue Base,SDValue Offset,ISD::MemIndexedMode AM)8625 SDValue SelectionDAG::getIndexedStridedLoadVP(SDValue OrigLoad, const SDLoc &DL,
8626 SDValue Base, SDValue Offset,
8627 ISD::MemIndexedMode AM) {
8628 auto *SLD = cast<VPStridedLoadSDNode>(OrigLoad);
8629 assert(SLD->getOffset().isUndef() &&
8630 "Strided load is already a indexed load!");
8631 // Don't propagate the invariant or dereferenceable flags.
8632 auto MMOFlags =
8633 SLD->getMemOperand()->getFlags() &
8634 ~(MachineMemOperand::MOInvariant | MachineMemOperand::MODereferenceable);
8635 return getStridedLoadVP(
8636 AM, SLD->getExtensionType(), OrigLoad.getValueType(), DL, SLD->getChain(),
8637 Base, Offset, SLD->getStride(), SLD->getMask(), SLD->getVectorLength(),
8638 SLD->getPointerInfo(), SLD->getMemoryVT(), SLD->getAlign(), MMOFlags,
8639 SLD->getAAInfo(), nullptr, SLD->isExpandingLoad());
8640 }
8641
getStridedStoreVP(SDValue Chain,const SDLoc & DL,SDValue Val,SDValue Ptr,SDValue Offset,SDValue Stride,SDValue Mask,SDValue EVL,EVT MemVT,MachineMemOperand * MMO,ISD::MemIndexedMode AM,bool IsTruncating,bool IsCompressing)8642 SDValue SelectionDAG::getStridedStoreVP(SDValue Chain, const SDLoc &DL,
8643 SDValue Val, SDValue Ptr,
8644 SDValue Offset, SDValue Stride,
8645 SDValue Mask, SDValue EVL, EVT MemVT,
8646 MachineMemOperand *MMO,
8647 ISD::MemIndexedMode AM,
8648 bool IsTruncating, bool IsCompressing) {
8649 assert(Chain.getValueType() == MVT::Other && "Invalid chain type");
8650 bool Indexed = AM != ISD::UNINDEXED;
8651 assert((Indexed || Offset.isUndef()) && "Unindexed vp_store with an offset!");
8652 SDVTList VTs = Indexed ? getVTList(Ptr.getValueType(), MVT::Other)
8653 : getVTList(MVT::Other);
8654 SDValue Ops[] = {Chain, Val, Ptr, Offset, Stride, Mask, EVL};
8655 FoldingSetNodeID ID;
8656 AddNodeIDNode(ID, ISD::EXPERIMENTAL_VP_STRIDED_STORE, VTs, Ops);
8657 ID.AddInteger(MemVT.getRawBits());
8658 ID.AddInteger(getSyntheticNodeSubclassData<VPStridedStoreSDNode>(
8659 DL.getIROrder(), VTs, AM, IsTruncating, IsCompressing, MemVT, MMO));
8660 ID.AddInteger(MMO->getPointerInfo().getAddrSpace());
8661 void *IP = nullptr;
8662 if (SDNode *E = FindNodeOrInsertPos(ID, DL, IP)) {
8663 cast<VPStridedStoreSDNode>(E)->refineAlignment(MMO);
8664 return SDValue(E, 0);
8665 }
8666 auto *N = newSDNode<VPStridedStoreSDNode>(DL.getIROrder(), DL.getDebugLoc(),
8667 VTs, AM, IsTruncating,
8668 IsCompressing, MemVT, MMO);
8669 createOperands(N, Ops);
8670
8671 CSEMap.InsertNode(N, IP);
8672 InsertNode(N);
8673 SDValue V(N, 0);
8674 NewSDValueDbgMsg(V, "Creating new node: ", this);
8675 return V;
8676 }
8677
getTruncStridedStoreVP(SDValue Chain,const SDLoc & DL,SDValue Val,SDValue Ptr,SDValue Stride,SDValue Mask,SDValue EVL,MachinePointerInfo PtrInfo,EVT SVT,Align Alignment,MachineMemOperand::Flags MMOFlags,const AAMDNodes & AAInfo,bool IsCompressing)8678 SDValue SelectionDAG::getTruncStridedStoreVP(
8679 SDValue Chain, const SDLoc &DL, SDValue Val, SDValue Ptr, SDValue Stride,
8680 SDValue Mask, SDValue EVL, MachinePointerInfo PtrInfo, EVT SVT,
8681 Align Alignment, MachineMemOperand::Flags MMOFlags, const AAMDNodes &AAInfo,
8682 bool IsCompressing) {
8683 assert(Chain.getValueType() == MVT::Other && "Invalid chain type");
8684
8685 MMOFlags |= MachineMemOperand::MOStore;
8686 assert((MMOFlags & MachineMemOperand::MOLoad) == 0);
8687
8688 if (PtrInfo.V.isNull())
8689 PtrInfo = InferPointerInfo(PtrInfo, *this, Ptr);
8690
8691 MachineFunction &MF = getMachineFunction();
8692 MachineMemOperand *MMO = MF.getMachineMemOperand(
8693 PtrInfo, MMOFlags, MemoryLocation::UnknownSize, Alignment, AAInfo);
8694 return getTruncStridedStoreVP(Chain, DL, Val, Ptr, Stride, Mask, EVL, SVT,
8695 MMO, IsCompressing);
8696 }
8697
getTruncStridedStoreVP(SDValue Chain,const SDLoc & DL,SDValue Val,SDValue Ptr,SDValue Stride,SDValue Mask,SDValue EVL,EVT SVT,MachineMemOperand * MMO,bool IsCompressing)8698 SDValue SelectionDAG::getTruncStridedStoreVP(SDValue Chain, const SDLoc &DL,
8699 SDValue Val, SDValue Ptr,
8700 SDValue Stride, SDValue Mask,
8701 SDValue EVL, EVT SVT,
8702 MachineMemOperand *MMO,
8703 bool IsCompressing) {
8704 EVT VT = Val.getValueType();
8705
8706 assert(Chain.getValueType() == MVT::Other && "Invalid chain type");
8707 if (VT == SVT)
8708 return getStridedStoreVP(Chain, DL, Val, Ptr, getUNDEF(Ptr.getValueType()),
8709 Stride, Mask, EVL, VT, MMO, ISD::UNINDEXED,
8710 /*IsTruncating*/ false, IsCompressing);
8711
8712 assert(SVT.getScalarType().bitsLT(VT.getScalarType()) &&
8713 "Should only be a truncating store, not extending!");
8714 assert(VT.isInteger() == SVT.isInteger() && "Can't do FP-INT conversion!");
8715 assert(VT.isVector() == SVT.isVector() &&
8716 "Cannot use trunc store to convert to or from a vector!");
8717 assert((!VT.isVector() ||
8718 VT.getVectorElementCount() == SVT.getVectorElementCount()) &&
8719 "Cannot use trunc store to change the number of vector elements!");
8720
8721 SDVTList VTs = getVTList(MVT::Other);
8722 SDValue Undef = getUNDEF(Ptr.getValueType());
8723 SDValue Ops[] = {Chain, Val, Ptr, Undef, Stride, Mask, EVL};
8724 FoldingSetNodeID ID;
8725 AddNodeIDNode(ID, ISD::EXPERIMENTAL_VP_STRIDED_STORE, VTs, Ops);
8726 ID.AddInteger(SVT.getRawBits());
8727 ID.AddInteger(getSyntheticNodeSubclassData<VPStridedStoreSDNode>(
8728 DL.getIROrder(), VTs, ISD::UNINDEXED, true, IsCompressing, SVT, MMO));
8729 ID.AddInteger(MMO->getPointerInfo().getAddrSpace());
8730 void *IP = nullptr;
8731 if (SDNode *E = FindNodeOrInsertPos(ID, DL, IP)) {
8732 cast<VPStridedStoreSDNode>(E)->refineAlignment(MMO);
8733 return SDValue(E, 0);
8734 }
8735 auto *N = newSDNode<VPStridedStoreSDNode>(DL.getIROrder(), DL.getDebugLoc(),
8736 VTs, ISD::UNINDEXED, true,
8737 IsCompressing, SVT, MMO);
8738 createOperands(N, Ops);
8739
8740 CSEMap.InsertNode(N, IP);
8741 InsertNode(N);
8742 SDValue V(N, 0);
8743 NewSDValueDbgMsg(V, "Creating new node: ", this);
8744 return V;
8745 }
8746
getIndexedStridedStoreVP(SDValue OrigStore,const SDLoc & DL,SDValue Base,SDValue Offset,ISD::MemIndexedMode AM)8747 SDValue SelectionDAG::getIndexedStridedStoreVP(SDValue OrigStore,
8748 const SDLoc &DL, SDValue Base,
8749 SDValue Offset,
8750 ISD::MemIndexedMode AM) {
8751 auto *SST = cast<VPStridedStoreSDNode>(OrigStore);
8752 assert(SST->getOffset().isUndef() &&
8753 "Strided store is already an indexed store!");
8754 SDVTList VTs = getVTList(Base.getValueType(), MVT::Other);
8755 SDValue Ops[] = {
8756 SST->getChain(), SST->getValue(), Base, Offset, SST->getStride(),
8757 SST->getMask(), SST->getVectorLength()};
8758 FoldingSetNodeID ID;
8759 AddNodeIDNode(ID, ISD::EXPERIMENTAL_VP_STRIDED_STORE, VTs, Ops);
8760 ID.AddInteger(SST->getMemoryVT().getRawBits());
8761 ID.AddInteger(SST->getRawSubclassData());
8762 ID.AddInteger(SST->getPointerInfo().getAddrSpace());
8763 void *IP = nullptr;
8764 if (SDNode *E = FindNodeOrInsertPos(ID, DL, IP))
8765 return SDValue(E, 0);
8766
8767 auto *N = newSDNode<VPStridedStoreSDNode>(
8768 DL.getIROrder(), DL.getDebugLoc(), VTs, AM, SST->isTruncatingStore(),
8769 SST->isCompressingStore(), SST->getMemoryVT(), SST->getMemOperand());
8770 createOperands(N, Ops);
8771
8772 CSEMap.InsertNode(N, IP);
8773 InsertNode(N);
8774 SDValue V(N, 0);
8775 NewSDValueDbgMsg(V, "Creating new node: ", this);
8776 return V;
8777 }
8778
getGatherVP(SDVTList VTs,EVT VT,const SDLoc & dl,ArrayRef<SDValue> Ops,MachineMemOperand * MMO,ISD::MemIndexType IndexType)8779 SDValue SelectionDAG::getGatherVP(SDVTList VTs, EVT VT, const SDLoc &dl,
8780 ArrayRef<SDValue> Ops, MachineMemOperand *MMO,
8781 ISD::MemIndexType IndexType) {
8782 assert(Ops.size() == 6 && "Incompatible number of operands");
8783
8784 FoldingSetNodeID ID;
8785 AddNodeIDNode(ID, ISD::VP_GATHER, VTs, Ops);
8786 ID.AddInteger(VT.getRawBits());
8787 ID.AddInteger(getSyntheticNodeSubclassData<VPGatherSDNode>(
8788 dl.getIROrder(), VTs, VT, MMO, IndexType));
8789 ID.AddInteger(MMO->getPointerInfo().getAddrSpace());
8790 ID.AddInteger(MMO->getFlags());
8791 void *IP = nullptr;
8792 if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) {
8793 cast<VPGatherSDNode>(E)->refineAlignment(MMO);
8794 return SDValue(E, 0);
8795 }
8796
8797 auto *N = newSDNode<VPGatherSDNode>(dl.getIROrder(), dl.getDebugLoc(), VTs,
8798 VT, MMO, IndexType);
8799 createOperands(N, Ops);
8800
8801 assert(N->getMask().getValueType().getVectorElementCount() ==
8802 N->getValueType(0).getVectorElementCount() &&
8803 "Vector width mismatch between mask and data");
8804 assert(N->getIndex().getValueType().getVectorElementCount().isScalable() ==
8805 N->getValueType(0).getVectorElementCount().isScalable() &&
8806 "Scalable flags of index and data do not match");
8807 assert(ElementCount::isKnownGE(
8808 N->getIndex().getValueType().getVectorElementCount(),
8809 N->getValueType(0).getVectorElementCount()) &&
8810 "Vector width mismatch between index and data");
8811 assert(isa<ConstantSDNode>(N->getScale()) &&
8812 cast<ConstantSDNode>(N->getScale())->getAPIntValue().isPowerOf2() &&
8813 "Scale should be a constant power of 2");
8814
8815 CSEMap.InsertNode(N, IP);
8816 InsertNode(N);
8817 SDValue V(N, 0);
8818 NewSDValueDbgMsg(V, "Creating new node: ", this);
8819 return V;
8820 }
8821
getScatterVP(SDVTList VTs,EVT VT,const SDLoc & dl,ArrayRef<SDValue> Ops,MachineMemOperand * MMO,ISD::MemIndexType IndexType)8822 SDValue SelectionDAG::getScatterVP(SDVTList VTs, EVT VT, const SDLoc &dl,
8823 ArrayRef<SDValue> Ops,
8824 MachineMemOperand *MMO,
8825 ISD::MemIndexType IndexType) {
8826 assert(Ops.size() == 7 && "Incompatible number of operands");
8827
8828 FoldingSetNodeID ID;
8829 AddNodeIDNode(ID, ISD::VP_SCATTER, VTs, Ops);
8830 ID.AddInteger(VT.getRawBits());
8831 ID.AddInteger(getSyntheticNodeSubclassData<VPScatterSDNode>(
8832 dl.getIROrder(), VTs, VT, MMO, IndexType));
8833 ID.AddInteger(MMO->getPointerInfo().getAddrSpace());
8834 ID.AddInteger(MMO->getFlags());
8835 void *IP = nullptr;
8836 if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) {
8837 cast<VPScatterSDNode>(E)->refineAlignment(MMO);
8838 return SDValue(E, 0);
8839 }
8840 auto *N = newSDNode<VPScatterSDNode>(dl.getIROrder(), dl.getDebugLoc(), VTs,
8841 VT, MMO, IndexType);
8842 createOperands(N, Ops);
8843
8844 assert(N->getMask().getValueType().getVectorElementCount() ==
8845 N->getValue().getValueType().getVectorElementCount() &&
8846 "Vector width mismatch between mask and data");
8847 assert(
8848 N->getIndex().getValueType().getVectorElementCount().isScalable() ==
8849 N->getValue().getValueType().getVectorElementCount().isScalable() &&
8850 "Scalable flags of index and data do not match");
8851 assert(ElementCount::isKnownGE(
8852 N->getIndex().getValueType().getVectorElementCount(),
8853 N->getValue().getValueType().getVectorElementCount()) &&
8854 "Vector width mismatch between index and data");
8855 assert(isa<ConstantSDNode>(N->getScale()) &&
8856 cast<ConstantSDNode>(N->getScale())->getAPIntValue().isPowerOf2() &&
8857 "Scale should be a constant power of 2");
8858
8859 CSEMap.InsertNode(N, IP);
8860 InsertNode(N);
8861 SDValue V(N, 0);
8862 NewSDValueDbgMsg(V, "Creating new node: ", this);
8863 return V;
8864 }
8865
getMaskedLoad(EVT VT,const SDLoc & dl,SDValue Chain,SDValue Base,SDValue Offset,SDValue Mask,SDValue PassThru,EVT MemVT,MachineMemOperand * MMO,ISD::MemIndexedMode AM,ISD::LoadExtType ExtTy,bool isExpanding)8866 SDValue SelectionDAG::getMaskedLoad(EVT VT, const SDLoc &dl, SDValue Chain,
8867 SDValue Base, SDValue Offset, SDValue Mask,
8868 SDValue PassThru, EVT MemVT,
8869 MachineMemOperand *MMO,
8870 ISD::MemIndexedMode AM,
8871 ISD::LoadExtType ExtTy, bool isExpanding) {
8872 bool Indexed = AM != ISD::UNINDEXED;
8873 assert((Indexed || Offset.isUndef()) &&
8874 "Unindexed masked load with an offset!");
8875 SDVTList VTs = Indexed ? getVTList(VT, Base.getValueType(), MVT::Other)
8876 : getVTList(VT, MVT::Other);
8877 SDValue Ops[] = {Chain, Base, Offset, Mask, PassThru};
8878 FoldingSetNodeID ID;
8879 AddNodeIDNode(ID, ISD::MLOAD, VTs, Ops);
8880 ID.AddInteger(MemVT.getRawBits());
8881 ID.AddInteger(getSyntheticNodeSubclassData<MaskedLoadSDNode>(
8882 dl.getIROrder(), VTs, AM, ExtTy, isExpanding, MemVT, MMO));
8883 ID.AddInteger(MMO->getPointerInfo().getAddrSpace());
8884 ID.AddInteger(MMO->getFlags());
8885 void *IP = nullptr;
8886 if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) {
8887 cast<MaskedLoadSDNode>(E)->refineAlignment(MMO);
8888 return SDValue(E, 0);
8889 }
8890 auto *N = newSDNode<MaskedLoadSDNode>(dl.getIROrder(), dl.getDebugLoc(), VTs,
8891 AM, ExtTy, isExpanding, MemVT, MMO);
8892 createOperands(N, Ops);
8893
8894 CSEMap.InsertNode(N, IP);
8895 InsertNode(N);
8896 SDValue V(N, 0);
8897 NewSDValueDbgMsg(V, "Creating new node: ", this);
8898 return V;
8899 }
8900
getIndexedMaskedLoad(SDValue OrigLoad,const SDLoc & dl,SDValue Base,SDValue Offset,ISD::MemIndexedMode AM)8901 SDValue SelectionDAG::getIndexedMaskedLoad(SDValue OrigLoad, const SDLoc &dl,
8902 SDValue Base, SDValue Offset,
8903 ISD::MemIndexedMode AM) {
8904 MaskedLoadSDNode *LD = cast<MaskedLoadSDNode>(OrigLoad);
8905 assert(LD->getOffset().isUndef() && "Masked load is already a indexed load!");
8906 return getMaskedLoad(OrigLoad.getValueType(), dl, LD->getChain(), Base,
8907 Offset, LD->getMask(), LD->getPassThru(),
8908 LD->getMemoryVT(), LD->getMemOperand(), AM,
8909 LD->getExtensionType(), LD->isExpandingLoad());
8910 }
8911
getMaskedStore(SDValue Chain,const SDLoc & dl,SDValue Val,SDValue Base,SDValue Offset,SDValue Mask,EVT MemVT,MachineMemOperand * MMO,ISD::MemIndexedMode AM,bool IsTruncating,bool IsCompressing)8912 SDValue SelectionDAG::getMaskedStore(SDValue Chain, const SDLoc &dl,
8913 SDValue Val, SDValue Base, SDValue Offset,
8914 SDValue Mask, EVT MemVT,
8915 MachineMemOperand *MMO,
8916 ISD::MemIndexedMode AM, bool IsTruncating,
8917 bool IsCompressing) {
8918 assert(Chain.getValueType() == MVT::Other &&
8919 "Invalid chain type");
8920 bool Indexed = AM != ISD::UNINDEXED;
8921 assert((Indexed || Offset.isUndef()) &&
8922 "Unindexed masked store with an offset!");
8923 SDVTList VTs = Indexed ? getVTList(Base.getValueType(), MVT::Other)
8924 : getVTList(MVT::Other);
8925 SDValue Ops[] = {Chain, Val, Base, Offset, Mask};
8926 FoldingSetNodeID ID;
8927 AddNodeIDNode(ID, ISD::MSTORE, VTs, Ops);
8928 ID.AddInteger(MemVT.getRawBits());
8929 ID.AddInteger(getSyntheticNodeSubclassData<MaskedStoreSDNode>(
8930 dl.getIROrder(), VTs, AM, IsTruncating, IsCompressing, MemVT, MMO));
8931 ID.AddInteger(MMO->getPointerInfo().getAddrSpace());
8932 ID.AddInteger(MMO->getFlags());
8933 void *IP = nullptr;
8934 if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) {
8935 cast<MaskedStoreSDNode>(E)->refineAlignment(MMO);
8936 return SDValue(E, 0);
8937 }
8938 auto *N =
8939 newSDNode<MaskedStoreSDNode>(dl.getIROrder(), dl.getDebugLoc(), VTs, AM,
8940 IsTruncating, IsCompressing, MemVT, MMO);
8941 createOperands(N, Ops);
8942
8943 CSEMap.InsertNode(N, IP);
8944 InsertNode(N);
8945 SDValue V(N, 0);
8946 NewSDValueDbgMsg(V, "Creating new node: ", this);
8947 return V;
8948 }
8949
getIndexedMaskedStore(SDValue OrigStore,const SDLoc & dl,SDValue Base,SDValue Offset,ISD::MemIndexedMode AM)8950 SDValue SelectionDAG::getIndexedMaskedStore(SDValue OrigStore, const SDLoc &dl,
8951 SDValue Base, SDValue Offset,
8952 ISD::MemIndexedMode AM) {
8953 MaskedStoreSDNode *ST = cast<MaskedStoreSDNode>(OrigStore);
8954 assert(ST->getOffset().isUndef() &&
8955 "Masked store is already a indexed store!");
8956 return getMaskedStore(ST->getChain(), dl, ST->getValue(), Base, Offset,
8957 ST->getMask(), ST->getMemoryVT(), ST->getMemOperand(),
8958 AM, ST->isTruncatingStore(), ST->isCompressingStore());
8959 }
8960
getMaskedGather(SDVTList VTs,EVT MemVT,const SDLoc & dl,ArrayRef<SDValue> Ops,MachineMemOperand * MMO,ISD::MemIndexType IndexType,ISD::LoadExtType ExtTy)8961 SDValue SelectionDAG::getMaskedGather(SDVTList VTs, EVT MemVT, const SDLoc &dl,
8962 ArrayRef<SDValue> Ops,
8963 MachineMemOperand *MMO,
8964 ISD::MemIndexType IndexType,
8965 ISD::LoadExtType ExtTy) {
8966 assert(Ops.size() == 6 && "Incompatible number of operands");
8967
8968 FoldingSetNodeID ID;
8969 AddNodeIDNode(ID, ISD::MGATHER, VTs, Ops);
8970 ID.AddInteger(MemVT.getRawBits());
8971 ID.AddInteger(getSyntheticNodeSubclassData<MaskedGatherSDNode>(
8972 dl.getIROrder(), VTs, MemVT, MMO, IndexType, ExtTy));
8973 ID.AddInteger(MMO->getPointerInfo().getAddrSpace());
8974 ID.AddInteger(MMO->getFlags());
8975 void *IP = nullptr;
8976 if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) {
8977 cast<MaskedGatherSDNode>(E)->refineAlignment(MMO);
8978 return SDValue(E, 0);
8979 }
8980
8981 auto *N = newSDNode<MaskedGatherSDNode>(dl.getIROrder(), dl.getDebugLoc(),
8982 VTs, MemVT, MMO, IndexType, ExtTy);
8983 createOperands(N, Ops);
8984
8985 assert(N->getPassThru().getValueType() == N->getValueType(0) &&
8986 "Incompatible type of the PassThru value in MaskedGatherSDNode");
8987 assert(N->getMask().getValueType().getVectorElementCount() ==
8988 N->getValueType(0).getVectorElementCount() &&
8989 "Vector width mismatch between mask and data");
8990 assert(N->getIndex().getValueType().getVectorElementCount().isScalable() ==
8991 N->getValueType(0).getVectorElementCount().isScalable() &&
8992 "Scalable flags of index and data do not match");
8993 assert(ElementCount::isKnownGE(
8994 N->getIndex().getValueType().getVectorElementCount(),
8995 N->getValueType(0).getVectorElementCount()) &&
8996 "Vector width mismatch between index and data");
8997 assert(isa<ConstantSDNode>(N->getScale()) &&
8998 cast<ConstantSDNode>(N->getScale())->getAPIntValue().isPowerOf2() &&
8999 "Scale should be a constant power of 2");
9000
9001 CSEMap.InsertNode(N, IP);
9002 InsertNode(N);
9003 SDValue V(N, 0);
9004 NewSDValueDbgMsg(V, "Creating new node: ", this);
9005 return V;
9006 }
9007
getMaskedScatter(SDVTList VTs,EVT MemVT,const SDLoc & dl,ArrayRef<SDValue> Ops,MachineMemOperand * MMO,ISD::MemIndexType IndexType,bool IsTrunc)9008 SDValue SelectionDAG::getMaskedScatter(SDVTList VTs, EVT MemVT, const SDLoc &dl,
9009 ArrayRef<SDValue> Ops,
9010 MachineMemOperand *MMO,
9011 ISD::MemIndexType IndexType,
9012 bool IsTrunc) {
9013 assert(Ops.size() == 6 && "Incompatible number of operands");
9014
9015 FoldingSetNodeID ID;
9016 AddNodeIDNode(ID, ISD::MSCATTER, VTs, Ops);
9017 ID.AddInteger(MemVT.getRawBits());
9018 ID.AddInteger(getSyntheticNodeSubclassData<MaskedScatterSDNode>(
9019 dl.getIROrder(), VTs, MemVT, MMO, IndexType, IsTrunc));
9020 ID.AddInteger(MMO->getPointerInfo().getAddrSpace());
9021 ID.AddInteger(MMO->getFlags());
9022 void *IP = nullptr;
9023 if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) {
9024 cast<MaskedScatterSDNode>(E)->refineAlignment(MMO);
9025 return SDValue(E, 0);
9026 }
9027
9028 auto *N = newSDNode<MaskedScatterSDNode>(dl.getIROrder(), dl.getDebugLoc(),
9029 VTs, MemVT, MMO, IndexType, IsTrunc);
9030 createOperands(N, Ops);
9031
9032 assert(N->getMask().getValueType().getVectorElementCount() ==
9033 N->getValue().getValueType().getVectorElementCount() &&
9034 "Vector width mismatch between mask and data");
9035 assert(
9036 N->getIndex().getValueType().getVectorElementCount().isScalable() ==
9037 N->getValue().getValueType().getVectorElementCount().isScalable() &&
9038 "Scalable flags of index and data do not match");
9039 assert(ElementCount::isKnownGE(
9040 N->getIndex().getValueType().getVectorElementCount(),
9041 N->getValue().getValueType().getVectorElementCount()) &&
9042 "Vector width mismatch between index and data");
9043 assert(isa<ConstantSDNode>(N->getScale()) &&
9044 cast<ConstantSDNode>(N->getScale())->getAPIntValue().isPowerOf2() &&
9045 "Scale should be a constant power of 2");
9046
9047 CSEMap.InsertNode(N, IP);
9048 InsertNode(N);
9049 SDValue V(N, 0);
9050 NewSDValueDbgMsg(V, "Creating new node: ", this);
9051 return V;
9052 }
9053
simplifySelect(SDValue Cond,SDValue T,SDValue F)9054 SDValue SelectionDAG::simplifySelect(SDValue Cond, SDValue T, SDValue F) {
9055 // select undef, T, F --> T (if T is a constant), otherwise F
9056 // select, ?, undef, F --> F
9057 // select, ?, T, undef --> T
9058 if (Cond.isUndef())
9059 return isConstantValueOfAnyType(T) ? T : F;
9060 if (T.isUndef())
9061 return F;
9062 if (F.isUndef())
9063 return T;
9064
9065 // select true, T, F --> T
9066 // select false, T, F --> F
9067 if (auto *CondC = dyn_cast<ConstantSDNode>(Cond))
9068 return CondC->isZero() ? F : T;
9069
9070 // TODO: This should simplify VSELECT with non-zero constant condition using
9071 // something like this (but check boolean contents to be complete?):
9072 if (ConstantSDNode *CondC = isConstOrConstSplat(Cond, /*AllowUndefs*/ false,
9073 /*AllowTruncation*/ true))
9074 if (CondC->isZero())
9075 return F;
9076
9077 // select ?, T, T --> T
9078 if (T == F)
9079 return T;
9080
9081 return SDValue();
9082 }
9083
simplifyShift(SDValue X,SDValue Y)9084 SDValue SelectionDAG::simplifyShift(SDValue X, SDValue Y) {
9085 // shift undef, Y --> 0 (can always assume that the undef value is 0)
9086 if (X.isUndef())
9087 return getConstant(0, SDLoc(X.getNode()), X.getValueType());
9088 // shift X, undef --> undef (because it may shift by the bitwidth)
9089 if (Y.isUndef())
9090 return getUNDEF(X.getValueType());
9091
9092 // shift 0, Y --> 0
9093 // shift X, 0 --> X
9094 if (isNullOrNullSplat(X) || isNullOrNullSplat(Y))
9095 return X;
9096
9097 // shift X, C >= bitwidth(X) --> undef
9098 // All vector elements must be too big (or undef) to avoid partial undefs.
9099 auto isShiftTooBig = [X](ConstantSDNode *Val) {
9100 return !Val || Val->getAPIntValue().uge(X.getScalarValueSizeInBits());
9101 };
9102 if (ISD::matchUnaryPredicate(Y, isShiftTooBig, true))
9103 return getUNDEF(X.getValueType());
9104
9105 return SDValue();
9106 }
9107
simplifyFPBinop(unsigned Opcode,SDValue X,SDValue Y,SDNodeFlags Flags)9108 SDValue SelectionDAG::simplifyFPBinop(unsigned Opcode, SDValue X, SDValue Y,
9109 SDNodeFlags Flags) {
9110 // If this operation has 'nnan' or 'ninf' and at least 1 disallowed operand
9111 // (an undef operand can be chosen to be Nan/Inf), then the result of this
9112 // operation is poison. That result can be relaxed to undef.
9113 ConstantFPSDNode *XC = isConstOrConstSplatFP(X, /* AllowUndefs */ true);
9114 ConstantFPSDNode *YC = isConstOrConstSplatFP(Y, /* AllowUndefs */ true);
9115 bool HasNan = (XC && XC->getValueAPF().isNaN()) ||
9116 (YC && YC->getValueAPF().isNaN());
9117 bool HasInf = (XC && XC->getValueAPF().isInfinity()) ||
9118 (YC && YC->getValueAPF().isInfinity());
9119
9120 if (Flags.hasNoNaNs() && (HasNan || X.isUndef() || Y.isUndef()))
9121 return getUNDEF(X.getValueType());
9122
9123 if (Flags.hasNoInfs() && (HasInf || X.isUndef() || Y.isUndef()))
9124 return getUNDEF(X.getValueType());
9125
9126 if (!YC)
9127 return SDValue();
9128
9129 // X + -0.0 --> X
9130 if (Opcode == ISD::FADD)
9131 if (YC->getValueAPF().isNegZero())
9132 return X;
9133
9134 // X - +0.0 --> X
9135 if (Opcode == ISD::FSUB)
9136 if (YC->getValueAPF().isPosZero())
9137 return X;
9138
9139 // X * 1.0 --> X
9140 // X / 1.0 --> X
9141 if (Opcode == ISD::FMUL || Opcode == ISD::FDIV)
9142 if (YC->getValueAPF().isExactlyValue(1.0))
9143 return X;
9144
9145 // X * 0.0 --> 0.0
9146 if (Opcode == ISD::FMUL && Flags.hasNoNaNs() && Flags.hasNoSignedZeros())
9147 if (YC->getValueAPF().isZero())
9148 return getConstantFP(0.0, SDLoc(Y), Y.getValueType());
9149
9150 return SDValue();
9151 }
9152
getVAArg(EVT VT,const SDLoc & dl,SDValue Chain,SDValue Ptr,SDValue SV,unsigned Align)9153 SDValue SelectionDAG::getVAArg(EVT VT, const SDLoc &dl, SDValue Chain,
9154 SDValue Ptr, SDValue SV, unsigned Align) {
9155 SDValue Ops[] = { Chain, Ptr, SV, getTargetConstant(Align, dl, MVT::i32) };
9156 return getNode(ISD::VAARG, dl, getVTList(VT, MVT::Other), Ops);
9157 }
9158
getNode(unsigned Opcode,const SDLoc & DL,EVT VT,ArrayRef<SDUse> Ops)9159 SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
9160 ArrayRef<SDUse> Ops) {
9161 switch (Ops.size()) {
9162 case 0: return getNode(Opcode, DL, VT);
9163 case 1: return getNode(Opcode, DL, VT, static_cast<const SDValue>(Ops[0]));
9164 case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1]);
9165 case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2]);
9166 default: break;
9167 }
9168
9169 // Copy from an SDUse array into an SDValue array for use with
9170 // the regular getNode logic.
9171 SmallVector<SDValue, 8> NewOps(Ops.begin(), Ops.end());
9172 return getNode(Opcode, DL, VT, NewOps);
9173 }
9174
getNode(unsigned Opcode,const SDLoc & DL,EVT VT,ArrayRef<SDValue> Ops)9175 SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
9176 ArrayRef<SDValue> Ops) {
9177 SDNodeFlags Flags;
9178 if (Inserter)
9179 Flags = Inserter->getFlags();
9180 return getNode(Opcode, DL, VT, Ops, Flags);
9181 }
9182
getNode(unsigned Opcode,const SDLoc & DL,EVT VT,ArrayRef<SDValue> Ops,const SDNodeFlags Flags)9183 SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
9184 ArrayRef<SDValue> Ops, const SDNodeFlags Flags) {
9185 unsigned NumOps = Ops.size();
9186 switch (NumOps) {
9187 case 0: return getNode(Opcode, DL, VT);
9188 case 1: return getNode(Opcode, DL, VT, Ops[0], Flags);
9189 case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Flags);
9190 case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2], Flags);
9191 default: break;
9192 }
9193
9194 #ifndef NDEBUG
9195 for (const auto &Op : Ops)
9196 assert(Op.getOpcode() != ISD::DELETED_NODE &&
9197 "Operand is DELETED_NODE!");
9198 #endif
9199
9200 switch (Opcode) {
9201 default: break;
9202 case ISD::BUILD_VECTOR:
9203 // Attempt to simplify BUILD_VECTOR.
9204 if (SDValue V = FoldBUILD_VECTOR(DL, VT, Ops, *this))
9205 return V;
9206 break;
9207 case ISD::CONCAT_VECTORS:
9208 if (SDValue V = foldCONCAT_VECTORS(DL, VT, Ops, *this))
9209 return V;
9210 break;
9211 case ISD::SELECT_CC:
9212 assert(NumOps == 5 && "SELECT_CC takes 5 operands!");
9213 assert(Ops[0].getValueType() == Ops[1].getValueType() &&
9214 "LHS and RHS of condition must have same type!");
9215 assert(Ops[2].getValueType() == Ops[3].getValueType() &&
9216 "True and False arms of SelectCC must have same type!");
9217 assert(Ops[2].getValueType() == VT &&
9218 "select_cc node must be of same type as true and false value!");
9219 assert((!Ops[0].getValueType().isVector() ||
9220 Ops[0].getValueType().getVectorElementCount() ==
9221 VT.getVectorElementCount()) &&
9222 "Expected select_cc with vector result to have the same sized "
9223 "comparison type!");
9224 break;
9225 case ISD::BR_CC:
9226 assert(NumOps == 5 && "BR_CC takes 5 operands!");
9227 assert(Ops[2].getValueType() == Ops[3].getValueType() &&
9228 "LHS/RHS of comparison should match types!");
9229 break;
9230 case ISD::VP_ADD:
9231 case ISD::VP_SUB:
9232 // If it is VP_ADD/VP_SUB mask operation then turn it to VP_XOR
9233 if (VT.isVector() && VT.getVectorElementType() == MVT::i1)
9234 Opcode = ISD::VP_XOR;
9235 break;
9236 case ISD::VP_MUL:
9237 // If it is VP_MUL mask operation then turn it to VP_AND
9238 if (VT.isVector() && VT.getVectorElementType() == MVT::i1)
9239 Opcode = ISD::VP_AND;
9240 break;
9241 case ISD::VP_REDUCE_MUL:
9242 // If it is VP_REDUCE_MUL mask operation then turn it to VP_REDUCE_AND
9243 if (VT == MVT::i1)
9244 Opcode = ISD::VP_REDUCE_AND;
9245 break;
9246 case ISD::VP_REDUCE_ADD:
9247 // If it is VP_REDUCE_ADD mask operation then turn it to VP_REDUCE_XOR
9248 if (VT == MVT::i1)
9249 Opcode = ISD::VP_REDUCE_XOR;
9250 break;
9251 case ISD::VP_REDUCE_SMAX:
9252 case ISD::VP_REDUCE_UMIN:
9253 // If it is VP_REDUCE_SMAX/VP_REDUCE_UMIN mask operation then turn it to
9254 // VP_REDUCE_AND.
9255 if (VT == MVT::i1)
9256 Opcode = ISD::VP_REDUCE_AND;
9257 break;
9258 case ISD::VP_REDUCE_SMIN:
9259 case ISD::VP_REDUCE_UMAX:
9260 // If it is VP_REDUCE_SMIN/VP_REDUCE_UMAX mask operation then turn it to
9261 // VP_REDUCE_OR.
9262 if (VT == MVT::i1)
9263 Opcode = ISD::VP_REDUCE_OR;
9264 break;
9265 }
9266
9267 // Memoize nodes.
9268 SDNode *N;
9269 SDVTList VTs = getVTList(VT);
9270
9271 if (VT != MVT::Glue) {
9272 FoldingSetNodeID ID;
9273 AddNodeIDNode(ID, Opcode, VTs, Ops);
9274 void *IP = nullptr;
9275
9276 if (SDNode *E = FindNodeOrInsertPos(ID, DL, IP))
9277 return SDValue(E, 0);
9278
9279 N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(), VTs);
9280 createOperands(N, Ops);
9281
9282 CSEMap.InsertNode(N, IP);
9283 } else {
9284 N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(), VTs);
9285 createOperands(N, Ops);
9286 }
9287
9288 N->setFlags(Flags);
9289 InsertNode(N);
9290 SDValue V(N, 0);
9291 NewSDValueDbgMsg(V, "Creating new node: ", this);
9292 return V;
9293 }
9294
getNode(unsigned Opcode,const SDLoc & DL,ArrayRef<EVT> ResultTys,ArrayRef<SDValue> Ops)9295 SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL,
9296 ArrayRef<EVT> ResultTys, ArrayRef<SDValue> Ops) {
9297 return getNode(Opcode, DL, getVTList(ResultTys), Ops);
9298 }
9299
getNode(unsigned Opcode,const SDLoc & DL,SDVTList VTList,ArrayRef<SDValue> Ops)9300 SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList,
9301 ArrayRef<SDValue> Ops) {
9302 SDNodeFlags Flags;
9303 if (Inserter)
9304 Flags = Inserter->getFlags();
9305 return getNode(Opcode, DL, VTList, Ops, Flags);
9306 }
9307
getNode(unsigned Opcode,const SDLoc & DL,SDVTList VTList,ArrayRef<SDValue> Ops,const SDNodeFlags Flags)9308 SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList,
9309 ArrayRef<SDValue> Ops, const SDNodeFlags Flags) {
9310 if (VTList.NumVTs == 1)
9311 return getNode(Opcode, DL, VTList.VTs[0], Ops, Flags);
9312
9313 #ifndef NDEBUG
9314 for (const auto &Op : Ops)
9315 assert(Op.getOpcode() != ISD::DELETED_NODE &&
9316 "Operand is DELETED_NODE!");
9317 #endif
9318
9319 switch (Opcode) {
9320 case ISD::SADDO:
9321 case ISD::UADDO:
9322 case ISD::SSUBO:
9323 case ISD::USUBO: {
9324 assert(VTList.NumVTs == 2 && Ops.size() == 2 &&
9325 "Invalid add/sub overflow op!");
9326 assert(VTList.VTs[0].isInteger() && VTList.VTs[1].isInteger() &&
9327 Ops[0].getValueType() == Ops[1].getValueType() &&
9328 Ops[0].getValueType() == VTList.VTs[0] &&
9329 "Binary operator types must match!");
9330 SDValue N1 = Ops[0], N2 = Ops[1];
9331 canonicalizeCommutativeBinop(Opcode, N1, N2);
9332
9333 // (X +- 0) -> X with zero-overflow.
9334 ConstantSDNode *N2CV = isConstOrConstSplat(N2, /*AllowUndefs*/ false,
9335 /*AllowTruncation*/ true);
9336 if (N2CV && N2CV->isZero()) {
9337 SDValue ZeroOverFlow = getConstant(0, DL, VTList.VTs[1]);
9338 return getNode(ISD::MERGE_VALUES, DL, VTList, {N1, ZeroOverFlow}, Flags);
9339 }
9340 break;
9341 }
9342 case ISD::SMUL_LOHI:
9343 case ISD::UMUL_LOHI: {
9344 assert(VTList.NumVTs == 2 && Ops.size() == 2 && "Invalid mul lo/hi op!");
9345 assert(VTList.VTs[0].isInteger() && VTList.VTs[0] == VTList.VTs[1] &&
9346 VTList.VTs[0] == Ops[0].getValueType() &&
9347 VTList.VTs[0] == Ops[1].getValueType() &&
9348 "Binary operator types must match!");
9349 break;
9350 }
9351 case ISD::STRICT_FP_EXTEND:
9352 assert(VTList.NumVTs == 2 && Ops.size() == 2 &&
9353 "Invalid STRICT_FP_EXTEND!");
9354 assert(VTList.VTs[0].isFloatingPoint() &&
9355 Ops[1].getValueType().isFloatingPoint() && "Invalid FP cast!");
9356 assert(VTList.VTs[0].isVector() == Ops[1].getValueType().isVector() &&
9357 "STRICT_FP_EXTEND result type should be vector iff the operand "
9358 "type is vector!");
9359 assert((!VTList.VTs[0].isVector() ||
9360 VTList.VTs[0].getVectorNumElements() ==
9361 Ops[1].getValueType().getVectorNumElements()) &&
9362 "Vector element count mismatch!");
9363 assert(Ops[1].getValueType().bitsLT(VTList.VTs[0]) &&
9364 "Invalid fpext node, dst <= src!");
9365 break;
9366 case ISD::STRICT_FP_ROUND:
9367 assert(VTList.NumVTs == 2 && Ops.size() == 3 && "Invalid STRICT_FP_ROUND!");
9368 assert(VTList.VTs[0].isVector() == Ops[1].getValueType().isVector() &&
9369 "STRICT_FP_ROUND result type should be vector iff the operand "
9370 "type is vector!");
9371 assert((!VTList.VTs[0].isVector() ||
9372 VTList.VTs[0].getVectorNumElements() ==
9373 Ops[1].getValueType().getVectorNumElements()) &&
9374 "Vector element count mismatch!");
9375 assert(VTList.VTs[0].isFloatingPoint() &&
9376 Ops[1].getValueType().isFloatingPoint() &&
9377 VTList.VTs[0].bitsLT(Ops[1].getValueType()) &&
9378 isa<ConstantSDNode>(Ops[2]) &&
9379 (cast<ConstantSDNode>(Ops[2])->getZExtValue() == 0 ||
9380 cast<ConstantSDNode>(Ops[2])->getZExtValue() == 1) &&
9381 "Invalid STRICT_FP_ROUND!");
9382 break;
9383 #if 0
9384 // FIXME: figure out how to safely handle things like
9385 // int foo(int x) { return 1 << (x & 255); }
9386 // int bar() { return foo(256); }
9387 case ISD::SRA_PARTS:
9388 case ISD::SRL_PARTS:
9389 case ISD::SHL_PARTS:
9390 if (N3.getOpcode() == ISD::SIGN_EXTEND_INREG &&
9391 cast<VTSDNode>(N3.getOperand(1))->getVT() != MVT::i1)
9392 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0));
9393 else if (N3.getOpcode() == ISD::AND)
9394 if (ConstantSDNode *AndRHS = dyn_cast<ConstantSDNode>(N3.getOperand(1))) {
9395 // If the and is only masking out bits that cannot effect the shift,
9396 // eliminate the and.
9397 unsigned NumBits = VT.getScalarSizeInBits()*2;
9398 if ((AndRHS->getValue() & (NumBits-1)) == NumBits-1)
9399 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0));
9400 }
9401 break;
9402 #endif
9403 }
9404
9405 // Memoize the node unless it returns a flag.
9406 SDNode *N;
9407 if (VTList.VTs[VTList.NumVTs-1] != MVT::Glue) {
9408 FoldingSetNodeID ID;
9409 AddNodeIDNode(ID, Opcode, VTList, Ops);
9410 void *IP = nullptr;
9411 if (SDNode *E = FindNodeOrInsertPos(ID, DL, IP))
9412 return SDValue(E, 0);
9413
9414 N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(), VTList);
9415 createOperands(N, Ops);
9416 CSEMap.InsertNode(N, IP);
9417 } else {
9418 N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(), VTList);
9419 createOperands(N, Ops);
9420 }
9421
9422 N->setFlags(Flags);
9423 InsertNode(N);
9424 SDValue V(N, 0);
9425 NewSDValueDbgMsg(V, "Creating new node: ", this);
9426 return V;
9427 }
9428
getNode(unsigned Opcode,const SDLoc & DL,SDVTList VTList)9429 SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL,
9430 SDVTList VTList) {
9431 return getNode(Opcode, DL, VTList, std::nullopt);
9432 }
9433
getNode(unsigned Opcode,const SDLoc & DL,SDVTList VTList,SDValue N1)9434 SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList,
9435 SDValue N1) {
9436 SDValue Ops[] = { N1 };
9437 return getNode(Opcode, DL, VTList, Ops);
9438 }
9439
getNode(unsigned Opcode,const SDLoc & DL,SDVTList VTList,SDValue N1,SDValue N2)9440 SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList,
9441 SDValue N1, SDValue N2) {
9442 SDValue Ops[] = { N1, N2 };
9443 return getNode(Opcode, DL, VTList, Ops);
9444 }
9445
getNode(unsigned Opcode,const SDLoc & DL,SDVTList VTList,SDValue N1,SDValue N2,SDValue N3)9446 SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList,
9447 SDValue N1, SDValue N2, SDValue N3) {
9448 SDValue Ops[] = { N1, N2, N3 };
9449 return getNode(Opcode, DL, VTList, Ops);
9450 }
9451
getNode(unsigned Opcode,const SDLoc & DL,SDVTList VTList,SDValue N1,SDValue N2,SDValue N3,SDValue N4)9452 SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList,
9453 SDValue N1, SDValue N2, SDValue N3, SDValue N4) {
9454 SDValue Ops[] = { N1, N2, N3, N4 };
9455 return getNode(Opcode, DL, VTList, Ops);
9456 }
9457
getNode(unsigned Opcode,const SDLoc & DL,SDVTList VTList,SDValue N1,SDValue N2,SDValue N3,SDValue N4,SDValue N5)9458 SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList,
9459 SDValue N1, SDValue N2, SDValue N3, SDValue N4,
9460 SDValue N5) {
9461 SDValue Ops[] = { N1, N2, N3, N4, N5 };
9462 return getNode(Opcode, DL, VTList, Ops);
9463 }
9464
getVTList(EVT VT)9465 SDVTList SelectionDAG::getVTList(EVT VT) {
9466 return makeVTList(SDNode::getValueTypeList(VT), 1);
9467 }
9468
getVTList(EVT VT1,EVT VT2)9469 SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2) {
9470 FoldingSetNodeID ID;
9471 ID.AddInteger(2U);
9472 ID.AddInteger(VT1.getRawBits());
9473 ID.AddInteger(VT2.getRawBits());
9474
9475 void *IP = nullptr;
9476 SDVTListNode *Result = VTListMap.FindNodeOrInsertPos(ID, IP);
9477 if (!Result) {
9478 EVT *Array = Allocator.Allocate<EVT>(2);
9479 Array[0] = VT1;
9480 Array[1] = VT2;
9481 Result = new (Allocator) SDVTListNode(ID.Intern(Allocator), Array, 2);
9482 VTListMap.InsertNode(Result, IP);
9483 }
9484 return Result->getSDVTList();
9485 }
9486
getVTList(EVT VT1,EVT VT2,EVT VT3)9487 SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3) {
9488 FoldingSetNodeID ID;
9489 ID.AddInteger(3U);
9490 ID.AddInteger(VT1.getRawBits());
9491 ID.AddInteger(VT2.getRawBits());
9492 ID.AddInteger(VT3.getRawBits());
9493
9494 void *IP = nullptr;
9495 SDVTListNode *Result = VTListMap.FindNodeOrInsertPos(ID, IP);
9496 if (!Result) {
9497 EVT *Array = Allocator.Allocate<EVT>(3);
9498 Array[0] = VT1;
9499 Array[1] = VT2;
9500 Array[2] = VT3;
9501 Result = new (Allocator) SDVTListNode(ID.Intern(Allocator), Array, 3);
9502 VTListMap.InsertNode(Result, IP);
9503 }
9504 return Result->getSDVTList();
9505 }
9506
getVTList(EVT VT1,EVT VT2,EVT VT3,EVT VT4)9507 SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3, EVT VT4) {
9508 FoldingSetNodeID ID;
9509 ID.AddInteger(4U);
9510 ID.AddInteger(VT1.getRawBits());
9511 ID.AddInteger(VT2.getRawBits());
9512 ID.AddInteger(VT3.getRawBits());
9513 ID.AddInteger(VT4.getRawBits());
9514
9515 void *IP = nullptr;
9516 SDVTListNode *Result = VTListMap.FindNodeOrInsertPos(ID, IP);
9517 if (!Result) {
9518 EVT *Array = Allocator.Allocate<EVT>(4);
9519 Array[0] = VT1;
9520 Array[1] = VT2;
9521 Array[2] = VT3;
9522 Array[3] = VT4;
9523 Result = new (Allocator) SDVTListNode(ID.Intern(Allocator), Array, 4);
9524 VTListMap.InsertNode(Result, IP);
9525 }
9526 return Result->getSDVTList();
9527 }
9528
getVTList(ArrayRef<EVT> VTs)9529 SDVTList SelectionDAG::getVTList(ArrayRef<EVT> VTs) {
9530 unsigned NumVTs = VTs.size();
9531 FoldingSetNodeID ID;
9532 ID.AddInteger(NumVTs);
9533 for (unsigned index = 0; index < NumVTs; index++) {
9534 ID.AddInteger(VTs[index].getRawBits());
9535 }
9536
9537 void *IP = nullptr;
9538 SDVTListNode *Result = VTListMap.FindNodeOrInsertPos(ID, IP);
9539 if (!Result) {
9540 EVT *Array = Allocator.Allocate<EVT>(NumVTs);
9541 llvm::copy(VTs, Array);
9542 Result = new (Allocator) SDVTListNode(ID.Intern(Allocator), Array, NumVTs);
9543 VTListMap.InsertNode(Result, IP);
9544 }
9545 return Result->getSDVTList();
9546 }
9547
9548
9549 /// UpdateNodeOperands - *Mutate* the specified node in-place to have the
9550 /// specified operands. If the resultant node already exists in the DAG,
9551 /// this does not modify the specified node, instead it returns the node that
9552 /// already exists. If the resultant node does not exist in the DAG, the
9553 /// input node is returned. As a degenerate case, if you specify the same
9554 /// input operands as the node already has, the input node is returned.
UpdateNodeOperands(SDNode * N,SDValue Op)9555 SDNode *SelectionDAG::UpdateNodeOperands(SDNode *N, SDValue Op) {
9556 assert(N->getNumOperands() == 1 && "Update with wrong number of operands");
9557
9558 // Check to see if there is no change.
9559 if (Op == N->getOperand(0)) return N;
9560
9561 // See if the modified node already exists.
9562 void *InsertPos = nullptr;
9563 if (SDNode *Existing = FindModifiedNodeSlot(N, Op, InsertPos))
9564 return Existing;
9565
9566 // Nope it doesn't. Remove the node from its current place in the maps.
9567 if (InsertPos)
9568 if (!RemoveNodeFromCSEMaps(N))
9569 InsertPos = nullptr;
9570
9571 // Now we update the operands.
9572 N->OperandList[0].set(Op);
9573
9574 updateDivergence(N);
9575 // If this gets put into a CSE map, add it.
9576 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
9577 return N;
9578 }
9579
UpdateNodeOperands(SDNode * N,SDValue Op1,SDValue Op2)9580 SDNode *SelectionDAG::UpdateNodeOperands(SDNode *N, SDValue Op1, SDValue Op2) {
9581 assert(N->getNumOperands() == 2 && "Update with wrong number of operands");
9582
9583 // Check to see if there is no change.
9584 if (Op1 == N->getOperand(0) && Op2 == N->getOperand(1))
9585 return N; // No operands changed, just return the input node.
9586
9587 // See if the modified node already exists.
9588 void *InsertPos = nullptr;
9589 if (SDNode *Existing = FindModifiedNodeSlot(N, Op1, Op2, InsertPos))
9590 return Existing;
9591
9592 // Nope it doesn't. Remove the node from its current place in the maps.
9593 if (InsertPos)
9594 if (!RemoveNodeFromCSEMaps(N))
9595 InsertPos = nullptr;
9596
9597 // Now we update the operands.
9598 if (N->OperandList[0] != Op1)
9599 N->OperandList[0].set(Op1);
9600 if (N->OperandList[1] != Op2)
9601 N->OperandList[1].set(Op2);
9602
9603 updateDivergence(N);
9604 // If this gets put into a CSE map, add it.
9605 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
9606 return N;
9607 }
9608
9609 SDNode *SelectionDAG::
UpdateNodeOperands(SDNode * N,SDValue Op1,SDValue Op2,SDValue Op3)9610 UpdateNodeOperands(SDNode *N, SDValue Op1, SDValue Op2, SDValue Op3) {
9611 SDValue Ops[] = { Op1, Op2, Op3 };
9612 return UpdateNodeOperands(N, Ops);
9613 }
9614
9615 SDNode *SelectionDAG::
UpdateNodeOperands(SDNode * N,SDValue Op1,SDValue Op2,SDValue Op3,SDValue Op4)9616 UpdateNodeOperands(SDNode *N, SDValue Op1, SDValue Op2,
9617 SDValue Op3, SDValue Op4) {
9618 SDValue Ops[] = { Op1, Op2, Op3, Op4 };
9619 return UpdateNodeOperands(N, Ops);
9620 }
9621
9622 SDNode *SelectionDAG::
UpdateNodeOperands(SDNode * N,SDValue Op1,SDValue Op2,SDValue Op3,SDValue Op4,SDValue Op5)9623 UpdateNodeOperands(SDNode *N, SDValue Op1, SDValue Op2,
9624 SDValue Op3, SDValue Op4, SDValue Op5) {
9625 SDValue Ops[] = { Op1, Op2, Op3, Op4, Op5 };
9626 return UpdateNodeOperands(N, Ops);
9627 }
9628
9629 SDNode *SelectionDAG::
UpdateNodeOperands(SDNode * N,ArrayRef<SDValue> Ops)9630 UpdateNodeOperands(SDNode *N, ArrayRef<SDValue> Ops) {
9631 unsigned NumOps = Ops.size();
9632 assert(N->getNumOperands() == NumOps &&
9633 "Update with wrong number of operands");
9634
9635 // If no operands changed just return the input node.
9636 if (std::equal(Ops.begin(), Ops.end(), N->op_begin()))
9637 return N;
9638
9639 // See if the modified node already exists.
9640 void *InsertPos = nullptr;
9641 if (SDNode *Existing = FindModifiedNodeSlot(N, Ops, InsertPos))
9642 return Existing;
9643
9644 // Nope it doesn't. Remove the node from its current place in the maps.
9645 if (InsertPos)
9646 if (!RemoveNodeFromCSEMaps(N))
9647 InsertPos = nullptr;
9648
9649 // Now we update the operands.
9650 for (unsigned i = 0; i != NumOps; ++i)
9651 if (N->OperandList[i] != Ops[i])
9652 N->OperandList[i].set(Ops[i]);
9653
9654 updateDivergence(N);
9655 // If this gets put into a CSE map, add it.
9656 if (InsertPos) CSEMap.InsertNode(N, InsertPos);
9657 return N;
9658 }
9659
9660 /// DropOperands - Release the operands and set this node to have
9661 /// zero operands.
DropOperands()9662 void SDNode::DropOperands() {
9663 // Unlike the code in MorphNodeTo that does this, we don't need to
9664 // watch for dead nodes here.
9665 for (op_iterator I = op_begin(), E = op_end(); I != E; ) {
9666 SDUse &Use = *I++;
9667 Use.set(SDValue());
9668 }
9669 }
9670
setNodeMemRefs(MachineSDNode * N,ArrayRef<MachineMemOperand * > NewMemRefs)9671 void SelectionDAG::setNodeMemRefs(MachineSDNode *N,
9672 ArrayRef<MachineMemOperand *> NewMemRefs) {
9673 if (NewMemRefs.empty()) {
9674 N->clearMemRefs();
9675 return;
9676 }
9677
9678 // Check if we can avoid allocating by storing a single reference directly.
9679 if (NewMemRefs.size() == 1) {
9680 N->MemRefs = NewMemRefs[0];
9681 N->NumMemRefs = 1;
9682 return;
9683 }
9684
9685 MachineMemOperand **MemRefsBuffer =
9686 Allocator.template Allocate<MachineMemOperand *>(NewMemRefs.size());
9687 llvm::copy(NewMemRefs, MemRefsBuffer);
9688 N->MemRefs = MemRefsBuffer;
9689 N->NumMemRefs = static_cast<int>(NewMemRefs.size());
9690 }
9691
9692 /// SelectNodeTo - These are wrappers around MorphNodeTo that accept a
9693 /// machine opcode.
9694 ///
SelectNodeTo(SDNode * N,unsigned MachineOpc,EVT VT)9695 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
9696 EVT VT) {
9697 SDVTList VTs = getVTList(VT);
9698 return SelectNodeTo(N, MachineOpc, VTs, std::nullopt);
9699 }
9700
SelectNodeTo(SDNode * N,unsigned MachineOpc,EVT VT,SDValue Op1)9701 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
9702 EVT VT, SDValue Op1) {
9703 SDVTList VTs = getVTList(VT);
9704 SDValue Ops[] = { Op1 };
9705 return SelectNodeTo(N, MachineOpc, VTs, Ops);
9706 }
9707
SelectNodeTo(SDNode * N,unsigned MachineOpc,EVT VT,SDValue Op1,SDValue Op2)9708 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
9709 EVT VT, SDValue Op1,
9710 SDValue Op2) {
9711 SDVTList VTs = getVTList(VT);
9712 SDValue Ops[] = { Op1, Op2 };
9713 return SelectNodeTo(N, MachineOpc, VTs, Ops);
9714 }
9715
SelectNodeTo(SDNode * N,unsigned MachineOpc,EVT VT,SDValue Op1,SDValue Op2,SDValue Op3)9716 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
9717 EVT VT, SDValue Op1,
9718 SDValue Op2, SDValue Op3) {
9719 SDVTList VTs = getVTList(VT);
9720 SDValue Ops[] = { Op1, Op2, Op3 };
9721 return SelectNodeTo(N, MachineOpc, VTs, Ops);
9722 }
9723
SelectNodeTo(SDNode * N,unsigned MachineOpc,EVT VT,ArrayRef<SDValue> Ops)9724 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
9725 EVT VT, ArrayRef<SDValue> Ops) {
9726 SDVTList VTs = getVTList(VT);
9727 return SelectNodeTo(N, MachineOpc, VTs, Ops);
9728 }
9729
SelectNodeTo(SDNode * N,unsigned MachineOpc,EVT VT1,EVT VT2,ArrayRef<SDValue> Ops)9730 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
9731 EVT VT1, EVT VT2, ArrayRef<SDValue> Ops) {
9732 SDVTList VTs = getVTList(VT1, VT2);
9733 return SelectNodeTo(N, MachineOpc, VTs, Ops);
9734 }
9735
SelectNodeTo(SDNode * N,unsigned MachineOpc,EVT VT1,EVT VT2)9736 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
9737 EVT VT1, EVT VT2) {
9738 SDVTList VTs = getVTList(VT1, VT2);
9739 return SelectNodeTo(N, MachineOpc, VTs, std::nullopt);
9740 }
9741
SelectNodeTo(SDNode * N,unsigned MachineOpc,EVT VT1,EVT VT2,EVT VT3,ArrayRef<SDValue> Ops)9742 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
9743 EVT VT1, EVT VT2, EVT VT3,
9744 ArrayRef<SDValue> Ops) {
9745 SDVTList VTs = getVTList(VT1, VT2, VT3);
9746 return SelectNodeTo(N, MachineOpc, VTs, Ops);
9747 }
9748
SelectNodeTo(SDNode * N,unsigned MachineOpc,EVT VT1,EVT VT2,SDValue Op1,SDValue Op2)9749 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
9750 EVT VT1, EVT VT2,
9751 SDValue Op1, SDValue Op2) {
9752 SDVTList VTs = getVTList(VT1, VT2);
9753 SDValue Ops[] = { Op1, Op2 };
9754 return SelectNodeTo(N, MachineOpc, VTs, Ops);
9755 }
9756
SelectNodeTo(SDNode * N,unsigned MachineOpc,SDVTList VTs,ArrayRef<SDValue> Ops)9757 SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc,
9758 SDVTList VTs,ArrayRef<SDValue> Ops) {
9759 SDNode *New = MorphNodeTo(N, ~MachineOpc, VTs, Ops);
9760 // Reset the NodeID to -1.
9761 New->setNodeId(-1);
9762 if (New != N) {
9763 ReplaceAllUsesWith(N, New);
9764 RemoveDeadNode(N);
9765 }
9766 return New;
9767 }
9768
9769 /// UpdateSDLocOnMergeSDNode - If the opt level is -O0 then it throws away
9770 /// the line number information on the merged node since it is not possible to
9771 /// preserve the information that operation is associated with multiple lines.
9772 /// This will make the debugger working better at -O0, were there is a higher
9773 /// probability having other instructions associated with that line.
9774 ///
9775 /// For IROrder, we keep the smaller of the two
UpdateSDLocOnMergeSDNode(SDNode * N,const SDLoc & OLoc)9776 SDNode *SelectionDAG::UpdateSDLocOnMergeSDNode(SDNode *N, const SDLoc &OLoc) {
9777 DebugLoc NLoc = N->getDebugLoc();
9778 if (NLoc && OptLevel == CodeGenOpt::None && OLoc.getDebugLoc() != NLoc) {
9779 N->setDebugLoc(DebugLoc());
9780 }
9781 unsigned Order = std::min(N->getIROrder(), OLoc.getIROrder());
9782 N->setIROrder(Order);
9783 return N;
9784 }
9785
9786 /// MorphNodeTo - This *mutates* the specified node to have the specified
9787 /// return type, opcode, and operands.
9788 ///
9789 /// Note that MorphNodeTo returns the resultant node. If there is already a
9790 /// node of the specified opcode and operands, it returns that node instead of
9791 /// the current one. Note that the SDLoc need not be the same.
9792 ///
9793 /// Using MorphNodeTo is faster than creating a new node and swapping it in
9794 /// with ReplaceAllUsesWith both because it often avoids allocating a new
9795 /// node, and because it doesn't require CSE recalculation for any of
9796 /// the node's users.
9797 ///
9798 /// However, note that MorphNodeTo recursively deletes dead nodes from the DAG.
9799 /// As a consequence it isn't appropriate to use from within the DAG combiner or
9800 /// the legalizer which maintain worklists that would need to be updated when
9801 /// deleting things.
MorphNodeTo(SDNode * N,unsigned Opc,SDVTList VTs,ArrayRef<SDValue> Ops)9802 SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
9803 SDVTList VTs, ArrayRef<SDValue> Ops) {
9804 // If an identical node already exists, use it.
9805 void *IP = nullptr;
9806 if (VTs.VTs[VTs.NumVTs-1] != MVT::Glue) {
9807 FoldingSetNodeID ID;
9808 AddNodeIDNode(ID, Opc, VTs, Ops);
9809 if (SDNode *ON = FindNodeOrInsertPos(ID, SDLoc(N), IP))
9810 return UpdateSDLocOnMergeSDNode(ON, SDLoc(N));
9811 }
9812
9813 if (!RemoveNodeFromCSEMaps(N))
9814 IP = nullptr;
9815
9816 // Start the morphing.
9817 N->NodeType = Opc;
9818 N->ValueList = VTs.VTs;
9819 N->NumValues = VTs.NumVTs;
9820
9821 // Clear the operands list, updating used nodes to remove this from their
9822 // use list. Keep track of any operands that become dead as a result.
9823 SmallPtrSet<SDNode*, 16> DeadNodeSet;
9824 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) {
9825 SDUse &Use = *I++;
9826 SDNode *Used = Use.getNode();
9827 Use.set(SDValue());
9828 if (Used->use_empty())
9829 DeadNodeSet.insert(Used);
9830 }
9831
9832 // For MachineNode, initialize the memory references information.
9833 if (MachineSDNode *MN = dyn_cast<MachineSDNode>(N))
9834 MN->clearMemRefs();
9835
9836 // Swap for an appropriately sized array from the recycler.
9837 removeOperands(N);
9838 createOperands(N, Ops);
9839
9840 // Delete any nodes that are still dead after adding the uses for the
9841 // new operands.
9842 if (!DeadNodeSet.empty()) {
9843 SmallVector<SDNode *, 16> DeadNodes;
9844 for (SDNode *N : DeadNodeSet)
9845 if (N->use_empty())
9846 DeadNodes.push_back(N);
9847 RemoveDeadNodes(DeadNodes);
9848 }
9849
9850 if (IP)
9851 CSEMap.InsertNode(N, IP); // Memoize the new node.
9852 return N;
9853 }
9854
mutateStrictFPToFP(SDNode * Node)9855 SDNode* SelectionDAG::mutateStrictFPToFP(SDNode *Node) {
9856 unsigned OrigOpc = Node->getOpcode();
9857 unsigned NewOpc;
9858 switch (OrigOpc) {
9859 default:
9860 llvm_unreachable("mutateStrictFPToFP called with unexpected opcode!");
9861 #define DAG_INSTRUCTION(NAME, NARG, ROUND_MODE, INTRINSIC, DAGN) \
9862 case ISD::STRICT_##DAGN: NewOpc = ISD::DAGN; break;
9863 #define CMP_INSTRUCTION(NAME, NARG, ROUND_MODE, INTRINSIC, DAGN) \
9864 case ISD::STRICT_##DAGN: NewOpc = ISD::SETCC; break;
9865 #include "llvm/IR/ConstrainedOps.def"
9866 }
9867
9868 assert(Node->getNumValues() == 2 && "Unexpected number of results!");
9869
9870 // We're taking this node out of the chain, so we need to re-link things.
9871 SDValue InputChain = Node->getOperand(0);
9872 SDValue OutputChain = SDValue(Node, 1);
9873 ReplaceAllUsesOfValueWith(OutputChain, InputChain);
9874
9875 SmallVector<SDValue, 3> Ops;
9876 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i)
9877 Ops.push_back(Node->getOperand(i));
9878
9879 SDVTList VTs = getVTList(Node->getValueType(0));
9880 SDNode *Res = MorphNodeTo(Node, NewOpc, VTs, Ops);
9881
9882 // MorphNodeTo can operate in two ways: if an existing node with the
9883 // specified operands exists, it can just return it. Otherwise, it
9884 // updates the node in place to have the requested operands.
9885 if (Res == Node) {
9886 // If we updated the node in place, reset the node ID. To the isel,
9887 // this should be just like a newly allocated machine node.
9888 Res->setNodeId(-1);
9889 } else {
9890 ReplaceAllUsesWith(Node, Res);
9891 RemoveDeadNode(Node);
9892 }
9893
9894 return Res;
9895 }
9896
9897 /// getMachineNode - These are used for target selectors to create a new node
9898 /// with specified return type(s), MachineInstr opcode, and operands.
9899 ///
9900 /// Note that getMachineNode returns the resultant node. If there is already a
9901 /// node of the specified opcode and operands, it returns that node instead of
9902 /// the current one.
getMachineNode(unsigned Opcode,const SDLoc & dl,EVT VT)9903 MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
9904 EVT VT) {
9905 SDVTList VTs = getVTList(VT);
9906 return getMachineNode(Opcode, dl, VTs, std::nullopt);
9907 }
9908
getMachineNode(unsigned Opcode,const SDLoc & dl,EVT VT,SDValue Op1)9909 MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
9910 EVT VT, SDValue Op1) {
9911 SDVTList VTs = getVTList(VT);
9912 SDValue Ops[] = { Op1 };
9913 return getMachineNode(Opcode, dl, VTs, Ops);
9914 }
9915
getMachineNode(unsigned Opcode,const SDLoc & dl,EVT VT,SDValue Op1,SDValue Op2)9916 MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
9917 EVT VT, SDValue Op1, SDValue Op2) {
9918 SDVTList VTs = getVTList(VT);
9919 SDValue Ops[] = { Op1, Op2 };
9920 return getMachineNode(Opcode, dl, VTs, Ops);
9921 }
9922
getMachineNode(unsigned Opcode,const SDLoc & dl,EVT VT,SDValue Op1,SDValue Op2,SDValue Op3)9923 MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
9924 EVT VT, SDValue Op1, SDValue Op2,
9925 SDValue Op3) {
9926 SDVTList VTs = getVTList(VT);
9927 SDValue Ops[] = { Op1, Op2, Op3 };
9928 return getMachineNode(Opcode, dl, VTs, Ops);
9929 }
9930
getMachineNode(unsigned Opcode,const SDLoc & dl,EVT VT,ArrayRef<SDValue> Ops)9931 MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
9932 EVT VT, ArrayRef<SDValue> Ops) {
9933 SDVTList VTs = getVTList(VT);
9934 return getMachineNode(Opcode, dl, VTs, Ops);
9935 }
9936
getMachineNode(unsigned Opcode,const SDLoc & dl,EVT VT1,EVT VT2,SDValue Op1,SDValue Op2)9937 MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
9938 EVT VT1, EVT VT2, SDValue Op1,
9939 SDValue Op2) {
9940 SDVTList VTs = getVTList(VT1, VT2);
9941 SDValue Ops[] = { Op1, Op2 };
9942 return getMachineNode(Opcode, dl, VTs, Ops);
9943 }
9944
getMachineNode(unsigned Opcode,const SDLoc & dl,EVT VT1,EVT VT2,SDValue Op1,SDValue Op2,SDValue Op3)9945 MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
9946 EVT VT1, EVT VT2, SDValue Op1,
9947 SDValue Op2, SDValue Op3) {
9948 SDVTList VTs = getVTList(VT1, VT2);
9949 SDValue Ops[] = { Op1, Op2, Op3 };
9950 return getMachineNode(Opcode, dl, VTs, Ops);
9951 }
9952
getMachineNode(unsigned Opcode,const SDLoc & dl,EVT VT1,EVT VT2,ArrayRef<SDValue> Ops)9953 MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
9954 EVT VT1, EVT VT2,
9955 ArrayRef<SDValue> Ops) {
9956 SDVTList VTs = getVTList(VT1, VT2);
9957 return getMachineNode(Opcode, dl, VTs, Ops);
9958 }
9959
getMachineNode(unsigned Opcode,const SDLoc & dl,EVT VT1,EVT VT2,EVT VT3,SDValue Op1,SDValue Op2)9960 MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
9961 EVT VT1, EVT VT2, EVT VT3,
9962 SDValue Op1, SDValue Op2) {
9963 SDVTList VTs = getVTList(VT1, VT2, VT3);
9964 SDValue Ops[] = { Op1, Op2 };
9965 return getMachineNode(Opcode, dl, VTs, Ops);
9966 }
9967
getMachineNode(unsigned Opcode,const SDLoc & dl,EVT VT1,EVT VT2,EVT VT3,SDValue Op1,SDValue Op2,SDValue Op3)9968 MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
9969 EVT VT1, EVT VT2, EVT VT3,
9970 SDValue Op1, SDValue Op2,
9971 SDValue Op3) {
9972 SDVTList VTs = getVTList(VT1, VT2, VT3);
9973 SDValue Ops[] = { Op1, Op2, Op3 };
9974 return getMachineNode(Opcode, dl, VTs, Ops);
9975 }
9976
getMachineNode(unsigned Opcode,const SDLoc & dl,EVT VT1,EVT VT2,EVT VT3,ArrayRef<SDValue> Ops)9977 MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
9978 EVT VT1, EVT VT2, EVT VT3,
9979 ArrayRef<SDValue> Ops) {
9980 SDVTList VTs = getVTList(VT1, VT2, VT3);
9981 return getMachineNode(Opcode, dl, VTs, Ops);
9982 }
9983
getMachineNode(unsigned Opcode,const SDLoc & dl,ArrayRef<EVT> ResultTys,ArrayRef<SDValue> Ops)9984 MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl,
9985 ArrayRef<EVT> ResultTys,
9986 ArrayRef<SDValue> Ops) {
9987 SDVTList VTs = getVTList(ResultTys);
9988 return getMachineNode(Opcode, dl, VTs, Ops);
9989 }
9990
getMachineNode(unsigned Opcode,const SDLoc & DL,SDVTList VTs,ArrayRef<SDValue> Ops)9991 MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &DL,
9992 SDVTList VTs,
9993 ArrayRef<SDValue> Ops) {
9994 bool DoCSE = VTs.VTs[VTs.NumVTs-1] != MVT::Glue;
9995 MachineSDNode *N;
9996 void *IP = nullptr;
9997
9998 if (DoCSE) {
9999 FoldingSetNodeID ID;
10000 AddNodeIDNode(ID, ~Opcode, VTs, Ops);
10001 IP = nullptr;
10002 if (SDNode *E = FindNodeOrInsertPos(ID, DL, IP)) {
10003 return cast<MachineSDNode>(UpdateSDLocOnMergeSDNode(E, DL));
10004 }
10005 }
10006
10007 // Allocate a new MachineSDNode.
10008 N = newSDNode<MachineSDNode>(~Opcode, DL.getIROrder(), DL.getDebugLoc(), VTs);
10009 createOperands(N, Ops);
10010
10011 if (DoCSE)
10012 CSEMap.InsertNode(N, IP);
10013
10014 InsertNode(N);
10015 NewSDValueDbgMsg(SDValue(N, 0), "Creating new machine node: ", this);
10016 return N;
10017 }
10018
10019 /// getTargetExtractSubreg - A convenience function for creating
10020 /// TargetOpcode::EXTRACT_SUBREG nodes.
getTargetExtractSubreg(int SRIdx,const SDLoc & DL,EVT VT,SDValue Operand)10021 SDValue SelectionDAG::getTargetExtractSubreg(int SRIdx, const SDLoc &DL, EVT VT,
10022 SDValue Operand) {
10023 SDValue SRIdxVal = getTargetConstant(SRIdx, DL, MVT::i32);
10024 SDNode *Subreg = getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL,
10025 VT, Operand, SRIdxVal);
10026 return SDValue(Subreg, 0);
10027 }
10028
10029 /// getTargetInsertSubreg - A convenience function for creating
10030 /// TargetOpcode::INSERT_SUBREG nodes.
getTargetInsertSubreg(int SRIdx,const SDLoc & DL,EVT VT,SDValue Operand,SDValue Subreg)10031 SDValue SelectionDAG::getTargetInsertSubreg(int SRIdx, const SDLoc &DL, EVT VT,
10032 SDValue Operand, SDValue Subreg) {
10033 SDValue SRIdxVal = getTargetConstant(SRIdx, DL, MVT::i32);
10034 SDNode *Result = getMachineNode(TargetOpcode::INSERT_SUBREG, DL,
10035 VT, Operand, Subreg, SRIdxVal);
10036 return SDValue(Result, 0);
10037 }
10038
10039 /// getNodeIfExists - Get the specified node if it's already available, or
10040 /// else return NULL.
getNodeIfExists(unsigned Opcode,SDVTList VTList,ArrayRef<SDValue> Ops)10041 SDNode *SelectionDAG::getNodeIfExists(unsigned Opcode, SDVTList VTList,
10042 ArrayRef<SDValue> Ops) {
10043 SDNodeFlags Flags;
10044 if (Inserter)
10045 Flags = Inserter->getFlags();
10046 return getNodeIfExists(Opcode, VTList, Ops, Flags);
10047 }
10048
getNodeIfExists(unsigned Opcode,SDVTList VTList,ArrayRef<SDValue> Ops,const SDNodeFlags Flags)10049 SDNode *SelectionDAG::getNodeIfExists(unsigned Opcode, SDVTList VTList,
10050 ArrayRef<SDValue> Ops,
10051 const SDNodeFlags Flags) {
10052 if (VTList.VTs[VTList.NumVTs - 1] != MVT::Glue) {
10053 FoldingSetNodeID ID;
10054 AddNodeIDNode(ID, Opcode, VTList, Ops);
10055 void *IP = nullptr;
10056 if (SDNode *E = FindNodeOrInsertPos(ID, SDLoc(), IP)) {
10057 E->intersectFlagsWith(Flags);
10058 return E;
10059 }
10060 }
10061 return nullptr;
10062 }
10063
10064 /// doesNodeExist - Check if a node exists without modifying its flags.
doesNodeExist(unsigned Opcode,SDVTList VTList,ArrayRef<SDValue> Ops)10065 bool SelectionDAG::doesNodeExist(unsigned Opcode, SDVTList VTList,
10066 ArrayRef<SDValue> Ops) {
10067 if (VTList.VTs[VTList.NumVTs - 1] != MVT::Glue) {
10068 FoldingSetNodeID ID;
10069 AddNodeIDNode(ID, Opcode, VTList, Ops);
10070 void *IP = nullptr;
10071 if (FindNodeOrInsertPos(ID, SDLoc(), IP))
10072 return true;
10073 }
10074 return false;
10075 }
10076
10077 /// getDbgValue - Creates a SDDbgValue node.
10078 ///
10079 /// SDNode
getDbgValue(DIVariable * Var,DIExpression * Expr,SDNode * N,unsigned R,bool IsIndirect,const DebugLoc & DL,unsigned O)10080 SDDbgValue *SelectionDAG::getDbgValue(DIVariable *Var, DIExpression *Expr,
10081 SDNode *N, unsigned R, bool IsIndirect,
10082 const DebugLoc &DL, unsigned O) {
10083 assert(cast<DILocalVariable>(Var)->isValidLocationForIntrinsic(DL) &&
10084 "Expected inlined-at fields to agree");
10085 return new (DbgInfo->getAlloc())
10086 SDDbgValue(DbgInfo->getAlloc(), Var, Expr, SDDbgOperand::fromNode(N, R),
10087 {}, IsIndirect, DL, O,
10088 /*IsVariadic=*/false);
10089 }
10090
10091 /// Constant
getConstantDbgValue(DIVariable * Var,DIExpression * Expr,const Value * C,const DebugLoc & DL,unsigned O)10092 SDDbgValue *SelectionDAG::getConstantDbgValue(DIVariable *Var,
10093 DIExpression *Expr,
10094 const Value *C,
10095 const DebugLoc &DL, unsigned O) {
10096 assert(cast<DILocalVariable>(Var)->isValidLocationForIntrinsic(DL) &&
10097 "Expected inlined-at fields to agree");
10098 return new (DbgInfo->getAlloc())
10099 SDDbgValue(DbgInfo->getAlloc(), Var, Expr, SDDbgOperand::fromConst(C), {},
10100 /*IsIndirect=*/false, DL, O,
10101 /*IsVariadic=*/false);
10102 }
10103
10104 /// FrameIndex
getFrameIndexDbgValue(DIVariable * Var,DIExpression * Expr,unsigned FI,bool IsIndirect,const DebugLoc & DL,unsigned O)10105 SDDbgValue *SelectionDAG::getFrameIndexDbgValue(DIVariable *Var,
10106 DIExpression *Expr, unsigned FI,
10107 bool IsIndirect,
10108 const DebugLoc &DL,
10109 unsigned O) {
10110 assert(cast<DILocalVariable>(Var)->isValidLocationForIntrinsic(DL) &&
10111 "Expected inlined-at fields to agree");
10112 return getFrameIndexDbgValue(Var, Expr, FI, {}, IsIndirect, DL, O);
10113 }
10114
10115 /// FrameIndex with dependencies
getFrameIndexDbgValue(DIVariable * Var,DIExpression * Expr,unsigned FI,ArrayRef<SDNode * > Dependencies,bool IsIndirect,const DebugLoc & DL,unsigned O)10116 SDDbgValue *SelectionDAG::getFrameIndexDbgValue(DIVariable *Var,
10117 DIExpression *Expr, unsigned FI,
10118 ArrayRef<SDNode *> Dependencies,
10119 bool IsIndirect,
10120 const DebugLoc &DL,
10121 unsigned O) {
10122 assert(cast<DILocalVariable>(Var)->isValidLocationForIntrinsic(DL) &&
10123 "Expected inlined-at fields to agree");
10124 return new (DbgInfo->getAlloc())
10125 SDDbgValue(DbgInfo->getAlloc(), Var, Expr, SDDbgOperand::fromFrameIdx(FI),
10126 Dependencies, IsIndirect, DL, O,
10127 /*IsVariadic=*/false);
10128 }
10129
10130 /// VReg
getVRegDbgValue(DIVariable * Var,DIExpression * Expr,unsigned VReg,bool IsIndirect,const DebugLoc & DL,unsigned O)10131 SDDbgValue *SelectionDAG::getVRegDbgValue(DIVariable *Var, DIExpression *Expr,
10132 unsigned VReg, bool IsIndirect,
10133 const DebugLoc &DL, unsigned O) {
10134 assert(cast<DILocalVariable>(Var)->isValidLocationForIntrinsic(DL) &&
10135 "Expected inlined-at fields to agree");
10136 return new (DbgInfo->getAlloc())
10137 SDDbgValue(DbgInfo->getAlloc(), Var, Expr, SDDbgOperand::fromVReg(VReg),
10138 {}, IsIndirect, DL, O,
10139 /*IsVariadic=*/false);
10140 }
10141
getDbgValueList(DIVariable * Var,DIExpression * Expr,ArrayRef<SDDbgOperand> Locs,ArrayRef<SDNode * > Dependencies,bool IsIndirect,const DebugLoc & DL,unsigned O,bool IsVariadic)10142 SDDbgValue *SelectionDAG::getDbgValueList(DIVariable *Var, DIExpression *Expr,
10143 ArrayRef<SDDbgOperand> Locs,
10144 ArrayRef<SDNode *> Dependencies,
10145 bool IsIndirect, const DebugLoc &DL,
10146 unsigned O, bool IsVariadic) {
10147 assert(cast<DILocalVariable>(Var)->isValidLocationForIntrinsic(DL) &&
10148 "Expected inlined-at fields to agree");
10149 return new (DbgInfo->getAlloc())
10150 SDDbgValue(DbgInfo->getAlloc(), Var, Expr, Locs, Dependencies, IsIndirect,
10151 DL, O, IsVariadic);
10152 }
10153
transferDbgValues(SDValue From,SDValue To,unsigned OffsetInBits,unsigned SizeInBits,bool InvalidateDbg)10154 void SelectionDAG::transferDbgValues(SDValue From, SDValue To,
10155 unsigned OffsetInBits, unsigned SizeInBits,
10156 bool InvalidateDbg) {
10157 SDNode *FromNode = From.getNode();
10158 SDNode *ToNode = To.getNode();
10159 assert(FromNode && ToNode && "Can't modify dbg values");
10160
10161 // PR35338
10162 // TODO: assert(From != To && "Redundant dbg value transfer");
10163 // TODO: assert(FromNode != ToNode && "Intranode dbg value transfer");
10164 if (From == To || FromNode == ToNode)
10165 return;
10166
10167 if (!FromNode->getHasDebugValue())
10168 return;
10169
10170 SDDbgOperand FromLocOp =
10171 SDDbgOperand::fromNode(From.getNode(), From.getResNo());
10172 SDDbgOperand ToLocOp = SDDbgOperand::fromNode(To.getNode(), To.getResNo());
10173
10174 SmallVector<SDDbgValue *, 2> ClonedDVs;
10175 for (SDDbgValue *Dbg : GetDbgValues(FromNode)) {
10176 if (Dbg->isInvalidated())
10177 continue;
10178
10179 // TODO: assert(!Dbg->isInvalidated() && "Transfer of invalid dbg value");
10180
10181 // Create a new location ops vector that is equal to the old vector, but
10182 // with each instance of FromLocOp replaced with ToLocOp.
10183 bool Changed = false;
10184 auto NewLocOps = Dbg->copyLocationOps();
10185 std::replace_if(
10186 NewLocOps.begin(), NewLocOps.end(),
10187 [&Changed, FromLocOp](const SDDbgOperand &Op) {
10188 bool Match = Op == FromLocOp;
10189 Changed |= Match;
10190 return Match;
10191 },
10192 ToLocOp);
10193 // Ignore this SDDbgValue if we didn't find a matching location.
10194 if (!Changed)
10195 continue;
10196
10197 DIVariable *Var = Dbg->getVariable();
10198 auto *Expr = Dbg->getExpression();
10199 // If a fragment is requested, update the expression.
10200 if (SizeInBits) {
10201 // When splitting a larger (e.g., sign-extended) value whose
10202 // lower bits are described with an SDDbgValue, do not attempt
10203 // to transfer the SDDbgValue to the upper bits.
10204 if (auto FI = Expr->getFragmentInfo())
10205 if (OffsetInBits + SizeInBits > FI->SizeInBits)
10206 continue;
10207 auto Fragment = DIExpression::createFragmentExpression(Expr, OffsetInBits,
10208 SizeInBits);
10209 if (!Fragment)
10210 continue;
10211 Expr = *Fragment;
10212 }
10213
10214 auto AdditionalDependencies = Dbg->getAdditionalDependencies();
10215 // Clone the SDDbgValue and move it to To.
10216 SDDbgValue *Clone = getDbgValueList(
10217 Var, Expr, NewLocOps, AdditionalDependencies, Dbg->isIndirect(),
10218 Dbg->getDebugLoc(), std::max(ToNode->getIROrder(), Dbg->getOrder()),
10219 Dbg->isVariadic());
10220 ClonedDVs.push_back(Clone);
10221
10222 if (InvalidateDbg) {
10223 // Invalidate value and indicate the SDDbgValue should not be emitted.
10224 Dbg->setIsInvalidated();
10225 Dbg->setIsEmitted();
10226 }
10227 }
10228
10229 for (SDDbgValue *Dbg : ClonedDVs) {
10230 assert(is_contained(Dbg->getSDNodes(), ToNode) &&
10231 "Transferred DbgValues should depend on the new SDNode");
10232 AddDbgValue(Dbg, false);
10233 }
10234 }
10235
salvageDebugInfo(SDNode & N)10236 void SelectionDAG::salvageDebugInfo(SDNode &N) {
10237 if (!N.getHasDebugValue())
10238 return;
10239
10240 SmallVector<SDDbgValue *, 2> ClonedDVs;
10241 for (auto *DV : GetDbgValues(&N)) {
10242 if (DV->isInvalidated())
10243 continue;
10244 switch (N.getOpcode()) {
10245 default:
10246 break;
10247 case ISD::ADD:
10248 SDValue N0 = N.getOperand(0);
10249 SDValue N1 = N.getOperand(1);
10250 if (!isConstantIntBuildVectorOrConstantInt(N0) &&
10251 isConstantIntBuildVectorOrConstantInt(N1)) {
10252 uint64_t Offset = N.getConstantOperandVal(1);
10253
10254 // Rewrite an ADD constant node into a DIExpression. Since we are
10255 // performing arithmetic to compute the variable's *value* in the
10256 // DIExpression, we need to mark the expression with a
10257 // DW_OP_stack_value.
10258 auto *DIExpr = DV->getExpression();
10259 auto NewLocOps = DV->copyLocationOps();
10260 bool Changed = false;
10261 for (size_t i = 0; i < NewLocOps.size(); ++i) {
10262 // We're not given a ResNo to compare against because the whole
10263 // node is going away. We know that any ISD::ADD only has one
10264 // result, so we can assume any node match is using the result.
10265 if (NewLocOps[i].getKind() != SDDbgOperand::SDNODE ||
10266 NewLocOps[i].getSDNode() != &N)
10267 continue;
10268 NewLocOps[i] = SDDbgOperand::fromNode(N0.getNode(), N0.getResNo());
10269 SmallVector<uint64_t, 3> ExprOps;
10270 DIExpression::appendOffset(ExprOps, Offset);
10271 DIExpr = DIExpression::appendOpsToArg(DIExpr, ExprOps, i, true);
10272 Changed = true;
10273 }
10274 (void)Changed;
10275 assert(Changed && "Salvage target doesn't use N");
10276
10277 auto AdditionalDependencies = DV->getAdditionalDependencies();
10278 SDDbgValue *Clone = getDbgValueList(DV->getVariable(), DIExpr,
10279 NewLocOps, AdditionalDependencies,
10280 DV->isIndirect(), DV->getDebugLoc(),
10281 DV->getOrder(), DV->isVariadic());
10282 ClonedDVs.push_back(Clone);
10283 DV->setIsInvalidated();
10284 DV->setIsEmitted();
10285 LLVM_DEBUG(dbgs() << "SALVAGE: Rewriting";
10286 N0.getNode()->dumprFull(this);
10287 dbgs() << " into " << *DIExpr << '\n');
10288 }
10289 }
10290 }
10291
10292 for (SDDbgValue *Dbg : ClonedDVs) {
10293 assert(!Dbg->getSDNodes().empty() &&
10294 "Salvaged DbgValue should depend on a new SDNode");
10295 AddDbgValue(Dbg, false);
10296 }
10297 }
10298
10299 /// Creates a SDDbgLabel node.
getDbgLabel(DILabel * Label,const DebugLoc & DL,unsigned O)10300 SDDbgLabel *SelectionDAG::getDbgLabel(DILabel *Label,
10301 const DebugLoc &DL, unsigned O) {
10302 assert(cast<DILabel>(Label)->isValidLocationForIntrinsic(DL) &&
10303 "Expected inlined-at fields to agree");
10304 return new (DbgInfo->getAlloc()) SDDbgLabel(Label, DL, O);
10305 }
10306
10307 namespace {
10308
10309 /// RAUWUpdateListener - Helper for ReplaceAllUsesWith - When the node
10310 /// pointed to by a use iterator is deleted, increment the use iterator
10311 /// so that it doesn't dangle.
10312 ///
10313 class RAUWUpdateListener : public SelectionDAG::DAGUpdateListener {
10314 SDNode::use_iterator &UI;
10315 SDNode::use_iterator &UE;
10316
NodeDeleted(SDNode * N,SDNode * E)10317 void NodeDeleted(SDNode *N, SDNode *E) override {
10318 // Increment the iterator as needed.
10319 while (UI != UE && N == *UI)
10320 ++UI;
10321 }
10322
10323 public:
RAUWUpdateListener(SelectionDAG & d,SDNode::use_iterator & ui,SDNode::use_iterator & ue)10324 RAUWUpdateListener(SelectionDAG &d,
10325 SDNode::use_iterator &ui,
10326 SDNode::use_iterator &ue)
10327 : SelectionDAG::DAGUpdateListener(d), UI(ui), UE(ue) {}
10328 };
10329
10330 } // end anonymous namespace
10331
10332 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
10333 /// This can cause recursive merging of nodes in the DAG.
10334 ///
10335 /// This version assumes From has a single result value.
10336 ///
ReplaceAllUsesWith(SDValue FromN,SDValue To)10337 void SelectionDAG::ReplaceAllUsesWith(SDValue FromN, SDValue To) {
10338 SDNode *From = FromN.getNode();
10339 assert(From->getNumValues() == 1 && FromN.getResNo() == 0 &&
10340 "Cannot replace with this method!");
10341 assert(From != To.getNode() && "Cannot replace uses of with self");
10342
10343 // Preserve Debug Values
10344 transferDbgValues(FromN, To);
10345 // Preserve extra info.
10346 copyExtraInfo(From, To.getNode());
10347
10348 // Iterate over all the existing uses of From. New uses will be added
10349 // to the beginning of the use list, which we avoid visiting.
10350 // This specifically avoids visiting uses of From that arise while the
10351 // replacement is happening, because any such uses would be the result
10352 // of CSE: If an existing node looks like From after one of its operands
10353 // is replaced by To, we don't want to replace of all its users with To
10354 // too. See PR3018 for more info.
10355 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
10356 RAUWUpdateListener Listener(*this, UI, UE);
10357 while (UI != UE) {
10358 SDNode *User = *UI;
10359
10360 // This node is about to morph, remove its old self from the CSE maps.
10361 RemoveNodeFromCSEMaps(User);
10362
10363 // A user can appear in a use list multiple times, and when this
10364 // happens the uses are usually next to each other in the list.
10365 // To help reduce the number of CSE recomputations, process all
10366 // the uses of this user that we can find this way.
10367 do {
10368 SDUse &Use = UI.getUse();
10369 ++UI;
10370 Use.set(To);
10371 if (To->isDivergent() != From->isDivergent())
10372 updateDivergence(User);
10373 } while (UI != UE && *UI == User);
10374 // Now that we have modified User, add it back to the CSE maps. If it
10375 // already exists there, recursively merge the results together.
10376 AddModifiedNodeToCSEMaps(User);
10377 }
10378
10379 // If we just RAUW'd the root, take note.
10380 if (FromN == getRoot())
10381 setRoot(To);
10382 }
10383
10384 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
10385 /// This can cause recursive merging of nodes in the DAG.
10386 ///
10387 /// This version assumes that for each value of From, there is a
10388 /// corresponding value in To in the same position with the same type.
10389 ///
ReplaceAllUsesWith(SDNode * From,SDNode * To)10390 void SelectionDAG::ReplaceAllUsesWith(SDNode *From, SDNode *To) {
10391 #ifndef NDEBUG
10392 for (unsigned i = 0, e = From->getNumValues(); i != e; ++i)
10393 assert((!From->hasAnyUseOfValue(i) ||
10394 From->getValueType(i) == To->getValueType(i)) &&
10395 "Cannot use this version of ReplaceAllUsesWith!");
10396 #endif
10397
10398 // Handle the trivial case.
10399 if (From == To)
10400 return;
10401
10402 // Preserve Debug Info. Only do this if there's a use.
10403 for (unsigned i = 0, e = From->getNumValues(); i != e; ++i)
10404 if (From->hasAnyUseOfValue(i)) {
10405 assert((i < To->getNumValues()) && "Invalid To location");
10406 transferDbgValues(SDValue(From, i), SDValue(To, i));
10407 }
10408 // Preserve extra info.
10409 copyExtraInfo(From, To);
10410
10411 // Iterate over just the existing users of From. See the comments in
10412 // the ReplaceAllUsesWith above.
10413 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
10414 RAUWUpdateListener Listener(*this, UI, UE);
10415 while (UI != UE) {
10416 SDNode *User = *UI;
10417
10418 // This node is about to morph, remove its old self from the CSE maps.
10419 RemoveNodeFromCSEMaps(User);
10420
10421 // A user can appear in a use list multiple times, and when this
10422 // happens the uses are usually next to each other in the list.
10423 // To help reduce the number of CSE recomputations, process all
10424 // the uses of this user that we can find this way.
10425 do {
10426 SDUse &Use = UI.getUse();
10427 ++UI;
10428 Use.setNode(To);
10429 if (To->isDivergent() != From->isDivergent())
10430 updateDivergence(User);
10431 } while (UI != UE && *UI == User);
10432
10433 // Now that we have modified User, add it back to the CSE maps. If it
10434 // already exists there, recursively merge the results together.
10435 AddModifiedNodeToCSEMaps(User);
10436 }
10437
10438 // If we just RAUW'd the root, take note.
10439 if (From == getRoot().getNode())
10440 setRoot(SDValue(To, getRoot().getResNo()));
10441 }
10442
10443 /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead.
10444 /// This can cause recursive merging of nodes in the DAG.
10445 ///
10446 /// This version can replace From with any result values. To must match the
10447 /// number and types of values returned by From.
ReplaceAllUsesWith(SDNode * From,const SDValue * To)10448 void SelectionDAG::ReplaceAllUsesWith(SDNode *From, const SDValue *To) {
10449 if (From->getNumValues() == 1) // Handle the simple case efficiently.
10450 return ReplaceAllUsesWith(SDValue(From, 0), To[0]);
10451
10452 for (unsigned i = 0, e = From->getNumValues(); i != e; ++i) {
10453 // Preserve Debug Info.
10454 transferDbgValues(SDValue(From, i), To[i]);
10455 // Preserve extra info.
10456 copyExtraInfo(From, To[i].getNode());
10457 }
10458
10459 // Iterate over just the existing users of From. See the comments in
10460 // the ReplaceAllUsesWith above.
10461 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end();
10462 RAUWUpdateListener Listener(*this, UI, UE);
10463 while (UI != UE) {
10464 SDNode *User = *UI;
10465
10466 // This node is about to morph, remove its old self from the CSE maps.
10467 RemoveNodeFromCSEMaps(User);
10468
10469 // A user can appear in a use list multiple times, and when this happens the
10470 // uses are usually next to each other in the list. To help reduce the
10471 // number of CSE and divergence recomputations, process all the uses of this
10472 // user that we can find this way.
10473 bool To_IsDivergent = false;
10474 do {
10475 SDUse &Use = UI.getUse();
10476 const SDValue &ToOp = To[Use.getResNo()];
10477 ++UI;
10478 Use.set(ToOp);
10479 To_IsDivergent |= ToOp->isDivergent();
10480 } while (UI != UE && *UI == User);
10481
10482 if (To_IsDivergent != From->isDivergent())
10483 updateDivergence(User);
10484
10485 // Now that we have modified User, add it back to the CSE maps. If it
10486 // already exists there, recursively merge the results together.
10487 AddModifiedNodeToCSEMaps(User);
10488 }
10489
10490 // If we just RAUW'd the root, take note.
10491 if (From == getRoot().getNode())
10492 setRoot(SDValue(To[getRoot().getResNo()]));
10493 }
10494
10495 /// ReplaceAllUsesOfValueWith - Replace any uses of From with To, leaving
10496 /// uses of other values produced by From.getNode() alone. The Deleted
10497 /// vector is handled the same way as for ReplaceAllUsesWith.
ReplaceAllUsesOfValueWith(SDValue From,SDValue To)10498 void SelectionDAG::ReplaceAllUsesOfValueWith(SDValue From, SDValue To){
10499 // Handle the really simple, really trivial case efficiently.
10500 if (From == To) return;
10501
10502 // Handle the simple, trivial, case efficiently.
10503 if (From.getNode()->getNumValues() == 1) {
10504 ReplaceAllUsesWith(From, To);
10505 return;
10506 }
10507
10508 // Preserve Debug Info.
10509 transferDbgValues(From, To);
10510 copyExtraInfo(From.getNode(), To.getNode());
10511
10512 // Iterate over just the existing users of From. See the comments in
10513 // the ReplaceAllUsesWith above.
10514 SDNode::use_iterator UI = From.getNode()->use_begin(),
10515 UE = From.getNode()->use_end();
10516 RAUWUpdateListener Listener(*this, UI, UE);
10517 while (UI != UE) {
10518 SDNode *User = *UI;
10519 bool UserRemovedFromCSEMaps = false;
10520
10521 // A user can appear in a use list multiple times, and when this
10522 // happens the uses are usually next to each other in the list.
10523 // To help reduce the number of CSE recomputations, process all
10524 // the uses of this user that we can find this way.
10525 do {
10526 SDUse &Use = UI.getUse();
10527
10528 // Skip uses of different values from the same node.
10529 if (Use.getResNo() != From.getResNo()) {
10530 ++UI;
10531 continue;
10532 }
10533
10534 // If this node hasn't been modified yet, it's still in the CSE maps,
10535 // so remove its old self from the CSE maps.
10536 if (!UserRemovedFromCSEMaps) {
10537 RemoveNodeFromCSEMaps(User);
10538 UserRemovedFromCSEMaps = true;
10539 }
10540
10541 ++UI;
10542 Use.set(To);
10543 if (To->isDivergent() != From->isDivergent())
10544 updateDivergence(User);
10545 } while (UI != UE && *UI == User);
10546 // We are iterating over all uses of the From node, so if a use
10547 // doesn't use the specific value, no changes are made.
10548 if (!UserRemovedFromCSEMaps)
10549 continue;
10550
10551 // Now that we have modified User, add it back to the CSE maps. If it
10552 // already exists there, recursively merge the results together.
10553 AddModifiedNodeToCSEMaps(User);
10554 }
10555
10556 // If we just RAUW'd the root, take note.
10557 if (From == getRoot())
10558 setRoot(To);
10559 }
10560
10561 namespace {
10562
10563 /// UseMemo - This class is used by SelectionDAG::ReplaceAllUsesOfValuesWith
10564 /// to record information about a use.
10565 struct UseMemo {
10566 SDNode *User;
10567 unsigned Index;
10568 SDUse *Use;
10569 };
10570
10571 /// operator< - Sort Memos by User.
operator <(const UseMemo & L,const UseMemo & R)10572 bool operator<(const UseMemo &L, const UseMemo &R) {
10573 return (intptr_t)L.User < (intptr_t)R.User;
10574 }
10575
10576 /// RAUOVWUpdateListener - Helper for ReplaceAllUsesOfValuesWith - When the node
10577 /// pointed to by a UseMemo is deleted, set the User to nullptr to indicate that
10578 /// the node already has been taken care of recursively.
10579 class RAUOVWUpdateListener : public SelectionDAG::DAGUpdateListener {
10580 SmallVector<UseMemo, 4> &Uses;
10581
NodeDeleted(SDNode * N,SDNode * E)10582 void NodeDeleted(SDNode *N, SDNode *E) override {
10583 for (UseMemo &Memo : Uses)
10584 if (Memo.User == N)
10585 Memo.User = nullptr;
10586 }
10587
10588 public:
RAUOVWUpdateListener(SelectionDAG & d,SmallVector<UseMemo,4> & uses)10589 RAUOVWUpdateListener(SelectionDAG &d, SmallVector<UseMemo, 4> &uses)
10590 : SelectionDAG::DAGUpdateListener(d), Uses(uses) {}
10591 };
10592
10593 } // end anonymous namespace
10594
calculateDivergence(SDNode * N)10595 bool SelectionDAG::calculateDivergence(SDNode *N) {
10596 if (TLI->isSDNodeAlwaysUniform(N)) {
10597 assert(!TLI->isSDNodeSourceOfDivergence(N, FLI, DA) &&
10598 "Conflicting divergence information!");
10599 return false;
10600 }
10601 if (TLI->isSDNodeSourceOfDivergence(N, FLI, DA))
10602 return true;
10603 for (const auto &Op : N->ops()) {
10604 if (Op.Val.getValueType() != MVT::Other && Op.getNode()->isDivergent())
10605 return true;
10606 }
10607 return false;
10608 }
10609
updateDivergence(SDNode * N)10610 void SelectionDAG::updateDivergence(SDNode *N) {
10611 SmallVector<SDNode *, 16> Worklist(1, N);
10612 do {
10613 N = Worklist.pop_back_val();
10614 bool IsDivergent = calculateDivergence(N);
10615 if (N->SDNodeBits.IsDivergent != IsDivergent) {
10616 N->SDNodeBits.IsDivergent = IsDivergent;
10617 llvm::append_range(Worklist, N->uses());
10618 }
10619 } while (!Worklist.empty());
10620 }
10621
CreateTopologicalOrder(std::vector<SDNode * > & Order)10622 void SelectionDAG::CreateTopologicalOrder(std::vector<SDNode *> &Order) {
10623 DenseMap<SDNode *, unsigned> Degree;
10624 Order.reserve(AllNodes.size());
10625 for (auto &N : allnodes()) {
10626 unsigned NOps = N.getNumOperands();
10627 Degree[&N] = NOps;
10628 if (0 == NOps)
10629 Order.push_back(&N);
10630 }
10631 for (size_t I = 0; I != Order.size(); ++I) {
10632 SDNode *N = Order[I];
10633 for (auto *U : N->uses()) {
10634 unsigned &UnsortedOps = Degree[U];
10635 if (0 == --UnsortedOps)
10636 Order.push_back(U);
10637 }
10638 }
10639 }
10640
10641 #ifndef NDEBUG
VerifyDAGDivergence()10642 void SelectionDAG::VerifyDAGDivergence() {
10643 std::vector<SDNode *> TopoOrder;
10644 CreateTopologicalOrder(TopoOrder);
10645 for (auto *N : TopoOrder) {
10646 assert(calculateDivergence(N) == N->isDivergent() &&
10647 "Divergence bit inconsistency detected");
10648 }
10649 }
10650 #endif
10651
10652 /// ReplaceAllUsesOfValuesWith - Replace any uses of From with To, leaving
10653 /// uses of other values produced by From.getNode() alone. The same value
10654 /// may appear in both the From and To list. The Deleted vector is
10655 /// handled the same way as for ReplaceAllUsesWith.
ReplaceAllUsesOfValuesWith(const SDValue * From,const SDValue * To,unsigned Num)10656 void SelectionDAG::ReplaceAllUsesOfValuesWith(const SDValue *From,
10657 const SDValue *To,
10658 unsigned Num){
10659 // Handle the simple, trivial case efficiently.
10660 if (Num == 1)
10661 return ReplaceAllUsesOfValueWith(*From, *To);
10662
10663 transferDbgValues(*From, *To);
10664 copyExtraInfo(From->getNode(), To->getNode());
10665
10666 // Read up all the uses and make records of them. This helps
10667 // processing new uses that are introduced during the
10668 // replacement process.
10669 SmallVector<UseMemo, 4> Uses;
10670 for (unsigned i = 0; i != Num; ++i) {
10671 unsigned FromResNo = From[i].getResNo();
10672 SDNode *FromNode = From[i].getNode();
10673 for (SDNode::use_iterator UI = FromNode->use_begin(),
10674 E = FromNode->use_end(); UI != E; ++UI) {
10675 SDUse &Use = UI.getUse();
10676 if (Use.getResNo() == FromResNo) {
10677 UseMemo Memo = { *UI, i, &Use };
10678 Uses.push_back(Memo);
10679 }
10680 }
10681 }
10682
10683 // Sort the uses, so that all the uses from a given User are together.
10684 llvm::sort(Uses);
10685 RAUOVWUpdateListener Listener(*this, Uses);
10686
10687 for (unsigned UseIndex = 0, UseIndexEnd = Uses.size();
10688 UseIndex != UseIndexEnd; ) {
10689 // We know that this user uses some value of From. If it is the right
10690 // value, update it.
10691 SDNode *User = Uses[UseIndex].User;
10692 // If the node has been deleted by recursive CSE updates when updating
10693 // another node, then just skip this entry.
10694 if (User == nullptr) {
10695 ++UseIndex;
10696 continue;
10697 }
10698
10699 // This node is about to morph, remove its old self from the CSE maps.
10700 RemoveNodeFromCSEMaps(User);
10701
10702 // The Uses array is sorted, so all the uses for a given User
10703 // are next to each other in the list.
10704 // To help reduce the number of CSE recomputations, process all
10705 // the uses of this user that we can find this way.
10706 do {
10707 unsigned i = Uses[UseIndex].Index;
10708 SDUse &Use = *Uses[UseIndex].Use;
10709 ++UseIndex;
10710
10711 Use.set(To[i]);
10712 } while (UseIndex != UseIndexEnd && Uses[UseIndex].User == User);
10713
10714 // Now that we have modified User, add it back to the CSE maps. If it
10715 // already exists there, recursively merge the results together.
10716 AddModifiedNodeToCSEMaps(User);
10717 }
10718 }
10719
10720 /// AssignTopologicalOrder - Assign a unique node id for each node in the DAG
10721 /// based on their topological order. It returns the maximum id and a vector
10722 /// of the SDNodes* in assigned order by reference.
AssignTopologicalOrder()10723 unsigned SelectionDAG::AssignTopologicalOrder() {
10724 unsigned DAGSize = 0;
10725
10726 // SortedPos tracks the progress of the algorithm. Nodes before it are
10727 // sorted, nodes after it are unsorted. When the algorithm completes
10728 // it is at the end of the list.
10729 allnodes_iterator SortedPos = allnodes_begin();
10730
10731 // Visit all the nodes. Move nodes with no operands to the front of
10732 // the list immediately. Annotate nodes that do have operands with their
10733 // operand count. Before we do this, the Node Id fields of the nodes
10734 // may contain arbitrary values. After, the Node Id fields for nodes
10735 // before SortedPos will contain the topological sort index, and the
10736 // Node Id fields for nodes At SortedPos and after will contain the
10737 // count of outstanding operands.
10738 for (SDNode &N : llvm::make_early_inc_range(allnodes())) {
10739 checkForCycles(&N, this);
10740 unsigned Degree = N.getNumOperands();
10741 if (Degree == 0) {
10742 // A node with no uses, add it to the result array immediately.
10743 N.setNodeId(DAGSize++);
10744 allnodes_iterator Q(&N);
10745 if (Q != SortedPos)
10746 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(Q));
10747 assert(SortedPos != AllNodes.end() && "Overran node list");
10748 ++SortedPos;
10749 } else {
10750 // Temporarily use the Node Id as scratch space for the degree count.
10751 N.setNodeId(Degree);
10752 }
10753 }
10754
10755 // Visit all the nodes. As we iterate, move nodes into sorted order,
10756 // such that by the time the end is reached all nodes will be sorted.
10757 for (SDNode &Node : allnodes()) {
10758 SDNode *N = &Node;
10759 checkForCycles(N, this);
10760 // N is in sorted position, so all its uses have one less operand
10761 // that needs to be sorted.
10762 for (SDNode *P : N->uses()) {
10763 unsigned Degree = P->getNodeId();
10764 assert(Degree != 0 && "Invalid node degree");
10765 --Degree;
10766 if (Degree == 0) {
10767 // All of P's operands are sorted, so P may sorted now.
10768 P->setNodeId(DAGSize++);
10769 if (P->getIterator() != SortedPos)
10770 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(P));
10771 assert(SortedPos != AllNodes.end() && "Overran node list");
10772 ++SortedPos;
10773 } else {
10774 // Update P's outstanding operand count.
10775 P->setNodeId(Degree);
10776 }
10777 }
10778 if (Node.getIterator() == SortedPos) {
10779 #ifndef NDEBUG
10780 allnodes_iterator I(N);
10781 SDNode *S = &*++I;
10782 dbgs() << "Overran sorted position:\n";
10783 S->dumprFull(this); dbgs() << "\n";
10784 dbgs() << "Checking if this is due to cycles\n";
10785 checkForCycles(this, true);
10786 #endif
10787 llvm_unreachable(nullptr);
10788 }
10789 }
10790
10791 assert(SortedPos == AllNodes.end() &&
10792 "Topological sort incomplete!");
10793 assert(AllNodes.front().getOpcode() == ISD::EntryToken &&
10794 "First node in topological sort is not the entry token!");
10795 assert(AllNodes.front().getNodeId() == 0 &&
10796 "First node in topological sort has non-zero id!");
10797 assert(AllNodes.front().getNumOperands() == 0 &&
10798 "First node in topological sort has operands!");
10799 assert(AllNodes.back().getNodeId() == (int)DAGSize-1 &&
10800 "Last node in topologic sort has unexpected id!");
10801 assert(AllNodes.back().use_empty() &&
10802 "Last node in topologic sort has users!");
10803 assert(DAGSize == allnodes_size() && "Node count mismatch!");
10804 return DAGSize;
10805 }
10806
10807 /// AddDbgValue - Add a dbg_value SDNode. If SD is non-null that means the
10808 /// value is produced by SD.
AddDbgValue(SDDbgValue * DB,bool isParameter)10809 void SelectionDAG::AddDbgValue(SDDbgValue *DB, bool isParameter) {
10810 for (SDNode *SD : DB->getSDNodes()) {
10811 if (!SD)
10812 continue;
10813 assert(DbgInfo->getSDDbgValues(SD).empty() || SD->getHasDebugValue());
10814 SD->setHasDebugValue(true);
10815 }
10816 DbgInfo->add(DB, isParameter);
10817 }
10818
AddDbgLabel(SDDbgLabel * DB)10819 void SelectionDAG::AddDbgLabel(SDDbgLabel *DB) { DbgInfo->add(DB); }
10820
makeEquivalentMemoryOrdering(SDValue OldChain,SDValue NewMemOpChain)10821 SDValue SelectionDAG::makeEquivalentMemoryOrdering(SDValue OldChain,
10822 SDValue NewMemOpChain) {
10823 assert(isa<MemSDNode>(NewMemOpChain) && "Expected a memop node");
10824 assert(NewMemOpChain.getValueType() == MVT::Other && "Expected a token VT");
10825 // The new memory operation must have the same position as the old load in
10826 // terms of memory dependency. Create a TokenFactor for the old load and new
10827 // memory operation and update uses of the old load's output chain to use that
10828 // TokenFactor.
10829 if (OldChain == NewMemOpChain || OldChain.use_empty())
10830 return NewMemOpChain;
10831
10832 SDValue TokenFactor = getNode(ISD::TokenFactor, SDLoc(OldChain), MVT::Other,
10833 OldChain, NewMemOpChain);
10834 ReplaceAllUsesOfValueWith(OldChain, TokenFactor);
10835 UpdateNodeOperands(TokenFactor.getNode(), OldChain, NewMemOpChain);
10836 return TokenFactor;
10837 }
10838
makeEquivalentMemoryOrdering(LoadSDNode * OldLoad,SDValue NewMemOp)10839 SDValue SelectionDAG::makeEquivalentMemoryOrdering(LoadSDNode *OldLoad,
10840 SDValue NewMemOp) {
10841 assert(isa<MemSDNode>(NewMemOp.getNode()) && "Expected a memop node");
10842 SDValue OldChain = SDValue(OldLoad, 1);
10843 SDValue NewMemOpChain = NewMemOp.getValue(1);
10844 return makeEquivalentMemoryOrdering(OldChain, NewMemOpChain);
10845 }
10846
getSymbolFunctionGlobalAddress(SDValue Op,Function ** OutFunction)10847 SDValue SelectionDAG::getSymbolFunctionGlobalAddress(SDValue Op,
10848 Function **OutFunction) {
10849 assert(isa<ExternalSymbolSDNode>(Op) && "Node should be an ExternalSymbol");
10850
10851 auto *Symbol = cast<ExternalSymbolSDNode>(Op)->getSymbol();
10852 auto *Module = MF->getFunction().getParent();
10853 auto *Function = Module->getFunction(Symbol);
10854
10855 if (OutFunction != nullptr)
10856 *OutFunction = Function;
10857
10858 if (Function != nullptr) {
10859 auto PtrTy = TLI->getPointerTy(getDataLayout(), Function->getAddressSpace());
10860 return getGlobalAddress(Function, SDLoc(Op), PtrTy);
10861 }
10862
10863 std::string ErrorStr;
10864 raw_string_ostream ErrorFormatter(ErrorStr);
10865 ErrorFormatter << "Undefined external symbol ";
10866 ErrorFormatter << '"' << Symbol << '"';
10867 report_fatal_error(Twine(ErrorFormatter.str()));
10868 }
10869
10870 //===----------------------------------------------------------------------===//
10871 // SDNode Class
10872 //===----------------------------------------------------------------------===//
10873
isNullConstant(SDValue V)10874 bool llvm::isNullConstant(SDValue V) {
10875 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(V);
10876 return Const != nullptr && Const->isZero();
10877 }
10878
isNullFPConstant(SDValue V)10879 bool llvm::isNullFPConstant(SDValue V) {
10880 ConstantFPSDNode *Const = dyn_cast<ConstantFPSDNode>(V);
10881 return Const != nullptr && Const->isZero() && !Const->isNegative();
10882 }
10883
isAllOnesConstant(SDValue V)10884 bool llvm::isAllOnesConstant(SDValue V) {
10885 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(V);
10886 return Const != nullptr && Const->isAllOnes();
10887 }
10888
isOneConstant(SDValue V)10889 bool llvm::isOneConstant(SDValue V) {
10890 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(V);
10891 return Const != nullptr && Const->isOne();
10892 }
10893
isMinSignedConstant(SDValue V)10894 bool llvm::isMinSignedConstant(SDValue V) {
10895 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(V);
10896 return Const != nullptr && Const->isMinSignedValue();
10897 }
10898
isNeutralConstant(unsigned Opcode,SDNodeFlags Flags,SDValue V,unsigned OperandNo)10899 bool llvm::isNeutralConstant(unsigned Opcode, SDNodeFlags Flags, SDValue V,
10900 unsigned OperandNo) {
10901 // NOTE: The cases should match with IR's ConstantExpr::getBinOpIdentity().
10902 // TODO: Target-specific opcodes could be added.
10903 if (auto *Const = isConstOrConstSplat(V)) {
10904 switch (Opcode) {
10905 case ISD::ADD:
10906 case ISD::OR:
10907 case ISD::XOR:
10908 case ISD::UMAX:
10909 return Const->isZero();
10910 case ISD::MUL:
10911 return Const->isOne();
10912 case ISD::AND:
10913 case ISD::UMIN:
10914 return Const->isAllOnes();
10915 case ISD::SMAX:
10916 return Const->isMinSignedValue();
10917 case ISD::SMIN:
10918 return Const->isMaxSignedValue();
10919 case ISD::SUB:
10920 case ISD::SHL:
10921 case ISD::SRA:
10922 case ISD::SRL:
10923 return OperandNo == 1 && Const->isZero();
10924 case ISD::UDIV:
10925 case ISD::SDIV:
10926 return OperandNo == 1 && Const->isOne();
10927 }
10928 } else if (auto *ConstFP = isConstOrConstSplatFP(V)) {
10929 switch (Opcode) {
10930 case ISD::FADD:
10931 return ConstFP->isZero() &&
10932 (Flags.hasNoSignedZeros() || ConstFP->isNegative());
10933 case ISD::FSUB:
10934 return OperandNo == 1 && ConstFP->isZero() &&
10935 (Flags.hasNoSignedZeros() || !ConstFP->isNegative());
10936 case ISD::FMUL:
10937 return ConstFP->isExactlyValue(1.0);
10938 case ISD::FDIV:
10939 return OperandNo == 1 && ConstFP->isExactlyValue(1.0);
10940 case ISD::FMINNUM:
10941 case ISD::FMAXNUM: {
10942 // Neutral element for fminnum is NaN, Inf or FLT_MAX, depending on FMF.
10943 EVT VT = V.getValueType();
10944 const fltSemantics &Semantics = SelectionDAG::EVTToAPFloatSemantics(VT);
10945 APFloat NeutralAF = !Flags.hasNoNaNs()
10946 ? APFloat::getQNaN(Semantics)
10947 : !Flags.hasNoInfs()
10948 ? APFloat::getInf(Semantics)
10949 : APFloat::getLargest(Semantics);
10950 if (Opcode == ISD::FMAXNUM)
10951 NeutralAF.changeSign();
10952
10953 return ConstFP->isExactlyValue(NeutralAF);
10954 }
10955 }
10956 }
10957 return false;
10958 }
10959
peekThroughBitcasts(SDValue V)10960 SDValue llvm::peekThroughBitcasts(SDValue V) {
10961 while (V.getOpcode() == ISD::BITCAST)
10962 V = V.getOperand(0);
10963 return V;
10964 }
10965
peekThroughOneUseBitcasts(SDValue V)10966 SDValue llvm::peekThroughOneUseBitcasts(SDValue V) {
10967 while (V.getOpcode() == ISD::BITCAST && V.getOperand(0).hasOneUse())
10968 V = V.getOperand(0);
10969 return V;
10970 }
10971
peekThroughExtractSubvectors(SDValue V)10972 SDValue llvm::peekThroughExtractSubvectors(SDValue V) {
10973 while (V.getOpcode() == ISD::EXTRACT_SUBVECTOR)
10974 V = V.getOperand(0);
10975 return V;
10976 }
10977
isBitwiseNot(SDValue V,bool AllowUndefs)10978 bool llvm::isBitwiseNot(SDValue V, bool AllowUndefs) {
10979 if (V.getOpcode() != ISD::XOR)
10980 return false;
10981 V = peekThroughBitcasts(V.getOperand(1));
10982 unsigned NumBits = V.getScalarValueSizeInBits();
10983 ConstantSDNode *C =
10984 isConstOrConstSplat(V, AllowUndefs, /*AllowTruncation*/ true);
10985 return C && (C->getAPIntValue().countTrailingOnes() >= NumBits);
10986 }
10987
isConstOrConstSplat(SDValue N,bool AllowUndefs,bool AllowTruncation)10988 ConstantSDNode *llvm::isConstOrConstSplat(SDValue N, bool AllowUndefs,
10989 bool AllowTruncation) {
10990 EVT VT = N.getValueType();
10991 APInt DemandedElts = VT.isFixedLengthVector()
10992 ? APInt::getAllOnes(VT.getVectorMinNumElements())
10993 : APInt(1, 1);
10994 return isConstOrConstSplat(N, DemandedElts, AllowUndefs, AllowTruncation);
10995 }
10996
isConstOrConstSplat(SDValue N,const APInt & DemandedElts,bool AllowUndefs,bool AllowTruncation)10997 ConstantSDNode *llvm::isConstOrConstSplat(SDValue N, const APInt &DemandedElts,
10998 bool AllowUndefs,
10999 bool AllowTruncation) {
11000 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N))
11001 return CN;
11002
11003 // SplatVectors can truncate their operands. Ignore that case here unless
11004 // AllowTruncation is set.
11005 if (N->getOpcode() == ISD::SPLAT_VECTOR) {
11006 EVT VecEltVT = N->getValueType(0).getVectorElementType();
11007 if (auto *CN = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
11008 EVT CVT = CN->getValueType(0);
11009 assert(CVT.bitsGE(VecEltVT) && "Illegal splat_vector element extension");
11010 if (AllowTruncation || CVT == VecEltVT)
11011 return CN;
11012 }
11013 }
11014
11015 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
11016 BitVector UndefElements;
11017 ConstantSDNode *CN = BV->getConstantSplatNode(DemandedElts, &UndefElements);
11018
11019 // BuildVectors can truncate their operands. Ignore that case here unless
11020 // AllowTruncation is set.
11021 // TODO: Look into whether we should allow UndefElements in non-DemandedElts
11022 if (CN && (UndefElements.none() || AllowUndefs)) {
11023 EVT CVT = CN->getValueType(0);
11024 EVT NSVT = N.getValueType().getScalarType();
11025 assert(CVT.bitsGE(NSVT) && "Illegal build vector element extension");
11026 if (AllowTruncation || (CVT == NSVT))
11027 return CN;
11028 }
11029 }
11030
11031 return nullptr;
11032 }
11033
isConstOrConstSplatFP(SDValue N,bool AllowUndefs)11034 ConstantFPSDNode *llvm::isConstOrConstSplatFP(SDValue N, bool AllowUndefs) {
11035 EVT VT = N.getValueType();
11036 APInt DemandedElts = VT.isFixedLengthVector()
11037 ? APInt::getAllOnes(VT.getVectorMinNumElements())
11038 : APInt(1, 1);
11039 return isConstOrConstSplatFP(N, DemandedElts, AllowUndefs);
11040 }
11041
isConstOrConstSplatFP(SDValue N,const APInt & DemandedElts,bool AllowUndefs)11042 ConstantFPSDNode *llvm::isConstOrConstSplatFP(SDValue N,
11043 const APInt &DemandedElts,
11044 bool AllowUndefs) {
11045 if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N))
11046 return CN;
11047
11048 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
11049 BitVector UndefElements;
11050 ConstantFPSDNode *CN =
11051 BV->getConstantFPSplatNode(DemandedElts, &UndefElements);
11052 // TODO: Look into whether we should allow UndefElements in non-DemandedElts
11053 if (CN && (UndefElements.none() || AllowUndefs))
11054 return CN;
11055 }
11056
11057 if (N.getOpcode() == ISD::SPLAT_VECTOR)
11058 if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N.getOperand(0)))
11059 return CN;
11060
11061 return nullptr;
11062 }
11063
isNullOrNullSplat(SDValue N,bool AllowUndefs)11064 bool llvm::isNullOrNullSplat(SDValue N, bool AllowUndefs) {
11065 // TODO: may want to use peekThroughBitcast() here.
11066 ConstantSDNode *C =
11067 isConstOrConstSplat(N, AllowUndefs, /*AllowTruncation=*/true);
11068 return C && C->isZero();
11069 }
11070
isOneOrOneSplat(SDValue N,bool AllowUndefs)11071 bool llvm::isOneOrOneSplat(SDValue N, bool AllowUndefs) {
11072 ConstantSDNode *C =
11073 isConstOrConstSplat(N, AllowUndefs, /*AllowTruncation*/ true);
11074 return C && C->isOne();
11075 }
11076
isAllOnesOrAllOnesSplat(SDValue N,bool AllowUndefs)11077 bool llvm::isAllOnesOrAllOnesSplat(SDValue N, bool AllowUndefs) {
11078 N = peekThroughBitcasts(N);
11079 unsigned BitWidth = N.getScalarValueSizeInBits();
11080 ConstantSDNode *C = isConstOrConstSplat(N, AllowUndefs);
11081 return C && C->isAllOnes() && C->getValueSizeInBits(0) == BitWidth;
11082 }
11083
~HandleSDNode()11084 HandleSDNode::~HandleSDNode() {
11085 DropOperands();
11086 }
11087
GlobalAddressSDNode(unsigned Opc,unsigned Order,const DebugLoc & DL,const GlobalValue * GA,EVT VT,int64_t o,unsigned TF)11088 GlobalAddressSDNode::GlobalAddressSDNode(unsigned Opc, unsigned Order,
11089 const DebugLoc &DL,
11090 const GlobalValue *GA, EVT VT,
11091 int64_t o, unsigned TF)
11092 : SDNode(Opc, Order, DL, getSDVTList(VT)), Offset(o), TargetFlags(TF) {
11093 TheGlobal = GA;
11094 }
11095
AddrSpaceCastSDNode(unsigned Order,const DebugLoc & dl,EVT VT,unsigned SrcAS,unsigned DestAS)11096 AddrSpaceCastSDNode::AddrSpaceCastSDNode(unsigned Order, const DebugLoc &dl,
11097 EVT VT, unsigned SrcAS,
11098 unsigned DestAS)
11099 : SDNode(ISD::ADDRSPACECAST, Order, dl, getSDVTList(VT)),
11100 SrcAddrSpace(SrcAS), DestAddrSpace(DestAS) {}
11101
MemSDNode(unsigned Opc,unsigned Order,const DebugLoc & dl,SDVTList VTs,EVT memvt,MachineMemOperand * mmo)11102 MemSDNode::MemSDNode(unsigned Opc, unsigned Order, const DebugLoc &dl,
11103 SDVTList VTs, EVT memvt, MachineMemOperand *mmo)
11104 : SDNode(Opc, Order, dl, VTs), MemoryVT(memvt), MMO(mmo) {
11105 MemSDNodeBits.IsVolatile = MMO->isVolatile();
11106 MemSDNodeBits.IsNonTemporal = MMO->isNonTemporal();
11107 MemSDNodeBits.IsDereferenceable = MMO->isDereferenceable();
11108 MemSDNodeBits.IsInvariant = MMO->isInvariant();
11109
11110 // We check here that the size of the memory operand fits within the size of
11111 // the MMO. This is because the MMO might indicate only a possible address
11112 // range instead of specifying the affected memory addresses precisely.
11113 // TODO: Make MachineMemOperands aware of scalable vectors.
11114 assert(memvt.getStoreSize().getKnownMinValue() <= MMO->getSize() &&
11115 "Size mismatch!");
11116 }
11117
11118 /// Profile - Gather unique data for the node.
11119 ///
Profile(FoldingSetNodeID & ID) const11120 void SDNode::Profile(FoldingSetNodeID &ID) const {
11121 AddNodeIDNode(ID, this);
11122 }
11123
11124 namespace {
11125
11126 struct EVTArray {
11127 std::vector<EVT> VTs;
11128
EVTArray__anon569ad0941511::EVTArray11129 EVTArray() {
11130 VTs.reserve(MVT::VALUETYPE_SIZE);
11131 for (unsigned i = 0; i < MVT::VALUETYPE_SIZE; ++i)
11132 VTs.push_back(MVT((MVT::SimpleValueType)i));
11133 }
11134 };
11135
11136 } // end anonymous namespace
11137
11138 /// getValueTypeList - Return a pointer to the specified value type.
11139 ///
getValueTypeList(EVT VT)11140 const EVT *SDNode::getValueTypeList(EVT VT) {
11141 static std::set<EVT, EVT::compareRawBits> EVTs;
11142 static EVTArray SimpleVTArray;
11143 static sys::SmartMutex<true> VTMutex;
11144
11145 if (VT.isExtended()) {
11146 sys::SmartScopedLock<true> Lock(VTMutex);
11147 return &(*EVTs.insert(VT).first);
11148 }
11149 assert(VT.getSimpleVT() < MVT::VALUETYPE_SIZE && "Value type out of range!");
11150 return &SimpleVTArray.VTs[VT.getSimpleVT().SimpleTy];
11151 }
11152
11153 /// hasNUsesOfValue - Return true if there are exactly NUSES uses of the
11154 /// indicated value. This method ignores uses of other values defined by this
11155 /// operation.
hasNUsesOfValue(unsigned NUses,unsigned Value) const11156 bool SDNode::hasNUsesOfValue(unsigned NUses, unsigned Value) const {
11157 assert(Value < getNumValues() && "Bad value!");
11158
11159 // TODO: Only iterate over uses of a given value of the node
11160 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI) {
11161 if (UI.getUse().getResNo() == Value) {
11162 if (NUses == 0)
11163 return false;
11164 --NUses;
11165 }
11166 }
11167
11168 // Found exactly the right number of uses?
11169 return NUses == 0;
11170 }
11171
11172 /// hasAnyUseOfValue - Return true if there are any use of the indicated
11173 /// value. This method ignores uses of other values defined by this operation.
hasAnyUseOfValue(unsigned Value) const11174 bool SDNode::hasAnyUseOfValue(unsigned Value) const {
11175 assert(Value < getNumValues() && "Bad value!");
11176
11177 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI)
11178 if (UI.getUse().getResNo() == Value)
11179 return true;
11180
11181 return false;
11182 }
11183
11184 /// isOnlyUserOf - Return true if this node is the only use of N.
isOnlyUserOf(const SDNode * N) const11185 bool SDNode::isOnlyUserOf(const SDNode *N) const {
11186 bool Seen = false;
11187 for (const SDNode *User : N->uses()) {
11188 if (User == this)
11189 Seen = true;
11190 else
11191 return false;
11192 }
11193
11194 return Seen;
11195 }
11196
11197 /// Return true if the only users of N are contained in Nodes.
areOnlyUsersOf(ArrayRef<const SDNode * > Nodes,const SDNode * N)11198 bool SDNode::areOnlyUsersOf(ArrayRef<const SDNode *> Nodes, const SDNode *N) {
11199 bool Seen = false;
11200 for (const SDNode *User : N->uses()) {
11201 if (llvm::is_contained(Nodes, User))
11202 Seen = true;
11203 else
11204 return false;
11205 }
11206
11207 return Seen;
11208 }
11209
11210 /// isOperand - Return true if this node is an operand of N.
isOperandOf(const SDNode * N) const11211 bool SDValue::isOperandOf(const SDNode *N) const {
11212 return is_contained(N->op_values(), *this);
11213 }
11214
isOperandOf(const SDNode * N) const11215 bool SDNode::isOperandOf(const SDNode *N) const {
11216 return any_of(N->op_values(),
11217 [this](SDValue Op) { return this == Op.getNode(); });
11218 }
11219
11220 /// reachesChainWithoutSideEffects - Return true if this operand (which must
11221 /// be a chain) reaches the specified operand without crossing any
11222 /// side-effecting instructions on any chain path. In practice, this looks
11223 /// through token factors and non-volatile loads. In order to remain efficient,
11224 /// this only looks a couple of nodes in, it does not do an exhaustive search.
11225 ///
11226 /// Note that we only need to examine chains when we're searching for
11227 /// side-effects; SelectionDAG requires that all side-effects are represented
11228 /// by chains, even if another operand would force a specific ordering. This
11229 /// constraint is necessary to allow transformations like splitting loads.
reachesChainWithoutSideEffects(SDValue Dest,unsigned Depth) const11230 bool SDValue::reachesChainWithoutSideEffects(SDValue Dest,
11231 unsigned Depth) const {
11232 if (*this == Dest) return true;
11233
11234 // Don't search too deeply, we just want to be able to see through
11235 // TokenFactor's etc.
11236 if (Depth == 0) return false;
11237
11238 // If this is a token factor, all inputs to the TF happen in parallel.
11239 if (getOpcode() == ISD::TokenFactor) {
11240 // First, try a shallow search.
11241 if (is_contained((*this)->ops(), Dest)) {
11242 // We found the chain we want as an operand of this TokenFactor.
11243 // Essentially, we reach the chain without side-effects if we could
11244 // serialize the TokenFactor into a simple chain of operations with
11245 // Dest as the last operation. This is automatically true if the
11246 // chain has one use: there are no other ordering constraints.
11247 // If the chain has more than one use, we give up: some other
11248 // use of Dest might force a side-effect between Dest and the current
11249 // node.
11250 if (Dest.hasOneUse())
11251 return true;
11252 }
11253 // Next, try a deep search: check whether every operand of the TokenFactor
11254 // reaches Dest.
11255 return llvm::all_of((*this)->ops(), [=](SDValue Op) {
11256 return Op.reachesChainWithoutSideEffects(Dest, Depth - 1);
11257 });
11258 }
11259
11260 // Loads don't have side effects, look through them.
11261 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(*this)) {
11262 if (Ld->isUnordered())
11263 return Ld->getChain().reachesChainWithoutSideEffects(Dest, Depth-1);
11264 }
11265 return false;
11266 }
11267
hasPredecessor(const SDNode * N) const11268 bool SDNode::hasPredecessor(const SDNode *N) const {
11269 SmallPtrSet<const SDNode *, 32> Visited;
11270 SmallVector<const SDNode *, 16> Worklist;
11271 Worklist.push_back(this);
11272 return hasPredecessorHelper(N, Visited, Worklist);
11273 }
11274
intersectFlagsWith(const SDNodeFlags Flags)11275 void SDNode::intersectFlagsWith(const SDNodeFlags Flags) {
11276 this->Flags.intersectWith(Flags);
11277 }
11278
11279 SDValue
matchBinOpReduction(SDNode * Extract,ISD::NodeType & BinOp,ArrayRef<ISD::NodeType> CandidateBinOps,bool AllowPartials)11280 SelectionDAG::matchBinOpReduction(SDNode *Extract, ISD::NodeType &BinOp,
11281 ArrayRef<ISD::NodeType> CandidateBinOps,
11282 bool AllowPartials) {
11283 // The pattern must end in an extract from index 0.
11284 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
11285 !isNullConstant(Extract->getOperand(1)))
11286 return SDValue();
11287
11288 // Match against one of the candidate binary ops.
11289 SDValue Op = Extract->getOperand(0);
11290 if (llvm::none_of(CandidateBinOps, [Op](ISD::NodeType BinOp) {
11291 return Op.getOpcode() == unsigned(BinOp);
11292 }))
11293 return SDValue();
11294
11295 // Floating-point reductions may require relaxed constraints on the final step
11296 // of the reduction because they may reorder intermediate operations.
11297 unsigned CandidateBinOp = Op.getOpcode();
11298 if (Op.getValueType().isFloatingPoint()) {
11299 SDNodeFlags Flags = Op->getFlags();
11300 switch (CandidateBinOp) {
11301 case ISD::FADD:
11302 if (!Flags.hasNoSignedZeros() || !Flags.hasAllowReassociation())
11303 return SDValue();
11304 break;
11305 default:
11306 llvm_unreachable("Unhandled FP opcode for binop reduction");
11307 }
11308 }
11309
11310 // Matching failed - attempt to see if we did enough stages that a partial
11311 // reduction from a subvector is possible.
11312 auto PartialReduction = [&](SDValue Op, unsigned NumSubElts) {
11313 if (!AllowPartials || !Op)
11314 return SDValue();
11315 EVT OpVT = Op.getValueType();
11316 EVT OpSVT = OpVT.getScalarType();
11317 EVT SubVT = EVT::getVectorVT(*getContext(), OpSVT, NumSubElts);
11318 if (!TLI->isExtractSubvectorCheap(SubVT, OpVT, 0))
11319 return SDValue();
11320 BinOp = (ISD::NodeType)CandidateBinOp;
11321 return getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(Op), SubVT, Op,
11322 getVectorIdxConstant(0, SDLoc(Op)));
11323 };
11324
11325 // At each stage, we're looking for something that looks like:
11326 // %s = shufflevector <8 x i32> %op, <8 x i32> undef,
11327 // <8 x i32> <i32 2, i32 3, i32 undef, i32 undef,
11328 // i32 undef, i32 undef, i32 undef, i32 undef>
11329 // %a = binop <8 x i32> %op, %s
11330 // Where the mask changes according to the stage. E.g. for a 3-stage pyramid,
11331 // we expect something like:
11332 // <4,5,6,7,u,u,u,u>
11333 // <2,3,u,u,u,u,u,u>
11334 // <1,u,u,u,u,u,u,u>
11335 // While a partial reduction match would be:
11336 // <2,3,u,u,u,u,u,u>
11337 // <1,u,u,u,u,u,u,u>
11338 unsigned Stages = Log2_32(Op.getValueType().getVectorNumElements());
11339 SDValue PrevOp;
11340 for (unsigned i = 0; i < Stages; ++i) {
11341 unsigned MaskEnd = (1 << i);
11342
11343 if (Op.getOpcode() != CandidateBinOp)
11344 return PartialReduction(PrevOp, MaskEnd);
11345
11346 SDValue Op0 = Op.getOperand(0);
11347 SDValue Op1 = Op.getOperand(1);
11348
11349 ShuffleVectorSDNode *Shuffle = dyn_cast<ShuffleVectorSDNode>(Op0);
11350 if (Shuffle) {
11351 Op = Op1;
11352 } else {
11353 Shuffle = dyn_cast<ShuffleVectorSDNode>(Op1);
11354 Op = Op0;
11355 }
11356
11357 // The first operand of the shuffle should be the same as the other operand
11358 // of the binop.
11359 if (!Shuffle || Shuffle->getOperand(0) != Op)
11360 return PartialReduction(PrevOp, MaskEnd);
11361
11362 // Verify the shuffle has the expected (at this stage of the pyramid) mask.
11363 for (int Index = 0; Index < (int)MaskEnd; ++Index)
11364 if (Shuffle->getMaskElt(Index) != (int)(MaskEnd + Index))
11365 return PartialReduction(PrevOp, MaskEnd);
11366
11367 PrevOp = Op;
11368 }
11369
11370 // Handle subvector reductions, which tend to appear after the shuffle
11371 // reduction stages.
11372 while (Op.getOpcode() == CandidateBinOp) {
11373 unsigned NumElts = Op.getValueType().getVectorNumElements();
11374 SDValue Op0 = Op.getOperand(0);
11375 SDValue Op1 = Op.getOperand(1);
11376 if (Op0.getOpcode() != ISD::EXTRACT_SUBVECTOR ||
11377 Op1.getOpcode() != ISD::EXTRACT_SUBVECTOR ||
11378 Op0.getOperand(0) != Op1.getOperand(0))
11379 break;
11380 SDValue Src = Op0.getOperand(0);
11381 unsigned NumSrcElts = Src.getValueType().getVectorNumElements();
11382 if (NumSrcElts != (2 * NumElts))
11383 break;
11384 if (!(Op0.getConstantOperandAPInt(1) == 0 &&
11385 Op1.getConstantOperandAPInt(1) == NumElts) &&
11386 !(Op1.getConstantOperandAPInt(1) == 0 &&
11387 Op0.getConstantOperandAPInt(1) == NumElts))
11388 break;
11389 Op = Src;
11390 }
11391
11392 BinOp = (ISD::NodeType)CandidateBinOp;
11393 return Op;
11394 }
11395
UnrollVectorOp(SDNode * N,unsigned ResNE)11396 SDValue SelectionDAG::UnrollVectorOp(SDNode *N, unsigned ResNE) {
11397 assert(N->getNumValues() == 1 &&
11398 "Can't unroll a vector with multiple results!");
11399
11400 EVT VT = N->getValueType(0);
11401 unsigned NE = VT.getVectorNumElements();
11402 EVT EltVT = VT.getVectorElementType();
11403 SDLoc dl(N);
11404
11405 SmallVector<SDValue, 8> Scalars;
11406 SmallVector<SDValue, 4> Operands(N->getNumOperands());
11407
11408 // If ResNE is 0, fully unroll the vector op.
11409 if (ResNE == 0)
11410 ResNE = NE;
11411 else if (NE > ResNE)
11412 NE = ResNE;
11413
11414 unsigned i;
11415 for (i= 0; i != NE; ++i) {
11416 for (unsigned j = 0, e = N->getNumOperands(); j != e; ++j) {
11417 SDValue Operand = N->getOperand(j);
11418 EVT OperandVT = Operand.getValueType();
11419 if (OperandVT.isVector()) {
11420 // A vector operand; extract a single element.
11421 EVT OperandEltVT = OperandVT.getVectorElementType();
11422 Operands[j] = getNode(ISD::EXTRACT_VECTOR_ELT, dl, OperandEltVT,
11423 Operand, getVectorIdxConstant(i, dl));
11424 } else {
11425 // A scalar operand; just use it as is.
11426 Operands[j] = Operand;
11427 }
11428 }
11429
11430 switch (N->getOpcode()) {
11431 default: {
11432 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT, Operands,
11433 N->getFlags()));
11434 break;
11435 }
11436 case ISD::VSELECT:
11437 Scalars.push_back(getNode(ISD::SELECT, dl, EltVT, Operands));
11438 break;
11439 case ISD::SHL:
11440 case ISD::SRA:
11441 case ISD::SRL:
11442 case ISD::ROTL:
11443 case ISD::ROTR:
11444 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT, Operands[0],
11445 getShiftAmountOperand(Operands[0].getValueType(),
11446 Operands[1])));
11447 break;
11448 case ISD::SIGN_EXTEND_INREG: {
11449 EVT ExtVT = cast<VTSDNode>(Operands[1])->getVT().getVectorElementType();
11450 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT,
11451 Operands[0],
11452 getValueType(ExtVT)));
11453 }
11454 }
11455 }
11456
11457 for (; i < ResNE; ++i)
11458 Scalars.push_back(getUNDEF(EltVT));
11459
11460 EVT VecVT = EVT::getVectorVT(*getContext(), EltVT, ResNE);
11461 return getBuildVector(VecVT, dl, Scalars);
11462 }
11463
UnrollVectorOverflowOp(SDNode * N,unsigned ResNE)11464 std::pair<SDValue, SDValue> SelectionDAG::UnrollVectorOverflowOp(
11465 SDNode *N, unsigned ResNE) {
11466 unsigned Opcode = N->getOpcode();
11467 assert((Opcode == ISD::UADDO || Opcode == ISD::SADDO ||
11468 Opcode == ISD::USUBO || Opcode == ISD::SSUBO ||
11469 Opcode == ISD::UMULO || Opcode == ISD::SMULO) &&
11470 "Expected an overflow opcode");
11471
11472 EVT ResVT = N->getValueType(0);
11473 EVT OvVT = N->getValueType(1);
11474 EVT ResEltVT = ResVT.getVectorElementType();
11475 EVT OvEltVT = OvVT.getVectorElementType();
11476 SDLoc dl(N);
11477
11478 // If ResNE is 0, fully unroll the vector op.
11479 unsigned NE = ResVT.getVectorNumElements();
11480 if (ResNE == 0)
11481 ResNE = NE;
11482 else if (NE > ResNE)
11483 NE = ResNE;
11484
11485 SmallVector<SDValue, 8> LHSScalars;
11486 SmallVector<SDValue, 8> RHSScalars;
11487 ExtractVectorElements(N->getOperand(0), LHSScalars, 0, NE);
11488 ExtractVectorElements(N->getOperand(1), RHSScalars, 0, NE);
11489
11490 EVT SVT = TLI->getSetCCResultType(getDataLayout(), *getContext(), ResEltVT);
11491 SDVTList VTs = getVTList(ResEltVT, SVT);
11492 SmallVector<SDValue, 8> ResScalars;
11493 SmallVector<SDValue, 8> OvScalars;
11494 for (unsigned i = 0; i < NE; ++i) {
11495 SDValue Res = getNode(Opcode, dl, VTs, LHSScalars[i], RHSScalars[i]);
11496 SDValue Ov =
11497 getSelect(dl, OvEltVT, Res.getValue(1),
11498 getBoolConstant(true, dl, OvEltVT, ResVT),
11499 getConstant(0, dl, OvEltVT));
11500
11501 ResScalars.push_back(Res);
11502 OvScalars.push_back(Ov);
11503 }
11504
11505 ResScalars.append(ResNE - NE, getUNDEF(ResEltVT));
11506 OvScalars.append(ResNE - NE, getUNDEF(OvEltVT));
11507
11508 EVT NewResVT = EVT::getVectorVT(*getContext(), ResEltVT, ResNE);
11509 EVT NewOvVT = EVT::getVectorVT(*getContext(), OvEltVT, ResNE);
11510 return std::make_pair(getBuildVector(NewResVT, dl, ResScalars),
11511 getBuildVector(NewOvVT, dl, OvScalars));
11512 }
11513
areNonVolatileConsecutiveLoads(LoadSDNode * LD,LoadSDNode * Base,unsigned Bytes,int Dist) const11514 bool SelectionDAG::areNonVolatileConsecutiveLoads(LoadSDNode *LD,
11515 LoadSDNode *Base,
11516 unsigned Bytes,
11517 int Dist) const {
11518 if (LD->isVolatile() || Base->isVolatile())
11519 return false;
11520 // TODO: probably too restrictive for atomics, revisit
11521 if (!LD->isSimple())
11522 return false;
11523 if (LD->isIndexed() || Base->isIndexed())
11524 return false;
11525 if (LD->getChain() != Base->getChain())
11526 return false;
11527 EVT VT = LD->getMemoryVT();
11528 if (VT.getSizeInBits() / 8 != Bytes)
11529 return false;
11530
11531 auto BaseLocDecomp = BaseIndexOffset::match(Base, *this);
11532 auto LocDecomp = BaseIndexOffset::match(LD, *this);
11533
11534 int64_t Offset = 0;
11535 if (BaseLocDecomp.equalBaseIndex(LocDecomp, *this, Offset))
11536 return (Dist * Bytes == Offset);
11537 return false;
11538 }
11539
11540 /// InferPtrAlignment - Infer alignment of a load / store address. Return
11541 /// std::nullopt if it cannot be inferred.
InferPtrAlign(SDValue Ptr) const11542 MaybeAlign SelectionDAG::InferPtrAlign(SDValue Ptr) const {
11543 // If this is a GlobalAddress + cst, return the alignment.
11544 const GlobalValue *GV = nullptr;
11545 int64_t GVOffset = 0;
11546 if (TLI->isGAPlusOffset(Ptr.getNode(), GV, GVOffset)) {
11547 unsigned PtrWidth = getDataLayout().getPointerTypeSizeInBits(GV->getType());
11548 KnownBits Known(PtrWidth);
11549 llvm::computeKnownBits(GV, Known, getDataLayout());
11550 unsigned AlignBits = Known.countMinTrailingZeros();
11551 if (AlignBits)
11552 return commonAlignment(Align(1ull << std::min(31U, AlignBits)), GVOffset);
11553 }
11554
11555 // If this is a direct reference to a stack slot, use information about the
11556 // stack slot's alignment.
11557 int FrameIdx = INT_MIN;
11558 int64_t FrameOffset = 0;
11559 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Ptr)) {
11560 FrameIdx = FI->getIndex();
11561 } else if (isBaseWithConstantOffset(Ptr) &&
11562 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
11563 // Handle FI+Cst
11564 FrameIdx = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
11565 FrameOffset = Ptr.getConstantOperandVal(1);
11566 }
11567
11568 if (FrameIdx != INT_MIN) {
11569 const MachineFrameInfo &MFI = getMachineFunction().getFrameInfo();
11570 return commonAlignment(MFI.getObjectAlign(FrameIdx), FrameOffset);
11571 }
11572
11573 return std::nullopt;
11574 }
11575
11576 /// GetSplitDestVTs - Compute the VTs needed for the low/hi parts of a type
11577 /// which is split (or expanded) into two not necessarily identical pieces.
GetSplitDestVTs(const EVT & VT) const11578 std::pair<EVT, EVT> SelectionDAG::GetSplitDestVTs(const EVT &VT) const {
11579 // Currently all types are split in half.
11580 EVT LoVT, HiVT;
11581 if (!VT.isVector())
11582 LoVT = HiVT = TLI->getTypeToTransformTo(*getContext(), VT);
11583 else
11584 LoVT = HiVT = VT.getHalfNumVectorElementsVT(*getContext());
11585
11586 return std::make_pair(LoVT, HiVT);
11587 }
11588
11589 /// GetDependentSplitDestVTs - Compute the VTs needed for the low/hi parts of a
11590 /// type, dependent on an enveloping VT that has been split into two identical
11591 /// pieces. Sets the HiIsEmpty flag when hi type has zero storage size.
11592 std::pair<EVT, EVT>
GetDependentSplitDestVTs(const EVT & VT,const EVT & EnvVT,bool * HiIsEmpty) const11593 SelectionDAG::GetDependentSplitDestVTs(const EVT &VT, const EVT &EnvVT,
11594 bool *HiIsEmpty) const {
11595 EVT EltTp = VT.getVectorElementType();
11596 // Examples:
11597 // custom VL=8 with enveloping VL=8/8 yields 8/0 (hi empty)
11598 // custom VL=9 with enveloping VL=8/8 yields 8/1
11599 // custom VL=10 with enveloping VL=8/8 yields 8/2
11600 // etc.
11601 ElementCount VTNumElts = VT.getVectorElementCount();
11602 ElementCount EnvNumElts = EnvVT.getVectorElementCount();
11603 assert(VTNumElts.isScalable() == EnvNumElts.isScalable() &&
11604 "Mixing fixed width and scalable vectors when enveloping a type");
11605 EVT LoVT, HiVT;
11606 if (VTNumElts.getKnownMinValue() > EnvNumElts.getKnownMinValue()) {
11607 LoVT = EVT::getVectorVT(*getContext(), EltTp, EnvNumElts);
11608 HiVT = EVT::getVectorVT(*getContext(), EltTp, VTNumElts - EnvNumElts);
11609 *HiIsEmpty = false;
11610 } else {
11611 // Flag that hi type has zero storage size, but return split envelop type
11612 // (this would be easier if vector types with zero elements were allowed).
11613 LoVT = EVT::getVectorVT(*getContext(), EltTp, VTNumElts);
11614 HiVT = EVT::getVectorVT(*getContext(), EltTp, EnvNumElts);
11615 *HiIsEmpty = true;
11616 }
11617 return std::make_pair(LoVT, HiVT);
11618 }
11619
11620 /// SplitVector - Split the vector with EXTRACT_SUBVECTOR and return the
11621 /// low/high part.
11622 std::pair<SDValue, SDValue>
SplitVector(const SDValue & N,const SDLoc & DL,const EVT & LoVT,const EVT & HiVT)11623 SelectionDAG::SplitVector(const SDValue &N, const SDLoc &DL, const EVT &LoVT,
11624 const EVT &HiVT) {
11625 assert(LoVT.isScalableVector() == HiVT.isScalableVector() &&
11626 LoVT.isScalableVector() == N.getValueType().isScalableVector() &&
11627 "Splitting vector with an invalid mixture of fixed and scalable "
11628 "vector types");
11629 assert(LoVT.getVectorMinNumElements() + HiVT.getVectorMinNumElements() <=
11630 N.getValueType().getVectorMinNumElements() &&
11631 "More vector elements requested than available!");
11632 SDValue Lo, Hi;
11633 Lo =
11634 getNode(ISD::EXTRACT_SUBVECTOR, DL, LoVT, N, getVectorIdxConstant(0, DL));
11635 // For scalable vectors it is safe to use LoVT.getVectorMinNumElements()
11636 // (rather than having to use ElementCount), because EXTRACT_SUBVECTOR scales
11637 // IDX with the runtime scaling factor of the result vector type. For
11638 // fixed-width result vectors, that runtime scaling factor is 1.
11639 Hi = getNode(ISD::EXTRACT_SUBVECTOR, DL, HiVT, N,
11640 getVectorIdxConstant(LoVT.getVectorMinNumElements(), DL));
11641 return std::make_pair(Lo, Hi);
11642 }
11643
SplitEVL(SDValue N,EVT VecVT,const SDLoc & DL)11644 std::pair<SDValue, SDValue> SelectionDAG::SplitEVL(SDValue N, EVT VecVT,
11645 const SDLoc &DL) {
11646 // Split the vector length parameter.
11647 // %evl -> umin(%evl, %halfnumelts) and usubsat(%evl - %halfnumelts).
11648 EVT VT = N.getValueType();
11649 assert(VecVT.getVectorElementCount().isKnownEven() &&
11650 "Expecting the mask to be an evenly-sized vector");
11651 unsigned HalfMinNumElts = VecVT.getVectorMinNumElements() / 2;
11652 SDValue HalfNumElts =
11653 VecVT.isFixedLengthVector()
11654 ? getConstant(HalfMinNumElts, DL, VT)
11655 : getVScale(DL, VT, APInt(VT.getScalarSizeInBits(), HalfMinNumElts));
11656 SDValue Lo = getNode(ISD::UMIN, DL, VT, N, HalfNumElts);
11657 SDValue Hi = getNode(ISD::USUBSAT, DL, VT, N, HalfNumElts);
11658 return std::make_pair(Lo, Hi);
11659 }
11660
11661 /// Widen the vector up to the next power of two using INSERT_SUBVECTOR.
WidenVector(const SDValue & N,const SDLoc & DL)11662 SDValue SelectionDAG::WidenVector(const SDValue &N, const SDLoc &DL) {
11663 EVT VT = N.getValueType();
11664 EVT WideVT = EVT::getVectorVT(*getContext(), VT.getVectorElementType(),
11665 NextPowerOf2(VT.getVectorNumElements()));
11666 return getNode(ISD::INSERT_SUBVECTOR, DL, WideVT, getUNDEF(WideVT), N,
11667 getVectorIdxConstant(0, DL));
11668 }
11669
ExtractVectorElements(SDValue Op,SmallVectorImpl<SDValue> & Args,unsigned Start,unsigned Count,EVT EltVT)11670 void SelectionDAG::ExtractVectorElements(SDValue Op,
11671 SmallVectorImpl<SDValue> &Args,
11672 unsigned Start, unsigned Count,
11673 EVT EltVT) {
11674 EVT VT = Op.getValueType();
11675 if (Count == 0)
11676 Count = VT.getVectorNumElements();
11677 if (EltVT == EVT())
11678 EltVT = VT.getVectorElementType();
11679 SDLoc SL(Op);
11680 for (unsigned i = Start, e = Start + Count; i != e; ++i) {
11681 Args.push_back(getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT, Op,
11682 getVectorIdxConstant(i, SL)));
11683 }
11684 }
11685
11686 // getAddressSpace - Return the address space this GlobalAddress belongs to.
getAddressSpace() const11687 unsigned GlobalAddressSDNode::getAddressSpace() const {
11688 return getGlobal()->getType()->getAddressSpace();
11689 }
11690
getType() const11691 Type *ConstantPoolSDNode::getType() const {
11692 if (isMachineConstantPoolEntry())
11693 return Val.MachineCPVal->getType();
11694 return Val.ConstVal->getType();
11695 }
11696
isConstantSplat(APInt & SplatValue,APInt & SplatUndef,unsigned & SplatBitSize,bool & HasAnyUndefs,unsigned MinSplatBits,bool IsBigEndian) const11697 bool BuildVectorSDNode::isConstantSplat(APInt &SplatValue, APInt &SplatUndef,
11698 unsigned &SplatBitSize,
11699 bool &HasAnyUndefs,
11700 unsigned MinSplatBits,
11701 bool IsBigEndian) const {
11702 EVT VT = getValueType(0);
11703 assert(VT.isVector() && "Expected a vector type");
11704 unsigned VecWidth = VT.getSizeInBits();
11705 if (MinSplatBits > VecWidth)
11706 return false;
11707
11708 // FIXME: The widths are based on this node's type, but build vectors can
11709 // truncate their operands.
11710 SplatValue = APInt(VecWidth, 0);
11711 SplatUndef = APInt(VecWidth, 0);
11712
11713 // Get the bits. Bits with undefined values (when the corresponding element
11714 // of the vector is an ISD::UNDEF value) are set in SplatUndef and cleared
11715 // in SplatValue. If any of the values are not constant, give up and return
11716 // false.
11717 unsigned int NumOps = getNumOperands();
11718 assert(NumOps > 0 && "isConstantSplat has 0-size build vector");
11719 unsigned EltWidth = VT.getScalarSizeInBits();
11720
11721 for (unsigned j = 0; j < NumOps; ++j) {
11722 unsigned i = IsBigEndian ? NumOps - 1 - j : j;
11723 SDValue OpVal = getOperand(i);
11724 unsigned BitPos = j * EltWidth;
11725
11726 if (OpVal.isUndef())
11727 SplatUndef.setBits(BitPos, BitPos + EltWidth);
11728 else if (auto *CN = dyn_cast<ConstantSDNode>(OpVal))
11729 SplatValue.insertBits(CN->getAPIntValue().zextOrTrunc(EltWidth), BitPos);
11730 else if (auto *CN = dyn_cast<ConstantFPSDNode>(OpVal))
11731 SplatValue.insertBits(CN->getValueAPF().bitcastToAPInt(), BitPos);
11732 else
11733 return false;
11734 }
11735
11736 // The build_vector is all constants or undefs. Find the smallest element
11737 // size that splats the vector.
11738 HasAnyUndefs = (SplatUndef != 0);
11739
11740 // FIXME: This does not work for vectors with elements less than 8 bits.
11741 while (VecWidth > 8) {
11742 unsigned HalfSize = VecWidth / 2;
11743 APInt HighValue = SplatValue.extractBits(HalfSize, HalfSize);
11744 APInt LowValue = SplatValue.extractBits(HalfSize, 0);
11745 APInt HighUndef = SplatUndef.extractBits(HalfSize, HalfSize);
11746 APInt LowUndef = SplatUndef.extractBits(HalfSize, 0);
11747
11748 // If the two halves do not match (ignoring undef bits), stop here.
11749 if ((HighValue & ~LowUndef) != (LowValue & ~HighUndef) ||
11750 MinSplatBits > HalfSize)
11751 break;
11752
11753 SplatValue = HighValue | LowValue;
11754 SplatUndef = HighUndef & LowUndef;
11755
11756 VecWidth = HalfSize;
11757 }
11758
11759 SplatBitSize = VecWidth;
11760 return true;
11761 }
11762
getSplatValue(const APInt & DemandedElts,BitVector * UndefElements) const11763 SDValue BuildVectorSDNode::getSplatValue(const APInt &DemandedElts,
11764 BitVector *UndefElements) const {
11765 unsigned NumOps = getNumOperands();
11766 if (UndefElements) {
11767 UndefElements->clear();
11768 UndefElements->resize(NumOps);
11769 }
11770 assert(NumOps == DemandedElts.getBitWidth() && "Unexpected vector size");
11771 if (!DemandedElts)
11772 return SDValue();
11773 SDValue Splatted;
11774 for (unsigned i = 0; i != NumOps; ++i) {
11775 if (!DemandedElts[i])
11776 continue;
11777 SDValue Op = getOperand(i);
11778 if (Op.isUndef()) {
11779 if (UndefElements)
11780 (*UndefElements)[i] = true;
11781 } else if (!Splatted) {
11782 Splatted = Op;
11783 } else if (Splatted != Op) {
11784 return SDValue();
11785 }
11786 }
11787
11788 if (!Splatted) {
11789 unsigned FirstDemandedIdx = DemandedElts.countTrailingZeros();
11790 assert(getOperand(FirstDemandedIdx).isUndef() &&
11791 "Can only have a splat without a constant for all undefs.");
11792 return getOperand(FirstDemandedIdx);
11793 }
11794
11795 return Splatted;
11796 }
11797
getSplatValue(BitVector * UndefElements) const11798 SDValue BuildVectorSDNode::getSplatValue(BitVector *UndefElements) const {
11799 APInt DemandedElts = APInt::getAllOnes(getNumOperands());
11800 return getSplatValue(DemandedElts, UndefElements);
11801 }
11802
getRepeatedSequence(const APInt & DemandedElts,SmallVectorImpl<SDValue> & Sequence,BitVector * UndefElements) const11803 bool BuildVectorSDNode::getRepeatedSequence(const APInt &DemandedElts,
11804 SmallVectorImpl<SDValue> &Sequence,
11805 BitVector *UndefElements) const {
11806 unsigned NumOps = getNumOperands();
11807 Sequence.clear();
11808 if (UndefElements) {
11809 UndefElements->clear();
11810 UndefElements->resize(NumOps);
11811 }
11812 assert(NumOps == DemandedElts.getBitWidth() && "Unexpected vector size");
11813 if (!DemandedElts || NumOps < 2 || !isPowerOf2_32(NumOps))
11814 return false;
11815
11816 // Set the undefs even if we don't find a sequence (like getSplatValue).
11817 if (UndefElements)
11818 for (unsigned I = 0; I != NumOps; ++I)
11819 if (DemandedElts[I] && getOperand(I).isUndef())
11820 (*UndefElements)[I] = true;
11821
11822 // Iteratively widen the sequence length looking for repetitions.
11823 for (unsigned SeqLen = 1; SeqLen < NumOps; SeqLen *= 2) {
11824 Sequence.append(SeqLen, SDValue());
11825 for (unsigned I = 0; I != NumOps; ++I) {
11826 if (!DemandedElts[I])
11827 continue;
11828 SDValue &SeqOp = Sequence[I % SeqLen];
11829 SDValue Op = getOperand(I);
11830 if (Op.isUndef()) {
11831 if (!SeqOp)
11832 SeqOp = Op;
11833 continue;
11834 }
11835 if (SeqOp && !SeqOp.isUndef() && SeqOp != Op) {
11836 Sequence.clear();
11837 break;
11838 }
11839 SeqOp = Op;
11840 }
11841 if (!Sequence.empty())
11842 return true;
11843 }
11844
11845 assert(Sequence.empty() && "Failed to empty non-repeating sequence pattern");
11846 return false;
11847 }
11848
getRepeatedSequence(SmallVectorImpl<SDValue> & Sequence,BitVector * UndefElements) const11849 bool BuildVectorSDNode::getRepeatedSequence(SmallVectorImpl<SDValue> &Sequence,
11850 BitVector *UndefElements) const {
11851 APInt DemandedElts = APInt::getAllOnes(getNumOperands());
11852 return getRepeatedSequence(DemandedElts, Sequence, UndefElements);
11853 }
11854
11855 ConstantSDNode *
getConstantSplatNode(const APInt & DemandedElts,BitVector * UndefElements) const11856 BuildVectorSDNode::getConstantSplatNode(const APInt &DemandedElts,
11857 BitVector *UndefElements) const {
11858 return dyn_cast_or_null<ConstantSDNode>(
11859 getSplatValue(DemandedElts, UndefElements));
11860 }
11861
11862 ConstantSDNode *
getConstantSplatNode(BitVector * UndefElements) const11863 BuildVectorSDNode::getConstantSplatNode(BitVector *UndefElements) const {
11864 return dyn_cast_or_null<ConstantSDNode>(getSplatValue(UndefElements));
11865 }
11866
11867 ConstantFPSDNode *
getConstantFPSplatNode(const APInt & DemandedElts,BitVector * UndefElements) const11868 BuildVectorSDNode::getConstantFPSplatNode(const APInt &DemandedElts,
11869 BitVector *UndefElements) const {
11870 return dyn_cast_or_null<ConstantFPSDNode>(
11871 getSplatValue(DemandedElts, UndefElements));
11872 }
11873
11874 ConstantFPSDNode *
getConstantFPSplatNode(BitVector * UndefElements) const11875 BuildVectorSDNode::getConstantFPSplatNode(BitVector *UndefElements) const {
11876 return dyn_cast_or_null<ConstantFPSDNode>(getSplatValue(UndefElements));
11877 }
11878
11879 int32_t
getConstantFPSplatPow2ToLog2Int(BitVector * UndefElements,uint32_t BitWidth) const11880 BuildVectorSDNode::getConstantFPSplatPow2ToLog2Int(BitVector *UndefElements,
11881 uint32_t BitWidth) const {
11882 if (ConstantFPSDNode *CN =
11883 dyn_cast_or_null<ConstantFPSDNode>(getSplatValue(UndefElements))) {
11884 bool IsExact;
11885 APSInt IntVal(BitWidth);
11886 const APFloat &APF = CN->getValueAPF();
11887 if (APF.convertToInteger(IntVal, APFloat::rmTowardZero, &IsExact) !=
11888 APFloat::opOK ||
11889 !IsExact)
11890 return -1;
11891
11892 return IntVal.exactLogBase2();
11893 }
11894 return -1;
11895 }
11896
getConstantRawBits(bool IsLittleEndian,unsigned DstEltSizeInBits,SmallVectorImpl<APInt> & RawBitElements,BitVector & UndefElements) const11897 bool BuildVectorSDNode::getConstantRawBits(
11898 bool IsLittleEndian, unsigned DstEltSizeInBits,
11899 SmallVectorImpl<APInt> &RawBitElements, BitVector &UndefElements) const {
11900 // Early-out if this contains anything but Undef/Constant/ConstantFP.
11901 if (!isConstant())
11902 return false;
11903
11904 unsigned NumSrcOps = getNumOperands();
11905 unsigned SrcEltSizeInBits = getValueType(0).getScalarSizeInBits();
11906 assert(((NumSrcOps * SrcEltSizeInBits) % DstEltSizeInBits) == 0 &&
11907 "Invalid bitcast scale");
11908
11909 // Extract raw src bits.
11910 SmallVector<APInt> SrcBitElements(NumSrcOps,
11911 APInt::getNullValue(SrcEltSizeInBits));
11912 BitVector SrcUndeElements(NumSrcOps, false);
11913
11914 for (unsigned I = 0; I != NumSrcOps; ++I) {
11915 SDValue Op = getOperand(I);
11916 if (Op.isUndef()) {
11917 SrcUndeElements.set(I);
11918 continue;
11919 }
11920 auto *CInt = dyn_cast<ConstantSDNode>(Op);
11921 auto *CFP = dyn_cast<ConstantFPSDNode>(Op);
11922 assert((CInt || CFP) && "Unknown constant");
11923 SrcBitElements[I] = CInt ? CInt->getAPIntValue().trunc(SrcEltSizeInBits)
11924 : CFP->getValueAPF().bitcastToAPInt();
11925 }
11926
11927 // Recast to dst width.
11928 recastRawBits(IsLittleEndian, DstEltSizeInBits, RawBitElements,
11929 SrcBitElements, UndefElements, SrcUndeElements);
11930 return true;
11931 }
11932
recastRawBits(bool IsLittleEndian,unsigned DstEltSizeInBits,SmallVectorImpl<APInt> & DstBitElements,ArrayRef<APInt> SrcBitElements,BitVector & DstUndefElements,const BitVector & SrcUndefElements)11933 void BuildVectorSDNode::recastRawBits(bool IsLittleEndian,
11934 unsigned DstEltSizeInBits,
11935 SmallVectorImpl<APInt> &DstBitElements,
11936 ArrayRef<APInt> SrcBitElements,
11937 BitVector &DstUndefElements,
11938 const BitVector &SrcUndefElements) {
11939 unsigned NumSrcOps = SrcBitElements.size();
11940 unsigned SrcEltSizeInBits = SrcBitElements[0].getBitWidth();
11941 assert(((NumSrcOps * SrcEltSizeInBits) % DstEltSizeInBits) == 0 &&
11942 "Invalid bitcast scale");
11943 assert(NumSrcOps == SrcUndefElements.size() &&
11944 "Vector size mismatch");
11945
11946 unsigned NumDstOps = (NumSrcOps * SrcEltSizeInBits) / DstEltSizeInBits;
11947 DstUndefElements.clear();
11948 DstUndefElements.resize(NumDstOps, false);
11949 DstBitElements.assign(NumDstOps, APInt::getNullValue(DstEltSizeInBits));
11950
11951 // Concatenate src elements constant bits together into dst element.
11952 if (SrcEltSizeInBits <= DstEltSizeInBits) {
11953 unsigned Scale = DstEltSizeInBits / SrcEltSizeInBits;
11954 for (unsigned I = 0; I != NumDstOps; ++I) {
11955 DstUndefElements.set(I);
11956 APInt &DstBits = DstBitElements[I];
11957 for (unsigned J = 0; J != Scale; ++J) {
11958 unsigned Idx = (I * Scale) + (IsLittleEndian ? J : (Scale - J - 1));
11959 if (SrcUndefElements[Idx])
11960 continue;
11961 DstUndefElements.reset(I);
11962 const APInt &SrcBits = SrcBitElements[Idx];
11963 assert(SrcBits.getBitWidth() == SrcEltSizeInBits &&
11964 "Illegal constant bitwidths");
11965 DstBits.insertBits(SrcBits, J * SrcEltSizeInBits);
11966 }
11967 }
11968 return;
11969 }
11970
11971 // Split src element constant bits into dst elements.
11972 unsigned Scale = SrcEltSizeInBits / DstEltSizeInBits;
11973 for (unsigned I = 0; I != NumSrcOps; ++I) {
11974 if (SrcUndefElements[I]) {
11975 DstUndefElements.set(I * Scale, (I + 1) * Scale);
11976 continue;
11977 }
11978 const APInt &SrcBits = SrcBitElements[I];
11979 for (unsigned J = 0; J != Scale; ++J) {
11980 unsigned Idx = (I * Scale) + (IsLittleEndian ? J : (Scale - J - 1));
11981 APInt &DstBits = DstBitElements[Idx];
11982 DstBits = SrcBits.extractBits(DstEltSizeInBits, J * DstEltSizeInBits);
11983 }
11984 }
11985 }
11986
isConstant() const11987 bool BuildVectorSDNode::isConstant() const {
11988 for (const SDValue &Op : op_values()) {
11989 unsigned Opc = Op.getOpcode();
11990 if (Opc != ISD::UNDEF && Opc != ISD::Constant && Opc != ISD::ConstantFP)
11991 return false;
11992 }
11993 return true;
11994 }
11995
11996 std::optional<std::pair<APInt, APInt>>
isConstantSequence() const11997 BuildVectorSDNode::isConstantSequence() const {
11998 unsigned NumOps = getNumOperands();
11999 if (NumOps < 2)
12000 return std::nullopt;
12001
12002 if (!isa<ConstantSDNode>(getOperand(0)) ||
12003 !isa<ConstantSDNode>(getOperand(1)))
12004 return std::nullopt;
12005
12006 unsigned EltSize = getValueType(0).getScalarSizeInBits();
12007 APInt Start = getConstantOperandAPInt(0).trunc(EltSize);
12008 APInt Stride = getConstantOperandAPInt(1).trunc(EltSize) - Start;
12009
12010 if (Stride.isZero())
12011 return std::nullopt;
12012
12013 for (unsigned i = 2; i < NumOps; ++i) {
12014 if (!isa<ConstantSDNode>(getOperand(i)))
12015 return std::nullopt;
12016
12017 APInt Val = getConstantOperandAPInt(i).trunc(EltSize);
12018 if (Val != (Start + (Stride * i)))
12019 return std::nullopt;
12020 }
12021
12022 return std::make_pair(Start, Stride);
12023 }
12024
isSplatMask(const int * Mask,EVT VT)12025 bool ShuffleVectorSDNode::isSplatMask(const int *Mask, EVT VT) {
12026 // Find the first non-undef value in the shuffle mask.
12027 unsigned i, e;
12028 for (i = 0, e = VT.getVectorNumElements(); i != e && Mask[i] < 0; ++i)
12029 /* search */;
12030
12031 // If all elements are undefined, this shuffle can be considered a splat
12032 // (although it should eventually get simplified away completely).
12033 if (i == e)
12034 return true;
12035
12036 // Make sure all remaining elements are either undef or the same as the first
12037 // non-undef value.
12038 for (int Idx = Mask[i]; i != e; ++i)
12039 if (Mask[i] >= 0 && Mask[i] != Idx)
12040 return false;
12041 return true;
12042 }
12043
12044 // Returns the SDNode if it is a constant integer BuildVector
12045 // or constant integer.
isConstantIntBuildVectorOrConstantInt(SDValue N) const12046 SDNode *SelectionDAG::isConstantIntBuildVectorOrConstantInt(SDValue N) const {
12047 if (isa<ConstantSDNode>(N))
12048 return N.getNode();
12049 if (ISD::isBuildVectorOfConstantSDNodes(N.getNode()))
12050 return N.getNode();
12051 // Treat a GlobalAddress supporting constant offset folding as a
12052 // constant integer.
12053 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N))
12054 if (GA->getOpcode() == ISD::GlobalAddress &&
12055 TLI->isOffsetFoldingLegal(GA))
12056 return GA;
12057 if ((N.getOpcode() == ISD::SPLAT_VECTOR) &&
12058 isa<ConstantSDNode>(N.getOperand(0)))
12059 return N.getNode();
12060 return nullptr;
12061 }
12062
12063 // Returns the SDNode if it is a constant float BuildVector
12064 // or constant float.
isConstantFPBuildVectorOrConstantFP(SDValue N) const12065 SDNode *SelectionDAG::isConstantFPBuildVectorOrConstantFP(SDValue N) const {
12066 if (isa<ConstantFPSDNode>(N))
12067 return N.getNode();
12068
12069 if (ISD::isBuildVectorOfConstantFPSDNodes(N.getNode()))
12070 return N.getNode();
12071
12072 if ((N.getOpcode() == ISD::SPLAT_VECTOR) &&
12073 isa<ConstantFPSDNode>(N.getOperand(0)))
12074 return N.getNode();
12075
12076 return nullptr;
12077 }
12078
createOperands(SDNode * Node,ArrayRef<SDValue> Vals)12079 void SelectionDAG::createOperands(SDNode *Node, ArrayRef<SDValue> Vals) {
12080 assert(!Node->OperandList && "Node already has operands");
12081 assert(SDNode::getMaxNumOperands() >= Vals.size() &&
12082 "too many operands to fit into SDNode");
12083 SDUse *Ops = OperandRecycler.allocate(
12084 ArrayRecycler<SDUse>::Capacity::get(Vals.size()), OperandAllocator);
12085
12086 bool IsDivergent = false;
12087 for (unsigned I = 0; I != Vals.size(); ++I) {
12088 Ops[I].setUser(Node);
12089 Ops[I].setInitial(Vals[I]);
12090 if (Ops[I].Val.getValueType() != MVT::Other) // Skip Chain. It does not carry divergence.
12091 IsDivergent |= Ops[I].getNode()->isDivergent();
12092 }
12093 Node->NumOperands = Vals.size();
12094 Node->OperandList = Ops;
12095 if (!TLI->isSDNodeAlwaysUniform(Node)) {
12096 IsDivergent |= TLI->isSDNodeSourceOfDivergence(Node, FLI, DA);
12097 Node->SDNodeBits.IsDivergent = IsDivergent;
12098 }
12099 checkForCycles(Node);
12100 }
12101
getTokenFactor(const SDLoc & DL,SmallVectorImpl<SDValue> & Vals)12102 SDValue SelectionDAG::getTokenFactor(const SDLoc &DL,
12103 SmallVectorImpl<SDValue> &Vals) {
12104 size_t Limit = SDNode::getMaxNumOperands();
12105 while (Vals.size() > Limit) {
12106 unsigned SliceIdx = Vals.size() - Limit;
12107 auto ExtractedTFs = ArrayRef<SDValue>(Vals).slice(SliceIdx, Limit);
12108 SDValue NewTF = getNode(ISD::TokenFactor, DL, MVT::Other, ExtractedTFs);
12109 Vals.erase(Vals.begin() + SliceIdx, Vals.end());
12110 Vals.emplace_back(NewTF);
12111 }
12112 return getNode(ISD::TokenFactor, DL, MVT::Other, Vals);
12113 }
12114
getNeutralElement(unsigned Opcode,const SDLoc & DL,EVT VT,SDNodeFlags Flags)12115 SDValue SelectionDAG::getNeutralElement(unsigned Opcode, const SDLoc &DL,
12116 EVT VT, SDNodeFlags Flags) {
12117 switch (Opcode) {
12118 default:
12119 return SDValue();
12120 case ISD::ADD:
12121 case ISD::OR:
12122 case ISD::XOR:
12123 case ISD::UMAX:
12124 return getConstant(0, DL, VT);
12125 case ISD::MUL:
12126 return getConstant(1, DL, VT);
12127 case ISD::AND:
12128 case ISD::UMIN:
12129 return getAllOnesConstant(DL, VT);
12130 case ISD::SMAX:
12131 return getConstant(APInt::getSignedMinValue(VT.getSizeInBits()), DL, VT);
12132 case ISD::SMIN:
12133 return getConstant(APInt::getSignedMaxValue(VT.getSizeInBits()), DL, VT);
12134 case ISD::FADD:
12135 return getConstantFP(-0.0, DL, VT);
12136 case ISD::FMUL:
12137 return getConstantFP(1.0, DL, VT);
12138 case ISD::FMINNUM:
12139 case ISD::FMAXNUM: {
12140 // Neutral element for fminnum is NaN, Inf or FLT_MAX, depending on FMF.
12141 const fltSemantics &Semantics = EVTToAPFloatSemantics(VT);
12142 APFloat NeutralAF = !Flags.hasNoNaNs() ? APFloat::getQNaN(Semantics) :
12143 !Flags.hasNoInfs() ? APFloat::getInf(Semantics) :
12144 APFloat::getLargest(Semantics);
12145 if (Opcode == ISD::FMAXNUM)
12146 NeutralAF.changeSign();
12147
12148 return getConstantFP(NeutralAF, DL, VT);
12149 }
12150 }
12151 }
12152
copyExtraInfo(SDNode * From,SDNode * To)12153 void SelectionDAG::copyExtraInfo(SDNode *From, SDNode *To) {
12154 assert(From && To && "Invalid SDNode; empty source SDValue?");
12155 auto I = SDEI.find(From);
12156 if (I == SDEI.end())
12157 return;
12158
12159 // Use of operator[] on the DenseMap may cause an insertion, which invalidates
12160 // the iterator, hence the need to make a copy to prevent a use-after-free.
12161 NodeExtraInfo Copy = I->second;
12162 SDEI[To] = std::move(Copy);
12163 }
12164
12165 #ifndef NDEBUG
checkForCyclesHelper(const SDNode * N,SmallPtrSetImpl<const SDNode * > & Visited,SmallPtrSetImpl<const SDNode * > & Checked,const llvm::SelectionDAG * DAG)12166 static void checkForCyclesHelper(const SDNode *N,
12167 SmallPtrSetImpl<const SDNode*> &Visited,
12168 SmallPtrSetImpl<const SDNode*> &Checked,
12169 const llvm::SelectionDAG *DAG) {
12170 // If this node has already been checked, don't check it again.
12171 if (Checked.count(N))
12172 return;
12173
12174 // If a node has already been visited on this depth-first walk, reject it as
12175 // a cycle.
12176 if (!Visited.insert(N).second) {
12177 errs() << "Detected cycle in SelectionDAG\n";
12178 dbgs() << "Offending node:\n";
12179 N->dumprFull(DAG); dbgs() << "\n";
12180 abort();
12181 }
12182
12183 for (const SDValue &Op : N->op_values())
12184 checkForCyclesHelper(Op.getNode(), Visited, Checked, DAG);
12185
12186 Checked.insert(N);
12187 Visited.erase(N);
12188 }
12189 #endif
12190
checkForCycles(const llvm::SDNode * N,const llvm::SelectionDAG * DAG,bool force)12191 void llvm::checkForCycles(const llvm::SDNode *N,
12192 const llvm::SelectionDAG *DAG,
12193 bool force) {
12194 #ifndef NDEBUG
12195 bool check = force;
12196 #ifdef EXPENSIVE_CHECKS
12197 check = true;
12198 #endif // EXPENSIVE_CHECKS
12199 if (check) {
12200 assert(N && "Checking nonexistent SDNode");
12201 SmallPtrSet<const SDNode*, 32> visited;
12202 SmallPtrSet<const SDNode*, 32> checked;
12203 checkForCyclesHelper(N, visited, checked, DAG);
12204 }
12205 #endif // !NDEBUG
12206 }
12207
checkForCycles(const llvm::SelectionDAG * DAG,bool force)12208 void llvm::checkForCycles(const llvm::SelectionDAG *DAG, bool force) {
12209 checkForCycles(DAG->getRoot().getNode(), DAG, force);
12210 }
12211