1 //===-- DAGCombiner.cpp - Implement a DAG node combiner -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass combines dag nodes to form fewer, simpler DAG nodes. It can be run
11 // both before and after the DAG is legalized.
13 // This pass is not a substitute for the LLVM IR instcombine pass. This pass is
14 // primarily intended to handle simplification opportunities that are implicit
15 // in the LLVM IR and exposed by the various codegen lowering phases.
17 //===----------------------------------------------------------------------===//
19 #include "llvm/ADT/SetVector.h"
20 #include "llvm/ADT/SmallBitVector.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/Analysis/AliasAnalysis.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/SelectionDAG.h"
28 #include "llvm/CodeGen/SelectionDAGTargetInfo.h"
29 #include "llvm/IR/DataLayout.h"
30 #include "llvm/IR/DerivedTypes.h"
31 #include "llvm/IR/Function.h"
32 #include "llvm/IR/LLVMContext.h"
33 #include "llvm/Support/CommandLine.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/Support/KnownBits.h"
37 #include "llvm/Support/MathExtras.h"
38 #include "llvm/Support/raw_ostream.h"
39 #include "llvm/Target/TargetLowering.h"
40 #include "llvm/Target/TargetOptions.h"
41 #include "llvm/Target/TargetRegisterInfo.h"
42 #include "llvm/Target/TargetSubtargetInfo.h"
46 #define DEBUG_TYPE "dagcombine"
48 STATISTIC(NodesCombined , "Number of dag nodes combined");
49 STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created");
50 STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created");
51 STATISTIC(OpsNarrowed , "Number of load/op/store narrowed");
52 STATISTIC(LdStFP2Int , "Number of fp load/store pairs transformed to int");
53 STATISTIC(SlicedLoads, "Number of load sliced");
57 CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden,
58 cl::desc("Enable DAG combiner's use of IR alias analysis"));
61 UseTBAA("combiner-use-tbaa", cl::Hidden, cl::init(true),
62 cl::desc("Enable DAG combiner's use of TBAA"));
65 static cl::opt<std::string>
66 CombinerAAOnlyFunc("combiner-aa-only-func", cl::Hidden,
67 cl::desc("Only use DAG-combiner alias analysis in this"
71 /// Hidden option to stress test load slicing, i.e., when this option
72 /// is enabled, load slicing bypasses most of its profitability guards.
74 StressLoadSlicing("combiner-stress-load-slicing", cl::Hidden,
75 cl::desc("Bypass the profitability model of load "
80 MaySplitLoadIndex("combiner-split-load-index", cl::Hidden, cl::init(true),
81 cl::desc("DAG combiner may split indexing from loads"));
83 //------------------------------ DAGCombiner ---------------------------------//
87 const TargetLowering &TLI;
89 CodeGenOpt::Level OptLevel;
94 /// \brief Worklist of all of the nodes that need to be simplified.
96 /// This must behave as a stack -- new nodes to process are pushed onto the
97 /// back and when processing we pop off of the back.
99 /// The worklist will not contain duplicates but may contain null entries
100 /// due to nodes being deleted from the underlying DAG.
101 SmallVector<SDNode *, 64> Worklist;
103 /// \brief Mapping from an SDNode to its position on the worklist.
105 /// This is used to find and remove nodes from the worklist (by nulling
106 /// them) when they are deleted from the underlying DAG. It relies on
107 /// stable indices of nodes within the worklist.
108 DenseMap<SDNode *, unsigned> WorklistMap;
110 /// \brief Set of nodes which have been combined (at least once).
112 /// This is used to allow us to reliably add any operands of a DAG node
113 /// which have not yet been combined to the worklist.
114 SmallPtrSet<SDNode *, 32> CombinedNodes;
116 // AA - Used for DAG load/store alias analysis.
119 /// When an instruction is simplified, add all users of the instruction to
120 /// the work lists because they might get more simplified now.
121 void AddUsersToWorklist(SDNode *N) {
122 for (SDNode *Node : N->uses())
126 /// Call the node-specific routine that folds each particular type of node.
127 SDValue visit(SDNode *N);
130 /// Add to the worklist making sure its instance is at the back (next to be
132 void AddToWorklist(SDNode *N) {
133 assert(N->getOpcode() != ISD::DELETED_NODE &&
134 "Deleted Node added to Worklist");
136 // Skip handle nodes as they can't usefully be combined and confuse the
137 // zero-use deletion strategy.
138 if (N->getOpcode() == ISD::HANDLENODE)
141 if (WorklistMap.insert(std::make_pair(N, Worklist.size())).second)
142 Worklist.push_back(N);
145 /// Remove all instances of N from the worklist.
146 void removeFromWorklist(SDNode *N) {
147 CombinedNodes.erase(N);
149 auto It = WorklistMap.find(N);
150 if (It == WorklistMap.end())
151 return; // Not in the worklist.
153 // Null out the entry rather than erasing it to avoid a linear operation.
154 Worklist[It->second] = nullptr;
155 WorklistMap.erase(It);
158 void deleteAndRecombine(SDNode *N);
159 bool recursivelyDeleteUnusedNodes(SDNode *N);
161 /// Replaces all uses of the results of one DAG node with new values.
162 SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
165 /// Replaces all uses of the results of one DAG node with new values.
166 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) {
167 return CombineTo(N, &Res, 1, AddTo);
170 /// Replaces all uses of the results of one DAG node with new values.
171 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1,
173 SDValue To[] = { Res0, Res1 };
174 return CombineTo(N, To, 2, AddTo);
177 void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO);
180 unsigned MaximumLegalStoreInBits;
182 /// Check the specified integer node value to see if it can be simplified or
183 /// if things it uses can be simplified by bit propagation.
184 /// If so, return true.
185 bool SimplifyDemandedBits(SDValue Op) {
186 unsigned BitWidth = Op.getScalarValueSizeInBits();
187 APInt Demanded = APInt::getAllOnesValue(BitWidth);
188 return SimplifyDemandedBits(Op, Demanded);
191 bool SimplifyDemandedBits(SDValue Op, const APInt &Demanded);
193 bool CombineToPreIndexedLoadStore(SDNode *N);
194 bool CombineToPostIndexedLoadStore(SDNode *N);
195 SDValue SplitIndexingFromLoad(LoadSDNode *LD);
196 bool SliceUpLoad(SDNode *N);
198 /// \brief Replace an ISD::EXTRACT_VECTOR_ELT of a load with a narrowed
201 /// \param EVE ISD::EXTRACT_VECTOR_ELT to be replaced.
202 /// \param InVecVT type of the input vector to EVE with bitcasts resolved.
203 /// \param EltNo index of the vector element to load.
204 /// \param OriginalLoad load that EVE came from to be replaced.
205 /// \returns EVE on success SDValue() on failure.
206 SDValue ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
207 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad);
208 void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad);
209 SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace);
210 SDValue SExtPromoteOperand(SDValue Op, EVT PVT);
211 SDValue ZExtPromoteOperand(SDValue Op, EVT PVT);
212 SDValue PromoteIntBinOp(SDValue Op);
213 SDValue PromoteIntShiftOp(SDValue Op);
214 SDValue PromoteExtend(SDValue Op);
215 bool PromoteLoad(SDValue Op);
217 void ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs, SDValue Trunc,
218 SDValue ExtLoad, const SDLoc &DL,
219 ISD::NodeType ExtType);
221 /// Call the node-specific routine that knows how to fold each
222 /// particular type of node. If that doesn't do anything, try the
223 /// target-specific DAG combines.
224 SDValue combine(SDNode *N);
226 // Visitation implementation - Implement dag node combining for different
227 // node types. The semantics are as follows:
229 // SDValue.getNode() == 0 - No change was made
230 // SDValue.getNode() == N - N was replaced, is dead and has been handled.
231 // otherwise - N should be replaced by the returned Operand.
233 SDValue visitTokenFactor(SDNode *N);
234 SDValue visitMERGE_VALUES(SDNode *N);
235 SDValue visitADD(SDNode *N);
236 SDValue visitADDLike(SDValue N0, SDValue N1, SDNode *LocReference);
237 SDValue visitSUB(SDNode *N);
238 SDValue visitADDC(SDNode *N);
239 SDValue visitUADDO(SDNode *N);
240 SDValue visitUADDOLike(SDValue N0, SDValue N1, SDNode *N);
241 SDValue visitSUBC(SDNode *N);
242 SDValue visitUSUBO(SDNode *N);
243 SDValue visitADDE(SDNode *N);
244 SDValue visitADDCARRY(SDNode *N);
245 SDValue visitADDCARRYLike(SDValue N0, SDValue N1, SDValue CarryIn, SDNode *N);
246 SDValue visitSUBE(SDNode *N);
247 SDValue visitSUBCARRY(SDNode *N);
248 SDValue visitMUL(SDNode *N);
249 SDValue useDivRem(SDNode *N);
250 SDValue visitSDIV(SDNode *N);
251 SDValue visitUDIV(SDNode *N);
252 SDValue visitREM(SDNode *N);
253 SDValue visitMULHU(SDNode *N);
254 SDValue visitMULHS(SDNode *N);
255 SDValue visitSMUL_LOHI(SDNode *N);
256 SDValue visitUMUL_LOHI(SDNode *N);
257 SDValue visitSMULO(SDNode *N);
258 SDValue visitUMULO(SDNode *N);
259 SDValue visitIMINMAX(SDNode *N);
260 SDValue visitAND(SDNode *N);
261 SDValue visitANDLike(SDValue N0, SDValue N1, SDNode *LocReference);
262 SDValue visitOR(SDNode *N);
263 SDValue visitORLike(SDValue N0, SDValue N1, SDNode *LocReference);
264 SDValue visitXOR(SDNode *N);
265 SDValue SimplifyVBinOp(SDNode *N);
266 SDValue visitSHL(SDNode *N);
267 SDValue visitSRA(SDNode *N);
268 SDValue visitSRL(SDNode *N);
269 SDValue visitRotate(SDNode *N);
270 SDValue visitABS(SDNode *N);
271 SDValue visitBSWAP(SDNode *N);
272 SDValue visitBITREVERSE(SDNode *N);
273 SDValue visitCTLZ(SDNode *N);
274 SDValue visitCTLZ_ZERO_UNDEF(SDNode *N);
275 SDValue visitCTTZ(SDNode *N);
276 SDValue visitCTTZ_ZERO_UNDEF(SDNode *N);
277 SDValue visitCTPOP(SDNode *N);
278 SDValue visitSELECT(SDNode *N);
279 SDValue visitVSELECT(SDNode *N);
280 SDValue visitSELECT_CC(SDNode *N);
281 SDValue visitSETCC(SDNode *N);
282 SDValue visitSETCCE(SDNode *N);
283 SDValue visitSETCCCARRY(SDNode *N);
284 SDValue visitSIGN_EXTEND(SDNode *N);
285 SDValue visitZERO_EXTEND(SDNode *N);
286 SDValue visitANY_EXTEND(SDNode *N);
287 SDValue visitAssertZext(SDNode *N);
288 SDValue visitSIGN_EXTEND_INREG(SDNode *N);
289 SDValue visitSIGN_EXTEND_VECTOR_INREG(SDNode *N);
290 SDValue visitZERO_EXTEND_VECTOR_INREG(SDNode *N);
291 SDValue visitTRUNCATE(SDNode *N);
292 SDValue visitBITCAST(SDNode *N);
293 SDValue visitBUILD_PAIR(SDNode *N);
294 SDValue visitFADD(SDNode *N);
295 SDValue visitFSUB(SDNode *N);
296 SDValue visitFMUL(SDNode *N);
297 SDValue visitFMA(SDNode *N);
298 SDValue visitFDIV(SDNode *N);
299 SDValue visitFREM(SDNode *N);
300 SDValue visitFSQRT(SDNode *N);
301 SDValue visitFCOPYSIGN(SDNode *N);
302 SDValue visitSINT_TO_FP(SDNode *N);
303 SDValue visitUINT_TO_FP(SDNode *N);
304 SDValue visitFP_TO_SINT(SDNode *N);
305 SDValue visitFP_TO_UINT(SDNode *N);
306 SDValue visitFP_ROUND(SDNode *N);
307 SDValue visitFP_ROUND_INREG(SDNode *N);
308 SDValue visitFP_EXTEND(SDNode *N);
309 SDValue visitFNEG(SDNode *N);
310 SDValue visitFABS(SDNode *N);
311 SDValue visitFCEIL(SDNode *N);
312 SDValue visitFTRUNC(SDNode *N);
313 SDValue visitFFLOOR(SDNode *N);
314 SDValue visitFMINNUM(SDNode *N);
315 SDValue visitFMAXNUM(SDNode *N);
316 SDValue visitBRCOND(SDNode *N);
317 SDValue visitBR_CC(SDNode *N);
318 SDValue visitLOAD(SDNode *N);
320 SDValue replaceStoreChain(StoreSDNode *ST, SDValue BetterChain);
321 SDValue replaceStoreOfFPConstant(StoreSDNode *ST);
323 SDValue visitSTORE(SDNode *N);
324 SDValue visitINSERT_VECTOR_ELT(SDNode *N);
325 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
326 SDValue visitBUILD_VECTOR(SDNode *N);
327 SDValue visitCONCAT_VECTORS(SDNode *N);
328 SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
329 SDValue visitVECTOR_SHUFFLE(SDNode *N);
330 SDValue visitSCALAR_TO_VECTOR(SDNode *N);
331 SDValue visitINSERT_SUBVECTOR(SDNode *N);
332 SDValue visitMLOAD(SDNode *N);
333 SDValue visitMSTORE(SDNode *N);
334 SDValue visitMGATHER(SDNode *N);
335 SDValue visitMSCATTER(SDNode *N);
336 SDValue visitFP_TO_FP16(SDNode *N);
337 SDValue visitFP16_TO_FP(SDNode *N);
339 SDValue visitFADDForFMACombine(SDNode *N);
340 SDValue visitFSUBForFMACombine(SDNode *N);
341 SDValue visitFMULForFMADistributiveCombine(SDNode *N);
343 SDValue XformToShuffleWithZero(SDNode *N);
344 SDValue ReassociateOps(unsigned Opc, const SDLoc &DL, SDValue LHS,
347 SDValue visitShiftByConstant(SDNode *N, ConstantSDNode *Amt);
349 SDValue foldSelectOfConstants(SDNode *N);
350 SDValue foldBinOpIntoSelect(SDNode *BO);
351 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
352 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
353 SDValue SimplifySelect(const SDLoc &DL, SDValue N0, SDValue N1, SDValue N2);
354 SDValue SimplifySelectCC(const SDLoc &DL, SDValue N0, SDValue N1,
355 SDValue N2, SDValue N3, ISD::CondCode CC,
356 bool NotExtCompare = false);
357 SDValue foldSelectCCToShiftAnd(const SDLoc &DL, SDValue N0, SDValue N1,
358 SDValue N2, SDValue N3, ISD::CondCode CC);
359 SDValue foldLogicOfSetCCs(bool IsAnd, SDValue N0, SDValue N1,
361 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
362 const SDLoc &DL, bool foldBooleans = true);
364 bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
366 bool isOneUseSetCC(SDValue N) const;
368 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
370 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
371 SDValue CombineExtLoad(SDNode *N);
372 SDValue combineRepeatedFPDivisors(SDNode *N);
373 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
374 SDValue BuildSDIV(SDNode *N);
375 SDValue BuildSDIVPow2(SDNode *N);
376 SDValue BuildUDIV(SDNode *N);
377 SDValue BuildLogBase2(SDValue Op, const SDLoc &DL);
378 SDValue BuildReciprocalEstimate(SDValue Op, SDNodeFlags Flags);
379 SDValue buildRsqrtEstimate(SDValue Op, SDNodeFlags Flags);
380 SDValue buildSqrtEstimate(SDValue Op, SDNodeFlags Flags);
381 SDValue buildSqrtEstimateImpl(SDValue Op, SDNodeFlags Flags, bool Recip);
382 SDValue buildSqrtNROneConst(SDValue Op, SDValue Est, unsigned Iterations,
383 SDNodeFlags Flags, bool Reciprocal);
384 SDValue buildSqrtNRTwoConst(SDValue Op, SDValue Est, unsigned Iterations,
385 SDNodeFlags Flags, bool Reciprocal);
386 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
387 bool DemandHighBits = true);
388 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
389 SDNode *MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg,
390 SDValue InnerPos, SDValue InnerNeg,
391 unsigned PosOpcode, unsigned NegOpcode,
393 SDNode *MatchRotate(SDValue LHS, SDValue RHS, const SDLoc &DL);
394 SDValue MatchLoadCombine(SDNode *N);
395 SDValue ReduceLoadWidth(SDNode *N);
396 SDValue ReduceLoadOpStoreWidth(SDNode *N);
397 SDValue splitMergedValStore(StoreSDNode *ST);
398 SDValue TransformFPLoadStorePair(SDNode *N);
399 SDValue reduceBuildVecExtToExtBuildVec(SDNode *N);
400 SDValue reduceBuildVecConvertToConvertBuildVec(SDNode *N);
401 SDValue reduceBuildVecToShuffle(SDNode *N);
402 SDValue createBuildVecShuffle(const SDLoc &DL, SDNode *N,
403 ArrayRef<int> VectorMask, SDValue VecIn1,
404 SDValue VecIn2, unsigned LeftIdx);
405 SDValue matchVSelectOpSizesWithSetCC(SDNode *N);
407 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
409 /// Walk up chain skipping non-aliasing memory nodes,
410 /// looking for aliasing nodes and adding them to the Aliases vector.
411 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
412 SmallVectorImpl<SDValue> &Aliases);
414 /// Return true if there is any possibility that the two addresses overlap.
415 bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const;
417 /// Walk up chain skipping non-aliasing memory nodes, looking for a better
418 /// chain (aliasing node.)
419 SDValue FindBetterChain(SDNode *N, SDValue Chain);
421 /// Try to replace a store and any possibly adjacent stores on
422 /// consecutive chains with better chains. Return true only if St is
425 /// Notice that other chains may still be replaced even if the function
427 bool findBetterNeighborChains(StoreSDNode *St);
429 /// Match "(X shl/srl V1) & V2" where V2 may not be present.
430 bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask);
432 /// Holds a pointer to an LSBaseSDNode as well as information on where it
433 /// is located in a sequence of memory operations connected by a chain.
435 MemOpLink(LSBaseSDNode *N, int64_t Offset)
436 : MemNode(N), OffsetFromBase(Offset) {}
437 // Ptr to the mem node.
438 LSBaseSDNode *MemNode;
439 // Offset from the base ptr.
440 int64_t OffsetFromBase;
443 /// This is a helper function for visitMUL to check the profitability
444 /// of folding (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2).
445 /// MulNode is the original multiply, AddNode is (add x, c1),
446 /// and ConstNode is c2.
447 bool isMulAddWithConstProfitable(SDNode *MulNode,
452 /// This is a helper function for visitAND and visitZERO_EXTEND. Returns
453 /// true if the (and (load x) c) pattern matches an extload. ExtVT returns
454 /// the type of the loaded value to be extended. LoadedVT returns the type
455 /// of the original loaded value. NarrowLoad returns whether the load would
456 /// need to be narrowed in order to match.
457 bool isAndLoadExtLoad(ConstantSDNode *AndC, LoadSDNode *LoadN,
458 EVT LoadResultTy, EVT &ExtVT, EVT &LoadedVT,
461 /// Helper function for MergeConsecutiveStores which merges the
462 /// component store chains.
463 SDValue getMergeStoreChains(SmallVectorImpl<MemOpLink> &StoreNodes,
466 /// This is a helper function for MergeConsecutiveStores. When the source
467 /// elements of the consecutive stores are all constants or all extracted
468 /// vector elements, try to merge them into one larger store.
469 /// \return True if a merged store was created.
470 bool MergeStoresOfConstantsOrVecElts(SmallVectorImpl<MemOpLink> &StoreNodes,
471 EVT MemVT, unsigned NumStores,
472 bool IsConstantSrc, bool UseVector);
474 /// This is a helper function for MergeConsecutiveStores.
475 /// Stores that may be merged are placed in StoreNodes.
476 void getStoreMergeCandidates(StoreSDNode *St,
477 SmallVectorImpl<MemOpLink> &StoreNodes);
479 /// Helper function for MergeConsecutiveStores. Checks if
480 /// Candidate stores have indirect dependency through their
481 /// operands. \return True if safe to merge
482 bool checkMergeStoreCandidatesForDependencies(
483 SmallVectorImpl<MemOpLink> &StoreNodes, unsigned NumStores);
485 /// Merge consecutive store operations into a wide store.
486 /// This optimization uses wide integers or vectors when possible.
487 /// \return number of stores that were merged into a merged store (the
488 /// affected nodes are stored as a prefix in \p StoreNodes).
489 bool MergeConsecutiveStores(StoreSDNode *N);
491 /// \brief Try to transform a truncation where C is a constant:
492 /// (trunc (and X, C)) -> (and (trunc X), (trunc C))
494 /// \p N needs to be a truncation and its first operand an AND. Other
495 /// requirements are checked by the function (e.g. that trunc is
496 /// single-use) and if missed an empty SDValue is returned.
497 SDValue distributeTruncateThroughAnd(SDNode *N);
500 DAGCombiner(SelectionDAG &D, AliasAnalysis *AA, CodeGenOpt::Level OL)
501 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
502 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(AA) {
503 ForCodeSize = DAG.getMachineFunction().getFunction()->optForSize();
505 MaximumLegalStoreInBits = 0;
506 for (MVT VT : MVT::all_valuetypes())
507 if (EVT(VT).isSimple() && VT != MVT::Other &&
508 TLI.isTypeLegal(EVT(VT)) &&
509 VT.getSizeInBits() >= MaximumLegalStoreInBits)
510 MaximumLegalStoreInBits = VT.getSizeInBits();
513 /// Runs the dag combiner on all nodes in the work list
514 void Run(CombineLevel AtLevel);
516 SelectionDAG &getDAG() const { return DAG; }
518 /// Returns a type large enough to hold any valid shift amount - before type
519 /// legalization these can be huge.
520 EVT getShiftAmountTy(EVT LHSTy) {
521 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
522 if (LHSTy.isVector())
524 auto &DL = DAG.getDataLayout();
525 return LegalTypes ? TLI.getScalarShiftAmountTy(DL, LHSTy)
526 : TLI.getPointerTy(DL);
529 /// This method returns true if we are running before type legalization or
530 /// if the specified VT is legal.
531 bool isTypeLegal(const EVT &VT) {
532 if (!LegalTypes) return true;
533 return TLI.isTypeLegal(VT);
536 /// Convenience wrapper around TargetLowering::getSetCCResultType
537 EVT getSetCCResultType(EVT VT) const {
538 return TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
545 /// This class is a DAGUpdateListener that removes any deleted
546 /// nodes from the worklist.
547 class WorklistRemover : public SelectionDAG::DAGUpdateListener {
550 explicit WorklistRemover(DAGCombiner &dc)
551 : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {}
553 void NodeDeleted(SDNode *N, SDNode *E) override {
554 DC.removeFromWorklist(N);
559 //===----------------------------------------------------------------------===//
560 // TargetLowering::DAGCombinerInfo implementation
561 //===----------------------------------------------------------------------===//
563 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
564 ((DAGCombiner*)DC)->AddToWorklist(N);
567 SDValue TargetLowering::DAGCombinerInfo::
568 CombineTo(SDNode *N, ArrayRef<SDValue> To, bool AddTo) {
569 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
572 SDValue TargetLowering::DAGCombinerInfo::
573 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
574 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
578 SDValue TargetLowering::DAGCombinerInfo::
579 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
580 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
583 void TargetLowering::DAGCombinerInfo::
584 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
585 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
588 //===----------------------------------------------------------------------===//
590 //===----------------------------------------------------------------------===//
592 void DAGCombiner::deleteAndRecombine(SDNode *N) {
593 removeFromWorklist(N);
595 // If the operands of this node are only used by the node, they will now be
596 // dead. Make sure to re-visit them and recursively delete dead nodes.
597 for (const SDValue &Op : N->ops())
598 // For an operand generating multiple values, one of the values may
599 // become dead allowing further simplification (e.g. split index
600 // arithmetic from an indexed load).
601 if (Op->hasOneUse() || Op->getNumValues() > 1)
602 AddToWorklist(Op.getNode());
607 /// Return 1 if we can compute the negated form of the specified expression for
608 /// the same cost as the expression itself, or 2 if we can compute the negated
609 /// form more cheaply than the expression itself.
610 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
611 const TargetLowering &TLI,
612 const TargetOptions *Options,
613 unsigned Depth = 0) {
614 // fneg is removable even if it has multiple uses.
615 if (Op.getOpcode() == ISD::FNEG) return 2;
617 // Don't allow anything with multiple uses.
618 if (!Op.hasOneUse()) return 0;
620 // Don't recurse exponentially.
621 if (Depth > 6) return 0;
623 switch (Op.getOpcode()) {
624 default: return false;
625 case ISD::ConstantFP: {
626 if (!LegalOperations)
629 // Don't invert constant FP values after legalization unless the target says
630 // the negated constant is legal.
631 EVT VT = Op.getValueType();
632 return TLI.isOperationLegal(ISD::ConstantFP, VT) ||
633 TLI.isFPImmLegal(neg(cast<ConstantFPSDNode>(Op)->getValueAPF()), VT);
636 // FIXME: determine better conditions for this xform.
637 if (!Options->UnsafeFPMath) return 0;
639 // After operation legalization, it might not be legal to create new FSUBs.
640 if (LegalOperations &&
641 !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType()))
644 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
645 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
648 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
649 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
652 // We can't turn -(A-B) into B-A when we honor signed zeros.
653 if (!Options->NoSignedZerosFPMath &&
654 !Op.getNode()->getFlags().hasNoSignedZeros())
657 // fold (fneg (fsub A, B)) -> (fsub B, A)
662 if (Options->HonorSignDependentRoundingFPMath()) return 0;
664 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
665 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
669 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
675 return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options,
680 /// If isNegatibleForFree returns true, return the newly negated expression.
681 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
682 bool LegalOperations, unsigned Depth = 0) {
683 const TargetOptions &Options = DAG.getTarget().Options;
684 // fneg is removable even if it has multiple uses.
685 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
687 // Don't allow anything with multiple uses.
688 assert(Op.hasOneUse() && "Unknown reuse!");
690 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
692 const SDNodeFlags Flags = Op.getNode()->getFlags();
694 switch (Op.getOpcode()) {
695 default: llvm_unreachable("Unknown code");
696 case ISD::ConstantFP: {
697 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
699 return DAG.getConstantFP(V, SDLoc(Op), Op.getValueType());
702 // FIXME: determine better conditions for this xform.
703 assert(Options.UnsafeFPMath);
705 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
706 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
707 DAG.getTargetLoweringInfo(), &Options, Depth+1))
708 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
709 GetNegatedExpression(Op.getOperand(0), DAG,
710 LegalOperations, Depth+1),
711 Op.getOperand(1), Flags);
712 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
713 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
714 GetNegatedExpression(Op.getOperand(1), DAG,
715 LegalOperations, Depth+1),
716 Op.getOperand(0), Flags);
718 // fold (fneg (fsub 0, B)) -> B
719 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
721 return Op.getOperand(1);
723 // fold (fneg (fsub A, B)) -> (fsub B, A)
724 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
725 Op.getOperand(1), Op.getOperand(0), Flags);
729 assert(!Options.HonorSignDependentRoundingFPMath());
731 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
732 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
733 DAG.getTargetLoweringInfo(), &Options, Depth+1))
734 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
735 GetNegatedExpression(Op.getOperand(0), DAG,
736 LegalOperations, Depth+1),
737 Op.getOperand(1), Flags);
739 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
740 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
742 GetNegatedExpression(Op.getOperand(1), DAG,
743 LegalOperations, Depth+1), Flags);
747 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
748 GetNegatedExpression(Op.getOperand(0), DAG,
749 LegalOperations, Depth+1));
751 return DAG.getNode(ISD::FP_ROUND, SDLoc(Op), Op.getValueType(),
752 GetNegatedExpression(Op.getOperand(0), DAG,
753 LegalOperations, Depth+1),
758 // APInts must be the same size for most operations, this helper
759 // function zero extends the shorter of the pair so that they match.
760 // We provide an Offset so that we can create bitwidths that won't overflow.
761 static void zeroExtendToMatch(APInt &LHS, APInt &RHS, unsigned Offset = 0) {
762 unsigned Bits = Offset + std::max(LHS.getBitWidth(), RHS.getBitWidth());
763 LHS = LHS.zextOrSelf(Bits);
764 RHS = RHS.zextOrSelf(Bits);
767 // Return true if this node is a setcc, or is a select_cc
768 // that selects between the target values used for true and false, making it
769 // equivalent to a setcc. Also, set the incoming LHS, RHS, and CC references to
770 // the appropriate nodes based on the type of node we are checking. This
771 // simplifies life a bit for the callers.
772 bool DAGCombiner::isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
774 if (N.getOpcode() == ISD::SETCC) {
775 LHS = N.getOperand(0);
776 RHS = N.getOperand(1);
777 CC = N.getOperand(2);
781 if (N.getOpcode() != ISD::SELECT_CC ||
782 !TLI.isConstTrueVal(N.getOperand(2).getNode()) ||
783 !TLI.isConstFalseVal(N.getOperand(3).getNode()))
786 if (TLI.getBooleanContents(N.getValueType()) ==
787 TargetLowering::UndefinedBooleanContent)
790 LHS = N.getOperand(0);
791 RHS = N.getOperand(1);
792 CC = N.getOperand(4);
796 /// Return true if this is a SetCC-equivalent operation with only one use.
797 /// If this is true, it allows the users to invert the operation for free when
798 /// it is profitable to do so.
799 bool DAGCombiner::isOneUseSetCC(SDValue N) const {
801 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
806 // \brief Returns the SDNode if it is a constant float BuildVector
807 // or constant float.
808 static SDNode *isConstantFPBuildVectorOrConstantFP(SDValue N) {
809 if (isa<ConstantFPSDNode>(N))
811 if (ISD::isBuildVectorOfConstantFPSDNodes(N.getNode()))
816 // Determines if it is a constant integer or a build vector of constant
817 // integers (and undefs).
818 // Do not permit build vector implicit truncation.
819 static bool isConstantOrConstantVector(SDValue N, bool NoOpaques = false) {
820 if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N))
821 return !(Const->isOpaque() && NoOpaques);
822 if (N.getOpcode() != ISD::BUILD_VECTOR)
824 unsigned BitWidth = N.getScalarValueSizeInBits();
825 for (const SDValue &Op : N->op_values()) {
828 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(Op);
829 if (!Const || Const->getAPIntValue().getBitWidth() != BitWidth ||
830 (Const->isOpaque() && NoOpaques))
836 // Determines if it is a constant null integer or a splatted vector of a
837 // constant null integer (with no undefs).
838 // Build vector implicit truncation is not an issue for null values.
839 static bool isNullConstantOrNullSplatConstant(SDValue N) {
840 if (ConstantSDNode *Splat = isConstOrConstSplat(N))
841 return Splat->isNullValue();
845 // Determines if it is a constant integer of one or a splatted vector of a
846 // constant integer of one (with no undefs).
847 // Do not permit build vector implicit truncation.
848 static bool isOneConstantOrOneSplatConstant(SDValue N) {
849 unsigned BitWidth = N.getScalarValueSizeInBits();
850 if (ConstantSDNode *Splat = isConstOrConstSplat(N))
851 return Splat->isOne() && Splat->getAPIntValue().getBitWidth() == BitWidth;
855 // Determines if it is a constant integer of all ones or a splatted vector of a
856 // constant integer of all ones (with no undefs).
857 // Do not permit build vector implicit truncation.
858 static bool isAllOnesConstantOrAllOnesSplatConstant(SDValue N) {
859 unsigned BitWidth = N.getScalarValueSizeInBits();
860 if (ConstantSDNode *Splat = isConstOrConstSplat(N))
861 return Splat->isAllOnesValue() &&
862 Splat->getAPIntValue().getBitWidth() == BitWidth;
866 // Determines if a BUILD_VECTOR is composed of all-constants possibly mixed with
868 static bool isAnyConstantBuildVector(const SDNode *N) {
869 return ISD::isBuildVectorOfConstantSDNodes(N) ||
870 ISD::isBuildVectorOfConstantFPSDNodes(N);
873 SDValue DAGCombiner::ReassociateOps(unsigned Opc, const SDLoc &DL, SDValue N0,
875 EVT VT = N0.getValueType();
876 if (N0.getOpcode() == Opc) {
877 if (SDNode *L = DAG.isConstantIntBuildVectorOrConstantInt(N0.getOperand(1))) {
878 if (SDNode *R = DAG.isConstantIntBuildVectorOrConstantInt(N1)) {
879 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
880 if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, DL, VT, L, R))
881 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
884 if (N0.hasOneUse()) {
885 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one
887 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1);
888 if (!OpNode.getNode())
890 AddToWorklist(OpNode.getNode());
891 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
896 if (N1.getOpcode() == Opc) {
897 if (SDNode *R = DAG.isConstantIntBuildVectorOrConstantInt(N1.getOperand(1))) {
898 if (SDNode *L = DAG.isConstantIntBuildVectorOrConstantInt(N0)) {
899 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
900 if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, DL, VT, R, L))
901 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
904 if (N1.hasOneUse()) {
905 // reassoc. (op x, (op y, c1)) -> (op (op x, y), c1) iff x+c1 has one
907 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0, N1.getOperand(0));
908 if (!OpNode.getNode())
910 AddToWorklist(OpNode.getNode());
911 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
919 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
921 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
923 DEBUG(dbgs() << "\nReplacing.1 ";
925 dbgs() << "\nWith: ";
926 To[0].getNode()->dump(&DAG);
927 dbgs() << " and " << NumTo-1 << " other values\n");
928 for (unsigned i = 0, e = NumTo; i != e; ++i)
929 assert((!To[i].getNode() ||
930 N->getValueType(i) == To[i].getValueType()) &&
931 "Cannot combine value to value of different type!");
933 WorklistRemover DeadNodes(*this);
934 DAG.ReplaceAllUsesWith(N, To);
936 // Push the new nodes and any users onto the worklist
937 for (unsigned i = 0, e = NumTo; i != e; ++i) {
938 if (To[i].getNode()) {
939 AddToWorklist(To[i].getNode());
940 AddUsersToWorklist(To[i].getNode());
945 // Finally, if the node is now dead, remove it from the graph. The node
946 // may not be dead if the replacement process recursively simplified to
947 // something else needing this node.
949 deleteAndRecombine(N);
950 return SDValue(N, 0);
954 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
955 // Replace all uses. If any nodes become isomorphic to other nodes and
956 // are deleted, make sure to remove them from our worklist.
957 WorklistRemover DeadNodes(*this);
958 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New);
960 // Push the new node and any (possibly new) users onto the worklist.
961 AddToWorklist(TLO.New.getNode());
962 AddUsersToWorklist(TLO.New.getNode());
964 // Finally, if the node is now dead, remove it from the graph. The node
965 // may not be dead if the replacement process recursively simplified to
966 // something else needing this node.
967 if (TLO.Old.getNode()->use_empty())
968 deleteAndRecombine(TLO.Old.getNode());
971 /// Check the specified integer node value to see if it can be simplified or if
972 /// things it uses can be simplified by bit propagation. If so, return true.
973 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
974 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
976 if (!TLI.SimplifyDemandedBits(Op, Demanded, Known, TLO))
980 AddToWorklist(Op.getNode());
982 // Replace the old value with the new one.
984 DEBUG(dbgs() << "\nReplacing.2 ";
985 TLO.Old.getNode()->dump(&DAG);
986 dbgs() << "\nWith: ";
987 TLO.New.getNode()->dump(&DAG);
990 CommitTargetLoweringOpt(TLO);
994 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
996 EVT VT = Load->getValueType(0);
997 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, VT, SDValue(ExtLoad, 0));
999 DEBUG(dbgs() << "\nReplacing.9 ";
1001 dbgs() << "\nWith: ";
1002 Trunc.getNode()->dump(&DAG);
1004 WorklistRemover DeadNodes(*this);
1005 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc);
1006 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1));
1007 deleteAndRecombine(Load);
1008 AddToWorklist(Trunc.getNode());
1011 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
1014 if (ISD::isUNINDEXEDLoad(Op.getNode())) {
1015 LoadSDNode *LD = cast<LoadSDNode>(Op);
1016 EVT MemVT = LD->getMemoryVT();
1017 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
1018 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, PVT, MemVT) ? ISD::ZEXTLOAD
1020 : LD->getExtensionType();
1022 return DAG.getExtLoad(ExtType, DL, PVT,
1023 LD->getChain(), LD->getBasePtr(),
1024 MemVT, LD->getMemOperand());
1027 unsigned Opc = Op.getOpcode();
1030 case ISD::AssertSext:
1031 return DAG.getNode(ISD::AssertSext, DL, PVT,
1032 SExtPromoteOperand(Op.getOperand(0), PVT),
1034 case ISD::AssertZext:
1035 return DAG.getNode(ISD::AssertZext, DL, PVT,
1036 ZExtPromoteOperand(Op.getOperand(0), PVT),
1038 case ISD::Constant: {
1040 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1041 return DAG.getNode(ExtOpc, DL, PVT, Op);
1045 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
1047 return DAG.getNode(ISD::ANY_EXTEND, DL, PVT, Op);
1050 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
1051 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
1053 EVT OldVT = Op.getValueType();
1055 bool Replace = false;
1056 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
1057 if (!NewOp.getNode())
1059 AddToWorklist(NewOp.getNode());
1062 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
1063 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, NewOp.getValueType(), NewOp,
1064 DAG.getValueType(OldVT));
1067 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
1068 EVT OldVT = Op.getValueType();
1070 bool Replace = false;
1071 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
1072 if (!NewOp.getNode())
1074 AddToWorklist(NewOp.getNode());
1077 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
1078 return DAG.getZeroExtendInReg(NewOp, DL, OldVT);
1081 /// Promote the specified integer binary operation if the target indicates it is
1082 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
1083 /// i32 since i16 instructions are longer.
1084 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
1085 if (!LegalOperations)
1088 EVT VT = Op.getValueType();
1089 if (VT.isVector() || !VT.isInteger())
1092 // If operation type is 'undesirable', e.g. i16 on x86, consider
1094 unsigned Opc = Op.getOpcode();
1095 if (TLI.isTypeDesirableForOp(Opc, VT))
1099 // Consult target whether it is a good idea to promote this operation and
1100 // what's the right type to promote it to.
1101 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1102 assert(PVT != VT && "Don't know what type to promote to!");
1104 DEBUG(dbgs() << "\nPromoting "; Op.getNode()->dump(&DAG));
1106 bool Replace0 = false;
1107 SDValue N0 = Op.getOperand(0);
1108 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
1110 bool Replace1 = false;
1111 SDValue N1 = Op.getOperand(1);
1112 SDValue NN1 = PromoteOperand(N1, PVT, Replace1);
1116 DAG.getNode(ISD::TRUNCATE, DL, VT, DAG.getNode(Opc, DL, PVT, NN0, NN1));
1118 // New replace instances of N0 and N1
1119 if (Replace0 && N0 && N0.getOpcode() != ISD::DELETED_NODE && NN0 &&
1120 NN0.getOpcode() != ISD::DELETED_NODE) {
1121 AddToWorklist(NN0.getNode());
1122 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
1125 if (Replace1 && N1 && N1.getOpcode() != ISD::DELETED_NODE && NN1 &&
1126 NN1.getOpcode() != ISD::DELETED_NODE) {
1127 AddToWorklist(NN1.getNode());
1128 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
1131 // Deal with Op being deleted.
1132 if (Op && Op.getOpcode() != ISD::DELETED_NODE)
1138 /// Promote the specified integer shift operation if the target indicates it is
1139 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
1140 /// i32 since i16 instructions are longer.
1141 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
1142 if (!LegalOperations)
1145 EVT VT = Op.getValueType();
1146 if (VT.isVector() || !VT.isInteger())
1149 // If operation type is 'undesirable', e.g. i16 on x86, consider
1151 unsigned Opc = Op.getOpcode();
1152 if (TLI.isTypeDesirableForOp(Opc, VT))
1156 // Consult target whether it is a good idea to promote this operation and
1157 // what's the right type to promote it to.
1158 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1159 assert(PVT != VT && "Don't know what type to promote to!");
1161 DEBUG(dbgs() << "\nPromoting "; Op.getNode()->dump(&DAG));
1163 bool Replace = false;
1164 SDValue N0 = Op.getOperand(0);
1165 SDValue N1 = Op.getOperand(1);
1166 if (Opc == ISD::SRA)
1167 N0 = SExtPromoteOperand(N0, PVT);
1168 else if (Opc == ISD::SRL)
1169 N0 = ZExtPromoteOperand(N0, PVT);
1171 N0 = PromoteOperand(N0, PVT, Replace);
1178 DAG.getNode(ISD::TRUNCATE, DL, VT, DAG.getNode(Opc, DL, PVT, N0, N1));
1180 AddToWorklist(N0.getNode());
1182 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
1184 // Deal with Op being deleted.
1185 if (Op && Op.getOpcode() != ISD::DELETED_NODE)
1191 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
1192 if (!LegalOperations)
1195 EVT VT = Op.getValueType();
1196 if (VT.isVector() || !VT.isInteger())
1199 // If operation type is 'undesirable', e.g. i16 on x86, consider
1201 unsigned Opc = Op.getOpcode();
1202 if (TLI.isTypeDesirableForOp(Opc, VT))
1206 // Consult target whether it is a good idea to promote this operation and
1207 // what's the right type to promote it to.
1208 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1209 assert(PVT != VT && "Don't know what type to promote to!");
1210 // fold (aext (aext x)) -> (aext x)
1211 // fold (aext (zext x)) -> (zext x)
1212 // fold (aext (sext x)) -> (sext x)
1213 DEBUG(dbgs() << "\nPromoting ";
1214 Op.getNode()->dump(&DAG));
1215 return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0));
1220 bool DAGCombiner::PromoteLoad(SDValue Op) {
1221 if (!LegalOperations)
1224 if (!ISD::isUNINDEXEDLoad(Op.getNode()))
1227 EVT VT = Op.getValueType();
1228 if (VT.isVector() || !VT.isInteger())
1231 // If operation type is 'undesirable', e.g. i16 on x86, consider
1233 unsigned Opc = Op.getOpcode();
1234 if (TLI.isTypeDesirableForOp(Opc, VT))
1238 // Consult target whether it is a good idea to promote this operation and
1239 // what's the right type to promote it to.
1240 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1241 assert(PVT != VT && "Don't know what type to promote to!");
1244 SDNode *N = Op.getNode();
1245 LoadSDNode *LD = cast<LoadSDNode>(N);
1246 EVT MemVT = LD->getMemoryVT();
1247 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
1248 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, PVT, MemVT) ? ISD::ZEXTLOAD
1250 : LD->getExtensionType();
1251 SDValue NewLD = DAG.getExtLoad(ExtType, DL, PVT,
1252 LD->getChain(), LD->getBasePtr(),
1253 MemVT, LD->getMemOperand());
1254 SDValue Result = DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD);
1256 DEBUG(dbgs() << "\nPromoting ";
1259 Result.getNode()->dump(&DAG);
1261 WorklistRemover DeadNodes(*this);
1262 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
1263 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1));
1264 deleteAndRecombine(N);
1265 AddToWorklist(Result.getNode());
1271 /// \brief Recursively delete a node which has no uses and any operands for
1272 /// which it is the only use.
1274 /// Note that this both deletes the nodes and removes them from the worklist.
1275 /// It also adds any nodes who have had a user deleted to the worklist as they
1276 /// may now have only one use and subject to other combines.
1277 bool DAGCombiner::recursivelyDeleteUnusedNodes(SDNode *N) {
1278 if (!N->use_empty())
1281 SmallSetVector<SDNode *, 16> Nodes;
1284 N = Nodes.pop_back_val();
1288 if (N->use_empty()) {
1289 for (const SDValue &ChildN : N->op_values())
1290 Nodes.insert(ChildN.getNode());
1292 removeFromWorklist(N);
1297 } while (!Nodes.empty());
1301 //===----------------------------------------------------------------------===//
1302 // Main DAG Combiner implementation
1303 //===----------------------------------------------------------------------===//
1305 void DAGCombiner::Run(CombineLevel AtLevel) {
1306 // set the instance variables, so that the various visit routines may use it.
1308 LegalOperations = Level >= AfterLegalizeVectorOps;
1309 LegalTypes = Level >= AfterLegalizeTypes;
1311 // Add all the dag nodes to the worklist.
1312 for (SDNode &Node : DAG.allnodes())
1313 AddToWorklist(&Node);
1315 // Create a dummy node (which is not added to allnodes), that adds a reference
1316 // to the root node, preventing it from being deleted, and tracking any
1317 // changes of the root.
1318 HandleSDNode Dummy(DAG.getRoot());
1320 // While the worklist isn't empty, find a node and try to combine it.
1321 while (!WorklistMap.empty()) {
1323 // The Worklist holds the SDNodes in order, but it may contain null entries.
1325 N = Worklist.pop_back_val();
1328 bool GoodWorklistEntry = WorklistMap.erase(N);
1329 (void)GoodWorklistEntry;
1330 assert(GoodWorklistEntry &&
1331 "Found a worklist entry without a corresponding map entry!");
1333 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1334 // N is deleted from the DAG, since they too may now be dead or may have a
1335 // reduced number of uses, allowing other xforms.
1336 if (recursivelyDeleteUnusedNodes(N))
1339 WorklistRemover DeadNodes(*this);
1341 // If this combine is running after legalizing the DAG, re-legalize any
1342 // nodes pulled off the worklist.
1343 if (Level == AfterLegalizeDAG) {
1344 SmallSetVector<SDNode *, 16> UpdatedNodes;
1345 bool NIsValid = DAG.LegalizeOp(N, UpdatedNodes);
1347 for (SDNode *LN : UpdatedNodes) {
1349 AddUsersToWorklist(LN);
1355 DEBUG(dbgs() << "\nCombining: "; N->dump(&DAG));
1357 // Add any operands of the new node which have not yet been combined to the
1358 // worklist as well. Because the worklist uniques things already, this
1359 // won't repeatedly process the same operand.
1360 CombinedNodes.insert(N);
1361 for (const SDValue &ChildN : N->op_values())
1362 if (!CombinedNodes.count(ChildN.getNode()))
1363 AddToWorklist(ChildN.getNode());
1365 SDValue RV = combine(N);
1372 // If we get back the same node we passed in, rather than a new node or
1373 // zero, we know that the node must have defined multiple values and
1374 // CombineTo was used. Since CombineTo takes care of the worklist
1375 // mechanics for us, we have no work to do in this case.
1376 if (RV.getNode() == N)
1379 assert(N->getOpcode() != ISD::DELETED_NODE &&
1380 RV.getOpcode() != ISD::DELETED_NODE &&
1381 "Node was deleted but visit returned new node!");
1383 DEBUG(dbgs() << " ... into: ";
1384 RV.getNode()->dump(&DAG));
1386 if (N->getNumValues() == RV.getNode()->getNumValues())
1387 DAG.ReplaceAllUsesWith(N, RV.getNode());
1389 assert(N->getValueType(0) == RV.getValueType() &&
1390 N->getNumValues() == 1 && "Type mismatch");
1391 DAG.ReplaceAllUsesWith(N, &RV);
1394 // Push the new node and any users onto the worklist
1395 AddToWorklist(RV.getNode());
1396 AddUsersToWorklist(RV.getNode());
1398 // Finally, if the node is now dead, remove it from the graph. The node
1399 // may not be dead if the replacement process recursively simplified to
1400 // something else needing this node. This will also take care of adding any
1401 // operands which have lost a user to the worklist.
1402 recursivelyDeleteUnusedNodes(N);
1405 // If the root changed (e.g. it was a dead load, update the root).
1406 DAG.setRoot(Dummy.getValue());
1407 DAG.RemoveDeadNodes();
1410 SDValue DAGCombiner::visit(SDNode *N) {
1411 switch (N->getOpcode()) {
1413 case ISD::TokenFactor: return visitTokenFactor(N);
1414 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1415 case ISD::ADD: return visitADD(N);
1416 case ISD::SUB: return visitSUB(N);
1417 case ISD::ADDC: return visitADDC(N);
1418 case ISD::UADDO: return visitUADDO(N);
1419 case ISD::SUBC: return visitSUBC(N);
1420 case ISD::USUBO: return visitUSUBO(N);
1421 case ISD::ADDE: return visitADDE(N);
1422 case ISD::ADDCARRY: return visitADDCARRY(N);
1423 case ISD::SUBE: return visitSUBE(N);
1424 case ISD::SUBCARRY: return visitSUBCARRY(N);
1425 case ISD::MUL: return visitMUL(N);
1426 case ISD::SDIV: return visitSDIV(N);
1427 case ISD::UDIV: return visitUDIV(N);
1429 case ISD::UREM: return visitREM(N);
1430 case ISD::MULHU: return visitMULHU(N);
1431 case ISD::MULHS: return visitMULHS(N);
1432 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1433 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1434 case ISD::SMULO: return visitSMULO(N);
1435 case ISD::UMULO: return visitUMULO(N);
1439 case ISD::UMAX: return visitIMINMAX(N);
1440 case ISD::AND: return visitAND(N);
1441 case ISD::OR: return visitOR(N);
1442 case ISD::XOR: return visitXOR(N);
1443 case ISD::SHL: return visitSHL(N);
1444 case ISD::SRA: return visitSRA(N);
1445 case ISD::SRL: return visitSRL(N);
1447 case ISD::ROTL: return visitRotate(N);
1448 case ISD::ABS: return visitABS(N);
1449 case ISD::BSWAP: return visitBSWAP(N);
1450 case ISD::BITREVERSE: return visitBITREVERSE(N);
1451 case ISD::CTLZ: return visitCTLZ(N);
1452 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1453 case ISD::CTTZ: return visitCTTZ(N);
1454 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1455 case ISD::CTPOP: return visitCTPOP(N);
1456 case ISD::SELECT: return visitSELECT(N);
1457 case ISD::VSELECT: return visitVSELECT(N);
1458 case ISD::SELECT_CC: return visitSELECT_CC(N);
1459 case ISD::SETCC: return visitSETCC(N);
1460 case ISD::SETCCE: return visitSETCCE(N);
1461 case ISD::SETCCCARRY: return visitSETCCCARRY(N);
1462 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1463 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1464 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1465 case ISD::AssertZext: return visitAssertZext(N);
1466 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1467 case ISD::SIGN_EXTEND_VECTOR_INREG: return visitSIGN_EXTEND_VECTOR_INREG(N);
1468 case ISD::ZERO_EXTEND_VECTOR_INREG: return visitZERO_EXTEND_VECTOR_INREG(N);
1469 case ISD::TRUNCATE: return visitTRUNCATE(N);
1470 case ISD::BITCAST: return visitBITCAST(N);
1471 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1472 case ISD::FADD: return visitFADD(N);
1473 case ISD::FSUB: return visitFSUB(N);
1474 case ISD::FMUL: return visitFMUL(N);
1475 case ISD::FMA: return visitFMA(N);
1476 case ISD::FDIV: return visitFDIV(N);
1477 case ISD::FREM: return visitFREM(N);
1478 case ISD::FSQRT: return visitFSQRT(N);
1479 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1480 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1481 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1482 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1483 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1484 case ISD::FP_ROUND: return visitFP_ROUND(N);
1485 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1486 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1487 case ISD::FNEG: return visitFNEG(N);
1488 case ISD::FABS: return visitFABS(N);
1489 case ISD::FFLOOR: return visitFFLOOR(N);
1490 case ISD::FMINNUM: return visitFMINNUM(N);
1491 case ISD::FMAXNUM: return visitFMAXNUM(N);
1492 case ISD::FCEIL: return visitFCEIL(N);
1493 case ISD::FTRUNC: return visitFTRUNC(N);
1494 case ISD::BRCOND: return visitBRCOND(N);
1495 case ISD::BR_CC: return visitBR_CC(N);
1496 case ISD::LOAD: return visitLOAD(N);
1497 case ISD::STORE: return visitSTORE(N);
1498 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1499 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1500 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1501 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1502 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1503 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1504 case ISD::SCALAR_TO_VECTOR: return visitSCALAR_TO_VECTOR(N);
1505 case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N);
1506 case ISD::MGATHER: return visitMGATHER(N);
1507 case ISD::MLOAD: return visitMLOAD(N);
1508 case ISD::MSCATTER: return visitMSCATTER(N);
1509 case ISD::MSTORE: return visitMSTORE(N);
1510 case ISD::FP_TO_FP16: return visitFP_TO_FP16(N);
1511 case ISD::FP16_TO_FP: return visitFP16_TO_FP(N);
1516 SDValue DAGCombiner::combine(SDNode *N) {
1517 SDValue RV = visit(N);
1519 // If nothing happened, try a target-specific DAG combine.
1520 if (!RV.getNode()) {
1521 assert(N->getOpcode() != ISD::DELETED_NODE &&
1522 "Node was deleted but visit returned NULL!");
1524 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1525 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1527 // Expose the DAG combiner to the target combiner impls.
1528 TargetLowering::DAGCombinerInfo
1529 DagCombineInfo(DAG, Level, false, this);
1531 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1535 // If nothing happened still, try promoting the operation.
1536 if (!RV.getNode()) {
1537 switch (N->getOpcode()) {
1545 RV = PromoteIntBinOp(SDValue(N, 0));
1550 RV = PromoteIntShiftOp(SDValue(N, 0));
1552 case ISD::SIGN_EXTEND:
1553 case ISD::ZERO_EXTEND:
1554 case ISD::ANY_EXTEND:
1555 RV = PromoteExtend(SDValue(N, 0));
1558 if (PromoteLoad(SDValue(N, 0)))
1564 // If N is a commutative binary node, try commuting it to enable more
1566 if (!RV.getNode() && SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1567 N->getNumValues() == 1) {
1568 SDValue N0 = N->getOperand(0);
1569 SDValue N1 = N->getOperand(1);
1571 // Constant operands are canonicalized to RHS.
1572 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1573 SDValue Ops[] = {N1, N0};
1574 SDNode *CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops,
1577 return SDValue(CSENode, 0);
1584 /// Given a node, return its input chain if it has one, otherwise return a null
1586 static SDValue getInputChainForNode(SDNode *N) {
1587 if (unsigned NumOps = N->getNumOperands()) {
1588 if (N->getOperand(0).getValueType() == MVT::Other)
1589 return N->getOperand(0);
1590 if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1591 return N->getOperand(NumOps-1);
1592 for (unsigned i = 1; i < NumOps-1; ++i)
1593 if (N->getOperand(i).getValueType() == MVT::Other)
1594 return N->getOperand(i);
1599 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1600 // If N has two operands, where one has an input chain equal to the other,
1601 // the 'other' chain is redundant.
1602 if (N->getNumOperands() == 2) {
1603 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1604 return N->getOperand(0);
1605 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1606 return N->getOperand(1);
1609 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1610 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1611 SmallPtrSet<SDNode*, 16> SeenOps;
1612 bool Changed = false; // If we should replace this token factor.
1614 // Start out with this token factor.
1617 // Iterate through token factors. The TFs grows when new token factors are
1619 for (unsigned i = 0; i < TFs.size(); ++i) {
1620 SDNode *TF = TFs[i];
1622 // Check each of the operands.
1623 for (const SDValue &Op : TF->op_values()) {
1625 switch (Op.getOpcode()) {
1626 case ISD::EntryToken:
1627 // Entry tokens don't need to be added to the list. They are
1632 case ISD::TokenFactor:
1633 if (Op.hasOneUse() && !is_contained(TFs, Op.getNode())) {
1634 // Queue up for processing.
1635 TFs.push_back(Op.getNode());
1636 // Clean up in case the token factor is removed.
1637 AddToWorklist(Op.getNode());
1644 // Only add if it isn't already in the list.
1645 if (SeenOps.insert(Op.getNode()).second)
1654 // Remove Nodes that are chained to another node in the list. Do so
1655 // by walking up chains breath-first stopping when we've seen
1656 // another operand. In general we must climb to the EntryNode, but we can exit
1657 // early if we find all remaining work is associated with just one operand as
1658 // no further pruning is possible.
1660 // List of nodes to search through and original Ops from which they originate.
1661 SmallVector<std::pair<SDNode *, unsigned>, 8> Worklist;
1662 SmallVector<unsigned, 8> OpWorkCount; // Count of work for each Op.
1663 SmallPtrSet<SDNode *, 16> SeenChains;
1664 bool DidPruneOps = false;
1666 unsigned NumLeftToConsider = 0;
1667 for (const SDValue &Op : Ops) {
1668 Worklist.push_back(std::make_pair(Op.getNode(), NumLeftToConsider++));
1669 OpWorkCount.push_back(1);
1672 auto AddToWorklist = [&](unsigned CurIdx, SDNode *Op, unsigned OpNumber) {
1673 // If this is an Op, we can remove the op from the list. Remark any
1674 // search associated with it as from the current OpNumber.
1675 if (SeenOps.count(Op) != 0) {
1678 unsigned OrigOpNumber = 0;
1679 while (OrigOpNumber < Ops.size() && Ops[OrigOpNumber].getNode() != Op)
1681 assert((OrigOpNumber != Ops.size()) &&
1682 "expected to find TokenFactor Operand");
1683 // Re-mark worklist from OrigOpNumber to OpNumber
1684 for (unsigned i = CurIdx + 1; i < Worklist.size(); ++i) {
1685 if (Worklist[i].second == OrigOpNumber) {
1686 Worklist[i].second = OpNumber;
1689 OpWorkCount[OpNumber] += OpWorkCount[OrigOpNumber];
1690 OpWorkCount[OrigOpNumber] = 0;
1691 NumLeftToConsider--;
1693 // Add if it's a new chain
1694 if (SeenChains.insert(Op).second) {
1695 OpWorkCount[OpNumber]++;
1696 Worklist.push_back(std::make_pair(Op, OpNumber));
1700 for (unsigned i = 0; i < Worklist.size() && i < 1024; ++i) {
1701 // We need at least be consider at least 2 Ops to prune.
1702 if (NumLeftToConsider <= 1)
1704 auto CurNode = Worklist[i].first;
1705 auto CurOpNumber = Worklist[i].second;
1706 assert((OpWorkCount[CurOpNumber] > 0) &&
1707 "Node should not appear in worklist");
1708 switch (CurNode->getOpcode()) {
1709 case ISD::EntryToken:
1710 // Hitting EntryToken is the only way for the search to terminate without
1712 // another operand's search. Prevent us from marking this operand
1714 NumLeftToConsider++;
1716 case ISD::TokenFactor:
1717 for (const SDValue &Op : CurNode->op_values())
1718 AddToWorklist(i, Op.getNode(), CurOpNumber);
1720 case ISD::CopyFromReg:
1721 case ISD::CopyToReg:
1722 AddToWorklist(i, CurNode->getOperand(0).getNode(), CurOpNumber);
1725 if (auto *MemNode = dyn_cast<MemSDNode>(CurNode))
1726 AddToWorklist(i, MemNode->getChain().getNode(), CurOpNumber);
1729 OpWorkCount[CurOpNumber]--;
1730 if (OpWorkCount[CurOpNumber] == 0)
1731 NumLeftToConsider--;
1734 // If we've changed things around then replace token factor.
1738 // The entry token is the only possible outcome.
1739 Result = DAG.getEntryNode();
1742 SmallVector<SDValue, 8> PrunedOps;
1744 for (const SDValue &Op : Ops) {
1745 if (SeenChains.count(Op.getNode()) == 0)
1746 PrunedOps.push_back(Op);
1748 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, PrunedOps);
1750 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Ops);
1758 /// MERGE_VALUES can always be eliminated.
1759 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1760 WorklistRemover DeadNodes(*this);
1761 // Replacing results may cause a different MERGE_VALUES to suddenly
1762 // be CSE'd with N, and carry its uses with it. Iterate until no
1763 // uses remain, to ensure that the node can be safely deleted.
1764 // First add the users of this node to the work list so that they
1765 // can be tried again once they have new operands.
1766 AddUsersToWorklist(N);
1768 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1769 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1770 } while (!N->use_empty());
1771 deleteAndRecombine(N);
1772 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1775 /// If \p N is a ConstantSDNode with isOpaque() == false return it casted to a
1776 /// ConstantSDNode pointer else nullptr.
1777 static ConstantSDNode *getAsNonOpaqueConstant(SDValue N) {
1778 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N);
1779 return Const != nullptr && !Const->isOpaque() ? Const : nullptr;
1782 SDValue DAGCombiner::foldBinOpIntoSelect(SDNode *BO) {
1783 auto BinOpcode = BO->getOpcode();
1784 assert((BinOpcode == ISD::ADD || BinOpcode == ISD::SUB ||
1785 BinOpcode == ISD::MUL || BinOpcode == ISD::SDIV ||
1786 BinOpcode == ISD::UDIV || BinOpcode == ISD::SREM ||
1787 BinOpcode == ISD::UREM || BinOpcode == ISD::AND ||
1788 BinOpcode == ISD::OR || BinOpcode == ISD::XOR ||
1789 BinOpcode == ISD::SHL || BinOpcode == ISD::SRL ||
1790 BinOpcode == ISD::SRA || BinOpcode == ISD::FADD ||
1791 BinOpcode == ISD::FSUB || BinOpcode == ISD::FMUL ||
1792 BinOpcode == ISD::FDIV || BinOpcode == ISD::FREM) &&
1793 "Unexpected binary operator");
1795 // Bail out if any constants are opaque because we can't constant fold those.
1796 SDValue C1 = BO->getOperand(1);
1797 if (!isConstantOrConstantVector(C1, true) &&
1798 !isConstantFPBuildVectorOrConstantFP(C1))
1801 // Don't do this unless the old select is going away. We want to eliminate the
1802 // binary operator, not replace a binop with a select.
1803 // TODO: Handle ISD::SELECT_CC.
1804 SDValue Sel = BO->getOperand(0);
1805 if (Sel.getOpcode() != ISD::SELECT || !Sel.hasOneUse())
1808 SDValue CT = Sel.getOperand(1);
1809 if (!isConstantOrConstantVector(CT, true) &&
1810 !isConstantFPBuildVectorOrConstantFP(CT))
1813 SDValue CF = Sel.getOperand(2);
1814 if (!isConstantOrConstantVector(CF, true) &&
1815 !isConstantFPBuildVectorOrConstantFP(CF))
1818 // We have a select-of-constants followed by a binary operator with a
1819 // constant. Eliminate the binop by pulling the constant math into the select.
1820 // Example: add (select Cond, CT, CF), C1 --> select Cond, CT + C1, CF + C1
1821 EVT VT = Sel.getValueType();
1823 SDValue NewCT = DAG.getNode(BinOpcode, DL, VT, CT, C1);
1824 assert((NewCT.isUndef() || isConstantOrConstantVector(NewCT) ||
1825 isConstantFPBuildVectorOrConstantFP(NewCT)) &&
1826 "Failed to constant fold a binop with constant operands");
1828 SDValue NewCF = DAG.getNode(BinOpcode, DL, VT, CF, C1);
1829 assert((NewCF.isUndef() || isConstantOrConstantVector(NewCF) ||
1830 isConstantFPBuildVectorOrConstantFP(NewCF)) &&
1831 "Failed to constant fold a binop with constant operands");
1833 return DAG.getSelect(DL, VT, Sel.getOperand(0), NewCT, NewCF);
1836 SDValue DAGCombiner::visitADD(SDNode *N) {
1837 SDValue N0 = N->getOperand(0);
1838 SDValue N1 = N->getOperand(1);
1839 EVT VT = N0.getValueType();
1843 if (VT.isVector()) {
1844 if (SDValue FoldedVOp = SimplifyVBinOp(N))
1847 // fold (add x, 0) -> x, vector edition
1848 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1850 if (ISD::isBuildVectorAllZeros(N0.getNode()))
1854 // fold (add x, undef) -> undef
1861 if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) {
1862 // canonicalize constant to RHS
1863 if (!DAG.isConstantIntBuildVectorOrConstantInt(N1))
1864 return DAG.getNode(ISD::ADD, DL, VT, N1, N0);
1865 // fold (add c1, c2) -> c1+c2
1866 return DAG.FoldConstantArithmetic(ISD::ADD, DL, VT, N0.getNode(),
1870 // fold (add x, 0) -> x
1871 if (isNullConstant(N1))
1874 if (isConstantOrConstantVector(N1, /* NoOpaque */ true)) {
1875 // fold ((c1-A)+c2) -> (c1+c2)-A
1876 if (N0.getOpcode() == ISD::SUB &&
1877 isConstantOrConstantVector(N0.getOperand(0), /* NoOpaque */ true)) {
1878 // FIXME: Adding 2 constants should be handled by FoldConstantArithmetic.
1879 return DAG.getNode(ISD::SUB, DL, VT,
1880 DAG.getNode(ISD::ADD, DL, VT, N1, N0.getOperand(0)),
1884 // add (sext i1 X), 1 -> zext (not i1 X)
1885 // We don't transform this pattern:
1886 // add (zext i1 X), -1 -> sext (not i1 X)
1887 // because most (?) targets generate better code for the zext form.
1888 if (N0.getOpcode() == ISD::SIGN_EXTEND && N0.hasOneUse() &&
1889 isOneConstantOrOneSplatConstant(N1)) {
1890 SDValue X = N0.getOperand(0);
1891 if ((!LegalOperations ||
1892 (TLI.isOperationLegal(ISD::XOR, X.getValueType()) &&
1893 TLI.isOperationLegal(ISD::ZERO_EXTEND, VT))) &&
1894 X.getScalarValueSizeInBits() == 1) {
1895 SDValue Not = DAG.getNOT(DL, X, X.getValueType());
1896 return DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Not);
1901 if (SDValue NewSel = foldBinOpIntoSelect(N))
1905 if (SDValue RADD = ReassociateOps(ISD::ADD, DL, N0, N1))
1908 // fold ((0-A) + B) -> B-A
1909 if (N0.getOpcode() == ISD::SUB &&
1910 isNullConstantOrNullSplatConstant(N0.getOperand(0)))
1911 return DAG.getNode(ISD::SUB, DL, VT, N1, N0.getOperand(1));
1913 // fold (A + (0-B)) -> A-B
1914 if (N1.getOpcode() == ISD::SUB &&
1915 isNullConstantOrNullSplatConstant(N1.getOperand(0)))
1916 return DAG.getNode(ISD::SUB, DL, VT, N0, N1.getOperand(1));
1918 // fold (A+(B-A)) -> B
1919 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1920 return N1.getOperand(0);
1922 // fold ((B-A)+A) -> B
1923 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1924 return N0.getOperand(0);
1926 // fold (A+(B-(A+C))) to (B-C)
1927 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1928 N0 == N1.getOperand(1).getOperand(0))
1929 return DAG.getNode(ISD::SUB, DL, VT, N1.getOperand(0),
1930 N1.getOperand(1).getOperand(1));
1932 // fold (A+(B-(C+A))) to (B-C)
1933 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1934 N0 == N1.getOperand(1).getOperand(1))
1935 return DAG.getNode(ISD::SUB, DL, VT, N1.getOperand(0),
1936 N1.getOperand(1).getOperand(0));
1938 // fold (A+((B-A)+or-C)) to (B+or-C)
1939 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1940 N1.getOperand(0).getOpcode() == ISD::SUB &&
1941 N0 == N1.getOperand(0).getOperand(1))
1942 return DAG.getNode(N1.getOpcode(), DL, VT, N1.getOperand(0).getOperand(0),
1945 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1946 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1947 SDValue N00 = N0.getOperand(0);
1948 SDValue N01 = N0.getOperand(1);
1949 SDValue N10 = N1.getOperand(0);
1950 SDValue N11 = N1.getOperand(1);
1952 if (isConstantOrConstantVector(N00) || isConstantOrConstantVector(N10))
1953 return DAG.getNode(ISD::SUB, DL, VT,
1954 DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10),
1955 DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11));
1958 if (SimplifyDemandedBits(SDValue(N, 0)))
1959 return SDValue(N, 0);
1961 // fold (a+b) -> (a|b) iff a and b share no bits.
1962 if ((!LegalOperations || TLI.isOperationLegal(ISD::OR, VT)) &&
1963 DAG.haveNoCommonBitsSet(N0, N1))
1964 return DAG.getNode(ISD::OR, DL, VT, N0, N1);
1966 if (SDValue Combined = visitADDLike(N0, N1, N))
1969 if (SDValue Combined = visitADDLike(N1, N0, N))
1975 static SDValue getAsCarry(const TargetLowering &TLI, SDValue V) {
1976 bool Masked = false;
1978 // First, peel away TRUNCATE/ZERO_EXTEND/AND nodes due to legalization.
1980 if (V.getOpcode() == ISD::TRUNCATE || V.getOpcode() == ISD::ZERO_EXTEND) {
1981 V = V.getOperand(0);
1985 if (V.getOpcode() == ISD::AND && isOneConstant(V.getOperand(1))) {
1987 V = V.getOperand(0);
1994 // If this is not a carry, return.
1995 if (V.getResNo() != 1)
1998 if (V.getOpcode() != ISD::ADDCARRY && V.getOpcode() != ISD::SUBCARRY &&
1999 V.getOpcode() != ISD::UADDO && V.getOpcode() != ISD::USUBO)
2002 // If the result is masked, then no matter what kind of bool it is we can
2003 // return. If it isn't, then we need to make sure the bool type is either 0 or
2004 // 1 and not other values.
2006 TLI.getBooleanContents(V.getValueType()) ==
2007 TargetLoweringBase::ZeroOrOneBooleanContent)
2013 SDValue DAGCombiner::visitADDLike(SDValue N0, SDValue N1, SDNode *LocReference) {
2014 EVT VT = N0.getValueType();
2015 SDLoc DL(LocReference);
2017 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
2018 if (N1.getOpcode() == ISD::SHL && N1.getOperand(0).getOpcode() == ISD::SUB &&
2019 isNullConstantOrNullSplatConstant(N1.getOperand(0).getOperand(0)))
2020 return DAG.getNode(ISD::SUB, DL, VT, N0,
2021 DAG.getNode(ISD::SHL, DL, VT,
2022 N1.getOperand(0).getOperand(1),
2025 if (N1.getOpcode() == ISD::AND) {
2026 SDValue AndOp0 = N1.getOperand(0);
2027 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
2028 unsigned DestBits = VT.getScalarSizeInBits();
2030 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
2031 // and similar xforms where the inner op is either ~0 or 0.
2032 if (NumSignBits == DestBits &&
2033 isOneConstantOrOneSplatConstant(N1->getOperand(1)))
2034 return DAG.getNode(ISD::SUB, DL, VT, N0, AndOp0);
2037 // add (sext i1), X -> sub X, (zext i1)
2038 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
2039 N0.getOperand(0).getValueType() == MVT::i1 &&
2040 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
2041 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
2042 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
2045 // add X, (sextinreg Y i1) -> sub X, (and Y 1)
2046 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
2047 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
2048 if (TN->getVT() == MVT::i1) {
2049 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
2050 DAG.getConstant(1, DL, VT));
2051 return DAG.getNode(ISD::SUB, DL, VT, N0, ZExt);
2055 // (add X, (addcarry Y, 0, Carry)) -> (addcarry X, Y, Carry)
2056 if (N1.getOpcode() == ISD::ADDCARRY && isNullConstant(N1.getOperand(1)))
2057 return DAG.getNode(ISD::ADDCARRY, DL, N1->getVTList(),
2058 N0, N1.getOperand(0), N1.getOperand(2));
2060 // (add X, Carry) -> (addcarry X, 0, Carry)
2061 if (TLI.isOperationLegalOrCustom(ISD::ADDCARRY, VT))
2062 if (SDValue Carry = getAsCarry(TLI, N1))
2063 return DAG.getNode(ISD::ADDCARRY, DL,
2064 DAG.getVTList(VT, Carry.getValueType()), N0,
2065 DAG.getConstant(0, DL, VT), Carry);
2070 SDValue DAGCombiner::visitADDC(SDNode *N) {
2071 SDValue N0 = N->getOperand(0);
2072 SDValue N1 = N->getOperand(1);
2073 EVT VT = N0.getValueType();
2076 // If the flag result is dead, turn this into an ADD.
2077 if (!N->hasAnyUseOfValue(1))
2078 return CombineTo(N, DAG.getNode(ISD::ADD, DL, VT, N0, N1),
2079 DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue));
2081 // canonicalize constant to RHS.
2082 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2083 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2085 return DAG.getNode(ISD::ADDC, DL, N->getVTList(), N1, N0);
2087 // fold (addc x, 0) -> x + no carry out
2088 if (isNullConstant(N1))
2089 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
2092 // If it cannot overflow, transform into an add.
2093 if (DAG.computeOverflowKind(N0, N1) == SelectionDAG::OFK_Never)
2094 return CombineTo(N, DAG.getNode(ISD::ADD, DL, VT, N0, N1),
2095 DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue));
2100 SDValue DAGCombiner::visitUADDO(SDNode *N) {
2101 SDValue N0 = N->getOperand(0);
2102 SDValue N1 = N->getOperand(1);
2103 EVT VT = N0.getValueType();
2107 EVT CarryVT = N->getValueType(1);
2110 // If the flag result is dead, turn this into an ADD.
2111 if (!N->hasAnyUseOfValue(1))
2112 return CombineTo(N, DAG.getNode(ISD::ADD, DL, VT, N0, N1),
2113 DAG.getUNDEF(CarryVT));
2115 // canonicalize constant to RHS.
2116 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2117 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2119 return DAG.getNode(ISD::UADDO, DL, N->getVTList(), N1, N0);
2121 // fold (uaddo x, 0) -> x + no carry out
2122 if (isNullConstant(N1))
2123 return CombineTo(N, N0, DAG.getConstant(0, DL, CarryVT));
2125 // If it cannot overflow, transform into an add.
2126 if (DAG.computeOverflowKind(N0, N1) == SelectionDAG::OFK_Never)
2127 return CombineTo(N, DAG.getNode(ISD::ADD, DL, VT, N0, N1),
2128 DAG.getConstant(0, DL, CarryVT));
2130 if (SDValue Combined = visitUADDOLike(N0, N1, N))
2133 if (SDValue Combined = visitUADDOLike(N1, N0, N))
2139 SDValue DAGCombiner::visitUADDOLike(SDValue N0, SDValue N1, SDNode *N) {
2140 auto VT = N0.getValueType();
2142 // (uaddo X, (addcarry Y, 0, Carry)) -> (addcarry X, Y, Carry)
2143 // If Y + 1 cannot overflow.
2144 if (N1.getOpcode() == ISD::ADDCARRY && isNullConstant(N1.getOperand(1))) {
2145 SDValue Y = N1.getOperand(0);
2146 SDValue One = DAG.getConstant(1, SDLoc(N), Y.getValueType());
2147 if (DAG.computeOverflowKind(Y, One) == SelectionDAG::OFK_Never)
2148 return DAG.getNode(ISD::ADDCARRY, SDLoc(N), N->getVTList(), N0, Y,
2152 // (uaddo X, Carry) -> (addcarry X, 0, Carry)
2153 if (TLI.isOperationLegalOrCustom(ISD::ADDCARRY, VT))
2154 if (SDValue Carry = getAsCarry(TLI, N1))
2155 return DAG.getNode(ISD::ADDCARRY, SDLoc(N), N->getVTList(), N0,
2156 DAG.getConstant(0, SDLoc(N), VT), Carry);
2161 SDValue DAGCombiner::visitADDE(SDNode *N) {
2162 SDValue N0 = N->getOperand(0);
2163 SDValue N1 = N->getOperand(1);
2164 SDValue CarryIn = N->getOperand(2);
2166 // canonicalize constant to RHS
2167 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2168 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2170 return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(),
2173 // fold (adde x, y, false) -> (addc x, y)
2174 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
2175 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1);
2180 SDValue DAGCombiner::visitADDCARRY(SDNode *N) {
2181 SDValue N0 = N->getOperand(0);
2182 SDValue N1 = N->getOperand(1);
2183 SDValue CarryIn = N->getOperand(2);
2186 // canonicalize constant to RHS
2187 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2188 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2190 return DAG.getNode(ISD::ADDCARRY, DL, N->getVTList(), N1, N0, CarryIn);
2192 // fold (addcarry x, y, false) -> (uaddo x, y)
2193 if (isNullConstant(CarryIn))
2194 return DAG.getNode(ISD::UADDO, DL, N->getVTList(), N0, N1);
2196 // fold (addcarry 0, 0, X) -> (and (ext/trunc X), 1) and no carry.
2197 if (isNullConstant(N0) && isNullConstant(N1)) {
2198 EVT VT = N0.getValueType();
2199 EVT CarryVT = CarryIn.getValueType();
2200 SDValue CarryExt = DAG.getBoolExtOrTrunc(CarryIn, DL, VT, CarryVT);
2201 AddToWorklist(CarryExt.getNode());
2202 return CombineTo(N, DAG.getNode(ISD::AND, DL, VT, CarryExt,
2203 DAG.getConstant(1, DL, VT)),
2204 DAG.getConstant(0, DL, CarryVT));
2207 if (SDValue Combined = visitADDCARRYLike(N0, N1, CarryIn, N))
2210 if (SDValue Combined = visitADDCARRYLike(N1, N0, CarryIn, N))
2216 SDValue DAGCombiner::visitADDCARRYLike(SDValue N0, SDValue N1, SDValue CarryIn,
2218 // Iff the flag result is dead:
2219 // (addcarry (add|uaddo X, Y), 0, Carry) -> (addcarry X, Y, Carry)
2220 if ((N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::UADDO) &&
2221 isNullConstant(N1) && !N->hasAnyUseOfValue(1))
2222 return DAG.getNode(ISD::ADDCARRY, SDLoc(N), N->getVTList(),
2223 N0.getOperand(0), N0.getOperand(1), CarryIn);
2226 * When one of the addcarry argument is itself a carry, we may be facing
2227 * a diamond carry propagation. In which case we try to transform the DAG
2228 * to ensure linear carry propagation if that is possible.
2230 * We are trying to get:
2231 * (addcarry X, 0, (addcarry A, B, Z):Carry)
2233 if (auto Y = getAsCarry(TLI, N1)) {
2239 * | (addcarry *, 0, Z)
2243 * (addcarry X, *, *)
2245 if (Y.getOpcode() == ISD::UADDO &&
2246 CarryIn.getResNo() == 1 &&
2247 CarryIn.getOpcode() == ISD::ADDCARRY &&
2248 isNullConstant(CarryIn.getOperand(1)) &&
2249 CarryIn.getOperand(0) == Y.getValue(0)) {
2250 auto NewY = DAG.getNode(ISD::ADDCARRY, SDLoc(N), Y->getVTList(),
2251 Y.getOperand(0), Y.getOperand(1),
2252 CarryIn.getOperand(2));
2253 AddToWorklist(NewY.getNode());
2254 return DAG.getNode(ISD::ADDCARRY, SDLoc(N), N->getVTList(), N0,
2255 DAG.getConstant(0, SDLoc(N), N0.getValueType()),
2263 // Since it may not be valid to emit a fold to zero for vector initializers
2264 // check if we can before folding.
2265 static SDValue tryFoldToZero(const SDLoc &DL, const TargetLowering &TLI, EVT VT,
2266 SelectionDAG &DAG, bool LegalOperations,
2269 return DAG.getConstant(0, DL, VT);
2270 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
2271 return DAG.getConstant(0, DL, VT);
2275 SDValue DAGCombiner::visitSUB(SDNode *N) {
2276 SDValue N0 = N->getOperand(0);
2277 SDValue N1 = N->getOperand(1);
2278 EVT VT = N0.getValueType();
2282 if (VT.isVector()) {
2283 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2286 // fold (sub x, 0) -> x, vector edition
2287 if (ISD::isBuildVectorAllZeros(N1.getNode()))
2291 // fold (sub x, x) -> 0
2292 // FIXME: Refactor this and xor and other similar operations together.
2294 return tryFoldToZero(DL, TLI, VT, DAG, LegalOperations, LegalTypes);
2295 if (DAG.isConstantIntBuildVectorOrConstantInt(N0) &&
2296 DAG.isConstantIntBuildVectorOrConstantInt(N1)) {
2297 // fold (sub c1, c2) -> c1-c2
2298 return DAG.FoldConstantArithmetic(ISD::SUB, DL, VT, N0.getNode(),
2302 if (SDValue NewSel = foldBinOpIntoSelect(N))
2305 ConstantSDNode *N1C = getAsNonOpaqueConstant(N1);
2307 // fold (sub x, c) -> (add x, -c)
2309 return DAG.getNode(ISD::ADD, DL, VT, N0,
2310 DAG.getConstant(-N1C->getAPIntValue(), DL, VT));
2313 if (isNullConstantOrNullSplatConstant(N0)) {
2314 unsigned BitWidth = VT.getScalarSizeInBits();
2315 // Right-shifting everything out but the sign bit followed by negation is
2316 // the same as flipping arithmetic/logical shift type without the negation:
2317 // -(X >>u 31) -> (X >>s 31)
2318 // -(X >>s 31) -> (X >>u 31)
2319 if (N1->getOpcode() == ISD::SRA || N1->getOpcode() == ISD::SRL) {
2320 ConstantSDNode *ShiftAmt = isConstOrConstSplat(N1.getOperand(1));
2321 if (ShiftAmt && ShiftAmt->getZExtValue() == BitWidth - 1) {
2322 auto NewSh = N1->getOpcode() == ISD::SRA ? ISD::SRL : ISD::SRA;
2323 if (!LegalOperations || TLI.isOperationLegal(NewSh, VT))
2324 return DAG.getNode(NewSh, DL, VT, N1.getOperand(0), N1.getOperand(1));
2328 // 0 - X --> 0 if the sub is NUW.
2329 if (N->getFlags().hasNoUnsignedWrap())
2332 if (DAG.MaskedValueIsZero(N1, ~APInt::getSignMask(BitWidth))) {
2333 // N1 is either 0 or the minimum signed value. If the sub is NSW, then
2334 // N1 must be 0 because negating the minimum signed value is undefined.
2335 if (N->getFlags().hasNoSignedWrap())
2338 // 0 - X --> X if X is 0 or the minimum signed value.
2343 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
2344 if (isAllOnesConstantOrAllOnesSplatConstant(N0))
2345 return DAG.getNode(ISD::XOR, DL, VT, N1, N0);
2347 // fold A-(A-B) -> B
2348 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
2349 return N1.getOperand(1);
2351 // fold (A+B)-A -> B
2352 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
2353 return N0.getOperand(1);
2355 // fold (A+B)-B -> A
2356 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
2357 return N0.getOperand(0);
2359 // fold C2-(A+C1) -> (C2-C1)-A
2360 if (N1.getOpcode() == ISD::ADD) {
2361 SDValue N11 = N1.getOperand(1);
2362 if (isConstantOrConstantVector(N0, /* NoOpaques */ true) &&
2363 isConstantOrConstantVector(N11, /* NoOpaques */ true)) {
2364 SDValue NewC = DAG.getNode(ISD::SUB, DL, VT, N0, N11);
2365 return DAG.getNode(ISD::SUB, DL, VT, NewC, N1.getOperand(0));
2369 // fold ((A+(B+or-C))-B) -> A+or-C
2370 if (N0.getOpcode() == ISD::ADD &&
2371 (N0.getOperand(1).getOpcode() == ISD::SUB ||
2372 N0.getOperand(1).getOpcode() == ISD::ADD) &&
2373 N0.getOperand(1).getOperand(0) == N1)
2374 return DAG.getNode(N0.getOperand(1).getOpcode(), DL, VT, N0.getOperand(0),
2375 N0.getOperand(1).getOperand(1));
2377 // fold ((A+(C+B))-B) -> A+C
2378 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1).getOpcode() == ISD::ADD &&
2379 N0.getOperand(1).getOperand(1) == N1)
2380 return DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0),
2381 N0.getOperand(1).getOperand(0));
2383 // fold ((A-(B-C))-C) -> A-B
2384 if (N0.getOpcode() == ISD::SUB && N0.getOperand(1).getOpcode() == ISD::SUB &&
2385 N0.getOperand(1).getOperand(1) == N1)
2386 return DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(0),
2387 N0.getOperand(1).getOperand(0));
2389 // If either operand of a sub is undef, the result is undef
2395 // If the relocation model supports it, consider symbol offsets.
2396 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
2397 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
2398 // fold (sub Sym, c) -> Sym-c
2399 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
2400 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
2402 (uint64_t)N1C->getSExtValue());
2403 // fold (sub Sym+c1, Sym+c2) -> c1-c2
2404 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
2405 if (GA->getGlobal() == GB->getGlobal())
2406 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
2410 // sub X, (sextinreg Y i1) -> add X, (and Y 1)
2411 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
2412 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
2413 if (TN->getVT() == MVT::i1) {
2414 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
2415 DAG.getConstant(1, DL, VT));
2416 return DAG.getNode(ISD::ADD, DL, VT, N0, ZExt);
2423 SDValue DAGCombiner::visitSUBC(SDNode *N) {
2424 SDValue N0 = N->getOperand(0);
2425 SDValue N1 = N->getOperand(1);
2426 EVT VT = N0.getValueType();
2429 // If the flag result is dead, turn this into an SUB.
2430 if (!N->hasAnyUseOfValue(1))
2431 return CombineTo(N, DAG.getNode(ISD::SUB, DL, VT, N0, N1),
2432 DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue));
2434 // fold (subc x, x) -> 0 + no borrow
2436 return CombineTo(N, DAG.getConstant(0, DL, VT),
2437 DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue));
2439 // fold (subc x, 0) -> x + no borrow
2440 if (isNullConstant(N1))
2441 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue));
2443 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
2444 if (isAllOnesConstant(N0))
2445 return CombineTo(N, DAG.getNode(ISD::XOR, DL, VT, N1, N0),
2446 DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue));
2451 SDValue DAGCombiner::visitUSUBO(SDNode *N) {
2452 SDValue N0 = N->getOperand(0);
2453 SDValue N1 = N->getOperand(1);
2454 EVT VT = N0.getValueType();
2458 EVT CarryVT = N->getValueType(1);
2461 // If the flag result is dead, turn this into an SUB.
2462 if (!N->hasAnyUseOfValue(1))
2463 return CombineTo(N, DAG.getNode(ISD::SUB, DL, VT, N0, N1),
2464 DAG.getUNDEF(CarryVT));
2466 // fold (usubo x, x) -> 0 + no borrow
2468 return CombineTo(N, DAG.getConstant(0, DL, VT),
2469 DAG.getConstant(0, DL, CarryVT));
2471 // fold (usubo x, 0) -> x + no borrow
2472 if (isNullConstant(N1))
2473 return CombineTo(N, N0, DAG.getConstant(0, DL, CarryVT));
2475 // Canonicalize (usubo -1, x) -> ~x, i.e. (xor x, -1) + no borrow
2476 if (isAllOnesConstant(N0))
2477 return CombineTo(N, DAG.getNode(ISD::XOR, DL, VT, N1, N0),
2478 DAG.getConstant(0, DL, CarryVT));
2483 SDValue DAGCombiner::visitSUBE(SDNode *N) {
2484 SDValue N0 = N->getOperand(0);
2485 SDValue N1 = N->getOperand(1);
2486 SDValue CarryIn = N->getOperand(2);
2488 // fold (sube x, y, false) -> (subc x, y)
2489 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
2490 return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1);
2495 SDValue DAGCombiner::visitSUBCARRY(SDNode *N) {
2496 SDValue N0 = N->getOperand(0);
2497 SDValue N1 = N->getOperand(1);
2498 SDValue CarryIn = N->getOperand(2);
2500 // fold (subcarry x, y, false) -> (usubo x, y)
2501 if (isNullConstant(CarryIn))
2502 return DAG.getNode(ISD::USUBO, SDLoc(N), N->getVTList(), N0, N1);
2507 SDValue DAGCombiner::visitMUL(SDNode *N) {
2508 SDValue N0 = N->getOperand(0);
2509 SDValue N1 = N->getOperand(1);
2510 EVT VT = N0.getValueType();
2512 // fold (mul x, undef) -> 0
2513 if (N0.isUndef() || N1.isUndef())
2514 return DAG.getConstant(0, SDLoc(N), VT);
2516 bool N0IsConst = false;
2517 bool N1IsConst = false;
2518 bool N1IsOpaqueConst = false;
2519 bool N0IsOpaqueConst = false;
2520 APInt ConstValue0, ConstValue1;
2522 if (VT.isVector()) {
2523 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2526 N0IsConst = ISD::isConstantSplatVector(N0.getNode(), ConstValue0);
2527 N1IsConst = ISD::isConstantSplatVector(N1.getNode(), ConstValue1);
2529 N0IsConst = isa<ConstantSDNode>(N0);
2531 ConstValue0 = cast<ConstantSDNode>(N0)->getAPIntValue();
2532 N0IsOpaqueConst = cast<ConstantSDNode>(N0)->isOpaque();
2534 N1IsConst = isa<ConstantSDNode>(N1);
2536 ConstValue1 = cast<ConstantSDNode>(N1)->getAPIntValue();
2537 N1IsOpaqueConst = cast<ConstantSDNode>(N1)->isOpaque();
2541 // fold (mul c1, c2) -> c1*c2
2542 if (N0IsConst && N1IsConst && !N0IsOpaqueConst && !N1IsOpaqueConst)
2543 return DAG.FoldConstantArithmetic(ISD::MUL, SDLoc(N), VT,
2544 N0.getNode(), N1.getNode());
2546 // canonicalize constant to RHS (vector doesn't have to splat)
2547 if (DAG.isConstantIntBuildVectorOrConstantInt(N0) &&
2548 !DAG.isConstantIntBuildVectorOrConstantInt(N1))
2549 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0);
2550 // fold (mul x, 0) -> 0
2551 if (N1IsConst && ConstValue1 == 0)
2553 // We require a splat of the entire scalar bit width for non-contiguous
2556 ConstValue1.getBitWidth() == VT.getScalarSizeInBits();
2557 // fold (mul x, 1) -> x
2558 if (N1IsConst && ConstValue1 == 1 && IsFullSplat)
2561 if (SDValue NewSel = foldBinOpIntoSelect(N))
2564 // fold (mul x, -1) -> 0-x
2565 if (N1IsConst && ConstValue1.isAllOnesValue()) {
2567 return DAG.getNode(ISD::SUB, DL, VT,
2568 DAG.getConstant(0, DL, VT), N0);
2570 // fold (mul x, (1 << c)) -> x << c
2571 if (N1IsConst && !N1IsOpaqueConst && ConstValue1.isPowerOf2() &&
2574 return DAG.getNode(ISD::SHL, DL, VT, N0,
2575 DAG.getConstant(ConstValue1.logBase2(), DL,
2576 getShiftAmountTy(N0.getValueType())));
2578 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
2579 if (N1IsConst && !N1IsOpaqueConst && (-ConstValue1).isPowerOf2() &&
2581 unsigned Log2Val = (-ConstValue1).logBase2();
2583 // FIXME: If the input is something that is easily negated (e.g. a
2584 // single-use add), we should put the negate there.
2585 return DAG.getNode(ISD::SUB, DL, VT,
2586 DAG.getConstant(0, DL, VT),
2587 DAG.getNode(ISD::SHL, DL, VT, N0,
2588 DAG.getConstant(Log2Val, DL,
2589 getShiftAmountTy(N0.getValueType()))));
2592 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
2593 if (N0.getOpcode() == ISD::SHL &&
2594 isConstantOrConstantVector(N1, /* NoOpaques */ true) &&
2595 isConstantOrConstantVector(N0.getOperand(1), /* NoOpaques */ true)) {
2596 SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT, N1, N0.getOperand(1));
2597 if (isConstantOrConstantVector(C3))
2598 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N0.getOperand(0), C3);
2601 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
2604 SDValue Sh(nullptr, 0), Y(nullptr, 0);
2606 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
2607 if (N0.getOpcode() == ISD::SHL &&
2608 isConstantOrConstantVector(N0.getOperand(1)) &&
2609 N0.getNode()->hasOneUse()) {
2611 } else if (N1.getOpcode() == ISD::SHL &&
2612 isConstantOrConstantVector(N1.getOperand(1)) &&
2613 N1.getNode()->hasOneUse()) {
2618 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT, Sh.getOperand(0), Y);
2619 return DAG.getNode(ISD::SHL, SDLoc(N), VT, Mul, Sh.getOperand(1));
2623 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
2624 if (DAG.isConstantIntBuildVectorOrConstantInt(N1) &&
2625 N0.getOpcode() == ISD::ADD &&
2626 DAG.isConstantIntBuildVectorOrConstantInt(N0.getOperand(1)) &&
2627 isMulAddWithConstProfitable(N, N0, N1))
2628 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
2629 DAG.getNode(ISD::MUL, SDLoc(N0), VT,
2630 N0.getOperand(0), N1),
2631 DAG.getNode(ISD::MUL, SDLoc(N1), VT,
2632 N0.getOperand(1), N1));
2635 if (SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1))
2641 /// Return true if divmod libcall is available.
2642 static bool isDivRemLibcallAvailable(SDNode *Node, bool isSigned,
2643 const TargetLowering &TLI) {
2645 EVT NodeType = Node->getValueType(0);
2646 if (!NodeType.isSimple())
2648 switch (NodeType.getSimpleVT().SimpleTy) {
2649 default: return false; // No libcall for vector types.
2650 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2651 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2652 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2653 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2654 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2657 return TLI.getLibcallName(LC) != nullptr;
2660 /// Issue divrem if both quotient and remainder are needed.
2661 SDValue DAGCombiner::useDivRem(SDNode *Node) {
2662 if (Node->use_empty())
2663 return SDValue(); // This is a dead node, leave it alone.
2665 unsigned Opcode = Node->getOpcode();
2666 bool isSigned = (Opcode == ISD::SDIV) || (Opcode == ISD::SREM);
2667 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
2669 // DivMod lib calls can still work on non-legal types if using lib-calls.
2670 EVT VT = Node->getValueType(0);
2671 if (VT.isVector() || !VT.isInteger())
2674 if (!TLI.isTypeLegal(VT) && !TLI.isOperationCustom(DivRemOpc, VT))
2677 // If DIVREM is going to get expanded into a libcall,
2678 // but there is no libcall available, then don't combine.
2679 if (!TLI.isOperationLegalOrCustom(DivRemOpc, VT) &&
2680 !isDivRemLibcallAvailable(Node, isSigned, TLI))
2683 // If div is legal, it's better to do the normal expansion
2684 unsigned OtherOpcode = 0;
2685 if ((Opcode == ISD::SDIV) || (Opcode == ISD::UDIV)) {
2686 OtherOpcode = isSigned ? ISD::SREM : ISD::UREM;
2687 if (TLI.isOperationLegalOrCustom(Opcode, VT))
2690 OtherOpcode = isSigned ? ISD::SDIV : ISD::UDIV;
2691 if (TLI.isOperationLegalOrCustom(OtherOpcode, VT))
2695 SDValue Op0 = Node->getOperand(0);
2696 SDValue Op1 = Node->getOperand(1);
2698 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2699 UE = Op0.getNode()->use_end(); UI != UE;) {
2700 SDNode *User = *UI++;
2701 if (User == Node || User->use_empty())
2703 // Convert the other matching node(s), too;
2704 // otherwise, the DIVREM may get target-legalized into something
2705 // target-specific that we won't be able to recognize.
2706 unsigned UserOpc = User->getOpcode();
2707 if ((UserOpc == Opcode || UserOpc == OtherOpcode || UserOpc == DivRemOpc) &&
2708 User->getOperand(0) == Op0 &&
2709 User->getOperand(1) == Op1) {
2711 if (UserOpc == OtherOpcode) {
2712 SDVTList VTs = DAG.getVTList(VT, VT);
2713 combined = DAG.getNode(DivRemOpc, SDLoc(Node), VTs, Op0, Op1);
2714 } else if (UserOpc == DivRemOpc) {
2715 combined = SDValue(User, 0);
2717 assert(UserOpc == Opcode);
2721 if (UserOpc == ISD::SDIV || UserOpc == ISD::UDIV)
2722 CombineTo(User, combined);
2723 else if (UserOpc == ISD::SREM || UserOpc == ISD::UREM)
2724 CombineTo(User, combined.getValue(1));
2730 static SDValue simplifyDivRem(SDNode *N, SelectionDAG &DAG) {
2731 SDValue N0 = N->getOperand(0);
2732 SDValue N1 = N->getOperand(1);
2733 EVT VT = N->getValueType(0);
2736 if (DAG.isUndef(N->getOpcode(), {N0, N1}))
2737 return DAG.getUNDEF(VT);
2742 return DAG.getConstant(0, DL, VT);
2747 SDValue DAGCombiner::visitSDIV(SDNode *N) {
2748 SDValue N0 = N->getOperand(0);
2749 SDValue N1 = N->getOperand(1);
2750 EVT VT = N->getValueType(0);
2754 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2759 // fold (sdiv c1, c2) -> c1/c2
2760 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2761 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2762 if (N0C && N1C && !N0C->isOpaque() && !N1C->isOpaque())
2763 return DAG.FoldConstantArithmetic(ISD::SDIV, DL, VT, N0C, N1C);
2764 // fold (sdiv X, 1) -> X
2765 if (N1C && N1C->isOne())
2767 // fold (sdiv X, -1) -> 0-X
2768 if (N1C && N1C->isAllOnesValue())
2769 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), N0);
2771 if (SDValue V = simplifyDivRem(N, DAG))
2774 if (SDValue NewSel = foldBinOpIntoSelect(N))
2777 // If we know the sign bits of both operands are zero, strength reduce to a
2778 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
2779 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2780 return DAG.getNode(ISD::UDIV, DL, N1.getValueType(), N0, N1);
2782 // fold (sdiv X, pow2) -> simple ops after legalize
2783 // FIXME: We check for the exact bit here because the generic lowering gives
2784 // better results in that case. The target-specific lowering should learn how
2785 // to handle exact sdivs efficiently.
2786 if (N1C && !N1C->isNullValue() && !N1C->isOpaque() &&
2787 !N->getFlags().hasExact() && (N1C->getAPIntValue().isPowerOf2() ||
2788 (-N1C->getAPIntValue()).isPowerOf2())) {
2789 // Target-specific implementation of sdiv x, pow2.
2790 if (SDValue Res = BuildSDIVPow2(N))
2793 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
2795 // Splat the sign bit into the register
2797 DAG.getNode(ISD::SRA, DL, VT, N0,
2798 DAG.getConstant(VT.getScalarSizeInBits() - 1, DL,
2799 getShiftAmountTy(N0.getValueType())));
2800 AddToWorklist(SGN.getNode());
2802 // Add (N0 < 0) ? abs2 - 1 : 0;
2804 DAG.getNode(ISD::SRL, DL, VT, SGN,
2805 DAG.getConstant(VT.getScalarSizeInBits() - lg2, DL,
2806 getShiftAmountTy(SGN.getValueType())));
2807 SDValue ADD = DAG.getNode(ISD::ADD, DL, VT, N0, SRL);
2808 AddToWorklist(SRL.getNode());
2809 AddToWorklist(ADD.getNode()); // Divide by pow2
2810 SDValue SRA = DAG.getNode(ISD::SRA, DL, VT, ADD,
2811 DAG.getConstant(lg2, DL,
2812 getShiftAmountTy(ADD.getValueType())));
2814 // If we're dividing by a positive value, we're done. Otherwise, we must
2815 // negate the result.
2816 if (N1C->getAPIntValue().isNonNegative())
2819 AddToWorklist(SRA.getNode());
2820 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), SRA);
2823 // If integer divide is expensive and we satisfy the requirements, emit an
2824 // alternate sequence. Targets may check function attributes for size/speed
2826 AttributeList Attr = DAG.getMachineFunction().getFunction()->getAttributes();
2827 if (N1C && !TLI.isIntDivCheap(N->getValueType(0), Attr))
2828 if (SDValue Op = BuildSDIV(N))
2831 // sdiv, srem -> sdivrem
2832 // If the divisor is constant, then return DIVREM only if isIntDivCheap() is
2833 // true. Otherwise, we break the simplification logic in visitREM().
2834 if (!N1C || TLI.isIntDivCheap(N->getValueType(0), Attr))
2835 if (SDValue DivRem = useDivRem(N))
2841 SDValue DAGCombiner::visitUDIV(SDNode *N) {
2842 SDValue N0 = N->getOperand(0);
2843 SDValue N1 = N->getOperand(1);
2844 EVT VT = N->getValueType(0);
2848 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2853 // fold (udiv c1, c2) -> c1/c2
2854 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2855 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2857 if (SDValue Folded = DAG.FoldConstantArithmetic(ISD::UDIV, DL, VT,
2861 if (SDValue V = simplifyDivRem(N, DAG))
2864 if (SDValue NewSel = foldBinOpIntoSelect(N))
2867 // fold (udiv x, (1 << c)) -> x >>u c
2868 if (isConstantOrConstantVector(N1, /*NoOpaques*/ true) &&
2869 DAG.isKnownToBeAPowerOfTwo(N1)) {
2870 SDValue LogBase2 = BuildLogBase2(N1, DL);
2871 AddToWorklist(LogBase2.getNode());
2873 EVT ShiftVT = getShiftAmountTy(N0.getValueType());
2874 SDValue Trunc = DAG.getZExtOrTrunc(LogBase2, DL, ShiftVT);
2875 AddToWorklist(Trunc.getNode());
2876 return DAG.getNode(ISD::SRL, DL, VT, N0, Trunc);
2879 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
2880 if (N1.getOpcode() == ISD::SHL) {
2881 SDValue N10 = N1.getOperand(0);
2882 if (isConstantOrConstantVector(N10, /*NoOpaques*/ true) &&
2883 DAG.isKnownToBeAPowerOfTwo(N10)) {
2884 SDValue LogBase2 = BuildLogBase2(N10, DL);
2885 AddToWorklist(LogBase2.getNode());
2887 EVT ADDVT = N1.getOperand(1).getValueType();
2888 SDValue Trunc = DAG.getZExtOrTrunc(LogBase2, DL, ADDVT);
2889 AddToWorklist(Trunc.getNode());
2890 SDValue Add = DAG.getNode(ISD::ADD, DL, ADDVT, N1.getOperand(1), Trunc);
2891 AddToWorklist(Add.getNode());
2892 return DAG.getNode(ISD::SRL, DL, VT, N0, Add);
2896 // fold (udiv x, c) -> alternate
2897 AttributeList Attr = DAG.getMachineFunction().getFunction()->getAttributes();
2898 if (N1C && !TLI.isIntDivCheap(N->getValueType(0), Attr))
2899 if (SDValue Op = BuildUDIV(N))
2902 // sdiv, srem -> sdivrem
2903 // If the divisor is constant, then return DIVREM only if isIntDivCheap() is
2904 // true. Otherwise, we break the simplification logic in visitREM().
2905 if (!N1C || TLI.isIntDivCheap(N->getValueType(0), Attr))
2906 if (SDValue DivRem = useDivRem(N))
2912 // handles ISD::SREM and ISD::UREM
2913 SDValue DAGCombiner::visitREM(SDNode *N) {
2914 unsigned Opcode = N->getOpcode();
2915 SDValue N0 = N->getOperand(0);
2916 SDValue N1 = N->getOperand(1);
2917 EVT VT = N->getValueType(0);
2918 bool isSigned = (Opcode == ISD::SREM);
2921 // fold (rem c1, c2) -> c1%c2
2922 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2923 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2925 if (SDValue Folded = DAG.FoldConstantArithmetic(Opcode, DL, VT, N0C, N1C))
2928 if (SDValue V = simplifyDivRem(N, DAG))
2931 if (SDValue NewSel = foldBinOpIntoSelect(N))
2935 // If we know the sign bits of both operands are zero, strength reduce to a
2936 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
2937 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2938 return DAG.getNode(ISD::UREM, DL, VT, N0, N1);
2940 SDValue NegOne = DAG.getAllOnesConstant(DL, VT);
2941 if (DAG.isKnownToBeAPowerOfTwo(N1)) {
2942 // fold (urem x, pow2) -> (and x, pow2-1)
2943 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N1, NegOne);
2944 AddToWorklist(Add.getNode());
2945 return DAG.getNode(ISD::AND, DL, VT, N0, Add);
2947 if (N1.getOpcode() == ISD::SHL &&
2948 DAG.isKnownToBeAPowerOfTwo(N1.getOperand(0))) {
2949 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2950 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N1, NegOne);
2951 AddToWorklist(Add.getNode());
2952 return DAG.getNode(ISD::AND, DL, VT, N0, Add);
2956 AttributeList Attr = DAG.getMachineFunction().getFunction()->getAttributes();
2958 // If X/C can be simplified by the division-by-constant logic, lower
2959 // X%C to the equivalent of X-X/C*C.
2960 // To avoid mangling nodes, this simplification requires that the combine()
2961 // call for the speculative DIV must not cause a DIVREM conversion. We guard
2962 // against this by skipping the simplification if isIntDivCheap(). When
2963 // div is not cheap, combine will not return a DIVREM. Regardless,
2964 // checking cheapness here makes sense since the simplification results in
2966 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap(VT, Attr)) {
2967 unsigned DivOpcode = isSigned ? ISD::SDIV : ISD::UDIV;
2968 SDValue Div = DAG.getNode(DivOpcode, DL, VT, N0, N1);
2969 AddToWorklist(Div.getNode());
2970 SDValue OptimizedDiv = combine(Div.getNode());
2971 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2972 assert((OptimizedDiv.getOpcode() != ISD::UDIVREM) &&
2973 (OptimizedDiv.getOpcode() != ISD::SDIVREM));
2974 SDValue Mul = DAG.getNode(ISD::MUL, DL, VT, OptimizedDiv, N1);
2975 SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0, Mul);
2976 AddToWorklist(Mul.getNode());
2981 // sdiv, srem -> sdivrem
2982 if (SDValue DivRem = useDivRem(N))
2983 return DivRem.getValue(1);
2988 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2989 SDValue N0 = N->getOperand(0);
2990 SDValue N1 = N->getOperand(1);
2991 EVT VT = N->getValueType(0);
2994 // fold (mulhs x, 0) -> 0
2995 if (isNullConstant(N1))
2997 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2998 if (isOneConstant(N1)) {
3000 return DAG.getNode(ISD::SRA, DL, N0.getValueType(), N0,
3001 DAG.getConstant(N0.getValueSizeInBits() - 1, DL,
3002 getShiftAmountTy(N0.getValueType())));
3004 // fold (mulhs x, undef) -> 0
3005 if (N0.isUndef() || N1.isUndef())
3006 return DAG.getConstant(0, SDLoc(N), VT);
3008 // If the type twice as wide is legal, transform the mulhs to a wider multiply
3010 if (VT.isSimple() && !VT.isVector()) {
3011 MVT Simple = VT.getSimpleVT();
3012 unsigned SimpleSize = Simple.getSizeInBits();
3013 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
3014 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
3015 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
3016 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
3017 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
3018 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
3019 DAG.getConstant(SimpleSize, DL,
3020 getShiftAmountTy(N1.getValueType())));
3021 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
3028 SDValue DAGCombiner::visitMULHU(SDNode *N) {
3029 SDValue N0 = N->getOperand(0);
3030 SDValue N1 = N->getOperand(1);
3031 EVT VT = N->getValueType(0);
3034 // fold (mulhu x, 0) -> 0
3035 if (isNullConstant(N1))
3037 // fold (mulhu x, 1) -> 0
3038 if (isOneConstant(N1))
3039 return DAG.getConstant(0, DL, N0.getValueType());
3040 // fold (mulhu x, undef) -> 0
3041 if (N0.isUndef() || N1.isUndef())
3042 return DAG.getConstant(0, DL, VT);
3044 // If the type twice as wide is legal, transform the mulhu to a wider multiply
3046 if (VT.isSimple() && !VT.isVector()) {
3047 MVT Simple = VT.getSimpleVT();
3048 unsigned SimpleSize = Simple.getSizeInBits();
3049 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
3050 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
3051 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
3052 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
3053 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
3054 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
3055 DAG.getConstant(SimpleSize, DL,
3056 getShiftAmountTy(N1.getValueType())));
3057 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
3064 /// Perform optimizations common to nodes that compute two values. LoOp and HiOp
3065 /// give the opcodes for the two computations that are being performed. Return
3066 /// true if a simplification was made.
3067 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
3069 // If the high half is not needed, just compute the low half.
3070 bool HiExists = N->hasAnyUseOfValue(1);
3072 (!LegalOperations ||
3073 TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) {
3074 SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
3075 return CombineTo(N, Res, Res);
3078 // If the low half is not needed, just compute the high half.
3079 bool LoExists = N->hasAnyUseOfValue(0);
3081 (!LegalOperations ||
3082 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
3083 SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
3084 return CombineTo(N, Res, Res);
3087 // If both halves are used, return as it is.
3088 if (LoExists && HiExists)
3091 // If the two computed results can be simplified separately, separate them.
3093 SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
3094 AddToWorklist(Lo.getNode());
3095 SDValue LoOpt = combine(Lo.getNode());
3096 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
3097 (!LegalOperations ||
3098 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
3099 return CombineTo(N, LoOpt, LoOpt);
3103 SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
3104 AddToWorklist(Hi.getNode());
3105 SDValue HiOpt = combine(Hi.getNode());
3106 if (HiOpt.getNode() && HiOpt != Hi &&
3107 (!LegalOperations ||
3108 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
3109 return CombineTo(N, HiOpt, HiOpt);
3115 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
3116 if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS))
3119 EVT VT = N->getValueType(0);
3122 // If the type is twice as wide is legal, transform the mulhu to a wider
3123 // multiply plus a shift.
3124 if (VT.isSimple() && !VT.isVector()) {
3125 MVT Simple = VT.getSimpleVT();
3126 unsigned SimpleSize = Simple.getSizeInBits();
3127 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
3128 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
3129 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
3130 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
3131 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
3132 // Compute the high part as N1.
3133 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
3134 DAG.getConstant(SimpleSize, DL,
3135 getShiftAmountTy(Lo.getValueType())));
3136 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
3137 // Compute the low part as N0.
3138 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
3139 return CombineTo(N, Lo, Hi);
3146 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
3147 if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU))
3150 EVT VT = N->getValueType(0);
3153 // If the type is twice as wide is legal, transform the mulhu to a wider
3154 // multiply plus a shift.
3155 if (VT.isSimple() && !VT.isVector()) {
3156 MVT Simple = VT.getSimpleVT();
3157 unsigned SimpleSize = Simple.getSizeInBits();
3158 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
3159 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
3160 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
3161 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
3162 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
3163 // Compute the high part as N1.
3164 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
3165 DAG.getConstant(SimpleSize, DL,
3166 getShiftAmountTy(Lo.getValueType())));
3167 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
3168 // Compute the low part as N0.
3169 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
3170 return CombineTo(N, Lo, Hi);
3177 SDValue DAGCombiner::visitSMULO(SDNode *N) {
3178 // (smulo x, 2) -> (saddo x, x)
3179 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
3180 if (C2->getAPIntValue() == 2)
3181 return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(),
3182 N->getOperand(0), N->getOperand(0));
3187 SDValue DAGCombiner::visitUMULO(SDNode *N) {
3188 // (umulo x, 2) -> (uaddo x, x)
3189 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
3190 if (C2->getAPIntValue() == 2)
3191 return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(),
3192 N->getOperand(0), N->getOperand(0));
3197 SDValue DAGCombiner::visitIMINMAX(SDNode *N) {
3198 SDValue N0 = N->getOperand(0);
3199 SDValue N1 = N->getOperand(1);
3200 EVT VT = N0.getValueType();
3204 if (SDValue FoldedVOp = SimplifyVBinOp(N))
3207 // fold (add c1, c2) -> c1+c2
3208 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
3209 ConstantSDNode *N1C = getAsNonOpaqueConstant(N1);
3211 return DAG.FoldConstantArithmetic(N->getOpcode(), SDLoc(N), VT, N0C, N1C);
3213 // canonicalize constant to RHS
3214 if (DAG.isConstantIntBuildVectorOrConstantInt(N0) &&
3215 !DAG.isConstantIntBuildVectorOrConstantInt(N1))
3216 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, N1, N0);
3221 /// If this is a binary operator with two operands of the same opcode, try to
3223 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
3224 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3225 EVT VT = N0.getValueType();
3226 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
3228 // Bail early if none of these transforms apply.
3229 if (N0.getNumOperands() == 0) return SDValue();
3231 // For each of OP in AND/OR/XOR:
3232 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
3233 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
3234 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
3235 // fold (OP (bswap x), (bswap y)) -> (bswap (OP x, y))
3236 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
3238 // do not sink logical op inside of a vector extend, since it may combine
3240 EVT Op0VT = N0.getOperand(0).getValueType();
3241 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
3242 N0.getOpcode() == ISD::SIGN_EXTEND ||
3243 N0.getOpcode() == ISD::BSWAP ||
3244 // Avoid infinite looping with PromoteIntBinOp.
3245 (N0.getOpcode() == ISD::ANY_EXTEND &&
3246 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
3247 (N0.getOpcode() == ISD::TRUNCATE &&
3248 (!TLI.isZExtFree(VT, Op0VT) ||
3249 !TLI.isTruncateFree(Op0VT, VT)) &&
3250 TLI.isTypeLegal(Op0VT))) &&
3252 Op0VT == N1.getOperand(0).getValueType() &&
3253 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
3254 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
3255 N0.getOperand(0).getValueType(),
3256 N0.getOperand(0), N1.getOperand(0));
3257 AddToWorklist(ORNode.getNode());
3258 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode);
3261 // For each of OP in SHL/SRL/SRA/AND...
3262 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
3263 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
3264 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
3265 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
3266 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
3267 N0.getOperand(1) == N1.getOperand(1)) {
3268 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
3269 N0.getOperand(0).getValueType(),
3270 N0.getOperand(0), N1.getOperand(0));
3271 AddToWorklist(ORNode.getNode());
3272 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
3273 ORNode, N0.getOperand(1));
3276 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
3277 // Only perform this optimization up until type legalization, before
3278 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
3279 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
3280 // we don't want to undo this promotion.
3281 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
3283 if ((N0.getOpcode() == ISD::BITCAST ||
3284 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) &&
3285 Level <= AfterLegalizeTypes) {
3286 SDValue In0 = N0.getOperand(0);
3287 SDValue In1 = N1.getOperand(0);
3288 EVT In0Ty = In0.getValueType();
3289 EVT In1Ty = In1.getValueType();
3291 // If both incoming values are integers, and the original types are the
3293 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
3294 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1);
3295 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op);
3296 AddToWorklist(Op.getNode());
3301 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
3302 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
3303 // If both shuffles use the same mask, and both shuffle within a single
3304 // vector, then it is worthwhile to move the swizzle after the operation.
3305 // The type-legalizer generates this pattern when loading illegal
3306 // vector types from memory. In many cases this allows additional shuffle
3308 // There are other cases where moving the shuffle after the xor/and/or
3309 // is profitable even if shuffles don't perform a swizzle.
3310 // If both shuffles use the same mask, and both shuffles have the same first
3311 // or second operand, then it might still be profitable to move the shuffle
3312 // after the xor/and/or operation.
3313 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) {
3314 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
3315 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
3317 assert(N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() &&
3318 "Inputs to shuffles are not the same type");
3320 // Check that both shuffles use the same mask. The masks are known to be of
3321 // the same length because the result vector type is the same.
3322 // Check also that shuffles have only one use to avoid introducing extra
3324 if (SVN0->hasOneUse() && SVN1->hasOneUse() &&
3325 SVN0->getMask().equals(SVN1->getMask())) {
3326 SDValue ShOp = N0->getOperand(1);
3328 // Don't try to fold this node if it requires introducing a
3329 // build vector of all zeros that might be illegal at this stage.
3330 if (N->getOpcode() == ISD::XOR && !ShOp.isUndef()) {
3332 ShOp = DAG.getConstant(0, SDLoc(N), VT);
3337 // (AND (shuf (A, C), shuf (B, C)) -> shuf (AND (A, B), C)
3338 // (OR (shuf (A, C), shuf (B, C)) -> shuf (OR (A, B), C)
3339 // (XOR (shuf (A, C), shuf (B, C)) -> shuf (XOR (A, B), V_0)
3340 if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) {
3341 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
3342 N0->getOperand(0), N1->getOperand(0));
3343 AddToWorklist(NewNode.getNode());
3344 return DAG.getVectorShuffle(VT, SDLoc(N), NewNode, ShOp,
3348 // Don't try to fold this node if it requires introducing a
3349 // build vector of all zeros that might be illegal at this stage.
3350 ShOp = N0->getOperand(0);
3351 if (N->getOpcode() == ISD::XOR && !ShOp.isUndef()) {
3353 ShOp = DAG.getConstant(0, SDLoc(N), VT);
3358 // (AND (shuf (C, A), shuf (C, B)) -> shuf (C, AND (A, B))
3359 // (OR (shuf (C, A), shuf (C, B)) -> shuf (C, OR (A, B))
3360 // (XOR (shuf (C, A), shuf (C, B)) -> shuf (V_0, XOR (A, B))
3361 if (N0->getOperand(0) == N1->getOperand(0) && ShOp.getNode()) {
3362 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
3363 N0->getOperand(1), N1->getOperand(1));
3364 AddToWorklist(NewNode.getNode());
3365 return DAG.getVectorShuffle(VT, SDLoc(N), ShOp, NewNode,
3374 /// Try to make (and/or setcc (LL, LR), setcc (RL, RR)) more efficient.
3375 SDValue DAGCombiner::foldLogicOfSetCCs(bool IsAnd, SDValue N0, SDValue N1,
3377 SDValue LL, LR, RL, RR, N0CC, N1CC;
3378 if (!isSetCCEquivalent(N0, LL, LR, N0CC) ||
3379 !isSetCCEquivalent(N1, RL, RR, N1CC))
3382 assert(N0.getValueType() == N1.getValueType() &&
3383 "Unexpected operand types for bitwise logic op");
3384 assert(LL.getValueType() == LR.getValueType() &&
3385 RL.getValueType() == RR.getValueType() &&
3386 "Unexpected operand types for setcc");
3388 // If we're here post-legalization or the logic op type is not i1, the logic
3389 // op type must match a setcc result type. Also, all folds require new
3390 // operations on the left and right operands, so those types must match.
3391 EVT VT = N0.getValueType();
3392 EVT OpVT = LL.getValueType();
3393 if (LegalOperations || VT != MVT::i1)
3394 if (VT != getSetCCResultType(OpVT))
3396 if (OpVT != RL.getValueType())
3399 ISD::CondCode CC0 = cast<CondCodeSDNode>(N0CC)->get();
3400 ISD::CondCode CC1 = cast<CondCodeSDNode>(N1CC)->get();
3401 bool IsInteger = OpVT.isInteger();
3402 if (LR == RR && CC0 == CC1 && IsInteger) {
3403 bool IsZero = isNullConstantOrNullSplatConstant(LR);
3404 bool IsNeg1 = isAllOnesConstantOrAllOnesSplatConstant(LR);
3407 bool AndEqZero = IsAnd && CC1 == ISD::SETEQ && IsZero;
3408 // All sign bits clear?
3409 bool AndGtNeg1 = IsAnd && CC1 == ISD::SETGT && IsNeg1;
3411 bool OrNeZero = !IsAnd && CC1 == ISD::SETNE && IsZero;
3412 // Any sign bits set?
3413 bool OrLtZero = !IsAnd && CC1 == ISD::SETLT && IsZero;
3415 // (and (seteq X, 0), (seteq Y, 0)) --> (seteq (or X, Y), 0)
3416 // (and (setgt X, -1), (setgt Y, -1)) --> (setgt (or X, Y), -1)
3417 // (or (setne X, 0), (setne Y, 0)) --> (setne (or X, Y), 0)
3418 // (or (setlt X, 0), (setlt Y, 0)) --> (setlt (or X, Y), 0)
3419 if (AndEqZero || AndGtNeg1 || OrNeZero || OrLtZero) {
3420 SDValue Or = DAG.getNode(ISD::OR, SDLoc(N0), OpVT, LL, RL);
3421 AddToWorklist(Or.getNode());
3422 return DAG.getSetCC(DL, VT, Or, LR, CC1);
3426 bool AndEqNeg1 = IsAnd && CC1 == ISD::SETEQ && IsNeg1;
3427 // All sign bits set?
3428 bool AndLtZero = IsAnd && CC1 == ISD::SETLT && IsZero;
3430 bool OrNeNeg1 = !IsAnd && CC1 == ISD::SETNE && IsNeg1;
3431 // Any sign bits clear?
3432 bool OrGtNeg1 = !IsAnd && CC1 == ISD::SETGT && IsNeg1;
3434 // (and (seteq X, -1), (seteq Y, -1)) --> (seteq (and X, Y), -1)
3435 // (and (setlt X, 0), (setlt Y, 0)) --> (setlt (and X, Y), 0)
3436 // (or (setne X, -1), (setne Y, -1)) --> (setne (and X, Y), -1)
3437 // (or (setgt X, -1), (setgt Y -1)) --> (setgt (and X, Y), -1)
3438 if (AndEqNeg1 || AndLtZero || OrNeNeg1 || OrGtNeg1) {
3439 SDValue And = DAG.getNode(ISD::AND, SDLoc(N0), OpVT, LL, RL);
3440 AddToWorklist(And.getNode());
3441 return DAG.getSetCC(DL, VT, And, LR, CC1);
3445 // TODO: What is the 'or' equivalent of this fold?
3446 // (and (setne X, 0), (setne X, -1)) --> (setuge (add X, 1), 2)
3447 if (IsAnd && LL == RL && CC0 == CC1 && IsInteger && CC0 == ISD::SETNE &&
3448 ((isNullConstant(LR) && isAllOnesConstant(RR)) ||
3449 (isAllOnesConstant(LR) && isNullConstant(RR)))) {
3450 SDValue One = DAG.getConstant(1, DL, OpVT);
3451 SDValue Two = DAG.getConstant(2, DL, OpVT);
3452 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0), OpVT, LL, One);
3453 AddToWorklist(Add.getNode());
3454 return DAG.getSetCC(DL, VT, Add, Two, ISD::SETUGE);
3457 // Try more general transforms if the predicates match and the only user of
3458 // the compares is the 'and' or 'or'.
3459 if (IsInteger && TLI.convertSetCCLogicToBitwiseLogic(OpVT) && CC0 == CC1 &&
3460 N0.hasOneUse() && N1.hasOneUse()) {
3461 // and (seteq A, B), (seteq C, D) --> seteq (or (xor A, B), (xor C, D)), 0
3462 // or (setne A, B), (setne C, D) --> setne (or (xor A, B), (xor C, D)), 0
3463 if ((IsAnd && CC1 == ISD::SETEQ) || (!IsAnd && CC1 == ISD::SETNE)) {
3464 SDValue XorL = DAG.getNode(ISD::XOR, SDLoc(N0), OpVT, LL, LR);
3465 SDValue XorR = DAG.getNode(ISD::XOR, SDLoc(N1), OpVT, RL, RR);
3466 SDValue Or = DAG.getNode(ISD::OR, DL, OpVT, XorL, XorR);
3467 SDValue Zero = DAG.getConstant(0, DL, OpVT);
3468 return DAG.getSetCC(DL, VT, Or, Zero, CC1);
3472 // Canonicalize equivalent operands to LL == RL.
3473 if (LL == RR && LR == RL) {
3474 CC1 = ISD::getSetCCSwappedOperands(CC1);
3478 // (and (setcc X, Y, CC0), (setcc X, Y, CC1)) --> (setcc X, Y, NewCC)
3479 // (or (setcc X, Y, CC0), (setcc X, Y, CC1)) --> (setcc X, Y, NewCC)
3480 if (LL == RL && LR == RR) {
3481 ISD::CondCode NewCC = IsAnd ? ISD::getSetCCAndOperation(CC0, CC1, IsInteger)
3482 : ISD::getSetCCOrOperation(CC0, CC1, IsInteger);
3483 if (NewCC != ISD::SETCC_INVALID &&
3484 (!LegalOperations ||
3485 (TLI.isCondCodeLegal(NewCC, LL.getSimpleValueType()) &&
3486 TLI.isOperationLegal(ISD::SETCC, OpVT))))
3487 return DAG.getSetCC(DL, VT, LL, LR, NewCC);
3493 /// This contains all DAGCombine rules which reduce two values combined by
3494 /// an And operation to a single value. This makes them reusable in the context
3495 /// of visitSELECT(). Rules involving constants are not included as
3496 /// visitSELECT() already handles those cases.
3497 SDValue DAGCombiner::visitANDLike(SDValue N0, SDValue N1, SDNode *N) {
3498 EVT VT = N1.getValueType();
3501 // fold (and x, undef) -> 0
3502 if (N0.isUndef() || N1.isUndef())
3503 return DAG.getConstant(0, DL, VT);
3505 if (SDValue V = foldLogicOfSetCCs(true, N0, N1, DL))
3508 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
3509 VT.getSizeInBits() <= 64) {
3510 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3511 APInt ADDC = ADDI->getAPIntValue();
3512 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
3513 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
3514 // immediate for an add, but it is legal if its top c2 bits are set,
3515 // transform the ADD so the immediate doesn't need to be materialized
3517 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
3518 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
3519 SRLI->getZExtValue());
3520 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
3522 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
3525 DAG.getNode(ISD::ADD, DL0, VT,
3526 N0.getOperand(0), DAG.getConstant(ADDC, DL, VT));
3527 CombineTo(N0.getNode(), NewAdd);
3528 // Return N so it doesn't get rechecked!
3529 return SDValue(N, 0);
3537 // Reduce bit extract of low half of an integer to the narrower type.
3538 // (and (srl i64:x, K), KMask) ->
3539 // (i64 zero_extend (and (srl (i32 (trunc i64:x)), K)), KMask)
3540 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
3541 if (ConstantSDNode *CAnd = dyn_cast<ConstantSDNode>(N1)) {
3542 if (ConstantSDNode *CShift = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3543 unsigned Size = VT.getSizeInBits();
3544 const APInt &AndMask = CAnd->getAPIntValue();
3545 unsigned ShiftBits = CShift->getZExtValue();
3547 // Bail out, this node will probably disappear anyway.
3551 unsigned MaskBits = AndMask.countTrailingOnes();
3552 EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), Size / 2);
3554 if (AndMask.isMask() &&
3555 // Required bits must not span the two halves of the integer and
3556 // must fit in the half size type.
3557 (ShiftBits + MaskBits <= Size / 2) &&
3558 TLI.isNarrowingProfitable(VT, HalfVT) &&
3559 TLI.isTypeDesirableForOp(ISD::AND, HalfVT) &&
3560 TLI.isTypeDesirableForOp(ISD::SRL, HalfVT) &&
3561 TLI.isTruncateFree(VT, HalfVT) &&
3562 TLI.isZExtFree(HalfVT, VT)) {
3563 // The isNarrowingProfitable is to avoid regressions on PPC and
3564 // AArch64 which match a few 64-bit bit insert / bit extract patterns
3565 // on downstream users of this. Those patterns could probably be
3566 // extended to handle extensions mixed in.
3569 assert(MaskBits <= Size);
3571 // Extracting the highest bit of the low half.
3572 EVT ShiftVT = TLI.getShiftAmountTy(HalfVT, DAG.getDataLayout());
3573 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SL, HalfVT,
3576 SDValue NewMask = DAG.getConstant(AndMask.trunc(Size / 2), SL, HalfVT);
3577 SDValue ShiftK = DAG.getConstant(ShiftBits, SL, ShiftVT);
3578 SDValue Shift = DAG.getNode(ISD::SRL, SL, HalfVT, Trunc, ShiftK);
3579 SDValue And = DAG.getNode(ISD::AND, SL, HalfVT, Shift, NewMask);
3580 return DAG.getNode(ISD::ZERO_EXTEND, SL, VT, And);
3589 bool DAGCombiner::isAndLoadExtLoad(ConstantSDNode *AndC, LoadSDNode *LoadN,
3590 EVT LoadResultTy, EVT &ExtVT, EVT &LoadedVT,
3592 uint32_t ActiveBits = AndC->getAPIntValue().getActiveBits();
3594 if (ActiveBits == 0 || !AndC->getAPIntValue().isMask(ActiveBits))
3597 ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
3598 LoadedVT = LoadN->getMemoryVT();
3600 if (ExtVT == LoadedVT &&
3601 (!LegalOperations ||
3602 TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy, ExtVT))) {
3603 // ZEXTLOAD will match without needing to change the size of the value being
3609 // Do not change the width of a volatile load.
3610 if (LoadN->isVolatile())
3613 // Do not generate loads of non-round integer types since these can
3614 // be expensive (and would be wrong if the type is not byte sized).
3615 if (!LoadedVT.bitsGT(ExtVT) || !ExtVT.isRound())
3618 if (LegalOperations &&
3619 !TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy, ExtVT))
3622 if (!TLI.shouldReduceLoadWidth(LoadN, ISD::ZEXTLOAD, ExtVT))
3629 SDValue DAGCombiner::visitAND(SDNode *N) {
3630 SDValue N0 = N->getOperand(0);
3631 SDValue N1 = N->getOperand(1);
3632 EVT VT = N1.getValueType();
3639 if (VT.isVector()) {
3640 if (SDValue FoldedVOp = SimplifyVBinOp(N))
3643 // fold (and x, 0) -> 0, vector edition
3644 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3645 // do not return N0, because undef node may exist in N0
3646 return DAG.getConstant(APInt::getNullValue(N0.getScalarValueSizeInBits()),
3647 SDLoc(N), N0.getValueType());
3648 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3649 // do not return N1, because undef node may exist in N1
3650 return DAG.getConstant(APInt::getNullValue(N1.getScalarValueSizeInBits()),
3651 SDLoc(N), N1.getValueType());
3653 // fold (and x, -1) -> x, vector edition
3654 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3656 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3660 // fold (and c1, c2) -> c1&c2
3661 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
3662 ConstantSDNode *N1C = isConstOrConstSplat(N1);
3663 if (N0C && N1C && !N1C->isOpaque())
3664 return DAG.FoldConstantArithmetic(ISD::AND, SDLoc(N), VT, N0C, N1C);
3665 // canonicalize constant to RHS
3666 if (DAG.isConstantIntBuildVectorOrConstantInt(N0) &&
3667 !DAG.isConstantIntBuildVectorOrConstantInt(N1))
3668 return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0);
3669 // fold (and x, -1) -> x
3670 if (isAllOnesConstant(N1))
3672 // if (and x, c) is known to be zero, return 0
3673 unsigned BitWidth = VT.getScalarSizeInBits();
3674 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
3675 APInt::getAllOnesValue(BitWidth)))
3676 return DAG.getConstant(0, SDLoc(N), VT);
3678 if (SDValue NewSel = foldBinOpIntoSelect(N))
3682 if (SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1))
3684 // fold (and (or x, C), D) -> D if (C & D) == D
3685 if (N1C && N0.getOpcode() == ISD::OR)
3686 if (ConstantSDNode *ORI = isConstOrConstSplat(N0.getOperand(1)))
3687 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
3689 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
3690 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
3691 SDValue N0Op0 = N0.getOperand(0);
3692 APInt Mask = ~N1C->getAPIntValue();
3693 Mask = Mask.trunc(N0Op0.getScalarValueSizeInBits());
3694 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
3695 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
3696 N0.getValueType(), N0Op0);
3698 // Replace uses of the AND with uses of the Zero extend node.
3701 // We actually want to replace all uses of the any_extend with the
3702 // zero_extend, to avoid duplicating things. This will later cause this
3703 // AND to be folded.
3704 CombineTo(N0.getNode(), Zext);
3705 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3708 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
3709 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
3710 // already be zero by virtue of the width of the base type of the load.
3712 // the 'X' node here can either be nothing or an extract_vector_elt to catch
3714 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
3715 N0.getValueSizeInBits() == N0.getOperand(0).getScalarValueSizeInBits() &&
3716 N0.getOperand(0).getOpcode() == ISD::LOAD &&
3717 N0.getOperand(0).getResNo() == 0) ||
3718 (N0.getOpcode() == ISD::LOAD && N0.getResNo() == 0)) {
3719 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
3720 N0 : N0.getOperand(0) );
3722 // Get the constant (if applicable) the zero'th operand is being ANDed with.
3723 // This can be a pure constant or a vector splat, in which case we treat the
3724 // vector as a scalar and use the splat value.
3725 APInt Constant = APInt::getNullValue(1);
3726 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
3727 Constant = C->getAPIntValue();
3728 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
3729 APInt SplatValue, SplatUndef;
3730 unsigned SplatBitSize;
3732 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
3733 SplatBitSize, HasAnyUndefs);
3735 // Undef bits can contribute to a possible optimisation if set, so
3737 SplatValue |= SplatUndef;
3739 // The splat value may be something like "0x00FFFFFF", which means 0 for
3740 // the first vector value and FF for the rest, repeating. We need a mask
3741 // that will apply equally to all members of the vector, so AND all the
3742 // lanes of the constant together.
3743 EVT VT = Vector->getValueType(0);
3744 unsigned BitWidth = VT.getScalarSizeInBits();
3746 // If the splat value has been compressed to a bitlength lower
3747 // than the size of the vector lane, we need to re-expand it to
3749 if (BitWidth > SplatBitSize)
3750 for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
3751 SplatBitSize < BitWidth;
3752 SplatBitSize = SplatBitSize * 2)
3753 SplatValue |= SplatValue.shl(SplatBitSize);
3755 // Make sure that variable 'Constant' is only set if 'SplatBitSize' is a
3756 // multiple of 'BitWidth'. Otherwise, we could propagate a wrong value.
3757 if (SplatBitSize % BitWidth == 0) {
3758 Constant = APInt::getAllOnesValue(BitWidth);
3759 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
3760 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
3765 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
3766 // actually legal and isn't going to get expanded, else this is a false
3768 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
3769 Load->getValueType(0),
3770 Load->getMemoryVT());
3772 // Resize the constant to the same size as the original memory access before
3773 // extension. If it is still the AllOnesValue then this AND is completely
3775 Constant = Constant.zextOrTrunc(Load->getMemoryVT().getScalarSizeInBits());
3778 switch (Load->getExtensionType()) {
3779 default: B = false; break;
3780 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
3782 case ISD::NON_EXTLOAD: B = true; break;
3785 if (B && Constant.isAllOnesValue()) {
3786 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
3787 // preserve semantics once we get rid of the AND.
3788 SDValue NewLoad(Load, 0);
3790 // Fold the AND away. NewLoad may get replaced immediately.
3791 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
3793 if (Load->getExtensionType() == ISD::EXTLOAD) {
3794 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
3795 Load->getValueType(0), SDLoc(Load),
3796 Load->getChain(), Load->getBasePtr(),
3797 Load->getOffset(), Load->getMemoryVT(),
3798 Load->getMemOperand());
3799 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
3800 if (Load->getNumValues() == 3) {
3801 // PRE/POST_INC loads have 3 values.
3802 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
3803 NewLoad.getValue(2) };
3804 CombineTo(Load, To, 3, true);
3806 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
3810 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3814 // fold (and (load x), 255) -> (zextload x, i8)
3815 // fold (and (extload x, i16), 255) -> (zextload x, i8)
3816 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
3817 if (!VT.isVector() && N1C && (N0.getOpcode() == ISD::LOAD ||
3818 (N0.getOpcode() == ISD::ANY_EXTEND &&
3819 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
3820 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
3821 LoadSDNode *LN0 = HasAnyExt
3822 ? cast<LoadSDNode>(N0.getOperand(0))
3823 : cast<LoadSDNode>(N0);
3824 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
3825 LN0->isUnindexed() && N0.hasOneUse() && SDValue(LN0, 0).hasOneUse()) {
3826 auto NarrowLoad = false;
3827 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
3828 EVT ExtVT, LoadedVT;
3829 if (isAndLoadExtLoad(N1C, LN0, LoadResultTy, ExtVT, LoadedVT,
3833 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
3834 LN0->getChain(), LN0->getBasePtr(), ExtVT,
3835 LN0->getMemOperand());
3837 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
3838 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3840 EVT PtrType = LN0->getOperand(1).getValueType();
3842 unsigned Alignment = LN0->getAlignment();
3843 SDValue NewPtr = LN0->getBasePtr();
3845 // For big endian targets, we need to add an offset to the pointer
3846 // to load the correct bytes. For little endian systems, we merely
3847 // need to read fewer bytes from the same pointer.
3848 if (DAG.getDataLayout().isBigEndian()) {
3849 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
3850 unsigned EVTStoreBytes = ExtVT.getStoreSize();
3851 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
3853 NewPtr = DAG.getNode(ISD::ADD, DL, PtrType,
3854 NewPtr, DAG.getConstant(PtrOff, DL, PtrType));
3855 Alignment = MinAlign(Alignment, PtrOff);
3858 AddToWorklist(NewPtr.getNode());
3860 SDValue Load = DAG.getExtLoad(
3861 ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy, LN0->getChain(), NewPtr,
3862 LN0->getPointerInfo(), ExtVT, Alignment,
3863 LN0->getMemOperand()->getFlags(), LN0->getAAInfo());
3865 CombineTo(LN0, Load, Load.getValue(1));
3866 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3872 if (SDValue Combined = visitANDLike(N0, N1, N))
3875 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
3876 if (N0.getOpcode() == N1.getOpcode())
3877 if (SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N))
3880 // Masking the negated extension of a boolean is just the zero-extended
3882 // and (sub 0, zext(bool X)), 1 --> zext(bool X)
3883 // and (sub 0, sext(bool X)), 1 --> zext(bool X)
3885 // Note: the SimplifyDemandedBits fold below can make an information-losing
3886 // transform, and then we have no way to find this better fold.
3887 if (N1C && N1C->isOne() && N0.getOpcode() == ISD::SUB) {
3888 ConstantSDNode *SubLHS = isConstOrConstSplat(N0.getOperand(0));
3889 SDValue SubRHS = N0.getOperand(1);
3890 if (SubLHS && SubLHS->isNullValue()) {
3891 if (SubRHS.getOpcode() == ISD::ZERO_EXTEND &&
3892 SubRHS.getOperand(0).getScalarValueSizeInBits() == 1)
3894 if (SubRHS.getOpcode() == ISD::SIGN_EXTEND &&
3895 SubRHS.getOperand(0).getScalarValueSizeInBits() == 1)
3896 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, SubRHS.getOperand(0));
3900 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
3901 // fold (and (sra)) -> (and (srl)) when possible.
3902 if (SimplifyDemandedBits(SDValue(N, 0)))
3903 return SDValue(N, 0);
3905 // fold (zext_inreg (extload x)) -> (zextload x)
3906 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
3907 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3908 EVT MemVT = LN0->getMemoryVT();
3909 // If we zero all the possible extended bits, then we can turn this into
3910 // a zextload if we are running before legalize or the operation is legal.
3911 unsigned BitWidth = N1.getScalarValueSizeInBits();
3912 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
3913 BitWidth - MemVT.getScalarSizeInBits())) &&
3914 ((!LegalOperations && !LN0->isVolatile()) ||
3915 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) {
3916 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
3917 LN0->getChain(), LN0->getBasePtr(),
3918 MemVT, LN0->getMemOperand());
3920 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
3921 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3924 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
3925 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
3927 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3928 EVT MemVT = LN0->getMemoryVT();
3929 // If we zero all the possible extended bits, then we can turn this into
3930 // a zextload if we are running before legalize or the operation is legal.
3931 unsigned BitWidth = N1.getScalarValueSizeInBits();
3932 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
3933 BitWidth - MemVT.getScalarSizeInBits())) &&
3934 ((!LegalOperations && !LN0->isVolatile()) ||
3935 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) {
3936 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
3937 LN0->getChain(), LN0->getBasePtr(),
3938 MemVT, LN0->getMemOperand());
3940 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
3941 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3944 // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const)
3945 if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) {
3946 if (SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
3947 N0.getOperand(1), false))
3954 /// Match (a >> 8) | (a << 8) as (bswap a) >> 16.
3955 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
3956 bool DemandHighBits) {
3957 if (!LegalOperations)
3960 EVT VT = N->getValueType(0);
3961 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
3963 if (!TLI.isOperationLegalOrCustom(ISD::BSWAP, VT))
3966 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
3967 bool LookPassAnd0 = false;
3968 bool LookPassAnd1 = false;
3969 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
3971 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
3973 if (N0.getOpcode() == ISD::AND) {
3974 if (!N0.getNode()->hasOneUse())
3976 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3977 if (!N01C || N01C->getZExtValue() != 0xFF00)
3979 N0 = N0.getOperand(0);
3980 LookPassAnd0 = true;
3983 if (N1.getOpcode() == ISD::AND) {
3984 if (!N1.getNode()->hasOneUse())
3986 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3987 if (!N11C || N11C->getZExtValue() != 0xFF)
3989 N1 = N1.getOperand(0);
3990 LookPassAnd1 = true;
3993 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
3995 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
3997 if (!N0.getNode()->hasOneUse() || !N1.getNode()->hasOneUse())
4000 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4001 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
4004 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
4007 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
4008 SDValue N00 = N0->getOperand(0);
4009 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
4010 if (!N00.getNode()->hasOneUse())
4012 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
4013 if (!N001C || N001C->getZExtValue() != 0xFF)
4015 N00 = N00.getOperand(0);
4016 LookPassAnd0 = true;
4019 SDValue N10 = N1->getOperand(0);
4020 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
4021 if (!N10.getNode()->hasOneUse())
4023 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
4024 if (!N101C || N101C->getZExtValue() != 0xFF00)
4026 N10 = N10.getOperand(0);
4027 LookPassAnd1 = true;
4033 // Make sure everything beyond the low halfword gets set to zero since the SRL
4034 // 16 will clear the top bits.
4035 unsigned OpSizeInBits = VT.getSizeInBits();
4036 if (DemandHighBits && OpSizeInBits > 16) {
4037 // If the left-shift isn't masked out then the only way this is a bswap is
4038 // if all bits beyond the low 8 are 0. In that case the entire pattern
4039 // reduces to a left shift anyway: leave it for other parts of the combiner.
4043 // However, if the right shift isn't masked out then it might be because
4044 // it's not needed. See if we can spot that too.
4045 if (!LookPassAnd1 &&
4046 !DAG.MaskedValueIsZero(
4047 N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16)))
4051 SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00);
4052 if (OpSizeInBits > 16) {
4054 Res = DAG.getNode(ISD::SRL, DL, VT, Res,
4055 DAG.getConstant(OpSizeInBits - 16, DL,
4056 getShiftAmountTy(VT)));
4061 /// Return true if the specified node is an element that makes up a 32-bit
4062 /// packed halfword byteswap.
4063 /// ((x & 0x000000ff) << 8) |
4064 /// ((x & 0x0000ff00) >> 8) |
4065 /// ((x & 0x00ff0000) << 8) |
4066 /// ((x & 0xff000000) >> 8)
4067 static bool isBSwapHWordElement(SDValue N, MutableArrayRef<SDNode *> Parts) {
4068 if (!N.getNode()->hasOneUse())
4071 unsigned Opc = N.getOpcode();
4072 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
4075 SDValue N0 = N.getOperand(0);
4076 unsigned Opc0 = N0.getOpcode();
4077 if (Opc0 != ISD::AND && Opc0 != ISD::SHL && Opc0 != ISD::SRL)
4080 ConstantSDNode *N1C = nullptr;
4081 // SHL or SRL: look upstream for AND mask operand
4082 if (Opc == ISD::AND)
4083 N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
4084 else if (Opc0 == ISD::AND)
4085 N1C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4089 unsigned MaskByteOffset;
4090 switch (N1C->getZExtValue()) {
4093 case 0xFF: MaskByteOffset = 0; break;
4094 case 0xFF00: MaskByteOffset = 1; break;
4095 case 0xFF0000: MaskByteOffset = 2; break;
4096 case 0xFF000000: MaskByteOffset = 3; break;
4099 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
4100 if (Opc == ISD::AND) {
4101 if (MaskByteOffset == 0 || MaskByteOffset == 2) {
4103 // (x >> 8) & 0xff0000
4104 if (Opc0 != ISD::SRL)
4106 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4107 if (!C || C->getZExtValue() != 8)
4110 // (x << 8) & 0xff00
4111 // (x << 8) & 0xff000000
4112 if (Opc0 != ISD::SHL)
4114 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4115 if (!C || C->getZExtValue() != 8)
4118 } else if (Opc == ISD::SHL) {
4120 // (x & 0xff0000) << 8
4121 if (MaskByteOffset != 0 && MaskByteOffset != 2)
4123 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
4124 if (!C || C->getZExtValue() != 8)
4126 } else { // Opc == ISD::SRL
4127 // (x & 0xff00) >> 8
4128 // (x & 0xff000000) >> 8
4129 if (MaskByteOffset != 1 && MaskByteOffset != 3)
4131 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
4132 if (!C || C->getZExtValue() != 8)
4136 if (Parts[MaskByteOffset])
4139 Parts[MaskByteOffset] = N0.getOperand(0).getNode();
4143 /// Match a 32-bit packed halfword bswap. That is
4144 /// ((x & 0x000000ff) << 8) |
4145 /// ((x & 0x0000ff00) >> 8) |
4146 /// ((x & 0x00ff0000) << 8) |
4147 /// ((x & 0xff000000) >> 8)
4148 /// => (rotl (bswap x), 16)
4149 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
4150 if (!LegalOperations)
4153 EVT VT = N->getValueType(0);
4156 if (!TLI.isOperationLegalOrCustom(ISD::BSWAP, VT))
4160 // (or (or (and), (and)), (or (and), (and)))
4161 // (or (or (or (and), (and)), (and)), (and))
4162 if (N0.getOpcode() != ISD::OR)
4164 SDValue N00 = N0.getOperand(0);
4165 SDValue N01 = N0.getOperand(1);
4166 SDNode *Parts[4] = {};
4168 if (N1.getOpcode() == ISD::OR &&
4169 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) {
4170 // (or (or (and), (and)), (or (and), (and)))
4171 if (!isBSwapHWordElement(N00, Parts))
4174 if (!isBSwapHWordElement(N01, Parts))
4176 SDValue N10 = N1.getOperand(0);
4177 if (!isBSwapHWordElement(N10, Parts))
4179 SDValue N11 = N1.getOperand(1);
4180 if (!isBSwapHWordElement(N11, Parts))
4183 // (or (or (or (and), (and)), (and)), (and))
4184 if (!isBSwapHWordElement(N1, Parts))
4186 if (!isBSwapHWordElement(N01, Parts))
4188 if (N00.getOpcode() != ISD::OR)
4190 SDValue N000 = N00.getOperand(0);
4191 if (!isBSwapHWordElement(N000, Parts))
4193 SDValue N001 = N00.getOperand(1);
4194 if (!isBSwapHWordElement(N001, Parts))
4198 // Make sure the parts are all coming from the same node.
4199 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
4203 SDValue BSwap = DAG.getNode(ISD::BSWAP, DL, VT,
4204 SDValue(Parts[0], 0));
4206 // Result of the bswap should be rotated by 16. If it's not legal, then
4207 // do (x << 16) | (x >> 16).
4208 SDValue ShAmt = DAG.getConstant(16, DL, getShiftAmountTy(VT));
4209 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
4210 return DAG.getNode(ISD::ROTL, DL, VT, BSwap, ShAmt);
4211 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
4212 return DAG.getNode(ISD::ROTR, DL, VT, BSwap, ShAmt);
4213 return DAG.getNode(ISD::OR, DL, VT,
4214 DAG.getNode(ISD::SHL, DL, VT, BSwap, ShAmt),
4215 DAG.getNode(ISD::SRL, DL, VT, BSwap, ShAmt));
4218 /// This contains all DAGCombine rules which reduce two values combined by
4219 /// an Or operation to a single value \see visitANDLike().
4220 SDValue DAGCombiner::visitORLike(SDValue N0, SDValue N1, SDNode *N) {
4221 EVT VT = N1.getValueType();
4224 // fold (or x, undef) -> -1
4225 if (!LegalOperations && (N0.isUndef() || N1.isUndef()))
4226 return DAG.getAllOnesConstant(DL, VT);
4228 if (SDValue V = foldLogicOfSetCCs(false, N0, N1, DL))
4231 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
4232 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == ISD::AND &&
4233 // Don't increase # computations.
4234 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
4235 // We can only do this xform if we know that bits from X that are set in C2
4236 // but not in C1 are already zero. Likewise for Y.
4237 if (const ConstantSDNode *N0O1C =
4238 getAsNonOpaqueConstant(N0.getOperand(1))) {
4239 if (const ConstantSDNode *N1O1C =
4240 getAsNonOpaqueConstant(N1.getOperand(1))) {
4241 // We can only do this xform if we know that bits from X that are set in
4242 // C2 but not in C1 are already zero. Likewise for Y.
4243 const APInt &LHSMask = N0O1C->getAPIntValue();
4244 const APInt &RHSMask = N1O1C->getAPIntValue();
4246 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
4247 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
4248 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
4249 N0.getOperand(0), N1.getOperand(0));
4250 return DAG.getNode(ISD::AND, DL, VT, X,
4251 DAG.getConstant(LHSMask | RHSMask, DL, VT));
4257 // (or (and X, M), (and X, N)) -> (and X, (or M, N))
4258 if (N0.getOpcode() == ISD::AND &&
4259 N1.getOpcode() == ISD::AND &&
4260 N0.getOperand(0) == N1.getOperand(0) &&
4261 // Don't increase # computations.
4262 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
4263 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
4264 N0.getOperand(1), N1.getOperand(1));
4265 return DAG.getNode(ISD::AND, DL, VT, N0.getOperand(0), X);
4271 SDValue DAGCombiner::visitOR(SDNode *N) {
4272 SDValue N0 = N->getOperand(0);
4273 SDValue N1 = N->getOperand(1);
4274 EVT VT = N1.getValueType();
4281 if (VT.isVector()) {
4282 if (SDValue FoldedVOp = SimplifyVBinOp(N))
4285 // fold (or x, 0) -> x, vector edition
4286 if (ISD::isBuildVectorAllZeros(N0.getNode()))
4288 if (ISD::isBuildVectorAllZeros(N1.getNode()))
4291 // fold (or x, -1) -> -1, vector edition
4292 if (ISD::isBuildVectorAllOnes(N0.getNode()))
4293 // do not return N0, because undef node may exist in N0
4294 return DAG.getAllOnesConstant(SDLoc(N), N0.getValueType());
4295 if (ISD::isBuildVectorAllOnes(N1.getNode()))
4296 // do not return N1, because undef node may exist in N1
4297 return DAG.getAllOnesConstant(SDLoc(N), N1.getValueType());
4299 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask)
4300 // Do this only if the resulting shuffle is legal.
4301 if (isa<ShuffleVectorSDNode>(N0) &&
4302 isa<ShuffleVectorSDNode>(N1) &&
4303 // Avoid folding a node with illegal type.
4304 TLI.isTypeLegal(VT)) {
4305 bool ZeroN00 = ISD::isBuildVectorAllZeros(N0.getOperand(0).getNode());
4306 bool ZeroN01 = ISD::isBuildVectorAllZeros(N0.getOperand(1).getNode());
4307 bool ZeroN10 = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode());
4308 bool ZeroN11 = ISD::isBuildVectorAllZeros(N1.getOperand(1).getNode());
4309 // Ensure both shuffles have a zero input.
4310 if ((ZeroN00 != ZeroN01) && (ZeroN10 != ZeroN11)) {
4311 assert((!ZeroN00 || !ZeroN01) && "Both inputs zero!");
4312 assert((!ZeroN10 || !ZeroN11) && "Both inputs zero!");
4313 const ShuffleVectorSDNode *SV0 = cast<ShuffleVectorSDNode>(N0);
4314 const ShuffleVectorSDNode *SV1 = cast<ShuffleVectorSDNode>(N1);
4315 bool CanFold = true;
4316 int NumElts = VT.getVectorNumElements();
4317 SmallVector<int, 4> Mask(NumElts);
4319 for (int i = 0; i != NumElts; ++i) {
4320 int M0 = SV0->getMaskElt(i);
4321 int M1 = SV1->getMaskElt(i);
4323 // Determine if either index is pointing to a zero vector.
4324 bool M0Zero = M0 < 0 || (ZeroN00 == (M0 < NumElts));
4325 bool M1Zero = M1 < 0 || (ZeroN10 == (M1 < NumElts));
4327 // If one element is zero and the otherside is undef, keep undef.
4328 // This also handles the case that both are undef.
4329 if ((M0Zero && M1 < 0) || (M1Zero && M0 < 0)) {
4334 // Make sure only one of the elements is zero.
4335 if (M0Zero == M1Zero) {
4340 assert((M0 >= 0 || M1 >= 0) && "Undef index!");
4342 // We have a zero and non-zero element. If the non-zero came from
4343 // SV0 make the index a LHS index. If it came from SV1, make it
4344 // a RHS index. We need to mod by NumElts because we don't care
4345 // which operand it came from in the original shuffles.
4346 Mask[i] = M1Zero ? M0 % NumElts : (M1 % NumElts) + NumElts;
4350 SDValue NewLHS = ZeroN00 ? N0.getOperand(1) : N0.getOperand(0);
4351 SDValue NewRHS = ZeroN10 ? N1.getOperand(1) : N1.getOperand(0);
4353 bool LegalMask = TLI.isShuffleMaskLegal(Mask, VT);
4355 std::swap(NewLHS, NewRHS);
4356 ShuffleVectorSDNode::commuteMask(Mask);
4357 LegalMask = TLI.isShuffleMaskLegal(Mask, VT);
4361 return DAG.getVectorShuffle(VT, SDLoc(N), NewLHS, NewRHS, Mask);
4367 // fold (or c1, c2) -> c1|c2
4368 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
4369 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4370 if (N0C && N1C && !N1C->isOpaque())
4371 return DAG.FoldConstantArithmetic(ISD::OR, SDLoc(N), VT, N0C, N1C);
4372 // canonicalize constant to RHS
4373 if (DAG.isConstantIntBuildVectorOrConstantInt(N0) &&
4374 !DAG.isConstantIntBuildVectorOrConstantInt(N1))
4375 return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0);
4376 // fold (or x, 0) -> x
4377 if (isNullConstant(N1))
4379 // fold (or x, -1) -> -1
4380 if (isAllOnesConstant(N1))
4383 if (SDValue NewSel = foldBinOpIntoSelect(N))
4386 // fold (or x, c) -> c iff (x & ~c) == 0
4387 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
4390 if (SDValue Combined = visitORLike(N0, N1, N))
4393 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
4394 if (SDValue BSwap = MatchBSwapHWord(N, N0, N1))
4396 if (SDValue BSwap = MatchBSwapHWordLow(N, N0, N1))
4400 if (SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1))
4403 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
4404 // iff (c1 & c2) != 0.
4405 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse()) {
4406 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
4407 if (C1->getAPIntValue().intersects(N1C->getAPIntValue())) {
4409 DAG.FoldConstantArithmetic(ISD::OR, SDLoc(N1), VT, N1C, C1))
4411 ISD::AND, SDLoc(N), VT,
4412 DAG.getNode(ISD::OR, SDLoc(N0), VT, N0.getOperand(0), N1), COR);
4418 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
4419 if (N0.getOpcode() == N1.getOpcode())
4420 if (SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N))
4423 // See if this is some rotate idiom.
4424 if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N)))
4425 return SDValue(Rot, 0);
4427 if (SDValue Load = MatchLoadCombine(N))
4430 // Simplify the operands using demanded-bits information.
4431 if (SimplifyDemandedBits(SDValue(N, 0)))
4432 return SDValue(N, 0);
4437 /// Match "(X shl/srl V1) & V2" where V2 may not be present.
4438 bool DAGCombiner::MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
4439 if (Op.getOpcode() == ISD::AND) {
4440 if (DAG.isConstantIntBuildVectorOrConstantInt(Op.getOperand(1))) {
4441 Mask = Op.getOperand(1);
4442 Op = Op.getOperand(0);
4448 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
4456 // Return true if we can prove that, whenever Neg and Pos are both in the
4457 // range [0, EltSize), Neg == (Pos == 0 ? 0 : EltSize - Pos). This means that
4458 // for two opposing shifts shift1 and shift2 and a value X with OpBits bits:
4460 // (or (shift1 X, Neg), (shift2 X, Pos))
4462 // reduces to a rotate in direction shift2 by Pos or (equivalently) a rotate
4463 // in direction shift1 by Neg. The range [0, EltSize) means that we only need
4464 // to consider shift amounts with defined behavior.
4465 static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned EltSize) {
4466 // If EltSize is a power of 2 then:
4468 // (a) (Pos == 0 ? 0 : EltSize - Pos) == (EltSize - Pos) & (EltSize - 1)
4469 // (b) Neg == Neg & (EltSize - 1) whenever Neg is in [0, EltSize).
4471 // So if EltSize is a power of 2 and Neg is (and Neg', EltSize-1), we check
4472 // for the stronger condition:
4474 // Neg & (EltSize - 1) == (EltSize - Pos) & (EltSize - 1) [A]
4476 // for all Neg and Pos. Since Neg & (EltSize - 1) == Neg' & (EltSize - 1)
4477 // we can just replace Neg with Neg' for the rest of the function.
4479 // In other cases we check for the even stronger condition:
4481 // Neg == EltSize - Pos [B]
4483 // for all Neg and Pos. Note that the (or ...) then invokes undefined
4484 // behavior if Pos == 0 (and consequently Neg == EltSize).
4486 // We could actually use [A] whenever EltSize is a power of 2, but the
4487 // only extra cases that it would match are those uninteresting ones
4488 // where Neg and Pos are never in range at the same time. E.g. for
4489 // EltSize == 32, using [A] would allow a Neg of the form (sub 64, Pos)
4490 // as well as (sub 32, Pos), but:
4492 // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos))
4494 // always invokes undefined behavior for 32-bit X.
4496 // Below, Mask == EltSize - 1 when using [A] and is all-ones otherwise.
4497 unsigned MaskLoBits = 0;
4498 if (Neg.getOpcode() == ISD::AND && isPowerOf2_64(EltSize)) {
4499 if (ConstantSDNode *NegC = isConstOrConstSplat(Neg.getOperand(1))) {
4500 if (NegC->getAPIntValue() == EltSize - 1) {
4501 Neg = Neg.getOperand(0);
4502 MaskLoBits = Log2_64(EltSize);
4507 // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1.
4508 if (Neg.getOpcode() != ISD::SUB)
4510 ConstantSDNode *NegC = isConstOrConstSplat(Neg.getOperand(0));
4513 SDValue NegOp1 = Neg.getOperand(1);
4515 // On the RHS of [A], if Pos is Pos' & (EltSize - 1), just replace Pos with
4516 // Pos'. The truncation is redundant for the purpose of the equality.
4517 if (MaskLoBits && Pos.getOpcode() == ISD::AND)
4518 if (ConstantSDNode *PosC = isConstOrConstSplat(Pos.getOperand(1)))
4519 if (PosC->getAPIntValue() == EltSize - 1)
4520 Pos = Pos.getOperand(0);
4522 // The condition we need is now:
4524 // (NegC - NegOp1) & Mask == (EltSize - Pos) & Mask
4526 // If NegOp1 == Pos then we need:
4528 // EltSize & Mask == NegC & Mask
4530 // (because "x & Mask" is a truncation and distributes through subtraction).
4533 Width = NegC->getAPIntValue();
4535 // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC.
4536 // Then the condition we want to prove becomes:
4538 // (NegC - NegOp1) & Mask == (EltSize - (NegOp1 + PosC)) & Mask
4540 // which, again because "x & Mask" is a truncation, becomes:
4542 // NegC & Mask == (EltSize - PosC) & Mask
4543 // EltSize & Mask == (NegC + PosC) & Mask
4544 else if (Pos.getOpcode() == ISD::ADD && Pos.getOperand(0) == NegOp1) {
4545 if (ConstantSDNode *PosC = isConstOrConstSplat(Pos.getOperand(1)))
4546 Width = PosC->getAPIntValue() + NegC->getAPIntValue();
4552 // Now we just need to check that EltSize & Mask == Width & Mask.
4554 // EltSize & Mask is 0 since Mask is EltSize - 1.
4555 return Width.getLoBits(MaskLoBits) == 0;
4556 return Width == EltSize;
4559 // A subroutine of MatchRotate used once we have found an OR of two opposite
4560 // shifts of Shifted. If Neg == <operand size> - Pos then the OR reduces
4561 // to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the
4562 // former being preferred if supported. InnerPos and InnerNeg are Pos and
4563 // Neg with outer conversions stripped away.
4564 SDNode *DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos,
4565 SDValue Neg, SDValue InnerPos,
4566 SDValue InnerNeg, unsigned PosOpcode,
4567 unsigned NegOpcode, const SDLoc &DL) {
4568 // fold (or (shl x, (*ext y)),
4569 // (srl x, (*ext (sub 32, y)))) ->
4570 // (rotl x, y) or (rotr x, (sub 32, y))
4572 // fold (or (shl x, (*ext (sub 32, y))),
4573 // (srl x, (*ext y))) ->
4574 // (rotr x, y) or (rotl x, (sub 32, y))
4575 EVT VT = Shifted.getValueType();
4576 if (matchRotateSub(InnerPos, InnerNeg, VT.getScalarSizeInBits())) {
4577 bool HasPos = TLI.isOperationLegalOrCustom(PosOpcode, VT);
4578 return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted,
4579 HasPos ? Pos : Neg).getNode();
4585 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
4586 // idioms for rotate, and if the target supports rotation instructions, generate
4588 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, const SDLoc &DL) {
4589 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
4590 EVT VT = LHS.getValueType();
4591 if (!TLI.isTypeLegal(VT)) return nullptr;
4593 // The target must have at least one rotate flavor.
4594 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
4595 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
4596 if (!HasROTL && !HasROTR) return nullptr;
4598 // Match "(X shl/srl V1) & V2" where V2 may not be present.
4599 SDValue LHSShift; // The shift.
4600 SDValue LHSMask; // AND value if any.
4601 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
4602 return nullptr; // Not part of a rotate.
4604 SDValue RHSShift; // The shift.
4605 SDValue RHSMask; // AND value if any.
4606 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
4607 return nullptr; // Not part of a rotate.
4609 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
4610 return nullptr; // Not shifting the same value.
4612 if (LHSShift.getOpcode() == RHSShift.getOpcode())
4613 return nullptr; // Shifts must disagree.
4615 // Canonicalize shl to left side in a shl/srl pair.
4616 if (RHSShift.getOpcode() == ISD::SHL) {
4617 std::swap(LHS, RHS);
4618 std::swap(LHSShift, RHSShift);
4619 std::swap(LHSMask, RHSMask);
4622 unsigned EltSizeInBits = VT.getScalarSizeInBits();
4623 SDValue LHSShiftArg = LHSShift.getOperand(0);
4624 SDValue LHSShiftAmt = LHSShift.getOperand(1);
4625 SDValue RHSShiftArg = RHSShift.getOperand(0);
4626 SDValue RHSShiftAmt = RHSShift.getOperand(1);
4628 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
4629 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
4630 if (isConstOrConstSplat(LHSShiftAmt) && isConstOrConstSplat(RHSShiftAmt)) {
4631 uint64_t LShVal = isConstOrConstSplat(LHSShiftAmt)->getZExtValue();
4632 uint64_t RShVal = isConstOrConstSplat(RHSShiftAmt)->getZExtValue();
4633 if ((LShVal + RShVal) != EltSizeInBits)
4636 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
4637 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
4639 // If there is an AND of either shifted operand, apply it to the result.
4640 if (LHSMask.getNode() || RHSMask.getNode()) {
4641 SDValue Mask = DAG.getAllOnesConstant(DL, VT);
4643 if (LHSMask.getNode()) {
4644 APInt RHSBits = APInt::getLowBitsSet(EltSizeInBits, LShVal);
4645 Mask = DAG.getNode(ISD::AND, DL, VT, Mask,
4646 DAG.getNode(ISD::OR, DL, VT, LHSMask,
4647 DAG.getConstant(RHSBits, DL, VT)));
4649 if (RHSMask.getNode()) {
4650 APInt LHSBits = APInt::getHighBitsSet(EltSizeInBits, RShVal);
4651 Mask = DAG.getNode(ISD::AND, DL, VT, Mask,
4652 DAG.getNode(ISD::OR, DL, VT, RHSMask,
4653 DAG.getConstant(LHSBits, DL, VT)));
4656 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, Mask);
4659 return Rot.getNode();
4662 // If there is a mask here, and we have a variable shift, we can't be sure
4663 // that we're masking out the right stuff.
4664 if (LHSMask.getNode() || RHSMask.getNode())
4667 // If the shift amount is sign/zext/any-extended just peel it off.
4668 SDValue LExtOp0 = LHSShiftAmt;
4669 SDValue RExtOp0 = RHSShiftAmt;
4670 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
4671 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
4672 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
4673 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
4674 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
4675 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
4676 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
4677 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
4678 LExtOp0 = LHSShiftAmt.getOperand(0);
4679 RExtOp0 = RHSShiftAmt.getOperand(0);
4682 SDNode *TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt,
4683 LExtOp0, RExtOp0, ISD::ROTL, ISD::ROTR, DL);
4687 SDNode *TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt,
4688 RExtOp0, LExtOp0, ISD::ROTR, ISD::ROTL, DL);
4696 /// Helper struct to parse and store a memory address as base + index + offset.
4697 /// We ignore sign extensions when it is safe to do so.
4698 /// The following two expressions are not equivalent. To differentiate we need
4699 /// to store whether there was a sign extension involved in the index
4701 /// (load (i64 add (i64 copyfromreg %c)
4702 /// (i64 signextend (add (i8 load %index)
4706 /// (load (i64 add (i64 copyfromreg %c)
4707 /// (i64 signextend (i32 add (i32 signextend (i8 load %index))
4709 struct BaseIndexOffset {
4713 bool IsIndexSignExt;
4715 BaseIndexOffset() : Offset(0), IsIndexSignExt(false) {}
4717 BaseIndexOffset(SDValue Base, SDValue Index, int64_t Offset,
4718 bool IsIndexSignExt) :
4719 Base(Base), Index(Index), Offset(Offset), IsIndexSignExt(IsIndexSignExt) {}
4721 bool equalBaseIndex(const BaseIndexOffset &Other) {
4722 return Other.Base == Base && Other.Index == Index &&
4723 Other.IsIndexSignExt == IsIndexSignExt;
4726 /// Parses tree in Ptr for base, index, offset addresses.
4727 static BaseIndexOffset match(SDValue Ptr, SelectionDAG &DAG,
4728 int64_t PartialOffset = 0) {
4729 bool IsIndexSignExt = false;
4731 // Split up a folded GlobalAddress+Offset into its component parts.
4732 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Ptr))
4733 if (GA->getOpcode() == ISD::GlobalAddress && GA->getOffset() != 0) {
4734 return BaseIndexOffset(DAG.getGlobalAddress(GA->getGlobal(),
4736 GA->getValueType(0),
4737 /*Offset=*/PartialOffset,
4738 /*isTargetGA=*/false,
4739 GA->getTargetFlags()),
4745 // We only can pattern match BASE + INDEX + OFFSET. If Ptr is not an ADD
4746 // instruction, then it could be just the BASE or everything else we don't
4747 // know how to handle. Just use Ptr as BASE and give up.
4748 if (Ptr->getOpcode() != ISD::ADD)
4749 return BaseIndexOffset(Ptr, SDValue(), PartialOffset, IsIndexSignExt);
4751 // We know that we have at least an ADD instruction. Try to pattern match
4752 // the simple case of BASE + OFFSET.
4753 if (isa<ConstantSDNode>(Ptr->getOperand(1))) {
4754 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
4755 return match(Ptr->getOperand(0), DAG, Offset + PartialOffset);
4758 // Inside a loop the current BASE pointer is calculated using an ADD and a
4759 // MUL instruction. In this case Ptr is the actual BASE pointer.
4760 // (i64 add (i64 %array_ptr)
4761 // (i64 mul (i64 %induction_var)
4762 // (i64 %element_size)))
4763 if (Ptr->getOperand(1)->getOpcode() == ISD::MUL)
4764 return BaseIndexOffset(Ptr, SDValue(), PartialOffset, IsIndexSignExt);
4766 // Look at Base + Index + Offset cases.
4767 SDValue Base = Ptr->getOperand(0);
4768 SDValue IndexOffset = Ptr->getOperand(1);
4770 // Skip signextends.
4771 if (IndexOffset->getOpcode() == ISD::SIGN_EXTEND) {
4772 IndexOffset = IndexOffset->getOperand(0);
4773 IsIndexSignExt = true;
4776 // Either the case of Base + Index (no offset) or something else.
4777 if (IndexOffset->getOpcode() != ISD::ADD)
4778 return BaseIndexOffset(Base, IndexOffset, PartialOffset, IsIndexSignExt);
4780 // Now we have the case of Base + Index + offset.
4781 SDValue Index = IndexOffset->getOperand(0);
4782 SDValue Offset = IndexOffset->getOperand(1);
4784 if (!isa<ConstantSDNode>(Offset))
4785 return BaseIndexOffset(Ptr, SDValue(), PartialOffset, IsIndexSignExt);
4787 // Ignore signextends.
4788 if (Index->getOpcode() == ISD::SIGN_EXTEND) {
4789 Index = Index->getOperand(0);
4790 IsIndexSignExt = true;
4791 } else IsIndexSignExt = false;
4793 int64_t Off = cast<ConstantSDNode>(Offset)->getSExtValue();
4794 return BaseIndexOffset(Base, Index, Off + PartialOffset, IsIndexSignExt);
4800 /// Represents known origin of an individual byte in load combine pattern. The
4801 /// value of the byte is either constant zero or comes from memory.
4802 struct ByteProvider {
4803 // For constant zero providers Load is set to nullptr. For memory providers
4804 // Load represents the node which loads the byte from memory.
4805 // ByteOffset is the offset of the byte in the value produced by the load.
4807 unsigned ByteOffset;
4809 ByteProvider() : Load(nullptr), ByteOffset(0) {}
4811 static ByteProvider getMemory(LoadSDNode *Load, unsigned ByteOffset) {
4812 return ByteProvider(Load, ByteOffset);
4814 static ByteProvider getConstantZero() { return ByteProvider(nullptr, 0); }
4816 bool isConstantZero() const { return !Load; }
4817 bool isMemory() const { return Load; }
4819 bool operator==(const ByteProvider &Other) const {
4820 return Other.Load == Load && Other.ByteOffset == ByteOffset;
4824 ByteProvider(LoadSDNode *Load, unsigned ByteOffset)
4825 : Load(Load), ByteOffset(ByteOffset) {}
4828 /// Recursively traverses the expression calculating the origin of the requested
4829 /// byte of the given value. Returns None if the provider can't be calculated.
4831 /// For all the values except the root of the expression verifies that the value
4832 /// has exactly one use and if it's not true return None. This way if the origin
4833 /// of the byte is returned it's guaranteed that the values which contribute to
4834 /// the byte are not used outside of this expression.
4836 /// Because the parts of the expression are not allowed to have more than one
4837 /// use this function iterates over trees, not DAGs. So it never visits the same
4838 /// node more than once.
4839 const Optional<ByteProvider> calculateByteProvider(SDValue Op, unsigned Index,
4841 bool Root = false) {
4842 // Typical i64 by i8 pattern requires recursion up to 8 calls depth
4846 if (!Root && !Op.hasOneUse())
4849 assert(Op.getValueType().isScalarInteger() && "can't handle other types");
4850 unsigned BitWidth = Op.getValueSizeInBits();
4851 if (BitWidth % 8 != 0)
4853 unsigned ByteWidth = BitWidth / 8;
4854 assert(Index < ByteWidth && "invalid index requested");
4857 switch (Op.getOpcode()) {
4859 auto LHS = calculateByteProvider(Op->getOperand(0), Index, Depth + 1);
4862 auto RHS = calculateByteProvider(Op->getOperand(1), Index, Depth + 1);
4866 if (LHS->isConstantZero())
4868 if (RHS->isConstantZero())
4873 auto ShiftOp = dyn_cast<ConstantSDNode>(Op->getOperand(1));
4877 uint64_t BitShift = ShiftOp->getZExtValue();
4878 if (BitShift % 8 != 0)
4880 uint64_t ByteShift = BitShift / 8;
4882 return Index < ByteShift
4883 ? ByteProvider::getConstantZero()
4884 : calculateByteProvider(Op->getOperand(0), Index - ByteShift,
4887 case ISD::ANY_EXTEND:
4888 case ISD::SIGN_EXTEND:
4889 case ISD::ZERO_EXTEND: {
4890 SDValue NarrowOp = Op->getOperand(0);
4891 unsigned NarrowBitWidth = NarrowOp.getScalarValueSizeInBits();
4892 if (NarrowBitWidth % 8 != 0)
4894 uint64_t NarrowByteWidth = NarrowBitWidth / 8;
4896 if (Index >= NarrowByteWidth)
4897 return Op.getOpcode() == ISD::ZERO_EXTEND
4898 ? Optional<ByteProvider>(ByteProvider::getConstantZero())
4900 return calculateByteProvider(NarrowOp, Index, Depth + 1);
4903 return calculateByteProvider(Op->getOperand(0), ByteWidth - Index - 1,
4906 auto L = cast<LoadSDNode>(Op.getNode());
4907 if (L->isVolatile() || L->isIndexed())
4910 unsigned NarrowBitWidth = L->getMemoryVT().getSizeInBits();
4911 if (NarrowBitWidth % 8 != 0)
4913 uint64_t NarrowByteWidth = NarrowBitWidth / 8;
4915 if (Index >= NarrowByteWidth)
4916 return L->getExtensionType() == ISD::ZEXTLOAD
4917 ? Optional<ByteProvider>(ByteProvider::getConstantZero())
4919 return ByteProvider::getMemory(L, Index);
4927 /// Match a pattern where a wide type scalar value is loaded by several narrow
4928 /// loads and combined by shifts and ors. Fold it into a single load or a load
4929 /// and a BSWAP if the targets supports it.
4931 /// Assuming little endian target:
4933 /// i32 val = a[0] | (a[1] << 8) | (a[2] << 16) | (a[3] << 24)
4935 /// i32 val = *((i32)a)
4938 /// i32 val = (a[0] << 24) | (a[1] << 16) | (a[2] << 8) | a[3]
4940 /// i32 val = BSWAP(*((i32)a))
4942 /// TODO: This rule matches complex patterns with OR node roots and doesn't
4943 /// interact well with the worklist mechanism. When a part of the pattern is
4944 /// updated (e.g. one of the loads) its direct users are put into the worklist,
4945 /// but the root node of the pattern which triggers the load combine is not
4946 /// necessarily a direct user of the changed node. For example, once the address
4947 /// of t28 load is reassociated load combine won't be triggered:
4948 /// t25: i32 = add t4, Constant:i32<2>
4949 /// t26: i64 = sign_extend t25
4950 /// t27: i64 = add t2, t26
4951 /// t28: i8,ch = load<LD1[%tmp9]> t0, t27, undef:i64
4952 /// t29: i32 = zero_extend t28
4953 /// t32: i32 = shl t29, Constant:i8<8>
4954 /// t33: i32 = or t23, t32
4955 /// As a possible fix visitLoad can check if the load can be a part of a load
4956 /// combine pattern and add corresponding OR roots to the worklist.
4957 SDValue DAGCombiner::MatchLoadCombine(SDNode *N) {
4958 assert(N->getOpcode() == ISD::OR &&
4959 "Can only match load combining against OR nodes");
4961 // Handles simple types only
4962 EVT VT = N->getValueType(0);
4963 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
4965 unsigned ByteWidth = VT.getSizeInBits() / 8;
4967 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4968 // Before legalize we can introduce too wide illegal loads which will be later
4969 // split into legal sized loads. This enables us to combine i64 load by i8
4970 // patterns to a couple of i32 loads on 32 bit targets.
4971 if (LegalOperations && !TLI.isOperationLegal(ISD::LOAD, VT))
4974 std::function<unsigned(unsigned, unsigned)> LittleEndianByteAt = [](
4975 unsigned BW, unsigned i) { return i; };
4976 std::function<unsigned(unsigned, unsigned)> BigEndianByteAt = [](
4977 unsigned BW, unsigned i) { return BW - i - 1; };
4979 bool IsBigEndianTarget = DAG.getDataLayout().isBigEndian();
4980 auto MemoryByteOffset = [&] (ByteProvider P) {
4981 assert(P.isMemory() && "Must be a memory byte provider");
4982 unsigned LoadBitWidth = P.Load->getMemoryVT().getSizeInBits();
4983 assert(LoadBitWidth % 8 == 0 &&
4984 "can only analyze providers for individual bytes not bit");
4985 unsigned LoadByteWidth = LoadBitWidth / 8;
4986 return IsBigEndianTarget
4987 ? BigEndianByteAt(LoadByteWidth, P.ByteOffset)
4988 : LittleEndianByteAt(LoadByteWidth, P.ByteOffset);
4991 Optional<BaseIndexOffset> Base;
4994 SmallSet<LoadSDNode *, 8> Loads;
4995 Optional<ByteProvider> FirstByteProvider;
4996 int64_t FirstOffset = INT64_MAX;
4998 // Check if all the bytes of the OR we are looking at are loaded from the same
4999 // base address. Collect bytes offsets from Base address in ByteOffsets.
5000 SmallVector<int64_t, 4> ByteOffsets(ByteWidth);
5001 for (unsigned i = 0; i < ByteWidth; i++) {
5002 auto P = calculateByteProvider(SDValue(N, 0), i, 0, /*Root=*/true);
5003 if (!P || !P->isMemory()) // All the bytes must be loaded from memory
5006 LoadSDNode *L = P->Load;
5007 assert(L->hasNUsesOfValue(1, 0) && !L->isVolatile() && !L->isIndexed() &&
5008 "Must be enforced by calculateByteProvider");
5009 assert(L->getOffset().isUndef() && "Unindexed load must have undef offset");
5011 // All loads must share the same chain
5012 SDValue LChain = L->getChain();
5015 else if (Chain != LChain)
5018 // Loads must share the same base address
5019 BaseIndexOffset Ptr = BaseIndexOffset::match(L->getBasePtr(), DAG);
5022 else if (!Base->equalBaseIndex(Ptr))
5025 // Calculate the offset of the current byte from the base address
5026 int64_t ByteOffsetFromBase = Ptr.Offset + MemoryByteOffset(*P);
5027 ByteOffsets[i] = ByteOffsetFromBase;
5029 // Remember the first byte load
5030 if (ByteOffsetFromBase < FirstOffset) {
5031 FirstByteProvider = P;
5032 FirstOffset = ByteOffsetFromBase;
5037 assert(Loads.size() > 0 && "All the bytes of the value must be loaded from "
5038 "memory, so there must be at least one load which produces the value");
5039 assert(Base && "Base address of the accessed memory location must be set");
5040 assert(FirstOffset != INT64_MAX && "First byte offset must be set");
5042 // Check if the bytes of the OR we are looking at match with either big or
5043 // little endian value load
5044 bool BigEndian = true, LittleEndian = true;
5045 for (unsigned i = 0; i < ByteWidth; i++) {
5046 int64_t CurrentByteOffset = ByteOffsets[i] - FirstOffset;
5047 LittleEndian &= CurrentByteOffset == LittleEndianByteAt(ByteWidth, i);
5048 BigEndian &= CurrentByteOffset == BigEndianByteAt(ByteWidth, i);
5049 if (!BigEndian && !LittleEndian)
5052 assert((BigEndian != LittleEndian) && "should be either or");
5053 assert(FirstByteProvider && "must be set");
5055 // Ensure that the first byte is loaded from zero offset of the first load.
5056 // So the combined value can be loaded from the first load address.
5057 if (MemoryByteOffset(*FirstByteProvider) != 0)
5059 LoadSDNode *FirstLoad = FirstByteProvider->Load;
5061 // The node we are looking at matches with the pattern, check if we can
5062 // replace it with a single load and bswap if needed.
5064 // If the load needs byte swap check if the target supports it
5065 bool NeedsBswap = IsBigEndianTarget != BigEndian;
5067 // Before legalize we can introduce illegal bswaps which will be later
5068 // converted to an explicit bswap sequence. This way we end up with a single
5069 // load and byte shuffling instead of several loads and byte shuffling.
5070 if (NeedsBswap && LegalOperations && !TLI.isOperationLegal(ISD::BSWAP, VT))
5073 // Check that a load of the wide type is both allowed and fast on the target
5075 bool Allowed = TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(),
5076 VT, FirstLoad->getAddressSpace(),
5077 FirstLoad->getAlignment(), &Fast);
5078 if (!Allowed || !Fast)
5082 DAG.getLoad(VT, SDLoc(N), Chain, FirstLoad->getBasePtr(),
5083 FirstLoad->getPointerInfo(), FirstLoad->getAlignment());
5085 // Transfer chain users from old loads to the new load.
5086 for (LoadSDNode *L : Loads)
5087 DAG.ReplaceAllUsesOfValueWith(SDValue(L, 1), SDValue(NewLoad.getNode(), 1));
5089 return NeedsBswap ? DAG.getNode(ISD::BSWAP, SDLoc(N), VT, NewLoad) : NewLoad;
5092 SDValue DAGCombiner::visitXOR(SDNode *N) {
5093 SDValue N0 = N->getOperand(0);
5094 SDValue N1 = N->getOperand(1);
5095 EVT VT = N0.getValueType();
5098 if (VT.isVector()) {
5099 if (SDValue FoldedVOp = SimplifyVBinOp(N))
5102 // fold (xor x, 0) -> x, vector edition
5103 if (ISD::isBuildVectorAllZeros(N0.getNode()))
5105 if (ISD::isBuildVectorAllZeros(N1.getNode()))
5109 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
5110 if (N0.isUndef() && N1.isUndef())
5111 return DAG.getConstant(0, SDLoc(N), VT);
5112 // fold (xor x, undef) -> undef
5117 // fold (xor c1, c2) -> c1^c2
5118 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
5119 ConstantSDNode *N1C = getAsNonOpaqueConstant(N1);
5121 return DAG.FoldConstantArithmetic(ISD::XOR, SDLoc(N), VT, N0C, N1C);
5122 // canonicalize constant to RHS
5123 if (DAG.isConstantIntBuildVectorOrConstantInt(N0) &&
5124 !DAG.isConstantIntBuildVectorOrConstantInt(N1))
5125 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
5126 // fold (xor x, 0) -> x
5127 if (isNullConstant(N1))
5130 if (SDValue NewSel = foldBinOpIntoSelect(N))
5134 if (SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1))
5137 // fold !(x cc y) -> (x !cc y)
5138 SDValue LHS, RHS, CC;
5139 if (TLI.isConstTrueVal(N1.getNode()) && isSetCCEquivalent(N0, LHS, RHS, CC)) {
5140 bool isInt = LHS.getValueType().isInteger();
5141 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
5144 if (!LegalOperations ||
5145 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) {
5146 switch (N0.getOpcode()) {
5148 llvm_unreachable("Unhandled SetCC Equivalent!");
5150 return DAG.getSetCC(SDLoc(N0), VT, LHS, RHS, NotCC);
5151 case ISD::SELECT_CC:
5152 return DAG.getSelectCC(SDLoc(N0), LHS, RHS, N0.getOperand(2),
5153 N0.getOperand(3), NotCC);
5158 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
5159 if (isOneConstant(N1) && N0.getOpcode() == ISD::ZERO_EXTEND &&
5160 N0.getNode()->hasOneUse() &&
5161 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
5162 SDValue V = N0.getOperand(0);
5164 V = DAG.getNode(ISD::XOR, DL, V.getValueType(), V,
5165 DAG.getConstant(1, DL, V.getValueType()));
5166 AddToWorklist(V.getNode());
5167 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V);
5170 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
5171 if (isOneConstant(N1) && VT == MVT::i1 &&
5172 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
5173 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
5174 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
5175 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
5176 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
5177 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
5178 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
5179 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
5182 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
5183 if (isAllOnesConstant(N1) &&
5184 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
5185 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
5186 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
5187 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
5188 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
5189 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
5190 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
5191 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
5194 // fold (xor (and x, y), y) -> (and (not x), y)
5195 if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
5196 N0->getOperand(1) == N1) {
5197 SDValue X = N0->getOperand(0);
5198 SDValue NotX = DAG.getNOT(SDLoc(X), X, VT);
5199 AddToWorklist(NotX.getNode());
5200 return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1);
5202 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
5203 if (N1C && N0.getOpcode() == ISD::XOR) {
5204 if (const ConstantSDNode *N00C = getAsNonOpaqueConstant(N0.getOperand(0))) {
5206 return DAG.getNode(ISD::XOR, DL, VT, N0.getOperand(1),
5207 DAG.getConstant(N1C->getAPIntValue() ^
5208 N00C->getAPIntValue(), DL, VT));
5210 if (const ConstantSDNode *N01C = getAsNonOpaqueConstant(N0.getOperand(1))) {
5212 return DAG.getNode(ISD::XOR, DL, VT, N0.getOperand(0),
5213 DAG.getConstant(N1C->getAPIntValue() ^
5214 N01C->getAPIntValue(), DL, VT));
5218 // fold Y = sra (X, size(X)-1); xor (add (X, Y), Y) -> (abs X)
5219 unsigned OpSizeInBits = VT.getScalarSizeInBits();
5220 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
5221 N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0) &&
5222 TLI.isOperationLegalOrCustom(ISD::ABS, VT)) {
5223 if (ConstantSDNode *C = isConstOrConstSplat(N1.getOperand(1)))
5224 if (C->getAPIntValue() == (OpSizeInBits - 1))
5225 return DAG.getNode(ISD::ABS, SDLoc(N), VT, N0.getOperand(0));
5228 // fold (xor x, x) -> 0
5230 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
5232 // fold (xor (shl 1, x), -1) -> (rotl ~1, x)
5233 // Here is a concrete example of this equivalence:
5235 // i16 shl == 1 << 14 == 16384 == 0b0100000000000000
5236 // i16 xor == ~(1 << 14) == 49151 == 0b1011111111111111
5240 // i16 ~1 == 0b1111111111111110
5241 // i16 rol(~1, 14) == 0b1011111111111111
5243 // Some additional tips to help conceptualize this transform:
5244 // - Try to see the operation as placing a single zero in a value of all ones.
5245 // - There exists no value for x which would allow the result to contain zero.
5246 // - Values of x larger than the bitwidth are undefined and do not require a
5247 // consistent result.
5248 // - Pushing the zero left requires shifting one bits in from the right.
5249 // A rotate left of ~1 is a nice way of achieving the desired result.
5250 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT) && N0.getOpcode() == ISD::SHL
5251 && isAllOnesConstant(N1) && isOneConstant(N0.getOperand(0))) {
5253 return DAG.getNode(ISD::ROTL, DL, VT, DAG.getConstant(~1, DL, VT),
5257 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
5258 if (N0.getOpcode() == N1.getOpcode())
5259 if (SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N))
5262 // Simplify the expression using non-local knowledge.
5263 if (SimplifyDemandedBits(SDValue(N, 0)))
5264 return SDValue(N, 0);
5269 /// Handle transforms common to the three shifts, when the shift amount is a
5271 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, ConstantSDNode *Amt) {
5272 SDNode *LHS = N->getOperand(0).getNode();
5273 if (!LHS->hasOneUse()) return SDValue();
5275 // We want to pull some binops through shifts, so that we have (and (shift))
5276 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
5277 // thing happens with address calculations, so it's important to canonicalize
5279 bool HighBitSet = false; // Can we transform this if the high bit is set?
5281 switch (LHS->getOpcode()) {
5282 default: return SDValue();
5285 HighBitSet = false; // We can only transform sra if the high bit is clear.
5288 HighBitSet = true; // We can only transform sra if the high bit is set.
5291 if (N->getOpcode() != ISD::SHL)
5292 return SDValue(); // only shl(add) not sr[al](add).
5293 HighBitSet = false; // We can only transform sra if the high bit is clear.
5297 // We require the RHS of the binop to be a constant and not opaque as well.
5298 ConstantSDNode *BinOpCst = getAsNonOpaqueConstant(LHS->getOperand(1));
5299 if (!BinOpCst) return SDValue();
5301 // FIXME: disable this unless the input to the binop is a shift by a constant
5302 // or is copy/select.Enable this in other cases when figure out it's exactly profitable.
5303 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
5304 bool isShift = BinOpLHSVal->getOpcode() == ISD::SHL ||
5305 BinOpLHSVal->getOpcode() == ISD::SRA ||
5306 BinOpLHSVal->getOpcode() == ISD::SRL;
5307 bool isCopyOrSelect = BinOpLHSVal->getOpcode() == ISD::CopyFromReg ||
5308 BinOpLHSVal->getOpcode() == ISD::SELECT;
5310 if ((!isShift || !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1))) &&
5314 if (isCopyOrSelect && N->hasOneUse())
5317 EVT VT = N->getValueType(0);
5319 // If this is a signed shift right, and the high bit is modified by the
5320 // logical operation, do not perform the transformation. The highBitSet
5321 // boolean indicates the value of the high bit of the constant which would
5322 // cause it to be modified for this operation.
5323 if (N->getOpcode() == ISD::SRA) {
5324 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
5325 if (BinOpRHSSignSet != HighBitSet)
5329 if (!TLI.isDesirableToCommuteWithShift(LHS))
5332 // Fold the constants, shifting the binop RHS by the shift amount.
5333 SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)),
5335 LHS->getOperand(1), N->getOperand(1));
5336 assert(isa<ConstantSDNode>(NewRHS) && "Folding was not successful!");
5338 // Create the new shift.
5339 SDValue NewShift = DAG.getNode(N->getOpcode(),
5340 SDLoc(LHS->getOperand(0)),
5341 VT, LHS->getOperand(0), N->getOperand(1));
5343 // Create the new binop.
5344 return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS);
5347 SDValue DAGCombiner::distributeTruncateThroughAnd(SDNode *N) {
5348 assert(N->getOpcode() == ISD::TRUNCATE);
5349 assert(N->getOperand(0).getOpcode() == ISD::AND);
5351 // (truncate:TruncVT (and N00, N01C)) -> (and (truncate:TruncVT N00), TruncC)
5352 if (N->hasOneUse() && N->getOperand(0).hasOneUse()) {
5353 SDValue N01 = N->getOperand(0).getOperand(1);
5354 if (isConstantOrConstantVector(N01, /* NoOpaques */ true)) {
5356 EVT TruncVT = N->getValueType(0);
5357 SDValue N00 = N->getOperand(0).getOperand(0);
5358 SDValue Trunc00 = DAG.getNode(ISD::TRUNCATE, DL, TruncVT, N00);
5359 SDValue Trunc01 = DAG.getNode(ISD::TRUNCATE, DL, TruncVT, N01);
5360 AddToWorklist(Trunc00.getNode());
5361 AddToWorklist(Trunc01.getNode());
5362 return DAG.getNode(ISD::AND, DL, TruncVT, Trunc00, Trunc01);
5369 SDValue DAGCombiner::visitRotate(SDNode *N) {
5370 // fold (rot* x, (trunc (and y, c))) -> (rot* x, (and (trunc y), (trunc c))).
5371 if (N->getOperand(1).getOpcode() == ISD::TRUNCATE &&
5372 N->getOperand(1).getOperand(0).getOpcode() == ISD::AND) {
5373 if (SDValue NewOp1 =
5374 distributeTruncateThroughAnd(N->getOperand(1).getNode()))
5375 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
5376 N->getOperand(0), NewOp1);
5381 SDValue DAGCombiner::visitSHL(SDNode *N) {
5382 SDValue N0 = N->getOperand(0);
5383 SDValue N1 = N->getOperand(1);
5384 EVT VT = N0.getValueType();
5385 unsigned OpSizeInBits = VT.getScalarSizeInBits();
5388 if (VT.isVector()) {
5389 if (SDValue FoldedVOp = SimplifyVBinOp(N))
5392 BuildVectorSDNode *N1CV = dyn_cast<BuildVectorSDNode>(N1);
5393 // If setcc produces all-one true value then:
5394 // (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<<N1CV)
5395 if (N1CV && N1CV->isConstant()) {
5396 if (N0.getOpcode() == ISD::AND) {
5397 SDValue N00 = N0->getOperand(0);
5398 SDValue N01 = N0->getOperand(1);
5399 BuildVectorSDNode *N01CV = dyn_cast<BuildVectorSDNode>(N01);
5401 if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC &&
5402 TLI.getBooleanContents(N00.getOperand(0).getValueType()) ==
5403 TargetLowering::ZeroOrNegativeOneBooleanContent) {
5404 if (SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, SDLoc(N), VT,
5406 return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C);
5412 ConstantSDNode *N1C = isConstOrConstSplat(N1);
5414 // fold (shl c1, c2) -> c1<<c2
5415 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
5416 if (N0C && N1C && !N1C->isOpaque())
5417 return DAG.FoldConstantArithmetic(ISD::SHL, SDLoc(N), VT, N0C, N1C);
5418 // fold (shl 0, x) -> 0
5419 if (isNullConstantOrNullSplatConstant(N0))
5421 // fold (shl x, c >= size(x)) -> undef
5422 if (N1C && N1C->getAPIntValue().uge(OpSizeInBits))
5423 return DAG.getUNDEF(VT);
5424 // fold (shl x, 0) -> x
5425 if (N1C && N1C->isNullValue())
5427 // fold (shl undef, x) -> 0
5429 return DAG.getConstant(0, SDLoc(N), VT);
5431 if (SDValue NewSel = foldBinOpIntoSelect(N))
5434 // if (shl x, c) is known to be zero, return 0
5435 if (DAG.MaskedValueIsZero(SDValue(N, 0),
5436 APInt::getAllOnesValue(OpSizeInBits)))
5437 return DAG.getConstant(0, SDLoc(N), VT);
5438 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
5439 if (N1.getOpcode() == ISD::TRUNCATE &&
5440 N1.getOperand(0).getOpcode() == ISD::AND) {
5441 if (SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode()))
5442 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, NewOp1);
5445 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
5446 return SDValue(N, 0);
5448 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
5449 if (N1C && N0.getOpcode() == ISD::SHL) {
5450 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
5452 APInt c1 = N0C1->getAPIntValue();
5453 APInt c2 = N1C->getAPIntValue();
5454 zeroExtendToMatch(c1, c2, 1 /* Overflow Bit */);
5456 APInt Sum = c1 + c2;
5457 if (Sum.uge(OpSizeInBits))
5458 return DAG.getConstant(0, DL, VT);
5461 ISD::SHL, DL, VT, N0.getOperand(0),
5462 DAG.getConstant(Sum.getZExtValue(), DL, N1.getValueType()));
5466 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
5467 // For this to be valid, the second form must not preserve any of the bits
5468 // that are shifted out by the inner shift in the first form. This means
5469 // the outer shift size must be >= the number of bits added by the ext.
5470 // As a corollary, we don't care what kind of ext it is.
5471 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
5472 N0.getOpcode() == ISD::ANY_EXTEND ||
5473 N0.getOpcode() == ISD::SIGN_EXTEND) &&
5474 N0.getOperand(0).getOpcode() == ISD::SHL) {
5475 SDValue N0Op0 = N0.getOperand(0);
5476 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
5477 APInt c1 = N0Op0C1->getAPIntValue();
5478 APInt c2 = N1C->getAPIntValue();
5479 zeroExtendToMatch(c1, c2, 1 /* Overflow Bit */);
5481 EVT InnerShiftVT = N0Op0.getValueType();
5482 uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits();
5483 if (c2.uge(OpSizeInBits - InnerShiftSize)) {
5485 APInt Sum = c1 + c2;
5486 if (Sum.uge(OpSizeInBits))
5487 return DAG.getConstant(0, DL, VT);
5491 DAG.getNode(N0.getOpcode(), DL, VT, N0Op0->getOperand(0)),
5492 DAG.getConstant(Sum.getZExtValue(), DL, N1.getValueType()));
5497 // fold (shl (zext (srl x, C)), C) -> (zext (shl (srl x, C), C))
5498 // Only fold this if the inner zext has no other uses to avoid increasing
5499 // the total number of instructions.
5500 if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() &&
5501 N0.getOperand(0).getOpcode() == ISD::SRL) {
5502 SDValue N0Op0 = N0.getOperand(0);
5503 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
5504 if (N0Op0C1->getAPIntValue().ult(VT.getScalarSizeInBits())) {
5505 uint64_t c1 = N0Op0C1->getZExtValue();
5506 uint64_t c2 = N1C->getZExtValue();
5508 SDValue NewOp0 = N0.getOperand(0);
5509 EVT CountVT = NewOp0.getOperand(1).getValueType();
5511 SDValue NewSHL = DAG.getNode(ISD::SHL, DL, NewOp0.getValueType(),
5513 DAG.getConstant(c2, DL, CountVT));
5514 AddToWorklist(NewSHL.getNode());
5515 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
5521 // fold (shl (sr[la] exact X, C1), C2) -> (shl X, (C2-C1)) if C1 <= C2
5522 // fold (shl (sr[la] exact X, C1), C2) -> (sr[la] X, (C2-C1)) if C1 > C2
5523 if (N1C && (N0.getOpcode() == ISD::SRL || N0.getOpcode() == ISD::SRA) &&
5524 N0->getFlags().hasExact()) {
5525 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
5526 uint64_t C1 = N0C1->getZExtValue();
5527 uint64_t C2 = N1C->getZExtValue();
5530 return DAG.getNode(ISD::SHL, DL, VT, N0.getOperand(0),
5531 DAG.getConstant(C2 - C1, DL, N1.getValueType()));
5532 return DAG.getNode(N0.getOpcode(), DL, VT, N0.getOperand(0),
5533 DAG.getConstant(C1 - C2, DL, N1.getValueType()));
5537 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
5538 // (and (srl x, (sub c1, c2), MASK)
5539 // Only fold this if the inner shift has no other uses -- if it does, folding
5540 // this will increase the total number of instructions.
5541 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
5542 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
5543 uint64_t c1 = N0C1->getZExtValue();
5544 if (c1 < OpSizeInBits) {
5545 uint64_t c2 = N1C->getZExtValue();
5546 APInt Mask = APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - c1);
5551 Shift = DAG.getNode(ISD::SHL, DL, VT, N0.getOperand(0),
5552 DAG.getConstant(c2 - c1, DL, N1.getValueType()));
5554 Mask.lshrInPlace(c1 - c2);
5556 Shift = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
5557 DAG.getConstant(c1 - c2, DL, N1.getValueType()));
5560 return DAG.getNode(ISD::AND, DL, VT, Shift,
5561 DAG.getConstant(Mask, DL, VT));
5566 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
5567 if (N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1) &&
5568 isConstantOrConstantVector(N1, /* No Opaques */ true)) {
5570 SDValue AllBits = DAG.getAllOnesConstant(DL, VT);
5571 SDValue HiBitsMask = DAG.getNode(ISD::SHL, DL, VT, AllBits, N1);
5572 return DAG.getNode(ISD::AND, DL, VT, N0.getOperand(0), HiBitsMask);
5575 // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
5576 // Variant of version done on multiply, except mul by a power of 2 is turned
5578 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
5579 isConstantOrConstantVector(N1, /* No Opaques */ true) &&
5580 isConstantOrConstantVector(N0.getOperand(1), /* No Opaques */ true)) {
5581 SDValue Shl0 = DAG.getNode(ISD::SHL, SDLoc(N0), VT, N0.getOperand(0), N1);
5582 SDValue Shl1 = DAG.getNode(ISD::SHL, SDLoc(N1), VT, N0.getOperand(1), N1);
5583 AddToWorklist(Shl0.getNode());
5584 AddToWorklist(Shl1.getNode());
5585 return DAG.getNode(ISD::ADD, SDLoc(N), VT, Shl0, Shl1);
5588 // fold (shl (mul x, c1), c2) -> (mul x, c1 << c2)
5589 if (N0.getOpcode() == ISD::MUL && N0.getNode()->hasOneUse() &&
5590 isConstantOrConstantVector(N1, /* No Opaques */ true) &&
5591 isConstantOrConstantVector(N0.getOperand(1), /* No Opaques */ true)) {
5592 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N1), VT, N0.getOperand(1), N1);
5593 if (isConstantOrConstantVector(Shl))
5594 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N0.getOperand(0), Shl);
5597 if (N1C && !N1C->isOpaque())
5598 if (SDValue NewSHL = visitShiftByConstant(N, N1C))
5604 SDValue DAGCombiner::visitSRA(SDNode *N) {
5605 SDValue N0 = N->getOperand(0);
5606 SDValue N1 = N->getOperand(1);
5607 EVT VT = N0.getValueType();
5608 unsigned OpSizeInBits = VT.getScalarSizeInBits();
5610 // Arithmetic shifting an all-sign-bit value is a no-op.
5611 // fold (sra 0, x) -> 0
5612 // fold (sra -1, x) -> -1
5613 if (DAG.ComputeNumSignBits(N0) == OpSizeInBits)
5618 if (SDValue FoldedVOp = SimplifyVBinOp(N))
5621 ConstantSDNode *N1C = isConstOrConstSplat(N1);
5623 // fold (sra c1, c2) -> (sra c1, c2)
5624 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
5625 if (N0C && N1C && !N1C->isOpaque())
5626 return DAG.FoldConstantArithmetic(ISD::SRA, SDLoc(N), VT, N0C, N1C);
5627 // fold (sra x, c >= size(x)) -> undef
5628 if (N1C && N1C->getAPIntValue().uge(OpSizeInBits))
5629 return DAG.getUNDEF(VT);
5630 // fold (sra x, 0) -> x
5631 if (N1C && N1C->isNullValue())
5634 if (SDValue NewSel = foldBinOpIntoSelect(N))
5637 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
5639 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
5640 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
5641 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
5643 ExtVT = EVT::getVectorVT(*DAG.getContext(),
5644 ExtVT, VT.getVectorNumElements());
5645 if ((!LegalOperations ||
5646 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
5647 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
5648 N0.getOperand(0), DAG.getValueType(ExtVT));
5651 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
5652 if (N1C && N0.getOpcode() == ISD::SRA) {
5653 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
5655 APInt c1 = N0C1->getAPIntValue();
5656 APInt c2 = N1C->getAPIntValue();
5657 zeroExtendToMatch(c1, c2, 1 /* Overflow Bit */);
5659 APInt Sum = c1 + c2;
5660 if (Sum.uge(OpSizeInBits))
5661 Sum = APInt(OpSizeInBits, OpSizeInBits - 1);
5664 ISD::SRA, DL, VT, N0.getOperand(0),
5665 DAG.getConstant(Sum.getZExtValue(), DL, N1.getValueType()));
5669 // fold (sra (shl X, m), (sub result_size, n))
5670 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
5671 // result_size - n != m.
5672 // If truncate is free for the target sext(shl) is likely to result in better
5674 if (N0.getOpcode() == ISD::SHL && N1C) {
5675 // Get the two constanst of the shifts, CN0 = m, CN = n.
5676 const ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1));
5678 LLVMContext &Ctx = *DAG.getContext();
5679 // Determine what the truncate's result bitsize and type would be.
5680 EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - N1C->getZExtValue());
5683 TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorNumElements());
5685 // Determine the residual right-shift amount.
5686 int ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
5688 // If the shift is not a no-op (in which case this should be just a sign
5689 // extend already), the truncated to type is legal, sign_extend is legal
5690 // on that type, and the truncate to that type is both legal and free,
5691 // perform the transform.
5692 if ((ShiftAmt > 0) &&
5693 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
5694 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
5695 TLI.isTruncateFree(VT, TruncVT)) {
5698 SDValue Amt = DAG.getConstant(ShiftAmt, DL,
5699 getShiftAmountTy(N0.getOperand(0).getValueType()));
5700 SDValue Shift = DAG.getNode(ISD::SRL, DL, VT,
5701 N0.getOperand(0), Amt);
5702 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, TruncVT,
5704 return DAG.getNode(ISD::SIGN_EXTEND, DL,
5705 N->getValueType(0), Trunc);
5710 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
5711 if (N1.getOpcode() == ISD::TRUNCATE &&
5712 N1.getOperand(0).getOpcode() == ISD::AND) {
5713 if (SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode()))
5714 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, NewOp1);
5717 // fold (sra (trunc (srl x, c1)), c2) -> (trunc (sra x, c1 + c2))
5718 // if c1 is equal to the number of bits the trunc removes
5719 if (N0.getOpcode() == ISD::TRUNCATE &&
5720 (N0.getOperand(0).getOpcode() == ISD::SRL ||
5721 N0.getOperand(0).getOpcode() == ISD::SRA) &&
5722 N0.getOperand(0).hasOneUse() &&
5723 N0.getOperand(0).getOperand(1).hasOneUse() &&
5725 SDValue N0Op0 = N0.getOperand(0);
5726 if (ConstantSDNode *LargeShift = isConstOrConstSplat(N0Op0.getOperand(1))) {
5727 unsigned LargeShiftVal = LargeShift->getZExtValue();
5728 EVT LargeVT = N0Op0.getValueType();
5730 if (LargeVT.getScalarSizeInBits() - OpSizeInBits == LargeShiftVal) {
5733 DAG.getConstant(LargeShiftVal + N1C->getZExtValue(), DL,
5734 getShiftAmountTy(N0Op0.getOperand(0).getValueType()));
5735 SDValue SRA = DAG.getNode(ISD::SRA, DL, LargeVT,
5736 N0Op0.getOperand(0), Amt);
5737 return DAG.getNode(ISD::TRUNCATE, DL, VT, SRA);
5742 // Simplify, based on bits shifted out of the LHS.
5743 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
5744 return SDValue(N, 0);
5747 // If the sign bit is known to be zero, switch this to a SRL.
5748 if (DAG.SignBitIsZero(N0))
5749 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1);
5751 if (N1C && !N1C->isOpaque())
5752 if (SDValue NewSRA = visitShiftByConstant(N, N1C))
5758 SDValue DAGCombiner::visitSRL(SDNode *N) {
5759 SDValue N0 = N->getOperand(0);
5760 SDValue N1 = N->getOperand(1);
5761 EVT VT = N0.getValueType();
5762 unsigned OpSizeInBits = VT.getScalarSizeInBits();
5766 if (SDValue FoldedVOp = SimplifyVBinOp(N))
5769 ConstantSDNode *N1C = isConstOrConstSplat(N1);
5771 // fold (srl c1, c2) -> c1 >>u c2
5772 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
5773 if (N0C && N1C && !N1C->isOpaque())
5774 return DAG.FoldConstantArithmetic(ISD::SRL, SDLoc(N), VT, N0C, N1C);
5775 // fold (srl 0, x) -> 0
5776 if (isNullConstantOrNullSplatConstant(N0))
5778 // fold (srl x, c >= size(x)) -> undef
5779 if (N1C && N1C->getAPIntValue().uge(OpSizeInBits))
5780 return DAG.getUNDEF(VT);
5781 // fold (srl x, 0) -> x
5782 if (N1C && N1C->isNullValue())
5785 if (SDValue NewSel = foldBinOpIntoSelect(N))
5788 // if (srl x, c) is known to be zero, return 0
5789 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
5790 APInt::getAllOnesValue(OpSizeInBits)))
5791 return DAG.getConstant(0, SDLoc(N), VT);
5793 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
5794 if (N1C && N0.getOpcode() == ISD::SRL) {
5795 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
5797 APInt c1 = N0C1->getAPIntValue();
5798 APInt c2 = N1C->getAPIntValue();
5799 zeroExtendToMatch(c1, c2, 1 /* Overflow Bit */);
5801 APInt Sum = c1 + c2;
5802 if (Sum.uge(OpSizeInBits))
5803 return DAG.getConstant(0, DL, VT);
5806 ISD::SRL, DL, VT, N0.getOperand(0),
5807 DAG.getConstant(Sum.getZExtValue(), DL, N1.getValueType()));
5811 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
5812 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
5813 N0.getOperand(0).getOpcode() == ISD::SRL) {
5814 if (auto N001C = isConstOrConstSplat(N0.getOperand(0).getOperand(1))) {
5815 uint64_t c1 = N001C->getZExtValue();
5816 uint64_t c2 = N1C->getZExtValue();
5817 EVT InnerShiftVT = N0.getOperand(0).getValueType();
5818 EVT ShiftCountVT = N0.getOperand(0).getOperand(1).getValueType();
5819 uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits();
5820 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
5821 if (c1 + OpSizeInBits == InnerShiftSize) {
5823 if (c1 + c2 >= InnerShiftSize)
5824 return DAG.getConstant(0, DL, VT);
5825 return DAG.getNode(ISD::TRUNCATE, DL, VT,
5826 DAG.getNode(ISD::SRL, DL, InnerShiftVT,
5827 N0.getOperand(0).getOperand(0),
5828 DAG.getConstant(c1 + c2, DL,
5834 // fold (srl (shl x, c), c) -> (and x, cst2)
5835 if (N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1 &&
5836 isConstantOrConstantVector(N1, /* NoOpaques */ true)) {
5839 DAG.getNode(ISD::SRL, DL, VT, DAG.getAllOnesConstant(DL, VT), N1);
5840 AddToWorklist(Mask.getNode());
5841 return DAG.getNode(ISD::AND, DL, VT, N0.getOperand(0), Mask);
5844 // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask)
5845 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
5846 // Shifting in all undef bits?
5847 EVT SmallVT = N0.getOperand(0).getValueType();
5848 unsigned BitSize = SmallVT.getScalarSizeInBits();
5849 if (N1C->getZExtValue() >= BitSize)
5850 return DAG.getUNDEF(VT);
5852 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
5853 uint64_t ShiftAmt = N1C->getZExtValue();
5855 SDValue SmallShift = DAG.getNode(ISD::SRL, DL0, SmallVT,
5857 DAG.getConstant(ShiftAmt, DL0,
5858 getShiftAmountTy(SmallVT)));
5859 AddToWorklist(SmallShift.getNode());
5860 APInt Mask = APInt::getLowBitsSet(OpSizeInBits, OpSizeInBits - ShiftAmt);
5862 return DAG.getNode(ISD::AND, DL, VT,
5863 DAG.getNode(ISD::ANY_EXTEND, DL, VT, SmallShift),
5864 DAG.getConstant(Mask, DL, VT));
5868 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
5869 // bit, which is unmodified by sra.
5870 if (N1C && N1C->getZExtValue() + 1 == OpSizeInBits) {
5871 if (N0.getOpcode() == ISD::SRA)
5872 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1);
5875 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
5876 if (N1C && N0.getOpcode() == ISD::CTLZ &&
5877 N1C->getAPIntValue() == Log2_32(OpSizeInBits)) {
5879 DAG.computeKnownBits(N0.getOperand(0), Known);
5881 // If any of the input bits are KnownOne, then the input couldn't be all
5882 // zeros, thus the result of the srl will always be zero.
5883 if (Known.One.getBoolValue()) return DAG.getConstant(0, SDLoc(N0), VT);
5885 // If all of the bits input the to ctlz node are known to be zero, then
5886 // the result of the ctlz is "32" and the result of the shift is one.
5887 APInt UnknownBits = ~Known.Zero;
5888 if (UnknownBits == 0) return DAG.getConstant(1, SDLoc(N0), VT);
5890 // Otherwise, check to see if there is exactly one bit input to the ctlz.
5891 if (UnknownBits.isPowerOf2()) {
5892 // Okay, we know that only that the single bit specified by UnknownBits
5893 // could be set on input to the CTLZ node. If this bit is set, the SRL
5894 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
5895 // to an SRL/XOR pair, which is likely to simplify more.
5896 unsigned ShAmt = UnknownBits.countTrailingZeros();
5897 SDValue Op = N0.getOperand(0);
5901 Op = DAG.getNode(ISD::SRL, DL, VT, Op,
5902 DAG.getConstant(ShAmt, DL,
5903 getShiftAmountTy(Op.getValueType())));
5904 AddToWorklist(Op.getNode());
5908 return DAG.getNode(ISD::XOR, DL, VT,
5909 Op, DAG.getConstant(1, DL, VT));
5913 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
5914 if (N1.getOpcode() == ISD::TRUNCATE &&
5915 N1.getOperand(0).getOpcode() == ISD::AND) {
5916 if (SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode()))
5917 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, NewOp1);
5920 // fold operands of srl based on knowledge that the low bits are not
5922 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
5923 return SDValue(N, 0);
5925 if (N1C && !N1C->isOpaque())
5926 if (SDValue NewSRL = visitShiftByConstant(N, N1C))
5929 // Attempt to convert a srl of a load into a narrower zero-extending load.
5930 if (SDValue NarrowLoad = ReduceLoadWidth(N))
5933 // Here is a common situation. We want to optimize:
5936 // %b = and i32 %a, 2
5937 // %c = srl i32 %b, 1
5938 // brcond i32 %c ...
5944 // %c = setcc eq %b, 0
5947 // However when after the source operand of SRL is optimized into AND, the SRL
5948 // itself may not be optimized further. Look for it and add the BRCOND into
5950 if (N->hasOneUse()) {
5951 SDNode *Use = *N->use_begin();
5952 if (Use->getOpcode() == ISD::BRCOND)
5954 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
5955 // Also look pass the truncate.
5956 Use = *Use->use_begin();
5957 if (Use->getOpcode() == ISD::BRCOND)
5965 SDValue DAGCombiner::visitABS(SDNode *N) {
5966 SDValue N0 = N->getOperand(0);
5967 EVT VT = N->getValueType(0);
5969 // fold (abs c1) -> c2
5970 if (DAG.isConstantIntBuildVectorOrConstantInt(N0))
5971 return DAG.getNode(ISD::ABS, SDLoc(N), VT, N0);
5972 // fold (abs (abs x)) -> (abs x)
5973 if (N0.getOpcode() == ISD::ABS)
5975 // fold (abs x) -> x iff not-negative
5976 if (DAG.SignBitIsZero(N0))
5981 SDValue DAGCombiner::visitBSWAP(SDNode *N) {
5982 SDValue N0 = N->getOperand(0);
5983 EVT VT = N->getValueType(0);
5985 // fold (bswap c1) -> c2
5986 if (DAG.isConstantIntBuildVectorOrConstantInt(N0))
5987 return DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N0);
5988 // fold (bswap (bswap x)) -> x
5989 if (N0.getOpcode() == ISD::BSWAP)
5990 return N0->getOperand(0);
5994 SDValue DAGCombiner::visitBITREVERSE(SDNode *N) {
5995 SDValue N0 = N->getOperand(0);
5996 EVT VT = N->getValueType(0);
5998 // fold (bitreverse c1) -> c2
5999 if (DAG.isConstantIntBuildVectorOrConstantInt(N0))
6000 return DAG.getNode(ISD::BITREVERSE, SDLoc(N), VT, N0);
6001 // fold (bitreverse (bitreverse x)) -> x
6002 if (N0.getOpcode() == ISD::BITREVERSE)
6003 return N0.getOperand(0);
6007 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
6008 SDValue N0 = N->getOperand(0);
6009 EVT VT = N->getValueType(0);
6011 // fold (ctlz c1) -> c2
6012 if (DAG.isConstantIntBuildVectorOrConstantInt(N0))
6013 return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0);
6017 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
6018 SDValue N0 = N->getOperand(0);
6019 EVT VT = N->getValueType(0);
6021 // fold (ctlz_zero_undef c1) -> c2
6022 if (DAG.isConstantIntBuildVectorOrConstantInt(N0))
6023 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0);
6027 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
6028 SDValue N0 = N->getOperand(0);
6029 EVT VT = N->getValueType(0);
6031 // fold (cttz c1) -> c2
6032 if (DAG.isConstantIntBuildVectorOrConstantInt(N0))
6033 return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0);
6037 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
6038 SDValue N0 = N->getOperand(0);
6039 EVT VT = N->getValueType(0);
6041 // fold (cttz_zero_undef c1) -> c2
6042 if (DAG.isConstantIntBuildVectorOrConstantInt(N0))
6043 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0);
6047 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
6048 SDValue N0 = N->getOperand(0);
6049 EVT VT = N->getValueType(0);
6051 // fold (ctpop c1) -> c2
6052 if (DAG.isConstantIntBuildVectorOrConstantInt(N0))
6053 return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0);
6058 /// \brief Generate Min/Max node
6059 static SDValue combineMinNumMaxNum(const SDLoc &DL, EVT VT, SDValue LHS,
6060 SDValue RHS, SDValue True, SDValue False,
6061 ISD::CondCode CC, const TargetLowering &TLI,
6062 SelectionDAG &DAG) {
6063 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
6073 unsigned Opcode = (LHS == True) ? ISD::FMINNUM : ISD::FMAXNUM;
6074 if (TLI.isOperationLegal(Opcode, VT))
6075 return DAG.getNode(Opcode, DL, VT, LHS, RHS);
6084 unsigned Opcode = (LHS == True) ? ISD::FMAXNUM : ISD::FMINNUM;
6085 if (TLI.isOperationLegal(Opcode, VT))
6086 return DAG.getNode(Opcode, DL, VT, LHS, RHS);
6094 SDValue DAGCombiner::foldSelectOfConstants(SDNode *N) {
6095 SDValue Cond = N->getOperand(0);
6096 SDValue N1 = N->getOperand(1);
6097 SDValue N2 = N->getOperand(2);
6098 EVT VT = N->getValueType(0);
6099 EVT CondVT = Cond.getValueType();
6102 if (!VT.isInteger())
6105 auto *C1 = dyn_cast<ConstantSDNode>(N1);
6106 auto *C2 = dyn_cast<ConstantSDNode>(N2);
6110 // Only do this before legalization to avoid conflicting with target-specific
6111 // transforms in the other direction (create a select from a zext/sext). There
6112 // is also a target-independent combine here in DAGCombiner in the other
6113 // direction for (select Cond, -1, 0) when the condition is not i1.
6114 if (CondVT == MVT::i1 && !LegalOperations) {
6115 if (C1->isNullValue() && C2->isOne()) {
6116 // select Cond, 0, 1 --> zext (!Cond)
6117 SDValue NotCond = DAG.getNOT(DL, Cond, MVT::i1);
6119 NotCond = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, NotCond);
6122 if (C1->isNullValue() && C2->isAllOnesValue()) {
6123 // select Cond, 0, -1 --> sext (!Cond)
6124 SDValue NotCond = DAG.getNOT(DL, Cond, MVT::i1);
6126 NotCond = DAG.getNode(ISD::SIGN_EXTEND, DL, VT, NotCond);
6129 if (C1->isOne() && C2->isNullValue()) {
6130 // select Cond, 1, 0 --> zext (Cond)
6132 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Cond);
6135 if (C1->isAllOnesValue() && C2->isNullValue()) {
6136 // select Cond, -1, 0 --> sext (Cond)
6138 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, VT, Cond);
6142 // For any constants that differ by 1, we can transform the select into an
6143 // extend and add. Use a target hook because some targets may prefer to
6144 // transform in the other direction.
6145 if (TLI.convertSelectOfConstantsToMath()) {
6146 if (C1->getAPIntValue() - 1 == C2->getAPIntValue()) {
6147 // select Cond, C1, C1-1 --> add (zext Cond), C1-1
6149 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Cond);
6150 return DAG.getNode(ISD::ADD, DL, VT, Cond, N2);
6152 if (C1->getAPIntValue() + 1 == C2->getAPIntValue()) {
6153 // select Cond, C1, C1+1 --> add (sext Cond), C1+1
6155 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, VT, Cond);
6156 return DAG.getNode(ISD::ADD, DL, VT, Cond, N2);
6163 // fold (select Cond, 0, 1) -> (xor Cond, 1)
6164 // We can't do this reliably if integer based booleans have different contents
6165 // to floating point based booleans. This is because we can't tell whether we
6166 // have an integer-based boolean or a floating-point-based boolean unless we
6167 // can find the SETCC that produced it and inspect its operands. This is
6168 // fairly easy if C is the SETCC node, but it can potentially be
6169 // undiscoverable (or not reasonably discoverable). For example, it could be
6170 // in another basic block or it could require searching a complicated
6172 if (CondVT.isInteger() &&
6173 TLI.getBooleanContents(false, true) ==
6174 TargetLowering::ZeroOrOneBooleanContent &&
6175 TLI.getBooleanContents(false, false) ==
6176 TargetLowering::ZeroOrOneBooleanContent &&
6177 C1->isNullValue() && C2->isOne()) {
6179 DAG.getNode(ISD::XOR, DL, CondVT, Cond, DAG.getConstant(1, DL, CondVT));
6180 if (VT.bitsEq(CondVT))
6182 return DAG.getZExtOrTrunc(NotCond, DL, VT);
6188 SDValue DAGCombiner::visitSELECT(SDNode *N) {
6189 SDValue N0 = N->getOperand(0);
6190 SDValue N1 = N->getOperand(1);
6191 SDValue N2 = N->getOperand(2);
6192 EVT VT = N->getValueType(0);
6193 EVT VT0 = N0.getValueType();
6195 // fold (select C, X, X) -> X
6198 if (const ConstantSDNode *N0C = dyn_cast<const ConstantSDNode>(N0)) {
6199 // fold (select true, X, Y) -> X
6200 // fold (select false, X, Y) -> Y
6201 return !N0C->isNullValue() ? N1 : N2;
6203 // fold (select X, X, Y) -> (or X, Y)
6204 // fold (select X, 1, Y) -> (or C, Y)
6205 if (VT == VT0 && VT == MVT::i1 && (N0 == N1 || isOneConstant(N1)))
6206 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
6208 if (SDValue V = foldSelectOfConstants(N))
6211 // fold (select C, 0, X) -> (and (not C), X)
6212 if (VT == VT0 && VT == MVT::i1 && isNullConstant(N1)) {
6213 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
6214 AddToWorklist(NOTNode.getNode());
6215 return DAG.getNode(ISD::AND, SDLoc(N), VT, NOTNode, N2);
6217 // fold (select C, X, 1) -> (or (not C), X)
6218 if (VT == VT0 && VT == MVT::i1 && isOneConstant(N2)) {
6219 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
6220 AddToWorklist(NOTNode.getNode());
6221 return DAG.getNode(ISD::OR, SDLoc(N), VT, NOTNode, N1);
6223 // fold (select X, Y, X) -> (and X, Y)
6224 // fold (select X, Y, 0) -> (and X, Y)
6225 if (VT == VT0 && VT == MVT::i1 && (N0 == N2 || isNullConstant(N2)))
6226 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
6228 // If we can fold this based on the true/false value, do so.
6229 if (SimplifySelectOps(N, N1, N2))
6230 return SDValue(N, 0); // Don't revisit N.
6232 if (VT0 == MVT::i1) {
6233 // The code in this block deals with the following 2 equivalences:
6234 // select(C0|C1, x, y) <=> select(C0, x, select(C1, x, y))
6235 // select(C0&C1, x, y) <=> select(C0, select(C1, x, y), y)
6236 // The target can specify its preferred form with the
6237 // shouldNormalizeToSelectSequence() callback. However we always transform
6238 // to the right anyway if we find the inner select exists in the DAG anyway
6239 // and we always transform to the left side if we know that we can further
6240 // optimize the combination of the conditions.
6241 bool normalizeToSequence
6242 = TLI.shouldNormalizeToSelectSequence(*DAG.getContext(), VT);
6243 // select (and Cond0, Cond1), X, Y
6244 // -> select Cond0, (select Cond1, X, Y), Y
6245 if (N0->getOpcode() == ISD::AND && N0->hasOneUse()) {
6246 SDValue Cond0 = N0->getOperand(0);
6247 SDValue Cond1 = N0->getOperand(1);
6248 SDValue InnerSelect = DAG.getNode(ISD::SELECT, SDLoc(N),
6249 N1.getValueType(), Cond1, N1, N2);
6250 if (normalizeToSequence || !InnerSelect.use_empty())
6251 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Cond0,
6254 // select (or Cond0, Cond1), X, Y -> select Cond0, X, (select Cond1, X, Y)
6255 if (N0->getOpcode() == ISD::OR && N0->hasOneUse()) {
6256 SDValue Cond0 = N0->getOperand(0);
6257 SDValue Cond1 = N0->getOperand(1);
6258 SDValue InnerSelect = DAG.getNode(ISD::SELECT, SDLoc(N),
6259 N1.getValueType(), Cond1, N1, N2);
6260 if (normalizeToSequence || !InnerSelect.use_empty())
6261 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Cond0, N1,
6265 // select Cond0, (select Cond1, X, Y), Y -> select (and Cond0, Cond1), X, Y
6266 if (N1->getOpcode() == ISD::SELECT && N1->hasOneUse()) {
6267 SDValue N1_0 = N1->getOperand(0);
6268 SDValue N1_1 = N1->getOperand(1);
6269 SDValue N1_2 = N1->getOperand(2);
6270 if (N1_2 == N2 && N0.getValueType() == N1_0.getValueType()) {
6271 // Create the actual and node if we can generate good code for it.
6272 if (!normalizeToSequence) {
6273 SDValue And = DAG.getNode(ISD::AND, SDLoc(N), N0.getValueType(),
6275 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), And,
6278 // Otherwise see if we can optimize the "and" to a better pattern.
6279 if (SDValue Combined = visitANDLike(N0, N1_0, N))
6280 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Combined,
6284 // select Cond0, X, (select Cond1, X, Y) -> select (or Cond0, Cond1), X, Y
6285 if (N2->getOpcode() == ISD::SELECT && N2->hasOneUse()) {
6286 SDValue N2_0 = N2->getOperand(0);
6287 SDValue N2_1 = N2->getOperand(1);
6288 SDValue N2_2 = N2->getOperand(2);
6289 if (N2_1 == N1 && N0.getValueType() == N2_0.getValueType()) {
6290 // Create the actual or node if we can generate good code for it.
6291 if (!normalizeToSequence) {
6292 SDValue Or = DAG.getNode(ISD::OR, SDLoc(N), N0.getValueType(),
6294 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Or,
6297 // Otherwise see if we can optimize to a better pattern.
6298 if (SDValue Combined = visitORLike(N0, N2_0, N))
6299 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Combined,
6305 // select (xor Cond, 1), X, Y -> select Cond, Y, X
6306 if (VT0 == MVT::i1) {
6307 if (N0->getOpcode() == ISD::XOR) {
6308 if (auto *C = dyn_cast<ConstantSDNode>(N0->getOperand(1))) {
6309 SDValue Cond0 = N0->getOperand(0);
6311 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(),
6317 // fold selects based on a setcc into other things, such as min/max/abs
6318 if (N0.getOpcode() == ISD::SETCC) {
6319 // select x, y (fcmp lt x, y) -> fminnum x, y
6320 // select x, y (fcmp gt x, y) -> fmaxnum x, y
6322 // This is OK if we don't care about what happens if either operand is a
6326 // FIXME: Instead of testing for UnsafeFPMath, this should be checking for
6327 // no signed zeros as well as no nans.
6328 const TargetOptions &Options = DAG.getTarget().Options;
6329 if (Options.UnsafeFPMath &&
6330 VT.isFloatingPoint() && N0.hasOneUse() &&
6331 DAG.isKnownNeverNaN(N1) && DAG.isKnownNeverNaN(N2)) {
6332 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
6334 if (SDValue FMinMax = combineMinNumMaxNum(SDLoc(N), VT, N0.getOperand(0),
6335 N0.getOperand(1), N1, N2, CC,
6340 if ((!LegalOperations &&
6341 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT)) ||
6342 TLI.isOperationLegal(ISD::SELECT_CC, VT))
6343 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT,
6344 N0.getOperand(0), N0.getOperand(1),
6345 N1, N2, N0.getOperand(2));
6346 return SimplifySelect(SDLoc(N), N0, N1, N2);
6353 std::pair<SDValue, SDValue> SplitVSETCC(const SDNode *N, SelectionDAG &DAG) {
6356 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
6358 // Split the inputs.
6359 SDValue Lo, Hi, LL, LH, RL, RH;
6360 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
6361 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
6363 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));
6364 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));
6366 return std::make_pair(Lo, Hi);
6369 // This function assumes all the vselect's arguments are CONCAT_VECTOR
6370 // nodes and that the condition is a BV of ConstantSDNodes (or undefs).
6371 static SDValue ConvertSelectToConcatVector(SDNode *N, SelectionDAG &DAG) {
6373 SDValue Cond = N->getOperand(0);
6374 SDValue LHS = N->getOperand(1);
6375 SDValue RHS = N->getOperand(2);
6376 EVT VT = N->getValueType(0);
6377 int NumElems = VT.getVectorNumElements();
6378 assert(LHS.getOpcode() == ISD::CONCAT_VECTORS &&
6379 RHS.getOpcode() == ISD::CONCAT_VECTORS &&
6380 Cond.getOpcode() == ISD::BUILD_VECTOR);
6382 // CONCAT_VECTOR can take an arbitrary number of arguments. We only care about
6383 // binary ones here.
6384 if (LHS->getNumOperands() != 2 || RHS->getNumOperands() != 2)
6387 // We're sure we have an even number of elements due to the
6388 // concat_vectors we have as arguments to vselect.
6389 // Skip BV elements until we find one that's not an UNDEF
6390 // After we find an UNDEF element, keep looping until we get to half the
6391 // length of the BV and see if all the non-undef nodes are the same.
6392 ConstantSDNode *BottomHalf = nullptr;
6393 for (int i = 0; i < NumElems / 2; ++i) {
6394 if (Cond->getOperand(i)->isUndef())
6397 if (BottomHalf == nullptr)
6398 BottomHalf = cast<ConstantSDNode>(Cond.getOperand(i));
6399 else if (Cond->getOperand(i).getNode() != BottomHalf)
6403 // Do the same for the second half of the BuildVector
6404 ConstantSDNode *TopHalf = nullptr;
6405 for (int i = NumElems / 2; i < NumElems; ++i) {
6406 if (Cond->getOperand(i)->isUndef())
6409 if (TopHalf == nullptr)
6410 TopHalf = cast<ConstantSDNode>(Cond.getOperand(i));
6411 else if (Cond->getOperand(i).getNode() != TopHalf)
6415 assert(TopHalf && BottomHalf &&
6416 "One half of the selector was all UNDEFs and the other was all the "
6417 "same value. This should have been addressed before this function.");
6419 ISD::CONCAT_VECTORS, DL, VT,
6420 BottomHalf->isNullValue() ? RHS->getOperand(0) : LHS->getOperand(0),
6421 TopHalf->isNullValue() ? RHS->getOperand(1) : LHS->getOperand(1));
6424 SDValue DAGCombiner::visitMSCATTER(SDNode *N) {
6426 if (Level >= AfterLegalizeTypes)
6429 MaskedScatterSDNode *MSC = cast<MaskedScatterSDNode>(N);
6430 SDValue Mask = MSC->getMask();
6431 SDValue Data = MSC->getValue();
6434 // If the MSCATTER data type requires splitting and the mask is provided by a
6435 // SETCC, then split both nodes and its operands before legalization. This
6436 // prevents the type legalizer from unrolling SETCC into scalar comparisons
6437 // and enables future optimizations (e.g. min/max pattern matching on X86).
6438 if (Mask.getOpcode() != ISD::SETCC)
6441 // Check if any splitting is required.
6442 if (TLI.getTypeAction(*DAG.getContext(), Data.getValueType()) !=
6443 TargetLowering::TypeSplitVector)
6445 SDValue MaskLo, MaskHi, Lo, Hi;
6446 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
6449 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MSC->getValueType(0));
6451 SDValue Chain = MSC->getChain();
6453 EVT MemoryVT = MSC->getMemoryVT();
6454 unsigned Alignment = MSC->getOriginalAlignment();
6456 EVT LoMemVT, HiMemVT;
6457 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
6459 SDValue DataLo, DataHi;
6460 std::tie(DataLo, DataHi) = DAG.SplitVector(Data, DL);
6462 SDValue BasePtr = MSC->getBasePtr();
6463 SDValue IndexLo, IndexHi;
6464 std::tie(IndexLo, IndexHi) = DAG.SplitVector(MSC->getIndex(), DL);
6466 MachineMemOperand *MMO = DAG.getMachineFunction().
6467 getMachineMemOperand(MSC->getPointerInfo(),
6468 MachineMemOperand::MOStore, LoMemVT.getStoreSize(),
6469 Alignment, MSC->getAAInfo(), MSC->getRanges());
6471 SDValue OpsLo[] = { Chain, DataLo, MaskLo, BasePtr, IndexLo };
6472 Lo = DAG.getMaskedScatter(DAG.getVTList(MVT::Other), DataLo.getValueType(),
6475 SDValue OpsHi[] = {Chain, DataHi, MaskHi, BasePtr, IndexHi};
6476 Hi = DAG.getMaskedScatter(DAG.getVTList(MVT::Other), DataHi.getValueType(),
6479 AddToWorklist(Lo.getNode());
6480 AddToWorklist(Hi.getNode());
6482 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);
6485 SDValue DAGCombiner::visitMSTORE(SDNode *N) {
6487 if (Level >= AfterLegalizeTypes)
6490 MaskedStoreSDNode *MST = dyn_cast<MaskedStoreSDNode>(N);
6491 SDValue Mask = MST->getMask();
6492 SDValue Data = MST->getValue();
6493 EVT VT = Data.getValueType();
6496 // If the MSTORE data type requires splitting and the mask is provided by a
6497 // SETCC, then split both nodes and its operands before legalization. This
6498 // prevents the type legalizer from unrolling SETCC into scalar comparisons
6499 // and enables future optimizations (e.g. min/max pattern matching on X86).
6500 if (Mask.getOpcode() == ISD::SETCC) {
6502 // Check if any splitting is required.
6503 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
6504 TargetLowering::TypeSplitVector)
6507 SDValue MaskLo, MaskHi, Lo, Hi;
6508 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
6510 SDValue Chain = MST->getChain();
6511 SDValue Ptr = MST->getBasePtr();
6513 EVT MemoryVT = MST->getMemoryVT();
6514 unsigned Alignment = MST->getOriginalAlignment();
6516 // if Alignment is equal to the vector size,
6517 // take the half of it for the second part
6518 unsigned SecondHalfAlignment =
6519 (Alignment == VT.getSizeInBits() / 8) ? Alignment / 2 : Alignment;
6521 EVT LoMemVT, HiMemVT;
6522 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
6524 SDValue DataLo, DataHi;
6525 std::tie(DataLo, DataHi) = DAG.SplitVector(Data, DL);
6527 MachineMemOperand *MMO = DAG.getMachineFunction().
6528 getMachineMemOperand(MST->getPointerInfo(),
6529 MachineMemOperand::MOStore, LoMemVT.getStoreSize(),
6530 Alignment, MST->getAAInfo(), MST->getRanges());
6532 Lo = DAG.getMaskedStore(Chain, DL, DataLo, Ptr, MaskLo, LoMemVT, MMO,
6533 MST->isTruncatingStore(),
6534 MST->isCompressingStore());
6536 Ptr = TLI.IncrementMemoryAddress(Ptr, MaskLo, DL, LoMemVT, DAG,
6537 MST->isCompressingStore());
6539 MMO = DAG.getMachineFunction().
6540 getMachineMemOperand(MST->getPointerInfo(),
6541 MachineMemOperand::MOStore, HiMemVT.getStoreSize(),
6542 SecondHalfAlignment, MST->getAAInfo(),
6545 Hi = DAG.getMaskedStore(Chain, DL, DataHi, Ptr, MaskHi, HiMemVT, MMO,
6546 MST->isTruncatingStore(),
6547 MST->isCompressingStore());
6549 AddToWorklist(Lo.getNode());
6550 AddToWorklist(Hi.getNode());
6552 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);
6557 SDValue DAGCombiner::visitMGATHER(SDNode *N) {
6559 if (Level >= AfterLegalizeTypes)
6562 MaskedGatherSDNode *MGT = dyn_cast<MaskedGatherSDNode>(N);
6563 SDValue Mask = MGT->getMask();
6566 // If the MGATHER result requires splitting and the mask is provided by a
6567 // SETCC, then split both nodes and its operands before legalization. This
6568 // prevents the type legalizer from unrolling SETCC into scalar comparisons
6569 // and enables future optimizations (e.g. min/max pattern matching on X86).
6571 if (Mask.getOpcode() != ISD::SETCC)
6574 EVT VT = N->getValueType(0);
6576 // Check if any splitting is required.
6577 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
6578 TargetLowering::TypeSplitVector)
6581 SDValue MaskLo, MaskHi, Lo, Hi;
6582 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
6584 SDValue Src0 = MGT->getValue();
6585 SDValue Src0Lo, Src0Hi;
6586 std::tie(Src0Lo, Src0Hi) = DAG.SplitVector(Src0, DL);
6589 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
6591 SDValue Chain = MGT->getChain();
6592 EVT MemoryVT = MGT->getMemoryVT();
6593 unsigned Alignment = MGT->getOriginalAlignment();
6595 EVT LoMemVT, HiMemVT;
6596 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
6598 SDValue BasePtr = MGT->getBasePtr();
6599 SDValue Index = MGT->getIndex();
6600 SDValue IndexLo, IndexHi;
6601 std::tie(IndexLo, IndexHi) = DAG.SplitVector(Index, DL);
6603 MachineMemOperand *MMO = DAG.getMachineFunction().
6604 getMachineMemOperand(MGT->getPointerInfo(),
6605 MachineMemOperand::MOLoad, LoMemVT.getStoreSize(),
6606 Alignment, MGT->getAAInfo(), MGT->getRanges());
6608 SDValue OpsLo[] = { Chain, Src0Lo, MaskLo, BasePtr, IndexLo };
6609 Lo = DAG.getMaskedGather(DAG.getVTList(LoVT, MVT::Other), LoVT, DL, OpsLo,
6612 SDValue OpsHi[] = {Chain, Src0Hi, MaskHi, BasePtr, IndexHi};
6613 Hi = DAG.getMaskedGather(DAG.getVTList(HiVT, MVT::Other), HiVT, DL, OpsHi,
6616 AddToWorklist(Lo.getNode());
6617 AddToWorklist(Hi.getNode());
6619 // Build a factor node to remember that this load is independent of the
6621 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo.getValue(1),
6624 // Legalized the chain result - switch anything that used the old chain to
6626 DAG.ReplaceAllUsesOfValueWith(SDValue(MGT, 1), Chain);
6628 SDValue GatherRes = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
6630 SDValue RetOps[] = { GatherRes, Chain };
6631 return DAG.getMergeValues(RetOps, DL);
6634 SDValue DAGCombiner::visitMLOAD(SDNode *N) {
6636 if (Level >= AfterLegalizeTypes)
6639 MaskedLoadSDNode *MLD = dyn_cast<MaskedLoadSDNode>(N);
6640 SDValue Mask = MLD->getMask();
6643 // If the MLOAD result requires splitting and the mask is provided by a
6644 // SETCC, then split both nodes and its operands before legalization. This
6645 // prevents the type legalizer from unrolling SETCC into scalar comparisons
6646 // and enables future optimizations (e.g. min/max pattern matching on X86).
6648 if (Mask.getOpcode() == ISD::SETCC) {
6649 EVT VT = N->getValueType(0);
6651 // Check if any splitting is required.
6652 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
6653 TargetLowering::TypeSplitVector)
6656 SDValue MaskLo, MaskHi, Lo, Hi;
6657 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
6659 SDValue Src0 = MLD->getSrc0();
6660 SDValue Src0Lo, Src0Hi;
6661 std::tie(Src0Lo, Src0Hi) = DAG.SplitVector(Src0, DL);
6664 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MLD->getValueType(0));
6666 SDValue Chain = MLD->getChain();
6667 SDValue Ptr = MLD->getBasePtr();
6668 EVT MemoryVT = MLD->getMemoryVT();
6669 unsigned Alignment = MLD->getOriginalAlignment();
6671 // if Alignment is equal to the vector size,
6672 // take the half of it for the second part
6673 unsigned SecondHalfAlignment =
6674 (Alignment == MLD->getValueType(0).getSizeInBits()/8) ?
6675 Alignment/2 : Alignment;
6677 EVT LoMemVT, HiMemVT;
6678 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
6680 MachineMemOperand *MMO = DAG.getMachineFunction().
6681 getMachineMemOperand(MLD->getPointerInfo(),
6682 MachineMemOperand::MOLoad, LoMemVT.getStoreSize(),
6683 Alignment, MLD->getAAInfo(), MLD->getRanges());
6685 Lo = DAG.getMaskedLoad(LoVT, DL, Chain, Ptr, MaskLo, Src0Lo, LoMemVT, MMO,
6686 ISD::NON_EXTLOAD, MLD->isExpandingLoad());
6688 Ptr = TLI.IncrementMemoryAddress(Ptr, MaskLo, DL, LoMemVT, DAG,
6689 MLD->isExpandingLoad());
6691 MMO = DAG.getMachineFunction().
6692 getMachineMemOperand(MLD->getPointerInfo(),
6693 MachineMemOperand::MOLoad, HiMemVT.getStoreSize(),
6694 SecondHalfAlignment, MLD->getAAInfo(), MLD->getRanges());
6696 Hi = DAG.getMaskedLoad(HiVT, DL, Chain, Ptr, MaskHi, Src0Hi, HiMemVT, MMO,
6697 ISD::NON_EXTLOAD, MLD->isExpandingLoad());
6699 AddToWorklist(Lo.getNode());
6700 AddToWorklist(Hi.getNode());
6702 // Build a factor node to remember that this load is independent of the
6704 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo.getValue(1),
6707 // Legalized the chain result - switch anything that used the old chain to
6709 DAG.ReplaceAllUsesOfValueWith(SDValue(MLD, 1), Chain);
6711 SDValue LoadRes = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
6713 SDValue RetOps[] = { LoadRes, Chain };
6714 return DAG.getMergeValues(RetOps, DL);
6719 SDValue DAGCombiner::visitVSELECT(SDNode *N) {
6720 SDValue N0 = N->getOperand(0);
6721 SDValue N1 = N->getOperand(1);
6722 SDValue N2 = N->getOperand(2);
6725 // fold (vselect C, X, X) -> X
6729 // Canonicalize integer abs.
6730 // vselect (setg[te] X, 0), X, -X ->
6731 // vselect (setgt X, -1), X, -X ->
6732 // vselect (setl[te] X, 0), -X, X ->
6733 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
6734 if (N0.getOpcode() == ISD::SETCC) {
6735 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
6736 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
6738 bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
6740 if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
6741 (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) &&
6742 N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1))
6743 isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode());
6744 else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) &&
6745 N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1))
6746 isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode());
6749 EVT VT = LHS.getValueType();
6750 if (TLI.isOperationLegalOrCustom(ISD::ABS, VT))
6751 return DAG.getNode(ISD::ABS, DL, VT, LHS);
6753 SDValue Shift = DAG.getNode(
6754 ISD::SRA, DL, VT, LHS,
6755 DAG.getConstant(VT.getScalarSizeInBits() - 1, DL, VT));
6756 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift);
6757 AddToWorklist(Shift.getNode());
6758 AddToWorklist(Add.getNode());
6759 return DAG.getNode(ISD::XOR, DL, VT, Add, Shift);
6763 if (SimplifySelectOps(N, N1, N2))
6764 return SDValue(N, 0); // Don't revisit N.
6766 // Fold (vselect (build_vector all_ones), N1, N2) -> N1
6767 if (ISD::isBuildVectorAllOnes(N0.getNode()))
6769 // Fold (vselect (build_vector all_zeros), N1, N2) -> N2
6770 if (ISD::isBuildVectorAllZeros(N0.getNode()))
6773 // The ConvertSelectToConcatVector function is assuming both the above
6774 // checks for (vselect (build_vector all{ones,zeros) ...) have been made
6776 if (N1.getOpcode() == ISD::CONCAT_VECTORS &&
6777 N2.getOpcode() == ISD::CONCAT_VECTORS &&
6778 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
6779 if (SDValue CV = ConvertSelectToConcatVector(N, DAG))
6786 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
6787 SDValue N0 = N->getOperand(0);
6788 SDValue N1 = N->getOperand(1);
6789 SDValue N2 = N->getOperand(2);
6790 SDValue N3 = N->getOperand(3);
6791 SDValue N4 = N->getOperand(4);
6792 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
6794 // fold select_cc lhs, rhs, x, x, cc -> x
6798 // Determine if the condition we're dealing with is constant
6799 if (SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()), N0, N1,
6800 CC, SDLoc(N), false)) {
6801 AddToWorklist(SCC.getNode());
6803 if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) {
6804 if (!SCCC->isNullValue())
6805 return N2; // cond always true -> true val
6807 return N3; // cond always false -> false val
6808 } else if (SCC->isUndef()) {
6809 // When the condition is UNDEF, just return the first operand. This is
6810 // coherent the DAG creation, no setcc node is created in this case
6812 } else if (SCC.getOpcode() == ISD::SETCC) {
6813 // Fold to a simpler select_cc
6814 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(),
6815 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
6820 // If we can fold this based on the true/false value, do so.
6821 if (SimplifySelectOps(N, N2, N3))
6822 return SDValue(N, 0); // Don't revisit N.
6824 // fold select_cc into other things, such as min/max/abs
6825 return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC);
6828 SDValue DAGCombiner::visitSETCC(SDNode *N) {
6829 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
6830 cast<CondCodeSDNode>(N->getOperand(2))->get(),
6834 SDValue DAGCombiner::visitSETCCE(SDNode *N) {
6835 SDValue LHS = N->getOperand(0);
6836 SDValue RHS = N->getOperand(1);
6837 SDValue Carry = N->getOperand(2);
6838 SDValue Cond = N->getOperand(3);
6840 // If Carry is false, fold to a regular SETCC.
6841 if (Carry.getOpcode() == ISD::CARRY_FALSE)
6842 return DAG.getNode(ISD::SETCC, SDLoc(N), N->getVTList(), LHS, RHS, Cond);
6847 SDValue DAGCombiner::visitSETCCCARRY(SDNode *N) {
6848 SDValue LHS = N->getOperand(0);
6849 SDValue RHS = N->getOperand(1);
6850 SDValue Carry = N->getOperand(2);
6851 SDValue Cond = N->getOperand(3);
6853 // If Carry is false, fold to a regular SETCC.
6854 if (isNullConstant(Carry))
6855 return DAG.getNode(ISD::SETCC, SDLoc(N), N->getVTList(), LHS, RHS, Cond);
6860 /// Try to fold a sext/zext/aext dag node into a ConstantSDNode or
6861 /// a build_vector of constants.
6862 /// This function is called by the DAGCombiner when visiting sext/zext/aext
6863 /// dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND).
6864 /// Vector extends are not folded if operations are legal; this is to
6865 /// avoid introducing illegal build_vector dag nodes.
6866 static SDNode *tryToFoldExtendOfConstant(SDNode *N, const TargetLowering &TLI,
6867 SelectionDAG &DAG, bool LegalTypes,
6868 bool LegalOperations) {
6869 unsigned Opcode = N->getOpcode();
6870 SDValue N0 = N->getOperand(0);
6871 EVT VT = N->getValueType(0);
6873 assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND ||
6874 Opcode == ISD::ANY_EXTEND || Opcode == ISD::SIGN_EXTEND_VECTOR_INREG ||
6875 Opcode == ISD::ZERO_EXTEND_VECTOR_INREG)
6876 && "Expected EXTEND dag node in input!");
6878 // fold (sext c1) -> c1
6879 // fold (zext c1) -> c1
6880 // fold (aext c1) -> c1
6881 if (isa<ConstantSDNode>(N0))
6882 return DAG.getNode(Opcode, SDLoc(N), VT, N0).getNode();
6884 // fold (sext (build_vector AllConstants) -> (build_vector AllConstants)
6885 // fold (zext (build_vector AllConstants) -> (build_vector AllConstants)
6886 // fold (aext (build_vector AllConstants) -> (build_vector AllConstants)
6887 EVT SVT = VT.getScalarType();
6888 if (!(VT.isVector() &&
6889 (!LegalTypes || (!LegalOperations && TLI.isTypeLegal(SVT))) &&
6890 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())))
6893 // We can fold this node into a build_vector.
6894 unsigned VTBits = SVT.getSizeInBits();
6895 unsigned EVTBits = N0->getValueType(0).getScalarSizeInBits();
6896 SmallVector<SDValue, 8> Elts;
6897 unsigned NumElts = VT.getVectorNumElements();
6900 for (unsigned i=0; i != NumElts; ++i) {
6901 SDValue Op = N0->getOperand(i);
6902 if (Op->isUndef()) {
6903 Elts.push_back(DAG.getUNDEF(SVT));
6908 // Get the constant value and if needed trunc it to the size of the type.
6909 // Nodes like build_vector might have constants wider than the scalar type.
6910 APInt C = cast<ConstantSDNode>(Op)->getAPIntValue().zextOrTrunc(EVTBits);
6911 if (Opcode == ISD::SIGN_EXTEND || Opcode == ISD::SIGN_EXTEND_VECTOR_INREG)
6912 Elts.push_back(DAG.getConstant(C.sext(VTBits), DL, SVT));
6914 Elts.push_back(DAG.getConstant(C.zext(VTBits), DL, SVT));
6917 return DAG.getBuildVector(VT, DL, Elts).getNode();
6920 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
6921 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
6922 // transformation. Returns true if extension are possible and the above
6923 // mentioned transformation is profitable.
6924 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
6926 SmallVectorImpl<SDNode *> &ExtendNodes,
6927 const TargetLowering &TLI) {
6928 bool HasCopyToRegUses = false;
6929 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
6930 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
6931 UE = N0.getNode()->use_end();
6936 if (UI.getUse().getResNo() != N0.getResNo())
6938 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
6939 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
6940 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
6941 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
6942 // Sign bits will be lost after a zext.
6945 for (unsigned i = 0; i != 2; ++i) {
6946 SDValue UseOp = User->getOperand(i);
6949 if (!isa<ConstantSDNode>(UseOp))
6954 ExtendNodes.push_back(User);
6957 // If truncates aren't free and there are users we can't
6958 // extend, it isn't worthwhile.
6961 // Remember if this value is live-out.
6962 if (User->getOpcode() == ISD::CopyToReg)
6963 HasCopyToRegUses = true;
6966 if (HasCopyToRegUses) {
6967 bool BothLiveOut = false;
6968 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
6970 SDUse &Use = UI.getUse();
6971 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
6977 // Both unextended and extended values are live out. There had better be
6978 // a good reason for the transformation.
6979 return ExtendNodes.size();
6984 void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
6985 SDValue Trunc, SDValue ExtLoad,
6986 const SDLoc &DL, ISD::NodeType ExtType) {
6987 // Extend SetCC uses if necessary.
6988 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
6989 SDNode *SetCC = SetCCs[i];
6990 SmallVector<SDValue, 4> Ops;
6992 for (unsigned j = 0; j != 2; ++j) {
6993 SDValue SOp = SetCC->getOperand(j);
6995 Ops.push_back(ExtLoad);
6997 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
7000 Ops.push_back(SetCC->getOperand(2));
7001 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), Ops));
7005 // FIXME: Bring more similar combines here, common to sext/zext (maybe aext?).
7006 SDValue DAGCombiner::CombineExtLoad(SDNode *N) {
7007 SDValue N0 = N->getOperand(0);
7008 EVT DstVT = N->getValueType(0);
7009 EVT SrcVT = N0.getValueType();
7011 assert((N->getOpcode() == ISD::SIGN_EXTEND ||
7012 N->getOpcode() == ISD::ZERO_EXTEND) &&
7013 "Unexpected node type (not an extend)!");
7015 // fold (sext (load x)) to multiple smaller sextloads; same for zext.
7016 // For example, on a target with legal v4i32, but illegal v8i32, turn:
7017 // (v8i32 (sext (v8i16 (load x))))
7019 // (v8i32 (concat_vectors (v4i32 (sextload x)),
7020 // (v4i32 (sextload (x + 16)))))
7021 // Where uses of the original load, i.e.:
7023 // are replaced with:
7025 // (v8i32 (concat_vectors (v4i32 (sextload x)),
7026 // (v4i32 (sextload (x + 16)))))))
7028 // This combine is only applicable to illegal, but splittable, vectors.
7029 // All legal types, and illegal non-vector types, are handled elsewhere.
7030 // This combine is controlled by TargetLowering::isVectorLoadExtDesirable.
7032 if (N0->getOpcode() != ISD::LOAD)
7035 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7037 if (!ISD::isNON_EXTLoad(LN0) || !ISD::isUNINDEXEDLoad(LN0) ||
7038 !N0.hasOneUse() || LN0->isVolatile() || !DstVT.isVector() ||
7039 !DstVT.isPow2VectorType() || !TLI.isVectorLoadExtDesirable(SDValue(N, 0)))
7042 SmallVector<SDNode *, 4> SetCCs;
7043 if (!ExtendUsesToFormExtLoad(N, N0, N->getOpcode(), SetCCs, TLI))
7046 ISD::LoadExtType ExtType =
7047 N->getOpcode() == ISD::SIGN_EXTEND ? ISD::SEXTLOAD : ISD::ZEXTLOAD;
7049 // Try to split the vector types to get down to legal types.
7050 EVT SplitSrcVT = SrcVT;
7051 EVT SplitDstVT = DstVT;
7052 while (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT) &&
7053 SplitSrcVT.getVectorNumElements() > 1) {
7054 SplitDstVT = DAG.GetSplitDestVTs(SplitDstVT).first;
7055 SplitSrcVT = DAG.GetSplitDestVTs(SplitSrcVT).first;
7058 if (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT))
7062 const unsigned NumSplits =
7063 DstVT.getVectorNumElements() / SplitDstVT.getVectorNumElements();
7064 const unsigned Stride = SplitSrcVT.getStoreSize();
7065 SmallVector<SDValue, 4> Loads;
7066 SmallVector<SDValue, 4> Chains;
7068 SDValue BasePtr = LN0->getBasePtr();
7069 for (unsigned Idx = 0; Idx < NumSplits; Idx++) {
7070 const unsigned Offset = Idx * Stride;
7071 const unsigned Align = MinAlign(LN0->getAlignment(), Offset);
7073 SDValue SplitLoad = DAG.getExtLoad(
7074 ExtType, DL, SplitDstVT, LN0->getChain(), BasePtr,
7075 LN0->getPointerInfo().getWithOffset(Offset), SplitSrcVT, Align,
7076 LN0->getMemOperand()->getFlags(), LN0->getAAInfo());
7078 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
7079 DAG.getConstant(Stride, DL, BasePtr.getValueType()));
7081 Loads.push_back(SplitLoad.getValue(0));
7082 Chains.push_back(SplitLoad.getValue(1));
7085 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
7086 SDValue NewValue = DAG.getNode(ISD::CONCAT_VECTORS, DL, DstVT, Loads);
7089 AddToWorklist(NewChain.getNode());
7091 CombineTo(N, NewValue);
7093 // Replace uses of the original load (before extension)
7094 // with a truncate of the concatenated sextloaded vectors.
7096 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(), NewValue);
7097 CombineTo(N0.getNode(), Trunc, NewChain);
7098 ExtendSetCCUses(SetCCs, Trunc, NewValue, DL,
7099 (ISD::NodeType)N->getOpcode());
7100 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7103 /// If we're narrowing or widening the result of a vector select and the final
7104 /// size is the same size as a setcc (compare) feeding the select, then try to
7105 /// apply the cast operation to the select's operands because matching vector
7106 /// sizes for a select condition and other operands should be more efficient.
7107 SDValue DAGCombiner::matchVSelectOpSizesWithSetCC(SDNode *Cast) {
7108 unsigned CastOpcode = Cast->getOpcode();
7109 assert((CastOpcode == ISD::SIGN_EXTEND || CastOpcode == ISD::ZERO_EXTEND ||
7110 CastOpcode == ISD::TRUNCATE || CastOpcode == ISD::FP_EXTEND ||
7111 CastOpcode == ISD::FP_ROUND) &&
7112 "Unexpected opcode for vector select narrowing/widening");
7114 // We only do this transform before legal ops because the pattern may be
7115 // obfuscated by target-specific operations after legalization. Do not create
7116 // an illegal select op, however, because that may be difficult to lower.
7117 EVT VT = Cast->getValueType(0);
7118 if (LegalOperations || !TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
7121 SDValue VSel = Cast->getOperand(0);
7122 if (VSel.getOpcode() != ISD::VSELECT || !VSel.hasOneUse() ||
7123 VSel.getOperand(0).getOpcode() != ISD::SETCC)
7126 // Does the setcc have the same vector size as the casted select?
7127 SDValue SetCC = VSel.getOperand(0);
7128 EVT SetCCVT = getSetCCResultType(SetCC.getOperand(0).getValueType());
7129 if (SetCCVT.getSizeInBits() != VT.getSizeInBits())
7132 // cast (vsel (setcc X), A, B) --> vsel (setcc X), (cast A), (cast B)
7133 SDValue A = VSel.getOperand(1);
7134 SDValue B = VSel.getOperand(2);
7135 SDValue CastA, CastB;
7137 if (CastOpcode == ISD::FP_ROUND) {
7138 // FP_ROUND (fptrunc) has an extra flag operand to pass along.
7139 CastA = DAG.getNode(CastOpcode, DL, VT, A, Cast->getOperand(1));
7140 CastB = DAG.getNode(CastOpcode, DL, VT, B, Cast->getOperand(1));
7142 CastA = DAG.getNode(CastOpcode, DL, VT, A);
7143 CastB = DAG.getNode(CastOpcode, DL, VT, B);
7145 return DAG.getNode(ISD::VSELECT, DL, VT, SetCC, CastA, CastB);
7148 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
7149 SDValue N0 = N->getOperand(0);
7150 EVT VT = N->getValueType(0);
7153 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
7155 return SDValue(Res, 0);
7157 // fold (sext (sext x)) -> (sext x)
7158 // fold (sext (aext x)) -> (sext x)
7159 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
7160 return DAG.getNode(ISD::SIGN_EXTEND, DL, VT, N0.getOperand(0));
7162 if (N0.getOpcode() == ISD::TRUNCATE) {
7163 // fold (sext (truncate (load x))) -> (sext (smaller load x))
7164 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
7165 if (SDValue NarrowLoad = ReduceLoadWidth(N0.getNode())) {
7166 SDNode *oye = N0.getOperand(0).getNode();
7167 if (NarrowLoad.getNode() != N0.getNode()) {
7168 CombineTo(N0.getNode(), NarrowLoad);
7169 // CombineTo deleted the truncate, if needed, but not what's under it.
7172 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7175 // See if the value being truncated is already sign extended. If so, just
7176 // eliminate the trunc/sext pair.
7177 SDValue Op = N0.getOperand(0);
7178 unsigned OpBits = Op.getScalarValueSizeInBits();
7179 unsigned MidBits = N0.getScalarValueSizeInBits();
7180 unsigned DestBits = VT.getScalarSizeInBits();
7181 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
7183 if (OpBits == DestBits) {
7184 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
7185 // bits, it is already ready.
7186 if (NumSignBits > DestBits-MidBits)
7188 } else if (OpBits < DestBits) {
7189 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
7190 // bits, just sext from i32.
7191 if (NumSignBits > OpBits-MidBits)
7192 return DAG.getNode(ISD::SIGN_EXTEND, DL, VT, Op);
7194 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
7195 // bits, just truncate to i32.
7196 if (NumSignBits > OpBits-MidBits)
7197 return DAG.getNode(ISD::TRUNCATE, DL, VT, Op);
7200 // fold (sext (truncate x)) -> (sextinreg x).
7201 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
7202 N0.getValueType())) {
7203 if (OpBits < DestBits)
7204 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op);
7205 else if (OpBits > DestBits)
7206 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op);
7207 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, Op,
7208 DAG.getValueType(N0.getValueType()));
7212 // fold (sext (load x)) -> (sext (truncate (sextload x)))
7213 // Only generate vector extloads when 1) they're legal, and 2) they are
7214 // deemed desirable by the target.
7215 if (ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
7216 ((!LegalOperations && !VT.isVector() &&
7217 !cast<LoadSDNode>(N0)->isVolatile()) ||
7218 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, N0.getValueType()))) {
7219 bool DoXform = true;
7220 SmallVector<SDNode*, 4> SetCCs;
7221 if (!N0.hasOneUse())
7222 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
7224 DoXform &= TLI.isVectorLoadExtDesirable(SDValue(N, 0));
7226 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7227 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, DL, VT, LN0->getChain(),
7228 LN0->getBasePtr(), N0.getValueType(),
7229 LN0->getMemOperand());
7230 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
7231 N0.getValueType(), ExtLoad);
7232 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, DL, ISD::SIGN_EXTEND);
7233 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
7234 return CombineTo(N, ExtLoad); // Return N so it doesn't get rechecked!
7238 // fold (sext (load x)) to multiple smaller sextloads.
7239 // Only on illegal but splittable vectors.
7240 if (SDValue ExtLoad = CombineExtLoad(N))
7243 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
7244 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
7245 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
7246 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
7247 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7248 EVT MemVT = LN0->getMemoryVT();
7249 if ((!LegalOperations && !LN0->isVolatile()) ||
7250 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, MemVT)) {
7251 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, DL, VT, LN0->getChain(),
7252 LN0->getBasePtr(), MemVT,
7253 LN0->getMemOperand());
7254 CombineTo(N, ExtLoad);
7255 CombineTo(N0.getNode(),
7256 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
7257 N0.getValueType(), ExtLoad),
7258 ExtLoad.getValue(1));
7259 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7263 // fold (sext (and/or/xor (load x), cst)) ->
7264 // (and/or/xor (sextload x), (sext cst))
7265 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
7266 N0.getOpcode() == ISD::XOR) &&
7267 isa<LoadSDNode>(N0.getOperand(0)) &&
7268 N0.getOperand(1).getOpcode() == ISD::Constant &&
7269 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, N0.getValueType()) &&
7270 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
7271 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
7272 if (LN0->getExtensionType() != ISD::ZEXTLOAD && LN0->isUnindexed()) {
7273 bool DoXform = true;
7274 SmallVector<SDNode*, 4> SetCCs;
7275 if (!N0.hasOneUse())
7276 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
7279 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN0), VT,
7280 LN0->getChain(), LN0->getBasePtr(),
7282 LN0->getMemOperand());
7283 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
7284 Mask = Mask.sext(VT.getSizeInBits());
7285 SDValue And = DAG.getNode(N0.getOpcode(), DL, VT,
7286 ExtLoad, DAG.getConstant(Mask, DL, VT));
7287 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
7288 SDLoc(N0.getOperand(0)),
7289 N0.getOperand(0).getValueType(), ExtLoad);
7290 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, DL, ISD::SIGN_EXTEND);
7291 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
7292 return CombineTo(N, And); // Return N so it doesn't get rechecked!
7297 if (N0.getOpcode() == ISD::SETCC) {
7298 SDValue N00 = N0.getOperand(0);
7299 SDValue N01 = N0.getOperand(1);
7300 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
7301 EVT N00VT = N0.getOperand(0).getValueType();
7303 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
7304 // Only do this before legalize for now.
7305 if (VT.isVector() && !LegalOperations &&
7306 TLI.getBooleanContents(N00VT) ==
7307 TargetLowering::ZeroOrNegativeOneBooleanContent) {
7308 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
7309 // of the same size as the compared operands. Only optimize sext(setcc())
7310 // if this is the case.
7311 EVT SVT = getSetCCResultType(N00VT);
7313 // We know that the # elements of the results is the same as the
7314 // # elements of the compare (and the # elements of the compare result
7315 // for that matter). Check to see that they are the same size. If so,
7316 // we know that the element size of the sext'd result matches the
7317 // element size of the compare operands.
7318 if (VT.getSizeInBits() == SVT.getSizeInBits())
7319 return DAG.getSetCC(DL, VT, N00, N01, CC);
7321 // If the desired elements are smaller or larger than the source
7322 // elements, we can use a matching integer vector type and then
7323 // truncate/sign extend.
7324 EVT MatchingVecType = N00VT.changeVectorElementTypeToInteger();
7325 if (SVT == MatchingVecType) {
7326 SDValue VsetCC = DAG.getSetCC(DL, MatchingVecType, N00, N01, CC);
7327 return DAG.getSExtOrTrunc(VsetCC, DL, VT);
7331 // sext(setcc x, y, cc) -> (select (setcc x, y, cc), T, 0)
7332 // Here, T can be 1 or -1, depending on the type of the setcc and
7333 // getBooleanContents().
7334 unsigned SetCCWidth = N0.getScalarValueSizeInBits();
7336 // To determine the "true" side of the select, we need to know the high bit
7337 // of the value returned by the setcc if it evaluates to true.
7338 // If the type of the setcc is i1, then the true case of the select is just
7339 // sext(i1 1), that is, -1.
7340 // If the type of the setcc is larger (say, i8) then the value of the high
7341 // bit depends on getBooleanContents(), so ask TLI for a real "true" value
7342 // of the appropriate width.
7343 SDValue ExtTrueVal = (SetCCWidth == 1) ? DAG.getAllOnesConstant(DL, VT)
7344 : TLI.getConstTrueVal(DAG, VT, DL);
7345 SDValue Zero = DAG.getConstant(0, DL, VT);
7347 SimplifySelectCC(DL, N00, N01, ExtTrueVal, Zero, CC, true))
7350 if (!VT.isVector()) {
7351 EVT SetCCVT = getSetCCResultType(N00VT);
7352 // Don't do this transform for i1 because there's a select transform
7353 // that would reverse it.
7354 // TODO: We should not do this transform at all without a target hook
7355 // because a sext is likely cheaper than a select?
7356 if (SetCCVT.getScalarSizeInBits() != 1 &&
7357 (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, N00VT))) {
7358 SDValue SetCC = DAG.getSetCC(DL, SetCCVT, N00, N01, CC);
7359 return DAG.getSelect(DL, VT, SetCC, ExtTrueVal, Zero);
7364 // fold (sext x) -> (zext x) if the sign bit is known zero.
7365 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
7366 DAG.SignBitIsZero(N0))
7367 return DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0);
7369 if (SDValue NewVSel = matchVSelectOpSizesWithSetCC(N))
7375 // isTruncateOf - If N is a truncate of some other value, return true, record
7376 // the value being truncated in Op and which of Op's bits are zero/one in Known.
7377 // This function computes KnownBits to avoid a duplicated call to
7378 // computeKnownBits in the caller.
7379 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
7381 if (N->getOpcode() == ISD::TRUNCATE) {
7382 Op = N->getOperand(0);
7383 DAG.computeKnownBits(Op, Known);
7387 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
7388 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
7391 SDValue Op0 = N->getOperand(0);
7392 SDValue Op1 = N->getOperand(1);
7393 assert(Op0.getValueType() == Op1.getValueType());
7395 if (isNullConstant(Op0))
7397 else if (isNullConstant(Op1))
7402 DAG.computeKnownBits(Op, Known);
7404 if (!(Known.Zero | 1).isAllOnesValue())
7410 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
7411 SDValue N0 = N->getOperand(0);
7412 EVT VT = N->getValueType(0);
7414 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
7416 return SDValue(Res, 0);
7418 // fold (zext (zext x)) -> (zext x)
7419 // fold (zext (aext x)) -> (zext x)
7420 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
7421 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT,
7424 // fold (zext (truncate x)) -> (zext x) or
7425 // (zext (truncate x)) -> (truncate x)
7426 // This is valid when the truncated bits of x are already zero.
7427 // FIXME: We should extend this to work for vectors too.
7430 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, Known)) {
7431 APInt TruncatedBits =
7432 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
7433 APInt(Op.getValueSizeInBits(), 0) :
7434 APInt::getBitsSet(Op.getValueSizeInBits(),
7435 N0.getValueSizeInBits(),
7436 std::min(Op.getValueSizeInBits(),
7437 VT.getSizeInBits()));
7438 if (TruncatedBits.isSubsetOf(Known.Zero))
7439 return DAG.getZExtOrTrunc(Op, SDLoc(N), VT);
7442 // fold (zext (truncate (load x))) -> (zext (smaller load x))
7443 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
7444 if (N0.getOpcode() == ISD::TRUNCATE) {
7445 if (SDValue NarrowLoad = ReduceLoadWidth(N0.getNode())) {
7446 SDNode *oye = N0.getOperand(0).getNode();
7447 if (NarrowLoad.getNode() != N0.getNode()) {
7448 CombineTo(N0.getNode(), NarrowLoad);
7449 // CombineTo deleted the truncate, if needed, but not what's under it.
7452 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7456 // fold (zext (truncate x)) -> (and x, mask)
7457 if (N0.getOpcode() == ISD::TRUNCATE) {
7458 // fold (zext (truncate (load x))) -> (zext (smaller load x))
7459 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
7460 if (SDValue NarrowLoad = ReduceLoadWidth(N0.getNode())) {
7461 SDNode *oye = N0.getOperand(0).getNode();
7462 if (NarrowLoad.getNode() != N0.getNode()) {
7463 CombineTo(N0.getNode(), NarrowLoad);
7464 // CombineTo deleted the truncate, if needed, but not what's under it.
7467 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7470 EVT SrcVT = N0.getOperand(0).getValueType();
7471 EVT MinVT = N0.getValueType();
7473 // Try to mask before the extension to avoid having to generate a larger mask,
7474 // possibly over several sub-vectors.
7475 if (SrcVT.bitsLT(VT)) {
7476 if (!LegalOperations || (TLI.isOperationLegal(ISD::AND, SrcVT) &&
7477 TLI.isOperationLegal(ISD::ZERO_EXTEND, VT))) {
7478 SDValue Op = N0.getOperand(0);
7479 Op = DAG.getZeroExtendInReg(Op, SDLoc(N), MinVT.getScalarType());
7480 AddToWorklist(Op.getNode());
7481 return DAG.getZExtOrTrunc(Op, SDLoc(N), VT);
7485 if (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT)) {
7486 SDValue Op = DAG.getAnyExtOrTrunc(N0.getOperand(0), SDLoc(N), VT);
7487 AddToWorklist(Op.getNode());
7488 return DAG.getZeroExtendInReg(Op, SDLoc(N), MinVT.getScalarType());
7492 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
7493 // if either of the casts is not free.
7494 if (N0.getOpcode() == ISD::AND &&
7495 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
7496 N0.getOperand(1).getOpcode() == ISD::Constant &&
7497 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
7498 N0.getValueType()) ||
7499 !TLI.isZExtFree(N0.getValueType(), VT))) {
7500 SDValue X = N0.getOperand(0).getOperand(0);
7501 X = DAG.getAnyExtOrTrunc(X, SDLoc(X), VT);
7502 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
7503 Mask = Mask.zext(VT.getSizeInBits());
7505 return DAG.getNode(ISD::AND, DL, VT,
7506 X, DAG.getConstant(Mask, DL, VT));
7509 // fold (zext (load x)) -> (zext (truncate (zextload x)))
7510 // Only generate vector extloads when 1) they're legal, and 2) they are
7511 // deemed desirable by the target.
7512 if (ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
7513 ((!LegalOperations && !VT.isVector() &&
7514 !cast<LoadSDNode>(N0)->isVolatile()) ||
7515 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, N0.getValueType()))) {
7516 bool DoXform = true;
7517 SmallVector<SDNode*, 4> SetCCs;
7518 if (!N0.hasOneUse())
7519 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
7521 DoXform &= TLI.isVectorLoadExtDesirable(SDValue(N, 0));
7523 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7524 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
7526 LN0->getBasePtr(), N0.getValueType(),
7527 LN0->getMemOperand());
7529 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
7530 N0.getValueType(), ExtLoad);
7531 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N), ISD::ZERO_EXTEND);
7532 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
7533 return CombineTo(N, ExtLoad); // Return N so it doesn't get rechecked!
7537 // fold (zext (load x)) to multiple smaller zextloads.
7538 // Only on illegal but splittable vectors.
7539 if (SDValue ExtLoad = CombineExtLoad(N))
7542 // fold (zext (and/or/xor (load x), cst)) ->
7543 // (and/or/xor (zextload x), (zext cst))
7544 // Unless (and (load x) cst) will match as a zextload already and has
7545 // additional users.
7546 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
7547 N0.getOpcode() == ISD::XOR) &&
7548 isa<LoadSDNode>(N0.getOperand(0)) &&
7549 N0.getOperand(1).getOpcode() == ISD::Constant &&
7550 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, N0.getValueType()) &&
7551 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
7552 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
7553 if (LN0->getExtensionType() != ISD::SEXTLOAD && LN0->isUnindexed()) {
7554 bool DoXform = true;
7555 SmallVector<SDNode*, 4> SetCCs;
7556 if (!N0.hasOneUse()) {
7557 if (N0.getOpcode() == ISD::AND) {
7558 auto *AndC = cast<ConstantSDNode>(N0.getOperand(1));
7559 auto NarrowLoad = false;
7560 EVT LoadResultTy = AndC->getValueType(0);
7561 EVT ExtVT, LoadedVT;
7562 if (isAndLoadExtLoad(AndC, LN0, LoadResultTy, ExtVT, LoadedVT,
7567 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0),
7568 ISD::ZERO_EXTEND, SetCCs, TLI);
7571 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), VT,
7572 LN0->getChain(), LN0->getBasePtr(),
7574 LN0->getMemOperand());
7575 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
7576 Mask = Mask.zext(VT.getSizeInBits());
7578 SDValue And = DAG.getNode(N0.getOpcode(), DL, VT,
7579 ExtLoad, DAG.getConstant(Mask, DL, VT));
7580 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
7581 SDLoc(N0.getOperand(0)),
7582 N0.getOperand(0).getValueType(), ExtLoad);
7583 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, DL, ISD::ZERO_EXTEND);
7584 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
7585 return CombineTo(N, And); // Return N so it doesn't get rechecked!
7590 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
7591 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
7592 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
7593 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
7594 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7595 EVT MemVT = LN0->getMemoryVT();
7596 if ((!LegalOperations && !LN0->isVolatile()) ||
7597 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT)) {
7598 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
7600 LN0->getBasePtr(), MemVT,
7601 LN0->getMemOperand());
7602 CombineTo(N, ExtLoad);
7603 CombineTo(N0.getNode(),
7604 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(),
7606 ExtLoad.getValue(1));
7607 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7611 if (N0.getOpcode() == ISD::SETCC) {
7612 // Only do this before legalize for now.
7613 if (!LegalOperations && VT.isVector() &&
7614 N0.getValueType().getVectorElementType() == MVT::i1) {
7615 EVT N00VT = N0.getOperand(0).getValueType();
7616 if (getSetCCResultType(N00VT) == N0.getValueType())
7619 // We know that the # elements of the results is the same as the #
7620 // elements of the compare (and the # elements of the compare result for
7621 // that matter). Check to see that they are the same size. If so, we know
7622 // that the element size of the sext'd result matches the element size of
7623 // the compare operands.
7625 SDValue VecOnes = DAG.getConstant(1, DL, VT);
7626 if (VT.getSizeInBits() == N00VT.getSizeInBits()) {
7627 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
7628 SDValue VSetCC = DAG.getNode(ISD::SETCC, DL, VT, N0.getOperand(0),
7629 N0.getOperand(1), N0.getOperand(2));
7630 return DAG.getNode(ISD::AND, DL, VT, VSetCC, VecOnes);
7633 // If the desired elements are smaller or larger than the source
7634 // elements we can use a matching integer vector type and then
7635 // truncate/sign extend.
7636 EVT MatchingElementType = EVT::getIntegerVT(
7637 *DAG.getContext(), N00VT.getScalarSizeInBits());
7638 EVT MatchingVectorType = EVT::getVectorVT(
7639 *DAG.getContext(), MatchingElementType, N00VT.getVectorNumElements());
7641 DAG.getNode(ISD::SETCC, DL, MatchingVectorType, N0.getOperand(0),
7642 N0.getOperand(1), N0.getOperand(2));
7643 return DAG.getNode(ISD::AND, DL, VT, DAG.getSExtOrTrunc(VsetCC, DL, VT),
7647 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
7649 if (SDValue SCC = SimplifySelectCC(
7650 DL, N0.getOperand(0), N0.getOperand(1), DAG.getConstant(1, DL, VT),
7651 DAG.getConstant(0, DL, VT),
7652 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true))
7656 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
7657 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
7658 isa<ConstantSDNode>(N0.getOperand(1)) &&
7659 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
7661 SDValue ShAmt = N0.getOperand(1);
7662 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
7663 if (N0.getOpcode() == ISD::SHL) {
7664 SDValue InnerZExt = N0.getOperand(0);
7665 // If the original shl may be shifting out bits, do not perform this
7667 unsigned KnownZeroBits = InnerZExt.getValueSizeInBits() -
7668 InnerZExt.getOperand(0).getValueSizeInBits();
7669 if (ShAmtVal > KnownZeroBits)
7675 // Ensure that the shift amount is wide enough for the shifted value.
7676 if (VT.getSizeInBits() >= 256)
7677 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
7679 return DAG.getNode(N0.getOpcode(), DL, VT,
7680 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
7684 if (SDValue NewVSel = matchVSelectOpSizesWithSetCC(N))
7690 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
7691 SDValue N0 = N->getOperand(0);
7692 EVT VT = N->getValueType(0);
7694 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
7696 return SDValue(Res, 0);
7698 // fold (aext (aext x)) -> (aext x)
7699 // fold (aext (zext x)) -> (zext x)
7700 // fold (aext (sext x)) -> (sext x)
7701 if (N0.getOpcode() == ISD::ANY_EXTEND ||
7702 N0.getOpcode() == ISD::ZERO_EXTEND ||
7703 N0.getOpcode() == ISD::SIGN_EXTEND)
7704 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0));
7706 // fold (aext (truncate (load x))) -> (aext (smaller load x))
7707 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
7708 if (N0.getOpcode() == ISD::TRUNCATE) {
7709 if (SDValue NarrowLoad = ReduceLoadWidth(N0.getNode())) {
7710 SDNode *oye = N0.getOperand(0).getNode();
7711 if (NarrowLoad.getNode() != N0.getNode()) {
7712 CombineTo(N0.getNode(), NarrowLoad);
7713 // CombineTo deleted the truncate, if needed, but not what's under it.
7716 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7720 // fold (aext (truncate x))
7721 if (N0.getOpcode() == ISD::TRUNCATE)
7722 return DAG.getAnyExtOrTrunc(N0.getOperand(0), SDLoc(N), VT);
7724 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
7725 // if the trunc is not free.
7726 if (N0.getOpcode() == ISD::AND &&
7727 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
7728 N0.getOperand(1).getOpcode() == ISD::Constant &&
7729 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
7730 N0.getValueType())) {
7732 SDValue X = N0.getOperand(0).getOperand(0);
7733 X = DAG.getAnyExtOrTrunc(X, DL, VT);
7734 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
7735 Mask = Mask.zext(VT.getSizeInBits());
7736 return DAG.getNode(ISD::AND, DL, VT,
7737 X, DAG.getConstant(Mask, DL, VT));
7740 // fold (aext (load x)) -> (aext (truncate (extload x)))
7741 // None of the supported targets knows how to perform load and any_ext
7742 // on vectors in one instruction. We only perform this transformation on
7744 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
7745 ISD::isUNINDEXEDLoad(N0.getNode()) &&
7746 TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) {
7747 bool DoXform = true;
7748 SmallVector<SDNode*, 4> SetCCs;
7749 if (!N0.hasOneUse())
7750 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
7752 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7753 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
7755 LN0->getBasePtr(), N0.getValueType(),
7756 LN0->getMemOperand());
7757 CombineTo(N, ExtLoad);
7758 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
7759 N0.getValueType(), ExtLoad);
7760 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
7761 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
7763 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7767 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
7768 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
7769 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
7770 if (N0.getOpcode() == ISD::LOAD &&
7771 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
7773 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7774 ISD::LoadExtType ExtType = LN0->getExtensionType();
7775 EVT MemVT = LN0->getMemoryVT();
7776 if (!LegalOperations || TLI.isLoadExtLegal(ExtType, VT, MemVT)) {
7777 SDValue ExtLoad = DAG.getExtLoad(ExtType, SDLoc(N),
7778 VT, LN0->getChain(), LN0->getBasePtr(),
7779 MemVT, LN0->getMemOperand());
7780 CombineTo(N, ExtLoad);
7781 CombineTo(N0.getNode(),
7782 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
7783 N0.getValueType(), ExtLoad),
7784 ExtLoad.getValue(1));
7785 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7789 if (N0.getOpcode() == ISD::SETCC) {
7791 // aext(setcc) -> vsetcc
7792 // aext(setcc) -> truncate(vsetcc)
7793 // aext(setcc) -> aext(vsetcc)
7794 // Only do this before legalize for now.
7795 if (VT.isVector() && !LegalOperations) {
7796 EVT N0VT = N0.getOperand(0).getValueType();
7797 // We know that the # elements of the results is the same as the
7798 // # elements of the compare (and the # elements of the compare result
7799 // for that matter). Check to see that they are the same size. If so,
7800 // we know that the element size of the sext'd result matches the
7801 // element size of the compare operands.
7802 if (VT.getSizeInBits() == N0VT.getSizeInBits())
7803 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
7805 cast<CondCodeSDNode>(N0.getOperand(2))->get());
7806 // If the desired elements are smaller or larger than the source
7807 // elements we can use a matching integer vector type and then
7808 // truncate/any extend
7810 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
7812 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
7814 cast<CondCodeSDNode>(N0.getOperand(2))->get());
7815 return DAG.getAnyExtOrTrunc(VsetCC, SDLoc(N), VT);
7819 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
7821 if (SDValue SCC = SimplifySelectCC(
7822 DL, N0.getOperand(0), N0.getOperand(1), DAG.getConstant(1, DL, VT),
7823 DAG.getConstant(0, DL, VT),
7824 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true))
7831 SDValue DAGCombiner::visitAssertZext(SDNode *N) {
7832 SDValue N0 = N->getOperand(0);
7833 SDValue N1 = N->getOperand(1);
7834 EVT EVT = cast<VTSDNode>(N1)->getVT();
7836 // fold (assertzext (assertzext x, vt), vt) -> (assertzext x, vt)
7837 if (N0.getOpcode() == ISD::AssertZext &&
7838 EVT == cast<VTSDNode>(N0.getOperand(1))->getVT())
7844 /// See if the specified operand can be simplified with the knowledge that only
7845 /// the bits specified by Mask are used. If so, return the simpler operand,
7846 /// otherwise return a null SDValue.
7848 /// (This exists alongside SimplifyDemandedBits because GetDemandedBits can
7849 /// simplify nodes with multiple uses more aggressively.)
7850 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
7851 switch (V.getOpcode()) {
7853 case ISD::Constant: {
7854 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
7855 assert(CV && "Const value should be ConstSDNode.");
7856 const APInt &CVal = CV->getAPIntValue();
7857 APInt NewVal = CVal & Mask;
7859 return DAG.getConstant(NewVal, SDLoc(V), V.getValueType());
7864 // If the LHS or RHS don't contribute bits to the or, drop them.
7865 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
7866 return V.getOperand(1);
7867 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
7868 return V.getOperand(0);
7871 // Only look at single-use SRLs.
7872 if (!V.getNode()->hasOneUse())
7874 if (ConstantSDNode *RHSC = getAsNonOpaqueConstant(V.getOperand(1))) {
7875 // See if we can recursively simplify the LHS.
7876 unsigned Amt = RHSC->getZExtValue();
7878 // Watch out for shift count overflow though.
7879 if (Amt >= Mask.getBitWidth()) break;
7880 APInt NewMask = Mask << Amt;
7881 if (SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask))
7882 return DAG.getNode(ISD::SRL, SDLoc(V), V.getValueType(),
7883 SimplifyLHS, V.getOperand(1));
7887 // X & -1 -> X (ignoring bits which aren't demanded).
7888 ConstantSDNode *AndVal = isConstOrConstSplat(V.getOperand(1));
7889 if (AndVal && (AndVal->getAPIntValue() & Mask) == Mask)
7890 return V.getOperand(0);
7897 /// If the result of a wider load is shifted to right of N bits and then
7898 /// truncated to a narrower type and where N is a multiple of number of bits of
7899 /// the narrower type, transform it to a narrower load from address + N / num of
7900 /// bits of new type. If the result is to be extended, also fold the extension
7901 /// to form a extending load.
7902 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
7903 unsigned Opc = N->getOpcode();
7905 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
7906 SDValue N0 = N->getOperand(0);
7907 EVT VT = N->getValueType(0);
7910 // This transformation isn't valid for vector loads.
7914 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
7916 if (Opc == ISD::SIGN_EXTEND_INREG) {
7917 ExtType = ISD::SEXTLOAD;
7918 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
7919 } else if (Opc == ISD::SRL) {
7920 // Another special-case: SRL is basically zero-extending a narrower value.
7921 ExtType = ISD::ZEXTLOAD;
7923 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
7924 if (!N01) return SDValue();
7925 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
7926 VT.getSizeInBits() - N01->getZExtValue());
7928 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, VT, ExtVT))
7931 unsigned EVTBits = ExtVT.getSizeInBits();
7933 // Do not generate loads of non-round integer types since these can
7934 // be expensive (and would be wrong if the type is not byte sized).
7935 if (!ExtVT.isRound())
7939 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
7940 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
7941 ShAmt = N01->getZExtValue();
7942 // Is the shift amount a multiple of size of VT?
7943 if ((ShAmt & (EVTBits-1)) == 0) {
7944 N0 = N0.getOperand(0);
7945 // Is the load width a multiple of size of VT?
7946 if ((N0.getValueSizeInBits() & (EVTBits-1)) != 0)
7950 // At this point, we must have a load or else we can't do the transform.
7951 if (!isa<LoadSDNode>(N0)) return SDValue();
7953 // Because a SRL must be assumed to *need* to zero-extend the high bits
7954 // (as opposed to anyext the high bits), we can't combine the zextload
7955 // lowering of SRL and an sextload.
7956 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
7959 // If the shift amount is larger than the input type then we're not
7960 // accessing any of the loaded bytes. If the load was a zextload/extload
7961 // then the result of the shift+trunc is zero/undef (handled elsewhere).
7962 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
7967 // If the load is shifted left (and the result isn't shifted back right),
7968 // we can fold the truncate through the shift.
7969 unsigned ShLeftAmt = 0;
7970 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
7971 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
7972 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
7973 ShLeftAmt = N01->getZExtValue();
7974 N0 = N0.getOperand(0);
7978 // If we haven't found a load, we can't narrow it. Don't transform one with
7979 // multiple uses, this would require adding a new load.
7980 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse())
7983 // Don't change the width of a volatile load.
7984 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7985 if (LN0->isVolatile())
7988 // Verify that we are actually reducing a load width here.
7989 if (LN0->getMemoryVT().getSizeInBits() < EVTBits)
7992 // For the transform to be legal, the load must produce only two values
7993 // (the value loaded and the chain). Don't transform a pre-increment
7994 // load, for example, which produces an extra value. Otherwise the
7995 // transformation is not equivalent, and the downstream logic to replace
7996 // uses gets things wrong.
7997 if (LN0->getNumValues() > 2)
8000 // If the load that we're shrinking is an extload and we're not just
8001 // discarding the extension we can't simply shrink the load. Bail.
8002 // TODO: It would be possible to merge the extensions in some cases.
8003 if (LN0->getExtensionType() != ISD::NON_EXTLOAD &&
8004 LN0->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits() + ShAmt)
8007 if (!TLI.shouldReduceLoadWidth(LN0, ExtType, ExtVT))
8010 EVT PtrType = N0.getOperand(1).getValueType();
8012 if (PtrType == MVT::Untyped || PtrType.isExtended())
8013 // It's not possible to generate a constant of extended or untyped type.
8016 // For big endian targets, we need to adjust the offset to the pointer to
8017 // load the correct bytes.
8018 if (DAG.getDataLayout().isBigEndian()) {
8019 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
8020 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
8021 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
8024 uint64_t PtrOff = ShAmt / 8;
8025 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
8027 // The original load itself didn't wrap, so an offset within it doesn't.
8029 Flags.setNoUnsignedWrap(true);
8030 SDValue NewPtr = DAG.getNode(ISD::ADD, DL,
8031 PtrType, LN0->getBasePtr(),
8032 DAG.getConstant(PtrOff, DL, PtrType),
8034 AddToWorklist(NewPtr.getNode());
8037 if (ExtType == ISD::NON_EXTLOAD)
8038 Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr,
8039 LN0->getPointerInfo().getWithOffset(PtrOff), NewAlign,
8040 LN0->getMemOperand()->getFlags(), LN0->getAAInfo());
8042 Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(), NewPtr,
8043 LN0->getPointerInfo().getWithOffset(PtrOff), ExtVT,
8044 NewAlign, LN0->getMemOperand()->getFlags(),
8047 // Replace the old load's chain with the new load's chain.
8048 WorklistRemover DeadNodes(*this);
8049 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
8051 // Shift the result left, if we've swallowed a left shift.
8052 SDValue Result = Load;
8053 if (ShLeftAmt != 0) {
8054 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
8055 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
8057 // If the shift amount is as large as the result size (but, presumably,
8058 // no larger than the source) then the useful bits of the result are
8059 // zero; we can't simply return the shortened shift, because the result
8060 // of that operation is undefined.
8062 if (ShLeftAmt >= VT.getSizeInBits())
8063 Result = DAG.getConstant(0, DL, VT);
8065 Result = DAG.getNode(ISD::SHL, DL, VT,
8066 Result, DAG.getConstant(ShLeftAmt, DL, ShImmTy));
8069 // Return the new loaded value.
8073 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
8074 SDValue N0 = N->getOperand(0);
8075 SDValue N1 = N->getOperand(1);
8076 EVT VT = N->getValueType(0);
8077 EVT EVT = cast<VTSDNode>(N1)->getVT();
8078 unsigned VTBits = VT.getScalarSizeInBits();
8079 unsigned EVTBits = EVT.getScalarSizeInBits();
8082 return DAG.getUNDEF(VT);
8084 // fold (sext_in_reg c1) -> c1
8085 if (DAG.isConstantIntBuildVectorOrConstantInt(N0))
8086 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1);
8088 // If the input is already sign extended, just drop the extension.
8089 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
8092 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
8093 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
8094 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT()))
8095 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
8096 N0.getOperand(0), N1);
8098 // fold (sext_in_reg (sext x)) -> (sext x)
8099 // fold (sext_in_reg (aext x)) -> (sext x)
8100 // if x is small enough.
8101 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
8102 SDValue N00 = N0.getOperand(0);
8103 if (N00.getScalarValueSizeInBits() <= EVTBits &&
8104 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
8105 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
8108 // fold (sext_in_reg (*_extend_vector_inreg x)) -> (sext_vector_in_reg x)
8109 if ((N0.getOpcode() == ISD::ANY_EXTEND_VECTOR_INREG ||
8110 N0.getOpcode() == ISD::SIGN_EXTEND_VECTOR_INREG ||
8111 N0.getOpcode() == ISD::ZERO_EXTEND_VECTOR_INREG) &&
8112 N0.getOperand(0).getScalarValueSizeInBits() == EVTBits) {
8113 if (!LegalOperations ||
8114 TLI.isOperationLegal(ISD::SIGN_EXTEND_VECTOR_INREG, VT))
8115 return DAG.getSignExtendVectorInReg(N0.getOperand(0), SDLoc(N), VT);
8118 // fold (sext_in_reg (zext x)) -> (sext x)
8119 // iff we are extending the source sign bit.
8120 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
8121 SDValue N00 = N0.getOperand(0);
8122 if (N00.getScalarValueSizeInBits() == EVTBits &&
8123 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
8124 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
8127 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
8128 if (DAG.MaskedValueIsZero(N0, APInt::getOneBitSet(VTBits, EVTBits - 1)))
8129 return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT.getScalarType());
8131 // fold operands of sext_in_reg based on knowledge that the top bits are not
8133 if (SimplifyDemandedBits(SDValue(N, 0)))
8134 return SDValue(N, 0);
8136 // fold (sext_in_reg (load x)) -> (smaller sextload x)
8137 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
8138 if (SDValue NarrowLoad = ReduceLoadWidth(N))
8141 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
8142 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
8143 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
8144 if (N0.getOpcode() == ISD::SRL) {
8145 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
8146 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
8147 // We can turn this into an SRA iff the input to the SRL is already sign
8149 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
8150 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
8151 return DAG.getNode(ISD::SRA, SDLoc(N), VT,
8152 N0.getOperand(0), N0.getOperand(1));
8156 // fold (sext_inreg (extload x)) -> (sextload x)
8157 if (ISD::isEXTLoad(N0.getNode()) &&
8158 ISD::isUNINDEXEDLoad(N0.getNode()) &&
8159 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
8160 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
8161 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) {
8162 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
8163 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
8165 LN0->getBasePtr(), EVT,
8166 LN0->getMemOperand());
8167 CombineTo(N, ExtLoad);
8168 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
8169 AddToWorklist(ExtLoad.getNode());
8170 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8172 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
8173 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
8175 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
8176 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
8177 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) {
8178 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
8179 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
8181 LN0->getBasePtr(), EVT,
8182 LN0->getMemOperand());
8183 CombineTo(N, ExtLoad);
8184 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
8185 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8188 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
8189 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
8190 if (SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
8191 N0.getOperand(1), false))
8192 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
8199 SDValue DAGCombiner::visitSIGN_EXTEND_VECTOR_INREG(SDNode *N) {
8200 SDValue N0 = N->getOperand(0);
8201 EVT VT = N->getValueType(0);
8204 return DAG.getUNDEF(VT);
8206 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
8208 return SDValue(Res, 0);
8213 SDValue DAGCombiner::visitZERO_EXTEND_VECTOR_INREG(SDNode *N) {
8214 SDValue N0 = N->getOperand(0);
8215 EVT VT = N->getValueType(0);
8218 return DAG.getUNDEF(VT);
8220 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
8222 return SDValue(Res, 0);
8227 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
8228 SDValue N0 = N->getOperand(0);
8229 EVT VT = N->getValueType(0);
8230 bool isLE = DAG.getDataLayout().isLittleEndian();
8233 if (N0.getValueType() == N->getValueType(0))
8235 // fold (truncate c1) -> c1
8236 if (DAG.isConstantIntBuildVectorOrConstantInt(N0))
8237 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0);
8238 // fold (truncate (truncate x)) -> (truncate x)
8239 if (N0.getOpcode() == ISD::TRUNCATE)
8240 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
8241 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
8242 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
8243 N0.getOpcode() == ISD::SIGN_EXTEND ||
8244 N0.getOpcode() == ISD::ANY_EXTEND) {
8245 // if the source is smaller than the dest, we still need an extend.
8246 if (N0.getOperand(0).getValueType().bitsLT(VT))
8247 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0));
8248 // if the source is larger than the dest, than we just need the truncate.
8249 if (N0.getOperand(0).getValueType().bitsGT(VT))
8250 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
8251 // if the source and dest are the same type, we can drop both the extend
8252 // and the truncate.
8253 return N0.getOperand(0);
8256 // If this is anyext(trunc), don't fold it, allow ourselves to be folded.
8257 if (N->hasOneUse() && (N->use_begin()->getOpcode() == ISD::ANY_EXTEND))
8260 // Fold extract-and-trunc into a narrow extract. For example:
8261 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
8262 // i32 y = TRUNCATE(i64 x)
8264 // v16i8 b = BITCAST (v2i64 val)
8265 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
8267 // Note: We only run this optimization after type legalization (which often
8268 // creates this pattern) and before operation legalization after which
8269 // we need to be more careful about the vector instructions that we generate.
8270 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
8271 LegalTypes && !LegalOperations && N0->hasOneUse() && VT != MVT::i1) {
8273 EVT VecTy = N0.getOperand(0).getValueType();
8274 EVT ExTy = N0.getValueType();
8275 EVT TrTy = N->getValueType(0);
8277 unsigned NumElem = VecTy.getVectorNumElements();
8278 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
8280 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
8281 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
8283 SDValue EltNo = N0->getOperand(1);
8284 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
8285 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
8286 EVT IndexTy = TLI.getVectorIdxTy(DAG.getDataLayout());
8287 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
8290 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, TrTy,
8291 DAG.getBitcast(NVT, N0.getOperand(0)),
8292 DAG.getConstant(Index, DL, IndexTy));
8296 // trunc (select c, a, b) -> select c, (trunc a), (trunc b)
8297 if (N0.getOpcode() == ISD::SELECT && N0.hasOneUse()) {
8298 EVT SrcVT = N0.getValueType();
8299 if ((!LegalOperations || TLI.isOperationLegal(ISD::SELECT, SrcVT)) &&
8300 TLI.isTruncateFree(SrcVT, VT)) {
8302 SDValue Cond = N0.getOperand(0);
8303 SDValue TruncOp0 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(1));
8304 SDValue TruncOp1 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(2));
8305 return DAG.getNode(ISD::SELECT, SDLoc(N), VT, Cond, TruncOp0, TruncOp1);
8309 // trunc (shl x, K) -> shl (trunc x), K => K < VT.getScalarSizeInBits()
8310 if (N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
8311 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::SHL, VT)) &&
8312 TLI.isTypeDesirableForOp(ISD::SHL, VT)) {
8313 if (const ConstantSDNode *CAmt = isConstOrConstSplat(N0.getOperand(1))) {
8314 uint64_t Amt = CAmt->getZExtValue();
8315 unsigned Size = VT.getScalarSizeInBits();
8319 EVT AmtVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
8321 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(0));
8322 return DAG.getNode(ISD::SHL, SL, VT, Trunc,
8323 DAG.getConstant(Amt, SL, AmtVT));
8328 // Fold a series of buildvector, bitcast, and truncate if possible.
8330 // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to
8331 // (2xi32 (buildvector x, y)).
8332 if (Level == AfterLegalizeVectorOps && VT.isVector() &&
8333 N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
8334 N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
8335 N0.getOperand(0).hasOneUse()) {
8337 SDValue BuildVect = N0.getOperand(0);
8338 EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType();
8339 EVT TruncVecEltTy = VT.getVectorElementType();
8341 // Check that the element types match.
8342 if (BuildVectEltTy == TruncVecEltTy) {
8343 // Now we only need to compute the offset of the truncated elements.
8344 unsigned BuildVecNumElts = BuildVect.getNumOperands();
8345 unsigned TruncVecNumElts = VT.getVectorNumElements();
8346 unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts;
8348 assert((BuildVecNumElts % TruncVecNumElts) == 0 &&
8349 "Invalid number of elements");
8351 SmallVector<SDValue, 8> Opnds;
8352 for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset)
8353 Opnds.push_back(BuildVect.getOperand(i));
8355 return DAG.getBuildVector(VT, SDLoc(N), Opnds);
8359 // See if we can simplify the input to this truncate through knowledge that
8360 // only the low bits are being used.
8361 // For example "trunc (or (shl x, 8), y)" // -> trunc y
8362 // Currently we only perform this optimization on scalars because vectors
8363 // may have different active low bits.
8364 if (!VT.isVector()) {
8365 if (SDValue Shorter =
8366 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
8367 VT.getSizeInBits())))
8368 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter);
8371 // fold (truncate (load x)) -> (smaller load x)
8372 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
8373 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
8374 if (SDValue Reduced = ReduceLoadWidth(N))
8377 // Handle the case where the load remains an extending load even
8378 // after truncation.
8379 if (N0.hasOneUse() && ISD::isUNINDEXEDLoad(N0.getNode())) {
8380 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
8381 if (!LN0->isVolatile() &&
8382 LN0->getMemoryVT().getStoreSizeInBits() < VT.getSizeInBits()) {
8383 SDValue NewLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(LN0),
8384 VT, LN0->getChain(), LN0->getBasePtr(),
8386 LN0->getMemOperand());
8387 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLoad.getValue(1));
8393 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
8394 // where ... are all 'undef'.
8395 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
8396 SmallVector<EVT, 8> VTs;
8399 unsigned NumDefs = 0;
8401 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
8402 SDValue X = N0.getOperand(i);
8408 // Stop if more than one members are non-undef.
8411 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
8412 VT.getVectorElementType(),
8413 X.getValueType().getVectorNumElements()));
8417 return DAG.getUNDEF(VT);
8420 assert(V.getNode() && "The single defined operand is empty!");
8421 SmallVector<SDValue, 8> Opnds;
8422 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
8424 Opnds.push_back(DAG.getUNDEF(VTs[i]));
8427 SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V);
8428 AddToWorklist(NV.getNode());
8429 Opnds.push_back(NV);
8431 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Opnds);
8435 // Fold truncate of a bitcast of a vector to an extract of the low vector
8438 // e.g. trunc (i64 (bitcast v2i32:x)) -> extract_vector_elt v2i32:x, 0
8439 if (N0.getOpcode() == ISD::BITCAST && !VT.isVector()) {
8440 SDValue VecSrc = N0.getOperand(0);
8441 EVT SrcVT = VecSrc.getValueType();
8442 if (SrcVT.isVector() && SrcVT.getScalarType() == VT &&
8443 (!LegalOperations ||
8444 TLI.isOperationLegal(ISD::EXTRACT_VECTOR_ELT, SrcVT))) {
8447 EVT IdxVT = TLI.getVectorIdxTy(DAG.getDataLayout());
8448 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, VT,
8449 VecSrc, DAG.getConstant(0, SL, IdxVT));
8453 // Simplify the operands using demanded-bits information.
8454 if (!VT.isVector() &&
8455 SimplifyDemandedBits(SDValue(N, 0)))
8456 return SDValue(N, 0);
8458 // (trunc adde(X, Y, Carry)) -> (adde trunc(X), trunc(Y), Carry)
8459 // (trunc addcarry(X, Y, Carry)) -> (addcarry trunc(X), trunc(Y), Carry)
8460 // When the adde's carry is not used.
8461 if ((N0.getOpcode() == ISD::ADDE || N0.getOpcode() == ISD::ADDCARRY) &&
8462 N0.hasOneUse() && !N0.getNode()->hasAnyUseOfValue(1) &&
8463 (!LegalOperations || TLI.isOperationLegal(N0.getOpcode(), VT))) {
8465 auto X = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(0));
8466 auto Y = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(1));
8467 auto VTs = DAG.getVTList(VT, N0->getValueType(1));
8468 return DAG.getNode(N0.getOpcode(), SL, VTs, X, Y, N0.getOperand(2));
8471 if (SDValue NewVSel = matchVSelectOpSizesWithSetCC(N))
8477 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
8478 SDValue Elt = N->getOperand(i);
8479 if (Elt.getOpcode() != ISD::MERGE_VALUES)
8480 return Elt.getNode();
8481 return Elt.getOperand(Elt.getResNo()).getNode();
8484 /// build_pair (load, load) -> load
8485 /// if load locations are consecutive.
8486 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
8487 assert(N->getOpcode() == ISD::BUILD_PAIR);
8489 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
8490 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
8491 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
8492 LD1->getAddressSpace() != LD2->getAddressSpace())
8494 EVT LD1VT = LD1->getValueType(0);
8495 unsigned LD1Bytes = LD1VT.getSizeInBits() / 8;
8496 if (ISD::isNON_EXTLoad(LD2) && LD2->hasOneUse() &&
8497 DAG.areNonVolatileConsecutiveLoads(LD2, LD1, LD1Bytes, 1)) {
8498 unsigned Align = LD1->getAlignment();
8499 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
8500 VT.getTypeForEVT(*DAG.getContext()));
8502 if (NewAlign <= Align &&
8503 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
8504 return DAG.getLoad(VT, SDLoc(N), LD1->getChain(), LD1->getBasePtr(),
8505 LD1->getPointerInfo(), Align);
8511 static unsigned getPPCf128HiElementSelector(const SelectionDAG &DAG) {
8512 // On little-endian machines, bitcasting from ppcf128 to i128 does swap the Hi
8513 // and Lo parts; on big-endian machines it doesn't.
8514 return DAG.getDataLayout().isBigEndian() ? 1 : 0;
8517 static SDValue foldBitcastedFPLogic(SDNode *N, SelectionDAG &DAG,
8518 const TargetLowering &TLI) {
8519 // If this is not a bitcast to an FP type or if the target doesn't have
8520 // IEEE754-compliant FP logic, we're done.
8521 EVT VT = N->getValueType(0);
8522 if (!VT.isFloatingPoint() || !TLI.hasBitPreservingFPLogic(VT))
8525 // TODO: Use splat values for the constant-checking below and remove this
8527 SDValue N0 = N->getOperand(0);
8528 EVT SourceVT = N0.getValueType();
8529 if (SourceVT.isVector())
8534 switch (N0.getOpcode()) {
8536 FPOpcode = ISD::FABS;
8537 SignMask = ~APInt::getSignMask(SourceVT.getSizeInBits());
8540 FPOpcode = ISD::FNEG;
8541 SignMask = APInt::getSignMask(SourceVT.getSizeInBits());
8543 // TODO: ISD::OR --> ISD::FNABS?
8548 // Fold (bitcast int (and (bitcast fp X to int), 0x7fff...) to fp) -> fabs X
8549 // Fold (bitcast int (xor (bitcast fp X to int), 0x8000...) to fp) -> fneg X
8550 SDValue LogicOp0 = N0.getOperand(0);
8551 ConstantSDNode *LogicOp1 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
8552 if (LogicOp1 && LogicOp1->getAPIntValue() == SignMask &&
8553 LogicOp0.getOpcode() == ISD::BITCAST &&
8554 LogicOp0->getOperand(0).getValueType() == VT)
8555 return DAG.getNode(FPOpcode, SDLoc(N), VT, LogicOp0->getOperand(0));
8560 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
8561 SDValue N0 = N->getOperand(0);
8562 EVT VT = N->getValueType(0);
8565 return DAG.getUNDEF(VT);
8567 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
8568 // Only do this before legalize, since afterward the target may be depending
8569 // on the bitconvert.
8570 // First check to see if this is all constant.
8572 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
8574 bool isSimple = cast<BuildVectorSDNode>(N0)->isConstant();
8576 EVT DestEltVT = N->getValueType(0).getVectorElementType();
8577 assert(!DestEltVT.isVector() &&
8578 "Element type of vector ValueType must not be vector!");
8580 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
8583 // If the input is a constant, let getNode fold it.
8584 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
8585 // If we can't allow illegal operations, we need to check that this is just
8586 // a fp -> int or int -> conversion and that the resulting operation will
8588 if (!LegalOperations ||
8589 (isa<ConstantSDNode>(N0) && VT.isFloatingPoint() && !VT.isVector() &&
8590 TLI.isOperationLegal(ISD::ConstantFP, VT)) ||
8591 (isa<ConstantFPSDNode>(N0) && VT.isInteger() && !VT.isVector() &&
8592 TLI.isOperationLegal(ISD::Constant, VT)))
8593 return DAG.getBitcast(VT, N0);
8596 // (conv (conv x, t1), t2) -> (conv x, t2)
8597 if (N0.getOpcode() == ISD::BITCAST)
8598 return DAG.getBitcast(VT, N0.getOperand(0));
8600 // fold (conv (load x)) -> (load (conv*)x)
8601 // If the resultant load doesn't need a higher alignment than the original!
8602 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
8603 // Do not change the width of a volatile load.
8604 !cast<LoadSDNode>(N0)->isVolatile() &&
8605 // Do not remove the cast if the types differ in endian layout.
8606 TLI.hasBigEndianPartOrdering(N0.getValueType(), DAG.getDataLayout()) ==
8607 TLI.hasBigEndianPartOrdering(VT, DAG.getDataLayout()) &&
8608 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)) &&
8609 TLI.isLoadBitCastBeneficial(N0.getValueType(), VT)) {
8610 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
8611 unsigned OrigAlign = LN0->getAlignment();
8614 if (TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT,
8615 LN0->getAddressSpace(), OrigAlign, &Fast) &&
8618 DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
8619 LN0->getPointerInfo(), OrigAlign,
8620 LN0->getMemOperand()->getFlags(), LN0->getAAInfo());
8621 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
8626 if (SDValue V = foldBitcastedFPLogic(N, DAG, TLI))
8629 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
8630 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
8633 // fold (bitcast (fneg x)) ->
8634 // flipbit = signbit
8635 // (xor (bitcast x) (build_pair flipbit, flipbit))
8637 // fold (bitcast (fabs x)) ->
8638 // flipbit = (and (extract_element (bitcast x), 0), signbit)
8639 // (xor (bitcast x) (build_pair flipbit, flipbit))
8640 // This often reduces constant pool loads.
8641 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) ||
8642 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) &&
8643 N0.getNode()->hasOneUse() && VT.isInteger() &&
8644 !VT.isVector() && !N0.getValueType().isVector()) {
8645 SDValue NewConv = DAG.getBitcast(VT, N0.getOperand(0));
8646 AddToWorklist(NewConv.getNode());
8649 if (N0.getValueType() == MVT::ppcf128 && !LegalTypes) {
8650 assert(VT.getSizeInBits() == 128);
8651 SDValue SignBit = DAG.getConstant(
8652 APInt::getSignMask(VT.getSizeInBits() / 2), SDLoc(N0), MVT::i64);
8654 if (N0.getOpcode() == ISD::FNEG) {
8656 AddToWorklist(FlipBit.getNode());
8658 assert(N0.getOpcode() == ISD::FABS);
8660 DAG.getNode(ISD::EXTRACT_ELEMENT, SDLoc(NewConv), MVT::i64, NewConv,
8661 DAG.getIntPtrConstant(getPPCf128HiElementSelector(DAG),
8663 AddToWorklist(Hi.getNode());
8664 FlipBit = DAG.getNode(ISD::AND, SDLoc(N0), MVT::i64, Hi, SignBit);
8665 AddToWorklist(FlipBit.getNode());
8668 DAG.getNode(ISD::BUILD_PAIR, SDLoc(N0), VT, FlipBit, FlipBit);
8669 AddToWorklist(FlipBits.getNode());
8670 return DAG.getNode(ISD::XOR, DL, VT, NewConv, FlipBits);
8672 APInt SignBit = APInt::getSignMask(VT.getSizeInBits());
8673 if (N0.getOpcode() == ISD::FNEG)
8674 return DAG.getNode(ISD::XOR, DL, VT,
8675 NewConv, DAG.getConstant(SignBit, DL, VT));
8676 assert(N0.getOpcode() == ISD::FABS);
8677 return DAG.getNode(ISD::AND, DL, VT,
8678 NewConv, DAG.getConstant(~SignBit, DL, VT));
8681 // fold (bitconvert (fcopysign cst, x)) ->
8682 // (or (and (bitconvert x), sign), (and cst, (not sign)))
8683 // Note that we don't handle (copysign x, cst) because this can always be
8684 // folded to an fneg or fabs.
8687 // fold (bitcast (fcopysign cst, x)) ->
8688 // flipbit = (and (extract_element
8689 // (xor (bitcast cst), (bitcast x)), 0),
8691 // (xor (bitcast cst) (build_pair flipbit, flipbit))
8692 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
8693 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
8694 VT.isInteger() && !VT.isVector()) {
8695 unsigned OrigXWidth = N0.getOperand(1).getValueSizeInBits();
8696 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
8697 if (isTypeLegal(IntXVT)) {
8698 SDValue X = DAG.getBitcast(IntXVT, N0.getOperand(1));
8699 AddToWorklist(X.getNode());
8701 // If X has a different width than the result/lhs, sext it or truncate it.
8702 unsigned VTWidth = VT.getSizeInBits();
8703 if (OrigXWidth < VTWidth) {
8704 X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X);
8705 AddToWorklist(X.getNode());
8706 } else if (OrigXWidth > VTWidth) {
8707 // To get the sign bit in the right place, we have to shift it right
8708 // before truncating.
8710 X = DAG.getNode(ISD::SRL, DL,
8711 X.getValueType(), X,
8712 DAG.getConstant(OrigXWidth-VTWidth, DL,
8714 AddToWorklist(X.getNode());
8715 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
8716 AddToWorklist(X.getNode());
8719 if (N0.getValueType() == MVT::ppcf128 && !LegalTypes) {
8720 APInt SignBit = APInt::getSignMask(VT.getSizeInBits() / 2);
8721 SDValue Cst = DAG.getBitcast(VT, N0.getOperand(0));
8722 AddToWorklist(Cst.getNode());
8723 SDValue X = DAG.getBitcast(VT, N0.getOperand(1));
8724 AddToWorklist(X.getNode());
8725 SDValue XorResult = DAG.getNode(ISD::XOR, SDLoc(N0), VT, Cst, X);
8726 AddToWorklist(XorResult.getNode());
8727 SDValue XorResult64 = DAG.getNode(
8728 ISD::EXTRACT_ELEMENT, SDLoc(XorResult), MVT::i64, XorResult,
8729 DAG.getIntPtrConstant(getPPCf128HiElementSelector(DAG),
8731 AddToWorklist(XorResult64.getNode());
8733 DAG.getNode(ISD::AND, SDLoc(XorResult64), MVT::i64, XorResult64,
8734 DAG.getConstant(SignBit, SDLoc(XorResult64), MVT::i64));
8735 AddToWorklist(FlipBit.getNode());
8737 DAG.getNode(ISD::BUILD_PAIR, SDLoc(N0), VT, FlipBit, FlipBit);
8738 AddToWorklist(FlipBits.getNode());
8739 return DAG.getNode(ISD::XOR, SDLoc(N), VT, Cst, FlipBits);
8741 APInt SignBit = APInt::getSignMask(VT.getSizeInBits());
8742 X = DAG.getNode(ISD::AND, SDLoc(X), VT,
8743 X, DAG.getConstant(SignBit, SDLoc(X), VT));
8744 AddToWorklist(X.getNode());
8746 SDValue Cst = DAG.getBitcast(VT, N0.getOperand(0));
8747 Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT,
8748 Cst, DAG.getConstant(~SignBit, SDLoc(Cst), VT));
8749 AddToWorklist(Cst.getNode());
8751 return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst);
8755 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
8756 if (N0.getOpcode() == ISD::BUILD_PAIR)
8757 if (SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT))
8760 // Remove double bitcasts from shuffles - this is often a legacy of
8761 // XformToShuffleWithZero being used to combine bitmaskings (of
8762 // float vectors bitcast to integer vectors) into shuffles.
8763 // bitcast(shuffle(bitcast(s0),bitcast(s1))) -> shuffle(s0,s1)
8764 if (Level < AfterLegalizeDAG && TLI.isTypeLegal(VT) && VT.isVector() &&
8765 N0->getOpcode() == ISD::VECTOR_SHUFFLE &&
8766 VT.getVectorNumElements() >= N0.getValueType().getVectorNumElements() &&
8767 !(VT.getVectorNumElements() % N0.getValueType().getVectorNumElements())) {
8768 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N0);
8770 // If operands are a bitcast, peek through if it casts the original VT.
8771 // If operands are a constant, just bitcast back to original VT.
8772 auto PeekThroughBitcast = [&](SDValue Op) {
8773 if (Op.getOpcode() == ISD::BITCAST &&
8774 Op.getOperand(0).getValueType() == VT)
8775 return SDValue(Op.getOperand(0));
8776 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) ||
8777 ISD::isBuildVectorOfConstantFPSDNodes(Op.getNode()))
8778 return DAG.getBitcast(VT, Op);
8782 SDValue SV0 = PeekThroughBitcast(N0->getOperand(0));
8783 SDValue SV1 = PeekThroughBitcast(N0->getOperand(1));
8788 VT.getVectorNumElements() / N0.getValueType().getVectorNumElements();
8789 SmallVector<int, 8> NewMask;
8790 for (int M : SVN->getMask())
8791 for (int i = 0; i != MaskScale; ++i)
8792 NewMask.push_back(M < 0 ? -1 : M * MaskScale + i);
8794 bool LegalMask = TLI.isShuffleMaskLegal(NewMask, VT);
8796 std::swap(SV0, SV1);
8797 ShuffleVectorSDNode::commuteMask(NewMask);
8798 LegalMask = TLI.isShuffleMaskLegal(NewMask, VT);
8802 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, NewMask);
8808 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
8809 EVT VT = N->getValueType(0);
8810 return CombineConsecutiveLoads(N, VT);
8813 /// We know that BV is a build_vector node with Constant, ConstantFP or Undef
8814 /// operands. DstEltVT indicates the destination element value type.
8815 SDValue DAGCombiner::
8816 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
8817 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
8819 // If this is already the right type, we're done.
8820 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
8822 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
8823 unsigned DstBitSize = DstEltVT.getSizeInBits();
8825 // If this is a conversion of N elements of one type to N elements of another
8826 // type, convert each element. This handles FP<->INT cases.
8827 if (SrcBitSize == DstBitSize) {
8828 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
8829 BV->getValueType(0).getVectorNumElements());
8831 // Due to the FP element handling below calling this routine recursively,
8832 // we can end up with a scalar-to-vector node here.
8833 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
8834 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
8835 DAG.getBitcast(DstEltVT, BV->getOperand(0)));
8837 SmallVector<SDValue, 8> Ops;
8838 for (SDValue Op : BV->op_values()) {
8839 // If the vector element type is not legal, the BUILD_VECTOR operands
8840 // are promoted and implicitly truncated. Make that explicit here.
8841 if (Op.getValueType() != SrcEltVT)
8842 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op);
8843 Ops.push_back(DAG.getBitcast(DstEltVT, Op));
8844 AddToWorklist(Ops.back().getNode());
8846 return DAG.getBuildVector(VT, SDLoc(BV), Ops);
8849 // Otherwise, we're growing or shrinking the elements. To avoid having to
8850 // handle annoying details of growing/shrinking FP values, we convert them to
8852 if (SrcEltVT.isFloatingPoint()) {
8853 // Convert the input float vector to a int vector where the elements are the
8855 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
8856 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
8860 // Now we know the input is an integer vector. If the output is a FP type,
8861 // convert to integer first, then to FP of the right size.
8862 if (DstEltVT.isFloatingPoint()) {
8863 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
8864 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
8866 // Next, convert to FP elements of the same size.
8867 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
8872 // Okay, we know the src/dst types are both integers of differing types.
8873 // Handling growing first.
8874 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
8875 if (SrcBitSize < DstBitSize) {
8876 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
8878 SmallVector<SDValue, 8> Ops;
8879 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
8880 i += NumInputsPerOutput) {
8881 bool isLE = DAG.getDataLayout().isLittleEndian();
8882 APInt NewBits = APInt(DstBitSize, 0);
8883 bool EltIsUndef = true;
8884 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
8885 // Shift the previously computed bits over.
8886 NewBits <<= SrcBitSize;
8887 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
8888 if (Op.isUndef()) continue;
8891 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
8892 zextOrTrunc(SrcBitSize).zext(DstBitSize);
8896 Ops.push_back(DAG.getUNDEF(DstEltVT));
8898 Ops.push_back(DAG.getConstant(NewBits, DL, DstEltVT));
8901 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
8902 return DAG.getBuildVector(VT, DL, Ops);
8905 // Finally, this must be the case where we are shrinking elements: each input
8906 // turns into multiple outputs.
8907 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
8908 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
8909 NumOutputsPerInput*BV->getNumOperands());
8910 SmallVector<SDValue, 8> Ops;
8912 for (const SDValue &Op : BV->op_values()) {
8914 Ops.append(NumOutputsPerInput, DAG.getUNDEF(DstEltVT));
8918 APInt OpVal = cast<ConstantSDNode>(Op)->
8919 getAPIntValue().zextOrTrunc(SrcBitSize);
8921 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
8922 APInt ThisVal = OpVal.trunc(DstBitSize);
8923 Ops.push_back(DAG.getConstant(ThisVal, DL, DstEltVT));
8924 OpVal.lshrInPlace(DstBitSize);
8927 // For big endian targets, swap the order of the pieces of each element.
8928 if (DAG.getDataLayout().isBigEndian())
8929 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
8932 return DAG.getBuildVector(VT, DL, Ops);
8935 static bool isContractable(SDNode *N) {
8936 SDNodeFlags F = N->getFlags();
8937 return F.hasAllowContract() || F.hasUnsafeAlgebra();
8940 /// Try to perform FMA combining on a given FADD node.
8941 SDValue DAGCombiner::visitFADDForFMACombine(SDNode *N) {
8942 SDValue N0 = N->getOperand(0);
8943 SDValue N1 = N->getOperand(1);
8944 EVT VT = N->getValueType(0);
8947 const TargetOptions &Options = DAG.getTarget().Options;
8949 // Floating-point multiply-add with intermediate rounding.
8950 bool HasFMAD = (LegalOperations && TLI.isOperationLegal(ISD::FMAD, VT));
8952 // Floating-point multiply-add without intermediate rounding.
8954 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
8955 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT));
8957 // No valid opcode, do not combine.
8958 if (!HasFMAD && !HasFMA)
8961 bool AllowFusionGlobally = (Options.AllowFPOpFusion == FPOpFusion::Fast ||
8962 Options.UnsafeFPMath || HasFMAD);
8963 // If the addition is not contractable, do not combine.
8964 if (!AllowFusionGlobally && !isContractable(N))
8967 const SelectionDAGTargetInfo *STI = DAG.getSubtarget().getSelectionDAGInfo();
8968 if (STI && STI->generateFMAsInMachineCombiner(OptLevel))
8971 // Always prefer FMAD to FMA for precision.
8972 unsigned PreferredFusedOpcode = HasFMAD ? ISD::FMAD : ISD::FMA;
8973 bool Aggressive = TLI.enableAggressiveFMAFusion(VT);
8974 bool LookThroughFPExt = TLI.isFPExtFree(VT);
8976 // Is the node an FMUL and contractable either due to global flags or
8978 auto isContractableFMUL = [AllowFusionGlobally](SDValue N) {
8979 if (N.getOpcode() != ISD::FMUL)
8981 return AllowFusionGlobally || isContractable(N.getNode());
8983 // If we have two choices trying to fold (fadd (fmul u, v), (fmul x, y)),
8984 // prefer to fold the multiply with fewer uses.
8985 if (Aggressive && isContractableFMUL(N0) && isContractableFMUL(N1)) {
8986 if (N0.getNode()->use_size() > N1.getNode()->use_size())
8990 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
8991 if (isContractableFMUL(N0) && (Aggressive || N0->hasOneUse())) {
8992 return DAG.getNode(PreferredFusedOpcode, SL, VT,
8993 N0.getOperand(0), N0.getOperand(1), N1);
8996 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
8997 // Note: Commutes FADD operands.
8998 if (isContractableFMUL(N1) && (Aggressive || N1->hasOneUse())) {
8999 return DAG.getNode(PreferredFusedOpcode, SL, VT,
9000 N1.getOperand(0), N1.getOperand(1), N0);
9003 // Look through FP_EXTEND nodes to do more combining.
9004 if (LookThroughFPExt) {
9005 // fold (fadd (fpext (fmul x, y)), z) -> (fma (fpext x), (fpext y), z)
9006 if (N0.getOpcode() == ISD::FP_EXTEND) {
9007 SDValue N00 = N0.getOperand(0);
9008 if (isContractableFMUL(N00))
9009 return DAG.getNode(PreferredFusedOpcode, SL, VT,
9010 DAG.getNode(ISD::FP_EXTEND, SL, VT,
9012 DAG.getNode(ISD::FP_EXTEND, SL, VT,
9013 N00.getOperand(1)), N1);
9016 // fold (fadd x, (fpext (fmul y, z))) -> (fma (fpext y), (fpext z), x)
9017 // Note: Commutes FADD operands.
9018 if (N1.getOpcode() == ISD::FP_EXTEND) {
9019 SDValue N10 = N1.getOperand(0);
9020 if (isContractableFMUL(N10))
9021 return DAG.getNode(PreferredFusedOpcode, SL, VT,
9022 DAG.getNode(ISD::FP_EXTEND, SL, VT,
9024 DAG.getNode(ISD::FP_EXTEND, SL, VT,
9025 N10.getOperand(1)), N0);
9029 // More folding opportunities when target permits.
9031 // fold (fadd (fma x, y, (fmul u, v)), z) -> (fma x, y (fma u, v, z))
9032 // FIXME: The UnsafeAlgebra flag should be propagated to FMA/FMAD, but FMF
9033 // are currently only supported on binary nodes.
9034 if (Options.UnsafeFPMath &&
9035 N0.getOpcode() == PreferredFusedOpcode &&
9036 N0.getOperand(2).getOpcode() == ISD::FMUL &&
9037 N0->hasOneUse() && N0.getOperand(2)->hasOneUse()) {
9038 return DAG.getNode(PreferredFusedOpcode, SL, VT,
9039 N0.getOperand(0), N0.getOperand(1),
9040 DAG.getNode(PreferredFusedOpcode, SL, VT,
9041 N0.getOperand(2).getOperand(0),
9042 N0.getOperand(2).getOperand(1),
9046 // fold (fadd x, (fma y, z, (fmul u, v)) -> (fma y, z (fma u, v, x))
9047 // FIXME: The UnsafeAlgebra flag should be propagated to FMA/FMAD, but FMF
9048 // are currently only supported on binary nodes.
9049 if (Options.UnsafeFPMath &&
9050 N1->getOpcode() == PreferredFusedOpcode &&
9051 N1.getOperand(2).getOpcode() == ISD::FMUL &&
9052 N1->hasOneUse() && N1.getOperand(2)->hasOneUse()) {
9053 return DAG.getNode(PreferredFusedOpcode, SL, VT,
9054 N1.getOperand(0), N1.getOperand(1),
9055 DAG.getNode(PreferredFusedOpcode, SL, VT,
9056 N1.getOperand(2).getOperand(0),
9057 N1.getOperand(2).getOperand(1),
9061 if (LookThroughFPExt) {
9062 // fold (fadd (fma x, y, (fpext (fmul u, v))), z)
9063 // -> (fma x, y, (fma (fpext u), (fpext v), z))
9064 auto FoldFAddFMAFPExtFMul = [&] (
9065 SDValue X, SDValue Y, SDValue U, SDValue V, SDValue Z) {
9066 return DAG.getNode(PreferredFusedOpcode, SL, VT, X, Y,
9067 DAG.getNode(PreferredFusedOpcode, SL, VT,
9068 DAG.getNode(ISD::FP_EXTEND, SL, VT, U),
9069 DAG.getNode(ISD::FP_EXTEND, SL, VT, V),
9072 if (N0.getOpcode() == PreferredFusedOpcode) {
9073 SDValue N02 = N0.getOperand(2);
9074 if (N02.getOpcode() == ISD::FP_EXTEND) {
9075 SDValue N020 = N02.getOperand(0);
9076 if (isContractableFMUL(N020))
9077 return FoldFAddFMAFPExtFMul(N0.getOperand(0), N0.getOperand(1),
9078 N020.getOperand(0), N020.getOperand(1),
9083 // fold (fadd (fpext (fma x, y, (fmul u, v))), z)
9084 // -> (fma (fpext x), (fpext y), (fma (fpext u), (fpext v), z))
9085 // FIXME: This turns two single-precision and one double-precision
9086 // operation into two double-precision operations, which might not be
9087 // interesting for all targets, especially GPUs.
9088 auto FoldFAddFPExtFMAFMul = [&] (
9089 SDValue X, SDValue Y, SDValue U, SDValue V, SDValue Z) {
9090 return DAG.getNode(PreferredFusedOpcode, SL, VT,
9091 DAG.getNode(ISD::FP_EXTEND, SL, VT, X),
9092 DAG.getNode(ISD::FP_EXTEND, SL, VT, Y),
9093 DAG.getNode(PreferredFusedOpcode, SL, VT,
9094 DAG.getNode(ISD::FP_EXTEND, SL, VT, U),
9095 DAG.getNode(ISD::FP_EXTEND, SL, VT, V),
9098 if (N0.getOpcode() == ISD::FP_EXTEND) {
9099 SDValue N00 = N0.getOperand(0);
9100 if (N00.getOpcode() == PreferredFusedOpcode) {
9101 SDValue N002 = N00.getOperand(2);
9102 if (isContractableFMUL(N002))
9103 return FoldFAddFPExtFMAFMul(N00.getOperand(0), N00.getOperand(1),
9104 N002.getOperand(0), N002.getOperand(1),
9109 // fold (fadd x, (fma y, z, (fpext (fmul u, v)))
9110 // -> (fma y, z, (fma (fpext u), (fpext v), x))
9111 if (N1.getOpcode() == PreferredFusedOpcode) {
9112 SDValue N12 = N1.getOperand(2);
9113 if (N12.getOpcode() == ISD::FP_EXTEND) {
9114 SDValue N120 = N12.getOperand(0);
9115 if (isContractableFMUL(N120))
9116 return FoldFAddFMAFPExtFMul(N1.getOperand(0), N1.getOperand(1),
9117 N120.getOperand(0), N120.getOperand(1),
9122 // fold (fadd x, (fpext (fma y, z, (fmul u, v)))
9123 // -> (fma (fpext y), (fpext z), (fma (fpext u), (fpext v), x))
9124 // FIXME: This turns two single-precision and one double-precision
9125 // operation into two double-precision operations, which might not be
9126 // interesting for all targets, especially GPUs.
9127 if (N1.getOpcode() == ISD::FP_EXTEND) {
9128 SDValue N10 = N1.getOperand(0);
9129 if (N10.getOpcode() == PreferredFusedOpcode) {
9130 SDValue N102 = N10.getOperand(2);
9131 if (isContractableFMUL(N102))
9132 return FoldFAddFPExtFMAFMul(N10.getOperand(0), N10.getOperand(1),
9133 N102.getOperand(0), N102.getOperand(1),
9143 /// Try to perform FMA combining on a given FSUB node.
9144 SDValue DAGCombiner::visitFSUBForFMACombine(SDNode *N) {
9145 SDValue N0 = N->getOperand(0);
9146 SDValue N1 = N->getOperand(1);
9147 EVT VT = N->getValueType(0);
9150 const TargetOptions &Options = DAG.getTarget().Options;
9151 // Floating-point multiply-add with intermediate rounding.
9152 bool HasFMAD = (LegalOperations && TLI.isOperationLegal(ISD::FMAD, VT));
9154 // Floating-point multiply-add without intermediate rounding.
9156 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
9157 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT));
9159 // No valid opcode, do not combine.
9160 if (!HasFMAD && !HasFMA)
9163 bool AllowFusionGlobally = (Options.AllowFPOpFusion == FPOpFusion::Fast ||
9164 Options.UnsafeFPMath || HasFMAD);
9165 // If the subtraction is not contractable, do not combine.
9166 if (!AllowFusionGlobally && !isContractable(N))
9169 const SelectionDAGTargetInfo *STI = DAG.getSubtarget().getSelectionDAGInfo();
9170 if (STI && STI->generateFMAsInMachineCombiner(OptLevel))
9173 // Always prefer FMAD to FMA for precision.
9174 unsigned PreferredFusedOpcode = HasFMAD ? ISD::FMAD : ISD::FMA;
9175 bool Aggressive = TLI.enableAggressiveFMAFusion(VT);
9176 bool LookThroughFPExt = TLI.isFPExtFree(VT);
9178 // Is the node an FMUL and contractable either due to global flags or
9180 auto isContractableFMUL = [AllowFusionGlobally](SDValue N) {
9181 if (N.getOpcode() != ISD::FMUL)
9183 return AllowFusionGlobally || isContractable(N.getNode());
9186 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
9187 if (isContractableFMUL(N0) && (Aggressive || N0->hasOneUse())) {
9188 return DAG.getNode(PreferredFusedOpcode, SL, VT,
9189 N0.getOperand(0), N0.getOperand(1),
9190 DAG.getNode(ISD::FNEG, SL, VT, N1));
9193 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
9194 // Note: Commutes FSUB operands.
9195 if (isContractableFMUL(N1) && (Aggressive || N1->hasOneUse()))
9196 return DAG.getNode(PreferredFusedOpcode, SL, VT,
9197 DAG.getNode(ISD::FNEG, SL, VT,
9199 N1.getOperand(1), N0);
9201 // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
9202 if (N0.getOpcode() == ISD::FNEG && isContractableFMUL(N0.getOperand(0)) &&
9203 (Aggressive || (N0->hasOneUse() && N0.getOperand(0).hasOneUse()))) {
9204 SDValue N00 = N0.getOperand(0).getOperand(0);
9205 SDValue N01 = N0.getOperand(0).getOperand(1);
9206 return DAG.getNode(PreferredFusedOpcode, SL, VT,
9207 DAG.getNode(ISD::FNEG, SL, VT, N00), N01,
9208 DAG.getNode(ISD::FNEG, SL, VT, N1));
9211 // Look through FP_EXTEND nodes to do more combining.
9212 if (LookThroughFPExt) {
9213 // fold (fsub (fpext (fmul x, y)), z)
9214 // -> (fma (fpext x), (fpext y), (fneg z))
9215 if (N0.getOpcode() == ISD::FP_EXTEND) {
9216 SDValue N00 = N0.getOperand(0);
9217 if (isContractableFMUL(N00))
9218 return DAG.getNode(PreferredFusedOpcode, SL, VT,
9219 DAG.getNode(ISD::FP_EXTEND, SL, VT,
9221 DAG.getNode(ISD::FP_EXTEND, SL, VT,
9223 DAG.getNode(ISD::FNEG, SL, VT, N1));
9226 // fold (fsub x, (fpext (fmul y, z)))
9227 // -> (fma (fneg (fpext y)), (fpext z), x)
9228 // Note: Commutes FSUB operands.
9229 if (N1.getOpcode() == ISD::FP_EXTEND) {
9230 SDValue N10 = N1.getOperand(0);
9231 if (isContractableFMUL(N10))
9232 return DAG.getNode(PreferredFusedOpcode, SL, VT,
9233 DAG.getNode(ISD::FNEG, SL, VT,
9234 DAG.getNode(ISD::FP_EXTEND, SL, VT,
9235 N10.getOperand(0))),
9236 DAG.getNode(ISD::FP_EXTEND, SL, VT,
9241 // fold (fsub (fpext (fneg (fmul, x, y))), z)
9242 // -> (fneg (fma (fpext x), (fpext y), z))
9243 // Note: This could be removed with appropriate canonicalization of the
9244 // input expression into (fneg (fadd (fpext (fmul, x, y)), z). However, the
9245 // orthogonal flags -fp-contract=fast and -enable-unsafe-fp-math prevent
9246 // from implementing the canonicalization in visitFSUB.
9247 if (N0.getOpcode() == ISD::FP_EXTEND) {
9248 SDValue N00 = N0.getOperand(0);
9249 if (N00.getOpcode() == ISD::FNEG) {
9250 SDValue N000 = N00.getOperand(0);
9251 if (isContractableFMUL(N000)) {
9252 return DAG.getNode(ISD::FNEG, SL, VT,
9253 DAG.getNode(PreferredFusedOpcode, SL, VT,
9254 DAG.getNode(ISD::FP_EXTEND, SL, VT,
9255 N000.getOperand(0)),
9256 DAG.getNode(ISD::FP_EXTEND, SL, VT,
9257 N000.getOperand(1)),
9263 // fold (fsub (fneg (fpext (fmul, x, y))), z)
9264 // -> (fneg (fma (fpext x)), (fpext y), z)
9265 // Note: This could be removed with appropriate canonicalization of the
9266 // input expression into (fneg (fadd (fpext (fmul, x, y)), z). However, the
9267 // orthogonal flags -fp-contract=fast and -enable-unsafe-fp-math prevent
9268 // from implementing the canonicalization in visitFSUB.
9269 if (N0.getOpcode() == ISD::FNEG) {
9270 SDValue N00 = N0.getOperand(0);
9271 if (N00.getOpcode() == ISD::FP_EXTEND) {
9272 SDValue N000 = N00.getOperand(0);
9273 if (isContractableFMUL(N000)) {
9274 return DAG.getNode(ISD::FNEG, SL, VT,
9275 DAG.getNode(PreferredFusedOpcode, SL, VT,
9276 DAG.getNode(ISD::FP_EXTEND, SL, VT,
9277 N000.getOperand(0)),
9278 DAG.getNode(ISD::FP_EXTEND, SL, VT,
9279 N000.getOperand(1)),
9287 // More folding opportunities when target permits.
9289 // fold (fsub (fma x, y, (fmul u, v)), z)
9290 // -> (fma x, y (fma u, v, (fneg z)))
9291 // FIXME: The UnsafeAlgebra flag should be propagated to FMA/FMAD, but FMF
9292 // are currently only supported on binary nodes.
9293 if (Options.UnsafeFPMath && N0.getOpcode() == PreferredFusedOpcode &&
9294 isContractableFMUL(N0.getOperand(2)) && N0->hasOneUse() &&
9295 N0.getOperand(2)->hasOneUse()) {
9296 return DAG.getNode(PreferredFusedOpcode, SL, VT,
9297 N0.getOperand(0), N0.getOperand(1),
9298 DAG.getNode(PreferredFusedOpcode, SL, VT,
9299 N0.getOperand(2).getOperand(0),
9300 N0.getOperand(2).getOperand(1),
9301 DAG.getNode(ISD::FNEG, SL, VT,
9305 // fold (fsub x, (fma y, z, (fmul u, v)))
9306 // -> (fma (fneg y), z, (fma (fneg u), v, x))
9307 // FIXME: The UnsafeAlgebra flag should be propagated to FMA/FMAD, but FMF
9308 // are currently only supported on binary nodes.
9309 if (Options.UnsafeFPMath && N1.getOpcode() == PreferredFusedOpcode &&
9310 isContractableFMUL(N1.getOperand(2))) {
9311 SDValue N20 = N1.getOperand(2).getOperand(0);
9312 SDValue N21 = N1.getOperand(2).getOperand(1);
9313 return DAG.getNode(PreferredFusedOpcode, SL, VT,
9314 DAG.getNode(ISD::FNEG, SL, VT,
9317 DAG.getNode(PreferredFusedOpcode, SL, VT,
9318 DAG.getNode(ISD::FNEG, SL, VT, N20),
9323 if (LookThroughFPExt) {
9324 // fold (fsub (fma x, y, (fpext (fmul u, v))), z)
9325 // -> (fma x, y (fma (fpext u), (fpext v), (fneg z)))
9326 if (N0.getOpcode() == PreferredFusedOpcode) {
9327 SDValue N02 = N0.getOperand(2);
9328 if (N02.getOpcode() == ISD::FP_EXTEND) {
9329 SDValue N020 = N02.getOperand(0);
9330 if (isContractableFMUL(N020))
9331 return DAG.getNode(PreferredFusedOpcode, SL, VT,
9332 N0.getOperand(0), N0.getOperand(1),
9333 DAG.getNode(PreferredFusedOpcode, SL, VT,
9334 DAG.getNode(ISD::FP_EXTEND, SL, VT,
9335 N020.getOperand(0)),
9336 DAG.getNode(ISD::FP_EXTEND, SL, VT,
9337 N020.getOperand(1)),
9338 DAG.getNode(ISD::FNEG, SL, VT,
9343 // fold (fsub (fpext (fma x, y, (fmul u, v))), z)
9344 // -> (fma (fpext x), (fpext y),
9345 // (fma (fpext u), (fpext v), (fneg z)))
9346 // FIXME: This turns two single-precision and one double-precision
9347 // operation into two double-precision operations, which might not be
9348 // interesting for all targets, especially GPUs.
9349 if (N0.getOpcode() == ISD::FP_EXTEND) {
9350 SDValue N00 = N0.getOperand(0);
9351 if (N00.getOpcode() == PreferredFusedOpcode) {
9352 SDValue N002 = N00.getOperand(2);
9353 if (isContractableFMUL(N002))
9354 return DAG.getNode(PreferredFusedOpcode, SL, VT,
9355 DAG.getNode(ISD::FP_EXTEND, SL, VT,
9357 DAG.getNode(ISD::FP_EXTEND, SL, VT,
9359 DAG.getNode(PreferredFusedOpcode, SL, VT,
9360 DAG.getNode(ISD::FP_EXTEND, SL, VT,
9361 N002.getOperand(0)),
9362 DAG.getNode(ISD::FP_EXTEND, SL, VT,
9363 N002.getOperand(1)),
9364 DAG.getNode(ISD::FNEG, SL, VT,
9369 // fold (fsub x, (fma y, z, (fpext (fmul u, v))))
9370 // -> (fma (fneg y), z, (fma (fneg (fpext u)), (fpext v), x))
9371 if (N1.getOpcode() == PreferredFusedOpcode &&
9372 N1.getOperand(2).getOpcode() == ISD::FP_EXTEND) {
9373 SDValue N120 = N1.getOperand(2).getOperand(0);
9374 if (isContractableFMUL(N120)) {
9375 SDValue N1200 = N120.getOperand(0);
9376 SDValue N1201 = N120.getOperand(1);
9377 return DAG.getNode(PreferredFusedOpcode, SL, VT,
9378 DAG.getNode(ISD::FNEG, SL, VT, N1.getOperand(0)),
9380 DAG.getNode(PreferredFusedOpcode, SL, VT,
9381 DAG.getNode(ISD::FNEG, SL, VT,
9382 DAG.getNode(ISD::FP_EXTEND, SL,
9384 DAG.getNode(ISD::FP_EXTEND, SL, VT,
9390 // fold (fsub x, (fpext (fma y, z, (fmul u, v))))
9391 // -> (fma (fneg (fpext y)), (fpext z),
9392 // (fma (fneg (fpext u)), (fpext v), x))
9393 // FIXME: This turns two single-precision and one double-precision
9394 // operation into two double-precision operations, which might not be
9395 // interesting for all targets, especially GPUs.
9396 if (N1.getOpcode() == ISD::FP_EXTEND &&
9397 N1.getOperand(0).getOpcode() == PreferredFusedOpcode) {
9398 SDValue N100 = N1.getOperand(0).getOperand(0);
9399 SDValue N101 = N1.getOperand(0).getOperand(1);
9400 SDValue N102 = N1.getOperand(0).getOperand(2);
9401 if (isContractableFMUL(N102)) {
9402 SDValue N1020 = N102.getOperand(0);
9403 SDValue N1021 = N102.getOperand(1);
9404 return DAG.getNode(PreferredFusedOpcode, SL, VT,
9405 DAG.getNode(ISD::FNEG, SL, VT,
9406 DAG.getNode(ISD::FP_EXTEND, SL, VT,
9408 DAG.getNode(ISD::FP_EXTEND, SL, VT, N101),
9409 DAG.getNode(PreferredFusedOpcode, SL, VT,
9410 DAG.getNode(ISD::FNEG, SL, VT,
9411 DAG.getNode(ISD::FP_EXTEND, SL,
9413 DAG.getNode(ISD::FP_EXTEND, SL, VT,
9424 /// Try to perform FMA combining on a given FMUL node based on the distributive
9425 /// law x * (y + 1) = x * y + x and variants thereof (commuted versions,
9426 /// subtraction instead of addition).
9427 SDValue DAGCombiner::visitFMULForFMADistributiveCombine(SDNode *N) {
9428 SDValue N0 = N->getOperand(0);
9429 SDValue N1 = N->getOperand(1);
9430 EVT VT = N->getValueType(0);
9433 assert(N->getOpcode() == ISD::FMUL && "Expected FMUL Operation");
9435 const TargetOptions &Options = DAG.getTarget().Options;
9437 // The transforms below are incorrect when x == 0 and y == inf, because the
9438 // intermediate multiplication produces a nan.
9439 if (!Options.NoInfsFPMath)
9442 // Floating-point multiply-add without intermediate rounding.
9444 (Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath) &&
9445 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
9446 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT));
9448 // Floating-point multiply-add with intermediate rounding. This can result
9449 // in a less precise result due to the changed rounding order.
9450 bool HasFMAD = Options.UnsafeFPMath &&
9451 (LegalOperations && TLI.isOperationLegal(ISD::FMAD, VT));
9453 // No valid opcode, do not combine.
9454 if (!HasFMAD && !HasFMA)
9457 // Always prefer FMAD to FMA for precision.
9458 unsigned PreferredFusedOpcode = HasFMAD ? ISD::FMAD : ISD::FMA;
9459 bool Aggressive = TLI.enableAggressiveFMAFusion(VT);
9461 // fold (fmul (fadd x, +1.0), y) -> (fma x, y, y)
9462 // fold (fmul (fadd x, -1.0), y) -> (fma x, y, (fneg y))
9463 auto FuseFADD = [&](SDValue X, SDValue Y) {
9464 if (X.getOpcode() == ISD::FADD && (Aggressive || X->hasOneUse())) {
9465 auto XC1 = isConstOrConstSplatFP(X.getOperand(1));
9466 if (XC1 && XC1->isExactlyValue(+1.0))
9467 return DAG.getNode(PreferredFusedOpcode, SL, VT, X.getOperand(0), Y, Y);
9468 if (XC1 && XC1->isExactlyValue(-1.0))
9469 return DAG.getNode(PreferredFusedOpcode, SL, VT, X.getOperand(0), Y,
9470 DAG.getNode(ISD::FNEG, SL, VT, Y));
9475 if (SDValue FMA = FuseFADD(N0, N1))
9477 if (SDValue FMA = FuseFADD(N1, N0))
9480 // fold (fmul (fsub +1.0, x), y) -> (fma (fneg x), y, y)
9481 // fold (fmul (fsub -1.0, x), y) -> (fma (fneg x), y, (fneg y))
9482 // fold (fmul (fsub x, +1.0), y) -> (fma x, y, (fneg y))
9483 // fold (fmul (fsub x, -1.0), y) -> (fma x, y, y)
9484 auto FuseFSUB = [&](SDValue X, SDValue Y) {
9485 if (X.getOpcode() == ISD::FSUB && (Aggressive || X->hasOneUse())) {
9486 auto XC0 = isConstOrConstSplatFP(X.getOperand(0));
9487 if (XC0 && XC0->isExactlyValue(+1.0))
9488 return DAG.getNode(PreferredFusedOpcode, SL, VT,
9489 DAG.getNode(ISD::FNEG, SL, VT, X.getOperand(1)), Y,
9491 if (XC0 && XC0->isExactlyValue(-1.0))
9492 return DAG.getNode(PreferredFusedOpcode, SL, VT,
9493 DAG.getNode(ISD::FNEG, SL, VT, X.getOperand(1)), Y,
9494 DAG.getNode(ISD::FNEG, SL, VT, Y));
9496 auto XC1 = isConstOrConstSplatFP(X.getOperand(1));
9497 if (XC1 && XC1->isExactlyValue(+1.0))
9498 return DAG.getNode(PreferredFusedOpcode, SL, VT, X.getOperand(0), Y,
9499 DAG.getNode(ISD::FNEG, SL, VT, Y));
9500 if (XC1 && XC1->isExactlyValue(-1.0))
9501 return DAG.getNode(PreferredFusedOpcode, SL, VT, X.getOperand(0), Y, Y);
9506 if (SDValue FMA = FuseFSUB(N0, N1))
9508 if (SDValue FMA = FuseFSUB(N1, N0))
9514 static bool isFMulNegTwo(SDValue &N) {
9515 if (N.getOpcode() != ISD::FMUL)
9517 if (ConstantFPSDNode *CFP = isConstOrConstSplatFP(N.getOperand(1)))
9518 return CFP->isExactlyValue(-2.0);
9522 SDValue DAGCombiner::visitFADD(SDNode *N) {
9523 SDValue N0 = N->getOperand(0);
9524 SDValue N1 = N->getOperand(1);
9525 bool N0CFP = isConstantFPBuildVectorOrConstantFP(N0);
9526 bool N1CFP = isConstantFPBuildVectorOrConstantFP(N1);
9527 EVT VT = N->getValueType(0);
9529 const TargetOptions &Options = DAG.getTarget().Options;
9530 const SDNodeFlags Flags = N->getFlags();
9534 if (SDValue FoldedVOp = SimplifyVBinOp(N))
9537 // fold (fadd c1, c2) -> c1 + c2
9539 return DAG.getNode(ISD::FADD, DL, VT, N0, N1, Flags);
9541 // canonicalize constant to RHS
9542 if (N0CFP && !N1CFP)
9543 return DAG.getNode(ISD::FADD, DL, VT, N1, N0, Flags);
9545 if (SDValue NewSel = foldBinOpIntoSelect(N))
9548 // fold (fadd A, (fneg B)) -> (fsub A, B)
9549 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
9550 isNegatibleForFree(N1, LegalOperations, TLI, &Options) == 2)
9551 return DAG.getNode(ISD::FSUB, DL, VT, N0,
9552 GetNegatedExpression(N1, DAG, LegalOperations), Flags);
9554 // fold (fadd (fneg A), B) -> (fsub B, A)
9555 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
9556 isNegatibleForFree(N0, LegalOperations, TLI, &Options) == 2)
9557 return DAG.getNode(ISD::FSUB, DL, VT, N1,
9558 GetNegatedExpression(N0, DAG, LegalOperations), Flags);
9560 // fold (fadd A, (fmul B, -2.0)) -> (fsub A, (fadd B, B))
9561 // fold (fadd (fmul B, -2.0), A) -> (fsub A, (fadd B, B))
9562 if ((isFMulNegTwo(N0) && N0.hasOneUse()) ||
9563 (isFMulNegTwo(N1) && N1.hasOneUse())) {
9564 bool N1IsFMul = isFMulNegTwo(N1);
9565 SDValue AddOp = N1IsFMul ? N1.getOperand(0) : N0.getOperand(0);
9566 SDValue Add = DAG.getNode(ISD::FADD, DL, VT, AddOp, AddOp, Flags);
9567 return DAG.getNode(ISD::FSUB, DL, VT, N1IsFMul ? N0 : N1, Add, Flags);
9570 // FIXME: Auto-upgrade the target/function-level option.
9571 if (Options.NoSignedZerosFPMath || N->getFlags().hasNoSignedZeros()) {
9572 // fold (fadd A, 0) -> A
9573 if (ConstantFPSDNode *N1C = isConstOrConstSplatFP(N1))
9578 // If 'unsafe math' is enabled, fold lots of things.
9579 if (Options.UnsafeFPMath) {
9580 // No FP constant should be created after legalization as Instruction
9581 // Selection pass has a hard time dealing with FP constants.
9582 bool AllowNewConst = (Level < AfterLegalizeDAG);
9584 // fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
9585 if (N1CFP && N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
9586 isConstantFPBuildVectorOrConstantFP(N0.getOperand(1)))
9587 return DAG.getNode(ISD::FADD, DL, VT, N0.getOperand(0),
9588 DAG.getNode(ISD::FADD, DL, VT, N0.getOperand(1), N1,
9592 // If allowed, fold (fadd (fneg x), x) -> 0.0
9593 if (AllowNewConst && N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1)
9594 return DAG.getConstantFP(0.0, DL, VT);
9596 // If allowed, fold (fadd x, (fneg x)) -> 0.0
9597 if (AllowNewConst && N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0)
9598 return DAG.getConstantFP(0.0, DL, VT);
9600 // We can fold chains of FADD's of the same value into multiplications.
9601 // This transform is not safe in general because we are reducing the number
9602 // of rounding steps.
9603 if (TLI.isOperationLegalOrCustom(ISD::FMUL, VT) && !N0CFP && !N1CFP) {
9604 if (N0.getOpcode() == ISD::FMUL) {
9605 bool CFP00 = isConstantFPBuildVectorOrConstantFP(N0.getOperand(0));
9606 bool CFP01 = isConstantFPBuildVectorOrConstantFP(N0.getOperand(1));
9608 // (fadd (fmul x, c), x) -> (fmul x, c+1)
9609 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
9610 SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, N0.getOperand(1),
9611 DAG.getConstantFP(1.0, DL, VT), Flags);
9612 return DAG.getNode(ISD::FMUL, DL, VT, N1, NewCFP, Flags);
9615 // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2)
9616 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
9617 N1.getOperand(0) == N1.getOperand(1) &&
9618 N0.getOperand(0) == N1.getOperand(0)) {
9619 SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, N0.getOperand(1),
9620 DAG.getConstantFP(2.0, DL, VT), Flags);
9621 return DAG.getNode(ISD::FMUL, DL, VT, N0.getOperand(0), NewCFP, Flags);
9625 if (N1.getOpcode() == ISD::FMUL) {
9626 bool CFP10 = isConstantFPBuildVectorOrConstantFP(N1.getOperand(0));
9627 bool CFP11 = isConstantFPBuildVectorOrConstantFP(N1.getOperand(1));
9629 // (fadd x, (fmul x, c)) -> (fmul x, c+1)
9630 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
9631 SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, N1.getOperand(1),
9632 DAG.getConstantFP(1.0, DL, VT), Flags);
9633 return DAG.getNode(ISD::FMUL, DL, VT, N0, NewCFP, Flags);
9636 // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2)
9637 if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD &&
9638 N0.getOperand(0) == N0.getOperand(1) &&
9639 N1.getOperand(0) == N0.getOperand(0)) {
9640 SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, N1.getOperand(1),
9641 DAG.getConstantFP(2.0, DL, VT), Flags);
9642 return DAG.getNode(ISD::FMUL, DL, VT, N1.getOperand(0), NewCFP, Flags);
9646 if (N0.getOpcode() == ISD::FADD && AllowNewConst) {
9647 bool CFP00 = isConstantFPBuildVectorOrConstantFP(N0.getOperand(0));
9648 // (fadd (fadd x, x), x) -> (fmul x, 3.0)
9649 if (!CFP00 && N0.getOperand(0) == N0.getOperand(1) &&
9650 (N0.getOperand(0) == N1)) {
9651 return DAG.getNode(ISD::FMUL, DL, VT,
9652 N1, DAG.getConstantFP(3.0, DL, VT), Flags);
9656 if (N1.getOpcode() == ISD::FADD && AllowNewConst) {
9657 bool CFP10 = isConstantFPBuildVectorOrConstantFP(N1.getOperand(0));
9658 // (fadd x, (fadd x, x)) -> (fmul x, 3.0)
9659 if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) &&
9660 N1.getOperand(0) == N0) {
9661 return DAG.getNode(ISD::FMUL, DL, VT,
9662 N0, DAG.getConstantFP(3.0, DL, VT), Flags);
9666 // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0)
9667 if (AllowNewConst &&
9668 N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
9669 N0.getOperand(0) == N0.getOperand(1) &&
9670 N1.getOperand(0) == N1.getOperand(1) &&
9671 N0.getOperand(0) == N1.getOperand(0)) {
9672 return DAG.getNode(ISD::FMUL, DL, VT, N0.getOperand(0),
9673 DAG.getConstantFP(4.0, DL, VT), Flags);
9676 } // enable-unsafe-fp-math
9678 // FADD -> FMA combines:
9679 if (SDValue Fused = visitFADDForFMACombine(N)) {
9680 AddToWorklist(Fused.getNode());
9686 SDValue DAGCombiner::visitFSUB(SDNode *N) {
9687 SDValue N0 = N->getOperand(0);
9688 SDValue N1 = N->getOperand(1);
9689 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
9690 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
9691 EVT VT = N->getValueType(0);
9693 const TargetOptions &Options = DAG.getTarget().Options;
9694 const SDNodeFlags Flags = N->getFlags();
9698 if (SDValue FoldedVOp = SimplifyVBinOp(N))
9701 // fold (fsub c1, c2) -> c1-c2
9703 return DAG.getNode(ISD::FSUB, DL, VT, N0, N1, Flags);
9705 if (SDValue NewSel = foldBinOpIntoSelect(N))
9708 // fold (fsub A, (fneg B)) -> (fadd A, B)
9709 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
9710 return DAG.getNode(ISD::FADD, DL, VT, N0,
9711 GetNegatedExpression(N1, DAG, LegalOperations), Flags);
9713 // FIXME: Auto-upgrade the target/function-level option.
9714 if (Options.NoSignedZerosFPMath || N->getFlags().hasNoSignedZeros()) {
9715 // (fsub 0, B) -> -B
9716 if (N0CFP && N0CFP->isZero()) {
9717 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
9718 return GetNegatedExpression(N1, DAG, LegalOperations);
9719 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
9720 return DAG.getNode(ISD::FNEG, DL, VT, N1, Flags);
9724 // If 'unsafe math' is enabled, fold lots of things.
9725 if (Options.UnsafeFPMath) {
9727 if (N1CFP && N1CFP->isZero())
9730 // (fsub x, x) -> 0.0
9732 return DAG.getConstantFP(0.0f, DL, VT);
9734 // (fsub x, (fadd x, y)) -> (fneg y)
9735 // (fsub x, (fadd y, x)) -> (fneg y)
9736 if (N1.getOpcode() == ISD::FADD) {
9737 SDValue N10 = N1->getOperand(0);
9738 SDValue N11 = N1->getOperand(1);
9740 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI, &Options))
9741 return GetNegatedExpression(N11, DAG, LegalOperations);
9743 if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI, &Options))
9744 return GetNegatedExpression(N10, DAG, LegalOperations);
9748 // FSUB -> FMA combines:
9749 if (SDValue Fused = visitFSUBForFMACombine(N)) {
9750 AddToWorklist(Fused.getNode());
9757 SDValue DAGCombiner::visitFMUL(SDNode *N) {
9758 SDValue N0 = N->getOperand(0);
9759 SDValue N1 = N->getOperand(1);
9760 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
9761 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
9762 EVT VT = N->getValueType(0);
9764 const TargetOptions &Options = DAG.getTarget().Options;
9765 const SDNodeFlags Flags = N->getFlags();
9768 if (VT.isVector()) {
9769 // This just handles C1 * C2 for vectors. Other vector folds are below.
9770 if (SDValue FoldedVOp = SimplifyVBinOp(N))
9774 // fold (fmul c1, c2) -> c1*c2
9776 return DAG.getNode(ISD::FMUL, DL, VT, N0, N1, Flags);
9778 // canonicalize constant to RHS
9779 if (isConstantFPBuildVectorOrConstantFP(N0) &&
9780 !isConstantFPBuildVectorOrConstantFP(N1))
9781 return DAG.getNode(ISD::FMUL, DL, VT, N1, N0, Flags);
9783 // fold (fmul A, 1.0) -> A
9784 if (N1CFP && N1CFP->isExactlyValue(1.0))
9787 if (SDValue NewSel = foldBinOpIntoSelect(N))
9790 if (Options.UnsafeFPMath) {
9791 // fold (fmul A, 0) -> 0
9792 if (N1CFP && N1CFP->isZero())
9795 // fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
9796 if (N0.getOpcode() == ISD::FMUL) {
9797 // Fold scalars or any vector constants (not just splats).
9798 // This fold is done in general by InstCombine, but extra fmul insts
9799 // may have been generated during lowering.
9800 SDValue N00 = N0.getOperand(0);
9801 SDValue N01 = N0.getOperand(1);
9802 auto *BV1 = dyn_cast<BuildVectorSDNode>(N1);
9803 auto *BV00 = dyn_cast<BuildVectorSDNode>(N00);
9804 auto *BV01 = dyn_cast<BuildVectorSDNode>(N01);
9806 // Check 1: Make sure that the first operand of the inner multiply is NOT
9807 // a constant. Otherwise, we may induce infinite looping.
9808 if (!(isConstOrConstSplatFP(N00) || (BV00 && BV00->isConstant()))) {
9809 // Check 2: Make sure that the second operand of the inner multiply and
9810 // the second operand of the outer multiply are constants.
9811 if ((N1CFP && isConstOrConstSplatFP(N01)) ||
9812 (BV1 && BV01 && BV1->isConstant() && BV01->isConstant())) {
9813 SDValue MulConsts = DAG.getNode(ISD::FMUL, DL, VT, N01, N1, Flags);
9814 return DAG.getNode(ISD::FMUL, DL, VT, N00, MulConsts, Flags);
9819 // fold (fmul (fadd x, x), c) -> (fmul x, (fmul 2.0, c))
9820 // Undo the fmul 2.0, x -> fadd x, x transformation, since if it occurs
9821 // during an early run of DAGCombiner can prevent folding with fmuls
9822 // inserted during lowering.
9823 if (N0.getOpcode() == ISD::FADD &&
9824 (N0.getOperand(0) == N0.getOperand(1)) &&
9826 const SDValue Two = DAG.getConstantFP(2.0, DL, VT);
9827 SDValue MulConsts = DAG.getNode(ISD::FMUL, DL, VT, Two, N1, Flags);
9828 return DAG.getNode(ISD::FMUL, DL, VT, N0.getOperand(0), MulConsts, Flags);
9832 // fold (fmul X, 2.0) -> (fadd X, X)
9833 if (N1CFP && N1CFP->isExactlyValue(+2.0))
9834 return DAG.getNode(ISD::FADD, DL, VT, N0, N0, Flags);
9836 // fold (fmul X, -1.0) -> (fneg X)
9837 if (N1CFP && N1CFP->isExactlyValue(-1.0))
9838 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
9839 return DAG.getNode(ISD::FNEG, DL, VT, N0);
9841 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
9842 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
9843 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
9844 // Both can be negated for free, check to see if at least one is cheaper
9846 if (LHSNeg == 2 || RHSNeg == 2)
9847 return DAG.getNode(ISD::FMUL, DL, VT,
9848 GetNegatedExpression(N0, DAG, LegalOperations),
9849 GetNegatedExpression(N1, DAG, LegalOperations),
9854 // FMUL -> FMA combines:
9855 if (SDValue Fused = visitFMULForFMADistributiveCombine(N)) {
9856 AddToWorklist(Fused.getNode());
9863 SDValue DAGCombiner::visitFMA(SDNode *N) {
9864 SDValue N0 = N->getOperand(0);
9865 SDValue N1 = N->getOperand(1);
9866 SDValue N2 = N->getOperand(2);
9867 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
9868 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
9869 EVT VT = N->getValueType(0);
9871 const TargetOptions &Options = DAG.getTarget().Options;
9873 // Constant fold FMA.
9874 if (isa<ConstantFPSDNode>(N0) &&
9875 isa<ConstantFPSDNode>(N1) &&
9876 isa<ConstantFPSDNode>(N2)) {
9877 return DAG.getNode(ISD::FMA, DL, VT, N0, N1, N2);
9880 if (Options.UnsafeFPMath) {
9881 if (N0CFP && N0CFP->isZero())
9883 if (N1CFP && N1CFP->isZero())
9886 // TODO: The FMA node should have flags that propagate to these nodes.
9887 if (N0CFP && N0CFP->isExactlyValue(1.0))
9888 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2);
9889 if (N1CFP && N1CFP->isExactlyValue(1.0))
9890 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2);
9892 // Canonicalize (fma c, x, y) -> (fma x, c, y)
9893 if (isConstantFPBuildVectorOrConstantFP(N0) &&
9894 !isConstantFPBuildVectorOrConstantFP(N1))
9895 return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2);
9897 // TODO: FMA nodes should have flags that propagate to the created nodes.
9898 // For now, create a Flags object for use with all unsafe math transforms.
9900 Flags.setUnsafeAlgebra(true);
9902 if (Options.UnsafeFPMath) {
9903 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
9904 if (N2.getOpcode() == ISD::FMUL && N0 == N2.getOperand(0) &&
9905 isConstantFPBuildVectorOrConstantFP(N1) &&
9906 isConstantFPBuildVectorOrConstantFP(N2.getOperand(1))) {
9907 return DAG.getNode(ISD::FMUL, DL, VT, N0,
9908 DAG.getNode(ISD::FADD, DL, VT, N1, N2.getOperand(1),
9912 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
9913 if (N0.getOpcode() == ISD::FMUL &&
9914 isConstantFPBuildVectorOrConstantFP(N1) &&
9915 isConstantFPBuildVectorOrConstantFP(N0.getOperand(1))) {
9916 return DAG.getNode(ISD::FMA, DL, VT,
9918 DAG.getNode(ISD::FMUL, DL, VT, N1, N0.getOperand(1),
9924 // (fma x, 1, y) -> (fadd x, y)
9925 // (fma x, -1, y) -> (fadd (fneg x), y)
9927 if (N1CFP->isExactlyValue(1.0))
9928 // TODO: The FMA node should have flags that propagate to this node.
9929 return DAG.getNode(ISD::FADD, DL, VT, N0, N2);
9931 if (N1CFP->isExactlyValue(-1.0) &&
9932 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
9933 SDValue RHSNeg = DAG.getNode(ISD::FNEG, DL, VT, N0);
9934 AddToWorklist(RHSNeg.getNode());
9935 // TODO: The FMA node should have flags that propagate to this node.
9936 return DAG.getNode(ISD::FADD, DL, VT, N2, RHSNeg);
9940 if (Options.UnsafeFPMath) {
9941 // (fma x, c, x) -> (fmul x, (c+1))
9942 if (N1CFP && N0 == N2) {
9943 return DAG.getNode(ISD::FMUL, DL, VT, N0,
9944 DAG.getNode(ISD::FADD, DL, VT, N1,
9945 DAG.getConstantFP(1.0, DL, VT), Flags),
9949 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
9950 if (N1CFP && N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0) {
9951 return DAG.getNode(ISD::FMUL, DL, VT, N0,
9952 DAG.getNode(ISD::FADD, DL, VT, N1,
9953 DAG.getConstantFP(-1.0, DL, VT), Flags),
9961 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
9963 // E.g., (a / D; b / D;) -> (recip = 1.0 / D; a * recip; b * recip)
9964 // Notice that this is not always beneficial. One reason is different targets
9965 // may have different costs for FDIV and FMUL, so sometimes the cost of two
9966 // FDIVs may be lower than the cost of one FDIV and two FMULs. Another reason
9967 // is the critical path is increased from "one FDIV" to "one FDIV + one FMUL".
9968 SDValue DAGCombiner::combineRepeatedFPDivisors(SDNode *N) {
9969 bool UnsafeMath = DAG.getTarget().Options.UnsafeFPMath;
9970 const SDNodeFlags Flags = N->getFlags();
9971 if (!UnsafeMath && !Flags.hasAllowReciprocal())
9974 // Skip if current node is a reciprocal.
9975 SDValue N0 = N->getOperand(0);
9976 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
9977 if (N0CFP && N0CFP->isExactlyValue(1.0))
9980 // Exit early if the target does not want this transform or if there can't
9981 // possibly be enough uses of the divisor to make the transform worthwhile.
9982 SDValue N1 = N->getOperand(1);
9983 unsigned MinUses = TLI.combineRepeatedFPDivisors();
9984 if (!MinUses || N1->use_size() < MinUses)
9987 // Find all FDIV users of the same divisor.
9988 // Use a set because duplicates may be present in the user list.
9989 SetVector<SDNode *> Users;
9990 for (auto *U : N1->uses()) {
9991 if (U->getOpcode() == ISD::FDIV && U->getOperand(1) == N1) {
9992 // This division is eligible for optimization only if global unsafe math
9993 // is enabled or if this division allows reciprocal formation.
9994 if (UnsafeMath || U->getFlags().hasAllowReciprocal())
9999 // Now that we have the actual number of divisor uses, make sure it meets
10000 // the minimum threshold specified by the target.
10001 if (Users.size() < MinUses)
10004 EVT VT = N->getValueType(0);
10006 SDValue FPOne = DAG.getConstantFP(1.0, DL, VT);
10007 SDValue Reciprocal = DAG.getNode(ISD::FDIV, DL, VT, FPOne, N1, Flags);
10009 // Dividend / Divisor -> Dividend * Reciprocal
10010 for (auto *U : Users) {
10011 SDValue Dividend = U->getOperand(0);
10012 if (Dividend != FPOne) {
10013 SDValue NewNode = DAG.getNode(ISD::FMUL, SDLoc(U), VT, Dividend,
10014 Reciprocal, Flags);
10015 CombineTo(U, NewNode);
10016 } else if (U != Reciprocal.getNode()) {
10017 // In the absence of fast-math-flags, this user node is always the
10018 // same node as Reciprocal, but with FMF they may be different nodes.
10019 CombineTo(U, Reciprocal);
10022 return SDValue(N, 0); // N was replaced.
10025 SDValue DAGCombiner::visitFDIV(SDNode *N) {
10026 SDValue N0 = N->getOperand(0);
10027 SDValue N1 = N->getOperand(1);
10028 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
10029 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
10030 EVT VT = N->getValueType(0);
10032 const TargetOptions &Options = DAG.getTarget().Options;
10033 SDNodeFlags Flags = N->getFlags();
10037 if (SDValue FoldedVOp = SimplifyVBinOp(N))
10040 // fold (fdiv c1, c2) -> c1/c2
10041 if (N0CFP && N1CFP)
10042 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1, Flags);
10044 if (SDValue NewSel = foldBinOpIntoSelect(N))
10047 if (Options.UnsafeFPMath) {
10048 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
10050 // Compute the reciprocal 1.0 / c2.
10051 const APFloat &N1APF = N1CFP->getValueAPF();
10052 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
10053 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
10054 // Only do the transform if the reciprocal is a legal fp immediate that
10055 // isn't too nasty (eg NaN, denormal, ...).
10056 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
10057 (!LegalOperations ||
10058 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
10059 // backend)... we should handle this gracefully after Legalize.
10060 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
10061 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
10062 TLI.isFPImmLegal(Recip, VT)))
10063 return DAG.getNode(ISD::FMUL, DL, VT, N0,
10064 DAG.getConstantFP(Recip, DL, VT), Flags);
10067 // If this FDIV is part of a reciprocal square root, it may be folded
10068 // into a target-specific square root estimate instruction.
10069 if (N1.getOpcode() == ISD::FSQRT) {
10070 if (SDValue RV = buildRsqrtEstimate(N1.getOperand(0), Flags)) {
10071 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV, Flags);
10073 } else if (N1.getOpcode() == ISD::FP_EXTEND &&
10074 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
10075 if (SDValue RV = buildRsqrtEstimate(N1.getOperand(0).getOperand(0),
10077 RV = DAG.getNode(ISD::FP_EXTEND, SDLoc(N1), VT, RV);
10078 AddToWorklist(RV.getNode());
10079 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV, Flags);
10081 } else if (N1.getOpcode() == ISD::FP_ROUND &&
10082 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
10083 if (SDValue RV = buildRsqrtEstimate(N1.getOperand(0).getOperand(0),
10085 RV = DAG.getNode(ISD::FP_ROUND, SDLoc(N1), VT, RV, N1.getOperand(1));
10086 AddToWorklist(RV.getNode());
10087 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV, Flags);
10089 } else if (N1.getOpcode() == ISD::FMUL) {
10090 // Look through an FMUL. Even though this won't remove the FDIV directly,
10091 // it's still worthwhile to get rid of the FSQRT if possible.
10094 if (N1.getOperand(0).getOpcode() == ISD::FSQRT) {
10095 SqrtOp = N1.getOperand(0);
10096 OtherOp = N1.getOperand(1);
10097 } else if (N1.getOperand(1).getOpcode() == ISD::FSQRT) {
10098 SqrtOp = N1.getOperand(1);
10099 OtherOp = N1.getOperand(0);
10101 if (SqrtOp.getNode()) {
10102 // We found a FSQRT, so try to make this fold:
10103 // x / (y * sqrt(z)) -> x * (rsqrt(z) / y)
10104 if (SDValue RV = buildRsqrtEstimate(SqrtOp.getOperand(0), Flags)) {
10105 RV = DAG.getNode(ISD::FDIV, SDLoc(N1), VT, RV, OtherOp, Flags);
10106 AddToWorklist(RV.getNode());
10107 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV, Flags);
10112 // Fold into a reciprocal estimate and multiply instead of a real divide.
10113 if (SDValue RV = BuildReciprocalEstimate(N1, Flags)) {
10114 AddToWorklist(RV.getNode());
10115 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV, Flags);
10119 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
10120 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
10121 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
10122 // Both can be negated for free, check to see if at least one is cheaper
10124 if (LHSNeg == 2 || RHSNeg == 2)
10125 return DAG.getNode(ISD::FDIV, SDLoc(N), VT,
10126 GetNegatedExpression(N0, DAG, LegalOperations),
10127 GetNegatedExpression(N1, DAG, LegalOperations),
10132 if (SDValue CombineRepeatedDivisors = combineRepeatedFPDivisors(N))
10133 return CombineRepeatedDivisors;
10138 SDValue DAGCombiner::visitFREM(SDNode *N) {
10139 SDValue N0 = N->getOperand(0);
10140 SDValue N1 = N->getOperand(1);
10141 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
10142 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
10143 EVT VT = N->getValueType(0);
10145 // fold (frem c1, c2) -> fmod(c1,c2)
10146 if (N0CFP && N1CFP)
10147 return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1, N->getFlags());
10149 if (SDValue NewSel = foldBinOpIntoSelect(N))
10155 SDValue DAGCombiner::visitFSQRT(SDNode *N) {
10156 if (!DAG.getTarget().Options.UnsafeFPMath)
10159 SDValue N0 = N->getOperand(0);
10160 if (TLI.isFsqrtCheap(N0, DAG))
10163 // TODO: FSQRT nodes should have flags that propagate to the created nodes.
10164 // For now, create a Flags object for use with all unsafe math transforms.
10166 Flags.setUnsafeAlgebra(true);
10167 return buildSqrtEstimate(N0, Flags);
10170 /// copysign(x, fp_extend(y)) -> copysign(x, y)
10171 /// copysign(x, fp_round(y)) -> copysign(x, y)
10172 static inline bool CanCombineFCOPYSIGN_EXTEND_ROUND(SDNode *N) {
10173 SDValue N1 = N->getOperand(1);
10174 if ((N1.getOpcode() == ISD::FP_EXTEND ||
10175 N1.getOpcode() == ISD::FP_ROUND)) {
10176 // Do not optimize out type conversion of f128 type yet.
10177 // For some targets like x86_64, configuration is changed to keep one f128
10178 // value in one SSE register, but instruction selection cannot handle
10179 // FCOPYSIGN on SSE registers yet.
10180 EVT N1VT = N1->getValueType(0);
10181 EVT N1Op0VT = N1->getOperand(0)->getValueType(0);
10182 return (N1VT == N1Op0VT || N1Op0VT != MVT::f128);
10187 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
10188 SDValue N0 = N->getOperand(0);
10189 SDValue N1 = N->getOperand(1);
10190 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
10191 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
10192 EVT VT = N->getValueType(0);
10194 if (N0CFP && N1CFP) // Constant fold
10195 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1);
10198 const APFloat &V = N1CFP->getValueAPF();
10199 // copysign(x, c1) -> fabs(x) iff ispos(c1)
10200 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
10201 if (!V.isNegative()) {
10202 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
10203 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
10205 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
10206 return DAG.getNode(ISD::FNEG, SDLoc(N), VT,
10207 DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0));
10211 // copysign(fabs(x), y) -> copysign(x, y)
10212 // copysign(fneg(x), y) -> copysign(x, y)
10213 // copysign(copysign(x,z), y) -> copysign(x, y)
10214 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
10215 N0.getOpcode() == ISD::FCOPYSIGN)
10216 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0.getOperand(0), N1);
10218 // copysign(x, abs(y)) -> abs(x)
10219 if (N1.getOpcode() == ISD::FABS)
10220 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
10222 // copysign(x, copysign(y,z)) -> copysign(x, z)
10223 if (N1.getOpcode() == ISD::FCOPYSIGN)
10224 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1.getOperand(1));
10226 // copysign(x, fp_extend(y)) -> copysign(x, y)
10227 // copysign(x, fp_round(y)) -> copysign(x, y)
10228 if (CanCombineFCOPYSIGN_EXTEND_ROUND(N))
10229 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1.getOperand(0));
10234 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
10235 SDValue N0 = N->getOperand(0);
10236 EVT VT = N->getValueType(0);
10237 EVT OpVT = N0.getValueType();
10239 // fold (sint_to_fp c1) -> c1fp
10240 if (DAG.isConstantIntBuildVectorOrConstantInt(N0) &&
10241 // ...but only if the target supports immediate floating-point values
10242 (!LegalOperations ||
10243 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
10244 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
10246 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
10247 // but UINT_TO_FP is legal on this target, try to convert.
10248 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
10249 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
10250 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
10251 if (DAG.SignBitIsZero(N0))
10252 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
10255 // The next optimizations are desirable only if SELECT_CC can be lowered.
10256 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
10257 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
10258 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
10260 (!LegalOperations ||
10261 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
10264 { N0.getOperand(0), N0.getOperand(1),
10265 DAG.getConstantFP(-1.0, DL, VT), DAG.getConstantFP(0.0, DL, VT),
10266 N0.getOperand(2) };
10267 return DAG.getNode(ISD::SELECT_CC, DL, VT, Ops);
10270 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
10271 // (select_cc x, y, 1.0, 0.0,, cc)
10272 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
10273 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
10274 (!LegalOperations ||
10275 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
10278 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
10279 DAG.getConstantFP(1.0, DL, VT), DAG.getConstantFP(0.0, DL, VT),
10280 N0.getOperand(0).getOperand(2) };
10281 return DAG.getNode(ISD::SELECT_CC, DL, VT, Ops);
10288 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
10289 SDValue N0 = N->getOperand(0);
10290 EVT VT = N->getValueType(0);
10291 EVT OpVT = N0.getValueType();
10293 // fold (uint_to_fp c1) -> c1fp
10294 if (DAG.isConstantIntBuildVectorOrConstantInt(N0) &&
10295 // ...but only if the target supports immediate floating-point values
10296 (!LegalOperations ||
10297 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
10298 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
10300 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
10301 // but SINT_TO_FP is legal on this target, try to convert.
10302 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
10303 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
10304 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
10305 if (DAG.SignBitIsZero(N0))
10306 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
10309 // The next optimizations are desirable only if SELECT_CC can be lowered.
10310 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
10311 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
10313 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
10314 (!LegalOperations ||
10315 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
10318 { N0.getOperand(0), N0.getOperand(1),
10319 DAG.getConstantFP(1.0, DL, VT), DAG.getConstantFP(0.0, DL, VT),
10320 N0.getOperand(2) };
10321 return DAG.getNode(ISD::SELECT_CC, DL, VT, Ops);
10328 // Fold (fp_to_{s/u}int ({s/u}int_to_fpx)) -> zext x, sext x, trunc x, or x
10329 static SDValue FoldIntToFPToInt(SDNode *N, SelectionDAG &DAG) {
10330 SDValue N0 = N->getOperand(0);
10331 EVT VT = N->getValueType(0);
10333 if (N0.getOpcode() != ISD::UINT_TO_FP && N0.getOpcode() != ISD::SINT_TO_FP)
10336 SDValue Src = N0.getOperand(0);
10337 EVT SrcVT = Src.getValueType();
10338 bool IsInputSigned = N0.getOpcode() == ISD::SINT_TO_FP;
10339 bool IsOutputSigned = N->getOpcode() == ISD::FP_TO_SINT;
10341 // We can safely assume the conversion won't overflow the output range,
10342 // because (for example) (uint8_t)18293.f is undefined behavior.
10344 // Since we can assume the conversion won't overflow, our decision as to
10345 // whether the input will fit in the float should depend on the minimum
10346 // of the input range and output range.
10348 // This means this is also safe for a signed input and unsigned output, since
10349 // a negative input would lead to undefined behavior.
10350 unsigned InputSize = (int)SrcVT.getScalarSizeInBits() - IsInputSigned;
10351 unsigned OutputSize = (int)VT.getScalarSizeInBits() - IsOutputSigned;
10352 unsigned ActualSize = std::min(InputSize, OutputSize);
10353 const fltSemantics &sem = DAG.EVTToAPFloatSemantics(N0.getValueType());
10355 // We can only fold away the float conversion if the input range can be
10356 // represented exactly in the float range.
10357 if (APFloat::semanticsPrecision(sem) >= ActualSize) {
10358 if (VT.getScalarSizeInBits() > SrcVT.getScalarSizeInBits()) {
10359 unsigned ExtOp = IsInputSigned && IsOutputSigned ? ISD::SIGN_EXTEND
10360 : ISD::ZERO_EXTEND;
10361 return DAG.getNode(ExtOp, SDLoc(N), VT, Src);
10363 if (VT.getScalarSizeInBits() < SrcVT.getScalarSizeInBits())
10364 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Src);
10365 return DAG.getBitcast(VT, Src);
10370 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
10371 SDValue N0 = N->getOperand(0);
10372 EVT VT = N->getValueType(0);
10374 // fold (fp_to_sint c1fp) -> c1
10375 if (isConstantFPBuildVectorOrConstantFP(N0))
10376 return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0);
10378 return FoldIntToFPToInt(N, DAG);
10381 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
10382 SDValue N0 = N->getOperand(0);
10383 EVT VT = N->getValueType(0);
10385 // fold (fp_to_uint c1fp) -> c1
10386 if (isConstantFPBuildVectorOrConstantFP(N0))
10387 return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0);
10389 return FoldIntToFPToInt(N, DAG);
10392 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
10393 SDValue N0 = N->getOperand(0);
10394 SDValue N1 = N->getOperand(1);
10395 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
10396 EVT VT = N->getValueType(0);
10398 // fold (fp_round c1fp) -> c1fp
10400 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1);
10402 // fold (fp_round (fp_extend x)) -> x
10403 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
10404 return N0.getOperand(0);
10406 // fold (fp_round (fp_round x)) -> (fp_round x)
10407 if (N0.getOpcode() == ISD::FP_ROUND) {
10408 const bool NIsTrunc = N->getConstantOperandVal(1) == 1;
10409 const bool N0IsTrunc = N0.getConstantOperandVal(1) == 1;
10411 // Skip this folding if it results in an fp_round from f80 to f16.
10413 // f80 to f16 always generates an expensive (and as yet, unimplemented)
10414 // libcall to __truncxfhf2 instead of selecting native f16 conversion
10415 // instructions from f32 or f64. Moreover, the first (value-preserving)
10416 // fp_round from f80 to either f32 or f64 may become a NOP in platforms like
10418 if (N0.getOperand(0).getValueType() == MVT::f80 && VT == MVT::f16)
10421 // If the first fp_round isn't a value preserving truncation, it might
10422 // introduce a tie in the second fp_round, that wouldn't occur in the
10423 // single-step fp_round we want to fold to.
10424 // In other words, double rounding isn't the same as rounding.
10425 // Also, this is a value preserving truncation iff both fp_round's are.
10426 if (DAG.getTarget().Options.UnsafeFPMath || N0IsTrunc) {
10428 return DAG.getNode(ISD::FP_ROUND, DL, VT, N0.getOperand(0),
10429 DAG.getIntPtrConstant(NIsTrunc && N0IsTrunc, DL));
10433 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
10434 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
10435 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT,
10436 N0.getOperand(0), N1);
10437 AddToWorklist(Tmp.getNode());
10438 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
10439 Tmp, N0.getOperand(1));
10442 if (SDValue NewVSel = matchVSelectOpSizesWithSetCC(N))
10448 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
10449 SDValue N0 = N->getOperand(0);
10450 EVT VT = N->getValueType(0);
10451 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
10452 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
10454 // fold (fp_round_inreg c1fp) -> c1fp
10455 if (N0CFP && isTypeLegal(EVT)) {
10457 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), DL, EVT);
10458 return DAG.getNode(ISD::FP_EXTEND, DL, VT, Round);
10464 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
10465 SDValue N0 = N->getOperand(0);
10466 EVT VT = N->getValueType(0);
10468 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
10469 if (N->hasOneUse() &&
10470 N->use_begin()->getOpcode() == ISD::FP_ROUND)
10473 // fold (fp_extend c1fp) -> c1fp
10474 if (isConstantFPBuildVectorOrConstantFP(N0))
10475 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0);
10477 // fold (fp_extend (fp16_to_fp op)) -> (fp16_to_fp op)
10478 if (N0.getOpcode() == ISD::FP16_TO_FP &&
10479 TLI.getOperationAction(ISD::FP16_TO_FP, VT) == TargetLowering::Legal)
10480 return DAG.getNode(ISD::FP16_TO_FP, SDLoc(N), VT, N0.getOperand(0));
10482 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
10484 if (N0.getOpcode() == ISD::FP_ROUND
10485 && N0.getConstantOperandVal(1) == 1) {
10486 SDValue In = N0.getOperand(0);
10487 if (In.getValueType() == VT) return In;
10488 if (VT.bitsLT(In.getValueType()))
10489 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT,
10490 In, N0.getOperand(1));
10491 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In);
10494 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
10495 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
10496 TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) {
10497 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
10498 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
10500 LN0->getBasePtr(), N0.getValueType(),
10501 LN0->getMemOperand());
10502 CombineTo(N, ExtLoad);
10503 CombineTo(N0.getNode(),
10504 DAG.getNode(ISD::FP_ROUND, SDLoc(N0),
10505 N0.getValueType(), ExtLoad,
10506 DAG.getIntPtrConstant(1, SDLoc(N0))),
10507 ExtLoad.getValue(1));
10508 return SDValue(N, 0); // Return N so it doesn't get rechecked!
10511 if (SDValue NewVSel = matchVSelectOpSizesWithSetCC(N))
10517 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
10518 SDValue N0 = N->getOperand(0);
10519 EVT VT = N->getValueType(0);
10521 // fold (fceil c1) -> fceil(c1)
10522 if (isConstantFPBuildVectorOrConstantFP(N0))
10523 return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0);
10528 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
10529 SDValue N0 = N->getOperand(0);
10530 EVT VT = N->getValueType(0);
10532 // fold (ftrunc c1) -> ftrunc(c1)
10533 if (isConstantFPBuildVectorOrConstantFP(N0))
10534 return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0);
10539 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
10540 SDValue N0 = N->getOperand(0);
10541 EVT VT = N->getValueType(0);
10543 // fold (ffloor c1) -> ffloor(c1)
10544 if (isConstantFPBuildVectorOrConstantFP(N0))
10545 return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0);
10550 // FIXME: FNEG and FABS have a lot in common; refactor.
10551 SDValue DAGCombiner::visitFNEG(SDNode *N) {
10552 SDValue N0 = N->getOperand(0);
10553 EVT VT = N->getValueType(0);
10555 // Constant fold FNEG.
10556 if (isConstantFPBuildVectorOrConstantFP(N0))
10557 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
10559 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
10560 &DAG.getTarget().Options))
10561 return GetNegatedExpression(N0, DAG, LegalOperations);
10563 // Transform fneg(bitconvert(x)) -> bitconvert(x ^ sign) to avoid loading
10564 // constant pool values.
10565 if (!TLI.isFNegFree(VT) &&
10566 N0.getOpcode() == ISD::BITCAST &&
10567 N0.getNode()->hasOneUse()) {
10568 SDValue Int = N0.getOperand(0);
10569 EVT IntVT = Int.getValueType();
10570 if (IntVT.isInteger() && !IntVT.isVector()) {
10572 if (N0.getValueType().isVector()) {
10573 // For a vector, get a mask such as 0x80... per scalar element
10575 SignMask = APInt::getSignMask(N0.getScalarValueSizeInBits());
10576 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
10578 // For a scalar, just generate 0x80...
10579 SignMask = APInt::getSignMask(IntVT.getSizeInBits());
10582 Int = DAG.getNode(ISD::XOR, DL0, IntVT, Int,
10583 DAG.getConstant(SignMask, DL0, IntVT));
10584 AddToWorklist(Int.getNode());
10585 return DAG.getBitcast(VT, Int);
10589 // (fneg (fmul c, x)) -> (fmul -c, x)
10590 if (N0.getOpcode() == ISD::FMUL &&
10591 (N0.getNode()->hasOneUse() || !TLI.isFNegFree(VT))) {
10592 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
10594 APFloat CVal = CFP1->getValueAPF();
10596 if (Level >= AfterLegalizeDAG &&
10597 (TLI.isFPImmLegal(CVal, VT) ||
10598 TLI.isOperationLegal(ISD::ConstantFP, VT)))
10599 return DAG.getNode(
10600 ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
10601 DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0.getOperand(1)),
10609 SDValue DAGCombiner::visitFMINNUM(SDNode *N) {
10610 SDValue N0 = N->getOperand(0);
10611 SDValue N1 = N->getOperand(1);
10612 EVT VT = N->getValueType(0);
10613 const ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
10614 const ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
10616 if (N0CFP && N1CFP) {
10617 const APFloat &C0 = N0CFP->getValueAPF();
10618 const APFloat &C1 = N1CFP->getValueAPF();
10619 return DAG.getConstantFP(minnum(C0, C1), SDLoc(N), VT);
10622 // Canonicalize to constant on RHS.
10623 if (isConstantFPBuildVectorOrConstantFP(N0) &&
10624 !isConstantFPBuildVectorOrConstantFP(N1))
10625 return DAG.getNode(ISD::FMINNUM, SDLoc(N), VT, N1, N0);
10630 SDValue DAGCombiner::visitFMAXNUM(SDNode *N) {
10631 SDValue N0 = N->getOperand(0);
10632 SDValue N1 = N->getOperand(1);
10633 EVT VT = N->getValueType(0);
10634 const ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
10635 const ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
10637 if (N0CFP && N1CFP) {
10638 const APFloat &C0 = N0CFP->getValueAPF();
10639 const APFloat &C1 = N1CFP->getValueAPF();
10640 return DAG.getConstantFP(maxnum(C0, C1), SDLoc(N), VT);
10643 // Canonicalize to constant on RHS.
10644 if (isConstantFPBuildVectorOrConstantFP(N0) &&
10645 !isConstantFPBuildVectorOrConstantFP(N1))
10646 return DAG.getNode(ISD::FMAXNUM, SDLoc(N), VT, N1, N0);
10651 SDValue DAGCombiner::visitFABS(SDNode *N) {
10652 SDValue N0 = N->getOperand(0);
10653 EVT VT = N->getValueType(0);
10655 // fold (fabs c1) -> fabs(c1)
10656 if (isConstantFPBuildVectorOrConstantFP(N0))
10657 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
10659 // fold (fabs (fabs x)) -> (fabs x)
10660 if (N0.getOpcode() == ISD::FABS)
10661 return N->getOperand(0);
10663 // fold (fabs (fneg x)) -> (fabs x)
10664 // fold (fabs (fcopysign x, y)) -> (fabs x)
10665 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
10666 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0));
10668 // Transform fabs(bitconvert(x)) -> bitconvert(x & ~sign) to avoid loading
10669 // constant pool values.
10670 if (!TLI.isFAbsFree(VT) &&
10671 N0.getOpcode() == ISD::BITCAST &&
10672 N0.getNode()->hasOneUse()) {
10673 SDValue Int = N0.getOperand(0);
10674 EVT IntVT = Int.getValueType();
10675 if (IntVT.isInteger() && !IntVT.isVector()) {
10677 if (N0.getValueType().isVector()) {
10678 // For a vector, get a mask such as 0x7f... per scalar element
10680 SignMask = ~APInt::getSignMask(N0.getScalarValueSizeInBits());
10681 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
10683 // For a scalar, just generate 0x7f...
10684 SignMask = ~APInt::getSignMask(IntVT.getSizeInBits());
10687 Int = DAG.getNode(ISD::AND, DL, IntVT, Int,
10688 DAG.getConstant(SignMask, DL, IntVT));
10689 AddToWorklist(Int.getNode());
10690 return DAG.getBitcast(N->getValueType(0), Int);
10697 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
10698 SDValue Chain = N->getOperand(0);
10699 SDValue N1 = N->getOperand(1);
10700 SDValue N2 = N->getOperand(2);
10702 // If N is a constant we could fold this into a fallthrough or unconditional
10703 // branch. However that doesn't happen very often in normal code, because
10704 // Instcombine/SimplifyCFG should have handled the available opportunities.
10705 // If we did this folding here, it would be necessary to update the
10706 // MachineBasicBlock CFG, which is awkward.
10708 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
10710 if (N1.getOpcode() == ISD::SETCC &&
10711 TLI.isOperationLegalOrCustom(ISD::BR_CC,
10712 N1.getOperand(0).getValueType())) {
10713 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
10714 Chain, N1.getOperand(2),
10715 N1.getOperand(0), N1.getOperand(1), N2);
10718 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
10719 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
10720 (N1.getOperand(0).hasOneUse() &&
10721 N1.getOperand(0).getOpcode() == ISD::SRL))) {
10722 SDNode *Trunc = nullptr;
10723 if (N1.getOpcode() == ISD::TRUNCATE) {
10724 // Look pass the truncate.
10725 Trunc = N1.getNode();
10726 N1 = N1.getOperand(0);
10729 // Match this pattern so that we can generate simpler code:
10732 // %b = and i32 %a, 2
10733 // %c = srl i32 %b, 1
10734 // brcond i32 %c ...
10739 // %b = and i32 %a, 2
10740 // %c = setcc eq %b, 0
10743 // This applies only when the AND constant value has one bit set and the
10744 // SRL constant is equal to the log2 of the AND constant. The back-end is
10745 // smart enough to convert the result into a TEST/JMP sequence.
10746 SDValue Op0 = N1.getOperand(0);
10747 SDValue Op1 = N1.getOperand(1);
10749 if (Op0.getOpcode() == ISD::AND &&
10750 Op1.getOpcode() == ISD::Constant) {
10751 SDValue AndOp1 = Op0.getOperand(1);
10753 if (AndOp1.getOpcode() == ISD::Constant) {
10754 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
10756 if (AndConst.isPowerOf2() &&
10757 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
10761 getSetCCResultType(Op0.getValueType()),
10762 Op0, DAG.getConstant(0, DL, Op0.getValueType()),
10765 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, DL,
10766 MVT::Other, Chain, SetCC, N2);
10767 // Don't add the new BRCond into the worklist or else SimplifySelectCC
10768 // will convert it back to (X & C1) >> C2.
10769 CombineTo(N, NewBRCond, false);
10770 // Truncate is dead.
10772 deleteAndRecombine(Trunc);
10773 // Replace the uses of SRL with SETCC
10774 WorklistRemover DeadNodes(*this);
10775 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
10776 deleteAndRecombine(N1.getNode());
10777 return SDValue(N, 0); // Return N so it doesn't get rechecked!
10783 // Restore N1 if the above transformation doesn't match.
10784 N1 = N->getOperand(1);
10787 // Transform br(xor(x, y)) -> br(x != y)
10788 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
10789 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
10790 SDNode *TheXor = N1.getNode();
10791 SDValue Op0 = TheXor->getOperand(0);
10792 SDValue Op1 = TheXor->getOperand(1);
10793 if (Op0.getOpcode() == Op1.getOpcode()) {
10794 // Avoid missing important xor optimizations.
10795 if (SDValue Tmp = visitXOR(TheXor)) {
10796 if (Tmp.getNode() != TheXor) {
10797 DEBUG(dbgs() << "\nReplacing.8 ";
10798 TheXor->dump(&DAG);
10799 dbgs() << "\nWith: ";
10800 Tmp.getNode()->dump(&DAG);
10802 WorklistRemover DeadNodes(*this);
10803 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
10804 deleteAndRecombine(TheXor);
10805 return DAG.getNode(ISD::BRCOND, SDLoc(N),
10806 MVT::Other, Chain, Tmp, N2);
10809 // visitXOR has changed XOR's operands or replaced the XOR completely,
10811 return SDValue(N, 0);
10815 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
10816 bool Equal = false;
10817 if (isOneConstant(Op0) && Op0.hasOneUse() &&
10818 Op0.getOpcode() == ISD::XOR) {
10819 TheXor = Op0.getNode();
10823 EVT SetCCVT = N1.getValueType();
10825 SetCCVT = getSetCCResultType(SetCCVT);
10826 SDValue SetCC = DAG.getSetCC(SDLoc(TheXor),
10829 Equal ? ISD::SETEQ : ISD::SETNE);
10830 // Replace the uses of XOR with SETCC
10831 WorklistRemover DeadNodes(*this);
10832 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
10833 deleteAndRecombine(N1.getNode());
10834 return DAG.getNode(ISD::BRCOND, SDLoc(N),
10835 MVT::Other, Chain, SetCC, N2);
10842 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
10844 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
10845 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
10846 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
10848 // If N is a constant we could fold this into a fallthrough or unconditional
10849 // branch. However that doesn't happen very often in normal code, because
10850 // Instcombine/SimplifyCFG should have handled the available opportunities.
10851 // If we did this folding here, it would be necessary to update the
10852 // MachineBasicBlock CFG, which is awkward.
10854 // Use SimplifySetCC to simplify SETCC's.
10855 SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()),
10856 CondLHS, CondRHS, CC->get(), SDLoc(N),
10858 if (Simp.getNode()) AddToWorklist(Simp.getNode());
10860 // fold to a simpler setcc
10861 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
10862 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
10863 N->getOperand(0), Simp.getOperand(2),
10864 Simp.getOperand(0), Simp.getOperand(1),
10870 /// Return true if 'Use' is a load or a store that uses N as its base pointer
10871 /// and that N may be folded in the load / store addressing mode.
10872 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
10874 const TargetLowering &TLI) {
10878 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
10879 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
10881 VT = LD->getMemoryVT();
10882 AS = LD->getAddressSpace();
10883 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
10884 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
10886 VT = ST->getMemoryVT();
10887 AS = ST->getAddressSpace();
10891 TargetLowering::AddrMode AM;
10892 if (N->getOpcode() == ISD::ADD) {
10893 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
10896 AM.BaseOffs = Offset->getSExtValue();
10900 } else if (N->getOpcode() == ISD::SUB) {
10901 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
10904 AM.BaseOffs = -Offset->getSExtValue();
10911 return TLI.isLegalAddressingMode(DAG.getDataLayout(), AM,
10912 VT.getTypeForEVT(*DAG.getContext()), AS);
10915 /// Try turning a load/store into a pre-indexed load/store when the base
10916 /// pointer is an add or subtract and it has other uses besides the load/store.
10917 /// After the transformation, the new indexed load/store has effectively folded
10918 /// the add/subtract in and all of its other uses are redirected to the
10919 /// new load/store.
10920 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
10921 if (Level < AfterLegalizeDAG)
10924 bool isLoad = true;
10927 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10928 if (LD->isIndexed())
10930 VT = LD->getMemoryVT();
10931 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
10932 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
10934 Ptr = LD->getBasePtr();
10935 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10936 if (ST->isIndexed())
10938 VT = ST->getMemoryVT();
10939 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
10940 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
10942 Ptr = ST->getBasePtr();
10948 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
10949 // out. There is no reason to make this a preinc/predec.
10950 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
10951 Ptr.getNode()->hasOneUse())
10954 // Ask the target to do addressing mode selection.
10957 ISD::MemIndexedMode AM = ISD::UNINDEXED;
10958 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
10961 // Backends without true r+i pre-indexed forms may need to pass a
10962 // constant base with a variable offset so that constant coercion
10963 // will work with the patterns in canonical form.
10964 bool Swapped = false;
10965 if (isa<ConstantSDNode>(BasePtr)) {
10966 std::swap(BasePtr, Offset);
10970 // Don't create a indexed load / store with zero offset.
10971 if (isNullConstant(Offset))
10974 // Try turning it into a pre-indexed load / store except when:
10975 // 1) The new base ptr is a frame index.
10976 // 2) If N is a store and the new base ptr is either the same as or is a
10977 // predecessor of the value being stored.
10978 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
10979 // that would create a cycle.
10980 // 4) All uses are load / store ops that use it as old base ptr.
10982 // Check #1. Preinc'ing a frame index would require copying the stack pointer
10983 // (plus the implicit offset) to a register to preinc anyway.
10984 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
10989 SDValue Val = cast<StoreSDNode>(N)->getValue();
10990 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
10994 // Caches for hasPredecessorHelper.
10995 SmallPtrSet<const SDNode *, 32> Visited;
10996 SmallVector<const SDNode *, 16> Worklist;
10997 Worklist.push_back(N);
10999 // If the offset is a constant, there may be other adds of constants that
11000 // can be folded with this one. We should do this to avoid having to keep
11001 // a copy of the original base pointer.
11002 SmallVector<SDNode *, 16> OtherUses;
11003 if (isa<ConstantSDNode>(Offset))
11004 for (SDNode::use_iterator UI = BasePtr.getNode()->use_begin(),
11005 UE = BasePtr.getNode()->use_end();
11007 SDUse &Use = UI.getUse();
11008 // Skip the use that is Ptr and uses of other results from BasePtr's
11009 // node (important for nodes that return multiple results).
11010 if (Use.getUser() == Ptr.getNode() || Use != BasePtr)
11013 if (SDNode::hasPredecessorHelper(Use.getUser(), Visited, Worklist))
11016 if (Use.getUser()->getOpcode() != ISD::ADD &&
11017 Use.getUser()->getOpcode() != ISD::SUB) {
11022 SDValue Op1 = Use.getUser()->getOperand((UI.getOperandNo() + 1) & 1);
11023 if (!isa<ConstantSDNode>(Op1)) {
11028 // FIXME: In some cases, we can be smarter about this.
11029 if (Op1.getValueType() != Offset.getValueType()) {
11034 OtherUses.push_back(Use.getUser());
11038 std::swap(BasePtr, Offset);
11040 // Now check for #3 and #4.
11041 bool RealUse = false;
11043 for (SDNode *Use : Ptr.getNode()->uses()) {
11046 if (SDNode::hasPredecessorHelper(Use, Visited, Worklist))
11049 // If Ptr may be folded in addressing mode of other use, then it's
11050 // not profitable to do this transformation.
11051 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
11060 Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
11061 BasePtr, Offset, AM);
11063 Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
11064 BasePtr, Offset, AM);
11067 DEBUG(dbgs() << "\nReplacing.4 ";
11069 dbgs() << "\nWith: ";
11070 Result.getNode()->dump(&DAG);
11072 WorklistRemover DeadNodes(*this);
11074 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
11075 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
11077 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
11080 // Finally, since the node is now dead, remove it from the graph.
11081 deleteAndRecombine(N);
11084 std::swap(BasePtr, Offset);
11086 // Replace other uses of BasePtr that can be updated to use Ptr
11087 for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) {
11088 unsigned OffsetIdx = 1;
11089 if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode())
11091 assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() ==
11092 BasePtr.getNode() && "Expected BasePtr operand");
11094 // We need to replace ptr0 in the following expression:
11095 // x0 * offset0 + y0 * ptr0 = t0
11097 // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store)
11099 // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the
11100 // indexed load/store and the expresion that needs to be re-written.
11102 // Therefore, we have:
11103 // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1
11105 ConstantSDNode *CN =
11106 cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx));
11107 int X0, X1, Y0, Y1;
11108 const APInt &Offset0 = CN->getAPIntValue();
11109 APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue();
11111 X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1;
11112 Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1;
11113 X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1;
11114 Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1;
11116 unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD;
11118 APInt CNV = Offset0;
11119 if (X0 < 0) CNV = -CNV;
11120 if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1;
11121 else CNV = CNV - Offset1;
11123 SDLoc DL(OtherUses[i]);
11125 // We can now generate the new expression.
11126 SDValue NewOp1 = DAG.getConstant(CNV, DL, CN->getValueType(0));
11127 SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0);
11129 SDValue NewUse = DAG.getNode(Opcode,
11131 OtherUses[i]->getValueType(0), NewOp1, NewOp2);
11132 DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse);
11133 deleteAndRecombine(OtherUses[i]);
11136 // Replace the uses of Ptr with uses of the updated base value.
11137 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
11138 deleteAndRecombine(Ptr.getNode());
11143 /// Try to combine a load/store with a add/sub of the base pointer node into a
11144 /// post-indexed load/store. The transformation folded the add/subtract into the
11145 /// new indexed load/store effectively and all of its uses are redirected to the
11146 /// new load/store.
11147 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
11148 if (Level < AfterLegalizeDAG)
11151 bool isLoad = true;
11154 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
11155 if (LD->isIndexed())
11157 VT = LD->getMemoryVT();
11158 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
11159 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
11161 Ptr = LD->getBasePtr();
11162 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
11163 if (ST->isIndexed())
11165 VT = ST->getMemoryVT();
11166 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
11167 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
11169 Ptr = ST->getBasePtr();
11175 if (Ptr.getNode()->hasOneUse())
11178 for (SDNode *Op : Ptr.getNode()->uses()) {
11180 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
11185 ISD::MemIndexedMode AM = ISD::UNINDEXED;
11186 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
11187 // Don't create a indexed load / store with zero offset.
11188 if (isNullConstant(Offset))
11191 // Try turning it into a post-indexed load / store except when
11192 // 1) All uses are load / store ops that use it as base ptr (and
11193 // it may be folded as addressing mmode).
11194 // 2) Op must be independent of N, i.e. Op is neither a predecessor
11195 // nor a successor of N. Otherwise, if Op is folded that would
11198 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
11202 bool TryNext = false;
11203 for (SDNode *Use : BasePtr.getNode()->uses()) {
11204 if (Use == Ptr.getNode())
11207 // If all the uses are load / store addresses, then don't do the
11209 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
11210 bool RealUse = false;
11211 for (SDNode *UseUse : Use->uses()) {
11212 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
11227 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
11228 SDValue Result = isLoad
11229 ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
11230 BasePtr, Offset, AM)
11231 : DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
11232 BasePtr, Offset, AM);
11233 ++PostIndexedNodes;
11235 DEBUG(dbgs() << "\nReplacing.5 ";
11237 dbgs() << "\nWith: ";
11238 Result.getNode()->dump(&DAG);
11240 WorklistRemover DeadNodes(*this);
11242 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
11243 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
11245 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
11248 // Finally, since the node is now dead, remove it from the graph.
11249 deleteAndRecombine(N);
11251 // Replace the uses of Use with uses of the updated base value.
11252 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
11253 Result.getValue(isLoad ? 1 : 0));
11254 deleteAndRecombine(Op);
11263 /// \brief Return the base-pointer arithmetic from an indexed \p LD.
11264 SDValue DAGCombiner::SplitIndexingFromLoad(LoadSDNode *LD) {
11265 ISD::MemIndexedMode AM = LD->getAddressingMode();
11266 assert(AM != ISD::UNINDEXED);
11267 SDValue BP = LD->getOperand(1);
11268 SDValue Inc = LD->getOperand(2);
11270 // Some backends use TargetConstants for load offsets, but don't expect
11271 // TargetConstants in general ADD nodes. We can convert these constants into
11272 // regular Constants (if the constant is not opaque).
11273 assert((Inc.getOpcode() != ISD::TargetConstant ||
11274 !cast<ConstantSDNode>(Inc)->isOpaque()) &&
11275 "Cannot split out indexing using opaque target constants");
11276 if (Inc.getOpcode() == ISD::TargetConstant) {
11277 ConstantSDNode *ConstInc = cast<ConstantSDNode>(Inc);
11278 Inc = DAG.getConstant(*ConstInc->getConstantIntValue(), SDLoc(Inc),
11279 ConstInc->getValueType(0));
11283 (AM == ISD::PRE_INC || AM == ISD::POST_INC ? ISD::ADD : ISD::SUB);
11284 return DAG.getNode(Opc, SDLoc(LD), BP.getSimpleValueType(), BP, Inc);
11287 SDValue DAGCombiner::visitLOAD(SDNode *N) {
11288 LoadSDNode *LD = cast<LoadSDNode>(N);
11289 SDValue Chain = LD->getChain();
11290 SDValue Ptr = LD->getBasePtr();
11292 // If load is not volatile and there are no uses of the loaded value (and
11293 // the updated indexed value in case of indexed loads), change uses of the
11294 // chain value into uses of the chain input (i.e. delete the dead load).
11295 if (!LD->isVolatile()) {
11296 if (N->getValueType(1) == MVT::Other) {
11297 // Unindexed loads.
11298 if (!N->hasAnyUseOfValue(0)) {
11299 // It's not safe to use the two value CombineTo variant here. e.g.
11300 // v1, chain2 = load chain1, loc
11301 // v2, chain3 = load chain2, loc
11303 // Now we replace use of chain2 with chain1. This makes the second load
11304 // isomorphic to the one we are deleting, and thus makes this load live.
11305 DEBUG(dbgs() << "\nReplacing.6 ";
11307 dbgs() << "\nWith chain: ";
11308 Chain.getNode()->dump(&DAG);
11310 WorklistRemover DeadNodes(*this);
11311 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
11312 AddUsersToWorklist(Chain.getNode());
11313 if (N->use_empty())
11314 deleteAndRecombine(N);
11316 return SDValue(N, 0); // Return N so it doesn't get rechecked!
11320 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
11322 // If this load has an opaque TargetConstant offset, then we cannot split
11323 // the indexing into an add/sub directly (that TargetConstant may not be
11324 // valid for a different type of node, and we cannot convert an opaque
11325 // target constant into a regular constant).
11326 bool HasOTCInc = LD->getOperand(2).getOpcode() == ISD::TargetConstant &&
11327 cast<ConstantSDNode>(LD->getOperand(2))->isOpaque();
11329 if (!N->hasAnyUseOfValue(0) &&
11330 ((MaySplitLoadIndex && !HasOTCInc) || !N->hasAnyUseOfValue(1))) {
11331 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
11333 if (N->hasAnyUseOfValue(1) && MaySplitLoadIndex && !HasOTCInc) {
11334 Index = SplitIndexingFromLoad(LD);
11335 // Try to fold the base pointer arithmetic into subsequent loads and
11337 AddUsersToWorklist(N);
11339 Index = DAG.getUNDEF(N->getValueType(1));
11340 DEBUG(dbgs() << "\nReplacing.7 ";
11342 dbgs() << "\nWith: ";
11343 Undef.getNode()->dump(&DAG);
11344 dbgs() << " and 2 other values\n");
11345 WorklistRemover DeadNodes(*this);
11346 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
11347 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Index);
11348 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
11349 deleteAndRecombine(N);
11350 return SDValue(N, 0); // Return N so it doesn't get rechecked!
11355 // If this load is directly stored, replace the load value with the stored
11357 // TODO: Handle store large -> read small portion.
11358 // TODO: Handle TRUNCSTORE/LOADEXT
11359 if (OptLevel != CodeGenOpt::None &&
11360 ISD::isNormalLoad(N) && !LD->isVolatile()) {
11361 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
11362 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
11363 if (PrevST->getBasePtr() == Ptr &&
11364 PrevST->getValue().getValueType() == N->getValueType(0))
11365 return CombineTo(N, PrevST->getOperand(1), Chain);
11369 // Try to infer better alignment information than the load already has.
11370 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
11371 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
11372 if (Align > LD->getMemOperand()->getBaseAlignment()) {
11373 SDValue NewLoad = DAG.getExtLoad(
11374 LD->getExtensionType(), SDLoc(N), LD->getValueType(0), Chain, Ptr,
11375 LD->getPointerInfo(), LD->getMemoryVT(), Align,
11376 LD->getMemOperand()->getFlags(), LD->getAAInfo());
11377 if (NewLoad.getNode() != N)
11378 return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true);
11383 if (LD->isUnindexed()) {
11384 // Walk up chain skipping non-aliasing memory nodes.
11385 SDValue BetterChain = FindBetterChain(N, Chain);
11387 // If there is a better chain.
11388 if (Chain != BetterChain) {
11391 // Replace the chain to void dependency.
11392 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
11393 ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD),
11394 BetterChain, Ptr, LD->getMemOperand());
11396 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD),
11397 LD->getValueType(0),
11398 BetterChain, Ptr, LD->getMemoryVT(),
11399 LD->getMemOperand());
11402 // Create token factor to keep old chain connected.
11403 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
11404 MVT::Other, Chain, ReplLoad.getValue(1));
11406 // Make sure the new and old chains are cleaned up.
11407 AddToWorklist(Token.getNode());
11409 // Replace uses with load result and token factor. Don't add users
11411 return CombineTo(N, ReplLoad.getValue(0), Token, false);
11415 // Try transforming N to an indexed load.
11416 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
11417 return SDValue(N, 0);
11419 // Try to slice up N to more direct loads if the slices are mapped to
11420 // different register banks or pairing can take place.
11421 if (SliceUpLoad(N))
11422 return SDValue(N, 0);
11428 /// \brief Helper structure used to slice a load in smaller loads.
11429 /// Basically a slice is obtained from the following sequence:
11430 /// Origin = load Ty1, Base
11431 /// Shift = srl Ty1 Origin, CstTy Amount
11432 /// Inst = trunc Shift to Ty2
11434 /// Then, it will be rewriten into:
11435 /// Slice = load SliceTy, Base + SliceOffset
11436 /// [Inst = zext Slice to Ty2], only if SliceTy <> Ty2
11438 /// SliceTy is deduced from the number of bits that are actually used to
11440 struct LoadedSlice {
11441 /// \brief Helper structure used to compute the cost of a slice.
11443 /// Are we optimizing for code size.
11447 unsigned Truncates;
11448 unsigned CrossRegisterBanksCopies;
11452 Cost(bool ForCodeSize = false)
11453 : ForCodeSize(ForCodeSize), Loads(0), Truncates(0),
11454 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {}
11456 /// \brief Get the cost of one isolated slice.
11457 Cost(const LoadedSlice &LS, bool ForCodeSize = false)
11458 : ForCodeSize(ForCodeSize), Loads(1), Truncates(0),
11459 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {
11460 EVT TruncType = LS.Inst->getValueType(0);
11461 EVT LoadedType = LS.getLoadedType();
11462 if (TruncType != LoadedType &&
11463 !LS.DAG->getTargetLoweringInfo().isZExtFree(LoadedType, TruncType))
11467 /// \brief Account for slicing gain in the current cost.
11468 /// Slicing provide a few gains like removing a shift or a
11469 /// truncate. This method allows to grow the cost of the original
11470 /// load with the gain from this slice.
11471 void addSliceGain(const LoadedSlice &LS) {
11472 // Each slice saves a truncate.
11473 const TargetLowering &TLI = LS.DAG->getTargetLoweringInfo();
11474 if (!TLI.isTruncateFree(LS.Inst->getOperand(0).getValueType(),
11475 LS.Inst->getValueType(0)))
11477 // If there is a shift amount, this slice gets rid of it.
11480 // If this slice can merge a cross register bank copy, account for it.
11481 if (LS.canMergeExpensiveCrossRegisterBankCopy())
11482 ++CrossRegisterBanksCopies;
11485 Cost &operator+=(const Cost &RHS) {
11486 Loads += RHS.Loads;
11487 Truncates += RHS.Truncates;
11488 CrossRegisterBanksCopies += RHS.CrossRegisterBanksCopies;
11489 ZExts += RHS.ZExts;
11490 Shift += RHS.Shift;
11494 bool operator==(const Cost &RHS) const {
11495 return Loads == RHS.Loads && Truncates == RHS.Truncates &&
11496 CrossRegisterBanksCopies == RHS.CrossRegisterBanksCopies &&
11497 ZExts == RHS.ZExts && Shift == RHS.Shift;
11500 bool operator!=(const Cost &RHS) const { return !(*this == RHS); }
11502 bool operator<(const Cost &RHS) const {
11503 // Assume cross register banks copies are as expensive as loads.
11504 // FIXME: Do we want some more target hooks?
11505 unsigned ExpensiveOpsLHS = Loads + CrossRegisterBanksCopies;
11506 unsigned ExpensiveOpsRHS = RHS.Loads + RHS.CrossRegisterBanksCopies;
11507 // Unless we are optimizing for code size, consider the
11508 // expensive operation first.
11509 if (!ForCodeSize && ExpensiveOpsLHS != ExpensiveOpsRHS)
11510 return ExpensiveOpsLHS < ExpensiveOpsRHS;
11511 return (Truncates + ZExts + Shift + ExpensiveOpsLHS) <
11512 (RHS.Truncates + RHS.ZExts + RHS.Shift + ExpensiveOpsRHS);
11515 bool operator>(const Cost &RHS) const { return RHS < *this; }
11517 bool operator<=(const Cost &RHS) const { return !(RHS < *this); }
11519 bool operator>=(const Cost &RHS) const { return !(*this < RHS); }
11521 // The last instruction that represent the slice. This should be a
11522 // truncate instruction.
11524 // The original load instruction.
11525 LoadSDNode *Origin;
11526 // The right shift amount in bits from the original load.
11528 // The DAG from which Origin came from.
11529 // This is used to get some contextual information about legal types, etc.
11532 LoadedSlice(SDNode *Inst = nullptr, LoadSDNode *Origin = nullptr,
11533 unsigned Shift = 0, SelectionDAG *DAG = nullptr)
11534 : Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {}
11536 /// \brief Get the bits used in a chunk of bits \p BitWidth large.
11537 /// \return Result is \p BitWidth and has used bits set to 1 and
11538 /// not used bits set to 0.
11539 APInt getUsedBits() const {
11540 // Reproduce the trunc(lshr) sequence:
11541 // - Start from the truncated value.
11542 // - Zero extend to the desired bit width.
11544 assert(Origin && "No original load to compare against.");
11545 unsigned BitWidth = Origin->getValueSizeInBits(0);
11546 assert(Inst && "This slice is not bound to an instruction");
11547 assert(Inst->getValueSizeInBits(0) <= BitWidth &&
11548 "Extracted slice is bigger than the whole type!");
11549 APInt UsedBits(Inst->getValueSizeInBits(0), 0);
11550 UsedBits.setAllBits();
11551 UsedBits = UsedBits.zext(BitWidth);
11552 UsedBits <<= Shift;
11556 /// \brief Get the size of the slice to be loaded in bytes.
11557 unsigned getLoadedSize() const {
11558 unsigned SliceSize = getUsedBits().countPopulation();
11559 assert(!(SliceSize & 0x7) && "Size is not a multiple of a byte.");
11560 return SliceSize / 8;
11563 /// \brief Get the type that will be loaded for this slice.
11564 /// Note: This may not be the final type for the slice.
11565 EVT getLoadedType() const {
11566 assert(DAG && "Missing context");
11567 LLVMContext &Ctxt = *DAG->getContext();
11568 return EVT::getIntegerVT(Ctxt, getLoadedSize() * 8);
11571 /// \brief Get the alignment of the load used for this slice.
11572 unsigned getAlignment() const {
11573 unsigned Alignment = Origin->getAlignment();
11574 unsigned Offset = getOffsetFromBase();
11576 Alignment = MinAlign(Alignment, Alignment + Offset);
11580 /// \brief Check if this slice can be rewritten with legal operations.
11581 bool isLegal() const {
11582 // An invalid slice is not legal.
11583 if (!Origin || !Inst || !DAG)
11586 // Offsets are for indexed load only, we do not handle that.
11587 if (!Origin->getOffset().isUndef())
11590 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
11592 // Check that the type is legal.
11593 EVT SliceType = getLoadedType();
11594 if (!TLI.isTypeLegal(SliceType))
11597 // Check that the load is legal for this type.
11598 if (!TLI.isOperationLegal(ISD::LOAD, SliceType))
11601 // Check that the offset can be computed.
11602 // 1. Check its type.
11603 EVT PtrType = Origin->getBasePtr().getValueType();
11604 if (PtrType == MVT::Untyped || PtrType.isExtended())
11607 // 2. Check that it fits in the immediate.
11608 if (!TLI.isLegalAddImmediate(getOffsetFromBase()))
11611 // 3. Check that the computation is legal.
11612 if (!TLI.isOperationLegal(ISD::ADD, PtrType))
11615 // Check that the zext is legal if it needs one.
11616 EVT TruncateType = Inst->getValueType(0);
11617 if (TruncateType != SliceType &&
11618 !TLI.isOperationLegal(ISD::ZERO_EXTEND, TruncateType))
11624 /// \brief Get the offset in bytes of this slice in the original chunk of
11626 /// \pre DAG != nullptr.
11627 uint64_t getOffsetFromBase() const {
11628 assert(DAG && "Missing context.");
11629 bool IsBigEndian = DAG->getDataLayout().isBigEndian();
11630 assert(!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported.");
11631 uint64_t Offset = Shift / 8;
11632 unsigned TySizeInBytes = Origin->getValueSizeInBits(0) / 8;
11633 assert(!(Origin->getValueSizeInBits(0) & 0x7) &&
11634 "The size of the original loaded type is not a multiple of a"
11636 // If Offset is bigger than TySizeInBytes, it means we are loading all
11637 // zeros. This should have been optimized before in the process.
11638 assert(TySizeInBytes > Offset &&
11639 "Invalid shift amount for given loaded size");
11641 Offset = TySizeInBytes - Offset - getLoadedSize();
11645 /// \brief Generate the sequence of instructions to load the slice
11646 /// represented by this object and redirect the uses of this slice to
11647 /// this new sequence of instructions.
11648 /// \pre this->Inst && this->Origin are valid Instructions and this
11649 /// object passed the legal check: LoadedSlice::isLegal returned true.
11650 /// \return The last instruction of the sequence used to load the slice.
11651 SDValue loadSlice() const {
11652 assert(Inst && Origin && "Unable to replace a non-existing slice.");
11653 const SDValue &OldBaseAddr = Origin->getBasePtr();
11654 SDValue BaseAddr = OldBaseAddr;
11655 // Get the offset in that chunk of bytes w.r.t. the endianness.
11656 int64_t Offset = static_cast<int64_t>(getOffsetFromBase());
11657 assert(Offset >= 0 && "Offset too big to fit in int64_t!");
11659 // BaseAddr = BaseAddr + Offset.
11660 EVT ArithType = BaseAddr.getValueType();
11662 BaseAddr = DAG->getNode(ISD::ADD, DL, ArithType, BaseAddr,
11663 DAG->getConstant(Offset, DL, ArithType));
11666 // Create the type of the loaded slice according to its size.
11667 EVT SliceType = getLoadedType();
11669 // Create the load for the slice.
11671 DAG->getLoad(SliceType, SDLoc(Origin), Origin->getChain(), BaseAddr,
11672 Origin->getPointerInfo().getWithOffset(Offset),
11673 getAlignment(), Origin->getMemOperand()->getFlags());
11674 // If the final type is not the same as the loaded type, this means that
11675 // we have to pad with zero. Create a zero extend for that.
11676 EVT FinalType = Inst->getValueType(0);
11677 if (SliceType != FinalType)
11679 DAG->getNode(ISD::ZERO_EXTEND, SDLoc(LastInst), FinalType, LastInst);
11683 /// \brief Check if this slice can be merged with an expensive cross register
11684 /// bank copy. E.g.,
11686 /// f = bitcast i32 i to float
11687 bool canMergeExpensiveCrossRegisterBankCopy() const {
11688 if (!Inst || !Inst->hasOneUse())
11690 SDNode *Use = *Inst->use_begin();
11691 if (Use->getOpcode() != ISD::BITCAST)
11693 assert(DAG && "Missing context");
11694 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
11695 EVT ResVT = Use->getValueType(0);
11696 const TargetRegisterClass *ResRC = TLI.getRegClassFor(ResVT.getSimpleVT());
11697 const TargetRegisterClass *ArgRC =
11698 TLI.getRegClassFor(Use->getOperand(0).getValueType().getSimpleVT());
11699 if (ArgRC == ResRC || !TLI.isOperationLegal(ISD::LOAD, ResVT))
11702 // At this point, we know that we perform a cross-register-bank copy.
11703 // Check if it is expensive.
11704 const TargetRegisterInfo *TRI = DAG->getSubtarget().getRegisterInfo();
11705 // Assume bitcasts are cheap, unless both register classes do not
11706 // explicitly share a common sub class.
11707 if (!TRI || TRI->getCommonSubClass(ArgRC, ResRC))
11710 // Check if it will be merged with the load.
11711 // 1. Check the alignment constraint.
11712 unsigned RequiredAlignment = DAG->getDataLayout().getABITypeAlignment(
11713 ResVT.getTypeForEVT(*DAG->getContext()));
11715 if (RequiredAlignment > getAlignment())
11718 // 2. Check that the load is a legal operation for that type.
11719 if (!TLI.isOperationLegal(ISD::LOAD, ResVT))
11722 // 3. Check that we do not have a zext in the way.
11723 if (Inst->getValueType(0) != getLoadedType())
11731 /// \brief Check that all bits set in \p UsedBits form a dense region, i.e.,
11732 /// \p UsedBits looks like 0..0 1..1 0..0.
11733 static bool areUsedBitsDense(const APInt &UsedBits) {
11734 // If all the bits are one, this is dense!
11735 if (UsedBits.isAllOnesValue())
11738 // Get rid of the unused bits on the right.
11739 APInt NarrowedUsedBits = UsedBits.lshr(UsedBits.countTrailingZeros());
11740 // Get rid of the unused bits on the left.
11741 if (NarrowedUsedBits.countLeadingZeros())
11742 NarrowedUsedBits = NarrowedUsedBits.trunc(NarrowedUsedBits.getActiveBits());
11743 // Check that the chunk of bits is completely used.
11744 return NarrowedUsedBits.isAllOnesValue();
11747 /// \brief Check whether or not \p First and \p Second are next to each other
11748 /// in memory. This means that there is no hole between the bits loaded
11749 /// by \p First and the bits loaded by \p Second.
11750 static bool areSlicesNextToEachOther(const LoadedSlice &First,
11751 const LoadedSlice &Second) {
11752 assert(First.Origin == Second.Origin && First.Origin &&
11753 "Unable to match different memory origins.");
11754 APInt UsedBits = First.getUsedBits();
11755 assert((UsedBits & Second.getUsedBits()) == 0 &&
11756 "Slices are not supposed to overlap.");
11757 UsedBits |= Second.getUsedBits();
11758 return areUsedBitsDense(UsedBits);
11761 /// \brief Adjust the \p GlobalLSCost according to the target
11762 /// paring capabilities and the layout of the slices.
11763 /// \pre \p GlobalLSCost should account for at least as many loads as
11764 /// there is in the slices in \p LoadedSlices.
11765 static void adjustCostForPairing(SmallVectorImpl<LoadedSlice> &LoadedSlices,
11766 LoadedSlice::Cost &GlobalLSCost) {
11767 unsigned NumberOfSlices = LoadedSlices.size();
11768 // If there is less than 2 elements, no pairing is possible.
11769 if (NumberOfSlices < 2)
11772 // Sort the slices so that elements that are likely to be next to each
11773 // other in memory are next to each other in the list.
11774 std::sort(LoadedSlices.begin(), LoadedSlices.end(),
11775 [](const LoadedSlice &LHS, const LoadedSlice &RHS) {
11776 assert(LHS.Origin == RHS.Origin && "Different bases not implemented.");
11777 return LHS.getOffsetFromBase() < RHS.getOffsetFromBase();
11779 const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo();
11780 // First (resp. Second) is the first (resp. Second) potentially candidate
11781 // to be placed in a paired load.
11782 const LoadedSlice *First = nullptr;
11783 const LoadedSlice *Second = nullptr;
11784 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice,
11785 // Set the beginning of the pair.
11788 Second = &LoadedSlices[CurrSlice];
11790 // If First is NULL, it means we start a new pair.
11791 // Get to the next slice.
11795 EVT LoadedType = First->getLoadedType();
11797 // If the types of the slices are different, we cannot pair them.
11798 if (LoadedType != Second->getLoadedType())
11801 // Check if the target supplies paired loads for this type.
11802 unsigned RequiredAlignment = 0;
11803 if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) {
11804 // move to the next pair, this type is hopeless.
11808 // Check if we meet the alignment requirement.
11809 if (RequiredAlignment > First->getAlignment())
11812 // Check that both loads are next to each other in memory.
11813 if (!areSlicesNextToEachOther(*First, *Second))
11816 assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!");
11817 --GlobalLSCost.Loads;
11818 // Move to the next pair.
11823 /// \brief Check the profitability of all involved LoadedSlice.
11824 /// Currently, it is considered profitable if there is exactly two
11825 /// involved slices (1) which are (2) next to each other in memory, and
11826 /// whose cost (\see LoadedSlice::Cost) is smaller than the original load (3).
11828 /// Note: The order of the elements in \p LoadedSlices may be modified, but not
11829 /// the elements themselves.
11831 /// FIXME: When the cost model will be mature enough, we can relax
11832 /// constraints (1) and (2).
11833 static bool isSlicingProfitable(SmallVectorImpl<LoadedSlice> &LoadedSlices,
11834 const APInt &UsedBits, bool ForCodeSize) {
11835 unsigned NumberOfSlices = LoadedSlices.size();
11836 if (StressLoadSlicing)
11837 return NumberOfSlices > 1;
11840 if (NumberOfSlices != 2)
11844 if (!areUsedBitsDense(UsedBits))
11848 LoadedSlice::Cost OrigCost(ForCodeSize), GlobalSlicingCost(ForCodeSize);
11849 // The original code has one big load.
11850 OrigCost.Loads = 1;
11851 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice) {
11852 const LoadedSlice &LS = LoadedSlices[CurrSlice];
11853 // Accumulate the cost of all the slices.
11854 LoadedSlice::Cost SliceCost(LS, ForCodeSize);
11855 GlobalSlicingCost += SliceCost;
11857 // Account as cost in the original configuration the gain obtained
11858 // with the current slices.
11859 OrigCost.addSliceGain(LS);
11862 // If the target supports paired load, adjust the cost accordingly.
11863 adjustCostForPairing(LoadedSlices, GlobalSlicingCost);
11864 return OrigCost > GlobalSlicingCost;
11867 /// \brief If the given load, \p LI, is used only by trunc or trunc(lshr)
11868 /// operations, split it in the various pieces being extracted.
11870 /// This sort of thing is introduced by SROA.
11871 /// This slicing takes care not to insert overlapping loads.
11872 /// \pre LI is a simple load (i.e., not an atomic or volatile load).
11873 bool DAGCombiner::SliceUpLoad(SDNode *N) {
11874 if (Level < AfterLegalizeDAG)
11877 LoadSDNode *LD = cast<LoadSDNode>(N);
11878 if (LD->isVolatile() || !ISD::isNormalLoad(LD) ||
11879 !LD->getValueType(0).isInteger())
11882 // Keep track of already used bits to detect overlapping values.
11883 // In that case, we will just abort the transformation.
11884 APInt UsedBits(LD->getValueSizeInBits(0), 0);
11886 SmallVector<LoadedSlice, 4> LoadedSlices;
11888 // Check if this load is used as several smaller chunks of bits.
11889 // Basically, look for uses in trunc or trunc(lshr) and record a new chain
11890 // of computation for each trunc.
11891 for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end();
11892 UI != UIEnd; ++UI) {
11893 // Skip the uses of the chain.
11894 if (UI.getUse().getResNo() != 0)
11897 SDNode *User = *UI;
11898 unsigned Shift = 0;
11900 // Check if this is a trunc(lshr).
11901 if (User->getOpcode() == ISD::SRL && User->hasOneUse() &&
11902 isa<ConstantSDNode>(User->getOperand(1))) {
11903 Shift = User->getConstantOperandVal(1);
11904 User = *User->use_begin();
11907 // At this point, User is a Truncate, iff we encountered, trunc or
11909 if (User->getOpcode() != ISD::TRUNCATE)
11912 // The width of the type must be a power of 2 and greater than 8-bits.
11913 // Otherwise the load cannot be represented in LLVM IR.
11914 // Moreover, if we shifted with a non-8-bits multiple, the slice
11915 // will be across several bytes. We do not support that.
11916 unsigned Width = User->getValueSizeInBits(0);
11917 if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7))
11920 // Build the slice for this chain of computations.
11921 LoadedSlice LS(User, LD, Shift, &DAG);
11922 APInt CurrentUsedBits = LS.getUsedBits();
11924 // Check if this slice overlaps with another.
11925 if ((CurrentUsedBits & UsedBits) != 0)
11927 // Update the bits used globally.
11928 UsedBits |= CurrentUsedBits;
11930 // Check if the new slice would be legal.
11934 // Record the slice.
11935 LoadedSlices.push_back(LS);
11938 // Abort slicing if it does not seem to be profitable.
11939 if (!isSlicingProfitable(LoadedSlices, UsedBits, ForCodeSize))
11944 // Rewrite each chain to use an independent load.
11945 // By construction, each chain can be represented by a unique load.
11947 // Prepare the argument for the new token factor for all the slices.
11948 SmallVector<SDValue, 8> ArgChains;
11949 for (SmallVectorImpl<LoadedSlice>::const_iterator
11950 LSIt = LoadedSlices.begin(),
11951 LSItEnd = LoadedSlices.end();
11952 LSIt != LSItEnd; ++LSIt) {
11953 SDValue SliceInst = LSIt->loadSlice();
11954 CombineTo(LSIt->Inst, SliceInst, true);
11955 if (SliceInst.getOpcode() != ISD::LOAD)
11956 SliceInst = SliceInst.getOperand(0);
11957 assert(SliceInst->getOpcode() == ISD::LOAD &&
11958 "It takes more than a zext to get to the loaded slice!!");
11959 ArgChains.push_back(SliceInst.getValue(1));
11962 SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other,
11964 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
11965 AddToWorklist(Chain.getNode());
11969 /// Check to see if V is (and load (ptr), imm), where the load is having
11970 /// specific bytes cleared out. If so, return the byte size being masked out
11971 /// and the shift amount.
11972 static std::pair<unsigned, unsigned>
11973 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
11974 std::pair<unsigned, unsigned> Result(0, 0);
11976 // Check for the structure we're looking for.
11977 if (V->getOpcode() != ISD::AND ||
11978 !isa<ConstantSDNode>(V->getOperand(1)) ||
11979 !ISD::isNormalLoad(V->getOperand(0).getNode()))
11982 // Check the chain and pointer.
11983 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
11984 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
11986 // The store should be chained directly to the load or be an operand of a
11988 if (LD == Chain.getNode())
11990 else if (Chain->getOpcode() != ISD::TokenFactor)
11991 return Result; // Fail.
11994 for (const SDValue &ChainOp : Chain->op_values())
11995 if (ChainOp.getNode() == LD) {
11999 if (!isOk) return Result;
12002 // This only handles simple types.
12003 if (V.getValueType() != MVT::i16 &&
12004 V.getValueType() != MVT::i32 &&
12005 V.getValueType() != MVT::i64)
12008 // Check the constant mask. Invert it so that the bits being masked out are
12009 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
12010 // follow the sign bit for uniformity.
12011 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
12012 unsigned NotMaskLZ = countLeadingZeros(NotMask);
12013 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
12014 unsigned NotMaskTZ = countTrailingZeros(NotMask);
12015 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
12016 if (NotMaskLZ == 64) return Result; // All zero mask.
12018 // See if we have a continuous run of bits. If so, we have 0*1+0*
12019 if (countTrailingOnes(NotMask >> NotMaskTZ) + NotMaskTZ + NotMaskLZ != 64)
12022 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
12023 if (V.getValueType() != MVT::i64 && NotMaskLZ)
12024 NotMaskLZ -= 64-V.getValueSizeInBits();
12026 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
12027 switch (MaskedBytes) {
12031 default: return Result; // All one mask, or 5-byte mask.
12034 // Verify that the first bit starts at a multiple of mask so that the access
12035 // is aligned the same as the access width.
12036 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
12038 Result.first = MaskedBytes;
12039 Result.second = NotMaskTZ/8;
12044 /// Check to see if IVal is something that provides a value as specified by
12045 /// MaskInfo. If so, replace the specified store with a narrower store of
12046 /// truncated IVal.
12048 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
12049 SDValue IVal, StoreSDNode *St,
12051 unsigned NumBytes = MaskInfo.first;
12052 unsigned ByteShift = MaskInfo.second;
12053 SelectionDAG &DAG = DC->getDAG();
12055 // Check to see if IVal is all zeros in the part being masked in by the 'or'
12056 // that uses this. If not, this is not a replacement.
12057 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
12058 ByteShift*8, (ByteShift+NumBytes)*8);
12059 if (!DAG.MaskedValueIsZero(IVal, Mask)) return nullptr;
12061 // Check that it is legal on the target to do this. It is legal if the new
12062 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
12064 MVT VT = MVT::getIntegerVT(NumBytes*8);
12065 if (!DC->isTypeLegal(VT))
12068 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
12069 // shifted by ByteShift and truncated down to NumBytes.
12072 IVal = DAG.getNode(ISD::SRL, DL, IVal.getValueType(), IVal,
12073 DAG.getConstant(ByteShift*8, DL,
12074 DC->getShiftAmountTy(IVal.getValueType())));
12077 // Figure out the offset for the store and the alignment of the access.
12079 unsigned NewAlign = St->getAlignment();
12081 if (DAG.getDataLayout().isLittleEndian())
12082 StOffset = ByteShift;
12084 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
12086 SDValue Ptr = St->getBasePtr();
12089 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(),
12090 Ptr, DAG.getConstant(StOffset, DL, Ptr.getValueType()));
12091 NewAlign = MinAlign(NewAlign, StOffset);
12094 // Truncate down to the new size.
12095 IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal);
12099 .getStore(St->getChain(), SDLoc(St), IVal, Ptr,
12100 St->getPointerInfo().getWithOffset(StOffset), NewAlign)
12105 /// Look for sequence of load / op / store where op is one of 'or', 'xor', and
12106 /// 'and' of immediates. If 'op' is only touching some of the loaded bits, try
12107 /// narrowing the load and store if it would end up being a win for performance
12109 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
12110 StoreSDNode *ST = cast<StoreSDNode>(N);
12111 if (ST->isVolatile())
12114 SDValue Chain = ST->getChain();
12115 SDValue Value = ST->getValue();
12116 SDValue Ptr = ST->getBasePtr();
12117 EVT VT = Value.getValueType();
12119 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
12122 unsigned Opc = Value.getOpcode();
12124 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
12125 // is a byte mask indicating a consecutive number of bytes, check to see if
12126 // Y is known to provide just those bytes. If so, we try to replace the
12127 // load + replace + store sequence with a single (narrower) store, which makes
12129 if (Opc == ISD::OR) {
12130 std::pair<unsigned, unsigned> MaskedLoad;
12131 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
12132 if (MaskedLoad.first)
12133 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
12134 Value.getOperand(1), ST,this))
12135 return SDValue(NewST, 0);
12137 // Or is commutative, so try swapping X and Y.
12138 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
12139 if (MaskedLoad.first)
12140 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
12141 Value.getOperand(0), ST,this))
12142 return SDValue(NewST, 0);
12145 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
12146 Value.getOperand(1).getOpcode() != ISD::Constant)
12149 SDValue N0 = Value.getOperand(0);
12150 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
12151 Chain == SDValue(N0.getNode(), 1)) {
12152 LoadSDNode *LD = cast<LoadSDNode>(N0);
12153 if (LD->getBasePtr() != Ptr ||
12154 LD->getPointerInfo().getAddrSpace() !=
12155 ST->getPointerInfo().getAddrSpace())
12158 // Find the type to narrow it the load / op / store to.
12159 SDValue N1 = Value.getOperand(1);
12160 unsigned BitWidth = N1.getValueSizeInBits();
12161 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
12162 if (Opc == ISD::AND)
12163 Imm ^= APInt::getAllOnesValue(BitWidth);
12164 if (Imm == 0 || Imm.isAllOnesValue())
12166 unsigned ShAmt = Imm.countTrailingZeros();
12167 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
12168 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
12169 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
12170 // The narrowing should be profitable, the load/store operation should be
12171 // legal (or custom) and the store size should be equal to the NewVT width.
12172 while (NewBW < BitWidth &&
12173 (NewVT.getStoreSizeInBits() != NewBW ||
12174 !TLI.isOperationLegalOrCustom(Opc, NewVT) ||
12175 !TLI.isNarrowingProfitable(VT, NewVT))) {
12176 NewBW = NextPowerOf2(NewBW);
12177 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
12179 if (NewBW >= BitWidth)
12182 // If the lsb changed does not start at the type bitwidth boundary,
12183 // start at the previous one.
12185 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
12186 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt,
12187 std::min(BitWidth, ShAmt + NewBW));
12188 if ((Imm & Mask) == Imm) {
12189 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
12190 if (Opc == ISD::AND)
12191 NewImm ^= APInt::getAllOnesValue(NewBW);
12192 uint64_t PtrOff = ShAmt / 8;
12193 // For big endian targets, we need to adjust the offset to the pointer to
12194 // load the correct bytes.
12195 if (DAG.getDataLayout().isBigEndian())
12196 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
12198 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
12199 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
12200 if (NewAlign < DAG.getDataLayout().getABITypeAlignment(NewVTTy))
12203 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD),
12204 Ptr.getValueType(), Ptr,
12205 DAG.getConstant(PtrOff, SDLoc(LD),
12206 Ptr.getValueType()));
12208 DAG.getLoad(NewVT, SDLoc(N0), LD->getChain(), NewPtr,
12209 LD->getPointerInfo().getWithOffset(PtrOff), NewAlign,
12210 LD->getMemOperand()->getFlags(), LD->getAAInfo());
12211 SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD,
12212 DAG.getConstant(NewImm, SDLoc(Value),
12215 DAG.getStore(Chain, SDLoc(N), NewVal, NewPtr,
12216 ST->getPointerInfo().getWithOffset(PtrOff), NewAlign);
12218 AddToWorklist(NewPtr.getNode());
12219 AddToWorklist(NewLD.getNode());
12220 AddToWorklist(NewVal.getNode());
12221 WorklistRemover DeadNodes(*this);
12222 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
12231 /// For a given floating point load / store pair, if the load value isn't used
12232 /// by any other operations, then consider transforming the pair to integer
12233 /// load / store operations if the target deems the transformation profitable.
12234 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
12235 StoreSDNode *ST = cast<StoreSDNode>(N);
12236 SDValue Chain = ST->getChain();
12237 SDValue Value = ST->getValue();
12238 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
12239 Value.hasOneUse() &&
12240 Chain == SDValue(Value.getNode(), 1)) {
12241 LoadSDNode *LD = cast<LoadSDNode>(Value);
12242 EVT VT = LD->getMemoryVT();
12243 if (!VT.isFloatingPoint() ||
12244 VT != ST->getMemoryVT() ||
12245 LD->isNonTemporal() ||
12246 ST->isNonTemporal() ||
12247 LD->getPointerInfo().getAddrSpace() != 0 ||
12248 ST->getPointerInfo().getAddrSpace() != 0)
12251 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
12252 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
12253 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
12254 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
12255 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
12258 unsigned LDAlign = LD->getAlignment();
12259 unsigned STAlign = ST->getAlignment();
12260 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
12261 unsigned ABIAlign = DAG.getDataLayout().getABITypeAlignment(IntVTTy);
12262 if (LDAlign < ABIAlign || STAlign < ABIAlign)
12266 DAG.getLoad(IntVT, SDLoc(Value), LD->getChain(), LD->getBasePtr(),
12267 LD->getPointerInfo(), LDAlign);
12270 DAG.getStore(NewLD.getValue(1), SDLoc(N), NewLD, ST->getBasePtr(),
12271 ST->getPointerInfo(), STAlign);
12273 AddToWorklist(NewLD.getNode());
12274 AddToWorklist(NewST.getNode());
12275 WorklistRemover DeadNodes(*this);
12276 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
12284 // This is a helper function for visitMUL to check the profitability
12285 // of folding (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2).
12286 // MulNode is the original multiply, AddNode is (add x, c1),
12287 // and ConstNode is c2.
12289 // If the (add x, c1) has multiple uses, we could increase
12290 // the number of adds if we make this transformation.
12291 // It would only be worth doing this if we can remove a
12292 // multiply in the process. Check for that here.
12296 // We're checking for cases where we have common "c3 * A" expressions.
12297 bool DAGCombiner::isMulAddWithConstProfitable(SDNode *MulNode,
12299 SDValue &ConstNode) {
12302 // If the add only has one use, this would be OK to do.
12303 if (AddNode.getNode()->hasOneUse())
12306 // Walk all the users of the constant with which we're multiplying.
12307 for (SDNode *Use : ConstNode->uses()) {
12309 if (Use == MulNode) // This use is the one we're on right now. Skip it.
12312 if (Use->getOpcode() == ISD::MUL) { // We have another multiply use.
12314 SDNode *MulVar = AddNode.getOperand(0).getNode();
12316 // OtherOp is what we're multiplying against the constant.
12317 if (Use->getOperand(0) == ConstNode)
12318 OtherOp = Use->getOperand(1).getNode();
12320 OtherOp = Use->getOperand(0).getNode();
12322 // Check to see if multiply is with the same operand of our "add".
12324 // ConstNode = CONST
12325 // Use = ConstNode * A <-- visiting Use. OtherOp is A.
12327 // AddNode = (A + c1) <-- MulVar is A.
12328 // = AddNode * ConstNode <-- current visiting instruction.
12330 // If we make this transformation, we will have a common
12331 // multiply (ConstNode * A) that we can save.
12332 if (OtherOp == MulVar)
12335 // Now check to see if a future expansion will give us a common
12338 // ConstNode = CONST
12339 // AddNode = (A + c1)
12340 // ... = AddNode * ConstNode <-- current visiting instruction.
12342 // OtherOp = (A + c2)
12343 // Use = OtherOp * ConstNode <-- visiting Use.
12345 // If we make this transformation, we will have a common
12346 // multiply (CONST * A) after we also do the same transformation
12347 // to the "t2" instruction.
12348 if (OtherOp->getOpcode() == ISD::ADD &&
12349 DAG.isConstantIntBuildVectorOrConstantInt(OtherOp->getOperand(1)) &&
12350 OtherOp->getOperand(0).getNode() == MulVar)
12355 // Didn't find a case where this would be profitable.
12359 SDValue DAGCombiner::getMergeStoreChains(SmallVectorImpl<MemOpLink> &StoreNodes,
12360 unsigned NumStores) {
12361 SmallVector<SDValue, 8> Chains;
12362 SmallPtrSet<const SDNode *, 8> Visited;
12363 SDLoc StoreDL(StoreNodes[0].MemNode);
12365 for (unsigned i = 0; i < NumStores; ++i) {
12366 Visited.insert(StoreNodes[i].MemNode);
12369 // don't include nodes that are children
12370 for (unsigned i = 0; i < NumStores; ++i) {
12371 if (Visited.count(StoreNodes[i].MemNode->getChain().getNode()) == 0)
12372 Chains.push_back(StoreNodes[i].MemNode->getChain());
12375 assert(Chains.size() > 0 && "Chain should have generated a chain");
12376 return DAG.getNode(ISD::TokenFactor, StoreDL, MVT::Other, Chains);
12379 bool DAGCombiner::MergeStoresOfConstantsOrVecElts(
12380 SmallVectorImpl<MemOpLink> &StoreNodes, EVT MemVT,
12381 unsigned NumStores, bool IsConstantSrc, bool UseVector) {
12382 // Make sure we have something to merge.
12386 int64_t ElementSizeBytes = MemVT.getSizeInBits() / 8;
12388 // The latest Node in the DAG.
12389 SDLoc DL(StoreNodes[0].MemNode);
12393 bool IsVec = MemVT.isVector();
12394 unsigned Elts = NumStores;
12396 // When merging vector stores, get the total number of elements.
12397 Elts *= MemVT.getVectorNumElements();
12399 // Get the type for the merged vector store.
12400 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(), Elts);
12401 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
12403 if (IsConstantSrc) {
12404 SmallVector<SDValue, 8> BuildVector;
12405 for (unsigned I = 0, E = Ty.getVectorNumElements(); I != E; ++I) {
12406 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[I].MemNode);
12407 SDValue Val = St->getValue();
12408 if (MemVT.getScalarType().isInteger())
12409 if (auto *CFP = dyn_cast<ConstantFPSDNode>(St->getValue()))
12410 Val = DAG.getConstant(
12411 (uint32_t)CFP->getValueAPF().bitcastToAPInt().getZExtValue(),
12412 SDLoc(CFP), MemVT);
12413 BuildVector.push_back(Val);
12415 StoredVal = DAG.getBuildVector(Ty, DL, BuildVector);
12417 SmallVector<SDValue, 8> Ops;
12418 for (unsigned i = 0; i < NumStores; ++i) {
12419 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
12420 SDValue Val = St->getValue();
12421 // All operands of BUILD_VECTOR / CONCAT_VECTOR must have the same type.
12422 if (Val.getValueType() != MemVT)
12424 Ops.push_back(Val);
12427 // Build the extracted vector elements back into a vector.
12428 StoredVal = DAG.getNode(IsVec ? ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR,
12431 // We should always use a vector store when merging extracted vector
12432 // elements, so this path implies a store of constants.
12433 assert(IsConstantSrc && "Merged vector elements should use vector store");
12435 unsigned SizeInBits = NumStores * ElementSizeBytes * 8;
12436 APInt StoreInt(SizeInBits, 0);
12438 // Construct a single integer constant which is made of the smaller
12439 // constant inputs.
12440 bool IsLE = DAG.getDataLayout().isLittleEndian();
12441 for (unsigned i = 0; i < NumStores; ++i) {
12442 unsigned Idx = IsLE ? (NumStores - 1 - i) : i;
12443 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
12445 SDValue Val = St->getValue();
12446 StoreInt <<= ElementSizeBytes * 8;
12447 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
12448 StoreInt |= C->getAPIntValue().zextOrTrunc(SizeInBits);
12449 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
12450 StoreInt |= C->getValueAPF().bitcastToAPInt().zextOrTrunc(SizeInBits);
12452 llvm_unreachable("Invalid constant element type");
12456 // Create the new Load and Store operations.
12457 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), SizeInBits);
12458 StoredVal = DAG.getConstant(StoreInt, DL, StoreTy);
12461 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
12462 SDValue NewChain = getMergeStoreChains(StoreNodes, NumStores);
12463 SDValue NewStore = DAG.getStore(NewChain, DL, StoredVal,
12464 FirstInChain->getBasePtr(),
12465 FirstInChain->getPointerInfo(),
12466 FirstInChain->getAlignment());
12468 // Replace all merged stores with the new store.
12469 for (unsigned i = 0; i < NumStores; ++i)
12470 CombineTo(StoreNodes[i].MemNode, NewStore);
12472 AddToWorklist(NewChain.getNode());
12476 void DAGCombiner::getStoreMergeCandidates(
12477 StoreSDNode *St, SmallVectorImpl<MemOpLink> &StoreNodes) {
12478 // This holds the base pointer, index, and the offset in bytes from the base
12480 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr(), DAG);
12481 EVT MemVT = St->getMemoryVT();
12483 // We must have a base and an offset.
12484 if (!BasePtr.Base.getNode())
12487 // Do not handle stores to undef base pointers.
12488 if (BasePtr.Base.isUndef())
12491 bool IsLoadSrc = isa<LoadSDNode>(St->getValue());
12492 bool IsConstantSrc = isa<ConstantSDNode>(St->getValue()) ||
12493 isa<ConstantFPSDNode>(St->getValue());
12494 bool IsExtractVecSrc =
12495 (St->getValue().getOpcode() == ISD::EXTRACT_VECTOR_ELT ||
12496 St->getValue().getOpcode() == ISD::EXTRACT_SUBVECTOR);
12497 auto CandidateMatch = [&](StoreSDNode *Other, BaseIndexOffset &Ptr) -> bool {
12498 if (Other->isVolatile() || Other->isIndexed())
12500 // We can merge constant floats to equivalent integers
12501 if (Other->getMemoryVT() != MemVT)
12502 if (!(MemVT.isInteger() && MemVT.bitsEq(Other->getMemoryVT()) &&
12503 isa<ConstantFPSDNode>(Other->getValue())))
12506 if (!isa<LoadSDNode>(Other->getValue()))
12509 if (!(isa<ConstantSDNode>(Other->getValue()) ||
12510 isa<ConstantFPSDNode>(Other->getValue())))
12512 if (IsExtractVecSrc)
12513 if (!(Other->getValue().getOpcode() == ISD::EXTRACT_VECTOR_ELT ||
12514 Other->getValue().getOpcode() == ISD::EXTRACT_SUBVECTOR))
12516 Ptr = BaseIndexOffset::match(Other->getBasePtr(), DAG);
12517 return (Ptr.equalBaseIndex(BasePtr));
12519 // We looking for a root node which is an ancestor to all mergable
12520 // stores. We search up through a load, to our root and then down
12521 // through all children. For instance we will find Store{1,2,3} if
12522 // St is Store1, Store2. or Store3 where the root is not a load
12523 // which always true for nonvolatile ops. TODO: Expand
12524 // the search to find all valid candidates through multiple layers of loads.
12527 // |-------|-------|
12528 // Load Load Store3
12532 // FIXME: We should be able to climb and
12533 // descend TokenFactors to find candidates as well.
12535 SDNode *RootNode = (St->getChain()).getNode();
12537 if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(RootNode)) {
12538 RootNode = Ldn->getChain().getNode();
12539 for (auto I = RootNode->use_begin(), E = RootNode->use_end(); I != E; ++I)
12540 if (I.getOperandNo() == 0 && isa<LoadSDNode>(*I)) // walk down chain
12541 for (auto I2 = (*I)->use_begin(), E2 = (*I)->use_end(); I2 != E2; ++I2)
12542 if (I2.getOperandNo() == 0)
12543 if (StoreSDNode *OtherST = dyn_cast<StoreSDNode>(*I2)) {
12544 BaseIndexOffset Ptr;
12545 if (CandidateMatch(OtherST, Ptr))
12546 StoreNodes.push_back(MemOpLink(OtherST, Ptr.Offset));
12549 for (auto I = RootNode->use_begin(), E = RootNode->use_end(); I != E; ++I)
12550 if (I.getOperandNo() == 0)
12551 if (StoreSDNode *OtherST = dyn_cast<StoreSDNode>(*I)) {
12552 BaseIndexOffset Ptr;
12553 if (CandidateMatch(OtherST, Ptr))
12554 StoreNodes.push_back(MemOpLink(OtherST, Ptr.Offset));
12558 // We need to check that merging these stores does not cause a loop
12559 // in the DAG. Any store candidate may depend on another candidate
12560 // indirectly through its operand (we already consider dependencies
12561 // through the chain). Check in parallel by searching up from
12562 // non-chain operands of candidates.
12563 bool DAGCombiner::checkMergeStoreCandidatesForDependencies(
12564 SmallVectorImpl<MemOpLink> &StoreNodes, unsigned NumStores) {
12565 SmallPtrSet<const SDNode *, 16> Visited;
12566 SmallVector<const SDNode *, 8> Worklist;
12567 // search ops of store candidates
12568 for (unsigned i = 0; i < NumStores; ++i) {
12569 SDNode *n = StoreNodes[i].MemNode;
12570 // Potential loops may happen only through non-chain operands
12571 for (unsigned j = 1; j < n->getNumOperands(); ++j)
12572 Worklist.push_back(n->getOperand(j).getNode());
12574 // search through DAG. We can stop early if we find a storenode
12575 for (unsigned i = 0; i < NumStores; ++i) {
12576 if (SDNode::hasPredecessorHelper(StoreNodes[i].MemNode, Visited, Worklist))
12582 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode *St) {
12583 if (OptLevel == CodeGenOpt::None)
12586 EVT MemVT = St->getMemoryVT();
12587 int64_t ElementSizeBytes = MemVT.getSizeInBits() / 8;
12589 if (MemVT.getSizeInBits() * 2 > MaximumLegalStoreInBits)
12592 bool NoVectors = DAG.getMachineFunction().getFunction()->hasFnAttribute(
12593 Attribute::NoImplicitFloat);
12595 // This function cannot currently deal with non-byte-sized memory sizes.
12596 if (ElementSizeBytes * 8 != MemVT.getSizeInBits())
12599 if (!MemVT.isSimple())
12602 // Perform an early exit check. Do not bother looking at stored values that
12603 // are not constants, loads, or extracted vector elements.
12604 SDValue StoredVal = St->getValue();
12605 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
12606 bool IsConstantSrc = isa<ConstantSDNode>(StoredVal) ||
12607 isa<ConstantFPSDNode>(StoredVal);
12608 bool IsExtractVecSrc = (StoredVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT ||
12609 StoredVal.getOpcode() == ISD::EXTRACT_SUBVECTOR);
12611 if (!IsConstantSrc && !IsLoadSrc && !IsExtractVecSrc)
12614 // Don't merge vectors into wider vectors if the source data comes from loads.
12615 // TODO: This restriction can be lifted by using logic similar to the
12616 // ExtractVecSrc case.
12617 if (MemVT.isVector() && IsLoadSrc)
12620 SmallVector<MemOpLink, 8> StoreNodes;
12621 // Find potential store merge candidates by searching through chain sub-DAG
12622 getStoreMergeCandidates(St, StoreNodes);
12624 // Check if there is anything to merge.
12625 if (StoreNodes.size() < 2)
12628 // Sort the memory operands according to their distance from the
12630 std::sort(StoreNodes.begin(), StoreNodes.end(),
12631 [](MemOpLink LHS, MemOpLink RHS) {
12632 return LHS.OffsetFromBase < RHS.OffsetFromBase;
12635 // Store Merge attempts to merge the lowest stores. This generally
12636 // works out as if successful, as the remaining stores are checked
12637 // after the first collection of stores is merged. However, in the
12638 // case that a non-mergeable store is found first, e.g., {p[-2],
12639 // p[0], p[1], p[2], p[3]}, we would fail and miss the subsequent
12640 // mergeable cases. To prevent this, we prune such stores from the
12641 // front of StoreNodes here.
12644 while (StoreNodes.size() > 1) {
12645 unsigned StartIdx = 0;
12646 while ((StartIdx + 1 < StoreNodes.size()) &&
12647 StoreNodes[StartIdx].OffsetFromBase + ElementSizeBytes !=
12648 StoreNodes[StartIdx + 1].OffsetFromBase)
12651 // Bail if we don't have enough candidates to merge.
12652 if (StartIdx + 1 >= StoreNodes.size())
12656 StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + StartIdx);
12658 // Scan the memory operations on the chain and find the first
12659 // non-consecutive store memory address.
12660 unsigned NumConsecutiveStores = 1;
12661 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
12662 // Check that the addresses are consecutive starting from the second
12663 // element in the list of stores.
12664 for (unsigned i = 1, e = StoreNodes.size(); i < e; ++i) {
12665 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
12666 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
12668 NumConsecutiveStores = i + 1;
12671 if (NumConsecutiveStores < 2) {
12672 StoreNodes.erase(StoreNodes.begin(),
12673 StoreNodes.begin() + NumConsecutiveStores);
12677 // Check that we can merge these candidates without causing a cycle
12678 if (!checkMergeStoreCandidatesForDependencies(StoreNodes,
12679 NumConsecutiveStores)) {
12680 StoreNodes.erase(StoreNodes.begin(),
12681 StoreNodes.begin() + NumConsecutiveStores);
12685 // The node with the lowest store address.
12686 LLVMContext &Context = *DAG.getContext();
12687 const DataLayout &DL = DAG.getDataLayout();
12689 // Store the constants into memory as one consecutive store.
12690 if (IsConstantSrc) {
12691 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
12692 unsigned FirstStoreAS = FirstInChain->getAddressSpace();
12693 unsigned FirstStoreAlign = FirstInChain->getAlignment();
12694 unsigned LastLegalType = 0;
12695 unsigned LastLegalVectorType = 0;
12696 bool NonZero = false;
12697 for (unsigned i = 0; i < NumConsecutiveStores; ++i) {
12698 StoreSDNode *ST = cast<StoreSDNode>(StoreNodes[i].MemNode);
12699 SDValue StoredVal = ST->getValue();
12701 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
12702 NonZero |= !C->isNullValue();
12703 } else if (ConstantFPSDNode *C =
12704 dyn_cast<ConstantFPSDNode>(StoredVal)) {
12705 NonZero |= !C->getConstantFPValue()->isNullValue();
12711 // Find a legal type for the constant store.
12712 unsigned SizeInBits = (i + 1) * ElementSizeBytes * 8;
12713 EVT StoreTy = EVT::getIntegerVT(Context, SizeInBits);
12714 bool IsFast = false;
12715 if (TLI.isTypeLegal(StoreTy) &&
12716 TLI.canMergeStoresTo(FirstStoreAS, StoreTy) &&
12717 TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstStoreAS,
12718 FirstStoreAlign, &IsFast) &&
12720 LastLegalType = i + 1;
12721 // Or check whether a truncstore is legal.
12722 } else if (!LegalTypes &&
12723 TLI.getTypeAction(Context, StoreTy) ==
12724 TargetLowering::TypePromoteInteger) {
12725 EVT LegalizedStoredValueTy =
12726 TLI.getTypeToTransformTo(Context, StoredVal.getValueType());
12727 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
12728 TLI.canMergeStoresTo(FirstStoreAS, LegalizedStoredValueTy) &&
12729 TLI.allowsMemoryAccess(Context, DL, LegalizedStoredValueTy,
12730 FirstStoreAS, FirstStoreAlign, &IsFast) &&
12732 LastLegalType = i + 1;
12736 // We only use vectors if the constant is known to be zero or the target
12737 // allows it and the function is not marked with the noimplicitfloat
12740 TLI.storeOfVectorConstantIsCheap(MemVT, i + 1, FirstStoreAS)) &&
12742 // Find a legal type for the vector store.
12743 EVT Ty = EVT::getVectorVT(Context, MemVT, i + 1);
12744 if (TLI.isTypeLegal(Ty) && TLI.canMergeStoresTo(FirstStoreAS, Ty) &&
12745 TLI.allowsMemoryAccess(Context, DL, Ty, FirstStoreAS,
12746 FirstStoreAlign, &IsFast) &&
12748 LastLegalVectorType = i + 1;
12752 // Check if we found a legal integer type that creates a meaningful merge.
12753 if (LastLegalType < 2 && LastLegalVectorType < 2) {
12754 StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + 1);
12758 bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors;
12759 unsigned NumElem = (UseVector) ? LastLegalVectorType : LastLegalType;
12761 bool Merged = MergeStoresOfConstantsOrVecElts(StoreNodes, MemVT, NumElem,
12764 StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + NumElem);
12767 // Remove merged stores for next iteration.
12769 StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + NumElem);
12773 // When extracting multiple vector elements, try to store them
12774 // in one vector store rather than a sequence of scalar stores.
12775 if (IsExtractVecSrc) {
12776 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
12777 unsigned FirstStoreAS = FirstInChain->getAddressSpace();
12778 unsigned FirstStoreAlign = FirstInChain->getAlignment();
12779 unsigned NumStoresToMerge = 1;
12780 bool IsVec = MemVT.isVector();
12781 for (unsigned i = 0; i < NumConsecutiveStores; ++i) {
12782 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
12783 unsigned StoreValOpcode = St->getValue().getOpcode();
12784 // This restriction could be loosened.
12785 // Bail out if any stored values are not elements extracted from a
12786 // vector. It should be possible to handle mixed sources, but load
12787 // sources need more careful handling (see the block of code below that
12788 // handles consecutive loads).
12789 if (StoreValOpcode != ISD::EXTRACT_VECTOR_ELT &&
12790 StoreValOpcode != ISD::EXTRACT_SUBVECTOR)
12793 // Find a legal type for the vector store.
12794 unsigned Elts = i + 1;
12796 // When merging vector stores, get the total number of elements.
12797 Elts *= MemVT.getVectorNumElements();
12800 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(), Elts);
12802 if (TLI.isTypeLegal(Ty) && TLI.canMergeStoresTo(FirstStoreAS, Ty) &&
12803 TLI.allowsMemoryAccess(Context, DL, Ty, FirstStoreAS,
12804 FirstStoreAlign, &IsFast) &&
12806 NumStoresToMerge = i + 1;
12809 bool Merged = MergeStoresOfConstantsOrVecElts(
12810 StoreNodes, MemVT, NumStoresToMerge, false, true);
12812 StoreNodes.erase(StoreNodes.begin(),
12813 StoreNodes.begin() + NumStoresToMerge);
12816 // Remove merged stores for next iteration.
12817 StoreNodes.erase(StoreNodes.begin(),
12818 StoreNodes.begin() + NumStoresToMerge);
12823 // Below we handle the case of multiple consecutive stores that
12824 // come from multiple consecutive loads. We merge them into a single
12825 // wide load and a single wide store.
12827 // Look for load nodes which are used by the stored values.
12828 SmallVector<MemOpLink, 8> LoadNodes;
12830 // Find acceptable loads. Loads need to have the same chain (token factor),
12831 // must not be zext, volatile, indexed, and they must be consecutive.
12832 BaseIndexOffset LdBasePtr;
12833 for (unsigned i = 0; i < NumConsecutiveStores; ++i) {
12834 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
12835 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
12839 // Loads must only have one use.
12840 if (!Ld->hasNUsesOfValue(1, 0))
12843 // The memory operands must not be volatile.
12844 if (Ld->isVolatile() || Ld->isIndexed())
12847 // We do not accept ext loads.
12848 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
12851 // The stored memory type must be the same.
12852 if (Ld->getMemoryVT() != MemVT)
12855 BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld->getBasePtr(), DAG);
12856 // If this is not the first ptr that we check.
12857 if (LdBasePtr.Base.getNode()) {
12858 // The base ptr must be the same.
12859 if (!LdPtr.equalBaseIndex(LdBasePtr))
12862 // Check that all other base pointers are the same as this one.
12866 // We found a potential memory operand to merge.
12867 LoadNodes.push_back(MemOpLink(Ld, LdPtr.Offset));
12870 if (LoadNodes.size() < 2) {
12871 StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + 1);
12875 // If we have load/store pair instructions and we only have two values,
12876 // don't bother merging.
12877 unsigned RequiredAlignment;
12878 if (LoadNodes.size() == 2 && TLI.hasPairedLoad(MemVT, RequiredAlignment) &&
12879 StoreNodes[0].MemNode->getAlignment() >= RequiredAlignment) {
12880 StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + 2);
12883 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
12884 unsigned FirstStoreAS = FirstInChain->getAddressSpace();
12885 unsigned FirstStoreAlign = FirstInChain->getAlignment();
12886 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
12887 unsigned FirstLoadAS = FirstLoad->getAddressSpace();
12888 unsigned FirstLoadAlign = FirstLoad->getAlignment();
12890 // Scan the memory operations on the chain and find the first
12891 // non-consecutive load memory address. These variables hold the index in
12892 // the store node array.
12893 unsigned LastConsecutiveLoad = 0;
12894 // This variable refers to the size and not index in the array.
12895 unsigned LastLegalVectorType = 0;
12896 unsigned LastLegalIntegerType = 0;
12897 StartAddress = LoadNodes[0].OffsetFromBase;
12898 SDValue FirstChain = FirstLoad->getChain();
12899 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
12900 // All loads must share the same chain.
12901 if (LoadNodes[i].MemNode->getChain() != FirstChain)
12904 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
12905 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
12907 LastConsecutiveLoad = i;
12908 // Find a legal type for the vector store.
12909 EVT StoreTy = EVT::getVectorVT(Context, MemVT, i + 1);
12910 bool IsFastSt, IsFastLd;
12911 if (TLI.isTypeLegal(StoreTy) &&
12912 TLI.canMergeStoresTo(FirstStoreAS, StoreTy) &&
12913 TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstStoreAS,
12914 FirstStoreAlign, &IsFastSt) &&
12916 TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstLoadAS,
12917 FirstLoadAlign, &IsFastLd) &&
12919 LastLegalVectorType = i + 1;
12922 // Find a legal type for the integer store.
12923 unsigned SizeInBits = (i + 1) * ElementSizeBytes * 8;
12924 StoreTy = EVT::getIntegerVT(Context, SizeInBits);
12925 if (TLI.isTypeLegal(StoreTy) &&
12926 TLI.canMergeStoresTo(FirstStoreAS, StoreTy) &&
12927 TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstStoreAS,
12928 FirstStoreAlign, &IsFastSt) &&
12930 TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstLoadAS,
12931 FirstLoadAlign, &IsFastLd) &&
12933 LastLegalIntegerType = i + 1;
12934 // Or check whether a truncstore and extload is legal.
12935 else if (TLI.getTypeAction(Context, StoreTy) ==
12936 TargetLowering::TypePromoteInteger) {
12937 EVT LegalizedStoredValueTy = TLI.getTypeToTransformTo(Context, StoreTy);
12939 TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
12940 TLI.canMergeStoresTo(FirstStoreAS, LegalizedStoredValueTy) &&
12941 TLI.isLoadExtLegal(ISD::ZEXTLOAD, LegalizedStoredValueTy,
12943 TLI.isLoadExtLegal(ISD::SEXTLOAD, LegalizedStoredValueTy,
12945 TLI.isLoadExtLegal(ISD::EXTLOAD, LegalizedStoredValueTy, StoreTy) &&
12946 TLI.allowsMemoryAccess(Context, DL, LegalizedStoredValueTy,
12947 FirstStoreAS, FirstStoreAlign, &IsFastSt) &&
12949 TLI.allowsMemoryAccess(Context, DL, LegalizedStoredValueTy,
12950 FirstLoadAS, FirstLoadAlign, &IsFastLd) &&
12952 LastLegalIntegerType = i + 1;
12956 // Only use vector types if the vector type is larger than the integer type.
12957 // If they are the same, use integers.
12958 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors;
12959 unsigned LastLegalType =
12960 std::max(LastLegalVectorType, LastLegalIntegerType);
12962 // We add +1 here because the LastXXX variables refer to location while
12963 // the NumElem refers to array/index size.
12964 unsigned NumElem = std::min(NumConsecutiveStores, LastConsecutiveLoad + 1);
12965 NumElem = std::min(LastLegalType, NumElem);
12968 StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + 1);
12972 // Find if it is better to use vectors or integers to load and store
12976 JointMemOpVT = EVT::getVectorVT(Context, MemVT, NumElem);
12978 unsigned SizeInBits = NumElem * ElementSizeBytes * 8;
12979 JointMemOpVT = EVT::getIntegerVT(Context, SizeInBits);
12982 SDLoc LoadDL(LoadNodes[0].MemNode);
12983 SDLoc StoreDL(StoreNodes[0].MemNode);
12985 // The merged loads are required to have the same incoming chain, so
12986 // using the first's chain is acceptable.
12987 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL, FirstLoad->getChain(),
12988 FirstLoad->getBasePtr(),
12989 FirstLoad->getPointerInfo(), FirstLoadAlign);
12991 SDValue NewStoreChain = getMergeStoreChains(StoreNodes, NumElem);
12993 AddToWorklist(NewStoreChain.getNode());
12995 SDValue NewStore = DAG.getStore(
12996 NewStoreChain, StoreDL, NewLoad, FirstInChain->getBasePtr(),
12997 FirstInChain->getPointerInfo(), FirstStoreAlign);
12999 // Transfer chain users from old loads to the new load.
13000 for (unsigned i = 0; i < NumElem; ++i) {
13001 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
13002 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
13003 SDValue(NewLoad.getNode(), 1));
13006 // Replace the all stores with the new store.
13007 for (unsigned i = 0; i < NumElem; ++i)
13008 CombineTo(StoreNodes[i].MemNode, NewStore);
13010 StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + NumElem);
13016 SDValue DAGCombiner::replaceStoreChain(StoreSDNode *ST, SDValue BetterChain) {
13020 // Replace the chain to avoid dependency.
13021 if (ST->isTruncatingStore()) {
13022 ReplStore = DAG.getTruncStore(BetterChain, SL, ST->getValue(),
13023 ST->getBasePtr(), ST->getMemoryVT(),
13024 ST->getMemOperand());
13026 ReplStore = DAG.getStore(BetterChain, SL, ST->getValue(), ST->getBasePtr(),
13027 ST->getMemOperand());
13030 // Create token to keep both nodes around.
13031 SDValue Token = DAG.getNode(ISD::TokenFactor, SL,
13032 MVT::Other, ST->getChain(), ReplStore);
13034 // Make sure the new and old chains are cleaned up.
13035 AddToWorklist(Token.getNode());
13037 // Don't add users to work list.
13038 return CombineTo(ST, Token, false);
13041 SDValue DAGCombiner::replaceStoreOfFPConstant(StoreSDNode *ST) {
13042 SDValue Value = ST->getValue();
13043 if (Value.getOpcode() == ISD::TargetConstantFP)
13048 SDValue Chain = ST->getChain();
13049 SDValue Ptr = ST->getBasePtr();
13051 const ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Value);
13053 // NOTE: If the original store is volatile, this transform must not increase
13054 // the number of stores. For example, on x86-32 an f64 can be stored in one
13055 // processor operation but an i64 (which is not legal) requires two. So the
13056 // transform should not be done in this case.
13059 switch (CFP->getSimpleValueType(0).SimpleTy) {
13061 llvm_unreachable("Unknown FP type");
13062 case MVT::f16: // We don't do this for these yet.
13068 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
13069 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
13071 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
13072 bitcastToAPInt().getZExtValue(), SDLoc(CFP),
13074 return DAG.getStore(Chain, DL, Tmp, Ptr, ST->getMemOperand());
13079 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
13080 !ST->isVolatile()) ||
13081 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
13083 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
13084 getZExtValue(), SDLoc(CFP), MVT::i64);
13085 return DAG.getStore(Chain, DL, Tmp,
13086 Ptr, ST->getMemOperand());
13089 if (!ST->isVolatile() &&
13090 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
13091 // Many FP stores are not made apparent until after legalize, e.g. for
13092 // argument passing. Since this is so common, custom legalize the
13093 // 64-bit integer store into two 32-bit stores.
13094 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
13095 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, SDLoc(CFP), MVT::i32);
13096 SDValue Hi = DAG.getConstant(Val >> 32, SDLoc(CFP), MVT::i32);
13097 if (DAG.getDataLayout().isBigEndian())
13100 unsigned Alignment = ST->getAlignment();
13101 MachineMemOperand::Flags MMOFlags = ST->getMemOperand()->getFlags();
13102 AAMDNodes AAInfo = ST->getAAInfo();
13104 SDValue St0 = DAG.getStore(Chain, DL, Lo, Ptr, ST->getPointerInfo(),
13105 ST->getAlignment(), MMOFlags, AAInfo);
13106 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
13107 DAG.getConstant(4, DL, Ptr.getValueType()));
13108 Alignment = MinAlign(Alignment, 4U);
13109 SDValue St1 = DAG.getStore(Chain, DL, Hi, Ptr,
13110 ST->getPointerInfo().getWithOffset(4),
13111 Alignment, MMOFlags, AAInfo);
13112 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
13120 SDValue DAGCombiner::visitSTORE(SDNode *N) {
13121 StoreSDNode *ST = cast<StoreSDNode>(N);
13122 SDValue Chain = ST->getChain();
13123 SDValue Value = ST->getValue();
13124 SDValue Ptr = ST->getBasePtr();
13126 // If this is a store of a bit convert, store the input value if the
13127 // resultant store does not need a higher alignment than the original.
13128 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
13129 ST->isUnindexed()) {
13130 EVT SVT = Value.getOperand(0).getValueType();
13131 if (((!LegalOperations && !ST->isVolatile()) ||
13132 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)) &&
13133 TLI.isStoreBitCastBeneficial(Value.getValueType(), SVT)) {
13134 unsigned OrigAlign = ST->getAlignment();
13136 if (TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), SVT,
13137 ST->getAddressSpace(), OrigAlign, &Fast) &&
13139 return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0), Ptr,
13140 ST->getPointerInfo(), OrigAlign,
13141 ST->getMemOperand()->getFlags(), ST->getAAInfo());
13146 // Turn 'store undef, Ptr' -> nothing.
13147 if (Value.isUndef() && ST->isUnindexed())
13150 // Try to infer better alignment information than the store already has.
13151 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
13152 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
13153 if (Align > ST->getAlignment()) {
13155 DAG.getTruncStore(Chain, SDLoc(N), Value, Ptr, ST->getPointerInfo(),
13156 ST->getMemoryVT(), Align,
13157 ST->getMemOperand()->getFlags(), ST->getAAInfo());
13158 if (NewStore.getNode() != N)
13159 return CombineTo(ST, NewStore, true);
13164 // Try transforming a pair floating point load / store ops to integer
13165 // load / store ops.
13166 if (SDValue NewST = TransformFPLoadStorePair(N))
13169 if (ST->isUnindexed()) {
13170 // Walk up chain skipping non-aliasing memory nodes, on this store and any
13171 // adjacent stores.
13172 if (findBetterNeighborChains(ST)) {
13173 // replaceStoreChain uses CombineTo, which handled all of the worklist
13174 // manipulation. Return the original node to not do anything else.
13175 return SDValue(ST, 0);
13177 Chain = ST->getChain();
13180 // Try transforming N to an indexed store.
13181 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
13182 return SDValue(N, 0);
13184 // FIXME: is there such a thing as a truncating indexed store?
13185 if (ST->isTruncatingStore() && ST->isUnindexed() &&
13186 Value.getValueType().isInteger()) {
13187 // See if we can simplify the input to this truncstore with knowledge that
13188 // only the low bits are being used. For example:
13189 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
13190 SDValue Shorter = GetDemandedBits(
13191 Value, APInt::getLowBitsSet(Value.getScalarValueSizeInBits(),
13192 ST->getMemoryVT().getScalarSizeInBits()));
13193 AddToWorklist(Value.getNode());
13194 if (Shorter.getNode())
13195 return DAG.getTruncStore(Chain, SDLoc(N), Shorter,
13196 Ptr, ST->getMemoryVT(), ST->getMemOperand());
13198 // Otherwise, see if we can simplify the operation with
13199 // SimplifyDemandedBits, which only works if the value has a single use.
13200 if (SimplifyDemandedBits(
13202 APInt::getLowBitsSet(Value.getScalarValueSizeInBits(),
13203 ST->getMemoryVT().getScalarSizeInBits()))) {
13204 // Re-visit the store if anything changed and the store hasn't been merged
13205 // with another node (N is deleted) SimplifyDemandedBits will add Value's
13206 // node back to the worklist if necessary, but we also need to re-visit
13207 // the Store node itself.
13208 if (N->getOpcode() != ISD::DELETED_NODE)
13210 return SDValue(N, 0);
13214 // If this is a load followed by a store to the same location, then the store
13216 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
13217 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
13218 ST->isUnindexed() && !ST->isVolatile() &&
13219 // There can't be any side effects between the load and store, such as
13220 // a call or store.
13221 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
13222 // The store is dead, remove it.
13227 if (StoreSDNode *ST1 = dyn_cast<StoreSDNode>(Chain)) {
13228 if (ST->isUnindexed() && !ST->isVolatile() && ST1->isUnindexed() &&
13229 !ST1->isVolatile() && ST1->getBasePtr() == Ptr &&
13230 ST->getMemoryVT() == ST1->getMemoryVT()) {
13231 // If this is a store followed by a store with the same value to the same
13232 // location, then the store is dead/noop.
13233 if (ST1->getValue() == Value) {
13234 // The store is dead, remove it.
13238 // If this is a store who's preceeding store to the same location
13239 // and no one other node is chained to that store we can effectively
13240 // drop the store. Do not remove stores to undef as they may be used as
13242 if (OptLevel != CodeGenOpt::None && ST1->hasOneUse() &&
13243 !ST1->getBasePtr().isUndef()) {
13244 // ST1 is fully overwritten and can be elided. Combine with it's chain
13246 CombineTo(ST1, ST1->getChain());
13252 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
13253 // truncating store. We can do this even if this is already a truncstore.
13254 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
13255 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
13256 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
13257 ST->getMemoryVT())) {
13258 return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0),
13259 Ptr, ST->getMemoryVT(), ST->getMemOperand());
13262 // Only perform this optimization before the types are legal, because we
13263 // don't want to perform this optimization on every DAGCombine invocation.
13266 // There can be multiple store sequences on the same chain.
13267 // Keep trying to merge store sequences until we are unable to do so
13268 // or until we merge the last store on the chain.
13269 bool Changed = MergeConsecutiveStores(ST);
13270 if (!Changed) break;
13271 // Return N as merge only uses CombineTo and no worklist clean
13272 // up is necessary.
13273 if (N->getOpcode() == ISD::DELETED_NODE || !isa<StoreSDNode>(N))
13274 return SDValue(N, 0);
13278 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
13280 // Make sure to do this only after attempting to merge stores in order to
13281 // avoid changing the types of some subset of stores due to visit order,
13282 // preventing their merging.
13283 if (isa<ConstantFPSDNode>(ST->getValue())) {
13284 if (SDValue NewSt = replaceStoreOfFPConstant(ST))
13288 if (SDValue NewSt = splitMergedValStore(ST))
13291 return ReduceLoadOpStoreWidth(N);
13294 /// For the instruction sequence of store below, F and I values
13295 /// are bundled together as an i64 value before being stored into memory.
13296 /// Sometimes it is more efficent to generate separate stores for F and I,
13297 /// which can remove the bitwise instructions or sink them to colder places.
13299 /// (store (or (zext (bitcast F to i32) to i64),
13300 /// (shl (zext I to i64), 32)), addr) -->
13301 /// (store F, addr) and (store I, addr+4)
13303 /// Similarly, splitting for other merged store can also be beneficial, like:
13304 /// For pair of {i32, i32}, i64 store --> two i32 stores.
13305 /// For pair of {i32, i16}, i64 store --> two i32 stores.
13306 /// For pair of {i16, i16}, i32 store --> two i16 stores.
13307 /// For pair of {i16, i8}, i32 store --> two i16 stores.
13308 /// For pair of {i8, i8}, i16 store --> two i8 stores.
13310 /// We allow each target to determine specifically which kind of splitting is
13313 /// The store patterns are commonly seen from the simple code snippet below
13314 /// if only std::make_pair(...) is sroa transformed before inlined into hoo.
13315 /// void goo(const std::pair<int, float> &);
13318 /// goo(std::make_pair(tmp, ftmp));
13322 SDValue DAGCombiner::splitMergedValStore(StoreSDNode *ST) {
13323 if (OptLevel == CodeGenOpt::None)
13326 SDValue Val = ST->getValue();
13329 // Match OR operand.
13330 if (!Val.getValueType().isScalarInteger() || Val.getOpcode() != ISD::OR)
13333 // Match SHL operand and get Lower and Higher parts of Val.
13334 SDValue Op1 = Val.getOperand(0);
13335 SDValue Op2 = Val.getOperand(1);
13337 if (Op1.getOpcode() != ISD::SHL) {
13338 std::swap(Op1, Op2);
13339 if (Op1.getOpcode() != ISD::SHL)
13343 Hi = Op1.getOperand(0);
13344 if (!Op1.hasOneUse())
13347 // Match shift amount to HalfValBitSize.
13348 unsigned HalfValBitSize = Val.getValueSizeInBits() / 2;
13349 ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(Op1.getOperand(1));
13350 if (!ShAmt || ShAmt->getAPIntValue() != HalfValBitSize)
13353 // Lo and Hi are zero-extended from int with size less equal than 32
13355 if (Lo.getOpcode() != ISD::ZERO_EXTEND || !Lo.hasOneUse() ||
13356 !Lo.getOperand(0).getValueType().isScalarInteger() ||
13357 Lo.getOperand(0).getValueSizeInBits() > HalfValBitSize ||
13358 Hi.getOpcode() != ISD::ZERO_EXTEND || !Hi.hasOneUse() ||
13359 !Hi.getOperand(0).getValueType().isScalarInteger() ||
13360 Hi.getOperand(0).getValueSizeInBits() > HalfValBitSize)
13363 // Use the EVT of low and high parts before bitcast as the input
13364 // of target query.
13365 EVT LowTy = (Lo.getOperand(0).getOpcode() == ISD::BITCAST)
13366 ? Lo.getOperand(0).getValueType()
13367 : Lo.getValueType();
13368 EVT HighTy = (Hi.getOperand(0).getOpcode() == ISD::BITCAST)
13369 ? Hi.getOperand(0).getValueType()
13370 : Hi.getValueType();
13371 if (!TLI.isMultiStoresCheaperThanBitsMerge(LowTy, HighTy))
13374 // Start to split store.
13375 unsigned Alignment = ST->getAlignment();
13376 MachineMemOperand::Flags MMOFlags = ST->getMemOperand()->getFlags();
13377 AAMDNodes AAInfo = ST->getAAInfo();
13379 // Change the sizes of Lo and Hi's value types to HalfValBitSize.
13380 EVT VT = EVT::getIntegerVT(*DAG.getContext(), HalfValBitSize);
13381 Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Lo.getOperand(0));
13382 Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Hi.getOperand(0));
13384 SDValue Chain = ST->getChain();
13385 SDValue Ptr = ST->getBasePtr();
13386 // Lower value store.
13387 SDValue St0 = DAG.getStore(Chain, DL, Lo, Ptr, ST->getPointerInfo(),
13388 ST->getAlignment(), MMOFlags, AAInfo);
13390 DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
13391 DAG.getConstant(HalfValBitSize / 8, DL, Ptr.getValueType()));
13392 // Higher value store.
13394 DAG.getStore(St0, DL, Hi, Ptr,
13395 ST->getPointerInfo().getWithOffset(HalfValBitSize / 8),
13396 Alignment / 2, MMOFlags, AAInfo);
13400 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
13401 SDValue InVec = N->getOperand(0);
13402 SDValue InVal = N->getOperand(1);
13403 SDValue EltNo = N->getOperand(2);
13406 // If the inserted element is an UNDEF, just use the input vector.
13407 if (InVal.isUndef())
13410 EVT VT = InVec.getValueType();
13412 // Check that we know which element is being inserted
13413 if (!isa<ConstantSDNode>(EltNo))
13415 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
13417 // Canonicalize insert_vector_elt dag nodes.
13419 // (insert_vector_elt (insert_vector_elt A, Idx0), Idx1)
13420 // -> (insert_vector_elt (insert_vector_elt A, Idx1), Idx0)
13422 // Do this only if the child insert_vector node has one use; also
13423 // do this only if indices are both constants and Idx1 < Idx0.
13424 if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT && InVec.hasOneUse()
13425 && isa<ConstantSDNode>(InVec.getOperand(2))) {
13426 unsigned OtherElt = InVec.getConstantOperandVal(2);
13427 if (Elt < OtherElt) {
13429 SDValue NewOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT,
13430 InVec.getOperand(0), InVal, EltNo);
13431 AddToWorklist(NewOp.getNode());
13432 return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(InVec.getNode()),
13433 VT, NewOp, InVec.getOperand(1), InVec.getOperand(2));
13437 // If we can't generate a legal BUILD_VECTOR, exit
13438 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
13441 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
13442 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
13443 // vector elements.
13444 SmallVector<SDValue, 8> Ops;
13445 // Do not combine these two vectors if the output vector will not replace
13446 // the input vector.
13447 if (InVec.getOpcode() == ISD::BUILD_VECTOR && InVec.hasOneUse()) {
13448 Ops.append(InVec.getNode()->op_begin(),
13449 InVec.getNode()->op_end());
13450 } else if (InVec.isUndef()) {
13451 unsigned NElts = VT.getVectorNumElements();
13452 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
13457 // Insert the element
13458 if (Elt < Ops.size()) {
13459 // All the operands of BUILD_VECTOR must have the same type;
13460 // we enforce that here.
13461 EVT OpVT = Ops[0].getValueType();
13462 Ops[Elt] = OpVT.isInteger() ? DAG.getAnyExtOrTrunc(InVal, DL, OpVT) : InVal;
13465 // Return the new vector
13466 return DAG.getBuildVector(VT, DL, Ops);
13469 SDValue DAGCombiner::ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
13470 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad) {
13471 assert(!OriginalLoad->isVolatile());
13473 EVT ResultVT = EVE->getValueType(0);
13474 EVT VecEltVT = InVecVT.getVectorElementType();
13475 unsigned Align = OriginalLoad->getAlignment();
13476 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
13477 VecEltVT.getTypeForEVT(*DAG.getContext()));
13479 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VecEltVT))
13482 ISD::LoadExtType ExtTy = ResultVT.bitsGT(VecEltVT) ?
13483 ISD::NON_EXTLOAD : ISD::EXTLOAD;
13484 if (!TLI.shouldReduceLoadWidth(OriginalLoad, ExtTy, VecEltVT))
13489 SDValue NewPtr = OriginalLoad->getBasePtr();
13491 EVT PtrType = NewPtr.getValueType();
13492 MachinePointerInfo MPI;
13494 if (auto *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo)) {
13495 int Elt = ConstEltNo->getZExtValue();
13496 unsigned PtrOff = VecEltVT.getSizeInBits() * Elt / 8;
13497 Offset = DAG.getConstant(PtrOff, DL, PtrType);
13498 MPI = OriginalLoad->getPointerInfo().getWithOffset(PtrOff);
13500 Offset = DAG.getZExtOrTrunc(EltNo, DL, PtrType);
13501 Offset = DAG.getNode(
13502 ISD::MUL, DL, PtrType, Offset,
13503 DAG.getConstant(VecEltVT.getStoreSize(), DL, PtrType));
13504 MPI = OriginalLoad->getPointerInfo();
13506 NewPtr = DAG.getNode(ISD::ADD, DL, PtrType, NewPtr, Offset);
13508 // The replacement we need to do here is a little tricky: we need to
13509 // replace an extractelement of a load with a load.
13510 // Use ReplaceAllUsesOfValuesWith to do the replacement.
13511 // Note that this replacement assumes that the extractvalue is the only
13512 // use of the load; that's okay because we don't want to perform this
13513 // transformation in other cases anyway.
13516 if (ResultVT.bitsGT(VecEltVT)) {
13517 // If the result type of vextract is wider than the load, then issue an
13518 // extending load instead.
13519 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, ResultVT,
13523 Load = DAG.getExtLoad(ExtType, SDLoc(EVE), ResultVT,
13524 OriginalLoad->getChain(), NewPtr, MPI, VecEltVT,
13525 Align, OriginalLoad->getMemOperand()->getFlags(),
13526 OriginalLoad->getAAInfo());
13527 Chain = Load.getValue(1);
13529 Load = DAG.getLoad(VecEltVT, SDLoc(EVE), OriginalLoad->getChain(), NewPtr,
13530 MPI, Align, OriginalLoad->getMemOperand()->getFlags(),
13531 OriginalLoad->getAAInfo());
13532 Chain = Load.getValue(1);
13533 if (ResultVT.bitsLT(VecEltVT))
13534 Load = DAG.getNode(ISD::TRUNCATE, SDLoc(EVE), ResultVT, Load);
13536 Load = DAG.getBitcast(ResultVT, Load);
13538 WorklistRemover DeadNodes(*this);
13539 SDValue From[] = { SDValue(EVE, 0), SDValue(OriginalLoad, 1) };
13540 SDValue To[] = { Load, Chain };
13541 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
13542 // Since we're explicitly calling ReplaceAllUses, add the new node to the
13543 // worklist explicitly as well.
13544 AddToWorklist(Load.getNode());
13545 AddUsersToWorklist(Load.getNode()); // Add users too
13546 // Make sure to revisit this node to clean it up; it will usually be dead.
13547 AddToWorklist(EVE);
13549 return SDValue(EVE, 0);
13552 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
13553 // (vextract (scalar_to_vector val, 0) -> val
13554 SDValue InVec = N->getOperand(0);
13555 EVT VT = InVec.getValueType();
13556 EVT NVT = N->getValueType(0);
13558 if (InVec.isUndef())
13559 return DAG.getUNDEF(NVT);
13561 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
13562 // Check if the result type doesn't match the inserted element type. A
13563 // SCALAR_TO_VECTOR may truncate the inserted element and the
13564 // EXTRACT_VECTOR_ELT may widen the extracted vector.
13565 SDValue InOp = InVec.getOperand(0);
13566 if (InOp.getValueType() != NVT) {
13567 assert(InOp.getValueType().isInteger() && NVT.isInteger());
13568 return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT);
13573 SDValue EltNo = N->getOperand(1);
13574 ConstantSDNode *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo);
13576 // extract_vector_elt (build_vector x, y), 1 -> y
13578 InVec.getOpcode() == ISD::BUILD_VECTOR &&
13579 TLI.isTypeLegal(VT) &&
13580 (InVec.hasOneUse() ||
13581 TLI.aggressivelyPreferBuildVectorSources(VT))) {
13582 SDValue Elt = InVec.getOperand(ConstEltNo->getZExtValue());
13583 EVT InEltVT = Elt.getValueType();
13585 // Sometimes build_vector's scalar input types do not match result type.
13586 if (NVT == InEltVT)
13589 // TODO: It may be useful to truncate if free if the build_vector implicitly
13593 // extract_vector_elt (v2i32 (bitcast i64:x)), 0 -> i32 (trunc i64:x)
13594 if (ConstEltNo && InVec.getOpcode() == ISD::BITCAST && InVec.hasOneUse() &&
13595 ConstEltNo->isNullValue() && VT.isInteger()) {
13596 SDValue BCSrc = InVec.getOperand(0);
13597 if (BCSrc.getValueType().isScalarInteger())
13598 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), NVT, BCSrc);
13601 // extract_vector_elt (insert_vector_elt vec, val, idx), idx) -> val
13603 // This only really matters if the index is non-constant since other combines
13604 // on the constant elements already work.
13605 if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT &&
13606 EltNo == InVec.getOperand(2)) {
13607 SDValue Elt = InVec.getOperand(1);
13608 return VT.isInteger() ? DAG.getAnyExtOrTrunc(Elt, SDLoc(N), NVT) : Elt;
13611 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
13612 // We only perform this optimization before the op legalization phase because
13613 // we may introduce new vector instructions which are not backed by TD
13614 // patterns. For example on AVX, extracting elements from a wide vector
13615 // without using extract_subvector. However, if we can find an underlying
13616 // scalar value, then we can always use that.
13617 if (ConstEltNo && InVec.getOpcode() == ISD::VECTOR_SHUFFLE) {
13618 int NumElem = VT.getVectorNumElements();
13619 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
13620 // Find the new index to extract from.
13621 int OrigElt = SVOp->getMaskElt(ConstEltNo->getZExtValue());
13623 // Extracting an undef index is undef.
13625 return DAG.getUNDEF(NVT);
13627 // Select the right vector half to extract from.
13629 if (OrigElt < NumElem) {
13630 SVInVec = InVec->getOperand(0);
13632 SVInVec = InVec->getOperand(1);
13633 OrigElt -= NumElem;
13636 if (SVInVec.getOpcode() == ISD::BUILD_VECTOR) {
13637 SDValue InOp = SVInVec.getOperand(OrigElt);
13638 if (InOp.getValueType() != NVT) {
13639 assert(InOp.getValueType().isInteger() && NVT.isInteger());
13640 InOp = DAG.getSExtOrTrunc(InOp, SDLoc(SVInVec), NVT);
13646 // FIXME: We should handle recursing on other vector shuffles and
13647 // scalar_to_vector here as well.
13649 if (!LegalOperations) {
13650 EVT IndexTy = TLI.getVectorIdxTy(DAG.getDataLayout());
13651 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT, SVInVec,
13652 DAG.getConstant(OrigElt, SDLoc(SVOp), IndexTy));
13656 bool BCNumEltsChanged = false;
13657 EVT ExtVT = VT.getVectorElementType();
13660 // If the result of load has to be truncated, then it's not necessarily
13662 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
13665 if (InVec.getOpcode() == ISD::BITCAST) {
13666 // Don't duplicate a load with other uses.
13667 if (!InVec.hasOneUse())
13670 EVT BCVT = InVec.getOperand(0).getValueType();
13671 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
13673 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
13674 BCNumEltsChanged = true;
13675 InVec = InVec.getOperand(0);
13676 ExtVT = BCVT.getVectorElementType();
13679 // (vextract (vN[if]M load $addr), i) -> ([if]M load $addr + i * size)
13680 if (!LegalOperations && !ConstEltNo && InVec.hasOneUse() &&
13681 ISD::isNormalLoad(InVec.getNode()) &&
13682 !N->getOperand(1)->hasPredecessor(InVec.getNode())) {
13683 SDValue Index = N->getOperand(1);
13684 if (LoadSDNode *OrigLoad = dyn_cast<LoadSDNode>(InVec)) {
13685 if (!OrigLoad->isVolatile()) {
13686 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, Index,
13692 // Perform only after legalization to ensure build_vector / vector_shuffle
13693 // optimizations have already been done.
13694 if (!LegalOperations) return SDValue();
13696 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
13697 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
13698 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
13701 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
13703 LoadSDNode *LN0 = nullptr;
13704 const ShuffleVectorSDNode *SVN = nullptr;
13705 if (ISD::isNormalLoad(InVec.getNode())) {
13706 LN0 = cast<LoadSDNode>(InVec);
13707 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
13708 InVec.getOperand(0).getValueType() == ExtVT &&
13709 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
13710 // Don't duplicate a load with other uses.
13711 if (!InVec.hasOneUse())
13714 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
13715 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
13716 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
13718 // (load $addr+1*size)
13720 // Don't duplicate a load with other uses.
13721 if (!InVec.hasOneUse())
13724 // If the bit convert changed the number of elements, it is unsafe
13725 // to examine the mask.
13726 if (BCNumEltsChanged)
13729 // Select the input vector, guarding against out of range extract vector.
13730 unsigned NumElems = VT.getVectorNumElements();
13731 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
13732 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
13734 if (InVec.getOpcode() == ISD::BITCAST) {
13735 // Don't duplicate a load with other uses.
13736 if (!InVec.hasOneUse())
13739 InVec = InVec.getOperand(0);
13741 if (ISD::isNormalLoad(InVec.getNode())) {
13742 LN0 = cast<LoadSDNode>(InVec);
13743 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
13744 EltNo = DAG.getConstant(Elt, SDLoc(EltNo), EltNo.getValueType());
13748 // Make sure we found a non-volatile load and the extractelement is
13750 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
13753 // If Idx was -1 above, Elt is going to be -1, so just return undef.
13755 return DAG.getUNDEF(LVT);
13757 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, EltNo, LN0);
13763 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
13764 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
13765 // We perform this optimization post type-legalization because
13766 // the type-legalizer often scalarizes integer-promoted vectors.
13767 // Performing this optimization before may create bit-casts which
13768 // will be type-legalized to complex code sequences.
13769 // We perform this optimization only before the operation legalizer because we
13770 // may introduce illegal operations.
13771 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
13774 unsigned NumInScalars = N->getNumOperands();
13776 EVT VT = N->getValueType(0);
13778 // Check to see if this is a BUILD_VECTOR of a bunch of values
13779 // which come from any_extend or zero_extend nodes. If so, we can create
13780 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
13781 // optimizations. We do not handle sign-extend because we can't fill the sign
13783 EVT SourceType = MVT::Other;
13784 bool AllAnyExt = true;
13786 for (unsigned i = 0; i != NumInScalars; ++i) {
13787 SDValue In = N->getOperand(i);
13788 // Ignore undef inputs.
13789 if (In.isUndef()) continue;
13791 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
13792 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
13794 // Abort if the element is not an extension.
13795 if (!ZeroExt && !AnyExt) {
13796 SourceType = MVT::Other;
13800 // The input is a ZeroExt or AnyExt. Check the original type.
13801 EVT InTy = In.getOperand(0).getValueType();
13803 // Check that all of the widened source types are the same.
13804 if (SourceType == MVT::Other)
13807 else if (InTy != SourceType) {
13808 // Multiple income types. Abort.
13809 SourceType = MVT::Other;
13813 // Check if all of the extends are ANY_EXTENDs.
13814 AllAnyExt &= AnyExt;
13817 // In order to have valid types, all of the inputs must be extended from the
13818 // same source type and all of the inputs must be any or zero extend.
13819 // Scalar sizes must be a power of two.
13820 EVT OutScalarTy = VT.getScalarType();
13821 bool ValidTypes = SourceType != MVT::Other &&
13822 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
13823 isPowerOf2_32(SourceType.getSizeInBits());
13825 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
13826 // turn into a single shuffle instruction.
13830 bool isLE = DAG.getDataLayout().isLittleEndian();
13831 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
13832 assert(ElemRatio > 1 && "Invalid element size ratio");
13833 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
13834 DAG.getConstant(0, DL, SourceType);
13836 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
13837 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
13839 // Populate the new build_vector
13840 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
13841 SDValue Cast = N->getOperand(i);
13842 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
13843 Cast.getOpcode() == ISD::ZERO_EXTEND ||
13844 Cast.isUndef()) && "Invalid cast opcode");
13846 if (Cast.isUndef())
13847 In = DAG.getUNDEF(SourceType);
13849 In = Cast->getOperand(0);
13850 unsigned Index = isLE ? (i * ElemRatio) :
13851 (i * ElemRatio + (ElemRatio - 1));
13853 assert(Index < Ops.size() && "Invalid index");
13857 // The type of the new BUILD_VECTOR node.
13858 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
13859 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
13860 "Invalid vector size");
13861 // Check if the new vector type is legal.
13862 if (!isTypeLegal(VecVT)) return SDValue();
13864 // Make the new BUILD_VECTOR.
13865 SDValue BV = DAG.getBuildVector(VecVT, DL, Ops);
13867 // The new BUILD_VECTOR node has the potential to be further optimized.
13868 AddToWorklist(BV.getNode());
13869 // Bitcast to the desired type.
13870 return DAG.getBitcast(VT, BV);
13873 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
13874 EVT VT = N->getValueType(0);
13876 unsigned NumInScalars = N->getNumOperands();
13879 EVT SrcVT = MVT::Other;
13880 unsigned Opcode = ISD::DELETED_NODE;
13881 unsigned NumDefs = 0;
13883 for (unsigned i = 0; i != NumInScalars; ++i) {
13884 SDValue In = N->getOperand(i);
13885 unsigned Opc = In.getOpcode();
13887 if (Opc == ISD::UNDEF)
13890 // If all scalar values are floats and converted from integers.
13891 if (Opcode == ISD::DELETED_NODE &&
13892 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
13899 EVT InVT = In.getOperand(0).getValueType();
13901 // If all scalar values are typed differently, bail out. It's chosen to
13902 // simplify BUILD_VECTOR of integer types.
13903 if (SrcVT == MVT::Other)
13910 // If the vector has just one element defined, it's not worth to fold it into
13911 // a vectorized one.
13915 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
13916 && "Should only handle conversion from integer to float.");
13917 assert(SrcVT != MVT::Other && "Cannot determine source type!");
13919 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
13921 if (!TLI.isOperationLegalOrCustom(Opcode, NVT))
13924 // Just because the floating-point vector type is legal does not necessarily
13925 // mean that the corresponding integer vector type is.
13926 if (!isTypeLegal(NVT))
13929 SmallVector<SDValue, 8> Opnds;
13930 for (unsigned i = 0; i != NumInScalars; ++i) {
13931 SDValue In = N->getOperand(i);
13934 Opnds.push_back(DAG.getUNDEF(SrcVT));
13936 Opnds.push_back(In.getOperand(0));
13938 SDValue BV = DAG.getBuildVector(NVT, DL, Opnds);
13939 AddToWorklist(BV.getNode());
13941 return DAG.getNode(Opcode, DL, VT, BV);
13944 SDValue DAGCombiner::createBuildVecShuffle(const SDLoc &DL, SDNode *N,
13945 ArrayRef<int> VectorMask,
13946 SDValue VecIn1, SDValue VecIn2,
13947 unsigned LeftIdx) {
13948 MVT IdxTy = TLI.getVectorIdxTy(DAG.getDataLayout());
13949 SDValue ZeroIdx = DAG.getConstant(0, DL, IdxTy);
13951 EVT VT = N->getValueType(0);
13952 EVT InVT1 = VecIn1.getValueType();
13953 EVT InVT2 = VecIn2.getNode() ? VecIn2.getValueType() : InVT1;
13955 unsigned Vec2Offset = InVT1.getVectorNumElements();
13956 unsigned NumElems = VT.getVectorNumElements();
13957 unsigned ShuffleNumElems = NumElems;
13959 // We can't generate a shuffle node with mismatched input and output types.
13960 // Try to make the types match the type of the output.
13961 if (InVT1 != VT || InVT2 != VT) {
13962 if ((VT.getSizeInBits() % InVT1.getSizeInBits() == 0) && InVT1 == InVT2) {
13963 // If the output vector length is a multiple of both input lengths,
13964 // we can concatenate them and pad the rest with undefs.
13965 unsigned NumConcats = VT.getSizeInBits() / InVT1.getSizeInBits();
13966 assert(NumConcats >= 2 && "Concat needs at least two inputs!");
13967 SmallVector<SDValue, 2> ConcatOps(NumConcats, DAG.getUNDEF(InVT1));
13968 ConcatOps[0] = VecIn1;
13969 ConcatOps[1] = VecIn2 ? VecIn2 : DAG.getUNDEF(InVT1);
13970 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, ConcatOps);
13971 VecIn2 = SDValue();
13972 } else if (InVT1.getSizeInBits() == VT.getSizeInBits() * 2) {
13973 if (!TLI.isExtractSubvectorCheap(VT, NumElems))
13976 if (!VecIn2.getNode()) {
13977 // If we only have one input vector, and it's twice the size of the
13978 // output, split it in two.
13979 VecIn2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, VecIn1,
13980 DAG.getConstant(NumElems, DL, IdxTy));
13981 VecIn1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, VecIn1, ZeroIdx);
13982 // Since we now have shorter input vectors, adjust the offset of the
13983 // second vector's start.
13984 Vec2Offset = NumElems;
13985 } else if (InVT2.getSizeInBits() <= InVT1.getSizeInBits()) {
13986 // VecIn1 is wider than the output, and we have another, possibly
13987 // smaller input. Pad the smaller input with undefs, shuffle at the
13988 // input vector width, and extract the output.
13989 // The shuffle type is different than VT, so check legality again.
13990 if (LegalOperations &&
13991 !TLI.isOperationLegal(ISD::VECTOR_SHUFFLE, InVT1))
13994 // Legalizing INSERT_SUBVECTOR is tricky - you basically have to
13995 // lower it back into a BUILD_VECTOR. So if the inserted type is
13996 // illegal, don't even try.
13997 if (InVT1 != InVT2) {
13998 if (!TLI.isTypeLegal(InVT2))
14000 VecIn2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, InVT1,
14001 DAG.getUNDEF(InVT1), VecIn2, ZeroIdx);
14003 ShuffleNumElems = NumElems * 2;
14005 // Both VecIn1 and VecIn2 are wider than the output, and VecIn2 is wider
14006 // than VecIn1. We can't handle this for now - this case will disappear
14007 // when we start sorting the vectors by type.
14011 // TODO: Support cases where the length mismatch isn't exactly by a
14013 // TODO: Move this check upwards, so that if we have bad type
14014 // mismatches, we don't create any DAG nodes.
14019 // Initialize mask to undef.
14020 SmallVector<int, 8> Mask(ShuffleNumElems, -1);
14022 // Only need to run up to the number of elements actually used, not the
14023 // total number of elements in the shuffle - if we are shuffling a wider
14024 // vector, the high lanes should be set to undef.
14025 for (unsigned i = 0; i != NumElems; ++i) {
14026 if (VectorMask[i] <= 0)
14029 unsigned ExtIndex = N->getOperand(i).getConstantOperandVal(1);
14030 if (VectorMask[i] == (int)LeftIdx) {
14031 Mask[i] = ExtIndex;
14032 } else if (VectorMask[i] == (int)LeftIdx + 1) {
14033 Mask[i] = Vec2Offset + ExtIndex;
14037 // The type the input vectors may have changed above.
14038 InVT1 = VecIn1.getValueType();
14040 // If we already have a VecIn2, it should have the same type as VecIn1.
14041 // If we don't, get an undef/zero vector of the appropriate type.
14042 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(InVT1);
14043 assert(InVT1 == VecIn2.getValueType() && "Unexpected second input type.");
14045 SDValue Shuffle = DAG.getVectorShuffle(InVT1, DL, VecIn1, VecIn2, Mask);
14046 if (ShuffleNumElems > NumElems)
14047 Shuffle = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Shuffle, ZeroIdx);
14052 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
14053 // operations. If the types of the vectors we're extracting from allow it,
14054 // turn this into a vector_shuffle node.
14055 SDValue DAGCombiner::reduceBuildVecToShuffle(SDNode *N) {
14057 EVT VT = N->getValueType(0);
14059 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
14060 if (!isTypeLegal(VT))
14063 // May only combine to shuffle after legalize if shuffle is legal.
14064 if (LegalOperations && !TLI.isOperationLegal(ISD::VECTOR_SHUFFLE, VT))
14067 bool UsesZeroVector = false;
14068 unsigned NumElems = N->getNumOperands();
14070 // Record, for each element of the newly built vector, which input vector
14071 // that element comes from. -1 stands for undef, 0 for the zero vector,
14072 // and positive values for the input vectors.
14073 // VectorMask maps each element to its vector number, and VecIn maps vector
14074 // numbers to their initial SDValues.
14076 SmallVector<int, 8> VectorMask(NumElems, -1);
14077 SmallVector<SDValue, 8> VecIn;
14078 VecIn.push_back(SDValue());
14080 for (unsigned i = 0; i != NumElems; ++i) {
14081 SDValue Op = N->getOperand(i);
14086 // See if we can use a blend with a zero vector.
14087 // TODO: Should we generalize this to a blend with an arbitrary constant
14089 if (isNullConstant(Op) || isNullFPConstant(Op)) {
14090 UsesZeroVector = true;
14095 // Not an undef or zero. If the input is something other than an
14096 // EXTRACT_VECTOR_ELT with a constant index, bail out.
14097 if (Op.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
14098 !isa<ConstantSDNode>(Op.getOperand(1)))
14101 SDValue ExtractedFromVec = Op.getOperand(0);
14103 // All inputs must have the same element type as the output.
14104 if (VT.getVectorElementType() !=
14105 ExtractedFromVec.getValueType().getVectorElementType())
14108 // Have we seen this input vector before?
14109 // The vectors are expected to be tiny (usually 1 or 2 elements), so using
14110 // a map back from SDValues to numbers isn't worth it.
14111 unsigned Idx = std::distance(
14112 VecIn.begin(), std::find(VecIn.begin(), VecIn.end(), ExtractedFromVec));
14113 if (Idx == VecIn.size())
14114 VecIn.push_back(ExtractedFromVec);
14116 VectorMask[i] = Idx;
14119 // If we didn't find at least one input vector, bail out.
14120 if (VecIn.size() < 2)
14123 // TODO: We want to sort the vectors by descending length, so that adjacent
14124 // pairs have similar length, and the longer vector is always first in the
14127 // TODO: Should this fire if some of the input vectors has illegal type (like
14128 // it does now), or should we let legalization run its course first?
14131 // Take pairs of vectors, and shuffle them so that the result has elements
14132 // from these vectors in the correct places.
14133 // For example, given:
14134 // t10: i32 = extract_vector_elt t1, Constant:i64<0>
14135 // t11: i32 = extract_vector_elt t2, Constant:i64<0>
14136 // t12: i32 = extract_vector_elt t3, Constant:i64<0>
14137 // t13: i32 = extract_vector_elt t1, Constant:i64<1>
14138 // t14: v4i32 = BUILD_VECTOR t10, t11, t12, t13
14139 // We will generate:
14140 // t20: v4i32 = vector_shuffle<0,4,u,1> t1, t2
14141 // t21: v4i32 = vector_shuffle<u,u,0,u> t3, undef
14142 SmallVector<SDValue, 4> Shuffles;
14143 for (unsigned In = 0, Len = (VecIn.size() / 2); In < Len; ++In) {
14144 unsigned LeftIdx = 2 * In + 1;
14145 SDValue VecLeft = VecIn[LeftIdx];
14147 (LeftIdx + 1) < VecIn.size() ? VecIn[LeftIdx + 1] : SDValue();
14149 if (SDValue Shuffle = createBuildVecShuffle(DL, N, VectorMask, VecLeft,
14150 VecRight, LeftIdx))
14151 Shuffles.push_back(Shuffle);
14156 // If we need the zero vector as an "ingredient" in the blend tree, add it
14157 // to the list of shuffles.
14158 if (UsesZeroVector)
14159 Shuffles.push_back(VT.isInteger() ? DAG.getConstant(0, DL, VT)
14160 : DAG.getConstantFP(0.0, DL, VT));
14162 // If we only have one shuffle, we're done.
14163 if (Shuffles.size() == 1)
14164 return Shuffles[0];
14166 // Update the vector mask to point to the post-shuffle vectors.
14167 for (int &Vec : VectorMask)
14169 Vec = Shuffles.size() - 1;
14171 Vec = (Vec - 1) / 2;
14173 // More than one shuffle. Generate a binary tree of blends, e.g. if from
14174 // the previous step we got the set of shuffles t10, t11, t12, t13, we will
14176 // t10: v8i32 = vector_shuffle<0,8,u,u,u,u,u,u> t1, t2
14177 // t11: v8i32 = vector_shuffle<u,u,0,8,u,u,u,u> t3, t4
14178 // t12: v8i32 = vector_shuffle<u,u,u,u,0,8,u,u> t5, t6
14179 // t13: v8i32 = vector_shuffle<u,u,u,u,u,u,0,8> t7, t8
14180 // t20: v8i32 = vector_shuffle<0,1,10,11,u,u,u,u> t10, t11
14181 // t21: v8i32 = vector_shuffle<u,u,u,u,4,5,14,15> t12, t13
14182 // t30: v8i32 = vector_shuffle<0,1,2,3,12,13,14,15> t20, t21
14184 // Make sure the initial size of the shuffle list is even.
14185 if (Shuffles.size() % 2)
14186 Shuffles.push_back(DAG.getUNDEF(VT));
14188 for (unsigned CurSize = Shuffles.size(); CurSize > 1; CurSize /= 2) {
14190 Shuffles[CurSize] = DAG.getUNDEF(VT);
14193 for (unsigned In = 0, Len = CurSize / 2; In < Len; ++In) {
14195 int Right = 2 * In + 1;
14196 SmallVector<int, 8> Mask(NumElems, -1);
14197 for (unsigned i = 0; i != NumElems; ++i) {
14198 if (VectorMask[i] == Left) {
14200 VectorMask[i] = In;
14201 } else if (VectorMask[i] == Right) {
14202 Mask[i] = i + NumElems;
14203 VectorMask[i] = In;
14208 DAG.getVectorShuffle(VT, DL, Shuffles[Left], Shuffles[Right], Mask);
14212 return Shuffles[0];
14215 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
14216 EVT VT = N->getValueType(0);
14218 // A vector built entirely of undefs is undef.
14219 if (ISD::allOperandsUndef(N))
14220 return DAG.getUNDEF(VT);
14222 // Check if we can express BUILD VECTOR via subvector extract.
14223 if (!LegalTypes && (N->getNumOperands() > 1)) {
14224 SDValue Op0 = N->getOperand(0);
14225 auto checkElem = [&](SDValue Op) -> uint64_t {
14226 if ((Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT) &&
14227 (Op0.getOperand(0) == Op.getOperand(0)))
14228 if (auto CNode = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
14229 return CNode->getZExtValue();
14233 int Offset = checkElem(Op0);
14234 for (unsigned i = 0; i < N->getNumOperands(); ++i) {
14235 if (Offset + i != checkElem(N->getOperand(i))) {
14241 if ((Offset == 0) &&
14242 (Op0.getOperand(0).getValueType() == N->getValueType(0)))
14243 return Op0.getOperand(0);
14244 if ((Offset != -1) &&
14245 ((Offset % N->getValueType(0).getVectorNumElements()) ==
14246 0)) // IDX must be multiple of output size.
14247 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N), N->getValueType(0),
14248 Op0.getOperand(0), Op0.getOperand(1));
14251 if (SDValue V = reduceBuildVecExtToExtBuildVec(N))
14254 if (SDValue V = reduceBuildVecConvertToConvertBuildVec(N))
14257 if (SDValue V = reduceBuildVecToShuffle(N))
14263 static SDValue combineConcatVectorOfScalars(SDNode *N, SelectionDAG &DAG) {
14264 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14265 EVT OpVT = N->getOperand(0).getValueType();
14267 // If the operands are legal vectors, leave them alone.
14268 if (TLI.isTypeLegal(OpVT))
14272 EVT VT = N->getValueType(0);
14273 SmallVector<SDValue, 8> Ops;
14275 EVT SVT = EVT::getIntegerVT(*DAG.getContext(), OpVT.getSizeInBits());
14276 SDValue ScalarUndef = DAG.getNode(ISD::UNDEF, DL, SVT);
14278 // Keep track of what we encounter.
14279 bool AnyInteger = false;
14280 bool AnyFP = false;
14281 for (const SDValue &Op : N->ops()) {
14282 if (ISD::BITCAST == Op.getOpcode() &&
14283 !Op.getOperand(0).getValueType().isVector())
14284 Ops.push_back(Op.getOperand(0));
14285 else if (ISD::UNDEF == Op.getOpcode())
14286 Ops.push_back(ScalarUndef);
14290 // Note whether we encounter an integer or floating point scalar.
14291 // If it's neither, bail out, it could be something weird like x86mmx.
14292 EVT LastOpVT = Ops.back().getValueType();
14293 if (LastOpVT.isFloatingPoint())
14295 else if (LastOpVT.isInteger())
14301 // If any of the operands is a floating point scalar bitcast to a vector,
14302 // use floating point types throughout, and bitcast everything.
14303 // Replace UNDEFs by another scalar UNDEF node, of the final desired type.
14305 SVT = EVT::getFloatingPointVT(OpVT.getSizeInBits());
14306 ScalarUndef = DAG.getNode(ISD::UNDEF, DL, SVT);
14308 for (SDValue &Op : Ops) {
14309 if (Op.getValueType() == SVT)
14314 Op = DAG.getBitcast(SVT, Op);
14319 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SVT,
14320 VT.getSizeInBits() / SVT.getSizeInBits());
14321 return DAG.getBitcast(VT, DAG.getBuildVector(VecVT, DL, Ops));
14324 // Check to see if this is a CONCAT_VECTORS of a bunch of EXTRACT_SUBVECTOR
14325 // operations. If so, and if the EXTRACT_SUBVECTOR vector inputs come from at
14326 // most two distinct vectors the same size as the result, attempt to turn this
14327 // into a legal shuffle.
14328 static SDValue combineConcatVectorOfExtracts(SDNode *N, SelectionDAG &DAG) {
14329 EVT VT = N->getValueType(0);
14330 EVT OpVT = N->getOperand(0).getValueType();
14331 int NumElts = VT.getVectorNumElements();
14332 int NumOpElts = OpVT.getVectorNumElements();
14334 SDValue SV0 = DAG.getUNDEF(VT), SV1 = DAG.getUNDEF(VT);
14335 SmallVector<int, 8> Mask;
14337 for (SDValue Op : N->ops()) {
14338 // Peek through any bitcast.
14339 while (Op.getOpcode() == ISD::BITCAST)
14340 Op = Op.getOperand(0);
14342 // UNDEF nodes convert to UNDEF shuffle mask values.
14343 if (Op.isUndef()) {
14344 Mask.append((unsigned)NumOpElts, -1);
14348 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
14351 // What vector are we extracting the subvector from and at what index?
14352 SDValue ExtVec = Op.getOperand(0);
14354 // We want the EVT of the original extraction to correctly scale the
14355 // extraction index.
14356 EVT ExtVT = ExtVec.getValueType();
14358 // Peek through any bitcast.
14359 while (ExtVec.getOpcode() == ISD::BITCAST)
14360 ExtVec = ExtVec.getOperand(0);
14362 // UNDEF nodes convert to UNDEF shuffle mask values.
14363 if (ExtVec.isUndef()) {
14364 Mask.append((unsigned)NumOpElts, -1);
14368 if (!isa<ConstantSDNode>(Op.getOperand(1)))
14370 int ExtIdx = Op.getConstantOperandVal(1);
14372 // Ensure that we are extracting a subvector from a vector the same
14373 // size as the result.
14374 if (ExtVT.getSizeInBits() != VT.getSizeInBits())
14377 // Scale the subvector index to account for any bitcast.
14378 int NumExtElts = ExtVT.getVectorNumElements();
14379 if (0 == (NumExtElts % NumElts))
14380 ExtIdx /= (NumExtElts / NumElts);
14381 else if (0 == (NumElts % NumExtElts))
14382 ExtIdx *= (NumElts / NumExtElts);
14386 // At most we can reference 2 inputs in the final shuffle.
14387 if (SV0.isUndef() || SV0 == ExtVec) {
14389 for (int i = 0; i != NumOpElts; ++i)
14390 Mask.push_back(i + ExtIdx);
14391 } else if (SV1.isUndef() || SV1 == ExtVec) {
14393 for (int i = 0; i != NumOpElts; ++i)
14394 Mask.push_back(i + ExtIdx + NumElts);
14400 if (!DAG.getTargetLoweringInfo().isShuffleMaskLegal(Mask, VT))
14403 return DAG.getVectorShuffle(VT, SDLoc(N), DAG.getBitcast(VT, SV0),
14404 DAG.getBitcast(VT, SV1), Mask);
14407 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
14408 // If we only have one input vector, we don't need to do any concatenation.
14409 if (N->getNumOperands() == 1)
14410 return N->getOperand(0);
14412 // Check if all of the operands are undefs.
14413 EVT VT = N->getValueType(0);
14414 if (ISD::allOperandsUndef(N))
14415 return DAG.getUNDEF(VT);
14417 // Optimize concat_vectors where all but the first of the vectors are undef.
14418 if (std::all_of(std::next(N->op_begin()), N->op_end(), [](const SDValue &Op) {
14419 return Op.isUndef();
14421 SDValue In = N->getOperand(0);
14422 assert(In.getValueType().isVector() && "Must concat vectors");
14424 // Transform: concat_vectors(scalar, undef) -> scalar_to_vector(sclr).
14425 if (In->getOpcode() == ISD::BITCAST &&
14426 !In->getOperand(0)->getValueType(0).isVector()) {
14427 SDValue Scalar = In->getOperand(0);
14429 // If the bitcast type isn't legal, it might be a trunc of a legal type;
14430 // look through the trunc so we can still do the transform:
14431 // concat_vectors(trunc(scalar), undef) -> scalar_to_vector(scalar)
14432 if (Scalar->getOpcode() == ISD::TRUNCATE &&
14433 !TLI.isTypeLegal(Scalar.getValueType()) &&
14434 TLI.isTypeLegal(Scalar->getOperand(0).getValueType()))
14435 Scalar = Scalar->getOperand(0);
14437 EVT SclTy = Scalar->getValueType(0);
14439 if (!SclTy.isFloatingPoint() && !SclTy.isInteger())
14442 unsigned VNTNumElms = VT.getSizeInBits() / SclTy.getSizeInBits();
14443 if (VNTNumElms < 2)
14446 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SclTy, VNTNumElms);
14447 if (!TLI.isTypeLegal(NVT) || !TLI.isTypeLegal(Scalar.getValueType()))
14450 SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(N), NVT, Scalar);
14451 return DAG.getBitcast(VT, Res);
14455 // Fold any combination of BUILD_VECTOR or UNDEF nodes into one BUILD_VECTOR.
14456 // We have already tested above for an UNDEF only concatenation.
14457 // fold (concat_vectors (BUILD_VECTOR A, B, ...), (BUILD_VECTOR C, D, ...))
14458 // -> (BUILD_VECTOR A, B, ..., C, D, ...)
14459 auto IsBuildVectorOrUndef = [](const SDValue &Op) {
14460 return ISD::UNDEF == Op.getOpcode() || ISD::BUILD_VECTOR == Op.getOpcode();
14462 if (llvm::all_of(N->ops(), IsBuildVectorOrUndef)) {
14463 SmallVector<SDValue, 8> Opnds;
14464 EVT SVT = VT.getScalarType();
14467 if (!SVT.isFloatingPoint()) {
14468 // If BUILD_VECTOR are from built from integer, they may have different
14469 // operand types. Get the smallest type and truncate all operands to it.
14470 bool FoundMinVT = false;
14471 for (const SDValue &Op : N->ops())
14472 if (ISD::BUILD_VECTOR == Op.getOpcode()) {
14473 EVT OpSVT = Op.getOperand(0)->getValueType(0);
14474 MinVT = (!FoundMinVT || OpSVT.bitsLE(MinVT)) ? OpSVT : MinVT;
14477 assert(FoundMinVT && "Concat vector type mismatch");
14480 for (const SDValue &Op : N->ops()) {
14481 EVT OpVT = Op.getValueType();
14482 unsigned NumElts = OpVT.getVectorNumElements();
14484 if (ISD::UNDEF == Op.getOpcode())
14485 Opnds.append(NumElts, DAG.getUNDEF(MinVT));
14487 if (ISD::BUILD_VECTOR == Op.getOpcode()) {
14488 if (SVT.isFloatingPoint()) {
14489 assert(SVT == OpVT.getScalarType() && "Concat vector type mismatch");
14490 Opnds.append(Op->op_begin(), Op->op_begin() + NumElts);
14492 for (unsigned i = 0; i != NumElts; ++i)
14494 DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinVT, Op.getOperand(i)));
14499 assert(VT.getVectorNumElements() == Opnds.size() &&
14500 "Concat vector type mismatch");
14501 return DAG.getBuildVector(VT, SDLoc(N), Opnds);
14504 // Fold CONCAT_VECTORS of only bitcast scalars (or undef) to BUILD_VECTOR.
14505 if (SDValue V = combineConcatVectorOfScalars(N, DAG))
14508 // Fold CONCAT_VECTORS of EXTRACT_SUBVECTOR (or undef) to VECTOR_SHUFFLE.
14509 if (Level < AfterLegalizeVectorOps && TLI.isTypeLegal(VT))
14510 if (SDValue V = combineConcatVectorOfExtracts(N, DAG))
14513 // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
14514 // nodes often generate nop CONCAT_VECTOR nodes.
14515 // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that
14516 // place the incoming vectors at the exact same location.
14517 SDValue SingleSource = SDValue();
14518 unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements();
14520 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
14521 SDValue Op = N->getOperand(i);
14526 // Check if this is the identity extract:
14527 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
14530 // Find the single incoming vector for the extract_subvector.
14531 if (SingleSource.getNode()) {
14532 if (Op.getOperand(0) != SingleSource)
14535 SingleSource = Op.getOperand(0);
14537 // Check the source type is the same as the type of the result.
14538 // If not, this concat may extend the vector, so we can not
14539 // optimize it away.
14540 if (SingleSource.getValueType() != N->getValueType(0))
14544 unsigned IdentityIndex = i * PartNumElem;
14545 ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1));
14546 // The extract index must be constant.
14550 // Check that we are reading from the identity index.
14551 if (CS->getZExtValue() != IdentityIndex)
14555 if (SingleSource.getNode())
14556 return SingleSource;
14561 /// If we are extracting a subvector produced by a wide binary operator with at
14562 /// at least one operand that was the result of a vector concatenation, then try
14563 /// to use the narrow vector operands directly to avoid the concatenation and
14565 static SDValue narrowExtractedVectorBinOp(SDNode *Extract, SelectionDAG &DAG) {
14566 // TODO: Refactor with the caller (visitEXTRACT_SUBVECTOR), so we can share
14567 // some of these bailouts with other transforms.
14569 // The extract index must be a constant, so we can map it to a concat operand.
14570 auto *ExtractIndex = dyn_cast<ConstantSDNode>(Extract->getOperand(1));
14574 // Only handle the case where we are doubling and then halving. A larger ratio
14575 // may require more than two narrow binops to replace the wide binop.
14576 EVT VT = Extract->getValueType(0);
14577 unsigned NumElems = VT.getVectorNumElements();
14578 assert((ExtractIndex->getZExtValue() % NumElems) == 0 &&
14579 "Extract index is not a multiple of the vector length.");
14580 if (Extract->getOperand(0).getValueSizeInBits() != VT.getSizeInBits() * 2)
14583 // We are looking for an optionally bitcasted wide vector binary operator
14584 // feeding an extract subvector.
14585 SDValue BinOp = Extract->getOperand(0);
14586 if (BinOp.getOpcode() == ISD::BITCAST)
14587 BinOp = BinOp.getOperand(0);
14589 // TODO: The motivating case for this transform is an x86 AVX1 target. That
14590 // target has temptingly almost legal versions of bitwise logic ops in 256-bit
14591 // flavors, but no other 256-bit integer support. This could be extended to
14592 // handle any binop, but that may require fixing/adding other folds to avoid
14593 // codegen regressions.
14594 unsigned BOpcode = BinOp.getOpcode();
14595 if (BOpcode != ISD::AND && BOpcode != ISD::OR && BOpcode != ISD::XOR)
14598 // The binop must be a vector type, so we can chop it in half.
14599 EVT WideBVT = BinOp.getValueType();
14600 if (!WideBVT.isVector())
14603 // Bail out if the target does not support a narrower version of the binop.
14604 EVT NarrowBVT = EVT::getVectorVT(*DAG.getContext(), WideBVT.getScalarType(),
14605 WideBVT.getVectorNumElements() / 2);
14606 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14607 if (!TLI.isOperationLegalOrCustomOrPromote(BOpcode, NarrowBVT))
14610 // Peek through bitcasts of the binary operator operands if needed.
14611 SDValue LHS = BinOp.getOperand(0);
14612 if (LHS.getOpcode() == ISD::BITCAST)
14613 LHS = LHS.getOperand(0);
14615 SDValue RHS = BinOp.getOperand(1);
14616 if (RHS.getOpcode() == ISD::BITCAST)
14617 RHS = RHS.getOperand(0);
14619 // We need at least one concatenation operation of a binop operand to make
14620 // this transform worthwhile. The concat must double the input vector sizes.
14621 // TODO: Should we also handle INSERT_SUBVECTOR patterns?
14623 LHS.getOpcode() == ISD::CONCAT_VECTORS && LHS.getNumOperands() == 2;
14625 RHS.getOpcode() == ISD::CONCAT_VECTORS && RHS.getNumOperands() == 2;
14626 if (!ConcatL && !ConcatR)
14629 // If one of the binop operands was not the result of a concat, we must
14630 // extract a half-sized operand for our new narrow binop. We can't just reuse
14631 // the original extract index operand because we may have bitcasted.
14632 unsigned ConcatOpNum = ExtractIndex->getZExtValue() / NumElems;
14633 unsigned ExtBOIdx = ConcatOpNum * NarrowBVT.getVectorNumElements();
14634 EVT ExtBOIdxVT = Extract->getOperand(1).getValueType();
14637 // extract (binop (concat X1, X2), (concat Y1, Y2)), N --> binop XN, YN
14638 // extract (binop (concat X1, X2), Y), N --> binop XN, (extract Y, N)
14639 // extract (binop X, (concat Y1, Y2)), N --> binop (extract X, N), YN
14640 SDValue X = ConcatL ? DAG.getBitcast(NarrowBVT, LHS.getOperand(ConcatOpNum))
14641 : DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, NarrowBVT,
14642 BinOp.getOperand(0),
14643 DAG.getConstant(ExtBOIdx, DL, ExtBOIdxVT));
14645 SDValue Y = ConcatR ? DAG.getBitcast(NarrowBVT, RHS.getOperand(ConcatOpNum))
14646 : DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, NarrowBVT,
14647 BinOp.getOperand(1),
14648 DAG.getConstant(ExtBOIdx, DL, ExtBOIdxVT));
14650 SDValue NarrowBinOp = DAG.getNode(BOpcode, DL, NarrowBVT, X, Y);
14651 return DAG.getBitcast(VT, NarrowBinOp);
14654 /// If we are extracting a subvector from a wide vector load, convert to a
14655 /// narrow load to eliminate the extraction:
14656 /// (extract_subvector (load wide vector)) --> (load narrow vector)
14657 static SDValue narrowExtractedVectorLoad(SDNode *Extract, SelectionDAG &DAG) {
14658 // TODO: Add support for big-endian. The offset calculation must be adjusted.
14659 if (DAG.getDataLayout().isBigEndian())
14662 // TODO: The one-use check is overly conservative. Check the cost of the
14663 // extract instead or remove that condition entirely.
14664 auto *Ld = dyn_cast<LoadSDNode>(Extract->getOperand(0));
14665 auto *ExtIdx = dyn_cast<ConstantSDNode>(Extract->getOperand(1));
14666 if (!Ld || !Ld->hasOneUse() || Ld->getExtensionType() || Ld->isVolatile() ||
14670 // The narrow load will be offset from the base address of the old load if
14671 // we are extracting from something besides index 0 (little-endian).
14672 EVT VT = Extract->getValueType(0);
14674 SDValue BaseAddr = Ld->getOperand(1);
14675 unsigned Offset = ExtIdx->getZExtValue() * VT.getScalarType().getStoreSize();
14677 // TODO: Use "BaseIndexOffset" to make this more effective.
14678 SDValue NewAddr = DAG.getMemBasePlusOffset(BaseAddr, Offset, DL);
14679 MachineFunction &MF = DAG.getMachineFunction();
14680 MachineMemOperand *MMO = MF.getMachineMemOperand(Ld->getMemOperand(), Offset,
14681 VT.getStoreSize());
14682 SDValue NewLd = DAG.getLoad(VT, DL, Ld->getChain(), NewAddr, MMO);
14684 // The new load must have the same position as the old load in terms of memory
14685 // dependency. Create a TokenFactor for Ld and NewLd and update uses of Ld's
14686 // output chain to use that TokenFactor.
14687 // TODO: This code is based on a similar sequence in x86 lowering. It should
14688 // be moved to a helper function, so it can be shared and reused.
14689 if (Ld->hasAnyUseOfValue(1)) {
14690 SDValue OldChain = SDValue(Ld, 1);
14691 SDValue NewChain = SDValue(NewLd.getNode(), 1);
14692 SDValue TokenFactor = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
14693 OldChain, NewChain);
14694 DAG.ReplaceAllUsesOfValueWith(OldChain, TokenFactor);
14695 DAG.UpdateNodeOperands(TokenFactor.getNode(), OldChain, NewChain);
14701 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
14702 EVT NVT = N->getValueType(0);
14703 SDValue V = N->getOperand(0);
14705 // Extract from UNDEF is UNDEF.
14707 return DAG.getUNDEF(NVT);
14709 if (TLI.isOperationLegalOrCustomOrPromote(ISD::LOAD, NVT))
14710 if (SDValue NarrowLoad = narrowExtractedVectorLoad(N, DAG))
14714 // (extract_subvec (concat V1, V2, ...), i)
14717 // Only operand 0 is checked as 'concat' assumes all inputs of the same
14719 if (V->getOpcode() == ISD::CONCAT_VECTORS &&
14720 isa<ConstantSDNode>(N->getOperand(1)) &&
14721 V->getOperand(0).getValueType() == NVT) {
14722 unsigned Idx = N->getConstantOperandVal(1);
14723 unsigned NumElems = NVT.getVectorNumElements();
14724 assert((Idx % NumElems) == 0 &&
14725 "IDX in concat is not a multiple of the result vector length.");
14726 return V->getOperand(Idx / NumElems);
14730 if (V->getOpcode() == ISD::BITCAST)
14731 V = V.getOperand(0);
14733 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
14734 // Handle only simple case where vector being inserted and vector
14735 // being extracted are of same size.
14736 EVT SmallVT = V->getOperand(1).getValueType();
14737 if (!NVT.bitsEq(SmallVT))
14740 // Only handle cases where both indexes are constants.
14741 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
14742 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
14744 if (InsIdx && ExtIdx) {
14746 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
14748 // indices are equal or bit offsets are equal => V1
14749 // otherwise => (extract_subvec V1, ExtIdx)
14750 if (InsIdx->getZExtValue() * SmallVT.getScalarSizeInBits() ==
14751 ExtIdx->getZExtValue() * NVT.getScalarSizeInBits())
14752 return DAG.getBitcast(NVT, V->getOperand(1));
14753 return DAG.getNode(
14754 ISD::EXTRACT_SUBVECTOR, SDLoc(N), NVT,
14755 DAG.getBitcast(N->getOperand(0).getValueType(), V->getOperand(0)),
14760 if (SDValue NarrowBOp = narrowExtractedVectorBinOp(N, DAG))
14766 static SDValue simplifyShuffleOperandRecursively(SmallBitVector &UsedElements,
14767 SDValue V, SelectionDAG &DAG) {
14769 EVT VT = V.getValueType();
14771 switch (V.getOpcode()) {
14775 case ISD::CONCAT_VECTORS: {
14776 EVT OpVT = V->getOperand(0).getValueType();
14777 int OpSize = OpVT.getVectorNumElements();
14778 SmallBitVector OpUsedElements(OpSize, false);
14779 bool FoundSimplification = false;
14780 SmallVector<SDValue, 4> NewOps;
14781 NewOps.reserve(V->getNumOperands());
14782 for (int i = 0, NumOps = V->getNumOperands(); i < NumOps; ++i) {
14783 SDValue Op = V->getOperand(i);
14784 bool OpUsed = false;
14785 for (int j = 0; j < OpSize; ++j)
14786 if (UsedElements[i * OpSize + j]) {
14787 OpUsedElements[j] = true;
14791 OpUsed ? simplifyShuffleOperandRecursively(OpUsedElements, Op, DAG)
14792 : DAG.getUNDEF(OpVT));
14793 FoundSimplification |= Op == NewOps.back();
14794 OpUsedElements.reset();
14796 if (FoundSimplification)
14797 V = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, NewOps);
14801 case ISD::INSERT_SUBVECTOR: {
14802 SDValue BaseV = V->getOperand(0);
14803 SDValue SubV = V->getOperand(1);
14804 auto *IdxN = dyn_cast<ConstantSDNode>(V->getOperand(2));
14808 int SubSize = SubV.getValueType().getVectorNumElements();
14809 int Idx = IdxN->getZExtValue();
14810 bool SubVectorUsed = false;
14811 SmallBitVector SubUsedElements(SubSize, false);
14812 for (int i = 0; i < SubSize; ++i)
14813 if (UsedElements[i + Idx]) {
14814 SubVectorUsed = true;
14815 SubUsedElements[i] = true;
14816 UsedElements[i + Idx] = false;
14819 // Now recurse on both the base and sub vectors.
14820 SDValue SimplifiedSubV =
14822 ? simplifyShuffleOperandRecursively(SubUsedElements, SubV, DAG)
14823 : DAG.getUNDEF(SubV.getValueType());
14824 SDValue SimplifiedBaseV = simplifyShuffleOperandRecursively(UsedElements, BaseV, DAG);
14825 if (SimplifiedSubV != SubV || SimplifiedBaseV != BaseV)
14826 V = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT,
14827 SimplifiedBaseV, SimplifiedSubV, V->getOperand(2));
14833 static SDValue simplifyShuffleOperands(ShuffleVectorSDNode *SVN, SDValue N0,
14834 SDValue N1, SelectionDAG &DAG) {
14835 EVT VT = SVN->getValueType(0);
14836 int NumElts = VT.getVectorNumElements();
14837 SmallBitVector N0UsedElements(NumElts, false), N1UsedElements(NumElts, false);
14838 for (int M : SVN->getMask())
14839 if (M >= 0 && M < NumElts)
14840 N0UsedElements[M] = true;
14841 else if (M >= NumElts)
14842 N1UsedElements[M - NumElts] = true;
14844 SDValue S0 = simplifyShuffleOperandRecursively(N0UsedElements, N0, DAG);
14845 SDValue S1 = simplifyShuffleOperandRecursively(N1UsedElements, N1, DAG);
14846 if (S0 == N0 && S1 == N1)
14849 return DAG.getVectorShuffle(VT, SDLoc(SVN), S0, S1, SVN->getMask());
14852 // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat,
14853 // or turn a shuffle of a single concat into simpler shuffle then concat.
14854 static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
14855 EVT VT = N->getValueType(0);
14856 unsigned NumElts = VT.getVectorNumElements();
14858 SDValue N0 = N->getOperand(0);
14859 SDValue N1 = N->getOperand(1);
14860 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
14862 SmallVector<SDValue, 4> Ops;
14863 EVT ConcatVT = N0.getOperand(0).getValueType();
14864 unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements();
14865 unsigned NumConcats = NumElts / NumElemsPerConcat;
14867 // Special case: shuffle(concat(A,B)) can be more efficiently represented
14868 // as concat(shuffle(A,B),UNDEF) if the shuffle doesn't set any of the high
14869 // half vector elements.
14870 if (NumElemsPerConcat * 2 == NumElts && N1.isUndef() &&
14871 std::all_of(SVN->getMask().begin() + NumElemsPerConcat,
14872 SVN->getMask().end(), [](int i) { return i == -1; })) {
14873 N0 = DAG.getVectorShuffle(ConcatVT, SDLoc(N), N0.getOperand(0), N0.getOperand(1),
14874 makeArrayRef(SVN->getMask().begin(), NumElemsPerConcat));
14875 N1 = DAG.getUNDEF(ConcatVT);
14876 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, N0, N1);
14879 // Look at every vector that's inserted. We're looking for exact
14880 // subvector-sized copies from a concatenated vector
14881 for (unsigned I = 0; I != NumConcats; ++I) {
14882 // Make sure we're dealing with a copy.
14883 unsigned Begin = I * NumElemsPerConcat;
14884 bool AllUndef = true, NoUndef = true;
14885 for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) {
14886 if (SVN->getMaskElt(J) >= 0)
14893 if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0)
14896 for (unsigned J = 1; J != NumElemsPerConcat; ++J)
14897 if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J))
14900 unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat;
14901 if (FirstElt < N0.getNumOperands())
14902 Ops.push_back(N0.getOperand(FirstElt));
14904 Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands()));
14906 } else if (AllUndef) {
14907 Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType()));
14908 } else { // Mixed with general masks and undefs, can't do optimization.
14913 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops);
14916 // Attempt to combine a shuffle of 2 inputs of 'scalar sources' -
14917 // BUILD_VECTOR or SCALAR_TO_VECTOR into a single BUILD_VECTOR.
14919 // SHUFFLE(BUILD_VECTOR(), BUILD_VECTOR()) -> BUILD_VECTOR() is always
14920 // a simplification in some sense, but it isn't appropriate in general: some
14921 // BUILD_VECTORs are substantially cheaper than others. The general case
14922 // of a BUILD_VECTOR requires inserting each element individually (or
14923 // performing the equivalent in a temporary stack variable). A BUILD_VECTOR of
14924 // all constants is a single constant pool load. A BUILD_VECTOR where each
14925 // element is identical is a splat. A BUILD_VECTOR where most of the operands
14926 // are undef lowers to a small number of element insertions.
14928 // To deal with this, we currently use a bunch of mostly arbitrary heuristics.
14929 // We don't fold shuffles where one side is a non-zero constant, and we don't
14930 // fold shuffles if the resulting BUILD_VECTOR would have duplicate
14931 // non-constant operands. This seems to work out reasonably well in practice.
14932 static SDValue combineShuffleOfScalars(ShuffleVectorSDNode *SVN,
14934 const TargetLowering &TLI) {
14935 EVT VT = SVN->getValueType(0);
14936 unsigned NumElts = VT.getVectorNumElements();
14937 SDValue N0 = SVN->getOperand(0);
14938 SDValue N1 = SVN->getOperand(1);
14940 if (!N0->hasOneUse() || !N1->hasOneUse())
14942 // If only one of N1,N2 is constant, bail out if it is not ALL_ZEROS as
14943 // discussed above.
14944 if (!N1.isUndef()) {
14945 bool N0AnyConst = isAnyConstantBuildVector(N0.getNode());
14946 bool N1AnyConst = isAnyConstantBuildVector(N1.getNode());
14947 if (N0AnyConst && !N1AnyConst && !ISD::isBuildVectorAllZeros(N0.getNode()))
14949 if (!N0AnyConst && N1AnyConst && !ISD::isBuildVectorAllZeros(N1.getNode()))
14953 SmallVector<SDValue, 8> Ops;
14954 SmallSet<SDValue, 16> DuplicateOps;
14955 for (int M : SVN->getMask()) {
14956 SDValue Op = DAG.getUNDEF(VT.getScalarType());
14958 int Idx = M < (int)NumElts ? M : M - NumElts;
14959 SDValue &S = (M < (int)NumElts ? N0 : N1);
14960 if (S.getOpcode() == ISD::BUILD_VECTOR) {
14961 Op = S.getOperand(Idx);
14962 } else if (S.getOpcode() == ISD::SCALAR_TO_VECTOR) {
14964 Op = S.getOperand(0);
14966 // Operand can't be combined - bail out.
14971 // Don't duplicate a non-constant BUILD_VECTOR operand; semantically, this is
14972 // fine, but it's likely to generate low-quality code if the target can't
14973 // reconstruct an appropriate shuffle.
14974 if (!Op.isUndef() && !isa<ConstantSDNode>(Op) && !isa<ConstantFPSDNode>(Op))
14975 if (!DuplicateOps.insert(Op).second)
14980 // BUILD_VECTOR requires all inputs to be of the same type, find the
14981 // maximum type and extend them all.
14982 EVT SVT = VT.getScalarType();
14983 if (SVT.isInteger())
14984 for (SDValue &Op : Ops)
14985 SVT = (SVT.bitsLT(Op.getValueType()) ? Op.getValueType() : SVT);
14986 if (SVT != VT.getScalarType())
14987 for (SDValue &Op : Ops)
14988 Op = TLI.isZExtFree(Op.getValueType(), SVT)
14989 ? DAG.getZExtOrTrunc(Op, SDLoc(SVN), SVT)
14990 : DAG.getSExtOrTrunc(Op, SDLoc(SVN), SVT);
14991 return DAG.getBuildVector(VT, SDLoc(SVN), Ops);
14994 // Match shuffles that can be converted to any_vector_extend_in_reg.
14995 // This is often generated during legalization.
14996 // e.g. v4i32 <0,u,1,u> -> (v2i64 any_vector_extend_in_reg(v4i32 src))
14997 // TODO Add support for ZERO_EXTEND_VECTOR_INREG when we have a test case.
14998 static SDValue combineShuffleToVectorExtend(ShuffleVectorSDNode *SVN,
15000 const TargetLowering &TLI,
15001 bool LegalOperations) {
15002 EVT VT = SVN->getValueType(0);
15003 bool IsBigEndian = DAG.getDataLayout().isBigEndian();
15005 // TODO Add support for big-endian when we have a test case.
15006 if (!VT.isInteger() || IsBigEndian)
15009 unsigned NumElts = VT.getVectorNumElements();
15010 unsigned EltSizeInBits = VT.getScalarSizeInBits();
15011 ArrayRef<int> Mask = SVN->getMask();
15012 SDValue N0 = SVN->getOperand(0);
15014 // shuffle<0,-1,1,-1> == (v2i64 anyextend_vector_inreg(v4i32))
15015 auto isAnyExtend = [&Mask, &NumElts](unsigned Scale) {
15016 for (unsigned i = 0; i != NumElts; ++i) {
15019 if ((i % Scale) == 0 && Mask[i] == (int)(i / Scale))
15026 // Attempt to match a '*_extend_vector_inreg' shuffle, we just search for
15027 // power-of-2 extensions as they are the most likely.
15028 for (unsigned Scale = 2; Scale < NumElts; Scale *= 2) {
15029 if (!isAnyExtend(Scale))
15032 EVT OutSVT = EVT::getIntegerVT(*DAG.getContext(), EltSizeInBits * Scale);
15033 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), OutSVT, NumElts / Scale);
15034 if (!LegalOperations ||
15035 TLI.isOperationLegalOrCustom(ISD::ANY_EXTEND_VECTOR_INREG, OutVT))
15036 return DAG.getBitcast(VT,
15037 DAG.getAnyExtendVectorInReg(N0, SDLoc(SVN), OutVT));
15043 // Detect 'truncate_vector_inreg' style shuffles that pack the lower parts of
15044 // each source element of a large type into the lowest elements of a smaller
15045 // destination type. This is often generated during legalization.
15046 // If the source node itself was a '*_extend_vector_inreg' node then we should
15047 // then be able to remove it.
15048 static SDValue combineTruncationShuffle(ShuffleVectorSDNode *SVN,
15049 SelectionDAG &DAG) {
15050 EVT VT = SVN->getValueType(0);
15051 bool IsBigEndian = DAG.getDataLayout().isBigEndian();
15053 // TODO Add support for big-endian when we have a test case.
15054 if (!VT.isInteger() || IsBigEndian)
15057 SDValue N0 = SVN->getOperand(0);
15058 while (N0.getOpcode() == ISD::BITCAST)
15059 N0 = N0.getOperand(0);
15061 unsigned Opcode = N0.getOpcode();
15062 if (Opcode != ISD::ANY_EXTEND_VECTOR_INREG &&
15063 Opcode != ISD::SIGN_EXTEND_VECTOR_INREG &&
15064 Opcode != ISD::ZERO_EXTEND_VECTOR_INREG)
15067 SDValue N00 = N0.getOperand(0);
15068 ArrayRef<int> Mask = SVN->getMask();
15069 unsigned NumElts = VT.getVectorNumElements();
15070 unsigned EltSizeInBits = VT.getScalarSizeInBits();
15071 unsigned ExtSrcSizeInBits = N00.getScalarValueSizeInBits();
15073 // (v4i32 truncate_vector_inreg(v2i64)) == shuffle<0,2-1,-1>
15074 // (v8i16 truncate_vector_inreg(v4i32)) == shuffle<0,2,4,6,-1,-1,-1,-1>
15075 // (v8i16 truncate_vector_inreg(v2i64)) == shuffle<0,4,-1,-1,-1,-1,-1,-1>
15076 auto isTruncate = [&Mask, &NumElts](unsigned Scale) {
15077 for (unsigned i = 0; i != NumElts; ++i) {
15080 if ((i * Scale) < NumElts && Mask[i] == (int)(i * Scale))
15087 // At the moment we just handle the case where we've truncated back to the
15088 // same size as before the extension.
15089 // TODO: handle more extension/truncation cases as cases arise.
15090 if (EltSizeInBits != ExtSrcSizeInBits)
15093 // Attempt to match a 'truncate_vector_inreg' shuffle, we just search for
15094 // power-of-2 truncations as they are the most likely.
15095 for (unsigned Scale = 2; Scale < NumElts; Scale *= 2)
15096 if (isTruncate(Scale))
15097 return DAG.getBitcast(VT, N00);
15102 // Combine shuffles of splat-shuffles of the form:
15103 // shuffle (shuffle V, undef, splat-mask), undef, M
15104 // If splat-mask contains undef elements, we need to be careful about
15105 // introducing undef's in the folded mask which are not the result of composing
15106 // the masks of the shuffles.
15107 static SDValue combineShuffleOfSplat(ArrayRef<int> UserMask,
15108 ShuffleVectorSDNode *Splat,
15109 SelectionDAG &DAG) {
15110 ArrayRef<int> SplatMask = Splat->getMask();
15111 assert(UserMask.size() == SplatMask.size() && "Mask length mismatch");
15113 // Prefer simplifying to the splat-shuffle, if possible. This is legal if
15114 // every undef mask element in the splat-shuffle has a corresponding undef
15115 // element in the user-shuffle's mask or if the composition of mask elements
15116 // would result in undef.
15117 // Examples for (shuffle (shuffle v, undef, SplatMask), undef, UserMask):
15118 // * UserMask=[0,2,u,u], SplatMask=[2,u,2,u] -> [2,2,u,u]
15119 // In this case it is not legal to simplify to the splat-shuffle because we
15120 // may be exposing the users of the shuffle an undef element at index 1
15121 // which was not there before the combine.
15122 // * UserMask=[0,u,2,u], SplatMask=[2,u,2,u] -> [2,u,2,u]
15123 // In this case the composition of masks yields SplatMask, so it's ok to
15124 // simplify to the splat-shuffle.
15125 // * UserMask=[3,u,2,u], SplatMask=[2,u,2,u] -> [u,u,2,u]
15126 // In this case the composed mask includes all undef elements of SplatMask
15127 // and in addition sets element zero to undef. It is safe to simplify to
15128 // the splat-shuffle.
15129 auto CanSimplifyToExistingSplat = [](ArrayRef<int> UserMask,
15130 ArrayRef<int> SplatMask) {
15131 for (unsigned i = 0, e = UserMask.size(); i != e; ++i)
15132 if (UserMask[i] != -1 && SplatMask[i] == -1 &&
15133 SplatMask[UserMask[i]] != -1)
15137 if (CanSimplifyToExistingSplat(UserMask, SplatMask))
15138 return SDValue(Splat, 0);
15140 // Create a new shuffle with a mask that is composed of the two shuffles'
15142 SmallVector<int, 32> NewMask;
15143 for (int Idx : UserMask)
15144 NewMask.push_back(Idx == -1 ? -1 : SplatMask[Idx]);
15146 return DAG.getVectorShuffle(Splat->getValueType(0), SDLoc(Splat),
15147 Splat->getOperand(0), Splat->getOperand(1),
15151 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
15152 EVT VT = N->getValueType(0);
15153 unsigned NumElts = VT.getVectorNumElements();
15155 SDValue N0 = N->getOperand(0);
15156 SDValue N1 = N->getOperand(1);
15158 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
15160 // Canonicalize shuffle undef, undef -> undef
15161 if (N0.isUndef() && N1.isUndef())
15162 return DAG.getUNDEF(VT);
15164 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
15166 // Canonicalize shuffle v, v -> v, undef
15168 SmallVector<int, 8> NewMask;
15169 for (unsigned i = 0; i != NumElts; ++i) {
15170 int Idx = SVN->getMaskElt(i);
15171 if (Idx >= (int)NumElts) Idx -= NumElts;
15172 NewMask.push_back(Idx);
15174 return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT), NewMask);
15177 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
15179 return DAG.getCommutedVectorShuffle(*SVN);
15181 // Remove references to rhs if it is undef
15182 if (N1.isUndef()) {
15183 bool Changed = false;
15184 SmallVector<int, 8> NewMask;
15185 for (unsigned i = 0; i != NumElts; ++i) {
15186 int Idx = SVN->getMaskElt(i);
15187 if (Idx >= (int)NumElts) {
15191 NewMask.push_back(Idx);
15194 return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, NewMask);
15197 // A shuffle of a single vector that is a splat can always be folded.
15198 if (auto *N0Shuf = dyn_cast<ShuffleVectorSDNode>(N0))
15199 if (N1->isUndef() && N0Shuf->isSplat())
15200 return combineShuffleOfSplat(SVN->getMask(), N0Shuf, DAG);
15202 // If it is a splat, check if the argument vector is another splat or a
15204 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
15205 SDNode *V = N0.getNode();
15207 // If this is a bit convert that changes the element type of the vector but
15208 // not the number of vector elements, look through it. Be careful not to
15209 // look though conversions that change things like v4f32 to v2f64.
15210 if (V->getOpcode() == ISD::BITCAST) {
15211 SDValue ConvInput = V->getOperand(0);
15212 if (ConvInput.getValueType().isVector() &&
15213 ConvInput.getValueType().getVectorNumElements() == NumElts)
15214 V = ConvInput.getNode();
15217 if (V->getOpcode() == ISD::BUILD_VECTOR) {
15218 assert(V->getNumOperands() == NumElts &&
15219 "BUILD_VECTOR has wrong number of operands");
15221 bool AllSame = true;
15222 for (unsigned i = 0; i != NumElts; ++i) {
15223 if (!V->getOperand(i).isUndef()) {
15224 Base = V->getOperand(i);
15228 // Splat of <u, u, u, u>, return <u, u, u, u>
15229 if (!Base.getNode())
15231 for (unsigned i = 0; i != NumElts; ++i) {
15232 if (V->getOperand(i) != Base) {
15237 // Splat of <x, x, x, x>, return <x, x, x, x>
15241 // Canonicalize any other splat as a build_vector.
15242 const SDValue &Splatted = V->getOperand(SVN->getSplatIndex());
15243 SmallVector<SDValue, 8> Ops(NumElts, Splatted);
15244 SDValue NewBV = DAG.getBuildVector(V->getValueType(0), SDLoc(N), Ops);
15246 // We may have jumped through bitcasts, so the type of the
15247 // BUILD_VECTOR may not match the type of the shuffle.
15248 if (V->getValueType(0) != VT)
15249 NewBV = DAG.getBitcast(VT, NewBV);
15254 // There are various patterns used to build up a vector from smaller vectors,
15255 // subvectors, or elements. Scan chains of these and replace unused insertions
15256 // or components with undef.
15257 if (SDValue S = simplifyShuffleOperands(SVN, N0, N1, DAG))
15260 // Match shuffles that can be converted to any_vector_extend_in_reg.
15261 if (SDValue V = combineShuffleToVectorExtend(SVN, DAG, TLI, LegalOperations))
15264 // Combine "truncate_vector_in_reg" style shuffles.
15265 if (SDValue V = combineTruncationShuffle(SVN, DAG))
15268 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
15269 Level < AfterLegalizeVectorOps &&
15271 (N1.getOpcode() == ISD::CONCAT_VECTORS &&
15272 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) {
15273 if (SDValue V = partitionShuffleOfConcats(N, DAG))
15277 // Attempt to combine a shuffle of 2 inputs of 'scalar sources' -
15278 // BUILD_VECTOR or SCALAR_TO_VECTOR into a single BUILD_VECTOR.
15279 if (Level < AfterLegalizeVectorOps && TLI.isTypeLegal(VT))
15280 if (SDValue Res = combineShuffleOfScalars(SVN, DAG, TLI))
15283 // If this shuffle only has a single input that is a bitcasted shuffle,
15284 // attempt to merge the 2 shuffles and suitably bitcast the inputs/output
15285 // back to their original types.
15286 if (N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
15287 N1.isUndef() && Level < AfterLegalizeVectorOps &&
15288 TLI.isTypeLegal(VT)) {
15290 // Peek through the bitcast only if there is one user.
15292 while (BC0.getOpcode() == ISD::BITCAST) {
15293 if (!BC0.hasOneUse())
15295 BC0 = BC0.getOperand(0);
15298 auto ScaleShuffleMask = [](ArrayRef<int> Mask, int Scale) {
15300 return SmallVector<int, 8>(Mask.begin(), Mask.end());
15302 SmallVector<int, 8> NewMask;
15304 for (int s = 0; s != Scale; ++s)
15305 NewMask.push_back(M < 0 ? -1 : Scale * M + s);
15309 if (BC0.getOpcode() == ISD::VECTOR_SHUFFLE && BC0.hasOneUse()) {
15310 EVT SVT = VT.getScalarType();
15311 EVT InnerVT = BC0->getValueType(0);
15312 EVT InnerSVT = InnerVT.getScalarType();
15314 // Determine which shuffle works with the smaller scalar type.
15315 EVT ScaleVT = SVT.bitsLT(InnerSVT) ? VT : InnerVT;
15316 EVT ScaleSVT = ScaleVT.getScalarType();
15318 if (TLI.isTypeLegal(ScaleVT) &&
15319 0 == (InnerSVT.getSizeInBits() % ScaleSVT.getSizeInBits()) &&
15320 0 == (SVT.getSizeInBits() % ScaleSVT.getSizeInBits())) {
15322 int InnerScale = InnerSVT.getSizeInBits() / ScaleSVT.getSizeInBits();
15323 int OuterScale = SVT.getSizeInBits() / ScaleSVT.getSizeInBits();
15325 // Scale the shuffle masks to the smaller scalar type.
15326 ShuffleVectorSDNode *InnerSVN = cast<ShuffleVectorSDNode>(BC0);
15327 SmallVector<int, 8> InnerMask =
15328 ScaleShuffleMask(InnerSVN->getMask(), InnerScale);
15329 SmallVector<int, 8> OuterMask =
15330 ScaleShuffleMask(SVN->getMask(), OuterScale);
15332 // Merge the shuffle masks.
15333 SmallVector<int, 8> NewMask;
15334 for (int M : OuterMask)
15335 NewMask.push_back(M < 0 ? -1 : InnerMask[M]);
15337 // Test for shuffle mask legality over both commutations.
15338 SDValue SV0 = BC0->getOperand(0);
15339 SDValue SV1 = BC0->getOperand(1);
15340 bool LegalMask = TLI.isShuffleMaskLegal(NewMask, ScaleVT);
15342 std::swap(SV0, SV1);
15343 ShuffleVectorSDNode::commuteMask(NewMask);
15344 LegalMask = TLI.isShuffleMaskLegal(NewMask, ScaleVT);
15348 SV0 = DAG.getBitcast(ScaleVT, SV0);
15349 SV1 = DAG.getBitcast(ScaleVT, SV1);
15350 return DAG.getBitcast(
15351 VT, DAG.getVectorShuffle(ScaleVT, SDLoc(N), SV0, SV1, NewMask));
15357 // Canonicalize shuffles according to rules:
15358 // shuffle(A, shuffle(A, B)) -> shuffle(shuffle(A,B), A)
15359 // shuffle(B, shuffle(A, B)) -> shuffle(shuffle(A,B), B)
15360 // shuffle(B, shuffle(A, Undef)) -> shuffle(shuffle(A, Undef), B)
15361 if (N1.getOpcode() == ISD::VECTOR_SHUFFLE &&
15362 N0.getOpcode() != ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
15363 TLI.isTypeLegal(VT)) {
15364 // The incoming shuffle must be of the same type as the result of the
15365 // current shuffle.
15366 assert(N1->getOperand(0).getValueType() == VT &&
15367 "Shuffle types don't match");
15369 SDValue SV0 = N1->getOperand(0);
15370 SDValue SV1 = N1->getOperand(1);
15371 bool HasSameOp0 = N0 == SV0;
15372 bool IsSV1Undef = SV1.isUndef();
15373 if (HasSameOp0 || IsSV1Undef || N0 == SV1)
15374 // Commute the operands of this shuffle so that next rule
15376 return DAG.getCommutedVectorShuffle(*SVN);
15379 // Try to fold according to rules:
15380 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2)
15381 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2)
15382 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2)
15383 // Don't try to fold shuffles with illegal type.
15384 // Only fold if this shuffle is the only user of the other shuffle.
15385 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && N->isOnlyUserOf(N0.getNode()) &&
15386 Level < AfterLegalizeDAG && TLI.isTypeLegal(VT)) {
15387 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
15389 // Don't try to fold splats; they're likely to simplify somehow, or they
15391 if (OtherSV->isSplat())
15394 // The incoming shuffle must be of the same type as the result of the
15395 // current shuffle.
15396 assert(OtherSV->getOperand(0).getValueType() == VT &&
15397 "Shuffle types don't match");
15400 SmallVector<int, 4> Mask;
15401 // Compute the combined shuffle mask for a shuffle with SV0 as the first
15402 // operand, and SV1 as the second operand.
15403 for (unsigned i = 0; i != NumElts; ++i) {
15404 int Idx = SVN->getMaskElt(i);
15406 // Propagate Undef.
15407 Mask.push_back(Idx);
15411 SDValue CurrentVec;
15412 if (Idx < (int)NumElts) {
15413 // This shuffle index refers to the inner shuffle N0. Lookup the inner
15414 // shuffle mask to identify which vector is actually referenced.
15415 Idx = OtherSV->getMaskElt(Idx);
15417 // Propagate Undef.
15418 Mask.push_back(Idx);
15422 CurrentVec = (Idx < (int) NumElts) ? OtherSV->getOperand(0)
15423 : OtherSV->getOperand(1);
15425 // This shuffle index references an element within N1.
15429 // Simple case where 'CurrentVec' is UNDEF.
15430 if (CurrentVec.isUndef()) {
15431 Mask.push_back(-1);
15435 // Canonicalize the shuffle index. We don't know yet if CurrentVec
15436 // will be the first or second operand of the combined shuffle.
15437 Idx = Idx % NumElts;
15438 if (!SV0.getNode() || SV0 == CurrentVec) {
15439 // Ok. CurrentVec is the left hand side.
15440 // Update the mask accordingly.
15442 Mask.push_back(Idx);
15446 // Bail out if we cannot convert the shuffle pair into a single shuffle.
15447 if (SV1.getNode() && SV1 != CurrentVec)
15450 // Ok. CurrentVec is the right hand side.
15451 // Update the mask accordingly.
15453 Mask.push_back(Idx + NumElts);
15456 // Check if all indices in Mask are Undef. In case, propagate Undef.
15457 bool isUndefMask = true;
15458 for (unsigned i = 0; i != NumElts && isUndefMask; ++i)
15459 isUndefMask &= Mask[i] < 0;
15462 return DAG.getUNDEF(VT);
15464 if (!SV0.getNode())
15465 SV0 = DAG.getUNDEF(VT);
15466 if (!SV1.getNode())
15467 SV1 = DAG.getUNDEF(VT);
15469 // Avoid introducing shuffles with illegal mask.
15470 if (!TLI.isShuffleMaskLegal(Mask, VT)) {
15471 ShuffleVectorSDNode::commuteMask(Mask);
15473 if (!TLI.isShuffleMaskLegal(Mask, VT))
15476 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, A, M2)
15477 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, A, M2)
15478 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, B, M2)
15479 std::swap(SV0, SV1);
15482 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2)
15483 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2)
15484 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2)
15485 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, Mask);
15491 SDValue DAGCombiner::visitSCALAR_TO_VECTOR(SDNode *N) {
15492 SDValue InVal = N->getOperand(0);
15493 EVT VT = N->getValueType(0);
15495 // Replace a SCALAR_TO_VECTOR(EXTRACT_VECTOR_ELT(V,C0)) pattern
15496 // with a VECTOR_SHUFFLE.
15497 if (InVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
15498 SDValue InVec = InVal->getOperand(0);
15499 SDValue EltNo = InVal->getOperand(1);
15501 // FIXME: We could support implicit truncation if the shuffle can be
15502 // scaled to a smaller vector scalar type.
15503 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(EltNo);
15504 if (C0 && VT == InVec.getValueType() &&
15505 VT.getScalarType() == InVal.getValueType()) {
15506 SmallVector<int, 8> NewMask(VT.getVectorNumElements(), -1);
15507 int Elt = C0->getZExtValue();
15510 if (TLI.isShuffleMaskLegal(NewMask, VT))
15511 return DAG.getVectorShuffle(VT, SDLoc(N), InVec, DAG.getUNDEF(VT),
15519 SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
15520 EVT VT = N->getValueType(0);
15521 SDValue N0 = N->getOperand(0);
15522 SDValue N1 = N->getOperand(1);
15523 SDValue N2 = N->getOperand(2);
15525 // If inserting an UNDEF, just return the original vector.
15529 // If this is an insert of an extracted vector into an undef vector, we can
15530 // just use the input to the extract.
15531 if (N0.isUndef() && N1.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
15532 N1.getOperand(1) == N2 && N1.getOperand(0).getValueType() == VT)
15533 return N1.getOperand(0);
15535 // Combine INSERT_SUBVECTORs where we are inserting to the same index.
15536 // INSERT_SUBVECTOR( INSERT_SUBVECTOR( Vec, SubOld, Idx ), SubNew, Idx )
15537 // --> INSERT_SUBVECTOR( Vec, SubNew, Idx )
15538 if (N0.getOpcode() == ISD::INSERT_SUBVECTOR &&
15539 N0.getOperand(1).getValueType() == N1.getValueType() &&
15540 N0.getOperand(2) == N2)
15541 return DAG.getNode(ISD::INSERT_SUBVECTOR, SDLoc(N), VT, N0.getOperand(0),
15544 if (!isa<ConstantSDNode>(N2))
15547 unsigned InsIdx = cast<ConstantSDNode>(N2)->getZExtValue();
15549 // Canonicalize insert_subvector dag nodes.
15551 // (insert_subvector (insert_subvector A, Idx0), Idx1)
15552 // -> (insert_subvector (insert_subvector A, Idx1), Idx0)
15553 if (N0.getOpcode() == ISD::INSERT_SUBVECTOR && N0.hasOneUse() &&
15554 N1.getValueType() == N0.getOperand(1).getValueType() &&
15555 isa<ConstantSDNode>(N0.getOperand(2))) {
15556 unsigned OtherIdx = N0.getConstantOperandVal(2);
15557 if (InsIdx < OtherIdx) {
15559 SDValue NewOp = DAG.getNode(ISD::INSERT_SUBVECTOR, SDLoc(N), VT,
15560 N0.getOperand(0), N1, N2);
15561 AddToWorklist(NewOp.getNode());
15562 return DAG.getNode(ISD::INSERT_SUBVECTOR, SDLoc(N0.getNode()),
15563 VT, NewOp, N0.getOperand(1), N0.getOperand(2));
15567 // If the input vector is a concatenation, and the insert replaces
15568 // one of the pieces, we can optimize into a single concat_vectors.
15569 if (N0.getOpcode() == ISD::CONCAT_VECTORS && N0.hasOneUse() &&
15570 N0.getOperand(0).getValueType() == N1.getValueType()) {
15571 unsigned Factor = N1.getValueType().getVectorNumElements();
15573 SmallVector<SDValue, 8> Ops(N0->op_begin(), N0->op_end());
15574 Ops[cast<ConstantSDNode>(N2)->getZExtValue() / Factor] = N1;
15576 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops);
15582 SDValue DAGCombiner::visitFP_TO_FP16(SDNode *N) {
15583 SDValue N0 = N->getOperand(0);
15585 // fold (fp_to_fp16 (fp16_to_fp op)) -> op
15586 if (N0->getOpcode() == ISD::FP16_TO_FP)
15587 return N0->getOperand(0);
15592 SDValue DAGCombiner::visitFP16_TO_FP(SDNode *N) {
15593 SDValue N0 = N->getOperand(0);
15595 // fold fp16_to_fp(op & 0xffff) -> fp16_to_fp(op)
15596 if (N0->getOpcode() == ISD::AND) {
15597 ConstantSDNode *AndConst = getAsNonOpaqueConstant(N0.getOperand(1));
15598 if (AndConst && AndConst->getAPIntValue() == 0xffff) {
15599 return DAG.getNode(ISD::FP16_TO_FP, SDLoc(N), N->getValueType(0),
15607 /// Returns a vector_shuffle if it able to transform an AND to a vector_shuffle
15608 /// with the destination vector and a zero vector.
15609 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
15610 /// vector_shuffle V, Zero, <0, 4, 2, 4>
15611 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
15612 EVT VT = N->getValueType(0);
15613 SDValue LHS = N->getOperand(0);
15614 SDValue RHS = N->getOperand(1);
15617 // Make sure we're not running after operation legalization where it
15618 // may have custom lowered the vector shuffles.
15619 if (LegalOperations)
15622 if (N->getOpcode() != ISD::AND)
15625 if (RHS.getOpcode() == ISD::BITCAST)
15626 RHS = RHS.getOperand(0);
15628 if (RHS.getOpcode() != ISD::BUILD_VECTOR)
15631 EVT RVT = RHS.getValueType();
15632 unsigned NumElts = RHS.getNumOperands();
15634 // Attempt to create a valid clear mask, splitting the mask into
15635 // sub elements and checking to see if each is
15636 // all zeros or all ones - suitable for shuffle masking.
15637 auto BuildClearMask = [&](int Split) {
15638 int NumSubElts = NumElts * Split;
15639 int NumSubBits = RVT.getScalarSizeInBits() / Split;
15641 SmallVector<int, 8> Indices;
15642 for (int i = 0; i != NumSubElts; ++i) {
15643 int EltIdx = i / Split;
15644 int SubIdx = i % Split;
15645 SDValue Elt = RHS.getOperand(EltIdx);
15646 if (Elt.isUndef()) {
15647 Indices.push_back(-1);
15652 if (isa<ConstantSDNode>(Elt))
15653 Bits = cast<ConstantSDNode>(Elt)->getAPIntValue();
15654 else if (isa<ConstantFPSDNode>(Elt))
15655 Bits = cast<ConstantFPSDNode>(Elt)->getValueAPF().bitcastToAPInt();
15659 // Extract the sub element from the constant bit mask.
15660 if (DAG.getDataLayout().isBigEndian()) {
15661 Bits.lshrInPlace((Split - SubIdx - 1) * NumSubBits);
15663 Bits.lshrInPlace(SubIdx * NumSubBits);
15667 Bits = Bits.trunc(NumSubBits);
15669 if (Bits.isAllOnesValue())
15670 Indices.push_back(i);
15671 else if (Bits == 0)
15672 Indices.push_back(i + NumSubElts);
15677 // Let's see if the target supports this vector_shuffle.
15678 EVT ClearSVT = EVT::getIntegerVT(*DAG.getContext(), NumSubBits);
15679 EVT ClearVT = EVT::getVectorVT(*DAG.getContext(), ClearSVT, NumSubElts);
15680 if (!TLI.isVectorClearMaskLegal(Indices, ClearVT))
15683 SDValue Zero = DAG.getConstant(0, DL, ClearVT);
15684 return DAG.getBitcast(VT, DAG.getVectorShuffle(ClearVT, DL,
15685 DAG.getBitcast(ClearVT, LHS),
15689 // Determine maximum split level (byte level masking).
15691 if (RVT.getScalarSizeInBits() % 8 == 0)
15692 MaxSplit = RVT.getScalarSizeInBits() / 8;
15694 for (int Split = 1; Split <= MaxSplit; ++Split)
15695 if (RVT.getScalarSizeInBits() % Split == 0)
15696 if (SDValue S = BuildClearMask(Split))
15702 /// Visit a binary vector operation, like ADD.
15703 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
15704 assert(N->getValueType(0).isVector() &&
15705 "SimplifyVBinOp only works on vectors!");
15707 SDValue LHS = N->getOperand(0);
15708 SDValue RHS = N->getOperand(1);
15709 SDValue Ops[] = {LHS, RHS};
15711 // See if we can constant fold the vector operation.
15712 if (SDValue Fold = DAG.FoldConstantVectorArithmetic(
15713 N->getOpcode(), SDLoc(LHS), LHS.getValueType(), Ops, N->getFlags()))
15716 // Try to convert a constant mask AND into a shuffle clear mask.
15717 if (SDValue Shuffle = XformToShuffleWithZero(N))
15720 // Type legalization might introduce new shuffles in the DAG.
15721 // Fold (VBinOp (shuffle (A, Undef, Mask)), (shuffle (B, Undef, Mask)))
15722 // -> (shuffle (VBinOp (A, B)), Undef, Mask).
15723 if (LegalTypes && isa<ShuffleVectorSDNode>(LHS) &&
15724 isa<ShuffleVectorSDNode>(RHS) && LHS.hasOneUse() && RHS.hasOneUse() &&
15725 LHS.getOperand(1).isUndef() &&
15726 RHS.getOperand(1).isUndef()) {
15727 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(LHS);
15728 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(RHS);
15730 if (SVN0->getMask().equals(SVN1->getMask())) {
15731 EVT VT = N->getValueType(0);
15732 SDValue UndefVector = LHS.getOperand(1);
15733 SDValue NewBinOp = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
15734 LHS.getOperand(0), RHS.getOperand(0),
15736 AddUsersToWorklist(N);
15737 return DAG.getVectorShuffle(VT, SDLoc(N), NewBinOp, UndefVector,
15745 SDValue DAGCombiner::SimplifySelect(const SDLoc &DL, SDValue N0, SDValue N1,
15747 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
15749 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
15750 cast<CondCodeSDNode>(N0.getOperand(2))->get());
15752 // If we got a simplified select_cc node back from SimplifySelectCC, then
15753 // break it down into a new SETCC node, and a new SELECT node, and then return
15754 // the SELECT node, since we were called with a SELECT node.
15755 if (SCC.getNode()) {
15756 // Check to see if we got a select_cc back (to turn into setcc/select).
15757 // Otherwise, just return whatever node we got back, like fabs.
15758 if (SCC.getOpcode() == ISD::SELECT_CC) {
15759 SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0),
15761 SCC.getOperand(0), SCC.getOperand(1),
15762 SCC.getOperand(4));
15763 AddToWorklist(SETCC.getNode());
15764 return DAG.getSelect(SDLoc(SCC), SCC.getValueType(), SETCC,
15765 SCC.getOperand(2), SCC.getOperand(3));
15773 /// Given a SELECT or a SELECT_CC node, where LHS and RHS are the two values
15774 /// being selected between, see if we can simplify the select. Callers of this
15775 /// should assume that TheSelect is deleted if this returns true. As such, they
15776 /// should return the appropriate thing (e.g. the node) back to the top-level of
15777 /// the DAG combiner loop to avoid it being looked at.
15778 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
15781 // fold (select (setcc x, [+-]0.0, *lt), NaN, (fsqrt x))
15782 // The select + setcc is redundant, because fsqrt returns NaN for X < 0.
15783 if (const ConstantFPSDNode *NaN = isConstOrConstSplatFP(LHS)) {
15784 if (NaN->isNaN() && RHS.getOpcode() == ISD::FSQRT) {
15785 // We have: (select (setcc ?, ?, ?), NaN, (fsqrt ?))
15786 SDValue Sqrt = RHS;
15789 const ConstantFPSDNode *Zero = nullptr;
15791 if (TheSelect->getOpcode() == ISD::SELECT_CC) {
15792 CC = dyn_cast<CondCodeSDNode>(TheSelect->getOperand(4))->get();
15793 CmpLHS = TheSelect->getOperand(0);
15794 Zero = isConstOrConstSplatFP(TheSelect->getOperand(1));
15796 // SELECT or VSELECT
15797 SDValue Cmp = TheSelect->getOperand(0);
15798 if (Cmp.getOpcode() == ISD::SETCC) {
15799 CC = dyn_cast<CondCodeSDNode>(Cmp.getOperand(2))->get();
15800 CmpLHS = Cmp.getOperand(0);
15801 Zero = isConstOrConstSplatFP(Cmp.getOperand(1));
15804 if (Zero && Zero->isZero() &&
15805 Sqrt.getOperand(0) == CmpLHS && (CC == ISD::SETOLT ||
15806 CC == ISD::SETULT || CC == ISD::SETLT)) {
15807 // We have: (select (setcc x, [+-]0.0, *lt), NaN, (fsqrt x))
15808 CombineTo(TheSelect, Sqrt);
15813 // Cannot simplify select with vector condition
15814 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
15816 // If this is a select from two identical things, try to pull the operation
15817 // through the select.
15818 if (LHS.getOpcode() != RHS.getOpcode() ||
15819 !LHS.hasOneUse() || !RHS.hasOneUse())
15822 // If this is a load and the token chain is identical, replace the select
15823 // of two loads with a load through a select of the address to load from.
15824 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
15825 // constants have been dropped into the constant pool.
15826 if (LHS.getOpcode() == ISD::LOAD) {
15827 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
15828 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
15830 // Token chains must be identical.
15831 if (LHS.getOperand(0) != RHS.getOperand(0) ||
15832 // Do not let this transformation reduce the number of volatile loads.
15833 LLD->isVolatile() || RLD->isVolatile() ||
15834 // FIXME: If either is a pre/post inc/dec load,
15835 // we'd need to split out the address adjustment.
15836 LLD->isIndexed() || RLD->isIndexed() ||
15837 // If this is an EXTLOAD, the VT's must match.
15838 LLD->getMemoryVT() != RLD->getMemoryVT() ||
15839 // If this is an EXTLOAD, the kind of extension must match.
15840 (LLD->getExtensionType() != RLD->getExtensionType() &&
15841 // The only exception is if one of the extensions is anyext.
15842 LLD->getExtensionType() != ISD::EXTLOAD &&
15843 RLD->getExtensionType() != ISD::EXTLOAD) ||
15844 // FIXME: this discards src value information. This is
15845 // over-conservative. It would be beneficial to be able to remember
15846 // both potential memory locations. Since we are discarding
15847 // src value info, don't do the transformation if the memory
15848 // locations are not in the default address space.
15849 LLD->getPointerInfo().getAddrSpace() != 0 ||
15850 RLD->getPointerInfo().getAddrSpace() != 0 ||
15851 !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(),
15852 LLD->getBasePtr().getValueType()))
15855 // Check that the select condition doesn't reach either load. If so,
15856 // folding this will induce a cycle into the DAG. If not, this is safe to
15857 // xform, so create a select of the addresses.
15859 if (TheSelect->getOpcode() == ISD::SELECT) {
15860 SDNode *CondNode = TheSelect->getOperand(0).getNode();
15861 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
15862 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
15864 // The loads must not depend on one another.
15865 if (LLD->isPredecessorOf(RLD) ||
15866 RLD->isPredecessorOf(LLD))
15868 Addr = DAG.getSelect(SDLoc(TheSelect),
15869 LLD->getBasePtr().getValueType(),
15870 TheSelect->getOperand(0), LLD->getBasePtr(),
15871 RLD->getBasePtr());
15872 } else { // Otherwise SELECT_CC
15873 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
15874 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
15876 if ((LLD->hasAnyUseOfValue(1) &&
15877 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
15878 (RLD->hasAnyUseOfValue(1) &&
15879 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
15882 Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect),
15883 LLD->getBasePtr().getValueType(),
15884 TheSelect->getOperand(0),
15885 TheSelect->getOperand(1),
15886 LLD->getBasePtr(), RLD->getBasePtr(),
15887 TheSelect->getOperand(4));
15891 // It is safe to replace the two loads if they have different alignments,
15892 // but the new load must be the minimum (most restrictive) alignment of the
15894 unsigned Alignment = std::min(LLD->getAlignment(), RLD->getAlignment());
15895 MachineMemOperand::Flags MMOFlags = LLD->getMemOperand()->getFlags();
15896 if (!RLD->isInvariant())
15897 MMOFlags &= ~MachineMemOperand::MOInvariant;
15898 if (!RLD->isDereferenceable())
15899 MMOFlags &= ~MachineMemOperand::MODereferenceable;
15900 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
15901 // FIXME: Discards pointer and AA info.
15902 Load = DAG.getLoad(TheSelect->getValueType(0), SDLoc(TheSelect),
15903 LLD->getChain(), Addr, MachinePointerInfo(), Alignment,
15906 // FIXME: Discards pointer and AA info.
15907 Load = DAG.getExtLoad(
15908 LLD->getExtensionType() == ISD::EXTLOAD ? RLD->getExtensionType()
15909 : LLD->getExtensionType(),
15910 SDLoc(TheSelect), TheSelect->getValueType(0), LLD->getChain(), Addr,
15911 MachinePointerInfo(), LLD->getMemoryVT(), Alignment, MMOFlags);
15914 // Users of the select now use the result of the load.
15915 CombineTo(TheSelect, Load);
15917 // Users of the old loads now use the new load's chain. We know the
15918 // old-load value is dead now.
15919 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
15920 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
15927 /// Try to fold an expression of the form (N0 cond N1) ? N2 : N3 to a shift and
15929 SDValue DAGCombiner::foldSelectCCToShiftAnd(const SDLoc &DL, SDValue N0,
15930 SDValue N1, SDValue N2, SDValue N3,
15931 ISD::CondCode CC) {
15932 // If this is a select where the false operand is zero and the compare is a
15933 // check of the sign bit, see if we can perform the "gzip trick":
15934 // select_cc setlt X, 0, A, 0 -> and (sra X, size(X)-1), A
15935 // select_cc setgt X, 0, A, 0 -> and (not (sra X, size(X)-1)), A
15936 EVT XType = N0.getValueType();
15937 EVT AType = N2.getValueType();
15938 if (!isNullConstant(N3) || !XType.bitsGE(AType))
15941 // If the comparison is testing for a positive value, we have to invert
15942 // the sign bit mask, so only do that transform if the target has a bitwise
15943 // 'and not' instruction (the invert is free).
15944 if (CC == ISD::SETGT && TLI.hasAndNot(N2)) {
15945 // (X > -1) ? A : 0
15946 // (X > 0) ? X : 0 <-- This is canonical signed max.
15947 if (!(isAllOnesConstant(N1) || (isNullConstant(N1) && N0 == N2)))
15949 } else if (CC == ISD::SETLT) {
15951 // (X < 1) ? X : 0 <-- This is un-canonicalized signed min.
15952 if (!(isNullConstant(N1) || (isOneConstant(N1) && N0 == N2)))
15958 // and (sra X, size(X)-1), A -> "and (srl X, C2), A" iff A is a single-bit
15960 EVT ShiftAmtTy = getShiftAmountTy(N0.getValueType());
15961 auto *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
15962 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue() - 1)) == 0)) {
15963 unsigned ShCt = XType.getSizeInBits() - N2C->getAPIntValue().logBase2() - 1;
15964 SDValue ShiftAmt = DAG.getConstant(ShCt, DL, ShiftAmtTy);
15965 SDValue Shift = DAG.getNode(ISD::SRL, DL, XType, N0, ShiftAmt);
15966 AddToWorklist(Shift.getNode());
15968 if (XType.bitsGT(AType)) {
15969 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
15970 AddToWorklist(Shift.getNode());
15973 if (CC == ISD::SETGT)
15974 Shift = DAG.getNOT(DL, Shift, AType);
15976 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
15979 SDValue ShiftAmt = DAG.getConstant(XType.getSizeInBits() - 1, DL, ShiftAmtTy);
15980 SDValue Shift = DAG.getNode(ISD::SRA, DL, XType, N0, ShiftAmt);
15981 AddToWorklist(Shift.getNode());
15983 if (XType.bitsGT(AType)) {
15984 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
15985 AddToWorklist(Shift.getNode());
15988 if (CC == ISD::SETGT)
15989 Shift = DAG.getNOT(DL, Shift, AType);
15991 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
15994 /// Simplify an expression of the form (N0 cond N1) ? N2 : N3
15995 /// where 'cond' is the comparison specified by CC.
15996 SDValue DAGCombiner::SimplifySelectCC(const SDLoc &DL, SDValue N0, SDValue N1,
15997 SDValue N2, SDValue N3, ISD::CondCode CC,
15998 bool NotExtCompare) {
15999 // (x ? y : y) -> y.
16000 if (N2 == N3) return N2;
16002 EVT VT = N2.getValueType();
16003 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
16004 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
16006 // Determine if the condition we're dealing with is constant
16007 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
16008 N0, N1, CC, DL, false);
16009 if (SCC.getNode()) AddToWorklist(SCC.getNode());
16011 if (ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode())) {
16012 // fold select_cc true, x, y -> x
16013 // fold select_cc false, x, y -> y
16014 return !SCCC->isNullValue() ? N2 : N3;
16017 // Check to see if we can simplify the select into an fabs node
16018 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
16019 // Allow either -0.0 or 0.0
16020 if (CFP->isZero()) {
16021 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
16022 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
16023 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
16024 N2 == N3.getOperand(0))
16025 return DAG.getNode(ISD::FABS, DL, VT, N0);
16027 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
16028 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
16029 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
16030 N2.getOperand(0) == N3)
16031 return DAG.getNode(ISD::FABS, DL, VT, N3);
16035 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
16036 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
16037 // in it. This is a win when the constant is not otherwise available because
16038 // it replaces two constant pool loads with one. We only do this if the FP
16039 // type is known to be legal, because if it isn't, then we are before legalize
16040 // types an we want the other legalization to happen first (e.g. to avoid
16041 // messing with soft float) and if the ConstantFP is not legal, because if
16042 // it is legal, we may not need to store the FP constant in a constant pool.
16043 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
16044 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
16045 if (TLI.isTypeLegal(N2.getValueType()) &&
16046 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
16047 TargetLowering::Legal &&
16048 !TLI.isFPImmLegal(TV->getValueAPF(), TV->getValueType(0)) &&
16049 !TLI.isFPImmLegal(FV->getValueAPF(), FV->getValueType(0))) &&
16050 // If both constants have multiple uses, then we won't need to do an
16051 // extra load, they are likely around in registers for other users.
16052 (TV->hasOneUse() || FV->hasOneUse())) {
16053 Constant *Elts[] = {
16054 const_cast<ConstantFP*>(FV->getConstantFPValue()),
16055 const_cast<ConstantFP*>(TV->getConstantFPValue())
16057 Type *FPTy = Elts[0]->getType();
16058 const DataLayout &TD = DAG.getDataLayout();
16060 // Create a ConstantArray of the two constants.
16061 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
16063 DAG.getConstantPool(CA, TLI.getPointerTy(DAG.getDataLayout()),
16064 TD.getPrefTypeAlignment(FPTy));
16065 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
16067 // Get the offsets to the 0 and 1 element of the array so that we can
16068 // select between them.
16069 SDValue Zero = DAG.getIntPtrConstant(0, DL);
16070 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
16071 SDValue One = DAG.getIntPtrConstant(EltSize, SDLoc(FV));
16073 SDValue Cond = DAG.getSetCC(DL,
16074 getSetCCResultType(N0.getValueType()),
16076 AddToWorklist(Cond.getNode());
16077 SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(),
16079 AddToWorklist(CstOffset.getNode());
16080 CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx,
16082 AddToWorklist(CPIdx.getNode());
16083 return DAG.getLoad(
16084 TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
16085 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
16090 if (SDValue V = foldSelectCCToShiftAnd(DL, N0, N1, N2, N3, CC))
16093 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
16094 // where y is has a single bit set.
16095 // A plaintext description would be, we can turn the SELECT_CC into an AND
16096 // when the condition can be materialized as an all-ones register. Any
16097 // single bit-test can be materialized as an all-ones register with
16098 // shift-left and shift-right-arith.
16099 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
16100 N0->getValueType(0) == VT && isNullConstant(N1) && isNullConstant(N2)) {
16101 SDValue AndLHS = N0->getOperand(0);
16102 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
16103 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
16104 // Shift the tested bit over the sign bit.
16105 const APInt &AndMask = ConstAndRHS->getAPIntValue();
16107 DAG.getConstant(AndMask.countLeadingZeros(), SDLoc(AndLHS),
16108 getShiftAmountTy(AndLHS.getValueType()));
16109 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt);
16111 // Now arithmetic right shift it all the way over, so the result is either
16112 // all-ones, or zero.
16114 DAG.getConstant(AndMask.getBitWidth() - 1, SDLoc(Shl),
16115 getShiftAmountTy(Shl.getValueType()));
16116 SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt);
16118 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
16122 // fold select C, 16, 0 -> shl C, 4
16123 if (N2C && isNullConstant(N3) && N2C->getAPIntValue().isPowerOf2() &&
16124 TLI.getBooleanContents(N0.getValueType()) ==
16125 TargetLowering::ZeroOrOneBooleanContent) {
16127 // If the caller doesn't want us to simplify this into a zext of a compare,
16129 if (NotExtCompare && N2C->isOne())
16132 // Get a SetCC of the condition
16133 // NOTE: Don't create a SETCC if it's not legal on this target.
16134 if (!LegalOperations ||
16135 TLI.isOperationLegal(ISD::SETCC, N0.getValueType())) {
16137 // cast from setcc result type to select result type
16139 SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()),
16141 if (N2.getValueType().bitsLT(SCC.getValueType()))
16142 Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2),
16143 N2.getValueType());
16145 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
16146 N2.getValueType(), SCC);
16148 SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC);
16149 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
16150 N2.getValueType(), SCC);
16153 AddToWorklist(SCC.getNode());
16154 AddToWorklist(Temp.getNode());
16159 // shl setcc result by log2 n2c
16160 return DAG.getNode(
16161 ISD::SHL, DL, N2.getValueType(), Temp,
16162 DAG.getConstant(N2C->getAPIntValue().logBase2(), SDLoc(Temp),
16163 getShiftAmountTy(Temp.getValueType())));
16167 // Check to see if this is an integer abs.
16168 // select_cc setg[te] X, 0, X, -X ->
16169 // select_cc setgt X, -1, X, -X ->
16170 // select_cc setl[te] X, 0, -X, X ->
16171 // select_cc setlt X, 1, -X, X ->
16172 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
16174 ConstantSDNode *SubC = nullptr;
16175 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
16176 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
16177 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
16178 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
16179 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
16180 (N1C->isOne() && CC == ISD::SETLT)) &&
16181 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
16182 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
16184 EVT XType = N0.getValueType();
16185 if (SubC && SubC->isNullValue() && XType.isInteger()) {
16187 SDValue Shift = DAG.getNode(ISD::SRA, DL, XType,
16189 DAG.getConstant(XType.getSizeInBits() - 1, DL,
16190 getShiftAmountTy(N0.getValueType())));
16191 SDValue Add = DAG.getNode(ISD::ADD, DL,
16193 AddToWorklist(Shift.getNode());
16194 AddToWorklist(Add.getNode());
16195 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
16199 // select_cc seteq X, 0, sizeof(X), ctlz(X) -> ctlz(X)
16200 // select_cc seteq X, 0, sizeof(X), ctlz_zero_undef(X) -> ctlz(X)
16201 // select_cc seteq X, 0, sizeof(X), cttz(X) -> cttz(X)
16202 // select_cc seteq X, 0, sizeof(X), cttz_zero_undef(X) -> cttz(X)
16203 // select_cc setne X, 0, ctlz(X), sizeof(X) -> ctlz(X)
16204 // select_cc setne X, 0, ctlz_zero_undef(X), sizeof(X) -> ctlz(X)
16205 // select_cc setne X, 0, cttz(X), sizeof(X) -> cttz(X)
16206 // select_cc setne X, 0, cttz_zero_undef(X), sizeof(X) -> cttz(X)
16207 if (N1C && N1C->isNullValue() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
16208 SDValue ValueOnZero = N2;
16209 SDValue Count = N3;
16210 // If the condition is NE instead of E, swap the operands.
16211 if (CC == ISD::SETNE)
16212 std::swap(ValueOnZero, Count);
16213 // Check if the value on zero is a constant equal to the bits in the type.
16214 if (auto *ValueOnZeroC = dyn_cast<ConstantSDNode>(ValueOnZero)) {
16215 if (ValueOnZeroC->getAPIntValue() == VT.getSizeInBits()) {
16216 // If the other operand is cttz/cttz_zero_undef of N0, and cttz is
16217 // legal, combine to just cttz.
16218 if ((Count.getOpcode() == ISD::CTTZ ||
16219 Count.getOpcode() == ISD::CTTZ_ZERO_UNDEF) &&
16220 N0 == Count.getOperand(0) &&
16221 (!LegalOperations || TLI.isOperationLegal(ISD::CTTZ, VT)))
16222 return DAG.getNode(ISD::CTTZ, DL, VT, N0);
16223 // If the other operand is ctlz/ctlz_zero_undef of N0, and ctlz is
16224 // legal, combine to just ctlz.
16225 if ((Count.getOpcode() == ISD::CTLZ ||
16226 Count.getOpcode() == ISD::CTLZ_ZERO_UNDEF) &&
16227 N0 == Count.getOperand(0) &&
16228 (!LegalOperations || TLI.isOperationLegal(ISD::CTLZ, VT)))
16229 return DAG.getNode(ISD::CTLZ, DL, VT, N0);
16237 /// This is a stub for TargetLowering::SimplifySetCC.
16238 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
16239 ISD::CondCode Cond, const SDLoc &DL,
16240 bool foldBooleans) {
16241 TargetLowering::DAGCombinerInfo
16242 DagCombineInfo(DAG, Level, false, this);
16243 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
16246 /// Given an ISD::SDIV node expressing a divide by constant, return
16247 /// a DAG expression to select that will generate the same value by multiplying
16248 /// by a magic number.
16249 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
16250 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
16251 // when optimising for minimum size, we don't want to expand a div to a mul
16253 if (DAG.getMachineFunction().getFunction()->optForMinSize())
16256 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
16260 // Avoid division by zero.
16261 if (C->isNullValue())
16264 std::vector<SDNode*> Built;
16266 TLI.BuildSDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
16268 for (SDNode *N : Built)
16273 /// Given an ISD::SDIV node expressing a divide by constant power of 2, return a
16274 /// DAG expression that will generate the same value by right shifting.
16275 SDValue DAGCombiner::BuildSDIVPow2(SDNode *N) {
16276 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
16280 // Avoid division by zero.
16281 if (C->isNullValue())
16284 std::vector<SDNode *> Built;
16285 SDValue S = TLI.BuildSDIVPow2(N, C->getAPIntValue(), DAG, &Built);
16287 for (SDNode *N : Built)
16292 /// Given an ISD::UDIV node expressing a divide by constant, return a DAG
16293 /// expression that will generate the same value by multiplying by a magic
16295 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
16296 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
16297 // when optimising for minimum size, we don't want to expand a div to a mul
16299 if (DAG.getMachineFunction().getFunction()->optForMinSize())
16302 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
16306 // Avoid division by zero.
16307 if (C->isNullValue())
16310 std::vector<SDNode*> Built;
16312 TLI.BuildUDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
16314 for (SDNode *N : Built)
16319 /// Determines the LogBase2 value for a non-null input value using the
16320 /// transform: LogBase2(V) = (EltBits - 1) - ctlz(V).
16321 SDValue DAGCombiner::BuildLogBase2(SDValue V, const SDLoc &DL) {
16322 EVT VT = V.getValueType();
16323 unsigned EltBits = VT.getScalarSizeInBits();
16324 SDValue Ctlz = DAG.getNode(ISD::CTLZ, DL, VT, V);
16325 SDValue Base = DAG.getConstant(EltBits - 1, DL, VT);
16326 SDValue LogBase2 = DAG.getNode(ISD::SUB, DL, VT, Base, Ctlz);
16330 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
16331 /// For the reciprocal, we need to find the zero of the function:
16332 /// F(X) = A X - 1 [which has a zero at X = 1/A]
16334 /// X_{i+1} = X_i (2 - A X_i) = X_i + X_i (1 - A X_i) [this second form
16335 /// does not require additional intermediate precision]
16336 SDValue DAGCombiner::BuildReciprocalEstimate(SDValue Op, SDNodeFlags Flags) {
16337 if (Level >= AfterLegalizeDAG)
16340 // TODO: Handle half and/or extended types?
16341 EVT VT = Op.getValueType();
16342 if (VT.getScalarType() != MVT::f32 && VT.getScalarType() != MVT::f64)
16345 // If estimates are explicitly disabled for this function, we're done.
16346 MachineFunction &MF = DAG.getMachineFunction();
16347 int Enabled = TLI.getRecipEstimateDivEnabled(VT, MF);
16348 if (Enabled == TLI.ReciprocalEstimate::Disabled)
16351 // Estimates may be explicitly enabled for this type with a custom number of
16352 // refinement steps.
16353 int Iterations = TLI.getDivRefinementSteps(VT, MF);
16354 if (SDValue Est = TLI.getRecipEstimate(Op, DAG, Enabled, Iterations)) {
16355 AddToWorklist(Est.getNode());
16358 EVT VT = Op.getValueType();
16360 SDValue FPOne = DAG.getConstantFP(1.0, DL, VT);
16362 // Newton iterations: Est = Est + Est (1 - Arg * Est)
16363 for (int i = 0; i < Iterations; ++i) {
16364 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Op, Est, Flags);
16365 AddToWorklist(NewEst.getNode());
16367 NewEst = DAG.getNode(ISD::FSUB, DL, VT, FPOne, NewEst, Flags);
16368 AddToWorklist(NewEst.getNode());
16370 NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst, Flags);
16371 AddToWorklist(NewEst.getNode());
16373 Est = DAG.getNode(ISD::FADD, DL, VT, Est, NewEst, Flags);
16374 AddToWorklist(Est.getNode());
16383 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
16384 /// For the reciprocal sqrt, we need to find the zero of the function:
16385 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
16387 /// X_{i+1} = X_i (1.5 - A X_i^2 / 2)
16388 /// As a result, we precompute A/2 prior to the iteration loop.
16389 SDValue DAGCombiner::buildSqrtNROneConst(SDValue Arg, SDValue Est,
16390 unsigned Iterations,
16391 SDNodeFlags Flags, bool Reciprocal) {
16392 EVT VT = Arg.getValueType();
16394 SDValue ThreeHalves = DAG.getConstantFP(1.5, DL, VT);
16396 // We now need 0.5 * Arg which we can write as (1.5 * Arg - Arg) so that
16397 // this entire sequence requires only one FP constant.
16398 SDValue HalfArg = DAG.getNode(ISD::FMUL, DL, VT, ThreeHalves, Arg, Flags);
16399 AddToWorklist(HalfArg.getNode());
16401 HalfArg = DAG.getNode(ISD::FSUB, DL, VT, HalfArg, Arg, Flags);
16402 AddToWorklist(HalfArg.getNode());
16404 // Newton iterations: Est = Est * (1.5 - HalfArg * Est * Est)
16405 for (unsigned i = 0; i < Iterations; ++i) {
16406 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, Est, Flags);
16407 AddToWorklist(NewEst.getNode());
16409 NewEst = DAG.getNode(ISD::FMUL, DL, VT, HalfArg, NewEst, Flags);
16410 AddToWorklist(NewEst.getNode());
16412 NewEst = DAG.getNode(ISD::FSUB, DL, VT, ThreeHalves, NewEst, Flags);
16413 AddToWorklist(NewEst.getNode());
16415 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst, Flags);
16416 AddToWorklist(Est.getNode());
16419 // If non-reciprocal square root is requested, multiply the result by Arg.
16421 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Arg, Flags);
16422 AddToWorklist(Est.getNode());
16428 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
16429 /// For the reciprocal sqrt, we need to find the zero of the function:
16430 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
16432 /// X_{i+1} = (-0.5 * X_i) * (A * X_i * X_i + (-3.0))
16433 SDValue DAGCombiner::buildSqrtNRTwoConst(SDValue Arg, SDValue Est,
16434 unsigned Iterations,
16435 SDNodeFlags Flags, bool Reciprocal) {
16436 EVT VT = Arg.getValueType();
16438 SDValue MinusThree = DAG.getConstantFP(-3.0, DL, VT);
16439 SDValue MinusHalf = DAG.getConstantFP(-0.5, DL, VT);
16441 // This routine must enter the loop below to work correctly
16442 // when (Reciprocal == false).
16443 assert(Iterations > 0);
16445 // Newton iterations for reciprocal square root:
16446 // E = (E * -0.5) * ((A * E) * E + -3.0)
16447 for (unsigned i = 0; i < Iterations; ++i) {
16448 SDValue AE = DAG.getNode(ISD::FMUL, DL, VT, Arg, Est, Flags);
16449 AddToWorklist(AE.getNode());
16451 SDValue AEE = DAG.getNode(ISD::FMUL, DL, VT, AE, Est, Flags);
16452 AddToWorklist(AEE.getNode());
16454 SDValue RHS = DAG.getNode(ISD::FADD, DL, VT, AEE, MinusThree, Flags);
16455 AddToWorklist(RHS.getNode());
16457 // When calculating a square root at the last iteration build:
16458 // S = ((A * E) * -0.5) * ((A * E) * E + -3.0)
16459 // (notice a common subexpression)
16461 if (Reciprocal || (i + 1) < Iterations) {
16462 // RSQRT: LHS = (E * -0.5)
16463 LHS = DAG.getNode(ISD::FMUL, DL, VT, Est, MinusHalf, Flags);
16465 // SQRT: LHS = (A * E) * -0.5
16466 LHS = DAG.getNode(ISD::FMUL, DL, VT, AE, MinusHalf, Flags);
16468 AddToWorklist(LHS.getNode());
16470 Est = DAG.getNode(ISD::FMUL, DL, VT, LHS, RHS, Flags);
16471 AddToWorklist(Est.getNode());
16477 /// Build code to calculate either rsqrt(Op) or sqrt(Op). In the latter case
16478 /// Op*rsqrt(Op) is actually computed, so additional postprocessing is needed if
16479 /// Op can be zero.
16480 SDValue DAGCombiner::buildSqrtEstimateImpl(SDValue Op, SDNodeFlags Flags,
16482 if (Level >= AfterLegalizeDAG)
16485 // TODO: Handle half and/or extended types?
16486 EVT VT = Op.getValueType();
16487 if (VT.getScalarType() != MVT::f32 && VT.getScalarType() != MVT::f64)
16490 // If estimates are explicitly disabled for this function, we're done.
16491 MachineFunction &MF = DAG.getMachineFunction();
16492 int Enabled = TLI.getRecipEstimateSqrtEnabled(VT, MF);
16493 if (Enabled == TLI.ReciprocalEstimate::Disabled)
16496 // Estimates may be explicitly enabled for this type with a custom number of
16497 // refinement steps.
16498 int Iterations = TLI.getSqrtRefinementSteps(VT, MF);
16500 bool UseOneConstNR = false;
16502 TLI.getSqrtEstimate(Op, DAG, Enabled, Iterations, UseOneConstNR,
16504 AddToWorklist(Est.getNode());
16507 Est = UseOneConstNR
16508 ? buildSqrtNROneConst(Op, Est, Iterations, Flags, Reciprocal)
16509 : buildSqrtNRTwoConst(Op, Est, Iterations, Flags, Reciprocal);
16512 // Unfortunately, Est is now NaN if the input was exactly 0.0.
16513 // Select out this case and force the answer to 0.0.
16514 EVT VT = Op.getValueType();
16517 SDValue FPZero = DAG.getConstantFP(0.0, DL, VT);
16518 EVT CCVT = getSetCCResultType(VT);
16519 SDValue ZeroCmp = DAG.getSetCC(DL, CCVT, Op, FPZero, ISD::SETEQ);
16520 AddToWorklist(ZeroCmp.getNode());
16522 Est = DAG.getNode(VT.isVector() ? ISD::VSELECT : ISD::SELECT, DL, VT,
16523 ZeroCmp, FPZero, Est);
16524 AddToWorklist(Est.getNode());
16533 SDValue DAGCombiner::buildRsqrtEstimate(SDValue Op, SDNodeFlags Flags) {
16534 return buildSqrtEstimateImpl(Op, Flags, true);
16537 SDValue DAGCombiner::buildSqrtEstimate(SDValue Op, SDNodeFlags Flags) {
16538 return buildSqrtEstimateImpl(Op, Flags, false);
16541 /// Return true if base is a frame index, which is known not to alias with
16542 /// anything but itself. Provides base object and offset as results.
16543 static bool findBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
16544 const GlobalValue *&GV, const void *&CV) {
16545 // Assume it is a primitive operation.
16546 Base = Ptr; Offset = 0; GV = nullptr; CV = nullptr;
16548 // If it's an adding a simple constant then integrate the offset.
16549 if (Base.getOpcode() == ISD::ADD) {
16550 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
16551 Base = Base.getOperand(0);
16552 Offset += C->getSExtValue();
16556 // Return the underlying GlobalValue, and update the Offset. Return false
16557 // for GlobalAddressSDNode since the same GlobalAddress may be represented
16558 // by multiple nodes with different offsets.
16559 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
16560 GV = G->getGlobal();
16561 Offset += G->getOffset();
16565 // Return the underlying Constant value, and update the Offset. Return false
16566 // for ConstantSDNodes since the same constant pool entry may be represented
16567 // by multiple nodes with different offsets.
16568 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
16569 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
16570 : (const void *)C->getConstVal();
16571 Offset += C->getOffset();
16574 // If it's any of the following then it can't alias with anything but itself.
16575 return isa<FrameIndexSDNode>(Base);
16578 /// Return true if there is any possibility that the two addresses overlap.
16579 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const {
16580 // If they are the same then they must be aliases.
16581 if (Op0->getBasePtr() == Op1->getBasePtr()) return true;
16583 // If they are both volatile then they cannot be reordered.
16584 if (Op0->isVolatile() && Op1->isVolatile()) return true;
16586 // If one operation reads from invariant memory, and the other may store, they
16587 // cannot alias. These should really be checking the equivalent of mayWrite,
16588 // but it only matters for memory nodes other than load /store.
16589 if (Op0->isInvariant() && Op1->writeMem())
16592 if (Op1->isInvariant() && Op0->writeMem())
16595 unsigned NumBytes0 = Op0->getMemoryVT().getSizeInBits() >> 3;
16596 unsigned NumBytes1 = Op1->getMemoryVT().getSizeInBits() >> 3;
16598 // Check for BaseIndexOffset matching.
16599 BaseIndexOffset BasePtr0 = BaseIndexOffset::match(Op0->getBasePtr(), DAG);
16600 BaseIndexOffset BasePtr1 = BaseIndexOffset::match(Op1->getBasePtr(), DAG);
16601 if (BasePtr0.equalBaseIndex(BasePtr1))
16602 return !((BasePtr0.Offset + NumBytes0 <= BasePtr1.Offset) ||
16603 (BasePtr1.Offset + NumBytes1 <= BasePtr0.Offset));
16605 // FIXME: findBaseOffset and ConstantValue/GlobalValue/FrameIndex analysis
16606 // modified to use BaseIndexOffset.
16608 // Gather base node and offset information.
16609 SDValue Base0, Base1;
16610 int64_t Offset0, Offset1;
16611 const GlobalValue *GV0, *GV1;
16612 const void *CV0, *CV1;
16613 bool IsFrameIndex0 = findBaseOffset(Op0->getBasePtr(),
16614 Base0, Offset0, GV0, CV0);
16615 bool IsFrameIndex1 = findBaseOffset(Op1->getBasePtr(),
16616 Base1, Offset1, GV1, CV1);
16618 // If they have the same base address, then check to see if they overlap.
16619 if (Base0 == Base1 || (GV0 && (GV0 == GV1)) || (CV0 && (CV0 == CV1)))
16620 return !((Offset0 + NumBytes0) <= Offset1 ||
16621 (Offset1 + NumBytes1) <= Offset0);
16623 // It is possible for different frame indices to alias each other, mostly
16624 // when tail call optimization reuses return address slots for arguments.
16625 // To catch this case, look up the actual index of frame indices to compute
16626 // the real alias relationship.
16627 if (IsFrameIndex0 && IsFrameIndex1) {
16628 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
16629 Offset0 += MFI.getObjectOffset(cast<FrameIndexSDNode>(Base0)->getIndex());
16630 Offset1 += MFI.getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
16631 return !((Offset0 + NumBytes0) <= Offset1 ||
16632 (Offset1 + NumBytes1) <= Offset0);
16635 // Otherwise, if we know what the bases are, and they aren't identical, then
16636 // we know they cannot alias.
16637 if ((IsFrameIndex0 || CV0 || GV0) && (IsFrameIndex1 || CV1 || GV1))
16640 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
16641 // compared to the size and offset of the access, we may be able to prove they
16642 // do not alias. This check is conservative for now to catch cases created by
16643 // splitting vector types.
16644 int64_t SrcValOffset0 = Op0->getSrcValueOffset();
16645 int64_t SrcValOffset1 = Op1->getSrcValueOffset();
16646 unsigned OrigAlignment0 = Op0->getOriginalAlignment();
16647 unsigned OrigAlignment1 = Op1->getOriginalAlignment();
16648 if (OrigAlignment0 == OrigAlignment1 && SrcValOffset0 != SrcValOffset1 &&
16649 NumBytes0 == NumBytes1 && OrigAlignment0 > NumBytes0) {
16650 int64_t OffAlign0 = SrcValOffset0 % OrigAlignment0;
16651 int64_t OffAlign1 = SrcValOffset1 % OrigAlignment1;
16653 // There is no overlap between these relatively aligned accesses of similar
16654 // size. Return no alias.
16655 if ((OffAlign0 + NumBytes0) <= OffAlign1 ||
16656 (OffAlign1 + NumBytes1) <= OffAlign0)
16660 bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0
16662 : DAG.getSubtarget().useAA();
16664 if (CombinerAAOnlyFunc.getNumOccurrences() &&
16665 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
16670 Op0->getMemOperand()->getValue() && Op1->getMemOperand()->getValue()) {
16671 // Use alias analysis information.
16672 int64_t MinOffset = std::min(SrcValOffset0, SrcValOffset1);
16673 int64_t Overlap0 = NumBytes0 + SrcValOffset0 - MinOffset;
16674 int64_t Overlap1 = NumBytes1 + SrcValOffset1 - MinOffset;
16675 AliasResult AAResult =
16676 AA->alias(MemoryLocation(Op0->getMemOperand()->getValue(), Overlap0,
16677 UseTBAA ? Op0->getAAInfo() : AAMDNodes()),
16678 MemoryLocation(Op1->getMemOperand()->getValue(), Overlap1,
16679 UseTBAA ? Op1->getAAInfo() : AAMDNodes()) );
16680 if (AAResult == NoAlias)
16684 // Otherwise we have to assume they alias.
16688 /// Walk up chain skipping non-aliasing memory nodes,
16689 /// looking for aliasing nodes and adding them to the Aliases vector.
16690 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
16691 SmallVectorImpl<SDValue> &Aliases) {
16692 SmallVector<SDValue, 8> Chains; // List of chains to visit.
16693 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
16695 // Get alias information for node.
16696 bool IsLoad = isa<LoadSDNode>(N) && !cast<LSBaseSDNode>(N)->isVolatile();
16699 Chains.push_back(OriginalChain);
16700 unsigned Depth = 0;
16702 // Look at each chain and determine if it is an alias. If so, add it to the
16703 // aliases list. If not, then continue up the chain looking for the next
16705 while (!Chains.empty()) {
16706 SDValue Chain = Chains.pop_back_val();
16708 // For TokenFactor nodes, look at each operand and only continue up the
16709 // chain until we reach the depth limit.
16711 // FIXME: The depth check could be made to return the last non-aliasing
16712 // chain we found before we hit a tokenfactor rather than the original
16714 if (Depth > TLI.getGatherAllAliasesMaxDepth()) {
16716 Aliases.push_back(OriginalChain);
16720 // Don't bother if we've been before.
16721 if (!Visited.insert(Chain.getNode()).second)
16724 switch (Chain.getOpcode()) {
16725 case ISD::EntryToken:
16726 // Entry token is ideal chain operand, but handled in FindBetterChain.
16731 // Get alias information for Chain.
16732 bool IsOpLoad = isa<LoadSDNode>(Chain.getNode()) &&
16733 !cast<LSBaseSDNode>(Chain.getNode())->isVolatile();
16735 // If chain is alias then stop here.
16736 if (!(IsLoad && IsOpLoad) &&
16737 isAlias(cast<LSBaseSDNode>(N), cast<LSBaseSDNode>(Chain.getNode()))) {
16738 Aliases.push_back(Chain);
16740 // Look further up the chain.
16741 Chains.push_back(Chain.getOperand(0));
16747 case ISD::TokenFactor:
16748 // We have to check each of the operands of the token factor for "small"
16749 // token factors, so we queue them up. Adding the operands to the queue
16750 // (stack) in reverse order maintains the original order and increases the
16751 // likelihood that getNode will find a matching token factor (CSE.)
16752 if (Chain.getNumOperands() > 16) {
16753 Aliases.push_back(Chain);
16756 for (unsigned n = Chain.getNumOperands(); n;)
16757 Chains.push_back(Chain.getOperand(--n));
16761 case ISD::CopyFromReg:
16762 // Forward past CopyFromReg.
16763 Chains.push_back(Chain.getOperand(0));
16768 // For all other instructions we will just have to take what we can get.
16769 Aliases.push_back(Chain);
16775 /// Walk up chain skipping non-aliasing memory nodes, looking for a better chain
16776 /// (aliasing node.)
16777 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
16778 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
16780 // Accumulate all the aliases to this node.
16781 GatherAllAliases(N, OldChain, Aliases);
16783 // If no operands then chain to entry token.
16784 if (Aliases.size() == 0)
16785 return DAG.getEntryNode();
16787 // If a single operand then chain to it. We don't need to revisit it.
16788 if (Aliases.size() == 1)
16791 // Construct a custom tailored token factor.
16792 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Aliases);
16795 // This function tries to collect a bunch of potentially interesting
16796 // nodes to improve the chains of, all at once. This might seem
16797 // redundant, as this function gets called when visiting every store
16798 // node, so why not let the work be done on each store as it's visited?
16800 // I believe this is mainly important because MergeConsecutiveStores
16801 // is unable to deal with merging stores of different sizes, so unless
16802 // we improve the chains of all the potential candidates up-front
16803 // before running MergeConsecutiveStores, it might only see some of
16804 // the nodes that will eventually be candidates, and then not be able
16805 // to go from a partially-merged state to the desired final
16806 // fully-merged state.
16807 bool DAGCombiner::findBetterNeighborChains(StoreSDNode *St) {
16808 // This holds the base pointer, index, and the offset in bytes from the base
16810 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr(), DAG);
16812 // We must have a base and an offset.
16813 if (!BasePtr.Base.getNode())
16816 // Do not handle stores to undef base pointers.
16817 if (BasePtr.Base.isUndef())
16820 SmallVector<StoreSDNode *, 8> ChainedStores;
16821 ChainedStores.push_back(St);
16823 // Walk up the chain and look for nodes with offsets from the same
16824 // base pointer. Stop when reaching an instruction with a different kind
16825 // or instruction which has a different base pointer.
16826 StoreSDNode *Index = St;
16828 // If the chain has more than one use, then we can't reorder the mem ops.
16829 if (Index != St && !SDValue(Index, 0)->hasOneUse())
16832 if (Index->isVolatile() || Index->isIndexed())
16835 // Find the base pointer and offset for this memory node.
16836 BaseIndexOffset Ptr = BaseIndexOffset::match(Index->getBasePtr(), DAG);
16838 // Check that the base pointer is the same as the original one.
16839 if (!Ptr.equalBaseIndex(BasePtr))
16842 // Walk up the chain to find the next store node, ignoring any
16843 // intermediate loads. Any other kind of node will halt the loop.
16844 SDNode *NextInChain = Index->getChain().getNode();
16846 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
16847 // We found a store node. Use it for the next iteration.
16848 if (STn->isVolatile() || STn->isIndexed()) {
16852 ChainedStores.push_back(STn);
16855 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
16856 NextInChain = Ldn->getChain().getNode();
16865 // At this point, ChainedStores lists all of the Store nodes
16866 // reachable by iterating up through chain nodes matching the above
16867 // conditions. For each such store identified, try to find an
16868 // earlier chain to attach the store to which won't violate the
16869 // required ordering.
16870 bool MadeChangeToSt = false;
16871 SmallVector<std::pair<StoreSDNode *, SDValue>, 8> BetterChains;
16873 for (StoreSDNode *ChainedStore : ChainedStores) {
16874 SDValue Chain = ChainedStore->getChain();
16875 SDValue BetterChain = FindBetterChain(ChainedStore, Chain);
16877 if (Chain != BetterChain) {
16878 if (ChainedStore == St)
16879 MadeChangeToSt = true;
16880 BetterChains.push_back(std::make_pair(ChainedStore, BetterChain));
16884 // Do all replacements after finding the replacements to make to avoid making
16885 // the chains more complicated by introducing new TokenFactors.
16886 for (auto Replacement : BetterChains)
16887 replaceStoreChain(Replacement.first, Replacement.second);
16889 return MadeChangeToSt;
16892 /// This is the entry point for the file.
16893 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis *AA,
16894 CodeGenOpt::Level OptLevel) {
16895 /// This is the main entry point to this class.
16896 DAGCombiner(*this, AA, OptLevel).Run(Level);