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/CodeGen/SelectionDAG.h"
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/ADT/SmallBitVector.h"
22 #include "llvm/ADT/SmallPtrSet.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/IR/DataLayout.h"
28 #include "llvm/IR/DerivedTypes.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/LLVMContext.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Target/TargetLowering.h"
37 #include "llvm/Target/TargetOptions.h"
38 #include "llvm/Target/TargetRegisterInfo.h"
39 #include "llvm/Target/TargetSubtargetInfo.h"
43 #define DEBUG_TYPE "dagcombine"
45 STATISTIC(NodesCombined , "Number of dag nodes combined");
46 STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created");
47 STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created");
48 STATISTIC(OpsNarrowed , "Number of load/op/store narrowed");
49 STATISTIC(LdStFP2Int , "Number of fp load/store pairs transformed to int");
50 STATISTIC(SlicedLoads, "Number of load sliced");
54 CombinerAA("combiner-alias-analysis", cl::Hidden,
55 cl::desc("Enable DAG combiner alias-analysis heuristics"));
58 CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden,
59 cl::desc("Enable DAG combiner's use of IR alias analysis"));
62 UseTBAA("combiner-use-tbaa", cl::Hidden, cl::init(true),
63 cl::desc("Enable DAG combiner's use of TBAA"));
66 static cl::opt<std::string>
67 CombinerAAOnlyFunc("combiner-aa-only-func", cl::Hidden,
68 cl::desc("Only use DAG-combiner alias analysis in this"
72 /// Hidden option to stress test load slicing, i.e., when this option
73 /// is enabled, load slicing bypasses most of its profitability guards.
75 StressLoadSlicing("combiner-stress-load-slicing", cl::Hidden,
76 cl::desc("Bypass the profitability model of load "
81 MaySplitLoadIndex("combiner-split-load-index", cl::Hidden, cl::init(true),
82 cl::desc("DAG combiner may split indexing from loads"));
84 //------------------------------ DAGCombiner ---------------------------------//
88 const TargetLowering &TLI;
90 CodeGenOpt::Level OptLevel;
95 /// \brief Worklist of all of the nodes that need to be simplified.
97 /// This must behave as a stack -- new nodes to process are pushed onto the
98 /// back and when processing we pop off of the back.
100 /// The worklist will not contain duplicates but may contain null entries
101 /// due to nodes being deleted from the underlying DAG.
102 SmallVector<SDNode *, 64> Worklist;
104 /// \brief Mapping from an SDNode to its position on the worklist.
106 /// This is used to find and remove nodes from the worklist (by nulling
107 /// them) when they are deleted from the underlying DAG. It relies on
108 /// stable indices of nodes within the worklist.
109 DenseMap<SDNode *, unsigned> WorklistMap;
111 /// \brief Set of nodes which have been combined (at least once).
113 /// This is used to allow us to reliably add any operands of a DAG node
114 /// which have not yet been combined to the worklist.
115 SmallPtrSet<SDNode *, 64> CombinedNodes;
117 // AA - Used for DAG load/store alias analysis.
120 /// When an instruction is simplified, add all users of the instruction to
121 /// the work lists because they might get more simplified now.
122 void AddUsersToWorklist(SDNode *N) {
123 for (SDNode *Node : N->uses())
127 /// Call the node-specific routine that folds each particular type of node.
128 SDValue visit(SDNode *N);
131 /// Add to the worklist making sure its instance is at the back (next to be
133 void AddToWorklist(SDNode *N) {
134 // Skip handle nodes as they can't usefully be combined and confuse the
135 // zero-use deletion strategy.
136 if (N->getOpcode() == ISD::HANDLENODE)
139 if (WorklistMap.insert(std::make_pair(N, Worklist.size())).second)
140 Worklist.push_back(N);
143 /// Remove all instances of N from the worklist.
144 void removeFromWorklist(SDNode *N) {
145 CombinedNodes.erase(N);
147 auto It = WorklistMap.find(N);
148 if (It == WorklistMap.end())
149 return; // Not in the worklist.
151 // Null out the entry rather than erasing it to avoid a linear operation.
152 Worklist[It->second] = nullptr;
153 WorklistMap.erase(It);
156 void deleteAndRecombine(SDNode *N);
157 bool recursivelyDeleteUnusedNodes(SDNode *N);
159 SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
162 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) {
163 return CombineTo(N, &Res, 1, AddTo);
166 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1,
168 SDValue To[] = { Res0, Res1 };
169 return CombineTo(N, To, 2, AddTo);
172 void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO);
176 /// Check the specified integer node value to see if it can be simplified or
177 /// if things it uses can be simplified by bit propagation.
178 /// If so, return true.
179 bool SimplifyDemandedBits(SDValue Op) {
180 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
181 APInt Demanded = APInt::getAllOnesValue(BitWidth);
182 return SimplifyDemandedBits(Op, Demanded);
185 bool SimplifyDemandedBits(SDValue Op, const APInt &Demanded);
187 bool CombineToPreIndexedLoadStore(SDNode *N);
188 bool CombineToPostIndexedLoadStore(SDNode *N);
189 SDValue SplitIndexingFromLoad(LoadSDNode *LD);
190 bool SliceUpLoad(SDNode *N);
192 /// \brief Replace an ISD::EXTRACT_VECTOR_ELT of a load with a narrowed
195 /// \param EVE ISD::EXTRACT_VECTOR_ELT to be replaced.
196 /// \param InVecVT type of the input vector to EVE with bitcasts resolved.
197 /// \param EltNo index of the vector element to load.
198 /// \param OriginalLoad load that EVE came from to be replaced.
199 /// \returns EVE on success SDValue() on failure.
200 SDValue ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
201 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad);
202 void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad);
203 SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace);
204 SDValue SExtPromoteOperand(SDValue Op, EVT PVT);
205 SDValue ZExtPromoteOperand(SDValue Op, EVT PVT);
206 SDValue PromoteIntBinOp(SDValue Op);
207 SDValue PromoteIntShiftOp(SDValue Op);
208 SDValue PromoteExtend(SDValue Op);
209 bool PromoteLoad(SDValue Op);
211 void ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
212 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
213 ISD::NodeType ExtType);
215 /// Call the node-specific routine that knows how to fold each
216 /// particular type of node. If that doesn't do anything, try the
217 /// target-specific DAG combines.
218 SDValue combine(SDNode *N);
220 // Visitation implementation - Implement dag node combining for different
221 // node types. The semantics are as follows:
223 // SDValue.getNode() == 0 - No change was made
224 // SDValue.getNode() == N - N was replaced, is dead and has been handled.
225 // otherwise - N should be replaced by the returned Operand.
227 SDValue visitTokenFactor(SDNode *N);
228 SDValue visitMERGE_VALUES(SDNode *N);
229 SDValue visitADD(SDNode *N);
230 SDValue visitSUB(SDNode *N);
231 SDValue visitADDC(SDNode *N);
232 SDValue visitSUBC(SDNode *N);
233 SDValue visitADDE(SDNode *N);
234 SDValue visitSUBE(SDNode *N);
235 SDValue visitMUL(SDNode *N);
236 SDValue visitSDIV(SDNode *N);
237 SDValue visitUDIV(SDNode *N);
238 SDValue visitSREM(SDNode *N);
239 SDValue visitUREM(SDNode *N);
240 SDValue visitMULHU(SDNode *N);
241 SDValue visitMULHS(SDNode *N);
242 SDValue visitSMUL_LOHI(SDNode *N);
243 SDValue visitUMUL_LOHI(SDNode *N);
244 SDValue visitSMULO(SDNode *N);
245 SDValue visitUMULO(SDNode *N);
246 SDValue visitSDIVREM(SDNode *N);
247 SDValue visitUDIVREM(SDNode *N);
248 SDValue visitIMINMAX(SDNode *N);
249 SDValue visitAND(SDNode *N);
250 SDValue visitANDLike(SDValue N0, SDValue N1, SDNode *LocReference);
251 SDValue visitOR(SDNode *N);
252 SDValue visitORLike(SDValue N0, SDValue N1, SDNode *LocReference);
253 SDValue visitXOR(SDNode *N);
254 SDValue SimplifyVBinOp(SDNode *N);
255 SDValue visitSHL(SDNode *N);
256 SDValue visitSRA(SDNode *N);
257 SDValue visitSRL(SDNode *N);
258 SDValue visitRotate(SDNode *N);
259 SDValue visitBSWAP(SDNode *N);
260 SDValue visitCTLZ(SDNode *N);
261 SDValue visitCTLZ_ZERO_UNDEF(SDNode *N);
262 SDValue visitCTTZ(SDNode *N);
263 SDValue visitCTTZ_ZERO_UNDEF(SDNode *N);
264 SDValue visitCTPOP(SDNode *N);
265 SDValue visitSELECT(SDNode *N);
266 SDValue visitVSELECT(SDNode *N);
267 SDValue visitSELECT_CC(SDNode *N);
268 SDValue visitSETCC(SDNode *N);
269 SDValue visitSIGN_EXTEND(SDNode *N);
270 SDValue visitZERO_EXTEND(SDNode *N);
271 SDValue visitANY_EXTEND(SDNode *N);
272 SDValue visitSIGN_EXTEND_INREG(SDNode *N);
273 SDValue visitSIGN_EXTEND_VECTOR_INREG(SDNode *N);
274 SDValue visitTRUNCATE(SDNode *N);
275 SDValue visitBITCAST(SDNode *N);
276 SDValue visitBUILD_PAIR(SDNode *N);
277 SDValue visitFADD(SDNode *N);
278 SDValue visitFSUB(SDNode *N);
279 SDValue visitFMUL(SDNode *N);
280 SDValue visitFMA(SDNode *N);
281 SDValue visitFDIV(SDNode *N);
282 SDValue visitFREM(SDNode *N);
283 SDValue visitFSQRT(SDNode *N);
284 SDValue visitFCOPYSIGN(SDNode *N);
285 SDValue visitSINT_TO_FP(SDNode *N);
286 SDValue visitUINT_TO_FP(SDNode *N);
287 SDValue visitFP_TO_SINT(SDNode *N);
288 SDValue visitFP_TO_UINT(SDNode *N);
289 SDValue visitFP_ROUND(SDNode *N);
290 SDValue visitFP_ROUND_INREG(SDNode *N);
291 SDValue visitFP_EXTEND(SDNode *N);
292 SDValue visitFNEG(SDNode *N);
293 SDValue visitFABS(SDNode *N);
294 SDValue visitFCEIL(SDNode *N);
295 SDValue visitFTRUNC(SDNode *N);
296 SDValue visitFFLOOR(SDNode *N);
297 SDValue visitFMINNUM(SDNode *N);
298 SDValue visitFMAXNUM(SDNode *N);
299 SDValue visitBRCOND(SDNode *N);
300 SDValue visitBR_CC(SDNode *N);
301 SDValue visitLOAD(SDNode *N);
302 SDValue visitSTORE(SDNode *N);
303 SDValue visitINSERT_VECTOR_ELT(SDNode *N);
304 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
305 SDValue visitBUILD_VECTOR(SDNode *N);
306 SDValue visitCONCAT_VECTORS(SDNode *N);
307 SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
308 SDValue visitVECTOR_SHUFFLE(SDNode *N);
309 SDValue visitSCALAR_TO_VECTOR(SDNode *N);
310 SDValue visitINSERT_SUBVECTOR(SDNode *N);
311 SDValue visitMLOAD(SDNode *N);
312 SDValue visitMSTORE(SDNode *N);
313 SDValue visitMGATHER(SDNode *N);
314 SDValue visitMSCATTER(SDNode *N);
315 SDValue visitFP_TO_FP16(SDNode *N);
316 SDValue visitFP16_TO_FP(SDNode *N);
318 SDValue visitFADDForFMACombine(SDNode *N);
319 SDValue visitFSUBForFMACombine(SDNode *N);
321 SDValue XformToShuffleWithZero(SDNode *N);
322 SDValue ReassociateOps(unsigned Opc, SDLoc DL, SDValue LHS, SDValue RHS);
324 SDValue visitShiftByConstant(SDNode *N, ConstantSDNode *Amt);
326 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
327 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
328 SDValue SimplifySelect(SDLoc DL, SDValue N0, SDValue N1, SDValue N2);
329 SDValue SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1, SDValue N2,
330 SDValue N3, ISD::CondCode CC,
331 bool NotExtCompare = false);
332 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
333 SDLoc DL, bool foldBooleans = true);
335 bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
337 bool isOneUseSetCC(SDValue N) const;
339 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
341 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
342 SDValue CombineExtLoad(SDNode *N);
343 SDValue combineRepeatedFPDivisors(SDNode *N);
344 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
345 SDValue BuildSDIV(SDNode *N);
346 SDValue BuildSDIVPow2(SDNode *N);
347 SDValue BuildUDIV(SDNode *N);
348 SDValue BuildReciprocalEstimate(SDValue Op);
349 SDValue BuildRsqrtEstimate(SDValue Op);
350 SDValue BuildRsqrtNROneConst(SDValue Op, SDValue Est, unsigned Iterations);
351 SDValue BuildRsqrtNRTwoConst(SDValue Op, SDValue Est, unsigned Iterations);
352 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
353 bool DemandHighBits = true);
354 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
355 SDNode *MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg,
356 SDValue InnerPos, SDValue InnerNeg,
357 unsigned PosOpcode, unsigned NegOpcode,
359 SDNode *MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL);
360 SDValue ReduceLoadWidth(SDNode *N);
361 SDValue ReduceLoadOpStoreWidth(SDNode *N);
362 SDValue TransformFPLoadStorePair(SDNode *N);
363 SDValue reduceBuildVecExtToExtBuildVec(SDNode *N);
364 SDValue reduceBuildVecConvertToConvertBuildVec(SDNode *N);
366 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
368 /// Walk up chain skipping non-aliasing memory nodes,
369 /// looking for aliasing nodes and adding them to the Aliases vector.
370 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
371 SmallVectorImpl<SDValue> &Aliases);
373 /// Return true if there is any possibility that the two addresses overlap.
374 bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const;
376 /// Walk up chain skipping non-aliasing memory nodes, looking for a better
377 /// chain (aliasing node.)
378 SDValue FindBetterChain(SDNode *N, SDValue Chain);
380 /// Holds a pointer to an LSBaseSDNode as well as information on where it
381 /// is located in a sequence of memory operations connected by a chain.
383 MemOpLink (LSBaseSDNode *N, int64_t Offset, unsigned Seq):
384 MemNode(N), OffsetFromBase(Offset), SequenceNum(Seq) { }
385 // Ptr to the mem node.
386 LSBaseSDNode *MemNode;
387 // Offset from the base ptr.
388 int64_t OffsetFromBase;
389 // What is the sequence number of this mem node.
390 // Lowest mem operand in the DAG starts at zero.
391 unsigned SequenceNum;
394 /// This is a helper function for MergeStoresOfConstantsOrVecElts. Returns a
395 /// constant build_vector of the stored constant values in Stores.
396 SDValue getMergedConstantVectorStore(SelectionDAG &DAG,
398 ArrayRef<MemOpLink> Stores,
401 /// This is a helper function for MergeConsecutiveStores. When the source
402 /// elements of the consecutive stores are all constants or all extracted
403 /// vector elements, try to merge them into one larger store.
404 /// \return True if a merged store was created.
405 bool MergeStoresOfConstantsOrVecElts(SmallVectorImpl<MemOpLink> &StoreNodes,
406 EVT MemVT, unsigned NumElem,
407 bool IsConstantSrc, bool UseVector);
409 /// This is a helper function for MergeConsecutiveStores.
410 /// Stores that may be merged are placed in StoreNodes.
411 /// Loads that may alias with those stores are placed in AliasLoadNodes.
412 void getStoreMergeAndAliasCandidates(
413 StoreSDNode* St, SmallVectorImpl<MemOpLink> &StoreNodes,
414 SmallVectorImpl<LSBaseSDNode*> &AliasLoadNodes);
416 /// Merge consecutive store operations into a wide store.
417 /// This optimization uses wide integers or vectors when possible.
418 /// \return True if some memory operations were changed.
419 bool MergeConsecutiveStores(StoreSDNode *N);
421 /// \brief Try to transform a truncation where C is a constant:
422 /// (trunc (and X, C)) -> (and (trunc X), (trunc C))
424 /// \p N needs to be a truncation and its first operand an AND. Other
425 /// requirements are checked by the function (e.g. that trunc is
426 /// single-use) and if missed an empty SDValue is returned.
427 SDValue distributeTruncateThroughAnd(SDNode *N);
430 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
431 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
432 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {
433 ForCodeSize = DAG.getMachineFunction().getFunction()->optForSize();
436 /// Runs the dag combiner on all nodes in the work list
437 void Run(CombineLevel AtLevel);
439 SelectionDAG &getDAG() const { return DAG; }
441 /// Returns a type large enough to hold any valid shift amount - before type
442 /// legalization these can be huge.
443 EVT getShiftAmountTy(EVT LHSTy) {
444 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
445 if (LHSTy.isVector())
447 auto &DL = DAG.getDataLayout();
448 return LegalTypes ? TLI.getScalarShiftAmountTy(DL, LHSTy)
449 : TLI.getPointerTy(DL);
452 /// This method returns true if we are running before type legalization or
453 /// if the specified VT is legal.
454 bool isTypeLegal(const EVT &VT) {
455 if (!LegalTypes) return true;
456 return TLI.isTypeLegal(VT);
459 /// Convenience wrapper around TargetLowering::getSetCCResultType
460 EVT getSetCCResultType(EVT VT) const {
461 return TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
468 /// This class is a DAGUpdateListener that removes any deleted
469 /// nodes from the worklist.
470 class WorklistRemover : public SelectionDAG::DAGUpdateListener {
473 explicit WorklistRemover(DAGCombiner &dc)
474 : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {}
476 void NodeDeleted(SDNode *N, SDNode *E) override {
477 DC.removeFromWorklist(N);
482 //===----------------------------------------------------------------------===//
483 // TargetLowering::DAGCombinerInfo implementation
484 //===----------------------------------------------------------------------===//
486 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
487 ((DAGCombiner*)DC)->AddToWorklist(N);
490 void TargetLowering::DAGCombinerInfo::RemoveFromWorklist(SDNode *N) {
491 ((DAGCombiner*)DC)->removeFromWorklist(N);
494 SDValue TargetLowering::DAGCombinerInfo::
495 CombineTo(SDNode *N, ArrayRef<SDValue> To, bool AddTo) {
496 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
499 SDValue TargetLowering::DAGCombinerInfo::
500 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
501 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
505 SDValue TargetLowering::DAGCombinerInfo::
506 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
507 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
510 void TargetLowering::DAGCombinerInfo::
511 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
512 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
515 //===----------------------------------------------------------------------===//
517 //===----------------------------------------------------------------------===//
519 void DAGCombiner::deleteAndRecombine(SDNode *N) {
520 removeFromWorklist(N);
522 // If the operands of this node are only used by the node, they will now be
523 // dead. Make sure to re-visit them and recursively delete dead nodes.
524 for (const SDValue &Op : N->ops())
525 // For an operand generating multiple values, one of the values may
526 // become dead allowing further simplification (e.g. split index
527 // arithmetic from an indexed load).
528 if (Op->hasOneUse() || Op->getNumValues() > 1)
529 AddToWorklist(Op.getNode());
534 /// Return 1 if we can compute the negated form of the specified expression for
535 /// the same cost as the expression itself, or 2 if we can compute the negated
536 /// form more cheaply than the expression itself.
537 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
538 const TargetLowering &TLI,
539 const TargetOptions *Options,
540 unsigned Depth = 0) {
541 // fneg is removable even if it has multiple uses.
542 if (Op.getOpcode() == ISD::FNEG) return 2;
544 // Don't allow anything with multiple uses.
545 if (!Op.hasOneUse()) return 0;
547 // Don't recurse exponentially.
548 if (Depth > 6) return 0;
550 switch (Op.getOpcode()) {
551 default: return false;
552 case ISD::ConstantFP:
553 // Don't invert constant FP values after legalize. The negated constant
554 // isn't necessarily legal.
555 return LegalOperations ? 0 : 1;
557 // FIXME: determine better conditions for this xform.
558 if (!Options->UnsafeFPMath) return 0;
560 // After operation legalization, it might not be legal to create new FSUBs.
561 if (LegalOperations &&
562 !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType()))
565 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
566 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
569 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
570 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
573 // We can't turn -(A-B) into B-A when we honor signed zeros.
574 if (!Options->UnsafeFPMath) return 0;
576 // fold (fneg (fsub A, B)) -> (fsub B, A)
581 if (Options->HonorSignDependentRoundingFPMath()) return 0;
583 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
584 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
588 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
594 return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options,
599 /// If isNegatibleForFree returns true, return the newly negated expression.
600 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
601 bool LegalOperations, unsigned Depth = 0) {
602 const TargetOptions &Options = DAG.getTarget().Options;
603 // fneg is removable even if it has multiple uses.
604 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
606 // Don't allow anything with multiple uses.
607 assert(Op.hasOneUse() && "Unknown reuse!");
609 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
610 switch (Op.getOpcode()) {
611 default: llvm_unreachable("Unknown code");
612 case ISD::ConstantFP: {
613 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
615 return DAG.getConstantFP(V, SDLoc(Op), Op.getValueType());
618 // FIXME: determine better conditions for this xform.
619 assert(Options.UnsafeFPMath);
621 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
622 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
623 DAG.getTargetLoweringInfo(), &Options, Depth+1))
624 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
625 GetNegatedExpression(Op.getOperand(0), DAG,
626 LegalOperations, Depth+1),
628 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
629 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
630 GetNegatedExpression(Op.getOperand(1), DAG,
631 LegalOperations, Depth+1),
634 // We can't turn -(A-B) into B-A when we honor signed zeros.
635 assert(Options.UnsafeFPMath);
637 // fold (fneg (fsub 0, B)) -> B
638 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
640 return Op.getOperand(1);
642 // fold (fneg (fsub A, B)) -> (fsub B, A)
643 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
644 Op.getOperand(1), Op.getOperand(0));
648 assert(!Options.HonorSignDependentRoundingFPMath());
650 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
651 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
652 DAG.getTargetLoweringInfo(), &Options, Depth+1))
653 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
654 GetNegatedExpression(Op.getOperand(0), DAG,
655 LegalOperations, Depth+1),
658 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
659 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
661 GetNegatedExpression(Op.getOperand(1), DAG,
662 LegalOperations, Depth+1));
666 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
667 GetNegatedExpression(Op.getOperand(0), DAG,
668 LegalOperations, Depth+1));
670 return DAG.getNode(ISD::FP_ROUND, SDLoc(Op), Op.getValueType(),
671 GetNegatedExpression(Op.getOperand(0), DAG,
672 LegalOperations, Depth+1),
677 // Return true if this node is a setcc, or is a select_cc
678 // that selects between the target values used for true and false, making it
679 // equivalent to a setcc. Also, set the incoming LHS, RHS, and CC references to
680 // the appropriate nodes based on the type of node we are checking. This
681 // simplifies life a bit for the callers.
682 bool DAGCombiner::isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
684 if (N.getOpcode() == ISD::SETCC) {
685 LHS = N.getOperand(0);
686 RHS = N.getOperand(1);
687 CC = N.getOperand(2);
691 if (N.getOpcode() != ISD::SELECT_CC ||
692 !TLI.isConstTrueVal(N.getOperand(2).getNode()) ||
693 !TLI.isConstFalseVal(N.getOperand(3).getNode()))
696 if (TLI.getBooleanContents(N.getValueType()) ==
697 TargetLowering::UndefinedBooleanContent)
700 LHS = N.getOperand(0);
701 RHS = N.getOperand(1);
702 CC = N.getOperand(4);
706 /// Return true if this is a SetCC-equivalent operation with only one use.
707 /// If this is true, it allows the users to invert the operation for free when
708 /// it is profitable to do so.
709 bool DAGCombiner::isOneUseSetCC(SDValue N) const {
711 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
716 /// Returns true if N is a BUILD_VECTOR node whose
717 /// elements are all the same constant or undefined.
718 static bool isConstantSplatVector(SDNode *N, APInt& SplatValue) {
719 BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(N);
724 unsigned SplatBitSize;
726 EVT EltVT = N->getValueType(0).getVectorElementType();
727 return (C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
729 EltVT.getSizeInBits() >= SplatBitSize);
732 // \brief Returns the SDNode if it is a constant integer BuildVector
733 // or constant integer.
734 static SDNode *isConstantIntBuildVectorOrConstantInt(SDValue N) {
735 if (isa<ConstantSDNode>(N))
737 if (ISD::isBuildVectorOfConstantSDNodes(N.getNode()))
742 // \brief Returns the SDNode if it is a constant float BuildVector
743 // or constant float.
744 static SDNode *isConstantFPBuildVectorOrConstantFP(SDValue N) {
745 if (isa<ConstantFPSDNode>(N))
747 if (ISD::isBuildVectorOfConstantFPSDNodes(N.getNode()))
752 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
754 static ConstantSDNode *isConstOrConstSplat(SDValue N) {
755 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N))
758 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
759 BitVector UndefElements;
760 ConstantSDNode *CN = BV->getConstantSplatNode(&UndefElements);
762 // BuildVectors can truncate their operands. Ignore that case here.
763 // FIXME: We blindly ignore splats which include undef which is overly
765 if (CN && UndefElements.none() &&
766 CN->getValueType(0) == N.getValueType().getScalarType())
773 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
775 static ConstantFPSDNode *isConstOrConstSplatFP(SDValue N) {
776 if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N))
779 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
780 BitVector UndefElements;
781 ConstantFPSDNode *CN = BV->getConstantFPSplatNode(&UndefElements);
783 if (CN && UndefElements.none())
790 SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL,
791 SDValue N0, SDValue N1) {
792 EVT VT = N0.getValueType();
793 if (N0.getOpcode() == Opc) {
794 if (SDNode *L = isConstantIntBuildVectorOrConstantInt(N0.getOperand(1))) {
795 if (SDNode *R = isConstantIntBuildVectorOrConstantInt(N1)) {
796 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
797 if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, DL, VT, L, R))
798 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
801 if (N0.hasOneUse()) {
802 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one
804 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1);
805 if (!OpNode.getNode())
807 AddToWorklist(OpNode.getNode());
808 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
813 if (N1.getOpcode() == Opc) {
814 if (SDNode *R = isConstantIntBuildVectorOrConstantInt(N1.getOperand(1))) {
815 if (SDNode *L = isConstantIntBuildVectorOrConstantInt(N0)) {
816 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
817 if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, DL, VT, R, L))
818 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
821 if (N1.hasOneUse()) {
822 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one
824 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N1.getOperand(0), N0);
825 if (!OpNode.getNode())
827 AddToWorklist(OpNode.getNode());
828 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
836 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
838 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
840 DEBUG(dbgs() << "\nReplacing.1 ";
842 dbgs() << "\nWith: ";
843 To[0].getNode()->dump(&DAG);
844 dbgs() << " and " << NumTo-1 << " other values\n");
845 for (unsigned i = 0, e = NumTo; i != e; ++i)
846 assert((!To[i].getNode() ||
847 N->getValueType(i) == To[i].getValueType()) &&
848 "Cannot combine value to value of different type!");
850 WorklistRemover DeadNodes(*this);
851 DAG.ReplaceAllUsesWith(N, To);
853 // Push the new nodes and any users onto the worklist
854 for (unsigned i = 0, e = NumTo; i != e; ++i) {
855 if (To[i].getNode()) {
856 AddToWorklist(To[i].getNode());
857 AddUsersToWorklist(To[i].getNode());
862 // Finally, if the node is now dead, remove it from the graph. The node
863 // may not be dead if the replacement process recursively simplified to
864 // something else needing this node.
866 deleteAndRecombine(N);
867 return SDValue(N, 0);
871 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
872 // Replace all uses. If any nodes become isomorphic to other nodes and
873 // are deleted, make sure to remove them from our worklist.
874 WorklistRemover DeadNodes(*this);
875 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New);
877 // Push the new node and any (possibly new) users onto the worklist.
878 AddToWorklist(TLO.New.getNode());
879 AddUsersToWorklist(TLO.New.getNode());
881 // Finally, if the node is now dead, remove it from the graph. The node
882 // may not be dead if the replacement process recursively simplified to
883 // something else needing this node.
884 if (TLO.Old.getNode()->use_empty())
885 deleteAndRecombine(TLO.Old.getNode());
888 /// Check the specified integer node value to see if it can be simplified or if
889 /// things it uses can be simplified by bit propagation. If so, return true.
890 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
891 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
892 APInt KnownZero, KnownOne;
893 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
897 AddToWorklist(Op.getNode());
899 // Replace the old value with the new one.
901 DEBUG(dbgs() << "\nReplacing.2 ";
902 TLO.Old.getNode()->dump(&DAG);
903 dbgs() << "\nWith: ";
904 TLO.New.getNode()->dump(&DAG);
907 CommitTargetLoweringOpt(TLO);
911 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
913 EVT VT = Load->getValueType(0);
914 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
916 DEBUG(dbgs() << "\nReplacing.9 ";
918 dbgs() << "\nWith: ";
919 Trunc.getNode()->dump(&DAG);
921 WorklistRemover DeadNodes(*this);
922 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc);
923 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1));
924 deleteAndRecombine(Load);
925 AddToWorklist(Trunc.getNode());
928 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
931 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
932 EVT MemVT = LD->getMemoryVT();
933 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
934 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, PVT, MemVT) ? ISD::ZEXTLOAD
936 : LD->getExtensionType();
938 return DAG.getExtLoad(ExtType, dl, PVT,
939 LD->getChain(), LD->getBasePtr(),
940 MemVT, LD->getMemOperand());
943 unsigned Opc = Op.getOpcode();
946 case ISD::AssertSext:
947 return DAG.getNode(ISD::AssertSext, dl, PVT,
948 SExtPromoteOperand(Op.getOperand(0), PVT),
950 case ISD::AssertZext:
951 return DAG.getNode(ISD::AssertZext, dl, PVT,
952 ZExtPromoteOperand(Op.getOperand(0), PVT),
954 case ISD::Constant: {
956 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
957 return DAG.getNode(ExtOpc, dl, PVT, Op);
961 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
963 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
966 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
967 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
969 EVT OldVT = Op.getValueType();
971 bool Replace = false;
972 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
973 if (!NewOp.getNode())
975 AddToWorklist(NewOp.getNode());
978 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
979 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
980 DAG.getValueType(OldVT));
983 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
984 EVT OldVT = Op.getValueType();
986 bool Replace = false;
987 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
988 if (!NewOp.getNode())
990 AddToWorklist(NewOp.getNode());
993 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
994 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
997 /// Promote the specified integer binary operation if the target indicates it is
998 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
999 /// i32 since i16 instructions are longer.
1000 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
1001 if (!LegalOperations)
1004 EVT VT = Op.getValueType();
1005 if (VT.isVector() || !VT.isInteger())
1008 // If operation type is 'undesirable', e.g. i16 on x86, consider
1010 unsigned Opc = Op.getOpcode();
1011 if (TLI.isTypeDesirableForOp(Opc, VT))
1015 // Consult target whether it is a good idea to promote this operation and
1016 // what's the right type to promote it to.
1017 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1018 assert(PVT != VT && "Don't know what type to promote to!");
1020 bool Replace0 = false;
1021 SDValue N0 = Op.getOperand(0);
1022 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
1026 bool Replace1 = false;
1027 SDValue N1 = Op.getOperand(1);
1032 NN1 = PromoteOperand(N1, PVT, Replace1);
1037 AddToWorklist(NN0.getNode());
1039 AddToWorklist(NN1.getNode());
1042 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
1044 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
1046 DEBUG(dbgs() << "\nPromoting ";
1047 Op.getNode()->dump(&DAG));
1049 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1050 DAG.getNode(Opc, dl, PVT, NN0, NN1));
1055 /// Promote the specified integer shift operation if the target indicates it is
1056 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
1057 /// i32 since i16 instructions are longer.
1058 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
1059 if (!LegalOperations)
1062 EVT VT = Op.getValueType();
1063 if (VT.isVector() || !VT.isInteger())
1066 // If operation type is 'undesirable', e.g. i16 on x86, consider
1068 unsigned Opc = Op.getOpcode();
1069 if (TLI.isTypeDesirableForOp(Opc, VT))
1073 // Consult target whether it is a good idea to promote this operation and
1074 // what's the right type to promote it to.
1075 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1076 assert(PVT != VT && "Don't know what type to promote to!");
1078 bool Replace = false;
1079 SDValue N0 = Op.getOperand(0);
1080 if (Opc == ISD::SRA)
1081 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
1082 else if (Opc == ISD::SRL)
1083 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
1085 N0 = PromoteOperand(N0, PVT, Replace);
1089 AddToWorklist(N0.getNode());
1091 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
1093 DEBUG(dbgs() << "\nPromoting ";
1094 Op.getNode()->dump(&DAG));
1096 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1097 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
1102 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
1103 if (!LegalOperations)
1106 EVT VT = Op.getValueType();
1107 if (VT.isVector() || !VT.isInteger())
1110 // If operation type is 'undesirable', e.g. i16 on x86, consider
1112 unsigned Opc = Op.getOpcode();
1113 if (TLI.isTypeDesirableForOp(Opc, VT))
1117 // Consult target whether it is a good idea to promote this operation and
1118 // what's the right type to promote it to.
1119 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1120 assert(PVT != VT && "Don't know what type to promote to!");
1121 // fold (aext (aext x)) -> (aext x)
1122 // fold (aext (zext x)) -> (zext x)
1123 // fold (aext (sext x)) -> (sext x)
1124 DEBUG(dbgs() << "\nPromoting ";
1125 Op.getNode()->dump(&DAG));
1126 return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0));
1131 bool DAGCombiner::PromoteLoad(SDValue Op) {
1132 if (!LegalOperations)
1135 EVT VT = Op.getValueType();
1136 if (VT.isVector() || !VT.isInteger())
1139 // If operation type is 'undesirable', e.g. i16 on x86, consider
1141 unsigned Opc = Op.getOpcode();
1142 if (TLI.isTypeDesirableForOp(Opc, VT))
1146 // Consult target whether it is a good idea to promote this operation and
1147 // what's the right type to promote it to.
1148 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1149 assert(PVT != VT && "Don't know what type to promote to!");
1152 SDNode *N = Op.getNode();
1153 LoadSDNode *LD = cast<LoadSDNode>(N);
1154 EVT MemVT = LD->getMemoryVT();
1155 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
1156 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, PVT, MemVT) ? ISD::ZEXTLOAD
1158 : LD->getExtensionType();
1159 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
1160 LD->getChain(), LD->getBasePtr(),
1161 MemVT, LD->getMemOperand());
1162 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
1164 DEBUG(dbgs() << "\nPromoting ";
1167 Result.getNode()->dump(&DAG);
1169 WorklistRemover DeadNodes(*this);
1170 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
1171 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1));
1172 deleteAndRecombine(N);
1173 AddToWorklist(Result.getNode());
1179 /// \brief Recursively delete a node which has no uses and any operands for
1180 /// which it is the only use.
1182 /// Note that this both deletes the nodes and removes them from the worklist.
1183 /// It also adds any nodes who have had a user deleted to the worklist as they
1184 /// may now have only one use and subject to other combines.
1185 bool DAGCombiner::recursivelyDeleteUnusedNodes(SDNode *N) {
1186 if (!N->use_empty())
1189 SmallSetVector<SDNode *, 16> Nodes;
1192 N = Nodes.pop_back_val();
1196 if (N->use_empty()) {
1197 for (const SDValue &ChildN : N->op_values())
1198 Nodes.insert(ChildN.getNode());
1200 removeFromWorklist(N);
1205 } while (!Nodes.empty());
1209 //===----------------------------------------------------------------------===//
1210 // Main DAG Combiner implementation
1211 //===----------------------------------------------------------------------===//
1213 void DAGCombiner::Run(CombineLevel AtLevel) {
1214 // set the instance variables, so that the various visit routines may use it.
1216 LegalOperations = Level >= AfterLegalizeVectorOps;
1217 LegalTypes = Level >= AfterLegalizeTypes;
1219 // Add all the dag nodes to the worklist.
1220 for (SDNode &Node : DAG.allnodes())
1221 AddToWorklist(&Node);
1223 // Create a dummy node (which is not added to allnodes), that adds a reference
1224 // to the root node, preventing it from being deleted, and tracking any
1225 // changes of the root.
1226 HandleSDNode Dummy(DAG.getRoot());
1228 // while the worklist isn't empty, find a node and
1229 // try and combine it.
1230 while (!WorklistMap.empty()) {
1232 // The Worklist holds the SDNodes in order, but it may contain null entries.
1234 N = Worklist.pop_back_val();
1237 bool GoodWorklistEntry = WorklistMap.erase(N);
1238 (void)GoodWorklistEntry;
1239 assert(GoodWorklistEntry &&
1240 "Found a worklist entry without a corresponding map entry!");
1242 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1243 // N is deleted from the DAG, since they too may now be dead or may have a
1244 // reduced number of uses, allowing other xforms.
1245 if (recursivelyDeleteUnusedNodes(N))
1248 WorklistRemover DeadNodes(*this);
1250 // If this combine is running after legalizing the DAG, re-legalize any
1251 // nodes pulled off the worklist.
1252 if (Level == AfterLegalizeDAG) {
1253 SmallSetVector<SDNode *, 16> UpdatedNodes;
1254 bool NIsValid = DAG.LegalizeOp(N, UpdatedNodes);
1256 for (SDNode *LN : UpdatedNodes) {
1258 AddUsersToWorklist(LN);
1264 DEBUG(dbgs() << "\nCombining: "; N->dump(&DAG));
1266 // Add any operands of the new node which have not yet been combined to the
1267 // worklist as well. Because the worklist uniques things already, this
1268 // won't repeatedly process the same operand.
1269 CombinedNodes.insert(N);
1270 for (const SDValue &ChildN : N->op_values())
1271 if (!CombinedNodes.count(ChildN.getNode()))
1272 AddToWorklist(ChildN.getNode());
1274 SDValue RV = combine(N);
1281 // If we get back the same node we passed in, rather than a new node or
1282 // zero, we know that the node must have defined multiple values and
1283 // CombineTo was used. Since CombineTo takes care of the worklist
1284 // mechanics for us, we have no work to do in this case.
1285 if (RV.getNode() == N)
1288 assert(N->getOpcode() != ISD::DELETED_NODE &&
1289 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1290 "Node was deleted but visit returned new node!");
1292 DEBUG(dbgs() << " ... into: ";
1293 RV.getNode()->dump(&DAG));
1295 // Transfer debug value.
1296 DAG.TransferDbgValues(SDValue(N, 0), RV);
1297 if (N->getNumValues() == RV.getNode()->getNumValues())
1298 DAG.ReplaceAllUsesWith(N, RV.getNode());
1300 assert(N->getValueType(0) == RV.getValueType() &&
1301 N->getNumValues() == 1 && "Type mismatch");
1303 DAG.ReplaceAllUsesWith(N, &OpV);
1306 // Push the new node and any users onto the worklist
1307 AddToWorklist(RV.getNode());
1308 AddUsersToWorklist(RV.getNode());
1310 // Finally, if the node is now dead, remove it from the graph. The node
1311 // may not be dead if the replacement process recursively simplified to
1312 // something else needing this node. This will also take care of adding any
1313 // operands which have lost a user to the worklist.
1314 recursivelyDeleteUnusedNodes(N);
1317 // If the root changed (e.g. it was a dead load, update the root).
1318 DAG.setRoot(Dummy.getValue());
1319 DAG.RemoveDeadNodes();
1322 SDValue DAGCombiner::visit(SDNode *N) {
1323 switch (N->getOpcode()) {
1325 case ISD::TokenFactor: return visitTokenFactor(N);
1326 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1327 case ISD::ADD: return visitADD(N);
1328 case ISD::SUB: return visitSUB(N);
1329 case ISD::ADDC: return visitADDC(N);
1330 case ISD::SUBC: return visitSUBC(N);
1331 case ISD::ADDE: return visitADDE(N);
1332 case ISD::SUBE: return visitSUBE(N);
1333 case ISD::MUL: return visitMUL(N);
1334 case ISD::SDIV: return visitSDIV(N);
1335 case ISD::UDIV: return visitUDIV(N);
1336 case ISD::SREM: return visitSREM(N);
1337 case ISD::UREM: return visitUREM(N);
1338 case ISD::MULHU: return visitMULHU(N);
1339 case ISD::MULHS: return visitMULHS(N);
1340 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1341 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1342 case ISD::SMULO: return visitSMULO(N);
1343 case ISD::UMULO: return visitUMULO(N);
1344 case ISD::SDIVREM: return visitSDIVREM(N);
1345 case ISD::UDIVREM: return visitUDIVREM(N);
1349 case ISD::UMAX: return visitIMINMAX(N);
1350 case ISD::AND: return visitAND(N);
1351 case ISD::OR: return visitOR(N);
1352 case ISD::XOR: return visitXOR(N);
1353 case ISD::SHL: return visitSHL(N);
1354 case ISD::SRA: return visitSRA(N);
1355 case ISD::SRL: return visitSRL(N);
1357 case ISD::ROTL: return visitRotate(N);
1358 case ISD::BSWAP: return visitBSWAP(N);
1359 case ISD::CTLZ: return visitCTLZ(N);
1360 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1361 case ISD::CTTZ: return visitCTTZ(N);
1362 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1363 case ISD::CTPOP: return visitCTPOP(N);
1364 case ISD::SELECT: return visitSELECT(N);
1365 case ISD::VSELECT: return visitVSELECT(N);
1366 case ISD::SELECT_CC: return visitSELECT_CC(N);
1367 case ISD::SETCC: return visitSETCC(N);
1368 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1369 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1370 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1371 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1372 case ISD::SIGN_EXTEND_VECTOR_INREG: return visitSIGN_EXTEND_VECTOR_INREG(N);
1373 case ISD::TRUNCATE: return visitTRUNCATE(N);
1374 case ISD::BITCAST: return visitBITCAST(N);
1375 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1376 case ISD::FADD: return visitFADD(N);
1377 case ISD::FSUB: return visitFSUB(N);
1378 case ISD::FMUL: return visitFMUL(N);
1379 case ISD::FMA: return visitFMA(N);
1380 case ISD::FDIV: return visitFDIV(N);
1381 case ISD::FREM: return visitFREM(N);
1382 case ISD::FSQRT: return visitFSQRT(N);
1383 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1384 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1385 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1386 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1387 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1388 case ISD::FP_ROUND: return visitFP_ROUND(N);
1389 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1390 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1391 case ISD::FNEG: return visitFNEG(N);
1392 case ISD::FABS: return visitFABS(N);
1393 case ISD::FFLOOR: return visitFFLOOR(N);
1394 case ISD::FMINNUM: return visitFMINNUM(N);
1395 case ISD::FMAXNUM: return visitFMAXNUM(N);
1396 case ISD::FCEIL: return visitFCEIL(N);
1397 case ISD::FTRUNC: return visitFTRUNC(N);
1398 case ISD::BRCOND: return visitBRCOND(N);
1399 case ISD::BR_CC: return visitBR_CC(N);
1400 case ISD::LOAD: return visitLOAD(N);
1401 case ISD::STORE: return visitSTORE(N);
1402 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1403 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1404 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1405 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1406 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1407 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1408 case ISD::SCALAR_TO_VECTOR: return visitSCALAR_TO_VECTOR(N);
1409 case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N);
1410 case ISD::MGATHER: return visitMGATHER(N);
1411 case ISD::MLOAD: return visitMLOAD(N);
1412 case ISD::MSCATTER: return visitMSCATTER(N);
1413 case ISD::MSTORE: return visitMSTORE(N);
1414 case ISD::FP_TO_FP16: return visitFP_TO_FP16(N);
1415 case ISD::FP16_TO_FP: return visitFP16_TO_FP(N);
1420 SDValue DAGCombiner::combine(SDNode *N) {
1421 SDValue RV = visit(N);
1423 // If nothing happened, try a target-specific DAG combine.
1424 if (!RV.getNode()) {
1425 assert(N->getOpcode() != ISD::DELETED_NODE &&
1426 "Node was deleted but visit returned NULL!");
1428 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1429 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1431 // Expose the DAG combiner to the target combiner impls.
1432 TargetLowering::DAGCombinerInfo
1433 DagCombineInfo(DAG, Level, false, this);
1435 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1439 // If nothing happened still, try promoting the operation.
1440 if (!RV.getNode()) {
1441 switch (N->getOpcode()) {
1449 RV = PromoteIntBinOp(SDValue(N, 0));
1454 RV = PromoteIntShiftOp(SDValue(N, 0));
1456 case ISD::SIGN_EXTEND:
1457 case ISD::ZERO_EXTEND:
1458 case ISD::ANY_EXTEND:
1459 RV = PromoteExtend(SDValue(N, 0));
1462 if (PromoteLoad(SDValue(N, 0)))
1468 // If N is a commutative binary node, try commuting it to enable more
1470 if (!RV.getNode() && SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1471 N->getNumValues() == 1) {
1472 SDValue N0 = N->getOperand(0);
1473 SDValue N1 = N->getOperand(1);
1475 // Constant operands are canonicalized to RHS.
1476 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1477 SDValue Ops[] = {N1, N0};
1479 if (const auto *BinNode = dyn_cast<BinaryWithFlagsSDNode>(N)) {
1480 CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops,
1483 CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops);
1486 return SDValue(CSENode, 0);
1493 /// Given a node, return its input chain if it has one, otherwise return a null
1495 static SDValue getInputChainForNode(SDNode *N) {
1496 if (unsigned NumOps = N->getNumOperands()) {
1497 if (N->getOperand(0).getValueType() == MVT::Other)
1498 return N->getOperand(0);
1499 if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1500 return N->getOperand(NumOps-1);
1501 for (unsigned i = 1; i < NumOps-1; ++i)
1502 if (N->getOperand(i).getValueType() == MVT::Other)
1503 return N->getOperand(i);
1508 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1509 // If N has two operands, where one has an input chain equal to the other,
1510 // the 'other' chain is redundant.
1511 if (N->getNumOperands() == 2) {
1512 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1513 return N->getOperand(0);
1514 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1515 return N->getOperand(1);
1518 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1519 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1520 SmallPtrSet<SDNode*, 16> SeenOps;
1521 bool Changed = false; // If we should replace this token factor.
1523 // Start out with this token factor.
1526 // Iterate through token factors. The TFs grows when new token factors are
1528 for (unsigned i = 0; i < TFs.size(); ++i) {
1529 SDNode *TF = TFs[i];
1531 // Check each of the operands.
1532 for (const SDValue &Op : TF->op_values()) {
1534 switch (Op.getOpcode()) {
1535 case ISD::EntryToken:
1536 // Entry tokens don't need to be added to the list. They are
1541 case ISD::TokenFactor:
1542 if (Op.hasOneUse() &&
1543 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1544 // Queue up for processing.
1545 TFs.push_back(Op.getNode());
1546 // Clean up in case the token factor is removed.
1547 AddToWorklist(Op.getNode());
1554 // Only add if it isn't already in the list.
1555 if (SeenOps.insert(Op.getNode()).second)
1566 // If we've changed things around then replace token factor.
1569 // The entry token is the only possible outcome.
1570 Result = DAG.getEntryNode();
1572 // New and improved token factor.
1573 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Ops);
1576 // Add users to worklist if AA is enabled, since it may introduce
1577 // a lot of new chained token factors while removing memory deps.
1578 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
1579 : DAG.getSubtarget().useAA();
1580 return CombineTo(N, Result, UseAA /*add to worklist*/);
1586 /// MERGE_VALUES can always be eliminated.
1587 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1588 WorklistRemover DeadNodes(*this);
1589 // Replacing results may cause a different MERGE_VALUES to suddenly
1590 // be CSE'd with N, and carry its uses with it. Iterate until no
1591 // uses remain, to ensure that the node can be safely deleted.
1592 // First add the users of this node to the work list so that they
1593 // can be tried again once they have new operands.
1594 AddUsersToWorklist(N);
1596 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1597 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1598 } while (!N->use_empty());
1599 deleteAndRecombine(N);
1600 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1603 static bool isNullConstant(SDValue V) {
1604 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(V);
1605 return Const != nullptr && Const->isNullValue();
1608 static bool isNullFPConstant(SDValue V) {
1609 ConstantFPSDNode *Const = dyn_cast<ConstantFPSDNode>(V);
1610 return Const != nullptr && Const->isZero() && !Const->isNegative();
1613 static bool isAllOnesConstant(SDValue V) {
1614 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(V);
1615 return Const != nullptr && Const->isAllOnesValue();
1618 static bool isOneConstant(SDValue V) {
1619 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(V);
1620 return Const != nullptr && Const->isOne();
1623 /// If \p N is a ContantSDNode with isOpaque() == false return it casted to a
1624 /// ContantSDNode pointer else nullptr.
1625 static ConstantSDNode *getAsNonOpaqueConstant(SDValue N) {
1626 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N);
1627 return Const != nullptr && !Const->isOpaque() ? Const : nullptr;
1630 SDValue DAGCombiner::visitADD(SDNode *N) {
1631 SDValue N0 = N->getOperand(0);
1632 SDValue N1 = N->getOperand(1);
1633 EVT VT = N0.getValueType();
1636 if (VT.isVector()) {
1637 if (SDValue FoldedVOp = SimplifyVBinOp(N))
1640 // fold (add x, 0) -> x, vector edition
1641 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1643 if (ISD::isBuildVectorAllZeros(N0.getNode()))
1647 // fold (add x, undef) -> undef
1648 if (N0.getOpcode() == ISD::UNDEF)
1650 if (N1.getOpcode() == ISD::UNDEF)
1652 // fold (add c1, c2) -> c1+c2
1653 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
1654 ConstantSDNode *N1C = getAsNonOpaqueConstant(N1);
1656 return DAG.FoldConstantArithmetic(ISD::ADD, SDLoc(N), VT, N0C, N1C);
1657 // canonicalize constant to RHS
1658 if (isConstantIntBuildVectorOrConstantInt(N0) &&
1659 !isConstantIntBuildVectorOrConstantInt(N1))
1660 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, N0);
1661 // fold (add x, 0) -> x
1662 if (isNullConstant(N1))
1664 // fold (add Sym, c) -> Sym+c
1665 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1666 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1667 GA->getOpcode() == ISD::GlobalAddress)
1668 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1670 (uint64_t)N1C->getSExtValue());
1671 // fold ((c1-A)+c2) -> (c1+c2)-A
1672 if (N1C && N0.getOpcode() == ISD::SUB)
1673 if (ConstantSDNode *N0C = getAsNonOpaqueConstant(N0.getOperand(0))) {
1675 return DAG.getNode(ISD::SUB, DL, VT,
1676 DAG.getConstant(N1C->getAPIntValue()+
1677 N0C->getAPIntValue(), DL, VT),
1681 if (SDValue RADD = ReassociateOps(ISD::ADD, SDLoc(N), N0, N1))
1683 // fold ((0-A) + B) -> B-A
1684 if (N0.getOpcode() == ISD::SUB && isNullConstant(N0.getOperand(0)))
1685 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, N0.getOperand(1));
1686 // fold (A + (0-B)) -> A-B
1687 if (N1.getOpcode() == ISD::SUB && isNullConstant(N1.getOperand(0)))
1688 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1.getOperand(1));
1689 // fold (A+(B-A)) -> B
1690 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1691 return N1.getOperand(0);
1692 // fold ((B-A)+A) -> B
1693 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1694 return N0.getOperand(0);
1695 // fold (A+(B-(A+C))) to (B-C)
1696 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1697 N0 == N1.getOperand(1).getOperand(0))
1698 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1699 N1.getOperand(1).getOperand(1));
1700 // fold (A+(B-(C+A))) to (B-C)
1701 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1702 N0 == N1.getOperand(1).getOperand(1))
1703 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1704 N1.getOperand(1).getOperand(0));
1705 // fold (A+((B-A)+or-C)) to (B+or-C)
1706 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1707 N1.getOperand(0).getOpcode() == ISD::SUB &&
1708 N0 == N1.getOperand(0).getOperand(1))
1709 return DAG.getNode(N1.getOpcode(), SDLoc(N), VT,
1710 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1712 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1713 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1714 SDValue N00 = N0.getOperand(0);
1715 SDValue N01 = N0.getOperand(1);
1716 SDValue N10 = N1.getOperand(0);
1717 SDValue N11 = N1.getOperand(1);
1719 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1720 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1721 DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10),
1722 DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11));
1725 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1726 return SDValue(N, 0);
1728 // fold (a+b) -> (a|b) iff a and b share no bits.
1729 if (VT.isInteger() && !VT.isVector()) {
1730 APInt LHSZero, LHSOne;
1731 APInt RHSZero, RHSOne;
1732 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1734 if (LHSZero.getBoolValue()) {
1735 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1737 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1738 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1739 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero){
1740 if (!LegalOperations || TLI.isOperationLegal(ISD::OR, VT))
1741 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1);
1746 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1747 if (N1.getOpcode() == ISD::SHL && N1.getOperand(0).getOpcode() == ISD::SUB &&
1748 isNullConstant(N1.getOperand(0).getOperand(0)))
1749 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0,
1750 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1751 N1.getOperand(0).getOperand(1),
1753 if (N0.getOpcode() == ISD::SHL && N0.getOperand(0).getOpcode() == ISD::SUB &&
1754 isNullConstant(N0.getOperand(0).getOperand(0)))
1755 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1,
1756 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1757 N0.getOperand(0).getOperand(1),
1760 if (N1.getOpcode() == ISD::AND) {
1761 SDValue AndOp0 = N1.getOperand(0);
1762 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1763 unsigned DestBits = VT.getScalarType().getSizeInBits();
1765 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1766 // and similar xforms where the inner op is either ~0 or 0.
1767 if (NumSignBits == DestBits && isOneConstant(N1->getOperand(1))) {
1769 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1773 // add (sext i1), X -> sub X, (zext i1)
1774 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1775 N0.getOperand(0).getValueType() == MVT::i1 &&
1776 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1778 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1779 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1782 // add X, (sextinreg Y i1) -> sub X, (and Y 1)
1783 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
1784 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
1785 if (TN->getVT() == MVT::i1) {
1787 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
1788 DAG.getConstant(1, DL, VT));
1789 return DAG.getNode(ISD::SUB, DL, VT, N0, ZExt);
1796 SDValue DAGCombiner::visitADDC(SDNode *N) {
1797 SDValue N0 = N->getOperand(0);
1798 SDValue N1 = N->getOperand(1);
1799 EVT VT = N0.getValueType();
1801 // If the flag result is dead, turn this into an ADD.
1802 if (!N->hasAnyUseOfValue(1))
1803 return CombineTo(N, DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N1),
1804 DAG.getNode(ISD::CARRY_FALSE,
1805 SDLoc(N), MVT::Glue));
1807 // canonicalize constant to RHS.
1808 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1809 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1811 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N1, N0);
1813 // fold (addc x, 0) -> x + no carry out
1814 if (isNullConstant(N1))
1815 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1816 SDLoc(N), MVT::Glue));
1818 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1819 APInt LHSZero, LHSOne;
1820 APInt RHSZero, RHSOne;
1821 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1823 if (LHSZero.getBoolValue()) {
1824 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1826 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1827 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1828 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1829 return CombineTo(N, DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1),
1830 DAG.getNode(ISD::CARRY_FALSE,
1831 SDLoc(N), MVT::Glue));
1837 SDValue DAGCombiner::visitADDE(SDNode *N) {
1838 SDValue N0 = N->getOperand(0);
1839 SDValue N1 = N->getOperand(1);
1840 SDValue CarryIn = N->getOperand(2);
1842 // canonicalize constant to RHS
1843 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1844 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1846 return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(),
1849 // fold (adde x, y, false) -> (addc x, y)
1850 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1851 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1);
1856 // Since it may not be valid to emit a fold to zero for vector initializers
1857 // check if we can before folding.
1858 static SDValue tryFoldToZero(SDLoc DL, const TargetLowering &TLI, EVT VT,
1860 bool LegalOperations, bool LegalTypes) {
1862 return DAG.getConstant(0, DL, VT);
1863 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
1864 return DAG.getConstant(0, DL, VT);
1868 SDValue DAGCombiner::visitSUB(SDNode *N) {
1869 SDValue N0 = N->getOperand(0);
1870 SDValue N1 = N->getOperand(1);
1871 EVT VT = N0.getValueType();
1874 if (VT.isVector()) {
1875 if (SDValue FoldedVOp = SimplifyVBinOp(N))
1878 // fold (sub x, 0) -> x, vector edition
1879 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1883 // fold (sub x, x) -> 0
1884 // FIXME: Refactor this and xor and other similar operations together.
1886 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
1887 // fold (sub c1, c2) -> c1-c2
1888 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
1889 ConstantSDNode *N1C = getAsNonOpaqueConstant(N1);
1891 return DAG.FoldConstantArithmetic(ISD::SUB, SDLoc(N), VT, N0C, N1C);
1892 // fold (sub x, c) -> (add x, -c)
1895 return DAG.getNode(ISD::ADD, DL, VT, N0,
1896 DAG.getConstant(-N1C->getAPIntValue(), DL, VT));
1898 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1899 if (isAllOnesConstant(N0))
1900 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
1901 // fold A-(A-B) -> B
1902 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1903 return N1.getOperand(1);
1904 // fold (A+B)-A -> B
1905 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1906 return N0.getOperand(1);
1907 // fold (A+B)-B -> A
1908 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1909 return N0.getOperand(0);
1910 // fold C2-(A+C1) -> (C2-C1)-A
1911 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? nullptr :
1912 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1913 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1915 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(),
1917 return DAG.getNode(ISD::SUB, DL, VT, NewC,
1920 // fold ((A+(B+or-C))-B) -> A+or-C
1921 if (N0.getOpcode() == ISD::ADD &&
1922 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1923 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1924 N0.getOperand(1).getOperand(0) == N1)
1925 return DAG.getNode(N0.getOperand(1).getOpcode(), SDLoc(N), VT,
1926 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1927 // fold ((A+(C+B))-B) -> A+C
1928 if (N0.getOpcode() == ISD::ADD &&
1929 N0.getOperand(1).getOpcode() == ISD::ADD &&
1930 N0.getOperand(1).getOperand(1) == N1)
1931 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1932 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1933 // fold ((A-(B-C))-C) -> A-B
1934 if (N0.getOpcode() == ISD::SUB &&
1935 N0.getOperand(1).getOpcode() == ISD::SUB &&
1936 N0.getOperand(1).getOperand(1) == N1)
1937 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1938 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1940 // If either operand of a sub is undef, the result is undef
1941 if (N0.getOpcode() == ISD::UNDEF)
1943 if (N1.getOpcode() == ISD::UNDEF)
1946 // If the relocation model supports it, consider symbol offsets.
1947 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1948 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1949 // fold (sub Sym, c) -> Sym-c
1950 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1951 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1953 (uint64_t)N1C->getSExtValue());
1954 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1955 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1956 if (GA->getGlobal() == GB->getGlobal())
1957 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1961 // sub X, (sextinreg Y i1) -> add X, (and Y 1)
1962 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
1963 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
1964 if (TN->getVT() == MVT::i1) {
1966 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
1967 DAG.getConstant(1, DL, VT));
1968 return DAG.getNode(ISD::ADD, DL, VT, N0, ZExt);
1975 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1976 SDValue N0 = N->getOperand(0);
1977 SDValue N1 = N->getOperand(1);
1978 EVT VT = N0.getValueType();
1980 // If the flag result is dead, turn this into an SUB.
1981 if (!N->hasAnyUseOfValue(1))
1982 return CombineTo(N, DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1),
1983 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1986 // fold (subc x, x) -> 0 + no borrow
1989 return CombineTo(N, DAG.getConstant(0, DL, VT),
1990 DAG.getNode(ISD::CARRY_FALSE, DL,
1994 // fold (subc x, 0) -> x + no borrow
1995 if (isNullConstant(N1))
1996 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1999 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
2000 if (isAllOnesConstant(N0))
2001 return CombineTo(N, DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0),
2002 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
2008 SDValue DAGCombiner::visitSUBE(SDNode *N) {
2009 SDValue N0 = N->getOperand(0);
2010 SDValue N1 = N->getOperand(1);
2011 SDValue CarryIn = N->getOperand(2);
2013 // fold (sube x, y, false) -> (subc x, y)
2014 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
2015 return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1);
2020 SDValue DAGCombiner::visitMUL(SDNode *N) {
2021 SDValue N0 = N->getOperand(0);
2022 SDValue N1 = N->getOperand(1);
2023 EVT VT = N0.getValueType();
2025 // fold (mul x, undef) -> 0
2026 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2027 return DAG.getConstant(0, SDLoc(N), VT);
2029 bool N0IsConst = false;
2030 bool N1IsConst = false;
2031 bool N1IsOpaqueConst = false;
2032 bool N0IsOpaqueConst = false;
2033 APInt ConstValue0, ConstValue1;
2035 if (VT.isVector()) {
2036 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2039 N0IsConst = isConstantSplatVector(N0.getNode(), ConstValue0);
2040 N1IsConst = isConstantSplatVector(N1.getNode(), ConstValue1);
2042 N0IsConst = isa<ConstantSDNode>(N0);
2044 ConstValue0 = cast<ConstantSDNode>(N0)->getAPIntValue();
2045 N0IsOpaqueConst = cast<ConstantSDNode>(N0)->isOpaque();
2047 N1IsConst = isa<ConstantSDNode>(N1);
2049 ConstValue1 = cast<ConstantSDNode>(N1)->getAPIntValue();
2050 N1IsOpaqueConst = cast<ConstantSDNode>(N1)->isOpaque();
2054 // fold (mul c1, c2) -> c1*c2
2055 if (N0IsConst && N1IsConst && !N0IsOpaqueConst && !N1IsOpaqueConst)
2056 return DAG.FoldConstantArithmetic(ISD::MUL, SDLoc(N), VT,
2057 N0.getNode(), N1.getNode());
2059 // canonicalize constant to RHS (vector doesn't have to splat)
2060 if (isConstantIntBuildVectorOrConstantInt(N0) &&
2061 !isConstantIntBuildVectorOrConstantInt(N1))
2062 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0);
2063 // fold (mul x, 0) -> 0
2064 if (N1IsConst && ConstValue1 == 0)
2066 // We require a splat of the entire scalar bit width for non-contiguous
2069 ConstValue1.getBitWidth() == VT.getScalarType().getSizeInBits();
2070 // fold (mul x, 1) -> x
2071 if (N1IsConst && ConstValue1 == 1 && IsFullSplat)
2073 // fold (mul x, -1) -> 0-x
2074 if (N1IsConst && ConstValue1.isAllOnesValue()) {
2076 return DAG.getNode(ISD::SUB, DL, VT,
2077 DAG.getConstant(0, DL, VT), N0);
2079 // fold (mul x, (1 << c)) -> x << c
2080 if (N1IsConst && !N1IsOpaqueConst && ConstValue1.isPowerOf2() &&
2083 return DAG.getNode(ISD::SHL, DL, VT, N0,
2084 DAG.getConstant(ConstValue1.logBase2(), DL,
2085 getShiftAmountTy(N0.getValueType())));
2087 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
2088 if (N1IsConst && !N1IsOpaqueConst && (-ConstValue1).isPowerOf2() &&
2090 unsigned Log2Val = (-ConstValue1).logBase2();
2092 // FIXME: If the input is something that is easily negated (e.g. a
2093 // single-use add), we should put the negate there.
2094 return DAG.getNode(ISD::SUB, DL, VT,
2095 DAG.getConstant(0, DL, VT),
2096 DAG.getNode(ISD::SHL, DL, VT, N0,
2097 DAG.getConstant(Log2Val, DL,
2098 getShiftAmountTy(N0.getValueType()))));
2102 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
2103 if (N1IsConst && N0.getOpcode() == ISD::SHL &&
2104 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2105 isa<ConstantSDNode>(N0.getOperand(1)))) {
2106 SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT,
2107 N1, N0.getOperand(1));
2108 AddToWorklist(C3.getNode());
2109 return DAG.getNode(ISD::MUL, SDLoc(N), VT,
2110 N0.getOperand(0), C3);
2113 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
2116 SDValue Sh(nullptr,0), Y(nullptr,0);
2117 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
2118 if (N0.getOpcode() == ISD::SHL &&
2119 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2120 isa<ConstantSDNode>(N0.getOperand(1))) &&
2121 N0.getNode()->hasOneUse()) {
2123 } else if (N1.getOpcode() == ISD::SHL &&
2124 isa<ConstantSDNode>(N1.getOperand(1)) &&
2125 N1.getNode()->hasOneUse()) {
2130 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2131 Sh.getOperand(0), Y);
2132 return DAG.getNode(ISD::SHL, SDLoc(N), VT,
2133 Mul, Sh.getOperand(1));
2137 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
2138 if (N1IsConst && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
2139 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2140 isa<ConstantSDNode>(N0.getOperand(1))))
2141 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
2142 DAG.getNode(ISD::MUL, SDLoc(N0), VT,
2143 N0.getOperand(0), N1),
2144 DAG.getNode(ISD::MUL, SDLoc(N1), VT,
2145 N0.getOperand(1), N1));
2148 if (SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1))
2154 SDValue DAGCombiner::visitSDIV(SDNode *N) {
2155 SDValue N0 = N->getOperand(0);
2156 SDValue N1 = N->getOperand(1);
2157 EVT VT = N->getValueType(0);
2161 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2164 // fold (sdiv c1, c2) -> c1/c2
2165 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2166 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2167 if (N0C && N1C && !N0C->isOpaque() && !N1C->isOpaque())
2168 return DAG.FoldConstantArithmetic(ISD::SDIV, SDLoc(N), VT, N0C, N1C);
2169 // fold (sdiv X, 1) -> X
2170 if (N1C && N1C->isOne())
2172 // fold (sdiv X, -1) -> 0-X
2173 if (N1C && N1C->isAllOnesValue()) {
2175 return DAG.getNode(ISD::SUB, DL, VT,
2176 DAG.getConstant(0, DL, VT), N0);
2178 // If we know the sign bits of both operands are zero, strength reduce to a
2179 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
2180 if (!VT.isVector()) {
2181 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2182 return DAG.getNode(ISD::UDIV, SDLoc(N), N1.getValueType(),
2186 // fold (sdiv X, pow2) -> simple ops after legalize
2187 // FIXME: We check for the exact bit here because the generic lowering gives
2188 // better results in that case. The target-specific lowering should learn how
2189 // to handle exact sdivs efficiently.
2190 if (N1C && !N1C->isNullValue() && !N1C->isOpaque() &&
2191 !cast<BinaryWithFlagsSDNode>(N)->Flags.hasExact() &&
2192 (N1C->getAPIntValue().isPowerOf2() ||
2193 (-N1C->getAPIntValue()).isPowerOf2())) {
2194 // Target-specific implementation of sdiv x, pow2.
2195 if (SDValue Res = BuildSDIVPow2(N))
2198 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
2201 // Splat the sign bit into the register
2203 DAG.getNode(ISD::SRA, DL, VT, N0,
2204 DAG.getConstant(VT.getScalarSizeInBits() - 1, DL,
2205 getShiftAmountTy(N0.getValueType())));
2206 AddToWorklist(SGN.getNode());
2208 // Add (N0 < 0) ? abs2 - 1 : 0;
2210 DAG.getNode(ISD::SRL, DL, VT, SGN,
2211 DAG.getConstant(VT.getScalarSizeInBits() - lg2, DL,
2212 getShiftAmountTy(SGN.getValueType())));
2213 SDValue ADD = DAG.getNode(ISD::ADD, DL, VT, N0, SRL);
2214 AddToWorklist(SRL.getNode());
2215 AddToWorklist(ADD.getNode()); // Divide by pow2
2216 SDValue SRA = DAG.getNode(ISD::SRA, DL, VT, ADD,
2217 DAG.getConstant(lg2, DL,
2218 getShiftAmountTy(ADD.getValueType())));
2220 // If we're dividing by a positive value, we're done. Otherwise, we must
2221 // negate the result.
2222 if (N1C->getAPIntValue().isNonNegative())
2225 AddToWorklist(SRA.getNode());
2226 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), SRA);
2229 // If integer divide is expensive and we satisfy the requirements, emit an
2230 // alternate sequence. Targets may check function attributes for size/speed
2232 AttributeSet Attr = DAG.getMachineFunction().getFunction()->getAttributes();
2233 if (N1C && !TLI.isIntDivCheap(N->getValueType(0), Attr))
2234 if (SDValue Op = BuildSDIV(N))
2238 if (N0.getOpcode() == ISD::UNDEF)
2239 return DAG.getConstant(0, SDLoc(N), VT);
2240 // X / undef -> undef
2241 if (N1.getOpcode() == ISD::UNDEF)
2247 SDValue DAGCombiner::visitUDIV(SDNode *N) {
2248 SDValue N0 = N->getOperand(0);
2249 SDValue N1 = N->getOperand(1);
2250 EVT VT = N->getValueType(0);
2254 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2257 // fold (udiv c1, c2) -> c1/c2
2258 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2259 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2261 if (SDValue Folded = DAG.FoldConstantArithmetic(ISD::UDIV, SDLoc(N), VT,
2264 // fold (udiv x, (1 << c)) -> x >>u c
2265 if (N1C && !N1C->isOpaque() && N1C->getAPIntValue().isPowerOf2()) {
2267 return DAG.getNode(ISD::SRL, DL, VT, N0,
2268 DAG.getConstant(N1C->getAPIntValue().logBase2(), DL,
2269 getShiftAmountTy(N0.getValueType())));
2271 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
2272 if (N1.getOpcode() == ISD::SHL) {
2273 if (ConstantSDNode *SHC = getAsNonOpaqueConstant(N1.getOperand(0))) {
2274 if (SHC->getAPIntValue().isPowerOf2()) {
2275 EVT ADDVT = N1.getOperand(1).getValueType();
2277 SDValue Add = DAG.getNode(ISD::ADD, DL, ADDVT,
2279 DAG.getConstant(SHC->getAPIntValue()
2282 AddToWorklist(Add.getNode());
2283 return DAG.getNode(ISD::SRL, DL, VT, N0, Add);
2288 // fold (udiv x, c) -> alternate
2289 AttributeSet Attr = DAG.getMachineFunction().getFunction()->getAttributes();
2290 if (N1C && !TLI.isIntDivCheap(N->getValueType(0), Attr))
2291 if (SDValue Op = BuildUDIV(N))
2295 if (N0.getOpcode() == ISD::UNDEF)
2296 return DAG.getConstant(0, SDLoc(N), VT);
2297 // X / undef -> undef
2298 if (N1.getOpcode() == ISD::UNDEF)
2304 SDValue DAGCombiner::visitSREM(SDNode *N) {
2305 SDValue N0 = N->getOperand(0);
2306 SDValue N1 = N->getOperand(1);
2307 EVT VT = N->getValueType(0);
2309 // fold (srem c1, c2) -> c1%c2
2310 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2311 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2313 if (SDValue Folded = DAG.FoldConstantArithmetic(ISD::SREM, SDLoc(N), VT,
2316 // If we know the sign bits of both operands are zero, strength reduce to a
2317 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
2318 if (!VT.isVector()) {
2319 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2320 return DAG.getNode(ISD::UREM, SDLoc(N), VT, N0, N1);
2323 // If X/C can be simplified by the division-by-constant logic, lower
2324 // X%C to the equivalent of X-X/C*C.
2325 if (N1C && !N1C->isNullValue()) {
2326 SDValue Div = DAG.getNode(ISD::SDIV, SDLoc(N), VT, N0, N1);
2327 AddToWorklist(Div.getNode());
2328 SDValue OptimizedDiv = combine(Div.getNode());
2329 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2330 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2332 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2333 AddToWorklist(Mul.getNode());
2339 if (N0.getOpcode() == ISD::UNDEF)
2340 return DAG.getConstant(0, SDLoc(N), VT);
2341 // X % undef -> undef
2342 if (N1.getOpcode() == ISD::UNDEF)
2348 SDValue DAGCombiner::visitUREM(SDNode *N) {
2349 SDValue N0 = N->getOperand(0);
2350 SDValue N1 = N->getOperand(1);
2351 EVT VT = N->getValueType(0);
2353 // fold (urem c1, c2) -> c1%c2
2354 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2355 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2357 if (SDValue Folded = DAG.FoldConstantArithmetic(ISD::UREM, SDLoc(N), VT,
2360 // fold (urem x, pow2) -> (and x, pow2-1)
2361 if (N1C && !N1C->isNullValue() && !N1C->isOpaque() &&
2362 N1C->getAPIntValue().isPowerOf2()) {
2364 return DAG.getNode(ISD::AND, DL, VT, N0,
2365 DAG.getConstant(N1C->getAPIntValue() - 1, DL, VT));
2367 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2368 if (N1.getOpcode() == ISD::SHL) {
2369 if (ConstantSDNode *SHC = getAsNonOpaqueConstant(N1.getOperand(0))) {
2370 if (SHC->getAPIntValue().isPowerOf2()) {
2373 DAG.getNode(ISD::ADD, DL, VT, N1,
2374 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL,
2376 AddToWorklist(Add.getNode());
2377 return DAG.getNode(ISD::AND, DL, VT, N0, Add);
2382 // If X/C can be simplified by the division-by-constant logic, lower
2383 // X%C to the equivalent of X-X/C*C.
2384 if (N1C && !N1C->isNullValue()) {
2385 SDValue Div = DAG.getNode(ISD::UDIV, SDLoc(N), VT, N0, N1);
2386 AddToWorklist(Div.getNode());
2387 SDValue OptimizedDiv = combine(Div.getNode());
2388 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2389 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2391 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2392 AddToWorklist(Mul.getNode());
2398 if (N0.getOpcode() == ISD::UNDEF)
2399 return DAG.getConstant(0, SDLoc(N), VT);
2400 // X % undef -> undef
2401 if (N1.getOpcode() == ISD::UNDEF)
2407 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2408 SDValue N0 = N->getOperand(0);
2409 SDValue N1 = N->getOperand(1);
2410 EVT VT = N->getValueType(0);
2413 // fold (mulhs x, 0) -> 0
2414 if (isNullConstant(N1))
2416 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2417 if (isOneConstant(N1)) {
2419 return DAG.getNode(ISD::SRA, DL, N0.getValueType(), N0,
2420 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2422 getShiftAmountTy(N0.getValueType())));
2424 // fold (mulhs x, undef) -> 0
2425 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2426 return DAG.getConstant(0, SDLoc(N), VT);
2428 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2430 if (VT.isSimple() && !VT.isVector()) {
2431 MVT Simple = VT.getSimpleVT();
2432 unsigned SimpleSize = Simple.getSizeInBits();
2433 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2434 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2435 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2436 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2437 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2438 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2439 DAG.getConstant(SimpleSize, DL,
2440 getShiftAmountTy(N1.getValueType())));
2441 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2448 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2449 SDValue N0 = N->getOperand(0);
2450 SDValue N1 = N->getOperand(1);
2451 EVT VT = N->getValueType(0);
2454 // fold (mulhu x, 0) -> 0
2455 if (isNullConstant(N1))
2457 // fold (mulhu x, 1) -> 0
2458 if (isOneConstant(N1))
2459 return DAG.getConstant(0, DL, N0.getValueType());
2460 // fold (mulhu x, undef) -> 0
2461 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2462 return DAG.getConstant(0, DL, VT);
2464 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2466 if (VT.isSimple() && !VT.isVector()) {
2467 MVT Simple = VT.getSimpleVT();
2468 unsigned SimpleSize = Simple.getSizeInBits();
2469 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2470 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2471 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2472 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2473 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2474 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2475 DAG.getConstant(SimpleSize, DL,
2476 getShiftAmountTy(N1.getValueType())));
2477 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2484 /// Perform optimizations common to nodes that compute two values. LoOp and HiOp
2485 /// give the opcodes for the two computations that are being performed. Return
2486 /// true if a simplification was made.
2487 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2489 // If the high half is not needed, just compute the low half.
2490 bool HiExists = N->hasAnyUseOfValue(1);
2492 (!LegalOperations ||
2493 TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) {
2494 SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2495 return CombineTo(N, Res, Res);
2498 // If the low half is not needed, just compute the high half.
2499 bool LoExists = N->hasAnyUseOfValue(0);
2501 (!LegalOperations ||
2502 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2503 SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2504 return CombineTo(N, Res, Res);
2507 // If both halves are used, return as it is.
2508 if (LoExists && HiExists)
2511 // If the two computed results can be simplified separately, separate them.
2513 SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2514 AddToWorklist(Lo.getNode());
2515 SDValue LoOpt = combine(Lo.getNode());
2516 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2517 (!LegalOperations ||
2518 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2519 return CombineTo(N, LoOpt, LoOpt);
2523 SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2524 AddToWorklist(Hi.getNode());
2525 SDValue HiOpt = combine(Hi.getNode());
2526 if (HiOpt.getNode() && HiOpt != Hi &&
2527 (!LegalOperations ||
2528 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2529 return CombineTo(N, HiOpt, HiOpt);
2535 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2536 if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS))
2539 EVT VT = N->getValueType(0);
2542 // If the type is twice as wide is legal, transform the mulhu to a wider
2543 // multiply plus a shift.
2544 if (VT.isSimple() && !VT.isVector()) {
2545 MVT Simple = VT.getSimpleVT();
2546 unsigned SimpleSize = Simple.getSizeInBits();
2547 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2548 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2549 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2550 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2551 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2552 // Compute the high part as N1.
2553 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2554 DAG.getConstant(SimpleSize, DL,
2555 getShiftAmountTy(Lo.getValueType())));
2556 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2557 // Compute the low part as N0.
2558 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2559 return CombineTo(N, Lo, Hi);
2566 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2567 if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU))
2570 EVT VT = N->getValueType(0);
2573 // If the type is twice as wide is legal, transform the mulhu to a wider
2574 // multiply plus a shift.
2575 if (VT.isSimple() && !VT.isVector()) {
2576 MVT Simple = VT.getSimpleVT();
2577 unsigned SimpleSize = Simple.getSizeInBits();
2578 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2579 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2580 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2581 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2582 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2583 // Compute the high part as N1.
2584 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2585 DAG.getConstant(SimpleSize, DL,
2586 getShiftAmountTy(Lo.getValueType())));
2587 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2588 // Compute the low part as N0.
2589 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2590 return CombineTo(N, Lo, Hi);
2597 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2598 // (smulo x, 2) -> (saddo x, x)
2599 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2600 if (C2->getAPIntValue() == 2)
2601 return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(),
2602 N->getOperand(0), N->getOperand(0));
2607 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2608 // (umulo x, 2) -> (uaddo x, x)
2609 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2610 if (C2->getAPIntValue() == 2)
2611 return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(),
2612 N->getOperand(0), N->getOperand(0));
2617 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2618 if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM))
2624 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2625 if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM))
2631 SDValue DAGCombiner::visitIMINMAX(SDNode *N) {
2632 SDValue N0 = N->getOperand(0);
2633 SDValue N1 = N->getOperand(1);
2634 EVT VT = N0.getValueType();
2638 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2641 // fold (add c1, c2) -> c1+c2
2642 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
2643 ConstantSDNode *N1C = getAsNonOpaqueConstant(N1);
2645 return DAG.FoldConstantArithmetic(N->getOpcode(), SDLoc(N), VT, N0C, N1C);
2647 // canonicalize constant to RHS
2648 if (isConstantIntBuildVectorOrConstantInt(N0) &&
2649 !isConstantIntBuildVectorOrConstantInt(N1))
2650 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, N1, N0);
2655 /// If this is a binary operator with two operands of the same opcode, try to
2657 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2658 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2659 EVT VT = N0.getValueType();
2660 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2662 // Bail early if none of these transforms apply.
2663 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2665 // For each of OP in AND/OR/XOR:
2666 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2667 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2668 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2669 // fold (OP (bswap x), (bswap y)) -> (bswap (OP x, y))
2670 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2672 // do not sink logical op inside of a vector extend, since it may combine
2674 EVT Op0VT = N0.getOperand(0).getValueType();
2675 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2676 N0.getOpcode() == ISD::SIGN_EXTEND ||
2677 N0.getOpcode() == ISD::BSWAP ||
2678 // Avoid infinite looping with PromoteIntBinOp.
2679 (N0.getOpcode() == ISD::ANY_EXTEND &&
2680 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2681 (N0.getOpcode() == ISD::TRUNCATE &&
2682 (!TLI.isZExtFree(VT, Op0VT) ||
2683 !TLI.isTruncateFree(Op0VT, VT)) &&
2684 TLI.isTypeLegal(Op0VT))) &&
2686 Op0VT == N1.getOperand(0).getValueType() &&
2687 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2688 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2689 N0.getOperand(0).getValueType(),
2690 N0.getOperand(0), N1.getOperand(0));
2691 AddToWorklist(ORNode.getNode());
2692 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode);
2695 // For each of OP in SHL/SRL/SRA/AND...
2696 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2697 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2698 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2699 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2700 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2701 N0.getOperand(1) == N1.getOperand(1)) {
2702 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2703 N0.getOperand(0).getValueType(),
2704 N0.getOperand(0), N1.getOperand(0));
2705 AddToWorklist(ORNode.getNode());
2706 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
2707 ORNode, N0.getOperand(1));
2710 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2711 // Only perform this optimization after type legalization and before
2712 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2713 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2714 // we don't want to undo this promotion.
2715 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2717 if ((N0.getOpcode() == ISD::BITCAST ||
2718 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) &&
2719 Level == AfterLegalizeTypes) {
2720 SDValue In0 = N0.getOperand(0);
2721 SDValue In1 = N1.getOperand(0);
2722 EVT In0Ty = In0.getValueType();
2723 EVT In1Ty = In1.getValueType();
2725 // If both incoming values are integers, and the original types are the
2727 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2728 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1);
2729 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op);
2730 AddToWorklist(Op.getNode());
2735 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2736 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2737 // If both shuffles use the same mask, and both shuffle within a single
2738 // vector, then it is worthwhile to move the swizzle after the operation.
2739 // The type-legalizer generates this pattern when loading illegal
2740 // vector types from memory. In many cases this allows additional shuffle
2742 // There are other cases where moving the shuffle after the xor/and/or
2743 // is profitable even if shuffles don't perform a swizzle.
2744 // If both shuffles use the same mask, and both shuffles have the same first
2745 // or second operand, then it might still be profitable to move the shuffle
2746 // after the xor/and/or operation.
2747 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) {
2748 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2749 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2751 assert(N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() &&
2752 "Inputs to shuffles are not the same type");
2754 // Check that both shuffles use the same mask. The masks are known to be of
2755 // the same length because the result vector type is the same.
2756 // Check also that shuffles have only one use to avoid introducing extra
2758 if (SVN0->hasOneUse() && SVN1->hasOneUse() &&
2759 SVN0->getMask().equals(SVN1->getMask())) {
2760 SDValue ShOp = N0->getOperand(1);
2762 // Don't try to fold this node if it requires introducing a
2763 // build vector of all zeros that might be illegal at this stage.
2764 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2766 ShOp = DAG.getConstant(0, SDLoc(N), VT);
2771 // (AND (shuf (A, C), shuf (B, C)) -> shuf (AND (A, B), C)
2772 // (OR (shuf (A, C), shuf (B, C)) -> shuf (OR (A, B), C)
2773 // (XOR (shuf (A, C), shuf (B, C)) -> shuf (XOR (A, B), V_0)
2774 if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) {
2775 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2776 N0->getOperand(0), N1->getOperand(0));
2777 AddToWorklist(NewNode.getNode());
2778 return DAG.getVectorShuffle(VT, SDLoc(N), NewNode, ShOp,
2779 &SVN0->getMask()[0]);
2782 // Don't try to fold this node if it requires introducing a
2783 // build vector of all zeros that might be illegal at this stage.
2784 ShOp = N0->getOperand(0);
2785 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2787 ShOp = DAG.getConstant(0, SDLoc(N), VT);
2792 // (AND (shuf (C, A), shuf (C, B)) -> shuf (C, AND (A, B))
2793 // (OR (shuf (C, A), shuf (C, B)) -> shuf (C, OR (A, B))
2794 // (XOR (shuf (C, A), shuf (C, B)) -> shuf (V_0, XOR (A, B))
2795 if (N0->getOperand(0) == N1->getOperand(0) && ShOp.getNode()) {
2796 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2797 N0->getOperand(1), N1->getOperand(1));
2798 AddToWorklist(NewNode.getNode());
2799 return DAG.getVectorShuffle(VT, SDLoc(N), ShOp, NewNode,
2800 &SVN0->getMask()[0]);
2808 /// This contains all DAGCombine rules which reduce two values combined by
2809 /// an And operation to a single value. This makes them reusable in the context
2810 /// of visitSELECT(). Rules involving constants are not included as
2811 /// visitSELECT() already handles those cases.
2812 SDValue DAGCombiner::visitANDLike(SDValue N0, SDValue N1,
2813 SDNode *LocReference) {
2814 EVT VT = N1.getValueType();
2816 // fold (and x, undef) -> 0
2817 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2818 return DAG.getConstant(0, SDLoc(LocReference), VT);
2819 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2820 SDValue LL, LR, RL, RR, CC0, CC1;
2821 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2822 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2823 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2825 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2826 LL.getValueType().isInteger()) {
2827 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2828 if (isNullConstant(LR) && Op1 == ISD::SETEQ) {
2829 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2830 LR.getValueType(), LL, RL);
2831 AddToWorklist(ORNode.getNode());
2832 return DAG.getSetCC(SDLoc(LocReference), VT, ORNode, LR, Op1);
2834 if (isAllOnesConstant(LR)) {
2835 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2836 if (Op1 == ISD::SETEQ) {
2837 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(N0),
2838 LR.getValueType(), LL, RL);
2839 AddToWorklist(ANDNode.getNode());
2840 return DAG.getSetCC(SDLoc(LocReference), VT, ANDNode, LR, Op1);
2842 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2843 if (Op1 == ISD::SETGT) {
2844 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2845 LR.getValueType(), LL, RL);
2846 AddToWorklist(ORNode.getNode());
2847 return DAG.getSetCC(SDLoc(LocReference), VT, ORNode, LR, Op1);
2851 // Simplify (and (setne X, 0), (setne X, -1)) -> (setuge (add X, 1), 2)
2852 if (LL == RL && isa<ConstantSDNode>(LR) && isa<ConstantSDNode>(RR) &&
2853 Op0 == Op1 && LL.getValueType().isInteger() &&
2854 Op0 == ISD::SETNE && ((isNullConstant(LR) && isAllOnesConstant(RR)) ||
2855 (isAllOnesConstant(LR) && isNullConstant(RR)))) {
2857 SDValue ADDNode = DAG.getNode(ISD::ADD, DL, LL.getValueType(),
2858 LL, DAG.getConstant(1, DL,
2859 LL.getValueType()));
2860 AddToWorklist(ADDNode.getNode());
2861 return DAG.getSetCC(SDLoc(LocReference), VT, ADDNode,
2862 DAG.getConstant(2, DL, LL.getValueType()),
2865 // canonicalize equivalent to ll == rl
2866 if (LL == RR && LR == RL) {
2867 Op1 = ISD::getSetCCSwappedOperands(Op1);
2870 if (LL == RL && LR == RR) {
2871 bool isInteger = LL.getValueType().isInteger();
2872 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2873 if (Result != ISD::SETCC_INVALID &&
2874 (!LegalOperations ||
2875 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
2876 TLI.isOperationLegal(ISD::SETCC,
2877 getSetCCResultType(N0.getSimpleValueType())))))
2878 return DAG.getSetCC(SDLoc(LocReference), N0.getValueType(),
2883 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
2884 VT.getSizeInBits() <= 64) {
2885 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2886 APInt ADDC = ADDI->getAPIntValue();
2887 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2888 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
2889 // immediate for an add, but it is legal if its top c2 bits are set,
2890 // transform the ADD so the immediate doesn't need to be materialized
2892 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
2893 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
2894 SRLI->getZExtValue());
2895 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
2897 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2900 DAG.getNode(ISD::ADD, DL, VT,
2901 N0.getOperand(0), DAG.getConstant(ADDC, DL, VT));
2902 CombineTo(N0.getNode(), NewAdd);
2903 // Return N so it doesn't get rechecked!
2904 return SDValue(LocReference, 0);
2915 SDValue DAGCombiner::visitAND(SDNode *N) {
2916 SDValue N0 = N->getOperand(0);
2917 SDValue N1 = N->getOperand(1);
2918 EVT VT = N1.getValueType();
2921 if (VT.isVector()) {
2922 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2925 // fold (and x, 0) -> 0, vector edition
2926 if (ISD::isBuildVectorAllZeros(N0.getNode()))
2927 // do not return N0, because undef node may exist in N0
2928 return DAG.getConstant(
2929 APInt::getNullValue(
2930 N0.getValueType().getScalarType().getSizeInBits()),
2931 SDLoc(N), N0.getValueType());
2932 if (ISD::isBuildVectorAllZeros(N1.getNode()))
2933 // do not return N1, because undef node may exist in N1
2934 return DAG.getConstant(
2935 APInt::getNullValue(
2936 N1.getValueType().getScalarType().getSizeInBits()),
2937 SDLoc(N), N1.getValueType());
2939 // fold (and x, -1) -> x, vector edition
2940 if (ISD::isBuildVectorAllOnes(N0.getNode()))
2942 if (ISD::isBuildVectorAllOnes(N1.getNode()))
2946 // fold (and c1, c2) -> c1&c2
2947 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
2948 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2949 if (N0C && N1C && !N1C->isOpaque())
2950 return DAG.FoldConstantArithmetic(ISD::AND, SDLoc(N), VT, N0C, N1C);
2951 // canonicalize constant to RHS
2952 if (isConstantIntBuildVectorOrConstantInt(N0) &&
2953 !isConstantIntBuildVectorOrConstantInt(N1))
2954 return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0);
2955 // fold (and x, -1) -> x
2956 if (isAllOnesConstant(N1))
2958 // if (and x, c) is known to be zero, return 0
2959 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2960 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2961 APInt::getAllOnesValue(BitWidth)))
2962 return DAG.getConstant(0, SDLoc(N), VT);
2964 if (SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1))
2966 // fold (and (or x, C), D) -> D if (C & D) == D
2967 if (N1C && N0.getOpcode() == ISD::OR)
2968 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2969 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2971 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2972 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2973 SDValue N0Op0 = N0.getOperand(0);
2974 APInt Mask = ~N1C->getAPIntValue();
2975 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2976 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2977 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
2978 N0.getValueType(), N0Op0);
2980 // Replace uses of the AND with uses of the Zero extend node.
2983 // We actually want to replace all uses of the any_extend with the
2984 // zero_extend, to avoid duplicating things. This will later cause this
2985 // AND to be folded.
2986 CombineTo(N0.getNode(), Zext);
2987 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2990 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2991 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2992 // already be zero by virtue of the width of the base type of the load.
2994 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2996 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2997 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2998 N0.getOpcode() == ISD::LOAD) {
2999 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
3000 N0 : N0.getOperand(0) );
3002 // Get the constant (if applicable) the zero'th operand is being ANDed with.
3003 // This can be a pure constant or a vector splat, in which case we treat the
3004 // vector as a scalar and use the splat value.
3005 APInt Constant = APInt::getNullValue(1);
3006 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
3007 Constant = C->getAPIntValue();
3008 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
3009 APInt SplatValue, SplatUndef;
3010 unsigned SplatBitSize;
3012 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
3013 SplatBitSize, HasAnyUndefs);
3015 // Undef bits can contribute to a possible optimisation if set, so
3017 SplatValue |= SplatUndef;
3019 // The splat value may be something like "0x00FFFFFF", which means 0 for
3020 // the first vector value and FF for the rest, repeating. We need a mask
3021 // that will apply equally to all members of the vector, so AND all the
3022 // lanes of the constant together.
3023 EVT VT = Vector->getValueType(0);
3024 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
3026 // If the splat value has been compressed to a bitlength lower
3027 // than the size of the vector lane, we need to re-expand it to
3029 if (BitWidth > SplatBitSize)
3030 for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
3031 SplatBitSize < BitWidth;
3032 SplatBitSize = SplatBitSize * 2)
3033 SplatValue |= SplatValue.shl(SplatBitSize);
3035 // Make sure that variable 'Constant' is only set if 'SplatBitSize' is a
3036 // multiple of 'BitWidth'. Otherwise, we could propagate a wrong value.
3037 if (SplatBitSize % BitWidth == 0) {
3038 Constant = APInt::getAllOnesValue(BitWidth);
3039 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
3040 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
3045 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
3046 // actually legal and isn't going to get expanded, else this is a false
3048 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
3049 Load->getValueType(0),
3050 Load->getMemoryVT());
3052 // Resize the constant to the same size as the original memory access before
3053 // extension. If it is still the AllOnesValue then this AND is completely
3056 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
3059 switch (Load->getExtensionType()) {
3060 default: B = false; break;
3061 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
3063 case ISD::NON_EXTLOAD: B = true; break;
3066 if (B && Constant.isAllOnesValue()) {
3067 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
3068 // preserve semantics once we get rid of the AND.
3069 SDValue NewLoad(Load, 0);
3070 if (Load->getExtensionType() == ISD::EXTLOAD) {
3071 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
3072 Load->getValueType(0), SDLoc(Load),
3073 Load->getChain(), Load->getBasePtr(),
3074 Load->getOffset(), Load->getMemoryVT(),
3075 Load->getMemOperand());
3076 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
3077 if (Load->getNumValues() == 3) {
3078 // PRE/POST_INC loads have 3 values.
3079 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
3080 NewLoad.getValue(2) };
3081 CombineTo(Load, To, 3, true);
3083 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
3087 // Fold the AND away, taking care not to fold to the old load node if we
3089 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
3091 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3095 // fold (and (load x), 255) -> (zextload x, i8)
3096 // fold (and (extload x, i16), 255) -> (zextload x, i8)
3097 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
3098 if (N1C && (N0.getOpcode() == ISD::LOAD ||
3099 (N0.getOpcode() == ISD::ANY_EXTEND &&
3100 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
3101 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
3102 LoadSDNode *LN0 = HasAnyExt
3103 ? cast<LoadSDNode>(N0.getOperand(0))
3104 : cast<LoadSDNode>(N0);
3105 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
3106 LN0->isUnindexed() && N0.hasOneUse() && SDValue(LN0, 0).hasOneUse()) {
3107 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
3108 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
3109 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
3110 EVT LoadedVT = LN0->getMemoryVT();
3111 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
3113 if (ExtVT == LoadedVT &&
3114 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy,
3118 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
3119 LN0->getChain(), LN0->getBasePtr(), ExtVT,
3120 LN0->getMemOperand());
3122 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
3123 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3126 // Do not change the width of a volatile load.
3127 // Do not generate loads of non-round integer types since these can
3128 // be expensive (and would be wrong if the type is not byte sized).
3129 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
3130 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy,
3132 EVT PtrType = LN0->getOperand(1).getValueType();
3134 unsigned Alignment = LN0->getAlignment();
3135 SDValue NewPtr = LN0->getBasePtr();
3137 // For big endian targets, we need to add an offset to the pointer
3138 // to load the correct bytes. For little endian systems, we merely
3139 // need to read fewer bytes from the same pointer.
3140 if (DAG.getDataLayout().isBigEndian()) {
3141 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
3142 unsigned EVTStoreBytes = ExtVT.getStoreSize();
3143 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
3145 NewPtr = DAG.getNode(ISD::ADD, DL, PtrType,
3146 NewPtr, DAG.getConstant(PtrOff, DL, PtrType));
3147 Alignment = MinAlign(Alignment, PtrOff);
3150 AddToWorklist(NewPtr.getNode());
3153 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
3154 LN0->getChain(), NewPtr,
3155 LN0->getPointerInfo(),
3156 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
3157 LN0->isInvariant(), Alignment, LN0->getAAInfo());
3159 CombineTo(LN0, Load, Load.getValue(1));
3160 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3166 if (SDValue Combined = visitANDLike(N0, N1, N))
3169 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
3170 if (N0.getOpcode() == N1.getOpcode())
3171 if (SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N))
3174 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
3175 // fold (and (sra)) -> (and (srl)) when possible.
3176 if (!VT.isVector() &&
3177 SimplifyDemandedBits(SDValue(N, 0)))
3178 return SDValue(N, 0);
3180 // fold (zext_inreg (extload x)) -> (zextload x)
3181 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
3182 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3183 EVT MemVT = LN0->getMemoryVT();
3184 // If we zero all the possible extended bits, then we can turn this into
3185 // a zextload if we are running before legalize or the operation is legal.
3186 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
3187 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
3188 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
3189 ((!LegalOperations && !LN0->isVolatile()) ||
3190 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) {
3191 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
3192 LN0->getChain(), LN0->getBasePtr(),
3193 MemVT, LN0->getMemOperand());
3195 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
3196 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3199 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
3200 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
3202 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3203 EVT MemVT = LN0->getMemoryVT();
3204 // If we zero all the possible extended bits, then we can turn this into
3205 // a zextload if we are running before legalize or the operation is legal.
3206 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
3207 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
3208 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
3209 ((!LegalOperations && !LN0->isVolatile()) ||
3210 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) {
3211 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
3212 LN0->getChain(), LN0->getBasePtr(),
3213 MemVT, LN0->getMemOperand());
3215 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
3216 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3219 // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const)
3220 if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) {
3221 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
3222 N0.getOperand(1), false);
3223 if (BSwap.getNode())
3230 /// Match (a >> 8) | (a << 8) as (bswap a) >> 16.
3231 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
3232 bool DemandHighBits) {
3233 if (!LegalOperations)
3236 EVT VT = N->getValueType(0);
3237 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
3239 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3242 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
3243 bool LookPassAnd0 = false;
3244 bool LookPassAnd1 = false;
3245 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
3247 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
3249 if (N0.getOpcode() == ISD::AND) {
3250 if (!N0.getNode()->hasOneUse())
3252 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3253 if (!N01C || N01C->getZExtValue() != 0xFF00)
3255 N0 = N0.getOperand(0);
3256 LookPassAnd0 = true;
3259 if (N1.getOpcode() == ISD::AND) {
3260 if (!N1.getNode()->hasOneUse())
3262 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3263 if (!N11C || N11C->getZExtValue() != 0xFF)
3265 N1 = N1.getOperand(0);
3266 LookPassAnd1 = true;
3269 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
3271 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
3273 if (!N0.getNode()->hasOneUse() ||
3274 !N1.getNode()->hasOneUse())
3277 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3278 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3281 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
3284 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
3285 SDValue N00 = N0->getOperand(0);
3286 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
3287 if (!N00.getNode()->hasOneUse())
3289 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
3290 if (!N001C || N001C->getZExtValue() != 0xFF)
3292 N00 = N00.getOperand(0);
3293 LookPassAnd0 = true;
3296 SDValue N10 = N1->getOperand(0);
3297 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
3298 if (!N10.getNode()->hasOneUse())
3300 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
3301 if (!N101C || N101C->getZExtValue() != 0xFF00)
3303 N10 = N10.getOperand(0);
3304 LookPassAnd1 = true;
3310 // Make sure everything beyond the low halfword gets set to zero since the SRL
3311 // 16 will clear the top bits.
3312 unsigned OpSizeInBits = VT.getSizeInBits();
3313 if (DemandHighBits && OpSizeInBits > 16) {
3314 // If the left-shift isn't masked out then the only way this is a bswap is
3315 // if all bits beyond the low 8 are 0. In that case the entire pattern
3316 // reduces to a left shift anyway: leave it for other parts of the combiner.
3320 // However, if the right shift isn't masked out then it might be because
3321 // it's not needed. See if we can spot that too.
3322 if (!LookPassAnd1 &&
3323 !DAG.MaskedValueIsZero(
3324 N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16)))
3328 SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00);
3329 if (OpSizeInBits > 16) {
3331 Res = DAG.getNode(ISD::SRL, DL, VT, Res,
3332 DAG.getConstant(OpSizeInBits - 16, DL,
3333 getShiftAmountTy(VT)));
3338 /// Return true if the specified node is an element that makes up a 32-bit
3339 /// packed halfword byteswap.
3340 /// ((x & 0x000000ff) << 8) |
3341 /// ((x & 0x0000ff00) >> 8) |
3342 /// ((x & 0x00ff0000) << 8) |
3343 /// ((x & 0xff000000) >> 8)
3344 static bool isBSwapHWordElement(SDValue N, MutableArrayRef<SDNode *> Parts) {
3345 if (!N.getNode()->hasOneUse())
3348 unsigned Opc = N.getOpcode();
3349 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
3352 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3357 switch (N1C->getZExtValue()) {
3360 case 0xFF: Num = 0; break;
3361 case 0xFF00: Num = 1; break;
3362 case 0xFF0000: Num = 2; break;
3363 case 0xFF000000: Num = 3; break;
3366 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
3367 SDValue N0 = N.getOperand(0);
3368 if (Opc == ISD::AND) {
3369 if (Num == 0 || Num == 2) {
3371 // (x >> 8) & 0xff0000
3372 if (N0.getOpcode() != ISD::SRL)
3374 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3375 if (!C || C->getZExtValue() != 8)
3378 // (x << 8) & 0xff00
3379 // (x << 8) & 0xff000000
3380 if (N0.getOpcode() != ISD::SHL)
3382 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3383 if (!C || C->getZExtValue() != 8)
3386 } else if (Opc == ISD::SHL) {
3388 // (x & 0xff0000) << 8
3389 if (Num != 0 && Num != 2)
3391 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3392 if (!C || C->getZExtValue() != 8)
3394 } else { // Opc == ISD::SRL
3395 // (x & 0xff00) >> 8
3396 // (x & 0xff000000) >> 8
3397 if (Num != 1 && Num != 3)
3399 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3400 if (!C || C->getZExtValue() != 8)
3407 Parts[Num] = N0.getOperand(0).getNode();
3411 /// Match a 32-bit packed halfword bswap. That is
3412 /// ((x & 0x000000ff) << 8) |
3413 /// ((x & 0x0000ff00) >> 8) |
3414 /// ((x & 0x00ff0000) << 8) |
3415 /// ((x & 0xff000000) >> 8)
3416 /// => (rotl (bswap x), 16)
3417 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
3418 if (!LegalOperations)
3421 EVT VT = N->getValueType(0);
3424 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3428 // (or (or (and), (and)), (or (and), (and)))
3429 // (or (or (or (and), (and)), (and)), (and))
3430 if (N0.getOpcode() != ISD::OR)
3432 SDValue N00 = N0.getOperand(0);
3433 SDValue N01 = N0.getOperand(1);
3434 SDNode *Parts[4] = {};
3436 if (N1.getOpcode() == ISD::OR &&
3437 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) {
3438 // (or (or (and), (and)), (or (and), (and)))
3439 SDValue N000 = N00.getOperand(0);
3440 if (!isBSwapHWordElement(N000, Parts))
3443 SDValue N001 = N00.getOperand(1);
3444 if (!isBSwapHWordElement(N001, Parts))
3446 SDValue N010 = N01.getOperand(0);
3447 if (!isBSwapHWordElement(N010, Parts))
3449 SDValue N011 = N01.getOperand(1);
3450 if (!isBSwapHWordElement(N011, Parts))
3453 // (or (or (or (and), (and)), (and)), (and))
3454 if (!isBSwapHWordElement(N1, Parts))
3456 if (!isBSwapHWordElement(N01, Parts))
3458 if (N00.getOpcode() != ISD::OR)
3460 SDValue N000 = N00.getOperand(0);
3461 if (!isBSwapHWordElement(N000, Parts))
3463 SDValue N001 = N00.getOperand(1);
3464 if (!isBSwapHWordElement(N001, Parts))
3468 // Make sure the parts are all coming from the same node.
3469 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
3473 SDValue BSwap = DAG.getNode(ISD::BSWAP, DL, VT,
3474 SDValue(Parts[0], 0));
3476 // Result of the bswap should be rotated by 16. If it's not legal, then
3477 // do (x << 16) | (x >> 16).
3478 SDValue ShAmt = DAG.getConstant(16, DL, getShiftAmountTy(VT));
3479 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3480 return DAG.getNode(ISD::ROTL, DL, VT, BSwap, ShAmt);
3481 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3482 return DAG.getNode(ISD::ROTR, DL, VT, BSwap, ShAmt);
3483 return DAG.getNode(ISD::OR, DL, VT,
3484 DAG.getNode(ISD::SHL, DL, VT, BSwap, ShAmt),
3485 DAG.getNode(ISD::SRL, DL, VT, BSwap, ShAmt));
3488 /// This contains all DAGCombine rules which reduce two values combined by
3489 /// an Or operation to a single value \see visitANDLike().
3490 SDValue DAGCombiner::visitORLike(SDValue N0, SDValue N1, SDNode *LocReference) {
3491 EVT VT = N1.getValueType();
3492 // fold (or x, undef) -> -1
3493 if (!LegalOperations &&
3494 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3495 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3496 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()),
3497 SDLoc(LocReference), VT);
3499 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3500 SDValue LL, LR, RL, RR, CC0, CC1;
3501 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3502 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3503 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3505 if (LR == RR && Op0 == Op1 && LL.getValueType().isInteger()) {
3506 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3507 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3508 if (isNullConstant(LR) && (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3509 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(LR),
3510 LR.getValueType(), LL, RL);
3511 AddToWorklist(ORNode.getNode());
3512 return DAG.getSetCC(SDLoc(LocReference), VT, ORNode, LR, Op1);
3514 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3515 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3516 if (isAllOnesConstant(LR) && (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3517 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(LR),
3518 LR.getValueType(), LL, RL);
3519 AddToWorklist(ANDNode.getNode());
3520 return DAG.getSetCC(SDLoc(LocReference), VT, ANDNode, LR, Op1);
3523 // canonicalize equivalent to ll == rl
3524 if (LL == RR && LR == RL) {
3525 Op1 = ISD::getSetCCSwappedOperands(Op1);
3528 if (LL == RL && LR == RR) {
3529 bool isInteger = LL.getValueType().isInteger();
3530 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3531 if (Result != ISD::SETCC_INVALID &&
3532 (!LegalOperations ||
3533 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
3534 TLI.isOperationLegal(ISD::SETCC,
3535 getSetCCResultType(N0.getValueType())))))
3536 return DAG.getSetCC(SDLoc(LocReference), N0.getValueType(),
3541 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3542 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == ISD::AND &&
3543 // Don't increase # computations.
3544 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3545 // We can only do this xform if we know that bits from X that are set in C2
3546 // but not in C1 are already zero. Likewise for Y.
3547 if (const ConstantSDNode *N0O1C =
3548 getAsNonOpaqueConstant(N0.getOperand(1))) {
3549 if (const ConstantSDNode *N1O1C =
3550 getAsNonOpaqueConstant(N1.getOperand(1))) {
3551 // We can only do this xform if we know that bits from X that are set in
3552 // C2 but not in C1 are already zero. Likewise for Y.
3553 const APInt &LHSMask = N0O1C->getAPIntValue();
3554 const APInt &RHSMask = N1O1C->getAPIntValue();
3556 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3557 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3558 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3559 N0.getOperand(0), N1.getOperand(0));
3560 SDLoc DL(LocReference);
3561 return DAG.getNode(ISD::AND, DL, VT, X,
3562 DAG.getConstant(LHSMask | RHSMask, DL, VT));
3568 // (or (and X, M), (and X, N)) -> (and X, (or M, N))
3569 if (N0.getOpcode() == ISD::AND &&
3570 N1.getOpcode() == ISD::AND &&
3571 N0.getOperand(0) == N1.getOperand(0) &&
3572 // Don't increase # computations.
3573 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3574 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3575 N0.getOperand(1), N1.getOperand(1));
3576 return DAG.getNode(ISD::AND, SDLoc(LocReference), VT, N0.getOperand(0), X);
3582 SDValue DAGCombiner::visitOR(SDNode *N) {
3583 SDValue N0 = N->getOperand(0);
3584 SDValue N1 = N->getOperand(1);
3585 EVT VT = N1.getValueType();
3588 if (VT.isVector()) {
3589 if (SDValue FoldedVOp = SimplifyVBinOp(N))
3592 // fold (or x, 0) -> x, vector edition
3593 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3595 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3598 // fold (or x, -1) -> -1, vector edition
3599 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3600 // do not return N0, because undef node may exist in N0
3601 return DAG.getConstant(
3602 APInt::getAllOnesValue(
3603 N0.getValueType().getScalarType().getSizeInBits()),
3604 SDLoc(N), N0.getValueType());
3605 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3606 // do not return N1, because undef node may exist in N1
3607 return DAG.getConstant(
3608 APInt::getAllOnesValue(
3609 N1.getValueType().getScalarType().getSizeInBits()),
3610 SDLoc(N), N1.getValueType());
3612 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask1)
3613 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf B, A, Mask2)
3614 // Do this only if the resulting shuffle is legal.
3615 if (isa<ShuffleVectorSDNode>(N0) &&
3616 isa<ShuffleVectorSDNode>(N1) &&
3617 // Avoid folding a node with illegal type.
3618 TLI.isTypeLegal(VT) &&
3619 N0->getOperand(1) == N1->getOperand(1) &&
3620 ISD::isBuildVectorAllZeros(N0.getOperand(1).getNode())) {
3621 bool CanFold = true;
3622 unsigned NumElts = VT.getVectorNumElements();
3623 const ShuffleVectorSDNode *SV0 = cast<ShuffleVectorSDNode>(N0);
3624 const ShuffleVectorSDNode *SV1 = cast<ShuffleVectorSDNode>(N1);
3625 // We construct two shuffle masks:
3626 // - Mask1 is a shuffle mask for a shuffle with N0 as the first operand
3627 // and N1 as the second operand.
3628 // - Mask2 is a shuffle mask for a shuffle with N1 as the first operand
3629 // and N0 as the second operand.
3630 // We do this because OR is commutable and therefore there might be
3631 // two ways to fold this node into a shuffle.
3632 SmallVector<int,4> Mask1;
3633 SmallVector<int,4> Mask2;
3635 for (unsigned i = 0; i != NumElts && CanFold; ++i) {
3636 int M0 = SV0->getMaskElt(i);
3637 int M1 = SV1->getMaskElt(i);
3639 // Both shuffle indexes are undef. Propagate Undef.
3640 if (M0 < 0 && M1 < 0) {
3641 Mask1.push_back(M0);
3642 Mask2.push_back(M0);
3646 if (M0 < 0 || M1 < 0 ||
3647 (M0 < (int)NumElts && M1 < (int)NumElts) ||
3648 (M0 >= (int)NumElts && M1 >= (int)NumElts)) {
3653 Mask1.push_back(M0 < (int)NumElts ? M0 : M1 + NumElts);
3654 Mask2.push_back(M1 < (int)NumElts ? M1 : M0 + NumElts);
3658 // Fold this sequence only if the resulting shuffle is 'legal'.
3659 if (TLI.isShuffleMaskLegal(Mask1, VT))
3660 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0),
3661 N1->getOperand(0), &Mask1[0]);
3662 if (TLI.isShuffleMaskLegal(Mask2, VT))
3663 return DAG.getVectorShuffle(VT, SDLoc(N), N1->getOperand(0),
3664 N0->getOperand(0), &Mask2[0]);
3669 // fold (or c1, c2) -> c1|c2
3670 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
3671 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3672 if (N0C && N1C && !N1C->isOpaque())
3673 return DAG.FoldConstantArithmetic(ISD::OR, SDLoc(N), VT, N0C, N1C);
3674 // canonicalize constant to RHS
3675 if (isConstantIntBuildVectorOrConstantInt(N0) &&
3676 !isConstantIntBuildVectorOrConstantInt(N1))
3677 return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0);
3678 // fold (or x, 0) -> x
3679 if (isNullConstant(N1))
3681 // fold (or x, -1) -> -1
3682 if (isAllOnesConstant(N1))
3684 // fold (or x, c) -> c iff (x & ~c) == 0
3685 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3688 if (SDValue Combined = visitORLike(N0, N1, N))
3691 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3692 if (SDValue BSwap = MatchBSwapHWord(N, N0, N1))
3694 if (SDValue BSwap = MatchBSwapHWordLow(N, N0, N1))
3698 if (SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1))
3700 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3701 // iff (c1 & c2) == 0.
3702 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3703 isa<ConstantSDNode>(N0.getOperand(1))) {
3704 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3705 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) {
3706 if (SDValue COR = DAG.FoldConstantArithmetic(ISD::OR, SDLoc(N1), VT,
3709 ISD::AND, SDLoc(N), VT,
3710 DAG.getNode(ISD::OR, SDLoc(N0), VT, N0.getOperand(0), N1), COR);
3714 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3715 if (N0.getOpcode() == N1.getOpcode())
3716 if (SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N))
3719 // See if this is some rotate idiom.
3720 if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N)))
3721 return SDValue(Rot, 0);
3723 // Simplify the operands using demanded-bits information.
3724 if (!VT.isVector() &&
3725 SimplifyDemandedBits(SDValue(N, 0)))
3726 return SDValue(N, 0);
3731 /// Match "(X shl/srl V1) & V2" where V2 may not be present.
3732 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3733 if (Op.getOpcode() == ISD::AND) {
3734 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3735 Mask = Op.getOperand(1);
3736 Op = Op.getOperand(0);
3742 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3750 // Return true if we can prove that, whenever Neg and Pos are both in the
3751 // range [0, OpSize), Neg == (Pos == 0 ? 0 : OpSize - Pos). This means that
3752 // for two opposing shifts shift1 and shift2 and a value X with OpBits bits:
3754 // (or (shift1 X, Neg), (shift2 X, Pos))
3756 // reduces to a rotate in direction shift2 by Pos or (equivalently) a rotate
3757 // in direction shift1 by Neg. The range [0, OpSize) means that we only need
3758 // to consider shift amounts with defined behavior.
3759 static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned OpSize) {
3760 // If OpSize is a power of 2 then:
3762 // (a) (Pos == 0 ? 0 : OpSize - Pos) == (OpSize - Pos) & (OpSize - 1)
3763 // (b) Neg == Neg & (OpSize - 1) whenever Neg is in [0, OpSize).
3765 // So if OpSize is a power of 2 and Neg is (and Neg', OpSize-1), we check
3766 // for the stronger condition:
3768 // Neg & (OpSize - 1) == (OpSize - Pos) & (OpSize - 1) [A]
3770 // for all Neg and Pos. Since Neg & (OpSize - 1) == Neg' & (OpSize - 1)
3771 // we can just replace Neg with Neg' for the rest of the function.
3773 // In other cases we check for the even stronger condition:
3775 // Neg == OpSize - Pos [B]
3777 // for all Neg and Pos. Note that the (or ...) then invokes undefined
3778 // behavior if Pos == 0 (and consequently Neg == OpSize).
3780 // We could actually use [A] whenever OpSize is a power of 2, but the
3781 // only extra cases that it would match are those uninteresting ones
3782 // where Neg and Pos are never in range at the same time. E.g. for
3783 // OpSize == 32, using [A] would allow a Neg of the form (sub 64, Pos)
3784 // as well as (sub 32, Pos), but:
3786 // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos))
3788 // always invokes undefined behavior for 32-bit X.
3790 // Below, Mask == OpSize - 1 when using [A] and is all-ones otherwise.
3791 unsigned MaskLoBits = 0;
3792 if (Neg.getOpcode() == ISD::AND &&
3793 isPowerOf2_64(OpSize) &&
3794 Neg.getOperand(1).getOpcode() == ISD::Constant &&
3795 cast<ConstantSDNode>(Neg.getOperand(1))->getAPIntValue() == OpSize - 1) {
3796 Neg = Neg.getOperand(0);
3797 MaskLoBits = Log2_64(OpSize);
3800 // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1.
3801 if (Neg.getOpcode() != ISD::SUB)
3803 ConstantSDNode *NegC = dyn_cast<ConstantSDNode>(Neg.getOperand(0));
3806 SDValue NegOp1 = Neg.getOperand(1);
3808 // On the RHS of [A], if Pos is Pos' & (OpSize - 1), just replace Pos with
3809 // Pos'. The truncation is redundant for the purpose of the equality.
3811 Pos.getOpcode() == ISD::AND &&
3812 Pos.getOperand(1).getOpcode() == ISD::Constant &&
3813 cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() == OpSize - 1)
3814 Pos = Pos.getOperand(0);
3816 // The condition we need is now:
3818 // (NegC - NegOp1) & Mask == (OpSize - Pos) & Mask
3820 // If NegOp1 == Pos then we need:
3822 // OpSize & Mask == NegC & Mask
3824 // (because "x & Mask" is a truncation and distributes through subtraction).
3827 Width = NegC->getAPIntValue();
3828 // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC.
3829 // Then the condition we want to prove becomes:
3831 // (NegC - NegOp1) & Mask == (OpSize - (NegOp1 + PosC)) & Mask
3833 // which, again because "x & Mask" is a truncation, becomes:
3835 // NegC & Mask == (OpSize - PosC) & Mask
3836 // OpSize & Mask == (NegC + PosC) & Mask
3837 else if (Pos.getOpcode() == ISD::ADD &&
3838 Pos.getOperand(0) == NegOp1 &&
3839 Pos.getOperand(1).getOpcode() == ISD::Constant)
3840 Width = (cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() +
3841 NegC->getAPIntValue());
3845 // Now we just need to check that OpSize & Mask == Width & Mask.
3847 // Opsize & Mask is 0 since Mask is Opsize - 1.
3848 return Width.getLoBits(MaskLoBits) == 0;
3849 return Width == OpSize;
3852 // A subroutine of MatchRotate used once we have found an OR of two opposite
3853 // shifts of Shifted. If Neg == <operand size> - Pos then the OR reduces
3854 // to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the
3855 // former being preferred if supported. InnerPos and InnerNeg are Pos and
3856 // Neg with outer conversions stripped away.
3857 SDNode *DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos,
3858 SDValue Neg, SDValue InnerPos,
3859 SDValue InnerNeg, unsigned PosOpcode,
3860 unsigned NegOpcode, SDLoc DL) {
3861 // fold (or (shl x, (*ext y)),
3862 // (srl x, (*ext (sub 32, y)))) ->
3863 // (rotl x, y) or (rotr x, (sub 32, y))
3865 // fold (or (shl x, (*ext (sub 32, y))),
3866 // (srl x, (*ext y))) ->
3867 // (rotr x, y) or (rotl x, (sub 32, y))
3868 EVT VT = Shifted.getValueType();
3869 if (matchRotateSub(InnerPos, InnerNeg, VT.getSizeInBits())) {
3870 bool HasPos = TLI.isOperationLegalOrCustom(PosOpcode, VT);
3871 return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted,
3872 HasPos ? Pos : Neg).getNode();
3878 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3879 // idioms for rotate, and if the target supports rotation instructions, generate
3881 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL) {
3882 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3883 EVT VT = LHS.getValueType();
3884 if (!TLI.isTypeLegal(VT)) return nullptr;
3886 // The target must have at least one rotate flavor.
3887 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3888 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3889 if (!HasROTL && !HasROTR) return nullptr;
3891 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3892 SDValue LHSShift; // The shift.
3893 SDValue LHSMask; // AND value if any.
3894 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3895 return nullptr; // Not part of a rotate.
3897 SDValue RHSShift; // The shift.
3898 SDValue RHSMask; // AND value if any.
3899 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3900 return nullptr; // Not part of a rotate.
3902 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3903 return nullptr; // Not shifting the same value.
3905 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3906 return nullptr; // Shifts must disagree.
3908 // Canonicalize shl to left side in a shl/srl pair.
3909 if (RHSShift.getOpcode() == ISD::SHL) {
3910 std::swap(LHS, RHS);
3911 std::swap(LHSShift, RHSShift);
3912 std::swap(LHSMask , RHSMask );
3915 unsigned OpSizeInBits = VT.getSizeInBits();
3916 SDValue LHSShiftArg = LHSShift.getOperand(0);
3917 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3918 SDValue RHSShiftArg = RHSShift.getOperand(0);
3919 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3921 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3922 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3923 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3924 RHSShiftAmt.getOpcode() == ISD::Constant) {
3925 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3926 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3927 if ((LShVal + RShVal) != OpSizeInBits)
3930 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3931 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
3933 // If there is an AND of either shifted operand, apply it to the result.
3934 if (LHSMask.getNode() || RHSMask.getNode()) {
3935 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3937 if (LHSMask.getNode()) {
3938 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3939 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3941 if (RHSMask.getNode()) {
3942 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3943 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3946 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, DL, VT));
3949 return Rot.getNode();
3952 // If there is a mask here, and we have a variable shift, we can't be sure
3953 // that we're masking out the right stuff.
3954 if (LHSMask.getNode() || RHSMask.getNode())
3957 // If the shift amount is sign/zext/any-extended just peel it off.
3958 SDValue LExtOp0 = LHSShiftAmt;
3959 SDValue RExtOp0 = RHSShiftAmt;
3960 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3961 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3962 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3963 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3964 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3965 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3966 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3967 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3968 LExtOp0 = LHSShiftAmt.getOperand(0);
3969 RExtOp0 = RHSShiftAmt.getOperand(0);
3972 SDNode *TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt,
3973 LExtOp0, RExtOp0, ISD::ROTL, ISD::ROTR, DL);
3977 SDNode *TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt,
3978 RExtOp0, LExtOp0, ISD::ROTR, ISD::ROTL, DL);
3985 SDValue DAGCombiner::visitXOR(SDNode *N) {
3986 SDValue N0 = N->getOperand(0);
3987 SDValue N1 = N->getOperand(1);
3988 EVT VT = N0.getValueType();
3991 if (VT.isVector()) {
3992 if (SDValue FoldedVOp = SimplifyVBinOp(N))
3995 // fold (xor x, 0) -> x, vector edition
3996 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3998 if (ISD::isBuildVectorAllZeros(N1.getNode()))
4002 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
4003 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
4004 return DAG.getConstant(0, SDLoc(N), VT);
4005 // fold (xor x, undef) -> undef
4006 if (N0.getOpcode() == ISD::UNDEF)
4008 if (N1.getOpcode() == ISD::UNDEF)
4010 // fold (xor c1, c2) -> c1^c2
4011 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
4012 ConstantSDNode *N1C = getAsNonOpaqueConstant(N1);
4014 return DAG.FoldConstantArithmetic(ISD::XOR, SDLoc(N), VT, N0C, N1C);
4015 // canonicalize constant to RHS
4016 if (isConstantIntBuildVectorOrConstantInt(N0) &&
4017 !isConstantIntBuildVectorOrConstantInt(N1))
4018 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
4019 // fold (xor x, 0) -> x
4020 if (isNullConstant(N1))
4023 if (SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1))
4026 // fold !(x cc y) -> (x !cc y)
4027 SDValue LHS, RHS, CC;
4028 if (TLI.isConstTrueVal(N1.getNode()) && isSetCCEquivalent(N0, LHS, RHS, CC)) {
4029 bool isInt = LHS.getValueType().isInteger();
4030 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
4033 if (!LegalOperations ||
4034 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) {
4035 switch (N0.getOpcode()) {
4037 llvm_unreachable("Unhandled SetCC Equivalent!");
4039 return DAG.getSetCC(SDLoc(N), VT, LHS, RHS, NotCC);
4040 case ISD::SELECT_CC:
4041 return DAG.getSelectCC(SDLoc(N), LHS, RHS, N0.getOperand(2),
4042 N0.getOperand(3), NotCC);
4047 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
4048 if (isOneConstant(N1) && N0.getOpcode() == ISD::ZERO_EXTEND &&
4049 N0.getNode()->hasOneUse() &&
4050 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
4051 SDValue V = N0.getOperand(0);
4053 V = DAG.getNode(ISD::XOR, DL, V.getValueType(), V,
4054 DAG.getConstant(1, DL, V.getValueType()));
4055 AddToWorklist(V.getNode());
4056 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V);
4059 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
4060 if (isOneConstant(N1) && VT == MVT::i1 &&
4061 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
4062 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
4063 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
4064 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
4065 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
4066 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
4067 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
4068 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
4071 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
4072 if (isAllOnesConstant(N1) &&
4073 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
4074 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
4075 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
4076 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
4077 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
4078 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
4079 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
4080 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
4083 // fold (xor (and x, y), y) -> (and (not x), y)
4084 if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
4085 N0->getOperand(1) == N1) {
4086 SDValue X = N0->getOperand(0);
4087 SDValue NotX = DAG.getNOT(SDLoc(X), X, VT);
4088 AddToWorklist(NotX.getNode());
4089 return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1);
4091 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
4092 if (N1C && N0.getOpcode() == ISD::XOR) {
4093 if (const ConstantSDNode *N00C = getAsNonOpaqueConstant(N0.getOperand(0))) {
4095 return DAG.getNode(ISD::XOR, DL, VT, N0.getOperand(1),
4096 DAG.getConstant(N1C->getAPIntValue() ^
4097 N00C->getAPIntValue(), DL, VT));
4099 if (const ConstantSDNode *N01C = getAsNonOpaqueConstant(N0.getOperand(1))) {
4101 return DAG.getNode(ISD::XOR, DL, VT, N0.getOperand(0),
4102 DAG.getConstant(N1C->getAPIntValue() ^
4103 N01C->getAPIntValue(), DL, VT));
4106 // fold (xor x, x) -> 0
4108 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
4110 // fold (xor (shl 1, x), -1) -> (rotl ~1, x)
4111 // Here is a concrete example of this equivalence:
4113 // i16 shl == 1 << 14 == 16384 == 0b0100000000000000
4114 // i16 xor == ~(1 << 14) == 49151 == 0b1011111111111111
4118 // i16 ~1 == 0b1111111111111110
4119 // i16 rol(~1, 14) == 0b1011111111111111
4121 // Some additional tips to help conceptualize this transform:
4122 // - Try to see the operation as placing a single zero in a value of all ones.
4123 // - There exists no value for x which would allow the result to contain zero.
4124 // - Values of x larger than the bitwidth are undefined and do not require a
4125 // consistent result.
4126 // - Pushing the zero left requires shifting one bits in from the right.
4127 // A rotate left of ~1 is a nice way of achieving the desired result.
4128 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT) && N0.getOpcode() == ISD::SHL
4129 && isAllOnesConstant(N1) && isOneConstant(N0.getOperand(0))) {
4131 return DAG.getNode(ISD::ROTL, DL, VT, DAG.getConstant(~1, DL, VT),
4135 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
4136 if (N0.getOpcode() == N1.getOpcode())
4137 if (SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N))
4140 // Simplify the expression using non-local knowledge.
4141 if (!VT.isVector() &&
4142 SimplifyDemandedBits(SDValue(N, 0)))
4143 return SDValue(N, 0);
4148 /// Handle transforms common to the three shifts, when the shift amount is a
4150 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, ConstantSDNode *Amt) {
4151 SDNode *LHS = N->getOperand(0).getNode();
4152 if (!LHS->hasOneUse()) return SDValue();
4154 // We want to pull some binops through shifts, so that we have (and (shift))
4155 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
4156 // thing happens with address calculations, so it's important to canonicalize
4158 bool HighBitSet = false; // Can we transform this if the high bit is set?
4160 switch (LHS->getOpcode()) {
4161 default: return SDValue();
4164 HighBitSet = false; // We can only transform sra if the high bit is clear.
4167 HighBitSet = true; // We can only transform sra if the high bit is set.
4170 if (N->getOpcode() != ISD::SHL)
4171 return SDValue(); // only shl(add) not sr[al](add).
4172 HighBitSet = false; // We can only transform sra if the high bit is clear.
4176 // We require the RHS of the binop to be a constant and not opaque as well.
4177 ConstantSDNode *BinOpCst = getAsNonOpaqueConstant(LHS->getOperand(1));
4178 if (!BinOpCst) return SDValue();
4180 // FIXME: disable this unless the input to the binop is a shift by a constant.
4181 // If it is not a shift, it pessimizes some common cases like:
4183 // void foo(int *X, int i) { X[i & 1235] = 1; }
4184 // int bar(int *X, int i) { return X[i & 255]; }
4185 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
4186 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
4187 BinOpLHSVal->getOpcode() != ISD::SRA &&
4188 BinOpLHSVal->getOpcode() != ISD::SRL) ||
4189 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
4192 EVT VT = N->getValueType(0);
4194 // If this is a signed shift right, and the high bit is modified by the
4195 // logical operation, do not perform the transformation. The highBitSet
4196 // boolean indicates the value of the high bit of the constant which would
4197 // cause it to be modified for this operation.
4198 if (N->getOpcode() == ISD::SRA) {
4199 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
4200 if (BinOpRHSSignSet != HighBitSet)
4204 if (!TLI.isDesirableToCommuteWithShift(LHS))
4207 // Fold the constants, shifting the binop RHS by the shift amount.
4208 SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)),
4210 LHS->getOperand(1), N->getOperand(1));
4211 assert(isa<ConstantSDNode>(NewRHS) && "Folding was not successful!");
4213 // Create the new shift.
4214 SDValue NewShift = DAG.getNode(N->getOpcode(),
4215 SDLoc(LHS->getOperand(0)),
4216 VT, LHS->getOperand(0), N->getOperand(1));
4218 // Create the new binop.
4219 return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS);
4222 SDValue DAGCombiner::distributeTruncateThroughAnd(SDNode *N) {
4223 assert(N->getOpcode() == ISD::TRUNCATE);
4224 assert(N->getOperand(0).getOpcode() == ISD::AND);
4226 // (truncate:TruncVT (and N00, N01C)) -> (and (truncate:TruncVT N00), TruncC)
4227 if (N->hasOneUse() && N->getOperand(0).hasOneUse()) {
4228 SDValue N01 = N->getOperand(0).getOperand(1);
4230 if (ConstantSDNode *N01C = isConstOrConstSplat(N01)) {
4231 if (!N01C->isOpaque()) {
4232 EVT TruncVT = N->getValueType(0);
4233 SDValue N00 = N->getOperand(0).getOperand(0);
4234 APInt TruncC = N01C->getAPIntValue();
4235 TruncC = TruncC.trunc(TruncVT.getScalarSizeInBits());
4238 return DAG.getNode(ISD::AND, DL, TruncVT,
4239 DAG.getNode(ISD::TRUNCATE, DL, TruncVT, N00),
4240 DAG.getConstant(TruncC, DL, TruncVT));
4248 SDValue DAGCombiner::visitRotate(SDNode *N) {
4249 // fold (rot* x, (trunc (and y, c))) -> (rot* x, (and (trunc y), (trunc c))).
4250 if (N->getOperand(1).getOpcode() == ISD::TRUNCATE &&
4251 N->getOperand(1).getOperand(0).getOpcode() == ISD::AND) {
4252 SDValue NewOp1 = distributeTruncateThroughAnd(N->getOperand(1).getNode());
4253 if (NewOp1.getNode())
4254 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
4255 N->getOperand(0), NewOp1);
4260 SDValue DAGCombiner::visitSHL(SDNode *N) {
4261 SDValue N0 = N->getOperand(0);
4262 SDValue N1 = N->getOperand(1);
4263 EVT VT = N0.getValueType();
4264 unsigned OpSizeInBits = VT.getScalarSizeInBits();
4267 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4268 if (VT.isVector()) {
4269 if (SDValue FoldedVOp = SimplifyVBinOp(N))
4272 BuildVectorSDNode *N1CV = dyn_cast<BuildVectorSDNode>(N1);
4273 // If setcc produces all-one true value then:
4274 // (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<<N1CV)
4275 if (N1CV && N1CV->isConstant()) {
4276 if (N0.getOpcode() == ISD::AND) {
4277 SDValue N00 = N0->getOperand(0);
4278 SDValue N01 = N0->getOperand(1);
4279 BuildVectorSDNode *N01CV = dyn_cast<BuildVectorSDNode>(N01);
4281 if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC &&
4282 TLI.getBooleanContents(N00.getOperand(0).getValueType()) ==
4283 TargetLowering::ZeroOrNegativeOneBooleanContent) {
4284 if (SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, SDLoc(N), VT,
4286 return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C);
4289 N1C = isConstOrConstSplat(N1);
4294 // fold (shl c1, c2) -> c1<<c2
4295 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
4296 if (N0C && N1C && !N1C->isOpaque())
4297 return DAG.FoldConstantArithmetic(ISD::SHL, SDLoc(N), VT, N0C, N1C);
4298 // fold (shl 0, x) -> 0
4299 if (isNullConstant(N0))
4301 // fold (shl x, c >= size(x)) -> undef
4302 if (N1C && N1C->getAPIntValue().uge(OpSizeInBits))
4303 return DAG.getUNDEF(VT);
4304 // fold (shl x, 0) -> x
4305 if (N1C && N1C->isNullValue())
4307 // fold (shl undef, x) -> 0
4308 if (N0.getOpcode() == ISD::UNDEF)
4309 return DAG.getConstant(0, SDLoc(N), VT);
4310 // if (shl x, c) is known to be zero, return 0
4311 if (DAG.MaskedValueIsZero(SDValue(N, 0),
4312 APInt::getAllOnesValue(OpSizeInBits)))
4313 return DAG.getConstant(0, SDLoc(N), VT);
4314 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
4315 if (N1.getOpcode() == ISD::TRUNCATE &&
4316 N1.getOperand(0).getOpcode() == ISD::AND) {
4317 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4318 if (NewOp1.getNode())
4319 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, NewOp1);
4322 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4323 return SDValue(N, 0);
4325 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
4326 if (N1C && N0.getOpcode() == ISD::SHL) {
4327 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4328 uint64_t c1 = N0C1->getZExtValue();
4329 uint64_t c2 = N1C->getZExtValue();
4331 if (c1 + c2 >= OpSizeInBits)
4332 return DAG.getConstant(0, DL, VT);
4333 return DAG.getNode(ISD::SHL, DL, VT, N0.getOperand(0),
4334 DAG.getConstant(c1 + c2, DL, N1.getValueType()));
4338 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
4339 // For this to be valid, the second form must not preserve any of the bits
4340 // that are shifted out by the inner shift in the first form. This means
4341 // the outer shift size must be >= the number of bits added by the ext.
4342 // As a corollary, we don't care what kind of ext it is.
4343 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
4344 N0.getOpcode() == ISD::ANY_EXTEND ||
4345 N0.getOpcode() == ISD::SIGN_EXTEND) &&
4346 N0.getOperand(0).getOpcode() == ISD::SHL) {
4347 SDValue N0Op0 = N0.getOperand(0);
4348 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4349 uint64_t c1 = N0Op0C1->getZExtValue();
4350 uint64_t c2 = N1C->getZExtValue();
4351 EVT InnerShiftVT = N0Op0.getValueType();
4352 uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits();
4353 if (c2 >= OpSizeInBits - InnerShiftSize) {
4355 if (c1 + c2 >= OpSizeInBits)
4356 return DAG.getConstant(0, DL, VT);
4357 return DAG.getNode(ISD::SHL, DL, VT,
4358 DAG.getNode(N0.getOpcode(), DL, VT,
4359 N0Op0->getOperand(0)),
4360 DAG.getConstant(c1 + c2, DL, N1.getValueType()));
4365 // fold (shl (zext (srl x, C)), C) -> (zext (shl (srl x, C), C))
4366 // Only fold this if the inner zext has no other uses to avoid increasing
4367 // the total number of instructions.
4368 if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() &&
4369 N0.getOperand(0).getOpcode() == ISD::SRL) {
4370 SDValue N0Op0 = N0.getOperand(0);
4371 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4372 uint64_t c1 = N0Op0C1->getZExtValue();
4373 if (c1 < VT.getScalarSizeInBits()) {
4374 uint64_t c2 = N1C->getZExtValue();
4376 SDValue NewOp0 = N0.getOperand(0);
4377 EVT CountVT = NewOp0.getOperand(1).getValueType();
4379 SDValue NewSHL = DAG.getNode(ISD::SHL, DL, NewOp0.getValueType(),
4381 DAG.getConstant(c2, DL, CountVT));
4382 AddToWorklist(NewSHL.getNode());
4383 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
4389 // fold (shl (sr[la] exact X, C1), C2) -> (shl X, (C2-C1)) if C1 <= C2
4390 // fold (shl (sr[la] exact X, C1), C2) -> (sr[la] X, (C2-C1)) if C1 > C2
4391 if (N1C && (N0.getOpcode() == ISD::SRL || N0.getOpcode() == ISD::SRA) &&
4392 cast<BinaryWithFlagsSDNode>(N0)->Flags.hasExact()) {
4393 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4394 uint64_t C1 = N0C1->getZExtValue();
4395 uint64_t C2 = N1C->getZExtValue();
4398 return DAG.getNode(ISD::SHL, DL, VT, N0.getOperand(0),
4399 DAG.getConstant(C2 - C1, DL, N1.getValueType()));
4400 return DAG.getNode(N0.getOpcode(), DL, VT, N0.getOperand(0),
4401 DAG.getConstant(C1 - C2, DL, N1.getValueType()));
4405 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
4406 // (and (srl x, (sub c1, c2), MASK)
4407 // Only fold this if the inner shift has no other uses -- if it does, folding
4408 // this will increase the total number of instructions.
4409 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
4410 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4411 uint64_t c1 = N0C1->getZExtValue();
4412 if (c1 < OpSizeInBits) {
4413 uint64_t c2 = N1C->getZExtValue();
4414 APInt Mask = APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - c1);
4417 Mask = Mask.shl(c2 - c1);
4419 Shift = DAG.getNode(ISD::SHL, DL, VT, N0.getOperand(0),
4420 DAG.getConstant(c2 - c1, DL, N1.getValueType()));
4422 Mask = Mask.lshr(c1 - c2);
4424 Shift = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4425 DAG.getConstant(c1 - c2, DL, N1.getValueType()));
4428 return DAG.getNode(ISD::AND, DL, VT, Shift,
4429 DAG.getConstant(Mask, DL, VT));
4433 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
4434 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
4435 unsigned BitSize = VT.getScalarSizeInBits();
4437 SDValue HiBitsMask =
4438 DAG.getConstant(APInt::getHighBitsSet(BitSize,
4439 BitSize - N1C->getZExtValue()),
4441 return DAG.getNode(ISD::AND, DL, VT, N0.getOperand(0),
4445 // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
4446 // Variant of version done on multiply, except mul by a power of 2 is turned
4449 if (N1C && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
4450 (isa<ConstantSDNode>(N0.getOperand(1)) ||
4451 isConstantSplatVector(N0.getOperand(1).getNode(), Val))) {
4452 SDValue Shl0 = DAG.getNode(ISD::SHL, SDLoc(N0), VT, N0.getOperand(0), N1);
4453 SDValue Shl1 = DAG.getNode(ISD::SHL, SDLoc(N1), VT, N0.getOperand(1), N1);
4454 return DAG.getNode(ISD::ADD, SDLoc(N), VT, Shl0, Shl1);
4457 // fold (shl (mul x, c1), c2) -> (mul x, c1 << c2)
4458 if (N1C && N0.getOpcode() == ISD::MUL && N0.getNode()->hasOneUse()) {
4459 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4460 if (SDValue Folded =
4461 DAG.FoldConstantArithmetic(ISD::SHL, SDLoc(N1), VT, N0C1, N1C))
4462 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N0.getOperand(0), Folded);
4466 if (N1C && !N1C->isOpaque())
4467 if (SDValue NewSHL = visitShiftByConstant(N, N1C))
4473 SDValue DAGCombiner::visitSRA(SDNode *N) {
4474 SDValue N0 = N->getOperand(0);
4475 SDValue N1 = N->getOperand(1);
4476 EVT VT = N0.getValueType();
4477 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4480 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4481 if (VT.isVector()) {
4482 if (SDValue FoldedVOp = SimplifyVBinOp(N))
4485 N1C = isConstOrConstSplat(N1);
4488 // fold (sra c1, c2) -> (sra c1, c2)
4489 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
4490 if (N0C && N1C && !N1C->isOpaque())
4491 return DAG.FoldConstantArithmetic(ISD::SRA, SDLoc(N), VT, N0C, N1C);
4492 // fold (sra 0, x) -> 0
4493 if (isNullConstant(N0))
4495 // fold (sra -1, x) -> -1
4496 if (isAllOnesConstant(N0))
4498 // fold (sra x, (setge c, size(x))) -> undef
4499 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4500 return DAG.getUNDEF(VT);
4501 // fold (sra x, 0) -> x
4502 if (N1C && N1C->isNullValue())
4504 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
4506 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
4507 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
4508 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
4510 ExtVT = EVT::getVectorVT(*DAG.getContext(),
4511 ExtVT, VT.getVectorNumElements());
4512 if ((!LegalOperations ||
4513 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
4514 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
4515 N0.getOperand(0), DAG.getValueType(ExtVT));
4518 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
4519 if (N1C && N0.getOpcode() == ISD::SRA) {
4520 if (ConstantSDNode *C1 = isConstOrConstSplat(N0.getOperand(1))) {
4521 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
4522 if (Sum >= OpSizeInBits)
4523 Sum = OpSizeInBits - 1;
4525 return DAG.getNode(ISD::SRA, DL, VT, N0.getOperand(0),
4526 DAG.getConstant(Sum, DL, N1.getValueType()));
4530 // fold (sra (shl X, m), (sub result_size, n))
4531 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
4532 // result_size - n != m.
4533 // If truncate is free for the target sext(shl) is likely to result in better
4535 if (N0.getOpcode() == ISD::SHL && N1C) {
4536 // Get the two constanst of the shifts, CN0 = m, CN = n.
4537 const ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1));
4539 LLVMContext &Ctx = *DAG.getContext();
4540 // Determine what the truncate's result bitsize and type would be.
4541 EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - N1C->getZExtValue());
4544 TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorNumElements());
4546 // Determine the residual right-shift amount.
4547 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
4549 // If the shift is not a no-op (in which case this should be just a sign
4550 // extend already), the truncated to type is legal, sign_extend is legal
4551 // on that type, and the truncate to that type is both legal and free,
4552 // perform the transform.
4553 if ((ShiftAmt > 0) &&
4554 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
4555 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
4556 TLI.isTruncateFree(VT, TruncVT)) {
4559 SDValue Amt = DAG.getConstant(ShiftAmt, DL,
4560 getShiftAmountTy(N0.getOperand(0).getValueType()));
4561 SDValue Shift = DAG.getNode(ISD::SRL, DL, VT,
4562 N0.getOperand(0), Amt);
4563 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, TruncVT,
4565 return DAG.getNode(ISD::SIGN_EXTEND, DL,
4566 N->getValueType(0), Trunc);
4571 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
4572 if (N1.getOpcode() == ISD::TRUNCATE &&
4573 N1.getOperand(0).getOpcode() == ISD::AND) {
4574 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4575 if (NewOp1.getNode())
4576 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, NewOp1);
4579 // fold (sra (trunc (srl x, c1)), c2) -> (trunc (sra x, c1 + c2))
4580 // if c1 is equal to the number of bits the trunc removes
4581 if (N0.getOpcode() == ISD::TRUNCATE &&
4582 (N0.getOperand(0).getOpcode() == ISD::SRL ||
4583 N0.getOperand(0).getOpcode() == ISD::SRA) &&
4584 N0.getOperand(0).hasOneUse() &&
4585 N0.getOperand(0).getOperand(1).hasOneUse() &&
4587 SDValue N0Op0 = N0.getOperand(0);
4588 if (ConstantSDNode *LargeShift = isConstOrConstSplat(N0Op0.getOperand(1))) {
4589 unsigned LargeShiftVal = LargeShift->getZExtValue();
4590 EVT LargeVT = N0Op0.getValueType();
4592 if (LargeVT.getScalarSizeInBits() - OpSizeInBits == LargeShiftVal) {
4595 DAG.getConstant(LargeShiftVal + N1C->getZExtValue(), DL,
4596 getShiftAmountTy(N0Op0.getOperand(0).getValueType()));
4597 SDValue SRA = DAG.getNode(ISD::SRA, DL, LargeVT,
4598 N0Op0.getOperand(0), Amt);
4599 return DAG.getNode(ISD::TRUNCATE, DL, VT, SRA);
4604 // Simplify, based on bits shifted out of the LHS.
4605 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4606 return SDValue(N, 0);
4609 // If the sign bit is known to be zero, switch this to a SRL.
4610 if (DAG.SignBitIsZero(N0))
4611 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1);
4613 if (N1C && !N1C->isOpaque())
4614 if (SDValue NewSRA = visitShiftByConstant(N, N1C))
4620 SDValue DAGCombiner::visitSRL(SDNode *N) {
4621 SDValue N0 = N->getOperand(0);
4622 SDValue N1 = N->getOperand(1);
4623 EVT VT = N0.getValueType();
4624 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4627 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4628 if (VT.isVector()) {
4629 if (SDValue FoldedVOp = SimplifyVBinOp(N))
4632 N1C = isConstOrConstSplat(N1);
4635 // fold (srl c1, c2) -> c1 >>u c2
4636 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
4637 if (N0C && N1C && !N1C->isOpaque())
4638 return DAG.FoldConstantArithmetic(ISD::SRL, SDLoc(N), VT, N0C, N1C);
4639 // fold (srl 0, x) -> 0
4640 if (isNullConstant(N0))
4642 // fold (srl x, c >= size(x)) -> undef
4643 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4644 return DAG.getUNDEF(VT);
4645 // fold (srl x, 0) -> x
4646 if (N1C && N1C->isNullValue())
4648 // if (srl x, c) is known to be zero, return 0
4649 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
4650 APInt::getAllOnesValue(OpSizeInBits)))
4651 return DAG.getConstant(0, SDLoc(N), VT);
4653 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
4654 if (N1C && N0.getOpcode() == ISD::SRL) {
4655 if (ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1))) {
4656 uint64_t c1 = N01C->getZExtValue();
4657 uint64_t c2 = N1C->getZExtValue();
4659 if (c1 + c2 >= OpSizeInBits)
4660 return DAG.getConstant(0, DL, VT);
4661 return DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4662 DAG.getConstant(c1 + c2, DL, N1.getValueType()));
4666 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
4667 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
4668 N0.getOperand(0).getOpcode() == ISD::SRL &&
4669 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
4671 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
4672 uint64_t c2 = N1C->getZExtValue();
4673 EVT InnerShiftVT = N0.getOperand(0).getValueType();
4674 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
4675 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
4676 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
4677 if (c1 + OpSizeInBits == InnerShiftSize) {
4679 if (c1 + c2 >= InnerShiftSize)
4680 return DAG.getConstant(0, DL, VT);
4681 return DAG.getNode(ISD::TRUNCATE, DL, VT,
4682 DAG.getNode(ISD::SRL, DL, InnerShiftVT,
4683 N0.getOperand(0)->getOperand(0),
4684 DAG.getConstant(c1 + c2, DL,
4689 // fold (srl (shl x, c), c) -> (and x, cst2)
4690 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1) {
4691 unsigned BitSize = N0.getScalarValueSizeInBits();
4692 if (BitSize <= 64) {
4693 uint64_t ShAmt = N1C->getZExtValue() + 64 - BitSize;
4695 return DAG.getNode(ISD::AND, DL, VT, N0.getOperand(0),
4696 DAG.getConstant(~0ULL >> ShAmt, DL, VT));
4700 // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask)
4701 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
4702 // Shifting in all undef bits?
4703 EVT SmallVT = N0.getOperand(0).getValueType();
4704 unsigned BitSize = SmallVT.getScalarSizeInBits();
4705 if (N1C->getZExtValue() >= BitSize)
4706 return DAG.getUNDEF(VT);
4708 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
4709 uint64_t ShiftAmt = N1C->getZExtValue();
4711 SDValue SmallShift = DAG.getNode(ISD::SRL, DL0, SmallVT,
4713 DAG.getConstant(ShiftAmt, DL0,
4714 getShiftAmountTy(SmallVT)));
4715 AddToWorklist(SmallShift.getNode());
4716 APInt Mask = APInt::getAllOnesValue(OpSizeInBits).lshr(ShiftAmt);
4718 return DAG.getNode(ISD::AND, DL, VT,
4719 DAG.getNode(ISD::ANY_EXTEND, DL, VT, SmallShift),
4720 DAG.getConstant(Mask, DL, VT));
4724 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
4725 // bit, which is unmodified by sra.
4726 if (N1C && N1C->getZExtValue() + 1 == OpSizeInBits) {
4727 if (N0.getOpcode() == ISD::SRA)
4728 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1);
4731 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
4732 if (N1C && N0.getOpcode() == ISD::CTLZ &&
4733 N1C->getAPIntValue() == Log2_32(OpSizeInBits)) {
4734 APInt KnownZero, KnownOne;
4735 DAG.computeKnownBits(N0.getOperand(0), KnownZero, KnownOne);
4737 // If any of the input bits are KnownOne, then the input couldn't be all
4738 // zeros, thus the result of the srl will always be zero.
4739 if (KnownOne.getBoolValue()) return DAG.getConstant(0, SDLoc(N0), VT);
4741 // If all of the bits input the to ctlz node are known to be zero, then
4742 // the result of the ctlz is "32" and the result of the shift is one.
4743 APInt UnknownBits = ~KnownZero;
4744 if (UnknownBits == 0) return DAG.getConstant(1, SDLoc(N0), VT);
4746 // Otherwise, check to see if there is exactly one bit input to the ctlz.
4747 if ((UnknownBits & (UnknownBits - 1)) == 0) {
4748 // Okay, we know that only that the single bit specified by UnknownBits
4749 // could be set on input to the CTLZ node. If this bit is set, the SRL
4750 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
4751 // to an SRL/XOR pair, which is likely to simplify more.
4752 unsigned ShAmt = UnknownBits.countTrailingZeros();
4753 SDValue Op = N0.getOperand(0);
4757 Op = DAG.getNode(ISD::SRL, DL, VT, Op,
4758 DAG.getConstant(ShAmt, DL,
4759 getShiftAmountTy(Op.getValueType())));
4760 AddToWorklist(Op.getNode());
4764 return DAG.getNode(ISD::XOR, DL, VT,
4765 Op, DAG.getConstant(1, DL, VT));
4769 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
4770 if (N1.getOpcode() == ISD::TRUNCATE &&
4771 N1.getOperand(0).getOpcode() == ISD::AND) {
4772 if (SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode()))
4773 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, NewOp1);
4776 // fold operands of srl based on knowledge that the low bits are not
4778 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4779 return SDValue(N, 0);
4781 if (N1C && !N1C->isOpaque())
4782 if (SDValue NewSRL = visitShiftByConstant(N, N1C))
4785 // Attempt to convert a srl of a load into a narrower zero-extending load.
4786 if (SDValue NarrowLoad = ReduceLoadWidth(N))
4789 // Here is a common situation. We want to optimize:
4792 // %b = and i32 %a, 2
4793 // %c = srl i32 %b, 1
4794 // brcond i32 %c ...
4800 // %c = setcc eq %b, 0
4803 // However when after the source operand of SRL is optimized into AND, the SRL
4804 // itself may not be optimized further. Look for it and add the BRCOND into
4806 if (N->hasOneUse()) {
4807 SDNode *Use = *N->use_begin();
4808 if (Use->getOpcode() == ISD::BRCOND)
4810 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
4811 // Also look pass the truncate.
4812 Use = *Use->use_begin();
4813 if (Use->getOpcode() == ISD::BRCOND)
4821 SDValue DAGCombiner::visitBSWAP(SDNode *N) {
4822 SDValue N0 = N->getOperand(0);
4823 EVT VT = N->getValueType(0);
4825 // fold (bswap c1) -> c2
4826 if (isConstantIntBuildVectorOrConstantInt(N0))
4827 return DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N0);
4828 // fold (bswap (bswap x)) -> x
4829 if (N0.getOpcode() == ISD::BSWAP)
4830 return N0->getOperand(0);
4834 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
4835 SDValue N0 = N->getOperand(0);
4836 EVT VT = N->getValueType(0);
4838 // fold (ctlz c1) -> c2
4839 if (isConstantIntBuildVectorOrConstantInt(N0))
4840 return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0);
4844 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4845 SDValue N0 = N->getOperand(0);
4846 EVT VT = N->getValueType(0);
4848 // fold (ctlz_zero_undef c1) -> c2
4849 if (isConstantIntBuildVectorOrConstantInt(N0))
4850 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4854 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4855 SDValue N0 = N->getOperand(0);
4856 EVT VT = N->getValueType(0);
4858 // fold (cttz c1) -> c2
4859 if (isConstantIntBuildVectorOrConstantInt(N0))
4860 return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0);
4864 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4865 SDValue N0 = N->getOperand(0);
4866 EVT VT = N->getValueType(0);
4868 // fold (cttz_zero_undef c1) -> c2
4869 if (isConstantIntBuildVectorOrConstantInt(N0))
4870 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4874 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4875 SDValue N0 = N->getOperand(0);
4876 EVT VT = N->getValueType(0);
4878 // fold (ctpop c1) -> c2
4879 if (isConstantIntBuildVectorOrConstantInt(N0))
4880 return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0);
4885 /// \brief Generate Min/Max node
4886 static SDValue combineMinNumMaxNum(SDLoc DL, EVT VT, SDValue LHS, SDValue RHS,
4887 SDValue True, SDValue False,
4888 ISD::CondCode CC, const TargetLowering &TLI,
4889 SelectionDAG &DAG) {
4890 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
4900 unsigned Opcode = (LHS == True) ? ISD::FMINNUM : ISD::FMAXNUM;
4901 if (TLI.isOperationLegal(Opcode, VT))
4902 return DAG.getNode(Opcode, DL, VT, LHS, RHS);
4911 unsigned Opcode = (LHS == True) ? ISD::FMAXNUM : ISD::FMINNUM;
4912 if (TLI.isOperationLegal(Opcode, VT))
4913 return DAG.getNode(Opcode, DL, VT, LHS, RHS);
4921 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4922 SDValue N0 = N->getOperand(0);
4923 SDValue N1 = N->getOperand(1);
4924 SDValue N2 = N->getOperand(2);
4925 EVT VT = N->getValueType(0);
4926 EVT VT0 = N0.getValueType();
4928 // fold (select C, X, X) -> X
4931 if (const ConstantSDNode *N0C = dyn_cast<const ConstantSDNode>(N0)) {
4932 // fold (select true, X, Y) -> X
4933 // fold (select false, X, Y) -> Y
4934 return !N0C->isNullValue() ? N1 : N2;
4936 // fold (select C, 1, X) -> (or C, X)
4937 if (VT == MVT::i1 && isOneConstant(N1))
4938 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4939 // fold (select C, 0, 1) -> (xor C, 1)
4940 // We can't do this reliably if integer based booleans have different contents
4941 // to floating point based booleans. This is because we can't tell whether we
4942 // have an integer-based boolean or a floating-point-based boolean unless we
4943 // can find the SETCC that produced it and inspect its operands. This is
4944 // fairly easy if C is the SETCC node, but it can potentially be
4945 // undiscoverable (or not reasonably discoverable). For example, it could be
4946 // in another basic block or it could require searching a complicated
4948 if (VT.isInteger() &&
4949 (VT0 == MVT::i1 || (VT0.isInteger() &&
4950 TLI.getBooleanContents(false, false) ==
4951 TLI.getBooleanContents(false, true) &&
4952 TLI.getBooleanContents(false, false) ==
4953 TargetLowering::ZeroOrOneBooleanContent)) &&
4954 isNullConstant(N1) && isOneConstant(N2)) {
4958 return DAG.getNode(ISD::XOR, DL, VT0,
4959 N0, DAG.getConstant(1, DL, VT0));
4962 XORNode = DAG.getNode(ISD::XOR, DL0, VT0,
4963 N0, DAG.getConstant(1, DL0, VT0));
4964 AddToWorklist(XORNode.getNode());
4966 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, XORNode);
4967 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, XORNode);
4969 // fold (select C, 0, X) -> (and (not C), X)
4970 if (VT == VT0 && VT == MVT::i1 && isNullConstant(N1)) {
4971 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4972 AddToWorklist(NOTNode.getNode());
4973 return DAG.getNode(ISD::AND, SDLoc(N), VT, NOTNode, N2);
4975 // fold (select C, X, 1) -> (or (not C), X)
4976 if (VT == VT0 && VT == MVT::i1 && isOneConstant(N2)) {
4977 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4978 AddToWorklist(NOTNode.getNode());
4979 return DAG.getNode(ISD::OR, SDLoc(N), VT, NOTNode, N1);
4981 // fold (select C, X, 0) -> (and C, X)
4982 if (VT == MVT::i1 && isNullConstant(N2))
4983 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4984 // fold (select X, X, Y) -> (or X, Y)
4985 // fold (select X, 1, Y) -> (or X, Y)
4986 if (VT == MVT::i1 && (N0 == N1 || isOneConstant(N1)))
4987 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4988 // fold (select X, Y, X) -> (and X, Y)
4989 // fold (select X, Y, 0) -> (and X, Y)
4990 if (VT == MVT::i1 && (N0 == N2 || isNullConstant(N2)))
4991 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4993 // If we can fold this based on the true/false value, do so.
4994 if (SimplifySelectOps(N, N1, N2))
4995 return SDValue(N, 0); // Don't revisit N.
4997 if (VT0 == MVT::i1) {
4998 // The code in this block deals with the following 2 equivalences:
4999 // select(C0|C1, x, y) <=> select(C0, x, select(C1, x, y))
5000 // select(C0&C1, x, y) <=> select(C0, select(C1, x, y), y)
5001 // The target can specify its prefered form with the
5002 // shouldNormalizeToSelectSequence() callback. However we always transform
5003 // to the right anyway if we find the inner select exists in the DAG anyway
5004 // and we always transform to the left side if we know that we can further
5005 // optimize the combination of the conditions.
5006 bool normalizeToSequence
5007 = TLI.shouldNormalizeToSelectSequence(*DAG.getContext(), VT);
5008 // select (and Cond0, Cond1), X, Y
5009 // -> select Cond0, (select Cond1, X, Y), Y
5010 if (N0->getOpcode() == ISD::AND && N0->hasOneUse()) {
5011 SDValue Cond0 = N0->getOperand(0);
5012 SDValue Cond1 = N0->getOperand(1);
5013 SDValue InnerSelect = DAG.getNode(ISD::SELECT, SDLoc(N),
5014 N1.getValueType(), Cond1, N1, N2);
5015 if (normalizeToSequence || !InnerSelect.use_empty())
5016 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Cond0,
5019 // select (or Cond0, Cond1), X, Y -> select Cond0, X, (select Cond1, X, Y)
5020 if (N0->getOpcode() == ISD::OR && N0->hasOneUse()) {
5021 SDValue Cond0 = N0->getOperand(0);
5022 SDValue Cond1 = N0->getOperand(1);
5023 SDValue InnerSelect = DAG.getNode(ISD::SELECT, SDLoc(N),
5024 N1.getValueType(), Cond1, N1, N2);
5025 if (normalizeToSequence || !InnerSelect.use_empty())
5026 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Cond0, N1,
5030 // select Cond0, (select Cond1, X, Y), Y -> select (and Cond0, Cond1), X, Y
5031 if (N1->getOpcode() == ISD::SELECT && N1->hasOneUse()) {
5032 SDValue N1_0 = N1->getOperand(0);
5033 SDValue N1_1 = N1->getOperand(1);
5034 SDValue N1_2 = N1->getOperand(2);
5035 if (N1_2 == N2 && N0.getValueType() == N1_0.getValueType()) {
5036 // Create the actual and node if we can generate good code for it.
5037 if (!normalizeToSequence) {
5038 SDValue And = DAG.getNode(ISD::AND, SDLoc(N), N0.getValueType(),
5040 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), And,
5043 // Otherwise see if we can optimize the "and" to a better pattern.
5044 if (SDValue Combined = visitANDLike(N0, N1_0, N))
5045 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Combined,
5049 // select Cond0, X, (select Cond1, X, Y) -> select (or Cond0, Cond1), X, Y
5050 if (N2->getOpcode() == ISD::SELECT && N2->hasOneUse()) {
5051 SDValue N2_0 = N2->getOperand(0);
5052 SDValue N2_1 = N2->getOperand(1);
5053 SDValue N2_2 = N2->getOperand(2);
5054 if (N2_1 == N1 && N0.getValueType() == N2_0.getValueType()) {
5055 // Create the actual or node if we can generate good code for it.
5056 if (!normalizeToSequence) {
5057 SDValue Or = DAG.getNode(ISD::OR, SDLoc(N), N0.getValueType(),
5059 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Or,
5062 // Otherwise see if we can optimize to a better pattern.
5063 if (SDValue Combined = visitORLike(N0, N2_0, N))
5064 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Combined,
5070 // fold selects based on a setcc into other things, such as min/max/abs
5071 if (N0.getOpcode() == ISD::SETCC) {
5072 // select x, y (fcmp lt x, y) -> fminnum x, y
5073 // select x, y (fcmp gt x, y) -> fmaxnum x, y
5075 // This is OK if we don't care about what happens if either operand is a
5079 // FIXME: Instead of testing for UnsafeFPMath, this should be checking for
5080 // no signed zeros as well as no nans.
5081 const TargetOptions &Options = DAG.getTarget().Options;
5082 if (Options.UnsafeFPMath &&
5083 VT.isFloatingPoint() && N0.hasOneUse() &&
5084 DAG.isKnownNeverNaN(N1) && DAG.isKnownNeverNaN(N2)) {
5085 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5087 if (SDValue FMinMax = combineMinNumMaxNum(SDLoc(N), VT, N0.getOperand(0),
5088 N0.getOperand(1), N1, N2, CC,
5093 if ((!LegalOperations &&
5094 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT)) ||
5095 TLI.isOperationLegal(ISD::SELECT_CC, VT))
5096 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT,
5097 N0.getOperand(0), N0.getOperand(1),
5098 N1, N2, N0.getOperand(2));
5099 return SimplifySelect(SDLoc(N), N0, N1, N2);
5106 std::pair<SDValue, SDValue> SplitVSETCC(const SDNode *N, SelectionDAG &DAG) {
5109 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
5111 // Split the inputs.
5112 SDValue Lo, Hi, LL, LH, RL, RH;
5113 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
5114 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
5116 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));
5117 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));
5119 return std::make_pair(Lo, Hi);
5122 // This function assumes all the vselect's arguments are CONCAT_VECTOR
5123 // nodes and that the condition is a BV of ConstantSDNodes (or undefs).
5124 static SDValue ConvertSelectToConcatVector(SDNode *N, SelectionDAG &DAG) {
5126 SDValue Cond = N->getOperand(0);
5127 SDValue LHS = N->getOperand(1);
5128 SDValue RHS = N->getOperand(2);
5129 EVT VT = N->getValueType(0);
5130 int NumElems = VT.getVectorNumElements();
5131 assert(LHS.getOpcode() == ISD::CONCAT_VECTORS &&
5132 RHS.getOpcode() == ISD::CONCAT_VECTORS &&
5133 Cond.getOpcode() == ISD::BUILD_VECTOR);
5135 // CONCAT_VECTOR can take an arbitrary number of arguments. We only care about
5136 // binary ones here.
5137 if (LHS->getNumOperands() != 2 || RHS->getNumOperands() != 2)
5140 // We're sure we have an even number of elements due to the
5141 // concat_vectors we have as arguments to vselect.
5142 // Skip BV elements until we find one that's not an UNDEF
5143 // After we find an UNDEF element, keep looping until we get to half the
5144 // length of the BV and see if all the non-undef nodes are the same.
5145 ConstantSDNode *BottomHalf = nullptr;
5146 for (int i = 0; i < NumElems / 2; ++i) {
5147 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
5150 if (BottomHalf == nullptr)
5151 BottomHalf = cast<ConstantSDNode>(Cond.getOperand(i));
5152 else if (Cond->getOperand(i).getNode() != BottomHalf)
5156 // Do the same for the second half of the BuildVector
5157 ConstantSDNode *TopHalf = nullptr;
5158 for (int i = NumElems / 2; i < NumElems; ++i) {
5159 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
5162 if (TopHalf == nullptr)
5163 TopHalf = cast<ConstantSDNode>(Cond.getOperand(i));
5164 else if (Cond->getOperand(i).getNode() != TopHalf)
5168 assert(TopHalf && BottomHalf &&
5169 "One half of the selector was all UNDEFs and the other was all the "
5170 "same value. This should have been addressed before this function.");
5172 ISD::CONCAT_VECTORS, dl, VT,
5173 BottomHalf->isNullValue() ? RHS->getOperand(0) : LHS->getOperand(0),
5174 TopHalf->isNullValue() ? RHS->getOperand(1) : LHS->getOperand(1));
5177 SDValue DAGCombiner::visitMSCATTER(SDNode *N) {
5179 if (Level >= AfterLegalizeTypes)
5182 MaskedScatterSDNode *MSC = cast<MaskedScatterSDNode>(N);
5183 SDValue Mask = MSC->getMask();
5184 SDValue Data = MSC->getValue();
5187 // If the MSCATTER data type requires splitting and the mask is provided by a
5188 // SETCC, then split both nodes and its operands before legalization. This
5189 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5190 // and enables future optimizations (e.g. min/max pattern matching on X86).
5191 if (Mask.getOpcode() != ISD::SETCC)
5194 // Check if any splitting is required.
5195 if (TLI.getTypeAction(*DAG.getContext(), Data.getValueType()) !=
5196 TargetLowering::TypeSplitVector)
5198 SDValue MaskLo, MaskHi, Lo, Hi;
5199 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
5202 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MSC->getValueType(0));
5204 SDValue Chain = MSC->getChain();
5206 EVT MemoryVT = MSC->getMemoryVT();
5207 unsigned Alignment = MSC->getOriginalAlignment();
5209 EVT LoMemVT, HiMemVT;
5210 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
5212 SDValue DataLo, DataHi;
5213 std::tie(DataLo, DataHi) = DAG.SplitVector(Data, DL);
5215 SDValue BasePtr = MSC->getBasePtr();
5216 SDValue IndexLo, IndexHi;
5217 std::tie(IndexLo, IndexHi) = DAG.SplitVector(MSC->getIndex(), DL);
5219 MachineMemOperand *MMO = DAG.getMachineFunction().
5220 getMachineMemOperand(MSC->getPointerInfo(),
5221 MachineMemOperand::MOStore, LoMemVT.getStoreSize(),
5222 Alignment, MSC->getAAInfo(), MSC->getRanges());
5224 SDValue OpsLo[] = { Chain, DataLo, MaskLo, BasePtr, IndexLo };
5225 Lo = DAG.getMaskedScatter(DAG.getVTList(MVT::Other), DataLo.getValueType(),
5228 SDValue OpsHi[] = {Chain, DataHi, MaskHi, BasePtr, IndexHi};
5229 Hi = DAG.getMaskedScatter(DAG.getVTList(MVT::Other), DataHi.getValueType(),
5232 AddToWorklist(Lo.getNode());
5233 AddToWorklist(Hi.getNode());
5235 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);
5238 SDValue DAGCombiner::visitMSTORE(SDNode *N) {
5240 if (Level >= AfterLegalizeTypes)
5243 MaskedStoreSDNode *MST = dyn_cast<MaskedStoreSDNode>(N);
5244 SDValue Mask = MST->getMask();
5245 SDValue Data = MST->getValue();
5248 // If the MSTORE data type requires splitting and the mask is provided by a
5249 // SETCC, then split both nodes and its operands before legalization. This
5250 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5251 // and enables future optimizations (e.g. min/max pattern matching on X86).
5252 if (Mask.getOpcode() == ISD::SETCC) {
5254 // Check if any splitting is required.
5255 if (TLI.getTypeAction(*DAG.getContext(), Data.getValueType()) !=
5256 TargetLowering::TypeSplitVector)
5259 SDValue MaskLo, MaskHi, Lo, Hi;
5260 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
5263 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MST->getValueType(0));
5265 SDValue Chain = MST->getChain();
5266 SDValue Ptr = MST->getBasePtr();
5268 EVT MemoryVT = MST->getMemoryVT();
5269 unsigned Alignment = MST->getOriginalAlignment();
5271 // if Alignment is equal to the vector size,
5272 // take the half of it for the second part
5273 unsigned SecondHalfAlignment =
5274 (Alignment == Data->getValueType(0).getSizeInBits()/8) ?
5275 Alignment/2 : Alignment;
5277 EVT LoMemVT, HiMemVT;
5278 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
5280 SDValue DataLo, DataHi;
5281 std::tie(DataLo, DataHi) = DAG.SplitVector(Data, DL);
5283 MachineMemOperand *MMO = DAG.getMachineFunction().
5284 getMachineMemOperand(MST->getPointerInfo(),
5285 MachineMemOperand::MOStore, LoMemVT.getStoreSize(),
5286 Alignment, MST->getAAInfo(), MST->getRanges());
5288 Lo = DAG.getMaskedStore(Chain, DL, DataLo, Ptr, MaskLo, LoMemVT, MMO,
5289 MST->isTruncatingStore());
5291 unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
5292 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5293 DAG.getConstant(IncrementSize, DL, Ptr.getValueType()));
5295 MMO = DAG.getMachineFunction().
5296 getMachineMemOperand(MST->getPointerInfo(),
5297 MachineMemOperand::MOStore, HiMemVT.getStoreSize(),
5298 SecondHalfAlignment, MST->getAAInfo(),
5301 Hi = DAG.getMaskedStore(Chain, DL, DataHi, Ptr, MaskHi, HiMemVT, MMO,
5302 MST->isTruncatingStore());
5304 AddToWorklist(Lo.getNode());
5305 AddToWorklist(Hi.getNode());
5307 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);
5312 SDValue DAGCombiner::visitMGATHER(SDNode *N) {
5314 if (Level >= AfterLegalizeTypes)
5317 MaskedGatherSDNode *MGT = dyn_cast<MaskedGatherSDNode>(N);
5318 SDValue Mask = MGT->getMask();
5321 // If the MGATHER result requires splitting and the mask is provided by a
5322 // SETCC, then split both nodes and its operands before legalization. This
5323 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5324 // and enables future optimizations (e.g. min/max pattern matching on X86).
5326 if (Mask.getOpcode() != ISD::SETCC)
5329 EVT VT = N->getValueType(0);
5331 // Check if any splitting is required.
5332 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
5333 TargetLowering::TypeSplitVector)
5336 SDValue MaskLo, MaskHi, Lo, Hi;
5337 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
5339 SDValue Src0 = MGT->getValue();
5340 SDValue Src0Lo, Src0Hi;
5341 std::tie(Src0Lo, Src0Hi) = DAG.SplitVector(Src0, DL);
5344 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
5346 SDValue Chain = MGT->getChain();
5347 EVT MemoryVT = MGT->getMemoryVT();
5348 unsigned Alignment = MGT->getOriginalAlignment();
5350 EVT LoMemVT, HiMemVT;
5351 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
5353 SDValue BasePtr = MGT->getBasePtr();
5354 SDValue Index = MGT->getIndex();
5355 SDValue IndexLo, IndexHi;
5356 std::tie(IndexLo, IndexHi) = DAG.SplitVector(Index, DL);
5358 MachineMemOperand *MMO = DAG.getMachineFunction().
5359 getMachineMemOperand(MGT->getPointerInfo(),
5360 MachineMemOperand::MOLoad, LoMemVT.getStoreSize(),
5361 Alignment, MGT->getAAInfo(), MGT->getRanges());
5363 SDValue OpsLo[] = { Chain, Src0Lo, MaskLo, BasePtr, IndexLo };
5364 Lo = DAG.getMaskedGather(DAG.getVTList(LoVT, MVT::Other), LoVT, DL, OpsLo,
5367 SDValue OpsHi[] = {Chain, Src0Hi, MaskHi, BasePtr, IndexHi};
5368 Hi = DAG.getMaskedGather(DAG.getVTList(HiVT, MVT::Other), HiVT, DL, OpsHi,
5371 AddToWorklist(Lo.getNode());
5372 AddToWorklist(Hi.getNode());
5374 // Build a factor node to remember that this load is independent of the
5376 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo.getValue(1),
5379 // Legalized the chain result - switch anything that used the old chain to
5381 DAG.ReplaceAllUsesOfValueWith(SDValue(MGT, 1), Chain);
5383 SDValue GatherRes = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
5385 SDValue RetOps[] = { GatherRes, Chain };
5386 return DAG.getMergeValues(RetOps, DL);
5389 SDValue DAGCombiner::visitMLOAD(SDNode *N) {
5391 if (Level >= AfterLegalizeTypes)
5394 MaskedLoadSDNode *MLD = dyn_cast<MaskedLoadSDNode>(N);
5395 SDValue Mask = MLD->getMask();
5398 // If the MLOAD result requires splitting and the mask is provided by a
5399 // SETCC, then split both nodes and its operands before legalization. This
5400 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5401 // and enables future optimizations (e.g. min/max pattern matching on X86).
5403 if (Mask.getOpcode() == ISD::SETCC) {
5404 EVT VT = N->getValueType(0);
5406 // Check if any splitting is required.
5407 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
5408 TargetLowering::TypeSplitVector)
5411 SDValue MaskLo, MaskHi, Lo, Hi;
5412 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
5414 SDValue Src0 = MLD->getSrc0();
5415 SDValue Src0Lo, Src0Hi;
5416 std::tie(Src0Lo, Src0Hi) = DAG.SplitVector(Src0, DL);
5419 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MLD->getValueType(0));
5421 SDValue Chain = MLD->getChain();
5422 SDValue Ptr = MLD->getBasePtr();
5423 EVT MemoryVT = MLD->getMemoryVT();
5424 unsigned Alignment = MLD->getOriginalAlignment();
5426 // if Alignment is equal to the vector size,
5427 // take the half of it for the second part
5428 unsigned SecondHalfAlignment =
5429 (Alignment == MLD->getValueType(0).getSizeInBits()/8) ?
5430 Alignment/2 : Alignment;
5432 EVT LoMemVT, HiMemVT;
5433 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
5435 MachineMemOperand *MMO = DAG.getMachineFunction().
5436 getMachineMemOperand(MLD->getPointerInfo(),
5437 MachineMemOperand::MOLoad, LoMemVT.getStoreSize(),
5438 Alignment, MLD->getAAInfo(), MLD->getRanges());
5440 Lo = DAG.getMaskedLoad(LoVT, DL, Chain, Ptr, MaskLo, Src0Lo, LoMemVT, MMO,
5443 unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
5444 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5445 DAG.getConstant(IncrementSize, DL, Ptr.getValueType()));
5447 MMO = DAG.getMachineFunction().
5448 getMachineMemOperand(MLD->getPointerInfo(),
5449 MachineMemOperand::MOLoad, HiMemVT.getStoreSize(),
5450 SecondHalfAlignment, MLD->getAAInfo(), MLD->getRanges());
5452 Hi = DAG.getMaskedLoad(HiVT, DL, Chain, Ptr, MaskHi, Src0Hi, HiMemVT, MMO,
5455 AddToWorklist(Lo.getNode());
5456 AddToWorklist(Hi.getNode());
5458 // Build a factor node to remember that this load is independent of the
5460 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo.getValue(1),
5463 // Legalized the chain result - switch anything that used the old chain to
5465 DAG.ReplaceAllUsesOfValueWith(SDValue(MLD, 1), Chain);
5467 SDValue LoadRes = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
5469 SDValue RetOps[] = { LoadRes, Chain };
5470 return DAG.getMergeValues(RetOps, DL);
5475 SDValue DAGCombiner::visitVSELECT(SDNode *N) {
5476 SDValue N0 = N->getOperand(0);
5477 SDValue N1 = N->getOperand(1);
5478 SDValue N2 = N->getOperand(2);
5481 // Canonicalize integer abs.
5482 // vselect (setg[te] X, 0), X, -X ->
5483 // vselect (setgt X, -1), X, -X ->
5484 // vselect (setl[te] X, 0), -X, X ->
5485 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
5486 if (N0.getOpcode() == ISD::SETCC) {
5487 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
5488 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5490 bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
5492 if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
5493 (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) &&
5494 N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1))
5495 isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode());
5496 else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) &&
5497 N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1))
5498 isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode());
5501 EVT VT = LHS.getValueType();
5502 SDValue Shift = DAG.getNode(
5503 ISD::SRA, DL, VT, LHS,
5504 DAG.getConstant(VT.getScalarType().getSizeInBits() - 1, DL, VT));
5505 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift);
5506 AddToWorklist(Shift.getNode());
5507 AddToWorklist(Add.getNode());
5508 return DAG.getNode(ISD::XOR, DL, VT, Add, Shift);
5512 if (SimplifySelectOps(N, N1, N2))
5513 return SDValue(N, 0); // Don't revisit N.
5515 // If the VSELECT result requires splitting and the mask is provided by a
5516 // SETCC, then split both nodes and its operands before legalization. This
5517 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5518 // and enables future optimizations (e.g. min/max pattern matching on X86).
5519 if (N0.getOpcode() == ISD::SETCC) {
5520 EVT VT = N->getValueType(0);
5522 // Check if any splitting is required.
5523 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
5524 TargetLowering::TypeSplitVector)
5527 SDValue Lo, Hi, CCLo, CCHi, LL, LH, RL, RH;
5528 std::tie(CCLo, CCHi) = SplitVSETCC(N0.getNode(), DAG);
5529 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 1);
5530 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 2);
5532 Lo = DAG.getNode(N->getOpcode(), DL, LL.getValueType(), CCLo, LL, RL);
5533 Hi = DAG.getNode(N->getOpcode(), DL, LH.getValueType(), CCHi, LH, RH);
5535 // Add the new VSELECT nodes to the work list in case they need to be split
5537 AddToWorklist(Lo.getNode());
5538 AddToWorklist(Hi.getNode());
5540 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
5543 // Fold (vselect (build_vector all_ones), N1, N2) -> N1
5544 if (ISD::isBuildVectorAllOnes(N0.getNode()))
5546 // Fold (vselect (build_vector all_zeros), N1, N2) -> N2
5547 if (ISD::isBuildVectorAllZeros(N0.getNode()))
5550 // The ConvertSelectToConcatVector function is assuming both the above
5551 // checks for (vselect (build_vector all{ones,zeros) ...) have been made
5553 if (N1.getOpcode() == ISD::CONCAT_VECTORS &&
5554 N2.getOpcode() == ISD::CONCAT_VECTORS &&
5555 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
5556 if (SDValue CV = ConvertSelectToConcatVector(N, DAG))
5563 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
5564 SDValue N0 = N->getOperand(0);
5565 SDValue N1 = N->getOperand(1);
5566 SDValue N2 = N->getOperand(2);
5567 SDValue N3 = N->getOperand(3);
5568 SDValue N4 = N->getOperand(4);
5569 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
5571 // fold select_cc lhs, rhs, x, x, cc -> x
5575 // Determine if the condition we're dealing with is constant
5576 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
5577 N0, N1, CC, SDLoc(N), false);
5578 if (SCC.getNode()) {
5579 AddToWorklist(SCC.getNode());
5581 if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) {
5582 if (!SCCC->isNullValue())
5583 return N2; // cond always true -> true val
5585 return N3; // cond always false -> false val
5586 } else if (SCC->getOpcode() == ISD::UNDEF) {
5587 // When the condition is UNDEF, just return the first operand. This is
5588 // coherent the DAG creation, no setcc node is created in this case
5590 } else if (SCC.getOpcode() == ISD::SETCC) {
5591 // Fold to a simpler select_cc
5592 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(),
5593 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
5598 // If we can fold this based on the true/false value, do so.
5599 if (SimplifySelectOps(N, N2, N3))
5600 return SDValue(N, 0); // Don't revisit N.
5602 // fold select_cc into other things, such as min/max/abs
5603 return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC);
5606 SDValue DAGCombiner::visitSETCC(SDNode *N) {
5607 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
5608 cast<CondCodeSDNode>(N->getOperand(2))->get(),
5612 /// Try to fold a sext/zext/aext dag node into a ConstantSDNode or
5613 /// a build_vector of constants.
5614 /// This function is called by the DAGCombiner when visiting sext/zext/aext
5615 /// dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND).
5616 /// Vector extends are not folded if operations are legal; this is to
5617 /// avoid introducing illegal build_vector dag nodes.
5618 static SDNode *tryToFoldExtendOfConstant(SDNode *N, const TargetLowering &TLI,
5619 SelectionDAG &DAG, bool LegalTypes,
5620 bool LegalOperations) {
5621 unsigned Opcode = N->getOpcode();
5622 SDValue N0 = N->getOperand(0);
5623 EVT VT = N->getValueType(0);
5625 assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND ||
5626 Opcode == ISD::ANY_EXTEND || Opcode == ISD::SIGN_EXTEND_VECTOR_INREG)
5627 && "Expected EXTEND dag node in input!");
5629 // fold (sext c1) -> c1
5630 // fold (zext c1) -> c1
5631 // fold (aext c1) -> c1
5632 if (isa<ConstantSDNode>(N0))
5633 return DAG.getNode(Opcode, SDLoc(N), VT, N0).getNode();
5635 // fold (sext (build_vector AllConstants) -> (build_vector AllConstants)
5636 // fold (zext (build_vector AllConstants) -> (build_vector AllConstants)
5637 // fold (aext (build_vector AllConstants) -> (build_vector AllConstants)
5638 EVT SVT = VT.getScalarType();
5639 if (!(VT.isVector() &&
5640 (!LegalTypes || (!LegalOperations && TLI.isTypeLegal(SVT))) &&
5641 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())))
5644 // We can fold this node into a build_vector.
5645 unsigned VTBits = SVT.getSizeInBits();
5646 unsigned EVTBits = N0->getValueType(0).getScalarType().getSizeInBits();
5647 SmallVector<SDValue, 8> Elts;
5648 unsigned NumElts = VT.getVectorNumElements();
5651 for (unsigned i=0; i != NumElts; ++i) {
5652 SDValue Op = N0->getOperand(i);
5653 if (Op->getOpcode() == ISD::UNDEF) {
5654 Elts.push_back(DAG.getUNDEF(SVT));
5659 // Get the constant value and if needed trunc it to the size of the type.
5660 // Nodes like build_vector might have constants wider than the scalar type.
5661 APInt C = cast<ConstantSDNode>(Op)->getAPIntValue().zextOrTrunc(EVTBits);
5662 if (Opcode == ISD::SIGN_EXTEND || Opcode == ISD::SIGN_EXTEND_VECTOR_INREG)
5663 Elts.push_back(DAG.getConstant(C.sext(VTBits), DL, SVT));
5665 Elts.push_back(DAG.getConstant(C.zext(VTBits), DL, SVT));
5668 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Elts).getNode();
5671 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
5672 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
5673 // transformation. Returns true if extension are possible and the above
5674 // mentioned transformation is profitable.
5675 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
5677 SmallVectorImpl<SDNode *> &ExtendNodes,
5678 const TargetLowering &TLI) {
5679 bool HasCopyToRegUses = false;
5680 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
5681 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
5682 UE = N0.getNode()->use_end();
5687 if (UI.getUse().getResNo() != N0.getResNo())
5689 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
5690 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
5691 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
5692 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
5693 // Sign bits will be lost after a zext.
5696 for (unsigned i = 0; i != 2; ++i) {
5697 SDValue UseOp = User->getOperand(i);
5700 if (!isa<ConstantSDNode>(UseOp))
5705 ExtendNodes.push_back(User);
5708 // If truncates aren't free and there are users we can't
5709 // extend, it isn't worthwhile.
5712 // Remember if this value is live-out.
5713 if (User->getOpcode() == ISD::CopyToReg)
5714 HasCopyToRegUses = true;
5717 if (HasCopyToRegUses) {
5718 bool BothLiveOut = false;
5719 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
5721 SDUse &Use = UI.getUse();
5722 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
5728 // Both unextended and extended values are live out. There had better be
5729 // a good reason for the transformation.
5730 return ExtendNodes.size();
5735 void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
5736 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
5737 ISD::NodeType ExtType) {
5738 // Extend SetCC uses if necessary.
5739 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
5740 SDNode *SetCC = SetCCs[i];
5741 SmallVector<SDValue, 4> Ops;
5743 for (unsigned j = 0; j != 2; ++j) {
5744 SDValue SOp = SetCC->getOperand(j);
5746 Ops.push_back(ExtLoad);
5748 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
5751 Ops.push_back(SetCC->getOperand(2));
5752 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), Ops));
5756 // FIXME: Bring more similar combines here, common to sext/zext (maybe aext?).
5757 SDValue DAGCombiner::CombineExtLoad(SDNode *N) {
5758 SDValue N0 = N->getOperand(0);
5759 EVT DstVT = N->getValueType(0);
5760 EVT SrcVT = N0.getValueType();
5762 assert((N->getOpcode() == ISD::SIGN_EXTEND ||
5763 N->getOpcode() == ISD::ZERO_EXTEND) &&
5764 "Unexpected node type (not an extend)!");
5766 // fold (sext (load x)) to multiple smaller sextloads; same for zext.
5767 // For example, on a target with legal v4i32, but illegal v8i32, turn:
5768 // (v8i32 (sext (v8i16 (load x))))
5770 // (v8i32 (concat_vectors (v4i32 (sextload x)),
5771 // (v4i32 (sextload (x + 16)))))
5772 // Where uses of the original load, i.e.:
5774 // are replaced with:
5776 // (v8i32 (concat_vectors (v4i32 (sextload x)),
5777 // (v4i32 (sextload (x + 16)))))))
5779 // This combine is only applicable to illegal, but splittable, vectors.
5780 // All legal types, and illegal non-vector types, are handled elsewhere.
5781 // This combine is controlled by TargetLowering::isVectorLoadExtDesirable.
5783 if (N0->getOpcode() != ISD::LOAD)
5786 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5788 if (!ISD::isNON_EXTLoad(LN0) || !ISD::isUNINDEXEDLoad(LN0) ||
5789 !N0.hasOneUse() || LN0->isVolatile() || !DstVT.isVector() ||
5790 !DstVT.isPow2VectorType() || !TLI.isVectorLoadExtDesirable(SDValue(N, 0)))
5793 SmallVector<SDNode *, 4> SetCCs;
5794 if (!ExtendUsesToFormExtLoad(N, N0, N->getOpcode(), SetCCs, TLI))
5797 ISD::LoadExtType ExtType =
5798 N->getOpcode() == ISD::SIGN_EXTEND ? ISD::SEXTLOAD : ISD::ZEXTLOAD;
5800 // Try to split the vector types to get down to legal types.
5801 EVT SplitSrcVT = SrcVT;
5802 EVT SplitDstVT = DstVT;
5803 while (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT) &&
5804 SplitSrcVT.getVectorNumElements() > 1) {
5805 SplitDstVT = DAG.GetSplitDestVTs(SplitDstVT).first;
5806 SplitSrcVT = DAG.GetSplitDestVTs(SplitSrcVT).first;
5809 if (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT))
5813 const unsigned NumSplits =
5814 DstVT.getVectorNumElements() / SplitDstVT.getVectorNumElements();
5815 const unsigned Stride = SplitSrcVT.getStoreSize();
5816 SmallVector<SDValue, 4> Loads;
5817 SmallVector<SDValue, 4> Chains;
5819 SDValue BasePtr = LN0->getBasePtr();
5820 for (unsigned Idx = 0; Idx < NumSplits; Idx++) {
5821 const unsigned Offset = Idx * Stride;
5822 const unsigned Align = MinAlign(LN0->getAlignment(), Offset);
5824 SDValue SplitLoad = DAG.getExtLoad(
5825 ExtType, DL, SplitDstVT, LN0->getChain(), BasePtr,
5826 LN0->getPointerInfo().getWithOffset(Offset), SplitSrcVT,
5827 LN0->isVolatile(), LN0->isNonTemporal(), LN0->isInvariant(),
5828 Align, LN0->getAAInfo());
5830 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
5831 DAG.getConstant(Stride, DL, BasePtr.getValueType()));
5833 Loads.push_back(SplitLoad.getValue(0));
5834 Chains.push_back(SplitLoad.getValue(1));
5837 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
5838 SDValue NewValue = DAG.getNode(ISD::CONCAT_VECTORS, DL, DstVT, Loads);
5840 CombineTo(N, NewValue);
5842 // Replace uses of the original load (before extension)
5843 // with a truncate of the concatenated sextloaded vectors.
5845 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(), NewValue);
5846 CombineTo(N0.getNode(), Trunc, NewChain);
5847 ExtendSetCCUses(SetCCs, Trunc, NewValue, DL,
5848 (ISD::NodeType)N->getOpcode());
5849 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5852 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
5853 SDValue N0 = N->getOperand(0);
5854 EVT VT = N->getValueType(0);
5856 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5858 return SDValue(Res, 0);
5860 // fold (sext (sext x)) -> (sext x)
5861 // fold (sext (aext x)) -> (sext x)
5862 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5863 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT,
5866 if (N0.getOpcode() == ISD::TRUNCATE) {
5867 // fold (sext (truncate (load x))) -> (sext (smaller load x))
5868 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
5869 if (SDValue NarrowLoad = ReduceLoadWidth(N0.getNode())) {
5870 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5871 if (NarrowLoad.getNode() != N0.getNode()) {
5872 CombineTo(N0.getNode(), NarrowLoad);
5873 // CombineTo deleted the truncate, if needed, but not what's under it.
5876 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5879 // See if the value being truncated is already sign extended. If so, just
5880 // eliminate the trunc/sext pair.
5881 SDValue Op = N0.getOperand(0);
5882 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
5883 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
5884 unsigned DestBits = VT.getScalarType().getSizeInBits();
5885 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
5887 if (OpBits == DestBits) {
5888 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
5889 // bits, it is already ready.
5890 if (NumSignBits > DestBits-MidBits)
5892 } else if (OpBits < DestBits) {
5893 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
5894 // bits, just sext from i32.
5895 if (NumSignBits > OpBits-MidBits)
5896 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, Op);
5898 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
5899 // bits, just truncate to i32.
5900 if (NumSignBits > OpBits-MidBits)
5901 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5904 // fold (sext (truncate x)) -> (sextinreg x).
5905 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
5906 N0.getValueType())) {
5907 if (OpBits < DestBits)
5908 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op);
5909 else if (OpBits > DestBits)
5910 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op);
5911 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, Op,
5912 DAG.getValueType(N0.getValueType()));
5916 // fold (sext (load x)) -> (sext (truncate (sextload x)))
5917 // Only generate vector extloads when 1) they're legal, and 2) they are
5918 // deemed desirable by the target.
5919 if (ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5920 ((!LegalOperations && !VT.isVector() &&
5921 !cast<LoadSDNode>(N0)->isVolatile()) ||
5922 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, N0.getValueType()))) {
5923 bool DoXform = true;
5924 SmallVector<SDNode*, 4> SetCCs;
5925 if (!N0.hasOneUse())
5926 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
5928 DoXform &= TLI.isVectorLoadExtDesirable(SDValue(N, 0));
5930 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5931 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5933 LN0->getBasePtr(), N0.getValueType(),
5934 LN0->getMemOperand());
5935 CombineTo(N, ExtLoad);
5936 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5937 N0.getValueType(), ExtLoad);
5938 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5939 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5941 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5945 // fold (sext (load x)) to multiple smaller sextloads.
5946 // Only on illegal but splittable vectors.
5947 if (SDValue ExtLoad = CombineExtLoad(N))
5950 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
5951 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
5952 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5953 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5954 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5955 EVT MemVT = LN0->getMemoryVT();
5956 if ((!LegalOperations && !LN0->isVolatile()) ||
5957 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, MemVT)) {
5958 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5960 LN0->getBasePtr(), MemVT,
5961 LN0->getMemOperand());
5962 CombineTo(N, ExtLoad);
5963 CombineTo(N0.getNode(),
5964 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5965 N0.getValueType(), ExtLoad),
5966 ExtLoad.getValue(1));
5967 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5971 // fold (sext (and/or/xor (load x), cst)) ->
5972 // (and/or/xor (sextload x), (sext cst))
5973 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5974 N0.getOpcode() == ISD::XOR) &&
5975 isa<LoadSDNode>(N0.getOperand(0)) &&
5976 N0.getOperand(1).getOpcode() == ISD::Constant &&
5977 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, N0.getValueType()) &&
5978 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5979 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5980 if (LN0->getExtensionType() != ISD::ZEXTLOAD && LN0->isUnindexed()) {
5981 bool DoXform = true;
5982 SmallVector<SDNode*, 4> SetCCs;
5983 if (!N0.hasOneUse())
5984 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
5987 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN0), VT,
5988 LN0->getChain(), LN0->getBasePtr(),
5990 LN0->getMemOperand());
5991 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5992 Mask = Mask.sext(VT.getSizeInBits());
5994 SDValue And = DAG.getNode(N0.getOpcode(), DL, VT,
5995 ExtLoad, DAG.getConstant(Mask, DL, VT));
5996 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5997 SDLoc(N0.getOperand(0)),
5998 N0.getOperand(0).getValueType(), ExtLoad);
6000 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
6001 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, DL,
6003 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6008 if (N0.getOpcode() == ISD::SETCC) {
6009 EVT N0VT = N0.getOperand(0).getValueType();
6010 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
6011 // Only do this before legalize for now.
6012 if (VT.isVector() && !LegalOperations &&
6013 TLI.getBooleanContents(N0VT) ==
6014 TargetLowering::ZeroOrNegativeOneBooleanContent) {
6015 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
6016 // of the same size as the compared operands. Only optimize sext(setcc())
6017 // if this is the case.
6018 EVT SVT = getSetCCResultType(N0VT);
6020 // We know that the # elements of the results is the same as the
6021 // # elements of the compare (and the # elements of the compare result
6022 // for that matter). Check to see that they are the same size. If so,
6023 // we know that the element size of the sext'd result matches the
6024 // element size of the compare operands.
6025 if (VT.getSizeInBits() == SVT.getSizeInBits())
6026 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
6028 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6030 // If the desired elements are smaller or larger than the source
6031 // elements we can use a matching integer vector type and then
6032 // truncate/sign extend
6033 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
6034 if (SVT == MatchingVectorType) {
6035 SDValue VsetCC = DAG.getSetCC(SDLoc(N), MatchingVectorType,
6036 N0.getOperand(0), N0.getOperand(1),
6037 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6038 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
6042 // sext(setcc x, y, cc) -> (select (setcc x, y, cc), -1, 0)
6043 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
6046 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), DL, VT);
6048 SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1),
6049 NegOne, DAG.getConstant(0, DL, VT),
6050 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
6051 if (SCC.getNode()) return SCC;
6053 if (!VT.isVector()) {
6054 EVT SetCCVT = getSetCCResultType(N0.getOperand(0).getValueType());
6055 if (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, SetCCVT)) {
6057 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
6058 SDValue SetCC = DAG.getSetCC(DL, SetCCVT,
6059 N0.getOperand(0), N0.getOperand(1), CC);
6060 return DAG.getSelect(DL, VT, SetCC,
6061 NegOne, DAG.getConstant(0, DL, VT));
6066 // fold (sext x) -> (zext x) if the sign bit is known zero.
6067 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
6068 DAG.SignBitIsZero(N0))
6069 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0);
6074 // isTruncateOf - If N is a truncate of some other value, return true, record
6075 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
6076 // This function computes KnownZero to avoid a duplicated call to
6077 // computeKnownBits in the caller.
6078 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
6081 if (N->getOpcode() == ISD::TRUNCATE) {
6082 Op = N->getOperand(0);
6083 DAG.computeKnownBits(Op, KnownZero, KnownOne);
6087 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
6088 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
6091 SDValue Op0 = N->getOperand(0);
6092 SDValue Op1 = N->getOperand(1);
6093 assert(Op0.getValueType() == Op1.getValueType());
6095 if (isNullConstant(Op0))
6097 else if (isNullConstant(Op1))
6102 DAG.computeKnownBits(Op, KnownZero, KnownOne);
6104 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
6110 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
6111 SDValue N0 = N->getOperand(0);
6112 EVT VT = N->getValueType(0);
6114 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
6116 return SDValue(Res, 0);
6118 // fold (zext (zext x)) -> (zext x)
6119 // fold (zext (aext x)) -> (zext x)
6120 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
6121 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT,
6124 // fold (zext (truncate x)) -> (zext x) or
6125 // (zext (truncate x)) -> (truncate x)
6126 // This is valid when the truncated bits of x are already zero.
6127 // FIXME: We should extend this to work for vectors too.
6130 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
6131 APInt TruncatedBits =
6132 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
6133 APInt(Op.getValueSizeInBits(), 0) :
6134 APInt::getBitsSet(Op.getValueSizeInBits(),
6135 N0.getValueSizeInBits(),
6136 std::min(Op.getValueSizeInBits(),
6137 VT.getSizeInBits()));
6138 if (TruncatedBits == (KnownZero & TruncatedBits)) {
6139 if (VT.bitsGT(Op.getValueType()))
6140 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, Op);
6141 if (VT.bitsLT(Op.getValueType()))
6142 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
6148 // fold (zext (truncate (load x))) -> (zext (smaller load x))
6149 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
6150 if (N0.getOpcode() == ISD::TRUNCATE) {
6151 if (SDValue NarrowLoad = ReduceLoadWidth(N0.getNode())) {
6152 SDNode* oye = N0.getNode()->getOperand(0).getNode();
6153 if (NarrowLoad.getNode() != N0.getNode()) {
6154 CombineTo(N0.getNode(), NarrowLoad);
6155 // CombineTo deleted the truncate, if needed, but not what's under it.
6158 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6162 // fold (zext (truncate x)) -> (and x, mask)
6163 if (N0.getOpcode() == ISD::TRUNCATE) {
6164 // fold (zext (truncate (load x))) -> (zext (smaller load x))
6165 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
6166 if (SDValue NarrowLoad = ReduceLoadWidth(N0.getNode())) {
6167 SDNode *oye = N0.getNode()->getOperand(0).getNode();
6168 if (NarrowLoad.getNode() != N0.getNode()) {
6169 CombineTo(N0.getNode(), NarrowLoad);
6170 // CombineTo deleted the truncate, if needed, but not what's under it.
6173 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6176 EVT SrcVT = N0.getOperand(0).getValueType();
6177 EVT MinVT = N0.getValueType();
6179 // Try to mask before the extension to avoid having to generate a larger mask,
6180 // possibly over several sub-vectors.
6181 if (SrcVT.bitsLT(VT)) {
6182 if (!LegalOperations || (TLI.isOperationLegal(ISD::AND, SrcVT) &&
6183 TLI.isOperationLegal(ISD::ZERO_EXTEND, VT))) {
6184 SDValue Op = N0.getOperand(0);
6185 Op = DAG.getZeroExtendInReg(Op, SDLoc(N), MinVT.getScalarType());
6186 AddToWorklist(Op.getNode());
6187 return DAG.getZExtOrTrunc(Op, SDLoc(N), VT);
6191 if (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT)) {
6192 SDValue Op = N0.getOperand(0);
6193 if (SrcVT.bitsLT(VT)) {
6194 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, Op);
6195 AddToWorklist(Op.getNode());
6196 } else if (SrcVT.bitsGT(VT)) {
6197 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
6198 AddToWorklist(Op.getNode());
6200 return DAG.getZeroExtendInReg(Op, SDLoc(N), MinVT.getScalarType());
6204 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
6205 // if either of the casts is not free.
6206 if (N0.getOpcode() == ISD::AND &&
6207 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
6208 N0.getOperand(1).getOpcode() == ISD::Constant &&
6209 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
6210 N0.getValueType()) ||
6211 !TLI.isZExtFree(N0.getValueType(), VT))) {
6212 SDValue X = N0.getOperand(0).getOperand(0);
6213 if (X.getValueType().bitsLT(VT)) {
6214 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(X), VT, X);
6215 } else if (X.getValueType().bitsGT(VT)) {
6216 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
6218 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
6219 Mask = Mask.zext(VT.getSizeInBits());
6221 return DAG.getNode(ISD::AND, DL, VT,
6222 X, DAG.getConstant(Mask, DL, VT));
6225 // fold (zext (load x)) -> (zext (truncate (zextload x)))
6226 // Only generate vector extloads when 1) they're legal, and 2) they are
6227 // deemed desirable by the target.
6228 if (ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
6229 ((!LegalOperations && !VT.isVector() &&
6230 !cast<LoadSDNode>(N0)->isVolatile()) ||
6231 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, N0.getValueType()))) {
6232 bool DoXform = true;
6233 SmallVector<SDNode*, 4> SetCCs;
6234 if (!N0.hasOneUse())
6235 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
6237 DoXform &= TLI.isVectorLoadExtDesirable(SDValue(N, 0));
6239 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6240 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
6242 LN0->getBasePtr(), N0.getValueType(),
6243 LN0->getMemOperand());
6244 CombineTo(N, ExtLoad);
6245 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
6246 N0.getValueType(), ExtLoad);
6247 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
6249 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
6251 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6255 // fold (zext (load x)) to multiple smaller zextloads.
6256 // Only on illegal but splittable vectors.
6257 if (SDValue ExtLoad = CombineExtLoad(N))
6260 // fold (zext (and/or/xor (load x), cst)) ->
6261 // (and/or/xor (zextload x), (zext cst))
6262 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
6263 N0.getOpcode() == ISD::XOR) &&
6264 isa<LoadSDNode>(N0.getOperand(0)) &&
6265 N0.getOperand(1).getOpcode() == ISD::Constant &&
6266 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, N0.getValueType()) &&
6267 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
6268 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
6269 if (LN0->getExtensionType() != ISD::SEXTLOAD && LN0->isUnindexed()) {
6270 bool DoXform = true;
6271 SmallVector<SDNode*, 4> SetCCs;
6272 if (!N0.hasOneUse())
6273 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
6276 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), VT,
6277 LN0->getChain(), LN0->getBasePtr(),
6279 LN0->getMemOperand());
6280 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
6281 Mask = Mask.zext(VT.getSizeInBits());
6283 SDValue And = DAG.getNode(N0.getOpcode(), DL, VT,
6284 ExtLoad, DAG.getConstant(Mask, DL, VT));
6285 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
6286 SDLoc(N0.getOperand(0)),
6287 N0.getOperand(0).getValueType(), ExtLoad);
6289 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
6290 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, DL,
6292 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6297 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
6298 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
6299 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
6300 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
6301 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6302 EVT MemVT = LN0->getMemoryVT();
6303 if ((!LegalOperations && !LN0->isVolatile()) ||
6304 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT)) {
6305 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
6307 LN0->getBasePtr(), MemVT,
6308 LN0->getMemOperand());
6309 CombineTo(N, ExtLoad);
6310 CombineTo(N0.getNode(),
6311 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(),
6313 ExtLoad.getValue(1));
6314 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6318 if (N0.getOpcode() == ISD::SETCC) {
6319 if (!LegalOperations && VT.isVector() &&
6320 N0.getValueType().getVectorElementType() == MVT::i1) {
6321 EVT N0VT = N0.getOperand(0).getValueType();
6322 if (getSetCCResultType(N0VT) == N0.getValueType())
6325 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
6326 // Only do this before legalize for now.
6327 EVT EltVT = VT.getVectorElementType();
6329 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
6330 DAG.getConstant(1, DL, EltVT));
6331 if (VT.getSizeInBits() == N0VT.getSizeInBits())
6332 // We know that the # elements of the results is the same as the
6333 // # elements of the compare (and the # elements of the compare result
6334 // for that matter). Check to see that they are the same size. If so,
6335 // we know that the element size of the sext'd result matches the
6336 // element size of the compare operands.
6337 return DAG.getNode(ISD::AND, DL, VT,
6338 DAG.getSetCC(DL, VT, N0.getOperand(0),
6340 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
6341 DAG.getNode(ISD::BUILD_VECTOR, DL, VT,
6344 // If the desired elements are smaller or larger than the source
6345 // elements we can use a matching integer vector type and then
6346 // truncate/sign extend
6347 EVT MatchingElementType =
6348 EVT::getIntegerVT(*DAG.getContext(),
6349 N0VT.getScalarType().getSizeInBits());
6350 EVT MatchingVectorType =
6351 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
6352 N0VT.getVectorNumElements());
6354 DAG.getSetCC(DL, MatchingVectorType, N0.getOperand(0),
6356 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6357 return DAG.getNode(ISD::AND, DL, VT,
6358 DAG.getSExtOrTrunc(VsetCC, DL, VT),
6359 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, OneOps));
6362 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
6365 SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1),
6366 DAG.getConstant(1, DL, VT), DAG.getConstant(0, DL, VT),
6367 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
6368 if (SCC.getNode()) return SCC;
6371 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
6372 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
6373 isa<ConstantSDNode>(N0.getOperand(1)) &&
6374 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
6376 SDValue ShAmt = N0.getOperand(1);
6377 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
6378 if (N0.getOpcode() == ISD::SHL) {
6379 SDValue InnerZExt = N0.getOperand(0);
6380 // If the original shl may be shifting out bits, do not perform this
6382 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
6383 InnerZExt.getOperand(0).getValueType().getSizeInBits();
6384 if (ShAmtVal > KnownZeroBits)
6390 // Ensure that the shift amount is wide enough for the shifted value.
6391 if (VT.getSizeInBits() >= 256)
6392 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
6394 return DAG.getNode(N0.getOpcode(), DL, VT,
6395 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
6402 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
6403 SDValue N0 = N->getOperand(0);
6404 EVT VT = N->getValueType(0);
6406 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
6408 return SDValue(Res, 0);
6410 // fold (aext (aext x)) -> (aext x)
6411 // fold (aext (zext x)) -> (zext x)
6412 // fold (aext (sext x)) -> (sext x)
6413 if (N0.getOpcode() == ISD::ANY_EXTEND ||
6414 N0.getOpcode() == ISD::ZERO_EXTEND ||
6415 N0.getOpcode() == ISD::SIGN_EXTEND)
6416 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0));
6418 // fold (aext (truncate (load x))) -> (aext (smaller load x))
6419 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
6420 if (N0.getOpcode() == ISD::TRUNCATE) {
6421 if (SDValue NarrowLoad = ReduceLoadWidth(N0.getNode())) {
6422 SDNode* oye = N0.getNode()->getOperand(0).getNode();
6423 if (NarrowLoad.getNode() != N0.getNode()) {
6424 CombineTo(N0.getNode(), NarrowLoad);
6425 // CombineTo deleted the truncate, if needed, but not what's under it.
6428 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6432 // fold (aext (truncate x))
6433 if (N0.getOpcode() == ISD::TRUNCATE) {
6434 SDValue TruncOp = N0.getOperand(0);
6435 if (TruncOp.getValueType() == VT)
6436 return TruncOp; // x iff x size == zext size.
6437 if (TruncOp.getValueType().bitsGT(VT))
6438 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, TruncOp);
6439 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, TruncOp);
6442 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
6443 // if the trunc is not free.
6444 if (N0.getOpcode() == ISD::AND &&
6445 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
6446 N0.getOperand(1).getOpcode() == ISD::Constant &&
6447 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
6448 N0.getValueType())) {
6449 SDValue X = N0.getOperand(0).getOperand(0);
6450 if (X.getValueType().bitsLT(VT)) {
6451 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, X);
6452 } else if (X.getValueType().bitsGT(VT)) {
6453 X = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, X);
6455 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
6456 Mask = Mask.zext(VT.getSizeInBits());
6458 return DAG.getNode(ISD::AND, DL, VT,
6459 X, DAG.getConstant(Mask, DL, VT));
6462 // fold (aext (load x)) -> (aext (truncate (extload x)))
6463 // None of the supported targets knows how to perform load and any_ext
6464 // on vectors in one instruction. We only perform this transformation on
6466 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
6467 ISD::isUNINDEXEDLoad(N0.getNode()) &&
6468 TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) {
6469 bool DoXform = true;
6470 SmallVector<SDNode*, 4> SetCCs;
6471 if (!N0.hasOneUse())
6472 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
6474 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6475 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
6477 LN0->getBasePtr(), N0.getValueType(),
6478 LN0->getMemOperand());
6479 CombineTo(N, ExtLoad);
6480 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
6481 N0.getValueType(), ExtLoad);
6482 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
6483 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
6485 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6489 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
6490 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
6491 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
6492 if (N0.getOpcode() == ISD::LOAD &&
6493 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
6495 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6496 ISD::LoadExtType ExtType = LN0->getExtensionType();
6497 EVT MemVT = LN0->getMemoryVT();
6498 if (!LegalOperations || TLI.isLoadExtLegal(ExtType, VT, MemVT)) {
6499 SDValue ExtLoad = DAG.getExtLoad(ExtType, SDLoc(N),
6500 VT, LN0->getChain(), LN0->getBasePtr(),
6501 MemVT, LN0->getMemOperand());
6502 CombineTo(N, ExtLoad);
6503 CombineTo(N0.getNode(),
6504 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
6505 N0.getValueType(), ExtLoad),
6506 ExtLoad.getValue(1));
6507 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6511 if (N0.getOpcode() == ISD::SETCC) {
6513 // aext(setcc) -> vsetcc
6514 // aext(setcc) -> truncate(vsetcc)
6515 // aext(setcc) -> aext(vsetcc)
6516 // Only do this before legalize for now.
6517 if (VT.isVector() && !LegalOperations) {
6518 EVT N0VT = N0.getOperand(0).getValueType();
6519 // We know that the # elements of the results is the same as the
6520 // # elements of the compare (and the # elements of the compare result
6521 // for that matter). Check to see that they are the same size. If so,
6522 // we know that the element size of the sext'd result matches the
6523 // element size of the compare operands.
6524 if (VT.getSizeInBits() == N0VT.getSizeInBits())
6525 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
6527 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6528 // If the desired elements are smaller or larger than the source
6529 // elements we can use a matching integer vector type and then
6530 // truncate/any extend
6532 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
6534 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
6536 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6537 return DAG.getAnyExtOrTrunc(VsetCC, SDLoc(N), VT);
6541 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
6544 SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1),
6545 DAG.getConstant(1, DL, VT), DAG.getConstant(0, DL, VT),
6546 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
6554 /// See if the specified operand can be simplified with the knowledge that only
6555 /// the bits specified by Mask are used. If so, return the simpler operand,
6556 /// otherwise return a null SDValue.
6557 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
6558 switch (V.getOpcode()) {
6560 case ISD::Constant: {
6561 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
6562 assert(CV && "Const value should be ConstSDNode.");
6563 const APInt &CVal = CV->getAPIntValue();
6564 APInt NewVal = CVal & Mask;
6566 return DAG.getConstant(NewVal, SDLoc(V), V.getValueType());
6571 // If the LHS or RHS don't contribute bits to the or, drop them.
6572 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
6573 return V.getOperand(1);
6574 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
6575 return V.getOperand(0);
6578 // Only look at single-use SRLs.
6579 if (!V.getNode()->hasOneUse())
6581 if (ConstantSDNode *RHSC = getAsNonOpaqueConstant(V.getOperand(1))) {
6582 // See if we can recursively simplify the LHS.
6583 unsigned Amt = RHSC->getZExtValue();
6585 // Watch out for shift count overflow though.
6586 if (Amt >= Mask.getBitWidth()) break;
6587 APInt NewMask = Mask << Amt;
6588 if (SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask))
6589 return DAG.getNode(ISD::SRL, SDLoc(V), V.getValueType(),
6590 SimplifyLHS, V.getOperand(1));
6596 /// If the result of a wider load is shifted to right of N bits and then
6597 /// truncated to a narrower type and where N is a multiple of number of bits of
6598 /// the narrower type, transform it to a narrower load from address + N / num of
6599 /// bits of new type. If the result is to be extended, also fold the extension
6600 /// to form a extending load.
6601 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
6602 unsigned Opc = N->getOpcode();
6604 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
6605 SDValue N0 = N->getOperand(0);
6606 EVT VT = N->getValueType(0);
6609 // This transformation isn't valid for vector loads.
6613 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
6615 if (Opc == ISD::SIGN_EXTEND_INREG) {
6616 ExtType = ISD::SEXTLOAD;
6617 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
6618 } else if (Opc == ISD::SRL) {
6619 // Another special-case: SRL is basically zero-extending a narrower value.
6620 ExtType = ISD::ZEXTLOAD;
6622 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
6623 if (!N01) return SDValue();
6624 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
6625 VT.getSizeInBits() - N01->getZExtValue());
6627 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, VT, ExtVT))
6630 unsigned EVTBits = ExtVT.getSizeInBits();
6632 // Do not generate loads of non-round integer types since these can
6633 // be expensive (and would be wrong if the type is not byte sized).
6634 if (!ExtVT.isRound())
6638 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
6639 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
6640 ShAmt = N01->getZExtValue();
6641 // Is the shift amount a multiple of size of VT?
6642 if ((ShAmt & (EVTBits-1)) == 0) {
6643 N0 = N0.getOperand(0);
6644 // Is the load width a multiple of size of VT?
6645 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
6649 // At this point, we must have a load or else we can't do the transform.
6650 if (!isa<LoadSDNode>(N0)) return SDValue();
6652 // Because a SRL must be assumed to *need* to zero-extend the high bits
6653 // (as opposed to anyext the high bits), we can't combine the zextload
6654 // lowering of SRL and an sextload.
6655 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
6658 // If the shift amount is larger than the input type then we're not
6659 // accessing any of the loaded bytes. If the load was a zextload/extload
6660 // then the result of the shift+trunc is zero/undef (handled elsewhere).
6661 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
6666 // If the load is shifted left (and the result isn't shifted back right),
6667 // we can fold the truncate through the shift.
6668 unsigned ShLeftAmt = 0;
6669 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
6670 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
6671 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
6672 ShLeftAmt = N01->getZExtValue();
6673 N0 = N0.getOperand(0);
6677 // If we haven't found a load, we can't narrow it. Don't transform one with
6678 // multiple uses, this would require adding a new load.
6679 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse())
6682 // Don't change the width of a volatile load.
6683 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6684 if (LN0->isVolatile())
6687 // Verify that we are actually reducing a load width here.
6688 if (LN0->getMemoryVT().getSizeInBits() < EVTBits)
6691 // For the transform to be legal, the load must produce only two values
6692 // (the value loaded and the chain). Don't transform a pre-increment
6693 // load, for example, which produces an extra value. Otherwise the
6694 // transformation is not equivalent, and the downstream logic to replace
6695 // uses gets things wrong.
6696 if (LN0->getNumValues() > 2)
6699 // If the load that we're shrinking is an extload and we're not just
6700 // discarding the extension we can't simply shrink the load. Bail.
6701 // TODO: It would be possible to merge the extensions in some cases.
6702 if (LN0->getExtensionType() != ISD::NON_EXTLOAD &&
6703 LN0->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits() + ShAmt)
6706 if (!TLI.shouldReduceLoadWidth(LN0, ExtType, ExtVT))
6709 EVT PtrType = N0.getOperand(1).getValueType();
6711 if (PtrType == MVT::Untyped || PtrType.isExtended())
6712 // It's not possible to generate a constant of extended or untyped type.
6715 // For big endian targets, we need to adjust the offset to the pointer to
6716 // load the correct bytes.
6717 if (DAG.getDataLayout().isBigEndian()) {
6718 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
6719 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
6720 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
6723 uint64_t PtrOff = ShAmt / 8;
6724 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
6726 SDValue NewPtr = DAG.getNode(ISD::ADD, DL,
6727 PtrType, LN0->getBasePtr(),
6728 DAG.getConstant(PtrOff, DL, PtrType));
6729 AddToWorklist(NewPtr.getNode());
6732 if (ExtType == ISD::NON_EXTLOAD)
6733 Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr,
6734 LN0->getPointerInfo().getWithOffset(PtrOff),
6735 LN0->isVolatile(), LN0->isNonTemporal(),
6736 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
6738 Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(),NewPtr,
6739 LN0->getPointerInfo().getWithOffset(PtrOff),
6740 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
6741 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
6743 // Replace the old load's chain with the new load's chain.
6744 WorklistRemover DeadNodes(*this);
6745 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
6747 // Shift the result left, if we've swallowed a left shift.
6748 SDValue Result = Load;
6749 if (ShLeftAmt != 0) {
6750 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
6751 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
6753 // If the shift amount is as large as the result size (but, presumably,
6754 // no larger than the source) then the useful bits of the result are
6755 // zero; we can't simply return the shortened shift, because the result
6756 // of that operation is undefined.
6758 if (ShLeftAmt >= VT.getSizeInBits())
6759 Result = DAG.getConstant(0, DL, VT);
6761 Result = DAG.getNode(ISD::SHL, DL, VT,
6762 Result, DAG.getConstant(ShLeftAmt, DL, ShImmTy));
6765 // Return the new loaded value.
6769 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
6770 SDValue N0 = N->getOperand(0);
6771 SDValue N1 = N->getOperand(1);
6772 EVT VT = N->getValueType(0);
6773 EVT EVT = cast<VTSDNode>(N1)->getVT();
6774 unsigned VTBits = VT.getScalarType().getSizeInBits();
6775 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
6777 // fold (sext_in_reg c1) -> c1
6778 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
6779 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1);
6781 // If the input is already sign extended, just drop the extension.
6782 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
6785 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
6786 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
6787 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT()))
6788 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
6789 N0.getOperand(0), N1);
6791 // fold (sext_in_reg (sext x)) -> (sext x)
6792 // fold (sext_in_reg (aext x)) -> (sext x)
6793 // if x is small enough.
6794 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
6795 SDValue N00 = N0.getOperand(0);
6796 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
6797 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
6798 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
6801 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
6802 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
6803 return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT);
6805 // fold operands of sext_in_reg based on knowledge that the top bits are not
6807 if (SimplifyDemandedBits(SDValue(N, 0)))
6808 return SDValue(N, 0);
6810 // fold (sext_in_reg (load x)) -> (smaller sextload x)
6811 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
6812 if (SDValue NarrowLoad = ReduceLoadWidth(N))
6815 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
6816 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
6817 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
6818 if (N0.getOpcode() == ISD::SRL) {
6819 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
6820 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
6821 // We can turn this into an SRA iff the input to the SRL is already sign
6823 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
6824 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
6825 return DAG.getNode(ISD::SRA, SDLoc(N), VT,
6826 N0.getOperand(0), N0.getOperand(1));
6830 // fold (sext_inreg (extload x)) -> (sextload x)
6831 if (ISD::isEXTLoad(N0.getNode()) &&
6832 ISD::isUNINDEXEDLoad(N0.getNode()) &&
6833 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
6834 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6835 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) {
6836 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6837 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
6839 LN0->getBasePtr(), EVT,
6840 LN0->getMemOperand());
6841 CombineTo(N, ExtLoad);
6842 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
6843 AddToWorklist(ExtLoad.getNode());
6844 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6846 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
6847 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
6849 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
6850 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6851 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) {
6852 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6853 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
6855 LN0->getBasePtr(), EVT,
6856 LN0->getMemOperand());
6857 CombineTo(N, ExtLoad);
6858 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
6859 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6862 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
6863 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
6864 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
6865 N0.getOperand(1), false);
6866 if (BSwap.getNode())
6867 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
6871 // Fold a sext_inreg of a build_vector of ConstantSDNodes or undefs
6872 // into a build_vector.
6873 if (ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
6874 SmallVector<SDValue, 8> Elts;
6875 unsigned NumElts = N0->getNumOperands();
6876 unsigned ShAmt = VTBits - EVTBits;
6878 for (unsigned i = 0; i != NumElts; ++i) {
6879 SDValue Op = N0->getOperand(i);
6880 if (Op->getOpcode() == ISD::UNDEF) {
6885 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
6886 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
6887 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
6888 SDLoc(Op), Op.getValueType()));
6891 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Elts);
6897 SDValue DAGCombiner::visitSIGN_EXTEND_VECTOR_INREG(SDNode *N) {
6898 SDValue N0 = N->getOperand(0);
6899 EVT VT = N->getValueType(0);
6901 if (N0.getOpcode() == ISD::UNDEF)
6902 return DAG.getUNDEF(VT);
6904 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
6906 return SDValue(Res, 0);
6911 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
6912 SDValue N0 = N->getOperand(0);
6913 EVT VT = N->getValueType(0);
6914 bool isLE = DAG.getDataLayout().isLittleEndian();
6917 if (N0.getValueType() == N->getValueType(0))
6919 // fold (truncate c1) -> c1
6920 if (isConstantIntBuildVectorOrConstantInt(N0))
6921 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0);
6922 // fold (truncate (truncate x)) -> (truncate x)
6923 if (N0.getOpcode() == ISD::TRUNCATE)
6924 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6925 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
6926 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
6927 N0.getOpcode() == ISD::SIGN_EXTEND ||
6928 N0.getOpcode() == ISD::ANY_EXTEND) {
6929 if (N0.getOperand(0).getValueType().bitsLT(VT))
6930 // if the source is smaller than the dest, we still need an extend
6931 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
6933 if (N0.getOperand(0).getValueType().bitsGT(VT))
6934 // if the source is larger than the dest, than we just need the truncate
6935 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6936 // if the source and dest are the same type, we can drop both the extend
6937 // and the truncate.
6938 return N0.getOperand(0);
6941 // Fold extract-and-trunc into a narrow extract. For example:
6942 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
6943 // i32 y = TRUNCATE(i64 x)
6945 // v16i8 b = BITCAST (v2i64 val)
6946 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
6948 // Note: We only run this optimization after type legalization (which often
6949 // creates this pattern) and before operation legalization after which
6950 // we need to be more careful about the vector instructions that we generate.
6951 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6952 LegalTypes && !LegalOperations && N0->hasOneUse() && VT != MVT::i1) {
6954 EVT VecTy = N0.getOperand(0).getValueType();
6955 EVT ExTy = N0.getValueType();
6956 EVT TrTy = N->getValueType(0);
6958 unsigned NumElem = VecTy.getVectorNumElements();
6959 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
6961 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
6962 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
6964 SDValue EltNo = N0->getOperand(1);
6965 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
6966 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
6967 EVT IndexTy = TLI.getVectorIdxTy(DAG.getDataLayout());
6968 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
6970 SDValue V = DAG.getNode(ISD::BITCAST, SDLoc(N),
6971 NVT, N0.getOperand(0));
6974 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
6976 DAG.getConstant(Index, DL, IndexTy));
6980 // trunc (select c, a, b) -> select c, (trunc a), (trunc b)
6981 if (N0.getOpcode() == ISD::SELECT) {
6982 EVT SrcVT = N0.getValueType();
6983 if ((!LegalOperations || TLI.isOperationLegal(ISD::SELECT, SrcVT)) &&
6984 TLI.isTruncateFree(SrcVT, VT)) {
6986 SDValue Cond = N0.getOperand(0);
6987 SDValue TruncOp0 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(1));
6988 SDValue TruncOp1 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(2));
6989 return DAG.getNode(ISD::SELECT, SDLoc(N), VT, Cond, TruncOp0, TruncOp1);
6993 // Fold a series of buildvector, bitcast, and truncate if possible.
6995 // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to
6996 // (2xi32 (buildvector x, y)).
6997 if (Level == AfterLegalizeVectorOps && VT.isVector() &&
6998 N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
6999 N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
7000 N0.getOperand(0).hasOneUse()) {
7002 SDValue BuildVect = N0.getOperand(0);
7003 EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType();
7004 EVT TruncVecEltTy = VT.getVectorElementType();
7006 // Check that the element types match.
7007 if (BuildVectEltTy == TruncVecEltTy) {
7008 // Now we only need to compute the offset of the truncated elements.
7009 unsigned BuildVecNumElts = BuildVect.getNumOperands();
7010 unsigned TruncVecNumElts = VT.getVectorNumElements();
7011 unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts;
7013 assert((BuildVecNumElts % TruncVecNumElts) == 0 &&
7014 "Invalid number of elements");
7016 SmallVector<SDValue, 8> Opnds;
7017 for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset)
7018 Opnds.push_back(BuildVect.getOperand(i));
7020 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
7024 // See if we can simplify the input to this truncate through knowledge that
7025 // only the low bits are being used.
7026 // For example "trunc (or (shl x, 8), y)" // -> trunc y
7027 // Currently we only perform this optimization on scalars because vectors
7028 // may have different active low bits.
7029 if (!VT.isVector()) {
7031 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
7032 VT.getSizeInBits()));
7033 if (Shorter.getNode())
7034 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter);
7036 // fold (truncate (load x)) -> (smaller load x)
7037 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
7038 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
7039 if (SDValue Reduced = ReduceLoadWidth(N))
7042 // Handle the case where the load remains an extending load even
7043 // after truncation.
7044 if (N0.hasOneUse() && ISD::isUNINDEXEDLoad(N0.getNode())) {
7045 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7046 if (!LN0->isVolatile() &&
7047 LN0->getMemoryVT().getStoreSizeInBits() < VT.getSizeInBits()) {
7048 SDValue NewLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(LN0),
7049 VT, LN0->getChain(), LN0->getBasePtr(),
7051 LN0->getMemOperand());
7052 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLoad.getValue(1));
7057 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
7058 // where ... are all 'undef'.
7059 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
7060 SmallVector<EVT, 8> VTs;
7063 unsigned NumDefs = 0;
7065 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
7066 SDValue X = N0.getOperand(i);
7067 if (X.getOpcode() != ISD::UNDEF) {
7072 // Stop if more than one members are non-undef.
7075 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
7076 VT.getVectorElementType(),
7077 X.getValueType().getVectorNumElements()));
7081 return DAG.getUNDEF(VT);
7084 assert(V.getNode() && "The single defined operand is empty!");
7085 SmallVector<SDValue, 8> Opnds;
7086 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
7088 Opnds.push_back(DAG.getUNDEF(VTs[i]));
7091 SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V);
7092 AddToWorklist(NV.getNode());
7093 Opnds.push_back(NV);
7095 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Opnds);
7099 // Simplify the operands using demanded-bits information.
7100 if (!VT.isVector() &&
7101 SimplifyDemandedBits(SDValue(N, 0)))
7102 return SDValue(N, 0);
7107 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
7108 SDValue Elt = N->getOperand(i);
7109 if (Elt.getOpcode() != ISD::MERGE_VALUES)
7110 return Elt.getNode();
7111 return Elt.getOperand(Elt.getResNo()).getNode();
7114 /// build_pair (load, load) -> load
7115 /// if load locations are consecutive.
7116 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
7117 assert(N->getOpcode() == ISD::BUILD_PAIR);
7119 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
7120 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
7121 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
7122 LD1->getAddressSpace() != LD2->getAddressSpace())
7124 EVT LD1VT = LD1->getValueType(0);
7126 if (ISD::isNON_EXTLoad(LD2) &&
7128 // If both are volatile this would reduce the number of volatile loads.
7129 // If one is volatile it might be ok, but play conservative and bail out.
7130 !LD1->isVolatile() &&
7131 !LD2->isVolatile() &&
7132 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
7133 unsigned Align = LD1->getAlignment();
7134 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
7135 VT.getTypeForEVT(*DAG.getContext()));
7137 if (NewAlign <= Align &&
7138 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
7139 return DAG.getLoad(VT, SDLoc(N), LD1->getChain(),
7140 LD1->getBasePtr(), LD1->getPointerInfo(),
7141 false, false, false, Align);
7147 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
7148 SDValue N0 = N->getOperand(0);
7149 EVT VT = N->getValueType(0);
7151 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
7152 // Only do this before legalize, since afterward the target may be depending
7153 // on the bitconvert.
7154 // First check to see if this is all constant.
7156 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
7158 bool isSimple = cast<BuildVectorSDNode>(N0)->isConstant();
7160 EVT DestEltVT = N->getValueType(0).getVectorElementType();
7161 assert(!DestEltVT.isVector() &&
7162 "Element type of vector ValueType must not be vector!");
7164 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
7167 // If the input is a constant, let getNode fold it.
7168 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
7169 // If we can't allow illegal operations, we need to check that this is just
7170 // a fp -> int or int -> conversion and that the resulting operation will
7172 if (!LegalOperations ||
7173 (isa<ConstantSDNode>(N0) && VT.isFloatingPoint() && !VT.isVector() &&
7174 TLI.isOperationLegal(ISD::ConstantFP, VT)) ||
7175 (isa<ConstantFPSDNode>(N0) && VT.isInteger() && !VT.isVector() &&
7176 TLI.isOperationLegal(ISD::Constant, VT)))
7177 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, N0);
7180 // (conv (conv x, t1), t2) -> (conv x, t2)
7181 if (N0.getOpcode() == ISD::BITCAST)
7182 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT,
7185 // fold (conv (load x)) -> (load (conv*)x)
7186 // If the resultant load doesn't need a higher alignment than the original!
7187 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7188 // Do not change the width of a volatile load.
7189 !cast<LoadSDNode>(N0)->isVolatile() &&
7190 // Do not remove the cast if the types differ in endian layout.
7191 TLI.hasBigEndianPartOrdering(N0.getValueType(), DAG.getDataLayout()) ==
7192 TLI.hasBigEndianPartOrdering(VT, DAG.getDataLayout()) &&
7193 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)) &&
7194 TLI.isLoadBitCastBeneficial(N0.getValueType(), VT)) {
7195 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7196 unsigned Align = DAG.getDataLayout().getABITypeAlignment(
7197 VT.getTypeForEVT(*DAG.getContext()));
7198 unsigned OrigAlign = LN0->getAlignment();
7200 if (Align <= OrigAlign) {
7201 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(),
7202 LN0->getBasePtr(), LN0->getPointerInfo(),
7203 LN0->isVolatile(), LN0->isNonTemporal(),
7204 LN0->isInvariant(), OrigAlign,
7206 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
7211 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
7212 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
7213 // This often reduces constant pool loads.
7214 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) ||
7215 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) &&
7216 N0.getNode()->hasOneUse() && VT.isInteger() &&
7217 !VT.isVector() && !N0.getValueType().isVector()) {
7218 SDValue NewConv = DAG.getNode(ISD::BITCAST, SDLoc(N0), VT,
7220 AddToWorklist(NewConv.getNode());
7223 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
7224 if (N0.getOpcode() == ISD::FNEG)
7225 return DAG.getNode(ISD::XOR, DL, VT,
7226 NewConv, DAG.getConstant(SignBit, DL, VT));
7227 assert(N0.getOpcode() == ISD::FABS);
7228 return DAG.getNode(ISD::AND, DL, VT,
7229 NewConv, DAG.getConstant(~SignBit, DL, VT));
7232 // fold (bitconvert (fcopysign cst, x)) ->
7233 // (or (and (bitconvert x), sign), (and cst, (not sign)))
7234 // Note that we don't handle (copysign x, cst) because this can always be
7235 // folded to an fneg or fabs.
7236 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
7237 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
7238 VT.isInteger() && !VT.isVector()) {
7239 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
7240 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
7241 if (isTypeLegal(IntXVT)) {
7242 SDValue X = DAG.getNode(ISD::BITCAST, SDLoc(N0),
7243 IntXVT, N0.getOperand(1));
7244 AddToWorklist(X.getNode());
7246 // If X has a different width than the result/lhs, sext it or truncate it.
7247 unsigned VTWidth = VT.getSizeInBits();
7248 if (OrigXWidth < VTWidth) {
7249 X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X);
7250 AddToWorklist(X.getNode());
7251 } else if (OrigXWidth > VTWidth) {
7252 // To get the sign bit in the right place, we have to shift it right
7253 // before truncating.
7255 X = DAG.getNode(ISD::SRL, DL,
7256 X.getValueType(), X,
7257 DAG.getConstant(OrigXWidth-VTWidth, DL,
7259 AddToWorklist(X.getNode());
7260 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
7261 AddToWorklist(X.getNode());
7264 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
7265 X = DAG.getNode(ISD::AND, SDLoc(X), VT,
7266 X, DAG.getConstant(SignBit, SDLoc(X), VT));
7267 AddToWorklist(X.getNode());
7269 SDValue Cst = DAG.getNode(ISD::BITCAST, SDLoc(N0),
7270 VT, N0.getOperand(0));
7271 Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT,
7272 Cst, DAG.getConstant(~SignBit, SDLoc(Cst), VT));
7273 AddToWorklist(Cst.getNode());
7275 return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst);
7279 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
7280 if (N0.getOpcode() == ISD::BUILD_PAIR)
7281 if (SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT))
7284 // Remove double bitcasts from shuffles - this is often a legacy of
7285 // XformToShuffleWithZero being used to combine bitmaskings (of
7286 // float vectors bitcast to integer vectors) into shuffles.
7287 // bitcast(shuffle(bitcast(s0),bitcast(s1))) -> shuffle(s0,s1)
7288 if (Level < AfterLegalizeDAG && TLI.isTypeLegal(VT) && VT.isVector() &&
7289 N0->getOpcode() == ISD::VECTOR_SHUFFLE &&
7290 VT.getVectorNumElements() >= N0.getValueType().getVectorNumElements() &&
7291 !(VT.getVectorNumElements() % N0.getValueType().getVectorNumElements())) {
7292 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N0);
7294 // If operands are a bitcast, peek through if it casts the original VT.
7295 // If operands are a constant, just bitcast back to original VT.
7296 auto PeekThroughBitcast = [&](SDValue Op) {
7297 if (Op.getOpcode() == ISD::BITCAST &&
7298 Op.getOperand(0).getValueType() == VT)
7299 return SDValue(Op.getOperand(0));
7300 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) ||
7301 ISD::isBuildVectorOfConstantFPSDNodes(Op.getNode()))
7302 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
7306 SDValue SV0 = PeekThroughBitcast(N0->getOperand(0));
7307 SDValue SV1 = PeekThroughBitcast(N0->getOperand(1));
7312 VT.getVectorNumElements() / N0.getValueType().getVectorNumElements();
7313 SmallVector<int, 8> NewMask;
7314 for (int M : SVN->getMask())
7315 for (int i = 0; i != MaskScale; ++i)
7316 NewMask.push_back(M < 0 ? -1 : M * MaskScale + i);
7318 bool LegalMask = TLI.isShuffleMaskLegal(NewMask, VT);
7320 std::swap(SV0, SV1);
7321 ShuffleVectorSDNode::commuteMask(NewMask);
7322 LegalMask = TLI.isShuffleMaskLegal(NewMask, VT);
7326 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, NewMask);
7332 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
7333 EVT VT = N->getValueType(0);
7334 return CombineConsecutiveLoads(N, VT);
7337 /// We know that BV is a build_vector node with Constant, ConstantFP or Undef
7338 /// operands. DstEltVT indicates the destination element value type.
7339 SDValue DAGCombiner::
7340 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
7341 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
7343 // If this is already the right type, we're done.
7344 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
7346 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
7347 unsigned DstBitSize = DstEltVT.getSizeInBits();
7349 // If this is a conversion of N elements of one type to N elements of another
7350 // type, convert each element. This handles FP<->INT cases.
7351 if (SrcBitSize == DstBitSize) {
7352 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
7353 BV->getValueType(0).getVectorNumElements());
7355 // Due to the FP element handling below calling this routine recursively,
7356 // we can end up with a scalar-to-vector node here.
7357 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
7358 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
7359 DAG.getNode(ISD::BITCAST, SDLoc(BV),
7360 DstEltVT, BV->getOperand(0)));
7362 SmallVector<SDValue, 8> Ops;
7363 for (SDValue Op : BV->op_values()) {
7364 // If the vector element type is not legal, the BUILD_VECTOR operands
7365 // are promoted and implicitly truncated. Make that explicit here.
7366 if (Op.getValueType() != SrcEltVT)
7367 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op);
7368 Ops.push_back(DAG.getNode(ISD::BITCAST, SDLoc(BV),
7370 AddToWorklist(Ops.back().getNode());
7372 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
7375 // Otherwise, we're growing or shrinking the elements. To avoid having to
7376 // handle annoying details of growing/shrinking FP values, we convert them to
7378 if (SrcEltVT.isFloatingPoint()) {
7379 // Convert the input float vector to a int vector where the elements are the
7381 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
7382 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
7386 // Now we know the input is an integer vector. If the output is a FP type,
7387 // convert to integer first, then to FP of the right size.
7388 if (DstEltVT.isFloatingPoint()) {
7389 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
7390 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
7392 // Next, convert to FP elements of the same size.
7393 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
7398 // Okay, we know the src/dst types are both integers of differing types.
7399 // Handling growing first.
7400 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
7401 if (SrcBitSize < DstBitSize) {
7402 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
7404 SmallVector<SDValue, 8> Ops;
7405 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
7406 i += NumInputsPerOutput) {
7407 bool isLE = DAG.getDataLayout().isLittleEndian();
7408 APInt NewBits = APInt(DstBitSize, 0);
7409 bool EltIsUndef = true;
7410 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
7411 // Shift the previously computed bits over.
7412 NewBits <<= SrcBitSize;
7413 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
7414 if (Op.getOpcode() == ISD::UNDEF) continue;
7417 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
7418 zextOrTrunc(SrcBitSize).zext(DstBitSize);
7422 Ops.push_back(DAG.getUNDEF(DstEltVT));
7424 Ops.push_back(DAG.getConstant(NewBits, DL, DstEltVT));
7427 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
7428 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Ops);
7431 // Finally, this must be the case where we are shrinking elements: each input
7432 // turns into multiple outputs.
7433 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
7434 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
7435 NumOutputsPerInput*BV->getNumOperands());
7436 SmallVector<SDValue, 8> Ops;
7438 for (const SDValue &Op : BV->op_values()) {
7439 if (Op.getOpcode() == ISD::UNDEF) {
7440 Ops.append(NumOutputsPerInput, DAG.getUNDEF(DstEltVT));
7444 APInt OpVal = cast<ConstantSDNode>(Op)->
7445 getAPIntValue().zextOrTrunc(SrcBitSize);
7447 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
7448 APInt ThisVal = OpVal.trunc(DstBitSize);
7449 Ops.push_back(DAG.getConstant(ThisVal, DL, DstEltVT));
7450 OpVal = OpVal.lshr(DstBitSize);
7453 // For big endian targets, swap the order of the pieces of each element.
7454 if (DAG.getDataLayout().isBigEndian())
7455 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
7458 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Ops);
7461 /// Try to perform FMA combining on a given FADD node.
7462 SDValue DAGCombiner::visitFADDForFMACombine(SDNode *N) {
7463 SDValue N0 = N->getOperand(0);
7464 SDValue N1 = N->getOperand(1);
7465 EVT VT = N->getValueType(0);
7468 const TargetOptions &Options = DAG.getTarget().Options;
7469 bool UnsafeFPMath = (Options.AllowFPOpFusion == FPOpFusion::Fast ||
7470 Options.UnsafeFPMath);
7472 // Floating-point multiply-add with intermediate rounding.
7473 bool HasFMAD = (LegalOperations &&
7474 TLI.isOperationLegal(ISD::FMAD, VT));
7476 // Floating-point multiply-add without intermediate rounding.
7477 bool HasFMA = ((!LegalOperations ||
7478 TLI.isOperationLegalOrCustom(ISD::FMA, VT)) &&
7479 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
7482 // No valid opcode, do not combine.
7483 if (!HasFMAD && !HasFMA)
7486 // Always prefer FMAD to FMA for precision.
7487 unsigned int PreferredFusedOpcode = HasFMAD ? ISD::FMAD : ISD::FMA;
7488 bool Aggressive = TLI.enableAggressiveFMAFusion(VT);
7489 bool LookThroughFPExt = TLI.isFPExtFree(VT);
7491 // If we have two choices trying to fold (fadd (fmul u, v), (fmul x, y)),
7492 // prefer to fold the multiply with fewer uses.
7493 if (Aggressive && N0.getOpcode() == ISD::FMUL &&
7494 N1.getOpcode() == ISD::FMUL) {
7495 if (N0.getNode()->use_size() > N1.getNode()->use_size())
7499 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
7500 if (N0.getOpcode() == ISD::FMUL &&
7501 (Aggressive || N0->hasOneUse())) {
7502 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7503 N0.getOperand(0), N0.getOperand(1), N1);
7506 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
7507 // Note: Commutes FADD operands.
7508 if (N1.getOpcode() == ISD::FMUL &&
7509 (Aggressive || N1->hasOneUse())) {
7510 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7511 N1.getOperand(0), N1.getOperand(1), N0);
7514 // Look through FP_EXTEND nodes to do more combining.
7515 if (UnsafeFPMath && LookThroughFPExt) {
7516 // fold (fadd (fpext (fmul x, y)), z) -> (fma (fpext x), (fpext y), z)
7517 if (N0.getOpcode() == ISD::FP_EXTEND) {
7518 SDValue N00 = N0.getOperand(0);
7519 if (N00.getOpcode() == ISD::FMUL)
7520 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7521 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7523 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7524 N00.getOperand(1)), N1);
7527 // fold (fadd x, (fpext (fmul y, z))) -> (fma (fpext y), (fpext z), x)
7528 // Note: Commutes FADD operands.
7529 if (N1.getOpcode() == ISD::FP_EXTEND) {
7530 SDValue N10 = N1.getOperand(0);
7531 if (N10.getOpcode() == ISD::FMUL)
7532 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7533 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7535 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7536 N10.getOperand(1)), N0);
7540 // More folding opportunities when target permits.
7541 if ((UnsafeFPMath || HasFMAD) && Aggressive) {
7542 // fold (fadd (fma x, y, (fmul u, v)), z) -> (fma x, y (fma u, v, z))
7543 if (N0.getOpcode() == PreferredFusedOpcode &&
7544 N0.getOperand(2).getOpcode() == ISD::FMUL) {
7545 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7546 N0.getOperand(0), N0.getOperand(1),
7547 DAG.getNode(PreferredFusedOpcode, SL, VT,
7548 N0.getOperand(2).getOperand(0),
7549 N0.getOperand(2).getOperand(1),
7553 // fold (fadd x, (fma y, z, (fmul u, v)) -> (fma y, z (fma u, v, x))
7554 if (N1->getOpcode() == PreferredFusedOpcode &&
7555 N1.getOperand(2).getOpcode() == ISD::FMUL) {
7556 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7557 N1.getOperand(0), N1.getOperand(1),
7558 DAG.getNode(PreferredFusedOpcode, SL, VT,
7559 N1.getOperand(2).getOperand(0),
7560 N1.getOperand(2).getOperand(1),
7564 if (UnsafeFPMath && LookThroughFPExt) {
7565 // fold (fadd (fma x, y, (fpext (fmul u, v))), z)
7566 // -> (fma x, y, (fma (fpext u), (fpext v), z))
7567 auto FoldFAddFMAFPExtFMul = [&] (
7568 SDValue X, SDValue Y, SDValue U, SDValue V, SDValue Z) {
7569 return DAG.getNode(PreferredFusedOpcode, SL, VT, X, Y,
7570 DAG.getNode(PreferredFusedOpcode, SL, VT,
7571 DAG.getNode(ISD::FP_EXTEND, SL, VT, U),
7572 DAG.getNode(ISD::FP_EXTEND, SL, VT, V),
7575 if (N0.getOpcode() == PreferredFusedOpcode) {
7576 SDValue N02 = N0.getOperand(2);
7577 if (N02.getOpcode() == ISD::FP_EXTEND) {
7578 SDValue N020 = N02.getOperand(0);
7579 if (N020.getOpcode() == ISD::FMUL)
7580 return FoldFAddFMAFPExtFMul(N0.getOperand(0), N0.getOperand(1),
7581 N020.getOperand(0), N020.getOperand(1),
7586 // fold (fadd (fpext (fma x, y, (fmul u, v))), z)
7587 // -> (fma (fpext x), (fpext y), (fma (fpext u), (fpext v), z))
7588 // FIXME: This turns two single-precision and one double-precision
7589 // operation into two double-precision operations, which might not be
7590 // interesting for all targets, especially GPUs.
7591 auto FoldFAddFPExtFMAFMul = [&] (
7592 SDValue X, SDValue Y, SDValue U, SDValue V, SDValue Z) {
7593 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7594 DAG.getNode(ISD::FP_EXTEND, SL, VT, X),
7595 DAG.getNode(ISD::FP_EXTEND, SL, VT, Y),
7596 DAG.getNode(PreferredFusedOpcode, SL, VT,
7597 DAG.getNode(ISD::FP_EXTEND, SL, VT, U),
7598 DAG.getNode(ISD::FP_EXTEND, SL, VT, V),
7601 if (N0.getOpcode() == ISD::FP_EXTEND) {
7602 SDValue N00 = N0.getOperand(0);
7603 if (N00.getOpcode() == PreferredFusedOpcode) {
7604 SDValue N002 = N00.getOperand(2);
7605 if (N002.getOpcode() == ISD::FMUL)
7606 return FoldFAddFPExtFMAFMul(N00.getOperand(0), N00.getOperand(1),
7607 N002.getOperand(0), N002.getOperand(1),
7612 // fold (fadd x, (fma y, z, (fpext (fmul u, v)))
7613 // -> (fma y, z, (fma (fpext u), (fpext v), x))
7614 if (N1.getOpcode() == PreferredFusedOpcode) {
7615 SDValue N12 = N1.getOperand(2);
7616 if (N12.getOpcode() == ISD::FP_EXTEND) {
7617 SDValue N120 = N12.getOperand(0);
7618 if (N120.getOpcode() == ISD::FMUL)
7619 return FoldFAddFMAFPExtFMul(N1.getOperand(0), N1.getOperand(1),
7620 N120.getOperand(0), N120.getOperand(1),
7625 // fold (fadd x, (fpext (fma y, z, (fmul u, v)))
7626 // -> (fma (fpext y), (fpext z), (fma (fpext u), (fpext v), x))
7627 // FIXME: This turns two single-precision and one double-precision
7628 // operation into two double-precision operations, which might not be
7629 // interesting for all targets, especially GPUs.
7630 if (N1.getOpcode() == ISD::FP_EXTEND) {
7631 SDValue N10 = N1.getOperand(0);
7632 if (N10.getOpcode() == PreferredFusedOpcode) {
7633 SDValue N102 = N10.getOperand(2);
7634 if (N102.getOpcode() == ISD::FMUL)
7635 return FoldFAddFPExtFMAFMul(N10.getOperand(0), N10.getOperand(1),
7636 N102.getOperand(0), N102.getOperand(1),
7646 /// Try to perform FMA combining on a given FSUB node.
7647 SDValue DAGCombiner::visitFSUBForFMACombine(SDNode *N) {
7648 SDValue N0 = N->getOperand(0);
7649 SDValue N1 = N->getOperand(1);
7650 EVT VT = N->getValueType(0);
7653 const TargetOptions &Options = DAG.getTarget().Options;
7654 bool UnsafeFPMath = (Options.AllowFPOpFusion == FPOpFusion::Fast ||
7655 Options.UnsafeFPMath);
7657 // Floating-point multiply-add with intermediate rounding.
7658 bool HasFMAD = (LegalOperations &&
7659 TLI.isOperationLegal(ISD::FMAD, VT));
7661 // Floating-point multiply-add without intermediate rounding.
7662 bool HasFMA = ((!LegalOperations ||
7663 TLI.isOperationLegalOrCustom(ISD::FMA, VT)) &&
7664 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
7667 // No valid opcode, do not combine.
7668 if (!HasFMAD && !HasFMA)
7671 // Always prefer FMAD to FMA for precision.
7672 unsigned int PreferredFusedOpcode = HasFMAD ? ISD::FMAD : ISD::FMA;
7673 bool Aggressive = TLI.enableAggressiveFMAFusion(VT);
7674 bool LookThroughFPExt = TLI.isFPExtFree(VT);
7676 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
7677 if (N0.getOpcode() == ISD::FMUL &&
7678 (Aggressive || N0->hasOneUse())) {
7679 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7680 N0.getOperand(0), N0.getOperand(1),
7681 DAG.getNode(ISD::FNEG, SL, VT, N1));
7684 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
7685 // Note: Commutes FSUB operands.
7686 if (N1.getOpcode() == ISD::FMUL &&
7687 (Aggressive || N1->hasOneUse()))
7688 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7689 DAG.getNode(ISD::FNEG, SL, VT,
7691 N1.getOperand(1), N0);
7693 // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
7694 if (N0.getOpcode() == ISD::FNEG &&
7695 N0.getOperand(0).getOpcode() == ISD::FMUL &&
7696 (Aggressive || (N0->hasOneUse() && N0.getOperand(0).hasOneUse()))) {
7697 SDValue N00 = N0.getOperand(0).getOperand(0);
7698 SDValue N01 = N0.getOperand(0).getOperand(1);
7699 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7700 DAG.getNode(ISD::FNEG, SL, VT, N00), N01,
7701 DAG.getNode(ISD::FNEG, SL, VT, N1));
7704 // Look through FP_EXTEND nodes to do more combining.
7705 if (UnsafeFPMath && LookThroughFPExt) {
7706 // fold (fsub (fpext (fmul x, y)), z)
7707 // -> (fma (fpext x), (fpext y), (fneg z))
7708 if (N0.getOpcode() == ISD::FP_EXTEND) {
7709 SDValue N00 = N0.getOperand(0);
7710 if (N00.getOpcode() == ISD::FMUL)
7711 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7712 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7714 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7716 DAG.getNode(ISD::FNEG, SL, VT, N1));
7719 // fold (fsub x, (fpext (fmul y, z)))
7720 // -> (fma (fneg (fpext y)), (fpext z), x)
7721 // Note: Commutes FSUB operands.
7722 if (N1.getOpcode() == ISD::FP_EXTEND) {
7723 SDValue N10 = N1.getOperand(0);
7724 if (N10.getOpcode() == ISD::FMUL)
7725 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7726 DAG.getNode(ISD::FNEG, SL, VT,
7727 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7728 N10.getOperand(0))),
7729 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7734 // fold (fsub (fpext (fneg (fmul, x, y))), z)
7735 // -> (fneg (fma (fpext x), (fpext y), z))
7736 // Note: This could be removed with appropriate canonicalization of the
7737 // input expression into (fneg (fadd (fpext (fmul, x, y)), z). However, the
7738 // orthogonal flags -fp-contract=fast and -enable-unsafe-fp-math prevent
7739 // from implementing the canonicalization in visitFSUB.
7740 if (N0.getOpcode() == ISD::FP_EXTEND) {
7741 SDValue N00 = N0.getOperand(0);
7742 if (N00.getOpcode() == ISD::FNEG) {
7743 SDValue N000 = N00.getOperand(0);
7744 if (N000.getOpcode() == ISD::FMUL) {
7745 return DAG.getNode(ISD::FNEG, SL, VT,
7746 DAG.getNode(PreferredFusedOpcode, SL, VT,
7747 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7748 N000.getOperand(0)),
7749 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7750 N000.getOperand(1)),
7756 // fold (fsub (fneg (fpext (fmul, x, y))), z)
7757 // -> (fneg (fma (fpext x)), (fpext y), z)
7758 // Note: This could be removed with appropriate canonicalization of the
7759 // input expression into (fneg (fadd (fpext (fmul, x, y)), z). However, the
7760 // orthogonal flags -fp-contract=fast and -enable-unsafe-fp-math prevent
7761 // from implementing the canonicalization in visitFSUB.
7762 if (N0.getOpcode() == ISD::FNEG) {
7763 SDValue N00 = N0.getOperand(0);
7764 if (N00.getOpcode() == ISD::FP_EXTEND) {
7765 SDValue N000 = N00.getOperand(0);
7766 if (N000.getOpcode() == ISD::FMUL) {
7767 return DAG.getNode(ISD::FNEG, SL, VT,
7768 DAG.getNode(PreferredFusedOpcode, SL, VT,
7769 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7770 N000.getOperand(0)),
7771 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7772 N000.getOperand(1)),
7780 // More folding opportunities when target permits.
7781 if ((UnsafeFPMath || HasFMAD) && Aggressive) {
7782 // fold (fsub (fma x, y, (fmul u, v)), z)
7783 // -> (fma x, y (fma u, v, (fneg z)))
7784 if (N0.getOpcode() == PreferredFusedOpcode &&
7785 N0.getOperand(2).getOpcode() == ISD::FMUL) {
7786 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7787 N0.getOperand(0), N0.getOperand(1),
7788 DAG.getNode(PreferredFusedOpcode, SL, VT,
7789 N0.getOperand(2).getOperand(0),
7790 N0.getOperand(2).getOperand(1),
7791 DAG.getNode(ISD::FNEG, SL, VT,
7795 // fold (fsub x, (fma y, z, (fmul u, v)))
7796 // -> (fma (fneg y), z, (fma (fneg u), v, x))
7797 if (N1.getOpcode() == PreferredFusedOpcode &&
7798 N1.getOperand(2).getOpcode() == ISD::FMUL) {
7799 SDValue N20 = N1.getOperand(2).getOperand(0);
7800 SDValue N21 = N1.getOperand(2).getOperand(1);
7801 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7802 DAG.getNode(ISD::FNEG, SL, VT,
7805 DAG.getNode(PreferredFusedOpcode, SL, VT,
7806 DAG.getNode(ISD::FNEG, SL, VT, N20),
7811 if (UnsafeFPMath && LookThroughFPExt) {
7812 // fold (fsub (fma x, y, (fpext (fmul u, v))), z)
7813 // -> (fma x, y (fma (fpext u), (fpext v), (fneg z)))
7814 if (N0.getOpcode() == PreferredFusedOpcode) {
7815 SDValue N02 = N0.getOperand(2);
7816 if (N02.getOpcode() == ISD::FP_EXTEND) {
7817 SDValue N020 = N02.getOperand(0);
7818 if (N020.getOpcode() == ISD::FMUL)
7819 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7820 N0.getOperand(0), N0.getOperand(1),
7821 DAG.getNode(PreferredFusedOpcode, SL, VT,
7822 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7823 N020.getOperand(0)),
7824 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7825 N020.getOperand(1)),
7826 DAG.getNode(ISD::FNEG, SL, VT,
7831 // fold (fsub (fpext (fma x, y, (fmul u, v))), z)
7832 // -> (fma (fpext x), (fpext y),
7833 // (fma (fpext u), (fpext v), (fneg z)))
7834 // FIXME: This turns two single-precision and one double-precision
7835 // operation into two double-precision operations, which might not be
7836 // interesting for all targets, especially GPUs.
7837 if (N0.getOpcode() == ISD::FP_EXTEND) {
7838 SDValue N00 = N0.getOperand(0);
7839 if (N00.getOpcode() == PreferredFusedOpcode) {
7840 SDValue N002 = N00.getOperand(2);
7841 if (N002.getOpcode() == ISD::FMUL)
7842 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7843 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7845 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7847 DAG.getNode(PreferredFusedOpcode, SL, VT,
7848 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7849 N002.getOperand(0)),
7850 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7851 N002.getOperand(1)),
7852 DAG.getNode(ISD::FNEG, SL, VT,
7857 // fold (fsub x, (fma y, z, (fpext (fmul u, v))))
7858 // -> (fma (fneg y), z, (fma (fneg (fpext u)), (fpext v), x))
7859 if (N1.getOpcode() == PreferredFusedOpcode &&
7860 N1.getOperand(2).getOpcode() == ISD::FP_EXTEND) {
7861 SDValue N120 = N1.getOperand(2).getOperand(0);
7862 if (N120.getOpcode() == ISD::FMUL) {
7863 SDValue N1200 = N120.getOperand(0);
7864 SDValue N1201 = N120.getOperand(1);
7865 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7866 DAG.getNode(ISD::FNEG, SL, VT, N1.getOperand(0)),
7868 DAG.getNode(PreferredFusedOpcode, SL, VT,
7869 DAG.getNode(ISD::FNEG, SL, VT,
7870 DAG.getNode(ISD::FP_EXTEND, SL,
7872 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7878 // fold (fsub x, (fpext (fma y, z, (fmul u, v))))
7879 // -> (fma (fneg (fpext y)), (fpext z),
7880 // (fma (fneg (fpext u)), (fpext v), x))
7881 // FIXME: This turns two single-precision and one double-precision
7882 // operation into two double-precision operations, which might not be
7883 // interesting for all targets, especially GPUs.
7884 if (N1.getOpcode() == ISD::FP_EXTEND &&
7885 N1.getOperand(0).getOpcode() == PreferredFusedOpcode) {
7886 SDValue N100 = N1.getOperand(0).getOperand(0);
7887 SDValue N101 = N1.getOperand(0).getOperand(1);
7888 SDValue N102 = N1.getOperand(0).getOperand(2);
7889 if (N102.getOpcode() == ISD::FMUL) {
7890 SDValue N1020 = N102.getOperand(0);
7891 SDValue N1021 = N102.getOperand(1);
7892 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7893 DAG.getNode(ISD::FNEG, SL, VT,
7894 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7896 DAG.getNode(ISD::FP_EXTEND, SL, VT, N101),
7897 DAG.getNode(PreferredFusedOpcode, SL, VT,
7898 DAG.getNode(ISD::FNEG, SL, VT,
7899 DAG.getNode(ISD::FP_EXTEND, SL,
7901 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7912 SDValue DAGCombiner::visitFADD(SDNode *N) {
7913 SDValue N0 = N->getOperand(0);
7914 SDValue N1 = N->getOperand(1);
7915 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7916 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7917 EVT VT = N->getValueType(0);
7919 const TargetOptions &Options = DAG.getTarget().Options;
7923 if (SDValue FoldedVOp = SimplifyVBinOp(N))
7926 // fold (fadd c1, c2) -> c1 + c2
7928 return DAG.getNode(ISD::FADD, DL, VT, N0, N1);
7930 // canonicalize constant to RHS
7931 if (N0CFP && !N1CFP)
7932 return DAG.getNode(ISD::FADD, DL, VT, N1, N0);
7934 // fold (fadd A, (fneg B)) -> (fsub A, B)
7935 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
7936 isNegatibleForFree(N1, LegalOperations, TLI, &Options) == 2)
7937 return DAG.getNode(ISD::FSUB, DL, VT, N0,
7938 GetNegatedExpression(N1, DAG, LegalOperations));
7940 // fold (fadd (fneg A), B) -> (fsub B, A)
7941 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
7942 isNegatibleForFree(N0, LegalOperations, TLI, &Options) == 2)
7943 return DAG.getNode(ISD::FSUB, DL, VT, N1,
7944 GetNegatedExpression(N0, DAG, LegalOperations));
7946 // If 'unsafe math' is enabled, fold lots of things.
7947 if (Options.UnsafeFPMath) {
7948 // No FP constant should be created after legalization as Instruction
7949 // Selection pass has a hard time dealing with FP constants.
7950 bool AllowNewConst = (Level < AfterLegalizeDAG);
7952 // fold (fadd A, 0) -> A
7953 if (N1CFP && N1CFP->isZero())
7956 // fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
7957 if (N1CFP && N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
7958 isa<ConstantFPSDNode>(N0.getOperand(1)))
7959 return DAG.getNode(ISD::FADD, DL, VT, N0.getOperand(0),
7960 DAG.getNode(ISD::FADD, DL, VT, N0.getOperand(1), N1));
7962 // If allowed, fold (fadd (fneg x), x) -> 0.0
7963 if (AllowNewConst && N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1)
7964 return DAG.getConstantFP(0.0, DL, VT);
7966 // If allowed, fold (fadd x, (fneg x)) -> 0.0
7967 if (AllowNewConst && N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0)
7968 return DAG.getConstantFP(0.0, DL, VT);
7970 // We can fold chains of FADD's of the same value into multiplications.
7971 // This transform is not safe in general because we are reducing the number
7972 // of rounding steps.
7973 if (TLI.isOperationLegalOrCustom(ISD::FMUL, VT) && !N0CFP && !N1CFP) {
7974 if (N0.getOpcode() == ISD::FMUL) {
7975 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
7976 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
7978 // (fadd (fmul x, c), x) -> (fmul x, c+1)
7979 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
7980 SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, SDValue(CFP01, 0),
7981 DAG.getConstantFP(1.0, DL, VT));
7982 return DAG.getNode(ISD::FMUL, DL, VT, N1, NewCFP);
7985 // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2)
7986 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
7987 N1.getOperand(0) == N1.getOperand(1) &&
7988 N0.getOperand(0) == N1.getOperand(0)) {
7989 SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, SDValue(CFP01, 0),
7990 DAG.getConstantFP(2.0, DL, VT));
7991 return DAG.getNode(ISD::FMUL, DL, VT, N0.getOperand(0), NewCFP);
7995 if (N1.getOpcode() == ISD::FMUL) {
7996 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
7997 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
7999 // (fadd x, (fmul x, c)) -> (fmul x, c+1)
8000 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
8001 SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, SDValue(CFP11, 0),
8002 DAG.getConstantFP(1.0, DL, VT));
8003 return DAG.getNode(ISD::FMUL, DL, VT, N0, NewCFP);
8006 // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2)
8007 if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD &&
8008 N0.getOperand(0) == N0.getOperand(1) &&
8009 N1.getOperand(0) == N0.getOperand(0)) {
8010 SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, SDValue(CFP11, 0),
8011 DAG.getConstantFP(2.0, DL, VT));
8012 return DAG.getNode(ISD::FMUL, DL, VT, N1.getOperand(0), NewCFP);
8016 if (N0.getOpcode() == ISD::FADD && AllowNewConst) {
8017 ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
8018 // (fadd (fadd x, x), x) -> (fmul x, 3.0)
8019 if (!CFP && N0.getOperand(0) == N0.getOperand(1) &&
8020 (N0.getOperand(0) == N1)) {
8021 return DAG.getNode(ISD::FMUL, DL, VT,
8022 N1, DAG.getConstantFP(3.0, DL, VT));
8026 if (N1.getOpcode() == ISD::FADD && AllowNewConst) {
8027 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
8028 // (fadd x, (fadd x, x)) -> (fmul x, 3.0)
8029 if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) &&
8030 N1.getOperand(0) == N0) {
8031 return DAG.getNode(ISD::FMUL, DL, VT,
8032 N0, DAG.getConstantFP(3.0, DL, VT));
8036 // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0)
8037 if (AllowNewConst &&
8038 N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
8039 N0.getOperand(0) == N0.getOperand(1) &&
8040 N1.getOperand(0) == N1.getOperand(1) &&
8041 N0.getOperand(0) == N1.getOperand(0)) {
8042 return DAG.getNode(ISD::FMUL, DL, VT,
8043 N0.getOperand(0), DAG.getConstantFP(4.0, DL, VT));
8046 } // enable-unsafe-fp-math
8048 // FADD -> FMA combines:
8049 if (SDValue Fused = visitFADDForFMACombine(N)) {
8050 AddToWorklist(Fused.getNode());
8057 SDValue DAGCombiner::visitFSUB(SDNode *N) {
8058 SDValue N0 = N->getOperand(0);
8059 SDValue N1 = N->getOperand(1);
8060 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
8061 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
8062 EVT VT = N->getValueType(0);
8064 const TargetOptions &Options = DAG.getTarget().Options;
8068 if (SDValue FoldedVOp = SimplifyVBinOp(N))
8071 // fold (fsub c1, c2) -> c1-c2
8073 return DAG.getNode(ISD::FSUB, dl, VT, N0, N1);
8075 // fold (fsub A, (fneg B)) -> (fadd A, B)
8076 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
8077 return DAG.getNode(ISD::FADD, dl, VT, N0,
8078 GetNegatedExpression(N1, DAG, LegalOperations));
8080 // If 'unsafe math' is enabled, fold lots of things.
8081 if (Options.UnsafeFPMath) {
8083 if (N1CFP && N1CFP->isZero())
8086 // (fsub 0, B) -> -B
8087 if (N0CFP && N0CFP->isZero()) {
8088 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
8089 return GetNegatedExpression(N1, DAG, LegalOperations);
8090 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
8091 return DAG.getNode(ISD::FNEG, dl, VT, N1);
8094 // (fsub x, x) -> 0.0
8096 return DAG.getConstantFP(0.0f, dl, VT);
8098 // (fsub x, (fadd x, y)) -> (fneg y)
8099 // (fsub x, (fadd y, x)) -> (fneg y)
8100 if (N1.getOpcode() == ISD::FADD) {
8101 SDValue N10 = N1->getOperand(0);
8102 SDValue N11 = N1->getOperand(1);
8104 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI, &Options))
8105 return GetNegatedExpression(N11, DAG, LegalOperations);
8107 if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI, &Options))
8108 return GetNegatedExpression(N10, DAG, LegalOperations);
8112 // FSUB -> FMA combines:
8113 if (SDValue Fused = visitFSUBForFMACombine(N)) {
8114 AddToWorklist(Fused.getNode());
8121 SDValue DAGCombiner::visitFMUL(SDNode *N) {
8122 SDValue N0 = N->getOperand(0);
8123 SDValue N1 = N->getOperand(1);
8124 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
8125 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
8126 EVT VT = N->getValueType(0);
8128 const TargetOptions &Options = DAG.getTarget().Options;
8131 if (VT.isVector()) {
8132 // This just handles C1 * C2 for vectors. Other vector folds are below.
8133 if (SDValue FoldedVOp = SimplifyVBinOp(N))
8137 // fold (fmul c1, c2) -> c1*c2
8139 return DAG.getNode(ISD::FMUL, DL, VT, N0, N1);
8141 // canonicalize constant to RHS
8142 if (isConstantFPBuildVectorOrConstantFP(N0) &&
8143 !isConstantFPBuildVectorOrConstantFP(N1))
8144 return DAG.getNode(ISD::FMUL, DL, VT, N1, N0);
8146 // fold (fmul A, 1.0) -> A
8147 if (N1CFP && N1CFP->isExactlyValue(1.0))
8150 if (Options.UnsafeFPMath) {
8151 // fold (fmul A, 0) -> 0
8152 if (N1CFP && N1CFP->isZero())
8155 // fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
8156 if (N0.getOpcode() == ISD::FMUL) {
8157 // Fold scalars or any vector constants (not just splats).
8158 // This fold is done in general by InstCombine, but extra fmul insts
8159 // may have been generated during lowering.
8160 SDValue N00 = N0.getOperand(0);
8161 SDValue N01 = N0.getOperand(1);
8162 auto *BV1 = dyn_cast<BuildVectorSDNode>(N1);
8163 auto *BV00 = dyn_cast<BuildVectorSDNode>(N00);
8164 auto *BV01 = dyn_cast<BuildVectorSDNode>(N01);
8166 // Check 1: Make sure that the first operand of the inner multiply is NOT
8167 // a constant. Otherwise, we may induce infinite looping.
8168 if (!(isConstOrConstSplatFP(N00) || (BV00 && BV00->isConstant()))) {
8169 // Check 2: Make sure that the second operand of the inner multiply and
8170 // the second operand of the outer multiply are constants.
8171 if ((N1CFP && isConstOrConstSplatFP(N01)) ||
8172 (BV1 && BV01 && BV1->isConstant() && BV01->isConstant())) {
8173 SDValue MulConsts = DAG.getNode(ISD::FMUL, DL, VT, N01, N1);
8174 return DAG.getNode(ISD::FMUL, DL, VT, N00, MulConsts);
8179 // fold (fmul (fadd x, x), c) -> (fmul x, (fmul 2.0, c))
8180 // Undo the fmul 2.0, x -> fadd x, x transformation, since if it occurs
8181 // during an early run of DAGCombiner can prevent folding with fmuls
8182 // inserted during lowering.
8183 if (N0.getOpcode() == ISD::FADD &&
8184 (N0.getOperand(0) == N0.getOperand(1)) &&
8186 const SDValue Two = DAG.getConstantFP(2.0, DL, VT);
8187 SDValue MulConsts = DAG.getNode(ISD::FMUL, DL, VT, Two, N1);
8188 return DAG.getNode(ISD::FMUL, DL, VT, N0.getOperand(0), MulConsts);
8192 // fold (fmul X, 2.0) -> (fadd X, X)
8193 if (N1CFP && N1CFP->isExactlyValue(+2.0))
8194 return DAG.getNode(ISD::FADD, DL, VT, N0, N0);
8196 // fold (fmul X, -1.0) -> (fneg X)
8197 if (N1CFP && N1CFP->isExactlyValue(-1.0))
8198 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
8199 return DAG.getNode(ISD::FNEG, DL, VT, N0);
8201 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
8202 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
8203 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
8204 // Both can be negated for free, check to see if at least one is cheaper
8206 if (LHSNeg == 2 || RHSNeg == 2)
8207 return DAG.getNode(ISD::FMUL, DL, VT,
8208 GetNegatedExpression(N0, DAG, LegalOperations),
8209 GetNegatedExpression(N1, DAG, LegalOperations));
8216 SDValue DAGCombiner::visitFMA(SDNode *N) {
8217 SDValue N0 = N->getOperand(0);
8218 SDValue N1 = N->getOperand(1);
8219 SDValue N2 = N->getOperand(2);
8220 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8221 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8222 EVT VT = N->getValueType(0);
8224 const TargetOptions &Options = DAG.getTarget().Options;
8226 // Constant fold FMA.
8227 if (isa<ConstantFPSDNode>(N0) &&
8228 isa<ConstantFPSDNode>(N1) &&
8229 isa<ConstantFPSDNode>(N2)) {
8230 return DAG.getNode(ISD::FMA, dl, VT, N0, N1, N2);
8233 if (Options.UnsafeFPMath) {
8234 if (N0CFP && N0CFP->isZero())
8236 if (N1CFP && N1CFP->isZero())
8239 if (N0CFP && N0CFP->isExactlyValue(1.0))
8240 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2);
8241 if (N1CFP && N1CFP->isExactlyValue(1.0))
8242 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2);
8244 // Canonicalize (fma c, x, y) -> (fma x, c, y)
8245 if (N0CFP && !N1CFP)
8246 return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2);
8248 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
8249 if (Options.UnsafeFPMath && N1CFP &&
8250 N2.getOpcode() == ISD::FMUL &&
8251 N0 == N2.getOperand(0) &&
8252 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
8253 return DAG.getNode(ISD::FMUL, dl, VT, N0,
8254 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
8258 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
8259 if (Options.UnsafeFPMath &&
8260 N0.getOpcode() == ISD::FMUL && N1CFP &&
8261 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
8262 return DAG.getNode(ISD::FMA, dl, VT,
8264 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
8268 // (fma x, 1, y) -> (fadd x, y)
8269 // (fma x, -1, y) -> (fadd (fneg x), y)
8271 if (N1CFP->isExactlyValue(1.0))
8272 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
8274 if (N1CFP->isExactlyValue(-1.0) &&
8275 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
8276 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
8277 AddToWorklist(RHSNeg.getNode());
8278 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
8282 // (fma x, c, x) -> (fmul x, (c+1))
8283 if (Options.UnsafeFPMath && N1CFP && N0 == N2)
8284 return DAG.getNode(ISD::FMUL, dl, VT, N0,
8285 DAG.getNode(ISD::FADD, dl, VT,
8286 N1, DAG.getConstantFP(1.0, dl, VT)));
8288 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
8289 if (Options.UnsafeFPMath && N1CFP &&
8290 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0)
8291 return DAG.getNode(ISD::FMUL, dl, VT, N0,
8292 DAG.getNode(ISD::FADD, dl, VT,
8293 N1, DAG.getConstantFP(-1.0, dl, VT)));
8299 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
8301 // E.g., (a / D; b / D;) -> (recip = 1.0 / D; a * recip; b * recip)
8302 // Notice that this is not always beneficial. One reason is different target
8303 // may have different costs for FDIV and FMUL, so sometimes the cost of two
8304 // FDIVs may be lower than the cost of one FDIV and two FMULs. Another reason
8305 // is the critical path is increased from "one FDIV" to "one FDIV + one FMUL".
8306 SDValue DAGCombiner::combineRepeatedFPDivisors(SDNode *N) {
8307 if (!DAG.getTarget().Options.UnsafeFPMath)
8310 // Skip if current node is a reciprocal.
8311 SDValue N0 = N->getOperand(0);
8312 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8313 if (N0CFP && N0CFP->isExactlyValue(1.0))
8316 // Exit early if the target does not want this transform or if there can't
8317 // possibly be enough uses of the divisor to make the transform worthwhile.
8318 SDValue N1 = N->getOperand(1);
8319 unsigned MinUses = TLI.combineRepeatedFPDivisors();
8320 if (!MinUses || N1->use_size() < MinUses)
8323 // Find all FDIV users of the same divisor.
8324 // Use a set because duplicates may be present in the user list.
8325 SetVector<SDNode *> Users;
8326 for (auto *U : N1->uses())
8327 if (U->getOpcode() == ISD::FDIV && U->getOperand(1) == N1)
8330 // Now that we have the actual number of divisor uses, make sure it meets
8331 // the minimum threshold specified by the target.
8332 if (Users.size() < MinUses)
8335 EVT VT = N->getValueType(0);
8337 SDValue FPOne = DAG.getConstantFP(1.0, DL, VT);
8338 // FIXME: This optimization requires some level of fast-math, so the
8339 // created reciprocal node should at least have the 'allowReciprocal'
8340 // fast-math-flag set.
8341 SDValue Reciprocal = DAG.getNode(ISD::FDIV, DL, VT, FPOne, N1);
8343 // Dividend / Divisor -> Dividend * Reciprocal
8344 for (auto *U : Users) {
8345 SDValue Dividend = U->getOperand(0);
8346 if (Dividend != FPOne) {
8347 SDValue NewNode = DAG.getNode(ISD::FMUL, SDLoc(U), VT, Dividend,
8349 CombineTo(U, NewNode);
8350 } else if (U != Reciprocal.getNode()) {
8351 // In the absence of fast-math-flags, this user node is always the
8352 // same node as Reciprocal, but with FMF they may be different nodes.
8353 CombineTo(U, Reciprocal);
8356 return SDValue(N, 0); // N was replaced.
8359 SDValue DAGCombiner::visitFDIV(SDNode *N) {
8360 SDValue N0 = N->getOperand(0);
8361 SDValue N1 = N->getOperand(1);
8362 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8363 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8364 EVT VT = N->getValueType(0);
8366 const TargetOptions &Options = DAG.getTarget().Options;
8370 if (SDValue FoldedVOp = SimplifyVBinOp(N))
8373 // fold (fdiv c1, c2) -> c1/c2
8375 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1);
8377 if (Options.UnsafeFPMath) {
8378 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
8380 // Compute the reciprocal 1.0 / c2.
8381 APFloat N1APF = N1CFP->getValueAPF();
8382 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
8383 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
8384 // Only do the transform if the reciprocal is a legal fp immediate that
8385 // isn't too nasty (eg NaN, denormal, ...).
8386 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
8387 (!LegalOperations ||
8388 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
8389 // backend)... we should handle this gracefully after Legalize.
8390 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
8391 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
8392 TLI.isFPImmLegal(Recip, VT)))
8393 return DAG.getNode(ISD::FMUL, DL, VT, N0,
8394 DAG.getConstantFP(Recip, DL, VT));
8397 // If this FDIV is part of a reciprocal square root, it may be folded
8398 // into a target-specific square root estimate instruction.
8399 if (N1.getOpcode() == ISD::FSQRT) {
8400 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0))) {
8401 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
8403 } else if (N1.getOpcode() == ISD::FP_EXTEND &&
8404 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
8405 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0).getOperand(0))) {
8406 RV = DAG.getNode(ISD::FP_EXTEND, SDLoc(N1), VT, RV);
8407 AddToWorklist(RV.getNode());
8408 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
8410 } else if (N1.getOpcode() == ISD::FP_ROUND &&
8411 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
8412 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0).getOperand(0))) {
8413 RV = DAG.getNode(ISD::FP_ROUND, SDLoc(N1), VT, RV, N1.getOperand(1));
8414 AddToWorklist(RV.getNode());
8415 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
8417 } else if (N1.getOpcode() == ISD::FMUL) {
8418 // Look through an FMUL. Even though this won't remove the FDIV directly,
8419 // it's still worthwhile to get rid of the FSQRT if possible.
8422 if (N1.getOperand(0).getOpcode() == ISD::FSQRT) {
8423 SqrtOp = N1.getOperand(0);
8424 OtherOp = N1.getOperand(1);
8425 } else if (N1.getOperand(1).getOpcode() == ISD::FSQRT) {
8426 SqrtOp = N1.getOperand(1);
8427 OtherOp = N1.getOperand(0);
8429 if (SqrtOp.getNode()) {
8430 // We found a FSQRT, so try to make this fold:
8431 // x / (y * sqrt(z)) -> x * (rsqrt(z) / y)
8432 if (SDValue RV = BuildRsqrtEstimate(SqrtOp.getOperand(0))) {
8433 RV = DAG.getNode(ISD::FDIV, SDLoc(N1), VT, RV, OtherOp);
8434 AddToWorklist(RV.getNode());
8435 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
8440 // Fold into a reciprocal estimate and multiply instead of a real divide.
8441 if (SDValue RV = BuildReciprocalEstimate(N1)) {
8442 AddToWorklist(RV.getNode());
8443 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
8447 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
8448 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
8449 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
8450 // Both can be negated for free, check to see if at least one is cheaper
8452 if (LHSNeg == 2 || RHSNeg == 2)
8453 return DAG.getNode(ISD::FDIV, SDLoc(N), VT,
8454 GetNegatedExpression(N0, DAG, LegalOperations),
8455 GetNegatedExpression(N1, DAG, LegalOperations));
8459 if (SDValue CombineRepeatedDivisors = combineRepeatedFPDivisors(N))
8460 return CombineRepeatedDivisors;
8465 SDValue DAGCombiner::visitFREM(SDNode *N) {
8466 SDValue N0 = N->getOperand(0);
8467 SDValue N1 = N->getOperand(1);
8468 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8469 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8470 EVT VT = N->getValueType(0);
8472 // fold (frem c1, c2) -> fmod(c1,c2)
8474 return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1);
8479 SDValue DAGCombiner::visitFSQRT(SDNode *N) {
8480 if (!DAG.getTarget().Options.UnsafeFPMath || TLI.isFsqrtCheap())
8483 // Compute this as X * (1/sqrt(X)) = X * (X ** -0.5)
8484 SDValue RV = BuildRsqrtEstimate(N->getOperand(0));
8488 EVT VT = RV.getValueType();
8490 RV = DAG.getNode(ISD::FMUL, DL, VT, N->getOperand(0), RV);
8491 AddToWorklist(RV.getNode());
8493 // Unfortunately, RV is now NaN if the input was exactly 0.
8494 // Select out this case and force the answer to 0.
8495 SDValue Zero = DAG.getConstantFP(0.0, DL, VT);
8496 EVT CCVT = TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
8497 SDValue ZeroCmp = DAG.getSetCC(DL, CCVT, N->getOperand(0), Zero, ISD::SETEQ);
8498 AddToWorklist(ZeroCmp.getNode());
8499 AddToWorklist(RV.getNode());
8501 return DAG.getNode(VT.isVector() ? ISD::VSELECT : ISD::SELECT, DL, VT,
8505 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
8506 SDValue N0 = N->getOperand(0);
8507 SDValue N1 = N->getOperand(1);
8508 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8509 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8510 EVT VT = N->getValueType(0);
8512 if (N0CFP && N1CFP) // Constant fold
8513 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1);
8516 const APFloat& V = N1CFP->getValueAPF();
8517 // copysign(x, c1) -> fabs(x) iff ispos(c1)
8518 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
8519 if (!V.isNegative()) {
8520 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
8521 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
8523 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
8524 return DAG.getNode(ISD::FNEG, SDLoc(N), VT,
8525 DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0));
8529 // copysign(fabs(x), y) -> copysign(x, y)
8530 // copysign(fneg(x), y) -> copysign(x, y)
8531 // copysign(copysign(x,z), y) -> copysign(x, y)
8532 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
8533 N0.getOpcode() == ISD::FCOPYSIGN)
8534 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
8535 N0.getOperand(0), N1);
8537 // copysign(x, abs(y)) -> abs(x)
8538 if (N1.getOpcode() == ISD::FABS)
8539 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
8541 // copysign(x, copysign(y,z)) -> copysign(x, z)
8542 if (N1.getOpcode() == ISD::FCOPYSIGN)
8543 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
8544 N0, N1.getOperand(1));
8546 // copysign(x, fp_extend(y)) -> copysign(x, y)
8547 // copysign(x, fp_round(y)) -> copysign(x, y)
8548 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
8549 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
8550 N0, N1.getOperand(0));
8555 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
8556 SDValue N0 = N->getOperand(0);
8557 EVT VT = N->getValueType(0);
8558 EVT OpVT = N0.getValueType();
8560 // fold (sint_to_fp c1) -> c1fp
8561 if (isConstantIntBuildVectorOrConstantInt(N0) &&
8562 // ...but only if the target supports immediate floating-point values
8563 (!LegalOperations ||
8564 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
8565 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
8567 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
8568 // but UINT_TO_FP is legal on this target, try to convert.
8569 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
8570 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
8571 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
8572 if (DAG.SignBitIsZero(N0))
8573 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
8576 // The next optimizations are desirable only if SELECT_CC can be lowered.
8577 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
8578 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
8579 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
8581 (!LegalOperations ||
8582 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
8585 { N0.getOperand(0), N0.getOperand(1),
8586 DAG.getConstantFP(-1.0, DL, VT), DAG.getConstantFP(0.0, DL, VT),
8588 return DAG.getNode(ISD::SELECT_CC, DL, VT, Ops);
8591 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
8592 // (select_cc x, y, 1.0, 0.0,, cc)
8593 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
8594 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
8595 (!LegalOperations ||
8596 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
8599 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
8600 DAG.getConstantFP(1.0, DL, VT), DAG.getConstantFP(0.0, DL, VT),
8601 N0.getOperand(0).getOperand(2) };
8602 return DAG.getNode(ISD::SELECT_CC, DL, VT, Ops);
8609 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
8610 SDValue N0 = N->getOperand(0);
8611 EVT VT = N->getValueType(0);
8612 EVT OpVT = N0.getValueType();
8614 // fold (uint_to_fp c1) -> c1fp
8615 if (isConstantIntBuildVectorOrConstantInt(N0) &&
8616 // ...but only if the target supports immediate floating-point values
8617 (!LegalOperations ||
8618 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
8619 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
8621 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
8622 // but SINT_TO_FP is legal on this target, try to convert.
8623 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
8624 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
8625 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
8626 if (DAG.SignBitIsZero(N0))
8627 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
8630 // The next optimizations are desirable only if SELECT_CC can be lowered.
8631 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
8632 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
8634 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
8635 (!LegalOperations ||
8636 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
8639 { N0.getOperand(0), N0.getOperand(1),
8640 DAG.getConstantFP(1.0, DL, VT), DAG.getConstantFP(0.0, DL, VT),
8642 return DAG.getNode(ISD::SELECT_CC, DL, VT, Ops);
8649 // Fold (fp_to_{s/u}int ({s/u}int_to_fpx)) -> zext x, sext x, trunc x, or x
8650 static SDValue FoldIntToFPToInt(SDNode *N, SelectionDAG &DAG) {
8651 SDValue N0 = N->getOperand(0);
8652 EVT VT = N->getValueType(0);
8654 if (N0.getOpcode() != ISD::UINT_TO_FP && N0.getOpcode() != ISD::SINT_TO_FP)
8657 SDValue Src = N0.getOperand(0);
8658 EVT SrcVT = Src.getValueType();
8659 bool IsInputSigned = N0.getOpcode() == ISD::SINT_TO_FP;
8660 bool IsOutputSigned = N->getOpcode() == ISD::FP_TO_SINT;
8662 // We can safely assume the conversion won't overflow the output range,
8663 // because (for example) (uint8_t)18293.f is undefined behavior.
8665 // Since we can assume the conversion won't overflow, our decision as to
8666 // whether the input will fit in the float should depend on the minimum
8667 // of the input range and output range.
8669 // This means this is also safe for a signed input and unsigned output, since
8670 // a negative input would lead to undefined behavior.
8671 unsigned InputSize = (int)SrcVT.getScalarSizeInBits() - IsInputSigned;
8672 unsigned OutputSize = (int)VT.getScalarSizeInBits() - IsOutputSigned;
8673 unsigned ActualSize = std::min(InputSize, OutputSize);
8674 const fltSemantics &sem = DAG.EVTToAPFloatSemantics(N0.getValueType());
8676 // We can only fold away the float conversion if the input range can be
8677 // represented exactly in the float range.
8678 if (APFloat::semanticsPrecision(sem) >= ActualSize) {
8679 if (VT.getScalarSizeInBits() > SrcVT.getScalarSizeInBits()) {
8680 unsigned ExtOp = IsInputSigned && IsOutputSigned ? ISD::SIGN_EXTEND
8682 return DAG.getNode(ExtOp, SDLoc(N), VT, Src);
8684 if (VT.getScalarSizeInBits() < SrcVT.getScalarSizeInBits())
8685 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Src);
8688 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Src);
8693 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
8694 SDValue N0 = N->getOperand(0);
8695 EVT VT = N->getValueType(0);
8697 // fold (fp_to_sint c1fp) -> c1
8698 if (isConstantFPBuildVectorOrConstantFP(N0))
8699 return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0);
8701 return FoldIntToFPToInt(N, DAG);
8704 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
8705 SDValue N0 = N->getOperand(0);
8706 EVT VT = N->getValueType(0);
8708 // fold (fp_to_uint c1fp) -> c1
8709 if (isConstantFPBuildVectorOrConstantFP(N0))
8710 return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0);
8712 return FoldIntToFPToInt(N, DAG);
8715 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
8716 SDValue N0 = N->getOperand(0);
8717 SDValue N1 = N->getOperand(1);
8718 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8719 EVT VT = N->getValueType(0);
8721 // fold (fp_round c1fp) -> c1fp
8723 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1);
8725 // fold (fp_round (fp_extend x)) -> x
8726 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
8727 return N0.getOperand(0);
8729 // fold (fp_round (fp_round x)) -> (fp_round x)
8730 if (N0.getOpcode() == ISD::FP_ROUND) {
8731 const bool NIsTrunc = N->getConstantOperandVal(1) == 1;
8732 const bool N0IsTrunc = N0.getNode()->getConstantOperandVal(1) == 1;
8733 // If the first fp_round isn't a value preserving truncation, it might
8734 // introduce a tie in the second fp_round, that wouldn't occur in the
8735 // single-step fp_round we want to fold to.
8736 // In other words, double rounding isn't the same as rounding.
8737 // Also, this is a value preserving truncation iff both fp_round's are.
8738 if (DAG.getTarget().Options.UnsafeFPMath || N0IsTrunc) {
8740 return DAG.getNode(ISD::FP_ROUND, DL, VT, N0.getOperand(0),
8741 DAG.getIntPtrConstant(NIsTrunc && N0IsTrunc, DL));
8745 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
8746 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
8747 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT,
8748 N0.getOperand(0), N1);
8749 AddToWorklist(Tmp.getNode());
8750 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
8751 Tmp, N0.getOperand(1));
8757 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
8758 SDValue N0 = N->getOperand(0);
8759 EVT VT = N->getValueType(0);
8760 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
8761 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8763 // fold (fp_round_inreg c1fp) -> c1fp
8764 if (N0CFP && isTypeLegal(EVT)) {
8766 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), DL, EVT);
8767 return DAG.getNode(ISD::FP_EXTEND, DL, VT, Round);
8773 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
8774 SDValue N0 = N->getOperand(0);
8775 EVT VT = N->getValueType(0);
8777 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
8778 if (N->hasOneUse() &&
8779 N->use_begin()->getOpcode() == ISD::FP_ROUND)
8782 // fold (fp_extend c1fp) -> c1fp
8783 if (isConstantFPBuildVectorOrConstantFP(N0))
8784 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0);
8786 // fold (fp_extend (fp16_to_fp op)) -> (fp16_to_fp op)
8787 if (N0.getOpcode() == ISD::FP16_TO_FP &&
8788 TLI.getOperationAction(ISD::FP16_TO_FP, VT) == TargetLowering::Legal)
8789 return DAG.getNode(ISD::FP16_TO_FP, SDLoc(N), VT, N0.getOperand(0));
8791 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
8793 if (N0.getOpcode() == ISD::FP_ROUND
8794 && N0.getNode()->getConstantOperandVal(1) == 1) {
8795 SDValue In = N0.getOperand(0);
8796 if (In.getValueType() == VT) return In;
8797 if (VT.bitsLT(In.getValueType()))
8798 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT,
8799 In, N0.getOperand(1));
8800 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In);
8803 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
8804 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
8805 TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) {
8806 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
8807 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
8809 LN0->getBasePtr(), N0.getValueType(),
8810 LN0->getMemOperand());
8811 CombineTo(N, ExtLoad);
8812 CombineTo(N0.getNode(),
8813 DAG.getNode(ISD::FP_ROUND, SDLoc(N0),
8814 N0.getValueType(), ExtLoad,
8815 DAG.getIntPtrConstant(1, SDLoc(N0))),
8816 ExtLoad.getValue(1));
8817 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8823 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
8824 SDValue N0 = N->getOperand(0);
8825 EVT VT = N->getValueType(0);
8827 // fold (fceil c1) -> fceil(c1)
8828 if (isConstantFPBuildVectorOrConstantFP(N0))
8829 return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0);
8834 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
8835 SDValue N0 = N->getOperand(0);
8836 EVT VT = N->getValueType(0);
8838 // fold (ftrunc c1) -> ftrunc(c1)
8839 if (isConstantFPBuildVectorOrConstantFP(N0))
8840 return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0);
8845 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
8846 SDValue N0 = N->getOperand(0);
8847 EVT VT = N->getValueType(0);
8849 // fold (ffloor c1) -> ffloor(c1)
8850 if (isConstantFPBuildVectorOrConstantFP(N0))
8851 return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0);
8856 // FIXME: FNEG and FABS have a lot in common; refactor.
8857 SDValue DAGCombiner::visitFNEG(SDNode *N) {
8858 SDValue N0 = N->getOperand(0);
8859 EVT VT = N->getValueType(0);
8861 // Constant fold FNEG.
8862 if (isConstantFPBuildVectorOrConstantFP(N0))
8863 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
8865 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
8866 &DAG.getTarget().Options))
8867 return GetNegatedExpression(N0, DAG, LegalOperations);
8869 // Transform fneg(bitconvert(x)) -> bitconvert(x ^ sign) to avoid loading
8870 // constant pool values.
8871 if (!TLI.isFNegFree(VT) &&
8872 N0.getOpcode() == ISD::BITCAST &&
8873 N0.getNode()->hasOneUse()) {
8874 SDValue Int = N0.getOperand(0);
8875 EVT IntVT = Int.getValueType();
8876 if (IntVT.isInteger() && !IntVT.isVector()) {
8878 if (N0.getValueType().isVector()) {
8879 // For a vector, get a mask such as 0x80... per scalar element
8881 SignMask = APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
8882 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
8884 // For a scalar, just generate 0x80...
8885 SignMask = APInt::getSignBit(IntVT.getSizeInBits());
8888 Int = DAG.getNode(ISD::XOR, DL0, IntVT, Int,
8889 DAG.getConstant(SignMask, DL0, IntVT));
8890 AddToWorklist(Int.getNode());
8891 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Int);
8895 // (fneg (fmul c, x)) -> (fmul -c, x)
8896 if (N0.getOpcode() == ISD::FMUL &&
8897 (N0.getNode()->hasOneUse() || !TLI.isFNegFree(VT))) {
8898 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
8900 APFloat CVal = CFP1->getValueAPF();
8902 if (Level >= AfterLegalizeDAG &&
8903 (TLI.isFPImmLegal(CVal, N->getValueType(0)) ||
8904 TLI.isOperationLegal(ISD::ConstantFP, N->getValueType(0))))
8906 ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
8907 DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0.getOperand(1)));
8914 SDValue DAGCombiner::visitFMINNUM(SDNode *N) {
8915 SDValue N0 = N->getOperand(0);
8916 SDValue N1 = N->getOperand(1);
8917 const ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8918 const ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8920 if (N0CFP && N1CFP) {
8921 const APFloat &C0 = N0CFP->getValueAPF();
8922 const APFloat &C1 = N1CFP->getValueAPF();
8923 return DAG.getConstantFP(minnum(C0, C1), SDLoc(N), N->getValueType(0));
8927 EVT VT = N->getValueType(0);
8928 // Canonicalize to constant on RHS.
8929 return DAG.getNode(ISD::FMINNUM, SDLoc(N), VT, N1, N0);
8935 SDValue DAGCombiner::visitFMAXNUM(SDNode *N) {
8936 SDValue N0 = N->getOperand(0);
8937 SDValue N1 = N->getOperand(1);
8938 const ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8939 const ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8941 if (N0CFP && N1CFP) {
8942 const APFloat &C0 = N0CFP->getValueAPF();
8943 const APFloat &C1 = N1CFP->getValueAPF();
8944 return DAG.getConstantFP(maxnum(C0, C1), SDLoc(N), N->getValueType(0));
8948 EVT VT = N->getValueType(0);
8949 // Canonicalize to constant on RHS.
8950 return DAG.getNode(ISD::FMAXNUM, SDLoc(N), VT, N1, N0);
8956 SDValue DAGCombiner::visitFABS(SDNode *N) {
8957 SDValue N0 = N->getOperand(0);
8958 EVT VT = N->getValueType(0);
8960 // fold (fabs c1) -> fabs(c1)
8961 if (isConstantFPBuildVectorOrConstantFP(N0))
8962 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
8964 // fold (fabs (fabs x)) -> (fabs x)
8965 if (N0.getOpcode() == ISD::FABS)
8966 return N->getOperand(0);
8968 // fold (fabs (fneg x)) -> (fabs x)
8969 // fold (fabs (fcopysign x, y)) -> (fabs x)
8970 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
8971 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0));
8973 // Transform fabs(bitconvert(x)) -> bitconvert(x & ~sign) to avoid loading
8974 // constant pool values.
8975 if (!TLI.isFAbsFree(VT) &&
8976 N0.getOpcode() == ISD::BITCAST &&
8977 N0.getNode()->hasOneUse()) {
8978 SDValue Int = N0.getOperand(0);
8979 EVT IntVT = Int.getValueType();
8980 if (IntVT.isInteger() && !IntVT.isVector()) {
8982 if (N0.getValueType().isVector()) {
8983 // For a vector, get a mask such as 0x7f... per scalar element
8985 SignMask = ~APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
8986 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
8988 // For a scalar, just generate 0x7f...
8989 SignMask = ~APInt::getSignBit(IntVT.getSizeInBits());
8992 Int = DAG.getNode(ISD::AND, DL, IntVT, Int,
8993 DAG.getConstant(SignMask, DL, IntVT));
8994 AddToWorklist(Int.getNode());
8995 return DAG.getNode(ISD::BITCAST, SDLoc(N), N->getValueType(0), Int);
9002 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
9003 SDValue Chain = N->getOperand(0);
9004 SDValue N1 = N->getOperand(1);
9005 SDValue N2 = N->getOperand(2);
9007 // If N is a constant we could fold this into a fallthrough or unconditional
9008 // branch. However that doesn't happen very often in normal code, because
9009 // Instcombine/SimplifyCFG should have handled the available opportunities.
9010 // If we did this folding here, it would be necessary to update the
9011 // MachineBasicBlock CFG, which is awkward.
9013 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
9015 if (N1.getOpcode() == ISD::SETCC &&
9016 TLI.isOperationLegalOrCustom(ISD::BR_CC,
9017 N1.getOperand(0).getValueType())) {
9018 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
9019 Chain, N1.getOperand(2),
9020 N1.getOperand(0), N1.getOperand(1), N2);
9023 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
9024 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
9025 (N1.getOperand(0).hasOneUse() &&
9026 N1.getOperand(0).getOpcode() == ISD::SRL))) {
9027 SDNode *Trunc = nullptr;
9028 if (N1.getOpcode() == ISD::TRUNCATE) {
9029 // Look pass the truncate.
9030 Trunc = N1.getNode();
9031 N1 = N1.getOperand(0);
9034 // Match this pattern so that we can generate simpler code:
9037 // %b = and i32 %a, 2
9038 // %c = srl i32 %b, 1
9039 // brcond i32 %c ...
9044 // %b = and i32 %a, 2
9045 // %c = setcc eq %b, 0
9048 // This applies only when the AND constant value has one bit set and the
9049 // SRL constant is equal to the log2 of the AND constant. The back-end is
9050 // smart enough to convert the result into a TEST/JMP sequence.
9051 SDValue Op0 = N1.getOperand(0);
9052 SDValue Op1 = N1.getOperand(1);
9054 if (Op0.getOpcode() == ISD::AND &&
9055 Op1.getOpcode() == ISD::Constant) {
9056 SDValue AndOp1 = Op0.getOperand(1);
9058 if (AndOp1.getOpcode() == ISD::Constant) {
9059 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
9061 if (AndConst.isPowerOf2() &&
9062 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
9066 getSetCCResultType(Op0.getValueType()),
9067 Op0, DAG.getConstant(0, DL, Op0.getValueType()),
9070 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, DL,
9071 MVT::Other, Chain, SetCC, N2);
9072 // Don't add the new BRCond into the worklist or else SimplifySelectCC
9073 // will convert it back to (X & C1) >> C2.
9074 CombineTo(N, NewBRCond, false);
9075 // Truncate is dead.
9077 deleteAndRecombine(Trunc);
9078 // Replace the uses of SRL with SETCC
9079 WorklistRemover DeadNodes(*this);
9080 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
9081 deleteAndRecombine(N1.getNode());
9082 return SDValue(N, 0); // Return N so it doesn't get rechecked!
9088 // Restore N1 if the above transformation doesn't match.
9089 N1 = N->getOperand(1);
9092 // Transform br(xor(x, y)) -> br(x != y)
9093 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
9094 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
9095 SDNode *TheXor = N1.getNode();
9096 SDValue Op0 = TheXor->getOperand(0);
9097 SDValue Op1 = TheXor->getOperand(1);
9098 if (Op0.getOpcode() == Op1.getOpcode()) {
9099 // Avoid missing important xor optimizations.
9100 if (SDValue Tmp = visitXOR(TheXor)) {
9101 if (Tmp.getNode() != TheXor) {
9102 DEBUG(dbgs() << "\nReplacing.8 ";
9104 dbgs() << "\nWith: ";
9105 Tmp.getNode()->dump(&DAG);
9107 WorklistRemover DeadNodes(*this);
9108 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
9109 deleteAndRecombine(TheXor);
9110 return DAG.getNode(ISD::BRCOND, SDLoc(N),
9111 MVT::Other, Chain, Tmp, N2);
9114 // visitXOR has changed XOR's operands or replaced the XOR completely,
9116 return SDValue(N, 0);
9120 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
9122 if (isOneConstant(Op0) && Op0.hasOneUse() &&
9123 Op0.getOpcode() == ISD::XOR) {
9124 TheXor = Op0.getNode();
9128 EVT SetCCVT = N1.getValueType();
9130 SetCCVT = getSetCCResultType(SetCCVT);
9131 SDValue SetCC = DAG.getSetCC(SDLoc(TheXor),
9134 Equal ? ISD::SETEQ : ISD::SETNE);
9135 // Replace the uses of XOR with SETCC
9136 WorklistRemover DeadNodes(*this);
9137 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
9138 deleteAndRecombine(N1.getNode());
9139 return DAG.getNode(ISD::BRCOND, SDLoc(N),
9140 MVT::Other, Chain, SetCC, N2);
9147 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
9149 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
9150 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
9151 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
9153 // If N is a constant we could fold this into a fallthrough or unconditional
9154 // branch. However that doesn't happen very often in normal code, because
9155 // Instcombine/SimplifyCFG should have handled the available opportunities.
9156 // If we did this folding here, it would be necessary to update the
9157 // MachineBasicBlock CFG, which is awkward.
9159 // Use SimplifySetCC to simplify SETCC's.
9160 SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()),
9161 CondLHS, CondRHS, CC->get(), SDLoc(N),
9163 if (Simp.getNode()) AddToWorklist(Simp.getNode());
9165 // fold to a simpler setcc
9166 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
9167 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
9168 N->getOperand(0), Simp.getOperand(2),
9169 Simp.getOperand(0), Simp.getOperand(1),
9175 /// Return true if 'Use' is a load or a store that uses N as its base pointer
9176 /// and that N may be folded in the load / store addressing mode.
9177 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
9179 const TargetLowering &TLI) {
9183 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
9184 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
9186 VT = LD->getMemoryVT();
9187 AS = LD->getAddressSpace();
9188 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
9189 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
9191 VT = ST->getMemoryVT();
9192 AS = ST->getAddressSpace();
9196 TargetLowering::AddrMode AM;
9197 if (N->getOpcode() == ISD::ADD) {
9198 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
9201 AM.BaseOffs = Offset->getSExtValue();
9205 } else if (N->getOpcode() == ISD::SUB) {
9206 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
9209 AM.BaseOffs = -Offset->getSExtValue();
9216 return TLI.isLegalAddressingMode(DAG.getDataLayout(), AM,
9217 VT.getTypeForEVT(*DAG.getContext()), AS);
9220 /// Try turning a load/store into a pre-indexed load/store when the base
9221 /// pointer is an add or subtract and it has other uses besides the load/store.
9222 /// After the transformation, the new indexed load/store has effectively folded
9223 /// the add/subtract in and all of its other uses are redirected to the
9225 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
9226 if (Level < AfterLegalizeDAG)
9232 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9233 if (LD->isIndexed())
9235 VT = LD->getMemoryVT();
9236 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
9237 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
9239 Ptr = LD->getBasePtr();
9240 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9241 if (ST->isIndexed())
9243 VT = ST->getMemoryVT();
9244 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
9245 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
9247 Ptr = ST->getBasePtr();
9253 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
9254 // out. There is no reason to make this a preinc/predec.
9255 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
9256 Ptr.getNode()->hasOneUse())
9259 // Ask the target to do addressing mode selection.
9262 ISD::MemIndexedMode AM = ISD::UNINDEXED;
9263 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
9266 // Backends without true r+i pre-indexed forms may need to pass a
9267 // constant base with a variable offset so that constant coercion
9268 // will work with the patterns in canonical form.
9269 bool Swapped = false;
9270 if (isa<ConstantSDNode>(BasePtr)) {
9271 std::swap(BasePtr, Offset);
9275 // Don't create a indexed load / store with zero offset.
9276 if (isNullConstant(Offset))
9279 // Try turning it into a pre-indexed load / store except when:
9280 // 1) The new base ptr is a frame index.
9281 // 2) If N is a store and the new base ptr is either the same as or is a
9282 // predecessor of the value being stored.
9283 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
9284 // that would create a cycle.
9285 // 4) All uses are load / store ops that use it as old base ptr.
9287 // Check #1. Preinc'ing a frame index would require copying the stack pointer
9288 // (plus the implicit offset) to a register to preinc anyway.
9289 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
9294 SDValue Val = cast<StoreSDNode>(N)->getValue();
9295 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
9299 // If the offset is a constant, there may be other adds of constants that
9300 // can be folded with this one. We should do this to avoid having to keep
9301 // a copy of the original base pointer.
9302 SmallVector<SDNode *, 16> OtherUses;
9303 if (isa<ConstantSDNode>(Offset))
9304 for (SDNode::use_iterator UI = BasePtr.getNode()->use_begin(),
9305 UE = BasePtr.getNode()->use_end();
9307 SDUse &Use = UI.getUse();
9308 // Skip the use that is Ptr and uses of other results from BasePtr's
9309 // node (important for nodes that return multiple results).
9310 if (Use.getUser() == Ptr.getNode() || Use != BasePtr)
9313 if (Use.getUser()->isPredecessorOf(N))
9316 if (Use.getUser()->getOpcode() != ISD::ADD &&
9317 Use.getUser()->getOpcode() != ISD::SUB) {
9322 SDValue Op1 = Use.getUser()->getOperand((UI.getOperandNo() + 1) & 1);
9323 if (!isa<ConstantSDNode>(Op1)) {
9328 // FIXME: In some cases, we can be smarter about this.
9329 if (Op1.getValueType() != Offset.getValueType()) {
9334 OtherUses.push_back(Use.getUser());
9338 std::swap(BasePtr, Offset);
9340 // Now check for #3 and #4.
9341 bool RealUse = false;
9343 // Caches for hasPredecessorHelper
9344 SmallPtrSet<const SDNode *, 32> Visited;
9345 SmallVector<const SDNode *, 16> Worklist;
9347 for (SDNode *Use : Ptr.getNode()->uses()) {
9350 if (N->hasPredecessorHelper(Use, Visited, Worklist))
9353 // If Ptr may be folded in addressing mode of other use, then it's
9354 // not profitable to do this transformation.
9355 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
9364 Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
9365 BasePtr, Offset, AM);
9367 Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
9368 BasePtr, Offset, AM);
9371 DEBUG(dbgs() << "\nReplacing.4 ";
9373 dbgs() << "\nWith: ";
9374 Result.getNode()->dump(&DAG);
9376 WorklistRemover DeadNodes(*this);
9378 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
9379 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
9381 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
9384 // Finally, since the node is now dead, remove it from the graph.
9385 deleteAndRecombine(N);
9388 std::swap(BasePtr, Offset);
9390 // Replace other uses of BasePtr that can be updated to use Ptr
9391 for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) {
9392 unsigned OffsetIdx = 1;
9393 if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode())
9395 assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() ==
9396 BasePtr.getNode() && "Expected BasePtr operand");
9398 // We need to replace ptr0 in the following expression:
9399 // x0 * offset0 + y0 * ptr0 = t0
9401 // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store)
9403 // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the
9404 // indexed load/store and the expresion that needs to be re-written.
9406 // Therefore, we have:
9407 // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1
9409 ConstantSDNode *CN =
9410 cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx));
9412 APInt Offset0 = CN->getAPIntValue();
9413 APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue();
9415 X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1;
9416 Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1;
9417 X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1;
9418 Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1;
9420 unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD;
9422 APInt CNV = Offset0;
9423 if (X0 < 0) CNV = -CNV;
9424 if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1;
9425 else CNV = CNV - Offset1;
9427 SDLoc DL(OtherUses[i]);
9429 // We can now generate the new expression.
9430 SDValue NewOp1 = DAG.getConstant(CNV, DL, CN->getValueType(0));
9431 SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0);
9433 SDValue NewUse = DAG.getNode(Opcode,
9435 OtherUses[i]->getValueType(0), NewOp1, NewOp2);
9436 DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse);
9437 deleteAndRecombine(OtherUses[i]);
9440 // Replace the uses of Ptr with uses of the updated base value.
9441 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
9442 deleteAndRecombine(Ptr.getNode());
9447 /// Try to combine a load/store with a add/sub of the base pointer node into a
9448 /// post-indexed load/store. The transformation folded the add/subtract into the
9449 /// new indexed load/store effectively and all of its uses are redirected to the
9451 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
9452 if (Level < AfterLegalizeDAG)
9458 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9459 if (LD->isIndexed())
9461 VT = LD->getMemoryVT();
9462 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
9463 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
9465 Ptr = LD->getBasePtr();
9466 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9467 if (ST->isIndexed())
9469 VT = ST->getMemoryVT();
9470 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
9471 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
9473 Ptr = ST->getBasePtr();
9479 if (Ptr.getNode()->hasOneUse())
9482 for (SDNode *Op : Ptr.getNode()->uses()) {
9484 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
9489 ISD::MemIndexedMode AM = ISD::UNINDEXED;
9490 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
9491 // Don't create a indexed load / store with zero offset.
9492 if (isNullConstant(Offset))
9495 // Try turning it into a post-indexed load / store except when
9496 // 1) All uses are load / store ops that use it as base ptr (and
9497 // it may be folded as addressing mmode).
9498 // 2) Op must be independent of N, i.e. Op is neither a predecessor
9499 // nor a successor of N. Otherwise, if Op is folded that would
9502 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
9506 bool TryNext = false;
9507 for (SDNode *Use : BasePtr.getNode()->uses()) {
9508 if (Use == Ptr.getNode())
9511 // If all the uses are load / store addresses, then don't do the
9513 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
9514 bool RealUse = false;
9515 for (SDNode *UseUse : Use->uses()) {
9516 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
9531 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
9532 SDValue Result = isLoad
9533 ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
9534 BasePtr, Offset, AM)
9535 : DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
9536 BasePtr, Offset, AM);
9539 DEBUG(dbgs() << "\nReplacing.5 ";
9541 dbgs() << "\nWith: ";
9542 Result.getNode()->dump(&DAG);
9544 WorklistRemover DeadNodes(*this);
9546 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
9547 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
9549 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
9552 // Finally, since the node is now dead, remove it from the graph.
9553 deleteAndRecombine(N);
9555 // Replace the uses of Use with uses of the updated base value.
9556 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
9557 Result.getValue(isLoad ? 1 : 0));
9558 deleteAndRecombine(Op);
9567 /// \brief Return the base-pointer arithmetic from an indexed \p LD.
9568 SDValue DAGCombiner::SplitIndexingFromLoad(LoadSDNode *LD) {
9569 ISD::MemIndexedMode AM = LD->getAddressingMode();
9570 assert(AM != ISD::UNINDEXED);
9571 SDValue BP = LD->getOperand(1);
9572 SDValue Inc = LD->getOperand(2);
9574 // Some backends use TargetConstants for load offsets, but don't expect
9575 // TargetConstants in general ADD nodes. We can convert these constants into
9576 // regular Constants (if the constant is not opaque).
9577 assert((Inc.getOpcode() != ISD::TargetConstant ||
9578 !cast<ConstantSDNode>(Inc)->isOpaque()) &&
9579 "Cannot split out indexing using opaque target constants");
9580 if (Inc.getOpcode() == ISD::TargetConstant) {
9581 ConstantSDNode *ConstInc = cast<ConstantSDNode>(Inc);
9582 Inc = DAG.getConstant(*ConstInc->getConstantIntValue(), SDLoc(Inc),
9583 ConstInc->getValueType(0));
9587 (AM == ISD::PRE_INC || AM == ISD::POST_INC ? ISD::ADD : ISD::SUB);
9588 return DAG.getNode(Opc, SDLoc(LD), BP.getSimpleValueType(), BP, Inc);
9591 SDValue DAGCombiner::visitLOAD(SDNode *N) {
9592 LoadSDNode *LD = cast<LoadSDNode>(N);
9593 SDValue Chain = LD->getChain();
9594 SDValue Ptr = LD->getBasePtr();
9596 // If load is not volatile and there are no uses of the loaded value (and
9597 // the updated indexed value in case of indexed loads), change uses of the
9598 // chain value into uses of the chain input (i.e. delete the dead load).
9599 if (!LD->isVolatile()) {
9600 if (N->getValueType(1) == MVT::Other) {
9602 if (!N->hasAnyUseOfValue(0)) {
9603 // It's not safe to use the two value CombineTo variant here. e.g.
9604 // v1, chain2 = load chain1, loc
9605 // v2, chain3 = load chain2, loc
9607 // Now we replace use of chain2 with chain1. This makes the second load
9608 // isomorphic to the one we are deleting, and thus makes this load live.
9609 DEBUG(dbgs() << "\nReplacing.6 ";
9611 dbgs() << "\nWith chain: ";
9612 Chain.getNode()->dump(&DAG);
9614 WorklistRemover DeadNodes(*this);
9615 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
9618 deleteAndRecombine(N);
9620 return SDValue(N, 0); // Return N so it doesn't get rechecked!
9624 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
9626 // If this load has an opaque TargetConstant offset, then we cannot split
9627 // the indexing into an add/sub directly (that TargetConstant may not be
9628 // valid for a different type of node, and we cannot convert an opaque
9629 // target constant into a regular constant).
9630 bool HasOTCInc = LD->getOperand(2).getOpcode() == ISD::TargetConstant &&
9631 cast<ConstantSDNode>(LD->getOperand(2))->isOpaque();
9633 if (!N->hasAnyUseOfValue(0) &&
9634 ((MaySplitLoadIndex && !HasOTCInc) || !N->hasAnyUseOfValue(1))) {
9635 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
9637 if (N->hasAnyUseOfValue(1) && MaySplitLoadIndex && !HasOTCInc) {
9638 Index = SplitIndexingFromLoad(LD);
9639 // Try to fold the base pointer arithmetic into subsequent loads and
9641 AddUsersToWorklist(N);
9643 Index = DAG.getUNDEF(N->getValueType(1));
9644 DEBUG(dbgs() << "\nReplacing.7 ";
9646 dbgs() << "\nWith: ";
9647 Undef.getNode()->dump(&DAG);
9648 dbgs() << " and 2 other values\n");
9649 WorklistRemover DeadNodes(*this);
9650 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
9651 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Index);
9652 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
9653 deleteAndRecombine(N);
9654 return SDValue(N, 0); // Return N so it doesn't get rechecked!
9659 // If this load is directly stored, replace the load value with the stored
9661 // TODO: Handle store large -> read small portion.
9662 // TODO: Handle TRUNCSTORE/LOADEXT
9663 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
9664 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
9665 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
9666 if (PrevST->getBasePtr() == Ptr &&
9667 PrevST->getValue().getValueType() == N->getValueType(0))
9668 return CombineTo(N, Chain.getOperand(1), Chain);
9672 // Try to infer better alignment information than the load already has.
9673 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
9674 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
9675 if (Align > LD->getMemOperand()->getBaseAlignment()) {
9677 DAG.getExtLoad(LD->getExtensionType(), SDLoc(N),
9678 LD->getValueType(0),
9679 Chain, Ptr, LD->getPointerInfo(),
9681 LD->isVolatile(), LD->isNonTemporal(),
9682 LD->isInvariant(), Align, LD->getAAInfo());
9683 if (NewLoad.getNode() != N)
9684 return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true);
9689 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
9690 : DAG.getSubtarget().useAA();
9692 if (CombinerAAOnlyFunc.getNumOccurrences() &&
9693 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
9696 if (UseAA && LD->isUnindexed()) {
9697 // Walk up chain skipping non-aliasing memory nodes.
9698 SDValue BetterChain = FindBetterChain(N, Chain);
9700 // If there is a better chain.
9701 if (Chain != BetterChain) {
9704 // Replace the chain to void dependency.
9705 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
9706 ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD),
9707 BetterChain, Ptr, LD->getMemOperand());
9709 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD),
9710 LD->getValueType(0),
9711 BetterChain, Ptr, LD->getMemoryVT(),
9712 LD->getMemOperand());
9715 // Create token factor to keep old chain connected.
9716 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
9717 MVT::Other, Chain, ReplLoad.getValue(1));
9719 // Make sure the new and old chains are cleaned up.
9720 AddToWorklist(Token.getNode());
9722 // Replace uses with load result and token factor. Don't add users
9724 return CombineTo(N, ReplLoad.getValue(0), Token, false);
9728 // Try transforming N to an indexed load.
9729 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
9730 return SDValue(N, 0);
9732 // Try to slice up N to more direct loads if the slices are mapped to
9733 // different register banks or pairing can take place.
9735 return SDValue(N, 0);
9741 /// \brief Helper structure used to slice a load in smaller loads.
9742 /// Basically a slice is obtained from the following sequence:
9743 /// Origin = load Ty1, Base
9744 /// Shift = srl Ty1 Origin, CstTy Amount
9745 /// Inst = trunc Shift to Ty2
9747 /// Then, it will be rewriten into:
9748 /// Slice = load SliceTy, Base + SliceOffset
9749 /// [Inst = zext Slice to Ty2], only if SliceTy <> Ty2
9751 /// SliceTy is deduced from the number of bits that are actually used to
9753 struct LoadedSlice {
9754 /// \brief Helper structure used to compute the cost of a slice.
9756 /// Are we optimizing for code size.
9761 unsigned CrossRegisterBanksCopies;
9765 Cost(bool ForCodeSize = false)
9766 : ForCodeSize(ForCodeSize), Loads(0), Truncates(0),
9767 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {}
9769 /// \brief Get the cost of one isolated slice.
9770 Cost(const LoadedSlice &LS, bool ForCodeSize = false)
9771 : ForCodeSize(ForCodeSize), Loads(1), Truncates(0),
9772 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {
9773 EVT TruncType = LS.Inst->getValueType(0);
9774 EVT LoadedType = LS.getLoadedType();
9775 if (TruncType != LoadedType &&
9776 !LS.DAG->getTargetLoweringInfo().isZExtFree(LoadedType, TruncType))
9780 /// \brief Account for slicing gain in the current cost.
9781 /// Slicing provide a few gains like removing a shift or a
9782 /// truncate. This method allows to grow the cost of the original
9783 /// load with the gain from this slice.
9784 void addSliceGain(const LoadedSlice &LS) {
9785 // Each slice saves a truncate.
9786 const TargetLowering &TLI = LS.DAG->getTargetLoweringInfo();
9787 if (!TLI.isTruncateFree(LS.Inst->getOperand(0).getValueType(),
9788 LS.Inst->getValueType(0)))
9790 // If there is a shift amount, this slice gets rid of it.
9793 // If this slice can merge a cross register bank copy, account for it.
9794 if (LS.canMergeExpensiveCrossRegisterBankCopy())
9795 ++CrossRegisterBanksCopies;
9798 Cost &operator+=(const Cost &RHS) {
9800 Truncates += RHS.Truncates;
9801 CrossRegisterBanksCopies += RHS.CrossRegisterBanksCopies;
9807 bool operator==(const Cost &RHS) const {
9808 return Loads == RHS.Loads && Truncates == RHS.Truncates &&
9809 CrossRegisterBanksCopies == RHS.CrossRegisterBanksCopies &&
9810 ZExts == RHS.ZExts && Shift == RHS.Shift;
9813 bool operator!=(const Cost &RHS) const { return !(*this == RHS); }
9815 bool operator<(const Cost &RHS) const {
9816 // Assume cross register banks copies are as expensive as loads.
9817 // FIXME: Do we want some more target hooks?
9818 unsigned ExpensiveOpsLHS = Loads + CrossRegisterBanksCopies;
9819 unsigned ExpensiveOpsRHS = RHS.Loads + RHS.CrossRegisterBanksCopies;
9820 // Unless we are optimizing for code size, consider the
9821 // expensive operation first.
9822 if (!ForCodeSize && ExpensiveOpsLHS != ExpensiveOpsRHS)
9823 return ExpensiveOpsLHS < ExpensiveOpsRHS;
9824 return (Truncates + ZExts + Shift + ExpensiveOpsLHS) <
9825 (RHS.Truncates + RHS.ZExts + RHS.Shift + ExpensiveOpsRHS);
9828 bool operator>(const Cost &RHS) const { return RHS < *this; }
9830 bool operator<=(const Cost &RHS) const { return !(RHS < *this); }
9832 bool operator>=(const Cost &RHS) const { return !(*this < RHS); }
9834 // The last instruction that represent the slice. This should be a
9835 // truncate instruction.
9837 // The original load instruction.
9839 // The right shift amount in bits from the original load.
9841 // The DAG from which Origin came from.
9842 // This is used to get some contextual information about legal types, etc.
9845 LoadedSlice(SDNode *Inst = nullptr, LoadSDNode *Origin = nullptr,
9846 unsigned Shift = 0, SelectionDAG *DAG = nullptr)
9847 : Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {}
9849 /// \brief Get the bits used in a chunk of bits \p BitWidth large.
9850 /// \return Result is \p BitWidth and has used bits set to 1 and
9851 /// not used bits set to 0.
9852 APInt getUsedBits() const {
9853 // Reproduce the trunc(lshr) sequence:
9854 // - Start from the truncated value.
9855 // - Zero extend to the desired bit width.
9857 assert(Origin && "No original load to compare against.");
9858 unsigned BitWidth = Origin->getValueSizeInBits(0);
9859 assert(Inst && "This slice is not bound to an instruction");
9860 assert(Inst->getValueSizeInBits(0) <= BitWidth &&
9861 "Extracted slice is bigger than the whole type!");
9862 APInt UsedBits(Inst->getValueSizeInBits(0), 0);
9863 UsedBits.setAllBits();
9864 UsedBits = UsedBits.zext(BitWidth);
9869 /// \brief Get the size of the slice to be loaded in bytes.
9870 unsigned getLoadedSize() const {
9871 unsigned SliceSize = getUsedBits().countPopulation();
9872 assert(!(SliceSize & 0x7) && "Size is not a multiple of a byte.");
9873 return SliceSize / 8;
9876 /// \brief Get the type that will be loaded for this slice.
9877 /// Note: This may not be the final type for the slice.
9878 EVT getLoadedType() const {
9879 assert(DAG && "Missing context");
9880 LLVMContext &Ctxt = *DAG->getContext();
9881 return EVT::getIntegerVT(Ctxt, getLoadedSize() * 8);
9884 /// \brief Get the alignment of the load used for this slice.
9885 unsigned getAlignment() const {
9886 unsigned Alignment = Origin->getAlignment();
9887 unsigned Offset = getOffsetFromBase();
9889 Alignment = MinAlign(Alignment, Alignment + Offset);
9893 /// \brief Check if this slice can be rewritten with legal operations.
9894 bool isLegal() const {
9895 // An invalid slice is not legal.
9896 if (!Origin || !Inst || !DAG)
9899 // Offsets are for indexed load only, we do not handle that.
9900 if (Origin->getOffset().getOpcode() != ISD::UNDEF)
9903 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
9905 // Check that the type is legal.
9906 EVT SliceType = getLoadedType();
9907 if (!TLI.isTypeLegal(SliceType))
9910 // Check that the load is legal for this type.
9911 if (!TLI.isOperationLegal(ISD::LOAD, SliceType))
9914 // Check that the offset can be computed.
9915 // 1. Check its type.
9916 EVT PtrType = Origin->getBasePtr().getValueType();
9917 if (PtrType == MVT::Untyped || PtrType.isExtended())
9920 // 2. Check that it fits in the immediate.
9921 if (!TLI.isLegalAddImmediate(getOffsetFromBase()))
9924 // 3. Check that the computation is legal.
9925 if (!TLI.isOperationLegal(ISD::ADD, PtrType))
9928 // Check that the zext is legal if it needs one.
9929 EVT TruncateType = Inst->getValueType(0);
9930 if (TruncateType != SliceType &&
9931 !TLI.isOperationLegal(ISD::ZERO_EXTEND, TruncateType))
9937 /// \brief Get the offset in bytes of this slice in the original chunk of
9939 /// \pre DAG != nullptr.
9940 uint64_t getOffsetFromBase() const {
9941 assert(DAG && "Missing context.");
9942 bool IsBigEndian = DAG->getDataLayout().isBigEndian();
9943 assert(!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported.");
9944 uint64_t Offset = Shift / 8;
9945 unsigned TySizeInBytes = Origin->getValueSizeInBits(0) / 8;
9946 assert(!(Origin->getValueSizeInBits(0) & 0x7) &&
9947 "The size of the original loaded type is not a multiple of a"
9949 // If Offset is bigger than TySizeInBytes, it means we are loading all
9950 // zeros. This should have been optimized before in the process.
9951 assert(TySizeInBytes > Offset &&
9952 "Invalid shift amount for given loaded size");
9954 Offset = TySizeInBytes - Offset - getLoadedSize();
9958 /// \brief Generate the sequence of instructions to load the slice
9959 /// represented by this object and redirect the uses of this slice to
9960 /// this new sequence of instructions.
9961 /// \pre this->Inst && this->Origin are valid Instructions and this
9962 /// object passed the legal check: LoadedSlice::isLegal returned true.
9963 /// \return The last instruction of the sequence used to load the slice.
9964 SDValue loadSlice() const {
9965 assert(Inst && Origin && "Unable to replace a non-existing slice.");
9966 const SDValue &OldBaseAddr = Origin->getBasePtr();
9967 SDValue BaseAddr = OldBaseAddr;
9968 // Get the offset in that chunk of bytes w.r.t. the endianess.
9969 int64_t Offset = static_cast<int64_t>(getOffsetFromBase());
9970 assert(Offset >= 0 && "Offset too big to fit in int64_t!");
9972 // BaseAddr = BaseAddr + Offset.
9973 EVT ArithType = BaseAddr.getValueType();
9975 BaseAddr = DAG->getNode(ISD::ADD, DL, ArithType, BaseAddr,
9976 DAG->getConstant(Offset, DL, ArithType));
9979 // Create the type of the loaded slice according to its size.
9980 EVT SliceType = getLoadedType();
9982 // Create the load for the slice.
9983 SDValue LastInst = DAG->getLoad(
9984 SliceType, SDLoc(Origin), Origin->getChain(), BaseAddr,
9985 Origin->getPointerInfo().getWithOffset(Offset), Origin->isVolatile(),
9986 Origin->isNonTemporal(), Origin->isInvariant(), getAlignment());
9987 // If the final type is not the same as the loaded type, this means that
9988 // we have to pad with zero. Create a zero extend for that.
9989 EVT FinalType = Inst->getValueType(0);
9990 if (SliceType != FinalType)
9992 DAG->getNode(ISD::ZERO_EXTEND, SDLoc(LastInst), FinalType, LastInst);
9996 /// \brief Check if this slice can be merged with an expensive cross register
9997 /// bank copy. E.g.,
9999 /// f = bitcast i32 i to float
10000 bool canMergeExpensiveCrossRegisterBankCopy() const {
10001 if (!Inst || !Inst->hasOneUse())
10003 SDNode *Use = *Inst->use_begin();
10004 if (Use->getOpcode() != ISD::BITCAST)
10006 assert(DAG && "Missing context");
10007 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
10008 EVT ResVT = Use->getValueType(0);
10009 const TargetRegisterClass *ResRC = TLI.getRegClassFor(ResVT.getSimpleVT());
10010 const TargetRegisterClass *ArgRC =
10011 TLI.getRegClassFor(Use->getOperand(0).getValueType().getSimpleVT());
10012 if (ArgRC == ResRC || !TLI.isOperationLegal(ISD::LOAD, ResVT))
10015 // At this point, we know that we perform a cross-register-bank copy.
10016 // Check if it is expensive.
10017 const TargetRegisterInfo *TRI = DAG->getSubtarget().getRegisterInfo();
10018 // Assume bitcasts are cheap, unless both register classes do not
10019 // explicitly share a common sub class.
10020 if (!TRI || TRI->getCommonSubClass(ArgRC, ResRC))
10023 // Check if it will be merged with the load.
10024 // 1. Check the alignment constraint.
10025 unsigned RequiredAlignment = DAG->getDataLayout().getABITypeAlignment(
10026 ResVT.getTypeForEVT(*DAG->getContext()));
10028 if (RequiredAlignment > getAlignment())
10031 // 2. Check that the load is a legal operation for that type.
10032 if (!TLI.isOperationLegal(ISD::LOAD, ResVT))
10035 // 3. Check that we do not have a zext in the way.
10036 if (Inst->getValueType(0) != getLoadedType())
10044 /// \brief Check that all bits set in \p UsedBits form a dense region, i.e.,
10045 /// \p UsedBits looks like 0..0 1..1 0..0.
10046 static bool areUsedBitsDense(const APInt &UsedBits) {
10047 // If all the bits are one, this is dense!
10048 if (UsedBits.isAllOnesValue())
10051 // Get rid of the unused bits on the right.
10052 APInt NarrowedUsedBits = UsedBits.lshr(UsedBits.countTrailingZeros());
10053 // Get rid of the unused bits on the left.
10054 if (NarrowedUsedBits.countLeadingZeros())
10055 NarrowedUsedBits = NarrowedUsedBits.trunc(NarrowedUsedBits.getActiveBits());
10056 // Check that the chunk of bits is completely used.
10057 return NarrowedUsedBits.isAllOnesValue();
10060 /// \brief Check whether or not \p First and \p Second are next to each other
10061 /// in memory. This means that there is no hole between the bits loaded
10062 /// by \p First and the bits loaded by \p Second.
10063 static bool areSlicesNextToEachOther(const LoadedSlice &First,
10064 const LoadedSlice &Second) {
10065 assert(First.Origin == Second.Origin && First.Origin &&
10066 "Unable to match different memory origins.");
10067 APInt UsedBits = First.getUsedBits();
10068 assert((UsedBits & Second.getUsedBits()) == 0 &&
10069 "Slices are not supposed to overlap.");
10070 UsedBits |= Second.getUsedBits();
10071 return areUsedBitsDense(UsedBits);
10074 /// \brief Adjust the \p GlobalLSCost according to the target
10075 /// paring capabilities and the layout of the slices.
10076 /// \pre \p GlobalLSCost should account for at least as many loads as
10077 /// there is in the slices in \p LoadedSlices.
10078 static void adjustCostForPairing(SmallVectorImpl<LoadedSlice> &LoadedSlices,
10079 LoadedSlice::Cost &GlobalLSCost) {
10080 unsigned NumberOfSlices = LoadedSlices.size();
10081 // If there is less than 2 elements, no pairing is possible.
10082 if (NumberOfSlices < 2)
10085 // Sort the slices so that elements that are likely to be next to each
10086 // other in memory are next to each other in the list.
10087 std::sort(LoadedSlices.begin(), LoadedSlices.end(),
10088 [](const LoadedSlice &LHS, const LoadedSlice &RHS) {
10089 assert(LHS.Origin == RHS.Origin && "Different bases not implemented.");
10090 return LHS.getOffsetFromBase() < RHS.getOffsetFromBase();
10092 const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo();
10093 // First (resp. Second) is the first (resp. Second) potentially candidate
10094 // to be placed in a paired load.
10095 const LoadedSlice *First = nullptr;
10096 const LoadedSlice *Second = nullptr;
10097 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice,
10098 // Set the beginning of the pair.
10101 Second = &LoadedSlices[CurrSlice];
10103 // If First is NULL, it means we start a new pair.
10104 // Get to the next slice.
10108 EVT LoadedType = First->getLoadedType();
10110 // If the types of the slices are different, we cannot pair them.
10111 if (LoadedType != Second->getLoadedType())
10114 // Check if the target supplies paired loads for this type.
10115 unsigned RequiredAlignment = 0;
10116 if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) {
10117 // move to the next pair, this type is hopeless.
10121 // Check if we meet the alignment requirement.
10122 if (RequiredAlignment > First->getAlignment())
10125 // Check that both loads are next to each other in memory.
10126 if (!areSlicesNextToEachOther(*First, *Second))
10129 assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!");
10130 --GlobalLSCost.Loads;
10131 // Move to the next pair.
10136 /// \brief Check the profitability of all involved LoadedSlice.
10137 /// Currently, it is considered profitable if there is exactly two
10138 /// involved slices (1) which are (2) next to each other in memory, and
10139 /// whose cost (\see LoadedSlice::Cost) is smaller than the original load (3).
10141 /// Note: The order of the elements in \p LoadedSlices may be modified, but not
10142 /// the elements themselves.
10144 /// FIXME: When the cost model will be mature enough, we can relax
10145 /// constraints (1) and (2).
10146 static bool isSlicingProfitable(SmallVectorImpl<LoadedSlice> &LoadedSlices,
10147 const APInt &UsedBits, bool ForCodeSize) {
10148 unsigned NumberOfSlices = LoadedSlices.size();
10149 if (StressLoadSlicing)
10150 return NumberOfSlices > 1;
10153 if (NumberOfSlices != 2)
10157 if (!areUsedBitsDense(UsedBits))
10161 LoadedSlice::Cost OrigCost(ForCodeSize), GlobalSlicingCost(ForCodeSize);
10162 // The original code has one big load.
10163 OrigCost.Loads = 1;
10164 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice) {
10165 const LoadedSlice &LS = LoadedSlices[CurrSlice];
10166 // Accumulate the cost of all the slices.
10167 LoadedSlice::Cost SliceCost(LS, ForCodeSize);
10168 GlobalSlicingCost += SliceCost;
10170 // Account as cost in the original configuration the gain obtained
10171 // with the current slices.
10172 OrigCost.addSliceGain(LS);
10175 // If the target supports paired load, adjust the cost accordingly.
10176 adjustCostForPairing(LoadedSlices, GlobalSlicingCost);
10177 return OrigCost > GlobalSlicingCost;
10180 /// \brief If the given load, \p LI, is used only by trunc or trunc(lshr)
10181 /// operations, split it in the various pieces being extracted.
10183 /// This sort of thing is introduced by SROA.
10184 /// This slicing takes care not to insert overlapping loads.
10185 /// \pre LI is a simple load (i.e., not an atomic or volatile load).
10186 bool DAGCombiner::SliceUpLoad(SDNode *N) {
10187 if (Level < AfterLegalizeDAG)
10190 LoadSDNode *LD = cast<LoadSDNode>(N);
10191 if (LD->isVolatile() || !ISD::isNormalLoad(LD) ||
10192 !LD->getValueType(0).isInteger())
10195 // Keep track of already used bits to detect overlapping values.
10196 // In that case, we will just abort the transformation.
10197 APInt UsedBits(LD->getValueSizeInBits(0), 0);
10199 SmallVector<LoadedSlice, 4> LoadedSlices;
10201 // Check if this load is used as several smaller chunks of bits.
10202 // Basically, look for uses in trunc or trunc(lshr) and record a new chain
10203 // of computation for each trunc.
10204 for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end();
10205 UI != UIEnd; ++UI) {
10206 // Skip the uses of the chain.
10207 if (UI.getUse().getResNo() != 0)
10210 SDNode *User = *UI;
10211 unsigned Shift = 0;
10213 // Check if this is a trunc(lshr).
10214 if (User->getOpcode() == ISD::SRL && User->hasOneUse() &&
10215 isa<ConstantSDNode>(User->getOperand(1))) {
10216 Shift = cast<ConstantSDNode>(User->getOperand(1))->getZExtValue();
10217 User = *User->use_begin();
10220 // At this point, User is a Truncate, iff we encountered, trunc or
10222 if (User->getOpcode() != ISD::TRUNCATE)
10225 // The width of the type must be a power of 2 and greater than 8-bits.
10226 // Otherwise the load cannot be represented in LLVM IR.
10227 // Moreover, if we shifted with a non-8-bits multiple, the slice
10228 // will be across several bytes. We do not support that.
10229 unsigned Width = User->getValueSizeInBits(0);
10230 if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7))
10233 // Build the slice for this chain of computations.
10234 LoadedSlice LS(User, LD, Shift, &DAG);
10235 APInt CurrentUsedBits = LS.getUsedBits();
10237 // Check if this slice overlaps with another.
10238 if ((CurrentUsedBits & UsedBits) != 0)
10240 // Update the bits used globally.
10241 UsedBits |= CurrentUsedBits;
10243 // Check if the new slice would be legal.
10247 // Record the slice.
10248 LoadedSlices.push_back(LS);
10251 // Abort slicing if it does not seem to be profitable.
10252 if (!isSlicingProfitable(LoadedSlices, UsedBits, ForCodeSize))
10257 // Rewrite each chain to use an independent load.
10258 // By construction, each chain can be represented by a unique load.
10260 // Prepare the argument for the new token factor for all the slices.
10261 SmallVector<SDValue, 8> ArgChains;
10262 for (SmallVectorImpl<LoadedSlice>::const_iterator
10263 LSIt = LoadedSlices.begin(),
10264 LSItEnd = LoadedSlices.end();
10265 LSIt != LSItEnd; ++LSIt) {
10266 SDValue SliceInst = LSIt->loadSlice();
10267 CombineTo(LSIt->Inst, SliceInst, true);
10268 if (SliceInst.getNode()->getOpcode() != ISD::LOAD)
10269 SliceInst = SliceInst.getOperand(0);
10270 assert(SliceInst->getOpcode() == ISD::LOAD &&
10271 "It takes more than a zext to get to the loaded slice!!");
10272 ArgChains.push_back(SliceInst.getValue(1));
10275 SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other,
10277 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
10281 /// Check to see if V is (and load (ptr), imm), where the load is having
10282 /// specific bytes cleared out. If so, return the byte size being masked out
10283 /// and the shift amount.
10284 static std::pair<unsigned, unsigned>
10285 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
10286 std::pair<unsigned, unsigned> Result(0, 0);
10288 // Check for the structure we're looking for.
10289 if (V->getOpcode() != ISD::AND ||
10290 !isa<ConstantSDNode>(V->getOperand(1)) ||
10291 !ISD::isNormalLoad(V->getOperand(0).getNode()))
10294 // Check the chain and pointer.
10295 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
10296 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
10298 // The store should be chained directly to the load or be an operand of a
10300 if (LD == Chain.getNode())
10302 else if (Chain->getOpcode() != ISD::TokenFactor)
10303 return Result; // Fail.
10306 for (const SDValue &ChainOp : Chain->op_values())
10307 if (ChainOp.getNode() == LD) {
10311 if (!isOk) return Result;
10314 // This only handles simple types.
10315 if (V.getValueType() != MVT::i16 &&
10316 V.getValueType() != MVT::i32 &&
10317 V.getValueType() != MVT::i64)
10320 // Check the constant mask. Invert it so that the bits being masked out are
10321 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
10322 // follow the sign bit for uniformity.
10323 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
10324 unsigned NotMaskLZ = countLeadingZeros(NotMask);
10325 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
10326 unsigned NotMaskTZ = countTrailingZeros(NotMask);
10327 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
10328 if (NotMaskLZ == 64) return Result; // All zero mask.
10330 // See if we have a continuous run of bits. If so, we have 0*1+0*
10331 if (countTrailingOnes(NotMask >> NotMaskTZ) + NotMaskTZ + NotMaskLZ != 64)
10334 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
10335 if (V.getValueType() != MVT::i64 && NotMaskLZ)
10336 NotMaskLZ -= 64-V.getValueSizeInBits();
10338 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
10339 switch (MaskedBytes) {
10343 default: return Result; // All one mask, or 5-byte mask.
10346 // Verify that the first bit starts at a multiple of mask so that the access
10347 // is aligned the same as the access width.
10348 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
10350 Result.first = MaskedBytes;
10351 Result.second = NotMaskTZ/8;
10356 /// Check to see if IVal is something that provides a value as specified by
10357 /// MaskInfo. If so, replace the specified store with a narrower store of
10358 /// truncated IVal.
10360 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
10361 SDValue IVal, StoreSDNode *St,
10363 unsigned NumBytes = MaskInfo.first;
10364 unsigned ByteShift = MaskInfo.second;
10365 SelectionDAG &DAG = DC->getDAG();
10367 // Check to see if IVal is all zeros in the part being masked in by the 'or'
10368 // that uses this. If not, this is not a replacement.
10369 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
10370 ByteShift*8, (ByteShift+NumBytes)*8);
10371 if (!DAG.MaskedValueIsZero(IVal, Mask)) return nullptr;
10373 // Check that it is legal on the target to do this. It is legal if the new
10374 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
10376 MVT VT = MVT::getIntegerVT(NumBytes*8);
10377 if (!DC->isTypeLegal(VT))
10380 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
10381 // shifted by ByteShift and truncated down to NumBytes.
10384 IVal = DAG.getNode(ISD::SRL, DL, IVal.getValueType(), IVal,
10385 DAG.getConstant(ByteShift*8, DL,
10386 DC->getShiftAmountTy(IVal.getValueType())));
10389 // Figure out the offset for the store and the alignment of the access.
10391 unsigned NewAlign = St->getAlignment();
10393 if (DAG.getDataLayout().isLittleEndian())
10394 StOffset = ByteShift;
10396 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
10398 SDValue Ptr = St->getBasePtr();
10401 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(),
10402 Ptr, DAG.getConstant(StOffset, DL, Ptr.getValueType()));
10403 NewAlign = MinAlign(NewAlign, StOffset);
10406 // Truncate down to the new size.
10407 IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal);
10410 return DAG.getStore(St->getChain(), SDLoc(St), IVal, Ptr,
10411 St->getPointerInfo().getWithOffset(StOffset),
10412 false, false, NewAlign).getNode();
10416 /// Look for sequence of load / op / store where op is one of 'or', 'xor', and
10417 /// 'and' of immediates. If 'op' is only touching some of the loaded bits, try
10418 /// narrowing the load and store if it would end up being a win for performance
10420 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
10421 StoreSDNode *ST = cast<StoreSDNode>(N);
10422 if (ST->isVolatile())
10425 SDValue Chain = ST->getChain();
10426 SDValue Value = ST->getValue();
10427 SDValue Ptr = ST->getBasePtr();
10428 EVT VT = Value.getValueType();
10430 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
10433 unsigned Opc = Value.getOpcode();
10435 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
10436 // is a byte mask indicating a consecutive number of bytes, check to see if
10437 // Y is known to provide just those bytes. If so, we try to replace the
10438 // load + replace + store sequence with a single (narrower) store, which makes
10440 if (Opc == ISD::OR) {
10441 std::pair<unsigned, unsigned> MaskedLoad;
10442 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
10443 if (MaskedLoad.first)
10444 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
10445 Value.getOperand(1), ST,this))
10446 return SDValue(NewST, 0);
10448 // Or is commutative, so try swapping X and Y.
10449 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
10450 if (MaskedLoad.first)
10451 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
10452 Value.getOperand(0), ST,this))
10453 return SDValue(NewST, 0);
10456 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
10457 Value.getOperand(1).getOpcode() != ISD::Constant)
10460 SDValue N0 = Value.getOperand(0);
10461 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
10462 Chain == SDValue(N0.getNode(), 1)) {
10463 LoadSDNode *LD = cast<LoadSDNode>(N0);
10464 if (LD->getBasePtr() != Ptr ||
10465 LD->getPointerInfo().getAddrSpace() !=
10466 ST->getPointerInfo().getAddrSpace())
10469 // Find the type to narrow it the load / op / store to.
10470 SDValue N1 = Value.getOperand(1);
10471 unsigned BitWidth = N1.getValueSizeInBits();
10472 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
10473 if (Opc == ISD::AND)
10474 Imm ^= APInt::getAllOnesValue(BitWidth);
10475 if (Imm == 0 || Imm.isAllOnesValue())
10477 unsigned ShAmt = Imm.countTrailingZeros();
10478 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
10479 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
10480 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
10481 // The narrowing should be profitable, the load/store operation should be
10482 // legal (or custom) and the store size should be equal to the NewVT width.
10483 while (NewBW < BitWidth &&
10484 (NewVT.getStoreSizeInBits() != NewBW ||
10485 !TLI.isOperationLegalOrCustom(Opc, NewVT) ||
10486 !TLI.isNarrowingProfitable(VT, NewVT))) {
10487 NewBW = NextPowerOf2(NewBW);
10488 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
10490 if (NewBW >= BitWidth)
10493 // If the lsb changed does not start at the type bitwidth boundary,
10494 // start at the previous one.
10496 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
10497 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt,
10498 std::min(BitWidth, ShAmt + NewBW));
10499 if ((Imm & Mask) == Imm) {
10500 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
10501 if (Opc == ISD::AND)
10502 NewImm ^= APInt::getAllOnesValue(NewBW);
10503 uint64_t PtrOff = ShAmt / 8;
10504 // For big endian targets, we need to adjust the offset to the pointer to
10505 // load the correct bytes.
10506 if (DAG.getDataLayout().isBigEndian())
10507 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
10509 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
10510 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
10511 if (NewAlign < DAG.getDataLayout().getABITypeAlignment(NewVTTy))
10514 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD),
10515 Ptr.getValueType(), Ptr,
10516 DAG.getConstant(PtrOff, SDLoc(LD),
10517 Ptr.getValueType()));
10518 SDValue NewLD = DAG.getLoad(NewVT, SDLoc(N0),
10519 LD->getChain(), NewPtr,
10520 LD->getPointerInfo().getWithOffset(PtrOff),
10521 LD->isVolatile(), LD->isNonTemporal(),
10522 LD->isInvariant(), NewAlign,
10524 SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD,
10525 DAG.getConstant(NewImm, SDLoc(Value),
10527 SDValue NewST = DAG.getStore(Chain, SDLoc(N),
10529 ST->getPointerInfo().getWithOffset(PtrOff),
10530 false, false, NewAlign);
10532 AddToWorklist(NewPtr.getNode());
10533 AddToWorklist(NewLD.getNode());
10534 AddToWorklist(NewVal.getNode());
10535 WorklistRemover DeadNodes(*this);
10536 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
10545 /// For a given floating point load / store pair, if the load value isn't used
10546 /// by any other operations, then consider transforming the pair to integer
10547 /// load / store operations if the target deems the transformation profitable.
10548 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
10549 StoreSDNode *ST = cast<StoreSDNode>(N);
10550 SDValue Chain = ST->getChain();
10551 SDValue Value = ST->getValue();
10552 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
10553 Value.hasOneUse() &&
10554 Chain == SDValue(Value.getNode(), 1)) {
10555 LoadSDNode *LD = cast<LoadSDNode>(Value);
10556 EVT VT = LD->getMemoryVT();
10557 if (!VT.isFloatingPoint() ||
10558 VT != ST->getMemoryVT() ||
10559 LD->isNonTemporal() ||
10560 ST->isNonTemporal() ||
10561 LD->getPointerInfo().getAddrSpace() != 0 ||
10562 ST->getPointerInfo().getAddrSpace() != 0)
10565 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
10566 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
10567 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
10568 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
10569 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
10572 unsigned LDAlign = LD->getAlignment();
10573 unsigned STAlign = ST->getAlignment();
10574 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
10575 unsigned ABIAlign = DAG.getDataLayout().getABITypeAlignment(IntVTTy);
10576 if (LDAlign < ABIAlign || STAlign < ABIAlign)
10579 SDValue NewLD = DAG.getLoad(IntVT, SDLoc(Value),
10580 LD->getChain(), LD->getBasePtr(),
10581 LD->getPointerInfo(),
10582 false, false, false, LDAlign);
10584 SDValue NewST = DAG.getStore(NewLD.getValue(1), SDLoc(N),
10585 NewLD, ST->getBasePtr(),
10586 ST->getPointerInfo(),
10587 false, false, STAlign);
10589 AddToWorklist(NewLD.getNode());
10590 AddToWorklist(NewST.getNode());
10591 WorklistRemover DeadNodes(*this);
10592 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
10601 /// Helper struct to parse and store a memory address as base + index + offset.
10602 /// We ignore sign extensions when it is safe to do so.
10603 /// The following two expressions are not equivalent. To differentiate we need
10604 /// to store whether there was a sign extension involved in the index
10606 /// (load (i64 add (i64 copyfromreg %c)
10607 /// (i64 signextend (add (i8 load %index)
10611 /// (load (i64 add (i64 copyfromreg %c)
10612 /// (i64 signextend (i32 add (i32 signextend (i8 load %index))
10614 struct BaseIndexOffset {
10618 bool IsIndexSignExt;
10620 BaseIndexOffset() : Offset(0), IsIndexSignExt(false) {}
10622 BaseIndexOffset(SDValue Base, SDValue Index, int64_t Offset,
10623 bool IsIndexSignExt) :
10624 Base(Base), Index(Index), Offset(Offset), IsIndexSignExt(IsIndexSignExt) {}
10626 bool equalBaseIndex(const BaseIndexOffset &Other) {
10627 return Other.Base == Base && Other.Index == Index &&
10628 Other.IsIndexSignExt == IsIndexSignExt;
10631 /// Parses tree in Ptr for base, index, offset addresses.
10632 static BaseIndexOffset match(SDValue Ptr) {
10633 bool IsIndexSignExt = false;
10635 // We only can pattern match BASE + INDEX + OFFSET. If Ptr is not an ADD
10636 // instruction, then it could be just the BASE or everything else we don't
10637 // know how to handle. Just use Ptr as BASE and give up.
10638 if (Ptr->getOpcode() != ISD::ADD)
10639 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
10641 // We know that we have at least an ADD instruction. Try to pattern match
10642 // the simple case of BASE + OFFSET.
10643 if (isa<ConstantSDNode>(Ptr->getOperand(1))) {
10644 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
10645 return BaseIndexOffset(Ptr->getOperand(0), SDValue(), Offset,
10649 // Inside a loop the current BASE pointer is calculated using an ADD and a
10650 // MUL instruction. In this case Ptr is the actual BASE pointer.
10651 // (i64 add (i64 %array_ptr)
10652 // (i64 mul (i64 %induction_var)
10653 // (i64 %element_size)))
10654 if (Ptr->getOperand(1)->getOpcode() == ISD::MUL)
10655 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
10657 // Look at Base + Index + Offset cases.
10658 SDValue Base = Ptr->getOperand(0);
10659 SDValue IndexOffset = Ptr->getOperand(1);
10661 // Skip signextends.
10662 if (IndexOffset->getOpcode() == ISD::SIGN_EXTEND) {
10663 IndexOffset = IndexOffset->getOperand(0);
10664 IsIndexSignExt = true;
10667 // Either the case of Base + Index (no offset) or something else.
10668 if (IndexOffset->getOpcode() != ISD::ADD)
10669 return BaseIndexOffset(Base, IndexOffset, 0, IsIndexSignExt);
10671 // Now we have the case of Base + Index + offset.
10672 SDValue Index = IndexOffset->getOperand(0);
10673 SDValue Offset = IndexOffset->getOperand(1);
10675 if (!isa<ConstantSDNode>(Offset))
10676 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
10678 // Ignore signextends.
10679 if (Index->getOpcode() == ISD::SIGN_EXTEND) {
10680 Index = Index->getOperand(0);
10681 IsIndexSignExt = true;
10682 } else IsIndexSignExt = false;
10684 int64_t Off = cast<ConstantSDNode>(Offset)->getSExtValue();
10685 return BaseIndexOffset(Base, Index, Off, IsIndexSignExt);
10690 SDValue DAGCombiner::getMergedConstantVectorStore(SelectionDAG &DAG,
10692 ArrayRef<MemOpLink> Stores,
10694 SmallVector<SDValue, 8> BuildVector;
10696 for (unsigned I = 0, E = Ty.getVectorNumElements(); I != E; ++I)
10697 BuildVector.push_back(cast<StoreSDNode>(Stores[I].MemNode)->getValue());
10699 return DAG.getNode(ISD::BUILD_VECTOR, SL, Ty, BuildVector);
10702 bool DAGCombiner::MergeStoresOfConstantsOrVecElts(
10703 SmallVectorImpl<MemOpLink> &StoreNodes, EVT MemVT,
10704 unsigned NumElem, bool IsConstantSrc, bool UseVector) {
10705 // Make sure we have something to merge.
10709 int64_t ElementSizeBytes = MemVT.getSizeInBits() / 8;
10710 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
10711 unsigned LatestNodeUsed = 0;
10713 for (unsigned i=0; i < NumElem; ++i) {
10714 // Find a chain for the new wide-store operand. Notice that some
10715 // of the store nodes that we found may not be selected for inclusion
10716 // in the wide store. The chain we use needs to be the chain of the
10717 // latest store node which is *used* and replaced by the wide store.
10718 if (StoreNodes[i].SequenceNum < StoreNodes[LatestNodeUsed].SequenceNum)
10719 LatestNodeUsed = i;
10722 // The latest Node in the DAG.
10723 LSBaseSDNode *LatestOp = StoreNodes[LatestNodeUsed].MemNode;
10724 SDLoc DL(StoreNodes[0].MemNode);
10728 // Find a legal type for the vector store.
10729 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
10730 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
10731 if (IsConstantSrc) {
10732 StoredVal = getMergedConstantVectorStore(DAG, DL, StoreNodes, Ty);
10734 SmallVector<SDValue, 8> Ops;
10735 for (unsigned i = 0; i < NumElem ; ++i) {
10736 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10737 SDValue Val = St->getValue();
10738 // All of the operands of a BUILD_VECTOR must have the same type.
10739 if (Val.getValueType() != MemVT)
10741 Ops.push_back(Val);
10744 // Build the extracted vector elements back into a vector.
10745 StoredVal = DAG.getNode(ISD::BUILD_VECTOR, DL, Ty, Ops);
10748 // We should always use a vector store when merging extracted vector
10749 // elements, so this path implies a store of constants.
10750 assert(IsConstantSrc && "Merged vector elements should use vector store");
10752 unsigned SizeInBits = NumElem * ElementSizeBytes * 8;
10753 APInt StoreInt(SizeInBits, 0);
10755 // Construct a single integer constant which is made of the smaller
10756 // constant inputs.
10757 bool IsLE = DAG.getDataLayout().isLittleEndian();
10758 for (unsigned i = 0; i < NumElem ; ++i) {
10759 unsigned Idx = IsLE ? (NumElem - 1 - i) : i;
10760 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
10761 SDValue Val = St->getValue();
10762 StoreInt <<= ElementSizeBytes * 8;
10763 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
10764 StoreInt |= C->getAPIntValue().zext(SizeInBits);
10765 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
10766 StoreInt |= C->getValueAPF().bitcastToAPInt().zext(SizeInBits);
10768 llvm_unreachable("Invalid constant element type");
10772 // Create the new Load and Store operations.
10773 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), SizeInBits);
10774 StoredVal = DAG.getConstant(StoreInt, DL, StoreTy);
10777 SDValue NewStore = DAG.getStore(LatestOp->getChain(), DL, StoredVal,
10778 FirstInChain->getBasePtr(),
10779 FirstInChain->getPointerInfo(),
10781 FirstInChain->getAlignment());
10783 // Replace the last store with the new store
10784 CombineTo(LatestOp, NewStore);
10785 // Erase all other stores.
10786 for (unsigned i = 0; i < NumElem ; ++i) {
10787 if (StoreNodes[i].MemNode == LatestOp)
10789 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10790 // ReplaceAllUsesWith will replace all uses that existed when it was
10791 // called, but graph optimizations may cause new ones to appear. For
10792 // example, the case in pr14333 looks like
10794 // St's chain -> St -> another store -> X
10796 // And the only difference from St to the other store is the chain.
10797 // When we change it's chain to be St's chain they become identical,
10798 // get CSEed and the net result is that X is now a use of St.
10799 // Since we know that St is redundant, just iterate.
10800 while (!St->use_empty())
10801 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain());
10802 deleteAndRecombine(St);
10808 void DAGCombiner::getStoreMergeAndAliasCandidates(
10809 StoreSDNode* St, SmallVectorImpl<MemOpLink> &StoreNodes,
10810 SmallVectorImpl<LSBaseSDNode*> &AliasLoadNodes) {
10811 // This holds the base pointer, index, and the offset in bytes from the base
10813 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr());
10815 // We must have a base and an offset.
10816 if (!BasePtr.Base.getNode())
10819 // Do not handle stores to undef base pointers.
10820 if (BasePtr.Base.getOpcode() == ISD::UNDEF)
10823 // Walk up the chain and look for nodes with offsets from the same
10824 // base pointer. Stop when reaching an instruction with a different kind
10825 // or instruction which has a different base pointer.
10826 EVT MemVT = St->getMemoryVT();
10828 StoreSDNode *Index = St;
10830 // If the chain has more than one use, then we can't reorder the mem ops.
10831 if (Index != St && !SDValue(Index, 0)->hasOneUse())
10834 // Find the base pointer and offset for this memory node.
10835 BaseIndexOffset Ptr = BaseIndexOffset::match(Index->getBasePtr());
10837 // Check that the base pointer is the same as the original one.
10838 if (!Ptr.equalBaseIndex(BasePtr))
10841 // The memory operands must not be volatile.
10842 if (Index->isVolatile() || Index->isIndexed())
10846 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
10847 if (St->isTruncatingStore())
10850 // The stored memory type must be the same.
10851 if (Index->getMemoryVT() != MemVT)
10854 // We found a potential memory operand to merge.
10855 StoreNodes.push_back(MemOpLink(Index, Ptr.Offset, Seq++));
10857 // Find the next memory operand in the chain. If the next operand in the
10858 // chain is a store then move up and continue the scan with the next
10859 // memory operand. If the next operand is a load save it and use alias
10860 // information to check if it interferes with anything.
10861 SDNode *NextInChain = Index->getChain().getNode();
10863 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
10864 // We found a store node. Use it for the next iteration.
10867 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
10868 if (Ldn->isVolatile()) {
10873 // Save the load node for later. Continue the scan.
10874 AliasLoadNodes.push_back(Ldn);
10875 NextInChain = Ldn->getChain().getNode();
10885 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
10886 if (OptLevel == CodeGenOpt::None)
10889 EVT MemVT = St->getMemoryVT();
10890 int64_t ElementSizeBytes = MemVT.getSizeInBits() / 8;
10891 bool NoVectors = DAG.getMachineFunction().getFunction()->hasFnAttribute(
10892 Attribute::NoImplicitFloat);
10894 // This function cannot currently deal with non-byte-sized memory sizes.
10895 if (ElementSizeBytes * 8 != MemVT.getSizeInBits())
10898 // Don't merge vectors into wider inputs.
10899 if (MemVT.isVector() || !MemVT.isSimple())
10902 // Perform an early exit check. Do not bother looking at stored values that
10903 // are not constants, loads, or extracted vector elements.
10904 SDValue StoredVal = St->getValue();
10905 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
10906 bool IsConstantSrc = isa<ConstantSDNode>(StoredVal) ||
10907 isa<ConstantFPSDNode>(StoredVal);
10908 bool IsExtractVecEltSrc = (StoredVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT);
10910 if (!IsConstantSrc && !IsLoadSrc && !IsExtractVecEltSrc)
10913 // Only look at ends of store sequences.
10914 SDValue Chain = SDValue(St, 0);
10915 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
10918 // Save the LoadSDNodes that we find in the chain.
10919 // We need to make sure that these nodes do not interfere with
10920 // any of the store nodes.
10921 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
10923 // Save the StoreSDNodes that we find in the chain.
10924 SmallVector<MemOpLink, 8> StoreNodes;
10926 getStoreMergeAndAliasCandidates(St, StoreNodes, AliasLoadNodes);
10928 // Check if there is anything to merge.
10929 if (StoreNodes.size() < 2)
10932 // Sort the memory operands according to their distance from the base pointer.
10933 std::sort(StoreNodes.begin(), StoreNodes.end(),
10934 [](MemOpLink LHS, MemOpLink RHS) {
10935 return LHS.OffsetFromBase < RHS.OffsetFromBase ||
10936 (LHS.OffsetFromBase == RHS.OffsetFromBase &&
10937 LHS.SequenceNum > RHS.SequenceNum);
10940 // Scan the memory operations on the chain and find the first non-consecutive
10941 // store memory address.
10942 unsigned LastConsecutiveStore = 0;
10943 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
10944 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) {
10946 // Check that the addresses are consecutive starting from the second
10947 // element in the list of stores.
10949 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
10950 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
10954 bool Alias = false;
10955 // Check if this store interferes with any of the loads that we found.
10956 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld)
10957 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) {
10961 // We found a load that alias with this store. Stop the sequence.
10965 // Mark this node as useful.
10966 LastConsecutiveStore = i;
10969 // The node with the lowest store address.
10970 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
10971 unsigned FirstStoreAS = FirstInChain->getAddressSpace();
10972 unsigned FirstStoreAlign = FirstInChain->getAlignment();
10973 LLVMContext &Context = *DAG.getContext();
10974 const DataLayout &DL = DAG.getDataLayout();
10976 // Store the constants into memory as one consecutive store.
10977 if (IsConstantSrc) {
10978 unsigned LastLegalType = 0;
10979 unsigned LastLegalVectorType = 0;
10980 bool NonZero = false;
10981 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
10982 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10983 SDValue StoredVal = St->getValue();
10985 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
10986 NonZero |= !C->isNullValue();
10987 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) {
10988 NonZero |= !C->getConstantFPValue()->isNullValue();
10994 // Find a legal type for the constant store.
10995 unsigned SizeInBits = (i+1) * ElementSizeBytes * 8;
10996 EVT StoreTy = EVT::getIntegerVT(Context, SizeInBits);
10997 if (TLI.isTypeLegal(StoreTy) &&
10998 TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstStoreAS,
10999 FirstStoreAlign)) {
11000 LastLegalType = i+1;
11001 // Or check whether a truncstore is legal.
11002 } else if (TLI.getTypeAction(Context, StoreTy) ==
11003 TargetLowering::TypePromoteInteger) {
11004 EVT LegalizedStoredValueTy =
11005 TLI.getTypeToTransformTo(Context, StoredVal.getValueType());
11006 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
11007 TLI.allowsMemoryAccess(Context, DL, LegalizedStoredValueTy,
11008 FirstStoreAS, FirstStoreAlign)) {
11009 LastLegalType = i + 1;
11013 // Find a legal type for the vector store.
11014 EVT Ty = EVT::getVectorVT(Context, MemVT, i+1);
11015 if (TLI.isTypeLegal(Ty) &&
11016 TLI.allowsMemoryAccess(Context, DL, Ty, FirstStoreAS,
11017 FirstStoreAlign)) {
11018 LastLegalVectorType = i + 1;
11023 // We only use vectors if the constant is known to be zero or the target
11024 // allows it and the function is not marked with the noimplicitfloat
11027 LastLegalVectorType = 0;
11028 } else if (NonZero && !TLI.storeOfVectorConstantIsCheap(MemVT,
11029 LastLegalVectorType,
11031 LastLegalVectorType = 0;
11034 // Check if we found a legal integer type to store.
11035 if (LastLegalType == 0 && LastLegalVectorType == 0)
11038 bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors;
11039 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType;
11041 return MergeStoresOfConstantsOrVecElts(StoreNodes, MemVT, NumElem,
11045 // When extracting multiple vector elements, try to store them
11046 // in one vector store rather than a sequence of scalar stores.
11047 if (IsExtractVecEltSrc) {
11048 unsigned NumElem = 0;
11049 for (unsigned i = 0; i < LastConsecutiveStore + 1; ++i) {
11050 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
11051 SDValue StoredVal = St->getValue();
11052 // This restriction could be loosened.
11053 // Bail out if any stored values are not elements extracted from a vector.
11054 // It should be possible to handle mixed sources, but load sources need
11055 // more careful handling (see the block of code below that handles
11056 // consecutive loads).
11057 if (StoredVal.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
11060 // Find a legal type for the vector store.
11061 EVT Ty = EVT::getVectorVT(Context, MemVT, i+1);
11062 if (TLI.isTypeLegal(Ty) &&
11063 TLI.allowsMemoryAccess(Context, DL, Ty, FirstStoreAS,
11068 return MergeStoresOfConstantsOrVecElts(StoreNodes, MemVT, NumElem,
11072 // Below we handle the case of multiple consecutive stores that
11073 // come from multiple consecutive loads. We merge them into a single
11074 // wide load and a single wide store.
11076 // Look for load nodes which are used by the stored values.
11077 SmallVector<MemOpLink, 8> LoadNodes;
11079 // Find acceptable loads. Loads need to have the same chain (token factor),
11080 // must not be zext, volatile, indexed, and they must be consecutive.
11081 BaseIndexOffset LdBasePtr;
11082 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
11083 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
11084 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
11087 // Loads must only have one use.
11088 if (!Ld->hasNUsesOfValue(1, 0))
11091 // The memory operands must not be volatile.
11092 if (Ld->isVolatile() || Ld->isIndexed())
11095 // We do not accept ext loads.
11096 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
11099 // The stored memory type must be the same.
11100 if (Ld->getMemoryVT() != MemVT)
11103 BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld->getBasePtr());
11104 // If this is not the first ptr that we check.
11105 if (LdBasePtr.Base.getNode()) {
11106 // The base ptr must be the same.
11107 if (!LdPtr.equalBaseIndex(LdBasePtr))
11110 // Check that all other base pointers are the same as this one.
11114 // We found a potential memory operand to merge.
11115 LoadNodes.push_back(MemOpLink(Ld, LdPtr.Offset, 0));
11118 if (LoadNodes.size() < 2)
11121 // If we have load/store pair instructions and we only have two values,
11123 unsigned RequiredAlignment;
11124 if (LoadNodes.size() == 2 && TLI.hasPairedLoad(MemVT, RequiredAlignment) &&
11125 St->getAlignment() >= RequiredAlignment)
11128 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
11129 unsigned FirstLoadAS = FirstLoad->getAddressSpace();
11130 unsigned FirstLoadAlign = FirstLoad->getAlignment();
11132 // Scan the memory operations on the chain and find the first non-consecutive
11133 // load memory address. These variables hold the index in the store node
11135 unsigned LastConsecutiveLoad = 0;
11136 // This variable refers to the size and not index in the array.
11137 unsigned LastLegalVectorType = 0;
11138 unsigned LastLegalIntegerType = 0;
11139 StartAddress = LoadNodes[0].OffsetFromBase;
11140 SDValue FirstChain = FirstLoad->getChain();
11141 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
11142 // All loads much share the same chain.
11143 if (LoadNodes[i].MemNode->getChain() != FirstChain)
11146 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
11147 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
11149 LastConsecutiveLoad = i;
11151 // Find a legal type for the vector store.
11152 EVT StoreTy = EVT::getVectorVT(Context, MemVT, i+1);
11153 if (TLI.isTypeLegal(StoreTy) &&
11154 TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstStoreAS,
11155 FirstStoreAlign) &&
11156 TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstLoadAS,
11158 LastLegalVectorType = i + 1;
11161 // Find a legal type for the integer store.
11162 unsigned SizeInBits = (i+1) * ElementSizeBytes * 8;
11163 StoreTy = EVT::getIntegerVT(Context, SizeInBits);
11164 if (TLI.isTypeLegal(StoreTy) &&
11165 TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstStoreAS,
11166 FirstStoreAlign) &&
11167 TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstLoadAS,
11169 LastLegalIntegerType = i + 1;
11170 // Or check whether a truncstore and extload is legal.
11171 else if (TLI.getTypeAction(Context, StoreTy) ==
11172 TargetLowering::TypePromoteInteger) {
11173 EVT LegalizedStoredValueTy =
11174 TLI.getTypeToTransformTo(Context, StoreTy);
11175 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
11176 TLI.isLoadExtLegal(ISD::ZEXTLOAD, LegalizedStoredValueTy, StoreTy) &&
11177 TLI.isLoadExtLegal(ISD::SEXTLOAD, LegalizedStoredValueTy, StoreTy) &&
11178 TLI.isLoadExtLegal(ISD::EXTLOAD, LegalizedStoredValueTy, StoreTy) &&
11179 TLI.allowsMemoryAccess(Context, DL, LegalizedStoredValueTy,
11180 FirstStoreAS, FirstStoreAlign) &&
11181 TLI.allowsMemoryAccess(Context, DL, LegalizedStoredValueTy,
11182 FirstLoadAS, FirstLoadAlign))
11183 LastLegalIntegerType = i+1;
11187 // Only use vector types if the vector type is larger than the integer type.
11188 // If they are the same, use integers.
11189 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors;
11190 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType);
11192 // We add +1 here because the LastXXX variables refer to location while
11193 // the NumElem refers to array/index size.
11194 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1;
11195 NumElem = std::min(LastLegalType, NumElem);
11200 // The latest Node in the DAG.
11201 unsigned LatestNodeUsed = 0;
11202 for (unsigned i=1; i<NumElem; ++i) {
11203 // Find a chain for the new wide-store operand. Notice that some
11204 // of the store nodes that we found may not be selected for inclusion
11205 // in the wide store. The chain we use needs to be the chain of the
11206 // latest store node which is *used* and replaced by the wide store.
11207 if (StoreNodes[i].SequenceNum < StoreNodes[LatestNodeUsed].SequenceNum)
11208 LatestNodeUsed = i;
11211 LSBaseSDNode *LatestOp = StoreNodes[LatestNodeUsed].MemNode;
11213 // Find if it is better to use vectors or integers to load and store
11217 JointMemOpVT = EVT::getVectorVT(Context, MemVT, NumElem);
11219 unsigned SizeInBits = NumElem * ElementSizeBytes * 8;
11220 JointMemOpVT = EVT::getIntegerVT(Context, SizeInBits);
11223 SDLoc LoadDL(LoadNodes[0].MemNode);
11224 SDLoc StoreDL(StoreNodes[0].MemNode);
11226 SDValue NewLoad = DAG.getLoad(
11227 JointMemOpVT, LoadDL, FirstLoad->getChain(), FirstLoad->getBasePtr(),
11228 FirstLoad->getPointerInfo(), false, false, false, FirstLoadAlign);
11230 SDValue NewStore = DAG.getStore(
11231 LatestOp->getChain(), StoreDL, NewLoad, FirstInChain->getBasePtr(),
11232 FirstInChain->getPointerInfo(), false, false, FirstStoreAlign);
11234 // Replace one of the loads with the new load.
11235 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode);
11236 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
11237 SDValue(NewLoad.getNode(), 1));
11239 // Remove the rest of the load chains.
11240 for (unsigned i = 1; i < NumElem ; ++i) {
11241 // Replace all chain users of the old load nodes with the chain of the new
11243 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
11244 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain());
11247 // Replace the last store with the new store.
11248 CombineTo(LatestOp, NewStore);
11249 // Erase all other stores.
11250 for (unsigned i = 0; i < NumElem ; ++i) {
11251 // Remove all Store nodes.
11252 if (StoreNodes[i].MemNode == LatestOp)
11254 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
11255 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain());
11256 deleteAndRecombine(St);
11262 SDValue DAGCombiner::visitSTORE(SDNode *N) {
11263 StoreSDNode *ST = cast<StoreSDNode>(N);
11264 SDValue Chain = ST->getChain();
11265 SDValue Value = ST->getValue();
11266 SDValue Ptr = ST->getBasePtr();
11268 // If this is a store of a bit convert, store the input value if the
11269 // resultant store does not need a higher alignment than the original.
11270 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
11271 ST->isUnindexed()) {
11272 unsigned OrigAlign = ST->getAlignment();
11273 EVT SVT = Value.getOperand(0).getValueType();
11274 unsigned Align = DAG.getDataLayout().getABITypeAlignment(
11275 SVT.getTypeForEVT(*DAG.getContext()));
11276 if (Align <= OrigAlign &&
11277 ((!LegalOperations && !ST->isVolatile()) ||
11278 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
11279 return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0),
11280 Ptr, ST->getPointerInfo(), ST->isVolatile(),
11281 ST->isNonTemporal(), OrigAlign,
11285 // Turn 'store undef, Ptr' -> nothing.
11286 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
11289 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
11290 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
11291 // NOTE: If the original store is volatile, this transform must not increase
11292 // the number of stores. For example, on x86-32 an f64 can be stored in one
11293 // processor operation but an i64 (which is not legal) requires two. So the
11294 // transform should not be done in this case.
11295 if (Value.getOpcode() != ISD::TargetConstantFP) {
11297 switch (CFP->getSimpleValueType(0).SimpleTy) {
11298 default: llvm_unreachable("Unknown FP type");
11299 case MVT::f16: // We don't do this for these yet.
11305 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
11306 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
11308 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
11309 bitcastToAPInt().getZExtValue(), SDLoc(CFP),
11311 return DAG.getStore(Chain, SDLoc(N), Tmp,
11312 Ptr, ST->getMemOperand());
11316 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
11317 !ST->isVolatile()) ||
11318 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
11320 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
11321 getZExtValue(), SDLoc(CFP), MVT::i64);
11322 return DAG.getStore(Chain, SDLoc(N), Tmp,
11323 Ptr, ST->getMemOperand());
11326 if (!ST->isVolatile() &&
11327 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
11328 // Many FP stores are not made apparent until after legalize, e.g. for
11329 // argument passing. Since this is so common, custom legalize the
11330 // 64-bit integer store into two 32-bit stores.
11331 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
11332 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, SDLoc(CFP), MVT::i32);
11333 SDValue Hi = DAG.getConstant(Val >> 32, SDLoc(CFP), MVT::i32);
11334 if (DAG.getDataLayout().isBigEndian())
11337 unsigned Alignment = ST->getAlignment();
11338 bool isVolatile = ST->isVolatile();
11339 bool isNonTemporal = ST->isNonTemporal();
11340 AAMDNodes AAInfo = ST->getAAInfo();
11344 SDValue St0 = DAG.getStore(Chain, SDLoc(ST), Lo,
11345 Ptr, ST->getPointerInfo(),
11346 isVolatile, isNonTemporal,
11347 ST->getAlignment(), AAInfo);
11348 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
11349 DAG.getConstant(4, DL, Ptr.getValueType()));
11350 Alignment = MinAlign(Alignment, 4U);
11351 SDValue St1 = DAG.getStore(Chain, SDLoc(ST), Hi,
11352 Ptr, ST->getPointerInfo().getWithOffset(4),
11353 isVolatile, isNonTemporal,
11354 Alignment, AAInfo);
11355 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
11364 // Try to infer better alignment information than the store already has.
11365 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
11366 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
11367 if (Align > ST->getAlignment()) {
11369 DAG.getTruncStore(Chain, SDLoc(N), Value,
11370 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
11371 ST->isVolatile(), ST->isNonTemporal(), Align,
11373 if (NewStore.getNode() != N)
11374 return CombineTo(ST, NewStore, true);
11379 // Try transforming a pair floating point load / store ops to integer
11380 // load / store ops.
11381 if (SDValue NewST = TransformFPLoadStorePair(N))
11384 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
11385 : DAG.getSubtarget().useAA();
11387 if (CombinerAAOnlyFunc.getNumOccurrences() &&
11388 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
11391 if (UseAA && ST->isUnindexed()) {
11392 // Walk up chain skipping non-aliasing memory nodes.
11393 SDValue BetterChain = FindBetterChain(N, Chain);
11395 // If there is a better chain.
11396 if (Chain != BetterChain) {
11399 // Replace the chain to avoid dependency.
11400 if (ST->isTruncatingStore()) {
11401 ReplStore = DAG.getTruncStore(BetterChain, SDLoc(N), Value, Ptr,
11402 ST->getMemoryVT(), ST->getMemOperand());
11404 ReplStore = DAG.getStore(BetterChain, SDLoc(N), Value, Ptr,
11405 ST->getMemOperand());
11408 // Create token to keep both nodes around.
11409 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
11410 MVT::Other, Chain, ReplStore);
11412 // Make sure the new and old chains are cleaned up.
11413 AddToWorklist(Token.getNode());
11415 // Don't add users to work list.
11416 return CombineTo(N, Token, false);
11420 // Try transforming N to an indexed store.
11421 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
11422 return SDValue(N, 0);
11424 // FIXME: is there such a thing as a truncating indexed store?
11425 if (ST->isTruncatingStore() && ST->isUnindexed() &&
11426 Value.getValueType().isInteger()) {
11427 // See if we can simplify the input to this truncstore with knowledge that
11428 // only the low bits are being used. For example:
11429 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
11431 GetDemandedBits(Value,
11432 APInt::getLowBitsSet(
11433 Value.getValueType().getScalarType().getSizeInBits(),
11434 ST->getMemoryVT().getScalarType().getSizeInBits()));
11435 AddToWorklist(Value.getNode());
11436 if (Shorter.getNode())
11437 return DAG.getTruncStore(Chain, SDLoc(N), Shorter,
11438 Ptr, ST->getMemoryVT(), ST->getMemOperand());
11440 // Otherwise, see if we can simplify the operation with
11441 // SimplifyDemandedBits, which only works if the value has a single use.
11442 if (SimplifyDemandedBits(Value,
11443 APInt::getLowBitsSet(
11444 Value.getValueType().getScalarType().getSizeInBits(),
11445 ST->getMemoryVT().getScalarType().getSizeInBits())))
11446 return SDValue(N, 0);
11449 // If this is a load followed by a store to the same location, then the store
11451 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
11452 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
11453 ST->isUnindexed() && !ST->isVolatile() &&
11454 // There can't be any side effects between the load and store, such as
11455 // a call or store.
11456 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
11457 // The store is dead, remove it.
11462 // If this is a store followed by a store with the same value to the same
11463 // location, then the store is dead/noop.
11464 if (StoreSDNode *ST1 = dyn_cast<StoreSDNode>(Chain)) {
11465 if (ST1->getBasePtr() == Ptr && ST->getMemoryVT() == ST1->getMemoryVT() &&
11466 ST1->getValue() == Value && ST->isUnindexed() && !ST->isVolatile() &&
11467 ST1->isUnindexed() && !ST1->isVolatile()) {
11468 // The store is dead, remove it.
11473 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
11474 // truncating store. We can do this even if this is already a truncstore.
11475 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
11476 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
11477 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
11478 ST->getMemoryVT())) {
11479 return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0),
11480 Ptr, ST->getMemoryVT(), ST->getMemOperand());
11483 // Only perform this optimization before the types are legal, because we
11484 // don't want to perform this optimization on every DAGCombine invocation.
11486 bool EverChanged = false;
11489 // There can be multiple store sequences on the same chain.
11490 // Keep trying to merge store sequences until we are unable to do so
11491 // or until we merge the last store on the chain.
11492 bool Changed = MergeConsecutiveStores(ST);
11493 EverChanged |= Changed;
11494 if (!Changed) break;
11495 } while (ST->getOpcode() != ISD::DELETED_NODE);
11498 return SDValue(N, 0);
11501 return ReduceLoadOpStoreWidth(N);
11504 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
11505 SDValue InVec = N->getOperand(0);
11506 SDValue InVal = N->getOperand(1);
11507 SDValue EltNo = N->getOperand(2);
11510 // If the inserted element is an UNDEF, just use the input vector.
11511 if (InVal.getOpcode() == ISD::UNDEF)
11514 EVT VT = InVec.getValueType();
11516 // If we can't generate a legal BUILD_VECTOR, exit
11517 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
11520 // Check that we know which element is being inserted
11521 if (!isa<ConstantSDNode>(EltNo))
11523 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
11525 // Canonicalize insert_vector_elt dag nodes.
11527 // (insert_vector_elt (insert_vector_elt A, Idx0), Idx1)
11528 // -> (insert_vector_elt (insert_vector_elt A, Idx1), Idx0)
11530 // Do this only if the child insert_vector node has one use; also
11531 // do this only if indices are both constants and Idx1 < Idx0.
11532 if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT && InVec.hasOneUse()
11533 && isa<ConstantSDNode>(InVec.getOperand(2))) {
11534 unsigned OtherElt =
11535 cast<ConstantSDNode>(InVec.getOperand(2))->getZExtValue();
11536 if (Elt < OtherElt) {
11538 SDValue NewOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(N), VT,
11539 InVec.getOperand(0), InVal, EltNo);
11540 AddToWorklist(NewOp.getNode());
11541 return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(InVec.getNode()),
11542 VT, NewOp, InVec.getOperand(1), InVec.getOperand(2));
11546 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
11547 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
11548 // vector elements.
11549 SmallVector<SDValue, 8> Ops;
11550 // Do not combine these two vectors if the output vector will not replace
11551 // the input vector.
11552 if (InVec.getOpcode() == ISD::BUILD_VECTOR && InVec.hasOneUse()) {
11553 Ops.append(InVec.getNode()->op_begin(),
11554 InVec.getNode()->op_end());
11555 } else if (InVec.getOpcode() == ISD::UNDEF) {
11556 unsigned NElts = VT.getVectorNumElements();
11557 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
11562 // Insert the element
11563 if (Elt < Ops.size()) {
11564 // All the operands of BUILD_VECTOR must have the same type;
11565 // we enforce that here.
11566 EVT OpVT = Ops[0].getValueType();
11567 if (InVal.getValueType() != OpVT)
11568 InVal = OpVT.bitsGT(InVal.getValueType()) ?
11569 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
11570 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
11574 // Return the new vector
11575 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
11578 SDValue DAGCombiner::ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
11579 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad) {
11580 EVT ResultVT = EVE->getValueType(0);
11581 EVT VecEltVT = InVecVT.getVectorElementType();
11582 unsigned Align = OriginalLoad->getAlignment();
11583 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
11584 VecEltVT.getTypeForEVT(*DAG.getContext()));
11586 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VecEltVT))
11591 SDValue NewPtr = OriginalLoad->getBasePtr();
11593 EVT PtrType = NewPtr.getValueType();
11594 MachinePointerInfo MPI;
11596 if (auto *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo)) {
11597 int Elt = ConstEltNo->getZExtValue();
11598 unsigned PtrOff = VecEltVT.getSizeInBits() * Elt / 8;
11599 Offset = DAG.getConstant(PtrOff, DL, PtrType);
11600 MPI = OriginalLoad->getPointerInfo().getWithOffset(PtrOff);
11602 Offset = DAG.getZExtOrTrunc(EltNo, DL, PtrType);
11603 Offset = DAG.getNode(
11604 ISD::MUL, DL, PtrType, Offset,
11605 DAG.getConstant(VecEltVT.getStoreSize(), DL, PtrType));
11606 MPI = OriginalLoad->getPointerInfo();
11608 NewPtr = DAG.getNode(ISD::ADD, DL, PtrType, NewPtr, Offset);
11610 // The replacement we need to do here is a little tricky: we need to
11611 // replace an extractelement of a load with a load.
11612 // Use ReplaceAllUsesOfValuesWith to do the replacement.
11613 // Note that this replacement assumes that the extractvalue is the only
11614 // use of the load; that's okay because we don't want to perform this
11615 // transformation in other cases anyway.
11618 if (ResultVT.bitsGT(VecEltVT)) {
11619 // If the result type of vextract is wider than the load, then issue an
11620 // extending load instead.
11621 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, ResultVT,
11625 Load = DAG.getExtLoad(
11626 ExtType, SDLoc(EVE), ResultVT, OriginalLoad->getChain(), NewPtr, MPI,
11627 VecEltVT, OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
11628 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
11629 Chain = Load.getValue(1);
11631 Load = DAG.getLoad(
11632 VecEltVT, SDLoc(EVE), OriginalLoad->getChain(), NewPtr, MPI,
11633 OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
11634 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
11635 Chain = Load.getValue(1);
11636 if (ResultVT.bitsLT(VecEltVT))
11637 Load = DAG.getNode(ISD::TRUNCATE, SDLoc(EVE), ResultVT, Load);
11639 Load = DAG.getNode(ISD::BITCAST, SDLoc(EVE), ResultVT, Load);
11641 WorklistRemover DeadNodes(*this);
11642 SDValue From[] = { SDValue(EVE, 0), SDValue(OriginalLoad, 1) };
11643 SDValue To[] = { Load, Chain };
11644 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
11645 // Since we're explicitly calling ReplaceAllUses, add the new node to the
11646 // worklist explicitly as well.
11647 AddToWorklist(Load.getNode());
11648 AddUsersToWorklist(Load.getNode()); // Add users too
11649 // Make sure to revisit this node to clean it up; it will usually be dead.
11650 AddToWorklist(EVE);
11652 return SDValue(EVE, 0);
11655 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
11656 // (vextract (scalar_to_vector val, 0) -> val
11657 SDValue InVec = N->getOperand(0);
11658 EVT VT = InVec.getValueType();
11659 EVT NVT = N->getValueType(0);
11661 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
11662 // Check if the result type doesn't match the inserted element type. A
11663 // SCALAR_TO_VECTOR may truncate the inserted element and the
11664 // EXTRACT_VECTOR_ELT may widen the extracted vector.
11665 SDValue InOp = InVec.getOperand(0);
11666 if (InOp.getValueType() != NVT) {
11667 assert(InOp.getValueType().isInteger() && NVT.isInteger());
11668 return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT);
11673 SDValue EltNo = N->getOperand(1);
11674 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
11676 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
11677 // We only perform this optimization before the op legalization phase because
11678 // we may introduce new vector instructions which are not backed by TD
11679 // patterns. For example on AVX, extracting elements from a wide vector
11680 // without using extract_subvector. However, if we can find an underlying
11681 // scalar value, then we can always use that.
11682 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
11684 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
11685 int NumElem = VT.getVectorNumElements();
11686 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
11687 // Find the new index to extract from.
11688 int OrigElt = SVOp->getMaskElt(Elt);
11690 // Extracting an undef index is undef.
11692 return DAG.getUNDEF(NVT);
11694 // Select the right vector half to extract from.
11696 if (OrigElt < NumElem) {
11697 SVInVec = InVec->getOperand(0);
11699 SVInVec = InVec->getOperand(1);
11700 OrigElt -= NumElem;
11703 if (SVInVec.getOpcode() == ISD::BUILD_VECTOR) {
11704 SDValue InOp = SVInVec.getOperand(OrigElt);
11705 if (InOp.getValueType() != NVT) {
11706 assert(InOp.getValueType().isInteger() && NVT.isInteger());
11707 InOp = DAG.getSExtOrTrunc(InOp, SDLoc(SVInVec), NVT);
11713 // FIXME: We should handle recursing on other vector shuffles and
11714 // scalar_to_vector here as well.
11716 if (!LegalOperations) {
11717 EVT IndexTy = TLI.getVectorIdxTy(DAG.getDataLayout());
11718 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT, SVInVec,
11719 DAG.getConstant(OrigElt, SDLoc(SVOp), IndexTy));
11723 bool BCNumEltsChanged = false;
11724 EVT ExtVT = VT.getVectorElementType();
11727 // If the result of load has to be truncated, then it's not necessarily
11729 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
11732 if (InVec.getOpcode() == ISD::BITCAST) {
11733 // Don't duplicate a load with other uses.
11734 if (!InVec.hasOneUse())
11737 EVT BCVT = InVec.getOperand(0).getValueType();
11738 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
11740 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
11741 BCNumEltsChanged = true;
11742 InVec = InVec.getOperand(0);
11743 ExtVT = BCVT.getVectorElementType();
11746 // (vextract (vN[if]M load $addr), i) -> ([if]M load $addr + i * size)
11747 if (!LegalOperations && !ConstEltNo && InVec.hasOneUse() &&
11748 ISD::isNormalLoad(InVec.getNode()) &&
11749 !N->getOperand(1)->hasPredecessor(InVec.getNode())) {
11750 SDValue Index = N->getOperand(1);
11751 if (LoadSDNode *OrigLoad = dyn_cast<LoadSDNode>(InVec))
11752 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, Index,
11756 // Perform only after legalization to ensure build_vector / vector_shuffle
11757 // optimizations have already been done.
11758 if (!LegalOperations) return SDValue();
11760 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
11761 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
11762 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
11765 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
11767 LoadSDNode *LN0 = nullptr;
11768 const ShuffleVectorSDNode *SVN = nullptr;
11769 if (ISD::isNormalLoad(InVec.getNode())) {
11770 LN0 = cast<LoadSDNode>(InVec);
11771 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
11772 InVec.getOperand(0).getValueType() == ExtVT &&
11773 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
11774 // Don't duplicate a load with other uses.
11775 if (!InVec.hasOneUse())
11778 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
11779 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
11780 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
11782 // (load $addr+1*size)
11784 // Don't duplicate a load with other uses.
11785 if (!InVec.hasOneUse())
11788 // If the bit convert changed the number of elements, it is unsafe
11789 // to examine the mask.
11790 if (BCNumEltsChanged)
11793 // Select the input vector, guarding against out of range extract vector.
11794 unsigned NumElems = VT.getVectorNumElements();
11795 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
11796 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
11798 if (InVec.getOpcode() == ISD::BITCAST) {
11799 // Don't duplicate a load with other uses.
11800 if (!InVec.hasOneUse())
11803 InVec = InVec.getOperand(0);
11805 if (ISD::isNormalLoad(InVec.getNode())) {
11806 LN0 = cast<LoadSDNode>(InVec);
11807 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
11808 EltNo = DAG.getConstant(Elt, SDLoc(EltNo), EltNo.getValueType());
11812 // Make sure we found a non-volatile load and the extractelement is
11814 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
11817 // If Idx was -1 above, Elt is going to be -1, so just return undef.
11819 return DAG.getUNDEF(LVT);
11821 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, EltNo, LN0);
11827 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
11828 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
11829 // We perform this optimization post type-legalization because
11830 // the type-legalizer often scalarizes integer-promoted vectors.
11831 // Performing this optimization before may create bit-casts which
11832 // will be type-legalized to complex code sequences.
11833 // We perform this optimization only before the operation legalizer because we
11834 // may introduce illegal operations.
11835 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
11838 unsigned NumInScalars = N->getNumOperands();
11840 EVT VT = N->getValueType(0);
11842 // Check to see if this is a BUILD_VECTOR of a bunch of values
11843 // which come from any_extend or zero_extend nodes. If so, we can create
11844 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
11845 // optimizations. We do not handle sign-extend because we can't fill the sign
11847 EVT SourceType = MVT::Other;
11848 bool AllAnyExt = true;
11850 for (unsigned i = 0; i != NumInScalars; ++i) {
11851 SDValue In = N->getOperand(i);
11852 // Ignore undef inputs.
11853 if (In.getOpcode() == ISD::UNDEF) continue;
11855 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
11856 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
11858 // Abort if the element is not an extension.
11859 if (!ZeroExt && !AnyExt) {
11860 SourceType = MVT::Other;
11864 // The input is a ZeroExt or AnyExt. Check the original type.
11865 EVT InTy = In.getOperand(0).getValueType();
11867 // Check that all of the widened source types are the same.
11868 if (SourceType == MVT::Other)
11871 else if (InTy != SourceType) {
11872 // Multiple income types. Abort.
11873 SourceType = MVT::Other;
11877 // Check if all of the extends are ANY_EXTENDs.
11878 AllAnyExt &= AnyExt;
11881 // In order to have valid types, all of the inputs must be extended from the
11882 // same source type and all of the inputs must be any or zero extend.
11883 // Scalar sizes must be a power of two.
11884 EVT OutScalarTy = VT.getScalarType();
11885 bool ValidTypes = SourceType != MVT::Other &&
11886 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
11887 isPowerOf2_32(SourceType.getSizeInBits());
11889 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
11890 // turn into a single shuffle instruction.
11894 bool isLE = DAG.getDataLayout().isLittleEndian();
11895 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
11896 assert(ElemRatio > 1 && "Invalid element size ratio");
11897 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
11898 DAG.getConstant(0, SDLoc(N), SourceType);
11900 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
11901 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
11903 // Populate the new build_vector
11904 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
11905 SDValue Cast = N->getOperand(i);
11906 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
11907 Cast.getOpcode() == ISD::ZERO_EXTEND ||
11908 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
11910 if (Cast.getOpcode() == ISD::UNDEF)
11911 In = DAG.getUNDEF(SourceType);
11913 In = Cast->getOperand(0);
11914 unsigned Index = isLE ? (i * ElemRatio) :
11915 (i * ElemRatio + (ElemRatio - 1));
11917 assert(Index < Ops.size() && "Invalid index");
11921 // The type of the new BUILD_VECTOR node.
11922 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
11923 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
11924 "Invalid vector size");
11925 // Check if the new vector type is legal.
11926 if (!isTypeLegal(VecVT)) return SDValue();
11928 // Make the new BUILD_VECTOR.
11929 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
11931 // The new BUILD_VECTOR node has the potential to be further optimized.
11932 AddToWorklist(BV.getNode());
11933 // Bitcast to the desired type.
11934 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
11937 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
11938 EVT VT = N->getValueType(0);
11940 unsigned NumInScalars = N->getNumOperands();
11943 EVT SrcVT = MVT::Other;
11944 unsigned Opcode = ISD::DELETED_NODE;
11945 unsigned NumDefs = 0;
11947 for (unsigned i = 0; i != NumInScalars; ++i) {
11948 SDValue In = N->getOperand(i);
11949 unsigned Opc = In.getOpcode();
11951 if (Opc == ISD::UNDEF)
11954 // If all scalar values are floats and converted from integers.
11955 if (Opcode == ISD::DELETED_NODE &&
11956 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
11963 EVT InVT = In.getOperand(0).getValueType();
11965 // If all scalar values are typed differently, bail out. It's chosen to
11966 // simplify BUILD_VECTOR of integer types.
11967 if (SrcVT == MVT::Other)
11974 // If the vector has just one element defined, it's not worth to fold it into
11975 // a vectorized one.
11979 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
11980 && "Should only handle conversion from integer to float.");
11981 assert(SrcVT != MVT::Other && "Cannot determine source type!");
11983 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
11985 if (!TLI.isOperationLegalOrCustom(Opcode, NVT))
11988 // Just because the floating-point vector type is legal does not necessarily
11989 // mean that the corresponding integer vector type is.
11990 if (!isTypeLegal(NVT))
11993 SmallVector<SDValue, 8> Opnds;
11994 for (unsigned i = 0; i != NumInScalars; ++i) {
11995 SDValue In = N->getOperand(i);
11997 if (In.getOpcode() == ISD::UNDEF)
11998 Opnds.push_back(DAG.getUNDEF(SrcVT));
12000 Opnds.push_back(In.getOperand(0));
12002 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Opnds);
12003 AddToWorklist(BV.getNode());
12005 return DAG.getNode(Opcode, dl, VT, BV);
12008 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
12009 unsigned NumInScalars = N->getNumOperands();
12011 EVT VT = N->getValueType(0);
12013 // A vector built entirely of undefs is undef.
12014 if (ISD::allOperandsUndef(N))
12015 return DAG.getUNDEF(VT);
12017 if (SDValue V = reduceBuildVecExtToExtBuildVec(N))
12020 if (SDValue V = reduceBuildVecConvertToConvertBuildVec(N))
12023 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
12024 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
12025 // at most two distinct vectors, turn this into a shuffle node.
12027 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
12028 if (!isTypeLegal(VT))
12031 // May only combine to shuffle after legalize if shuffle is legal.
12032 if (LegalOperations && !TLI.isOperationLegal(ISD::VECTOR_SHUFFLE, VT))
12035 SDValue VecIn1, VecIn2;
12036 bool UsesZeroVector = false;
12037 for (unsigned i = 0; i != NumInScalars; ++i) {
12038 SDValue Op = N->getOperand(i);
12039 // Ignore undef inputs.
12040 if (Op.getOpcode() == ISD::UNDEF) continue;
12042 // See if we can combine this build_vector into a blend with a zero vector.
12043 if (!VecIn2.getNode() && (isNullConstant(Op) || isNullFPConstant(Op))) {
12044 UsesZeroVector = true;
12048 // If this input is something other than a EXTRACT_VECTOR_ELT with a
12049 // constant index, bail out.
12050 if (Op.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
12051 !isa<ConstantSDNode>(Op.getOperand(1))) {
12052 VecIn1 = VecIn2 = SDValue(nullptr, 0);
12056 // We allow up to two distinct input vectors.
12057 SDValue ExtractedFromVec = Op.getOperand(0);
12058 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
12061 if (!VecIn1.getNode()) {
12062 VecIn1 = ExtractedFromVec;
12063 } else if (!VecIn2.getNode() && !UsesZeroVector) {
12064 VecIn2 = ExtractedFromVec;
12066 // Too many inputs.
12067 VecIn1 = VecIn2 = SDValue(nullptr, 0);
12072 // If everything is good, we can make a shuffle operation.
12073 if (VecIn1.getNode()) {
12074 unsigned InNumElements = VecIn1.getValueType().getVectorNumElements();
12075 SmallVector<int, 8> Mask;
12076 for (unsigned i = 0; i != NumInScalars; ++i) {
12077 unsigned Opcode = N->getOperand(i).getOpcode();
12078 if (Opcode == ISD::UNDEF) {
12079 Mask.push_back(-1);
12083 // Operands can also be zero.
12084 if (Opcode != ISD::EXTRACT_VECTOR_ELT) {
12085 assert(UsesZeroVector &&
12086 (Opcode == ISD::Constant || Opcode == ISD::ConstantFP) &&
12087 "Unexpected node found!");
12088 Mask.push_back(NumInScalars+i);
12092 // If extracting from the first vector, just use the index directly.
12093 SDValue Extract = N->getOperand(i);
12094 SDValue ExtVal = Extract.getOperand(1);
12095 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
12096 if (Extract.getOperand(0) == VecIn1) {
12097 Mask.push_back(ExtIndex);
12101 // Otherwise, use InIdx + InputVecSize
12102 Mask.push_back(InNumElements + ExtIndex);
12105 // Avoid introducing illegal shuffles with zero.
12106 if (UsesZeroVector && !TLI.isVectorClearMaskLegal(Mask, VT))
12109 // We can't generate a shuffle node with mismatched input and output types.
12110 // Attempt to transform a single input vector to the correct type.
12111 if ((VT != VecIn1.getValueType())) {
12112 // If the input vector type has a different base type to the output
12113 // vector type, bail out.
12114 EVT VTElemType = VT.getVectorElementType();
12115 if ((VecIn1.getValueType().getVectorElementType() != VTElemType) ||
12116 (VecIn2.getNode() &&
12117 (VecIn2.getValueType().getVectorElementType() != VTElemType)))
12120 // If the input vector is too small, widen it.
12121 // We only support widening of vectors which are half the size of the
12122 // output registers. For example XMM->YMM widening on X86 with AVX.
12123 EVT VecInT = VecIn1.getValueType();
12124 if (VecInT.getSizeInBits() * 2 == VT.getSizeInBits()) {
12125 // If we only have one small input, widen it by adding undef values.
12126 if (!VecIn2.getNode())
12127 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, VecIn1,
12128 DAG.getUNDEF(VecIn1.getValueType()));
12129 else if (VecIn1.getValueType() == VecIn2.getValueType()) {
12130 // If we have two small inputs of the same type, try to concat them.
12131 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, VecIn1, VecIn2);
12132 VecIn2 = SDValue(nullptr, 0);
12135 } else if (VecInT.getSizeInBits() == VT.getSizeInBits() * 2) {
12136 // If the input vector is too large, try to split it.
12137 // We don't support having two input vectors that are too large.
12138 // If the zero vector was used, we can not split the vector,
12139 // since we'd need 3 inputs.
12140 if (UsesZeroVector || VecIn2.getNode())
12143 if (!TLI.isExtractSubvectorCheap(VT, VT.getVectorNumElements()))
12146 // Try to replace VecIn1 with two extract_subvectors
12147 // No need to update the masks, they should still be correct.
12148 VecIn2 = DAG.getNode(
12149 ISD::EXTRACT_SUBVECTOR, dl, VT, VecIn1,
12150 DAG.getConstant(VT.getVectorNumElements(), dl,
12151 TLI.getVectorIdxTy(DAG.getDataLayout())));
12152 VecIn1 = DAG.getNode(
12153 ISD::EXTRACT_SUBVECTOR, dl, VT, VecIn1,
12154 DAG.getConstant(0, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
12159 if (UsesZeroVector)
12160 VecIn2 = VT.isInteger() ? DAG.getConstant(0, dl, VT) :
12161 DAG.getConstantFP(0.0, dl, VT);
12163 // If VecIn2 is unused then change it to undef.
12164 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
12166 // Check that we were able to transform all incoming values to the same
12168 if (VecIn2.getValueType() != VecIn1.getValueType() ||
12169 VecIn1.getValueType() != VT)
12172 // Return the new VECTOR_SHUFFLE node.
12176 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]);
12182 static SDValue combineConcatVectorOfScalars(SDNode *N, SelectionDAG &DAG) {
12183 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12184 EVT OpVT = N->getOperand(0).getValueType();
12186 // If the operands are legal vectors, leave them alone.
12187 if (TLI.isTypeLegal(OpVT))
12191 EVT VT = N->getValueType(0);
12192 SmallVector<SDValue, 8> Ops;
12194 EVT SVT = EVT::getIntegerVT(*DAG.getContext(), OpVT.getSizeInBits());
12195 SDValue ScalarUndef = DAG.getNode(ISD::UNDEF, DL, SVT);
12197 // Keep track of what we encounter.
12198 bool AnyInteger = false;
12199 bool AnyFP = false;
12200 for (const SDValue &Op : N->ops()) {
12201 if (ISD::BITCAST == Op.getOpcode() &&
12202 !Op.getOperand(0).getValueType().isVector())
12203 Ops.push_back(Op.getOperand(0));
12204 else if (ISD::UNDEF == Op.getOpcode())
12205 Ops.push_back(ScalarUndef);
12209 // Note whether we encounter an integer or floating point scalar.
12210 // If it's neither, bail out, it could be something weird like x86mmx.
12211 EVT LastOpVT = Ops.back().getValueType();
12212 if (LastOpVT.isFloatingPoint())
12214 else if (LastOpVT.isInteger())
12220 // If any of the operands is a floating point scalar bitcast to a vector,
12221 // use floating point types throughout, and bitcast everything.
12222 // Replace UNDEFs by another scalar UNDEF node, of the final desired type.
12224 SVT = EVT::getFloatingPointVT(OpVT.getSizeInBits());
12225 ScalarUndef = DAG.getNode(ISD::UNDEF, DL, SVT);
12227 for (SDValue &Op : Ops) {
12228 if (Op.getValueType() == SVT)
12230 if (Op.getOpcode() == ISD::UNDEF)
12233 Op = DAG.getNode(ISD::BITCAST, DL, SVT, Op);
12238 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SVT,
12239 VT.getSizeInBits() / SVT.getSizeInBits());
12240 return DAG.getNode(ISD::BITCAST, DL, VT,
12241 DAG.getNode(ISD::BUILD_VECTOR, DL, VecVT, Ops));
12244 // Check to see if this is a CONCAT_VECTORS of a bunch of EXTRACT_SUBVECTOR
12245 // operations. If so, and if the EXTRACT_SUBVECTOR vector inputs come from at
12246 // most two distinct vectors the same size as the result, attempt to turn this
12247 // into a legal shuffle.
12248 static SDValue combineConcatVectorOfExtracts(SDNode *N, SelectionDAG &DAG) {
12249 EVT VT = N->getValueType(0);
12250 EVT OpVT = N->getOperand(0).getValueType();
12251 int NumElts = VT.getVectorNumElements();
12252 int NumOpElts = OpVT.getVectorNumElements();
12254 SDValue SV0 = DAG.getUNDEF(VT), SV1 = DAG.getUNDEF(VT);
12255 SmallVector<int, 8> Mask;
12257 for (SDValue Op : N->ops()) {
12258 // Peek through any bitcast.
12259 while (Op.getOpcode() == ISD::BITCAST)
12260 Op = Op.getOperand(0);
12262 // UNDEF nodes convert to UNDEF shuffle mask values.
12263 if (Op.getOpcode() == ISD::UNDEF) {
12264 Mask.append((unsigned)NumOpElts, -1);
12268 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
12271 // What vector are we extracting the subvector from and at what index?
12272 SDValue ExtVec = Op.getOperand(0);
12274 // We want the EVT of the original extraction to correctly scale the
12275 // extraction index.
12276 EVT ExtVT = ExtVec.getValueType();
12278 // Peek through any bitcast.
12279 while (ExtVec.getOpcode() == ISD::BITCAST)
12280 ExtVec = ExtVec.getOperand(0);
12282 // UNDEF nodes convert to UNDEF shuffle mask values.
12283 if (ExtVec.getOpcode() == ISD::UNDEF) {
12284 Mask.append((unsigned)NumOpElts, -1);
12288 if (!isa<ConstantSDNode>(Op.getOperand(1)))
12290 int ExtIdx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
12292 // Ensure that we are extracting a subvector from a vector the same
12293 // size as the result.
12294 if (ExtVT.getSizeInBits() != VT.getSizeInBits())
12297 // Scale the subvector index to account for any bitcast.
12298 int NumExtElts = ExtVT.getVectorNumElements();
12299 if (0 == (NumExtElts % NumElts))
12300 ExtIdx /= (NumExtElts / NumElts);
12301 else if (0 == (NumElts % NumExtElts))
12302 ExtIdx *= (NumElts / NumExtElts);
12306 // At most we can reference 2 inputs in the final shuffle.
12307 if (SV0.getOpcode() == ISD::UNDEF || SV0 == ExtVec) {
12309 for (int i = 0; i != NumOpElts; ++i)
12310 Mask.push_back(i + ExtIdx);
12311 } else if (SV1.getOpcode() == ISD::UNDEF || SV1 == ExtVec) {
12313 for (int i = 0; i != NumOpElts; ++i)
12314 Mask.push_back(i + ExtIdx + NumElts);
12320 if (!DAG.getTargetLoweringInfo().isShuffleMaskLegal(Mask, VT))
12323 return DAG.getVectorShuffle(VT, SDLoc(N), DAG.getBitcast(VT, SV0),
12324 DAG.getBitcast(VT, SV1), Mask);
12327 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
12328 // If we only have one input vector, we don't need to do any concatenation.
12329 if (N->getNumOperands() == 1)
12330 return N->getOperand(0);
12332 // Check if all of the operands are undefs.
12333 EVT VT = N->getValueType(0);
12334 if (ISD::allOperandsUndef(N))
12335 return DAG.getUNDEF(VT);
12337 // Optimize concat_vectors where all but the first of the vectors are undef.
12338 if (std::all_of(std::next(N->op_begin()), N->op_end(), [](const SDValue &Op) {
12339 return Op.getOpcode() == ISD::UNDEF;
12341 SDValue In = N->getOperand(0);
12342 assert(In.getValueType().isVector() && "Must concat vectors");
12344 // Transform: concat_vectors(scalar, undef) -> scalar_to_vector(sclr).
12345 if (In->getOpcode() == ISD::BITCAST &&
12346 !In->getOperand(0)->getValueType(0).isVector()) {
12347 SDValue Scalar = In->getOperand(0);
12349 // If the bitcast type isn't legal, it might be a trunc of a legal type;
12350 // look through the trunc so we can still do the transform:
12351 // concat_vectors(trunc(scalar), undef) -> scalar_to_vector(scalar)
12352 if (Scalar->getOpcode() == ISD::TRUNCATE &&
12353 !TLI.isTypeLegal(Scalar.getValueType()) &&
12354 TLI.isTypeLegal(Scalar->getOperand(0).getValueType()))
12355 Scalar = Scalar->getOperand(0);
12357 EVT SclTy = Scalar->getValueType(0);
12359 if (!SclTy.isFloatingPoint() && !SclTy.isInteger())
12362 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SclTy,
12363 VT.getSizeInBits() / SclTy.getSizeInBits());
12364 if (!TLI.isTypeLegal(NVT) || !TLI.isTypeLegal(Scalar.getValueType()))
12367 SDLoc dl = SDLoc(N);
12368 SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NVT, Scalar);
12369 return DAG.getNode(ISD::BITCAST, dl, VT, Res);
12373 // Fold any combination of BUILD_VECTOR or UNDEF nodes into one BUILD_VECTOR.
12374 // We have already tested above for an UNDEF only concatenation.
12375 // fold (concat_vectors (BUILD_VECTOR A, B, ...), (BUILD_VECTOR C, D, ...))
12376 // -> (BUILD_VECTOR A, B, ..., C, D, ...)
12377 auto IsBuildVectorOrUndef = [](const SDValue &Op) {
12378 return ISD::UNDEF == Op.getOpcode() || ISD::BUILD_VECTOR == Op.getOpcode();
12380 bool AllBuildVectorsOrUndefs =
12381 std::all_of(N->op_begin(), N->op_end(), IsBuildVectorOrUndef);
12382 if (AllBuildVectorsOrUndefs) {
12383 SmallVector<SDValue, 8> Opnds;
12384 EVT SVT = VT.getScalarType();
12387 if (!SVT.isFloatingPoint()) {
12388 // If BUILD_VECTOR are from built from integer, they may have different
12389 // operand types. Get the smallest type and truncate all operands to it.
12390 bool FoundMinVT = false;
12391 for (const SDValue &Op : N->ops())
12392 if (ISD::BUILD_VECTOR == Op.getOpcode()) {
12393 EVT OpSVT = Op.getOperand(0)->getValueType(0);
12394 MinVT = (!FoundMinVT || OpSVT.bitsLE(MinVT)) ? OpSVT : MinVT;
12397 assert(FoundMinVT && "Concat vector type mismatch");
12400 for (const SDValue &Op : N->ops()) {
12401 EVT OpVT = Op.getValueType();
12402 unsigned NumElts = OpVT.getVectorNumElements();
12404 if (ISD::UNDEF == Op.getOpcode())
12405 Opnds.append(NumElts, DAG.getUNDEF(MinVT));
12407 if (ISD::BUILD_VECTOR == Op.getOpcode()) {
12408 if (SVT.isFloatingPoint()) {
12409 assert(SVT == OpVT.getScalarType() && "Concat vector type mismatch");
12410 Opnds.append(Op->op_begin(), Op->op_begin() + NumElts);
12412 for (unsigned i = 0; i != NumElts; ++i)
12414 DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinVT, Op.getOperand(i)));
12419 assert(VT.getVectorNumElements() == Opnds.size() &&
12420 "Concat vector type mismatch");
12421 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
12424 // Fold CONCAT_VECTORS of only bitcast scalars (or undef) to BUILD_VECTOR.
12425 if (SDValue V = combineConcatVectorOfScalars(N, DAG))
12428 // Fold CONCAT_VECTORS of EXTRACT_SUBVECTOR (or undef) to VECTOR_SHUFFLE.
12429 if (Level < AfterLegalizeVectorOps && TLI.isTypeLegal(VT))
12430 if (SDValue V = combineConcatVectorOfExtracts(N, DAG))
12433 // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
12434 // nodes often generate nop CONCAT_VECTOR nodes.
12435 // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that
12436 // place the incoming vectors at the exact same location.
12437 SDValue SingleSource = SDValue();
12438 unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements();
12440 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
12441 SDValue Op = N->getOperand(i);
12443 if (Op.getOpcode() == ISD::UNDEF)
12446 // Check if this is the identity extract:
12447 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
12450 // Find the single incoming vector for the extract_subvector.
12451 if (SingleSource.getNode()) {
12452 if (Op.getOperand(0) != SingleSource)
12455 SingleSource = Op.getOperand(0);
12457 // Check the source type is the same as the type of the result.
12458 // If not, this concat may extend the vector, so we can not
12459 // optimize it away.
12460 if (SingleSource.getValueType() != N->getValueType(0))
12464 unsigned IdentityIndex = i * PartNumElem;
12465 ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1));
12466 // The extract index must be constant.
12470 // Check that we are reading from the identity index.
12471 if (CS->getZExtValue() != IdentityIndex)
12475 if (SingleSource.getNode())
12476 return SingleSource;
12481 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
12482 EVT NVT = N->getValueType(0);
12483 SDValue V = N->getOperand(0);
12485 if (V->getOpcode() == ISD::CONCAT_VECTORS) {
12487 // (extract_subvec (concat V1, V2, ...), i)
12490 // Only operand 0 is checked as 'concat' assumes all inputs of the same
12492 if (V->getOperand(0).getValueType() != NVT)
12494 unsigned Idx = N->getConstantOperandVal(1);
12495 unsigned NumElems = NVT.getVectorNumElements();
12496 assert((Idx % NumElems) == 0 &&
12497 "IDX in concat is not a multiple of the result vector length.");
12498 return V->getOperand(Idx / NumElems);
12502 if (V->getOpcode() == ISD::BITCAST)
12503 V = V.getOperand(0);
12505 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
12507 // Handle only simple case where vector being inserted and vector
12508 // being extracted are of same type, and are half size of larger vectors.
12509 EVT BigVT = V->getOperand(0).getValueType();
12510 EVT SmallVT = V->getOperand(1).getValueType();
12511 if (!NVT.bitsEq(SmallVT) || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
12514 // Only handle cases where both indexes are constants with the same type.
12515 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
12516 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
12518 if (InsIdx && ExtIdx &&
12519 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
12520 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
12522 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
12524 // indices are equal or bit offsets are equal => V1
12525 // otherwise => (extract_subvec V1, ExtIdx)
12526 if (InsIdx->getZExtValue() * SmallVT.getScalarType().getSizeInBits() ==
12527 ExtIdx->getZExtValue() * NVT.getScalarType().getSizeInBits())
12528 return DAG.getNode(ISD::BITCAST, dl, NVT, V->getOperand(1));
12529 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT,
12530 DAG.getNode(ISD::BITCAST, dl,
12531 N->getOperand(0).getValueType(),
12532 V->getOperand(0)), N->getOperand(1));
12539 static SDValue simplifyShuffleOperandRecursively(SmallBitVector &UsedElements,
12540 SDValue V, SelectionDAG &DAG) {
12542 EVT VT = V.getValueType();
12544 switch (V.getOpcode()) {
12548 case ISD::CONCAT_VECTORS: {
12549 EVT OpVT = V->getOperand(0).getValueType();
12550 int OpSize = OpVT.getVectorNumElements();
12551 SmallBitVector OpUsedElements(OpSize, false);
12552 bool FoundSimplification = false;
12553 SmallVector<SDValue, 4> NewOps;
12554 NewOps.reserve(V->getNumOperands());
12555 for (int i = 0, NumOps = V->getNumOperands(); i < NumOps; ++i) {
12556 SDValue Op = V->getOperand(i);
12557 bool OpUsed = false;
12558 for (int j = 0; j < OpSize; ++j)
12559 if (UsedElements[i * OpSize + j]) {
12560 OpUsedElements[j] = true;
12564 OpUsed ? simplifyShuffleOperandRecursively(OpUsedElements, Op, DAG)
12565 : DAG.getUNDEF(OpVT));
12566 FoundSimplification |= Op == NewOps.back();
12567 OpUsedElements.reset();
12569 if (FoundSimplification)
12570 V = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, NewOps);
12574 case ISD::INSERT_SUBVECTOR: {
12575 SDValue BaseV = V->getOperand(0);
12576 SDValue SubV = V->getOperand(1);
12577 auto *IdxN = dyn_cast<ConstantSDNode>(V->getOperand(2));
12581 int SubSize = SubV.getValueType().getVectorNumElements();
12582 int Idx = IdxN->getZExtValue();
12583 bool SubVectorUsed = false;
12584 SmallBitVector SubUsedElements(SubSize, false);
12585 for (int i = 0; i < SubSize; ++i)
12586 if (UsedElements[i + Idx]) {
12587 SubVectorUsed = true;
12588 SubUsedElements[i] = true;
12589 UsedElements[i + Idx] = false;
12592 // Now recurse on both the base and sub vectors.
12593 SDValue SimplifiedSubV =
12595 ? simplifyShuffleOperandRecursively(SubUsedElements, SubV, DAG)
12596 : DAG.getUNDEF(SubV.getValueType());
12597 SDValue SimplifiedBaseV = simplifyShuffleOperandRecursively(UsedElements, BaseV, DAG);
12598 if (SimplifiedSubV != SubV || SimplifiedBaseV != BaseV)
12599 V = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT,
12600 SimplifiedBaseV, SimplifiedSubV, V->getOperand(2));
12606 static SDValue simplifyShuffleOperands(ShuffleVectorSDNode *SVN, SDValue N0,
12607 SDValue N1, SelectionDAG &DAG) {
12608 EVT VT = SVN->getValueType(0);
12609 int NumElts = VT.getVectorNumElements();
12610 SmallBitVector N0UsedElements(NumElts, false), N1UsedElements(NumElts, false);
12611 for (int M : SVN->getMask())
12612 if (M >= 0 && M < NumElts)
12613 N0UsedElements[M] = true;
12614 else if (M >= NumElts)
12615 N1UsedElements[M - NumElts] = true;
12617 SDValue S0 = simplifyShuffleOperandRecursively(N0UsedElements, N0, DAG);
12618 SDValue S1 = simplifyShuffleOperandRecursively(N1UsedElements, N1, DAG);
12619 if (S0 == N0 && S1 == N1)
12622 return DAG.getVectorShuffle(VT, SDLoc(SVN), S0, S1, SVN->getMask());
12625 // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat,
12626 // or turn a shuffle of a single concat into simpler shuffle then concat.
12627 static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
12628 EVT VT = N->getValueType(0);
12629 unsigned NumElts = VT.getVectorNumElements();
12631 SDValue N0 = N->getOperand(0);
12632 SDValue N1 = N->getOperand(1);
12633 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
12635 SmallVector<SDValue, 4> Ops;
12636 EVT ConcatVT = N0.getOperand(0).getValueType();
12637 unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements();
12638 unsigned NumConcats = NumElts / NumElemsPerConcat;
12640 // Special case: shuffle(concat(A,B)) can be more efficiently represented
12641 // as concat(shuffle(A,B),UNDEF) if the shuffle doesn't set any of the high
12642 // half vector elements.
12643 if (NumElemsPerConcat * 2 == NumElts && N1.getOpcode() == ISD::UNDEF &&
12644 std::all_of(SVN->getMask().begin() + NumElemsPerConcat,
12645 SVN->getMask().end(), [](int i) { return i == -1; })) {
12646 N0 = DAG.getVectorShuffle(ConcatVT, SDLoc(N), N0.getOperand(0), N0.getOperand(1),
12647 ArrayRef<int>(SVN->getMask().begin(), NumElemsPerConcat));
12648 N1 = DAG.getUNDEF(ConcatVT);
12649 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, N0, N1);
12652 // Look at every vector that's inserted. We're looking for exact
12653 // subvector-sized copies from a concatenated vector
12654 for (unsigned I = 0; I != NumConcats; ++I) {
12655 // Make sure we're dealing with a copy.
12656 unsigned Begin = I * NumElemsPerConcat;
12657 bool AllUndef = true, NoUndef = true;
12658 for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) {
12659 if (SVN->getMaskElt(J) >= 0)
12666 if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0)
12669 for (unsigned J = 1; J != NumElemsPerConcat; ++J)
12670 if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J))
12673 unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat;
12674 if (FirstElt < N0.getNumOperands())
12675 Ops.push_back(N0.getOperand(FirstElt));
12677 Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands()));
12679 } else if (AllUndef) {
12680 Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType()));
12681 } else { // Mixed with general masks and undefs, can't do optimization.
12686 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops);
12689 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
12690 EVT VT = N->getValueType(0);
12691 unsigned NumElts = VT.getVectorNumElements();
12693 SDValue N0 = N->getOperand(0);
12694 SDValue N1 = N->getOperand(1);
12696 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
12698 // Canonicalize shuffle undef, undef -> undef
12699 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
12700 return DAG.getUNDEF(VT);
12702 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
12704 // Canonicalize shuffle v, v -> v, undef
12706 SmallVector<int, 8> NewMask;
12707 for (unsigned i = 0; i != NumElts; ++i) {
12708 int Idx = SVN->getMaskElt(i);
12709 if (Idx >= (int)NumElts) Idx -= NumElts;
12710 NewMask.push_back(Idx);
12712 return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT),
12716 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
12717 if (N0.getOpcode() == ISD::UNDEF) {
12718 SmallVector<int, 8> NewMask;
12719 for (unsigned i = 0; i != NumElts; ++i) {
12720 int Idx = SVN->getMaskElt(i);
12722 if (Idx >= (int)NumElts)
12725 Idx = -1; // remove reference to lhs
12727 NewMask.push_back(Idx);
12729 return DAG.getVectorShuffle(VT, SDLoc(N), N1, DAG.getUNDEF(VT),
12733 // Remove references to rhs if it is undef
12734 if (N1.getOpcode() == ISD::UNDEF) {
12735 bool Changed = false;
12736 SmallVector<int, 8> NewMask;
12737 for (unsigned i = 0; i != NumElts; ++i) {
12738 int Idx = SVN->getMaskElt(i);
12739 if (Idx >= (int)NumElts) {
12743 NewMask.push_back(Idx);
12746 return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, &NewMask[0]);
12749 // If it is a splat, check if the argument vector is another splat or a
12751 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
12752 SDNode *V = N0.getNode();
12754 // If this is a bit convert that changes the element type of the vector but
12755 // not the number of vector elements, look through it. Be careful not to
12756 // look though conversions that change things like v4f32 to v2f64.
12757 if (V->getOpcode() == ISD::BITCAST) {
12758 SDValue ConvInput = V->getOperand(0);
12759 if (ConvInput.getValueType().isVector() &&
12760 ConvInput.getValueType().getVectorNumElements() == NumElts)
12761 V = ConvInput.getNode();
12764 if (V->getOpcode() == ISD::BUILD_VECTOR) {
12765 assert(V->getNumOperands() == NumElts &&
12766 "BUILD_VECTOR has wrong number of operands");
12768 bool AllSame = true;
12769 for (unsigned i = 0; i != NumElts; ++i) {
12770 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
12771 Base = V->getOperand(i);
12775 // Splat of <u, u, u, u>, return <u, u, u, u>
12776 if (!Base.getNode())
12778 for (unsigned i = 0; i != NumElts; ++i) {
12779 if (V->getOperand(i) != Base) {
12784 // Splat of <x, x, x, x>, return <x, x, x, x>
12788 // Canonicalize any other splat as a build_vector.
12789 const SDValue &Splatted = V->getOperand(SVN->getSplatIndex());
12790 SmallVector<SDValue, 8> Ops(NumElts, Splatted);
12791 SDValue NewBV = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
12792 V->getValueType(0), Ops);
12794 // We may have jumped through bitcasts, so the type of the
12795 // BUILD_VECTOR may not match the type of the shuffle.
12796 if (V->getValueType(0) != VT)
12797 NewBV = DAG.getNode(ISD::BITCAST, SDLoc(N), VT, NewBV);
12802 // There are various patterns used to build up a vector from smaller vectors,
12803 // subvectors, or elements. Scan chains of these and replace unused insertions
12804 // or components with undef.
12805 if (SDValue S = simplifyShuffleOperands(SVN, N0, N1, DAG))
12808 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
12809 Level < AfterLegalizeVectorOps &&
12810 (N1.getOpcode() == ISD::UNDEF ||
12811 (N1.getOpcode() == ISD::CONCAT_VECTORS &&
12812 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) {
12813 SDValue V = partitionShuffleOfConcats(N, DAG);
12819 // Attempt to combine a shuffle of 2 inputs of 'scalar sources' -
12820 // BUILD_VECTOR or SCALAR_TO_VECTOR into a single BUILD_VECTOR.
12821 if (Level < AfterLegalizeVectorOps && TLI.isTypeLegal(VT)) {
12822 SmallVector<SDValue, 8> Ops;
12823 for (int M : SVN->getMask()) {
12824 SDValue Op = DAG.getUNDEF(VT.getScalarType());
12826 int Idx = M % NumElts;
12827 SDValue &S = (M < (int)NumElts ? N0 : N1);
12828 if (S.getOpcode() == ISD::BUILD_VECTOR && S.hasOneUse()) {
12829 Op = S.getOperand(Idx);
12830 } else if (S.getOpcode() == ISD::SCALAR_TO_VECTOR && S.hasOneUse()) {
12832 Op = S.getOperand(0);
12834 // Operand can't be combined - bail out.
12840 if (Ops.size() == VT.getVectorNumElements()) {
12841 // BUILD_VECTOR requires all inputs to be of the same type, find the
12842 // maximum type and extend them all.
12843 EVT SVT = VT.getScalarType();
12844 if (SVT.isInteger())
12845 for (SDValue &Op : Ops)
12846 SVT = (SVT.bitsLT(Op.getValueType()) ? Op.getValueType() : SVT);
12847 if (SVT != VT.getScalarType())
12848 for (SDValue &Op : Ops)
12849 Op = TLI.isZExtFree(Op.getValueType(), SVT)
12850 ? DAG.getZExtOrTrunc(Op, SDLoc(N), SVT)
12851 : DAG.getSExtOrTrunc(Op, SDLoc(N), SVT);
12852 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Ops);
12856 // If this shuffle only has a single input that is a bitcasted shuffle,
12857 // attempt to merge the 2 shuffles and suitably bitcast the inputs/output
12858 // back to their original types.
12859 if (N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
12860 N1.getOpcode() == ISD::UNDEF && Level < AfterLegalizeVectorOps &&
12861 TLI.isTypeLegal(VT)) {
12863 // Peek through the bitcast only if there is one user.
12865 while (BC0.getOpcode() == ISD::BITCAST) {
12866 if (!BC0.hasOneUse())
12868 BC0 = BC0.getOperand(0);
12871 auto ScaleShuffleMask = [](ArrayRef<int> Mask, int Scale) {
12873 return SmallVector<int, 8>(Mask.begin(), Mask.end());
12875 SmallVector<int, 8> NewMask;
12877 for (int s = 0; s != Scale; ++s)
12878 NewMask.push_back(M < 0 ? -1 : Scale * M + s);
12882 if (BC0.getOpcode() == ISD::VECTOR_SHUFFLE && BC0.hasOneUse()) {
12883 EVT SVT = VT.getScalarType();
12884 EVT InnerVT = BC0->getValueType(0);
12885 EVT InnerSVT = InnerVT.getScalarType();
12887 // Determine which shuffle works with the smaller scalar type.
12888 EVT ScaleVT = SVT.bitsLT(InnerSVT) ? VT : InnerVT;
12889 EVT ScaleSVT = ScaleVT.getScalarType();
12891 if (TLI.isTypeLegal(ScaleVT) &&
12892 0 == (InnerSVT.getSizeInBits() % ScaleSVT.getSizeInBits()) &&
12893 0 == (SVT.getSizeInBits() % ScaleSVT.getSizeInBits())) {
12895 int InnerScale = InnerSVT.getSizeInBits() / ScaleSVT.getSizeInBits();
12896 int OuterScale = SVT.getSizeInBits() / ScaleSVT.getSizeInBits();
12898 // Scale the shuffle masks to the smaller scalar type.
12899 ShuffleVectorSDNode *InnerSVN = cast<ShuffleVectorSDNode>(BC0);
12900 SmallVector<int, 8> InnerMask =
12901 ScaleShuffleMask(InnerSVN->getMask(), InnerScale);
12902 SmallVector<int, 8> OuterMask =
12903 ScaleShuffleMask(SVN->getMask(), OuterScale);
12905 // Merge the shuffle masks.
12906 SmallVector<int, 8> NewMask;
12907 for (int M : OuterMask)
12908 NewMask.push_back(M < 0 ? -1 : InnerMask[M]);
12910 // Test for shuffle mask legality over both commutations.
12911 SDValue SV0 = BC0->getOperand(0);
12912 SDValue SV1 = BC0->getOperand(1);
12913 bool LegalMask = TLI.isShuffleMaskLegal(NewMask, ScaleVT);
12915 std::swap(SV0, SV1);
12916 ShuffleVectorSDNode::commuteMask(NewMask);
12917 LegalMask = TLI.isShuffleMaskLegal(NewMask, ScaleVT);
12921 SV0 = DAG.getNode(ISD::BITCAST, SDLoc(N), ScaleVT, SV0);
12922 SV1 = DAG.getNode(ISD::BITCAST, SDLoc(N), ScaleVT, SV1);
12923 return DAG.getNode(
12924 ISD::BITCAST, SDLoc(N), VT,
12925 DAG.getVectorShuffle(ScaleVT, SDLoc(N), SV0, SV1, NewMask));
12931 // Canonicalize shuffles according to rules:
12932 // shuffle(A, shuffle(A, B)) -> shuffle(shuffle(A,B), A)
12933 // shuffle(B, shuffle(A, B)) -> shuffle(shuffle(A,B), B)
12934 // shuffle(B, shuffle(A, Undef)) -> shuffle(shuffle(A, Undef), B)
12935 if (N1.getOpcode() == ISD::VECTOR_SHUFFLE &&
12936 N0.getOpcode() != ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
12937 TLI.isTypeLegal(VT)) {
12938 // The incoming shuffle must be of the same type as the result of the
12939 // current shuffle.
12940 assert(N1->getOperand(0).getValueType() == VT &&
12941 "Shuffle types don't match");
12943 SDValue SV0 = N1->getOperand(0);
12944 SDValue SV1 = N1->getOperand(1);
12945 bool HasSameOp0 = N0 == SV0;
12946 bool IsSV1Undef = SV1.getOpcode() == ISD::UNDEF;
12947 if (HasSameOp0 || IsSV1Undef || N0 == SV1)
12948 // Commute the operands of this shuffle so that next rule
12950 return DAG.getCommutedVectorShuffle(*SVN);
12953 // Try to fold according to rules:
12954 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2)
12955 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2)
12956 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2)
12957 // Don't try to fold shuffles with illegal type.
12958 // Only fold if this shuffle is the only user of the other shuffle.
12959 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && N->isOnlyUserOf(N0.getNode()) &&
12960 Level < AfterLegalizeDAG && TLI.isTypeLegal(VT)) {
12961 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
12963 // The incoming shuffle must be of the same type as the result of the
12964 // current shuffle.
12965 assert(OtherSV->getOperand(0).getValueType() == VT &&
12966 "Shuffle types don't match");
12969 SmallVector<int, 4> Mask;
12970 // Compute the combined shuffle mask for a shuffle with SV0 as the first
12971 // operand, and SV1 as the second operand.
12972 for (unsigned i = 0; i != NumElts; ++i) {
12973 int Idx = SVN->getMaskElt(i);
12975 // Propagate Undef.
12976 Mask.push_back(Idx);
12980 SDValue CurrentVec;
12981 if (Idx < (int)NumElts) {
12982 // This shuffle index refers to the inner shuffle N0. Lookup the inner
12983 // shuffle mask to identify which vector is actually referenced.
12984 Idx = OtherSV->getMaskElt(Idx);
12986 // Propagate Undef.
12987 Mask.push_back(Idx);
12991 CurrentVec = (Idx < (int) NumElts) ? OtherSV->getOperand(0)
12992 : OtherSV->getOperand(1);
12994 // This shuffle index references an element within N1.
12998 // Simple case where 'CurrentVec' is UNDEF.
12999 if (CurrentVec.getOpcode() == ISD::UNDEF) {
13000 Mask.push_back(-1);
13004 // Canonicalize the shuffle index. We don't know yet if CurrentVec
13005 // will be the first or second operand of the combined shuffle.
13006 Idx = Idx % NumElts;
13007 if (!SV0.getNode() || SV0 == CurrentVec) {
13008 // Ok. CurrentVec is the left hand side.
13009 // Update the mask accordingly.
13011 Mask.push_back(Idx);
13015 // Bail out if we cannot convert the shuffle pair into a single shuffle.
13016 if (SV1.getNode() && SV1 != CurrentVec)
13019 // Ok. CurrentVec is the right hand side.
13020 // Update the mask accordingly.
13022 Mask.push_back(Idx + NumElts);
13025 // Check if all indices in Mask are Undef. In case, propagate Undef.
13026 bool isUndefMask = true;
13027 for (unsigned i = 0; i != NumElts && isUndefMask; ++i)
13028 isUndefMask &= Mask[i] < 0;
13031 return DAG.getUNDEF(VT);
13033 if (!SV0.getNode())
13034 SV0 = DAG.getUNDEF(VT);
13035 if (!SV1.getNode())
13036 SV1 = DAG.getUNDEF(VT);
13038 // Avoid introducing shuffles with illegal mask.
13039 if (!TLI.isShuffleMaskLegal(Mask, VT)) {
13040 ShuffleVectorSDNode::commuteMask(Mask);
13042 if (!TLI.isShuffleMaskLegal(Mask, VT))
13045 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, A, M2)
13046 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, A, M2)
13047 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, B, M2)
13048 std::swap(SV0, SV1);
13051 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2)
13052 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2)
13053 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2)
13054 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, &Mask[0]);
13060 SDValue DAGCombiner::visitSCALAR_TO_VECTOR(SDNode *N) {
13061 SDValue InVal = N->getOperand(0);
13062 EVT VT = N->getValueType(0);
13064 // Replace a SCALAR_TO_VECTOR(EXTRACT_VECTOR_ELT(V,C0)) pattern
13065 // with a VECTOR_SHUFFLE.
13066 if (InVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
13067 SDValue InVec = InVal->getOperand(0);
13068 SDValue EltNo = InVal->getOperand(1);
13070 // FIXME: We could support implicit truncation if the shuffle can be
13071 // scaled to a smaller vector scalar type.
13072 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(EltNo);
13073 if (C0 && VT == InVec.getValueType() &&
13074 VT.getScalarType() == InVal.getValueType()) {
13075 SmallVector<int, 8> NewMask(VT.getVectorNumElements(), -1);
13076 int Elt = C0->getZExtValue();
13079 if (TLI.isShuffleMaskLegal(NewMask, VT))
13080 return DAG.getVectorShuffle(VT, SDLoc(N), InVec, DAG.getUNDEF(VT),
13088 SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
13089 SDValue N0 = N->getOperand(0);
13090 SDValue N2 = N->getOperand(2);
13092 // If the input vector is a concatenation, and the insert replaces
13093 // one of the halves, we can optimize into a single concat_vectors.
13094 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
13095 N0->getNumOperands() == 2 && N2.getOpcode() == ISD::Constant) {
13096 APInt InsIdx = cast<ConstantSDNode>(N2)->getAPIntValue();
13097 EVT VT = N->getValueType(0);
13099 // Lower half: fold (insert_subvector (concat_vectors X, Y), Z) ->
13100 // (concat_vectors Z, Y)
13102 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
13103 N->getOperand(1), N0.getOperand(1));
13105 // Upper half: fold (insert_subvector (concat_vectors X, Y), Z) ->
13106 // (concat_vectors X, Z)
13107 if (InsIdx == VT.getVectorNumElements()/2)
13108 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
13109 N0.getOperand(0), N->getOperand(1));
13115 SDValue DAGCombiner::visitFP_TO_FP16(SDNode *N) {
13116 SDValue N0 = N->getOperand(0);
13118 // fold (fp_to_fp16 (fp16_to_fp op)) -> op
13119 if (N0->getOpcode() == ISD::FP16_TO_FP)
13120 return N0->getOperand(0);
13125 SDValue DAGCombiner::visitFP16_TO_FP(SDNode *N) {
13126 SDValue N0 = N->getOperand(0);
13128 // fold fp16_to_fp(op & 0xffff) -> fp16_to_fp(op)
13129 if (N0->getOpcode() == ISD::AND) {
13130 ConstantSDNode *AndConst = getAsNonOpaqueConstant(N0.getOperand(1));
13131 if (AndConst && AndConst->getAPIntValue() == 0xffff) {
13132 return DAG.getNode(ISD::FP16_TO_FP, SDLoc(N), N->getValueType(0),
13140 /// Returns a vector_shuffle if it able to transform an AND to a vector_shuffle
13141 /// with the destination vector and a zero vector.
13142 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
13143 /// vector_shuffle V, Zero, <0, 4, 2, 4>
13144 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
13145 EVT VT = N->getValueType(0);
13146 SDValue LHS = N->getOperand(0);
13147 SDValue RHS = N->getOperand(1);
13150 // Make sure we're not running after operation legalization where it
13151 // may have custom lowered the vector shuffles.
13152 if (LegalOperations)
13155 if (N->getOpcode() != ISD::AND)
13158 if (RHS.getOpcode() == ISD::BITCAST)
13159 RHS = RHS.getOperand(0);
13161 if (RHS.getOpcode() != ISD::BUILD_VECTOR)
13164 EVT RVT = RHS.getValueType();
13165 unsigned NumElts = RHS.getNumOperands();
13167 // Attempt to create a valid clear mask, splitting the mask into
13168 // sub elements and checking to see if each is
13169 // all zeros or all ones - suitable for shuffle masking.
13170 auto BuildClearMask = [&](int Split) {
13171 int NumSubElts = NumElts * Split;
13172 int NumSubBits = RVT.getScalarSizeInBits() / Split;
13174 SmallVector<int, 8> Indices;
13175 for (int i = 0; i != NumSubElts; ++i) {
13176 int EltIdx = i / Split;
13177 int SubIdx = i % Split;
13178 SDValue Elt = RHS.getOperand(EltIdx);
13179 if (Elt.getOpcode() == ISD::UNDEF) {
13180 Indices.push_back(-1);
13185 if (isa<ConstantSDNode>(Elt))
13186 Bits = cast<ConstantSDNode>(Elt)->getAPIntValue();
13187 else if (isa<ConstantFPSDNode>(Elt))
13188 Bits = cast<ConstantFPSDNode>(Elt)->getValueAPF().bitcastToAPInt();
13192 // Extract the sub element from the constant bit mask.
13193 if (DAG.getDataLayout().isBigEndian()) {
13194 Bits = Bits.lshr((Split - SubIdx - 1) * NumSubBits);
13196 Bits = Bits.lshr(SubIdx * NumSubBits);
13200 Bits = Bits.trunc(NumSubBits);
13202 if (Bits.isAllOnesValue())
13203 Indices.push_back(i);
13204 else if (Bits == 0)
13205 Indices.push_back(i + NumSubElts);
13210 // Let's see if the target supports this vector_shuffle.
13211 EVT ClearSVT = EVT::getIntegerVT(*DAG.getContext(), NumSubBits);
13212 EVT ClearVT = EVT::getVectorVT(*DAG.getContext(), ClearSVT, NumSubElts);
13213 if (!TLI.isVectorClearMaskLegal(Indices, ClearVT))
13216 SDValue Zero = DAG.getConstant(0, dl, ClearVT);
13217 return DAG.getBitcast(VT, DAG.getVectorShuffle(ClearVT, dl,
13218 DAG.getBitcast(ClearVT, LHS),
13219 Zero, &Indices[0]));
13222 // Determine maximum split level (byte level masking).
13224 if (RVT.getScalarSizeInBits() % 8 == 0)
13225 MaxSplit = RVT.getScalarSizeInBits() / 8;
13227 for (int Split = 1; Split <= MaxSplit; ++Split)
13228 if (RVT.getScalarSizeInBits() % Split == 0)
13229 if (SDValue S = BuildClearMask(Split))
13235 /// Visit a binary vector operation, like ADD.
13236 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
13237 assert(N->getValueType(0).isVector() &&
13238 "SimplifyVBinOp only works on vectors!");
13240 SDValue LHS = N->getOperand(0);
13241 SDValue RHS = N->getOperand(1);
13243 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
13245 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
13246 RHS.getOpcode() == ISD::BUILD_VECTOR) {
13247 // Check if both vectors are constants. If not bail out.
13248 if (!(cast<BuildVectorSDNode>(LHS)->isConstant() &&
13249 cast<BuildVectorSDNode>(RHS)->isConstant()))
13252 SmallVector<SDValue, 8> Ops;
13253 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
13254 SDValue LHSOp = LHS.getOperand(i);
13255 SDValue RHSOp = RHS.getOperand(i);
13257 // Can't fold divide by zero.
13258 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
13259 N->getOpcode() == ISD::FDIV) {
13260 if (isNullConstant(RHSOp) || (RHSOp.getOpcode() == ISD::ConstantFP &&
13261 cast<ConstantFPSDNode>(RHSOp.getNode())->isZero()))
13265 EVT VT = LHSOp.getValueType();
13266 EVT RVT = RHSOp.getValueType();
13268 // Integer BUILD_VECTOR operands may have types larger than the element
13269 // size (e.g., when the element type is not legal). Prior to type
13270 // legalization, the types may not match between the two BUILD_VECTORS.
13271 // Truncate one of the operands to make them match.
13272 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
13273 RHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, RHSOp);
13275 LHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), RVT, LHSOp);
13279 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(LHS), VT,
13281 if (FoldOp.getOpcode() != ISD::UNDEF &&
13282 FoldOp.getOpcode() != ISD::Constant &&
13283 FoldOp.getOpcode() != ISD::ConstantFP)
13285 Ops.push_back(FoldOp);
13286 AddToWorklist(FoldOp.getNode());
13289 if (Ops.size() == LHS.getNumOperands())
13290 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), LHS.getValueType(), Ops);
13293 // Try to convert a constant mask AND into a shuffle clear mask.
13294 if (SDValue Shuffle = XformToShuffleWithZero(N))
13297 // Type legalization might introduce new shuffles in the DAG.
13298 // Fold (VBinOp (shuffle (A, Undef, Mask)), (shuffle (B, Undef, Mask)))
13299 // -> (shuffle (VBinOp (A, B)), Undef, Mask).
13300 if (LegalTypes && isa<ShuffleVectorSDNode>(LHS) &&
13301 isa<ShuffleVectorSDNode>(RHS) && LHS.hasOneUse() && RHS.hasOneUse() &&
13302 LHS.getOperand(1).getOpcode() == ISD::UNDEF &&
13303 RHS.getOperand(1).getOpcode() == ISD::UNDEF) {
13304 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(LHS);
13305 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(RHS);
13307 if (SVN0->getMask().equals(SVN1->getMask())) {
13308 EVT VT = N->getValueType(0);
13309 SDValue UndefVector = LHS.getOperand(1);
13310 SDValue NewBinOp = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
13311 LHS.getOperand(0), RHS.getOperand(0));
13312 AddUsersToWorklist(N);
13313 return DAG.getVectorShuffle(VT, SDLoc(N), NewBinOp, UndefVector,
13314 &SVN0->getMask()[0]);
13321 SDValue DAGCombiner::SimplifySelect(SDLoc DL, SDValue N0,
13322 SDValue N1, SDValue N2){
13323 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
13325 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
13326 cast<CondCodeSDNode>(N0.getOperand(2))->get());
13328 // If we got a simplified select_cc node back from SimplifySelectCC, then
13329 // break it down into a new SETCC node, and a new SELECT node, and then return
13330 // the SELECT node, since we were called with a SELECT node.
13331 if (SCC.getNode()) {
13332 // Check to see if we got a select_cc back (to turn into setcc/select).
13333 // Otherwise, just return whatever node we got back, like fabs.
13334 if (SCC.getOpcode() == ISD::SELECT_CC) {
13335 SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0),
13337 SCC.getOperand(0), SCC.getOperand(1),
13338 SCC.getOperand(4));
13339 AddToWorklist(SETCC.getNode());
13340 return DAG.getSelect(SDLoc(SCC), SCC.getValueType(), SETCC,
13341 SCC.getOperand(2), SCC.getOperand(3));
13349 /// Given a SELECT or a SELECT_CC node, where LHS and RHS are the two values
13350 /// being selected between, see if we can simplify the select. Callers of this
13351 /// should assume that TheSelect is deleted if this returns true. As such, they
13352 /// should return the appropriate thing (e.g. the node) back to the top-level of
13353 /// the DAG combiner loop to avoid it being looked at.
13354 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
13357 // fold (select (setcc x, -0.0, *lt), NaN, (fsqrt x))
13358 // The select + setcc is redundant, because fsqrt returns NaN for X < -0.
13359 if (const ConstantFPSDNode *NaN = isConstOrConstSplatFP(LHS)) {
13360 if (NaN->isNaN() && RHS.getOpcode() == ISD::FSQRT) {
13361 // We have: (select (setcc ?, ?, ?), NaN, (fsqrt ?))
13362 SDValue Sqrt = RHS;
13365 const ConstantFPSDNode *NegZero = nullptr;
13367 if (TheSelect->getOpcode() == ISD::SELECT_CC) {
13368 CC = dyn_cast<CondCodeSDNode>(TheSelect->getOperand(4))->get();
13369 CmpLHS = TheSelect->getOperand(0);
13370 NegZero = isConstOrConstSplatFP(TheSelect->getOperand(1));
13372 // SELECT or VSELECT
13373 SDValue Cmp = TheSelect->getOperand(0);
13374 if (Cmp.getOpcode() == ISD::SETCC) {
13375 CC = dyn_cast<CondCodeSDNode>(Cmp.getOperand(2))->get();
13376 CmpLHS = Cmp.getOperand(0);
13377 NegZero = isConstOrConstSplatFP(Cmp.getOperand(1));
13380 if (NegZero && NegZero->isNegative() && NegZero->isZero() &&
13381 Sqrt.getOperand(0) == CmpLHS && (CC == ISD::SETOLT ||
13382 CC == ISD::SETULT || CC == ISD::SETLT)) {
13383 // We have: (select (setcc x, -0.0, *lt), NaN, (fsqrt x))
13384 CombineTo(TheSelect, Sqrt);
13389 // Cannot simplify select with vector condition
13390 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
13392 // If this is a select from two identical things, try to pull the operation
13393 // through the select.
13394 if (LHS.getOpcode() != RHS.getOpcode() ||
13395 !LHS.hasOneUse() || !RHS.hasOneUse())
13398 // If this is a load and the token chain is identical, replace the select
13399 // of two loads with a load through a select of the address to load from.
13400 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
13401 // constants have been dropped into the constant pool.
13402 if (LHS.getOpcode() == ISD::LOAD) {
13403 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
13404 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
13406 // Token chains must be identical.
13407 if (LHS.getOperand(0) != RHS.getOperand(0) ||
13408 // Do not let this transformation reduce the number of volatile loads.
13409 LLD->isVolatile() || RLD->isVolatile() ||
13410 // FIXME: If either is a pre/post inc/dec load,
13411 // we'd need to split out the address adjustment.
13412 LLD->isIndexed() || RLD->isIndexed() ||
13413 // If this is an EXTLOAD, the VT's must match.
13414 LLD->getMemoryVT() != RLD->getMemoryVT() ||
13415 // If this is an EXTLOAD, the kind of extension must match.
13416 (LLD->getExtensionType() != RLD->getExtensionType() &&
13417 // The only exception is if one of the extensions is anyext.
13418 LLD->getExtensionType() != ISD::EXTLOAD &&
13419 RLD->getExtensionType() != ISD::EXTLOAD) ||
13420 // FIXME: this discards src value information. This is
13421 // over-conservative. It would be beneficial to be able to remember
13422 // both potential memory locations. Since we are discarding
13423 // src value info, don't do the transformation if the memory
13424 // locations are not in the default address space.
13425 LLD->getPointerInfo().getAddrSpace() != 0 ||
13426 RLD->getPointerInfo().getAddrSpace() != 0 ||
13427 !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(),
13428 LLD->getBasePtr().getValueType()))
13431 // Check that the select condition doesn't reach either load. If so,
13432 // folding this will induce a cycle into the DAG. If not, this is safe to
13433 // xform, so create a select of the addresses.
13435 if (TheSelect->getOpcode() == ISD::SELECT) {
13436 SDNode *CondNode = TheSelect->getOperand(0).getNode();
13437 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
13438 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
13440 // The loads must not depend on one another.
13441 if (LLD->isPredecessorOf(RLD) ||
13442 RLD->isPredecessorOf(LLD))
13444 Addr = DAG.getSelect(SDLoc(TheSelect),
13445 LLD->getBasePtr().getValueType(),
13446 TheSelect->getOperand(0), LLD->getBasePtr(),
13447 RLD->getBasePtr());
13448 } else { // Otherwise SELECT_CC
13449 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
13450 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
13452 if ((LLD->hasAnyUseOfValue(1) &&
13453 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
13454 (RLD->hasAnyUseOfValue(1) &&
13455 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
13458 Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect),
13459 LLD->getBasePtr().getValueType(),
13460 TheSelect->getOperand(0),
13461 TheSelect->getOperand(1),
13462 LLD->getBasePtr(), RLD->getBasePtr(),
13463 TheSelect->getOperand(4));
13467 // It is safe to replace the two loads if they have different alignments,
13468 // but the new load must be the minimum (most restrictive) alignment of the
13470 bool isInvariant = LLD->isInvariant() & RLD->isInvariant();
13471 unsigned Alignment = std::min(LLD->getAlignment(), RLD->getAlignment());
13472 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
13473 Load = DAG.getLoad(TheSelect->getValueType(0),
13475 // FIXME: Discards pointer and AA info.
13476 LLD->getChain(), Addr, MachinePointerInfo(),
13477 LLD->isVolatile(), LLD->isNonTemporal(),
13478 isInvariant, Alignment);
13480 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
13481 RLD->getExtensionType() : LLD->getExtensionType(),
13483 TheSelect->getValueType(0),
13484 // FIXME: Discards pointer and AA info.
13485 LLD->getChain(), Addr, MachinePointerInfo(),
13486 LLD->getMemoryVT(), LLD->isVolatile(),
13487 LLD->isNonTemporal(), isInvariant, Alignment);
13490 // Users of the select now use the result of the load.
13491 CombineTo(TheSelect, Load);
13493 // Users of the old loads now use the new load's chain. We know the
13494 // old-load value is dead now.
13495 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
13496 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
13503 /// Simplify an expression of the form (N0 cond N1) ? N2 : N3
13504 /// where 'cond' is the comparison specified by CC.
13505 SDValue DAGCombiner::SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1,
13506 SDValue N2, SDValue N3,
13507 ISD::CondCode CC, bool NotExtCompare) {
13508 // (x ? y : y) -> y.
13509 if (N2 == N3) return N2;
13511 EVT VT = N2.getValueType();
13512 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
13513 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
13515 // Determine if the condition we're dealing with is constant
13516 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
13517 N0, N1, CC, DL, false);
13518 if (SCC.getNode()) AddToWorklist(SCC.getNode());
13520 if (ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode())) {
13521 // fold select_cc true, x, y -> x
13522 // fold select_cc false, x, y -> y
13523 return !SCCC->isNullValue() ? N2 : N3;
13526 // Check to see if we can simplify the select into an fabs node
13527 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
13528 // Allow either -0.0 or 0.0
13529 if (CFP->isZero()) {
13530 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
13531 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
13532 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
13533 N2 == N3.getOperand(0))
13534 return DAG.getNode(ISD::FABS, DL, VT, N0);
13536 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
13537 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
13538 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
13539 N2.getOperand(0) == N3)
13540 return DAG.getNode(ISD::FABS, DL, VT, N3);
13544 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
13545 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
13546 // in it. This is a win when the constant is not otherwise available because
13547 // it replaces two constant pool loads with one. We only do this if the FP
13548 // type is known to be legal, because if it isn't, then we are before legalize
13549 // types an we want the other legalization to happen first (e.g. to avoid
13550 // messing with soft float) and if the ConstantFP is not legal, because if
13551 // it is legal, we may not need to store the FP constant in a constant pool.
13552 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
13553 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
13554 if (TLI.isTypeLegal(N2.getValueType()) &&
13555 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
13556 TargetLowering::Legal &&
13557 !TLI.isFPImmLegal(TV->getValueAPF(), TV->getValueType(0)) &&
13558 !TLI.isFPImmLegal(FV->getValueAPF(), FV->getValueType(0))) &&
13559 // If both constants have multiple uses, then we won't need to do an
13560 // extra load, they are likely around in registers for other users.
13561 (TV->hasOneUse() || FV->hasOneUse())) {
13562 Constant *Elts[] = {
13563 const_cast<ConstantFP*>(FV->getConstantFPValue()),
13564 const_cast<ConstantFP*>(TV->getConstantFPValue())
13566 Type *FPTy = Elts[0]->getType();
13567 const DataLayout &TD = DAG.getDataLayout();
13569 // Create a ConstantArray of the two constants.
13570 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
13572 DAG.getConstantPool(CA, TLI.getPointerTy(DAG.getDataLayout()),
13573 TD.getPrefTypeAlignment(FPTy));
13574 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
13576 // Get the offsets to the 0 and 1 element of the array so that we can
13577 // select between them.
13578 SDValue Zero = DAG.getIntPtrConstant(0, DL);
13579 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
13580 SDValue One = DAG.getIntPtrConstant(EltSize, SDLoc(FV));
13582 SDValue Cond = DAG.getSetCC(DL,
13583 getSetCCResultType(N0.getValueType()),
13585 AddToWorklist(Cond.getNode());
13586 SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(),
13588 AddToWorklist(CstOffset.getNode());
13589 CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx,
13591 AddToWorklist(CPIdx.getNode());
13592 return DAG.getLoad(
13593 TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
13594 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13595 false, false, false, Alignment);
13599 // Check to see if we can perform the "gzip trick", transforming
13600 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
13601 if (isNullConstant(N3) && CC == ISD::SETLT &&
13602 (isNullConstant(N1) || // (a < 0) ? b : 0
13603 (isOneConstant(N1) && N0 == N2))) { // (a < 1) ? a : 0
13604 EVT XType = N0.getValueType();
13605 EVT AType = N2.getValueType();
13606 if (XType.bitsGE(AType)) {
13607 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
13608 // single-bit constant.
13609 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue() - 1)) == 0)) {
13610 unsigned ShCtV = N2C->getAPIntValue().logBase2();
13611 ShCtV = XType.getSizeInBits() - ShCtV - 1;
13612 SDValue ShCt = DAG.getConstant(ShCtV, SDLoc(N0),
13613 getShiftAmountTy(N0.getValueType()));
13614 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0),
13616 AddToWorklist(Shift.getNode());
13618 if (XType.bitsGT(AType)) {
13619 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
13620 AddToWorklist(Shift.getNode());
13623 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
13626 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0),
13628 DAG.getConstant(XType.getSizeInBits() - 1,
13630 getShiftAmountTy(N0.getValueType())));
13631 AddToWorklist(Shift.getNode());
13633 if (XType.bitsGT(AType)) {
13634 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
13635 AddToWorklist(Shift.getNode());
13638 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
13642 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
13643 // where y is has a single bit set.
13644 // A plaintext description would be, we can turn the SELECT_CC into an AND
13645 // when the condition can be materialized as an all-ones register. Any
13646 // single bit-test can be materialized as an all-ones register with
13647 // shift-left and shift-right-arith.
13648 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
13649 N0->getValueType(0) == VT && isNullConstant(N1) && isNullConstant(N2)) {
13650 SDValue AndLHS = N0->getOperand(0);
13651 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
13652 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
13653 // Shift the tested bit over the sign bit.
13654 APInt AndMask = ConstAndRHS->getAPIntValue();
13656 DAG.getConstant(AndMask.countLeadingZeros(), SDLoc(AndLHS),
13657 getShiftAmountTy(AndLHS.getValueType()));
13658 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt);
13660 // Now arithmetic right shift it all the way over, so the result is either
13661 // all-ones, or zero.
13663 DAG.getConstant(AndMask.getBitWidth() - 1, SDLoc(Shl),
13664 getShiftAmountTy(Shl.getValueType()));
13665 SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt);
13667 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
13671 // fold select C, 16, 0 -> shl C, 4
13672 if (N2C && isNullConstant(N3) && N2C->getAPIntValue().isPowerOf2() &&
13673 TLI.getBooleanContents(N0.getValueType()) ==
13674 TargetLowering::ZeroOrOneBooleanContent) {
13676 // If the caller doesn't want us to simplify this into a zext of a compare,
13678 if (NotExtCompare && N2C->isOne())
13681 // Get a SetCC of the condition
13682 // NOTE: Don't create a SETCC if it's not legal on this target.
13683 if (!LegalOperations ||
13684 TLI.isOperationLegal(ISD::SETCC,
13685 LegalTypes ? getSetCCResultType(N0.getValueType()) : MVT::i1)) {
13687 // cast from setcc result type to select result type
13689 SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()),
13691 if (N2.getValueType().bitsLT(SCC.getValueType()))
13692 Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2),
13693 N2.getValueType());
13695 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
13696 N2.getValueType(), SCC);
13698 SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC);
13699 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
13700 N2.getValueType(), SCC);
13703 AddToWorklist(SCC.getNode());
13704 AddToWorklist(Temp.getNode());
13709 // shl setcc result by log2 n2c
13710 return DAG.getNode(
13711 ISD::SHL, DL, N2.getValueType(), Temp,
13712 DAG.getConstant(N2C->getAPIntValue().logBase2(), SDLoc(Temp),
13713 getShiftAmountTy(Temp.getValueType())));
13717 // Check to see if this is the equivalent of setcc
13718 // FIXME: Turn all of these into setcc if setcc if setcc is legal
13719 // otherwise, go ahead with the folds.
13720 if (0 && isNullConstant(N3) && isOneConstant(N2)) {
13721 EVT XType = N0.getValueType();
13722 if (!LegalOperations ||
13723 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(XType))) {
13724 SDValue Res = DAG.getSetCC(DL, getSetCCResultType(XType), N0, N1, CC);
13725 if (Res.getValueType() != VT)
13726 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
13730 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
13731 if (isNullConstant(N1) && CC == ISD::SETEQ &&
13732 (!LegalOperations ||
13733 TLI.isOperationLegal(ISD::CTLZ, XType))) {
13734 SDValue Ctlz = DAG.getNode(ISD::CTLZ, SDLoc(N0), XType, N0);
13735 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
13736 DAG.getConstant(Log2_32(XType.getSizeInBits()),
13738 getShiftAmountTy(Ctlz.getValueType())));
13740 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
13741 if (isNullConstant(N1) && CC == ISD::SETGT) {
13743 SDValue NegN0 = DAG.getNode(ISD::SUB, DL,
13744 XType, DAG.getConstant(0, DL, XType), N0);
13745 SDValue NotN0 = DAG.getNOT(DL, N0, XType);
13746 return DAG.getNode(ISD::SRL, DL, XType,
13747 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
13748 DAG.getConstant(XType.getSizeInBits() - 1, DL,
13749 getShiftAmountTy(XType)));
13751 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
13752 if (isAllOnesConstant(N1) && CC == ISD::SETGT) {
13754 SDValue Sign = DAG.getNode(ISD::SRL, DL, XType, N0,
13755 DAG.getConstant(XType.getSizeInBits() - 1, DL,
13756 getShiftAmountTy(N0.getValueType())));
13757 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, DL,
13762 // Check to see if this is an integer abs.
13763 // select_cc setg[te] X, 0, X, -X ->
13764 // select_cc setgt X, -1, X, -X ->
13765 // select_cc setl[te] X, 0, -X, X ->
13766 // select_cc setlt X, 1, -X, X ->
13767 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
13769 ConstantSDNode *SubC = nullptr;
13770 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
13771 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
13772 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
13773 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
13774 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
13775 (N1C->isOne() && CC == ISD::SETLT)) &&
13776 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
13777 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
13779 EVT XType = N0.getValueType();
13780 if (SubC && SubC->isNullValue() && XType.isInteger()) {
13782 SDValue Shift = DAG.getNode(ISD::SRA, DL, XType,
13784 DAG.getConstant(XType.getSizeInBits() - 1, DL,
13785 getShiftAmountTy(N0.getValueType())));
13786 SDValue Add = DAG.getNode(ISD::ADD, DL,
13788 AddToWorklist(Shift.getNode());
13789 AddToWorklist(Add.getNode());
13790 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
13797 /// This is a stub for TargetLowering::SimplifySetCC.
13798 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
13799 SDValue N1, ISD::CondCode Cond,
13800 SDLoc DL, bool foldBooleans) {
13801 TargetLowering::DAGCombinerInfo
13802 DagCombineInfo(DAG, Level, false, this);
13803 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
13806 /// Given an ISD::SDIV node expressing a divide by constant, return
13807 /// a DAG expression to select that will generate the same value by multiplying
13808 /// by a magic number.
13809 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
13810 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
13811 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
13815 // Avoid division by zero.
13816 if (C->isNullValue())
13819 std::vector<SDNode*> Built;
13821 TLI.BuildSDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
13823 for (SDNode *N : Built)
13828 /// Given an ISD::SDIV node expressing a divide by constant power of 2, return a
13829 /// DAG expression that will generate the same value by right shifting.
13830 SDValue DAGCombiner::BuildSDIVPow2(SDNode *N) {
13831 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
13835 // Avoid division by zero.
13836 if (C->isNullValue())
13839 std::vector<SDNode *> Built;
13840 SDValue S = TLI.BuildSDIVPow2(N, C->getAPIntValue(), DAG, &Built);
13842 for (SDNode *N : Built)
13847 /// Given an ISD::UDIV node expressing a divide by constant, return a DAG
13848 /// expression that will generate the same value by multiplying by a magic
13850 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
13851 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
13852 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
13856 // Avoid division by zero.
13857 if (C->isNullValue())
13860 std::vector<SDNode*> Built;
13862 TLI.BuildUDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
13864 for (SDNode *N : Built)
13869 SDValue DAGCombiner::BuildReciprocalEstimate(SDValue Op) {
13870 if (Level >= AfterLegalizeDAG)
13873 // Expose the DAG combiner to the target combiner implementations.
13874 TargetLowering::DAGCombinerInfo DCI(DAG, Level, false, this);
13876 unsigned Iterations = 0;
13877 if (SDValue Est = TLI.getRecipEstimate(Op, DCI, Iterations)) {
13879 // Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
13880 // For the reciprocal, we need to find the zero of the function:
13881 // F(X) = A X - 1 [which has a zero at X = 1/A]
13883 // X_{i+1} = X_i (2 - A X_i) = X_i + X_i (1 - A X_i) [this second form
13884 // does not require additional intermediate precision]
13885 EVT VT = Op.getValueType();
13887 SDValue FPOne = DAG.getConstantFP(1.0, DL, VT);
13889 AddToWorklist(Est.getNode());
13891 // Newton iterations: Est = Est + Est (1 - Arg * Est)
13892 for (unsigned i = 0; i < Iterations; ++i) {
13893 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Op, Est);
13894 AddToWorklist(NewEst.getNode());
13896 NewEst = DAG.getNode(ISD::FSUB, DL, VT, FPOne, NewEst);
13897 AddToWorklist(NewEst.getNode());
13899 NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst);
13900 AddToWorklist(NewEst.getNode());
13902 Est = DAG.getNode(ISD::FADD, DL, VT, Est, NewEst);
13903 AddToWorklist(Est.getNode());
13912 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
13913 /// For the reciprocal sqrt, we need to find the zero of the function:
13914 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
13916 /// X_{i+1} = X_i (1.5 - A X_i^2 / 2)
13917 /// As a result, we precompute A/2 prior to the iteration loop.
13918 SDValue DAGCombiner::BuildRsqrtNROneConst(SDValue Arg, SDValue Est,
13919 unsigned Iterations) {
13920 EVT VT = Arg.getValueType();
13922 SDValue ThreeHalves = DAG.getConstantFP(1.5, DL, VT);
13924 // We now need 0.5 * Arg which we can write as (1.5 * Arg - Arg) so that
13925 // this entire sequence requires only one FP constant.
13926 SDValue HalfArg = DAG.getNode(ISD::FMUL, DL, VT, ThreeHalves, Arg);
13927 AddToWorklist(HalfArg.getNode());
13929 HalfArg = DAG.getNode(ISD::FSUB, DL, VT, HalfArg, Arg);
13930 AddToWorklist(HalfArg.getNode());
13932 // Newton iterations: Est = Est * (1.5 - HalfArg * Est * Est)
13933 for (unsigned i = 0; i < Iterations; ++i) {
13934 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, Est);
13935 AddToWorklist(NewEst.getNode());
13937 NewEst = DAG.getNode(ISD::FMUL, DL, VT, HalfArg, NewEst);
13938 AddToWorklist(NewEst.getNode());
13940 NewEst = DAG.getNode(ISD::FSUB, DL, VT, ThreeHalves, NewEst);
13941 AddToWorklist(NewEst.getNode());
13943 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst);
13944 AddToWorklist(Est.getNode());
13949 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
13950 /// For the reciprocal sqrt, we need to find the zero of the function:
13951 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
13953 /// X_{i+1} = (-0.5 * X_i) * (A * X_i * X_i + (-3.0))
13954 SDValue DAGCombiner::BuildRsqrtNRTwoConst(SDValue Arg, SDValue Est,
13955 unsigned Iterations) {
13956 EVT VT = Arg.getValueType();
13958 SDValue MinusThree = DAG.getConstantFP(-3.0, DL, VT);
13959 SDValue MinusHalf = DAG.getConstantFP(-0.5, DL, VT);
13961 // Newton iterations: Est = -0.5 * Est * (-3.0 + Arg * Est * Est)
13962 for (unsigned i = 0; i < Iterations; ++i) {
13963 SDValue HalfEst = DAG.getNode(ISD::FMUL, DL, VT, Est, MinusHalf);
13964 AddToWorklist(HalfEst.getNode());
13966 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Est);
13967 AddToWorklist(Est.getNode());
13969 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Arg);
13970 AddToWorklist(Est.getNode());
13972 Est = DAG.getNode(ISD::FADD, DL, VT, Est, MinusThree);
13973 AddToWorklist(Est.getNode());
13975 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, HalfEst);
13976 AddToWorklist(Est.getNode());
13981 SDValue DAGCombiner::BuildRsqrtEstimate(SDValue Op) {
13982 if (Level >= AfterLegalizeDAG)
13985 // Expose the DAG combiner to the target combiner implementations.
13986 TargetLowering::DAGCombinerInfo DCI(DAG, Level, false, this);
13987 unsigned Iterations = 0;
13988 bool UseOneConstNR = false;
13989 if (SDValue Est = TLI.getRsqrtEstimate(Op, DCI, Iterations, UseOneConstNR)) {
13990 AddToWorklist(Est.getNode());
13992 Est = UseOneConstNR ?
13993 BuildRsqrtNROneConst(Op, Est, Iterations) :
13994 BuildRsqrtNRTwoConst(Op, Est, Iterations);
14002 /// Return true if base is a frame index, which is known not to alias with
14003 /// anything but itself. Provides base object and offset as results.
14004 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
14005 const GlobalValue *&GV, const void *&CV) {
14006 // Assume it is a primitive operation.
14007 Base = Ptr; Offset = 0; GV = nullptr; CV = nullptr;
14009 // If it's an adding a simple constant then integrate the offset.
14010 if (Base.getOpcode() == ISD::ADD) {
14011 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
14012 Base = Base.getOperand(0);
14013 Offset += C->getZExtValue();
14017 // Return the underlying GlobalValue, and update the Offset. Return false
14018 // for GlobalAddressSDNode since the same GlobalAddress may be represented
14019 // by multiple nodes with different offsets.
14020 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
14021 GV = G->getGlobal();
14022 Offset += G->getOffset();
14026 // Return the underlying Constant value, and update the Offset. Return false
14027 // for ConstantSDNodes since the same constant pool entry may be represented
14028 // by multiple nodes with different offsets.
14029 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
14030 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
14031 : (const void *)C->getConstVal();
14032 Offset += C->getOffset();
14035 // If it's any of the following then it can't alias with anything but itself.
14036 return isa<FrameIndexSDNode>(Base);
14039 /// Return true if there is any possibility that the two addresses overlap.
14040 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const {
14041 // If they are the same then they must be aliases.
14042 if (Op0->getBasePtr() == Op1->getBasePtr()) return true;
14044 // If they are both volatile then they cannot be reordered.
14045 if (Op0->isVolatile() && Op1->isVolatile()) return true;
14047 // If one operation reads from invariant memory, and the other may store, they
14048 // cannot alias. These should really be checking the equivalent of mayWrite,
14049 // but it only matters for memory nodes other than load /store.
14050 if (Op0->isInvariant() && Op1->writeMem())
14053 if (Op1->isInvariant() && Op0->writeMem())
14056 // Gather base node and offset information.
14057 SDValue Base1, Base2;
14058 int64_t Offset1, Offset2;
14059 const GlobalValue *GV1, *GV2;
14060 const void *CV1, *CV2;
14061 bool isFrameIndex1 = FindBaseOffset(Op0->getBasePtr(),
14062 Base1, Offset1, GV1, CV1);
14063 bool isFrameIndex2 = FindBaseOffset(Op1->getBasePtr(),
14064 Base2, Offset2, GV2, CV2);
14066 // If they have a same base address then check to see if they overlap.
14067 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
14068 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
14069 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
14071 // It is possible for different frame indices to alias each other, mostly
14072 // when tail call optimization reuses return address slots for arguments.
14073 // To catch this case, look up the actual index of frame indices to compute
14074 // the real alias relationship.
14075 if (isFrameIndex1 && isFrameIndex2) {
14076 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
14077 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
14078 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
14079 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
14080 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
14083 // Otherwise, if we know what the bases are, and they aren't identical, then
14084 // we know they cannot alias.
14085 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
14088 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
14089 // compared to the size and offset of the access, we may be able to prove they
14090 // do not alias. This check is conservative for now to catch cases created by
14091 // splitting vector types.
14092 if ((Op0->getOriginalAlignment() == Op1->getOriginalAlignment()) &&
14093 (Op0->getSrcValueOffset() != Op1->getSrcValueOffset()) &&
14094 (Op0->getMemoryVT().getSizeInBits() >> 3 ==
14095 Op1->getMemoryVT().getSizeInBits() >> 3) &&
14096 (Op0->getOriginalAlignment() > Op0->getMemoryVT().getSizeInBits()) >> 3) {
14097 int64_t OffAlign1 = Op0->getSrcValueOffset() % Op0->getOriginalAlignment();
14098 int64_t OffAlign2 = Op1->getSrcValueOffset() % Op1->getOriginalAlignment();
14100 // There is no overlap between these relatively aligned accesses of similar
14101 // size, return no alias.
14102 if ((OffAlign1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign2 ||
14103 (OffAlign2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign1)
14107 bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0
14109 : DAG.getSubtarget().useAA();
14111 if (CombinerAAOnlyFunc.getNumOccurrences() &&
14112 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
14116 Op0->getMemOperand()->getValue() && Op1->getMemOperand()->getValue()) {
14117 // Use alias analysis information.
14118 int64_t MinOffset = std::min(Op0->getSrcValueOffset(),
14119 Op1->getSrcValueOffset());
14120 int64_t Overlap1 = (Op0->getMemoryVT().getSizeInBits() >> 3) +
14121 Op0->getSrcValueOffset() - MinOffset;
14122 int64_t Overlap2 = (Op1->getMemoryVT().getSizeInBits() >> 3) +
14123 Op1->getSrcValueOffset() - MinOffset;
14124 AliasResult AAResult =
14125 AA.alias(MemoryLocation(Op0->getMemOperand()->getValue(), Overlap1,
14126 UseTBAA ? Op0->getAAInfo() : AAMDNodes()),
14127 MemoryLocation(Op1->getMemOperand()->getValue(), Overlap2,
14128 UseTBAA ? Op1->getAAInfo() : AAMDNodes()));
14129 if (AAResult == NoAlias)
14133 // Otherwise we have to assume they alias.
14137 /// Walk up chain skipping non-aliasing memory nodes,
14138 /// looking for aliasing nodes and adding them to the Aliases vector.
14139 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
14140 SmallVectorImpl<SDValue> &Aliases) {
14141 SmallVector<SDValue, 8> Chains; // List of chains to visit.
14142 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
14144 // Get alias information for node.
14145 bool IsLoad = isa<LoadSDNode>(N) && !cast<LSBaseSDNode>(N)->isVolatile();
14148 Chains.push_back(OriginalChain);
14149 unsigned Depth = 0;
14151 // Look at each chain and determine if it is an alias. If so, add it to the
14152 // aliases list. If not, then continue up the chain looking for the next
14154 while (!Chains.empty()) {
14155 SDValue Chain = Chains.pop_back_val();
14157 // For TokenFactor nodes, look at each operand and only continue up the
14158 // chain until we find two aliases. If we've seen two aliases, assume we'll
14159 // find more and revert to original chain since the xform is unlikely to be
14162 // FIXME: The depth check could be made to return the last non-aliasing
14163 // chain we found before we hit a tokenfactor rather than the original
14165 if (Depth > 6 || Aliases.size() == 2) {
14167 Aliases.push_back(OriginalChain);
14171 // Don't bother if we've been before.
14172 if (!Visited.insert(Chain.getNode()).second)
14175 switch (Chain.getOpcode()) {
14176 case ISD::EntryToken:
14177 // Entry token is ideal chain operand, but handled in FindBetterChain.
14182 // Get alias information for Chain.
14183 bool IsOpLoad = isa<LoadSDNode>(Chain.getNode()) &&
14184 !cast<LSBaseSDNode>(Chain.getNode())->isVolatile();
14186 // If chain is alias then stop here.
14187 if (!(IsLoad && IsOpLoad) &&
14188 isAlias(cast<LSBaseSDNode>(N), cast<LSBaseSDNode>(Chain.getNode()))) {
14189 Aliases.push_back(Chain);
14191 // Look further up the chain.
14192 Chains.push_back(Chain.getOperand(0));
14198 case ISD::TokenFactor:
14199 // We have to check each of the operands of the token factor for "small"
14200 // token factors, so we queue them up. Adding the operands to the queue
14201 // (stack) in reverse order maintains the original order and increases the
14202 // likelihood that getNode will find a matching token factor (CSE.)
14203 if (Chain.getNumOperands() > 16) {
14204 Aliases.push_back(Chain);
14207 for (unsigned n = Chain.getNumOperands(); n;)
14208 Chains.push_back(Chain.getOperand(--n));
14213 // For all other instructions we will just have to take what we can get.
14214 Aliases.push_back(Chain);
14219 // We need to be careful here to also search for aliases through the
14220 // value operand of a store, etc. Consider the following situation:
14222 // L1 = load Token1, %52
14223 // S1 = store Token1, L1, %51
14224 // L2 = load Token1, %52+8
14225 // S2 = store Token1, L2, %51+8
14226 // Token2 = Token(S1, S2)
14227 // L3 = load Token2, %53
14228 // S3 = store Token2, L3, %52
14229 // L4 = load Token2, %53+8
14230 // S4 = store Token2, L4, %52+8
14231 // If we search for aliases of S3 (which loads address %52), and we look
14232 // only through the chain, then we'll miss the trivial dependence on L1
14233 // (which also loads from %52). We then might change all loads and
14234 // stores to use Token1 as their chain operand, which could result in
14235 // copying %53 into %52 before copying %52 into %51 (which should
14238 // The problem is, however, that searching for such data dependencies
14239 // can become expensive, and the cost is not directly related to the
14240 // chain depth. Instead, we'll rule out such configurations here by
14241 // insisting that we've visited all chain users (except for users
14242 // of the original chain, which is not necessary). When doing this,
14243 // we need to look through nodes we don't care about (otherwise, things
14244 // like register copies will interfere with trivial cases).
14246 SmallVector<const SDNode *, 16> Worklist;
14247 for (const SDNode *N : Visited)
14248 if (N != OriginalChain.getNode())
14249 Worklist.push_back(N);
14251 while (!Worklist.empty()) {
14252 const SDNode *M = Worklist.pop_back_val();
14254 // We have already visited M, and want to make sure we've visited any uses
14255 // of M that we care about. For uses that we've not visisted, and don't
14256 // care about, queue them to the worklist.
14258 for (SDNode::use_iterator UI = M->use_begin(),
14259 UIE = M->use_end(); UI != UIE; ++UI)
14260 if (UI.getUse().getValueType() == MVT::Other &&
14261 Visited.insert(*UI).second) {
14262 if (isa<MemSDNode>(*UI)) {
14263 // We've not visited this use, and we care about it (it could have an
14264 // ordering dependency with the original node).
14266 Aliases.push_back(OriginalChain);
14270 // We've not visited this use, but we don't care about it. Mark it as
14271 // visited and enqueue it to the worklist.
14272 Worklist.push_back(*UI);
14277 /// Walk up chain skipping non-aliasing memory nodes, looking for a better chain
14278 /// (aliasing node.)
14279 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
14280 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
14282 // Accumulate all the aliases to this node.
14283 GatherAllAliases(N, OldChain, Aliases);
14285 // If no operands then chain to entry token.
14286 if (Aliases.size() == 0)
14287 return DAG.getEntryNode();
14289 // If a single operand then chain to it. We don't need to revisit it.
14290 if (Aliases.size() == 1)
14293 // Construct a custom tailored token factor.
14294 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Aliases);
14297 /// This is the entry point for the file.
14298 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
14299 CodeGenOpt::Level OptLevel) {
14300 /// This is the main entry point to this class.
14301 DAGCombiner(*this, AA, OptLevel).Run(Level);